File: | src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/include/llvm/Support/Alignment.h |
Warning: | line 85, column 47 The result of the left shift is undefined due to shifting by '255', which is greater or equal to the width of type 'uint64_t' |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | //===- CodeGenPrepare.cpp - Prepare a function for code generation --------===// | ||||
2 | // | ||||
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. | ||||
4 | // See https://llvm.org/LICENSE.txt for license information. | ||||
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception | ||||
6 | // | ||||
7 | //===----------------------------------------------------------------------===// | ||||
8 | // | ||||
9 | // This pass munges the code in the input function to better prepare it for | ||||
10 | // SelectionDAG-based code generation. This works around limitations in it's | ||||
11 | // basic-block-at-a-time approach. It should eventually be removed. | ||||
12 | // | ||||
13 | //===----------------------------------------------------------------------===// | ||||
14 | |||||
15 | #include "llvm/ADT/APInt.h" | ||||
16 | #include "llvm/ADT/ArrayRef.h" | ||||
17 | #include "llvm/ADT/DenseMap.h" | ||||
18 | #include "llvm/ADT/MapVector.h" | ||||
19 | #include "llvm/ADT/PointerIntPair.h" | ||||
20 | #include "llvm/ADT/STLExtras.h" | ||||
21 | #include "llvm/ADT/SmallPtrSet.h" | ||||
22 | #include "llvm/ADT/SmallVector.h" | ||||
23 | #include "llvm/ADT/Statistic.h" | ||||
24 | #include "llvm/Analysis/BlockFrequencyInfo.h" | ||||
25 | #include "llvm/Analysis/BranchProbabilityInfo.h" | ||||
26 | #include "llvm/Analysis/ConstantFolding.h" | ||||
27 | #include "llvm/Analysis/InstructionSimplify.h" | ||||
28 | #include "llvm/Analysis/LoopInfo.h" | ||||
29 | #include "llvm/Analysis/MemoryBuiltins.h" | ||||
30 | #include "llvm/Analysis/ProfileSummaryInfo.h" | ||||
31 | #include "llvm/Analysis/TargetLibraryInfo.h" | ||||
32 | #include "llvm/Analysis/TargetTransformInfo.h" | ||||
33 | #include "llvm/Analysis/ValueTracking.h" | ||||
34 | #include "llvm/Analysis/VectorUtils.h" | ||||
35 | #include "llvm/CodeGen/Analysis.h" | ||||
36 | #include "llvm/CodeGen/ISDOpcodes.h" | ||||
37 | #include "llvm/CodeGen/SelectionDAGNodes.h" | ||||
38 | #include "llvm/CodeGen/TargetLowering.h" | ||||
39 | #include "llvm/CodeGen/TargetPassConfig.h" | ||||
40 | #include "llvm/CodeGen/TargetSubtargetInfo.h" | ||||
41 | #include "llvm/CodeGen/ValueTypes.h" | ||||
42 | #include "llvm/Config/llvm-config.h" | ||||
43 | #include "llvm/IR/Argument.h" | ||||
44 | #include "llvm/IR/Attributes.h" | ||||
45 | #include "llvm/IR/BasicBlock.h" | ||||
46 | #include "llvm/IR/Constant.h" | ||||
47 | #include "llvm/IR/Constants.h" | ||||
48 | #include "llvm/IR/DataLayout.h" | ||||
49 | #include "llvm/IR/DebugInfo.h" | ||||
50 | #include "llvm/IR/DerivedTypes.h" | ||||
51 | #include "llvm/IR/Dominators.h" | ||||
52 | #include "llvm/IR/Function.h" | ||||
53 | #include "llvm/IR/GetElementPtrTypeIterator.h" | ||||
54 | #include "llvm/IR/GlobalValue.h" | ||||
55 | #include "llvm/IR/GlobalVariable.h" | ||||
56 | #include "llvm/IR/IRBuilder.h" | ||||
57 | #include "llvm/IR/InlineAsm.h" | ||||
58 | #include "llvm/IR/InstrTypes.h" | ||||
59 | #include "llvm/IR/Instruction.h" | ||||
60 | #include "llvm/IR/Instructions.h" | ||||
61 | #include "llvm/IR/IntrinsicInst.h" | ||||
62 | #include "llvm/IR/Intrinsics.h" | ||||
63 | #include "llvm/IR/IntrinsicsAArch64.h" | ||||
64 | #include "llvm/IR/LLVMContext.h" | ||||
65 | #include "llvm/IR/MDBuilder.h" | ||||
66 | #include "llvm/IR/Module.h" | ||||
67 | #include "llvm/IR/Operator.h" | ||||
68 | #include "llvm/IR/PatternMatch.h" | ||||
69 | #include "llvm/IR/Statepoint.h" | ||||
70 | #include "llvm/IR/Type.h" | ||||
71 | #include "llvm/IR/Use.h" | ||||
72 | #include "llvm/IR/User.h" | ||||
73 | #include "llvm/IR/Value.h" | ||||
74 | #include "llvm/IR/ValueHandle.h" | ||||
75 | #include "llvm/IR/ValueMap.h" | ||||
76 | #include "llvm/InitializePasses.h" | ||||
77 | #include "llvm/Pass.h" | ||||
78 | #include "llvm/Support/BlockFrequency.h" | ||||
79 | #include "llvm/Support/BranchProbability.h" | ||||
80 | #include "llvm/Support/Casting.h" | ||||
81 | #include "llvm/Support/CommandLine.h" | ||||
82 | #include "llvm/Support/Compiler.h" | ||||
83 | #include "llvm/Support/Debug.h" | ||||
84 | #include "llvm/Support/ErrorHandling.h" | ||||
85 | #include "llvm/Support/MachineValueType.h" | ||||
86 | #include "llvm/Support/MathExtras.h" | ||||
87 | #include "llvm/Support/raw_ostream.h" | ||||
88 | #include "llvm/Target/TargetMachine.h" | ||||
89 | #include "llvm/Target/TargetOptions.h" | ||||
90 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" | ||||
91 | #include "llvm/Transforms/Utils/BypassSlowDivision.h" | ||||
92 | #include "llvm/Transforms/Utils/Local.h" | ||||
93 | #include "llvm/Transforms/Utils/SimplifyLibCalls.h" | ||||
94 | #include "llvm/Transforms/Utils/SizeOpts.h" | ||||
95 | #include <algorithm> | ||||
96 | #include <cassert> | ||||
97 | #include <cstdint> | ||||
98 | #include <iterator> | ||||
99 | #include <limits> | ||||
100 | #include <memory> | ||||
101 | #include <utility> | ||||
102 | #include <vector> | ||||
103 | |||||
104 | using namespace llvm; | ||||
105 | using namespace llvm::PatternMatch; | ||||
106 | |||||
107 | #define DEBUG_TYPE"codegenprepare" "codegenprepare" | ||||
108 | |||||
109 | STATISTIC(NumBlocksElim, "Number of blocks eliminated")static llvm::Statistic NumBlocksElim = {"codegenprepare", "NumBlocksElim" , "Number of blocks eliminated"}; | ||||
110 | STATISTIC(NumPHIsElim, "Number of trivial PHIs eliminated")static llvm::Statistic NumPHIsElim = {"codegenprepare", "NumPHIsElim" , "Number of trivial PHIs eliminated"}; | ||||
111 | STATISTIC(NumGEPsElim, "Number of GEPs converted to casts")static llvm::Statistic NumGEPsElim = {"codegenprepare", "NumGEPsElim" , "Number of GEPs converted to casts"}; | ||||
112 | STATISTIC(NumCmpUses, "Number of uses of Cmp expressions replaced with uses of "static llvm::Statistic NumCmpUses = {"codegenprepare", "NumCmpUses" , "Number of uses of Cmp expressions replaced with uses of " "sunken Cmps" } | ||||
113 | "sunken Cmps")static llvm::Statistic NumCmpUses = {"codegenprepare", "NumCmpUses" , "Number of uses of Cmp expressions replaced with uses of " "sunken Cmps" }; | ||||
114 | STATISTIC(NumCastUses, "Number of uses of Cast expressions replaced with uses "static llvm::Statistic NumCastUses = {"codegenprepare", "NumCastUses" , "Number of uses of Cast expressions replaced with uses " "of sunken Casts" } | ||||
115 | "of sunken Casts")static llvm::Statistic NumCastUses = {"codegenprepare", "NumCastUses" , "Number of uses of Cast expressions replaced with uses " "of sunken Casts" }; | ||||
116 | STATISTIC(NumMemoryInsts, "Number of memory instructions whose address "static llvm::Statistic NumMemoryInsts = {"codegenprepare", "NumMemoryInsts" , "Number of memory instructions whose address " "computations were sunk" } | ||||
117 | "computations were sunk")static llvm::Statistic NumMemoryInsts = {"codegenprepare", "NumMemoryInsts" , "Number of memory instructions whose address " "computations were sunk" }; | ||||
118 | STATISTIC(NumMemoryInstsPhiCreated,static llvm::Statistic NumMemoryInstsPhiCreated = {"codegenprepare" , "NumMemoryInstsPhiCreated", "Number of phis created when address " "computations were sunk to memory instructions"} | ||||
119 | "Number of phis created when address "static llvm::Statistic NumMemoryInstsPhiCreated = {"codegenprepare" , "NumMemoryInstsPhiCreated", "Number of phis created when address " "computations were sunk to memory instructions"} | ||||
120 | "computations were sunk to memory instructions")static llvm::Statistic NumMemoryInstsPhiCreated = {"codegenprepare" , "NumMemoryInstsPhiCreated", "Number of phis created when address " "computations were sunk to memory instructions"}; | ||||
121 | STATISTIC(NumMemoryInstsSelectCreated,static llvm::Statistic NumMemoryInstsSelectCreated = {"codegenprepare" , "NumMemoryInstsSelectCreated", "Number of select created when address " "computations were sunk to memory instructions"} | ||||
122 | "Number of select created when address "static llvm::Statistic NumMemoryInstsSelectCreated = {"codegenprepare" , "NumMemoryInstsSelectCreated", "Number of select created when address " "computations were sunk to memory instructions"} | ||||
123 | "computations were sunk to memory instructions")static llvm::Statistic NumMemoryInstsSelectCreated = {"codegenprepare" , "NumMemoryInstsSelectCreated", "Number of select created when address " "computations were sunk to memory instructions"}; | ||||
124 | STATISTIC(NumExtsMoved, "Number of [s|z]ext instructions combined with loads")static llvm::Statistic NumExtsMoved = {"codegenprepare", "NumExtsMoved" , "Number of [s|z]ext instructions combined with loads"}; | ||||
125 | STATISTIC(NumExtUses, "Number of uses of [s|z]ext instructions optimized")static llvm::Statistic NumExtUses = {"codegenprepare", "NumExtUses" , "Number of uses of [s|z]ext instructions optimized"}; | ||||
126 | STATISTIC(NumAndsAdded,static llvm::Statistic NumAndsAdded = {"codegenprepare", "NumAndsAdded" , "Number of and mask instructions added to form ext loads"} | ||||
127 | "Number of and mask instructions added to form ext loads")static llvm::Statistic NumAndsAdded = {"codegenprepare", "NumAndsAdded" , "Number of and mask instructions added to form ext loads"}; | ||||
128 | STATISTIC(NumAndUses, "Number of uses of and mask instructions optimized")static llvm::Statistic NumAndUses = {"codegenprepare", "NumAndUses" , "Number of uses of and mask instructions optimized"}; | ||||
129 | STATISTIC(NumRetsDup, "Number of return instructions duplicated")static llvm::Statistic NumRetsDup = {"codegenprepare", "NumRetsDup" , "Number of return instructions duplicated"}; | ||||
130 | STATISTIC(NumDbgValueMoved, "Number of debug value instructions moved")static llvm::Statistic NumDbgValueMoved = {"codegenprepare", "NumDbgValueMoved" , "Number of debug value instructions moved"}; | ||||
131 | STATISTIC(NumSelectsExpanded, "Number of selects turned into branches")static llvm::Statistic NumSelectsExpanded = {"codegenprepare" , "NumSelectsExpanded", "Number of selects turned into branches" }; | ||||
132 | STATISTIC(NumStoreExtractExposed, "Number of store(extractelement) exposed")static llvm::Statistic NumStoreExtractExposed = {"codegenprepare" , "NumStoreExtractExposed", "Number of store(extractelement) exposed" }; | ||||
133 | |||||
134 | static cl::opt<bool> DisableBranchOpts( | ||||
135 | "disable-cgp-branch-opts", cl::Hidden, cl::init(false), | ||||
136 | cl::desc("Disable branch optimizations in CodeGenPrepare")); | ||||
137 | |||||
138 | static cl::opt<bool> | ||||
139 | DisableGCOpts("disable-cgp-gc-opts", cl::Hidden, cl::init(false), | ||||
140 | cl::desc("Disable GC optimizations in CodeGenPrepare")); | ||||
141 | |||||
142 | static cl::opt<bool> DisableSelectToBranch( | ||||
143 | "disable-cgp-select2branch", cl::Hidden, cl::init(false), | ||||
144 | cl::desc("Disable select to branch conversion.")); | ||||
145 | |||||
146 | static cl::opt<bool> AddrSinkUsingGEPs( | ||||
147 | "addr-sink-using-gep", cl::Hidden, cl::init(true), | ||||
148 | cl::desc("Address sinking in CGP using GEPs.")); | ||||
149 | |||||
150 | static cl::opt<bool> EnableAndCmpSinking( | ||||
151 | "enable-andcmp-sinking", cl::Hidden, cl::init(true), | ||||
152 | cl::desc("Enable sinkinig and/cmp into branches.")); | ||||
153 | |||||
154 | static cl::opt<bool> DisableStoreExtract( | ||||
155 | "disable-cgp-store-extract", cl::Hidden, cl::init(false), | ||||
156 | cl::desc("Disable store(extract) optimizations in CodeGenPrepare")); | ||||
157 | |||||
158 | static cl::opt<bool> StressStoreExtract( | ||||
159 | "stress-cgp-store-extract", cl::Hidden, cl::init(false), | ||||
160 | cl::desc("Stress test store(extract) optimizations in CodeGenPrepare")); | ||||
161 | |||||
162 | static cl::opt<bool> DisableExtLdPromotion( | ||||
163 | "disable-cgp-ext-ld-promotion", cl::Hidden, cl::init(false), | ||||
164 | cl::desc("Disable ext(promotable(ld)) -> promoted(ext(ld)) optimization in " | ||||
165 | "CodeGenPrepare")); | ||||
166 | |||||
167 | static cl::opt<bool> StressExtLdPromotion( | ||||
168 | "stress-cgp-ext-ld-promotion", cl::Hidden, cl::init(false), | ||||
169 | cl::desc("Stress test ext(promotable(ld)) -> promoted(ext(ld)) " | ||||
170 | "optimization in CodeGenPrepare")); | ||||
171 | |||||
172 | static cl::opt<bool> DisablePreheaderProtect( | ||||
173 | "disable-preheader-prot", cl::Hidden, cl::init(false), | ||||
174 | cl::desc("Disable protection against removing loop preheaders")); | ||||
175 | |||||
176 | static cl::opt<bool> ProfileGuidedSectionPrefix( | ||||
177 | "profile-guided-section-prefix", cl::Hidden, cl::init(true), cl::ZeroOrMore, | ||||
178 | cl::desc("Use profile info to add section prefix for hot/cold functions")); | ||||
179 | |||||
180 | static cl::opt<bool> ProfileUnknownInSpecialSection( | ||||
181 | "profile-unknown-in-special-section", cl::Hidden, cl::init(false), | ||||
182 | cl::ZeroOrMore, | ||||
183 | cl::desc("In profiling mode like sampleFDO, if a function doesn't have " | ||||
184 | "profile, we cannot tell the function is cold for sure because " | ||||
185 | "it may be a function newly added without ever being sampled. " | ||||
186 | "With the flag enabled, compiler can put such profile unknown " | ||||
187 | "functions into a special section, so runtime system can choose " | ||||
188 | "to handle it in a different way than .text section, to save " | ||||
189 | "RAM for example. ")); | ||||
190 | |||||
191 | static cl::opt<unsigned> FreqRatioToSkipMerge( | ||||
192 | "cgp-freq-ratio-to-skip-merge", cl::Hidden, cl::init(2), | ||||
193 | cl::desc("Skip merging empty blocks if (frequency of empty block) / " | ||||
194 | "(frequency of destination block) is greater than this ratio")); | ||||
195 | |||||
196 | static cl::opt<bool> ForceSplitStore( | ||||
197 | "force-split-store", cl::Hidden, cl::init(false), | ||||
198 | cl::desc("Force store splitting no matter what the target query says.")); | ||||
199 | |||||
200 | static cl::opt<bool> | ||||
201 | EnableTypePromotionMerge("cgp-type-promotion-merge", cl::Hidden, | ||||
202 | cl::desc("Enable merging of redundant sexts when one is dominating" | ||||
203 | " the other."), cl::init(true)); | ||||
204 | |||||
205 | static cl::opt<bool> DisableComplexAddrModes( | ||||
206 | "disable-complex-addr-modes", cl::Hidden, cl::init(false), | ||||
207 | cl::desc("Disables combining addressing modes with different parts " | ||||
208 | "in optimizeMemoryInst.")); | ||||
209 | |||||
210 | static cl::opt<bool> | ||||
211 | AddrSinkNewPhis("addr-sink-new-phis", cl::Hidden, cl::init(false), | ||||
212 | cl::desc("Allow creation of Phis in Address sinking.")); | ||||
213 | |||||
214 | static cl::opt<bool> | ||||
215 | AddrSinkNewSelects("addr-sink-new-select", cl::Hidden, cl::init(true), | ||||
216 | cl::desc("Allow creation of selects in Address sinking.")); | ||||
217 | |||||
218 | static cl::opt<bool> AddrSinkCombineBaseReg( | ||||
219 | "addr-sink-combine-base-reg", cl::Hidden, cl::init(true), | ||||
220 | cl::desc("Allow combining of BaseReg field in Address sinking.")); | ||||
221 | |||||
222 | static cl::opt<bool> AddrSinkCombineBaseGV( | ||||
223 | "addr-sink-combine-base-gv", cl::Hidden, cl::init(true), | ||||
224 | cl::desc("Allow combining of BaseGV field in Address sinking.")); | ||||
225 | |||||
226 | static cl::opt<bool> AddrSinkCombineBaseOffs( | ||||
227 | "addr-sink-combine-base-offs", cl::Hidden, cl::init(true), | ||||
228 | cl::desc("Allow combining of BaseOffs field in Address sinking.")); | ||||
229 | |||||
230 | static cl::opt<bool> AddrSinkCombineScaledReg( | ||||
231 | "addr-sink-combine-scaled-reg", cl::Hidden, cl::init(true), | ||||
232 | cl::desc("Allow combining of ScaledReg field in Address sinking.")); | ||||
233 | |||||
234 | static cl::opt<bool> | ||||
235 | EnableGEPOffsetSplit("cgp-split-large-offset-gep", cl::Hidden, | ||||
236 | cl::init(true), | ||||
237 | cl::desc("Enable splitting large offset of GEP.")); | ||||
238 | |||||
239 | static cl::opt<bool> EnableICMP_EQToICMP_ST( | ||||
240 | "cgp-icmp-eq2icmp-st", cl::Hidden, cl::init(false), | ||||
241 | cl::desc("Enable ICMP_EQ to ICMP_S(L|G)T conversion.")); | ||||
242 | |||||
243 | static cl::opt<bool> | ||||
244 | VerifyBFIUpdates("cgp-verify-bfi-updates", cl::Hidden, cl::init(false), | ||||
245 | cl::desc("Enable BFI update verification for " | ||||
246 | "CodeGenPrepare.")); | ||||
247 | |||||
248 | static cl::opt<bool> OptimizePhiTypes( | ||||
249 | "cgp-optimize-phi-types", cl::Hidden, cl::init(false), | ||||
250 | cl::desc("Enable converting phi types in CodeGenPrepare")); | ||||
251 | |||||
252 | namespace { | ||||
253 | |||||
254 | enum ExtType { | ||||
255 | ZeroExtension, // Zero extension has been seen. | ||||
256 | SignExtension, // Sign extension has been seen. | ||||
257 | BothExtension // This extension type is used if we saw sext after | ||||
258 | // ZeroExtension had been set, or if we saw zext after | ||||
259 | // SignExtension had been set. It makes the type | ||||
260 | // information of a promoted instruction invalid. | ||||
261 | }; | ||||
262 | |||||
263 | using SetOfInstrs = SmallPtrSet<Instruction *, 16>; | ||||
264 | using TypeIsSExt = PointerIntPair<Type *, 2, ExtType>; | ||||
265 | using InstrToOrigTy = DenseMap<Instruction *, TypeIsSExt>; | ||||
266 | using SExts = SmallVector<Instruction *, 16>; | ||||
267 | using ValueToSExts = DenseMap<Value *, SExts>; | ||||
268 | |||||
269 | class TypePromotionTransaction; | ||||
270 | |||||
271 | class CodeGenPrepare : public FunctionPass { | ||||
272 | const TargetMachine *TM = nullptr; | ||||
273 | const TargetSubtargetInfo *SubtargetInfo; | ||||
274 | const TargetLowering *TLI = nullptr; | ||||
275 | const TargetRegisterInfo *TRI; | ||||
276 | const TargetTransformInfo *TTI = nullptr; | ||||
277 | const TargetLibraryInfo *TLInfo; | ||||
278 | const LoopInfo *LI; | ||||
279 | std::unique_ptr<BlockFrequencyInfo> BFI; | ||||
280 | std::unique_ptr<BranchProbabilityInfo> BPI; | ||||
281 | ProfileSummaryInfo *PSI; | ||||
282 | |||||
283 | /// As we scan instructions optimizing them, this is the next instruction | ||||
284 | /// to optimize. Transforms that can invalidate this should update it. | ||||
285 | BasicBlock::iterator CurInstIterator; | ||||
286 | |||||
287 | /// Keeps track of non-local addresses that have been sunk into a block. | ||||
288 | /// This allows us to avoid inserting duplicate code for blocks with | ||||
289 | /// multiple load/stores of the same address. The usage of WeakTrackingVH | ||||
290 | /// enables SunkAddrs to be treated as a cache whose entries can be | ||||
291 | /// invalidated if a sunken address computation has been erased. | ||||
292 | ValueMap<Value*, WeakTrackingVH> SunkAddrs; | ||||
293 | |||||
294 | /// Keeps track of all instructions inserted for the current function. | ||||
295 | SetOfInstrs InsertedInsts; | ||||
296 | |||||
297 | /// Keeps track of the type of the related instruction before their | ||||
298 | /// promotion for the current function. | ||||
299 | InstrToOrigTy PromotedInsts; | ||||
300 | |||||
301 | /// Keep track of instructions removed during promotion. | ||||
302 | SetOfInstrs RemovedInsts; | ||||
303 | |||||
304 | /// Keep track of sext chains based on their initial value. | ||||
305 | DenseMap<Value *, Instruction *> SeenChainsForSExt; | ||||
306 | |||||
307 | /// Keep track of GEPs accessing the same data structures such as structs or | ||||
308 | /// arrays that are candidates to be split later because of their large | ||||
309 | /// size. | ||||
310 | MapVector< | ||||
311 | AssertingVH<Value>, | ||||
312 | SmallVector<std::pair<AssertingVH<GetElementPtrInst>, int64_t>, 32>> | ||||
313 | LargeOffsetGEPMap; | ||||
314 | |||||
315 | /// Keep track of new GEP base after splitting the GEPs having large offset. | ||||
316 | SmallSet<AssertingVH<Value>, 2> NewGEPBases; | ||||
317 | |||||
318 | /// Map serial numbers to Large offset GEPs. | ||||
319 | DenseMap<AssertingVH<GetElementPtrInst>, int> LargeOffsetGEPID; | ||||
320 | |||||
321 | /// Keep track of SExt promoted. | ||||
322 | ValueToSExts ValToSExtendedUses; | ||||
323 | |||||
324 | /// True if the function has the OptSize attribute. | ||||
325 | bool OptSize; | ||||
326 | |||||
327 | /// DataLayout for the Function being processed. | ||||
328 | const DataLayout *DL = nullptr; | ||||
329 | |||||
330 | /// Building the dominator tree can be expensive, so we only build it | ||||
331 | /// lazily and update it when required. | ||||
332 | std::unique_ptr<DominatorTree> DT; | ||||
333 | |||||
334 | public: | ||||
335 | static char ID; // Pass identification, replacement for typeid | ||||
336 | |||||
337 | CodeGenPrepare() : FunctionPass(ID) { | ||||
338 | initializeCodeGenPreparePass(*PassRegistry::getPassRegistry()); | ||||
339 | } | ||||
340 | |||||
341 | bool runOnFunction(Function &F) override; | ||||
342 | |||||
343 | StringRef getPassName() const override { return "CodeGen Prepare"; } | ||||
344 | |||||
345 | void getAnalysisUsage(AnalysisUsage &AU) const override { | ||||
346 | // FIXME: When we can selectively preserve passes, preserve the domtree. | ||||
347 | AU.addRequired<ProfileSummaryInfoWrapperPass>(); | ||||
348 | AU.addRequired<TargetLibraryInfoWrapperPass>(); | ||||
349 | AU.addRequired<TargetPassConfig>(); | ||||
350 | AU.addRequired<TargetTransformInfoWrapperPass>(); | ||||
351 | AU.addRequired<LoopInfoWrapperPass>(); | ||||
352 | } | ||||
353 | |||||
354 | private: | ||||
355 | template <typename F> | ||||
356 | void resetIteratorIfInvalidatedWhileCalling(BasicBlock *BB, F f) { | ||||
357 | // Substituting can cause recursive simplifications, which can invalidate | ||||
358 | // our iterator. Use a WeakTrackingVH to hold onto it in case this | ||||
359 | // happens. | ||||
360 | Value *CurValue = &*CurInstIterator; | ||||
361 | WeakTrackingVH IterHandle(CurValue); | ||||
362 | |||||
363 | f(); | ||||
364 | |||||
365 | // If the iterator instruction was recursively deleted, start over at the | ||||
366 | // start of the block. | ||||
367 | if (IterHandle != CurValue) { | ||||
368 | CurInstIterator = BB->begin(); | ||||
369 | SunkAddrs.clear(); | ||||
370 | } | ||||
371 | } | ||||
372 | |||||
373 | // Get the DominatorTree, building if necessary. | ||||
374 | DominatorTree &getDT(Function &F) { | ||||
375 | if (!DT) | ||||
376 | DT = std::make_unique<DominatorTree>(F); | ||||
377 | return *DT; | ||||
378 | } | ||||
379 | |||||
380 | void removeAllAssertingVHReferences(Value *V); | ||||
381 | bool eliminateAssumptions(Function &F); | ||||
382 | bool eliminateFallThrough(Function &F); | ||||
383 | bool eliminateMostlyEmptyBlocks(Function &F); | ||||
384 | BasicBlock *findDestBlockOfMergeableEmptyBlock(BasicBlock *BB); | ||||
385 | bool canMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const; | ||||
386 | void eliminateMostlyEmptyBlock(BasicBlock *BB); | ||||
387 | bool isMergingEmptyBlockProfitable(BasicBlock *BB, BasicBlock *DestBB, | ||||
388 | bool isPreheader); | ||||
389 | bool makeBitReverse(Instruction &I); | ||||
390 | bool optimizeBlock(BasicBlock &BB, bool &ModifiedDT); | ||||
391 | bool optimizeInst(Instruction *I, bool &ModifiedDT); | ||||
392 | bool optimizeMemoryInst(Instruction *MemoryInst, Value *Addr, | ||||
393 | Type *AccessTy, unsigned AddrSpace); | ||||
394 | bool optimizeGatherScatterInst(Instruction *MemoryInst, Value *Ptr); | ||||
395 | bool optimizeInlineAsmInst(CallInst *CS); | ||||
396 | bool optimizeCallInst(CallInst *CI, bool &ModifiedDT); | ||||
397 | bool optimizeExt(Instruction *&I); | ||||
398 | bool optimizeExtUses(Instruction *I); | ||||
399 | bool optimizeLoadExt(LoadInst *Load); | ||||
400 | bool optimizeShiftInst(BinaryOperator *BO); | ||||
401 | bool optimizeFunnelShift(IntrinsicInst *Fsh); | ||||
402 | bool optimizeSelectInst(SelectInst *SI); | ||||
403 | bool optimizeShuffleVectorInst(ShuffleVectorInst *SVI); | ||||
404 | bool optimizeSwitchInst(SwitchInst *SI); | ||||
405 | bool optimizeExtractElementInst(Instruction *Inst); | ||||
406 | bool dupRetToEnableTailCallOpts(BasicBlock *BB, bool &ModifiedDT); | ||||
407 | bool fixupDbgValue(Instruction *I); | ||||
408 | bool placeDbgValues(Function &F); | ||||
409 | bool placePseudoProbes(Function &F); | ||||
410 | bool canFormExtLd(const SmallVectorImpl<Instruction *> &MovedExts, | ||||
411 | LoadInst *&LI, Instruction *&Inst, bool HasPromoted); | ||||
412 | bool tryToPromoteExts(TypePromotionTransaction &TPT, | ||||
413 | const SmallVectorImpl<Instruction *> &Exts, | ||||
414 | SmallVectorImpl<Instruction *> &ProfitablyMovedExts, | ||||
415 | unsigned CreatedInstsCost = 0); | ||||
416 | bool mergeSExts(Function &F); | ||||
417 | bool splitLargeGEPOffsets(); | ||||
418 | bool optimizePhiType(PHINode *Inst, SmallPtrSetImpl<PHINode *> &Visited, | ||||
419 | SmallPtrSetImpl<Instruction *> &DeletedInstrs); | ||||
420 | bool optimizePhiTypes(Function &F); | ||||
421 | bool performAddressTypePromotion( | ||||
422 | Instruction *&Inst, | ||||
423 | bool AllowPromotionWithoutCommonHeader, | ||||
424 | bool HasPromoted, TypePromotionTransaction &TPT, | ||||
425 | SmallVectorImpl<Instruction *> &SpeculativelyMovedExts); | ||||
426 | bool splitBranchCondition(Function &F, bool &ModifiedDT); | ||||
427 | bool simplifyOffsetableRelocate(GCStatepointInst &I); | ||||
428 | |||||
429 | bool tryToSinkFreeOperands(Instruction *I); | ||||
430 | bool replaceMathCmpWithIntrinsic(BinaryOperator *BO, Value *Arg0, | ||||
431 | Value *Arg1, CmpInst *Cmp, | ||||
432 | Intrinsic::ID IID); | ||||
433 | bool optimizeCmp(CmpInst *Cmp, bool &ModifiedDT); | ||||
434 | bool combineToUSubWithOverflow(CmpInst *Cmp, bool &ModifiedDT); | ||||
435 | bool combineToUAddWithOverflow(CmpInst *Cmp, bool &ModifiedDT); | ||||
436 | void verifyBFIUpdates(Function &F); | ||||
437 | }; | ||||
438 | |||||
439 | } // end anonymous namespace | ||||
440 | |||||
441 | char CodeGenPrepare::ID = 0; | ||||
442 | |||||
443 | INITIALIZE_PASS_BEGIN(CodeGenPrepare, DEBUG_TYPE,static void *initializeCodeGenPreparePassOnce(PassRegistry & Registry) { | ||||
444 | "Optimize for code generation", false, false)static void *initializeCodeGenPreparePassOnce(PassRegistry & Registry) { | ||||
445 | INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)initializeLoopInfoWrapperPassPass(Registry); | ||||
446 | INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass)initializeProfileSummaryInfoWrapperPassPass(Registry); | ||||
447 | INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)initializeTargetLibraryInfoWrapperPassPass(Registry); | ||||
448 | INITIALIZE_PASS_DEPENDENCY(TargetPassConfig)initializeTargetPassConfigPass(Registry); | ||||
449 | INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)initializeTargetTransformInfoWrapperPassPass(Registry); | ||||
450 | INITIALIZE_PASS_END(CodeGenPrepare, DEBUG_TYPE,PassInfo *PI = new PassInfo( "Optimize for code generation", "codegenprepare" , &CodeGenPrepare::ID, PassInfo::NormalCtor_t(callDefaultCtor <CodeGenPrepare>), false, false); Registry.registerPass (*PI, true); return PI; } static llvm::once_flag InitializeCodeGenPreparePassFlag ; void llvm::initializeCodeGenPreparePass(PassRegistry &Registry ) { llvm::call_once(InitializeCodeGenPreparePassFlag, initializeCodeGenPreparePassOnce , std::ref(Registry)); } | ||||
451 | "Optimize for code generation", false, false)PassInfo *PI = new PassInfo( "Optimize for code generation", "codegenprepare" , &CodeGenPrepare::ID, PassInfo::NormalCtor_t(callDefaultCtor <CodeGenPrepare>), false, false); Registry.registerPass (*PI, true); return PI; } static llvm::once_flag InitializeCodeGenPreparePassFlag ; void llvm::initializeCodeGenPreparePass(PassRegistry &Registry ) { llvm::call_once(InitializeCodeGenPreparePassFlag, initializeCodeGenPreparePassOnce , std::ref(Registry)); } | ||||
452 | |||||
453 | FunctionPass *llvm::createCodeGenPreparePass() { return new CodeGenPrepare(); } | ||||
454 | |||||
455 | bool CodeGenPrepare::runOnFunction(Function &F) { | ||||
456 | if (skipFunction(F)) | ||||
457 | return false; | ||||
458 | |||||
459 | DL = &F.getParent()->getDataLayout(); | ||||
460 | |||||
461 | bool EverMadeChange = false; | ||||
462 | // Clear per function information. | ||||
463 | InsertedInsts.clear(); | ||||
464 | PromotedInsts.clear(); | ||||
465 | |||||
466 | TM = &getAnalysis<TargetPassConfig>().getTM<TargetMachine>(); | ||||
467 | SubtargetInfo = TM->getSubtargetImpl(F); | ||||
468 | TLI = SubtargetInfo->getTargetLowering(); | ||||
469 | TRI = SubtargetInfo->getRegisterInfo(); | ||||
470 | TLInfo = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F); | ||||
471 | TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); | ||||
472 | LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); | ||||
473 | BPI.reset(new BranchProbabilityInfo(F, *LI)); | ||||
474 | BFI.reset(new BlockFrequencyInfo(F, *BPI, *LI)); | ||||
475 | PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI(); | ||||
476 | OptSize = F.hasOptSize(); | ||||
477 | if (ProfileGuidedSectionPrefix) { | ||||
478 | // The hot attribute overwrites profile count based hotness while profile | ||||
479 | // counts based hotness overwrite the cold attribute. | ||||
480 | // This is a conservative behabvior. | ||||
481 | if (F.hasFnAttribute(Attribute::Hot) || | ||||
482 | PSI->isFunctionHotInCallGraph(&F, *BFI)) | ||||
483 | F.setSectionPrefix("hot"); | ||||
484 | // If PSI shows this function is not hot, we will placed the function | ||||
485 | // into unlikely section if (1) PSI shows this is a cold function, or | ||||
486 | // (2) the function has a attribute of cold. | ||||
487 | else if (PSI->isFunctionColdInCallGraph(&F, *BFI) || | ||||
488 | F.hasFnAttribute(Attribute::Cold)) | ||||
489 | F.setSectionPrefix("unlikely"); | ||||
490 | else if (ProfileUnknownInSpecialSection && PSI->hasPartialSampleProfile() && | ||||
491 | PSI->isFunctionHotnessUnknown(F)) | ||||
492 | F.setSectionPrefix("unknown"); | ||||
493 | } | ||||
494 | |||||
495 | /// This optimization identifies DIV instructions that can be | ||||
496 | /// profitably bypassed and carried out with a shorter, faster divide. | ||||
497 | if (!OptSize && !PSI->hasHugeWorkingSetSize() && TLI->isSlowDivBypassed()) { | ||||
498 | const DenseMap<unsigned int, unsigned int> &BypassWidths = | ||||
499 | TLI->getBypassSlowDivWidths(); | ||||
500 | BasicBlock* BB = &*F.begin(); | ||||
501 | while (BB != nullptr) { | ||||
502 | // bypassSlowDivision may create new BBs, but we don't want to reapply the | ||||
503 | // optimization to those blocks. | ||||
504 | BasicBlock* Next = BB->getNextNode(); | ||||
505 | // F.hasOptSize is already checked in the outer if statement. | ||||
506 | if (!llvm::shouldOptimizeForSize(BB, PSI, BFI.get())) | ||||
507 | EverMadeChange |= bypassSlowDivision(BB, BypassWidths); | ||||
508 | BB = Next; | ||||
509 | } | ||||
510 | } | ||||
511 | |||||
512 | // Get rid of @llvm.assume builtins before attempting to eliminate empty | ||||
513 | // blocks, since there might be blocks that only contain @llvm.assume calls | ||||
514 | // (plus arguments that we can get rid of). | ||||
515 | EverMadeChange |= eliminateAssumptions(F); | ||||
516 | |||||
517 | // Eliminate blocks that contain only PHI nodes and an | ||||
518 | // unconditional branch. | ||||
519 | EverMadeChange |= eliminateMostlyEmptyBlocks(F); | ||||
520 | |||||
521 | bool ModifiedDT = false; | ||||
522 | if (!DisableBranchOpts) | ||||
523 | EverMadeChange |= splitBranchCondition(F, ModifiedDT); | ||||
524 | |||||
525 | // Split some critical edges where one of the sources is an indirect branch, | ||||
526 | // to help generate sane code for PHIs involving such edges. | ||||
527 | EverMadeChange |= SplitIndirectBrCriticalEdges(F); | ||||
528 | |||||
529 | bool MadeChange = true; | ||||
530 | while (MadeChange) { | ||||
531 | MadeChange = false; | ||||
532 | DT.reset(); | ||||
533 | for (Function::iterator I = F.begin(); I != F.end(); ) { | ||||
534 | BasicBlock *BB = &*I++; | ||||
535 | bool ModifiedDTOnIteration = false; | ||||
536 | MadeChange |= optimizeBlock(*BB, ModifiedDTOnIteration); | ||||
537 | |||||
538 | // Restart BB iteration if the dominator tree of the Function was changed | ||||
539 | if (ModifiedDTOnIteration) | ||||
540 | break; | ||||
541 | } | ||||
542 | if (EnableTypePromotionMerge && !ValToSExtendedUses.empty()) | ||||
543 | MadeChange |= mergeSExts(F); | ||||
544 | if (!LargeOffsetGEPMap.empty()) | ||||
545 | MadeChange |= splitLargeGEPOffsets(); | ||||
546 | MadeChange |= optimizePhiTypes(F); | ||||
547 | |||||
548 | if (MadeChange) | ||||
549 | eliminateFallThrough(F); | ||||
550 | |||||
551 | // Really free removed instructions during promotion. | ||||
552 | for (Instruction *I : RemovedInsts) | ||||
553 | I->deleteValue(); | ||||
554 | |||||
555 | EverMadeChange |= MadeChange; | ||||
556 | SeenChainsForSExt.clear(); | ||||
557 | ValToSExtendedUses.clear(); | ||||
558 | RemovedInsts.clear(); | ||||
559 | LargeOffsetGEPMap.clear(); | ||||
560 | LargeOffsetGEPID.clear(); | ||||
561 | } | ||||
562 | |||||
563 | NewGEPBases.clear(); | ||||
564 | SunkAddrs.clear(); | ||||
565 | |||||
566 | if (!DisableBranchOpts) { | ||||
567 | MadeChange = false; | ||||
568 | // Use a set vector to get deterministic iteration order. The order the | ||||
569 | // blocks are removed may affect whether or not PHI nodes in successors | ||||
570 | // are removed. | ||||
571 | SmallSetVector<BasicBlock*, 8> WorkList; | ||||
572 | for (BasicBlock &BB : F) { | ||||
573 | SmallVector<BasicBlock *, 2> Successors(successors(&BB)); | ||||
574 | MadeChange |= ConstantFoldTerminator(&BB, true); | ||||
575 | if (!MadeChange) continue; | ||||
576 | |||||
577 | for (BasicBlock *Succ : Successors) | ||||
578 | if (pred_empty(Succ)) | ||||
579 | WorkList.insert(Succ); | ||||
580 | } | ||||
581 | |||||
582 | // Delete the dead blocks and any of their dead successors. | ||||
583 | MadeChange |= !WorkList.empty(); | ||||
584 | while (!WorkList.empty()) { | ||||
585 | BasicBlock *BB = WorkList.pop_back_val(); | ||||
586 | SmallVector<BasicBlock*, 2> Successors(successors(BB)); | ||||
587 | |||||
588 | DeleteDeadBlock(BB); | ||||
589 | |||||
590 | for (BasicBlock *Succ : Successors) | ||||
591 | if (pred_empty(Succ)) | ||||
592 | WorkList.insert(Succ); | ||||
593 | } | ||||
594 | |||||
595 | // Merge pairs of basic blocks with unconditional branches, connected by | ||||
596 | // a single edge. | ||||
597 | if (EverMadeChange || MadeChange) | ||||
598 | MadeChange |= eliminateFallThrough(F); | ||||
599 | |||||
600 | EverMadeChange |= MadeChange; | ||||
601 | } | ||||
602 | |||||
603 | if (!DisableGCOpts) { | ||||
604 | SmallVector<GCStatepointInst *, 2> Statepoints; | ||||
605 | for (BasicBlock &BB : F) | ||||
606 | for (Instruction &I : BB) | ||||
607 | if (auto *SP = dyn_cast<GCStatepointInst>(&I)) | ||||
608 | Statepoints.push_back(SP); | ||||
609 | for (auto &I : Statepoints) | ||||
610 | EverMadeChange |= simplifyOffsetableRelocate(*I); | ||||
611 | } | ||||
612 | |||||
613 | // Do this last to clean up use-before-def scenarios introduced by other | ||||
614 | // preparatory transforms. | ||||
615 | EverMadeChange |= placeDbgValues(F); | ||||
616 | EverMadeChange |= placePseudoProbes(F); | ||||
617 | |||||
618 | #ifndef NDEBUG1 | ||||
619 | if (VerifyBFIUpdates) | ||||
620 | verifyBFIUpdates(F); | ||||
621 | #endif | ||||
622 | |||||
623 | return EverMadeChange; | ||||
624 | } | ||||
625 | |||||
626 | bool CodeGenPrepare::eliminateAssumptions(Function &F) { | ||||
627 | bool MadeChange = false; | ||||
628 | for (BasicBlock &BB : F) { | ||||
629 | CurInstIterator = BB.begin(); | ||||
630 | while (CurInstIterator != BB.end()) { | ||||
631 | Instruction *I = &*(CurInstIterator++); | ||||
632 | if (auto *Assume = dyn_cast<AssumeInst>(I)) { | ||||
633 | MadeChange = true; | ||||
634 | Value *Operand = Assume->getOperand(0); | ||||
635 | Assume->eraseFromParent(); | ||||
636 | |||||
637 | resetIteratorIfInvalidatedWhileCalling(&BB, [&]() { | ||||
638 | RecursivelyDeleteTriviallyDeadInstructions(Operand, TLInfo, nullptr); | ||||
639 | }); | ||||
640 | } | ||||
641 | } | ||||
642 | } | ||||
643 | return MadeChange; | ||||
644 | } | ||||
645 | |||||
646 | /// An instruction is about to be deleted, so remove all references to it in our | ||||
647 | /// GEP-tracking data strcutures. | ||||
648 | void CodeGenPrepare::removeAllAssertingVHReferences(Value *V) { | ||||
649 | LargeOffsetGEPMap.erase(V); | ||||
650 | NewGEPBases.erase(V); | ||||
651 | |||||
652 | auto GEP = dyn_cast<GetElementPtrInst>(V); | ||||
653 | if (!GEP) | ||||
654 | return; | ||||
655 | |||||
656 | LargeOffsetGEPID.erase(GEP); | ||||
657 | |||||
658 | auto VecI = LargeOffsetGEPMap.find(GEP->getPointerOperand()); | ||||
659 | if (VecI == LargeOffsetGEPMap.end()) | ||||
660 | return; | ||||
661 | |||||
662 | auto &GEPVector = VecI->second; | ||||
663 | const auto &I = | ||||
664 | llvm::find_if(GEPVector, [=](auto &Elt) { return Elt.first == GEP; }); | ||||
665 | if (I == GEPVector.end()) | ||||
666 | return; | ||||
667 | |||||
668 | GEPVector.erase(I); | ||||
669 | if (GEPVector.empty()) | ||||
670 | LargeOffsetGEPMap.erase(VecI); | ||||
671 | } | ||||
672 | |||||
673 | // Verify BFI has been updated correctly by recomputing BFI and comparing them. | ||||
674 | void LLVM_ATTRIBUTE_UNUSED__attribute__((__unused__)) CodeGenPrepare::verifyBFIUpdates(Function &F) { | ||||
675 | DominatorTree NewDT(F); | ||||
676 | LoopInfo NewLI(NewDT); | ||||
677 | BranchProbabilityInfo NewBPI(F, NewLI, TLInfo); | ||||
678 | BlockFrequencyInfo NewBFI(F, NewBPI, NewLI); | ||||
679 | NewBFI.verifyMatch(*BFI); | ||||
680 | } | ||||
681 | |||||
682 | /// Merge basic blocks which are connected by a single edge, where one of the | ||||
683 | /// basic blocks has a single successor pointing to the other basic block, | ||||
684 | /// which has a single predecessor. | ||||
685 | bool CodeGenPrepare::eliminateFallThrough(Function &F) { | ||||
686 | bool Changed = false; | ||||
687 | // Scan all of the blocks in the function, except for the entry block. | ||||
688 | // Use a temporary array to avoid iterator being invalidated when | ||||
689 | // deleting blocks. | ||||
690 | SmallVector<WeakTrackingVH, 16> Blocks; | ||||
691 | for (auto &Block : llvm::drop_begin(F)) | ||||
692 | Blocks.push_back(&Block); | ||||
693 | |||||
694 | SmallSet<WeakTrackingVH, 16> Preds; | ||||
695 | for (auto &Block : Blocks) { | ||||
696 | auto *BB = cast_or_null<BasicBlock>(Block); | ||||
697 | if (!BB) | ||||
698 | continue; | ||||
699 | // If the destination block has a single pred, then this is a trivial | ||||
700 | // edge, just collapse it. | ||||
701 | BasicBlock *SinglePred = BB->getSinglePredecessor(); | ||||
702 | |||||
703 | // Don't merge if BB's address is taken. | ||||
704 | if (!SinglePred || SinglePred == BB || BB->hasAddressTaken()) continue; | ||||
705 | |||||
706 | BranchInst *Term = dyn_cast<BranchInst>(SinglePred->getTerminator()); | ||||
707 | if (Term && !Term->isConditional()) { | ||||
708 | Changed = true; | ||||
709 | LLVM_DEBUG(dbgs() << "To merge:\n" << *BB << "\n\n\n")do { } while (false); | ||||
710 | |||||
711 | // Merge BB into SinglePred and delete it. | ||||
712 | MergeBlockIntoPredecessor(BB); | ||||
713 | Preds.insert(SinglePred); | ||||
714 | } | ||||
715 | } | ||||
716 | |||||
717 | // (Repeatedly) merging blocks into their predecessors can create redundant | ||||
718 | // debug intrinsics. | ||||
719 | for (auto &Pred : Preds) | ||||
720 | if (auto *BB = cast_or_null<BasicBlock>(Pred)) | ||||
721 | RemoveRedundantDbgInstrs(BB); | ||||
722 | |||||
723 | return Changed; | ||||
724 | } | ||||
725 | |||||
726 | /// Find a destination block from BB if BB is mergeable empty block. | ||||
727 | BasicBlock *CodeGenPrepare::findDestBlockOfMergeableEmptyBlock(BasicBlock *BB) { | ||||
728 | // If this block doesn't end with an uncond branch, ignore it. | ||||
729 | BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()); | ||||
730 | if (!BI || !BI->isUnconditional()) | ||||
731 | return nullptr; | ||||
732 | |||||
733 | // If the instruction before the branch (skipping debug info) isn't a phi | ||||
734 | // node, then other stuff is happening here. | ||||
735 | BasicBlock::iterator BBI = BI->getIterator(); | ||||
736 | if (BBI != BB->begin()) { | ||||
737 | --BBI; | ||||
738 | while (isa<DbgInfoIntrinsic>(BBI)) { | ||||
739 | if (BBI == BB->begin()) | ||||
740 | break; | ||||
741 | --BBI; | ||||
742 | } | ||||
743 | if (!isa<DbgInfoIntrinsic>(BBI) && !isa<PHINode>(BBI)) | ||||
744 | return nullptr; | ||||
745 | } | ||||
746 | |||||
747 | // Do not break infinite loops. | ||||
748 | BasicBlock *DestBB = BI->getSuccessor(0); | ||||
749 | if (DestBB == BB) | ||||
750 | return nullptr; | ||||
751 | |||||
752 | if (!canMergeBlocks(BB, DestBB)) | ||||
753 | DestBB = nullptr; | ||||
754 | |||||
755 | return DestBB; | ||||
756 | } | ||||
757 | |||||
758 | /// Eliminate blocks that contain only PHI nodes, debug info directives, and an | ||||
759 | /// unconditional branch. Passes before isel (e.g. LSR/loopsimplify) often split | ||||
760 | /// edges in ways that are non-optimal for isel. Start by eliminating these | ||||
761 | /// blocks so we can split them the way we want them. | ||||
762 | bool CodeGenPrepare::eliminateMostlyEmptyBlocks(Function &F) { | ||||
763 | SmallPtrSet<BasicBlock *, 16> Preheaders; | ||||
764 | SmallVector<Loop *, 16> LoopList(LI->begin(), LI->end()); | ||||
765 | while (!LoopList.empty()) { | ||||
766 | Loop *L = LoopList.pop_back_val(); | ||||
767 | llvm::append_range(LoopList, *L); | ||||
768 | if (BasicBlock *Preheader = L->getLoopPreheader()) | ||||
769 | Preheaders.insert(Preheader); | ||||
770 | } | ||||
771 | |||||
772 | bool MadeChange = false; | ||||
773 | // Copy blocks into a temporary array to avoid iterator invalidation issues | ||||
774 | // as we remove them. | ||||
775 | // Note that this intentionally skips the entry block. | ||||
776 | SmallVector<WeakTrackingVH, 16> Blocks; | ||||
777 | for (auto &Block : llvm::drop_begin(F)) | ||||
778 | Blocks.push_back(&Block); | ||||
779 | |||||
780 | for (auto &Block : Blocks) { | ||||
781 | BasicBlock *BB = cast_or_null<BasicBlock>(Block); | ||||
782 | if (!BB) | ||||
783 | continue; | ||||
784 | BasicBlock *DestBB = findDestBlockOfMergeableEmptyBlock(BB); | ||||
785 | if (!DestBB || | ||||
786 | !isMergingEmptyBlockProfitable(BB, DestBB, Preheaders.count(BB))) | ||||
787 | continue; | ||||
788 | |||||
789 | eliminateMostlyEmptyBlock(BB); | ||||
790 | MadeChange = true; | ||||
791 | } | ||||
792 | return MadeChange; | ||||
793 | } | ||||
794 | |||||
795 | bool CodeGenPrepare::isMergingEmptyBlockProfitable(BasicBlock *BB, | ||||
796 | BasicBlock *DestBB, | ||||
797 | bool isPreheader) { | ||||
798 | // Do not delete loop preheaders if doing so would create a critical edge. | ||||
799 | // Loop preheaders can be good locations to spill registers. If the | ||||
800 | // preheader is deleted and we create a critical edge, registers may be | ||||
801 | // spilled in the loop body instead. | ||||
802 | if (!DisablePreheaderProtect && isPreheader && | ||||
803 | !(BB->getSinglePredecessor() && | ||||
804 | BB->getSinglePredecessor()->getSingleSuccessor())) | ||||
805 | return false; | ||||
806 | |||||
807 | // Skip merging if the block's successor is also a successor to any callbr | ||||
808 | // that leads to this block. | ||||
809 | // FIXME: Is this really needed? Is this a correctness issue? | ||||
810 | for (BasicBlock *Pred : predecessors(BB)) { | ||||
811 | if (auto *CBI = dyn_cast<CallBrInst>((Pred)->getTerminator())) | ||||
812 | for (unsigned i = 0, e = CBI->getNumSuccessors(); i != e; ++i) | ||||
813 | if (DestBB == CBI->getSuccessor(i)) | ||||
814 | return false; | ||||
815 | } | ||||
816 | |||||
817 | // Try to skip merging if the unique predecessor of BB is terminated by a | ||||
818 | // switch or indirect branch instruction, and BB is used as an incoming block | ||||
819 | // of PHIs in DestBB. In such case, merging BB and DestBB would cause ISel to | ||||
820 | // add COPY instructions in the predecessor of BB instead of BB (if it is not | ||||
821 | // merged). Note that the critical edge created by merging such blocks wont be | ||||
822 | // split in MachineSink because the jump table is not analyzable. By keeping | ||||
823 | // such empty block (BB), ISel will place COPY instructions in BB, not in the | ||||
824 | // predecessor of BB. | ||||
825 | BasicBlock *Pred = BB->getUniquePredecessor(); | ||||
826 | if (!Pred || | ||||
827 | !(isa<SwitchInst>(Pred->getTerminator()) || | ||||
828 | isa<IndirectBrInst>(Pred->getTerminator()))) | ||||
829 | return true; | ||||
830 | |||||
831 | if (BB->getTerminator() != BB->getFirstNonPHIOrDbg()) | ||||
832 | return true; | ||||
833 | |||||
834 | // We use a simple cost heuristic which determine skipping merging is | ||||
835 | // profitable if the cost of skipping merging is less than the cost of | ||||
836 | // merging : Cost(skipping merging) < Cost(merging BB), where the | ||||
837 | // Cost(skipping merging) is Freq(BB) * (Cost(Copy) + Cost(Branch)), and | ||||
838 | // the Cost(merging BB) is Freq(Pred) * Cost(Copy). | ||||
839 | // Assuming Cost(Copy) == Cost(Branch), we could simplify it to : | ||||
840 | // Freq(Pred) / Freq(BB) > 2. | ||||
841 | // Note that if there are multiple empty blocks sharing the same incoming | ||||
842 | // value for the PHIs in the DestBB, we consider them together. In such | ||||
843 | // case, Cost(merging BB) will be the sum of their frequencies. | ||||
844 | |||||
845 | if (!isa<PHINode>(DestBB->begin())) | ||||
846 | return true; | ||||
847 | |||||
848 | SmallPtrSet<BasicBlock *, 16> SameIncomingValueBBs; | ||||
849 | |||||
850 | // Find all other incoming blocks from which incoming values of all PHIs in | ||||
851 | // DestBB are the same as the ones from BB. | ||||
852 | for (BasicBlock *DestBBPred : predecessors(DestBB)) { | ||||
853 | if (DestBBPred == BB) | ||||
854 | continue; | ||||
855 | |||||
856 | if (llvm::all_of(DestBB->phis(), [&](const PHINode &DestPN) { | ||||
857 | return DestPN.getIncomingValueForBlock(BB) == | ||||
858 | DestPN.getIncomingValueForBlock(DestBBPred); | ||||
859 | })) | ||||
860 | SameIncomingValueBBs.insert(DestBBPred); | ||||
861 | } | ||||
862 | |||||
863 | // See if all BB's incoming values are same as the value from Pred. In this | ||||
864 | // case, no reason to skip merging because COPYs are expected to be place in | ||||
865 | // Pred already. | ||||
866 | if (SameIncomingValueBBs.count(Pred)) | ||||
867 | return true; | ||||
868 | |||||
869 | BlockFrequency PredFreq = BFI->getBlockFreq(Pred); | ||||
870 | BlockFrequency BBFreq = BFI->getBlockFreq(BB); | ||||
871 | |||||
872 | for (auto *SameValueBB : SameIncomingValueBBs) | ||||
873 | if (SameValueBB->getUniquePredecessor() == Pred && | ||||
874 | DestBB == findDestBlockOfMergeableEmptyBlock(SameValueBB)) | ||||
875 | BBFreq += BFI->getBlockFreq(SameValueBB); | ||||
876 | |||||
877 | return PredFreq.getFrequency() <= | ||||
878 | BBFreq.getFrequency() * FreqRatioToSkipMerge; | ||||
879 | } | ||||
880 | |||||
881 | /// Return true if we can merge BB into DestBB if there is a single | ||||
882 | /// unconditional branch between them, and BB contains no other non-phi | ||||
883 | /// instructions. | ||||
884 | bool CodeGenPrepare::canMergeBlocks(const BasicBlock *BB, | ||||
885 | const BasicBlock *DestBB) const { | ||||
886 | // We only want to eliminate blocks whose phi nodes are used by phi nodes in | ||||
887 | // the successor. If there are more complex condition (e.g. preheaders), | ||||
888 | // don't mess around with them. | ||||
889 | for (const PHINode &PN : BB->phis()) { | ||||
890 | for (const User *U : PN.users()) { | ||||
891 | const Instruction *UI = cast<Instruction>(U); | ||||
892 | if (UI->getParent() != DestBB || !isa<PHINode>(UI)) | ||||
893 | return false; | ||||
894 | // If User is inside DestBB block and it is a PHINode then check | ||||
895 | // incoming value. If incoming value is not from BB then this is | ||||
896 | // a complex condition (e.g. preheaders) we want to avoid here. | ||||
897 | if (UI->getParent() == DestBB) { | ||||
898 | if (const PHINode *UPN = dyn_cast<PHINode>(UI)) | ||||
899 | for (unsigned I = 0, E = UPN->getNumIncomingValues(); I != E; ++I) { | ||||
900 | Instruction *Insn = dyn_cast<Instruction>(UPN->getIncomingValue(I)); | ||||
901 | if (Insn && Insn->getParent() == BB && | ||||
902 | Insn->getParent() != UPN->getIncomingBlock(I)) | ||||
903 | return false; | ||||
904 | } | ||||
905 | } | ||||
906 | } | ||||
907 | } | ||||
908 | |||||
909 | // If BB and DestBB contain any common predecessors, then the phi nodes in BB | ||||
910 | // and DestBB may have conflicting incoming values for the block. If so, we | ||||
911 | // can't merge the block. | ||||
912 | const PHINode *DestBBPN = dyn_cast<PHINode>(DestBB->begin()); | ||||
913 | if (!DestBBPN) return true; // no conflict. | ||||
914 | |||||
915 | // Collect the preds of BB. | ||||
916 | SmallPtrSet<const BasicBlock*, 16> BBPreds; | ||||
917 | if (const PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) { | ||||
918 | // It is faster to get preds from a PHI than with pred_iterator. | ||||
919 | for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i) | ||||
920 | BBPreds.insert(BBPN->getIncomingBlock(i)); | ||||
921 | } else { | ||||
922 | BBPreds.insert(pred_begin(BB), pred_end(BB)); | ||||
923 | } | ||||
924 | |||||
925 | // Walk the preds of DestBB. | ||||
926 | for (unsigned i = 0, e = DestBBPN->getNumIncomingValues(); i != e; ++i) { | ||||
927 | BasicBlock *Pred = DestBBPN->getIncomingBlock(i); | ||||
928 | if (BBPreds.count(Pred)) { // Common predecessor? | ||||
929 | for (const PHINode &PN : DestBB->phis()) { | ||||
930 | const Value *V1 = PN.getIncomingValueForBlock(Pred); | ||||
931 | const Value *V2 = PN.getIncomingValueForBlock(BB); | ||||
932 | |||||
933 | // If V2 is a phi node in BB, look up what the mapped value will be. | ||||
934 | if (const PHINode *V2PN = dyn_cast<PHINode>(V2)) | ||||
935 | if (V2PN->getParent() == BB) | ||||
936 | V2 = V2PN->getIncomingValueForBlock(Pred); | ||||
937 | |||||
938 | // If there is a conflict, bail out. | ||||
939 | if (V1 != V2) return false; | ||||
940 | } | ||||
941 | } | ||||
942 | } | ||||
943 | |||||
944 | return true; | ||||
945 | } | ||||
946 | |||||
947 | /// Eliminate a basic block that has only phi's and an unconditional branch in | ||||
948 | /// it. | ||||
949 | void CodeGenPrepare::eliminateMostlyEmptyBlock(BasicBlock *BB) { | ||||
950 | BranchInst *BI = cast<BranchInst>(BB->getTerminator()); | ||||
951 | BasicBlock *DestBB = BI->getSuccessor(0); | ||||
952 | |||||
953 | LLVM_DEBUG(dbgs() << "MERGING MOSTLY EMPTY BLOCKS - BEFORE:\n"do { } while (false) | ||||
954 | << *BB << *DestBB)do { } while (false); | ||||
955 | |||||
956 | // If the destination block has a single pred, then this is a trivial edge, | ||||
957 | // just collapse it. | ||||
958 | if (BasicBlock *SinglePred = DestBB->getSinglePredecessor()) { | ||||
959 | if (SinglePred != DestBB) { | ||||
960 | assert(SinglePred == BB &&((void)0) | ||||
961 | "Single predecessor not the same as predecessor")((void)0); | ||||
962 | // Merge DestBB into SinglePred/BB and delete it. | ||||
963 | MergeBlockIntoPredecessor(DestBB); | ||||
964 | // Note: BB(=SinglePred) will not be deleted on this path. | ||||
965 | // DestBB(=its single successor) is the one that was deleted. | ||||
966 | LLVM_DEBUG(dbgs() << "AFTER:\n" << *SinglePred << "\n\n\n")do { } while (false); | ||||
967 | return; | ||||
968 | } | ||||
969 | } | ||||
970 | |||||
971 | // Otherwise, we have multiple predecessors of BB. Update the PHIs in DestBB | ||||
972 | // to handle the new incoming edges it is about to have. | ||||
973 | for (PHINode &PN : DestBB->phis()) { | ||||
974 | // Remove the incoming value for BB, and remember it. | ||||
975 | Value *InVal = PN.removeIncomingValue(BB, false); | ||||
976 | |||||
977 | // Two options: either the InVal is a phi node defined in BB or it is some | ||||
978 | // value that dominates BB. | ||||
979 | PHINode *InValPhi = dyn_cast<PHINode>(InVal); | ||||
980 | if (InValPhi && InValPhi->getParent() == BB) { | ||||
981 | // Add all of the input values of the input PHI as inputs of this phi. | ||||
982 | for (unsigned i = 0, e = InValPhi->getNumIncomingValues(); i != e; ++i) | ||||
983 | PN.addIncoming(InValPhi->getIncomingValue(i), | ||||
984 | InValPhi->getIncomingBlock(i)); | ||||
985 | } else { | ||||
986 | // Otherwise, add one instance of the dominating value for each edge that | ||||
987 | // we will be adding. | ||||
988 | if (PHINode *BBPN = dyn_cast<PHINode>(BB->begin())) { | ||||
989 | for (unsigned i = 0, e = BBPN->getNumIncomingValues(); i != e; ++i) | ||||
990 | PN.addIncoming(InVal, BBPN->getIncomingBlock(i)); | ||||
991 | } else { | ||||
992 | for (BasicBlock *Pred : predecessors(BB)) | ||||
993 | PN.addIncoming(InVal, Pred); | ||||
994 | } | ||||
995 | } | ||||
996 | } | ||||
997 | |||||
998 | // The PHIs are now updated, change everything that refers to BB to use | ||||
999 | // DestBB and remove BB. | ||||
1000 | BB->replaceAllUsesWith(DestBB); | ||||
1001 | BB->eraseFromParent(); | ||||
1002 | ++NumBlocksElim; | ||||
1003 | |||||
1004 | LLVM_DEBUG(dbgs() << "AFTER:\n" << *DestBB << "\n\n\n")do { } while (false); | ||||
1005 | } | ||||
1006 | |||||
1007 | // Computes a map of base pointer relocation instructions to corresponding | ||||
1008 | // derived pointer relocation instructions given a vector of all relocate calls | ||||
1009 | static void computeBaseDerivedRelocateMap( | ||||
1010 | const SmallVectorImpl<GCRelocateInst *> &AllRelocateCalls, | ||||
1011 | DenseMap<GCRelocateInst *, SmallVector<GCRelocateInst *, 2>> | ||||
1012 | &RelocateInstMap) { | ||||
1013 | // Collect information in two maps: one primarily for locating the base object | ||||
1014 | // while filling the second map; the second map is the final structure holding | ||||
1015 | // a mapping between Base and corresponding Derived relocate calls | ||||
1016 | DenseMap<std::pair<unsigned, unsigned>, GCRelocateInst *> RelocateIdxMap; | ||||
1017 | for (auto *ThisRelocate : AllRelocateCalls) { | ||||
1018 | auto K = std::make_pair(ThisRelocate->getBasePtrIndex(), | ||||
1019 | ThisRelocate->getDerivedPtrIndex()); | ||||
1020 | RelocateIdxMap.insert(std::make_pair(K, ThisRelocate)); | ||||
1021 | } | ||||
1022 | for (auto &Item : RelocateIdxMap) { | ||||
1023 | std::pair<unsigned, unsigned> Key = Item.first; | ||||
1024 | if (Key.first == Key.second) | ||||
1025 | // Base relocation: nothing to insert | ||||
1026 | continue; | ||||
1027 | |||||
1028 | GCRelocateInst *I = Item.second; | ||||
1029 | auto BaseKey = std::make_pair(Key.first, Key.first); | ||||
1030 | |||||
1031 | // We're iterating over RelocateIdxMap so we cannot modify it. | ||||
1032 | auto MaybeBase = RelocateIdxMap.find(BaseKey); | ||||
1033 | if (MaybeBase == RelocateIdxMap.end()) | ||||
1034 | // TODO: We might want to insert a new base object relocate and gep off | ||||
1035 | // that, if there are enough derived object relocates. | ||||
1036 | continue; | ||||
1037 | |||||
1038 | RelocateInstMap[MaybeBase->second].push_back(I); | ||||
1039 | } | ||||
1040 | } | ||||
1041 | |||||
1042 | // Accepts a GEP and extracts the operands into a vector provided they're all | ||||
1043 | // small integer constants | ||||
1044 | static bool getGEPSmallConstantIntOffsetV(GetElementPtrInst *GEP, | ||||
1045 | SmallVectorImpl<Value *> &OffsetV) { | ||||
1046 | for (unsigned i = 1; i < GEP->getNumOperands(); i++) { | ||||
1047 | // Only accept small constant integer operands | ||||
1048 | auto *Op = dyn_cast<ConstantInt>(GEP->getOperand(i)); | ||||
1049 | if (!Op || Op->getZExtValue() > 20) | ||||
1050 | return false; | ||||
1051 | } | ||||
1052 | |||||
1053 | for (unsigned i = 1; i < GEP->getNumOperands(); i++) | ||||
1054 | OffsetV.push_back(GEP->getOperand(i)); | ||||
1055 | return true; | ||||
1056 | } | ||||
1057 | |||||
1058 | // Takes a RelocatedBase (base pointer relocation instruction) and Targets to | ||||
1059 | // replace, computes a replacement, and affects it. | ||||
1060 | static bool | ||||
1061 | simplifyRelocatesOffABase(GCRelocateInst *RelocatedBase, | ||||
1062 | const SmallVectorImpl<GCRelocateInst *> &Targets) { | ||||
1063 | bool MadeChange = false; | ||||
1064 | // We must ensure the relocation of derived pointer is defined after | ||||
1065 | // relocation of base pointer. If we find a relocation corresponding to base | ||||
1066 | // defined earlier than relocation of base then we move relocation of base | ||||
1067 | // right before found relocation. We consider only relocation in the same | ||||
1068 | // basic block as relocation of base. Relocations from other basic block will | ||||
1069 | // be skipped by optimization and we do not care about them. | ||||
1070 | for (auto R = RelocatedBase->getParent()->getFirstInsertionPt(); | ||||
1071 | &*R != RelocatedBase; ++R) | ||||
1072 | if (auto *RI = dyn_cast<GCRelocateInst>(R)) | ||||
1073 | if (RI->getStatepoint() == RelocatedBase->getStatepoint()) | ||||
1074 | if (RI->getBasePtrIndex() == RelocatedBase->getBasePtrIndex()) { | ||||
1075 | RelocatedBase->moveBefore(RI); | ||||
1076 | break; | ||||
1077 | } | ||||
1078 | |||||
1079 | for (GCRelocateInst *ToReplace : Targets) { | ||||
1080 | assert(ToReplace->getBasePtrIndex() == RelocatedBase->getBasePtrIndex() &&((void)0) | ||||
1081 | "Not relocating a derived object of the original base object")((void)0); | ||||
1082 | if (ToReplace->getBasePtrIndex() == ToReplace->getDerivedPtrIndex()) { | ||||
1083 | // A duplicate relocate call. TODO: coalesce duplicates. | ||||
1084 | continue; | ||||
1085 | } | ||||
1086 | |||||
1087 | if (RelocatedBase->getParent() != ToReplace->getParent()) { | ||||
1088 | // Base and derived relocates are in different basic blocks. | ||||
1089 | // In this case transform is only valid when base dominates derived | ||||
1090 | // relocate. However it would be too expensive to check dominance | ||||
1091 | // for each such relocate, so we skip the whole transformation. | ||||
1092 | continue; | ||||
1093 | } | ||||
1094 | |||||
1095 | Value *Base = ToReplace->getBasePtr(); | ||||
1096 | auto *Derived = dyn_cast<GetElementPtrInst>(ToReplace->getDerivedPtr()); | ||||
1097 | if (!Derived || Derived->getPointerOperand() != Base) | ||||
1098 | continue; | ||||
1099 | |||||
1100 | SmallVector<Value *, 2> OffsetV; | ||||
1101 | if (!getGEPSmallConstantIntOffsetV(Derived, OffsetV)) | ||||
1102 | continue; | ||||
1103 | |||||
1104 | // Create a Builder and replace the target callsite with a gep | ||||
1105 | assert(RelocatedBase->getNextNode() &&((void)0) | ||||
1106 | "Should always have one since it's not a terminator")((void)0); | ||||
1107 | |||||
1108 | // Insert after RelocatedBase | ||||
1109 | IRBuilder<> Builder(RelocatedBase->getNextNode()); | ||||
1110 | Builder.SetCurrentDebugLocation(ToReplace->getDebugLoc()); | ||||
1111 | |||||
1112 | // If gc_relocate does not match the actual type, cast it to the right type. | ||||
1113 | // In theory, there must be a bitcast after gc_relocate if the type does not | ||||
1114 | // match, and we should reuse it to get the derived pointer. But it could be | ||||
1115 | // cases like this: | ||||
1116 | // bb1: | ||||
1117 | // ... | ||||
1118 | // %g1 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...) | ||||
1119 | // br label %merge | ||||
1120 | // | ||||
1121 | // bb2: | ||||
1122 | // ... | ||||
1123 | // %g2 = call coldcc i8 addrspace(1)* @llvm.experimental.gc.relocate.p1i8(...) | ||||
1124 | // br label %merge | ||||
1125 | // | ||||
1126 | // merge: | ||||
1127 | // %p1 = phi i8 addrspace(1)* [ %g1, %bb1 ], [ %g2, %bb2 ] | ||||
1128 | // %cast = bitcast i8 addrspace(1)* %p1 in to i32 addrspace(1)* | ||||
1129 | // | ||||
1130 | // In this case, we can not find the bitcast any more. So we insert a new bitcast | ||||
1131 | // no matter there is already one or not. In this way, we can handle all cases, and | ||||
1132 | // the extra bitcast should be optimized away in later passes. | ||||
1133 | Value *ActualRelocatedBase = RelocatedBase; | ||||
1134 | if (RelocatedBase->getType() != Base->getType()) { | ||||
1135 | ActualRelocatedBase = | ||||
1136 | Builder.CreateBitCast(RelocatedBase, Base->getType()); | ||||
1137 | } | ||||
1138 | Value *Replacement = Builder.CreateGEP( | ||||
1139 | Derived->getSourceElementType(), ActualRelocatedBase, makeArrayRef(OffsetV)); | ||||
1140 | Replacement->takeName(ToReplace); | ||||
1141 | // If the newly generated derived pointer's type does not match the original derived | ||||
1142 | // pointer's type, cast the new derived pointer to match it. Same reasoning as above. | ||||
1143 | Value *ActualReplacement = Replacement; | ||||
1144 | if (Replacement->getType() != ToReplace->getType()) { | ||||
1145 | ActualReplacement = | ||||
1146 | Builder.CreateBitCast(Replacement, ToReplace->getType()); | ||||
1147 | } | ||||
1148 | ToReplace->replaceAllUsesWith(ActualReplacement); | ||||
1149 | ToReplace->eraseFromParent(); | ||||
1150 | |||||
1151 | MadeChange = true; | ||||
1152 | } | ||||
1153 | return MadeChange; | ||||
1154 | } | ||||
1155 | |||||
1156 | // Turns this: | ||||
1157 | // | ||||
1158 | // %base = ... | ||||
1159 | // %ptr = gep %base + 15 | ||||
1160 | // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr) | ||||
1161 | // %base' = relocate(%tok, i32 4, i32 4) | ||||
1162 | // %ptr' = relocate(%tok, i32 4, i32 5) | ||||
1163 | // %val = load %ptr' | ||||
1164 | // | ||||
1165 | // into this: | ||||
1166 | // | ||||
1167 | // %base = ... | ||||
1168 | // %ptr = gep %base + 15 | ||||
1169 | // %tok = statepoint (%fun, i32 0, i32 0, i32 0, %base, %ptr) | ||||
1170 | // %base' = gc.relocate(%tok, i32 4, i32 4) | ||||
1171 | // %ptr' = gep %base' + 15 | ||||
1172 | // %val = load %ptr' | ||||
1173 | bool CodeGenPrepare::simplifyOffsetableRelocate(GCStatepointInst &I) { | ||||
1174 | bool MadeChange = false; | ||||
1175 | SmallVector<GCRelocateInst *, 2> AllRelocateCalls; | ||||
1176 | for (auto *U : I.users()) | ||||
1177 | if (GCRelocateInst *Relocate = dyn_cast<GCRelocateInst>(U)) | ||||
1178 | // Collect all the relocate calls associated with a statepoint | ||||
1179 | AllRelocateCalls.push_back(Relocate); | ||||
1180 | |||||
1181 | // We need at least one base pointer relocation + one derived pointer | ||||
1182 | // relocation to mangle | ||||
1183 | if (AllRelocateCalls.size() < 2) | ||||
1184 | return false; | ||||
1185 | |||||
1186 | // RelocateInstMap is a mapping from the base relocate instruction to the | ||||
1187 | // corresponding derived relocate instructions | ||||
1188 | DenseMap<GCRelocateInst *, SmallVector<GCRelocateInst *, 2>> RelocateInstMap; | ||||
1189 | computeBaseDerivedRelocateMap(AllRelocateCalls, RelocateInstMap); | ||||
1190 | if (RelocateInstMap.empty()) | ||||
1191 | return false; | ||||
1192 | |||||
1193 | for (auto &Item : RelocateInstMap) | ||||
1194 | // Item.first is the RelocatedBase to offset against | ||||
1195 | // Item.second is the vector of Targets to replace | ||||
1196 | MadeChange = simplifyRelocatesOffABase(Item.first, Item.second); | ||||
1197 | return MadeChange; | ||||
1198 | } | ||||
1199 | |||||
1200 | /// Sink the specified cast instruction into its user blocks. | ||||
1201 | static bool SinkCast(CastInst *CI) { | ||||
1202 | BasicBlock *DefBB = CI->getParent(); | ||||
1203 | |||||
1204 | /// InsertedCasts - Only insert a cast in each block once. | ||||
1205 | DenseMap<BasicBlock*, CastInst*> InsertedCasts; | ||||
1206 | |||||
1207 | bool MadeChange = false; | ||||
1208 | for (Value::user_iterator UI = CI->user_begin(), E = CI->user_end(); | ||||
1209 | UI != E; ) { | ||||
1210 | Use &TheUse = UI.getUse(); | ||||
1211 | Instruction *User = cast<Instruction>(*UI); | ||||
1212 | |||||
1213 | // Figure out which BB this cast is used in. For PHI's this is the | ||||
1214 | // appropriate predecessor block. | ||||
1215 | BasicBlock *UserBB = User->getParent(); | ||||
1216 | if (PHINode *PN = dyn_cast<PHINode>(User)) { | ||||
1217 | UserBB = PN->getIncomingBlock(TheUse); | ||||
1218 | } | ||||
1219 | |||||
1220 | // Preincrement use iterator so we don't invalidate it. | ||||
1221 | ++UI; | ||||
1222 | |||||
1223 | // The first insertion point of a block containing an EH pad is after the | ||||
1224 | // pad. If the pad is the user, we cannot sink the cast past the pad. | ||||
1225 | if (User->isEHPad()) | ||||
1226 | continue; | ||||
1227 | |||||
1228 | // If the block selected to receive the cast is an EH pad that does not | ||||
1229 | // allow non-PHI instructions before the terminator, we can't sink the | ||||
1230 | // cast. | ||||
1231 | if (UserBB->getTerminator()->isEHPad()) | ||||
1232 | continue; | ||||
1233 | |||||
1234 | // If this user is in the same block as the cast, don't change the cast. | ||||
1235 | if (UserBB == DefBB) continue; | ||||
1236 | |||||
1237 | // If we have already inserted a cast into this block, use it. | ||||
1238 | CastInst *&InsertedCast = InsertedCasts[UserBB]; | ||||
1239 | |||||
1240 | if (!InsertedCast) { | ||||
1241 | BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt(); | ||||
1242 | assert(InsertPt != UserBB->end())((void)0); | ||||
1243 | InsertedCast = CastInst::Create(CI->getOpcode(), CI->getOperand(0), | ||||
1244 | CI->getType(), "", &*InsertPt); | ||||
1245 | InsertedCast->setDebugLoc(CI->getDebugLoc()); | ||||
1246 | } | ||||
1247 | |||||
1248 | // Replace a use of the cast with a use of the new cast. | ||||
1249 | TheUse = InsertedCast; | ||||
1250 | MadeChange = true; | ||||
1251 | ++NumCastUses; | ||||
1252 | } | ||||
1253 | |||||
1254 | // If we removed all uses, nuke the cast. | ||||
1255 | if (CI->use_empty()) { | ||||
1256 | salvageDebugInfo(*CI); | ||||
1257 | CI->eraseFromParent(); | ||||
1258 | MadeChange = true; | ||||
1259 | } | ||||
1260 | |||||
1261 | return MadeChange; | ||||
1262 | } | ||||
1263 | |||||
1264 | /// If the specified cast instruction is a noop copy (e.g. it's casting from | ||||
1265 | /// one pointer type to another, i32->i8 on PPC), sink it into user blocks to | ||||
1266 | /// reduce the number of virtual registers that must be created and coalesced. | ||||
1267 | /// | ||||
1268 | /// Return true if any changes are made. | ||||
1269 | static bool OptimizeNoopCopyExpression(CastInst *CI, const TargetLowering &TLI, | ||||
1270 | const DataLayout &DL) { | ||||
1271 | // Sink only "cheap" (or nop) address-space casts. This is a weaker condition | ||||
1272 | // than sinking only nop casts, but is helpful on some platforms. | ||||
1273 | if (auto *ASC = dyn_cast<AddrSpaceCastInst>(CI)) { | ||||
1274 | if (!TLI.isFreeAddrSpaceCast(ASC->getSrcAddressSpace(), | ||||
1275 | ASC->getDestAddressSpace())) | ||||
1276 | return false; | ||||
1277 | } | ||||
1278 | |||||
1279 | // If this is a noop copy, | ||||
1280 | EVT SrcVT = TLI.getValueType(DL, CI->getOperand(0)->getType()); | ||||
1281 | EVT DstVT = TLI.getValueType(DL, CI->getType()); | ||||
1282 | |||||
1283 | // This is an fp<->int conversion? | ||||
1284 | if (SrcVT.isInteger() != DstVT.isInteger()) | ||||
1285 | return false; | ||||
1286 | |||||
1287 | // If this is an extension, it will be a zero or sign extension, which | ||||
1288 | // isn't a noop. | ||||
1289 | if (SrcVT.bitsLT(DstVT)) return false; | ||||
1290 | |||||
1291 | // If these values will be promoted, find out what they will be promoted | ||||
1292 | // to. This helps us consider truncates on PPC as noop copies when they | ||||
1293 | // are. | ||||
1294 | if (TLI.getTypeAction(CI->getContext(), SrcVT) == | ||||
1295 | TargetLowering::TypePromoteInteger) | ||||
1296 | SrcVT = TLI.getTypeToTransformTo(CI->getContext(), SrcVT); | ||||
1297 | if (TLI.getTypeAction(CI->getContext(), DstVT) == | ||||
1298 | TargetLowering::TypePromoteInteger) | ||||
1299 | DstVT = TLI.getTypeToTransformTo(CI->getContext(), DstVT); | ||||
1300 | |||||
1301 | // If, after promotion, these are the same types, this is a noop copy. | ||||
1302 | if (SrcVT != DstVT) | ||||
1303 | return false; | ||||
1304 | |||||
1305 | return SinkCast(CI); | ||||
1306 | } | ||||
1307 | |||||
1308 | // Match a simple increment by constant operation. Note that if a sub is | ||||
1309 | // matched, the step is negated (as if the step had been canonicalized to | ||||
1310 | // an add, even though we leave the instruction alone.) | ||||
1311 | bool matchIncrement(const Instruction* IVInc, Instruction *&LHS, | ||||
1312 | Constant *&Step) { | ||||
1313 | if (match(IVInc, m_Add(m_Instruction(LHS), m_Constant(Step))) || | ||||
1314 | match(IVInc, m_ExtractValue<0>(m_Intrinsic<Intrinsic::uadd_with_overflow>( | ||||
1315 | m_Instruction(LHS), m_Constant(Step))))) | ||||
1316 | return true; | ||||
1317 | if (match(IVInc, m_Sub(m_Instruction(LHS), m_Constant(Step))) || | ||||
1318 | match(IVInc, m_ExtractValue<0>(m_Intrinsic<Intrinsic::usub_with_overflow>( | ||||
1319 | m_Instruction(LHS), m_Constant(Step))))) { | ||||
1320 | Step = ConstantExpr::getNeg(Step); | ||||
1321 | return true; | ||||
1322 | } | ||||
1323 | return false; | ||||
1324 | } | ||||
1325 | |||||
1326 | /// If given \p PN is an inductive variable with value IVInc coming from the | ||||
1327 | /// backedge, and on each iteration it gets increased by Step, return pair | ||||
1328 | /// <IVInc, Step>. Otherwise, return None. | ||||
1329 | static Optional<std::pair<Instruction *, Constant *> > | ||||
1330 | getIVIncrement(const PHINode *PN, const LoopInfo *LI) { | ||||
1331 | const Loop *L = LI->getLoopFor(PN->getParent()); | ||||
1332 | if (!L || L->getHeader() != PN->getParent() || !L->getLoopLatch()) | ||||
1333 | return None; | ||||
1334 | auto *IVInc = | ||||
1335 | dyn_cast<Instruction>(PN->getIncomingValueForBlock(L->getLoopLatch())); | ||||
1336 | if (!IVInc || LI->getLoopFor(IVInc->getParent()) != L) | ||||
1337 | return None; | ||||
1338 | Instruction *LHS = nullptr; | ||||
1339 | Constant *Step = nullptr; | ||||
1340 | if (matchIncrement(IVInc, LHS, Step) && LHS == PN) | ||||
1341 | return std::make_pair(IVInc, Step); | ||||
1342 | return None; | ||||
1343 | } | ||||
1344 | |||||
1345 | static bool isIVIncrement(const Value *V, const LoopInfo *LI) { | ||||
1346 | auto *I = dyn_cast<Instruction>(V); | ||||
1347 | if (!I) | ||||
1348 | return false; | ||||
1349 | Instruction *LHS = nullptr; | ||||
1350 | Constant *Step = nullptr; | ||||
1351 | if (!matchIncrement(I, LHS, Step)) | ||||
1352 | return false; | ||||
1353 | if (auto *PN = dyn_cast<PHINode>(LHS)) | ||||
1354 | if (auto IVInc = getIVIncrement(PN, LI)) | ||||
1355 | return IVInc->first == I; | ||||
1356 | return false; | ||||
1357 | } | ||||
1358 | |||||
1359 | bool CodeGenPrepare::replaceMathCmpWithIntrinsic(BinaryOperator *BO, | ||||
1360 | Value *Arg0, Value *Arg1, | ||||
1361 | CmpInst *Cmp, | ||||
1362 | Intrinsic::ID IID) { | ||||
1363 | auto IsReplacableIVIncrement = [this, &Cmp](BinaryOperator *BO) { | ||||
1364 | if (!isIVIncrement(BO, LI)) | ||||
1365 | return false; | ||||
1366 | const Loop *L = LI->getLoopFor(BO->getParent()); | ||||
1367 | assert(L && "L should not be null after isIVIncrement()")((void)0); | ||||
1368 | // Do not risk on moving increment into a child loop. | ||||
1369 | if (LI->getLoopFor(Cmp->getParent()) != L) | ||||
1370 | return false; | ||||
1371 | |||||
1372 | // Finally, we need to ensure that the insert point will dominate all | ||||
1373 | // existing uses of the increment. | ||||
1374 | |||||
1375 | auto &DT = getDT(*BO->getParent()->getParent()); | ||||
1376 | if (DT.dominates(Cmp->getParent(), BO->getParent())) | ||||
1377 | // If we're moving up the dom tree, all uses are trivially dominated. | ||||
1378 | // (This is the common case for code produced by LSR.) | ||||
1379 | return true; | ||||
1380 | |||||
1381 | // Otherwise, special case the single use in the phi recurrence. | ||||
1382 | return BO->hasOneUse() && DT.dominates(Cmp->getParent(), L->getLoopLatch()); | ||||
1383 | }; | ||||
1384 | if (BO->getParent() != Cmp->getParent() && !IsReplacableIVIncrement(BO)) { | ||||
1385 | // We used to use a dominator tree here to allow multi-block optimization. | ||||
1386 | // But that was problematic because: | ||||
1387 | // 1. It could cause a perf regression by hoisting the math op into the | ||||
1388 | // critical path. | ||||
1389 | // 2. It could cause a perf regression by creating a value that was live | ||||
1390 | // across multiple blocks and increasing register pressure. | ||||
1391 | // 3. Use of a dominator tree could cause large compile-time regression. | ||||
1392 | // This is because we recompute the DT on every change in the main CGP | ||||
1393 | // run-loop. The recomputing is probably unnecessary in many cases, so if | ||||
1394 | // that was fixed, using a DT here would be ok. | ||||
1395 | // | ||||
1396 | // There is one important particular case we still want to handle: if BO is | ||||
1397 | // the IV increment. Important properties that make it profitable: | ||||
1398 | // - We can speculate IV increment anywhere in the loop (as long as the | ||||
1399 | // indvar Phi is its only user); | ||||
1400 | // - Upon computing Cmp, we effectively compute something equivalent to the | ||||
1401 | // IV increment (despite it loops differently in the IR). So moving it up | ||||
1402 | // to the cmp point does not really increase register pressure. | ||||
1403 | return false; | ||||
1404 | } | ||||
1405 | |||||
1406 | // We allow matching the canonical IR (add X, C) back to (usubo X, -C). | ||||
1407 | if (BO->getOpcode() == Instruction::Add && | ||||
1408 | IID == Intrinsic::usub_with_overflow) { | ||||
1409 | assert(isa<Constant>(Arg1) && "Unexpected input for usubo")((void)0); | ||||
1410 | Arg1 = ConstantExpr::getNeg(cast<Constant>(Arg1)); | ||||
1411 | } | ||||
1412 | |||||
1413 | // Insert at the first instruction of the pair. | ||||
1414 | Instruction *InsertPt = nullptr; | ||||
1415 | for (Instruction &Iter : *Cmp->getParent()) { | ||||
1416 | // If BO is an XOR, it is not guaranteed that it comes after both inputs to | ||||
1417 | // the overflow intrinsic are defined. | ||||
1418 | if ((BO->getOpcode() != Instruction::Xor && &Iter == BO) || &Iter == Cmp) { | ||||
1419 | InsertPt = &Iter; | ||||
1420 | break; | ||||
1421 | } | ||||
1422 | } | ||||
1423 | assert(InsertPt != nullptr && "Parent block did not contain cmp or binop")((void)0); | ||||
1424 | |||||
1425 | IRBuilder<> Builder(InsertPt); | ||||
1426 | Value *MathOV = Builder.CreateBinaryIntrinsic(IID, Arg0, Arg1); | ||||
1427 | if (BO->getOpcode() != Instruction::Xor) { | ||||
1428 | Value *Math = Builder.CreateExtractValue(MathOV, 0, "math"); | ||||
1429 | BO->replaceAllUsesWith(Math); | ||||
1430 | } else | ||||
1431 | assert(BO->hasOneUse() &&((void)0) | ||||
1432 | "Patterns with XOr should use the BO only in the compare")((void)0); | ||||
1433 | Value *OV = Builder.CreateExtractValue(MathOV, 1, "ov"); | ||||
1434 | Cmp->replaceAllUsesWith(OV); | ||||
1435 | Cmp->eraseFromParent(); | ||||
1436 | BO->eraseFromParent(); | ||||
1437 | return true; | ||||
1438 | } | ||||
1439 | |||||
1440 | /// Match special-case patterns that check for unsigned add overflow. | ||||
1441 | static bool matchUAddWithOverflowConstantEdgeCases(CmpInst *Cmp, | ||||
1442 | BinaryOperator *&Add) { | ||||
1443 | // Add = add A, 1; Cmp = icmp eq A,-1 (overflow if A is max val) | ||||
1444 | // Add = add A,-1; Cmp = icmp ne A, 0 (overflow if A is non-zero) | ||||
1445 | Value *A = Cmp->getOperand(0), *B = Cmp->getOperand(1); | ||||
1446 | |||||
1447 | // We are not expecting non-canonical/degenerate code. Just bail out. | ||||
1448 | if (isa<Constant>(A)) | ||||
1449 | return false; | ||||
1450 | |||||
1451 | ICmpInst::Predicate Pred = Cmp->getPredicate(); | ||||
1452 | if (Pred == ICmpInst::ICMP_EQ && match(B, m_AllOnes())) | ||||
1453 | B = ConstantInt::get(B->getType(), 1); | ||||
1454 | else if (Pred == ICmpInst::ICMP_NE && match(B, m_ZeroInt())) | ||||
1455 | B = ConstantInt::get(B->getType(), -1); | ||||
1456 | else | ||||
1457 | return false; | ||||
1458 | |||||
1459 | // Check the users of the variable operand of the compare looking for an add | ||||
1460 | // with the adjusted constant. | ||||
1461 | for (User *U : A->users()) { | ||||
1462 | if (match(U, m_Add(m_Specific(A), m_Specific(B)))) { | ||||
1463 | Add = cast<BinaryOperator>(U); | ||||
1464 | return true; | ||||
1465 | } | ||||
1466 | } | ||||
1467 | return false; | ||||
1468 | } | ||||
1469 | |||||
1470 | /// Try to combine the compare into a call to the llvm.uadd.with.overflow | ||||
1471 | /// intrinsic. Return true if any changes were made. | ||||
1472 | bool CodeGenPrepare::combineToUAddWithOverflow(CmpInst *Cmp, | ||||
1473 | bool &ModifiedDT) { | ||||
1474 | Value *A, *B; | ||||
1475 | BinaryOperator *Add; | ||||
1476 | if (!match(Cmp, m_UAddWithOverflow(m_Value(A), m_Value(B), m_BinOp(Add)))) { | ||||
1477 | if (!matchUAddWithOverflowConstantEdgeCases(Cmp, Add)) | ||||
1478 | return false; | ||||
1479 | // Set A and B in case we match matchUAddWithOverflowConstantEdgeCases. | ||||
1480 | A = Add->getOperand(0); | ||||
1481 | B = Add->getOperand(1); | ||||
1482 | } | ||||
1483 | |||||
1484 | if (!TLI->shouldFormOverflowOp(ISD::UADDO, | ||||
1485 | TLI->getValueType(*DL, Add->getType()), | ||||
1486 | Add->hasNUsesOrMore(2))) | ||||
1487 | return false; | ||||
1488 | |||||
1489 | // We don't want to move around uses of condition values this late, so we | ||||
1490 | // check if it is legal to create the call to the intrinsic in the basic | ||||
1491 | // block containing the icmp. | ||||
1492 | if (Add->getParent() != Cmp->getParent() && !Add->hasOneUse()) | ||||
1493 | return false; | ||||
1494 | |||||
1495 | if (!replaceMathCmpWithIntrinsic(Add, A, B, Cmp, | ||||
1496 | Intrinsic::uadd_with_overflow)) | ||||
1497 | return false; | ||||
1498 | |||||
1499 | // Reset callers - do not crash by iterating over a dead instruction. | ||||
1500 | ModifiedDT = true; | ||||
1501 | return true; | ||||
1502 | } | ||||
1503 | |||||
1504 | bool CodeGenPrepare::combineToUSubWithOverflow(CmpInst *Cmp, | ||||
1505 | bool &ModifiedDT) { | ||||
1506 | // We are not expecting non-canonical/degenerate code. Just bail out. | ||||
1507 | Value *A = Cmp->getOperand(0), *B = Cmp->getOperand(1); | ||||
1508 | if (isa<Constant>(A) && isa<Constant>(B)) | ||||
1509 | return false; | ||||
1510 | |||||
1511 | // Convert (A u> B) to (A u< B) to simplify pattern matching. | ||||
1512 | ICmpInst::Predicate Pred = Cmp->getPredicate(); | ||||
1513 | if (Pred == ICmpInst::ICMP_UGT) { | ||||
1514 | std::swap(A, B); | ||||
1515 | Pred = ICmpInst::ICMP_ULT; | ||||
1516 | } | ||||
1517 | // Convert special-case: (A == 0) is the same as (A u< 1). | ||||
1518 | if (Pred == ICmpInst::ICMP_EQ && match(B, m_ZeroInt())) { | ||||
1519 | B = ConstantInt::get(B->getType(), 1); | ||||
1520 | Pred = ICmpInst::ICMP_ULT; | ||||
1521 | } | ||||
1522 | // Convert special-case: (A != 0) is the same as (0 u< A). | ||||
1523 | if (Pred == ICmpInst::ICMP_NE && match(B, m_ZeroInt())) { | ||||
1524 | std::swap(A, B); | ||||
1525 | Pred = ICmpInst::ICMP_ULT; | ||||
1526 | } | ||||
1527 | if (Pred != ICmpInst::ICMP_ULT) | ||||
1528 | return false; | ||||
1529 | |||||
1530 | // Walk the users of a variable operand of a compare looking for a subtract or | ||||
1531 | // add with that same operand. Also match the 2nd operand of the compare to | ||||
1532 | // the add/sub, but that may be a negated constant operand of an add. | ||||
1533 | Value *CmpVariableOperand = isa<Constant>(A) ? B : A; | ||||
1534 | BinaryOperator *Sub = nullptr; | ||||
1535 | for (User *U : CmpVariableOperand->users()) { | ||||
1536 | // A - B, A u< B --> usubo(A, B) | ||||
1537 | if (match(U, m_Sub(m_Specific(A), m_Specific(B)))) { | ||||
1538 | Sub = cast<BinaryOperator>(U); | ||||
1539 | break; | ||||
1540 | } | ||||
1541 | |||||
1542 | // A + (-C), A u< C (canonicalized form of (sub A, C)) | ||||
1543 | const APInt *CmpC, *AddC; | ||||
1544 | if (match(U, m_Add(m_Specific(A), m_APInt(AddC))) && | ||||
1545 | match(B, m_APInt(CmpC)) && *AddC == -(*CmpC)) { | ||||
1546 | Sub = cast<BinaryOperator>(U); | ||||
1547 | break; | ||||
1548 | } | ||||
1549 | } | ||||
1550 | if (!Sub) | ||||
1551 | return false; | ||||
1552 | |||||
1553 | if (!TLI->shouldFormOverflowOp(ISD::USUBO, | ||||
1554 | TLI->getValueType(*DL, Sub->getType()), | ||||
1555 | Sub->hasNUsesOrMore(2))) | ||||
1556 | return false; | ||||
1557 | |||||
1558 | if (!replaceMathCmpWithIntrinsic(Sub, Sub->getOperand(0), Sub->getOperand(1), | ||||
1559 | Cmp, Intrinsic::usub_with_overflow)) | ||||
1560 | return false; | ||||
1561 | |||||
1562 | // Reset callers - do not crash by iterating over a dead instruction. | ||||
1563 | ModifiedDT = true; | ||||
1564 | return true; | ||||
1565 | } | ||||
1566 | |||||
1567 | /// Sink the given CmpInst into user blocks to reduce the number of virtual | ||||
1568 | /// registers that must be created and coalesced. This is a clear win except on | ||||
1569 | /// targets with multiple condition code registers (PowerPC), where it might | ||||
1570 | /// lose; some adjustment may be wanted there. | ||||
1571 | /// | ||||
1572 | /// Return true if any changes are made. | ||||
1573 | static bool sinkCmpExpression(CmpInst *Cmp, const TargetLowering &TLI) { | ||||
1574 | if (TLI.hasMultipleConditionRegisters()) | ||||
1575 | return false; | ||||
1576 | |||||
1577 | // Avoid sinking soft-FP comparisons, since this can move them into a loop. | ||||
1578 | if (TLI.useSoftFloat() && isa<FCmpInst>(Cmp)) | ||||
1579 | return false; | ||||
1580 | |||||
1581 | // Only insert a cmp in each block once. | ||||
1582 | DenseMap<BasicBlock*, CmpInst*> InsertedCmps; | ||||
1583 | |||||
1584 | bool MadeChange = false; | ||||
1585 | for (Value::user_iterator UI = Cmp->user_begin(), E = Cmp->user_end(); | ||||
1586 | UI != E; ) { | ||||
1587 | Use &TheUse = UI.getUse(); | ||||
1588 | Instruction *User = cast<Instruction>(*UI); | ||||
1589 | |||||
1590 | // Preincrement use iterator so we don't invalidate it. | ||||
1591 | ++UI; | ||||
1592 | |||||
1593 | // Don't bother for PHI nodes. | ||||
1594 | if (isa<PHINode>(User)) | ||||
1595 | continue; | ||||
1596 | |||||
1597 | // Figure out which BB this cmp is used in. | ||||
1598 | BasicBlock *UserBB = User->getParent(); | ||||
1599 | BasicBlock *DefBB = Cmp->getParent(); | ||||
1600 | |||||
1601 | // If this user is in the same block as the cmp, don't change the cmp. | ||||
1602 | if (UserBB == DefBB) continue; | ||||
1603 | |||||
1604 | // If we have already inserted a cmp into this block, use it. | ||||
1605 | CmpInst *&InsertedCmp = InsertedCmps[UserBB]; | ||||
1606 | |||||
1607 | if (!InsertedCmp) { | ||||
1608 | BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt(); | ||||
1609 | assert(InsertPt != UserBB->end())((void)0); | ||||
1610 | InsertedCmp = | ||||
1611 | CmpInst::Create(Cmp->getOpcode(), Cmp->getPredicate(), | ||||
1612 | Cmp->getOperand(0), Cmp->getOperand(1), "", | ||||
1613 | &*InsertPt); | ||||
1614 | // Propagate the debug info. | ||||
1615 | InsertedCmp->setDebugLoc(Cmp->getDebugLoc()); | ||||
1616 | } | ||||
1617 | |||||
1618 | // Replace a use of the cmp with a use of the new cmp. | ||||
1619 | TheUse = InsertedCmp; | ||||
1620 | MadeChange = true; | ||||
1621 | ++NumCmpUses; | ||||
1622 | } | ||||
1623 | |||||
1624 | // If we removed all uses, nuke the cmp. | ||||
1625 | if (Cmp->use_empty()) { | ||||
1626 | Cmp->eraseFromParent(); | ||||
1627 | MadeChange = true; | ||||
1628 | } | ||||
1629 | |||||
1630 | return MadeChange; | ||||
1631 | } | ||||
1632 | |||||
1633 | /// For pattern like: | ||||
1634 | /// | ||||
1635 | /// DomCond = icmp sgt/slt CmpOp0, CmpOp1 (might not be in DomBB) | ||||
1636 | /// ... | ||||
1637 | /// DomBB: | ||||
1638 | /// ... | ||||
1639 | /// br DomCond, TrueBB, CmpBB | ||||
1640 | /// CmpBB: (with DomBB being the single predecessor) | ||||
1641 | /// ... | ||||
1642 | /// Cmp = icmp eq CmpOp0, CmpOp1 | ||||
1643 | /// ... | ||||
1644 | /// | ||||
1645 | /// It would use two comparison on targets that lowering of icmp sgt/slt is | ||||
1646 | /// different from lowering of icmp eq (PowerPC). This function try to convert | ||||
1647 | /// 'Cmp = icmp eq CmpOp0, CmpOp1' to ' Cmp = icmp slt/sgt CmpOp0, CmpOp1'. | ||||
1648 | /// After that, DomCond and Cmp can use the same comparison so reduce one | ||||
1649 | /// comparison. | ||||
1650 | /// | ||||
1651 | /// Return true if any changes are made. | ||||
1652 | static bool foldICmpWithDominatingICmp(CmpInst *Cmp, | ||||
1653 | const TargetLowering &TLI) { | ||||
1654 | if (!EnableICMP_EQToICMP_ST && TLI.isEqualityCmpFoldedWithSignedCmp()) | ||||
1655 | return false; | ||||
1656 | |||||
1657 | ICmpInst::Predicate Pred = Cmp->getPredicate(); | ||||
1658 | if (Pred != ICmpInst::ICMP_EQ) | ||||
1659 | return false; | ||||
1660 | |||||
1661 | // If icmp eq has users other than BranchInst and SelectInst, converting it to | ||||
1662 | // icmp slt/sgt would introduce more redundant LLVM IR. | ||||
1663 | for (User *U : Cmp->users()) { | ||||
1664 | if (isa<BranchInst>(U)) | ||||
1665 | continue; | ||||
1666 | if (isa<SelectInst>(U) && cast<SelectInst>(U)->getCondition() == Cmp) | ||||
1667 | continue; | ||||
1668 | return false; | ||||
1669 | } | ||||
1670 | |||||
1671 | // This is a cheap/incomplete check for dominance - just match a single | ||||
1672 | // predecessor with a conditional branch. | ||||
1673 | BasicBlock *CmpBB = Cmp->getParent(); | ||||
1674 | BasicBlock *DomBB = CmpBB->getSinglePredecessor(); | ||||
1675 | if (!DomBB) | ||||
1676 | return false; | ||||
1677 | |||||
1678 | // We want to ensure that the only way control gets to the comparison of | ||||
1679 | // interest is that a less/greater than comparison on the same operands is | ||||
1680 | // false. | ||||
1681 | Value *DomCond; | ||||
1682 | BasicBlock *TrueBB, *FalseBB; | ||||
1683 | if (!match(DomBB->getTerminator(), m_Br(m_Value(DomCond), TrueBB, FalseBB))) | ||||
1684 | return false; | ||||
1685 | if (CmpBB != FalseBB) | ||||
1686 | return false; | ||||
1687 | |||||
1688 | Value *CmpOp0 = Cmp->getOperand(0), *CmpOp1 = Cmp->getOperand(1); | ||||
1689 | ICmpInst::Predicate DomPred; | ||||
1690 | if (!match(DomCond, m_ICmp(DomPred, m_Specific(CmpOp0), m_Specific(CmpOp1)))) | ||||
1691 | return false; | ||||
1692 | if (DomPred != ICmpInst::ICMP_SGT && DomPred != ICmpInst::ICMP_SLT) | ||||
1693 | return false; | ||||
1694 | |||||
1695 | // Convert the equality comparison to the opposite of the dominating | ||||
1696 | // comparison and swap the direction for all branch/select users. | ||||
1697 | // We have conceptually converted: | ||||
1698 | // Res = (a < b) ? <LT_RES> : (a == b) ? <EQ_RES> : <GT_RES>; | ||||
1699 | // to | ||||
1700 | // Res = (a < b) ? <LT_RES> : (a > b) ? <GT_RES> : <EQ_RES>; | ||||
1701 | // And similarly for branches. | ||||
1702 | for (User *U : Cmp->users()) { | ||||
1703 | if (auto *BI = dyn_cast<BranchInst>(U)) { | ||||
1704 | assert(BI->isConditional() && "Must be conditional")((void)0); | ||||
1705 | BI->swapSuccessors(); | ||||
1706 | continue; | ||||
1707 | } | ||||
1708 | if (auto *SI = dyn_cast<SelectInst>(U)) { | ||||
1709 | // Swap operands | ||||
1710 | SI->swapValues(); | ||||
1711 | SI->swapProfMetadata(); | ||||
1712 | continue; | ||||
1713 | } | ||||
1714 | llvm_unreachable("Must be a branch or a select")__builtin_unreachable(); | ||||
1715 | } | ||||
1716 | Cmp->setPredicate(CmpInst::getSwappedPredicate(DomPred)); | ||||
1717 | return true; | ||||
1718 | } | ||||
1719 | |||||
1720 | bool CodeGenPrepare::optimizeCmp(CmpInst *Cmp, bool &ModifiedDT) { | ||||
1721 | if (sinkCmpExpression(Cmp, *TLI)) | ||||
1722 | return true; | ||||
1723 | |||||
1724 | if (combineToUAddWithOverflow(Cmp, ModifiedDT)) | ||||
1725 | return true; | ||||
1726 | |||||
1727 | if (combineToUSubWithOverflow(Cmp, ModifiedDT)) | ||||
1728 | return true; | ||||
1729 | |||||
1730 | if (foldICmpWithDominatingICmp(Cmp, *TLI)) | ||||
1731 | return true; | ||||
1732 | |||||
1733 | return false; | ||||
1734 | } | ||||
1735 | |||||
1736 | /// Duplicate and sink the given 'and' instruction into user blocks where it is | ||||
1737 | /// used in a compare to allow isel to generate better code for targets where | ||||
1738 | /// this operation can be combined. | ||||
1739 | /// | ||||
1740 | /// Return true if any changes are made. | ||||
1741 | static bool sinkAndCmp0Expression(Instruction *AndI, | ||||
1742 | const TargetLowering &TLI, | ||||
1743 | SetOfInstrs &InsertedInsts) { | ||||
1744 | // Double-check that we're not trying to optimize an instruction that was | ||||
1745 | // already optimized by some other part of this pass. | ||||
1746 | assert(!InsertedInsts.count(AndI) &&((void)0) | ||||
1747 | "Attempting to optimize already optimized and instruction")((void)0); | ||||
1748 | (void) InsertedInsts; | ||||
1749 | |||||
1750 | // Nothing to do for single use in same basic block. | ||||
1751 | if (AndI->hasOneUse() && | ||||
1752 | AndI->getParent() == cast<Instruction>(*AndI->user_begin())->getParent()) | ||||
1753 | return false; | ||||
1754 | |||||
1755 | // Try to avoid cases where sinking/duplicating is likely to increase register | ||||
1756 | // pressure. | ||||
1757 | if (!isa<ConstantInt>(AndI->getOperand(0)) && | ||||
1758 | !isa<ConstantInt>(AndI->getOperand(1)) && | ||||
1759 | AndI->getOperand(0)->hasOneUse() && AndI->getOperand(1)->hasOneUse()) | ||||
1760 | return false; | ||||
1761 | |||||
1762 | for (auto *U : AndI->users()) { | ||||
1763 | Instruction *User = cast<Instruction>(U); | ||||
1764 | |||||
1765 | // Only sink 'and' feeding icmp with 0. | ||||
1766 | if (!isa<ICmpInst>(User)) | ||||
1767 | return false; | ||||
1768 | |||||
1769 | auto *CmpC = dyn_cast<ConstantInt>(User->getOperand(1)); | ||||
1770 | if (!CmpC || !CmpC->isZero()) | ||||
1771 | return false; | ||||
1772 | } | ||||
1773 | |||||
1774 | if (!TLI.isMaskAndCmp0FoldingBeneficial(*AndI)) | ||||
1775 | return false; | ||||
1776 | |||||
1777 | LLVM_DEBUG(dbgs() << "found 'and' feeding only icmp 0;\n")do { } while (false); | ||||
1778 | LLVM_DEBUG(AndI->getParent()->dump())do { } while (false); | ||||
1779 | |||||
1780 | // Push the 'and' into the same block as the icmp 0. There should only be | ||||
1781 | // one (icmp (and, 0)) in each block, since CSE/GVN should have removed any | ||||
1782 | // others, so we don't need to keep track of which BBs we insert into. | ||||
1783 | for (Value::user_iterator UI = AndI->user_begin(), E = AndI->user_end(); | ||||
1784 | UI != E; ) { | ||||
1785 | Use &TheUse = UI.getUse(); | ||||
1786 | Instruction *User = cast<Instruction>(*UI); | ||||
1787 | |||||
1788 | // Preincrement use iterator so we don't invalidate it. | ||||
1789 | ++UI; | ||||
1790 | |||||
1791 | LLVM_DEBUG(dbgs() << "sinking 'and' use: " << *User << "\n")do { } while (false); | ||||
1792 | |||||
1793 | // Keep the 'and' in the same place if the use is already in the same block. | ||||
1794 | Instruction *InsertPt = | ||||
1795 | User->getParent() == AndI->getParent() ? AndI : User; | ||||
1796 | Instruction *InsertedAnd = | ||||
1797 | BinaryOperator::Create(Instruction::And, AndI->getOperand(0), | ||||
1798 | AndI->getOperand(1), "", InsertPt); | ||||
1799 | // Propagate the debug info. | ||||
1800 | InsertedAnd->setDebugLoc(AndI->getDebugLoc()); | ||||
1801 | |||||
1802 | // Replace a use of the 'and' with a use of the new 'and'. | ||||
1803 | TheUse = InsertedAnd; | ||||
1804 | ++NumAndUses; | ||||
1805 | LLVM_DEBUG(User->getParent()->dump())do { } while (false); | ||||
1806 | } | ||||
1807 | |||||
1808 | // We removed all uses, nuke the and. | ||||
1809 | AndI->eraseFromParent(); | ||||
1810 | return true; | ||||
1811 | } | ||||
1812 | |||||
1813 | /// Check if the candidates could be combined with a shift instruction, which | ||||
1814 | /// includes: | ||||
1815 | /// 1. Truncate instruction | ||||
1816 | /// 2. And instruction and the imm is a mask of the low bits: | ||||
1817 | /// imm & (imm+1) == 0 | ||||
1818 | static bool isExtractBitsCandidateUse(Instruction *User) { | ||||
1819 | if (!isa<TruncInst>(User)) { | ||||
1820 | if (User->getOpcode() != Instruction::And || | ||||
1821 | !isa<ConstantInt>(User->getOperand(1))) | ||||
1822 | return false; | ||||
1823 | |||||
1824 | const APInt &Cimm = cast<ConstantInt>(User->getOperand(1))->getValue(); | ||||
1825 | |||||
1826 | if ((Cimm & (Cimm + 1)).getBoolValue()) | ||||
1827 | return false; | ||||
1828 | } | ||||
1829 | return true; | ||||
1830 | } | ||||
1831 | |||||
1832 | /// Sink both shift and truncate instruction to the use of truncate's BB. | ||||
1833 | static bool | ||||
1834 | SinkShiftAndTruncate(BinaryOperator *ShiftI, Instruction *User, ConstantInt *CI, | ||||
1835 | DenseMap<BasicBlock *, BinaryOperator *> &InsertedShifts, | ||||
1836 | const TargetLowering &TLI, const DataLayout &DL) { | ||||
1837 | BasicBlock *UserBB = User->getParent(); | ||||
1838 | DenseMap<BasicBlock *, CastInst *> InsertedTruncs; | ||||
1839 | auto *TruncI = cast<TruncInst>(User); | ||||
1840 | bool MadeChange = false; | ||||
1841 | |||||
1842 | for (Value::user_iterator TruncUI = TruncI->user_begin(), | ||||
1843 | TruncE = TruncI->user_end(); | ||||
1844 | TruncUI != TruncE;) { | ||||
1845 | |||||
1846 | Use &TruncTheUse = TruncUI.getUse(); | ||||
1847 | Instruction *TruncUser = cast<Instruction>(*TruncUI); | ||||
1848 | // Preincrement use iterator so we don't invalidate it. | ||||
1849 | |||||
1850 | ++TruncUI; | ||||
1851 | |||||
1852 | int ISDOpcode = TLI.InstructionOpcodeToISD(TruncUser->getOpcode()); | ||||
1853 | if (!ISDOpcode) | ||||
1854 | continue; | ||||
1855 | |||||
1856 | // If the use is actually a legal node, there will not be an | ||||
1857 | // implicit truncate. | ||||
1858 | // FIXME: always querying the result type is just an | ||||
1859 | // approximation; some nodes' legality is determined by the | ||||
1860 | // operand or other means. There's no good way to find out though. | ||||
1861 | if (TLI.isOperationLegalOrCustom( | ||||
1862 | ISDOpcode, TLI.getValueType(DL, TruncUser->getType(), true))) | ||||
1863 | continue; | ||||
1864 | |||||
1865 | // Don't bother for PHI nodes. | ||||
1866 | if (isa<PHINode>(TruncUser)) | ||||
1867 | continue; | ||||
1868 | |||||
1869 | BasicBlock *TruncUserBB = TruncUser->getParent(); | ||||
1870 | |||||
1871 | if (UserBB == TruncUserBB) | ||||
1872 | continue; | ||||
1873 | |||||
1874 | BinaryOperator *&InsertedShift = InsertedShifts[TruncUserBB]; | ||||
1875 | CastInst *&InsertedTrunc = InsertedTruncs[TruncUserBB]; | ||||
1876 | |||||
1877 | if (!InsertedShift && !InsertedTrunc) { | ||||
1878 | BasicBlock::iterator InsertPt = TruncUserBB->getFirstInsertionPt(); | ||||
1879 | assert(InsertPt != TruncUserBB->end())((void)0); | ||||
1880 | // Sink the shift | ||||
1881 | if (ShiftI->getOpcode() == Instruction::AShr) | ||||
1882 | InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI, | ||||
1883 | "", &*InsertPt); | ||||
1884 | else | ||||
1885 | InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI, | ||||
1886 | "", &*InsertPt); | ||||
1887 | InsertedShift->setDebugLoc(ShiftI->getDebugLoc()); | ||||
1888 | |||||
1889 | // Sink the trunc | ||||
1890 | BasicBlock::iterator TruncInsertPt = TruncUserBB->getFirstInsertionPt(); | ||||
1891 | TruncInsertPt++; | ||||
1892 | assert(TruncInsertPt != TruncUserBB->end())((void)0); | ||||
1893 | |||||
1894 | InsertedTrunc = CastInst::Create(TruncI->getOpcode(), InsertedShift, | ||||
1895 | TruncI->getType(), "", &*TruncInsertPt); | ||||
1896 | InsertedTrunc->setDebugLoc(TruncI->getDebugLoc()); | ||||
1897 | |||||
1898 | MadeChange = true; | ||||
1899 | |||||
1900 | TruncTheUse = InsertedTrunc; | ||||
1901 | } | ||||
1902 | } | ||||
1903 | return MadeChange; | ||||
1904 | } | ||||
1905 | |||||
1906 | /// Sink the shift *right* instruction into user blocks if the uses could | ||||
1907 | /// potentially be combined with this shift instruction and generate BitExtract | ||||
1908 | /// instruction. It will only be applied if the architecture supports BitExtract | ||||
1909 | /// instruction. Here is an example: | ||||
1910 | /// BB1: | ||||
1911 | /// %x.extract.shift = lshr i64 %arg1, 32 | ||||
1912 | /// BB2: | ||||
1913 | /// %x.extract.trunc = trunc i64 %x.extract.shift to i16 | ||||
1914 | /// ==> | ||||
1915 | /// | ||||
1916 | /// BB2: | ||||
1917 | /// %x.extract.shift.1 = lshr i64 %arg1, 32 | ||||
1918 | /// %x.extract.trunc = trunc i64 %x.extract.shift.1 to i16 | ||||
1919 | /// | ||||
1920 | /// CodeGen will recognize the pattern in BB2 and generate BitExtract | ||||
1921 | /// instruction. | ||||
1922 | /// Return true if any changes are made. | ||||
1923 | static bool OptimizeExtractBits(BinaryOperator *ShiftI, ConstantInt *CI, | ||||
1924 | const TargetLowering &TLI, | ||||
1925 | const DataLayout &DL) { | ||||
1926 | BasicBlock *DefBB = ShiftI->getParent(); | ||||
1927 | |||||
1928 | /// Only insert instructions in each block once. | ||||
1929 | DenseMap<BasicBlock *, BinaryOperator *> InsertedShifts; | ||||
1930 | |||||
1931 | bool shiftIsLegal = TLI.isTypeLegal(TLI.getValueType(DL, ShiftI->getType())); | ||||
1932 | |||||
1933 | bool MadeChange = false; | ||||
1934 | for (Value::user_iterator UI = ShiftI->user_begin(), E = ShiftI->user_end(); | ||||
1935 | UI != E;) { | ||||
1936 | Use &TheUse = UI.getUse(); | ||||
1937 | Instruction *User = cast<Instruction>(*UI); | ||||
1938 | // Preincrement use iterator so we don't invalidate it. | ||||
1939 | ++UI; | ||||
1940 | |||||
1941 | // Don't bother for PHI nodes. | ||||
1942 | if (isa<PHINode>(User)) | ||||
1943 | continue; | ||||
1944 | |||||
1945 | if (!isExtractBitsCandidateUse(User)) | ||||
1946 | continue; | ||||
1947 | |||||
1948 | BasicBlock *UserBB = User->getParent(); | ||||
1949 | |||||
1950 | if (UserBB == DefBB) { | ||||
1951 | // If the shift and truncate instruction are in the same BB. The use of | ||||
1952 | // the truncate(TruncUse) may still introduce another truncate if not | ||||
1953 | // legal. In this case, we would like to sink both shift and truncate | ||||
1954 | // instruction to the BB of TruncUse. | ||||
1955 | // for example: | ||||
1956 | // BB1: | ||||
1957 | // i64 shift.result = lshr i64 opnd, imm | ||||
1958 | // trunc.result = trunc shift.result to i16 | ||||
1959 | // | ||||
1960 | // BB2: | ||||
1961 | // ----> We will have an implicit truncate here if the architecture does | ||||
1962 | // not have i16 compare. | ||||
1963 | // cmp i16 trunc.result, opnd2 | ||||
1964 | // | ||||
1965 | if (isa<TruncInst>(User) && shiftIsLegal | ||||
1966 | // If the type of the truncate is legal, no truncate will be | ||||
1967 | // introduced in other basic blocks. | ||||
1968 | && | ||||
1969 | (!TLI.isTypeLegal(TLI.getValueType(DL, User->getType())))) | ||||
1970 | MadeChange = | ||||
1971 | SinkShiftAndTruncate(ShiftI, User, CI, InsertedShifts, TLI, DL); | ||||
1972 | |||||
1973 | continue; | ||||
1974 | } | ||||
1975 | // If we have already inserted a shift into this block, use it. | ||||
1976 | BinaryOperator *&InsertedShift = InsertedShifts[UserBB]; | ||||
1977 | |||||
1978 | if (!InsertedShift) { | ||||
1979 | BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt(); | ||||
1980 | assert(InsertPt != UserBB->end())((void)0); | ||||
1981 | |||||
1982 | if (ShiftI->getOpcode() == Instruction::AShr) | ||||
1983 | InsertedShift = BinaryOperator::CreateAShr(ShiftI->getOperand(0), CI, | ||||
1984 | "", &*InsertPt); | ||||
1985 | else | ||||
1986 | InsertedShift = BinaryOperator::CreateLShr(ShiftI->getOperand(0), CI, | ||||
1987 | "", &*InsertPt); | ||||
1988 | InsertedShift->setDebugLoc(ShiftI->getDebugLoc()); | ||||
1989 | |||||
1990 | MadeChange = true; | ||||
1991 | } | ||||
1992 | |||||
1993 | // Replace a use of the shift with a use of the new shift. | ||||
1994 | TheUse = InsertedShift; | ||||
1995 | } | ||||
1996 | |||||
1997 | // If we removed all uses, or there are none, nuke the shift. | ||||
1998 | if (ShiftI->use_empty()) { | ||||
1999 | salvageDebugInfo(*ShiftI); | ||||
2000 | ShiftI->eraseFromParent(); | ||||
2001 | MadeChange = true; | ||||
2002 | } | ||||
2003 | |||||
2004 | return MadeChange; | ||||
2005 | } | ||||
2006 | |||||
2007 | /// If counting leading or trailing zeros is an expensive operation and a zero | ||||
2008 | /// input is defined, add a check for zero to avoid calling the intrinsic. | ||||
2009 | /// | ||||
2010 | /// We want to transform: | ||||
2011 | /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 false) | ||||
2012 | /// | ||||
2013 | /// into: | ||||
2014 | /// entry: | ||||
2015 | /// %cmpz = icmp eq i64 %A, 0 | ||||
2016 | /// br i1 %cmpz, label %cond.end, label %cond.false | ||||
2017 | /// cond.false: | ||||
2018 | /// %z = call i64 @llvm.cttz.i64(i64 %A, i1 true) | ||||
2019 | /// br label %cond.end | ||||
2020 | /// cond.end: | ||||
2021 | /// %ctz = phi i64 [ 64, %entry ], [ %z, %cond.false ] | ||||
2022 | /// | ||||
2023 | /// If the transform is performed, return true and set ModifiedDT to true. | ||||
2024 | static bool despeculateCountZeros(IntrinsicInst *CountZeros, | ||||
2025 | const TargetLowering *TLI, | ||||
2026 | const DataLayout *DL, | ||||
2027 | bool &ModifiedDT) { | ||||
2028 | // If a zero input is undefined, it doesn't make sense to despeculate that. | ||||
2029 | if (match(CountZeros->getOperand(1), m_One())) | ||||
2030 | return false; | ||||
2031 | |||||
2032 | // If it's cheap to speculate, there's nothing to do. | ||||
2033 | auto IntrinsicID = CountZeros->getIntrinsicID(); | ||||
2034 | if ((IntrinsicID == Intrinsic::cttz && TLI->isCheapToSpeculateCttz()) || | ||||
2035 | (IntrinsicID == Intrinsic::ctlz && TLI->isCheapToSpeculateCtlz())) | ||||
2036 | return false; | ||||
2037 | |||||
2038 | // Only handle legal scalar cases. Anything else requires too much work. | ||||
2039 | Type *Ty = CountZeros->getType(); | ||||
2040 | unsigned SizeInBits = Ty->getPrimitiveSizeInBits(); | ||||
2041 | if (Ty->isVectorTy() || SizeInBits > DL->getLargestLegalIntTypeSizeInBits()) | ||||
2042 | return false; | ||||
2043 | |||||
2044 | // Bail if the value is never zero. | ||||
2045 | if (llvm::isKnownNonZero(CountZeros->getOperand(0), *DL)) | ||||
2046 | return false; | ||||
2047 | |||||
2048 | // The intrinsic will be sunk behind a compare against zero and branch. | ||||
2049 | BasicBlock *StartBlock = CountZeros->getParent(); | ||||
2050 | BasicBlock *CallBlock = StartBlock->splitBasicBlock(CountZeros, "cond.false"); | ||||
2051 | |||||
2052 | // Create another block after the count zero intrinsic. A PHI will be added | ||||
2053 | // in this block to select the result of the intrinsic or the bit-width | ||||
2054 | // constant if the input to the intrinsic is zero. | ||||
2055 | BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(CountZeros)); | ||||
2056 | BasicBlock *EndBlock = CallBlock->splitBasicBlock(SplitPt, "cond.end"); | ||||
2057 | |||||
2058 | // Set up a builder to create a compare, conditional branch, and PHI. | ||||
2059 | IRBuilder<> Builder(CountZeros->getContext()); | ||||
2060 | Builder.SetInsertPoint(StartBlock->getTerminator()); | ||||
2061 | Builder.SetCurrentDebugLocation(CountZeros->getDebugLoc()); | ||||
2062 | |||||
2063 | // Replace the unconditional branch that was created by the first split with | ||||
2064 | // a compare against zero and a conditional branch. | ||||
2065 | Value *Zero = Constant::getNullValue(Ty); | ||||
2066 | Value *Cmp = Builder.CreateICmpEQ(CountZeros->getOperand(0), Zero, "cmpz"); | ||||
2067 | Builder.CreateCondBr(Cmp, EndBlock, CallBlock); | ||||
2068 | StartBlock->getTerminator()->eraseFromParent(); | ||||
2069 | |||||
2070 | // Create a PHI in the end block to select either the output of the intrinsic | ||||
2071 | // or the bit width of the operand. | ||||
2072 | Builder.SetInsertPoint(&EndBlock->front()); | ||||
2073 | PHINode *PN = Builder.CreatePHI(Ty, 2, "ctz"); | ||||
2074 | CountZeros->replaceAllUsesWith(PN); | ||||
2075 | Value *BitWidth = Builder.getInt(APInt(SizeInBits, SizeInBits)); | ||||
2076 | PN->addIncoming(BitWidth, StartBlock); | ||||
2077 | PN->addIncoming(CountZeros, CallBlock); | ||||
2078 | |||||
2079 | // We are explicitly handling the zero case, so we can set the intrinsic's | ||||
2080 | // undefined zero argument to 'true'. This will also prevent reprocessing the | ||||
2081 | // intrinsic; we only despeculate when a zero input is defined. | ||||
2082 | CountZeros->setArgOperand(1, Builder.getTrue()); | ||||
2083 | ModifiedDT = true; | ||||
2084 | return true; | ||||
2085 | } | ||||
2086 | |||||
2087 | bool CodeGenPrepare::optimizeCallInst(CallInst *CI, bool &ModifiedDT) { | ||||
2088 | BasicBlock *BB = CI->getParent(); | ||||
2089 | |||||
2090 | // Lower inline assembly if we can. | ||||
2091 | // If we found an inline asm expession, and if the target knows how to | ||||
2092 | // lower it to normal LLVM code, do so now. | ||||
2093 | if (CI->isInlineAsm()) { | ||||
2094 | if (TLI->ExpandInlineAsm(CI)) { | ||||
2095 | // Avoid invalidating the iterator. | ||||
2096 | CurInstIterator = BB->begin(); | ||||
2097 | // Avoid processing instructions out of order, which could cause | ||||
2098 | // reuse before a value is defined. | ||||
2099 | SunkAddrs.clear(); | ||||
2100 | return true; | ||||
2101 | } | ||||
2102 | // Sink address computing for memory operands into the block. | ||||
2103 | if (optimizeInlineAsmInst(CI)) | ||||
2104 | return true; | ||||
2105 | } | ||||
2106 | |||||
2107 | // Align the pointer arguments to this call if the target thinks it's a good | ||||
2108 | // idea | ||||
2109 | unsigned MinSize, PrefAlign; | ||||
2110 | if (TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) { | ||||
2111 | for (auto &Arg : CI->arg_operands()) { | ||||
2112 | // We want to align both objects whose address is used directly and | ||||
2113 | // objects whose address is used in casts and GEPs, though it only makes | ||||
2114 | // sense for GEPs if the offset is a multiple of the desired alignment and | ||||
2115 | // if size - offset meets the size threshold. | ||||
2116 | if (!Arg->getType()->isPointerTy()) | ||||
2117 | continue; | ||||
2118 | APInt Offset(DL->getIndexSizeInBits( | ||||
2119 | cast<PointerType>(Arg->getType())->getAddressSpace()), | ||||
2120 | 0); | ||||
2121 | Value *Val = Arg->stripAndAccumulateInBoundsConstantOffsets(*DL, Offset); | ||||
2122 | uint64_t Offset2 = Offset.getLimitedValue(); | ||||
2123 | if ((Offset2 & (PrefAlign-1)) != 0) | ||||
2124 | continue; | ||||
2125 | AllocaInst *AI; | ||||
2126 | if ((AI = dyn_cast<AllocaInst>(Val)) && AI->getAlignment() < PrefAlign && | ||||
2127 | DL->getTypeAllocSize(AI->getAllocatedType()) >= MinSize + Offset2) | ||||
2128 | AI->setAlignment(Align(PrefAlign)); | ||||
2129 | // Global variables can only be aligned if they are defined in this | ||||
2130 | // object (i.e. they are uniquely initialized in this object), and | ||||
2131 | // over-aligning global variables that have an explicit section is | ||||
2132 | // forbidden. | ||||
2133 | GlobalVariable *GV; | ||||
2134 | if ((GV = dyn_cast<GlobalVariable>(Val)) && GV->canIncreaseAlignment() && | ||||
2135 | GV->getPointerAlignment(*DL) < PrefAlign && | ||||
2136 | DL->getTypeAllocSize(GV->getValueType()) >= | ||||
2137 | MinSize + Offset2) | ||||
2138 | GV->setAlignment(MaybeAlign(PrefAlign)); | ||||
2139 | } | ||||
2140 | // If this is a memcpy (or similar) then we may be able to improve the | ||||
2141 | // alignment | ||||
2142 | if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(CI)) { | ||||
2143 | Align DestAlign = getKnownAlignment(MI->getDest(), *DL); | ||||
2144 | MaybeAlign MIDestAlign = MI->getDestAlign(); | ||||
2145 | if (!MIDestAlign || DestAlign > *MIDestAlign) | ||||
2146 | MI->setDestAlignment(DestAlign); | ||||
2147 | if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI)) { | ||||
2148 | MaybeAlign MTISrcAlign = MTI->getSourceAlign(); | ||||
2149 | Align SrcAlign = getKnownAlignment(MTI->getSource(), *DL); | ||||
2150 | if (!MTISrcAlign || SrcAlign > *MTISrcAlign) | ||||
2151 | MTI->setSourceAlignment(SrcAlign); | ||||
2152 | } | ||||
2153 | } | ||||
2154 | } | ||||
2155 | |||||
2156 | // If we have a cold call site, try to sink addressing computation into the | ||||
2157 | // cold block. This interacts with our handling for loads and stores to | ||||
2158 | // ensure that we can fold all uses of a potential addressing computation | ||||
2159 | // into their uses. TODO: generalize this to work over profiling data | ||||
2160 | if (CI->hasFnAttr(Attribute::Cold) && | ||||
2161 | !OptSize && !llvm::shouldOptimizeForSize(BB, PSI, BFI.get())) | ||||
2162 | for (auto &Arg : CI->arg_operands()) { | ||||
2163 | if (!Arg->getType()->isPointerTy()) | ||||
2164 | continue; | ||||
2165 | unsigned AS = Arg->getType()->getPointerAddressSpace(); | ||||
2166 | return optimizeMemoryInst(CI, Arg, Arg->getType(), AS); | ||||
2167 | } | ||||
2168 | |||||
2169 | IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI); | ||||
2170 | if (II) { | ||||
2171 | switch (II->getIntrinsicID()) { | ||||
2172 | default: break; | ||||
2173 | case Intrinsic::assume: | ||||
2174 | llvm_unreachable("llvm.assume should have been removed already")__builtin_unreachable(); | ||||
2175 | case Intrinsic::experimental_widenable_condition: { | ||||
2176 | // Give up on future widening oppurtunties so that we can fold away dead | ||||
2177 | // paths and merge blocks before going into block-local instruction | ||||
2178 | // selection. | ||||
2179 | if (II->use_empty()) { | ||||
2180 | II->eraseFromParent(); | ||||
2181 | return true; | ||||
2182 | } | ||||
2183 | Constant *RetVal = ConstantInt::getTrue(II->getContext()); | ||||
2184 | resetIteratorIfInvalidatedWhileCalling(BB, [&]() { | ||||
2185 | replaceAndRecursivelySimplify(CI, RetVal, TLInfo, nullptr); | ||||
2186 | }); | ||||
2187 | return true; | ||||
2188 | } | ||||
2189 | case Intrinsic::objectsize: | ||||
2190 | llvm_unreachable("llvm.objectsize.* should have been lowered already")__builtin_unreachable(); | ||||
2191 | case Intrinsic::is_constant: | ||||
2192 | llvm_unreachable("llvm.is.constant.* should have been lowered already")__builtin_unreachable(); | ||||
2193 | case Intrinsic::aarch64_stlxr: | ||||
2194 | case Intrinsic::aarch64_stxr: { | ||||
2195 | ZExtInst *ExtVal = dyn_cast<ZExtInst>(CI->getArgOperand(0)); | ||||
2196 | if (!ExtVal || !ExtVal->hasOneUse() || | ||||
2197 | ExtVal->getParent() == CI->getParent()) | ||||
2198 | return false; | ||||
2199 | // Sink a zext feeding stlxr/stxr before it, so it can be folded into it. | ||||
2200 | ExtVal->moveBefore(CI); | ||||
2201 | // Mark this instruction as "inserted by CGP", so that other | ||||
2202 | // optimizations don't touch it. | ||||
2203 | InsertedInsts.insert(ExtVal); | ||||
2204 | return true; | ||||
2205 | } | ||||
2206 | |||||
2207 | case Intrinsic::launder_invariant_group: | ||||
2208 | case Intrinsic::strip_invariant_group: { | ||||
2209 | Value *ArgVal = II->getArgOperand(0); | ||||
2210 | auto it = LargeOffsetGEPMap.find(II); | ||||
2211 | if (it != LargeOffsetGEPMap.end()) { | ||||
2212 | // Merge entries in LargeOffsetGEPMap to reflect the RAUW. | ||||
2213 | // Make sure not to have to deal with iterator invalidation | ||||
2214 | // after possibly adding ArgVal to LargeOffsetGEPMap. | ||||
2215 | auto GEPs = std::move(it->second); | ||||
2216 | LargeOffsetGEPMap[ArgVal].append(GEPs.begin(), GEPs.end()); | ||||
2217 | LargeOffsetGEPMap.erase(II); | ||||
2218 | } | ||||
2219 | |||||
2220 | II->replaceAllUsesWith(ArgVal); | ||||
2221 | II->eraseFromParent(); | ||||
2222 | return true; | ||||
2223 | } | ||||
2224 | case Intrinsic::cttz: | ||||
2225 | case Intrinsic::ctlz: | ||||
2226 | // If counting zeros is expensive, try to avoid it. | ||||
2227 | return despeculateCountZeros(II, TLI, DL, ModifiedDT); | ||||
2228 | case Intrinsic::fshl: | ||||
2229 | case Intrinsic::fshr: | ||||
2230 | return optimizeFunnelShift(II); | ||||
2231 | case Intrinsic::dbg_value: | ||||
2232 | return fixupDbgValue(II); | ||||
2233 | case Intrinsic::vscale: { | ||||
2234 | // If datalayout has no special restrictions on vector data layout, | ||||
2235 | // replace `llvm.vscale` by an equivalent constant expression | ||||
2236 | // to benefit from cheap constant propagation. | ||||
2237 | Type *ScalableVectorTy = | ||||
2238 | VectorType::get(Type::getInt8Ty(II->getContext()), 1, true); | ||||
2239 | if (DL->getTypeAllocSize(ScalableVectorTy).getKnownMinSize() == 8) { | ||||
2240 | auto *Null = Constant::getNullValue(ScalableVectorTy->getPointerTo()); | ||||
2241 | auto *One = ConstantInt::getSigned(II->getType(), 1); | ||||
2242 | auto *CGep = | ||||
2243 | ConstantExpr::getGetElementPtr(ScalableVectorTy, Null, One); | ||||
2244 | II->replaceAllUsesWith(ConstantExpr::getPtrToInt(CGep, II->getType())); | ||||
2245 | II->eraseFromParent(); | ||||
2246 | return true; | ||||
2247 | } | ||||
2248 | break; | ||||
2249 | } | ||||
2250 | case Intrinsic::masked_gather: | ||||
2251 | return optimizeGatherScatterInst(II, II->getArgOperand(0)); | ||||
2252 | case Intrinsic::masked_scatter: | ||||
2253 | return optimizeGatherScatterInst(II, II->getArgOperand(1)); | ||||
2254 | } | ||||
2255 | |||||
2256 | SmallVector<Value *, 2> PtrOps; | ||||
2257 | Type *AccessTy; | ||||
2258 | if (TLI->getAddrModeArguments(II, PtrOps, AccessTy)) | ||||
2259 | while (!PtrOps.empty()) { | ||||
2260 | Value *PtrVal = PtrOps.pop_back_val(); | ||||
2261 | unsigned AS = PtrVal->getType()->getPointerAddressSpace(); | ||||
2262 | if (optimizeMemoryInst(II, PtrVal, AccessTy, AS)) | ||||
2263 | return true; | ||||
2264 | } | ||||
2265 | } | ||||
2266 | |||||
2267 | // From here on out we're working with named functions. | ||||
2268 | if (!CI->getCalledFunction()) return false; | ||||
2269 | |||||
2270 | // Lower all default uses of _chk calls. This is very similar | ||||
2271 | // to what InstCombineCalls does, but here we are only lowering calls | ||||
2272 | // to fortified library functions (e.g. __memcpy_chk) that have the default | ||||
2273 | // "don't know" as the objectsize. Anything else should be left alone. | ||||
2274 | FortifiedLibCallSimplifier Simplifier(TLInfo, true); | ||||
2275 | IRBuilder<> Builder(CI); | ||||
2276 | if (Value *V = Simplifier.optimizeCall(CI, Builder)) { | ||||
2277 | CI->replaceAllUsesWith(V); | ||||
2278 | CI->eraseFromParent(); | ||||
2279 | return true; | ||||
2280 | } | ||||
2281 | |||||
2282 | return false; | ||||
2283 | } | ||||
2284 | |||||
2285 | /// Look for opportunities to duplicate return instructions to the predecessor | ||||
2286 | /// to enable tail call optimizations. The case it is currently looking for is: | ||||
2287 | /// @code | ||||
2288 | /// bb0: | ||||
2289 | /// %tmp0 = tail call i32 @f0() | ||||
2290 | /// br label %return | ||||
2291 | /// bb1: | ||||
2292 | /// %tmp1 = tail call i32 @f1() | ||||
2293 | /// br label %return | ||||
2294 | /// bb2: | ||||
2295 | /// %tmp2 = tail call i32 @f2() | ||||
2296 | /// br label %return | ||||
2297 | /// return: | ||||
2298 | /// %retval = phi i32 [ %tmp0, %bb0 ], [ %tmp1, %bb1 ], [ %tmp2, %bb2 ] | ||||
2299 | /// ret i32 %retval | ||||
2300 | /// @endcode | ||||
2301 | /// | ||||
2302 | /// => | ||||
2303 | /// | ||||
2304 | /// @code | ||||
2305 | /// bb0: | ||||
2306 | /// %tmp0 = tail call i32 @f0() | ||||
2307 | /// ret i32 %tmp0 | ||||
2308 | /// bb1: | ||||
2309 | /// %tmp1 = tail call i32 @f1() | ||||
2310 | /// ret i32 %tmp1 | ||||
2311 | /// bb2: | ||||
2312 | /// %tmp2 = tail call i32 @f2() | ||||
2313 | /// ret i32 %tmp2 | ||||
2314 | /// @endcode | ||||
2315 | bool CodeGenPrepare::dupRetToEnableTailCallOpts(BasicBlock *BB, bool &ModifiedDT) { | ||||
2316 | ReturnInst *RetI = dyn_cast<ReturnInst>(BB->getTerminator()); | ||||
2317 | if (!RetI) | ||||
2318 | return false; | ||||
2319 | |||||
2320 | PHINode *PN = nullptr; | ||||
2321 | ExtractValueInst *EVI = nullptr; | ||||
2322 | BitCastInst *BCI = nullptr; | ||||
2323 | Value *V = RetI->getReturnValue(); | ||||
2324 | if (V) { | ||||
2325 | BCI = dyn_cast<BitCastInst>(V); | ||||
2326 | if (BCI) | ||||
2327 | V = BCI->getOperand(0); | ||||
2328 | |||||
2329 | EVI = dyn_cast<ExtractValueInst>(V); | ||||
2330 | if (EVI) { | ||||
2331 | V = EVI->getOperand(0); | ||||
2332 | if (!llvm::all_of(EVI->indices(), [](unsigned idx) { return idx == 0; })) | ||||
2333 | return false; | ||||
2334 | } | ||||
2335 | |||||
2336 | PN = dyn_cast<PHINode>(V); | ||||
2337 | if (!PN) | ||||
2338 | return false; | ||||
2339 | } | ||||
2340 | |||||
2341 | if (PN && PN->getParent() != BB) | ||||
2342 | return false; | ||||
2343 | |||||
2344 | auto isLifetimeEndOrBitCastFor = [](const Instruction *Inst) { | ||||
2345 | const BitCastInst *BC = dyn_cast<BitCastInst>(Inst); | ||||
2346 | if (BC && BC->hasOneUse()) | ||||
2347 | Inst = BC->user_back(); | ||||
2348 | |||||
2349 | if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) | ||||
2350 | return II->getIntrinsicID() == Intrinsic::lifetime_end; | ||||
2351 | return false; | ||||
2352 | }; | ||||
2353 | |||||
2354 | // Make sure there are no instructions between the first instruction | ||||
2355 | // and return. | ||||
2356 | const Instruction *BI = BB->getFirstNonPHI(); | ||||
2357 | // Skip over debug and the bitcast. | ||||
2358 | while (isa<DbgInfoIntrinsic>(BI) || BI == BCI || BI == EVI || | ||||
2359 | isa<PseudoProbeInst>(BI) || isLifetimeEndOrBitCastFor(BI)) | ||||
2360 | BI = BI->getNextNode(); | ||||
2361 | if (BI != RetI) | ||||
2362 | return false; | ||||
2363 | |||||
2364 | /// Only dup the ReturnInst if the CallInst is likely to be emitted as a tail | ||||
2365 | /// call. | ||||
2366 | const Function *F = BB->getParent(); | ||||
2367 | SmallVector<BasicBlock*, 4> TailCallBBs; | ||||
2368 | if (PN) { | ||||
2369 | for (unsigned I = 0, E = PN->getNumIncomingValues(); I != E; ++I) { | ||||
2370 | // Look through bitcasts. | ||||
2371 | Value *IncomingVal = PN->getIncomingValue(I)->stripPointerCasts(); | ||||
2372 | CallInst *CI = dyn_cast<CallInst>(IncomingVal); | ||||
2373 | BasicBlock *PredBB = PN->getIncomingBlock(I); | ||||
2374 | // Make sure the phi value is indeed produced by the tail call. | ||||
2375 | if (CI && CI->hasOneUse() && CI->getParent() == PredBB && | ||||
2376 | TLI->mayBeEmittedAsTailCall(CI) && | ||||
2377 | attributesPermitTailCall(F, CI, RetI, *TLI)) | ||||
2378 | TailCallBBs.push_back(PredBB); | ||||
2379 | } | ||||
2380 | } else { | ||||
2381 | SmallPtrSet<BasicBlock*, 4> VisitedBBs; | ||||
2382 | for (BasicBlock *Pred : predecessors(BB)) { | ||||
2383 | if (!VisitedBBs.insert(Pred).second) | ||||
2384 | continue; | ||||
2385 | if (Instruction *I = Pred->rbegin()->getPrevNonDebugInstruction(true)) { | ||||
2386 | CallInst *CI = dyn_cast<CallInst>(I); | ||||
2387 | if (CI && CI->use_empty() && TLI->mayBeEmittedAsTailCall(CI) && | ||||
2388 | attributesPermitTailCall(F, CI, RetI, *TLI)) | ||||
2389 | TailCallBBs.push_back(Pred); | ||||
2390 | } | ||||
2391 | } | ||||
2392 | } | ||||
2393 | |||||
2394 | bool Changed = false; | ||||
2395 | for (auto const &TailCallBB : TailCallBBs) { | ||||
2396 | // Make sure the call instruction is followed by an unconditional branch to | ||||
2397 | // the return block. | ||||
2398 | BranchInst *BI = dyn_cast<BranchInst>(TailCallBB->getTerminator()); | ||||
2399 | if (!BI || !BI->isUnconditional() || BI->getSuccessor(0) != BB) | ||||
2400 | continue; | ||||
2401 | |||||
2402 | // Duplicate the return into TailCallBB. | ||||
2403 | (void)FoldReturnIntoUncondBranch(RetI, BB, TailCallBB); | ||||
2404 | assert(!VerifyBFIUpdates ||((void)0) | ||||
2405 | BFI->getBlockFreq(BB) >= BFI->getBlockFreq(TailCallBB))((void)0); | ||||
2406 | BFI->setBlockFreq( | ||||
2407 | BB, | ||||
2408 | (BFI->getBlockFreq(BB) - BFI->getBlockFreq(TailCallBB)).getFrequency()); | ||||
2409 | ModifiedDT = Changed = true; | ||||
2410 | ++NumRetsDup; | ||||
2411 | } | ||||
2412 | |||||
2413 | // If we eliminated all predecessors of the block, delete the block now. | ||||
2414 | if (Changed && !BB->hasAddressTaken() && pred_empty(BB)) | ||||
2415 | BB->eraseFromParent(); | ||||
2416 | |||||
2417 | return Changed; | ||||
2418 | } | ||||
2419 | |||||
2420 | //===----------------------------------------------------------------------===// | ||||
2421 | // Memory Optimization | ||||
2422 | //===----------------------------------------------------------------------===// | ||||
2423 | |||||
2424 | namespace { | ||||
2425 | |||||
2426 | /// This is an extended version of TargetLowering::AddrMode | ||||
2427 | /// which holds actual Value*'s for register values. | ||||
2428 | struct ExtAddrMode : public TargetLowering::AddrMode { | ||||
2429 | Value *BaseReg = nullptr; | ||||
2430 | Value *ScaledReg = nullptr; | ||||
2431 | Value *OriginalValue = nullptr; | ||||
2432 | bool InBounds = true; | ||||
2433 | |||||
2434 | enum FieldName { | ||||
2435 | NoField = 0x00, | ||||
2436 | BaseRegField = 0x01, | ||||
2437 | BaseGVField = 0x02, | ||||
2438 | BaseOffsField = 0x04, | ||||
2439 | ScaledRegField = 0x08, | ||||
2440 | ScaleField = 0x10, | ||||
2441 | MultipleFields = 0xff | ||||
2442 | }; | ||||
2443 | |||||
2444 | |||||
2445 | ExtAddrMode() = default; | ||||
2446 | |||||
2447 | void print(raw_ostream &OS) const; | ||||
2448 | void dump() const; | ||||
2449 | |||||
2450 | FieldName compare(const ExtAddrMode &other) { | ||||
2451 | // First check that the types are the same on each field, as differing types | ||||
2452 | // is something we can't cope with later on. | ||||
2453 | if (BaseReg && other.BaseReg && | ||||
2454 | BaseReg->getType() != other.BaseReg->getType()) | ||||
2455 | return MultipleFields; | ||||
2456 | if (BaseGV && other.BaseGV && | ||||
2457 | BaseGV->getType() != other.BaseGV->getType()) | ||||
2458 | return MultipleFields; | ||||
2459 | if (ScaledReg && other.ScaledReg && | ||||
2460 | ScaledReg->getType() != other.ScaledReg->getType()) | ||||
2461 | return MultipleFields; | ||||
2462 | |||||
2463 | // Conservatively reject 'inbounds' mismatches. | ||||
2464 | if (InBounds != other.InBounds) | ||||
2465 | return MultipleFields; | ||||
2466 | |||||
2467 | // Check each field to see if it differs. | ||||
2468 | unsigned Result = NoField; | ||||
2469 | if (BaseReg != other.BaseReg) | ||||
2470 | Result |= BaseRegField; | ||||
2471 | if (BaseGV != other.BaseGV) | ||||
2472 | Result |= BaseGVField; | ||||
2473 | if (BaseOffs != other.BaseOffs) | ||||
2474 | Result |= BaseOffsField; | ||||
2475 | if (ScaledReg != other.ScaledReg) | ||||
2476 | Result |= ScaledRegField; | ||||
2477 | // Don't count 0 as being a different scale, because that actually means | ||||
2478 | // unscaled (which will already be counted by having no ScaledReg). | ||||
2479 | if (Scale && other.Scale && Scale != other.Scale) | ||||
2480 | Result |= ScaleField; | ||||
2481 | |||||
2482 | if (countPopulation(Result) > 1) | ||||
2483 | return MultipleFields; | ||||
2484 | else | ||||
2485 | return static_cast<FieldName>(Result); | ||||
2486 | } | ||||
2487 | |||||
2488 | // An AddrMode is trivial if it involves no calculation i.e. it is just a base | ||||
2489 | // with no offset. | ||||
2490 | bool isTrivial() { | ||||
2491 | // An AddrMode is (BaseGV + BaseReg + BaseOffs + ScaleReg * Scale) so it is | ||||
2492 | // trivial if at most one of these terms is nonzero, except that BaseGV and | ||||
2493 | // BaseReg both being zero actually means a null pointer value, which we | ||||
2494 | // consider to be 'non-zero' here. | ||||
2495 | return !BaseOffs && !Scale && !(BaseGV && BaseReg); | ||||
2496 | } | ||||
2497 | |||||
2498 | Value *GetFieldAsValue(FieldName Field, Type *IntPtrTy) { | ||||
2499 | switch (Field) { | ||||
2500 | default: | ||||
2501 | return nullptr; | ||||
2502 | case BaseRegField: | ||||
2503 | return BaseReg; | ||||
2504 | case BaseGVField: | ||||
2505 | return BaseGV; | ||||
2506 | case ScaledRegField: | ||||
2507 | return ScaledReg; | ||||
2508 | case BaseOffsField: | ||||
2509 | return ConstantInt::get(IntPtrTy, BaseOffs); | ||||
2510 | } | ||||
2511 | } | ||||
2512 | |||||
2513 | void SetCombinedField(FieldName Field, Value *V, | ||||
2514 | const SmallVectorImpl<ExtAddrMode> &AddrModes) { | ||||
2515 | switch (Field) { | ||||
2516 | default: | ||||
2517 | llvm_unreachable("Unhandled fields are expected to be rejected earlier")__builtin_unreachable(); | ||||
2518 | break; | ||||
2519 | case ExtAddrMode::BaseRegField: | ||||
2520 | BaseReg = V; | ||||
2521 | break; | ||||
2522 | case ExtAddrMode::BaseGVField: | ||||
2523 | // A combined BaseGV is an Instruction, not a GlobalValue, so it goes | ||||
2524 | // in the BaseReg field. | ||||
2525 | assert(BaseReg == nullptr)((void)0); | ||||
2526 | BaseReg = V; | ||||
2527 | BaseGV = nullptr; | ||||
2528 | break; | ||||
2529 | case ExtAddrMode::ScaledRegField: | ||||
2530 | ScaledReg = V; | ||||
2531 | // If we have a mix of scaled and unscaled addrmodes then we want scale | ||||
2532 | // to be the scale and not zero. | ||||
2533 | if (!Scale) | ||||
2534 | for (const ExtAddrMode &AM : AddrModes) | ||||
2535 | if (AM.Scale) { | ||||
2536 | Scale = AM.Scale; | ||||
2537 | break; | ||||
2538 | } | ||||
2539 | break; | ||||
2540 | case ExtAddrMode::BaseOffsField: | ||||
2541 | // The offset is no longer a constant, so it goes in ScaledReg with a | ||||
2542 | // scale of 1. | ||||
2543 | assert(ScaledReg == nullptr)((void)0); | ||||
2544 | ScaledReg = V; | ||||
2545 | Scale = 1; | ||||
2546 | BaseOffs = 0; | ||||
2547 | break; | ||||
2548 | } | ||||
2549 | } | ||||
2550 | }; | ||||
2551 | |||||
2552 | } // end anonymous namespace | ||||
2553 | |||||
2554 | #ifndef NDEBUG1 | ||||
2555 | static inline raw_ostream &operator<<(raw_ostream &OS, const ExtAddrMode &AM) { | ||||
2556 | AM.print(OS); | ||||
2557 | return OS; | ||||
2558 | } | ||||
2559 | #endif | ||||
2560 | |||||
2561 | #if !defined(NDEBUG1) || defined(LLVM_ENABLE_DUMP) | ||||
2562 | void ExtAddrMode::print(raw_ostream &OS) const { | ||||
2563 | bool NeedPlus = false; | ||||
2564 | OS << "["; | ||||
2565 | if (InBounds) | ||||
2566 | OS << "inbounds "; | ||||
2567 | if (BaseGV) { | ||||
2568 | OS << (NeedPlus ? " + " : "") | ||||
2569 | << "GV:"; | ||||
2570 | BaseGV->printAsOperand(OS, /*PrintType=*/false); | ||||
2571 | NeedPlus = true; | ||||
2572 | } | ||||
2573 | |||||
2574 | if (BaseOffs) { | ||||
2575 | OS << (NeedPlus ? " + " : "") | ||||
2576 | << BaseOffs; | ||||
2577 | NeedPlus = true; | ||||
2578 | } | ||||
2579 | |||||
2580 | if (BaseReg) { | ||||
2581 | OS << (NeedPlus ? " + " : "") | ||||
2582 | << "Base:"; | ||||
2583 | BaseReg->printAsOperand(OS, /*PrintType=*/false); | ||||
2584 | NeedPlus = true; | ||||
2585 | } | ||||
2586 | if (Scale) { | ||||
2587 | OS << (NeedPlus ? " + " : "") | ||||
2588 | << Scale << "*"; | ||||
2589 | ScaledReg->printAsOperand(OS, /*PrintType=*/false); | ||||
2590 | } | ||||
2591 | |||||
2592 | OS << ']'; | ||||
2593 | } | ||||
2594 | |||||
2595 | LLVM_DUMP_METHOD__attribute__((noinline)) void ExtAddrMode::dump() const { | ||||
2596 | print(dbgs()); | ||||
2597 | dbgs() << '\n'; | ||||
2598 | } | ||||
2599 | #endif | ||||
2600 | |||||
2601 | namespace { | ||||
2602 | |||||
2603 | /// This class provides transaction based operation on the IR. | ||||
2604 | /// Every change made through this class is recorded in the internal state and | ||||
2605 | /// can be undone (rollback) until commit is called. | ||||
2606 | /// CGP does not check if instructions could be speculatively executed when | ||||
2607 | /// moved. Preserving the original location would pessimize the debugging | ||||
2608 | /// experience, as well as negatively impact the quality of sample PGO. | ||||
2609 | class TypePromotionTransaction { | ||||
2610 | /// This represents the common interface of the individual transaction. | ||||
2611 | /// Each class implements the logic for doing one specific modification on | ||||
2612 | /// the IR via the TypePromotionTransaction. | ||||
2613 | class TypePromotionAction { | ||||
2614 | protected: | ||||
2615 | /// The Instruction modified. | ||||
2616 | Instruction *Inst; | ||||
2617 | |||||
2618 | public: | ||||
2619 | /// Constructor of the action. | ||||
2620 | /// The constructor performs the related action on the IR. | ||||
2621 | TypePromotionAction(Instruction *Inst) : Inst(Inst) {} | ||||
2622 | |||||
2623 | virtual ~TypePromotionAction() = default; | ||||
2624 | |||||
2625 | /// Undo the modification done by this action. | ||||
2626 | /// When this method is called, the IR must be in the same state as it was | ||||
2627 | /// before this action was applied. | ||||
2628 | /// \pre Undoing the action works if and only if the IR is in the exact same | ||||
2629 | /// state as it was directly after this action was applied. | ||||
2630 | virtual void undo() = 0; | ||||
2631 | |||||
2632 | /// Advocate every change made by this action. | ||||
2633 | /// When the results on the IR of the action are to be kept, it is important | ||||
2634 | /// to call this function, otherwise hidden information may be kept forever. | ||||
2635 | virtual void commit() { | ||||
2636 | // Nothing to be done, this action is not doing anything. | ||||
2637 | } | ||||
2638 | }; | ||||
2639 | |||||
2640 | /// Utility to remember the position of an instruction. | ||||
2641 | class InsertionHandler { | ||||
2642 | /// Position of an instruction. | ||||
2643 | /// Either an instruction: | ||||
2644 | /// - Is the first in a basic block: BB is used. | ||||
2645 | /// - Has a previous instruction: PrevInst is used. | ||||
2646 | union { | ||||
2647 | Instruction *PrevInst; | ||||
2648 | BasicBlock *BB; | ||||
2649 | } Point; | ||||
2650 | |||||
2651 | /// Remember whether or not the instruction had a previous instruction. | ||||
2652 | bool HasPrevInstruction; | ||||
2653 | |||||
2654 | public: | ||||
2655 | /// Record the position of \p Inst. | ||||
2656 | InsertionHandler(Instruction *Inst) { | ||||
2657 | BasicBlock::iterator It = Inst->getIterator(); | ||||
2658 | HasPrevInstruction = (It != (Inst->getParent()->begin())); | ||||
2659 | if (HasPrevInstruction) | ||||
2660 | Point.PrevInst = &*--It; | ||||
2661 | else | ||||
2662 | Point.BB = Inst->getParent(); | ||||
2663 | } | ||||
2664 | |||||
2665 | /// Insert \p Inst at the recorded position. | ||||
2666 | void insert(Instruction *Inst) { | ||||
2667 | if (HasPrevInstruction) { | ||||
2668 | if (Inst->getParent()) | ||||
2669 | Inst->removeFromParent(); | ||||
2670 | Inst->insertAfter(Point.PrevInst); | ||||
2671 | } else { | ||||
2672 | Instruction *Position = &*Point.BB->getFirstInsertionPt(); | ||||
2673 | if (Inst->getParent()) | ||||
2674 | Inst->moveBefore(Position); | ||||
2675 | else | ||||
2676 | Inst->insertBefore(Position); | ||||
2677 | } | ||||
2678 | } | ||||
2679 | }; | ||||
2680 | |||||
2681 | /// Move an instruction before another. | ||||
2682 | class InstructionMoveBefore : public TypePromotionAction { | ||||
2683 | /// Original position of the instruction. | ||||
2684 | InsertionHandler Position; | ||||
2685 | |||||
2686 | public: | ||||
2687 | /// Move \p Inst before \p Before. | ||||
2688 | InstructionMoveBefore(Instruction *Inst, Instruction *Before) | ||||
2689 | : TypePromotionAction(Inst), Position(Inst) { | ||||
2690 | LLVM_DEBUG(dbgs() << "Do: move: " << *Inst << "\nbefore: " << *Beforedo { } while (false) | ||||
2691 | << "\n")do { } while (false); | ||||
2692 | Inst->moveBefore(Before); | ||||
2693 | } | ||||
2694 | |||||
2695 | /// Move the instruction back to its original position. | ||||
2696 | void undo() override { | ||||
2697 | LLVM_DEBUG(dbgs() << "Undo: moveBefore: " << *Inst << "\n")do { } while (false); | ||||
2698 | Position.insert(Inst); | ||||
2699 | } | ||||
2700 | }; | ||||
2701 | |||||
2702 | /// Set the operand of an instruction with a new value. | ||||
2703 | class OperandSetter : public TypePromotionAction { | ||||
2704 | /// Original operand of the instruction. | ||||
2705 | Value *Origin; | ||||
2706 | |||||
2707 | /// Index of the modified instruction. | ||||
2708 | unsigned Idx; | ||||
2709 | |||||
2710 | public: | ||||
2711 | /// Set \p Idx operand of \p Inst with \p NewVal. | ||||
2712 | OperandSetter(Instruction *Inst, unsigned Idx, Value *NewVal) | ||||
2713 | : TypePromotionAction(Inst), Idx(Idx) { | ||||
2714 | LLVM_DEBUG(dbgs() << "Do: setOperand: " << Idx << "\n"do { } while (false) | ||||
2715 | << "for:" << *Inst << "\n"do { } while (false) | ||||
2716 | << "with:" << *NewVal << "\n")do { } while (false); | ||||
2717 | Origin = Inst->getOperand(Idx); | ||||
2718 | Inst->setOperand(Idx, NewVal); | ||||
2719 | } | ||||
2720 | |||||
2721 | /// Restore the original value of the instruction. | ||||
2722 | void undo() override { | ||||
2723 | LLVM_DEBUG(dbgs() << "Undo: setOperand:" << Idx << "\n"do { } while (false) | ||||
2724 | << "for: " << *Inst << "\n"do { } while (false) | ||||
2725 | << "with: " << *Origin << "\n")do { } while (false); | ||||
2726 | Inst->setOperand(Idx, Origin); | ||||
2727 | } | ||||
2728 | }; | ||||
2729 | |||||
2730 | /// Hide the operands of an instruction. | ||||
2731 | /// Do as if this instruction was not using any of its operands. | ||||
2732 | class OperandsHider : public TypePromotionAction { | ||||
2733 | /// The list of original operands. | ||||
2734 | SmallVector<Value *, 4> OriginalValues; | ||||
2735 | |||||
2736 | public: | ||||
2737 | /// Remove \p Inst from the uses of the operands of \p Inst. | ||||
2738 | OperandsHider(Instruction *Inst) : TypePromotionAction(Inst) { | ||||
2739 | LLVM_DEBUG(dbgs() << "Do: OperandsHider: " << *Inst << "\n")do { } while (false); | ||||
2740 | unsigned NumOpnds = Inst->getNumOperands(); | ||||
2741 | OriginalValues.reserve(NumOpnds); | ||||
2742 | for (unsigned It = 0; It < NumOpnds; ++It) { | ||||
2743 | // Save the current operand. | ||||
2744 | Value *Val = Inst->getOperand(It); | ||||
2745 | OriginalValues.push_back(Val); | ||||
2746 | // Set a dummy one. | ||||
2747 | // We could use OperandSetter here, but that would imply an overhead | ||||
2748 | // that we are not willing to pay. | ||||
2749 | Inst->setOperand(It, UndefValue::get(Val->getType())); | ||||
2750 | } | ||||
2751 | } | ||||
2752 | |||||
2753 | /// Restore the original list of uses. | ||||
2754 | void undo() override { | ||||
2755 | LLVM_DEBUG(dbgs() << "Undo: OperandsHider: " << *Inst << "\n")do { } while (false); | ||||
2756 | for (unsigned It = 0, EndIt = OriginalValues.size(); It != EndIt; ++It) | ||||
2757 | Inst->setOperand(It, OriginalValues[It]); | ||||
2758 | } | ||||
2759 | }; | ||||
2760 | |||||
2761 | /// Build a truncate instruction. | ||||
2762 | class TruncBuilder : public TypePromotionAction { | ||||
2763 | Value *Val; | ||||
2764 | |||||
2765 | public: | ||||
2766 | /// Build a truncate instruction of \p Opnd producing a \p Ty | ||||
2767 | /// result. | ||||
2768 | /// trunc Opnd to Ty. | ||||
2769 | TruncBuilder(Instruction *Opnd, Type *Ty) : TypePromotionAction(Opnd) { | ||||
2770 | IRBuilder<> Builder(Opnd); | ||||
2771 | Builder.SetCurrentDebugLocation(DebugLoc()); | ||||
2772 | Val = Builder.CreateTrunc(Opnd, Ty, "promoted"); | ||||
2773 | LLVM_DEBUG(dbgs() << "Do: TruncBuilder: " << *Val << "\n")do { } while (false); | ||||
2774 | } | ||||
2775 | |||||
2776 | /// Get the built value. | ||||
2777 | Value *getBuiltValue() { return Val; } | ||||
2778 | |||||
2779 | /// Remove the built instruction. | ||||
2780 | void undo() override { | ||||
2781 | LLVM_DEBUG(dbgs() << "Undo: TruncBuilder: " << *Val << "\n")do { } while (false); | ||||
2782 | if (Instruction *IVal = dyn_cast<Instruction>(Val)) | ||||
2783 | IVal->eraseFromParent(); | ||||
2784 | } | ||||
2785 | }; | ||||
2786 | |||||
2787 | /// Build a sign extension instruction. | ||||
2788 | class SExtBuilder : public TypePromotionAction { | ||||
2789 | Value *Val; | ||||
2790 | |||||
2791 | public: | ||||
2792 | /// Build a sign extension instruction of \p Opnd producing a \p Ty | ||||
2793 | /// result. | ||||
2794 | /// sext Opnd to Ty. | ||||
2795 | SExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty) | ||||
2796 | : TypePromotionAction(InsertPt) { | ||||
2797 | IRBuilder<> Builder(InsertPt); | ||||
2798 | Val = Builder.CreateSExt(Opnd, Ty, "promoted"); | ||||
2799 | LLVM_DEBUG(dbgs() << "Do: SExtBuilder: " << *Val << "\n")do { } while (false); | ||||
2800 | } | ||||
2801 | |||||
2802 | /// Get the built value. | ||||
2803 | Value *getBuiltValue() { return Val; } | ||||
2804 | |||||
2805 | /// Remove the built instruction. | ||||
2806 | void undo() override { | ||||
2807 | LLVM_DEBUG(dbgs() << "Undo: SExtBuilder: " << *Val << "\n")do { } while (false); | ||||
2808 | if (Instruction *IVal = dyn_cast<Instruction>(Val)) | ||||
2809 | IVal->eraseFromParent(); | ||||
2810 | } | ||||
2811 | }; | ||||
2812 | |||||
2813 | /// Build a zero extension instruction. | ||||
2814 | class ZExtBuilder : public TypePromotionAction { | ||||
2815 | Value *Val; | ||||
2816 | |||||
2817 | public: | ||||
2818 | /// Build a zero extension instruction of \p Opnd producing a \p Ty | ||||
2819 | /// result. | ||||
2820 | /// zext Opnd to Ty. | ||||
2821 | ZExtBuilder(Instruction *InsertPt, Value *Opnd, Type *Ty) | ||||
2822 | : TypePromotionAction(InsertPt) { | ||||
2823 | IRBuilder<> Builder(InsertPt); | ||||
2824 | Builder.SetCurrentDebugLocation(DebugLoc()); | ||||
2825 | Val = Builder.CreateZExt(Opnd, Ty, "promoted"); | ||||
2826 | LLVM_DEBUG(dbgs() << "Do: ZExtBuilder: " << *Val << "\n")do { } while (false); | ||||
2827 | } | ||||
2828 | |||||
2829 | /// Get the built value. | ||||
2830 | Value *getBuiltValue() { return Val; } | ||||
2831 | |||||
2832 | /// Remove the built instruction. | ||||
2833 | void undo() override { | ||||
2834 | LLVM_DEBUG(dbgs() << "Undo: ZExtBuilder: " << *Val << "\n")do { } while (false); | ||||
2835 | if (Instruction *IVal = dyn_cast<Instruction>(Val)) | ||||
2836 | IVal->eraseFromParent(); | ||||
2837 | } | ||||
2838 | }; | ||||
2839 | |||||
2840 | /// Mutate an instruction to another type. | ||||
2841 | class TypeMutator : public TypePromotionAction { | ||||
2842 | /// Record the original type. | ||||
2843 | Type *OrigTy; | ||||
2844 | |||||
2845 | public: | ||||
2846 | /// Mutate the type of \p Inst into \p NewTy. | ||||
2847 | TypeMutator(Instruction *Inst, Type *NewTy) | ||||
2848 | : TypePromotionAction(Inst), OrigTy(Inst->getType()) { | ||||
2849 | LLVM_DEBUG(dbgs() << "Do: MutateType: " << *Inst << " with " << *NewTydo { } while (false) | ||||
2850 | << "\n")do { } while (false); | ||||
2851 | Inst->mutateType(NewTy); | ||||
2852 | } | ||||
2853 | |||||
2854 | /// Mutate the instruction back to its original type. | ||||
2855 | void undo() override { | ||||
2856 | LLVM_DEBUG(dbgs() << "Undo: MutateType: " << *Inst << " with " << *OrigTydo { } while (false) | ||||
2857 | << "\n")do { } while (false); | ||||
2858 | Inst->mutateType(OrigTy); | ||||
2859 | } | ||||
2860 | }; | ||||
2861 | |||||
2862 | /// Replace the uses of an instruction by another instruction. | ||||
2863 | class UsesReplacer : public TypePromotionAction { | ||||
2864 | /// Helper structure to keep track of the replaced uses. | ||||
2865 | struct InstructionAndIdx { | ||||
2866 | /// The instruction using the instruction. | ||||
2867 | Instruction *Inst; | ||||
2868 | |||||
2869 | /// The index where this instruction is used for Inst. | ||||
2870 | unsigned Idx; | ||||
2871 | |||||
2872 | InstructionAndIdx(Instruction *Inst, unsigned Idx) | ||||
2873 | : Inst(Inst), Idx(Idx) {} | ||||
2874 | }; | ||||
2875 | |||||
2876 | /// Keep track of the original uses (pair Instruction, Index). | ||||
2877 | SmallVector<InstructionAndIdx, 4> OriginalUses; | ||||
2878 | /// Keep track of the debug users. | ||||
2879 | SmallVector<DbgValueInst *, 1> DbgValues; | ||||
2880 | |||||
2881 | /// Keep track of the new value so that we can undo it by replacing | ||||
2882 | /// instances of the new value with the original value. | ||||
2883 | Value *New; | ||||
2884 | |||||
2885 | using use_iterator = SmallVectorImpl<InstructionAndIdx>::iterator; | ||||
2886 | |||||
2887 | public: | ||||
2888 | /// Replace all the use of \p Inst by \p New. | ||||
2889 | UsesReplacer(Instruction *Inst, Value *New) | ||||
2890 | : TypePromotionAction(Inst), New(New) { | ||||
2891 | LLVM_DEBUG(dbgs() << "Do: UsersReplacer: " << *Inst << " with " << *Newdo { } while (false) | ||||
2892 | << "\n")do { } while (false); | ||||
2893 | // Record the original uses. | ||||
2894 | for (Use &U : Inst->uses()) { | ||||
2895 | Instruction *UserI = cast<Instruction>(U.getUser()); | ||||
2896 | OriginalUses.push_back(InstructionAndIdx(UserI, U.getOperandNo())); | ||||
2897 | } | ||||
2898 | // Record the debug uses separately. They are not in the instruction's | ||||
2899 | // use list, but they are replaced by RAUW. | ||||
2900 | findDbgValues(DbgValues, Inst); | ||||
2901 | |||||
2902 | // Now, we can replace the uses. | ||||
2903 | Inst->replaceAllUsesWith(New); | ||||
2904 | } | ||||
2905 | |||||
2906 | /// Reassign the original uses of Inst to Inst. | ||||
2907 | void undo() override { | ||||
2908 | LLVM_DEBUG(dbgs() << "Undo: UsersReplacer: " << *Inst << "\n")do { } while (false); | ||||
2909 | for (InstructionAndIdx &Use : OriginalUses) | ||||
2910 | Use.Inst->setOperand(Use.Idx, Inst); | ||||
2911 | // RAUW has replaced all original uses with references to the new value, | ||||
2912 | // including the debug uses. Since we are undoing the replacements, | ||||
2913 | // the original debug uses must also be reinstated to maintain the | ||||
2914 | // correctness and utility of debug value instructions. | ||||
2915 | for (auto *DVI : DbgValues) | ||||
2916 | DVI->replaceVariableLocationOp(New, Inst); | ||||
2917 | } | ||||
2918 | }; | ||||
2919 | |||||
2920 | /// Remove an instruction from the IR. | ||||
2921 | class InstructionRemover : public TypePromotionAction { | ||||
2922 | /// Original position of the instruction. | ||||
2923 | InsertionHandler Inserter; | ||||
2924 | |||||
2925 | /// Helper structure to hide all the link to the instruction. In other | ||||
2926 | /// words, this helps to do as if the instruction was removed. | ||||
2927 | OperandsHider Hider; | ||||
2928 | |||||
2929 | /// Keep track of the uses replaced, if any. | ||||
2930 | UsesReplacer *Replacer = nullptr; | ||||
2931 | |||||
2932 | /// Keep track of instructions removed. | ||||
2933 | SetOfInstrs &RemovedInsts; | ||||
2934 | |||||
2935 | public: | ||||
2936 | /// Remove all reference of \p Inst and optionally replace all its | ||||
2937 | /// uses with New. | ||||
2938 | /// \p RemovedInsts Keep track of the instructions removed by this Action. | ||||
2939 | /// \pre If !Inst->use_empty(), then New != nullptr | ||||
2940 | InstructionRemover(Instruction *Inst, SetOfInstrs &RemovedInsts, | ||||
2941 | Value *New = nullptr) | ||||
2942 | : TypePromotionAction(Inst), Inserter(Inst), Hider(Inst), | ||||
2943 | RemovedInsts(RemovedInsts) { | ||||
2944 | if (New) | ||||
2945 | Replacer = new UsesReplacer(Inst, New); | ||||
2946 | LLVM_DEBUG(dbgs() << "Do: InstructionRemover: " << *Inst << "\n")do { } while (false); | ||||
2947 | RemovedInsts.insert(Inst); | ||||
2948 | /// The instructions removed here will be freed after completing | ||||
2949 | /// optimizeBlock() for all blocks as we need to keep track of the | ||||
2950 | /// removed instructions during promotion. | ||||
2951 | Inst->removeFromParent(); | ||||
2952 | } | ||||
2953 | |||||
2954 | ~InstructionRemover() override { delete Replacer; } | ||||
2955 | |||||
2956 | /// Resurrect the instruction and reassign it to the proper uses if | ||||
2957 | /// new value was provided when build this action. | ||||
2958 | void undo() override { | ||||
2959 | LLVM_DEBUG(dbgs() << "Undo: InstructionRemover: " << *Inst << "\n")do { } while (false); | ||||
2960 | Inserter.insert(Inst); | ||||
2961 | if (Replacer) | ||||
2962 | Replacer->undo(); | ||||
2963 | Hider.undo(); | ||||
2964 | RemovedInsts.erase(Inst); | ||||
2965 | } | ||||
2966 | }; | ||||
2967 | |||||
2968 | public: | ||||
2969 | /// Restoration point. | ||||
2970 | /// The restoration point is a pointer to an action instead of an iterator | ||||
2971 | /// because the iterator may be invalidated but not the pointer. | ||||
2972 | using ConstRestorationPt = const TypePromotionAction *; | ||||
2973 | |||||
2974 | TypePromotionTransaction(SetOfInstrs &RemovedInsts) | ||||
2975 | : RemovedInsts(RemovedInsts) {} | ||||
2976 | |||||
2977 | /// Advocate every changes made in that transaction. Return true if any change | ||||
2978 | /// happen. | ||||
2979 | bool commit(); | ||||
2980 | |||||
2981 | /// Undo all the changes made after the given point. | ||||
2982 | void rollback(ConstRestorationPt Point); | ||||
2983 | |||||
2984 | /// Get the current restoration point. | ||||
2985 | ConstRestorationPt getRestorationPoint() const; | ||||
2986 | |||||
2987 | /// \name API for IR modification with state keeping to support rollback. | ||||
2988 | /// @{ | ||||
2989 | /// Same as Instruction::setOperand. | ||||
2990 | void setOperand(Instruction *Inst, unsigned Idx, Value *NewVal); | ||||
2991 | |||||
2992 | /// Same as Instruction::eraseFromParent. | ||||
2993 | void eraseInstruction(Instruction *Inst, Value *NewVal = nullptr); | ||||
2994 | |||||
2995 | /// Same as Value::replaceAllUsesWith. | ||||
2996 | void replaceAllUsesWith(Instruction *Inst, Value *New); | ||||
2997 | |||||
2998 | /// Same as Value::mutateType. | ||||
2999 | void mutateType(Instruction *Inst, Type *NewTy); | ||||
3000 | |||||
3001 | /// Same as IRBuilder::createTrunc. | ||||
3002 | Value *createTrunc(Instruction *Opnd, Type *Ty); | ||||
3003 | |||||
3004 | /// Same as IRBuilder::createSExt. | ||||
3005 | Value *createSExt(Instruction *Inst, Value *Opnd, Type *Ty); | ||||
3006 | |||||
3007 | /// Same as IRBuilder::createZExt. | ||||
3008 | Value *createZExt(Instruction *Inst, Value *Opnd, Type *Ty); | ||||
3009 | |||||
3010 | /// Same as Instruction::moveBefore. | ||||
3011 | void moveBefore(Instruction *Inst, Instruction *Before); | ||||
3012 | /// @} | ||||
3013 | |||||
3014 | private: | ||||
3015 | /// The ordered list of actions made so far. | ||||
3016 | SmallVector<std::unique_ptr<TypePromotionAction>, 16> Actions; | ||||
3017 | |||||
3018 | using CommitPt = SmallVectorImpl<std::unique_ptr<TypePromotionAction>>::iterator; | ||||
3019 | |||||
3020 | SetOfInstrs &RemovedInsts; | ||||
3021 | }; | ||||
3022 | |||||
3023 | } // end anonymous namespace | ||||
3024 | |||||
3025 | void TypePromotionTransaction::setOperand(Instruction *Inst, unsigned Idx, | ||||
3026 | Value *NewVal) { | ||||
3027 | Actions.push_back(std::make_unique<TypePromotionTransaction::OperandSetter>( | ||||
3028 | Inst, Idx, NewVal)); | ||||
3029 | } | ||||
3030 | |||||
3031 | void TypePromotionTransaction::eraseInstruction(Instruction *Inst, | ||||
3032 | Value *NewVal) { | ||||
3033 | Actions.push_back( | ||||
3034 | std::make_unique<TypePromotionTransaction::InstructionRemover>( | ||||
3035 | Inst, RemovedInsts, NewVal)); | ||||
3036 | } | ||||
3037 | |||||
3038 | void TypePromotionTransaction::replaceAllUsesWith(Instruction *Inst, | ||||
3039 | Value *New) { | ||||
3040 | Actions.push_back( | ||||
3041 | std::make_unique<TypePromotionTransaction::UsesReplacer>(Inst, New)); | ||||
3042 | } | ||||
3043 | |||||
3044 | void TypePromotionTransaction::mutateType(Instruction *Inst, Type *NewTy) { | ||||
3045 | Actions.push_back( | ||||
3046 | std::make_unique<TypePromotionTransaction::TypeMutator>(Inst, NewTy)); | ||||
3047 | } | ||||
3048 | |||||
3049 | Value *TypePromotionTransaction::createTrunc(Instruction *Opnd, | ||||
3050 | Type *Ty) { | ||||
3051 | std::unique_ptr<TruncBuilder> Ptr(new TruncBuilder(Opnd, Ty)); | ||||
3052 | Value *Val = Ptr->getBuiltValue(); | ||||
3053 | Actions.push_back(std::move(Ptr)); | ||||
3054 | return Val; | ||||
3055 | } | ||||
3056 | |||||
3057 | Value *TypePromotionTransaction::createSExt(Instruction *Inst, | ||||
3058 | Value *Opnd, Type *Ty) { | ||||
3059 | std::unique_ptr<SExtBuilder> Ptr(new SExtBuilder(Inst, Opnd, Ty)); | ||||
3060 | Value *Val = Ptr->getBuiltValue(); | ||||
3061 | Actions.push_back(std::move(Ptr)); | ||||
3062 | return Val; | ||||
3063 | } | ||||
3064 | |||||
3065 | Value *TypePromotionTransaction::createZExt(Instruction *Inst, | ||||
3066 | Value *Opnd, Type *Ty) { | ||||
3067 | std::unique_ptr<ZExtBuilder> Ptr(new ZExtBuilder(Inst, Opnd, Ty)); | ||||
3068 | Value *Val = Ptr->getBuiltValue(); | ||||
3069 | Actions.push_back(std::move(Ptr)); | ||||
3070 | return Val; | ||||
3071 | } | ||||
3072 | |||||
3073 | void TypePromotionTransaction::moveBefore(Instruction *Inst, | ||||
3074 | Instruction *Before) { | ||||
3075 | Actions.push_back( | ||||
3076 | std::make_unique<TypePromotionTransaction::InstructionMoveBefore>( | ||||
3077 | Inst, Before)); | ||||
3078 | } | ||||
3079 | |||||
3080 | TypePromotionTransaction::ConstRestorationPt | ||||
3081 | TypePromotionTransaction::getRestorationPoint() const { | ||||
3082 | return !Actions.empty() ? Actions.back().get() : nullptr; | ||||
3083 | } | ||||
3084 | |||||
3085 | bool TypePromotionTransaction::commit() { | ||||
3086 | for (std::unique_ptr<TypePromotionAction> &Action : Actions) | ||||
3087 | Action->commit(); | ||||
3088 | bool Modified = !Actions.empty(); | ||||
3089 | Actions.clear(); | ||||
3090 | return Modified; | ||||
3091 | } | ||||
3092 | |||||
3093 | void TypePromotionTransaction::rollback( | ||||
3094 | TypePromotionTransaction::ConstRestorationPt Point) { | ||||
3095 | while (!Actions.empty() && Point != Actions.back().get()) { | ||||
3096 | std::unique_ptr<TypePromotionAction> Curr = Actions.pop_back_val(); | ||||
3097 | Curr->undo(); | ||||
3098 | } | ||||
3099 | } | ||||
3100 | |||||
3101 | namespace { | ||||
3102 | |||||
3103 | /// A helper class for matching addressing modes. | ||||
3104 | /// | ||||
3105 | /// This encapsulates the logic for matching the target-legal addressing modes. | ||||
3106 | class AddressingModeMatcher { | ||||
3107 | SmallVectorImpl<Instruction*> &AddrModeInsts; | ||||
3108 | const TargetLowering &TLI; | ||||
3109 | const TargetRegisterInfo &TRI; | ||||
3110 | const DataLayout &DL; | ||||
3111 | const LoopInfo &LI; | ||||
3112 | const std::function<const DominatorTree &()> getDTFn; | ||||
3113 | |||||
3114 | /// AccessTy/MemoryInst - This is the type for the access (e.g. double) and | ||||
3115 | /// the memory instruction that we're computing this address for. | ||||
3116 | Type *AccessTy; | ||||
3117 | unsigned AddrSpace; | ||||
3118 | Instruction *MemoryInst; | ||||
3119 | |||||
3120 | /// This is the addressing mode that we're building up. This is | ||||
3121 | /// part of the return value of this addressing mode matching stuff. | ||||
3122 | ExtAddrMode &AddrMode; | ||||
3123 | |||||
3124 | /// The instructions inserted by other CodeGenPrepare optimizations. | ||||
3125 | const SetOfInstrs &InsertedInsts; | ||||
3126 | |||||
3127 | /// A map from the instructions to their type before promotion. | ||||
3128 | InstrToOrigTy &PromotedInsts; | ||||
3129 | |||||
3130 | /// The ongoing transaction where every action should be registered. | ||||
3131 | TypePromotionTransaction &TPT; | ||||
3132 | |||||
3133 | // A GEP which has too large offset to be folded into the addressing mode. | ||||
3134 | std::pair<AssertingVH<GetElementPtrInst>, int64_t> &LargeOffsetGEP; | ||||
3135 | |||||
3136 | /// This is set to true when we should not do profitability checks. | ||||
3137 | /// When true, IsProfitableToFoldIntoAddressingMode always returns true. | ||||
3138 | bool IgnoreProfitability; | ||||
3139 | |||||
3140 | /// True if we are optimizing for size. | ||||
3141 | bool OptSize; | ||||
3142 | |||||
3143 | ProfileSummaryInfo *PSI; | ||||
3144 | BlockFrequencyInfo *BFI; | ||||
3145 | |||||
3146 | AddressingModeMatcher( | ||||
3147 | SmallVectorImpl<Instruction *> &AMI, const TargetLowering &TLI, | ||||
3148 | const TargetRegisterInfo &TRI, const LoopInfo &LI, | ||||
3149 | const std::function<const DominatorTree &()> getDTFn, | ||||
3150 | Type *AT, unsigned AS, Instruction *MI, ExtAddrMode &AM, | ||||
3151 | const SetOfInstrs &InsertedInsts, InstrToOrigTy &PromotedInsts, | ||||
3152 | TypePromotionTransaction &TPT, | ||||
3153 | std::pair<AssertingVH<GetElementPtrInst>, int64_t> &LargeOffsetGEP, | ||||
3154 | bool OptSize, ProfileSummaryInfo *PSI, BlockFrequencyInfo *BFI) | ||||
3155 | : AddrModeInsts(AMI), TLI(TLI), TRI(TRI), | ||||
3156 | DL(MI->getModule()->getDataLayout()), LI(LI), getDTFn(getDTFn), | ||||
3157 | AccessTy(AT), AddrSpace(AS), MemoryInst(MI), AddrMode(AM), | ||||
3158 | InsertedInsts(InsertedInsts), PromotedInsts(PromotedInsts), TPT(TPT), | ||||
3159 | LargeOffsetGEP(LargeOffsetGEP), OptSize(OptSize), PSI(PSI), BFI(BFI) { | ||||
3160 | IgnoreProfitability = false; | ||||
3161 | } | ||||
3162 | |||||
3163 | public: | ||||
3164 | /// Find the maximal addressing mode that a load/store of V can fold, | ||||
3165 | /// give an access type of AccessTy. This returns a list of involved | ||||
3166 | /// instructions in AddrModeInsts. | ||||
3167 | /// \p InsertedInsts The instructions inserted by other CodeGenPrepare | ||||
3168 | /// optimizations. | ||||
3169 | /// \p PromotedInsts maps the instructions to their type before promotion. | ||||
3170 | /// \p The ongoing transaction where every action should be registered. | ||||
3171 | static ExtAddrMode | ||||
3172 | Match(Value *V, Type *AccessTy, unsigned AS, Instruction *MemoryInst, | ||||
3173 | SmallVectorImpl<Instruction *> &AddrModeInsts, | ||||
3174 | const TargetLowering &TLI, const LoopInfo &LI, | ||||
3175 | const std::function<const DominatorTree &()> getDTFn, | ||||
3176 | const TargetRegisterInfo &TRI, const SetOfInstrs &InsertedInsts, | ||||
3177 | InstrToOrigTy &PromotedInsts, TypePromotionTransaction &TPT, | ||||
3178 | std::pair<AssertingVH<GetElementPtrInst>, int64_t> &LargeOffsetGEP, | ||||
3179 | bool OptSize, ProfileSummaryInfo *PSI, BlockFrequencyInfo *BFI) { | ||||
3180 | ExtAddrMode Result; | ||||
3181 | |||||
3182 | bool Success = AddressingModeMatcher( | ||||
3183 | AddrModeInsts, TLI, TRI, LI, getDTFn, AccessTy, AS, MemoryInst, Result, | ||||
3184 | InsertedInsts, PromotedInsts, TPT, LargeOffsetGEP, OptSize, PSI, | ||||
3185 | BFI).matchAddr(V, 0); | ||||
3186 | (void)Success; assert(Success && "Couldn't select *anything*?")((void)0); | ||||
3187 | return Result; | ||||
3188 | } | ||||
3189 | |||||
3190 | private: | ||||
3191 | bool matchScaledValue(Value *ScaleReg, int64_t Scale, unsigned Depth); | ||||
3192 | bool matchAddr(Value *Addr, unsigned Depth); | ||||
3193 | bool matchOperationAddr(User *AddrInst, unsigned Opcode, unsigned Depth, | ||||
3194 | bool *MovedAway = nullptr); | ||||
3195 | bool isProfitableToFoldIntoAddressingMode(Instruction *I, | ||||
3196 | ExtAddrMode &AMBefore, | ||||
3197 | ExtAddrMode &AMAfter); | ||||
3198 | bool valueAlreadyLiveAtInst(Value *Val, Value *KnownLive1, Value *KnownLive2); | ||||
3199 | bool isPromotionProfitable(unsigned NewCost, unsigned OldCost, | ||||
3200 | Value *PromotedOperand) const; | ||||
3201 | }; | ||||
3202 | |||||
3203 | class PhiNodeSet; | ||||
3204 | |||||
3205 | /// An iterator for PhiNodeSet. | ||||
3206 | class PhiNodeSetIterator { | ||||
3207 | PhiNodeSet * const Set; | ||||
3208 | size_t CurrentIndex = 0; | ||||
3209 | |||||
3210 | public: | ||||
3211 | /// The constructor. Start should point to either a valid element, or be equal | ||||
3212 | /// to the size of the underlying SmallVector of the PhiNodeSet. | ||||
3213 | PhiNodeSetIterator(PhiNodeSet * const Set, size_t Start); | ||||
3214 | PHINode * operator*() const; | ||||
3215 | PhiNodeSetIterator& operator++(); | ||||
3216 | bool operator==(const PhiNodeSetIterator &RHS) const; | ||||
3217 | bool operator!=(const PhiNodeSetIterator &RHS) const; | ||||
3218 | }; | ||||
3219 | |||||
3220 | /// Keeps a set of PHINodes. | ||||
3221 | /// | ||||
3222 | /// This is a minimal set implementation for a specific use case: | ||||
3223 | /// It is very fast when there are very few elements, but also provides good | ||||
3224 | /// performance when there are many. It is similar to SmallPtrSet, but also | ||||
3225 | /// provides iteration by insertion order, which is deterministic and stable | ||||
3226 | /// across runs. It is also similar to SmallSetVector, but provides removing | ||||
3227 | /// elements in O(1) time. This is achieved by not actually removing the element | ||||
3228 | /// from the underlying vector, so comes at the cost of using more memory, but | ||||
3229 | /// that is fine, since PhiNodeSets are used as short lived objects. | ||||
3230 | class PhiNodeSet { | ||||
3231 | friend class PhiNodeSetIterator; | ||||
3232 | |||||
3233 | using MapType = SmallDenseMap<PHINode *, size_t, 32>; | ||||
3234 | using iterator = PhiNodeSetIterator; | ||||
3235 | |||||
3236 | /// Keeps the elements in the order of their insertion in the underlying | ||||
3237 | /// vector. To achieve constant time removal, it never deletes any element. | ||||
3238 | SmallVector<PHINode *, 32> NodeList; | ||||
3239 | |||||
3240 | /// Keeps the elements in the underlying set implementation. This (and not the | ||||
3241 | /// NodeList defined above) is the source of truth on whether an element | ||||
3242 | /// is actually in the collection. | ||||
3243 | MapType NodeMap; | ||||
3244 | |||||
3245 | /// Points to the first valid (not deleted) element when the set is not empty | ||||
3246 | /// and the value is not zero. Equals to the size of the underlying vector | ||||
3247 | /// when the set is empty. When the value is 0, as in the beginning, the | ||||
3248 | /// first element may or may not be valid. | ||||
3249 | size_t FirstValidElement = 0; | ||||
3250 | |||||
3251 | public: | ||||
3252 | /// Inserts a new element to the collection. | ||||
3253 | /// \returns true if the element is actually added, i.e. was not in the | ||||
3254 | /// collection before the operation. | ||||
3255 | bool insert(PHINode *Ptr) { | ||||
3256 | if (NodeMap.insert(std::make_pair(Ptr, NodeList.size())).second) { | ||||
3257 | NodeList.push_back(Ptr); | ||||
3258 | return true; | ||||
3259 | } | ||||
3260 | return false; | ||||
3261 | } | ||||
3262 | |||||
3263 | /// Removes the element from the collection. | ||||
3264 | /// \returns whether the element is actually removed, i.e. was in the | ||||
3265 | /// collection before the operation. | ||||
3266 | bool erase(PHINode *Ptr) { | ||||
3267 | if (NodeMap.erase(Ptr)) { | ||||
3268 | SkipRemovedElements(FirstValidElement); | ||||
3269 | return true; | ||||
3270 | } | ||||
3271 | return false; | ||||
3272 | } | ||||
3273 | |||||
3274 | /// Removes all elements and clears the collection. | ||||
3275 | void clear() { | ||||
3276 | NodeMap.clear(); | ||||
3277 | NodeList.clear(); | ||||
3278 | FirstValidElement = 0; | ||||
3279 | } | ||||
3280 | |||||
3281 | /// \returns an iterator that will iterate the elements in the order of | ||||
3282 | /// insertion. | ||||
3283 | iterator begin() { | ||||
3284 | if (FirstValidElement == 0) | ||||
3285 | SkipRemovedElements(FirstValidElement); | ||||
3286 | return PhiNodeSetIterator(this, FirstValidElement); | ||||
3287 | } | ||||
3288 | |||||
3289 | /// \returns an iterator that points to the end of the collection. | ||||
3290 | iterator end() { return PhiNodeSetIterator(this, NodeList.size()); } | ||||
3291 | |||||
3292 | /// Returns the number of elements in the collection. | ||||
3293 | size_t size() const { | ||||
3294 | return NodeMap.size(); | ||||
3295 | } | ||||
3296 | |||||
3297 | /// \returns 1 if the given element is in the collection, and 0 if otherwise. | ||||
3298 | size_t count(PHINode *Ptr) const { | ||||
3299 | return NodeMap.count(Ptr); | ||||
3300 | } | ||||
3301 | |||||
3302 | private: | ||||
3303 | /// Updates the CurrentIndex so that it will point to a valid element. | ||||
3304 | /// | ||||
3305 | /// If the element of NodeList at CurrentIndex is valid, it does not | ||||
3306 | /// change it. If there are no more valid elements, it updates CurrentIndex | ||||
3307 | /// to point to the end of the NodeList. | ||||
3308 | void SkipRemovedElements(size_t &CurrentIndex) { | ||||
3309 | while (CurrentIndex < NodeList.size()) { | ||||
3310 | auto it = NodeMap.find(NodeList[CurrentIndex]); | ||||
3311 | // If the element has been deleted and added again later, NodeMap will | ||||
3312 | // point to a different index, so CurrentIndex will still be invalid. | ||||
3313 | if (it != NodeMap.end() && it->second == CurrentIndex) | ||||
3314 | break; | ||||
3315 | ++CurrentIndex; | ||||
3316 | } | ||||
3317 | } | ||||
3318 | }; | ||||
3319 | |||||
3320 | PhiNodeSetIterator::PhiNodeSetIterator(PhiNodeSet *const Set, size_t Start) | ||||
3321 | : Set(Set), CurrentIndex(Start) {} | ||||
3322 | |||||
3323 | PHINode * PhiNodeSetIterator::operator*() const { | ||||
3324 | assert(CurrentIndex < Set->NodeList.size() &&((void)0) | ||||
3325 | "PhiNodeSet access out of range")((void)0); | ||||
3326 | return Set->NodeList[CurrentIndex]; | ||||
3327 | } | ||||
3328 | |||||
3329 | PhiNodeSetIterator& PhiNodeSetIterator::operator++() { | ||||
3330 | assert(CurrentIndex < Set->NodeList.size() &&((void)0) | ||||
3331 | "PhiNodeSet access out of range")((void)0); | ||||
3332 | ++CurrentIndex; | ||||
3333 | Set->SkipRemovedElements(CurrentIndex); | ||||
3334 | return *this; | ||||
3335 | } | ||||
3336 | |||||
3337 | bool PhiNodeSetIterator::operator==(const PhiNodeSetIterator &RHS) const { | ||||
3338 | return CurrentIndex == RHS.CurrentIndex; | ||||
3339 | } | ||||
3340 | |||||
3341 | bool PhiNodeSetIterator::operator!=(const PhiNodeSetIterator &RHS) const { | ||||
3342 | return !((*this) == RHS); | ||||
3343 | } | ||||
3344 | |||||
3345 | /// Keep track of simplification of Phi nodes. | ||||
3346 | /// Accept the set of all phi nodes and erase phi node from this set | ||||
3347 | /// if it is simplified. | ||||
3348 | class SimplificationTracker { | ||||
3349 | DenseMap<Value *, Value *> Storage; | ||||
3350 | const SimplifyQuery &SQ; | ||||
3351 | // Tracks newly created Phi nodes. The elements are iterated by insertion | ||||
3352 | // order. | ||||
3353 | PhiNodeSet AllPhiNodes; | ||||
3354 | // Tracks newly created Select nodes. | ||||
3355 | SmallPtrSet<SelectInst *, 32> AllSelectNodes; | ||||
3356 | |||||
3357 | public: | ||||
3358 | SimplificationTracker(const SimplifyQuery &sq) | ||||
3359 | : SQ(sq) {} | ||||
3360 | |||||
3361 | Value *Get(Value *V) { | ||||
3362 | do { | ||||
3363 | auto SV = Storage.find(V); | ||||
3364 | if (SV == Storage.end()) | ||||
3365 | return V; | ||||
3366 | V = SV->second; | ||||
3367 | } while (true); | ||||
3368 | } | ||||
3369 | |||||
3370 | Value *Simplify(Value *Val) { | ||||
3371 | SmallVector<Value *, 32> WorkList; | ||||
3372 | SmallPtrSet<Value *, 32> Visited; | ||||
3373 | WorkList.push_back(Val); | ||||
3374 | while (!WorkList.empty()) { | ||||
3375 | auto *P = WorkList.pop_back_val(); | ||||
3376 | if (!Visited.insert(P).second) | ||||
3377 | continue; | ||||
3378 | if (auto *PI = dyn_cast<Instruction>(P)) | ||||
3379 | if (Value *V = SimplifyInstruction(cast<Instruction>(PI), SQ)) { | ||||
3380 | for (auto *U : PI->users()) | ||||
3381 | WorkList.push_back(cast<Value>(U)); | ||||
3382 | Put(PI, V); | ||||
3383 | PI->replaceAllUsesWith(V); | ||||
3384 | if (auto *PHI = dyn_cast<PHINode>(PI)) | ||||
3385 | AllPhiNodes.erase(PHI); | ||||
3386 | if (auto *Select = dyn_cast<SelectInst>(PI)) | ||||
3387 | AllSelectNodes.erase(Select); | ||||
3388 | PI->eraseFromParent(); | ||||
3389 | } | ||||
3390 | } | ||||
3391 | return Get(Val); | ||||
3392 | } | ||||
3393 | |||||
3394 | void Put(Value *From, Value *To) { | ||||
3395 | Storage.insert({ From, To }); | ||||
3396 | } | ||||
3397 | |||||
3398 | void ReplacePhi(PHINode *From, PHINode *To) { | ||||
3399 | Value* OldReplacement = Get(From); | ||||
3400 | while (OldReplacement != From) { | ||||
3401 | From = To; | ||||
3402 | To = dyn_cast<PHINode>(OldReplacement); | ||||
3403 | OldReplacement = Get(From); | ||||
3404 | } | ||||
3405 | assert(To && Get(To) == To && "Replacement PHI node is already replaced.")((void)0); | ||||
3406 | Put(From, To); | ||||
3407 | From->replaceAllUsesWith(To); | ||||
3408 | AllPhiNodes.erase(From); | ||||
3409 | From->eraseFromParent(); | ||||
3410 | } | ||||
3411 | |||||
3412 | PhiNodeSet& newPhiNodes() { return AllPhiNodes; } | ||||
3413 | |||||
3414 | void insertNewPhi(PHINode *PN) { AllPhiNodes.insert(PN); } | ||||
3415 | |||||
3416 | void insertNewSelect(SelectInst *SI) { AllSelectNodes.insert(SI); } | ||||
3417 | |||||
3418 | unsigned countNewPhiNodes() const { return AllPhiNodes.size(); } | ||||
3419 | |||||
3420 | unsigned countNewSelectNodes() const { return AllSelectNodes.size(); } | ||||
3421 | |||||
3422 | void destroyNewNodes(Type *CommonType) { | ||||
3423 | // For safe erasing, replace the uses with dummy value first. | ||||
3424 | auto *Dummy = UndefValue::get(CommonType); | ||||
3425 | for (auto *I : AllPhiNodes) { | ||||
3426 | I->replaceAllUsesWith(Dummy); | ||||
3427 | I->eraseFromParent(); | ||||
3428 | } | ||||
3429 | AllPhiNodes.clear(); | ||||
3430 | for (auto *I : AllSelectNodes) { | ||||
3431 | I->replaceAllUsesWith(Dummy); | ||||
3432 | I->eraseFromParent(); | ||||
3433 | } | ||||
3434 | AllSelectNodes.clear(); | ||||
3435 | } | ||||
3436 | }; | ||||
3437 | |||||
3438 | /// A helper class for combining addressing modes. | ||||
3439 | class AddressingModeCombiner { | ||||
3440 | typedef DenseMap<Value *, Value *> FoldAddrToValueMapping; | ||||
3441 | typedef std::pair<PHINode *, PHINode *> PHIPair; | ||||
3442 | |||||
3443 | private: | ||||
3444 | /// The addressing modes we've collected. | ||||
3445 | SmallVector<ExtAddrMode, 16> AddrModes; | ||||
3446 | |||||
3447 | /// The field in which the AddrModes differ, when we have more than one. | ||||
3448 | ExtAddrMode::FieldName DifferentField = ExtAddrMode::NoField; | ||||
3449 | |||||
3450 | /// Are the AddrModes that we have all just equal to their original values? | ||||
3451 | bool AllAddrModesTrivial = true; | ||||
3452 | |||||
3453 | /// Common Type for all different fields in addressing modes. | ||||
3454 | Type *CommonType; | ||||
3455 | |||||
3456 | /// SimplifyQuery for simplifyInstruction utility. | ||||
3457 | const SimplifyQuery &SQ; | ||||
3458 | |||||
3459 | /// Original Address. | ||||
3460 | Value *Original; | ||||
3461 | |||||
3462 | public: | ||||
3463 | AddressingModeCombiner(const SimplifyQuery &_SQ, Value *OriginalValue) | ||||
3464 | : CommonType(nullptr), SQ(_SQ), Original(OriginalValue) {} | ||||
3465 | |||||
3466 | /// Get the combined AddrMode | ||||
3467 | const ExtAddrMode &getAddrMode() const { | ||||
3468 | return AddrModes[0]; | ||||
3469 | } | ||||
3470 | |||||
3471 | /// Add a new AddrMode if it's compatible with the AddrModes we already | ||||
3472 | /// have. | ||||
3473 | /// \return True iff we succeeded in doing so. | ||||
3474 | bool addNewAddrMode(ExtAddrMode &NewAddrMode) { | ||||
3475 | // Take note of if we have any non-trivial AddrModes, as we need to detect | ||||
3476 | // when all AddrModes are trivial as then we would introduce a phi or select | ||||
3477 | // which just duplicates what's already there. | ||||
3478 | AllAddrModesTrivial = AllAddrModesTrivial && NewAddrMode.isTrivial(); | ||||
3479 | |||||
3480 | // If this is the first addrmode then everything is fine. | ||||
3481 | if (AddrModes.empty()) { | ||||
3482 | AddrModes.emplace_back(NewAddrMode); | ||||
3483 | return true; | ||||
3484 | } | ||||
3485 | |||||
3486 | // Figure out how different this is from the other address modes, which we | ||||
3487 | // can do just by comparing against the first one given that we only care | ||||
3488 | // about the cumulative difference. | ||||
3489 | ExtAddrMode::FieldName ThisDifferentField = | ||||
3490 | AddrModes[0].compare(NewAddrMode); | ||||
3491 | if (DifferentField == ExtAddrMode::NoField) | ||||
3492 | DifferentField = ThisDifferentField; | ||||
3493 | else if (DifferentField != ThisDifferentField) | ||||
3494 | DifferentField = ExtAddrMode::MultipleFields; | ||||
3495 | |||||
3496 | // If NewAddrMode differs in more than one dimension we cannot handle it. | ||||
3497 | bool CanHandle = DifferentField != ExtAddrMode::MultipleFields; | ||||
3498 | |||||
3499 | // If Scale Field is different then we reject. | ||||
3500 | CanHandle = CanHandle && DifferentField != ExtAddrMode::ScaleField; | ||||
3501 | |||||
3502 | // We also must reject the case when base offset is different and | ||||
3503 | // scale reg is not null, we cannot handle this case due to merge of | ||||
3504 | // different offsets will be used as ScaleReg. | ||||
3505 | CanHandle = CanHandle && (DifferentField != ExtAddrMode::BaseOffsField || | ||||
3506 | !NewAddrMode.ScaledReg); | ||||
3507 | |||||
3508 | // We also must reject the case when GV is different and BaseReg installed | ||||
3509 | // due to we want to use base reg as a merge of GV values. | ||||
3510 | CanHandle = CanHandle && (DifferentField != ExtAddrMode::BaseGVField || | ||||
3511 | !NewAddrMode.HasBaseReg); | ||||
3512 | |||||
3513 | // Even if NewAddMode is the same we still need to collect it due to | ||||
3514 | // original value is different. And later we will need all original values | ||||
3515 | // as anchors during finding the common Phi node. | ||||
3516 | if (CanHandle) | ||||
3517 | AddrModes.emplace_back(NewAddrMode); | ||||
3518 | else | ||||
3519 | AddrModes.clear(); | ||||
3520 | |||||
3521 | return CanHandle; | ||||
3522 | } | ||||
3523 | |||||
3524 | /// Combine the addressing modes we've collected into a single | ||||
3525 | /// addressing mode. | ||||
3526 | /// \return True iff we successfully combined them or we only had one so | ||||
3527 | /// didn't need to combine them anyway. | ||||
3528 | bool combineAddrModes() { | ||||
3529 | // If we have no AddrModes then they can't be combined. | ||||
3530 | if (AddrModes.size() == 0) | ||||
3531 | return false; | ||||
3532 | |||||
3533 | // A single AddrMode can trivially be combined. | ||||
3534 | if (AddrModes.size() == 1 || DifferentField == ExtAddrMode::NoField) | ||||
3535 | return true; | ||||
3536 | |||||
3537 | // If the AddrModes we collected are all just equal to the value they are | ||||
3538 | // derived from then combining them wouldn't do anything useful. | ||||
3539 | if (AllAddrModesTrivial) | ||||
3540 | return false; | ||||
3541 | |||||
3542 | if (!addrModeCombiningAllowed()) | ||||
3543 | return false; | ||||
3544 | |||||
3545 | // Build a map between <original value, basic block where we saw it> to | ||||
3546 | // value of base register. | ||||
3547 | // Bail out if there is no common type. | ||||
3548 | FoldAddrToValueMapping Map; | ||||
3549 | if (!initializeMap(Map)) | ||||
3550 | return false; | ||||
3551 | |||||
3552 | Value *CommonValue = findCommon(Map); | ||||
3553 | if (CommonValue) | ||||
3554 | AddrModes[0].SetCombinedField(DifferentField, CommonValue, AddrModes); | ||||
3555 | return CommonValue != nullptr; | ||||
3556 | } | ||||
3557 | |||||
3558 | private: | ||||
3559 | /// Initialize Map with anchor values. For address seen | ||||
3560 | /// we set the value of different field saw in this address. | ||||
3561 | /// At the same time we find a common type for different field we will | ||||
3562 | /// use to create new Phi/Select nodes. Keep it in CommonType field. | ||||
3563 | /// Return false if there is no common type found. | ||||
3564 | bool initializeMap(FoldAddrToValueMapping &Map) { | ||||
3565 | // Keep track of keys where the value is null. We will need to replace it | ||||
3566 | // with constant null when we know the common type. | ||||
3567 | SmallVector<Value *, 2> NullValue; | ||||
3568 | Type *IntPtrTy = SQ.DL.getIntPtrType(AddrModes[0].OriginalValue->getType()); | ||||
3569 | for (auto &AM : AddrModes) { | ||||
3570 | Value *DV = AM.GetFieldAsValue(DifferentField, IntPtrTy); | ||||
3571 | if (DV) { | ||||
3572 | auto *Type = DV->getType(); | ||||
3573 | if (CommonType && CommonType != Type) | ||||
3574 | return false; | ||||
3575 | CommonType = Type; | ||||
3576 | Map[AM.OriginalValue] = DV; | ||||
3577 | } else { | ||||
3578 | NullValue.push_back(AM.OriginalValue); | ||||
3579 | } | ||||
3580 | } | ||||
3581 | assert(CommonType && "At least one non-null value must be!")((void)0); | ||||
3582 | for (auto *V : NullValue) | ||||
3583 | Map[V] = Constant::getNullValue(CommonType); | ||||
3584 | return true; | ||||
3585 | } | ||||
3586 | |||||
3587 | /// We have mapping between value A and other value B where B was a field in | ||||
3588 | /// addressing mode represented by A. Also we have an original value C | ||||
3589 | /// representing an address we start with. Traversing from C through phi and | ||||
3590 | /// selects we ended up with A's in a map. This utility function tries to find | ||||
3591 | /// a value V which is a field in addressing mode C and traversing through phi | ||||
3592 | /// nodes and selects we will end up in corresponded values B in a map. | ||||
3593 | /// The utility will create a new Phi/Selects if needed. | ||||
3594 | // The simple example looks as follows: | ||||
3595 | // BB1: | ||||
3596 | // p1 = b1 + 40 | ||||
3597 | // br cond BB2, BB3 | ||||
3598 | // BB2: | ||||
3599 | // p2 = b2 + 40 | ||||
3600 | // br BB3 | ||||
3601 | // BB3: | ||||
3602 | // p = phi [p1, BB1], [p2, BB2] | ||||
3603 | // v = load p | ||||
3604 | // Map is | ||||
3605 | // p1 -> b1 | ||||
3606 | // p2 -> b2 | ||||
3607 | // Request is | ||||
3608 | // p -> ? | ||||
3609 | // The function tries to find or build phi [b1, BB1], [b2, BB2] in BB3. | ||||
3610 | Value *findCommon(FoldAddrToValueMapping &Map) { | ||||
3611 | // Tracks the simplification of newly created phi nodes. The reason we use | ||||
3612 | // this mapping is because we will add new created Phi nodes in AddrToBase. | ||||
3613 | // Simplification of Phi nodes is recursive, so some Phi node may | ||||
3614 | // be simplified after we added it to AddrToBase. In reality this | ||||
3615 | // simplification is possible only if original phi/selects were not | ||||
3616 | // simplified yet. | ||||
3617 | // Using this mapping we can find the current value in AddrToBase. | ||||
3618 | SimplificationTracker ST(SQ); | ||||
3619 | |||||
3620 | // First step, DFS to create PHI nodes for all intermediate blocks. | ||||
3621 | // Also fill traverse order for the second step. | ||||
3622 | SmallVector<Value *, 32> TraverseOrder; | ||||
3623 | InsertPlaceholders(Map, TraverseOrder, ST); | ||||
3624 | |||||
3625 | // Second Step, fill new nodes by merged values and simplify if possible. | ||||
3626 | FillPlaceholders(Map, TraverseOrder, ST); | ||||
3627 | |||||
3628 | if (!AddrSinkNewSelects && ST.countNewSelectNodes() > 0) { | ||||
3629 | ST.destroyNewNodes(CommonType); | ||||
3630 | return nullptr; | ||||
3631 | } | ||||
3632 | |||||
3633 | // Now we'd like to match New Phi nodes to existed ones. | ||||
3634 | unsigned PhiNotMatchedCount = 0; | ||||
3635 | if (!MatchPhiSet(ST, AddrSinkNewPhis, PhiNotMatchedCount)) { | ||||
3636 | ST.destroyNewNodes(CommonType); | ||||
3637 | return nullptr; | ||||
3638 | } | ||||
3639 | |||||
3640 | auto *Result = ST.Get(Map.find(Original)->second); | ||||
3641 | if (Result) { | ||||
3642 | NumMemoryInstsPhiCreated += ST.countNewPhiNodes() + PhiNotMatchedCount; | ||||
3643 | NumMemoryInstsSelectCreated += ST.countNewSelectNodes(); | ||||
3644 | } | ||||
3645 | return Result; | ||||
3646 | } | ||||
3647 | |||||
3648 | /// Try to match PHI node to Candidate. | ||||
3649 | /// Matcher tracks the matched Phi nodes. | ||||
3650 | bool MatchPhiNode(PHINode *PHI, PHINode *Candidate, | ||||
3651 | SmallSetVector<PHIPair, 8> &Matcher, | ||||
3652 | PhiNodeSet &PhiNodesToMatch) { | ||||
3653 | SmallVector<PHIPair, 8> WorkList; | ||||
3654 | Matcher.insert({ PHI, Candidate }); | ||||
3655 | SmallSet<PHINode *, 8> MatchedPHIs; | ||||
3656 | MatchedPHIs.insert(PHI); | ||||
3657 | WorkList.push_back({ PHI, Candidate }); | ||||
3658 | SmallSet<PHIPair, 8> Visited; | ||||
3659 | while (!WorkList.empty()) { | ||||
3660 | auto Item = WorkList.pop_back_val(); | ||||
3661 | if (!Visited.insert(Item).second) | ||||
3662 | continue; | ||||
3663 | // We iterate over all incoming values to Phi to compare them. | ||||
3664 | // If values are different and both of them Phi and the first one is a | ||||
3665 | // Phi we added (subject to match) and both of them is in the same basic | ||||
3666 | // block then we can match our pair if values match. So we state that | ||||
3667 | // these values match and add it to work list to verify that. | ||||
3668 | for (auto B : Item.first->blocks()) { | ||||
3669 | Value *FirstValue = Item.first->getIncomingValueForBlock(B); | ||||
3670 | Value *SecondValue = Item.second->getIncomingValueForBlock(B); | ||||
3671 | if (FirstValue == SecondValue) | ||||
3672 | continue; | ||||
3673 | |||||
3674 | PHINode *FirstPhi = dyn_cast<PHINode>(FirstValue); | ||||
3675 | PHINode *SecondPhi = dyn_cast<PHINode>(SecondValue); | ||||
3676 | |||||
3677 | // One of them is not Phi or | ||||
3678 | // The first one is not Phi node from the set we'd like to match or | ||||
3679 | // Phi nodes from different basic blocks then | ||||
3680 | // we will not be able to match. | ||||
3681 | if (!FirstPhi || !SecondPhi || !PhiNodesToMatch.count(FirstPhi) || | ||||
3682 | FirstPhi->getParent() != SecondPhi->getParent()) | ||||
3683 | return false; | ||||
3684 | |||||
3685 | // If we already matched them then continue. | ||||
3686 | if (Matcher.count({ FirstPhi, SecondPhi })) | ||||
3687 | continue; | ||||
3688 | // So the values are different and does not match. So we need them to | ||||
3689 | // match. (But we register no more than one match per PHI node, so that | ||||
3690 | // we won't later try to replace them twice.) | ||||
3691 | if (MatchedPHIs.insert(FirstPhi).second) | ||||
3692 | Matcher.insert({ FirstPhi, SecondPhi }); | ||||
3693 | // But me must check it. | ||||
3694 | WorkList.push_back({ FirstPhi, SecondPhi }); | ||||
3695 | } | ||||
3696 | } | ||||
3697 | return true; | ||||
3698 | } | ||||
3699 | |||||
3700 | /// For the given set of PHI nodes (in the SimplificationTracker) try | ||||
3701 | /// to find their equivalents. | ||||
3702 | /// Returns false if this matching fails and creation of new Phi is disabled. | ||||
3703 | bool MatchPhiSet(SimplificationTracker &ST, bool AllowNewPhiNodes, | ||||
3704 | unsigned &PhiNotMatchedCount) { | ||||
3705 | // Matched and PhiNodesToMatch iterate their elements in a deterministic | ||||
3706 | // order, so the replacements (ReplacePhi) are also done in a deterministic | ||||
3707 | // order. | ||||
3708 | SmallSetVector<PHIPair, 8> Matched; | ||||
3709 | SmallPtrSet<PHINode *, 8> WillNotMatch; | ||||
3710 | PhiNodeSet &PhiNodesToMatch = ST.newPhiNodes(); | ||||
3711 | while (PhiNodesToMatch.size()) { | ||||
3712 | PHINode *PHI = *PhiNodesToMatch.begin(); | ||||
3713 | |||||
3714 | // Add us, if no Phi nodes in the basic block we do not match. | ||||
3715 | WillNotMatch.clear(); | ||||
3716 | WillNotMatch.insert(PHI); | ||||
3717 | |||||
3718 | // Traverse all Phis until we found equivalent or fail to do that. | ||||
3719 | bool IsMatched = false; | ||||
3720 | for (auto &P : PHI->getParent()->phis()) { | ||||
3721 | if (&P == PHI) | ||||
3722 | continue; | ||||
3723 | if ((IsMatched = MatchPhiNode(PHI, &P, Matched, PhiNodesToMatch))) | ||||
3724 | break; | ||||
3725 | // If it does not match, collect all Phi nodes from matcher. | ||||
3726 | // if we end up with no match, them all these Phi nodes will not match | ||||
3727 | // later. | ||||
3728 | for (auto M : Matched) | ||||
3729 | WillNotMatch.insert(M.first); | ||||
3730 | Matched.clear(); | ||||
3731 | } | ||||
3732 | if (IsMatched) { | ||||
3733 | // Replace all matched values and erase them. | ||||
3734 | for (auto MV : Matched) | ||||
3735 | ST.ReplacePhi(MV.first, MV.second); | ||||
3736 | Matched.clear(); | ||||
3737 | continue; | ||||
3738 | } | ||||
3739 | // If we are not allowed to create new nodes then bail out. | ||||
3740 | if (!AllowNewPhiNodes) | ||||
3741 | return false; | ||||
3742 | // Just remove all seen values in matcher. They will not match anything. | ||||
3743 | PhiNotMatchedCount += WillNotMatch.size(); | ||||
3744 | for (auto *P : WillNotMatch) | ||||
3745 | PhiNodesToMatch.erase(P); | ||||
3746 | } | ||||
3747 | return true; | ||||
3748 | } | ||||
3749 | /// Fill the placeholders with values from predecessors and simplify them. | ||||
3750 | void FillPlaceholders(FoldAddrToValueMapping &Map, | ||||
3751 | SmallVectorImpl<Value *> &TraverseOrder, | ||||
3752 | SimplificationTracker &ST) { | ||||
3753 | while (!TraverseOrder.empty()) { | ||||
3754 | Value *Current = TraverseOrder.pop_back_val(); | ||||
3755 | assert(Map.find(Current) != Map.end() && "No node to fill!!!")((void)0); | ||||
3756 | Value *V = Map[Current]; | ||||
3757 | |||||
3758 | if (SelectInst *Select = dyn_cast<SelectInst>(V)) { | ||||
3759 | // CurrentValue also must be Select. | ||||
3760 | auto *CurrentSelect = cast<SelectInst>(Current); | ||||
3761 | auto *TrueValue = CurrentSelect->getTrueValue(); | ||||
3762 | assert(Map.find(TrueValue) != Map.end() && "No True Value!")((void)0); | ||||
3763 | Select->setTrueValue(ST.Get(Map[TrueValue])); | ||||
3764 | auto *FalseValue = CurrentSelect->getFalseValue(); | ||||
3765 | assert(Map.find(FalseValue) != Map.end() && "No False Value!")((void)0); | ||||
3766 | Select->setFalseValue(ST.Get(Map[FalseValue])); | ||||
3767 | } else { | ||||
3768 | // Must be a Phi node then. | ||||
3769 | auto *PHI = cast<PHINode>(V); | ||||
3770 | // Fill the Phi node with values from predecessors. | ||||
3771 | for (auto *B : predecessors(PHI->getParent())) { | ||||
3772 | Value *PV = cast<PHINode>(Current)->getIncomingValueForBlock(B); | ||||
3773 | assert(Map.find(PV) != Map.end() && "No predecessor Value!")((void)0); | ||||
3774 | PHI->addIncoming(ST.Get(Map[PV]), B); | ||||
3775 | } | ||||
3776 | } | ||||
3777 | Map[Current] = ST.Simplify(V); | ||||
3778 | } | ||||
3779 | } | ||||
3780 | |||||
3781 | /// Starting from original value recursively iterates over def-use chain up to | ||||
3782 | /// known ending values represented in a map. For each traversed phi/select | ||||
3783 | /// inserts a placeholder Phi or Select. | ||||
3784 | /// Reports all new created Phi/Select nodes by adding them to set. | ||||
3785 | /// Also reports and order in what values have been traversed. | ||||
3786 | void InsertPlaceholders(FoldAddrToValueMapping &Map, | ||||
3787 | SmallVectorImpl<Value *> &TraverseOrder, | ||||
3788 | SimplificationTracker &ST) { | ||||
3789 | SmallVector<Value *, 32> Worklist; | ||||
3790 | assert((isa<PHINode>(Original) || isa<SelectInst>(Original)) &&((void)0) | ||||
3791 | "Address must be a Phi or Select node")((void)0); | ||||
3792 | auto *Dummy = UndefValue::get(CommonType); | ||||
3793 | Worklist.push_back(Original); | ||||
3794 | while (!Worklist.empty()) { | ||||
3795 | Value *Current = Worklist.pop_back_val(); | ||||
3796 | // if it is already visited or it is an ending value then skip it. | ||||
3797 | if (Map.find(Current) != Map.end()) | ||||
3798 | continue; | ||||
3799 | TraverseOrder.push_back(Current); | ||||
3800 | |||||
3801 | // CurrentValue must be a Phi node or select. All others must be covered | ||||
3802 | // by anchors. | ||||
3803 | if (SelectInst *CurrentSelect = dyn_cast<SelectInst>(Current)) { | ||||
3804 | // Is it OK to get metadata from OrigSelect?! | ||||
3805 | // Create a Select placeholder with dummy value. | ||||
3806 | SelectInst *Select = SelectInst::Create( | ||||
3807 | CurrentSelect->getCondition(), Dummy, Dummy, | ||||
3808 | CurrentSelect->getName(), CurrentSelect, CurrentSelect); | ||||
3809 | Map[Current] = Select; | ||||
3810 | ST.insertNewSelect(Select); | ||||
3811 | // We are interested in True and False values. | ||||
3812 | Worklist.push_back(CurrentSelect->getTrueValue()); | ||||
3813 | Worklist.push_back(CurrentSelect->getFalseValue()); | ||||
3814 | } else { | ||||
3815 | // It must be a Phi node then. | ||||
3816 | PHINode *CurrentPhi = cast<PHINode>(Current); | ||||
3817 | unsigned PredCount = CurrentPhi->getNumIncomingValues(); | ||||
3818 | PHINode *PHI = | ||||
3819 | PHINode::Create(CommonType, PredCount, "sunk_phi", CurrentPhi); | ||||
3820 | Map[Current] = PHI; | ||||
3821 | ST.insertNewPhi(PHI); | ||||
3822 | append_range(Worklist, CurrentPhi->incoming_values()); | ||||
3823 | } | ||||
3824 | } | ||||
3825 | } | ||||
3826 | |||||
3827 | bool addrModeCombiningAllowed() { | ||||
3828 | if (DisableComplexAddrModes) | ||||
3829 | return false; | ||||
3830 | switch (DifferentField) { | ||||
3831 | default: | ||||
3832 | return false; | ||||
3833 | case ExtAddrMode::BaseRegField: | ||||
3834 | return AddrSinkCombineBaseReg; | ||||
3835 | case ExtAddrMode::BaseGVField: | ||||
3836 | return AddrSinkCombineBaseGV; | ||||
3837 | case ExtAddrMode::BaseOffsField: | ||||
3838 | return AddrSinkCombineBaseOffs; | ||||
3839 | case ExtAddrMode::ScaledRegField: | ||||
3840 | return AddrSinkCombineScaledReg; | ||||
3841 | } | ||||
3842 | } | ||||
3843 | }; | ||||
3844 | } // end anonymous namespace | ||||
3845 | |||||
3846 | /// Try adding ScaleReg*Scale to the current addressing mode. | ||||
3847 | /// Return true and update AddrMode if this addr mode is legal for the target, | ||||
3848 | /// false if not. | ||||
3849 | bool AddressingModeMatcher::matchScaledValue(Value *ScaleReg, int64_t Scale, | ||||
3850 | unsigned Depth) { | ||||
3851 | // If Scale is 1, then this is the same as adding ScaleReg to the addressing | ||||
3852 | // mode. Just process that directly. | ||||
3853 | if (Scale == 1) | ||||
3854 | return matchAddr(ScaleReg, Depth); | ||||
3855 | |||||
3856 | // If the scale is 0, it takes nothing to add this. | ||||
3857 | if (Scale == 0) | ||||
3858 | return true; | ||||
3859 | |||||
3860 | // If we already have a scale of this value, we can add to it, otherwise, we | ||||
3861 | // need an available scale field. | ||||
3862 | if (AddrMode.Scale != 0 && AddrMode.ScaledReg != ScaleReg) | ||||
3863 | return false; | ||||
3864 | |||||
3865 | ExtAddrMode TestAddrMode = AddrMode; | ||||
3866 | |||||
3867 | // Add scale to turn X*4+X*3 -> X*7. This could also do things like | ||||
3868 | // [A+B + A*7] -> [B+A*8]. | ||||
3869 | TestAddrMode.Scale += Scale; | ||||
3870 | TestAddrMode.ScaledReg = ScaleReg; | ||||
3871 | |||||
3872 | // If the new address isn't legal, bail out. | ||||
3873 | if (!TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace)) | ||||
3874 | return false; | ||||
3875 | |||||
3876 | // It was legal, so commit it. | ||||
3877 | AddrMode = TestAddrMode; | ||||
3878 | |||||
3879 | // Okay, we decided that we can add ScaleReg+Scale to AddrMode. Check now | ||||
3880 | // to see if ScaleReg is actually X+C. If so, we can turn this into adding | ||||
3881 | // X*Scale + C*Scale to addr mode. If we found available IV increment, do not | ||||
3882 | // go any further: we can reuse it and cannot eliminate it. | ||||
3883 | ConstantInt *CI = nullptr; Value *AddLHS = nullptr; | ||||
3884 | if (isa<Instruction>(ScaleReg) && // not a constant expr. | ||||
3885 | match(ScaleReg, m_Add(m_Value(AddLHS), m_ConstantInt(CI))) && | ||||
3886 | !isIVIncrement(ScaleReg, &LI) && CI->getValue().isSignedIntN(64)) { | ||||
3887 | TestAddrMode.InBounds = false; | ||||
3888 | TestAddrMode.ScaledReg = AddLHS; | ||||
3889 | TestAddrMode.BaseOffs += CI->getSExtValue() * TestAddrMode.Scale; | ||||
3890 | |||||
3891 | // If this addressing mode is legal, commit it and remember that we folded | ||||
3892 | // this instruction. | ||||
3893 | if (TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace)) { | ||||
3894 | AddrModeInsts.push_back(cast<Instruction>(ScaleReg)); | ||||
3895 | AddrMode = TestAddrMode; | ||||
3896 | return true; | ||||
3897 | } | ||||
3898 | // Restore status quo. | ||||
3899 | TestAddrMode = AddrMode; | ||||
3900 | } | ||||
3901 | |||||
3902 | // If this is an add recurrence with a constant step, return the increment | ||||
3903 | // instruction and the canonicalized step. | ||||
3904 | auto GetConstantStep = [this](const Value * V) | ||||
3905 | ->Optional<std::pair<Instruction *, APInt> > { | ||||
3906 | auto *PN = dyn_cast<PHINode>(V); | ||||
3907 | if (!PN) | ||||
3908 | return None; | ||||
3909 | auto IVInc = getIVIncrement(PN, &LI); | ||||
3910 | if (!IVInc) | ||||
3911 | return None; | ||||
3912 | // TODO: The result of the intrinsics above is two-compliment. However when | ||||
3913 | // IV inc is expressed as add or sub, iv.next is potentially a poison value. | ||||
3914 | // If it has nuw or nsw flags, we need to make sure that these flags are | ||||
3915 | // inferrable at the point of memory instruction. Otherwise we are replacing | ||||
3916 | // well-defined two-compliment computation with poison. Currently, to avoid | ||||
3917 | // potentially complex analysis needed to prove this, we reject such cases. | ||||
3918 | if (auto *OIVInc = dyn_cast<OverflowingBinaryOperator>(IVInc->first)) | ||||
3919 | if (OIVInc->hasNoSignedWrap() || OIVInc->hasNoUnsignedWrap()) | ||||
3920 | return None; | ||||
3921 | if (auto *ConstantStep = dyn_cast<ConstantInt>(IVInc->second)) | ||||
3922 | return std::make_pair(IVInc->first, ConstantStep->getValue()); | ||||
3923 | return None; | ||||
3924 | }; | ||||
3925 | |||||
3926 | // Try to account for the following special case: | ||||
3927 | // 1. ScaleReg is an inductive variable; | ||||
3928 | // 2. We use it with non-zero offset; | ||||
3929 | // 3. IV's increment is available at the point of memory instruction. | ||||
3930 | // | ||||
3931 | // In this case, we may reuse the IV increment instead of the IV Phi to | ||||
3932 | // achieve the following advantages: | ||||
3933 | // 1. If IV step matches the offset, we will have no need in the offset; | ||||
3934 | // 2. Even if they don't match, we will reduce the overlap of living IV | ||||
3935 | // and IV increment, that will potentially lead to better register | ||||
3936 | // assignment. | ||||
3937 | if (AddrMode.BaseOffs) { | ||||
3938 | if (auto IVStep = GetConstantStep(ScaleReg)) { | ||||
3939 | Instruction *IVInc = IVStep->first; | ||||
3940 | // The following assert is important to ensure a lack of infinite loops. | ||||
3941 | // This transforms is (intentionally) the inverse of the one just above. | ||||
3942 | // If they don't agree on the definition of an increment, we'd alternate | ||||
3943 | // back and forth indefinitely. | ||||
3944 | assert(isIVIncrement(IVInc, &LI) && "implied by GetConstantStep")((void)0); | ||||
3945 | APInt Step = IVStep->second; | ||||
3946 | APInt Offset = Step * AddrMode.Scale; | ||||
3947 | if (Offset.isSignedIntN(64)) { | ||||
3948 | TestAddrMode.InBounds = false; | ||||
3949 | TestAddrMode.ScaledReg = IVInc; | ||||
3950 | TestAddrMode.BaseOffs -= Offset.getLimitedValue(); | ||||
3951 | // If this addressing mode is legal, commit it.. | ||||
3952 | // (Note that we defer the (expensive) domtree base legality check | ||||
3953 | // to the very last possible point.) | ||||
3954 | if (TLI.isLegalAddressingMode(DL, TestAddrMode, AccessTy, AddrSpace) && | ||||
3955 | getDTFn().dominates(IVInc, MemoryInst)) { | ||||
3956 | AddrModeInsts.push_back(cast<Instruction>(IVInc)); | ||||
3957 | AddrMode = TestAddrMode; | ||||
3958 | return true; | ||||
3959 | } | ||||
3960 | // Restore status quo. | ||||
3961 | TestAddrMode = AddrMode; | ||||
3962 | } | ||||
3963 | } | ||||
3964 | } | ||||
3965 | |||||
3966 | // Otherwise, just return what we have. | ||||
3967 | return true; | ||||
3968 | } | ||||
3969 | |||||
3970 | /// This is a little filter, which returns true if an addressing computation | ||||
3971 | /// involving I might be folded into a load/store accessing it. | ||||
3972 | /// This doesn't need to be perfect, but needs to accept at least | ||||
3973 | /// the set of instructions that MatchOperationAddr can. | ||||
3974 | static bool MightBeFoldableInst(Instruction *I) { | ||||
3975 | switch (I->getOpcode()) { | ||||
3976 | case Instruction::BitCast: | ||||
3977 | case Instruction::AddrSpaceCast: | ||||
3978 | // Don't touch identity bitcasts. | ||||
3979 | if (I->getType() == I->getOperand(0)->getType()) | ||||
3980 | return false; | ||||
3981 | return I->getType()->isIntOrPtrTy(); | ||||
3982 | case Instruction::PtrToInt: | ||||
3983 | // PtrToInt is always a noop, as we know that the int type is pointer sized. | ||||
3984 | return true; | ||||
3985 | case Instruction::IntToPtr: | ||||
3986 | // We know the input is intptr_t, so this is foldable. | ||||
3987 | return true; | ||||
3988 | case Instruction::Add: | ||||
3989 | return true; | ||||
3990 | case Instruction::Mul: | ||||
3991 | case Instruction::Shl: | ||||
3992 | // Can only handle X*C and X << C. | ||||
3993 | return isa<ConstantInt>(I->getOperand(1)); | ||||
3994 | case Instruction::GetElementPtr: | ||||
3995 | return true; | ||||
3996 | default: | ||||
3997 | return false; | ||||
3998 | } | ||||
3999 | } | ||||
4000 | |||||
4001 | /// Check whether or not \p Val is a legal instruction for \p TLI. | ||||
4002 | /// \note \p Val is assumed to be the product of some type promotion. | ||||
4003 | /// Therefore if \p Val has an undefined state in \p TLI, this is assumed | ||||
4004 | /// to be legal, as the non-promoted value would have had the same state. | ||||
4005 | static bool isPromotedInstructionLegal(const TargetLowering &TLI, | ||||
4006 | const DataLayout &DL, Value *Val) { | ||||
4007 | Instruction *PromotedInst = dyn_cast<Instruction>(Val); | ||||
4008 | if (!PromotedInst) | ||||
4009 | return false; | ||||
4010 | int ISDOpcode = TLI.InstructionOpcodeToISD(PromotedInst->getOpcode()); | ||||
4011 | // If the ISDOpcode is undefined, it was undefined before the promotion. | ||||
4012 | if (!ISDOpcode) | ||||
4013 | return true; | ||||
4014 | // Otherwise, check if the promoted instruction is legal or not. | ||||
4015 | return TLI.isOperationLegalOrCustom( | ||||
4016 | ISDOpcode, TLI.getValueType(DL, PromotedInst->getType())); | ||||
4017 | } | ||||
4018 | |||||
4019 | namespace { | ||||
4020 | |||||
4021 | /// Hepler class to perform type promotion. | ||||
4022 | class TypePromotionHelper { | ||||
4023 | /// Utility function to add a promoted instruction \p ExtOpnd to | ||||
4024 | /// \p PromotedInsts and record the type of extension we have seen. | ||||
4025 | static void addPromotedInst(InstrToOrigTy &PromotedInsts, | ||||
4026 | Instruction *ExtOpnd, | ||||
4027 | bool IsSExt) { | ||||
4028 | ExtType ExtTy = IsSExt ? SignExtension : ZeroExtension; | ||||
4029 | InstrToOrigTy::iterator It = PromotedInsts.find(ExtOpnd); | ||||
4030 | if (It != PromotedInsts.end()) { | ||||
4031 | // If the new extension is same as original, the information in | ||||
4032 | // PromotedInsts[ExtOpnd] is still correct. | ||||
4033 | if (It->second.getInt() == ExtTy) | ||||
4034 | return; | ||||
4035 | |||||
4036 | // Now the new extension is different from old extension, we make | ||||
4037 | // the type information invalid by setting extension type to | ||||
4038 | // BothExtension. | ||||
4039 | ExtTy = BothExtension; | ||||
4040 | } | ||||
4041 | PromotedInsts[ExtOpnd] = TypeIsSExt(ExtOpnd->getType(), ExtTy); | ||||
4042 | } | ||||
4043 | |||||
4044 | /// Utility function to query the original type of instruction \p Opnd | ||||
4045 | /// with a matched extension type. If the extension doesn't match, we | ||||
4046 | /// cannot use the information we had on the original type. | ||||
4047 | /// BothExtension doesn't match any extension type. | ||||
4048 | static const Type *getOrigType(const InstrToOrigTy &PromotedInsts, | ||||
4049 | Instruction *Opnd, | ||||
4050 | bool IsSExt) { | ||||
4051 | ExtType ExtTy = IsSExt ? SignExtension : ZeroExtension; | ||||
4052 | InstrToOrigTy::const_iterator It = PromotedInsts.find(Opnd); | ||||
4053 | if (It != PromotedInsts.end() && It->second.getInt() == ExtTy) | ||||
4054 | return It->second.getPointer(); | ||||
4055 | return nullptr; | ||||
4056 | } | ||||
4057 | |||||
4058 | /// Utility function to check whether or not a sign or zero extension | ||||
4059 | /// of \p Inst with \p ConsideredExtType can be moved through \p Inst by | ||||
4060 | /// either using the operands of \p Inst or promoting \p Inst. | ||||
4061 | /// The type of the extension is defined by \p IsSExt. | ||||
4062 | /// In other words, check if: | ||||
4063 | /// ext (Ty Inst opnd1 opnd2 ... opndN) to ConsideredExtType. | ||||
4064 | /// #1 Promotion applies: | ||||
4065 | /// ConsideredExtType Inst (ext opnd1 to ConsideredExtType, ...). | ||||
4066 | /// #2 Operand reuses: | ||||
4067 | /// ext opnd1 to ConsideredExtType. | ||||
4068 | /// \p PromotedInsts maps the instructions to their type before promotion. | ||||
4069 | static bool canGetThrough(const Instruction *Inst, Type *ConsideredExtType, | ||||
4070 | const InstrToOrigTy &PromotedInsts, bool IsSExt); | ||||
4071 | |||||
4072 | /// Utility function to determine if \p OpIdx should be promoted when | ||||
4073 | /// promoting \p Inst. | ||||
4074 | static bool shouldExtOperand(const Instruction *Inst, int OpIdx) { | ||||
4075 | return !(isa<SelectInst>(Inst) && OpIdx == 0); | ||||
4076 | } | ||||
4077 | |||||
4078 | /// Utility function to promote the operand of \p Ext when this | ||||
4079 | /// operand is a promotable trunc or sext or zext. | ||||
4080 | /// \p PromotedInsts maps the instructions to their type before promotion. | ||||
4081 | /// \p CreatedInstsCost[out] contains the cost of all instructions | ||||
4082 | /// created to promote the operand of Ext. | ||||
4083 | /// Newly added extensions are inserted in \p Exts. | ||||
4084 | /// Newly added truncates are inserted in \p Truncs. | ||||
4085 | /// Should never be called directly. | ||||
4086 | /// \return The promoted value which is used instead of Ext. | ||||
4087 | static Value *promoteOperandForTruncAndAnyExt( | ||||
4088 | Instruction *Ext, TypePromotionTransaction &TPT, | ||||
4089 | InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost, | ||||
4090 | SmallVectorImpl<Instruction *> *Exts, | ||||
4091 | SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI); | ||||
4092 | |||||
4093 | /// Utility function to promote the operand of \p Ext when this | ||||
4094 | /// operand is promotable and is not a supported trunc or sext. | ||||
4095 | /// \p PromotedInsts maps the instructions to their type before promotion. | ||||
4096 | /// \p CreatedInstsCost[out] contains the cost of all the instructions | ||||
4097 | /// created to promote the operand of Ext. | ||||
4098 | /// Newly added extensions are inserted in \p Exts. | ||||
4099 | /// Newly added truncates are inserted in \p Truncs. | ||||
4100 | /// Should never be called directly. | ||||
4101 | /// \return The promoted value which is used instead of Ext. | ||||
4102 | static Value *promoteOperandForOther(Instruction *Ext, | ||||
4103 | TypePromotionTransaction &TPT, | ||||
4104 | InstrToOrigTy &PromotedInsts, | ||||
4105 | unsigned &CreatedInstsCost, | ||||
4106 | SmallVectorImpl<Instruction *> *Exts, | ||||
4107 | SmallVectorImpl<Instruction *> *Truncs, | ||||
4108 | const TargetLowering &TLI, bool IsSExt); | ||||
4109 | |||||
4110 | /// \see promoteOperandForOther. | ||||
4111 | static Value *signExtendOperandForOther( | ||||
4112 | Instruction *Ext, TypePromotionTransaction &TPT, | ||||
4113 | InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost, | ||||
4114 | SmallVectorImpl<Instruction *> *Exts, | ||||
4115 | SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) { | ||||
4116 | return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost, | ||||
4117 | Exts, Truncs, TLI, true); | ||||
4118 | } | ||||
4119 | |||||
4120 | /// \see promoteOperandForOther. | ||||
4121 | static Value *zeroExtendOperandForOther( | ||||
4122 | Instruction *Ext, TypePromotionTransaction &TPT, | ||||
4123 | InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost, | ||||
4124 | SmallVectorImpl<Instruction *> *Exts, | ||||
4125 | SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) { | ||||
4126 | return promoteOperandForOther(Ext, TPT, PromotedInsts, CreatedInstsCost, | ||||
4127 | Exts, Truncs, TLI, false); | ||||
4128 | } | ||||
4129 | |||||
4130 | public: | ||||
4131 | /// Type for the utility function that promotes the operand of Ext. | ||||
4132 | using Action = Value *(*)(Instruction *Ext, TypePromotionTransaction &TPT, | ||||
4133 | InstrToOrigTy &PromotedInsts, | ||||
4134 | unsigned &CreatedInstsCost, | ||||
4135 | SmallVectorImpl<Instruction *> *Exts, | ||||
4136 | SmallVectorImpl<Instruction *> *Truncs, | ||||
4137 | const TargetLowering &TLI); | ||||
4138 | |||||
4139 | /// Given a sign/zero extend instruction \p Ext, return the appropriate | ||||
4140 | /// action to promote the operand of \p Ext instead of using Ext. | ||||
4141 | /// \return NULL if no promotable action is possible with the current | ||||
4142 | /// sign extension. | ||||
4143 | /// \p InsertedInsts keeps track of all the instructions inserted by the | ||||
4144 | /// other CodeGenPrepare optimizations. This information is important | ||||
4145 | /// because we do not want to promote these instructions as CodeGenPrepare | ||||
4146 | /// will reinsert them later. Thus creating an infinite loop: create/remove. | ||||
4147 | /// \p PromotedInsts maps the instructions to their type before promotion. | ||||
4148 | static Action getAction(Instruction *Ext, const SetOfInstrs &InsertedInsts, | ||||
4149 | const TargetLowering &TLI, | ||||
4150 | const InstrToOrigTy &PromotedInsts); | ||||
4151 | }; | ||||
4152 | |||||
4153 | } // end anonymous namespace | ||||
4154 | |||||
4155 | bool TypePromotionHelper::canGetThrough(const Instruction *Inst, | ||||
4156 | Type *ConsideredExtType, | ||||
4157 | const InstrToOrigTy &PromotedInsts, | ||||
4158 | bool IsSExt) { | ||||
4159 | // The promotion helper does not know how to deal with vector types yet. | ||||
4160 | // To be able to fix that, we would need to fix the places where we | ||||
4161 | // statically extend, e.g., constants and such. | ||||
4162 | if (Inst->getType()->isVectorTy()) | ||||
4163 | return false; | ||||
4164 | |||||
4165 | // We can always get through zext. | ||||
4166 | if (isa<ZExtInst>(Inst)) | ||||
4167 | return true; | ||||
4168 | |||||
4169 | // sext(sext) is ok too. | ||||
4170 | if (IsSExt && isa<SExtInst>(Inst)) | ||||
4171 | return true; | ||||
4172 | |||||
4173 | // We can get through binary operator, if it is legal. In other words, the | ||||
4174 | // binary operator must have a nuw or nsw flag. | ||||
4175 | const BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst); | ||||
4176 | if (isa_and_nonnull<OverflowingBinaryOperator>(BinOp) && | ||||
4177 | ((!IsSExt && BinOp->hasNoUnsignedWrap()) || | ||||
4178 | (IsSExt && BinOp->hasNoSignedWrap()))) | ||||
4179 | return true; | ||||
4180 | |||||
4181 | // ext(and(opnd, cst)) --> and(ext(opnd), ext(cst)) | ||||
4182 | if ((Inst->getOpcode() == Instruction::And || | ||||
4183 | Inst->getOpcode() == Instruction::Or)) | ||||
4184 | return true; | ||||
4185 | |||||
4186 | // ext(xor(opnd, cst)) --> xor(ext(opnd), ext(cst)) | ||||
4187 | if (Inst->getOpcode() == Instruction::Xor) { | ||||
4188 | const ConstantInt *Cst = dyn_cast<ConstantInt>(Inst->getOperand(1)); | ||||
4189 | // Make sure it is not a NOT. | ||||
4190 | if (Cst && !Cst->getValue().isAllOnesValue()) | ||||
4191 | return true; | ||||
4192 | } | ||||
4193 | |||||
4194 | // zext(shrl(opnd, cst)) --> shrl(zext(opnd), zext(cst)) | ||||
4195 | // It may change a poisoned value into a regular value, like | ||||
4196 | // zext i32 (shrl i8 %val, 12) --> shrl i32 (zext i8 %val), 12 | ||||
4197 | // poisoned value regular value | ||||
4198 | // It should be OK since undef covers valid value. | ||||
4199 | if (Inst->getOpcode() == Instruction::LShr && !IsSExt) | ||||
4200 | return true; | ||||
4201 | |||||
4202 | // and(ext(shl(opnd, cst)), cst) --> and(shl(ext(opnd), ext(cst)), cst) | ||||
4203 | // It may change a poisoned value into a regular value, like | ||||
4204 | // zext i32 (shl i8 %val, 12) --> shl i32 (zext i8 %val), 12 | ||||
4205 | // poisoned value regular value | ||||
4206 | // It should be OK since undef covers valid value. | ||||
4207 | if (Inst->getOpcode() == Instruction::Shl && Inst->hasOneUse()) { | ||||
4208 | const auto *ExtInst = cast<const Instruction>(*Inst->user_begin()); | ||||
4209 | if (ExtInst->hasOneUse()) { | ||||
4210 | const auto *AndInst = dyn_cast<const Instruction>(*ExtInst->user_begin()); | ||||
4211 | if (AndInst && AndInst->getOpcode() == Instruction::And) { | ||||
4212 | const auto *Cst = dyn_cast<ConstantInt>(AndInst->getOperand(1)); | ||||
4213 | if (Cst && | ||||
4214 | Cst->getValue().isIntN(Inst->getType()->getIntegerBitWidth())) | ||||
4215 | return true; | ||||
4216 | } | ||||
4217 | } | ||||
4218 | } | ||||
4219 | |||||
4220 | // Check if we can do the following simplification. | ||||
4221 | // ext(trunc(opnd)) --> ext(opnd) | ||||
4222 | if (!isa<TruncInst>(Inst)) | ||||
4223 | return false; | ||||
4224 | |||||
4225 | Value *OpndVal = Inst->getOperand(0); | ||||
4226 | // Check if we can use this operand in the extension. | ||||
4227 | // If the type is larger than the result type of the extension, we cannot. | ||||
4228 | if (!OpndVal->getType()->isIntegerTy() || | ||||
4229 | OpndVal->getType()->getIntegerBitWidth() > | ||||
4230 | ConsideredExtType->getIntegerBitWidth()) | ||||
4231 | return false; | ||||
4232 | |||||
4233 | // If the operand of the truncate is not an instruction, we will not have | ||||
4234 | // any information on the dropped bits. | ||||
4235 | // (Actually we could for constant but it is not worth the extra logic). | ||||
4236 | Instruction *Opnd = dyn_cast<Instruction>(OpndVal); | ||||
4237 | if (!Opnd) | ||||
4238 | return false; | ||||
4239 | |||||
4240 | // Check if the source of the type is narrow enough. | ||||
4241 | // I.e., check that trunc just drops extended bits of the same kind of | ||||
4242 | // the extension. | ||||
4243 | // #1 get the type of the operand and check the kind of the extended bits. | ||||
4244 | const Type *OpndType = getOrigType(PromotedInsts, Opnd, IsSExt); | ||||
4245 | if (OpndType) | ||||
4246 | ; | ||||
4247 | else if ((IsSExt && isa<SExtInst>(Opnd)) || (!IsSExt && isa<ZExtInst>(Opnd))) | ||||
4248 | OpndType = Opnd->getOperand(0)->getType(); | ||||
4249 | else | ||||
4250 | return false; | ||||
4251 | |||||
4252 | // #2 check that the truncate just drops extended bits. | ||||
4253 | return Inst->getType()->getIntegerBitWidth() >= | ||||
4254 | OpndType->getIntegerBitWidth(); | ||||
4255 | } | ||||
4256 | |||||
4257 | TypePromotionHelper::Action TypePromotionHelper::getAction( | ||||
4258 | Instruction *Ext, const SetOfInstrs &InsertedInsts, | ||||
4259 | const TargetLowering &TLI, const InstrToOrigTy &PromotedInsts) { | ||||
4260 | assert((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&((void)0) | ||||
4261 | "Unexpected instruction type")((void)0); | ||||
4262 | Instruction *ExtOpnd = dyn_cast<Instruction>(Ext->getOperand(0)); | ||||
4263 | Type *ExtTy = Ext->getType(); | ||||
4264 | bool IsSExt = isa<SExtInst>(Ext); | ||||
4265 | // If the operand of the extension is not an instruction, we cannot | ||||
4266 | // get through. | ||||
4267 | // If it, check we can get through. | ||||
4268 | if (!ExtOpnd || !canGetThrough(ExtOpnd, ExtTy, PromotedInsts, IsSExt)) | ||||
4269 | return nullptr; | ||||
4270 | |||||
4271 | // Do not promote if the operand has been added by codegenprepare. | ||||
4272 | // Otherwise, it means we are undoing an optimization that is likely to be | ||||
4273 | // redone, thus causing potential infinite loop. | ||||
4274 | if (isa<TruncInst>(ExtOpnd) && InsertedInsts.count(ExtOpnd)) | ||||
4275 | return nullptr; | ||||
4276 | |||||
4277 | // SExt or Trunc instructions. | ||||
4278 | // Return the related handler. | ||||
4279 | if (isa<SExtInst>(ExtOpnd) || isa<TruncInst>(ExtOpnd) || | ||||
4280 | isa<ZExtInst>(ExtOpnd)) | ||||
4281 | return promoteOperandForTruncAndAnyExt; | ||||
4282 | |||||
4283 | // Regular instruction. | ||||
4284 | // Abort early if we will have to insert non-free instructions. | ||||
4285 | if (!ExtOpnd->hasOneUse() && !TLI.isTruncateFree(ExtTy, ExtOpnd->getType())) | ||||
4286 | return nullptr; | ||||
4287 | return IsSExt ? signExtendOperandForOther : zeroExtendOperandForOther; | ||||
4288 | } | ||||
4289 | |||||
4290 | Value *TypePromotionHelper::promoteOperandForTruncAndAnyExt( | ||||
4291 | Instruction *SExt, TypePromotionTransaction &TPT, | ||||
4292 | InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost, | ||||
4293 | SmallVectorImpl<Instruction *> *Exts, | ||||
4294 | SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI) { | ||||
4295 | // By construction, the operand of SExt is an instruction. Otherwise we cannot | ||||
4296 | // get through it and this method should not be called. | ||||
4297 | Instruction *SExtOpnd = cast<Instruction>(SExt->getOperand(0)); | ||||
4298 | Value *ExtVal = SExt; | ||||
4299 | bool HasMergedNonFreeExt = false; | ||||
4300 | if (isa<ZExtInst>(SExtOpnd)) { | ||||
4301 | // Replace s|zext(zext(opnd)) | ||||
4302 | // => zext(opnd). | ||||
4303 | HasMergedNonFreeExt = !TLI.isExtFree(SExtOpnd); | ||||
4304 | Value *ZExt = | ||||
4305 | TPT.createZExt(SExt, SExtOpnd->getOperand(0), SExt->getType()); | ||||
4306 | TPT.replaceAllUsesWith(SExt, ZExt); | ||||
4307 | TPT.eraseInstruction(SExt); | ||||
4308 | ExtVal = ZExt; | ||||
4309 | } else { | ||||
4310 | // Replace z|sext(trunc(opnd)) or sext(sext(opnd)) | ||||
4311 | // => z|sext(opnd). | ||||
4312 | TPT.setOperand(SExt, 0, SExtOpnd->getOperand(0)); | ||||
4313 | } | ||||
4314 | CreatedInstsCost = 0; | ||||
4315 | |||||
4316 | // Remove dead code. | ||||
4317 | if (SExtOpnd->use_empty()) | ||||
4318 | TPT.eraseInstruction(SExtOpnd); | ||||
4319 | |||||
4320 | // Check if the extension is still needed. | ||||
4321 | Instruction *ExtInst = dyn_cast<Instruction>(ExtVal); | ||||
4322 | if (!ExtInst || ExtInst->getType() != ExtInst->getOperand(0)->getType()) { | ||||
4323 | if (ExtInst) { | ||||
4324 | if (Exts) | ||||
4325 | Exts->push_back(ExtInst); | ||||
4326 | CreatedInstsCost = !TLI.isExtFree(ExtInst) && !HasMergedNonFreeExt; | ||||
4327 | } | ||||
4328 | return ExtVal; | ||||
4329 | } | ||||
4330 | |||||
4331 | // At this point we have: ext ty opnd to ty. | ||||
4332 | // Reassign the uses of ExtInst to the opnd and remove ExtInst. | ||||
4333 | Value *NextVal = ExtInst->getOperand(0); | ||||
4334 | TPT.eraseInstruction(ExtInst, NextVal); | ||||
4335 | return NextVal; | ||||
4336 | } | ||||
4337 | |||||
4338 | Value *TypePromotionHelper::promoteOperandForOther( | ||||
4339 | Instruction *Ext, TypePromotionTransaction &TPT, | ||||
4340 | InstrToOrigTy &PromotedInsts, unsigned &CreatedInstsCost, | ||||
4341 | SmallVectorImpl<Instruction *> *Exts, | ||||
4342 | SmallVectorImpl<Instruction *> *Truncs, const TargetLowering &TLI, | ||||
4343 | bool IsSExt) { | ||||
4344 | // By construction, the operand of Ext is an instruction. Otherwise we cannot | ||||
4345 | // get through it and this method should not be called. | ||||
4346 | Instruction *ExtOpnd = cast<Instruction>(Ext->getOperand(0)); | ||||
4347 | CreatedInstsCost = 0; | ||||
4348 | if (!ExtOpnd->hasOneUse()) { | ||||
4349 | // ExtOpnd will be promoted. | ||||
4350 | // All its uses, but Ext, will need to use a truncated value of the | ||||
4351 | // promoted version. | ||||
4352 | // Create the truncate now. | ||||
4353 | Value *Trunc = TPT.createTrunc(Ext, ExtOpnd->getType()); | ||||
4354 | if (Instruction *ITrunc = dyn_cast<Instruction>(Trunc)) { | ||||
4355 | // Insert it just after the definition. | ||||
4356 | ITrunc->moveAfter(ExtOpnd); | ||||
4357 | if (Truncs) | ||||
4358 | Truncs->push_back(ITrunc); | ||||
4359 | } | ||||
4360 | |||||
4361 | TPT.replaceAllUsesWith(ExtOpnd, Trunc); | ||||
4362 | // Restore the operand of Ext (which has been replaced by the previous call | ||||
4363 | // to replaceAllUsesWith) to avoid creating a cycle trunc <-> sext. | ||||
4364 | TPT.setOperand(Ext, 0, ExtOpnd); | ||||
4365 | } | ||||
4366 | |||||
4367 | // Get through the Instruction: | ||||
4368 | // 1. Update its type. | ||||
4369 | // 2. Replace the uses of Ext by Inst. | ||||
4370 | // 3. Extend each operand that needs to be extended. | ||||
4371 | |||||
4372 | // Remember the original type of the instruction before promotion. | ||||
4373 | // This is useful to know that the high bits are sign extended bits. | ||||
4374 | addPromotedInst(PromotedInsts, ExtOpnd, IsSExt); | ||||
4375 | // Step #1. | ||||
4376 | TPT.mutateType(ExtOpnd, Ext->getType()); | ||||
4377 | // Step #2. | ||||
4378 | TPT.replaceAllUsesWith(Ext, ExtOpnd); | ||||
4379 | // Step #3. | ||||
4380 | Instruction *ExtForOpnd = Ext; | ||||
4381 | |||||
4382 | LLVM_DEBUG(dbgs() << "Propagate Ext to operands\n")do { } while (false); | ||||
4383 | for (int OpIdx = 0, EndOpIdx = ExtOpnd->getNumOperands(); OpIdx != EndOpIdx; | ||||
4384 | ++OpIdx) { | ||||
4385 | LLVM_DEBUG(dbgs() << "Operand:\n" << *(ExtOpnd->getOperand(OpIdx)) << '\n')do { } while (false); | ||||
4386 | if (ExtOpnd->getOperand(OpIdx)->getType() == Ext->getType() || | ||||
4387 | !shouldExtOperand(ExtOpnd, OpIdx)) { | ||||
4388 | LLVM_DEBUG(dbgs() << "No need to propagate\n")do { } while (false); | ||||
4389 | continue; | ||||
4390 | } | ||||
4391 | // Check if we can statically extend the operand. | ||||
4392 | Value *Opnd = ExtOpnd->getOperand(OpIdx); | ||||
4393 | if (const ConstantInt *Cst = dyn_cast<ConstantInt>(Opnd)) { | ||||
4394 | LLVM_DEBUG(dbgs() << "Statically extend\n")do { } while (false); | ||||
4395 | unsigned BitWidth = Ext->getType()->getIntegerBitWidth(); | ||||
4396 | APInt CstVal = IsSExt ? Cst->getValue().sext(BitWidth) | ||||
4397 | : Cst->getValue().zext(BitWidth); | ||||
4398 | TPT.setOperand(ExtOpnd, OpIdx, ConstantInt::get(Ext->getType(), CstVal)); | ||||
4399 | continue; | ||||
4400 | } | ||||
4401 | // UndefValue are typed, so we have to statically sign extend them. | ||||
4402 | if (isa<UndefValue>(Opnd)) { | ||||
4403 | LLVM_DEBUG(dbgs() << "Statically extend\n")do { } while (false); | ||||
4404 | TPT.setOperand(ExtOpnd, OpIdx, UndefValue::get(Ext->getType())); | ||||
4405 | continue; | ||||
4406 | } | ||||
4407 | |||||
4408 | // Otherwise we have to explicitly sign extend the operand. | ||||
4409 | // Check if Ext was reused to extend an operand. | ||||
4410 | if (!ExtForOpnd) { | ||||
4411 | // If yes, create a new one. | ||||
4412 | LLVM_DEBUG(dbgs() << "More operands to ext\n")do { } while (false); | ||||
4413 | Value *ValForExtOpnd = IsSExt ? TPT.createSExt(Ext, Opnd, Ext->getType()) | ||||
4414 | : TPT.createZExt(Ext, Opnd, Ext->getType()); | ||||
4415 | if (!isa<Instruction>(ValForExtOpnd)) { | ||||
4416 | TPT.setOperand(ExtOpnd, OpIdx, ValForExtOpnd); | ||||
4417 | continue; | ||||
4418 | } | ||||
4419 | ExtForOpnd = cast<Instruction>(ValForExtOpnd); | ||||
4420 | } | ||||
4421 | if (Exts) | ||||
4422 | Exts->push_back(ExtForOpnd); | ||||
4423 | TPT.setOperand(ExtForOpnd, 0, Opnd); | ||||
4424 | |||||
4425 | // Move the sign extension before the insertion point. | ||||
4426 | TPT.moveBefore(ExtForOpnd, ExtOpnd); | ||||
4427 | TPT.setOperand(ExtOpnd, OpIdx, ExtForOpnd); | ||||
4428 | CreatedInstsCost += !TLI.isExtFree(ExtForOpnd); | ||||
4429 | // If more sext are required, new instructions will have to be created. | ||||
4430 | ExtForOpnd = nullptr; | ||||
4431 | } | ||||
4432 | if (ExtForOpnd == Ext) { | ||||
4433 | LLVM_DEBUG(dbgs() << "Extension is useless now\n")do { } while (false); | ||||
4434 | TPT.eraseInstruction(Ext); | ||||
4435 | } | ||||
4436 | return ExtOpnd; | ||||
4437 | } | ||||
4438 | |||||
4439 | /// Check whether or not promoting an instruction to a wider type is profitable. | ||||
4440 | /// \p NewCost gives the cost of extension instructions created by the | ||||
4441 | /// promotion. | ||||
4442 | /// \p OldCost gives the cost of extension instructions before the promotion | ||||
4443 | /// plus the number of instructions that have been | ||||
4444 | /// matched in the addressing mode the promotion. | ||||
4445 | /// \p PromotedOperand is the value that has been promoted. | ||||
4446 | /// \return True if the promotion is profitable, false otherwise. | ||||
4447 | bool AddressingModeMatcher::isPromotionProfitable( | ||||
4448 | unsigned NewCost, unsigned OldCost, Value *PromotedOperand) const { | ||||
4449 | LLVM_DEBUG(dbgs() << "OldCost: " << OldCost << "\tNewCost: " << NewCostdo { } while (false) | ||||
4450 | << '\n')do { } while (false); | ||||
4451 | // The cost of the new extensions is greater than the cost of the | ||||
4452 | // old extension plus what we folded. | ||||
4453 | // This is not profitable. | ||||
4454 | if (NewCost > OldCost) | ||||
4455 | return false; | ||||
4456 | if (NewCost < OldCost) | ||||
4457 | return true; | ||||
4458 | // The promotion is neutral but it may help folding the sign extension in | ||||
4459 | // loads for instance. | ||||
4460 | // Check that we did not create an illegal instruction. | ||||
4461 | return isPromotedInstructionLegal(TLI, DL, PromotedOperand); | ||||
4462 | } | ||||
4463 | |||||
4464 | /// Given an instruction or constant expr, see if we can fold the operation | ||||
4465 | /// into the addressing mode. If so, update the addressing mode and return | ||||
4466 | /// true, otherwise return false without modifying AddrMode. | ||||
4467 | /// If \p MovedAway is not NULL, it contains the information of whether or | ||||
4468 | /// not AddrInst has to be folded into the addressing mode on success. | ||||
4469 | /// If \p MovedAway == true, \p AddrInst will not be part of the addressing | ||||
4470 | /// because it has been moved away. | ||||
4471 | /// Thus AddrInst must not be added in the matched instructions. | ||||
4472 | /// This state can happen when AddrInst is a sext, since it may be moved away. | ||||
4473 | /// Therefore, AddrInst may not be valid when MovedAway is true and it must | ||||
4474 | /// not be referenced anymore. | ||||
4475 | bool AddressingModeMatcher::matchOperationAddr(User *AddrInst, unsigned Opcode, | ||||
4476 | unsigned Depth, | ||||
4477 | bool *MovedAway) { | ||||
4478 | // Avoid exponential behavior on extremely deep expression trees. | ||||
4479 | if (Depth >= 5) return false; | ||||
4480 | |||||
4481 | // By default, all matched instructions stay in place. | ||||
4482 | if (MovedAway) | ||||
4483 | *MovedAway = false; | ||||
4484 | |||||
4485 | switch (Opcode) { | ||||
4486 | case Instruction::PtrToInt: | ||||
4487 | // PtrToInt is always a noop, as we know that the int type is pointer sized. | ||||
4488 | return matchAddr(AddrInst->getOperand(0), Depth); | ||||
4489 | case Instruction::IntToPtr: { | ||||
4490 | auto AS = AddrInst->getType()->getPointerAddressSpace(); | ||||
4491 | auto PtrTy = MVT::getIntegerVT(DL.getPointerSizeInBits(AS)); | ||||
4492 | // This inttoptr is a no-op if the integer type is pointer sized. | ||||
4493 | if (TLI.getValueType(DL, AddrInst->getOperand(0)->getType()) == PtrTy) | ||||
4494 | return matchAddr(AddrInst->getOperand(0), Depth); | ||||
4495 | return false; | ||||
4496 | } | ||||
4497 | case Instruction::BitCast: | ||||
4498 | // BitCast is always a noop, and we can handle it as long as it is | ||||
4499 | // int->int or pointer->pointer (we don't want int<->fp or something). | ||||
4500 | if (AddrInst->getOperand(0)->getType()->isIntOrPtrTy() && | ||||
4501 | // Don't touch identity bitcasts. These were probably put here by LSR, | ||||
4502 | // and we don't want to mess around with them. Assume it knows what it | ||||
4503 | // is doing. | ||||
4504 | AddrInst->getOperand(0)->getType() != AddrInst->getType()) | ||||
4505 | return matchAddr(AddrInst->getOperand(0), Depth); | ||||
4506 | return false; | ||||
4507 | case Instruction::AddrSpaceCast: { | ||||
4508 | unsigned SrcAS | ||||
4509 | = AddrInst->getOperand(0)->getType()->getPointerAddressSpace(); | ||||
4510 | unsigned DestAS = AddrInst->getType()->getPointerAddressSpace(); | ||||
4511 | if (TLI.getTargetMachine().isNoopAddrSpaceCast(SrcAS, DestAS)) | ||||
4512 | return matchAddr(AddrInst->getOperand(0), Depth); | ||||
4513 | return false; | ||||
4514 | } | ||||
4515 | case Instruction::Add: { | ||||
4516 | // Check to see if we can merge in the RHS then the LHS. If so, we win. | ||||
4517 | ExtAddrMode BackupAddrMode = AddrMode; | ||||
4518 | unsigned OldSize = AddrModeInsts.size(); | ||||
4519 | // Start a transaction at this point. | ||||
4520 | // The LHS may match but not the RHS. | ||||
4521 | // Therefore, we need a higher level restoration point to undo partially | ||||
4522 | // matched operation. | ||||
4523 | TypePromotionTransaction::ConstRestorationPt LastKnownGood = | ||||
4524 | TPT.getRestorationPoint(); | ||||
4525 | |||||
4526 | AddrMode.InBounds = false; | ||||
4527 | if (matchAddr(AddrInst->getOperand(1), Depth+1) && | ||||
4528 | matchAddr(AddrInst->getOperand(0), Depth+1)) | ||||
4529 | return true; | ||||
4530 | |||||
4531 | // Restore the old addr mode info. | ||||
4532 | AddrMode = BackupAddrMode; | ||||
4533 | AddrModeInsts.resize(OldSize); | ||||
4534 | TPT.rollback(LastKnownGood); | ||||
4535 | |||||
4536 | // Otherwise this was over-aggressive. Try merging in the LHS then the RHS. | ||||
4537 | if (matchAddr(AddrInst->getOperand(0), Depth+1) && | ||||
4538 | matchAddr(AddrInst->getOperand(1), Depth+1)) | ||||
4539 | return true; | ||||
4540 | |||||
4541 | // Otherwise we definitely can't merge the ADD in. | ||||
4542 | AddrMode = BackupAddrMode; | ||||
4543 | AddrModeInsts.resize(OldSize); | ||||
4544 | TPT.rollback(LastKnownGood); | ||||
4545 | break; | ||||
4546 | } | ||||
4547 | //case Instruction::Or: | ||||
4548 | // TODO: We can handle "Or Val, Imm" iff this OR is equivalent to an ADD. | ||||
4549 | //break; | ||||
4550 | case Instruction::Mul: | ||||
4551 | case Instruction::Shl: { | ||||
4552 | // Can only handle X*C and X << C. | ||||
4553 | AddrMode.InBounds = false; | ||||
4554 | ConstantInt *RHS = dyn_cast<ConstantInt>(AddrInst->getOperand(1)); | ||||
4555 | if (!RHS || RHS->getBitWidth() > 64) | ||||
4556 | return false; | ||||
4557 | int64_t Scale = RHS->getSExtValue(); | ||||
4558 | if (Opcode == Instruction::Shl) | ||||
4559 | Scale = 1LL << Scale; | ||||
4560 | |||||
4561 | return matchScaledValue(AddrInst->getOperand(0), Scale, Depth); | ||||
4562 | } | ||||
4563 | case Instruction::GetElementPtr: { | ||||
4564 | // Scan the GEP. We check it if it contains constant offsets and at most | ||||
4565 | // one variable offset. | ||||
4566 | int VariableOperand = -1; | ||||
4567 | unsigned VariableScale = 0; | ||||
4568 | |||||
4569 | int64_t ConstantOffset = 0; | ||||
4570 | gep_type_iterator GTI = gep_type_begin(AddrInst); | ||||
4571 | for (unsigned i = 1, e = AddrInst->getNumOperands(); i != e; ++i, ++GTI) { | ||||
4572 | if (StructType *STy = GTI.getStructTypeOrNull()) { | ||||
4573 | const StructLayout *SL = DL.getStructLayout(STy); | ||||
4574 | unsigned Idx = | ||||
4575 | cast<ConstantInt>(AddrInst->getOperand(i))->getZExtValue(); | ||||
4576 | ConstantOffset += SL->getElementOffset(Idx); | ||||
4577 | } else { | ||||
4578 | TypeSize TS = DL.getTypeAllocSize(GTI.getIndexedType()); | ||||
4579 | if (TS.isNonZero()) { | ||||
4580 | // The optimisations below currently only work for fixed offsets. | ||||
4581 | if (TS.isScalable()) | ||||
4582 | return false; | ||||
4583 | int64_t TypeSize = TS.getFixedSize(); | ||||
4584 | if (ConstantInt *CI = | ||||
4585 | dyn_cast<ConstantInt>(AddrInst->getOperand(i))) { | ||||
4586 | const APInt &CVal = CI->getValue(); | ||||
4587 | if (CVal.getMinSignedBits() <= 64) { | ||||
4588 | ConstantOffset += CVal.getSExtValue() * TypeSize; | ||||
4589 | continue; | ||||
4590 | } | ||||
4591 | } | ||||
4592 | // We only allow one variable index at the moment. | ||||
4593 | if (VariableOperand != -1) | ||||
4594 | return false; | ||||
4595 | |||||
4596 | // Remember the variable index. | ||||
4597 | VariableOperand = i; | ||||
4598 | VariableScale = TypeSize; | ||||
4599 | } | ||||
4600 | } | ||||
4601 | } | ||||
4602 | |||||
4603 | // A common case is for the GEP to only do a constant offset. In this case, | ||||
4604 | // just add it to the disp field and check validity. | ||||
4605 | if (VariableOperand == -1) { | ||||
4606 | AddrMode.BaseOffs += ConstantOffset; | ||||
4607 | if (ConstantOffset == 0 || | ||||
4608 | TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) { | ||||
4609 | // Check to see if we can fold the base pointer in too. | ||||
4610 | if (matchAddr(AddrInst->getOperand(0), Depth+1)) { | ||||
4611 | if (!cast<GEPOperator>(AddrInst)->isInBounds()) | ||||
4612 | AddrMode.InBounds = false; | ||||
4613 | return true; | ||||
4614 | } | ||||
4615 | } else if (EnableGEPOffsetSplit && isa<GetElementPtrInst>(AddrInst) && | ||||
4616 | TLI.shouldConsiderGEPOffsetSplit() && Depth == 0 && | ||||
4617 | ConstantOffset > 0) { | ||||
4618 | // Record GEPs with non-zero offsets as candidates for splitting in the | ||||
4619 | // event that the offset cannot fit into the r+i addressing mode. | ||||
4620 | // Simple and common case that only one GEP is used in calculating the | ||||
4621 | // address for the memory access. | ||||
4622 | Value *Base = AddrInst->getOperand(0); | ||||
4623 | auto *BaseI = dyn_cast<Instruction>(Base); | ||||
4624 | auto *GEP = cast<GetElementPtrInst>(AddrInst); | ||||
4625 | if (isa<Argument>(Base) || isa<GlobalValue>(Base) || | ||||
4626 | (BaseI && !isa<CastInst>(BaseI) && | ||||
4627 | !isa<GetElementPtrInst>(BaseI))) { | ||||
4628 | // Make sure the parent block allows inserting non-PHI instructions | ||||
4629 | // before the terminator. | ||||
4630 | BasicBlock *Parent = | ||||
4631 | BaseI ? BaseI->getParent() : &GEP->getFunction()->getEntryBlock(); | ||||
4632 | if (!Parent->getTerminator()->isEHPad()) | ||||
4633 | LargeOffsetGEP = std::make_pair(GEP, ConstantOffset); | ||||
4634 | } | ||||
4635 | } | ||||
4636 | AddrMode.BaseOffs -= ConstantOffset; | ||||
4637 | return false; | ||||
4638 | } | ||||
4639 | |||||
4640 | // Save the valid addressing mode in case we can't match. | ||||
4641 | ExtAddrMode BackupAddrMode = AddrMode; | ||||
4642 | unsigned OldSize = AddrModeInsts.size(); | ||||
4643 | |||||
4644 | // See if the scale and offset amount is valid for this target. | ||||
4645 | AddrMode.BaseOffs += ConstantOffset; | ||||
4646 | if (!cast<GEPOperator>(AddrInst)->isInBounds()) | ||||
4647 | AddrMode.InBounds = false; | ||||
4648 | |||||
4649 | // Match the base operand of the GEP. | ||||
4650 | if (!matchAddr(AddrInst->getOperand(0), Depth+1)) { | ||||
4651 | // If it couldn't be matched, just stuff the value in a register. | ||||
4652 | if (AddrMode.HasBaseReg) { | ||||
4653 | AddrMode = BackupAddrMode; | ||||
4654 | AddrModeInsts.resize(OldSize); | ||||
4655 | return false; | ||||
4656 | } | ||||
4657 | AddrMode.HasBaseReg = true; | ||||
4658 | AddrMode.BaseReg = AddrInst->getOperand(0); | ||||
4659 | } | ||||
4660 | |||||
4661 | // Match the remaining variable portion of the GEP. | ||||
4662 | if (!matchScaledValue(AddrInst->getOperand(VariableOperand), VariableScale, | ||||
4663 | Depth)) { | ||||
4664 | // If it couldn't be matched, try stuffing the base into a register | ||||
4665 | // instead of matching it, and retrying the match of the scale. | ||||
4666 | AddrMode = BackupAddrMode; | ||||
4667 | AddrModeInsts.resize(OldSize); | ||||
4668 | if (AddrMode.HasBaseReg) | ||||
4669 | return false; | ||||
4670 | AddrMode.HasBaseReg = true; | ||||
4671 | AddrMode.BaseReg = AddrInst->getOperand(0); | ||||
4672 | AddrMode.BaseOffs += ConstantOffset; | ||||
4673 | if (!matchScaledValue(AddrInst->getOperand(VariableOperand), | ||||
4674 | VariableScale, Depth)) { | ||||
4675 | // If even that didn't work, bail. | ||||
4676 | AddrMode = BackupAddrMode; | ||||
4677 | AddrModeInsts.resize(OldSize); | ||||
4678 | return false; | ||||
4679 | } | ||||
4680 | } | ||||
4681 | |||||
4682 | return true; | ||||
4683 | } | ||||
4684 | case Instruction::SExt: | ||||
4685 | case Instruction::ZExt: { | ||||
4686 | Instruction *Ext = dyn_cast<Instruction>(AddrInst); | ||||
4687 | if (!Ext) | ||||
4688 | return false; | ||||
4689 | |||||
4690 | // Try to move this ext out of the way of the addressing mode. | ||||
4691 | // Ask for a method for doing so. | ||||
4692 | TypePromotionHelper::Action TPH = | ||||
4693 | TypePromotionHelper::getAction(Ext, InsertedInsts, TLI, PromotedInsts); | ||||
4694 | if (!TPH) | ||||
4695 | return false; | ||||
4696 | |||||
4697 | TypePromotionTransaction::ConstRestorationPt LastKnownGood = | ||||
4698 | TPT.getRestorationPoint(); | ||||
4699 | unsigned CreatedInstsCost = 0; | ||||
4700 | unsigned ExtCost = !TLI.isExtFree(Ext); | ||||
4701 | Value *PromotedOperand = | ||||
4702 | TPH(Ext, TPT, PromotedInsts, CreatedInstsCost, nullptr, nullptr, TLI); | ||||
4703 | // SExt has been moved away. | ||||
4704 | // Thus either it will be rematched later in the recursive calls or it is | ||||
4705 | // gone. Anyway, we must not fold it into the addressing mode at this point. | ||||
4706 | // E.g., | ||||
4707 | // op = add opnd, 1 | ||||
4708 | // idx = ext op | ||||
4709 | // addr = gep base, idx | ||||
4710 | // is now: | ||||
4711 | // promotedOpnd = ext opnd <- no match here | ||||
4712 | // op = promoted_add promotedOpnd, 1 <- match (later in recursive calls) | ||||
4713 | // addr = gep base, op <- match | ||||
4714 | if (MovedAway) | ||||
4715 | *MovedAway = true; | ||||
4716 | |||||
4717 | assert(PromotedOperand &&((void)0) | ||||
4718 | "TypePromotionHelper should have filtered out those cases")((void)0); | ||||
4719 | |||||
4720 | ExtAddrMode BackupAddrMode = AddrMode; | ||||
4721 | unsigned OldSize = AddrModeInsts.size(); | ||||
4722 | |||||
4723 | if (!matchAddr(PromotedOperand, Depth) || | ||||
4724 | // The total of the new cost is equal to the cost of the created | ||||
4725 | // instructions. | ||||
4726 | // The total of the old cost is equal to the cost of the extension plus | ||||
4727 | // what we have saved in the addressing mode. | ||||
4728 | !isPromotionProfitable(CreatedInstsCost, | ||||
4729 | ExtCost + (AddrModeInsts.size() - OldSize), | ||||
4730 | PromotedOperand)) { | ||||
4731 | AddrMode = BackupAddrMode; | ||||
4732 | AddrModeInsts.resize(OldSize); | ||||
4733 | LLVM_DEBUG(dbgs() << "Sign extension does not pay off: rollback\n")do { } while (false); | ||||
4734 | TPT.rollback(LastKnownGood); | ||||
4735 | return false; | ||||
4736 | } | ||||
4737 | return true; | ||||
4738 | } | ||||
4739 | } | ||||
4740 | return false; | ||||
4741 | } | ||||
4742 | |||||
4743 | /// If we can, try to add the value of 'Addr' into the current addressing mode. | ||||
4744 | /// If Addr can't be added to AddrMode this returns false and leaves AddrMode | ||||
4745 | /// unmodified. This assumes that Addr is either a pointer type or intptr_t | ||||
4746 | /// for the target. | ||||
4747 | /// | ||||
4748 | bool AddressingModeMatcher::matchAddr(Value *Addr, unsigned Depth) { | ||||
4749 | // Start a transaction at this point that we will rollback if the matching | ||||
4750 | // fails. | ||||
4751 | TypePromotionTransaction::ConstRestorationPt LastKnownGood = | ||||
4752 | TPT.getRestorationPoint(); | ||||
4753 | if (ConstantInt *CI = dyn_cast<ConstantInt>(Addr)) { | ||||
4754 | if (CI->getValue().isSignedIntN(64)) { | ||||
4755 | // Fold in immediates if legal for the target. | ||||
4756 | AddrMode.BaseOffs += CI->getSExtValue(); | ||||
4757 | if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) | ||||
4758 | return true; | ||||
4759 | AddrMode.BaseOffs -= CI->getSExtValue(); | ||||
4760 | } | ||||
4761 | } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) { | ||||
4762 | // If this is a global variable, try to fold it into the addressing mode. | ||||
4763 | if (!AddrMode.BaseGV) { | ||||
4764 | AddrMode.BaseGV = GV; | ||||
4765 | if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) | ||||
4766 | return true; | ||||
4767 | AddrMode.BaseGV = nullptr; | ||||
4768 | } | ||||
4769 | } else if (Instruction *I = dyn_cast<Instruction>(Addr)) { | ||||
4770 | ExtAddrMode BackupAddrMode = AddrMode; | ||||
4771 | unsigned OldSize = AddrModeInsts.size(); | ||||
4772 | |||||
4773 | // Check to see if it is possible to fold this operation. | ||||
4774 | bool MovedAway = false; | ||||
4775 | if (matchOperationAddr(I, I->getOpcode(), Depth, &MovedAway)) { | ||||
4776 | // This instruction may have been moved away. If so, there is nothing | ||||
4777 | // to check here. | ||||
4778 | if (MovedAway) | ||||
4779 | return true; | ||||
4780 | // Okay, it's possible to fold this. Check to see if it is actually | ||||
4781 | // *profitable* to do so. We use a simple cost model to avoid increasing | ||||
4782 | // register pressure too much. | ||||
4783 | if (I->hasOneUse() || | ||||
4784 | isProfitableToFoldIntoAddressingMode(I, BackupAddrMode, AddrMode)) { | ||||
4785 | AddrModeInsts.push_back(I); | ||||
4786 | return true; | ||||
4787 | } | ||||
4788 | |||||
4789 | // It isn't profitable to do this, roll back. | ||||
4790 | //cerr << "NOT FOLDING: " << *I; | ||||
4791 | AddrMode = BackupAddrMode; | ||||
4792 | AddrModeInsts.resize(OldSize); | ||||
4793 | TPT.rollback(LastKnownGood); | ||||
4794 | } | ||||
4795 | } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) { | ||||
4796 | if (matchOperationAddr(CE, CE->getOpcode(), Depth)) | ||||
4797 | return true; | ||||
4798 | TPT.rollback(LastKnownGood); | ||||
4799 | } else if (isa<ConstantPointerNull>(Addr)) { | ||||
4800 | // Null pointer gets folded without affecting the addressing mode. | ||||
4801 | return true; | ||||
4802 | } | ||||
4803 | |||||
4804 | // Worse case, the target should support [reg] addressing modes. :) | ||||
4805 | if (!AddrMode.HasBaseReg) { | ||||
4806 | AddrMode.HasBaseReg = true; | ||||
4807 | AddrMode.BaseReg = Addr; | ||||
4808 | // Still check for legality in case the target supports [imm] but not [i+r]. | ||||
4809 | if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) | ||||
4810 | return true; | ||||
4811 | AddrMode.HasBaseReg = false; | ||||
4812 | AddrMode.BaseReg = nullptr; | ||||
4813 | } | ||||
4814 | |||||
4815 | // If the base register is already taken, see if we can do [r+r]. | ||||
4816 | if (AddrMode.Scale == 0) { | ||||
4817 | AddrMode.Scale = 1; | ||||
4818 | AddrMode.ScaledReg = Addr; | ||||
4819 | if (TLI.isLegalAddressingMode(DL, AddrMode, AccessTy, AddrSpace)) | ||||
4820 | return true; | ||||
4821 | AddrMode.Scale = 0; | ||||
4822 | AddrMode.ScaledReg = nullptr; | ||||
4823 | } | ||||
4824 | // Couldn't match. | ||||
4825 | TPT.rollback(LastKnownGood); | ||||
4826 | return false; | ||||
4827 | } | ||||
4828 | |||||
4829 | /// Check to see if all uses of OpVal by the specified inline asm call are due | ||||
4830 | /// to memory operands. If so, return true, otherwise return false. | ||||
4831 | static bool IsOperandAMemoryOperand(CallInst *CI, InlineAsm *IA, Value *OpVal, | ||||
4832 | const TargetLowering &TLI, | ||||
4833 | const TargetRegisterInfo &TRI) { | ||||
4834 | const Function *F = CI->getFunction(); | ||||
4835 | TargetLowering::AsmOperandInfoVector TargetConstraints = | ||||
4836 | TLI.ParseConstraints(F->getParent()->getDataLayout(), &TRI, *CI); | ||||
4837 | |||||
4838 | for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) { | ||||
4839 | TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i]; | ||||
4840 | |||||
4841 | // Compute the constraint code and ConstraintType to use. | ||||
4842 | TLI.ComputeConstraintToUse(OpInfo, SDValue()); | ||||
4843 | |||||
4844 | // If this asm operand is our Value*, and if it isn't an indirect memory | ||||
4845 | // operand, we can't fold it! | ||||
4846 | if (OpInfo.CallOperandVal == OpVal && | ||||
4847 | (OpInfo.ConstraintType != TargetLowering::C_Memory || | ||||
4848 | !OpInfo.isIndirect)) | ||||
4849 | return false; | ||||
4850 | } | ||||
4851 | |||||
4852 | return true; | ||||
4853 | } | ||||
4854 | |||||
4855 | // Max number of memory uses to look at before aborting the search to conserve | ||||
4856 | // compile time. | ||||
4857 | static constexpr int MaxMemoryUsesToScan = 20; | ||||
4858 | |||||
4859 | /// Recursively walk all the uses of I until we find a memory use. | ||||
4860 | /// If we find an obviously non-foldable instruction, return true. | ||||
4861 | /// Add the ultimately found memory instructions to MemoryUses. | ||||
4862 | static bool FindAllMemoryUses( | ||||
4863 | Instruction *I, | ||||
4864 | SmallVectorImpl<std::pair<Instruction *, unsigned>> &MemoryUses, | ||||
4865 | SmallPtrSetImpl<Instruction *> &ConsideredInsts, const TargetLowering &TLI, | ||||
4866 | const TargetRegisterInfo &TRI, bool OptSize, ProfileSummaryInfo *PSI, | ||||
4867 | BlockFrequencyInfo *BFI, int SeenInsts = 0) { | ||||
4868 | // If we already considered this instruction, we're done. | ||||
4869 | if (!ConsideredInsts.insert(I).second) | ||||
4870 | return false; | ||||
4871 | |||||
4872 | // If this is an obviously unfoldable instruction, bail out. | ||||
4873 | if (!MightBeFoldableInst(I)) | ||||
4874 | return true; | ||||
4875 | |||||
4876 | // Loop over all the uses, recursively processing them. | ||||
4877 | for (Use &U : I->uses()) { | ||||
4878 | // Conservatively return true if we're seeing a large number or a deep chain | ||||
4879 | // of users. This avoids excessive compilation times in pathological cases. | ||||
4880 | if (SeenInsts++ >= MaxMemoryUsesToScan) | ||||
4881 | return true; | ||||
4882 | |||||
4883 | Instruction *UserI = cast<Instruction>(U.getUser()); | ||||
4884 | if (LoadInst *LI = dyn_cast<LoadInst>(UserI)) { | ||||
4885 | MemoryUses.push_back(std::make_pair(LI, U.getOperandNo())); | ||||
4886 | continue; | ||||
4887 | } | ||||
4888 | |||||
4889 | if (StoreInst *SI = dyn_cast<StoreInst>(UserI)) { | ||||
4890 | unsigned opNo = U.getOperandNo(); | ||||
4891 | if (opNo != StoreInst::getPointerOperandIndex()) | ||||
4892 | return true; // Storing addr, not into addr. | ||||
4893 | MemoryUses.push_back(std::make_pair(SI, opNo)); | ||||
4894 | continue; | ||||
4895 | } | ||||
4896 | |||||
4897 | if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(UserI)) { | ||||
4898 | unsigned opNo = U.getOperandNo(); | ||||
4899 | if (opNo != AtomicRMWInst::getPointerOperandIndex()) | ||||
4900 | return true; // Storing addr, not into addr. | ||||
4901 | MemoryUses.push_back(std::make_pair(RMW, opNo)); | ||||
4902 | continue; | ||||
4903 | } | ||||
4904 | |||||
4905 | if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(UserI)) { | ||||
4906 | unsigned opNo = U.getOperandNo(); | ||||
4907 | if (opNo != AtomicCmpXchgInst::getPointerOperandIndex()) | ||||
4908 | return true; // Storing addr, not into addr. | ||||
4909 | MemoryUses.push_back(std::make_pair(CmpX, opNo)); | ||||
4910 | continue; | ||||
4911 | } | ||||
4912 | |||||
4913 | if (CallInst *CI = dyn_cast<CallInst>(UserI)) { | ||||
4914 | if (CI->hasFnAttr(Attribute::Cold)) { | ||||
4915 | // If this is a cold call, we can sink the addressing calculation into | ||||
4916 | // the cold path. See optimizeCallInst | ||||
4917 | bool OptForSize = OptSize || | ||||
4918 | llvm::shouldOptimizeForSize(CI->getParent(), PSI, BFI); | ||||
4919 | if (!OptForSize) | ||||
4920 | continue; | ||||
4921 | } | ||||
4922 | |||||
4923 | InlineAsm *IA = dyn_cast<InlineAsm>(CI->getCalledOperand()); | ||||
4924 | if (!IA) return true; | ||||
4925 | |||||
4926 | // If this is a memory operand, we're cool, otherwise bail out. | ||||
4927 | if (!IsOperandAMemoryOperand(CI, IA, I, TLI, TRI)) | ||||
4928 | return true; | ||||
4929 | continue; | ||||
4930 | } | ||||
4931 | |||||
4932 | if (FindAllMemoryUses(UserI, MemoryUses, ConsideredInsts, TLI, TRI, OptSize, | ||||
4933 | PSI, BFI, SeenInsts)) | ||||
4934 | return true; | ||||
4935 | } | ||||
4936 | |||||
4937 | return false; | ||||
4938 | } | ||||
4939 | |||||
4940 | /// Return true if Val is already known to be live at the use site that we're | ||||
4941 | /// folding it into. If so, there is no cost to include it in the addressing | ||||
4942 | /// mode. KnownLive1 and KnownLive2 are two values that we know are live at the | ||||
4943 | /// instruction already. | ||||
4944 | bool AddressingModeMatcher::valueAlreadyLiveAtInst(Value *Val,Value *KnownLive1, | ||||
4945 | Value *KnownLive2) { | ||||
4946 | // If Val is either of the known-live values, we know it is live! | ||||
4947 | if (Val == nullptr || Val == KnownLive1 || Val == KnownLive2) | ||||
4948 | return true; | ||||
4949 | |||||
4950 | // All values other than instructions and arguments (e.g. constants) are live. | ||||
4951 | if (!isa<Instruction>(Val) && !isa<Argument>(Val)) return true; | ||||
4952 | |||||
4953 | // If Val is a constant sized alloca in the entry block, it is live, this is | ||||
4954 | // true because it is just a reference to the stack/frame pointer, which is | ||||
4955 | // live for the whole function. | ||||
4956 | if (AllocaInst *AI = dyn_cast<AllocaInst>(Val)) | ||||
4957 | if (AI->isStaticAlloca()) | ||||
4958 | return true; | ||||
4959 | |||||
4960 | // Check to see if this value is already used in the memory instruction's | ||||
4961 | // block. If so, it's already live into the block at the very least, so we | ||||
4962 | // can reasonably fold it. | ||||
4963 | return Val->isUsedInBasicBlock(MemoryInst->getParent()); | ||||
4964 | } | ||||
4965 | |||||
4966 | /// It is possible for the addressing mode of the machine to fold the specified | ||||
4967 | /// instruction into a load or store that ultimately uses it. | ||||
4968 | /// However, the specified instruction has multiple uses. | ||||
4969 | /// Given this, it may actually increase register pressure to fold it | ||||
4970 | /// into the load. For example, consider this code: | ||||
4971 | /// | ||||
4972 | /// X = ... | ||||
4973 | /// Y = X+1 | ||||
4974 | /// use(Y) -> nonload/store | ||||
4975 | /// Z = Y+1 | ||||
4976 | /// load Z | ||||
4977 | /// | ||||
4978 | /// In this case, Y has multiple uses, and can be folded into the load of Z | ||||
4979 | /// (yielding load [X+2]). However, doing this will cause both "X" and "X+1" to | ||||
4980 | /// be live at the use(Y) line. If we don't fold Y into load Z, we use one | ||||
4981 | /// fewer register. Since Y can't be folded into "use(Y)" we don't increase the | ||||
4982 | /// number of computations either. | ||||
4983 | /// | ||||
4984 | /// Note that this (like most of CodeGenPrepare) is just a rough heuristic. If | ||||
4985 | /// X was live across 'load Z' for other reasons, we actually *would* want to | ||||
4986 | /// fold the addressing mode in the Z case. This would make Y die earlier. | ||||
4987 | bool AddressingModeMatcher:: | ||||
4988 | isProfitableToFoldIntoAddressingMode(Instruction *I, ExtAddrMode &AMBefore, | ||||
4989 | ExtAddrMode &AMAfter) { | ||||
4990 | if (IgnoreProfitability) return true; | ||||
4991 | |||||
4992 | // AMBefore is the addressing mode before this instruction was folded into it, | ||||
4993 | // and AMAfter is the addressing mode after the instruction was folded. Get | ||||
4994 | // the set of registers referenced by AMAfter and subtract out those | ||||
4995 | // referenced by AMBefore: this is the set of values which folding in this | ||||
4996 | // address extends the lifetime of. | ||||
4997 | // | ||||
4998 | // Note that there are only two potential values being referenced here, | ||||
4999 | // BaseReg and ScaleReg (global addresses are always available, as are any | ||||
5000 | // folded immediates). | ||||
5001 | Value *BaseReg = AMAfter.BaseReg, *ScaledReg = AMAfter.ScaledReg; | ||||
5002 | |||||
5003 | // If the BaseReg or ScaledReg was referenced by the previous addrmode, their | ||||
5004 | // lifetime wasn't extended by adding this instruction. | ||||
5005 | if (valueAlreadyLiveAtInst(BaseReg, AMBefore.BaseReg, AMBefore.ScaledReg)) | ||||
5006 | BaseReg = nullptr; | ||||
5007 | if (valueAlreadyLiveAtInst(ScaledReg, AMBefore.BaseReg, AMBefore.ScaledReg)) | ||||
5008 | ScaledReg = nullptr; | ||||
5009 | |||||
5010 | // If folding this instruction (and it's subexprs) didn't extend any live | ||||
5011 | // ranges, we're ok with it. | ||||
5012 | if (!BaseReg && !ScaledReg) | ||||
5013 | return true; | ||||
5014 | |||||
5015 | // If all uses of this instruction can have the address mode sunk into them, | ||||
5016 | // we can remove the addressing mode and effectively trade one live register | ||||
5017 | // for another (at worst.) In this context, folding an addressing mode into | ||||
5018 | // the use is just a particularly nice way of sinking it. | ||||
5019 | SmallVector<std::pair<Instruction*,unsigned>, 16> MemoryUses; | ||||
5020 | SmallPtrSet<Instruction*, 16> ConsideredInsts; | ||||
5021 | if (FindAllMemoryUses(I, MemoryUses, ConsideredInsts, TLI, TRI, OptSize, | ||||
5022 | PSI, BFI)) | ||||
5023 | return false; // Has a non-memory, non-foldable use! | ||||
5024 | |||||
5025 | // Now that we know that all uses of this instruction are part of a chain of | ||||
5026 | // computation involving only operations that could theoretically be folded | ||||
5027 | // into a memory use, loop over each of these memory operation uses and see | ||||
5028 | // if they could *actually* fold the instruction. The assumption is that | ||||
5029 | // addressing modes are cheap and that duplicating the computation involved | ||||
5030 | // many times is worthwhile, even on a fastpath. For sinking candidates | ||||
5031 | // (i.e. cold call sites), this serves as a way to prevent excessive code | ||||
5032 | // growth since most architectures have some reasonable small and fast way to | ||||
5033 | // compute an effective address. (i.e LEA on x86) | ||||
5034 | SmallVector<Instruction*, 32> MatchedAddrModeInsts; | ||||
5035 | for (unsigned i = 0, e = MemoryUses.size(); i != e; ++i) { | ||||
5036 | Instruction *User = MemoryUses[i].first; | ||||
5037 | unsigned OpNo = MemoryUses[i].second; | ||||
5038 | |||||
5039 | // Get the access type of this use. If the use isn't a pointer, we don't | ||||
5040 | // know what it accesses. | ||||
5041 | Value *Address = User->getOperand(OpNo); | ||||
5042 | PointerType *AddrTy = dyn_cast<PointerType>(Address->getType()); | ||||
5043 | if (!AddrTy) | ||||
5044 | return false; | ||||
5045 | Type *AddressAccessTy = AddrTy->getElementType(); | ||||
5046 | unsigned AS = AddrTy->getAddressSpace(); | ||||
5047 | |||||
5048 | // Do a match against the root of this address, ignoring profitability. This | ||||
5049 | // will tell us if the addressing mode for the memory operation will | ||||
5050 | // *actually* cover the shared instruction. | ||||
5051 | ExtAddrMode Result; | ||||
5052 | std::pair<AssertingVH<GetElementPtrInst>, int64_t> LargeOffsetGEP(nullptr, | ||||
5053 | 0); | ||||
5054 | TypePromotionTransaction::ConstRestorationPt LastKnownGood = | ||||
5055 | TPT.getRestorationPoint(); | ||||
5056 | AddressingModeMatcher Matcher(MatchedAddrModeInsts, TLI, TRI, LI, getDTFn, | ||||
5057 | AddressAccessTy, AS, MemoryInst, Result, | ||||
5058 | InsertedInsts, PromotedInsts, TPT, | ||||
5059 | LargeOffsetGEP, OptSize, PSI, BFI); | ||||
5060 | Matcher.IgnoreProfitability = true; | ||||
5061 | bool Success = Matcher.matchAddr(Address, 0); | ||||
5062 | (void)Success; assert(Success && "Couldn't select *anything*?")((void)0); | ||||
5063 | |||||
5064 | // The match was to check the profitability, the changes made are not | ||||
5065 | // part of the original matcher. Therefore, they should be dropped | ||||
5066 | // otherwise the original matcher will not present the right state. | ||||
5067 | TPT.rollback(LastKnownGood); | ||||
5068 | |||||
5069 | // If the match didn't cover I, then it won't be shared by it. | ||||
5070 | if (!is_contained(MatchedAddrModeInsts, I)) | ||||
5071 | return false; | ||||
5072 | |||||
5073 | MatchedAddrModeInsts.clear(); | ||||
5074 | } | ||||
5075 | |||||
5076 | return true; | ||||
5077 | } | ||||
5078 | |||||
5079 | /// Return true if the specified values are defined in a | ||||
5080 | /// different basic block than BB. | ||||
5081 | static bool IsNonLocalValue(Value *V, BasicBlock *BB) { | ||||
5082 | if (Instruction *I = dyn_cast<Instruction>(V)) | ||||
5083 | return I->getParent() != BB; | ||||
5084 | return false; | ||||
5085 | } | ||||
5086 | |||||
5087 | /// Sink addressing mode computation immediate before MemoryInst if doing so | ||||
5088 | /// can be done without increasing register pressure. The need for the | ||||
5089 | /// register pressure constraint means this can end up being an all or nothing | ||||
5090 | /// decision for all uses of the same addressing computation. | ||||
5091 | /// | ||||
5092 | /// Load and Store Instructions often have addressing modes that can do | ||||
5093 | /// significant amounts of computation. As such, instruction selection will try | ||||
5094 | /// to get the load or store to do as much computation as possible for the | ||||
5095 | /// program. The problem is that isel can only see within a single block. As | ||||
5096 | /// such, we sink as much legal addressing mode work into the block as possible. | ||||
5097 | /// | ||||
5098 | /// This method is used to optimize both load/store and inline asms with memory | ||||
5099 | /// operands. It's also used to sink addressing computations feeding into cold | ||||
5100 | /// call sites into their (cold) basic block. | ||||
5101 | /// | ||||
5102 | /// The motivation for handling sinking into cold blocks is that doing so can | ||||
5103 | /// both enable other address mode sinking (by satisfying the register pressure | ||||
5104 | /// constraint above), and reduce register pressure globally (by removing the | ||||
5105 | /// addressing mode computation from the fast path entirely.). | ||||
5106 | bool CodeGenPrepare::optimizeMemoryInst(Instruction *MemoryInst, Value *Addr, | ||||
5107 | Type *AccessTy, unsigned AddrSpace) { | ||||
5108 | Value *Repl = Addr; | ||||
5109 | |||||
5110 | // Try to collapse single-value PHI nodes. This is necessary to undo | ||||
5111 | // unprofitable PRE transformations. | ||||
5112 | SmallVector<Value*, 8> worklist; | ||||
5113 | SmallPtrSet<Value*, 16> Visited; | ||||
5114 | worklist.push_back(Addr); | ||||
5115 | |||||
5116 | // Use a worklist to iteratively look through PHI and select nodes, and | ||||
5117 | // ensure that the addressing mode obtained from the non-PHI/select roots of | ||||
5118 | // the graph are compatible. | ||||
5119 | bool PhiOrSelectSeen = false; | ||||
5120 | SmallVector<Instruction*, 16> AddrModeInsts; | ||||
5121 | const SimplifyQuery SQ(*DL, TLInfo); | ||||
5122 | AddressingModeCombiner AddrModes(SQ, Addr); | ||||
5123 | TypePromotionTransaction TPT(RemovedInsts); | ||||
5124 | TypePromotionTransaction::ConstRestorationPt LastKnownGood = | ||||
5125 | TPT.getRestorationPoint(); | ||||
5126 | while (!worklist.empty()) { | ||||
5127 | Value *V = worklist.back(); | ||||
5128 | worklist.pop_back(); | ||||
5129 | |||||
5130 | // We allow traversing cyclic Phi nodes. | ||||
5131 | // In case of success after this loop we ensure that traversing through | ||||
5132 | // Phi nodes ends up with all cases to compute address of the form | ||||
5133 | // BaseGV + Base + Scale * Index + Offset | ||||
5134 | // where Scale and Offset are constans and BaseGV, Base and Index | ||||
5135 | // are exactly the same Values in all cases. | ||||
5136 | // It means that BaseGV, Scale and Offset dominate our memory instruction | ||||
5137 | // and have the same value as they had in address computation represented | ||||
5138 | // as Phi. So we can safely sink address computation to memory instruction. | ||||
5139 | if (!Visited.insert(V).second) | ||||
5140 | continue; | ||||
5141 | |||||
5142 | // For a PHI node, push all of its incoming values. | ||||
5143 | if (PHINode *P = dyn_cast<PHINode>(V)) { | ||||
5144 | append_range(worklist, P->incoming_values()); | ||||
5145 | PhiOrSelectSeen = true; | ||||
5146 | continue; | ||||
5147 | } | ||||
5148 | // Similar for select. | ||||
5149 | if (SelectInst *SI = dyn_cast<SelectInst>(V)) { | ||||
5150 | worklist.push_back(SI->getFalseValue()); | ||||
5151 | worklist.push_back(SI->getTrueValue()); | ||||
5152 | PhiOrSelectSeen = true; | ||||
5153 | continue; | ||||
5154 | } | ||||
5155 | |||||
5156 | // For non-PHIs, determine the addressing mode being computed. Note that | ||||
5157 | // the result may differ depending on what other uses our candidate | ||||
5158 | // addressing instructions might have. | ||||
5159 | AddrModeInsts.clear(); | ||||
5160 | std::pair<AssertingVH<GetElementPtrInst>, int64_t> LargeOffsetGEP(nullptr, | ||||
5161 | 0); | ||||
5162 | // Defer the query (and possible computation of) the dom tree to point of | ||||
5163 | // actual use. It's expected that most address matches don't actually need | ||||
5164 | // the domtree. | ||||
5165 | auto getDTFn = [MemoryInst, this]() -> const DominatorTree & { | ||||
5166 | Function *F = MemoryInst->getParent()->getParent(); | ||||
5167 | return this->getDT(*F); | ||||
5168 | }; | ||||
5169 | ExtAddrMode NewAddrMode = AddressingModeMatcher::Match( | ||||
5170 | V, AccessTy, AddrSpace, MemoryInst, AddrModeInsts, *TLI, *LI, getDTFn, | ||||
5171 | *TRI, InsertedInsts, PromotedInsts, TPT, LargeOffsetGEP, OptSize, PSI, | ||||
5172 | BFI.get()); | ||||
5173 | |||||
5174 | GetElementPtrInst *GEP = LargeOffsetGEP.first; | ||||
5175 | if (GEP && !NewGEPBases.count(GEP)) { | ||||
5176 | // If splitting the underlying data structure can reduce the offset of a | ||||
5177 | // GEP, collect the GEP. Skip the GEPs that are the new bases of | ||||
5178 | // previously split data structures. | ||||
5179 | LargeOffsetGEPMap[GEP->getPointerOperand()].push_back(LargeOffsetGEP); | ||||
5180 | if (LargeOffsetGEPID.find(GEP) == LargeOffsetGEPID.end()) | ||||
5181 | LargeOffsetGEPID[GEP] = LargeOffsetGEPID.size(); | ||||
5182 | } | ||||
5183 | |||||
5184 | NewAddrMode.OriginalValue = V; | ||||
5185 | if (!AddrModes.addNewAddrMode(NewAddrMode)) | ||||
5186 | break; | ||||
5187 | } | ||||
5188 | |||||
5189 | // Try to combine the AddrModes we've collected. If we couldn't collect any, | ||||
5190 | // or we have multiple but either couldn't combine them or combining them | ||||
5191 | // wouldn't do anything useful, bail out now. | ||||
5192 | if (!AddrModes.combineAddrModes()) { | ||||
5193 | TPT.rollback(LastKnownGood); | ||||
5194 | return false; | ||||
5195 | } | ||||
5196 | bool Modified = TPT.commit(); | ||||
5197 | |||||
5198 | // Get the combined AddrMode (or the only AddrMode, if we only had one). | ||||
5199 | ExtAddrMode AddrMode = AddrModes.getAddrMode(); | ||||
5200 | |||||
5201 | // If all the instructions matched are already in this BB, don't do anything. | ||||
5202 | // If we saw a Phi node then it is not local definitely, and if we saw a select | ||||
5203 | // then we want to push the address calculation past it even if it's already | ||||
5204 | // in this BB. | ||||
5205 | if (!PhiOrSelectSeen && none_of(AddrModeInsts, [&](Value *V) { | ||||
5206 | return IsNonLocalValue(V, MemoryInst->getParent()); | ||||
5207 | })) { | ||||
5208 | LLVM_DEBUG(dbgs() << "CGP: Found local addrmode: " << AddrModedo { } while (false) | ||||
5209 | << "\n")do { } while (false); | ||||
5210 | return Modified; | ||||
5211 | } | ||||
5212 | |||||
5213 | // Insert this computation right after this user. Since our caller is | ||||
5214 | // scanning from the top of the BB to the bottom, reuse of the expr are | ||||
5215 | // guaranteed to happen later. | ||||
5216 | IRBuilder<> Builder(MemoryInst); | ||||
5217 | |||||
5218 | // Now that we determined the addressing expression we want to use and know | ||||
5219 | // that we have to sink it into this block. Check to see if we have already | ||||
5220 | // done this for some other load/store instr in this block. If so, reuse | ||||
5221 | // the computation. Before attempting reuse, check if the address is valid | ||||
5222 | // as it may have been erased. | ||||
5223 | |||||
5224 | WeakTrackingVH SunkAddrVH = SunkAddrs[Addr]; | ||||
5225 | |||||
5226 | Value * SunkAddr = SunkAddrVH.pointsToAliveValue() ? SunkAddrVH : nullptr; | ||||
5227 | if (SunkAddr) { | ||||
5228 | LLVM_DEBUG(dbgs() << "CGP: Reusing nonlocal addrmode: " << AddrModedo { } while (false) | ||||
5229 | << " for " << *MemoryInst << "\n")do { } while (false); | ||||
5230 | if (SunkAddr->getType() != Addr->getType()) | ||||
5231 | SunkAddr = Builder.CreatePointerCast(SunkAddr, Addr->getType()); | ||||
5232 | } else if (AddrSinkUsingGEPs || (!AddrSinkUsingGEPs.getNumOccurrences() && | ||||
5233 | SubtargetInfo->addrSinkUsingGEPs())) { | ||||
5234 | // By default, we use the GEP-based method when AA is used later. This | ||||
5235 | // prevents new inttoptr/ptrtoint pairs from degrading AA capabilities. | ||||
5236 | LLVM_DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrModedo { } while (false) | ||||
5237 | << " for " << *MemoryInst << "\n")do { } while (false); | ||||
5238 | Type *IntPtrTy = DL->getIntPtrType(Addr->getType()); | ||||
5239 | Value *ResultPtr = nullptr, *ResultIndex = nullptr; | ||||
5240 | |||||
5241 | // First, find the pointer. | ||||
5242 | if (AddrMode.BaseReg && AddrMode.BaseReg->getType()->isPointerTy()) { | ||||
5243 | ResultPtr = AddrMode.BaseReg; | ||||
5244 | AddrMode.BaseReg = nullptr; | ||||
5245 | } | ||||
5246 | |||||
5247 | if (AddrMode.Scale && AddrMode.ScaledReg->getType()->isPointerTy()) { | ||||
5248 | // We can't add more than one pointer together, nor can we scale a | ||||
5249 | // pointer (both of which seem meaningless). | ||||
5250 | if (ResultPtr || AddrMode.Scale != 1) | ||||
5251 | return Modified; | ||||
5252 | |||||
5253 | ResultPtr = AddrMode.ScaledReg; | ||||
5254 | AddrMode.Scale = 0; | ||||
5255 | } | ||||
5256 | |||||
5257 | // It is only safe to sign extend the BaseReg if we know that the math | ||||
5258 | // required to create it did not overflow before we extend it. Since | ||||
5259 | // the original IR value was tossed in favor of a constant back when | ||||
5260 | // the AddrMode was created we need to bail out gracefully if widths | ||||
5261 | // do not match instead of extending it. | ||||
5262 | // | ||||
5263 | // (See below for code to add the scale.) | ||||
5264 | if (AddrMode.Scale) { | ||||
5265 | Type *ScaledRegTy = AddrMode.ScaledReg->getType(); | ||||
5266 | if (cast<IntegerType>(IntPtrTy)->getBitWidth() > | ||||
5267 | cast<IntegerType>(ScaledRegTy)->getBitWidth()) | ||||
5268 | return Modified; | ||||
5269 | } | ||||
5270 | |||||
5271 | if (AddrMode.BaseGV) { | ||||
5272 | if (ResultPtr) | ||||
5273 | return Modified; | ||||
5274 | |||||
5275 | ResultPtr = AddrMode.BaseGV; | ||||
5276 | } | ||||
5277 | |||||
5278 | // If the real base value actually came from an inttoptr, then the matcher | ||||
5279 | // will look through it and provide only the integer value. In that case, | ||||
5280 | // use it here. | ||||
5281 | if (!DL->isNonIntegralPointerType(Addr->getType())) { | ||||
5282 | if (!ResultPtr && AddrMode.BaseReg) { | ||||
5283 | ResultPtr = Builder.CreateIntToPtr(AddrMode.BaseReg, Addr->getType(), | ||||
5284 | "sunkaddr"); | ||||
5285 | AddrMode.BaseReg = nullptr; | ||||
5286 | } else if (!ResultPtr && AddrMode.Scale == 1) { | ||||
5287 | ResultPtr = Builder.CreateIntToPtr(AddrMode.ScaledReg, Addr->getType(), | ||||
5288 | "sunkaddr"); | ||||
5289 | AddrMode.Scale = 0; | ||||
5290 | } | ||||
5291 | } | ||||
5292 | |||||
5293 | if (!ResultPtr && | ||||
5294 | !AddrMode.BaseReg && !AddrMode.Scale && !AddrMode.BaseOffs) { | ||||
5295 | SunkAddr = Constant::getNullValue(Addr->getType()); | ||||
5296 | } else if (!ResultPtr) { | ||||
5297 | return Modified; | ||||
5298 | } else { | ||||
5299 | Type *I8PtrTy = | ||||
5300 | Builder.getInt8PtrTy(Addr->getType()->getPointerAddressSpace()); | ||||
5301 | Type *I8Ty = Builder.getInt8Ty(); | ||||
5302 | |||||
5303 | // Start with the base register. Do this first so that subsequent address | ||||
5304 | // matching finds it last, which will prevent it from trying to match it | ||||
5305 | // as the scaled value in case it happens to be a mul. That would be | ||||
5306 | // problematic if we've sunk a different mul for the scale, because then | ||||
5307 | // we'd end up sinking both muls. | ||||
5308 | if (AddrMode.BaseReg) { | ||||
5309 | Value *V = AddrMode.BaseReg; | ||||
5310 | if (V->getType() != IntPtrTy) | ||||
5311 | V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr"); | ||||
5312 | |||||
5313 | ResultIndex = V; | ||||
5314 | } | ||||
5315 | |||||
5316 | // Add the scale value. | ||||
5317 | if (AddrMode.Scale) { | ||||
5318 | Value *V = AddrMode.ScaledReg; | ||||
5319 | if (V->getType() == IntPtrTy) { | ||||
5320 | // done. | ||||
5321 | } else { | ||||
5322 | assert(cast<IntegerType>(IntPtrTy)->getBitWidth() <((void)0) | ||||
5323 | cast<IntegerType>(V->getType())->getBitWidth() &&((void)0) | ||||
5324 | "We can't transform if ScaledReg is too narrow")((void)0); | ||||
5325 | V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr"); | ||||
5326 | } | ||||
5327 | |||||
5328 | if (AddrMode.Scale != 1) | ||||
5329 | V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale), | ||||
5330 | "sunkaddr"); | ||||
5331 | if (ResultIndex) | ||||
5332 | ResultIndex = Builder.CreateAdd(ResultIndex, V, "sunkaddr"); | ||||
5333 | else | ||||
5334 | ResultIndex = V; | ||||
5335 | } | ||||
5336 | |||||
5337 | // Add in the Base Offset if present. | ||||
5338 | if (AddrMode.BaseOffs) { | ||||
5339 | Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs); | ||||
5340 | if (ResultIndex) { | ||||
5341 | // We need to add this separately from the scale above to help with | ||||
5342 | // SDAG consecutive load/store merging. | ||||
5343 | if (ResultPtr->getType() != I8PtrTy) | ||||
5344 | ResultPtr = Builder.CreatePointerCast(ResultPtr, I8PtrTy); | ||||
5345 | ResultPtr = | ||||
5346 | AddrMode.InBounds | ||||
5347 | ? Builder.CreateInBoundsGEP(I8Ty, ResultPtr, ResultIndex, | ||||
5348 | "sunkaddr") | ||||
5349 | : Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr"); | ||||
5350 | } | ||||
5351 | |||||
5352 | ResultIndex = V; | ||||
5353 | } | ||||
5354 | |||||
5355 | if (!ResultIndex) { | ||||
5356 | SunkAddr = ResultPtr; | ||||
5357 | } else { | ||||
5358 | if (ResultPtr->getType() != I8PtrTy) | ||||
5359 | ResultPtr = Builder.CreatePointerCast(ResultPtr, I8PtrTy); | ||||
5360 | SunkAddr = | ||||
5361 | AddrMode.InBounds | ||||
5362 | ? Builder.CreateInBoundsGEP(I8Ty, ResultPtr, ResultIndex, | ||||
5363 | "sunkaddr") | ||||
5364 | : Builder.CreateGEP(I8Ty, ResultPtr, ResultIndex, "sunkaddr"); | ||||
5365 | } | ||||
5366 | |||||
5367 | if (SunkAddr->getType() != Addr->getType()) | ||||
5368 | SunkAddr = Builder.CreatePointerCast(SunkAddr, Addr->getType()); | ||||
5369 | } | ||||
5370 | } else { | ||||
5371 | // We'd require a ptrtoint/inttoptr down the line, which we can't do for | ||||
5372 | // non-integral pointers, so in that case bail out now. | ||||
5373 | Type *BaseTy = AddrMode.BaseReg ? AddrMode.BaseReg->getType() : nullptr; | ||||
5374 | Type *ScaleTy = AddrMode.Scale ? AddrMode.ScaledReg->getType() : nullptr; | ||||
5375 | PointerType *BasePtrTy = dyn_cast_or_null<PointerType>(BaseTy); | ||||
5376 | PointerType *ScalePtrTy = dyn_cast_or_null<PointerType>(ScaleTy); | ||||
5377 | if (DL->isNonIntegralPointerType(Addr->getType()) || | ||||
5378 | (BasePtrTy && DL->isNonIntegralPointerType(BasePtrTy)) || | ||||
5379 | (ScalePtrTy && DL->isNonIntegralPointerType(ScalePtrTy)) || | ||||
5380 | (AddrMode.BaseGV && | ||||
5381 | DL->isNonIntegralPointerType(AddrMode.BaseGV->getType()))) | ||||
5382 | return Modified; | ||||
5383 | |||||
5384 | LLVM_DEBUG(dbgs() << "CGP: SINKING nonlocal addrmode: " << AddrModedo { } while (false) | ||||
5385 | << " for " << *MemoryInst << "\n")do { } while (false); | ||||
5386 | Type *IntPtrTy = DL->getIntPtrType(Addr->getType()); | ||||
5387 | Value *Result = nullptr; | ||||
5388 | |||||
5389 | // Start with the base register. Do this first so that subsequent address | ||||
5390 | // matching finds it last, which will prevent it from trying to match it | ||||
5391 | // as the scaled value in case it happens to be a mul. That would be | ||||
5392 | // problematic if we've sunk a different mul for the scale, because then | ||||
5393 | // we'd end up sinking both muls. | ||||
5394 | if (AddrMode.BaseReg) { | ||||
5395 | Value *V = AddrMode.BaseReg; | ||||
5396 | if (V->getType()->isPointerTy()) | ||||
5397 | V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr"); | ||||
5398 | if (V->getType() != IntPtrTy) | ||||
5399 | V = Builder.CreateIntCast(V, IntPtrTy, /*isSigned=*/true, "sunkaddr"); | ||||
5400 | Result = V; | ||||
5401 | } | ||||
5402 | |||||
5403 | // Add the scale value. | ||||
5404 | if (AddrMode.Scale) { | ||||
5405 | Value *V = AddrMode.ScaledReg; | ||||
5406 | if (V->getType() == IntPtrTy) { | ||||
5407 | // done. | ||||
5408 | } else if (V->getType()->isPointerTy()) { | ||||
5409 | V = Builder.CreatePtrToInt(V, IntPtrTy, "sunkaddr"); | ||||
5410 | } else if (cast<IntegerType>(IntPtrTy)->getBitWidth() < | ||||
5411 | cast<IntegerType>(V->getType())->getBitWidth()) { | ||||
5412 | V = Builder.CreateTrunc(V, IntPtrTy, "sunkaddr"); | ||||
5413 | } else { | ||||
5414 | // It is only safe to sign extend the BaseReg if we know that the math | ||||
5415 | // required to create it did not overflow before we extend it. Since | ||||
5416 | // the original IR value was tossed in favor of a constant back when | ||||
5417 | // the AddrMode was created we need to bail out gracefully if widths | ||||
5418 | // do not match instead of extending it. | ||||
5419 | Instruction *I = dyn_cast_or_null<Instruction>(Result); | ||||
5420 | if (I && (Result != AddrMode.BaseReg)) | ||||
5421 | I->eraseFromParent(); | ||||
5422 | return Modified; | ||||
5423 | } | ||||
5424 | if (AddrMode.Scale != 1) | ||||
5425 | V = Builder.CreateMul(V, ConstantInt::get(IntPtrTy, AddrMode.Scale), | ||||
5426 | "sunkaddr"); | ||||
5427 | if (Result) | ||||
5428 | Result = Builder.CreateAdd(Result, V, "sunkaddr"); | ||||
5429 | else | ||||
5430 | Result = V; | ||||
5431 | } | ||||
5432 | |||||
5433 | // Add in the BaseGV if present. | ||||
5434 | if (AddrMode.BaseGV) { | ||||
5435 | Value *V = Builder.CreatePtrToInt(AddrMode.BaseGV, IntPtrTy, "sunkaddr"); | ||||
5436 | if (Result) | ||||
5437 | Result = Builder.CreateAdd(Result, V, "sunkaddr"); | ||||
5438 | else | ||||
5439 | Result = V; | ||||
5440 | } | ||||
5441 | |||||
5442 | // Add in the Base Offset if present. | ||||
5443 | if (AddrMode.BaseOffs) { | ||||
5444 | Value *V = ConstantInt::get(IntPtrTy, AddrMode.BaseOffs); | ||||
5445 | if (Result) | ||||
5446 | Result = Builder.CreateAdd(Result, V, "sunkaddr"); | ||||
5447 | else | ||||
5448 | Result = V; | ||||
5449 | } | ||||
5450 | |||||
5451 | if (!Result) | ||||
5452 | SunkAddr = Constant::getNullValue(Addr->getType()); | ||||
5453 | else | ||||
5454 | SunkAddr = Builder.CreateIntToPtr(Result, Addr->getType(), "sunkaddr"); | ||||
5455 | } | ||||
5456 | |||||
5457 | MemoryInst->replaceUsesOfWith(Repl, SunkAddr); | ||||
5458 | // Store the newly computed address into the cache. In the case we reused a | ||||
5459 | // value, this should be idempotent. | ||||
5460 | SunkAddrs[Addr] = WeakTrackingVH(SunkAddr); | ||||
5461 | |||||
5462 | // If we have no uses, recursively delete the value and all dead instructions | ||||
5463 | // using it. | ||||
5464 | if (Repl->use_empty()) { | ||||
5465 | resetIteratorIfInvalidatedWhileCalling(CurInstIterator->getParent(), [&]() { | ||||
5466 | RecursivelyDeleteTriviallyDeadInstructions( | ||||
5467 | Repl, TLInfo, nullptr, | ||||
5468 | [&](Value *V) { removeAllAssertingVHReferences(V); }); | ||||
5469 | }); | ||||
5470 | } | ||||
5471 | ++NumMemoryInsts; | ||||
5472 | return true; | ||||
5473 | } | ||||
5474 | |||||
5475 | /// Rewrite GEP input to gather/scatter to enable SelectionDAGBuilder to find | ||||
5476 | /// a uniform base to use for ISD::MGATHER/MSCATTER. SelectionDAGBuilder can | ||||
5477 | /// only handle a 2 operand GEP in the same basic block or a splat constant | ||||
5478 | /// vector. The 2 operands to the GEP must have a scalar pointer and a vector | ||||
5479 | /// index. | ||||
5480 | /// | ||||
5481 | /// If the existing GEP has a vector base pointer that is splat, we can look | ||||
5482 | /// through the splat to find the scalar pointer. If we can't find a scalar | ||||
5483 | /// pointer there's nothing we can do. | ||||
5484 | /// | ||||
5485 | /// If we have a GEP with more than 2 indices where the middle indices are all | ||||
5486 | /// zeroes, we can replace it with 2 GEPs where the second has 2 operands. | ||||
5487 | /// | ||||
5488 | /// If the final index isn't a vector or is a splat, we can emit a scalar GEP | ||||
5489 | /// followed by a GEP with an all zeroes vector index. This will enable | ||||
5490 | /// SelectionDAGBuilder to use the scalar GEP as the uniform base and have a | ||||
5491 | /// zero index. | ||||
5492 | bool CodeGenPrepare::optimizeGatherScatterInst(Instruction *MemoryInst, | ||||
5493 | Value *Ptr) { | ||||
5494 | Value *NewAddr; | ||||
5495 | |||||
5496 | if (const auto *GEP = dyn_cast<GetElementPtrInst>(Ptr)) { | ||||
5497 | // Don't optimize GEPs that don't have indices. | ||||
5498 | if (!GEP->hasIndices()) | ||||
5499 | return false; | ||||
5500 | |||||
5501 | // If the GEP and the gather/scatter aren't in the same BB, don't optimize. | ||||
5502 | // FIXME: We should support this by sinking the GEP. | ||||
5503 | if (MemoryInst->getParent() != GEP->getParent()) | ||||
5504 | return false; | ||||
5505 | |||||
5506 | SmallVector<Value *, 2> Ops(GEP->operands()); | ||||
5507 | |||||
5508 | bool RewriteGEP = false; | ||||
5509 | |||||
5510 | if (Ops[0]->getType()->isVectorTy()) { | ||||
5511 | Ops[0] = getSplatValue(Ops[0]); | ||||
5512 | if (!Ops[0]) | ||||
5513 | return false; | ||||
5514 | RewriteGEP = true; | ||||
5515 | } | ||||
5516 | |||||
5517 | unsigned FinalIndex = Ops.size() - 1; | ||||
5518 | |||||
5519 | // Ensure all but the last index is 0. | ||||
5520 | // FIXME: This isn't strictly required. All that's required is that they are | ||||
5521 | // all scalars or splats. | ||||
5522 | for (unsigned i = 1; i < FinalIndex; ++i) { | ||||
5523 | auto *C = dyn_cast<Constant>(Ops[i]); | ||||
5524 | if (!C) | ||||
5525 | return false; | ||||
5526 | if (isa<VectorType>(C->getType())) | ||||
5527 | C = C->getSplatValue(); | ||||
5528 | auto *CI = dyn_cast_or_null<ConstantInt>(C); | ||||
5529 | if (!CI || !CI->isZero()) | ||||
5530 | return false; | ||||
5531 | // Scalarize the index if needed. | ||||
5532 | Ops[i] = CI; | ||||
5533 | } | ||||
5534 | |||||
5535 | // Try to scalarize the final index. | ||||
5536 | if (Ops[FinalIndex]->getType()->isVectorTy()) { | ||||
5537 | if (Value *V = getSplatValue(Ops[FinalIndex])) { | ||||
5538 | auto *C = dyn_cast<ConstantInt>(V); | ||||
5539 | // Don't scalarize all zeros vector. | ||||
5540 | if (!C || !C->isZero()) { | ||||
5541 | Ops[FinalIndex] = V; | ||||
5542 | RewriteGEP = true; | ||||
5543 | } | ||||
5544 | } | ||||
5545 | } | ||||
5546 | |||||
5547 | // If we made any changes or the we have extra operands, we need to generate | ||||
5548 | // new instructions. | ||||
5549 | if (!RewriteGEP && Ops.size() == 2) | ||||
5550 | return false; | ||||
5551 | |||||
5552 | auto NumElts = cast<VectorType>(Ptr->getType())->getElementCount(); | ||||
5553 | |||||
5554 | IRBuilder<> Builder(MemoryInst); | ||||
5555 | |||||
5556 | Type *SourceTy = GEP->getSourceElementType(); | ||||
5557 | Type *ScalarIndexTy = DL->getIndexType(Ops[0]->getType()->getScalarType()); | ||||
5558 | |||||
5559 | // If the final index isn't a vector, emit a scalar GEP containing all ops | ||||
5560 | // and a vector GEP with all zeroes final index. | ||||
5561 | if (!Ops[FinalIndex]->getType()->isVectorTy()) { | ||||
5562 | NewAddr = Builder.CreateGEP(SourceTy, Ops[0], | ||||
5563 | makeArrayRef(Ops).drop_front()); | ||||
5564 | auto *IndexTy = VectorType::get(ScalarIndexTy, NumElts); | ||||
5565 | auto *SecondTy = GetElementPtrInst::getIndexedType( | ||||
5566 | SourceTy, makeArrayRef(Ops).drop_front()); | ||||
5567 | NewAddr = | ||||
5568 | Builder.CreateGEP(SecondTy, NewAddr, Constant::getNullValue(IndexTy)); | ||||
5569 | } else { | ||||
5570 | Value *Base = Ops[0]; | ||||
5571 | Value *Index = Ops[FinalIndex]; | ||||
5572 | |||||
5573 | // Create a scalar GEP if there are more than 2 operands. | ||||
5574 | if (Ops.size() != 2) { | ||||
5575 | // Replace the last index with 0. | ||||
5576 | Ops[FinalIndex] = Constant::getNullValue(ScalarIndexTy); | ||||
5577 | Base = Builder.CreateGEP(SourceTy, Base, | ||||
5578 | makeArrayRef(Ops).drop_front()); | ||||
5579 | SourceTy = GetElementPtrInst::getIndexedType( | ||||
5580 | SourceTy, makeArrayRef(Ops).drop_front()); | ||||
5581 | } | ||||
5582 | |||||
5583 | // Now create the GEP with scalar pointer and vector index. | ||||
5584 | NewAddr = Builder.CreateGEP(SourceTy, Base, Index); | ||||
5585 | } | ||||
5586 | } else if (!isa<Constant>(Ptr)) { | ||||
5587 | // Not a GEP, maybe its a splat and we can create a GEP to enable | ||||
5588 | // SelectionDAGBuilder to use it as a uniform base. | ||||
5589 | Value *V = getSplatValue(Ptr); | ||||
5590 | if (!V) | ||||
5591 | return false; | ||||
5592 | |||||
5593 | auto NumElts = cast<VectorType>(Ptr->getType())->getElementCount(); | ||||
5594 | |||||
5595 | IRBuilder<> Builder(MemoryInst); | ||||
5596 | |||||
5597 | // Emit a vector GEP with a scalar pointer and all 0s vector index. | ||||
5598 | Type *ScalarIndexTy = DL->getIndexType(V->getType()->getScalarType()); | ||||
5599 | auto *IndexTy = VectorType::get(ScalarIndexTy, NumElts); | ||||
5600 | Type *ScalarTy; | ||||
5601 | if (cast<IntrinsicInst>(MemoryInst)->getIntrinsicID() == | ||||
5602 | Intrinsic::masked_gather) { | ||||
5603 | ScalarTy = MemoryInst->getType()->getScalarType(); | ||||
5604 | } else { | ||||
5605 | assert(cast<IntrinsicInst>(MemoryInst)->getIntrinsicID() ==((void)0) | ||||
5606 | Intrinsic::masked_scatter)((void)0); | ||||
5607 | ScalarTy = MemoryInst->getOperand(0)->getType()->getScalarType(); | ||||
5608 | } | ||||
5609 | NewAddr = Builder.CreateGEP(ScalarTy, V, Constant::getNullValue(IndexTy)); | ||||
5610 | } else { | ||||
5611 | // Constant, SelectionDAGBuilder knows to check if its a splat. | ||||
5612 | return false; | ||||
5613 | } | ||||
5614 | |||||
5615 | MemoryInst->replaceUsesOfWith(Ptr, NewAddr); | ||||
5616 | |||||
5617 | // If we have no uses, recursively delete the value and all dead instructions | ||||
5618 | // using it. | ||||
5619 | if (Ptr->use_empty()) | ||||
5620 | RecursivelyDeleteTriviallyDeadInstructions( | ||||
5621 | Ptr, TLInfo, nullptr, | ||||
5622 | [&](Value *V) { removeAllAssertingVHReferences(V); }); | ||||
5623 | |||||
5624 | return true; | ||||
5625 | } | ||||
5626 | |||||
5627 | /// If there are any memory operands, use OptimizeMemoryInst to sink their | ||||
5628 | /// address computing into the block when possible / profitable. | ||||
5629 | bool CodeGenPrepare::optimizeInlineAsmInst(CallInst *CS) { | ||||
5630 | bool MadeChange = false; | ||||
5631 | |||||
5632 | const TargetRegisterInfo *TRI = | ||||
5633 | TM->getSubtargetImpl(*CS->getFunction())->getRegisterInfo(); | ||||
5634 | TargetLowering::AsmOperandInfoVector TargetConstraints = | ||||
5635 | TLI->ParseConstraints(*DL, TRI, *CS); | ||||
5636 | unsigned ArgNo = 0; | ||||
5637 | for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) { | ||||
5638 | TargetLowering::AsmOperandInfo &OpInfo = TargetConstraints[i]; | ||||
5639 | |||||
5640 | // Compute the constraint code and ConstraintType to use. | ||||
5641 | TLI->ComputeConstraintToUse(OpInfo, SDValue()); | ||||
5642 | |||||
5643 | if (OpInfo.ConstraintType == TargetLowering::C_Memory && | ||||
5644 | OpInfo.isIndirect) { | ||||
5645 | Value *OpVal = CS->getArgOperand(ArgNo++); | ||||
5646 | MadeChange |= optimizeMemoryInst(CS, OpVal, OpVal->getType(), ~0u); | ||||
5647 | } else if (OpInfo.Type == InlineAsm::isInput) | ||||
5648 | ArgNo++; | ||||
5649 | } | ||||
5650 | |||||
5651 | return MadeChange; | ||||
5652 | } | ||||
5653 | |||||
5654 | /// Check if all the uses of \p Val are equivalent (or free) zero or | ||||
5655 | /// sign extensions. | ||||
5656 | static bool hasSameExtUse(Value *Val, const TargetLowering &TLI) { | ||||
5657 | assert(!Val->use_empty() && "Input must have at least one use")((void)0); | ||||
5658 | const Instruction *FirstUser = cast<Instruction>(*Val->user_begin()); | ||||
5659 | bool IsSExt = isa<SExtInst>(FirstUser); | ||||
5660 | Type *ExtTy = FirstUser->getType(); | ||||
5661 | for (const User *U : Val->users()) { | ||||
5662 | const Instruction *UI = cast<Instruction>(U); | ||||
5663 | if ((IsSExt && !isa<SExtInst>(UI)) || (!IsSExt && !isa<ZExtInst>(UI))) | ||||
5664 | return false; | ||||
5665 | Type *CurTy = UI->getType(); | ||||
5666 | // Same input and output types: Same instruction after CSE. | ||||
5667 | if (CurTy == ExtTy) | ||||
5668 | continue; | ||||
5669 | |||||
5670 | // If IsSExt is true, we are in this situation: | ||||
5671 | // a = Val | ||||
5672 | // b = sext ty1 a to ty2 | ||||
5673 | // c = sext ty1 a to ty3 | ||||
5674 | // Assuming ty2 is shorter than ty3, this could be turned into: | ||||
5675 | // a = Val | ||||
5676 | // b = sext ty1 a to ty2 | ||||
5677 | // c = sext ty2 b to ty3 | ||||
5678 | // However, the last sext is not free. | ||||
5679 | if (IsSExt) | ||||
5680 | return false; | ||||
5681 | |||||
5682 | // This is a ZExt, maybe this is free to extend from one type to another. | ||||
5683 | // In that case, we would not account for a different use. | ||||
5684 | Type *NarrowTy; | ||||
5685 | Type *LargeTy; | ||||
5686 | if (ExtTy->getScalarType()->getIntegerBitWidth() > | ||||
5687 | CurTy->getScalarType()->getIntegerBitWidth()) { | ||||
5688 | NarrowTy = CurTy; | ||||
5689 | LargeTy = ExtTy; | ||||
5690 | } else { | ||||
5691 | NarrowTy = ExtTy; | ||||
5692 | LargeTy = CurTy; | ||||
5693 | } | ||||
5694 | |||||
5695 | if (!TLI.isZExtFree(NarrowTy, LargeTy)) | ||||
5696 | return false; | ||||
5697 | } | ||||
5698 | // All uses are the same or can be derived from one another for free. | ||||
5699 | return true; | ||||
5700 | } | ||||
5701 | |||||
5702 | /// Try to speculatively promote extensions in \p Exts and continue | ||||
5703 | /// promoting through newly promoted operands recursively as far as doing so is | ||||
5704 | /// profitable. Save extensions profitably moved up, in \p ProfitablyMovedExts. | ||||
5705 | /// When some promotion happened, \p TPT contains the proper state to revert | ||||
5706 | /// them. | ||||
5707 | /// | ||||
5708 | /// \return true if some promotion happened, false otherwise. | ||||
5709 | bool CodeGenPrepare::tryToPromoteExts( | ||||
5710 | TypePromotionTransaction &TPT, const SmallVectorImpl<Instruction *> &Exts, | ||||
5711 | SmallVectorImpl<Instruction *> &ProfitablyMovedExts, | ||||
5712 | unsigned CreatedInstsCost) { | ||||
5713 | bool Promoted = false; | ||||
5714 | |||||
5715 | // Iterate over all the extensions to try to promote them. | ||||
5716 | for (auto *I : Exts) { | ||||
5717 | // Early check if we directly have ext(load). | ||||
5718 | if (isa<LoadInst>(I->getOperand(0))) { | ||||
5719 | ProfitablyMovedExts.push_back(I); | ||||
5720 | continue; | ||||
5721 | } | ||||
5722 | |||||
5723 | // Check whether or not we want to do any promotion. The reason we have | ||||
5724 | // this check inside the for loop is to catch the case where an extension | ||||
5725 | // is directly fed by a load because in such case the extension can be moved | ||||
5726 | // up without any promotion on its operands. | ||||
5727 | if (!TLI->enableExtLdPromotion() || DisableExtLdPromotion) | ||||
5728 | return false; | ||||
5729 | |||||
5730 | // Get the action to perform the promotion. | ||||
5731 | TypePromotionHelper::Action TPH = | ||||
5732 | TypePromotionHelper::getAction(I, InsertedInsts, *TLI, PromotedInsts); | ||||
5733 | // Check if we can promote. | ||||
5734 | if (!TPH) { | ||||
5735 | // Save the current extension as we cannot move up through its operand. | ||||
5736 | ProfitablyMovedExts.push_back(I); | ||||
5737 | continue; | ||||
5738 | } | ||||
5739 | |||||
5740 | // Save the current state. | ||||
5741 | TypePromotionTransaction::ConstRestorationPt LastKnownGood = | ||||
5742 | TPT.getRestorationPoint(); | ||||
5743 | SmallVector<Instruction *, 4> NewExts; | ||||
5744 | unsigned NewCreatedInstsCost = 0; | ||||
5745 | unsigned ExtCost = !TLI->isExtFree(I); | ||||
5746 | // Promote. | ||||
5747 | Value *PromotedVal = TPH(I, TPT, PromotedInsts, NewCreatedInstsCost, | ||||
5748 | &NewExts, nullptr, *TLI); | ||||
5749 | assert(PromotedVal &&((void)0) | ||||
5750 | "TypePromotionHelper should have filtered out those cases")((void)0); | ||||
5751 | |||||
5752 | // We would be able to merge only one extension in a load. | ||||
5753 | // Therefore, if we have more than 1 new extension we heuristically | ||||
5754 | // cut this search path, because it means we degrade the code quality. | ||||
5755 | // With exactly 2, the transformation is neutral, because we will merge | ||||
5756 | // one extension but leave one. However, we optimistically keep going, | ||||
5757 | // because the new extension may be removed too. | ||||
5758 | long long TotalCreatedInstsCost = CreatedInstsCost + NewCreatedInstsCost; | ||||
5759 | // FIXME: It would be possible to propagate a negative value instead of | ||||
5760 | // conservatively ceiling it to 0. | ||||
5761 | TotalCreatedInstsCost = | ||||
5762 | std::max((long long)0, (TotalCreatedInstsCost - ExtCost)); | ||||
5763 | if (!StressExtLdPromotion && | ||||
5764 | (TotalCreatedInstsCost > 1 || | ||||
5765 | !isPromotedInstructionLegal(*TLI, *DL, PromotedVal))) { | ||||
5766 | // This promotion is not profitable, rollback to the previous state, and | ||||
5767 | // save the current extension in ProfitablyMovedExts as the latest | ||||
5768 | // speculative promotion turned out to be unprofitable. | ||||
5769 | TPT.rollback(LastKnownGood); | ||||
5770 | ProfitablyMovedExts.push_back(I); | ||||
5771 | continue; | ||||
5772 | } | ||||
5773 | // Continue promoting NewExts as far as doing so is profitable. | ||||
5774 | SmallVector<Instruction *, 2> NewlyMovedExts; | ||||
5775 | (void)tryToPromoteExts(TPT, NewExts, NewlyMovedExts, TotalCreatedInstsCost); | ||||
5776 | bool NewPromoted = false; | ||||
5777 | for (auto *ExtInst : NewlyMovedExts) { | ||||
5778 | Instruction *MovedExt = cast<Instruction>(ExtInst); | ||||
5779 | Value *ExtOperand = MovedExt->getOperand(0); | ||||
5780 | // If we have reached to a load, we need this extra profitability check | ||||
5781 | // as it could potentially be merged into an ext(load). | ||||
5782 | if (isa<LoadInst>(ExtOperand) && | ||||
5783 | !(StressExtLdPromotion || NewCreatedInstsCost <= ExtCost || | ||||
5784 | (ExtOperand->hasOneUse() || hasSameExtUse(ExtOperand, *TLI)))) | ||||
5785 | continue; | ||||
5786 | |||||
5787 | ProfitablyMovedExts.push_back(MovedExt); | ||||
5788 | NewPromoted = true; | ||||
5789 | } | ||||
5790 | |||||
5791 | // If none of speculative promotions for NewExts is profitable, rollback | ||||
5792 | // and save the current extension (I) as the last profitable extension. | ||||
5793 | if (!NewPromoted) { | ||||
5794 | TPT.rollback(LastKnownGood); | ||||
5795 | ProfitablyMovedExts.push_back(I); | ||||
5796 | continue; | ||||
5797 | } | ||||
5798 | // The promotion is profitable. | ||||
5799 | Promoted = true; | ||||
5800 | } | ||||
5801 | return Promoted; | ||||
5802 | } | ||||
5803 | |||||
5804 | /// Merging redundant sexts when one is dominating the other. | ||||
5805 | bool CodeGenPrepare::mergeSExts(Function &F) { | ||||
5806 | bool Changed = false; | ||||
5807 | for (auto &Entry : ValToSExtendedUses) { | ||||
5808 | SExts &Insts = Entry.second; | ||||
5809 | SExts CurPts; | ||||
5810 | for (Instruction *Inst : Insts) { | ||||
5811 | if (RemovedInsts.count(Inst) || !isa<SExtInst>(Inst) || | ||||
5812 | Inst->getOperand(0) != Entry.first) | ||||
5813 | continue; | ||||
5814 | bool inserted = false; | ||||
5815 | for (auto &Pt : CurPts) { | ||||
5816 | if (getDT(F).dominates(Inst, Pt)) { | ||||
5817 | Pt->replaceAllUsesWith(Inst); | ||||
5818 | RemovedInsts.insert(Pt); | ||||
5819 | Pt->removeFromParent(); | ||||
5820 | Pt = Inst; | ||||
5821 | inserted = true; | ||||
5822 | Changed = true; | ||||
5823 | break; | ||||
5824 | } | ||||
5825 | if (!getDT(F).dominates(Pt, Inst)) | ||||
5826 | // Give up if we need to merge in a common dominator as the | ||||
5827 | // experiments show it is not profitable. | ||||
5828 | continue; | ||||
5829 | Inst->replaceAllUsesWith(Pt); | ||||
5830 | RemovedInsts.insert(Inst); | ||||
5831 | Inst->removeFromParent(); | ||||
5832 | inserted = true; | ||||
5833 | Changed = true; | ||||
5834 | break; | ||||
5835 | } | ||||
5836 | if (!inserted) | ||||
5837 | CurPts.push_back(Inst); | ||||
5838 | } | ||||
5839 | } | ||||
5840 | return Changed; | ||||
5841 | } | ||||
5842 | |||||
5843 | // Splitting large data structures so that the GEPs accessing them can have | ||||
5844 | // smaller offsets so that they can be sunk to the same blocks as their users. | ||||
5845 | // For example, a large struct starting from %base is split into two parts | ||||
5846 | // where the second part starts from %new_base. | ||||
5847 | // | ||||
5848 | // Before: | ||||
5849 | // BB0: | ||||
5850 | // %base = | ||||
5851 | // | ||||
5852 | // BB1: | ||||
5853 | // %gep0 = gep %base, off0 | ||||
5854 | // %gep1 = gep %base, off1 | ||||
5855 | // %gep2 = gep %base, off2 | ||||
5856 | // | ||||
5857 | // BB2: | ||||
5858 | // %load1 = load %gep0 | ||||
5859 | // %load2 = load %gep1 | ||||
5860 | // %load3 = load %gep2 | ||||
5861 | // | ||||
5862 | // After: | ||||
5863 | // BB0: | ||||
5864 | // %base = | ||||
5865 | // %new_base = gep %base, off0 | ||||
5866 | // | ||||
5867 | // BB1: | ||||
5868 | // %new_gep0 = %new_base | ||||
5869 | // %new_gep1 = gep %new_base, off1 - off0 | ||||
5870 | // %new_gep2 = gep %new_base, off2 - off0 | ||||
5871 | // | ||||
5872 | // BB2: | ||||
5873 | // %load1 = load i32, i32* %new_gep0 | ||||
5874 | // %load2 = load i32, i32* %new_gep1 | ||||
5875 | // %load3 = load i32, i32* %new_gep2 | ||||
5876 | // | ||||
5877 | // %new_gep1 and %new_gep2 can be sunk to BB2 now after the splitting because | ||||
5878 | // their offsets are smaller enough to fit into the addressing mode. | ||||
5879 | bool CodeGenPrepare::splitLargeGEPOffsets() { | ||||
5880 | bool Changed = false; | ||||
5881 | for (auto &Entry : LargeOffsetGEPMap) { | ||||
5882 | Value *OldBase = Entry.first; | ||||
5883 | SmallVectorImpl<std::pair<AssertingVH<GetElementPtrInst>, int64_t>> | ||||
5884 | &LargeOffsetGEPs = Entry.second; | ||||
5885 | auto compareGEPOffset = | ||||
5886 | [&](const std::pair<GetElementPtrInst *, int64_t> &LHS, | ||||
5887 | const std::pair<GetElementPtrInst *, int64_t> &RHS) { | ||||
5888 | if (LHS.first == RHS.first) | ||||
5889 | return false; | ||||
5890 | if (LHS.second != RHS.second) | ||||
5891 | return LHS.second < RHS.second; | ||||
5892 | return LargeOffsetGEPID[LHS.first] < LargeOffsetGEPID[RHS.first]; | ||||
5893 | }; | ||||
5894 | // Sorting all the GEPs of the same data structures based on the offsets. | ||||
5895 | llvm::sort(LargeOffsetGEPs, compareGEPOffset); | ||||
5896 | LargeOffsetGEPs.erase( | ||||
5897 | std::unique(LargeOffsetGEPs.begin(), LargeOffsetGEPs.end()), | ||||
5898 | LargeOffsetGEPs.end()); | ||||
5899 | // Skip if all the GEPs have the same offsets. | ||||
5900 | if (LargeOffsetGEPs.front().second == LargeOffsetGEPs.back().second) | ||||
5901 | continue; | ||||
5902 | GetElementPtrInst *BaseGEP = LargeOffsetGEPs.begin()->first; | ||||
5903 | int64_t BaseOffset = LargeOffsetGEPs.begin()->second; | ||||
5904 | Value *NewBaseGEP = nullptr; | ||||
5905 | |||||
5906 | auto *LargeOffsetGEP = LargeOffsetGEPs.begin(); | ||||
5907 | while (LargeOffsetGEP != LargeOffsetGEPs.end()) { | ||||
5908 | GetElementPtrInst *GEP = LargeOffsetGEP->first; | ||||
5909 | int64_t Offset = LargeOffsetGEP->second; | ||||
5910 | if (Offset != BaseOffset) { | ||||
5911 | TargetLowering::AddrMode AddrMode; | ||||
5912 | AddrMode.BaseOffs = Offset - BaseOffset; | ||||
5913 | // The result type of the GEP might not be the type of the memory | ||||
5914 | // access. | ||||
5915 | if (!TLI->isLegalAddressingMode(*DL, AddrMode, | ||||
5916 | GEP->getResultElementType(), | ||||
5917 | GEP->getAddressSpace())) { | ||||
5918 | // We need to create a new base if the offset to the current base is | ||||
5919 | // too large to fit into the addressing mode. So, a very large struct | ||||
5920 | // may be split into several parts. | ||||
5921 | BaseGEP = GEP; | ||||
5922 | BaseOffset = Offset; | ||||
5923 | NewBaseGEP = nullptr; | ||||
5924 | } | ||||
5925 | } | ||||
5926 | |||||
5927 | // Generate a new GEP to replace the current one. | ||||
5928 | LLVMContext &Ctx = GEP->getContext(); | ||||
5929 | Type *IntPtrTy = DL->getIntPtrType(GEP->getType()); | ||||
5930 | Type *I8PtrTy = | ||||
5931 | Type::getInt8PtrTy(Ctx, GEP->getType()->getPointerAddressSpace()); | ||||
5932 | Type *I8Ty = Type::getInt8Ty(Ctx); | ||||
5933 | |||||
5934 | if (!NewBaseGEP) { | ||||
5935 | // Create a new base if we don't have one yet. Find the insertion | ||||
5936 | // pointer for the new base first. | ||||
5937 | BasicBlock::iterator NewBaseInsertPt; | ||||
5938 | BasicBlock *NewBaseInsertBB; | ||||
5939 | if (auto *BaseI = dyn_cast<Instruction>(OldBase)) { | ||||
5940 | // If the base of the struct is an instruction, the new base will be | ||||
5941 | // inserted close to it. | ||||
5942 | NewBaseInsertBB = BaseI->getParent(); | ||||
5943 | if (isa<PHINode>(BaseI)) | ||||
5944 | NewBaseInsertPt = NewBaseInsertBB->getFirstInsertionPt(); | ||||
5945 | else if (InvokeInst *Invoke = dyn_cast<InvokeInst>(BaseI)) { | ||||
5946 | NewBaseInsertBB = | ||||
5947 | SplitEdge(NewBaseInsertBB, Invoke->getNormalDest()); | ||||
5948 | NewBaseInsertPt = NewBaseInsertBB->getFirstInsertionPt(); | ||||
5949 | } else | ||||
5950 | NewBaseInsertPt = std::next(BaseI->getIterator()); | ||||
5951 | } else { | ||||
5952 | // If the current base is an argument or global value, the new base | ||||
5953 | // will be inserted to the entry block. | ||||
5954 | NewBaseInsertBB = &BaseGEP->getFunction()->getEntryBlock(); | ||||
5955 | NewBaseInsertPt = NewBaseInsertBB->getFirstInsertionPt(); | ||||
5956 | } | ||||
5957 | IRBuilder<> NewBaseBuilder(NewBaseInsertBB, NewBaseInsertPt); | ||||
5958 | // Create a new base. | ||||
5959 | Value *BaseIndex = ConstantInt::get(IntPtrTy, BaseOffset); | ||||
5960 | NewBaseGEP = OldBase; | ||||
5961 | if (NewBaseGEP->getType() != I8PtrTy) | ||||
5962 | NewBaseGEP = NewBaseBuilder.CreatePointerCast(NewBaseGEP, I8PtrTy); | ||||
5963 | NewBaseGEP = | ||||
5964 | NewBaseBuilder.CreateGEP(I8Ty, NewBaseGEP, BaseIndex, "splitgep"); | ||||
5965 | NewGEPBases.insert(NewBaseGEP); | ||||
5966 | } | ||||
5967 | |||||
5968 | IRBuilder<> Builder(GEP); | ||||
5969 | Value *NewGEP = NewBaseGEP; | ||||
5970 | if (Offset == BaseOffset) { | ||||
5971 | if (GEP->getType() != I8PtrTy) | ||||
5972 | NewGEP = Builder.CreatePointerCast(NewGEP, GEP->getType()); | ||||
5973 | } else { | ||||
5974 | // Calculate the new offset for the new GEP. | ||||
5975 | Value *Index = ConstantInt::get(IntPtrTy, Offset - BaseOffset); | ||||
5976 | NewGEP = Builder.CreateGEP(I8Ty, NewBaseGEP, Index); | ||||
5977 | |||||
5978 | if (GEP->getType() != I8PtrTy) | ||||
5979 | NewGEP = Builder.CreatePointerCast(NewGEP, GEP->getType()); | ||||
5980 | } | ||||
5981 | GEP->replaceAllUsesWith(NewGEP); | ||||
5982 | LargeOffsetGEPID.erase(GEP); | ||||
5983 | LargeOffsetGEP = LargeOffsetGEPs.erase(LargeOffsetGEP); | ||||
5984 | GEP->eraseFromParent(); | ||||
5985 | Changed = true; | ||||
5986 | } | ||||
5987 | } | ||||
5988 | return Changed; | ||||
5989 | } | ||||
5990 | |||||
5991 | bool CodeGenPrepare::optimizePhiType( | ||||
5992 | PHINode *I, SmallPtrSetImpl<PHINode *> &Visited, | ||||
5993 | SmallPtrSetImpl<Instruction *> &DeletedInstrs) { | ||||
5994 | // We are looking for a collection on interconnected phi nodes that together | ||||
5995 | // only use loads/bitcasts and are used by stores/bitcasts, and the bitcasts | ||||
5996 | // are of the same type. Convert the whole set of nodes to the type of the | ||||
5997 | // bitcast. | ||||
5998 | Type *PhiTy = I->getType(); | ||||
5999 | Type *ConvertTy = nullptr; | ||||
6000 | if (Visited.count(I) || | ||||
6001 | (!I->getType()->isIntegerTy() && !I->getType()->isFloatingPointTy())) | ||||
6002 | return false; | ||||
6003 | |||||
6004 | SmallVector<Instruction *, 4> Worklist; | ||||
6005 | Worklist.push_back(cast<Instruction>(I)); | ||||
6006 | SmallPtrSet<PHINode *, 4> PhiNodes; | ||||
6007 | PhiNodes.insert(I); | ||||
6008 | Visited.insert(I); | ||||
6009 | SmallPtrSet<Instruction *, 4> Defs; | ||||
6010 | SmallPtrSet<Instruction *, 4> Uses; | ||||
6011 | // This works by adding extra bitcasts between load/stores and removing | ||||
6012 | // existing bicasts. If we have a phi(bitcast(load)) or a store(bitcast(phi)) | ||||
6013 | // we can get in the situation where we remove a bitcast in one iteration | ||||
6014 | // just to add it again in the next. We need to ensure that at least one | ||||
6015 | // bitcast we remove are anchored to something that will not change back. | ||||
6016 | bool AnyAnchored = false; | ||||
6017 | |||||
6018 | while (!Worklist.empty()) { | ||||
6019 | Instruction *II = Worklist.pop_back_val(); | ||||
6020 | |||||
6021 | if (auto *Phi = dyn_cast<PHINode>(II)) { | ||||
6022 | // Handle Defs, which might also be PHI's | ||||
6023 | for (Value *V : Phi->incoming_values()) { | ||||
6024 | if (auto *OpPhi = dyn_cast<PHINode>(V)) { | ||||
6025 | if (!PhiNodes.count(OpPhi)) { | ||||
6026 | if (Visited.count(OpPhi)) | ||||
6027 | return false; | ||||
6028 | PhiNodes.insert(OpPhi); | ||||
6029 | Visited.insert(OpPhi); | ||||
6030 | Worklist.push_back(OpPhi); | ||||
6031 | } | ||||
6032 | } else if (auto *OpLoad = dyn_cast<LoadInst>(V)) { | ||||
6033 | if (!OpLoad->isSimple()) | ||||
6034 | return false; | ||||
6035 | if (!Defs.count(OpLoad)) { | ||||
6036 | Defs.insert(OpLoad); | ||||
6037 | Worklist.push_back(OpLoad); | ||||
6038 | } | ||||
6039 | } else if (auto *OpEx = dyn_cast<ExtractElementInst>(V)) { | ||||
6040 | if (!Defs.count(OpEx)) { | ||||
6041 | Defs.insert(OpEx); | ||||
6042 | Worklist.push_back(OpEx); | ||||
6043 | } | ||||
6044 | } else if (auto *OpBC = dyn_cast<BitCastInst>(V)) { | ||||
6045 | if (!ConvertTy) | ||||
6046 | ConvertTy = OpBC->getOperand(0)->getType(); | ||||
6047 | if (OpBC->getOperand(0)->getType() != ConvertTy) | ||||
6048 | return false; | ||||
6049 | if (!Defs.count(OpBC)) { | ||||
6050 | Defs.insert(OpBC); | ||||
6051 | Worklist.push_back(OpBC); | ||||
6052 | AnyAnchored |= !isa<LoadInst>(OpBC->getOperand(0)) && | ||||
6053 | !isa<ExtractElementInst>(OpBC->getOperand(0)); | ||||
6054 | } | ||||
6055 | } else if (!isa<UndefValue>(V)) { | ||||
6056 | return false; | ||||
6057 | } | ||||
6058 | } | ||||
6059 | } | ||||
6060 | |||||
6061 | // Handle uses which might also be phi's | ||||
6062 | for (User *V : II->users()) { | ||||
6063 | if (auto *OpPhi = dyn_cast<PHINode>(V)) { | ||||
6064 | if (!PhiNodes.count(OpPhi)) { | ||||
6065 | if (Visited.count(OpPhi)) | ||||
6066 | return false; | ||||
6067 | PhiNodes.insert(OpPhi); | ||||
6068 | Visited.insert(OpPhi); | ||||
6069 | Worklist.push_back(OpPhi); | ||||
6070 | } | ||||
6071 | } else if (auto *OpStore = dyn_cast<StoreInst>(V)) { | ||||
6072 | if (!OpStore->isSimple() || OpStore->getOperand(0) != II) | ||||
6073 | return false; | ||||
6074 | Uses.insert(OpStore); | ||||
6075 | } else if (auto *OpBC = dyn_cast<BitCastInst>(V)) { | ||||
6076 | if (!ConvertTy) | ||||
6077 | ConvertTy = OpBC->getType(); | ||||
6078 | if (OpBC->getType() != ConvertTy) | ||||
6079 | return false; | ||||
6080 | Uses.insert(OpBC); | ||||
6081 | AnyAnchored |= | ||||
6082 | any_of(OpBC->users(), [](User *U) { return !isa<StoreInst>(U); }); | ||||
6083 | } else { | ||||
6084 | return false; | ||||
6085 | } | ||||
6086 | } | ||||
6087 | } | ||||
6088 | |||||
6089 | if (!ConvertTy || !AnyAnchored || !TLI->shouldConvertPhiType(PhiTy, ConvertTy)) | ||||
6090 | return false; | ||||
6091 | |||||
6092 | LLVM_DEBUG(dbgs() << "Converting " << *I << "\n and connected nodes to "do { } while (false) | ||||
6093 | << *ConvertTy << "\n")do { } while (false); | ||||
6094 | |||||
6095 | // Create all the new phi nodes of the new type, and bitcast any loads to the | ||||
6096 | // correct type. | ||||
6097 | ValueToValueMap ValMap; | ||||
6098 | ValMap[UndefValue::get(PhiTy)] = UndefValue::get(ConvertTy); | ||||
6099 | for (Instruction *D : Defs) { | ||||
6100 | if (isa<BitCastInst>(D)) { | ||||
6101 | ValMap[D] = D->getOperand(0); | ||||
6102 | DeletedInstrs.insert(D); | ||||
6103 | } else { | ||||
6104 | ValMap[D] = | ||||
6105 | new BitCastInst(D, ConvertTy, D->getName() + ".bc", D->getNextNode()); | ||||
6106 | } | ||||
6107 | } | ||||
6108 | for (PHINode *Phi : PhiNodes) | ||||
6109 | ValMap[Phi] = PHINode::Create(ConvertTy, Phi->getNumIncomingValues(), | ||||
6110 | Phi->getName() + ".tc", Phi); | ||||
6111 | // Pipe together all the PhiNodes. | ||||
6112 | for (PHINode *Phi : PhiNodes) { | ||||
6113 | PHINode *NewPhi = cast<PHINode>(ValMap[Phi]); | ||||
6114 | for (int i = 0, e = Phi->getNumIncomingValues(); i < e; i++) | ||||
6115 | NewPhi->addIncoming(ValMap[Phi->getIncomingValue(i)], | ||||
6116 | Phi->getIncomingBlock(i)); | ||||
6117 | Visited.insert(NewPhi); | ||||
6118 | } | ||||
6119 | // And finally pipe up the stores and bitcasts | ||||
6120 | for (Instruction *U : Uses) { | ||||
6121 | if (isa<BitCastInst>(U)) { | ||||
6122 | DeletedInstrs.insert(U); | ||||
6123 | U->replaceAllUsesWith(ValMap[U->getOperand(0)]); | ||||
6124 | } else { | ||||
6125 | U->setOperand(0, | ||||
6126 | new BitCastInst(ValMap[U->getOperand(0)], PhiTy, "bc", U)); | ||||
6127 | } | ||||
6128 | } | ||||
6129 | |||||
6130 | // Save the removed phis to be deleted later. | ||||
6131 | for (PHINode *Phi : PhiNodes) | ||||
6132 | DeletedInstrs.insert(Phi); | ||||
6133 | return true; | ||||
6134 | } | ||||
6135 | |||||
6136 | bool CodeGenPrepare::optimizePhiTypes(Function &F) { | ||||
6137 | if (!OptimizePhiTypes) | ||||
6138 | return false; | ||||
6139 | |||||
6140 | bool Changed = false; | ||||
6141 | SmallPtrSet<PHINode *, 4> Visited; | ||||
6142 | SmallPtrSet<Instruction *, 4> DeletedInstrs; | ||||
6143 | |||||
6144 | // Attempt to optimize all the phis in the functions to the correct type. | ||||
6145 | for (auto &BB : F) | ||||
6146 | for (auto &Phi : BB.phis()) | ||||
6147 | Changed |= optimizePhiType(&Phi, Visited, DeletedInstrs); | ||||
6148 | |||||
6149 | // Remove any old phi's that have been converted. | ||||
6150 | for (auto *I : DeletedInstrs) { | ||||
6151 | I->replaceAllUsesWith(UndefValue::get(I->getType())); | ||||
6152 | I->eraseFromParent(); | ||||
6153 | } | ||||
6154 | |||||
6155 | return Changed; | ||||
6156 | } | ||||
6157 | |||||
6158 | /// Return true, if an ext(load) can be formed from an extension in | ||||
6159 | /// \p MovedExts. | ||||
6160 | bool CodeGenPrepare::canFormExtLd( | ||||
6161 | const SmallVectorImpl<Instruction *> &MovedExts, LoadInst *&LI, | ||||
6162 | Instruction *&Inst, bool HasPromoted) { | ||||
6163 | for (auto *MovedExtInst : MovedExts) { | ||||
6164 | if (isa<LoadInst>(MovedExtInst->getOperand(0))) { | ||||
6165 | LI = cast<LoadInst>(MovedExtInst->getOperand(0)); | ||||
6166 | Inst = MovedExtInst; | ||||
6167 | break; | ||||
6168 | } | ||||
6169 | } | ||||
6170 | if (!LI) | ||||
6171 | return false; | ||||
6172 | |||||
6173 | // If they're already in the same block, there's nothing to do. | ||||
6174 | // Make the cheap checks first if we did not promote. | ||||
6175 | // If we promoted, we need to check if it is indeed profitable. | ||||
6176 | if (!HasPromoted && LI->getParent() == Inst->getParent()) | ||||
6177 | return false; | ||||
6178 | |||||
6179 | return TLI->isExtLoad(LI, Inst, *DL); | ||||
6180 | } | ||||
6181 | |||||
6182 | /// Move a zext or sext fed by a load into the same basic block as the load, | ||||
6183 | /// unless conditions are unfavorable. This allows SelectionDAG to fold the | ||||
6184 | /// extend into the load. | ||||
6185 | /// | ||||
6186 | /// E.g., | ||||
6187 | /// \code | ||||
6188 | /// %ld = load i32* %addr | ||||
6189 | /// %add = add nuw i32 %ld, 4 | ||||
6190 | /// %zext = zext i32 %add to i64 | ||||
6191 | // \endcode | ||||
6192 | /// => | ||||
6193 | /// \code | ||||
6194 | /// %ld = load i32* %addr | ||||
6195 | /// %zext = zext i32 %ld to i64 | ||||
6196 | /// %add = add nuw i64 %zext, 4 | ||||
6197 | /// \encode | ||||
6198 | /// Note that the promotion in %add to i64 is done in tryToPromoteExts(), which | ||||
6199 | /// allow us to match zext(load i32*) to i64. | ||||
6200 | /// | ||||
6201 | /// Also, try to promote the computations used to obtain a sign extended | ||||
6202 | /// value used into memory accesses. | ||||
6203 | /// E.g., | ||||
6204 | /// \code | ||||
6205 | /// a = add nsw i32 b, 3 | ||||
6206 | /// d = sext i32 a to i64 | ||||
6207 | /// e = getelementptr ..., i64 d | ||||
6208 | /// \endcode | ||||
6209 | /// => | ||||
6210 | /// \code | ||||
6211 | /// f = sext i32 b to i64 | ||||
6212 | /// a = add nsw i64 f, 3 | ||||
6213 | /// e = getelementptr ..., i64 a | ||||
6214 | /// \endcode | ||||
6215 | /// | ||||
6216 | /// \p Inst[in/out] the extension may be modified during the process if some | ||||
6217 | /// promotions apply. | ||||
6218 | bool CodeGenPrepare::optimizeExt(Instruction *&Inst) { | ||||
6219 | bool AllowPromotionWithoutCommonHeader = false; | ||||
6220 | /// See if it is an interesting sext operations for the address type | ||||
6221 | /// promotion before trying to promote it, e.g., the ones with the right | ||||
6222 | /// type and used in memory accesses. | ||||
6223 | bool ATPConsiderable = TTI->shouldConsiderAddressTypePromotion( | ||||
6224 | *Inst, AllowPromotionWithoutCommonHeader); | ||||
6225 | TypePromotionTransaction TPT(RemovedInsts); | ||||
6226 | TypePromotionTransaction::ConstRestorationPt LastKnownGood = | ||||
6227 | TPT.getRestorationPoint(); | ||||
6228 | SmallVector<Instruction *, 1> Exts; | ||||
6229 | SmallVector<Instruction *, 2> SpeculativelyMovedExts; | ||||
6230 | Exts.push_back(Inst); | ||||
6231 | |||||
6232 | bool HasPromoted = tryToPromoteExts(TPT, Exts, SpeculativelyMovedExts); | ||||
6233 | |||||
6234 | // Look for a load being extended. | ||||
6235 | LoadInst *LI = nullptr; | ||||
6236 | Instruction *ExtFedByLoad; | ||||
6237 | |||||
6238 | // Try to promote a chain of computation if it allows to form an extended | ||||
6239 | // load. | ||||
6240 | if (canFormExtLd(SpeculativelyMovedExts, LI, ExtFedByLoad, HasPromoted)) { | ||||
6241 | assert(LI && ExtFedByLoad && "Expect a valid load and extension")((void)0); | ||||
6242 | TPT.commit(); | ||||
6243 | // Move the extend into the same block as the load. | ||||
6244 | ExtFedByLoad->moveAfter(LI); | ||||
6245 | ++NumExtsMoved; | ||||
6246 | Inst = ExtFedByLoad; | ||||
6247 | return true; | ||||
6248 | } | ||||
6249 | |||||
6250 | // Continue promoting SExts if known as considerable depending on targets. | ||||
6251 | if (ATPConsiderable && | ||||
6252 | performAddressTypePromotion(Inst, AllowPromotionWithoutCommonHeader, | ||||
6253 | HasPromoted, TPT, SpeculativelyMovedExts)) | ||||
6254 | return true; | ||||
6255 | |||||
6256 | TPT.rollback(LastKnownGood); | ||||
6257 | return false; | ||||
6258 | } | ||||
6259 | |||||
6260 | // Perform address type promotion if doing so is profitable. | ||||
6261 | // If AllowPromotionWithoutCommonHeader == false, we should find other sext | ||||
6262 | // instructions that sign extended the same initial value. However, if | ||||
6263 | // AllowPromotionWithoutCommonHeader == true, we expect promoting the | ||||
6264 | // extension is just profitable. | ||||
6265 | bool CodeGenPrepare::performAddressTypePromotion( | ||||
6266 | Instruction *&Inst, bool AllowPromotionWithoutCommonHeader, | ||||
6267 | bool HasPromoted, TypePromotionTransaction &TPT, | ||||
6268 | SmallVectorImpl<Instruction *> &SpeculativelyMovedExts) { | ||||
6269 | bool Promoted = false; | ||||
6270 | SmallPtrSet<Instruction *, 1> UnhandledExts; | ||||
6271 | bool AllSeenFirst = true; | ||||
6272 | for (auto *I : SpeculativelyMovedExts) { | ||||
6273 | Value *HeadOfChain = I->getOperand(0); | ||||
6274 | DenseMap<Value *, Instruction *>::iterator AlreadySeen = | ||||
6275 | SeenChainsForSExt.find(HeadOfChain); | ||||
6276 | // If there is an unhandled SExt which has the same header, try to promote | ||||
6277 | // it as well. | ||||
6278 | if (AlreadySeen != SeenChainsForSExt.end()) { | ||||
6279 | if (AlreadySeen->second != nullptr) | ||||
6280 | UnhandledExts.insert(AlreadySeen->second); | ||||
6281 | AllSeenFirst = false; | ||||
6282 | } | ||||
6283 | } | ||||
6284 | |||||
6285 | if (!AllSeenFirst || (AllowPromotionWithoutCommonHeader && | ||||
6286 | SpeculativelyMovedExts.size() == 1)) { | ||||
6287 | TPT.commit(); | ||||
6288 | if (HasPromoted) | ||||
6289 | Promoted = true; | ||||
6290 | for (auto *I : SpeculativelyMovedExts) { | ||||
6291 | Value *HeadOfChain = I->getOperand(0); | ||||
6292 | SeenChainsForSExt[HeadOfChain] = nullptr; | ||||
6293 | ValToSExtendedUses[HeadOfChain].push_back(I); | ||||
6294 | } | ||||
6295 | // Update Inst as promotion happen. | ||||
6296 | Inst = SpeculativelyMovedExts.pop_back_val(); | ||||
6297 | } else { | ||||
6298 | // This is the first chain visited from the header, keep the current chain | ||||
6299 | // as unhandled. Defer to promote this until we encounter another SExt | ||||
6300 | // chain derived from the same header. | ||||
6301 | for (auto *I : SpeculativelyMovedExts) { | ||||
6302 | Value *HeadOfChain = I->getOperand(0); | ||||
6303 | SeenChainsForSExt[HeadOfChain] = Inst; | ||||
6304 | } | ||||
6305 | return false; | ||||
6306 | } | ||||
6307 | |||||
6308 | if (!AllSeenFirst && !UnhandledExts.empty()) | ||||
6309 | for (auto *VisitedSExt : UnhandledExts) { | ||||
6310 | if (RemovedInsts.count(VisitedSExt)) | ||||
6311 | continue; | ||||
6312 | TypePromotionTransaction TPT(RemovedInsts); | ||||
6313 | SmallVector<Instruction *, 1> Exts; | ||||
6314 | SmallVector<Instruction *, 2> Chains; | ||||
6315 | Exts.push_back(VisitedSExt); | ||||
6316 | bool HasPromoted = tryToPromoteExts(TPT, Exts, Chains); | ||||
6317 | TPT.commit(); | ||||
6318 | if (HasPromoted) | ||||
6319 | Promoted = true; | ||||
6320 | for (auto *I : Chains) { | ||||
6321 | Value *HeadOfChain = I->getOperand(0); | ||||
6322 | // Mark this as handled. | ||||
6323 | SeenChainsForSExt[HeadOfChain] = nullptr; | ||||
6324 | ValToSExtendedUses[HeadOfChain].push_back(I); | ||||
6325 | } | ||||
6326 | } | ||||
6327 | return Promoted; | ||||
6328 | } | ||||
6329 | |||||
6330 | bool CodeGenPrepare::optimizeExtUses(Instruction *I) { | ||||
6331 | BasicBlock *DefBB = I->getParent(); | ||||
6332 | |||||
6333 | // If the result of a {s|z}ext and its source are both live out, rewrite all | ||||
6334 | // other uses of the source with result of extension. | ||||
6335 | Value *Src = I->getOperand(0); | ||||
6336 | if (Src->hasOneUse()) | ||||
6337 | return false; | ||||
6338 | |||||
6339 | // Only do this xform if truncating is free. | ||||
6340 | if (!TLI->isTruncateFree(I->getType(), Src->getType())) | ||||
6341 | return false; | ||||
6342 | |||||
6343 | // Only safe to perform the optimization if the source is also defined in | ||||
6344 | // this block. | ||||
6345 | if (!isa<Instruction>(Src) || DefBB != cast<Instruction>(Src)->getParent()) | ||||
6346 | return false; | ||||
6347 | |||||
6348 | bool DefIsLiveOut = false; | ||||
6349 | for (User *U : I->users()) { | ||||
6350 | Instruction *UI = cast<Instruction>(U); | ||||
6351 | |||||
6352 | // Figure out which BB this ext is used in. | ||||
6353 | BasicBlock *UserBB = UI->getParent(); | ||||
6354 | if (UserBB == DefBB) continue; | ||||
6355 | DefIsLiveOut = true; | ||||
6356 | break; | ||||
6357 | } | ||||
6358 | if (!DefIsLiveOut) | ||||
6359 | return false; | ||||
6360 | |||||
6361 | // Make sure none of the uses are PHI nodes. | ||||
6362 | for (User *U : Src->users()) { | ||||
6363 | Instruction *UI = cast<Instruction>(U); | ||||
6364 | BasicBlock *UserBB = UI->getParent(); | ||||
6365 | if (UserBB == DefBB) continue; | ||||
6366 | // Be conservative. We don't want this xform to end up introducing | ||||
6367 | // reloads just before load / store instructions. | ||||
6368 | if (isa<PHINode>(UI) || isa<LoadInst>(UI) || isa<StoreInst>(UI)) | ||||
6369 | return false; | ||||
6370 | } | ||||
6371 | |||||
6372 | // InsertedTruncs - Only insert one trunc in each block once. | ||||
6373 | DenseMap<BasicBlock*, Instruction*> InsertedTruncs; | ||||
6374 | |||||
6375 | bool MadeChange = false; | ||||
6376 | for (Use &U : Src->uses()) { | ||||
6377 | Instruction *User = cast<Instruction>(U.getUser()); | ||||
6378 | |||||
6379 | // Figure out which BB this ext is used in. | ||||
6380 | BasicBlock *UserBB = User->getParent(); | ||||
6381 | if (UserBB == DefBB) continue; | ||||
6382 | |||||
6383 | // Both src and def are live in this block. Rewrite the use. | ||||
6384 | Instruction *&InsertedTrunc = InsertedTruncs[UserBB]; | ||||
6385 | |||||
6386 | if (!InsertedTrunc) { | ||||
6387 | BasicBlock::iterator InsertPt = UserBB->getFirstInsertionPt(); | ||||
6388 | assert(InsertPt != UserBB->end())((void)0); | ||||
6389 | InsertedTrunc = new TruncInst(I, Src->getType(), "", &*InsertPt); | ||||
6390 | InsertedInsts.insert(InsertedTrunc); | ||||
6391 | } | ||||
6392 | |||||
6393 | // Replace a use of the {s|z}ext source with a use of the result. | ||||
6394 | U = InsertedTrunc; | ||||
6395 | ++NumExtUses; | ||||
6396 | MadeChange = true; | ||||
6397 | } | ||||
6398 | |||||
6399 | return MadeChange; | ||||
6400 | } | ||||
6401 | |||||
6402 | // Find loads whose uses only use some of the loaded value's bits. Add an "and" | ||||
6403 | // just after the load if the target can fold this into one extload instruction, | ||||
6404 | // with the hope of eliminating some of the other later "and" instructions using | ||||
6405 | // the loaded value. "and"s that are made trivially redundant by the insertion | ||||
6406 | // of the new "and" are removed by this function, while others (e.g. those whose | ||||
6407 | // path from the load goes through a phi) are left for isel to potentially | ||||
6408 | // remove. | ||||
6409 | // | ||||
6410 | // For example: | ||||
6411 | // | ||||
6412 | // b0: | ||||
6413 | // x = load i32 | ||||
6414 | // ... | ||||
6415 | // b1: | ||||
6416 | // y = and x, 0xff | ||||
6417 | // z = use y | ||||
6418 | // | ||||
6419 | // becomes: | ||||
6420 | // | ||||
6421 | // b0: | ||||
6422 | // x = load i32 | ||||
6423 | // x' = and x, 0xff | ||||
6424 | // ... | ||||
6425 | // b1: | ||||
6426 | // z = use x' | ||||
6427 | // | ||||
6428 | // whereas: | ||||
6429 | // | ||||
6430 | // b0: | ||||
6431 | // x1 = load i32 | ||||
6432 | // ... | ||||
6433 | // b1: | ||||
6434 | // x2 = load i32 | ||||
6435 | // ... | ||||
6436 | // b2: | ||||
6437 | // x = phi x1, x2 | ||||
6438 | // y = and x, 0xff | ||||
6439 | // | ||||
6440 | // becomes (after a call to optimizeLoadExt for each load): | ||||
6441 | // | ||||
6442 | // b0: | ||||
6443 | // x1 = load i32 | ||||
6444 | // x1' = and x1, 0xff | ||||
6445 | // ... | ||||
6446 | // b1: | ||||
6447 | // x2 = load i32 | ||||
6448 | // x2' = and x2, 0xff | ||||
6449 | // ... | ||||
6450 | // b2: | ||||
6451 | // x = phi x1', x2' | ||||
6452 | // y = and x, 0xff | ||||
6453 | bool CodeGenPrepare::optimizeLoadExt(LoadInst *Load) { | ||||
6454 | if (!Load->isSimple() || !Load->getType()->isIntOrPtrTy()) | ||||
6455 | return false; | ||||
6456 | |||||
6457 | // Skip loads we've already transformed. | ||||
6458 | if (Load->hasOneUse() && | ||||
6459 | InsertedInsts.count(cast<Instruction>(*Load->user_begin()))) | ||||
6460 | return false; | ||||
6461 | |||||
6462 | // Look at all uses of Load, looking through phis, to determine how many bits | ||||
6463 | // of the loaded value are needed. | ||||
6464 | SmallVector<Instruction *, 8> WorkList; | ||||
6465 | SmallPtrSet<Instruction *, 16> Visited; | ||||
6466 | SmallVector<Instruction *, 8> AndsToMaybeRemove; | ||||
6467 | for (auto *U : Load->users()) | ||||
6468 | WorkList.push_back(cast<Instruction>(U)); | ||||
6469 | |||||
6470 | EVT LoadResultVT = TLI->getValueType(*DL, Load->getType()); | ||||
6471 | unsigned BitWidth = LoadResultVT.getSizeInBits(); | ||||
6472 | // If the BitWidth is 0, do not try to optimize the type | ||||
6473 | if (BitWidth == 0) | ||||
6474 | return false; | ||||
6475 | |||||
6476 | APInt DemandBits(BitWidth, 0); | ||||
6477 | APInt WidestAndBits(BitWidth, 0); | ||||
6478 | |||||
6479 | while (!WorkList.empty()) { | ||||
6480 | Instruction *I = WorkList.back(); | ||||
6481 | WorkList.pop_back(); | ||||
6482 | |||||
6483 | // Break use-def graph loops. | ||||
6484 | if (!Visited.insert(I).second) | ||||
6485 | continue; | ||||
6486 | |||||
6487 | // For a PHI node, push all of its users. | ||||
6488 | if (auto *Phi = dyn_cast<PHINode>(I)) { | ||||
6489 | for (auto *U : Phi->users()) | ||||
6490 | WorkList.push_back(cast<Instruction>(U)); | ||||
6491 | continue; | ||||
6492 | } | ||||
6493 | |||||
6494 | switch (I->getOpcode()) { | ||||
6495 | case Instruction::And: { | ||||
6496 | auto *AndC = dyn_cast<ConstantInt>(I->getOperand(1)); | ||||
6497 | if (!AndC) | ||||
6498 | return false; | ||||
6499 | APInt AndBits = AndC->getValue(); | ||||
6500 | DemandBits |= AndBits; | ||||
6501 | // Keep track of the widest and mask we see. | ||||
6502 | if (AndBits.ugt(WidestAndBits)) | ||||
6503 | WidestAndBits = AndBits; | ||||
6504 | if (AndBits == WidestAndBits && I->getOperand(0) == Load) | ||||
6505 | AndsToMaybeRemove.push_back(I); | ||||
6506 | break; | ||||
6507 | } | ||||
6508 | |||||
6509 | case Instruction::Shl: { | ||||
6510 | auto *ShlC = dyn_cast<ConstantInt>(I->getOperand(1)); | ||||
6511 | if (!ShlC) | ||||
6512 | return false; | ||||
6513 | uint64_t ShiftAmt = ShlC->getLimitedValue(BitWidth - 1); | ||||
6514 | DemandBits.setLowBits(BitWidth - ShiftAmt); | ||||
6515 | break; | ||||
6516 | } | ||||
6517 | |||||
6518 | case Instruction::Trunc: { | ||||
6519 | EVT TruncVT = TLI->getValueType(*DL, I->getType()); | ||||
6520 | unsigned TruncBitWidth = TruncVT.getSizeInBits(); | ||||
6521 | DemandBits.setLowBits(TruncBitWidth); | ||||
6522 | break; | ||||
6523 | } | ||||
6524 | |||||
6525 | default: | ||||
6526 | return false; | ||||
6527 | } | ||||
6528 | } | ||||
6529 | |||||
6530 | uint32_t ActiveBits = DemandBits.getActiveBits(); | ||||
6531 | // Avoid hoisting (and (load x) 1) since it is unlikely to be folded by the | ||||
6532 | // target even if isLoadExtLegal says an i1 EXTLOAD is valid. For example, | ||||
6533 | // for the AArch64 target isLoadExtLegal(ZEXTLOAD, i32, i1) returns true, but | ||||
6534 | // (and (load x) 1) is not matched as a single instruction, rather as a LDR | ||||
6535 | // followed by an AND. | ||||
6536 | // TODO: Look into removing this restriction by fixing backends to either | ||||
6537 | // return false for isLoadExtLegal for i1 or have them select this pattern to | ||||
6538 | // a single instruction. | ||||
6539 | // | ||||
6540 | // Also avoid hoisting if we didn't see any ands with the exact DemandBits | ||||
6541 | // mask, since these are the only ands that will be removed by isel. | ||||
6542 | if (ActiveBits <= 1 || !DemandBits.isMask(ActiveBits) || | ||||
6543 | WidestAndBits != DemandBits) | ||||
6544 | return false; | ||||
6545 | |||||
6546 | LLVMContext &Ctx = Load->getType()->getContext(); | ||||
6547 | Type *TruncTy = Type::getIntNTy(Ctx, ActiveBits); | ||||
6548 | EVT TruncVT = TLI->getValueType(*DL, TruncTy); | ||||
6549 | |||||
6550 | // Reject cases that won't be matched as extloads. | ||||
6551 | if (!LoadResultVT.bitsGT(TruncVT) || !TruncVT.isRound() || | ||||
6552 | !TLI->isLoadExtLegal(ISD::ZEXTLOAD, LoadResultVT, TruncVT)) | ||||
6553 | return false; | ||||
6554 | |||||
6555 | IRBuilder<> Builder(Load->getNextNode()); | ||||
6556 | auto *NewAnd = cast<Instruction>( | ||||
6557 | Builder.CreateAnd(Load, ConstantInt::get(Ctx, DemandBits))); | ||||
6558 | // Mark this instruction as "inserted by CGP", so that other | ||||
6559 | // optimizations don't touch it. | ||||
6560 | InsertedInsts.insert(NewAnd); | ||||
6561 | |||||
6562 | // Replace all uses of load with new and (except for the use of load in the | ||||
6563 | // new and itself). | ||||
6564 | Load->replaceAllUsesWith(NewAnd); | ||||
6565 | NewAnd->setOperand(0, Load); | ||||
6566 | |||||
6567 | // Remove any and instructions that are now redundant. | ||||
6568 | for (auto *And : AndsToMaybeRemove) | ||||
6569 | // Check that the and mask is the same as the one we decided to put on the | ||||
6570 | // new and. | ||||
6571 | if (cast<ConstantInt>(And->getOperand(1))->getValue() == DemandBits) { | ||||
6572 | And->replaceAllUsesWith(NewAnd); | ||||
6573 | if (&*CurInstIterator == And) | ||||
6574 | CurInstIterator = std::next(And->getIterator()); | ||||
6575 | And->eraseFromParent(); | ||||
6576 | ++NumAndUses; | ||||
6577 | } | ||||
6578 | |||||
6579 | ++NumAndsAdded; | ||||
6580 | return true; | ||||
6581 | } | ||||
6582 | |||||
6583 | /// Check if V (an operand of a select instruction) is an expensive instruction | ||||
6584 | /// that is only used once. | ||||
6585 | static bool sinkSelectOperand(const TargetTransformInfo *TTI, Value *V) { | ||||
6586 | auto *I = dyn_cast<Instruction>(V); | ||||
6587 | // If it's safe to speculatively execute, then it should not have side | ||||
6588 | // effects; therefore, it's safe to sink and possibly *not* execute. | ||||
6589 | return I && I->hasOneUse() && isSafeToSpeculativelyExecute(I) && | ||||
6590 | TTI->getUserCost(I, TargetTransformInfo::TCK_SizeAndLatency) >= | ||||
6591 | TargetTransformInfo::TCC_Expensive; | ||||
6592 | } | ||||
6593 | |||||
6594 | /// Returns true if a SelectInst should be turned into an explicit branch. | ||||
6595 | static bool isFormingBranchFromSelectProfitable(const TargetTransformInfo *TTI, | ||||
6596 | const TargetLowering *TLI, | ||||
6597 | SelectInst *SI) { | ||||
6598 | // If even a predictable select is cheap, then a branch can't be cheaper. | ||||
6599 | if (!TLI->isPredictableSelectExpensive()) | ||||
6600 | return false; | ||||
6601 | |||||
6602 | // FIXME: This should use the same heuristics as IfConversion to determine | ||||
6603 | // whether a select is better represented as a branch. | ||||
6604 | |||||
6605 | // If metadata tells us that the select condition is obviously predictable, | ||||
6606 | // then we want to replace the select with a branch. | ||||
6607 | uint64_t TrueWeight, FalseWeight; | ||||
6608 | if (SI->extractProfMetadata(TrueWeight, FalseWeight)) { | ||||
6609 | uint64_t Max = std::max(TrueWeight, FalseWeight); | ||||
6610 | uint64_t Sum = TrueWeight + FalseWeight; | ||||
6611 | if (Sum != 0) { | ||||
6612 | auto Probability = BranchProbability::getBranchProbability(Max, Sum); | ||||
6613 | if (Probability > TTI->getPredictableBranchThreshold()) | ||||
6614 | return true; | ||||
6615 | } | ||||
6616 | } | ||||
6617 | |||||
6618 | CmpInst *Cmp = dyn_cast<CmpInst>(SI->getCondition()); | ||||
6619 | |||||
6620 | // If a branch is predictable, an out-of-order CPU can avoid blocking on its | ||||
6621 | // comparison condition. If the compare has more than one use, there's | ||||
6622 | // probably another cmov or setcc around, so it's not worth emitting a branch. | ||||
6623 | if (!Cmp || !Cmp->hasOneUse()) | ||||
6624 | return false; | ||||
6625 | |||||
6626 | // If either operand of the select is expensive and only needed on one side | ||||
6627 | // of the select, we should form a branch. | ||||
6628 | if (sinkSelectOperand(TTI, SI->getTrueValue()) || | ||||
6629 | sinkSelectOperand(TTI, SI->getFalseValue())) | ||||
6630 | return true; | ||||
6631 | |||||
6632 | return false; | ||||
6633 | } | ||||
6634 | |||||
6635 | /// If \p isTrue is true, return the true value of \p SI, otherwise return | ||||
6636 | /// false value of \p SI. If the true/false value of \p SI is defined by any | ||||
6637 | /// select instructions in \p Selects, look through the defining select | ||||
6638 | /// instruction until the true/false value is not defined in \p Selects. | ||||
6639 | static Value *getTrueOrFalseValue( | ||||
6640 | SelectInst *SI, bool isTrue, | ||||
6641 | const SmallPtrSet<const Instruction *, 2> &Selects) { | ||||
6642 | Value *V = nullptr; | ||||
6643 | |||||
6644 | for (SelectInst *DefSI = SI; DefSI != nullptr && Selects.count(DefSI); | ||||
6645 | DefSI = dyn_cast<SelectInst>(V)) { | ||||
6646 | assert(DefSI->getCondition() == SI->getCondition() &&((void)0) | ||||
6647 | "The condition of DefSI does not match with SI")((void)0); | ||||
6648 | V = (isTrue ? DefSI->getTrueValue() : DefSI->getFalseValue()); | ||||
6649 | } | ||||
6650 | |||||
6651 | assert(V && "Failed to get select true/false value")((void)0); | ||||
6652 | return V; | ||||
6653 | } | ||||
6654 | |||||
6655 | bool CodeGenPrepare::optimizeShiftInst(BinaryOperator *Shift) { | ||||
6656 | assert(Shift->isShift() && "Expected a shift")((void)0); | ||||
6657 | |||||
6658 | // If this is (1) a vector shift, (2) shifts by scalars are cheaper than | ||||
6659 | // general vector shifts, and (3) the shift amount is a select-of-splatted | ||||
6660 | // values, hoist the shifts before the select: | ||||
6661 | // shift Op0, (select Cond, TVal, FVal) --> | ||||
6662 | // select Cond, (shift Op0, TVal), (shift Op0, FVal) | ||||
6663 | // | ||||
6664 | // This is inverting a generic IR transform when we know that the cost of a | ||||
6665 | // general vector shift is more than the cost of 2 shift-by-scalars. | ||||
6666 | // We can't do this effectively in SDAG because we may not be able to | ||||
6667 | // determine if the select operands are splats from within a basic block. | ||||
6668 | Type *Ty = Shift->getType(); | ||||
6669 | if (!Ty->isVectorTy() || !TLI->isVectorShiftByScalarCheap(Ty)) | ||||
6670 | return false; | ||||
6671 | Value *Cond, *TVal, *FVal; | ||||
6672 | if (!match(Shift->getOperand(1), | ||||
6673 | m_OneUse(m_Select(m_Value(Cond), m_Value(TVal), m_Value(FVal))))) | ||||
6674 | return false; | ||||
6675 | if (!isSplatValue(TVal) || !isSplatValue(FVal)) | ||||
6676 | return false; | ||||
6677 | |||||
6678 | IRBuilder<> Builder(Shift); | ||||
6679 | BinaryOperator::BinaryOps Opcode = Shift->getOpcode(); | ||||
6680 | Value *NewTVal = Builder.CreateBinOp(Opcode, Shift->getOperand(0), TVal); | ||||
6681 | Value *NewFVal = Builder.CreateBinOp(Opcode, Shift->getOperand(0), FVal); | ||||
6682 | Value *NewSel = Builder.CreateSelect(Cond, NewTVal, NewFVal); | ||||
6683 | Shift->replaceAllUsesWith(NewSel); | ||||
6684 | Shift->eraseFromParent(); | ||||
6685 | return true; | ||||
6686 | } | ||||
6687 | |||||
6688 | bool CodeGenPrepare::optimizeFunnelShift(IntrinsicInst *Fsh) { | ||||
6689 | Intrinsic::ID Opcode = Fsh->getIntrinsicID(); | ||||
6690 | assert((Opcode == Intrinsic::fshl || Opcode == Intrinsic::fshr) &&((void)0) | ||||
6691 | "Expected a funnel shift")((void)0); | ||||
6692 | |||||
6693 | // If this is (1) a vector funnel shift, (2) shifts by scalars are cheaper | ||||
6694 | // than general vector shifts, and (3) the shift amount is select-of-splatted | ||||
6695 | // values, hoist the funnel shifts before the select: | ||||
6696 | // fsh Op0, Op1, (select Cond, TVal, FVal) --> | ||||
6697 | // select Cond, (fsh Op0, Op1, TVal), (fsh Op0, Op1, FVal) | ||||
6698 | // | ||||
6699 | // This is inverting a generic IR transform when we know that the cost of a | ||||
6700 | // general vector shift is more than the cost of 2 shift-by-scalars. | ||||
6701 | // We can't do this effectively in SDAG because we may not be able to | ||||
6702 | // determine if the select operands are splats from within a basic block. | ||||
6703 | Type *Ty = Fsh->getType(); | ||||
6704 | if (!Ty->isVectorTy() || !TLI->isVectorShiftByScalarCheap(Ty)) | ||||
6705 | return false; | ||||
6706 | Value *Cond, *TVal, *FVal; | ||||
6707 | if (!match(Fsh->getOperand(2), | ||||
6708 | m_OneUse(m_Select(m_Value(Cond), m_Value(TVal), m_Value(FVal))))) | ||||
6709 | return false; | ||||
6710 | if (!isSplatValue(TVal) || !isSplatValue(FVal)) | ||||
6711 | return false; | ||||
6712 | |||||
6713 | IRBuilder<> Builder(Fsh); | ||||
6714 | Value *X = Fsh->getOperand(0), *Y = Fsh->getOperand(1); | ||||
6715 | Value *NewTVal = Builder.CreateIntrinsic(Opcode, Ty, { X, Y, TVal }); | ||||
6716 | Value *NewFVal = Builder.CreateIntrinsic(Opcode, Ty, { X, Y, FVal }); | ||||
6717 | Value *NewSel = Builder.CreateSelect(Cond, NewTVal, NewFVal); | ||||
6718 | Fsh->replaceAllUsesWith(NewSel); | ||||
6719 | Fsh->eraseFromParent(); | ||||
6720 | return true; | ||||
6721 | } | ||||
6722 | |||||
6723 | /// If we have a SelectInst that will likely profit from branch prediction, | ||||
6724 | /// turn it into a branch. | ||||
6725 | bool CodeGenPrepare::optimizeSelectInst(SelectInst *SI) { | ||||
6726 | if (DisableSelectToBranch) | ||||
6727 | return false; | ||||
6728 | |||||
6729 | // Find all consecutive select instructions that share the same condition. | ||||
6730 | SmallVector<SelectInst *, 2> ASI; | ||||
6731 | ASI.push_back(SI); | ||||
6732 | for (BasicBlock::iterator It = ++BasicBlock::iterator(SI); | ||||
6733 | It != SI->getParent()->end(); ++It) { | ||||
6734 | SelectInst *I = dyn_cast<SelectInst>(&*It); | ||||
6735 | if (I && SI->getCondition() == I->getCondition()) { | ||||
6736 | ASI.push_back(I); | ||||
6737 | } else { | ||||
6738 | break; | ||||
6739 | } | ||||
6740 | } | ||||
6741 | |||||
6742 | SelectInst *LastSI = ASI.back(); | ||||
6743 | // Increment the current iterator to skip all the rest of select instructions | ||||
6744 | // because they will be either "not lowered" or "all lowered" to branch. | ||||
6745 | CurInstIterator = std::next(LastSI->getIterator()); | ||||
6746 | |||||
6747 | bool VectorCond = !SI->getCondition()->getType()->isIntegerTy(1); | ||||
6748 | |||||
6749 | // Can we convert the 'select' to CF ? | ||||
6750 | if (VectorCond || SI->getMetadata(LLVMContext::MD_unpredictable)) | ||||
6751 | return false; | ||||
6752 | |||||
6753 | TargetLowering::SelectSupportKind SelectKind; | ||||
6754 | if (VectorCond) | ||||
6755 | SelectKind = TargetLowering::VectorMaskSelect; | ||||
6756 | else if (SI->getType()->isVectorTy()) | ||||
6757 | SelectKind = TargetLowering::ScalarCondVectorVal; | ||||
6758 | else | ||||
6759 | SelectKind = TargetLowering::ScalarValSelect; | ||||
6760 | |||||
6761 | if (TLI->isSelectSupported(SelectKind) && | ||||
6762 | (!isFormingBranchFromSelectProfitable(TTI, TLI, SI) || OptSize || | ||||
6763 | llvm::shouldOptimizeForSize(SI->getParent(), PSI, BFI.get()))) | ||||
6764 | return false; | ||||
6765 | |||||
6766 | // The DominatorTree needs to be rebuilt by any consumers after this | ||||
6767 | // transformation. We simply reset here rather than setting the ModifiedDT | ||||
6768 | // flag to avoid restarting the function walk in runOnFunction for each | ||||
6769 | // select optimized. | ||||
6770 | DT.reset(); | ||||
6771 | |||||
6772 | // Transform a sequence like this: | ||||
6773 | // start: | ||||
6774 | // %cmp = cmp uge i32 %a, %b | ||||
6775 | // %sel = select i1 %cmp, i32 %c, i32 %d | ||||
6776 | // | ||||
6777 | // Into: | ||||
6778 | // start: | ||||
6779 | // %cmp = cmp uge i32 %a, %b | ||||
6780 | // %cmp.frozen = freeze %cmp | ||||
6781 | // br i1 %cmp.frozen, label %select.true, label %select.false | ||||
6782 | // select.true: | ||||
6783 | // br label %select.end | ||||
6784 | // select.false: | ||||
6785 | // br label %select.end | ||||
6786 | // select.end: | ||||
6787 | // %sel = phi i32 [ %c, %select.true ], [ %d, %select.false ] | ||||
6788 | // | ||||
6789 | // %cmp should be frozen, otherwise it may introduce undefined behavior. | ||||
6790 | // In addition, we may sink instructions that produce %c or %d from | ||||
6791 | // the entry block into the destination(s) of the new branch. | ||||
6792 | // If the true or false blocks do not contain a sunken instruction, that | ||||
6793 | // block and its branch may be optimized away. In that case, one side of the | ||||
6794 | // first branch will point directly to select.end, and the corresponding PHI | ||||
6795 | // predecessor block will be the start block. | ||||
6796 | |||||
6797 | // First, we split the block containing the select into 2 blocks. | ||||
6798 | BasicBlock *StartBlock = SI->getParent(); | ||||
6799 | BasicBlock::iterator SplitPt = ++(BasicBlock::iterator(LastSI)); | ||||
6800 | BasicBlock *EndBlock = StartBlock->splitBasicBlock(SplitPt, "select.end"); | ||||
6801 | BFI->setBlockFreq(EndBlock, BFI->getBlockFreq(StartBlock).getFrequency()); | ||||
6802 | |||||
6803 | // Delete the unconditional branch that was just created by the split. | ||||
6804 | StartBlock->getTerminator()->eraseFromParent(); | ||||
6805 | |||||
6806 | // These are the new basic blocks for the conditional branch. | ||||
6807 | // At least one will become an actual new basic block. | ||||
6808 | BasicBlock *TrueBlock = nullptr; | ||||
6809 | BasicBlock *FalseBlock = nullptr; | ||||
6810 | BranchInst *TrueBranch = nullptr; | ||||
6811 | BranchInst *FalseBranch = nullptr; | ||||
6812 | |||||
6813 | // Sink expensive instructions into the conditional blocks to avoid executing | ||||
6814 | // them speculatively. | ||||
6815 | for (SelectInst *SI : ASI) { | ||||
6816 | if (sinkSelectOperand(TTI, SI->getTrueValue())) { | ||||
6817 | if (TrueBlock == nullptr) { | ||||
6818 | TrueBlock = BasicBlock::Create(SI->getContext(), "select.true.sink", | ||||
6819 | EndBlock->getParent(), EndBlock); | ||||
6820 | TrueBranch = BranchInst::Create(EndBlock, TrueBlock); | ||||
6821 | TrueBranch->setDebugLoc(SI->getDebugLoc()); | ||||
6822 | } | ||||
6823 | auto *TrueInst = cast<Instruction>(SI->getTrueValue()); | ||||
6824 | TrueInst->moveBefore(TrueBranch); | ||||
6825 | } | ||||
6826 | if (sinkSelectOperand(TTI, SI->getFalseValue())) { | ||||
6827 | if (FalseBlock == nullptr) { | ||||
6828 | FalseBlock = BasicBlock::Create(SI->getContext(), "select.false.sink", | ||||
6829 | EndBlock->getParent(), EndBlock); | ||||
6830 | FalseBranch = BranchInst::Create(EndBlock, FalseBlock); | ||||
6831 | FalseBranch->setDebugLoc(SI->getDebugLoc()); | ||||
6832 | } | ||||
6833 | auto *FalseInst = cast<Instruction>(SI->getFalseValue()); | ||||
6834 | FalseInst->moveBefore(FalseBranch); | ||||
6835 | } | ||||
6836 | } | ||||
6837 | |||||
6838 | // If there was nothing to sink, then arbitrarily choose the 'false' side | ||||
6839 | // for a new input value to the PHI. | ||||
6840 | if (TrueBlock == FalseBlock) { | ||||
6841 | assert(TrueBlock == nullptr &&((void)0) | ||||
6842 | "Unexpected basic block transform while optimizing select")((void)0); | ||||
6843 | |||||
6844 | FalseBlock = BasicBlock::Create(SI->getContext(), "select.false", | ||||
6845 | EndBlock->getParent(), EndBlock); | ||||
6846 | auto *FalseBranch = BranchInst::Create(EndBlock, FalseBlock); | ||||
6847 | FalseBranch->setDebugLoc(SI->getDebugLoc()); | ||||
6848 | } | ||||
6849 | |||||
6850 | // Insert the real conditional branch based on the original condition. | ||||
6851 | // If we did not create a new block for one of the 'true' or 'false' paths | ||||
6852 | // of the condition, it means that side of the branch goes to the end block | ||||
6853 | // directly and the path originates from the start block from the point of | ||||
6854 | // view of the new PHI. | ||||
6855 | BasicBlock *TT, *FT; | ||||
6856 | if (TrueBlock == nullptr) { | ||||
6857 | TT = EndBlock; | ||||
6858 | FT = FalseBlock; | ||||
6859 | TrueBlock = StartBlock; | ||||
6860 | } else if (FalseBlock == nullptr) { | ||||
6861 | TT = TrueBlock; | ||||
6862 | FT = EndBlock; | ||||
6863 | FalseBlock = StartBlock; | ||||
6864 | } else { | ||||
6865 | TT = TrueBlock; | ||||
6866 | FT = FalseBlock; | ||||
6867 | } | ||||
6868 | IRBuilder<> IB(SI); | ||||
6869 | auto *CondFr = IB.CreateFreeze(SI->getCondition(), SI->getName() + ".frozen"); | ||||
6870 | IB.CreateCondBr(CondFr, TT, FT, SI); | ||||
6871 | |||||
6872 | SmallPtrSet<const Instruction *, 2> INS; | ||||
6873 | INS.insert(ASI.begin(), ASI.end()); | ||||
6874 | // Use reverse iterator because later select may use the value of the | ||||
6875 | // earlier select, and we need to propagate value through earlier select | ||||
6876 | // to get the PHI operand. | ||||
6877 | for (auto It = ASI.rbegin(); It != ASI.rend(); ++It) { | ||||
6878 | SelectInst *SI = *It; | ||||
6879 | // The select itself is replaced with a PHI Node. | ||||
6880 | PHINode *PN = PHINode::Create(SI->getType(), 2, "", &EndBlock->front()); | ||||
6881 | PN->takeName(SI); | ||||
6882 | PN->addIncoming(getTrueOrFalseValue(SI, true, INS), TrueBlock); | ||||
6883 | PN->addIncoming(getTrueOrFalseValue(SI, false, INS), FalseBlock); | ||||
6884 | PN->setDebugLoc(SI->getDebugLoc()); | ||||
6885 | |||||
6886 | SI->replaceAllUsesWith(PN); | ||||
6887 | SI->eraseFromParent(); | ||||
6888 | INS.erase(SI); | ||||
6889 | ++NumSelectsExpanded; | ||||
6890 | } | ||||
6891 | |||||
6892 | // Instruct OptimizeBlock to skip to the next block. | ||||
6893 | CurInstIterator = StartBlock->end(); | ||||
6894 | return true; | ||||
6895 | } | ||||
6896 | |||||
6897 | /// Some targets only accept certain types for splat inputs. For example a VDUP | ||||
6898 | /// in MVE takes a GPR (integer) register, and the instruction that incorporate | ||||
6899 | /// a VDUP (such as a VADD qd, qm, rm) also require a gpr register. | ||||
6900 | bool CodeGenPrepare::optimizeShuffleVectorInst(ShuffleVectorInst *SVI) { | ||||
6901 | // Accept shuf(insertelem(undef/poison, val, 0), undef/poison, <0,0,..>) only | ||||
6902 | if (!match(SVI, m_Shuffle(m_InsertElt(m_Undef(), m_Value(), m_ZeroInt()), | ||||
6903 | m_Undef(), m_ZeroMask()))) | ||||
6904 | return false; | ||||
6905 | Type *NewType = TLI->shouldConvertSplatType(SVI); | ||||
6906 | if (!NewType) | ||||
6907 | return false; | ||||
6908 | |||||
6909 | auto *SVIVecType = cast<FixedVectorType>(SVI->getType()); | ||||
6910 | assert(!NewType->isVectorTy() && "Expected a scalar type!")((void)0); | ||||
6911 | assert(NewType->getScalarSizeInBits() == SVIVecType->getScalarSizeInBits() &&((void)0) | ||||
6912 | "Expected a type of the same size!")((void)0); | ||||
6913 | auto *NewVecType = | ||||
6914 | FixedVectorType::get(NewType, SVIVecType->getNumElements()); | ||||
6915 | |||||
6916 | // Create a bitcast (shuffle (insert (bitcast(..)))) | ||||
6917 | IRBuilder<> Builder(SVI->getContext()); | ||||
6918 | Builder.SetInsertPoint(SVI); | ||||
6919 | Value *BC1 = Builder.CreateBitCast( | ||||
6920 | cast<Instruction>(SVI->getOperand(0))->getOperand(1), NewType); | ||||
6921 | Value *Shuffle = Builder.CreateVectorSplat(NewVecType->getNumElements(), BC1); | ||||
6922 | Value *BC2 = Builder.CreateBitCast(Shuffle, SVIVecType); | ||||
6923 | |||||
6924 | SVI->replaceAllUsesWith(BC2); | ||||
6925 | RecursivelyDeleteTriviallyDeadInstructions( | ||||
6926 | SVI, TLInfo, nullptr, [&](Value *V) { removeAllAssertingVHReferences(V); }); | ||||
6927 | |||||
6928 | // Also hoist the bitcast up to its operand if it they are not in the same | ||||
6929 | // block. | ||||
6930 | if (auto *BCI = dyn_cast<Instruction>(BC1)) | ||||
6931 | if (auto *Op = dyn_cast<Instruction>(BCI->getOperand(0))) | ||||
6932 | if (BCI->getParent() != Op->getParent() && !isa<PHINode>(Op) && | ||||
6933 | !Op->isTerminator() && !Op->isEHPad()) | ||||
6934 | BCI->moveAfter(Op); | ||||
6935 | |||||
6936 | return true; | ||||
6937 | } | ||||
6938 | |||||
6939 | bool CodeGenPrepare::tryToSinkFreeOperands(Instruction *I) { | ||||
6940 | // If the operands of I can be folded into a target instruction together with | ||||
6941 | // I, duplicate and sink them. | ||||
6942 | SmallVector<Use *, 4> OpsToSink; | ||||
6943 | if (!TLI->shouldSinkOperands(I, OpsToSink)) | ||||
6944 | return false; | ||||
6945 | |||||
6946 | // OpsToSink can contain multiple uses in a use chain (e.g. | ||||
6947 | // (%u1 with %u1 = shufflevector), (%u2 with %u2 = zext %u1)). The dominating | ||||
6948 | // uses must come first, so we process the ops in reverse order so as to not | ||||
6949 | // create invalid IR. | ||||
6950 | BasicBlock *TargetBB = I->getParent(); | ||||
6951 | bool Changed = false; | ||||
6952 | SmallVector<Use *, 4> ToReplace; | ||||
6953 | for (Use *U : reverse(OpsToSink)) { | ||||
6954 | auto *UI = cast<Instruction>(U->get()); | ||||
6955 | if (UI->getParent() == TargetBB || isa<PHINode>(UI)) | ||||
6956 | continue; | ||||
6957 | ToReplace.push_back(U); | ||||
6958 | } | ||||
6959 | |||||
6960 | SetVector<Instruction *> MaybeDead; | ||||
6961 | DenseMap<Instruction *, Instruction *> NewInstructions; | ||||
6962 | Instruction *InsertPoint = I; | ||||
6963 | for (Use *U : ToReplace) { | ||||
6964 | auto *UI = cast<Instruction>(U->get()); | ||||
6965 | Instruction *NI = UI->clone(); | ||||
6966 | NewInstructions[UI] = NI; | ||||
6967 | MaybeDead.insert(UI); | ||||
6968 | LLVM_DEBUG(dbgs() << "Sinking " << *UI << " to user " << *I << "\n")do { } while (false); | ||||
6969 | NI->insertBefore(InsertPoint); | ||||
6970 | InsertPoint = NI; | ||||
6971 | InsertedInsts.insert(NI); | ||||
6972 | |||||
6973 | // Update the use for the new instruction, making sure that we update the | ||||
6974 | // sunk instruction uses, if it is part of a chain that has already been | ||||
6975 | // sunk. | ||||
6976 | Instruction *OldI = cast<Instruction>(U->getUser()); | ||||
6977 | if (NewInstructions.count(OldI)) | ||||
6978 | NewInstructions[OldI]->setOperand(U->getOperandNo(), NI); | ||||
6979 | else | ||||
6980 | U->set(NI); | ||||
6981 | Changed = true; | ||||
6982 | } | ||||
6983 | |||||
6984 | // Remove instructions that are dead after sinking. | ||||
6985 | for (auto *I : MaybeDead) { | ||||
6986 | if (!I->hasNUsesOrMore(1)) { | ||||
6987 | LLVM_DEBUG(dbgs() << "Removing dead instruction: " << *I << "\n")do { } while (false); | ||||
6988 | I->eraseFromParent(); | ||||
6989 | } | ||||
6990 | } | ||||
6991 | |||||
6992 | return Changed; | ||||
6993 | } | ||||
6994 | |||||
6995 | bool CodeGenPrepare::optimizeSwitchInst(SwitchInst *SI) { | ||||
6996 | Value *Cond = SI->getCondition(); | ||||
6997 | Type *OldType = Cond->getType(); | ||||
6998 | LLVMContext &Context = Cond->getContext(); | ||||
6999 | EVT OldVT = TLI->getValueType(*DL, OldType); | ||||
7000 | MVT RegType = TLI->getRegisterType(Context, OldVT); | ||||
7001 | unsigned RegWidth = RegType.getSizeInBits(); | ||||
7002 | |||||
7003 | if (RegWidth <= cast<IntegerType>(OldType)->getBitWidth()) | ||||
7004 | return false; | ||||
7005 | |||||
7006 | // If the register width is greater than the type width, expand the condition | ||||
7007 | // of the switch instruction and each case constant to the width of the | ||||
7008 | // register. By widening the type of the switch condition, subsequent | ||||
7009 | // comparisons (for case comparisons) will not need to be extended to the | ||||
7010 | // preferred register width, so we will potentially eliminate N-1 extends, | ||||
7011 | // where N is the number of cases in the switch. | ||||
7012 | auto *NewType = Type::getIntNTy(Context, RegWidth); | ||||
7013 | |||||
7014 | // Extend the switch condition and case constants using the target preferred | ||||
7015 | // extend unless the switch condition is a function argument with an extend | ||||
7016 | // attribute. In that case, we can avoid an unnecessary mask/extension by | ||||
7017 | // matching the argument extension instead. | ||||
7018 | Instruction::CastOps ExtType = Instruction::ZExt; | ||||
7019 | // Some targets prefer SExt over ZExt. | ||||
7020 | if (TLI->isSExtCheaperThanZExt(OldVT, RegType)) | ||||
7021 | ExtType = Instruction::SExt; | ||||
7022 | |||||
7023 | if (auto *Arg = dyn_cast<Argument>(Cond)) { | ||||
7024 | if (Arg->hasSExtAttr()) | ||||
7025 | ExtType = Instruction::SExt; | ||||
7026 | if (Arg->hasZExtAttr()) | ||||
7027 | ExtType = Instruction::ZExt; | ||||
7028 | } | ||||
7029 | |||||
7030 | auto *ExtInst = CastInst::Create(ExtType, Cond, NewType); | ||||
7031 | ExtInst->insertBefore(SI); | ||||
7032 | ExtInst->setDebugLoc(SI->getDebugLoc()); | ||||
7033 | SI->setCondition(ExtInst); | ||||
7034 | for (auto Case : SI->cases()) { | ||||
7035 | APInt NarrowConst = Case.getCaseValue()->getValue(); | ||||
7036 | APInt WideConst = (ExtType == Instruction::ZExt) ? | ||||
7037 | NarrowConst.zext(RegWidth) : NarrowConst.sext(RegWidth); | ||||
7038 | Case.setValue(ConstantInt::get(Context, WideConst)); | ||||
7039 | } | ||||
7040 | |||||
7041 | return true; | ||||
7042 | } | ||||
7043 | |||||
7044 | |||||
7045 | namespace { | ||||
7046 | |||||
7047 | /// Helper class to promote a scalar operation to a vector one. | ||||
7048 | /// This class is used to move downward extractelement transition. | ||||
7049 | /// E.g., | ||||
7050 | /// a = vector_op <2 x i32> | ||||
7051 | /// b = extractelement <2 x i32> a, i32 0 | ||||
7052 | /// c = scalar_op b | ||||
7053 | /// store c | ||||
7054 | /// | ||||
7055 | /// => | ||||
7056 | /// a = vector_op <2 x i32> | ||||
7057 | /// c = vector_op a (equivalent to scalar_op on the related lane) | ||||
7058 | /// * d = extractelement <2 x i32> c, i32 0 | ||||
7059 | /// * store d | ||||
7060 | /// Assuming both extractelement and store can be combine, we get rid of the | ||||
7061 | /// transition. | ||||
7062 | class VectorPromoteHelper { | ||||
7063 | /// DataLayout associated with the current module. | ||||
7064 | const DataLayout &DL; | ||||
7065 | |||||
7066 | /// Used to perform some checks on the legality of vector operations. | ||||
7067 | const TargetLowering &TLI; | ||||
7068 | |||||
7069 | /// Used to estimated the cost of the promoted chain. | ||||
7070 | const TargetTransformInfo &TTI; | ||||
7071 | |||||
7072 | /// The transition being moved downwards. | ||||
7073 | Instruction *Transition; | ||||
7074 | |||||
7075 | /// The sequence of instructions to be promoted. | ||||
7076 | SmallVector<Instruction *, 4> InstsToBePromoted; | ||||
7077 | |||||
7078 | /// Cost of combining a store and an extract. | ||||
7079 | unsigned StoreExtractCombineCost; | ||||
7080 | |||||
7081 | /// Instruction that will be combined with the transition. | ||||
7082 | Instruction *CombineInst = nullptr; | ||||
7083 | |||||
7084 | /// The instruction that represents the current end of the transition. | ||||
7085 | /// Since we are faking the promotion until we reach the end of the chain | ||||
7086 | /// of computation, we need a way to get the current end of the transition. | ||||
7087 | Instruction *getEndOfTransition() const { | ||||
7088 | if (InstsToBePromoted.empty()) | ||||
7089 | return Transition; | ||||
7090 | return InstsToBePromoted.back(); | ||||
7091 | } | ||||
7092 | |||||
7093 | /// Return the index of the original value in the transition. | ||||
7094 | /// E.g., for "extractelement <2 x i32> c, i32 1" the original value, | ||||
7095 | /// c, is at index 0. | ||||
7096 | unsigned getTransitionOriginalValueIdx() const { | ||||
7097 | assert(isa<ExtractElementInst>(Transition) &&((void)0) | ||||
7098 | "Other kind of transitions are not supported yet")((void)0); | ||||
7099 | return 0; | ||||
7100 | } | ||||
7101 | |||||
7102 | /// Return the index of the index in the transition. | ||||
7103 | /// E.g., for "extractelement <2 x i32> c, i32 0" the index | ||||
7104 | /// is at index 1. | ||||
7105 | unsigned getTransitionIdx() const { | ||||
7106 | assert(isa<ExtractElementInst>(Transition) &&((void)0) | ||||
7107 | "Other kind of transitions are not supported yet")((void)0); | ||||
7108 | return 1; | ||||
7109 | } | ||||
7110 | |||||
7111 | /// Get the type of the transition. | ||||
7112 | /// This is the type of the original value. | ||||
7113 | /// E.g., for "extractelement <2 x i32> c, i32 1" the type of the | ||||
7114 | /// transition is <2 x i32>. | ||||
7115 | Type *getTransitionType() const { | ||||
7116 | return Transition->getOperand(getTransitionOriginalValueIdx())->getType(); | ||||
7117 | } | ||||
7118 | |||||
7119 | /// Promote \p ToBePromoted by moving \p Def downward through. | ||||
7120 | /// I.e., we have the following sequence: | ||||
7121 | /// Def = Transition <ty1> a to <ty2> | ||||
7122 | /// b = ToBePromoted <ty2> Def, ... | ||||
7123 | /// => | ||||
7124 | /// b = ToBePromoted <ty1> a, ... | ||||
7125 | /// Def = Transition <ty1> ToBePromoted to <ty2> | ||||
7126 | void promoteImpl(Instruction *ToBePromoted); | ||||
7127 | |||||
7128 | /// Check whether or not it is profitable to promote all the | ||||
7129 | /// instructions enqueued to be promoted. | ||||
7130 | bool isProfitableToPromote() { | ||||
7131 | Value *ValIdx = Transition->getOperand(getTransitionOriginalValueIdx()); | ||||
7132 | unsigned Index = isa<ConstantInt>(ValIdx) | ||||
7133 | ? cast<ConstantInt>(ValIdx)->getZExtValue() | ||||
7134 | : -1; | ||||
7135 | Type *PromotedType = getTransitionType(); | ||||
7136 | |||||
7137 | StoreInst *ST = cast<StoreInst>(CombineInst); | ||||
7138 | unsigned AS = ST->getPointerAddressSpace(); | ||||
7139 | // Check if this store is supported. | ||||
7140 | if (!TLI.allowsMisalignedMemoryAccesses( | ||||
7141 | TLI.getValueType(DL, ST->getValueOperand()->getType()), AS, | ||||
7142 | ST->getAlign())) { | ||||
7143 | // If this is not supported, there is no way we can combine | ||||
7144 | // the extract with the store. | ||||
7145 | return false; | ||||
7146 | } | ||||
7147 | |||||
7148 | // The scalar chain of computation has to pay for the transition | ||||
7149 | // scalar to vector. | ||||
7150 | // The vector chain has to account for the combining cost. | ||||
7151 | InstructionCost ScalarCost = | ||||
7152 | TTI.getVectorInstrCost(Transition->getOpcode(), PromotedType, Index); | ||||
7153 | InstructionCost VectorCost = StoreExtractCombineCost; | ||||
7154 | enum TargetTransformInfo::TargetCostKind CostKind = | ||||
7155 | TargetTransformInfo::TCK_RecipThroughput; | ||||
7156 | for (const auto &Inst : InstsToBePromoted) { | ||||
7157 | // Compute the cost. | ||||
7158 | // By construction, all instructions being promoted are arithmetic ones. | ||||
7159 | // Moreover, one argument is a constant that can be viewed as a splat | ||||
7160 | // constant. | ||||
7161 | Value *Arg0 = Inst->getOperand(0); | ||||
7162 | bool IsArg0Constant = isa<UndefValue>(Arg0) || isa<ConstantInt>(Arg0) || | ||||
7163 | isa<ConstantFP>(Arg0); | ||||
7164 | TargetTransformInfo::OperandValueKind Arg0OVK = | ||||
7165 | IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue | ||||
7166 | : TargetTransformInfo::OK_AnyValue; | ||||
7167 | TargetTransformInfo::OperandValueKind Arg1OVK = | ||||
7168 | !IsArg0Constant ? TargetTransformInfo::OK_UniformConstantValue | ||||
7169 | : TargetTransformInfo::OK_AnyValue; | ||||
7170 | ScalarCost += TTI.getArithmeticInstrCost( | ||||
7171 | Inst->getOpcode(), Inst->getType(), CostKind, Arg0OVK, Arg1OVK); | ||||
7172 | VectorCost += TTI.getArithmeticInstrCost(Inst->getOpcode(), PromotedType, | ||||
7173 | CostKind, | ||||
7174 | Arg0OVK, Arg1OVK); | ||||
7175 | } | ||||
7176 | LLVM_DEBUG(do { } while (false) | ||||
7177 | dbgs() << "Estimated cost of computation to be promoted:\nScalar: "do { } while (false) | ||||
7178 | << ScalarCost << "\nVector: " << VectorCost << '\n')do { } while (false); | ||||
7179 | return ScalarCost > VectorCost; | ||||
7180 | } | ||||
7181 | |||||
7182 | /// Generate a constant vector with \p Val with the same | ||||
7183 | /// number of elements as the transition. | ||||
7184 | /// \p UseSplat defines whether or not \p Val should be replicated | ||||
7185 | /// across the whole vector. | ||||
7186 | /// In other words, if UseSplat == true, we generate <Val, Val, ..., Val>, | ||||
7187 | /// otherwise we generate a vector with as many undef as possible: | ||||
7188 | /// <undef, ..., undef, Val, undef, ..., undef> where \p Val is only | ||||
7189 | /// used at the index of the extract. | ||||
7190 | Value *getConstantVector(Constant *Val, bool UseSplat) const { | ||||
7191 | unsigned ExtractIdx = std::numeric_limits<unsigned>::max(); | ||||
7192 | if (!UseSplat) { | ||||
7193 | // If we cannot determine where the constant must be, we have to | ||||
7194 | // use a splat constant. | ||||
7195 | Value *ValExtractIdx = Transition->getOperand(getTransitionIdx()); | ||||
7196 | if (ConstantInt *CstVal = dyn_cast<ConstantInt>(ValExtractIdx)) | ||||
7197 | ExtractIdx = CstVal->getSExtValue(); | ||||
7198 | else | ||||
7199 | UseSplat = true; | ||||
7200 | } | ||||
7201 | |||||
7202 | ElementCount EC = cast<VectorType>(getTransitionType())->getElementCount(); | ||||
7203 | if (UseSplat) | ||||
7204 | return ConstantVector::getSplat(EC, Val); | ||||
7205 | |||||
7206 | if (!EC.isScalable()) { | ||||
7207 | SmallVector<Constant *, 4> ConstVec; | ||||
7208 | UndefValue *UndefVal = UndefValue::get(Val->getType()); | ||||
7209 | for (unsigned Idx = 0; Idx != EC.getKnownMinValue(); ++Idx) { | ||||
7210 | if (Idx == ExtractIdx) | ||||
7211 | ConstVec.push_back(Val); | ||||
7212 | else | ||||
7213 | ConstVec.push_back(UndefVal); | ||||
7214 | } | ||||
7215 | return ConstantVector::get(ConstVec); | ||||
7216 | } else | ||||
7217 | llvm_unreachable(__builtin_unreachable() | ||||
7218 | "Generate scalable vector for non-splat is unimplemented")__builtin_unreachable(); | ||||
7219 | } | ||||
7220 | |||||
7221 | /// Check if promoting to a vector type an operand at \p OperandIdx | ||||
7222 | /// in \p Use can trigger undefined behavior. | ||||
7223 | static bool canCauseUndefinedBehavior(const Instruction *Use, | ||||
7224 | unsigned OperandIdx) { | ||||
7225 | // This is not safe to introduce undef when the operand is on | ||||
7226 | // the right hand side of a division-like instruction. | ||||
7227 | if (OperandIdx != 1) | ||||
7228 | return false; | ||||
7229 | switch (Use->getOpcode()) { | ||||
7230 | default: | ||||
7231 | return false; | ||||
7232 | case Instruction::SDiv: | ||||
7233 | case Instruction::UDiv: | ||||
7234 | case Instruction::SRem: | ||||
7235 | case Instruction::URem: | ||||
7236 | return true; | ||||
7237 | case Instruction::FDiv: | ||||
7238 | case Instruction::FRem: | ||||
7239 | return !Use->hasNoNaNs(); | ||||
7240 | } | ||||
7241 | llvm_unreachable(nullptr)__builtin_unreachable(); | ||||
7242 | } | ||||
7243 | |||||
7244 | public: | ||||
7245 | VectorPromoteHelper(const DataLayout &DL, const TargetLowering &TLI, | ||||
7246 | const TargetTransformInfo &TTI, Instruction *Transition, | ||||
7247 | unsigned CombineCost) | ||||
7248 | : DL(DL), TLI(TLI), TTI(TTI), Transition(Transition), | ||||
7249 | StoreExtractCombineCost(CombineCost) { | ||||
7250 | assert(Transition && "Do not know how to promote null")((void)0); | ||||
7251 | } | ||||
7252 | |||||
7253 | /// Check if we can promote \p ToBePromoted to \p Type. | ||||
7254 | bool canPromote(const Instruction *ToBePromoted) const { | ||||
7255 | // We could support CastInst too. | ||||
7256 | return isa<BinaryOperator>(ToBePromoted); | ||||
7257 | } | ||||
7258 | |||||
7259 | /// Check if it is profitable to promote \p ToBePromoted | ||||
7260 | /// by moving downward the transition through. | ||||
7261 | bool shouldPromote(const Instruction *ToBePromoted) const { | ||||
7262 | // Promote only if all the operands can be statically expanded. | ||||
7263 | // Indeed, we do not want to introduce any new kind of transitions. | ||||
7264 | for (const Use &U : ToBePromoted->operands()) { | ||||
7265 | const Value *Val = U.get(); | ||||
7266 | if (Val == getEndOfTransition()) { | ||||
7267 | // If the use is a division and the transition is on the rhs, | ||||
7268 | // we cannot promote the operation, otherwise we may create a | ||||
7269 | // division by zero. | ||||
7270 | if (canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo())) | ||||
7271 | return false; | ||||
7272 | continue; | ||||
7273 | } | ||||
7274 | if (!isa<ConstantInt>(Val) && !isa<UndefValue>(Val) && | ||||
7275 | !isa<ConstantFP>(Val)) | ||||
7276 | return false; | ||||
7277 | } | ||||
7278 | // Check that the resulting operation is legal. | ||||
7279 | int ISDOpcode = TLI.InstructionOpcodeToISD(ToBePromoted->getOpcode()); | ||||
7280 | if (!ISDOpcode) | ||||
7281 | return false; | ||||
7282 | return StressStoreExtract || | ||||
7283 | TLI.isOperationLegalOrCustom( | ||||
7284 | ISDOpcode, TLI.getValueType(DL, getTransitionType(), true)); | ||||
7285 | } | ||||
7286 | |||||
7287 | /// Check whether or not \p Use can be combined | ||||
7288 | /// with the transition. | ||||
7289 | /// I.e., is it possible to do Use(Transition) => AnotherUse? | ||||
7290 | bool canCombine(const Instruction *Use) { return isa<StoreInst>(Use); } | ||||
7291 | |||||
7292 | /// Record \p ToBePromoted as part of the chain to be promoted. | ||||
7293 | void enqueueForPromotion(Instruction *ToBePromoted) { | ||||
7294 | InstsToBePromoted.push_back(ToBePromoted); | ||||
7295 | } | ||||
7296 | |||||
7297 | /// Set the instruction that will be combined with the transition. | ||||
7298 | void recordCombineInstruction(Instruction *ToBeCombined) { | ||||
7299 | assert(canCombine(ToBeCombined) && "Unsupported instruction to combine")((void)0); | ||||
7300 | CombineInst = ToBeCombined; | ||||
7301 | } | ||||
7302 | |||||
7303 | /// Promote all the instructions enqueued for promotion if it is | ||||
7304 | /// is profitable. | ||||
7305 | /// \return True if the promotion happened, false otherwise. | ||||
7306 | bool promote() { | ||||
7307 | // Check if there is something to promote. | ||||
7308 | // Right now, if we do not have anything to combine with, | ||||
7309 | // we assume the promotion is not profitable. | ||||
7310 | if (InstsToBePromoted.empty() || !CombineInst) | ||||
7311 | return false; | ||||
7312 | |||||
7313 | // Check cost. | ||||
7314 | if (!StressStoreExtract && !isProfitableToPromote()) | ||||
7315 | return false; | ||||
7316 | |||||
7317 | // Promote. | ||||
7318 | for (auto &ToBePromoted : InstsToBePromoted) | ||||
7319 | promoteImpl(ToBePromoted); | ||||
7320 | InstsToBePromoted.clear(); | ||||
7321 | return true; | ||||
7322 | } | ||||
7323 | }; | ||||
7324 | |||||
7325 | } // end anonymous namespace | ||||
7326 | |||||
7327 | void VectorPromoteHelper::promoteImpl(Instruction *ToBePromoted) { | ||||
7328 | // At this point, we know that all the operands of ToBePromoted but Def | ||||
7329 | // can be statically promoted. | ||||
7330 | // For Def, we need to use its parameter in ToBePromoted: | ||||
7331 | // b = ToBePromoted ty1 a | ||||
7332 | // Def = Transition ty1 b to ty2 | ||||
7333 | // Move the transition down. | ||||
7334 | // 1. Replace all uses of the promoted operation by the transition. | ||||
7335 | // = ... b => = ... Def. | ||||
7336 | assert(ToBePromoted->getType() == Transition->getType() &&((void)0) | ||||
7337 | "The type of the result of the transition does not match "((void)0) | ||||
7338 | "the final type")((void)0); | ||||
7339 | ToBePromoted->replaceAllUsesWith(Transition); | ||||
7340 | // 2. Update the type of the uses. | ||||
7341 | // b = ToBePromoted ty2 Def => b = ToBePromoted ty1 Def. | ||||
7342 | Type *TransitionTy = getTransitionType(); | ||||
7343 | ToBePromoted->mutateType(TransitionTy); | ||||
7344 | // 3. Update all the operands of the promoted operation with promoted | ||||
7345 | // operands. | ||||
7346 | // b = ToBePromoted ty1 Def => b = ToBePromoted ty1 a. | ||||
7347 | for (Use &U : ToBePromoted->operands()) { | ||||
7348 | Value *Val = U.get(); | ||||
7349 | Value *NewVal = nullptr; | ||||
7350 | if (Val == Transition) | ||||
7351 | NewVal = Transition->getOperand(getTransitionOriginalValueIdx()); | ||||
7352 | else if (isa<UndefValue>(Val) || isa<ConstantInt>(Val) || | ||||
7353 | isa<ConstantFP>(Val)) { | ||||
7354 | // Use a splat constant if it is not safe to use undef. | ||||
7355 | NewVal = getConstantVector( | ||||
7356 | cast<Constant>(Val), | ||||
7357 | isa<UndefValue>(Val) || | ||||
7358 | canCauseUndefinedBehavior(ToBePromoted, U.getOperandNo())); | ||||
7359 | } else | ||||
7360 | llvm_unreachable("Did you modified shouldPromote and forgot to update "__builtin_unreachable() | ||||
7361 | "this?")__builtin_unreachable(); | ||||
7362 | ToBePromoted->setOperand(U.getOperandNo(), NewVal); | ||||
7363 | } | ||||
7364 | Transition->moveAfter(ToBePromoted); | ||||
7365 | Transition->setOperand(getTransitionOriginalValueIdx(), ToBePromoted); | ||||
7366 | } | ||||
7367 | |||||
7368 | /// Some targets can do store(extractelement) with one instruction. | ||||
7369 | /// Try to push the extractelement towards the stores when the target | ||||
7370 | /// has this feature and this is profitable. | ||||
7371 | bool CodeGenPrepare::optimizeExtractElementInst(Instruction *Inst) { | ||||
7372 | unsigned CombineCost = std::numeric_limits<unsigned>::max(); | ||||
7373 | if (DisableStoreExtract || | ||||
7374 | (!StressStoreExtract && | ||||
7375 | !TLI->canCombineStoreAndExtract(Inst->getOperand(0)->getType(), | ||||
7376 | Inst->getOperand(1), CombineCost))) | ||||
7377 | return false; | ||||
7378 | |||||
7379 | // At this point we know that Inst is a vector to scalar transition. | ||||
7380 | // Try to move it down the def-use chain, until: | ||||
7381 | // - We can combine the transition with its single use | ||||
7382 | // => we got rid of the transition. | ||||
7383 | // - We escape the current basic block | ||||
7384 | // => we would need to check that we are moving it at a cheaper place and | ||||
7385 | // we do not do that for now. | ||||
7386 | BasicBlock *Parent = Inst->getParent(); | ||||
7387 | LLVM_DEBUG(dbgs() << "Found an interesting transition: " << *Inst << '\n')do { } while (false); | ||||
7388 | VectorPromoteHelper VPH(*DL, *TLI, *TTI, Inst, CombineCost); | ||||
7389 | // If the transition has more than one use, assume this is not going to be | ||||
7390 | // beneficial. | ||||
7391 | while (Inst->hasOneUse()) { | ||||
7392 | Instruction *ToBePromoted = cast<Instruction>(*Inst->user_begin()); | ||||
7393 | LLVM_DEBUG(dbgs() << "Use: " << *ToBePromoted << '\n')do { } while (false); | ||||
7394 | |||||
7395 | if (ToBePromoted->getParent() != Parent) { | ||||
7396 | LLVM_DEBUG(dbgs() << "Instruction to promote is in a different block ("do { } while (false) | ||||
7397 | << ToBePromoted->getParent()->getName()do { } while (false) | ||||
7398 | << ") than the transition (" << Parent->getName()do { } while (false) | ||||
7399 | << ").\n")do { } while (false); | ||||
7400 | return false; | ||||
7401 | } | ||||
7402 | |||||
7403 | if (VPH.canCombine(ToBePromoted)) { | ||||
7404 | LLVM_DEBUG(dbgs() << "Assume " << *Inst << '\n'do { } while (false) | ||||
7405 | << "will be combined with: " << *ToBePromoted << '\n')do { } while (false); | ||||
7406 | VPH.recordCombineInstruction(ToBePromoted); | ||||
7407 | bool Changed = VPH.promote(); | ||||
7408 | NumStoreExtractExposed += Changed; | ||||
7409 | return Changed; | ||||
7410 | } | ||||
7411 | |||||
7412 | LLVM_DEBUG(dbgs() << "Try promoting.\n")do { } while (false); | ||||
7413 | if (!VPH.canPromote(ToBePromoted) || !VPH.shouldPromote(ToBePromoted)) | ||||
7414 | return false; | ||||
7415 | |||||
7416 | LLVM_DEBUG(dbgs() << "Promoting is possible... Enqueue for promotion!\n")do { } while (false); | ||||
7417 | |||||
7418 | VPH.enqueueForPromotion(ToBePromoted); | ||||
7419 | Inst = ToBePromoted; | ||||
7420 | } | ||||
7421 | return false; | ||||
7422 | } | ||||
7423 | |||||
7424 | /// For the instruction sequence of store below, F and I values | ||||
7425 | /// are bundled together as an i64 value before being stored into memory. | ||||
7426 | /// Sometimes it is more efficient to generate separate stores for F and I, | ||||
7427 | /// which can remove the bitwise instructions or sink them to colder places. | ||||
7428 | /// | ||||
7429 | /// (store (or (zext (bitcast F to i32) to i64), | ||||
7430 | /// (shl (zext I to i64), 32)), addr) --> | ||||
7431 | /// (store F, addr) and (store I, addr+4) | ||||
7432 | /// | ||||
7433 | /// Similarly, splitting for other merged store can also be beneficial, like: | ||||
7434 | /// For pair of {i32, i32}, i64 store --> two i32 stores. | ||||
7435 | /// For pair of {i32, i16}, i64 store --> two i32 stores. | ||||
7436 | /// For pair of {i16, i16}, i32 store --> two i16 stores. | ||||
7437 | /// For pair of {i16, i8}, i32 store --> two i16 stores. | ||||
7438 | /// For pair of {i8, i8}, i16 store --> two i8 stores. | ||||
7439 | /// | ||||
7440 | /// We allow each target to determine specifically which kind of splitting is | ||||
7441 | /// supported. | ||||
7442 | /// | ||||
7443 | /// The store patterns are commonly seen from the simple code snippet below | ||||
7444 | /// if only std::make_pair(...) is sroa transformed before inlined into hoo. | ||||
7445 | /// void goo(const std::pair<int, float> &); | ||||
7446 | /// hoo() { | ||||
7447 | /// ... | ||||
7448 | /// goo(std::make_pair(tmp, ftmp)); | ||||
7449 | /// ... | ||||
7450 | /// } | ||||
7451 | /// | ||||
7452 | /// Although we already have similar splitting in DAG Combine, we duplicate | ||||
7453 | /// it in CodeGenPrepare to catch the case in which pattern is across | ||||
7454 | /// multiple BBs. The logic in DAG Combine is kept to catch case generated | ||||
7455 | /// during code expansion. | ||||
7456 | static bool splitMergedValStore(StoreInst &SI, const DataLayout &DL, | ||||
7457 | const TargetLowering &TLI) { | ||||
7458 | // Handle simple but common cases only. | ||||
7459 | Type *StoreType = SI.getValueOperand()->getType(); | ||||
7460 | |||||
7461 | // The code below assumes shifting a value by <number of bits>, | ||||
7462 | // whereas scalable vectors would have to be shifted by | ||||
7463 | // <2log(vscale) + number of bits> in order to store the | ||||
7464 | // low/high parts. Bailing out for now. | ||||
7465 | if (isa<ScalableVectorType>(StoreType)) | ||||
7466 | return false; | ||||
7467 | |||||
7468 | if (!DL.typeSizeEqualsStoreSize(StoreType) || | ||||
7469 | DL.getTypeSizeInBits(StoreType) == 0) | ||||
7470 | return false; | ||||
7471 | |||||
7472 | unsigned HalfValBitSize = DL.getTypeSizeInBits(StoreType) / 2; | ||||
7473 | Type *SplitStoreType = Type::getIntNTy(SI.getContext(), HalfValBitSize); | ||||
7474 | if (!DL.typeSizeEqualsStoreSize(SplitStoreType)) | ||||
7475 | return false; | ||||
7476 | |||||
7477 | // Don't split the store if it is volatile. | ||||
7478 | if (SI.isVolatile()) | ||||
7479 | return false; | ||||
7480 | |||||
7481 | // Match the following patterns: | ||||
7482 | // (store (or (zext LValue to i64), | ||||
7483 | // (shl (zext HValue to i64), 32)), HalfValBitSize) | ||||
7484 | // or | ||||
7485 | // (store (or (shl (zext HValue to i64), 32)), HalfValBitSize) | ||||
7486 | // (zext LValue to i64), | ||||
7487 | // Expect both operands of OR and the first operand of SHL have only | ||||
7488 | // one use. | ||||
7489 | Value *LValue, *HValue; | ||||
7490 | if (!match(SI.getValueOperand(), | ||||
7491 | m_c_Or(m_OneUse(m_ZExt(m_Value(LValue))), | ||||
7492 | m_OneUse(m_Shl(m_OneUse(m_ZExt(m_Value(HValue))), | ||||
7493 | m_SpecificInt(HalfValBitSize)))))) | ||||
7494 | return false; | ||||
7495 | |||||
7496 | // Check LValue and HValue are int with size less or equal than 32. | ||||
7497 | if (!LValue->getType()->isIntegerTy() || | ||||
7498 | DL.getTypeSizeInBits(LValue->getType()) > HalfValBitSize || | ||||
7499 | !HValue->getType()->isIntegerTy() || | ||||
7500 | DL.getTypeSizeInBits(HValue->getType()) > HalfValBitSize) | ||||
7501 | return false; | ||||
7502 | |||||
7503 | // If LValue/HValue is a bitcast instruction, use the EVT before bitcast | ||||
7504 | // as the input of target query. | ||||
7505 | auto *LBC = dyn_cast<BitCastInst>(LValue); | ||||
7506 | auto *HBC = dyn_cast<BitCastInst>(HValue); | ||||
7507 | EVT LowTy = LBC ? EVT::getEVT(LBC->getOperand(0)->getType()) | ||||
7508 | : EVT::getEVT(LValue->getType()); | ||||
7509 | EVT HighTy = HBC ? EVT::getEVT(HBC->getOperand(0)->getType()) | ||||
7510 | : EVT::getEVT(HValue->getType()); | ||||
7511 | if (!ForceSplitStore && !TLI.isMultiStoresCheaperThanBitsMerge(LowTy, HighTy)) | ||||
7512 | return false; | ||||
7513 | |||||
7514 | // Start to split store. | ||||
7515 | IRBuilder<> Builder(SI.getContext()); | ||||
7516 | Builder.SetInsertPoint(&SI); | ||||
7517 | |||||
7518 | // If LValue/HValue is a bitcast in another BB, create a new one in current | ||||
7519 | // BB so it may be merged with the splitted stores by dag combiner. | ||||
7520 | if (LBC && LBC->getParent() != SI.getParent()) | ||||
7521 | LValue = Builder.CreateBitCast(LBC->getOperand(0), LBC->getType()); | ||||
7522 | if (HBC && HBC->getParent() != SI.getParent()) | ||||
7523 | HValue = Builder.CreateBitCast(HBC->getOperand(0), HBC->getType()); | ||||
7524 | |||||
7525 | bool IsLE = SI.getModule()->getDataLayout().isLittleEndian(); | ||||
7526 | auto CreateSplitStore = [&](Value *V, bool Upper) { | ||||
7527 | V = Builder.CreateZExtOrBitCast(V, SplitStoreType); | ||||
7528 | Value *Addr = Builder.CreateBitCast( | ||||
7529 | SI.getOperand(1), | ||||
7530 | SplitStoreType->getPointerTo(SI.getPointerAddressSpace())); | ||||
7531 | Align Alignment = SI.getAlign(); | ||||
7532 | const bool IsOffsetStore = (IsLE && Upper) || (!IsLE && !Upper); | ||||
| |||||
7533 | if (IsOffsetStore
| ||||
7534 | Addr = Builder.CreateGEP( | ||||
7535 | SplitStoreType, Addr, | ||||
7536 | ConstantInt::get(Type::getInt32Ty(SI.getContext()), 1)); | ||||
7537 | |||||
7538 | // When splitting the store in half, naturally one half will retain the | ||||
7539 | // alignment of the original wider store, regardless of whether it was | ||||
7540 | // over-aligned or not, while the other will require adjustment. | ||||
7541 | Alignment = commonAlignment(Alignment, HalfValBitSize / 8); | ||||
7542 | } | ||||
7543 | Builder.CreateAlignedStore(V, Addr, Alignment); | ||||
7544 | }; | ||||
7545 | |||||
7546 | CreateSplitStore(LValue, false); | ||||
7547 | CreateSplitStore(HValue, true); | ||||
7548 | |||||
7549 | // Delete the old store. | ||||
7550 | SI.eraseFromParent(); | ||||
7551 | return true; | ||||
7552 | } | ||||
7553 | |||||
7554 | // Return true if the GEP has two operands, the first operand is of a sequential | ||||
7555 | // type, and the second operand is a constant. | ||||
7556 | static bool GEPSequentialConstIndexed(GetElementPtrInst *GEP) { | ||||
7557 | gep_type_iterator I = gep_type_begin(*GEP); | ||||
7558 | return GEP->getNumOperands() == 2 && | ||||
7559 | I.isSequential() && | ||||
7560 | isa<ConstantInt>(GEP->getOperand(1)); | ||||
7561 | } | ||||
7562 | |||||
7563 | // Try unmerging GEPs to reduce liveness interference (register pressure) across | ||||
7564 | // IndirectBr edges. Since IndirectBr edges tend to touch on many blocks, | ||||
7565 | // reducing liveness interference across those edges benefits global register | ||||
7566 | // allocation. Currently handles only certain cases. | ||||
7567 | // | ||||
7568 | // For example, unmerge %GEPI and %UGEPI as below. | ||||
7569 | // | ||||
7570 | // ---------- BEFORE ---------- | ||||
7571 | // SrcBlock: | ||||
7572 | // ... | ||||
7573 | // %GEPIOp = ... | ||||
7574 | // ... | ||||
7575 | // %GEPI = gep %GEPIOp, Idx | ||||
7576 | // ... | ||||
7577 | // indirectbr ... [ label %DstB0, label %DstB1, ... label %DstBi ... ] | ||||
7578 | // (* %GEPI is alive on the indirectbr edges due to other uses ahead) | ||||
7579 | // (* %GEPIOp is alive on the indirectbr edges only because of it's used by | ||||
7580 | // %UGEPI) | ||||
7581 | // | ||||
7582 | // DstB0: ... (there may be a gep similar to %UGEPI to be unmerged) | ||||
7583 | // DstB1: ... (there may be a gep similar to %UGEPI to be unmerged) | ||||
7584 | // ... | ||||
7585 | // | ||||
7586 | // DstBi: | ||||
7587 | // ... | ||||
7588 | // %UGEPI = gep %GEPIOp, UIdx | ||||
7589 | // ... | ||||
7590 | // --------------------------- | ||||
7591 | // | ||||
7592 | // ---------- AFTER ---------- | ||||
7593 | // SrcBlock: | ||||
7594 | // ... (same as above) | ||||
7595 | // (* %GEPI is still alive on the indirectbr edges) | ||||
7596 | // (* %GEPIOp is no longer alive on the indirectbr edges as a result of the | ||||
7597 | // unmerging) | ||||
7598 | // ... | ||||
7599 | // | ||||
7600 | // DstBi: | ||||
7601 | // ... | ||||
7602 | // %UGEPI = gep %GEPI, (UIdx-Idx) | ||||
7603 | // ... | ||||
7604 | // --------------------------- | ||||
7605 | // | ||||
7606 | // The register pressure on the IndirectBr edges is reduced because %GEPIOp is | ||||
7607 | // no longer alive on them. | ||||
7608 | // | ||||
7609 | // We try to unmerge GEPs here in CodGenPrepare, as opposed to limiting merging | ||||
7610 | // of GEPs in the first place in InstCombiner::visitGetElementPtrInst() so as | ||||
7611 | // not to disable further simplications and optimizations as a result of GEP | ||||
7612 | // merging. | ||||
7613 | // | ||||
7614 | // Note this unmerging may increase the length of the data flow critical path | ||||
7615 | // (the path from %GEPIOp to %UGEPI would go through %GEPI), which is a tradeoff | ||||
7616 | // between the register pressure and the length of data-flow critical | ||||
7617 | // path. Restricting this to the uncommon IndirectBr case would minimize the | ||||
7618 | // impact of potentially longer critical path, if any, and the impact on compile | ||||
7619 | // time. | ||||
7620 | static bool tryUnmergingGEPsAcrossIndirectBr(GetElementPtrInst *GEPI, | ||||
7621 | const TargetTransformInfo *TTI) { | ||||
7622 | BasicBlock *SrcBlock = GEPI->getParent(); | ||||
7623 | // Check that SrcBlock ends with an IndirectBr. If not, give up. The common | ||||
7624 | // (non-IndirectBr) cases exit early here. | ||||
7625 | if (!isa<IndirectBrInst>(SrcBlock->getTerminator())) | ||||
7626 | return false; | ||||
7627 | // Check that GEPI is a simple gep with a single constant index. | ||||
7628 | if (!GEPSequentialConstIndexed(GEPI)) | ||||
7629 | return false; | ||||
7630 | ConstantInt *GEPIIdx = cast<ConstantInt>(GEPI->getOperand(1)); | ||||
7631 | // Check that GEPI is a cheap one. | ||||
7632 | if (TTI->getIntImmCost(GEPIIdx->getValue(), GEPIIdx->getType(), | ||||
7633 | TargetTransformInfo::TCK_SizeAndLatency) | ||||
7634 | > TargetTransformInfo::TCC_Basic) | ||||
7635 | return false; | ||||
7636 | Value *GEPIOp = GEPI->getOperand(0); | ||||
7637 | // Check that GEPIOp is an instruction that's also defined in SrcBlock. | ||||
7638 | if (!isa<Instruction>(GEPIOp)) | ||||
7639 | return false; | ||||
7640 | auto *GEPIOpI = cast<Instruction>(GEPIOp); | ||||
7641 | if (GEPIOpI->getParent() != SrcBlock) | ||||
7642 | return false; | ||||
7643 | // Check that GEP is used outside the block, meaning it's alive on the | ||||
7644 | // IndirectBr edge(s). | ||||
7645 | if (find_if(GEPI->users(), [&](User *Usr) { | ||||
7646 | if (auto *I = dyn_cast<Instruction>(Usr)) { | ||||
7647 | if (I->getParent() != SrcBlock) { | ||||
7648 | return true; | ||||
7649 | } | ||||
7650 | } | ||||
7651 | return false; | ||||
7652 | }) == GEPI->users().end()) | ||||
7653 | return false; | ||||
7654 | // The second elements of the GEP chains to be unmerged. | ||||
7655 | std::vector<GetElementPtrInst *> UGEPIs; | ||||
7656 | // Check each user of GEPIOp to check if unmerging would make GEPIOp not alive | ||||
7657 | // on IndirectBr edges. | ||||
7658 | for (User *Usr : GEPIOp->users()) { | ||||
7659 | if (Usr == GEPI) continue; | ||||
7660 | // Check if Usr is an Instruction. If not, give up. | ||||
7661 | if (!isa<Instruction>(Usr)) | ||||
7662 | return false; | ||||
7663 | auto *UI = cast<Instruction>(Usr); | ||||
7664 | // Check if Usr in the same block as GEPIOp, which is fine, skip. | ||||
7665 | if (UI->getParent() == SrcBlock) | ||||
7666 | continue; | ||||
7667 | // Check if Usr is a GEP. If not, give up. | ||||
7668 | if (!isa<GetElementPtrInst>(Usr)) | ||||
7669 | return false; | ||||
7670 | auto *UGEPI = cast<GetElementPtrInst>(Usr); | ||||
7671 | // Check if UGEPI is a simple gep with a single constant index and GEPIOp is | ||||
7672 | // the pointer operand to it. If so, record it in the vector. If not, give | ||||
7673 | // up. | ||||
7674 | if (!GEPSequentialConstIndexed(UGEPI)) | ||||
7675 | return false; | ||||
7676 | if (UGEPI->getOperand(0) != GEPIOp) | ||||
7677 | return false; | ||||
7678 | if (GEPIIdx->getType() != | ||||
7679 | cast<ConstantInt>(UGEPI->getOperand(1))->getType()) | ||||
7680 | return false; | ||||
7681 | ConstantInt *UGEPIIdx = cast<ConstantInt>(UGEPI->getOperand(1)); | ||||
7682 | if (TTI->getIntImmCost(UGEPIIdx->getValue(), UGEPIIdx->getType(), | ||||
7683 | TargetTransformInfo::TCK_SizeAndLatency) | ||||
7684 | > TargetTransformInfo::TCC_Basic) | ||||
7685 | return false; | ||||
7686 | UGEPIs.push_back(UGEPI); | ||||
7687 | } | ||||
7688 | if (UGEPIs.size() == 0) | ||||
7689 | return false; | ||||
7690 | // Check the materializing cost of (Uidx-Idx). | ||||
7691 | for (GetElementPtrInst *UGEPI : UGEPIs) { | ||||
7692 | ConstantInt *UGEPIIdx = cast<ConstantInt>(UGEPI->getOperand(1)); | ||||
7693 | APInt NewIdx = UGEPIIdx->getValue() - GEPIIdx->getValue(); | ||||
7694 | InstructionCost ImmCost = TTI->getIntImmCost( | ||||
7695 | NewIdx, GEPIIdx->getType(), TargetTransformInfo::TCK_SizeAndLatency); | ||||
7696 | if (ImmCost > TargetTransformInfo::TCC_Basic) | ||||
7697 | return false; | ||||
7698 | } | ||||
7699 | // Now unmerge between GEPI and UGEPIs. | ||||
7700 | for (GetElementPtrInst *UGEPI : UGEPIs) { | ||||
7701 | UGEPI->setOperand(0, GEPI); | ||||
7702 | ConstantInt *UGEPIIdx = cast<ConstantInt>(UGEPI->getOperand(1)); | ||||
7703 | Constant *NewUGEPIIdx = | ||||
7704 | ConstantInt::get(GEPIIdx->getType(), | ||||
7705 | UGEPIIdx->getValue() - GEPIIdx->getValue()); | ||||
7706 | UGEPI->setOperand(1, NewUGEPIIdx); | ||||
7707 | // If GEPI is not inbounds but UGEPI is inbounds, change UGEPI to not | ||||
7708 | // inbounds to avoid UB. | ||||
7709 | if (!GEPI->isInBounds()) { | ||||
7710 | UGEPI->setIsInBounds(false); | ||||
7711 | } | ||||
7712 | } | ||||
7713 | // After unmerging, verify that GEPIOp is actually only used in SrcBlock (not | ||||
7714 | // alive on IndirectBr edges). | ||||
7715 | assert(find_if(GEPIOp->users(), [&](User *Usr) {((void)0) | ||||
7716 | return cast<Instruction>(Usr)->getParent() != SrcBlock;((void)0) | ||||
7717 | }) == GEPIOp->users().end() && "GEPIOp is used outside SrcBlock")((void)0); | ||||
7718 | return true; | ||||
7719 | } | ||||
7720 | |||||
7721 | static bool optimizeBranch(BranchInst *Branch, const TargetLowering &TLI) { | ||||
7722 | // Try and convert | ||||
7723 | // %c = icmp ult %x, 8 | ||||
7724 | // br %c, bla, blb | ||||
7725 | // %tc = lshr %x, 3 | ||||
7726 | // to | ||||
7727 | // %tc = lshr %x, 3 | ||||
7728 | // %c = icmp eq %tc, 0 | ||||
7729 | // br %c, bla, blb | ||||
7730 | // Creating the cmp to zero can be better for the backend, especially if the | ||||
7731 | // lshr produces flags that can be used automatically. | ||||
7732 | if (!TLI.preferZeroCompareBranch() || !Branch->isConditional()) | ||||
7733 | return false; | ||||
7734 | |||||
7735 | ICmpInst *Cmp = dyn_cast<ICmpInst>(Branch->getCondition()); | ||||
7736 | if (!Cmp || !isa<ConstantInt>(Cmp->getOperand(1)) || !Cmp->hasOneUse()) | ||||
7737 | return false; | ||||
7738 | |||||
7739 | Value *X = Cmp->getOperand(0); | ||||
7740 | APInt CmpC = cast<ConstantInt>(Cmp->getOperand(1))->getValue(); | ||||
7741 | |||||
7742 | for (auto *U : X->users()) { | ||||
7743 | Instruction *UI = dyn_cast<Instruction>(U); | ||||
7744 | // A quick dominance check | ||||
7745 | if (!UI || | ||||
7746 | (UI->getParent() != Branch->getParent() && | ||||
7747 | UI->getParent() != Branch->getSuccessor(0) && | ||||
7748 | UI->getParent() != Branch->getSuccessor(1)) || | ||||
7749 | (UI->getParent() != Branch->getParent() && | ||||
7750 | !UI->getParent()->getSinglePredecessor())) | ||||
7751 | continue; | ||||
7752 | |||||
7753 | if (CmpC.isPowerOf2() && Cmp->getPredicate() == ICmpInst::ICMP_ULT && | ||||
7754 | match(UI, m_Shr(m_Specific(X), m_SpecificInt(CmpC.logBase2())))) { | ||||
7755 | IRBuilder<> Builder(Branch); | ||||
7756 | if (UI->getParent() != Branch->getParent()) | ||||
7757 | UI->moveBefore(Branch); | ||||
7758 | Value *NewCmp = Builder.CreateCmp(ICmpInst::ICMP_EQ, UI, | ||||
7759 | ConstantInt::get(UI->getType(), 0)); | ||||
7760 | LLVM_DEBUG(dbgs() << "Converting " << *Cmp << "\n")do { } while (false); | ||||
7761 | LLVM_DEBUG(dbgs() << " to compare on zero: " << *NewCmp << "\n")do { } while (false); | ||||
7762 | Cmp->replaceAllUsesWith(NewCmp); | ||||
7763 | return true; | ||||
7764 | } | ||||
7765 | if (Cmp->isEquality() && | ||||
7766 | (match(UI, m_Add(m_Specific(X), m_SpecificInt(-CmpC))) || | ||||
7767 | match(UI, m_Sub(m_Specific(X), m_SpecificInt(CmpC))))) { | ||||
7768 | IRBuilder<> Builder(Branch); | ||||
7769 | if (UI->getParent() != Branch->getParent()) | ||||
7770 | UI->moveBefore(Branch); | ||||
7771 | Value *NewCmp = Builder.CreateCmp(Cmp->getPredicate(), UI, | ||||
7772 | ConstantInt::get(UI->getType(), 0)); | ||||
7773 | LLVM_DEBUG(dbgs() << "Converting " << *Cmp << "\n")do { } while (false); | ||||
7774 | LLVM_DEBUG(dbgs() << " to compare on zero: " << *NewCmp << "\n")do { } while (false); | ||||
7775 | Cmp->replaceAllUsesWith(NewCmp); | ||||
7776 | return true; | ||||
7777 | } | ||||
7778 | } | ||||
7779 | return false; | ||||
7780 | } | ||||
7781 | |||||
7782 | bool CodeGenPrepare::optimizeInst(Instruction *I, bool &ModifiedDT) { | ||||
7783 | // Bail out if we inserted the instruction to prevent optimizations from | ||||
7784 | // stepping on each other's toes. | ||||
7785 | if (InsertedInsts.count(I)) | ||||
7786 | return false; | ||||
7787 | |||||
7788 | // TODO: Move into the switch on opcode below here. | ||||
7789 | if (PHINode *P = dyn_cast<PHINode>(I)) { | ||||
7790 | // It is possible for very late stage optimizations (such as SimplifyCFG) | ||||
7791 | // to introduce PHI nodes too late to be cleaned up. If we detect such a | ||||
7792 | // trivial PHI, go ahead and zap it here. | ||||
7793 | if (Value *V = SimplifyInstruction(P, {*DL, TLInfo})) { | ||||
7794 | LargeOffsetGEPMap.erase(P); | ||||
7795 | P->replaceAllUsesWith(V); | ||||
7796 | P->eraseFromParent(); | ||||
7797 | ++NumPHIsElim; | ||||
7798 | return true; | ||||
7799 | } | ||||
7800 | return false; | ||||
7801 | } | ||||
7802 | |||||
7803 | if (CastInst *CI = dyn_cast<CastInst>(I)) { | ||||
7804 | // If the source of the cast is a constant, then this should have | ||||
7805 | // already been constant folded. The only reason NOT to constant fold | ||||
7806 | // it is if something (e.g. LSR) was careful to place the constant | ||||
7807 | // evaluation in a block other than then one that uses it (e.g. to hoist | ||||
7808 | // the address of globals out of a loop). If this is the case, we don't | ||||
7809 | // want to forward-subst the cast. | ||||
7810 | if (isa<Constant>(CI->getOperand(0))) | ||||
7811 | return false; | ||||
7812 | |||||
7813 | if (OptimizeNoopCopyExpression(CI, *TLI, *DL)) | ||||
7814 | return true; | ||||
7815 | |||||
7816 | if (isa<ZExtInst>(I) || isa<SExtInst>(I)) { | ||||
7817 | /// Sink a zext or sext into its user blocks if the target type doesn't | ||||
7818 | /// fit in one register | ||||
7819 | if (TLI->getTypeAction(CI->getContext(), | ||||
7820 | TLI->getValueType(*DL, CI->getType())) == | ||||
7821 | TargetLowering::TypeExpandInteger) { | ||||
7822 | return SinkCast(CI); | ||||
7823 | } else { | ||||
7824 | bool MadeChange = optimizeExt(I); | ||||
7825 | return MadeChange | optimizeExtUses(I); | ||||
7826 | } | ||||
7827 | } | ||||
7828 | return false; | ||||
7829 | } | ||||
7830 | |||||
7831 | if (auto *Cmp = dyn_cast<CmpInst>(I)) | ||||
7832 | if (optimizeCmp(Cmp, ModifiedDT)) | ||||
7833 | return true; | ||||
7834 | |||||
7835 | if (LoadInst *LI = dyn_cast<LoadInst>(I)) { | ||||
7836 | LI->setMetadata(LLVMContext::MD_invariant_group, nullptr); | ||||
7837 | bool Modified = optimizeLoadExt(LI); | ||||
7838 | unsigned AS = LI->getPointerAddressSpace(); | ||||
7839 | Modified |= optimizeMemoryInst(I, I->getOperand(0), LI->getType(), AS); | ||||
7840 | return Modified; | ||||
7841 | } | ||||
7842 | |||||
7843 | if (StoreInst *SI = dyn_cast<StoreInst>(I)) { | ||||
7844 | if (splitMergedValStore(*SI, *DL, *TLI)) | ||||
7845 | return true; | ||||
7846 | SI->setMetadata(LLVMContext::MD_invariant_group, nullptr); | ||||
7847 | unsigned AS = SI->getPointerAddressSpace(); | ||||
7848 | return optimizeMemoryInst(I, SI->getOperand(1), | ||||
7849 | SI->getOperand(0)->getType(), AS); | ||||
7850 | } | ||||
7851 | |||||
7852 | if (AtomicRMWInst *RMW = dyn_cast<AtomicRMWInst>(I)) { | ||||
7853 | unsigned AS = RMW->getPointerAddressSpace(); | ||||
7854 | return optimizeMemoryInst(I, RMW->getPointerOperand(), | ||||
7855 | RMW->getType(), AS); | ||||
7856 | } | ||||
7857 | |||||
7858 | if (AtomicCmpXchgInst *CmpX = dyn_cast<AtomicCmpXchgInst>(I)) { | ||||
7859 | unsigned AS = CmpX->getPointerAddressSpace(); | ||||
7860 | return optimizeMemoryInst(I, CmpX->getPointerOperand(), | ||||
7861 | CmpX->getCompareOperand()->getType(), AS); | ||||
7862 | } | ||||
7863 | |||||
7864 | BinaryOperator *BinOp = dyn_cast<BinaryOperator>(I); | ||||
7865 | |||||
7866 | if (BinOp && (BinOp->getOpcode() == Instruction::And) && EnableAndCmpSinking) | ||||
7867 | return sinkAndCmp0Expression(BinOp, *TLI, InsertedInsts); | ||||
7868 | |||||
7869 | // TODO: Move this into the switch on opcode - it handles shifts already. | ||||
7870 | if (BinOp && (BinOp->getOpcode() == Instruction::AShr || | ||||
7871 | BinOp->getOpcode() == Instruction::LShr)) { | ||||
7872 | ConstantInt *CI = dyn_cast<ConstantInt>(BinOp->getOperand(1)); | ||||
7873 | if (CI && TLI->hasExtractBitsInsn()) | ||||
7874 | if (OptimizeExtractBits(BinOp, CI, *TLI, *DL)) | ||||
7875 | return true; | ||||
7876 | } | ||||
7877 | |||||
7878 | if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) { | ||||
7879 | if (GEPI->hasAllZeroIndices()) { | ||||
7880 | /// The GEP operand must be a pointer, so must its result -> BitCast | ||||
7881 | Instruction *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(), | ||||
7882 | GEPI->getName(), GEPI); | ||||
7883 | NC->setDebugLoc(GEPI->getDebugLoc()); | ||||
7884 | GEPI->replaceAllUsesWith(NC); | ||||
7885 | GEPI->eraseFromParent(); | ||||
7886 | ++NumGEPsElim; | ||||
7887 | optimizeInst(NC, ModifiedDT); | ||||
7888 | return true; | ||||
7889 | } | ||||
7890 | if (tryUnmergingGEPsAcrossIndirectBr(GEPI, TTI)) { | ||||
7891 | return true; | ||||
7892 | } | ||||
7893 | return false; | ||||
7894 | } | ||||
7895 | |||||
7896 | if (FreezeInst *FI = dyn_cast<FreezeInst>(I)) { | ||||
7897 | // freeze(icmp a, const)) -> icmp (freeze a), const | ||||
7898 | // This helps generate efficient conditional jumps. | ||||
7899 | Instruction *CmpI = nullptr; | ||||
7900 | if (ICmpInst *II = dyn_cast<ICmpInst>(FI->getOperand(0))) | ||||
7901 | CmpI = II; | ||||
7902 | else if (FCmpInst *F = dyn_cast<FCmpInst>(FI->getOperand(0))) | ||||
7903 | CmpI = F->getFastMathFlags().none() ? F : nullptr; | ||||
7904 | |||||
7905 | if (CmpI && CmpI->hasOneUse()) { | ||||
7906 | auto Op0 = CmpI->getOperand(0), Op1 = CmpI->getOperand(1); | ||||
7907 | bool Const0 = isa<ConstantInt>(Op0) || isa<ConstantFP>(Op0) || | ||||
7908 | isa<ConstantPointerNull>(Op0); | ||||
7909 | bool Const1 = isa<ConstantInt>(Op1) || isa<ConstantFP>(Op1) || | ||||
7910 | isa<ConstantPointerNull>(Op1); | ||||
7911 | if (Const0 || Const1) { | ||||
7912 | if (!Const0 || !Const1) { | ||||
7913 | auto *F = new FreezeInst(Const0 ? Op1 : Op0, "", CmpI); | ||||
7914 | F->takeName(FI); | ||||
7915 | CmpI->setOperand(Const0 ? 1 : 0, F); | ||||
7916 | } | ||||
7917 | FI->replaceAllUsesWith(CmpI); | ||||
7918 | FI->eraseFromParent(); | ||||
7919 | return true; | ||||
7920 | } | ||||
7921 | } | ||||
7922 | return false; | ||||
7923 | } | ||||
7924 | |||||
7925 | if (tryToSinkFreeOperands(I)) | ||||
7926 | return true; | ||||
7927 | |||||
7928 | switch (I->getOpcode()) { | ||||
7929 | case Instruction::Shl: | ||||
7930 | case Instruction::LShr: | ||||
7931 | case Instruction::AShr: | ||||
7932 | return optimizeShiftInst(cast<BinaryOperator>(I)); | ||||
7933 | case Instruction::Call: | ||||
7934 | return optimizeCallInst(cast<CallInst>(I), ModifiedDT); | ||||
7935 | case Instruction::Select: | ||||
7936 | return optimizeSelectInst(cast<SelectInst>(I)); | ||||
7937 | case Instruction::ShuffleVector: | ||||
7938 | return optimizeShuffleVectorInst(cast<ShuffleVectorInst>(I)); | ||||
7939 | case Instruction::Switch: | ||||
7940 | return optimizeSwitchInst(cast<SwitchInst>(I)); | ||||
7941 | case Instruction::ExtractElement: | ||||
7942 | return optimizeExtractElementInst(cast<ExtractElementInst>(I)); | ||||
7943 | case Instruction::Br: | ||||
7944 | return optimizeBranch(cast<BranchInst>(I), *TLI); | ||||
7945 | } | ||||
7946 | |||||
7947 | return false; | ||||
7948 | } | ||||
7949 | |||||
7950 | /// Given an OR instruction, check to see if this is a bitreverse | ||||
7951 | /// idiom. If so, insert the new intrinsic and return true. | ||||
7952 | bool CodeGenPrepare::makeBitReverse(Instruction &I) { | ||||
7953 | if (!I.getType()->isIntegerTy() || | ||||
7954 | !TLI->isOperationLegalOrCustom(ISD::BITREVERSE, | ||||
7955 | TLI->getValueType(*DL, I.getType(), true))) | ||||
7956 | return false; | ||||
7957 | |||||
7958 | SmallVector<Instruction*, 4> Insts; | ||||
7959 | if (!recognizeBSwapOrBitReverseIdiom(&I, false, true, Insts)) | ||||
7960 | return false; | ||||
7961 | Instruction *LastInst = Insts.back(); | ||||
7962 | I.replaceAllUsesWith(LastInst); | ||||
7963 | RecursivelyDeleteTriviallyDeadInstructions( | ||||
7964 | &I, TLInfo, nullptr, [&](Value *V) { removeAllAssertingVHReferences(V); }); | ||||
7965 | return true; | ||||
7966 | } | ||||
7967 | |||||
7968 | // In this pass we look for GEP and cast instructions that are used | ||||
7969 | // across basic blocks and rewrite them to improve basic-block-at-a-time | ||||
7970 | // selection. | ||||
7971 | bool CodeGenPrepare::optimizeBlock(BasicBlock &BB, bool &ModifiedDT) { | ||||
7972 | SunkAddrs.clear(); | ||||
7973 | bool MadeChange = false; | ||||
7974 | |||||
7975 | CurInstIterator = BB.begin(); | ||||
7976 | while (CurInstIterator != BB.end()) { | ||||
7977 | MadeChange |= optimizeInst(&*CurInstIterator++, ModifiedDT); | ||||
7978 | if (ModifiedDT) | ||||
7979 | return true; | ||||
7980 | } | ||||
7981 | |||||
7982 | bool MadeBitReverse = true; | ||||
7983 | while (MadeBitReverse) { | ||||
7984 | MadeBitReverse = false; | ||||
7985 | for (auto &I : reverse(BB)) { | ||||
7986 | if (makeBitReverse(I)) { | ||||
7987 | MadeBitReverse = MadeChange = true; | ||||
7988 | break; | ||||
7989 | } | ||||
7990 | } | ||||
7991 | } | ||||
7992 | MadeChange |= dupRetToEnableTailCallOpts(&BB, ModifiedDT); | ||||
7993 | |||||
7994 | return MadeChange; | ||||
7995 | } | ||||
7996 | |||||
7997 | // Some CGP optimizations may move or alter what's computed in a block. Check | ||||
7998 | // whether a dbg.value intrinsic could be pointed at a more appropriate operand. | ||||
7999 | bool CodeGenPrepare::fixupDbgValue(Instruction *I) { | ||||
8000 | assert(isa<DbgValueInst>(I))((void)0); | ||||
8001 | DbgValueInst &DVI = *cast<DbgValueInst>(I); | ||||
8002 | |||||
8003 | // Does this dbg.value refer to a sunk address calculation? | ||||
8004 | bool AnyChange = false; | ||||
8005 | SmallDenseSet<Value *> LocationOps(DVI.location_ops().begin(), | ||||
8006 | DVI.location_ops().end()); | ||||
8007 | for (Value *Location : LocationOps) { | ||||
8008 | WeakTrackingVH SunkAddrVH = SunkAddrs[Location]; | ||||
8009 | Value *SunkAddr = SunkAddrVH.pointsToAliveValue() ? SunkAddrVH : nullptr; | ||||
8010 | if (SunkAddr) { | ||||
8011 | // Point dbg.value at locally computed address, which should give the best | ||||
8012 | // opportunity to be accurately lowered. This update may change the type | ||||
8013 | // of pointer being referred to; however this makes no difference to | ||||
8014 | // debugging information, and we can't generate bitcasts that may affect | ||||
8015 | // codegen. | ||||
8016 | DVI.replaceVariableLocationOp(Location, SunkAddr); | ||||
8017 | AnyChange = true; | ||||
8018 | } | ||||
8019 | } | ||||
8020 | return AnyChange; | ||||
8021 | } | ||||
8022 | |||||
8023 | // A llvm.dbg.value may be using a value before its definition, due to | ||||
8024 | // optimizations in this pass and others. Scan for such dbg.values, and rescue | ||||
8025 | // them by moving the dbg.value to immediately after the value definition. | ||||
8026 | // FIXME: Ideally this should never be necessary, and this has the potential | ||||
8027 | // to re-order dbg.value intrinsics. | ||||
8028 | bool CodeGenPrepare::placeDbgValues(Function &F) { | ||||
8029 | bool MadeChange = false; | ||||
8030 | DominatorTree DT(F); | ||||
8031 | |||||
8032 | for (BasicBlock &BB : F) { | ||||
8033 | for (BasicBlock::iterator BI = BB.begin(), BE = BB.end(); BI != BE;) { | ||||
8034 | Instruction *Insn = &*BI++; | ||||
8035 | DbgValueInst *DVI = dyn_cast<DbgValueInst>(Insn); | ||||
8036 | if (!DVI) | ||||
8037 | continue; | ||||
8038 | |||||
8039 | SmallVector<Instruction *, 4> VIs; | ||||
8040 | for (Value *V : DVI->getValues()) | ||||
8041 | if (Instruction *VI = dyn_cast_or_null<Instruction>(V)) | ||||
8042 | VIs.push_back(VI); | ||||
8043 | |||||
8044 | // This DVI may depend on multiple instructions, complicating any | ||||
8045 | // potential sink. This block takes the defensive approach, opting to | ||||
8046 | // "undef" the DVI if it has more than one instruction and any of them do | ||||
8047 | // not dominate DVI. | ||||
8048 | for (Instruction *VI : VIs) { | ||||
8049 | if (VI->isTerminator()) | ||||
8050 | continue; | ||||
8051 | |||||
8052 | // If VI is a phi in a block with an EHPad terminator, we can't insert | ||||
8053 | // after it. | ||||
8054 | if (isa<PHINode>(VI) && VI->getParent()->getTerminator()->isEHPad()) | ||||
8055 | continue; | ||||
8056 | |||||
8057 | // If the defining instruction dominates the dbg.value, we do not need | ||||
8058 | // to move the dbg.value. | ||||
8059 | if (DT.dominates(VI, DVI)) | ||||
8060 | continue; | ||||
8061 | |||||
8062 | // If we depend on multiple instructions and any of them doesn't | ||||
8063 | // dominate this DVI, we probably can't salvage it: moving it to | ||||
8064 | // after any of the instructions could cause us to lose the others. | ||||
8065 | if (VIs.size() > 1) { | ||||
8066 | LLVM_DEBUG(do { } while (false) | ||||
8067 | dbgs()do { } while (false) | ||||
8068 | << "Unable to find valid location for Debug Value, undefing:\n"do { } while (false) | ||||
8069 | << *DVI)do { } while (false); | ||||
8070 | DVI->setUndef(); | ||||
8071 | break; | ||||
8072 | } | ||||
8073 | |||||
8074 | LLVM_DEBUG(dbgs() << "Moving Debug Value before :\n"do { } while (false) | ||||
8075 | << *DVI << ' ' << *VI)do { } while (false); | ||||
8076 | DVI->removeFromParent(); | ||||
8077 | if (isa<PHINode>(VI)) | ||||
8078 | DVI->insertBefore(&*VI->getParent()->getFirstInsertionPt()); | ||||
8079 | else | ||||
8080 | DVI->insertAfter(VI); | ||||
8081 | MadeChange = true; | ||||
8082 | ++NumDbgValueMoved; | ||||
8083 | } | ||||
8084 | } | ||||
8085 | } | ||||
8086 | return MadeChange; | ||||
8087 | } | ||||
8088 | |||||
8089 | // Group scattered pseudo probes in a block to favor SelectionDAG. Scattered | ||||
8090 | // probes can be chained dependencies of other regular DAG nodes and block DAG | ||||
8091 | // combine optimizations. | ||||
8092 | bool CodeGenPrepare::placePseudoProbes(Function &F) { | ||||
8093 | bool MadeChange = false; | ||||
8094 | for (auto &Block : F) { | ||||
8095 | // Move the rest probes to the beginning of the block. | ||||
8096 | auto FirstInst = Block.getFirstInsertionPt(); | ||||
8097 | while (FirstInst != Block.end() && FirstInst->isDebugOrPseudoInst()) | ||||
8098 | ++FirstInst; | ||||
8099 | BasicBlock::iterator I(FirstInst); | ||||
8100 | I++; | ||||
8101 | while (I != Block.end()) { | ||||
8102 | if (auto *II = dyn_cast<PseudoProbeInst>(I++)) { | ||||
8103 | II->moveBefore(&*FirstInst); | ||||
8104 | MadeChange = true; | ||||
8105 | } | ||||
8106 | } | ||||
8107 | } | ||||
8108 | return MadeChange; | ||||
8109 | } | ||||
8110 | |||||
8111 | /// Scale down both weights to fit into uint32_t. | ||||
8112 | static void scaleWeights(uint64_t &NewTrue, uint64_t &NewFalse) { | ||||
8113 | uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse; | ||||
8114 | uint32_t Scale = (NewMax / std::numeric_limits<uint32_t>::max()) + 1; | ||||
8115 | NewTrue = NewTrue / Scale; | ||||
8116 | NewFalse = NewFalse / Scale; | ||||
8117 | } | ||||
8118 | |||||
8119 | /// Some targets prefer to split a conditional branch like: | ||||
8120 | /// \code | ||||
8121 | /// %0 = icmp ne i32 %a, 0 | ||||
8122 | /// %1 = icmp ne i32 %b, 0 | ||||
8123 | /// %or.cond = or i1 %0, %1 | ||||
8124 | /// br i1 %or.cond, label %TrueBB, label %FalseBB | ||||
8125 | /// \endcode | ||||
8126 | /// into multiple branch instructions like: | ||||
8127 | /// \code | ||||
8128 | /// bb1: | ||||
8129 | /// %0 = icmp ne i32 %a, 0 | ||||
8130 | /// br i1 %0, label %TrueBB, label %bb2 | ||||
8131 | /// bb2: | ||||
8132 | /// %1 = icmp ne i32 %b, 0 | ||||
8133 | /// br i1 %1, label %TrueBB, label %FalseBB | ||||
8134 | /// \endcode | ||||
8135 | /// This usually allows instruction selection to do even further optimizations | ||||
8136 | /// and combine the compare with the branch instruction. Currently this is | ||||
8137 | /// applied for targets which have "cheap" jump instructions. | ||||
8138 | /// | ||||
8139 | /// FIXME: Remove the (equivalent?) implementation in SelectionDAG. | ||||
8140 | /// | ||||
8141 | bool CodeGenPrepare::splitBranchCondition(Function &F, bool &ModifiedDT) { | ||||
8142 | if (!TM->Options.EnableFastISel || TLI->isJumpExpensive()) | ||||
8143 | return false; | ||||
8144 | |||||
8145 | bool MadeChange = false; | ||||
8146 | for (auto &BB : F) { | ||||
8147 | // Does this BB end with the following? | ||||
8148 | // %cond1 = icmp|fcmp|binary instruction ... | ||||
8149 | // %cond2 = icmp|fcmp|binary instruction ... | ||||
8150 | // %cond.or = or|and i1 %cond1, cond2 | ||||
8151 | // br i1 %cond.or label %dest1, label %dest2" | ||||
8152 | Instruction *LogicOp; | ||||
8153 | BasicBlock *TBB, *FBB; | ||||
8154 | if (!match(BB.getTerminator(), | ||||
8155 | m_Br(m_OneUse(m_Instruction(LogicOp)), TBB, FBB))) | ||||
8156 | continue; | ||||
8157 | |||||
8158 | auto *Br1 = cast<BranchInst>(BB.getTerminator()); | ||||
8159 | if (Br1->getMetadata(LLVMContext::MD_unpredictable)) | ||||
8160 | continue; | ||||
8161 | |||||
8162 | // The merging of mostly empty BB can cause a degenerate branch. | ||||
8163 | if (TBB == FBB) | ||||
8164 | continue; | ||||
8165 | |||||
8166 | unsigned Opc; | ||||
8167 | Value *Cond1, *Cond2; | ||||
8168 | if (match(LogicOp, | ||||
8169 | m_LogicalAnd(m_OneUse(m_Value(Cond1)), m_OneUse(m_Value(Cond2))))) | ||||
8170 | Opc = Instruction::And; | ||||
8171 | else if (match(LogicOp, m_LogicalOr(m_OneUse(m_Value(Cond1)), | ||||
8172 | m_OneUse(m_Value(Cond2))))) | ||||
8173 | Opc = Instruction::Or; | ||||
8174 | else | ||||
8175 | continue; | ||||
8176 | |||||
8177 | auto IsGoodCond = [](Value *Cond) { | ||||
8178 | return match( | ||||
8179 | Cond, | ||||
8180 | m_CombineOr(m_Cmp(), m_CombineOr(m_LogicalAnd(m_Value(), m_Value()), | ||||
8181 | m_LogicalOr(m_Value(), m_Value())))); | ||||
8182 | }; | ||||
8183 | if (!IsGoodCond(Cond1) || !IsGoodCond(Cond2)) | ||||
8184 | continue; | ||||
8185 | |||||
8186 | LLVM_DEBUG(dbgs() << "Before branch condition splitting\n"; BB.dump())do { } while (false); | ||||
8187 | |||||
8188 | // Create a new BB. | ||||
8189 | auto *TmpBB = | ||||
8190 | BasicBlock::Create(BB.getContext(), BB.getName() + ".cond.split", | ||||
8191 | BB.getParent(), BB.getNextNode()); | ||||
8192 | |||||
8193 | // Update original basic block by using the first condition directly by the | ||||
8194 | // branch instruction and removing the no longer needed and/or instruction. | ||||
8195 | Br1->setCondition(Cond1); | ||||
8196 | LogicOp->eraseFromParent(); | ||||
8197 | |||||
8198 | // Depending on the condition we have to either replace the true or the | ||||
8199 | // false successor of the original branch instruction. | ||||
8200 | if (Opc == Instruction::And) | ||||
8201 | Br1->setSuccessor(0, TmpBB); | ||||
8202 | else | ||||
8203 | Br1->setSuccessor(1, TmpBB); | ||||
8204 | |||||
8205 | // Fill in the new basic block. | ||||
8206 | auto *Br2 = IRBuilder<>(TmpBB).CreateCondBr(Cond2, TBB, FBB); | ||||
8207 | if (auto *I = dyn_cast<Instruction>(Cond2)) { | ||||
8208 | I->removeFromParent(); | ||||
8209 | I->insertBefore(Br2); | ||||
8210 | } | ||||
8211 | |||||
8212 | // Update PHI nodes in both successors. The original BB needs to be | ||||
8213 | // replaced in one successor's PHI nodes, because the branch comes now from | ||||
8214 | // the newly generated BB (NewBB). In the other successor we need to add one | ||||
8215 | // incoming edge to the PHI nodes, because both branch instructions target | ||||
8216 | // now the same successor. Depending on the original branch condition | ||||
8217 | // (and/or) we have to swap the successors (TrueDest, FalseDest), so that | ||||
8218 | // we perform the correct update for the PHI nodes. | ||||
8219 | // This doesn't change the successor order of the just created branch | ||||
8220 | // instruction (or any other instruction). | ||||
8221 | if (Opc == Instruction::Or) | ||||
8222 | std::swap(TBB, FBB); | ||||
8223 | |||||
8224 | // Replace the old BB with the new BB. | ||||
8225 | TBB->replacePhiUsesWith(&BB, TmpBB); | ||||
8226 | |||||
8227 | // Add another incoming edge form the new BB. | ||||
8228 | for (PHINode &PN : FBB->phis()) { | ||||
8229 | auto *Val = PN.getIncomingValueForBlock(&BB); | ||||
8230 | PN.addIncoming(Val, TmpBB); | ||||
8231 | } | ||||
8232 | |||||
8233 | // Update the branch weights (from SelectionDAGBuilder:: | ||||
8234 | // FindMergedConditions). | ||||
8235 | if (Opc == Instruction::Or) { | ||||
8236 | // Codegen X | Y as: | ||||
8237 | // BB1: | ||||
8238 | // jmp_if_X TBB | ||||
8239 | // jmp TmpBB | ||||
8240 | // TmpBB: | ||||
8241 | // jmp_if_Y TBB | ||||
8242 | // jmp FBB | ||||
8243 | // | ||||
8244 | |||||
8245 | // We have flexibility in setting Prob for BB1 and Prob for NewBB. | ||||
8246 | // The requirement is that | ||||
8247 | // TrueProb for BB1 + (FalseProb for BB1 * TrueProb for TmpBB) | ||||
8248 | // = TrueProb for original BB. | ||||
8249 | // Assuming the original weights are A and B, one choice is to set BB1's | ||||
8250 | // weights to A and A+2B, and set TmpBB's weights to A and 2B. This choice | ||||
8251 | // assumes that | ||||
8252 | // TrueProb for BB1 == FalseProb for BB1 * TrueProb for TmpBB. | ||||
8253 | // Another choice is to assume TrueProb for BB1 equals to TrueProb for | ||||
8254 | // TmpBB, but the math is more complicated. | ||||
8255 | uint64_t TrueWeight, FalseWeight; | ||||
8256 | if (Br1->extractProfMetadata(TrueWeight, FalseWeight)) { | ||||
8257 | uint64_t NewTrueWeight = TrueWeight; | ||||
8258 | uint64_t NewFalseWeight = TrueWeight + 2 * FalseWeight; | ||||
8259 | scaleWeights(NewTrueWeight, NewFalseWeight); | ||||
8260 | Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext()) | ||||
8261 | .createBranchWeights(TrueWeight, FalseWeight)); | ||||
8262 | |||||
8263 | NewTrueWeight = TrueWeight; | ||||
8264 | NewFalseWeight = 2 * FalseWeight; | ||||
8265 | scaleWeights(NewTrueWeight, NewFalseWeight); | ||||
8266 | Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext()) | ||||
8267 | .createBranchWeights(TrueWeight, FalseWeight)); | ||||
8268 | } | ||||
8269 | } else { | ||||
8270 | // Codegen X & Y as: | ||||
8271 | // BB1: | ||||
8272 | // jmp_if_X TmpBB | ||||
8273 | // jmp FBB | ||||
8274 | // TmpBB: | ||||
8275 | // jmp_if_Y TBB | ||||
8276 | // jmp FBB | ||||
8277 | // | ||||
8278 | // This requires creation of TmpBB after CurBB. | ||||
8279 | |||||
8280 | // We have flexibility in setting Prob for BB1 and Prob for TmpBB. | ||||
8281 | // The requirement is that | ||||
8282 | // FalseProb for BB1 + (TrueProb for BB1 * FalseProb for TmpBB) | ||||
8283 | // = FalseProb for original BB. | ||||
8284 | // Assuming the original weights are A and B, one choice is to set BB1's | ||||
8285 | // weights to 2A+B and B, and set TmpBB's weights to 2A and B. This choice | ||||
8286 | // assumes that | ||||
8287 | // FalseProb for BB1 == TrueProb for BB1 * FalseProb for TmpBB. | ||||
8288 | uint64_t TrueWeight, FalseWeight; | ||||
8289 | if (Br1->extractProfMetadata(TrueWeight, FalseWeight)) { | ||||
8290 | uint64_t NewTrueWeight = 2 * TrueWeight + FalseWeight; | ||||
8291 | uint64_t NewFalseWeight = FalseWeight; | ||||
8292 | scaleWeights(NewTrueWeight, NewFalseWeight); | ||||
8293 | Br1->setMetadata(LLVMContext::MD_prof, MDBuilder(Br1->getContext()) | ||||
8294 | .createBranchWeights(TrueWeight, FalseWeight)); | ||||
8295 | |||||
8296 | NewTrueWeight = 2 * TrueWeight; | ||||
8297 | NewFalseWeight = FalseWeight; | ||||
8298 | scaleWeights(NewTrueWeight, NewFalseWeight); | ||||
8299 | Br2->setMetadata(LLVMContext::MD_prof, MDBuilder(Br2->getContext()) | ||||
8300 | .createBranchWeights(TrueWeight, FalseWeight)); | ||||
8301 | } | ||||
8302 | } | ||||
8303 | |||||
8304 | ModifiedDT = true; | ||||
8305 | MadeChange = true; | ||||
8306 | |||||
8307 | LLVM_DEBUG(dbgs() << "After branch condition splitting\n"; BB.dump();do { } while (false) | ||||
8308 | TmpBB->dump())do { } while (false); | ||||
8309 | } | ||||
8310 | return MadeChange; | ||||
8311 | } |
1 | //===-- llvm/Support/Alignment.h - Useful alignment functions ---*- C++ -*-===// | |||
2 | // | |||
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. | |||
4 | // See https://llvm.org/LICENSE.txt for license information. | |||
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception | |||
6 | // | |||
7 | //===----------------------------------------------------------------------===// | |||
8 | // | |||
9 | // This file contains types to represent alignments. | |||
10 | // They are instrumented to guarantee some invariants are preserved and prevent | |||
11 | // invalid manipulations. | |||
12 | // | |||
13 | // - Align represents an alignment in bytes, it is always set and always a valid | |||
14 | // power of two, its minimum value is 1 which means no alignment requirements. | |||
15 | // | |||
16 | // - MaybeAlign is an optional type, it may be undefined or set. When it's set | |||
17 | // you can get the underlying Align type by using the getValue() method. | |||
18 | // | |||
19 | //===----------------------------------------------------------------------===// | |||
20 | ||||
21 | #ifndef LLVM_SUPPORT_ALIGNMENT_H_ | |||
22 | #define LLVM_SUPPORT_ALIGNMENT_H_ | |||
23 | ||||
24 | #include "llvm/ADT/Optional.h" | |||
25 | #include "llvm/Support/MathExtras.h" | |||
26 | #include <cassert> | |||
27 | #ifndef NDEBUG1 | |||
28 | #include <string> | |||
29 | #endif // NDEBUG | |||
30 | ||||
31 | namespace llvm { | |||
32 | ||||
33 | #define ALIGN_CHECK_ISPOSITIVE(decl) \ | |||
34 | assert(decl > 0 && (#decl " should be defined"))((void)0) | |||
35 | ||||
36 | /// This struct is a compact representation of a valid (non-zero power of two) | |||
37 | /// alignment. | |||
38 | /// It is suitable for use as static global constants. | |||
39 | struct Align { | |||
40 | private: | |||
41 | uint8_t ShiftValue = 0; /// The log2 of the required alignment. | |||
42 | /// ShiftValue is less than 64 by construction. | |||
43 | ||||
44 | friend struct MaybeAlign; | |||
45 | friend unsigned Log2(Align); | |||
46 | friend bool operator==(Align Lhs, Align Rhs); | |||
47 | friend bool operator!=(Align Lhs, Align Rhs); | |||
48 | friend bool operator<=(Align Lhs, Align Rhs); | |||
49 | friend bool operator>=(Align Lhs, Align Rhs); | |||
50 | friend bool operator<(Align Lhs, Align Rhs); | |||
51 | friend bool operator>(Align Lhs, Align Rhs); | |||
52 | friend unsigned encode(struct MaybeAlign A); | |||
53 | friend struct MaybeAlign decodeMaybeAlign(unsigned Value); | |||
54 | ||||
55 | /// A trivial type to allow construction of constexpr Align. | |||
56 | /// This is currently needed to workaround a bug in GCC 5.3 which prevents | |||
57 | /// definition of constexpr assign operators. | |||
58 | /// https://stackoverflow.com/questions/46756288/explicitly-defaulted-function-cannot-be-declared-as-constexpr-because-the-implic | |||
59 | /// FIXME: Remove this, make all assign operators constexpr and introduce user | |||
60 | /// defined literals when we don't have to support GCC 5.3 anymore. | |||
61 | /// https://llvm.org/docs/GettingStarted.html#getting-a-modern-host-c-toolchain | |||
62 | struct LogValue { | |||
63 | uint8_t Log; | |||
64 | }; | |||
65 | ||||
66 | public: | |||
67 | /// Default is byte-aligned. | |||
68 | constexpr Align() = default; | |||
69 | /// Do not perform checks in case of copy/move construct/assign, because the | |||
70 | /// checks have been performed when building `Other`. | |||
71 | constexpr Align(const Align &Other) = default; | |||
72 | constexpr Align(Align &&Other) = default; | |||
73 | Align &operator=(const Align &Other) = default; | |||
74 | Align &operator=(Align &&Other) = default; | |||
75 | ||||
76 | explicit Align(uint64_t Value) { | |||
77 | assert(Value > 0 && "Value must not be 0")((void)0); | |||
78 | assert(llvm::isPowerOf2_64(Value) && "Alignment is not a power of 2")((void)0); | |||
79 | ShiftValue = Log2_64(Value); | |||
80 | assert(ShiftValue < 64 && "Broken invariant")((void)0); | |||
81 | } | |||
82 | ||||
83 | /// This is a hole in the type system and should not be abused. | |||
84 | /// Needed to interact with C for instance. | |||
85 | uint64_t value() const { return uint64_t(1) << ShiftValue; } | |||
| ||||
86 | ||||
87 | /// Allow constructions of constexpr Align. | |||
88 | template <size_t kValue> constexpr static LogValue Constant() { | |||
89 | return LogValue{static_cast<uint8_t>(CTLog2<kValue>())}; | |||
90 | } | |||
91 | ||||
92 | /// Allow constructions of constexpr Align from types. | |||
93 | /// Compile time equivalent to Align(alignof(T)). | |||
94 | template <typename T> constexpr static LogValue Of() { | |||
95 | return Constant<std::alignment_of<T>::value>(); | |||
96 | } | |||
97 | ||||
98 | /// Constexpr constructor from LogValue type. | |||
99 | constexpr Align(LogValue CA) : ShiftValue(CA.Log) {} | |||
100 | }; | |||
101 | ||||
102 | /// Treats the value 0 as a 1, so Align is always at least 1. | |||
103 | inline Align assumeAligned(uint64_t Value) { | |||
104 | return Value ? Align(Value) : Align(); | |||
105 | } | |||
106 | ||||
107 | /// This struct is a compact representation of a valid (power of two) or | |||
108 | /// undefined (0) alignment. | |||
109 | struct MaybeAlign : public llvm::Optional<Align> { | |||
110 | private: | |||
111 | using UP = llvm::Optional<Align>; | |||
112 | ||||
113 | public: | |||
114 | /// Default is undefined. | |||
115 | MaybeAlign() = default; | |||
116 | /// Do not perform checks in case of copy/move construct/assign, because the | |||
117 | /// checks have been performed when building `Other`. | |||
118 | MaybeAlign(const MaybeAlign &Other) = default; | |||
119 | MaybeAlign &operator=(const MaybeAlign &Other) = default; | |||
120 | MaybeAlign(MaybeAlign &&Other) = default; | |||
121 | MaybeAlign &operator=(MaybeAlign &&Other) = default; | |||
122 | ||||
123 | /// Use llvm::Optional<Align> constructor. | |||
124 | using UP::UP; | |||
125 | ||||
126 | explicit MaybeAlign(uint64_t Value) { | |||
127 | assert((Value == 0 || llvm::isPowerOf2_64(Value)) &&((void)0) | |||
128 | "Alignment is neither 0 nor a power of 2")((void)0); | |||
129 | if (Value) | |||
130 | emplace(Value); | |||
131 | } | |||
132 | ||||
133 | /// For convenience, returns a valid alignment or 1 if undefined. | |||
134 | Align valueOrOne() const { return hasValue() ? getValue() : Align(); } | |||
135 | }; | |||
136 | ||||
137 | /// Checks that SizeInBytes is a multiple of the alignment. | |||
138 | inline bool isAligned(Align Lhs, uint64_t SizeInBytes) { | |||
139 | return SizeInBytes % Lhs.value() == 0; | |||
140 | } | |||
141 | ||||
142 | /// Checks that Addr is a multiple of the alignment. | |||
143 | inline bool isAddrAligned(Align Lhs, const void *Addr) { | |||
144 | return isAligned(Lhs, reinterpret_cast<uintptr_t>(Addr)); | |||
145 | } | |||
146 | ||||
147 | /// Returns a multiple of A needed to store `Size` bytes. | |||
148 | inline uint64_t alignTo(uint64_t Size, Align A) { | |||
149 | const uint64_t Value = A.value(); | |||
150 | // The following line is equivalent to `(Size + Value - 1) / Value * Value`. | |||
151 | ||||
152 | // The division followed by a multiplication can be thought of as a right | |||
153 | // shift followed by a left shift which zeros out the extra bits produced in | |||
154 | // the bump; `~(Value - 1)` is a mask where all those bits being zeroed out | |||
155 | // are just zero. | |||
156 | ||||
157 | // Most compilers can generate this code but the pattern may be missed when | |||
158 | // multiple functions gets inlined. | |||
159 | return (Size + Value - 1) & ~(Value - 1U); | |||
160 | } | |||
161 | ||||
162 | /// If non-zero \p Skew is specified, the return value will be a minimal integer | |||
163 | /// that is greater than or equal to \p Size and equal to \p A * N + \p Skew for | |||
164 | /// some integer N. If \p Skew is larger than \p A, its value is adjusted to '\p | |||
165 | /// Skew mod \p A'. | |||
166 | /// | |||
167 | /// Examples: | |||
168 | /// \code | |||
169 | /// alignTo(5, Align(8), 7) = 7 | |||
170 | /// alignTo(17, Align(8), 1) = 17 | |||
171 | /// alignTo(~0LL, Align(8), 3) = 3 | |||
172 | /// \endcode | |||
173 | inline uint64_t alignTo(uint64_t Size, Align A, uint64_t Skew) { | |||
174 | const uint64_t Value = A.value(); | |||
175 | Skew %= Value; | |||
176 | return ((Size + Value - 1 - Skew) & ~(Value - 1U)) + Skew; | |||
177 | } | |||
178 | ||||
179 | /// Returns a multiple of A needed to store `Size` bytes. | |||
180 | /// Returns `Size` if current alignment is undefined. | |||
181 | inline uint64_t alignTo(uint64_t Size, MaybeAlign A) { | |||
182 | return A ? alignTo(Size, A.getValue()) : Size; | |||
183 | } | |||
184 | ||||
185 | /// Aligns `Addr` to `Alignment` bytes, rounding up. | |||
186 | inline uintptr_t alignAddr(const void *Addr, Align Alignment) { | |||
187 | uintptr_t ArithAddr = reinterpret_cast<uintptr_t>(Addr); | |||
188 | assert(static_cast<uintptr_t>(ArithAddr + Alignment.value() - 1) >=((void)0) | |||
189 | ArithAddr &&((void)0) | |||
190 | "Overflow")((void)0); | |||
191 | return alignTo(ArithAddr, Alignment); | |||
192 | } | |||
193 | ||||
194 | /// Returns the offset to the next integer (mod 2**64) that is greater than | |||
195 | /// or equal to \p Value and is a multiple of \p Align. | |||
196 | inline uint64_t offsetToAlignment(uint64_t Value, Align Alignment) { | |||
197 | return alignTo(Value, Alignment) - Value; | |||
198 | } | |||
199 | ||||
200 | /// Returns the necessary adjustment for aligning `Addr` to `Alignment` | |||
201 | /// bytes, rounding up. | |||
202 | inline uint64_t offsetToAlignedAddr(const void *Addr, Align Alignment) { | |||
203 | return offsetToAlignment(reinterpret_cast<uintptr_t>(Addr), Alignment); | |||
204 | } | |||
205 | ||||
206 | /// Returns the log2 of the alignment. | |||
207 | inline unsigned Log2(Align A) { return A.ShiftValue; } | |||
208 | ||||
209 | /// Returns the alignment that satisfies both alignments. | |||
210 | /// Same semantic as MinAlign. | |||
211 | inline Align commonAlignment(Align A, Align B) { return std::min(A, B); } | |||
212 | ||||
213 | /// Returns the alignment that satisfies both alignments. | |||
214 | /// Same semantic as MinAlign. | |||
215 | inline Align commonAlignment(Align A, uint64_t Offset) { | |||
216 | return Align(MinAlign(A.value(), Offset)); | |||
217 | } | |||
218 | ||||
219 | /// Returns the alignment that satisfies both alignments. | |||
220 | /// Same semantic as MinAlign. | |||
221 | inline MaybeAlign commonAlignment(MaybeAlign A, MaybeAlign B) { | |||
222 | return A && B ? commonAlignment(*A, *B) : A ? A : B; | |||
223 | } | |||
224 | ||||
225 | /// Returns the alignment that satisfies both alignments. | |||
226 | /// Same semantic as MinAlign. | |||
227 | inline MaybeAlign commonAlignment(MaybeAlign A, uint64_t Offset) { | |||
228 | return MaybeAlign(MinAlign((*A).value(), Offset)); | |||
229 | } | |||
230 | ||||
231 | /// Returns a representation of the alignment that encodes undefined as 0. | |||
232 | inline unsigned encode(MaybeAlign A) { return A ? A->ShiftValue + 1 : 0; } | |||
233 | ||||
234 | /// Dual operation of the encode function above. | |||
235 | inline MaybeAlign decodeMaybeAlign(unsigned Value) { | |||
236 | if (Value == 0) | |||
237 | return MaybeAlign(); | |||
238 | Align Out; | |||
239 | Out.ShiftValue = Value - 1; | |||
240 | return Out; | |||
241 | } | |||
242 | ||||
243 | /// Returns a representation of the alignment, the encoded value is positive by | |||
244 | /// definition. | |||
245 | inline unsigned encode(Align A) { return encode(MaybeAlign(A)); } | |||
246 | ||||
247 | /// Comparisons between Align and scalars. Rhs must be positive. | |||
248 | inline bool operator==(Align Lhs, uint64_t Rhs) { | |||
249 | ALIGN_CHECK_ISPOSITIVE(Rhs); | |||
250 | return Lhs.value() == Rhs; | |||
251 | } | |||
252 | inline bool operator!=(Align Lhs, uint64_t Rhs) { | |||
253 | ALIGN_CHECK_ISPOSITIVE(Rhs); | |||
254 | return Lhs.value() != Rhs; | |||
255 | } | |||
256 | inline bool operator<=(Align Lhs, uint64_t Rhs) { | |||
257 | ALIGN_CHECK_ISPOSITIVE(Rhs); | |||
258 | return Lhs.value() <= Rhs; | |||
259 | } | |||
260 | inline bool operator>=(Align Lhs, uint64_t Rhs) { | |||
261 | ALIGN_CHECK_ISPOSITIVE(Rhs); | |||
262 | return Lhs.value() >= Rhs; | |||
263 | } | |||
264 | inline bool operator<(Align Lhs, uint64_t Rhs) { | |||
265 | ALIGN_CHECK_ISPOSITIVE(Rhs); | |||
266 | return Lhs.value() < Rhs; | |||
267 | } | |||
268 | inline bool operator>(Align Lhs, uint64_t Rhs) { | |||
269 | ALIGN_CHECK_ISPOSITIVE(Rhs); | |||
270 | return Lhs.value() > Rhs; | |||
271 | } | |||
272 | ||||
273 | /// Comparisons between MaybeAlign and scalars. | |||
274 | inline bool operator==(MaybeAlign Lhs, uint64_t Rhs) { | |||
275 | return Lhs ? (*Lhs).value() == Rhs : Rhs == 0; | |||
276 | } | |||
277 | inline bool operator!=(MaybeAlign Lhs, uint64_t Rhs) { | |||
278 | return Lhs ? (*Lhs).value() != Rhs : Rhs != 0; | |||
279 | } | |||
280 | ||||
281 | /// Comparisons operators between Align. | |||
282 | inline bool operator==(Align Lhs, Align Rhs) { | |||
283 | return Lhs.ShiftValue == Rhs.ShiftValue; | |||
284 | } | |||
285 | inline bool operator!=(Align Lhs, Align Rhs) { | |||
286 | return Lhs.ShiftValue != Rhs.ShiftValue; | |||
287 | } | |||
288 | inline bool operator<=(Align Lhs, Align Rhs) { | |||
289 | return Lhs.ShiftValue <= Rhs.ShiftValue; | |||
290 | } | |||
291 | inline bool operator>=(Align Lhs, Align Rhs) { | |||
292 | return Lhs.ShiftValue >= Rhs.ShiftValue; | |||
293 | } | |||
294 | inline bool operator<(Align Lhs, Align Rhs) { | |||
295 | return Lhs.ShiftValue < Rhs.ShiftValue; | |||
296 | } | |||
297 | inline bool operator>(Align Lhs, Align Rhs) { | |||
298 | return Lhs.ShiftValue > Rhs.ShiftValue; | |||
299 | } | |||
300 | ||||
301 | // Don't allow relational comparisons with MaybeAlign. | |||
302 | bool operator<=(Align Lhs, MaybeAlign Rhs) = delete; | |||
303 | bool operator>=(Align Lhs, MaybeAlign Rhs) = delete; | |||
304 | bool operator<(Align Lhs, MaybeAlign Rhs) = delete; | |||
305 | bool operator>(Align Lhs, MaybeAlign Rhs) = delete; | |||
306 | ||||
307 | bool operator<=(MaybeAlign Lhs, Align Rhs) = delete; | |||
308 | bool operator>=(MaybeAlign Lhs, Align Rhs) = delete; | |||
309 | bool operator<(MaybeAlign Lhs, Align Rhs) = delete; | |||
310 | bool operator>(MaybeAlign Lhs, Align Rhs) = delete; | |||
311 | ||||
312 | bool operator<=(MaybeAlign Lhs, MaybeAlign Rhs) = delete; | |||
313 | bool operator>=(MaybeAlign Lhs, MaybeAlign Rhs) = delete; | |||
314 | bool operator<(MaybeAlign Lhs, MaybeAlign Rhs) = delete; | |||
315 | bool operator>(MaybeAlign Lhs, MaybeAlign Rhs) = delete; | |||
316 | ||||
317 | inline Align operator*(Align Lhs, uint64_t Rhs) { | |||
318 | assert(Rhs > 0 && "Rhs must be positive")((void)0); | |||
319 | return Align(Lhs.value() * Rhs); | |||
320 | } | |||
321 | ||||
322 | inline MaybeAlign operator*(MaybeAlign Lhs, uint64_t Rhs) { | |||
323 | assert(Rhs > 0 && "Rhs must be positive")((void)0); | |||
324 | return Lhs ? Lhs.getValue() * Rhs : MaybeAlign(); | |||
325 | } | |||
326 | ||||
327 | inline Align operator/(Align Lhs, uint64_t Divisor) { | |||
328 | assert(llvm::isPowerOf2_64(Divisor) &&((void)0) | |||
329 | "Divisor must be positive and a power of 2")((void)0); | |||
330 | assert(Lhs != 1 && "Can't halve byte alignment")((void)0); | |||
331 | return Align(Lhs.value() / Divisor); | |||
332 | } | |||
333 | ||||
334 | inline MaybeAlign operator/(MaybeAlign Lhs, uint64_t Divisor) { | |||
335 | assert(llvm::isPowerOf2_64(Divisor) &&((void)0) | |||
336 | "Divisor must be positive and a power of 2")((void)0); | |||
337 | return Lhs ? Lhs.getValue() / Divisor : MaybeAlign(); | |||
338 | } | |||
339 | ||||
340 | inline Align max(MaybeAlign Lhs, Align Rhs) { | |||
341 | return Lhs && *Lhs > Rhs ? *Lhs : Rhs; | |||
342 | } | |||
343 | ||||
344 | inline Align max(Align Lhs, MaybeAlign Rhs) { | |||
345 | return Rhs && *Rhs > Lhs ? *Rhs : Lhs; | |||
346 | } | |||
347 | ||||
348 | #ifndef NDEBUG1 | |||
349 | // For usage in LLVM_DEBUG macros. | |||
350 | inline std::string DebugStr(const Align &A) { | |||
351 | return std::to_string(A.value()); | |||
352 | } | |||
353 | // For usage in LLVM_DEBUG macros. | |||
354 | inline std::string DebugStr(const MaybeAlign &MA) { | |||
355 | if (MA) | |||
356 | return std::to_string(MA->value()); | |||
357 | return "None"; | |||
358 | } | |||
359 | #endif // NDEBUG | |||
360 | ||||
361 | #undef ALIGN_CHECK_ISPOSITIVE | |||
362 | ||||
363 | } // namespace llvm | |||
364 | ||||
365 | #endif // LLVM_SUPPORT_ALIGNMENT_H_ |