File: | src/gnu/usr.bin/clang/libLLVM/../../../llvm/llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp |
Warning: | line 4093, column 7 Value stored to 'NumLeftover' is never read |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | //===-- llvm/CodeGen/GlobalISel/LegalizerHelper.cpp -----------------------===// |
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 | /// \file This file implements the LegalizerHelper class to legalize |
10 | /// individual instructions and the LegalizeMachineIR wrapper pass for the |
11 | /// primary legalization. |
12 | // |
13 | //===----------------------------------------------------------------------===// |
14 | |
15 | #include "llvm/CodeGen/GlobalISel/LegalizerHelper.h" |
16 | #include "llvm/CodeGen/GlobalISel/CallLowering.h" |
17 | #include "llvm/CodeGen/GlobalISel/GISelChangeObserver.h" |
18 | #include "llvm/CodeGen/GlobalISel/LegalizerInfo.h" |
19 | #include "llvm/CodeGen/GlobalISel/LostDebugLocObserver.h" |
20 | #include "llvm/CodeGen/GlobalISel/MIPatternMatch.h" |
21 | #include "llvm/CodeGen/GlobalISel/Utils.h" |
22 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
23 | #include "llvm/CodeGen/TargetFrameLowering.h" |
24 | #include "llvm/CodeGen/TargetInstrInfo.h" |
25 | #include "llvm/CodeGen/TargetLowering.h" |
26 | #include "llvm/CodeGen/TargetOpcodes.h" |
27 | #include "llvm/CodeGen/TargetSubtargetInfo.h" |
28 | #include "llvm/IR/Instructions.h" |
29 | #include "llvm/Support/Debug.h" |
30 | #include "llvm/Support/MathExtras.h" |
31 | #include "llvm/Support/raw_ostream.h" |
32 | |
33 | #define DEBUG_TYPE"legalizer" "legalizer" |
34 | |
35 | using namespace llvm; |
36 | using namespace LegalizeActions; |
37 | using namespace MIPatternMatch; |
38 | |
39 | /// Try to break down \p OrigTy into \p NarrowTy sized pieces. |
40 | /// |
41 | /// Returns the number of \p NarrowTy elements needed to reconstruct \p OrigTy, |
42 | /// with any leftover piece as type \p LeftoverTy |
43 | /// |
44 | /// Returns -1 in the first element of the pair if the breakdown is not |
45 | /// satisfiable. |
46 | static std::pair<int, int> |
47 | getNarrowTypeBreakDown(LLT OrigTy, LLT NarrowTy, LLT &LeftoverTy) { |
48 | assert(!LeftoverTy.isValid() && "this is an out argument")((void)0); |
49 | |
50 | unsigned Size = OrigTy.getSizeInBits(); |
51 | unsigned NarrowSize = NarrowTy.getSizeInBits(); |
52 | unsigned NumParts = Size / NarrowSize; |
53 | unsigned LeftoverSize = Size - NumParts * NarrowSize; |
54 | assert(Size > NarrowSize)((void)0); |
55 | |
56 | if (LeftoverSize == 0) |
57 | return {NumParts, 0}; |
58 | |
59 | if (NarrowTy.isVector()) { |
60 | unsigned EltSize = OrigTy.getScalarSizeInBits(); |
61 | if (LeftoverSize % EltSize != 0) |
62 | return {-1, -1}; |
63 | LeftoverTy = LLT::scalarOrVector( |
64 | ElementCount::getFixed(LeftoverSize / EltSize), EltSize); |
65 | } else { |
66 | LeftoverTy = LLT::scalar(LeftoverSize); |
67 | } |
68 | |
69 | int NumLeftover = LeftoverSize / LeftoverTy.getSizeInBits(); |
70 | return std::make_pair(NumParts, NumLeftover); |
71 | } |
72 | |
73 | static Type *getFloatTypeForLLT(LLVMContext &Ctx, LLT Ty) { |
74 | |
75 | if (!Ty.isScalar()) |
76 | return nullptr; |
77 | |
78 | switch (Ty.getSizeInBits()) { |
79 | case 16: |
80 | return Type::getHalfTy(Ctx); |
81 | case 32: |
82 | return Type::getFloatTy(Ctx); |
83 | case 64: |
84 | return Type::getDoubleTy(Ctx); |
85 | case 80: |
86 | return Type::getX86_FP80Ty(Ctx); |
87 | case 128: |
88 | return Type::getFP128Ty(Ctx); |
89 | default: |
90 | return nullptr; |
91 | } |
92 | } |
93 | |
94 | LegalizerHelper::LegalizerHelper(MachineFunction &MF, |
95 | GISelChangeObserver &Observer, |
96 | MachineIRBuilder &Builder) |
97 | : MIRBuilder(Builder), Observer(Observer), MRI(MF.getRegInfo()), |
98 | LI(*MF.getSubtarget().getLegalizerInfo()), |
99 | TLI(*MF.getSubtarget().getTargetLowering()) { } |
100 | |
101 | LegalizerHelper::LegalizerHelper(MachineFunction &MF, const LegalizerInfo &LI, |
102 | GISelChangeObserver &Observer, |
103 | MachineIRBuilder &B) |
104 | : MIRBuilder(B), Observer(Observer), MRI(MF.getRegInfo()), LI(LI), |
105 | TLI(*MF.getSubtarget().getTargetLowering()) { } |
106 | |
107 | LegalizerHelper::LegalizeResult |
108 | LegalizerHelper::legalizeInstrStep(MachineInstr &MI, |
109 | LostDebugLocObserver &LocObserver) { |
110 | LLVM_DEBUG(dbgs() << "Legalizing: " << MI)do { } while (false); |
111 | |
112 | MIRBuilder.setInstrAndDebugLoc(MI); |
113 | |
114 | if (MI.getOpcode() == TargetOpcode::G_INTRINSIC || |
115 | MI.getOpcode() == TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS) |
116 | return LI.legalizeIntrinsic(*this, MI) ? Legalized : UnableToLegalize; |
117 | auto Step = LI.getAction(MI, MRI); |
118 | switch (Step.Action) { |
119 | case Legal: |
120 | LLVM_DEBUG(dbgs() << ".. Already legal\n")do { } while (false); |
121 | return AlreadyLegal; |
122 | case Libcall: |
123 | LLVM_DEBUG(dbgs() << ".. Convert to libcall\n")do { } while (false); |
124 | return libcall(MI, LocObserver); |
125 | case NarrowScalar: |
126 | LLVM_DEBUG(dbgs() << ".. Narrow scalar\n")do { } while (false); |
127 | return narrowScalar(MI, Step.TypeIdx, Step.NewType); |
128 | case WidenScalar: |
129 | LLVM_DEBUG(dbgs() << ".. Widen scalar\n")do { } while (false); |
130 | return widenScalar(MI, Step.TypeIdx, Step.NewType); |
131 | case Bitcast: |
132 | LLVM_DEBUG(dbgs() << ".. Bitcast type\n")do { } while (false); |
133 | return bitcast(MI, Step.TypeIdx, Step.NewType); |
134 | case Lower: |
135 | LLVM_DEBUG(dbgs() << ".. Lower\n")do { } while (false); |
136 | return lower(MI, Step.TypeIdx, Step.NewType); |
137 | case FewerElements: |
138 | LLVM_DEBUG(dbgs() << ".. Reduce number of elements\n")do { } while (false); |
139 | return fewerElementsVector(MI, Step.TypeIdx, Step.NewType); |
140 | case MoreElements: |
141 | LLVM_DEBUG(dbgs() << ".. Increase number of elements\n")do { } while (false); |
142 | return moreElementsVector(MI, Step.TypeIdx, Step.NewType); |
143 | case Custom: |
144 | LLVM_DEBUG(dbgs() << ".. Custom legalization\n")do { } while (false); |
145 | return LI.legalizeCustom(*this, MI) ? Legalized : UnableToLegalize; |
146 | default: |
147 | LLVM_DEBUG(dbgs() << ".. Unable to legalize\n")do { } while (false); |
148 | return UnableToLegalize; |
149 | } |
150 | } |
151 | |
152 | void LegalizerHelper::extractParts(Register Reg, LLT Ty, int NumParts, |
153 | SmallVectorImpl<Register> &VRegs) { |
154 | for (int i = 0; i < NumParts; ++i) |
155 | VRegs.push_back(MRI.createGenericVirtualRegister(Ty)); |
156 | MIRBuilder.buildUnmerge(VRegs, Reg); |
157 | } |
158 | |
159 | bool LegalizerHelper::extractParts(Register Reg, LLT RegTy, |
160 | LLT MainTy, LLT &LeftoverTy, |
161 | SmallVectorImpl<Register> &VRegs, |
162 | SmallVectorImpl<Register> &LeftoverRegs) { |
163 | assert(!LeftoverTy.isValid() && "this is an out argument")((void)0); |
164 | |
165 | unsigned RegSize = RegTy.getSizeInBits(); |
166 | unsigned MainSize = MainTy.getSizeInBits(); |
167 | unsigned NumParts = RegSize / MainSize; |
168 | unsigned LeftoverSize = RegSize - NumParts * MainSize; |
169 | |
170 | // Use an unmerge when possible. |
171 | if (LeftoverSize == 0) { |
172 | for (unsigned I = 0; I < NumParts; ++I) |
173 | VRegs.push_back(MRI.createGenericVirtualRegister(MainTy)); |
174 | MIRBuilder.buildUnmerge(VRegs, Reg); |
175 | return true; |
176 | } |
177 | |
178 | if (MainTy.isVector()) { |
179 | unsigned EltSize = MainTy.getScalarSizeInBits(); |
180 | if (LeftoverSize % EltSize != 0) |
181 | return false; |
182 | LeftoverTy = LLT::scalarOrVector( |
183 | ElementCount::getFixed(LeftoverSize / EltSize), EltSize); |
184 | } else { |
185 | LeftoverTy = LLT::scalar(LeftoverSize); |
186 | } |
187 | |
188 | // For irregular sizes, extract the individual parts. |
189 | for (unsigned I = 0; I != NumParts; ++I) { |
190 | Register NewReg = MRI.createGenericVirtualRegister(MainTy); |
191 | VRegs.push_back(NewReg); |
192 | MIRBuilder.buildExtract(NewReg, Reg, MainSize * I); |
193 | } |
194 | |
195 | for (unsigned Offset = MainSize * NumParts; Offset < RegSize; |
196 | Offset += LeftoverSize) { |
197 | Register NewReg = MRI.createGenericVirtualRegister(LeftoverTy); |
198 | LeftoverRegs.push_back(NewReg); |
199 | MIRBuilder.buildExtract(NewReg, Reg, Offset); |
200 | } |
201 | |
202 | return true; |
203 | } |
204 | |
205 | void LegalizerHelper::insertParts(Register DstReg, |
206 | LLT ResultTy, LLT PartTy, |
207 | ArrayRef<Register> PartRegs, |
208 | LLT LeftoverTy, |
209 | ArrayRef<Register> LeftoverRegs) { |
210 | if (!LeftoverTy.isValid()) { |
211 | assert(LeftoverRegs.empty())((void)0); |
212 | |
213 | if (!ResultTy.isVector()) { |
214 | MIRBuilder.buildMerge(DstReg, PartRegs); |
215 | return; |
216 | } |
217 | |
218 | if (PartTy.isVector()) |
219 | MIRBuilder.buildConcatVectors(DstReg, PartRegs); |
220 | else |
221 | MIRBuilder.buildBuildVector(DstReg, PartRegs); |
222 | return; |
223 | } |
224 | |
225 | SmallVector<Register> GCDRegs; |
226 | LLT GCDTy = getGCDType(getGCDType(ResultTy, LeftoverTy), PartTy); |
227 | for (auto PartReg : concat<const Register>(PartRegs, LeftoverRegs)) |
228 | extractGCDType(GCDRegs, GCDTy, PartReg); |
229 | LLT ResultLCMTy = buildLCMMergePieces(ResultTy, LeftoverTy, GCDTy, GCDRegs); |
230 | buildWidenedRemergeToDst(DstReg, ResultLCMTy, GCDRegs); |
231 | } |
232 | |
233 | /// Append the result registers of G_UNMERGE_VALUES \p MI to \p Regs. |
234 | static void getUnmergeResults(SmallVectorImpl<Register> &Regs, |
235 | const MachineInstr &MI) { |
236 | assert(MI.getOpcode() == TargetOpcode::G_UNMERGE_VALUES)((void)0); |
237 | |
238 | const int StartIdx = Regs.size(); |
239 | const int NumResults = MI.getNumOperands() - 1; |
240 | Regs.resize(Regs.size() + NumResults); |
241 | for (int I = 0; I != NumResults; ++I) |
242 | Regs[StartIdx + I] = MI.getOperand(I).getReg(); |
243 | } |
244 | |
245 | void LegalizerHelper::extractGCDType(SmallVectorImpl<Register> &Parts, |
246 | LLT GCDTy, Register SrcReg) { |
247 | LLT SrcTy = MRI.getType(SrcReg); |
248 | if (SrcTy == GCDTy) { |
249 | // If the source already evenly divides the result type, we don't need to do |
250 | // anything. |
251 | Parts.push_back(SrcReg); |
252 | } else { |
253 | // Need to split into common type sized pieces. |
254 | auto Unmerge = MIRBuilder.buildUnmerge(GCDTy, SrcReg); |
255 | getUnmergeResults(Parts, *Unmerge); |
256 | } |
257 | } |
258 | |
259 | LLT LegalizerHelper::extractGCDType(SmallVectorImpl<Register> &Parts, LLT DstTy, |
260 | LLT NarrowTy, Register SrcReg) { |
261 | LLT SrcTy = MRI.getType(SrcReg); |
262 | LLT GCDTy = getGCDType(getGCDType(SrcTy, NarrowTy), DstTy); |
263 | extractGCDType(Parts, GCDTy, SrcReg); |
264 | return GCDTy; |
265 | } |
266 | |
267 | LLT LegalizerHelper::buildLCMMergePieces(LLT DstTy, LLT NarrowTy, LLT GCDTy, |
268 | SmallVectorImpl<Register> &VRegs, |
269 | unsigned PadStrategy) { |
270 | LLT LCMTy = getLCMType(DstTy, NarrowTy); |
271 | |
272 | int NumParts = LCMTy.getSizeInBits() / NarrowTy.getSizeInBits(); |
273 | int NumSubParts = NarrowTy.getSizeInBits() / GCDTy.getSizeInBits(); |
274 | int NumOrigSrc = VRegs.size(); |
275 | |
276 | Register PadReg; |
277 | |
278 | // Get a value we can use to pad the source value if the sources won't evenly |
279 | // cover the result type. |
280 | if (NumOrigSrc < NumParts * NumSubParts) { |
281 | if (PadStrategy == TargetOpcode::G_ZEXT) |
282 | PadReg = MIRBuilder.buildConstant(GCDTy, 0).getReg(0); |
283 | else if (PadStrategy == TargetOpcode::G_ANYEXT) |
284 | PadReg = MIRBuilder.buildUndef(GCDTy).getReg(0); |
285 | else { |
286 | assert(PadStrategy == TargetOpcode::G_SEXT)((void)0); |
287 | |
288 | // Shift the sign bit of the low register through the high register. |
289 | auto ShiftAmt = |
290 | MIRBuilder.buildConstant(LLT::scalar(64), GCDTy.getSizeInBits() - 1); |
291 | PadReg = MIRBuilder.buildAShr(GCDTy, VRegs.back(), ShiftAmt).getReg(0); |
292 | } |
293 | } |
294 | |
295 | // Registers for the final merge to be produced. |
296 | SmallVector<Register, 4> Remerge(NumParts); |
297 | |
298 | // Registers needed for intermediate merges, which will be merged into a |
299 | // source for Remerge. |
300 | SmallVector<Register, 4> SubMerge(NumSubParts); |
301 | |
302 | // Once we've fully read off the end of the original source bits, we can reuse |
303 | // the same high bits for remaining padding elements. |
304 | Register AllPadReg; |
305 | |
306 | // Build merges to the LCM type to cover the original result type. |
307 | for (int I = 0; I != NumParts; ++I) { |
308 | bool AllMergePartsArePadding = true; |
309 | |
310 | // Build the requested merges to the requested type. |
311 | for (int J = 0; J != NumSubParts; ++J) { |
312 | int Idx = I * NumSubParts + J; |
313 | if (Idx >= NumOrigSrc) { |
314 | SubMerge[J] = PadReg; |
315 | continue; |
316 | } |
317 | |
318 | SubMerge[J] = VRegs[Idx]; |
319 | |
320 | // There are meaningful bits here we can't reuse later. |
321 | AllMergePartsArePadding = false; |
322 | } |
323 | |
324 | // If we've filled up a complete piece with padding bits, we can directly |
325 | // emit the natural sized constant if applicable, rather than a merge of |
326 | // smaller constants. |
327 | if (AllMergePartsArePadding && !AllPadReg) { |
328 | if (PadStrategy == TargetOpcode::G_ANYEXT) |
329 | AllPadReg = MIRBuilder.buildUndef(NarrowTy).getReg(0); |
330 | else if (PadStrategy == TargetOpcode::G_ZEXT) |
331 | AllPadReg = MIRBuilder.buildConstant(NarrowTy, 0).getReg(0); |
332 | |
333 | // If this is a sign extension, we can't materialize a trivial constant |
334 | // with the right type and have to produce a merge. |
335 | } |
336 | |
337 | if (AllPadReg) { |
338 | // Avoid creating additional instructions if we're just adding additional |
339 | // copies of padding bits. |
340 | Remerge[I] = AllPadReg; |
341 | continue; |
342 | } |
343 | |
344 | if (NumSubParts == 1) |
345 | Remerge[I] = SubMerge[0]; |
346 | else |
347 | Remerge[I] = MIRBuilder.buildMerge(NarrowTy, SubMerge).getReg(0); |
348 | |
349 | // In the sign extend padding case, re-use the first all-signbit merge. |
350 | if (AllMergePartsArePadding && !AllPadReg) |
351 | AllPadReg = Remerge[I]; |
352 | } |
353 | |
354 | VRegs = std::move(Remerge); |
355 | return LCMTy; |
356 | } |
357 | |
358 | void LegalizerHelper::buildWidenedRemergeToDst(Register DstReg, LLT LCMTy, |
359 | ArrayRef<Register> RemergeRegs) { |
360 | LLT DstTy = MRI.getType(DstReg); |
361 | |
362 | // Create the merge to the widened source, and extract the relevant bits into |
363 | // the result. |
364 | |
365 | if (DstTy == LCMTy) { |
366 | MIRBuilder.buildMerge(DstReg, RemergeRegs); |
367 | return; |
368 | } |
369 | |
370 | auto Remerge = MIRBuilder.buildMerge(LCMTy, RemergeRegs); |
371 | if (DstTy.isScalar() && LCMTy.isScalar()) { |
372 | MIRBuilder.buildTrunc(DstReg, Remerge); |
373 | return; |
374 | } |
375 | |
376 | if (LCMTy.isVector()) { |
377 | unsigned NumDefs = LCMTy.getSizeInBits() / DstTy.getSizeInBits(); |
378 | SmallVector<Register, 8> UnmergeDefs(NumDefs); |
379 | UnmergeDefs[0] = DstReg; |
380 | for (unsigned I = 1; I != NumDefs; ++I) |
381 | UnmergeDefs[I] = MRI.createGenericVirtualRegister(DstTy); |
382 | |
383 | MIRBuilder.buildUnmerge(UnmergeDefs, |
384 | MIRBuilder.buildMerge(LCMTy, RemergeRegs)); |
385 | return; |
386 | } |
387 | |
388 | llvm_unreachable("unhandled case")__builtin_unreachable(); |
389 | } |
390 | |
391 | static RTLIB::Libcall getRTLibDesc(unsigned Opcode, unsigned Size) { |
392 | #define RTLIBCASE_INT(LibcallPrefix)do { switch (Size) { case 32: return RTLIB::LibcallPrefix32; case 64: return RTLIB::LibcallPrefix64; case 128: return RTLIB::LibcallPrefix128 ; default: __builtin_unreachable(); } } while (0) \ |
393 | do { \ |
394 | switch (Size) { \ |
395 | case 32: \ |
396 | return RTLIB::LibcallPrefix##32; \ |
397 | case 64: \ |
398 | return RTLIB::LibcallPrefix##64; \ |
399 | case 128: \ |
400 | return RTLIB::LibcallPrefix##128; \ |
401 | default: \ |
402 | llvm_unreachable("unexpected size")__builtin_unreachable(); \ |
403 | } \ |
404 | } while (0) |
405 | |
406 | #define RTLIBCASE(LibcallPrefix)do { switch (Size) { case 32: return RTLIB::LibcallPrefix32; case 64: return RTLIB::LibcallPrefix64; case 80: return RTLIB::LibcallPrefix80 ; case 128: return RTLIB::LibcallPrefix128; default: __builtin_unreachable (); } } while (0) \ |
407 | do { \ |
408 | switch (Size) { \ |
409 | case 32: \ |
410 | return RTLIB::LibcallPrefix##32; \ |
411 | case 64: \ |
412 | return RTLIB::LibcallPrefix##64; \ |
413 | case 80: \ |
414 | return RTLIB::LibcallPrefix##80; \ |
415 | case 128: \ |
416 | return RTLIB::LibcallPrefix##128; \ |
417 | default: \ |
418 | llvm_unreachable("unexpected size")__builtin_unreachable(); \ |
419 | } \ |
420 | } while (0) |
421 | |
422 | switch (Opcode) { |
423 | case TargetOpcode::G_SDIV: |
424 | RTLIBCASE_INT(SDIV_I)do { switch (Size) { case 32: return RTLIB::SDIV_I32; case 64 : return RTLIB::SDIV_I64; case 128: return RTLIB::SDIV_I128; default : __builtin_unreachable(); } } while (0); |
425 | case TargetOpcode::G_UDIV: |
426 | RTLIBCASE_INT(UDIV_I)do { switch (Size) { case 32: return RTLIB::UDIV_I32; case 64 : return RTLIB::UDIV_I64; case 128: return RTLIB::UDIV_I128; default : __builtin_unreachable(); } } while (0); |
427 | case TargetOpcode::G_SREM: |
428 | RTLIBCASE_INT(SREM_I)do { switch (Size) { case 32: return RTLIB::SREM_I32; case 64 : return RTLIB::SREM_I64; case 128: return RTLIB::SREM_I128; default : __builtin_unreachable(); } } while (0); |
429 | case TargetOpcode::G_UREM: |
430 | RTLIBCASE_INT(UREM_I)do { switch (Size) { case 32: return RTLIB::UREM_I32; case 64 : return RTLIB::UREM_I64; case 128: return RTLIB::UREM_I128; default : __builtin_unreachable(); } } while (0); |
431 | case TargetOpcode::G_CTLZ_ZERO_UNDEF: |
432 | RTLIBCASE_INT(CTLZ_I)do { switch (Size) { case 32: return RTLIB::CTLZ_I32; case 64 : return RTLIB::CTLZ_I64; case 128: return RTLIB::CTLZ_I128; default : __builtin_unreachable(); } } while (0); |
433 | case TargetOpcode::G_FADD: |
434 | RTLIBCASE(ADD_F)do { switch (Size) { case 32: return RTLIB::ADD_F32; case 64: return RTLIB::ADD_F64; case 80: return RTLIB::ADD_F80; case 128 : return RTLIB::ADD_F128; default: __builtin_unreachable(); } } while (0); |
435 | case TargetOpcode::G_FSUB: |
436 | RTLIBCASE(SUB_F)do { switch (Size) { case 32: return RTLIB::SUB_F32; case 64: return RTLIB::SUB_F64; case 80: return RTLIB::SUB_F80; case 128 : return RTLIB::SUB_F128; default: __builtin_unreachable(); } } while (0); |
437 | case TargetOpcode::G_FMUL: |
438 | RTLIBCASE(MUL_F)do { switch (Size) { case 32: return RTLIB::MUL_F32; case 64: return RTLIB::MUL_F64; case 80: return RTLIB::MUL_F80; case 128 : return RTLIB::MUL_F128; default: __builtin_unreachable(); } } while (0); |
439 | case TargetOpcode::G_FDIV: |
440 | RTLIBCASE(DIV_F)do { switch (Size) { case 32: return RTLIB::DIV_F32; case 64: return RTLIB::DIV_F64; case 80: return RTLIB::DIV_F80; case 128 : return RTLIB::DIV_F128; default: __builtin_unreachable(); } } while (0); |
441 | case TargetOpcode::G_FEXP: |
442 | RTLIBCASE(EXP_F)do { switch (Size) { case 32: return RTLIB::EXP_F32; case 64: return RTLIB::EXP_F64; case 80: return RTLIB::EXP_F80; case 128 : return RTLIB::EXP_F128; default: __builtin_unreachable(); } } while (0); |
443 | case TargetOpcode::G_FEXP2: |
444 | RTLIBCASE(EXP2_F)do { switch (Size) { case 32: return RTLIB::EXP2_F32; case 64 : return RTLIB::EXP2_F64; case 80: return RTLIB::EXP2_F80; case 128: return RTLIB::EXP2_F128; default: __builtin_unreachable (); } } while (0); |
445 | case TargetOpcode::G_FREM: |
446 | RTLIBCASE(REM_F)do { switch (Size) { case 32: return RTLIB::REM_F32; case 64: return RTLIB::REM_F64; case 80: return RTLIB::REM_F80; case 128 : return RTLIB::REM_F128; default: __builtin_unreachable(); } } while (0); |
447 | case TargetOpcode::G_FPOW: |
448 | RTLIBCASE(POW_F)do { switch (Size) { case 32: return RTLIB::POW_F32; case 64: return RTLIB::POW_F64; case 80: return RTLIB::POW_F80; case 128 : return RTLIB::POW_F128; default: __builtin_unreachable(); } } while (0); |
449 | case TargetOpcode::G_FMA: |
450 | RTLIBCASE(FMA_F)do { switch (Size) { case 32: return RTLIB::FMA_F32; case 64: return RTLIB::FMA_F64; case 80: return RTLIB::FMA_F80; case 128 : return RTLIB::FMA_F128; default: __builtin_unreachable(); } } while (0); |
451 | case TargetOpcode::G_FSIN: |
452 | RTLIBCASE(SIN_F)do { switch (Size) { case 32: return RTLIB::SIN_F32; case 64: return RTLIB::SIN_F64; case 80: return RTLIB::SIN_F80; case 128 : return RTLIB::SIN_F128; default: __builtin_unreachable(); } } while (0); |
453 | case TargetOpcode::G_FCOS: |
454 | RTLIBCASE(COS_F)do { switch (Size) { case 32: return RTLIB::COS_F32; case 64: return RTLIB::COS_F64; case 80: return RTLIB::COS_F80; case 128 : return RTLIB::COS_F128; default: __builtin_unreachable(); } } while (0); |
455 | case TargetOpcode::G_FLOG10: |
456 | RTLIBCASE(LOG10_F)do { switch (Size) { case 32: return RTLIB::LOG10_F32; case 64 : return RTLIB::LOG10_F64; case 80: return RTLIB::LOG10_F80; case 128: return RTLIB::LOG10_F128; default: __builtin_unreachable (); } } while (0); |
457 | case TargetOpcode::G_FLOG: |
458 | RTLIBCASE(LOG_F)do { switch (Size) { case 32: return RTLIB::LOG_F32; case 64: return RTLIB::LOG_F64; case 80: return RTLIB::LOG_F80; case 128 : return RTLIB::LOG_F128; default: __builtin_unreachable(); } } while (0); |
459 | case TargetOpcode::G_FLOG2: |
460 | RTLIBCASE(LOG2_F)do { switch (Size) { case 32: return RTLIB::LOG2_F32; case 64 : return RTLIB::LOG2_F64; case 80: return RTLIB::LOG2_F80; case 128: return RTLIB::LOG2_F128; default: __builtin_unreachable (); } } while (0); |
461 | case TargetOpcode::G_FCEIL: |
462 | RTLIBCASE(CEIL_F)do { switch (Size) { case 32: return RTLIB::CEIL_F32; case 64 : return RTLIB::CEIL_F64; case 80: return RTLIB::CEIL_F80; case 128: return RTLIB::CEIL_F128; default: __builtin_unreachable (); } } while (0); |
463 | case TargetOpcode::G_FFLOOR: |
464 | RTLIBCASE(FLOOR_F)do { switch (Size) { case 32: return RTLIB::FLOOR_F32; case 64 : return RTLIB::FLOOR_F64; case 80: return RTLIB::FLOOR_F80; case 128: return RTLIB::FLOOR_F128; default: __builtin_unreachable (); } } while (0); |
465 | case TargetOpcode::G_FMINNUM: |
466 | RTLIBCASE(FMIN_F)do { switch (Size) { case 32: return RTLIB::FMIN_F32; case 64 : return RTLIB::FMIN_F64; case 80: return RTLIB::FMIN_F80; case 128: return RTLIB::FMIN_F128; default: __builtin_unreachable (); } } while (0); |
467 | case TargetOpcode::G_FMAXNUM: |
468 | RTLIBCASE(FMAX_F)do { switch (Size) { case 32: return RTLIB::FMAX_F32; case 64 : return RTLIB::FMAX_F64; case 80: return RTLIB::FMAX_F80; case 128: return RTLIB::FMAX_F128; default: __builtin_unreachable (); } } while (0); |
469 | case TargetOpcode::G_FSQRT: |
470 | RTLIBCASE(SQRT_F)do { switch (Size) { case 32: return RTLIB::SQRT_F32; case 64 : return RTLIB::SQRT_F64; case 80: return RTLIB::SQRT_F80; case 128: return RTLIB::SQRT_F128; default: __builtin_unreachable (); } } while (0); |
471 | case TargetOpcode::G_FRINT: |
472 | RTLIBCASE(RINT_F)do { switch (Size) { case 32: return RTLIB::RINT_F32; case 64 : return RTLIB::RINT_F64; case 80: return RTLIB::RINT_F80; case 128: return RTLIB::RINT_F128; default: __builtin_unreachable (); } } while (0); |
473 | case TargetOpcode::G_FNEARBYINT: |
474 | RTLIBCASE(NEARBYINT_F)do { switch (Size) { case 32: return RTLIB::NEARBYINT_F32; case 64: return RTLIB::NEARBYINT_F64; case 80: return RTLIB::NEARBYINT_F80 ; case 128: return RTLIB::NEARBYINT_F128; default: __builtin_unreachable (); } } while (0); |
475 | case TargetOpcode::G_INTRINSIC_ROUNDEVEN: |
476 | RTLIBCASE(ROUNDEVEN_F)do { switch (Size) { case 32: return RTLIB::ROUNDEVEN_F32; case 64: return RTLIB::ROUNDEVEN_F64; case 80: return RTLIB::ROUNDEVEN_F80 ; case 128: return RTLIB::ROUNDEVEN_F128; default: __builtin_unreachable (); } } while (0); |
477 | } |
478 | llvm_unreachable("Unknown libcall function")__builtin_unreachable(); |
479 | } |
480 | |
481 | /// True if an instruction is in tail position in its caller. Intended for |
482 | /// legalizing libcalls as tail calls when possible. |
483 | static bool isLibCallInTailPosition(MachineInstr &MI, |
484 | const TargetInstrInfo &TII, |
485 | MachineRegisterInfo &MRI) { |
486 | MachineBasicBlock &MBB = *MI.getParent(); |
487 | const Function &F = MBB.getParent()->getFunction(); |
488 | |
489 | // Conservatively require the attributes of the call to match those of |
490 | // the return. Ignore NoAlias and NonNull because they don't affect the |
491 | // call sequence. |
492 | AttributeList CallerAttrs = F.getAttributes(); |
493 | if (AttrBuilder(CallerAttrs, AttributeList::ReturnIndex) |
494 | .removeAttribute(Attribute::NoAlias) |
495 | .removeAttribute(Attribute::NonNull) |
496 | .hasAttributes()) |
497 | return false; |
498 | |
499 | // It's not safe to eliminate the sign / zero extension of the return value. |
500 | if (CallerAttrs.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt) || |
501 | CallerAttrs.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt)) |
502 | return false; |
503 | |
504 | // Only tail call if the following instruction is a standard return or if we |
505 | // have a `thisreturn` callee, and a sequence like: |
506 | // |
507 | // G_MEMCPY %0, %1, %2 |
508 | // $x0 = COPY %0 |
509 | // RET_ReallyLR implicit $x0 |
510 | auto Next = next_nodbg(MI.getIterator(), MBB.instr_end()); |
511 | if (Next != MBB.instr_end() && Next->isCopy()) { |
512 | switch (MI.getOpcode()) { |
513 | default: |
514 | llvm_unreachable("unsupported opcode")__builtin_unreachable(); |
515 | case TargetOpcode::G_BZERO: |
516 | return false; |
517 | case TargetOpcode::G_MEMCPY: |
518 | case TargetOpcode::G_MEMMOVE: |
519 | case TargetOpcode::G_MEMSET: |
520 | break; |
521 | } |
522 | |
523 | Register VReg = MI.getOperand(0).getReg(); |
524 | if (!VReg.isVirtual() || VReg != Next->getOperand(1).getReg()) |
525 | return false; |
526 | |
527 | Register PReg = Next->getOperand(0).getReg(); |
528 | if (!PReg.isPhysical()) |
529 | return false; |
530 | |
531 | auto Ret = next_nodbg(Next, MBB.instr_end()); |
532 | if (Ret == MBB.instr_end() || !Ret->isReturn()) |
533 | return false; |
534 | |
535 | if (Ret->getNumImplicitOperands() != 1) |
536 | return false; |
537 | |
538 | if (PReg != Ret->getOperand(0).getReg()) |
539 | return false; |
540 | |
541 | // Skip over the COPY that we just validated. |
542 | Next = Ret; |
543 | } |
544 | |
545 | if (Next == MBB.instr_end() || TII.isTailCall(*Next) || !Next->isReturn()) |
546 | return false; |
547 | |
548 | return true; |
549 | } |
550 | |
551 | LegalizerHelper::LegalizeResult |
552 | llvm::createLibcall(MachineIRBuilder &MIRBuilder, const char *Name, |
553 | const CallLowering::ArgInfo &Result, |
554 | ArrayRef<CallLowering::ArgInfo> Args, |
555 | const CallingConv::ID CC) { |
556 | auto &CLI = *MIRBuilder.getMF().getSubtarget().getCallLowering(); |
557 | |
558 | CallLowering::CallLoweringInfo Info; |
559 | Info.CallConv = CC; |
560 | Info.Callee = MachineOperand::CreateES(Name); |
561 | Info.OrigRet = Result; |
562 | std::copy(Args.begin(), Args.end(), std::back_inserter(Info.OrigArgs)); |
563 | if (!CLI.lowerCall(MIRBuilder, Info)) |
564 | return LegalizerHelper::UnableToLegalize; |
565 | |
566 | return LegalizerHelper::Legalized; |
567 | } |
568 | |
569 | LegalizerHelper::LegalizeResult |
570 | llvm::createLibcall(MachineIRBuilder &MIRBuilder, RTLIB::Libcall Libcall, |
571 | const CallLowering::ArgInfo &Result, |
572 | ArrayRef<CallLowering::ArgInfo> Args) { |
573 | auto &TLI = *MIRBuilder.getMF().getSubtarget().getTargetLowering(); |
574 | const char *Name = TLI.getLibcallName(Libcall); |
575 | const CallingConv::ID CC = TLI.getLibcallCallingConv(Libcall); |
576 | return createLibcall(MIRBuilder, Name, Result, Args, CC); |
577 | } |
578 | |
579 | // Useful for libcalls where all operands have the same type. |
580 | static LegalizerHelper::LegalizeResult |
581 | simpleLibcall(MachineInstr &MI, MachineIRBuilder &MIRBuilder, unsigned Size, |
582 | Type *OpType) { |
583 | auto Libcall = getRTLibDesc(MI.getOpcode(), Size); |
584 | |
585 | // FIXME: What does the original arg index mean here? |
586 | SmallVector<CallLowering::ArgInfo, 3> Args; |
587 | for (unsigned i = 1; i < MI.getNumOperands(); i++) |
588 | Args.push_back({MI.getOperand(i).getReg(), OpType, 0}); |
589 | return createLibcall(MIRBuilder, Libcall, |
590 | {MI.getOperand(0).getReg(), OpType, 0}, Args); |
591 | } |
592 | |
593 | LegalizerHelper::LegalizeResult |
594 | llvm::createMemLibcall(MachineIRBuilder &MIRBuilder, MachineRegisterInfo &MRI, |
595 | MachineInstr &MI, LostDebugLocObserver &LocObserver) { |
596 | auto &Ctx = MIRBuilder.getMF().getFunction().getContext(); |
597 | |
598 | SmallVector<CallLowering::ArgInfo, 3> Args; |
599 | // Add all the args, except for the last which is an imm denoting 'tail'. |
600 | for (unsigned i = 0; i < MI.getNumOperands() - 1; ++i) { |
601 | Register Reg = MI.getOperand(i).getReg(); |
602 | |
603 | // Need derive an IR type for call lowering. |
604 | LLT OpLLT = MRI.getType(Reg); |
605 | Type *OpTy = nullptr; |
606 | if (OpLLT.isPointer()) |
607 | OpTy = Type::getInt8PtrTy(Ctx, OpLLT.getAddressSpace()); |
608 | else |
609 | OpTy = IntegerType::get(Ctx, OpLLT.getSizeInBits()); |
610 | Args.push_back({Reg, OpTy, 0}); |
611 | } |
612 | |
613 | auto &CLI = *MIRBuilder.getMF().getSubtarget().getCallLowering(); |
614 | auto &TLI = *MIRBuilder.getMF().getSubtarget().getTargetLowering(); |
615 | RTLIB::Libcall RTLibcall; |
616 | unsigned Opc = MI.getOpcode(); |
617 | switch (Opc) { |
618 | case TargetOpcode::G_BZERO: |
619 | RTLibcall = RTLIB::BZERO; |
620 | break; |
621 | case TargetOpcode::G_MEMCPY: |
622 | RTLibcall = RTLIB::MEMCPY; |
623 | Args[0].Flags[0].setReturned(); |
624 | break; |
625 | case TargetOpcode::G_MEMMOVE: |
626 | RTLibcall = RTLIB::MEMMOVE; |
627 | Args[0].Flags[0].setReturned(); |
628 | break; |
629 | case TargetOpcode::G_MEMSET: |
630 | RTLibcall = RTLIB::MEMSET; |
631 | Args[0].Flags[0].setReturned(); |
632 | break; |
633 | default: |
634 | llvm_unreachable("unsupported opcode")__builtin_unreachable(); |
635 | } |
636 | const char *Name = TLI.getLibcallName(RTLibcall); |
637 | |
638 | // Unsupported libcall on the target. |
639 | if (!Name) { |
640 | LLVM_DEBUG(dbgs() << ".. .. Could not find libcall name for "do { } while (false) |
641 | << MIRBuilder.getTII().getName(Opc) << "\n")do { } while (false); |
642 | return LegalizerHelper::UnableToLegalize; |
643 | } |
644 | |
645 | CallLowering::CallLoweringInfo Info; |
646 | Info.CallConv = TLI.getLibcallCallingConv(RTLibcall); |
647 | Info.Callee = MachineOperand::CreateES(Name); |
648 | Info.OrigRet = CallLowering::ArgInfo({0}, Type::getVoidTy(Ctx), 0); |
649 | Info.IsTailCall = MI.getOperand(MI.getNumOperands() - 1).getImm() && |
650 | isLibCallInTailPosition(MI, MIRBuilder.getTII(), MRI); |
651 | |
652 | std::copy(Args.begin(), Args.end(), std::back_inserter(Info.OrigArgs)); |
653 | if (!CLI.lowerCall(MIRBuilder, Info)) |
654 | return LegalizerHelper::UnableToLegalize; |
655 | |
656 | if (Info.LoweredTailCall) { |
657 | assert(Info.IsTailCall && "Lowered tail call when it wasn't a tail call?")((void)0); |
658 | |
659 | // Check debug locations before removing the return. |
660 | LocObserver.checkpoint(true); |
661 | |
662 | // We must have a return following the call (or debug insts) to get past |
663 | // isLibCallInTailPosition. |
664 | do { |
665 | MachineInstr *Next = MI.getNextNode(); |
666 | assert(Next &&((void)0) |
667 | (Next->isCopy() || Next->isReturn() || Next->isDebugInstr()) &&((void)0) |
668 | "Expected instr following MI to be return or debug inst?")((void)0); |
669 | // We lowered a tail call, so the call is now the return from the block. |
670 | // Delete the old return. |
671 | Next->eraseFromParent(); |
672 | } while (MI.getNextNode()); |
673 | |
674 | // We expect to lose the debug location from the return. |
675 | LocObserver.checkpoint(false); |
676 | } |
677 | |
678 | return LegalizerHelper::Legalized; |
679 | } |
680 | |
681 | static RTLIB::Libcall getConvRTLibDesc(unsigned Opcode, Type *ToType, |
682 | Type *FromType) { |
683 | auto ToMVT = MVT::getVT(ToType); |
684 | auto FromMVT = MVT::getVT(FromType); |
685 | |
686 | switch (Opcode) { |
687 | case TargetOpcode::G_FPEXT: |
688 | return RTLIB::getFPEXT(FromMVT, ToMVT); |
689 | case TargetOpcode::G_FPTRUNC: |
690 | return RTLIB::getFPROUND(FromMVT, ToMVT); |
691 | case TargetOpcode::G_FPTOSI: |
692 | return RTLIB::getFPTOSINT(FromMVT, ToMVT); |
693 | case TargetOpcode::G_FPTOUI: |
694 | return RTLIB::getFPTOUINT(FromMVT, ToMVT); |
695 | case TargetOpcode::G_SITOFP: |
696 | return RTLIB::getSINTTOFP(FromMVT, ToMVT); |
697 | case TargetOpcode::G_UITOFP: |
698 | return RTLIB::getUINTTOFP(FromMVT, ToMVT); |
699 | } |
700 | llvm_unreachable("Unsupported libcall function")__builtin_unreachable(); |
701 | } |
702 | |
703 | static LegalizerHelper::LegalizeResult |
704 | conversionLibcall(MachineInstr &MI, MachineIRBuilder &MIRBuilder, Type *ToType, |
705 | Type *FromType) { |
706 | RTLIB::Libcall Libcall = getConvRTLibDesc(MI.getOpcode(), ToType, FromType); |
707 | return createLibcall(MIRBuilder, Libcall, |
708 | {MI.getOperand(0).getReg(), ToType, 0}, |
709 | {{MI.getOperand(1).getReg(), FromType, 0}}); |
710 | } |
711 | |
712 | LegalizerHelper::LegalizeResult |
713 | LegalizerHelper::libcall(MachineInstr &MI, LostDebugLocObserver &LocObserver) { |
714 | LLT LLTy = MRI.getType(MI.getOperand(0).getReg()); |
715 | unsigned Size = LLTy.getSizeInBits(); |
716 | auto &Ctx = MIRBuilder.getMF().getFunction().getContext(); |
717 | |
718 | switch (MI.getOpcode()) { |
719 | default: |
720 | return UnableToLegalize; |
721 | case TargetOpcode::G_SDIV: |
722 | case TargetOpcode::G_UDIV: |
723 | case TargetOpcode::G_SREM: |
724 | case TargetOpcode::G_UREM: |
725 | case TargetOpcode::G_CTLZ_ZERO_UNDEF: { |
726 | Type *HLTy = IntegerType::get(Ctx, Size); |
727 | auto Status = simpleLibcall(MI, MIRBuilder, Size, HLTy); |
728 | if (Status != Legalized) |
729 | return Status; |
730 | break; |
731 | } |
732 | case TargetOpcode::G_FADD: |
733 | case TargetOpcode::G_FSUB: |
734 | case TargetOpcode::G_FMUL: |
735 | case TargetOpcode::G_FDIV: |
736 | case TargetOpcode::G_FMA: |
737 | case TargetOpcode::G_FPOW: |
738 | case TargetOpcode::G_FREM: |
739 | case TargetOpcode::G_FCOS: |
740 | case TargetOpcode::G_FSIN: |
741 | case TargetOpcode::G_FLOG10: |
742 | case TargetOpcode::G_FLOG: |
743 | case TargetOpcode::G_FLOG2: |
744 | case TargetOpcode::G_FEXP: |
745 | case TargetOpcode::G_FEXP2: |
746 | case TargetOpcode::G_FCEIL: |
747 | case TargetOpcode::G_FFLOOR: |
748 | case TargetOpcode::G_FMINNUM: |
749 | case TargetOpcode::G_FMAXNUM: |
750 | case TargetOpcode::G_FSQRT: |
751 | case TargetOpcode::G_FRINT: |
752 | case TargetOpcode::G_FNEARBYINT: |
753 | case TargetOpcode::G_INTRINSIC_ROUNDEVEN: { |
754 | Type *HLTy = getFloatTypeForLLT(Ctx, LLTy); |
755 | if (!HLTy || (Size != 32 && Size != 64 && Size != 80 && Size != 128)) { |
756 | LLVM_DEBUG(dbgs() << "No libcall available for type " << LLTy << ".\n")do { } while (false); |
757 | return UnableToLegalize; |
758 | } |
759 | auto Status = simpleLibcall(MI, MIRBuilder, Size, HLTy); |
760 | if (Status != Legalized) |
761 | return Status; |
762 | break; |
763 | } |
764 | case TargetOpcode::G_FPEXT: |
765 | case TargetOpcode::G_FPTRUNC: { |
766 | Type *FromTy = getFloatTypeForLLT(Ctx, MRI.getType(MI.getOperand(1).getReg())); |
767 | Type *ToTy = getFloatTypeForLLT(Ctx, MRI.getType(MI.getOperand(0).getReg())); |
768 | if (!FromTy || !ToTy) |
769 | return UnableToLegalize; |
770 | LegalizeResult Status = conversionLibcall(MI, MIRBuilder, ToTy, FromTy ); |
771 | if (Status != Legalized) |
772 | return Status; |
773 | break; |
774 | } |
775 | case TargetOpcode::G_FPTOSI: |
776 | case TargetOpcode::G_FPTOUI: { |
777 | // FIXME: Support other types |
778 | unsigned FromSize = MRI.getType(MI.getOperand(1).getReg()).getSizeInBits(); |
779 | unsigned ToSize = MRI.getType(MI.getOperand(0).getReg()).getSizeInBits(); |
780 | if ((ToSize != 32 && ToSize != 64) || (FromSize != 32 && FromSize != 64)) |
781 | return UnableToLegalize; |
782 | LegalizeResult Status = conversionLibcall( |
783 | MI, MIRBuilder, |
784 | ToSize == 32 ? Type::getInt32Ty(Ctx) : Type::getInt64Ty(Ctx), |
785 | FromSize == 64 ? Type::getDoubleTy(Ctx) : Type::getFloatTy(Ctx)); |
786 | if (Status != Legalized) |
787 | return Status; |
788 | break; |
789 | } |
790 | case TargetOpcode::G_SITOFP: |
791 | case TargetOpcode::G_UITOFP: { |
792 | // FIXME: Support other types |
793 | unsigned FromSize = MRI.getType(MI.getOperand(1).getReg()).getSizeInBits(); |
794 | unsigned ToSize = MRI.getType(MI.getOperand(0).getReg()).getSizeInBits(); |
795 | if ((FromSize != 32 && FromSize != 64) || (ToSize != 32 && ToSize != 64)) |
796 | return UnableToLegalize; |
797 | LegalizeResult Status = conversionLibcall( |
798 | MI, MIRBuilder, |
799 | ToSize == 64 ? Type::getDoubleTy(Ctx) : Type::getFloatTy(Ctx), |
800 | FromSize == 32 ? Type::getInt32Ty(Ctx) : Type::getInt64Ty(Ctx)); |
801 | if (Status != Legalized) |
802 | return Status; |
803 | break; |
804 | } |
805 | case TargetOpcode::G_BZERO: |
806 | case TargetOpcode::G_MEMCPY: |
807 | case TargetOpcode::G_MEMMOVE: |
808 | case TargetOpcode::G_MEMSET: { |
809 | LegalizeResult Result = |
810 | createMemLibcall(MIRBuilder, *MIRBuilder.getMRI(), MI, LocObserver); |
811 | if (Result != Legalized) |
812 | return Result; |
813 | MI.eraseFromParent(); |
814 | return Result; |
815 | } |
816 | } |
817 | |
818 | MI.eraseFromParent(); |
819 | return Legalized; |
820 | } |
821 | |
822 | LegalizerHelper::LegalizeResult LegalizerHelper::narrowScalar(MachineInstr &MI, |
823 | unsigned TypeIdx, |
824 | LLT NarrowTy) { |
825 | uint64_t SizeOp0 = MRI.getType(MI.getOperand(0).getReg()).getSizeInBits(); |
826 | uint64_t NarrowSize = NarrowTy.getSizeInBits(); |
827 | |
828 | switch (MI.getOpcode()) { |
829 | default: |
830 | return UnableToLegalize; |
831 | case TargetOpcode::G_IMPLICIT_DEF: { |
832 | Register DstReg = MI.getOperand(0).getReg(); |
833 | LLT DstTy = MRI.getType(DstReg); |
834 | |
835 | // If SizeOp0 is not an exact multiple of NarrowSize, emit |
836 | // G_ANYEXT(G_IMPLICIT_DEF). Cast result to vector if needed. |
837 | // FIXME: Although this would also be legal for the general case, it causes |
838 | // a lot of regressions in the emitted code (superfluous COPYs, artifact |
839 | // combines not being hit). This seems to be a problem related to the |
840 | // artifact combiner. |
841 | if (SizeOp0 % NarrowSize != 0) { |
842 | LLT ImplicitTy = NarrowTy; |
843 | if (DstTy.isVector()) |
844 | ImplicitTy = LLT::vector(DstTy.getElementCount(), ImplicitTy); |
845 | |
846 | Register ImplicitReg = MIRBuilder.buildUndef(ImplicitTy).getReg(0); |
847 | MIRBuilder.buildAnyExt(DstReg, ImplicitReg); |
848 | |
849 | MI.eraseFromParent(); |
850 | return Legalized; |
851 | } |
852 | |
853 | int NumParts = SizeOp0 / NarrowSize; |
854 | |
855 | SmallVector<Register, 2> DstRegs; |
856 | for (int i = 0; i < NumParts; ++i) |
857 | DstRegs.push_back(MIRBuilder.buildUndef(NarrowTy).getReg(0)); |
858 | |
859 | if (DstTy.isVector()) |
860 | MIRBuilder.buildBuildVector(DstReg, DstRegs); |
861 | else |
862 | MIRBuilder.buildMerge(DstReg, DstRegs); |
863 | MI.eraseFromParent(); |
864 | return Legalized; |
865 | } |
866 | case TargetOpcode::G_CONSTANT: { |
867 | LLT Ty = MRI.getType(MI.getOperand(0).getReg()); |
868 | const APInt &Val = MI.getOperand(1).getCImm()->getValue(); |
869 | unsigned TotalSize = Ty.getSizeInBits(); |
870 | unsigned NarrowSize = NarrowTy.getSizeInBits(); |
871 | int NumParts = TotalSize / NarrowSize; |
872 | |
873 | SmallVector<Register, 4> PartRegs; |
874 | for (int I = 0; I != NumParts; ++I) { |
875 | unsigned Offset = I * NarrowSize; |
876 | auto K = MIRBuilder.buildConstant(NarrowTy, |
877 | Val.lshr(Offset).trunc(NarrowSize)); |
878 | PartRegs.push_back(K.getReg(0)); |
879 | } |
880 | |
881 | LLT LeftoverTy; |
882 | unsigned LeftoverBits = TotalSize - NumParts * NarrowSize; |
883 | SmallVector<Register, 1> LeftoverRegs; |
884 | if (LeftoverBits != 0) { |
885 | LeftoverTy = LLT::scalar(LeftoverBits); |
886 | auto K = MIRBuilder.buildConstant( |
887 | LeftoverTy, |
888 | Val.lshr(NumParts * NarrowSize).trunc(LeftoverBits)); |
889 | LeftoverRegs.push_back(K.getReg(0)); |
890 | } |
891 | |
892 | insertParts(MI.getOperand(0).getReg(), |
893 | Ty, NarrowTy, PartRegs, LeftoverTy, LeftoverRegs); |
894 | |
895 | MI.eraseFromParent(); |
896 | return Legalized; |
897 | } |
898 | case TargetOpcode::G_SEXT: |
899 | case TargetOpcode::G_ZEXT: |
900 | case TargetOpcode::G_ANYEXT: |
901 | return narrowScalarExt(MI, TypeIdx, NarrowTy); |
902 | case TargetOpcode::G_TRUNC: { |
903 | if (TypeIdx != 1) |
904 | return UnableToLegalize; |
905 | |
906 | uint64_t SizeOp1 = MRI.getType(MI.getOperand(1).getReg()).getSizeInBits(); |
907 | if (NarrowTy.getSizeInBits() * 2 != SizeOp1) { |
908 | LLVM_DEBUG(dbgs() << "Can't narrow trunc to type " << NarrowTy << "\n")do { } while (false); |
909 | return UnableToLegalize; |
910 | } |
911 | |
912 | auto Unmerge = MIRBuilder.buildUnmerge(NarrowTy, MI.getOperand(1)); |
913 | MIRBuilder.buildCopy(MI.getOperand(0), Unmerge.getReg(0)); |
914 | MI.eraseFromParent(); |
915 | return Legalized; |
916 | } |
917 | |
918 | case TargetOpcode::G_FREEZE: |
919 | return reduceOperationWidth(MI, TypeIdx, NarrowTy); |
920 | case TargetOpcode::G_ADD: |
921 | case TargetOpcode::G_SUB: |
922 | case TargetOpcode::G_SADDO: |
923 | case TargetOpcode::G_SSUBO: |
924 | case TargetOpcode::G_SADDE: |
925 | case TargetOpcode::G_SSUBE: |
926 | case TargetOpcode::G_UADDO: |
927 | case TargetOpcode::G_USUBO: |
928 | case TargetOpcode::G_UADDE: |
929 | case TargetOpcode::G_USUBE: |
930 | return narrowScalarAddSub(MI, TypeIdx, NarrowTy); |
931 | case TargetOpcode::G_MUL: |
932 | case TargetOpcode::G_UMULH: |
933 | return narrowScalarMul(MI, NarrowTy); |
934 | case TargetOpcode::G_EXTRACT: |
935 | return narrowScalarExtract(MI, TypeIdx, NarrowTy); |
936 | case TargetOpcode::G_INSERT: |
937 | return narrowScalarInsert(MI, TypeIdx, NarrowTy); |
938 | case TargetOpcode::G_LOAD: { |
939 | auto &LoadMI = cast<GLoad>(MI); |
940 | Register DstReg = LoadMI.getDstReg(); |
941 | LLT DstTy = MRI.getType(DstReg); |
942 | if (DstTy.isVector()) |
943 | return UnableToLegalize; |
944 | |
945 | if (8 * LoadMI.getMemSize() != DstTy.getSizeInBits()) { |
946 | Register TmpReg = MRI.createGenericVirtualRegister(NarrowTy); |
947 | MIRBuilder.buildLoad(TmpReg, LoadMI.getPointerReg(), LoadMI.getMMO()); |
948 | MIRBuilder.buildAnyExt(DstReg, TmpReg); |
949 | LoadMI.eraseFromParent(); |
950 | return Legalized; |
951 | } |
952 | |
953 | return reduceLoadStoreWidth(LoadMI, TypeIdx, NarrowTy); |
954 | } |
955 | case TargetOpcode::G_ZEXTLOAD: |
956 | case TargetOpcode::G_SEXTLOAD: { |
957 | auto &LoadMI = cast<GExtLoad>(MI); |
958 | Register DstReg = LoadMI.getDstReg(); |
959 | Register PtrReg = LoadMI.getPointerReg(); |
960 | |
961 | Register TmpReg = MRI.createGenericVirtualRegister(NarrowTy); |
962 | auto &MMO = LoadMI.getMMO(); |
963 | unsigned MemSize = MMO.getSizeInBits(); |
964 | |
965 | if (MemSize == NarrowSize) { |
966 | MIRBuilder.buildLoad(TmpReg, PtrReg, MMO); |
967 | } else if (MemSize < NarrowSize) { |
968 | MIRBuilder.buildLoadInstr(LoadMI.getOpcode(), TmpReg, PtrReg, MMO); |
969 | } else if (MemSize > NarrowSize) { |
970 | // FIXME: Need to split the load. |
971 | return UnableToLegalize; |
972 | } |
973 | |
974 | if (isa<GZExtLoad>(LoadMI)) |
975 | MIRBuilder.buildZExt(DstReg, TmpReg); |
976 | else |
977 | MIRBuilder.buildSExt(DstReg, TmpReg); |
978 | |
979 | LoadMI.eraseFromParent(); |
980 | return Legalized; |
981 | } |
982 | case TargetOpcode::G_STORE: { |
983 | auto &StoreMI = cast<GStore>(MI); |
984 | |
985 | Register SrcReg = StoreMI.getValueReg(); |
986 | LLT SrcTy = MRI.getType(SrcReg); |
987 | if (SrcTy.isVector()) |
988 | return UnableToLegalize; |
989 | |
990 | int NumParts = SizeOp0 / NarrowSize; |
991 | unsigned HandledSize = NumParts * NarrowTy.getSizeInBits(); |
992 | unsigned LeftoverBits = SrcTy.getSizeInBits() - HandledSize; |
993 | if (SrcTy.isVector() && LeftoverBits != 0) |
994 | return UnableToLegalize; |
995 | |
996 | if (8 * StoreMI.getMemSize() != SrcTy.getSizeInBits()) { |
997 | Register TmpReg = MRI.createGenericVirtualRegister(NarrowTy); |
998 | MIRBuilder.buildTrunc(TmpReg, SrcReg); |
999 | MIRBuilder.buildStore(TmpReg, StoreMI.getPointerReg(), StoreMI.getMMO()); |
1000 | StoreMI.eraseFromParent(); |
1001 | return Legalized; |
1002 | } |
1003 | |
1004 | return reduceLoadStoreWidth(StoreMI, 0, NarrowTy); |
1005 | } |
1006 | case TargetOpcode::G_SELECT: |
1007 | return narrowScalarSelect(MI, TypeIdx, NarrowTy); |
1008 | case TargetOpcode::G_AND: |
1009 | case TargetOpcode::G_OR: |
1010 | case TargetOpcode::G_XOR: { |
1011 | // Legalize bitwise operation: |
1012 | // A = BinOp<Ty> B, C |
1013 | // into: |
1014 | // B1, ..., BN = G_UNMERGE_VALUES B |
1015 | // C1, ..., CN = G_UNMERGE_VALUES C |
1016 | // A1 = BinOp<Ty/N> B1, C2 |
1017 | // ... |
1018 | // AN = BinOp<Ty/N> BN, CN |
1019 | // A = G_MERGE_VALUES A1, ..., AN |
1020 | return narrowScalarBasic(MI, TypeIdx, NarrowTy); |
1021 | } |
1022 | case TargetOpcode::G_SHL: |
1023 | case TargetOpcode::G_LSHR: |
1024 | case TargetOpcode::G_ASHR: |
1025 | return narrowScalarShift(MI, TypeIdx, NarrowTy); |
1026 | case TargetOpcode::G_CTLZ: |
1027 | case TargetOpcode::G_CTLZ_ZERO_UNDEF: |
1028 | case TargetOpcode::G_CTTZ: |
1029 | case TargetOpcode::G_CTTZ_ZERO_UNDEF: |
1030 | case TargetOpcode::G_CTPOP: |
1031 | if (TypeIdx == 1) |
1032 | switch (MI.getOpcode()) { |
1033 | case TargetOpcode::G_CTLZ: |
1034 | case TargetOpcode::G_CTLZ_ZERO_UNDEF: |
1035 | return narrowScalarCTLZ(MI, TypeIdx, NarrowTy); |
1036 | case TargetOpcode::G_CTTZ: |
1037 | case TargetOpcode::G_CTTZ_ZERO_UNDEF: |
1038 | return narrowScalarCTTZ(MI, TypeIdx, NarrowTy); |
1039 | case TargetOpcode::G_CTPOP: |
1040 | return narrowScalarCTPOP(MI, TypeIdx, NarrowTy); |
1041 | default: |
1042 | return UnableToLegalize; |
1043 | } |
1044 | |
1045 | Observer.changingInstr(MI); |
1046 | narrowScalarDst(MI, NarrowTy, 0, TargetOpcode::G_ZEXT); |
1047 | Observer.changedInstr(MI); |
1048 | return Legalized; |
1049 | case TargetOpcode::G_INTTOPTR: |
1050 | if (TypeIdx != 1) |
1051 | return UnableToLegalize; |
1052 | |
1053 | Observer.changingInstr(MI); |
1054 | narrowScalarSrc(MI, NarrowTy, 1); |
1055 | Observer.changedInstr(MI); |
1056 | return Legalized; |
1057 | case TargetOpcode::G_PTRTOINT: |
1058 | if (TypeIdx != 0) |
1059 | return UnableToLegalize; |
1060 | |
1061 | Observer.changingInstr(MI); |
1062 | narrowScalarDst(MI, NarrowTy, 0, TargetOpcode::G_ZEXT); |
1063 | Observer.changedInstr(MI); |
1064 | return Legalized; |
1065 | case TargetOpcode::G_PHI: { |
1066 | // FIXME: add support for when SizeOp0 isn't an exact multiple of |
1067 | // NarrowSize. |
1068 | if (SizeOp0 % NarrowSize != 0) |
1069 | return UnableToLegalize; |
1070 | |
1071 | unsigned NumParts = SizeOp0 / NarrowSize; |
1072 | SmallVector<Register, 2> DstRegs(NumParts); |
1073 | SmallVector<SmallVector<Register, 2>, 2> SrcRegs(MI.getNumOperands() / 2); |
1074 | Observer.changingInstr(MI); |
1075 | for (unsigned i = 1; i < MI.getNumOperands(); i += 2) { |
1076 | MachineBasicBlock &OpMBB = *MI.getOperand(i + 1).getMBB(); |
1077 | MIRBuilder.setInsertPt(OpMBB, OpMBB.getFirstTerminator()); |
1078 | extractParts(MI.getOperand(i).getReg(), NarrowTy, NumParts, |
1079 | SrcRegs[i / 2]); |
1080 | } |
1081 | MachineBasicBlock &MBB = *MI.getParent(); |
1082 | MIRBuilder.setInsertPt(MBB, MI); |
1083 | for (unsigned i = 0; i < NumParts; ++i) { |
1084 | DstRegs[i] = MRI.createGenericVirtualRegister(NarrowTy); |
1085 | MachineInstrBuilder MIB = |
1086 | MIRBuilder.buildInstr(TargetOpcode::G_PHI).addDef(DstRegs[i]); |
1087 | for (unsigned j = 1; j < MI.getNumOperands(); j += 2) |
1088 | MIB.addUse(SrcRegs[j / 2][i]).add(MI.getOperand(j + 1)); |
1089 | } |
1090 | MIRBuilder.setInsertPt(MBB, MBB.getFirstNonPHI()); |
1091 | MIRBuilder.buildMerge(MI.getOperand(0), DstRegs); |
1092 | Observer.changedInstr(MI); |
1093 | MI.eraseFromParent(); |
1094 | return Legalized; |
1095 | } |
1096 | case TargetOpcode::G_EXTRACT_VECTOR_ELT: |
1097 | case TargetOpcode::G_INSERT_VECTOR_ELT: { |
1098 | if (TypeIdx != 2) |
1099 | return UnableToLegalize; |
1100 | |
1101 | int OpIdx = MI.getOpcode() == TargetOpcode::G_EXTRACT_VECTOR_ELT ? 2 : 3; |
1102 | Observer.changingInstr(MI); |
1103 | narrowScalarSrc(MI, NarrowTy, OpIdx); |
1104 | Observer.changedInstr(MI); |
1105 | return Legalized; |
1106 | } |
1107 | case TargetOpcode::G_ICMP: { |
1108 | Register LHS = MI.getOperand(2).getReg(); |
1109 | LLT SrcTy = MRI.getType(LHS); |
1110 | uint64_t SrcSize = SrcTy.getSizeInBits(); |
1111 | CmpInst::Predicate Pred = |
1112 | static_cast<CmpInst::Predicate>(MI.getOperand(1).getPredicate()); |
1113 | |
1114 | // TODO: Handle the non-equality case for weird sizes. |
1115 | if (NarrowSize * 2 != SrcSize && !ICmpInst::isEquality(Pred)) |
1116 | return UnableToLegalize; |
1117 | |
1118 | LLT LeftoverTy; // Example: s88 -> s64 (NarrowTy) + s24 (leftover) |
1119 | SmallVector<Register, 4> LHSPartRegs, LHSLeftoverRegs; |
1120 | if (!extractParts(LHS, SrcTy, NarrowTy, LeftoverTy, LHSPartRegs, |
1121 | LHSLeftoverRegs)) |
1122 | return UnableToLegalize; |
1123 | |
1124 | LLT Unused; // Matches LeftoverTy; G_ICMP LHS and RHS are the same type. |
1125 | SmallVector<Register, 4> RHSPartRegs, RHSLeftoverRegs; |
1126 | if (!extractParts(MI.getOperand(3).getReg(), SrcTy, NarrowTy, Unused, |
1127 | RHSPartRegs, RHSLeftoverRegs)) |
1128 | return UnableToLegalize; |
1129 | |
1130 | // We now have the LHS and RHS of the compare split into narrow-type |
1131 | // registers, plus potentially some leftover type. |
1132 | Register Dst = MI.getOperand(0).getReg(); |
1133 | LLT ResTy = MRI.getType(Dst); |
1134 | if (ICmpInst::isEquality(Pred)) { |
1135 | // For each part on the LHS and RHS, keep track of the result of XOR-ing |
1136 | // them together. For each equal part, the result should be all 0s. For |
1137 | // each non-equal part, we'll get at least one 1. |
1138 | auto Zero = MIRBuilder.buildConstant(NarrowTy, 0); |
1139 | SmallVector<Register, 4> Xors; |
1140 | for (auto LHSAndRHS : zip(LHSPartRegs, RHSPartRegs)) { |
1141 | auto LHS = std::get<0>(LHSAndRHS); |
1142 | auto RHS = std::get<1>(LHSAndRHS); |
1143 | auto Xor = MIRBuilder.buildXor(NarrowTy, LHS, RHS).getReg(0); |
1144 | Xors.push_back(Xor); |
1145 | } |
1146 | |
1147 | // Build a G_XOR for each leftover register. Each G_XOR must be widened |
1148 | // to the desired narrow type so that we can OR them together later. |
1149 | SmallVector<Register, 4> WidenedXors; |
1150 | for (auto LHSAndRHS : zip(LHSLeftoverRegs, RHSLeftoverRegs)) { |
1151 | auto LHS = std::get<0>(LHSAndRHS); |
1152 | auto RHS = std::get<1>(LHSAndRHS); |
1153 | auto Xor = MIRBuilder.buildXor(LeftoverTy, LHS, RHS).getReg(0); |
1154 | LLT GCDTy = extractGCDType(WidenedXors, NarrowTy, LeftoverTy, Xor); |
1155 | buildLCMMergePieces(LeftoverTy, NarrowTy, GCDTy, WidenedXors, |
1156 | /* PadStrategy = */ TargetOpcode::G_ZEXT); |
1157 | Xors.insert(Xors.end(), WidenedXors.begin(), WidenedXors.end()); |
1158 | } |
1159 | |
1160 | // Now, for each part we broke up, we know if they are equal/not equal |
1161 | // based off the G_XOR. We can OR these all together and compare against |
1162 | // 0 to get the result. |
1163 | assert(Xors.size() >= 2 && "Should have gotten at least two Xors?")((void)0); |
1164 | auto Or = MIRBuilder.buildOr(NarrowTy, Xors[0], Xors[1]); |
1165 | for (unsigned I = 2, E = Xors.size(); I < E; ++I) |
1166 | Or = MIRBuilder.buildOr(NarrowTy, Or, Xors[I]); |
1167 | MIRBuilder.buildICmp(Pred, Dst, Or, Zero); |
1168 | } else { |
1169 | // TODO: Handle non-power-of-two types. |
1170 | assert(LHSPartRegs.size() == 2 && "Expected exactly 2 LHS part regs?")((void)0); |
1171 | assert(RHSPartRegs.size() == 2 && "Expected exactly 2 RHS part regs?")((void)0); |
1172 | Register LHSL = LHSPartRegs[0]; |
1173 | Register LHSH = LHSPartRegs[1]; |
1174 | Register RHSL = RHSPartRegs[0]; |
1175 | Register RHSH = RHSPartRegs[1]; |
1176 | MachineInstrBuilder CmpH = MIRBuilder.buildICmp(Pred, ResTy, LHSH, RHSH); |
1177 | MachineInstrBuilder CmpHEQ = |
1178 | MIRBuilder.buildICmp(CmpInst::Predicate::ICMP_EQ, ResTy, LHSH, RHSH); |
1179 | MachineInstrBuilder CmpLU = MIRBuilder.buildICmp( |
1180 | ICmpInst::getUnsignedPredicate(Pred), ResTy, LHSL, RHSL); |
1181 | MIRBuilder.buildSelect(Dst, CmpHEQ, CmpLU, CmpH); |
1182 | } |
1183 | MI.eraseFromParent(); |
1184 | return Legalized; |
1185 | } |
1186 | case TargetOpcode::G_SEXT_INREG: { |
1187 | if (TypeIdx != 0) |
1188 | return UnableToLegalize; |
1189 | |
1190 | int64_t SizeInBits = MI.getOperand(2).getImm(); |
1191 | |
1192 | // So long as the new type has more bits than the bits we're extending we |
1193 | // don't need to break it apart. |
1194 | if (NarrowTy.getScalarSizeInBits() >= SizeInBits) { |
1195 | Observer.changingInstr(MI); |
1196 | // We don't lose any non-extension bits by truncating the src and |
1197 | // sign-extending the dst. |
1198 | MachineOperand &MO1 = MI.getOperand(1); |
1199 | auto TruncMIB = MIRBuilder.buildTrunc(NarrowTy, MO1); |
1200 | MO1.setReg(TruncMIB.getReg(0)); |
1201 | |
1202 | MachineOperand &MO2 = MI.getOperand(0); |
1203 | Register DstExt = MRI.createGenericVirtualRegister(NarrowTy); |
1204 | MIRBuilder.setInsertPt(MIRBuilder.getMBB(), ++MIRBuilder.getInsertPt()); |
1205 | MIRBuilder.buildSExt(MO2, DstExt); |
1206 | MO2.setReg(DstExt); |
1207 | Observer.changedInstr(MI); |
1208 | return Legalized; |
1209 | } |
1210 | |
1211 | // Break it apart. Components below the extension point are unmodified. The |
1212 | // component containing the extension point becomes a narrower SEXT_INREG. |
1213 | // Components above it are ashr'd from the component containing the |
1214 | // extension point. |
1215 | if (SizeOp0 % NarrowSize != 0) |
1216 | return UnableToLegalize; |
1217 | int NumParts = SizeOp0 / NarrowSize; |
1218 | |
1219 | // List the registers where the destination will be scattered. |
1220 | SmallVector<Register, 2> DstRegs; |
1221 | // List the registers where the source will be split. |
1222 | SmallVector<Register, 2> SrcRegs; |
1223 | |
1224 | // Create all the temporary registers. |
1225 | for (int i = 0; i < NumParts; ++i) { |
1226 | Register SrcReg = MRI.createGenericVirtualRegister(NarrowTy); |
1227 | |
1228 | SrcRegs.push_back(SrcReg); |
1229 | } |
1230 | |
1231 | // Explode the big arguments into smaller chunks. |
1232 | MIRBuilder.buildUnmerge(SrcRegs, MI.getOperand(1)); |
1233 | |
1234 | Register AshrCstReg = |
1235 | MIRBuilder.buildConstant(NarrowTy, NarrowTy.getScalarSizeInBits() - 1) |
1236 | .getReg(0); |
1237 | Register FullExtensionReg = 0; |
1238 | Register PartialExtensionReg = 0; |
1239 | |
1240 | // Do the operation on each small part. |
1241 | for (int i = 0; i < NumParts; ++i) { |
1242 | if ((i + 1) * NarrowTy.getScalarSizeInBits() < SizeInBits) |
1243 | DstRegs.push_back(SrcRegs[i]); |
1244 | else if (i * NarrowTy.getScalarSizeInBits() > SizeInBits) { |
1245 | assert(PartialExtensionReg &&((void)0) |
1246 | "Expected to visit partial extension before full")((void)0); |
1247 | if (FullExtensionReg) { |
1248 | DstRegs.push_back(FullExtensionReg); |
1249 | continue; |
1250 | } |
1251 | DstRegs.push_back( |
1252 | MIRBuilder.buildAShr(NarrowTy, PartialExtensionReg, AshrCstReg) |
1253 | .getReg(0)); |
1254 | FullExtensionReg = DstRegs.back(); |
1255 | } else { |
1256 | DstRegs.push_back( |
1257 | MIRBuilder |
1258 | .buildInstr( |
1259 | TargetOpcode::G_SEXT_INREG, {NarrowTy}, |
1260 | {SrcRegs[i], SizeInBits % NarrowTy.getScalarSizeInBits()}) |
1261 | .getReg(0)); |
1262 | PartialExtensionReg = DstRegs.back(); |
1263 | } |
1264 | } |
1265 | |
1266 | // Gather the destination registers into the final destination. |
1267 | Register DstReg = MI.getOperand(0).getReg(); |
1268 | MIRBuilder.buildMerge(DstReg, DstRegs); |
1269 | MI.eraseFromParent(); |
1270 | return Legalized; |
1271 | } |
1272 | case TargetOpcode::G_BSWAP: |
1273 | case TargetOpcode::G_BITREVERSE: { |
1274 | if (SizeOp0 % NarrowSize != 0) |
1275 | return UnableToLegalize; |
1276 | |
1277 | Observer.changingInstr(MI); |
1278 | SmallVector<Register, 2> SrcRegs, DstRegs; |
1279 | unsigned NumParts = SizeOp0 / NarrowSize; |
1280 | extractParts(MI.getOperand(1).getReg(), NarrowTy, NumParts, SrcRegs); |
1281 | |
1282 | for (unsigned i = 0; i < NumParts; ++i) { |
1283 | auto DstPart = MIRBuilder.buildInstr(MI.getOpcode(), {NarrowTy}, |
1284 | {SrcRegs[NumParts - 1 - i]}); |
1285 | DstRegs.push_back(DstPart.getReg(0)); |
1286 | } |
1287 | |
1288 | MIRBuilder.buildMerge(MI.getOperand(0), DstRegs); |
1289 | |
1290 | Observer.changedInstr(MI); |
1291 | MI.eraseFromParent(); |
1292 | return Legalized; |
1293 | } |
1294 | case TargetOpcode::G_PTR_ADD: |
1295 | case TargetOpcode::G_PTRMASK: { |
1296 | if (TypeIdx != 1) |
1297 | return UnableToLegalize; |
1298 | Observer.changingInstr(MI); |
1299 | narrowScalarSrc(MI, NarrowTy, 2); |
1300 | Observer.changedInstr(MI); |
1301 | return Legalized; |
1302 | } |
1303 | case TargetOpcode::G_FPTOUI: |
1304 | case TargetOpcode::G_FPTOSI: |
1305 | return narrowScalarFPTOI(MI, TypeIdx, NarrowTy); |
1306 | case TargetOpcode::G_FPEXT: |
1307 | if (TypeIdx != 0) |
1308 | return UnableToLegalize; |
1309 | Observer.changingInstr(MI); |
1310 | narrowScalarDst(MI, NarrowTy, 0, TargetOpcode::G_FPEXT); |
1311 | Observer.changedInstr(MI); |
1312 | return Legalized; |
1313 | } |
1314 | } |
1315 | |
1316 | Register LegalizerHelper::coerceToScalar(Register Val) { |
1317 | LLT Ty = MRI.getType(Val); |
1318 | if (Ty.isScalar()) |
1319 | return Val; |
1320 | |
1321 | const DataLayout &DL = MIRBuilder.getDataLayout(); |
1322 | LLT NewTy = LLT::scalar(Ty.getSizeInBits()); |
1323 | if (Ty.isPointer()) { |
1324 | if (DL.isNonIntegralAddressSpace(Ty.getAddressSpace())) |
1325 | return Register(); |
1326 | return MIRBuilder.buildPtrToInt(NewTy, Val).getReg(0); |
1327 | } |
1328 | |
1329 | Register NewVal = Val; |
1330 | |
1331 | assert(Ty.isVector())((void)0); |
1332 | LLT EltTy = Ty.getElementType(); |
1333 | if (EltTy.isPointer()) |
1334 | NewVal = MIRBuilder.buildPtrToInt(NewTy, NewVal).getReg(0); |
1335 | return MIRBuilder.buildBitcast(NewTy, NewVal).getReg(0); |
1336 | } |
1337 | |
1338 | void LegalizerHelper::widenScalarSrc(MachineInstr &MI, LLT WideTy, |
1339 | unsigned OpIdx, unsigned ExtOpcode) { |
1340 | MachineOperand &MO = MI.getOperand(OpIdx); |
1341 | auto ExtB = MIRBuilder.buildInstr(ExtOpcode, {WideTy}, {MO}); |
1342 | MO.setReg(ExtB.getReg(0)); |
1343 | } |
1344 | |
1345 | void LegalizerHelper::narrowScalarSrc(MachineInstr &MI, LLT NarrowTy, |
1346 | unsigned OpIdx) { |
1347 | MachineOperand &MO = MI.getOperand(OpIdx); |
1348 | auto ExtB = MIRBuilder.buildTrunc(NarrowTy, MO); |
1349 | MO.setReg(ExtB.getReg(0)); |
1350 | } |
1351 | |
1352 | void LegalizerHelper::widenScalarDst(MachineInstr &MI, LLT WideTy, |
1353 | unsigned OpIdx, unsigned TruncOpcode) { |
1354 | MachineOperand &MO = MI.getOperand(OpIdx); |
1355 | Register DstExt = MRI.createGenericVirtualRegister(WideTy); |
1356 | MIRBuilder.setInsertPt(MIRBuilder.getMBB(), ++MIRBuilder.getInsertPt()); |
1357 | MIRBuilder.buildInstr(TruncOpcode, {MO}, {DstExt}); |
1358 | MO.setReg(DstExt); |
1359 | } |
1360 | |
1361 | void LegalizerHelper::narrowScalarDst(MachineInstr &MI, LLT NarrowTy, |
1362 | unsigned OpIdx, unsigned ExtOpcode) { |
1363 | MachineOperand &MO = MI.getOperand(OpIdx); |
1364 | Register DstTrunc = MRI.createGenericVirtualRegister(NarrowTy); |
1365 | MIRBuilder.setInsertPt(MIRBuilder.getMBB(), ++MIRBuilder.getInsertPt()); |
1366 | MIRBuilder.buildInstr(ExtOpcode, {MO}, {DstTrunc}); |
1367 | MO.setReg(DstTrunc); |
1368 | } |
1369 | |
1370 | void LegalizerHelper::moreElementsVectorDst(MachineInstr &MI, LLT WideTy, |
1371 | unsigned OpIdx) { |
1372 | MachineOperand &MO = MI.getOperand(OpIdx); |
1373 | MIRBuilder.setInsertPt(MIRBuilder.getMBB(), ++MIRBuilder.getInsertPt()); |
1374 | MO.setReg(widenWithUnmerge(WideTy, MO.getReg())); |
1375 | } |
1376 | |
1377 | void LegalizerHelper::moreElementsVectorSrc(MachineInstr &MI, LLT MoreTy, |
1378 | unsigned OpIdx) { |
1379 | MachineOperand &MO = MI.getOperand(OpIdx); |
1380 | |
1381 | LLT OldTy = MRI.getType(MO.getReg()); |
1382 | unsigned OldElts = OldTy.getNumElements(); |
1383 | unsigned NewElts = MoreTy.getNumElements(); |
1384 | |
1385 | unsigned NumParts = NewElts / OldElts; |
1386 | |
1387 | // Use concat_vectors if the result is a multiple of the number of elements. |
1388 | if (NumParts * OldElts == NewElts) { |
1389 | SmallVector<Register, 8> Parts; |
1390 | Parts.push_back(MO.getReg()); |
1391 | |
1392 | Register ImpDef = MIRBuilder.buildUndef(OldTy).getReg(0); |
1393 | for (unsigned I = 1; I != NumParts; ++I) |
1394 | Parts.push_back(ImpDef); |
1395 | |
1396 | auto Concat = MIRBuilder.buildConcatVectors(MoreTy, Parts); |
1397 | MO.setReg(Concat.getReg(0)); |
1398 | return; |
1399 | } |
1400 | |
1401 | Register MoreReg = MRI.createGenericVirtualRegister(MoreTy); |
1402 | Register ImpDef = MIRBuilder.buildUndef(MoreTy).getReg(0); |
1403 | MIRBuilder.buildInsert(MoreReg, ImpDef, MO.getReg(), 0); |
1404 | MO.setReg(MoreReg); |
1405 | } |
1406 | |
1407 | void LegalizerHelper::bitcastSrc(MachineInstr &MI, LLT CastTy, unsigned OpIdx) { |
1408 | MachineOperand &Op = MI.getOperand(OpIdx); |
1409 | Op.setReg(MIRBuilder.buildBitcast(CastTy, Op).getReg(0)); |
1410 | } |
1411 | |
1412 | void LegalizerHelper::bitcastDst(MachineInstr &MI, LLT CastTy, unsigned OpIdx) { |
1413 | MachineOperand &MO = MI.getOperand(OpIdx); |
1414 | Register CastDst = MRI.createGenericVirtualRegister(CastTy); |
1415 | MIRBuilder.setInsertPt(MIRBuilder.getMBB(), ++MIRBuilder.getInsertPt()); |
1416 | MIRBuilder.buildBitcast(MO, CastDst); |
1417 | MO.setReg(CastDst); |
1418 | } |
1419 | |
1420 | LegalizerHelper::LegalizeResult |
1421 | LegalizerHelper::widenScalarMergeValues(MachineInstr &MI, unsigned TypeIdx, |
1422 | LLT WideTy) { |
1423 | if (TypeIdx != 1) |
1424 | return UnableToLegalize; |
1425 | |
1426 | Register DstReg = MI.getOperand(0).getReg(); |
1427 | LLT DstTy = MRI.getType(DstReg); |
1428 | if (DstTy.isVector()) |
1429 | return UnableToLegalize; |
1430 | |
1431 | Register Src1 = MI.getOperand(1).getReg(); |
1432 | LLT SrcTy = MRI.getType(Src1); |
1433 | const int DstSize = DstTy.getSizeInBits(); |
1434 | const int SrcSize = SrcTy.getSizeInBits(); |
1435 | const int WideSize = WideTy.getSizeInBits(); |
1436 | const int NumMerge = (DstSize + WideSize - 1) / WideSize; |
1437 | |
1438 | unsigned NumOps = MI.getNumOperands(); |
1439 | unsigned NumSrc = MI.getNumOperands() - 1; |
1440 | unsigned PartSize = DstTy.getSizeInBits() / NumSrc; |
1441 | |
1442 | if (WideSize >= DstSize) { |
1443 | // Directly pack the bits in the target type. |
1444 | Register ResultReg = MIRBuilder.buildZExt(WideTy, Src1).getReg(0); |
1445 | |
1446 | for (unsigned I = 2; I != NumOps; ++I) { |
1447 | const unsigned Offset = (I - 1) * PartSize; |
1448 | |
1449 | Register SrcReg = MI.getOperand(I).getReg(); |
1450 | assert(MRI.getType(SrcReg) == LLT::scalar(PartSize))((void)0); |
1451 | |
1452 | auto ZextInput = MIRBuilder.buildZExt(WideTy, SrcReg); |
1453 | |
1454 | Register NextResult = I + 1 == NumOps && WideTy == DstTy ? DstReg : |
1455 | MRI.createGenericVirtualRegister(WideTy); |
1456 | |
1457 | auto ShiftAmt = MIRBuilder.buildConstant(WideTy, Offset); |
1458 | auto Shl = MIRBuilder.buildShl(WideTy, ZextInput, ShiftAmt); |
1459 | MIRBuilder.buildOr(NextResult, ResultReg, Shl); |
1460 | ResultReg = NextResult; |
1461 | } |
1462 | |
1463 | if (WideSize > DstSize) |
1464 | MIRBuilder.buildTrunc(DstReg, ResultReg); |
1465 | else if (DstTy.isPointer()) |
1466 | MIRBuilder.buildIntToPtr(DstReg, ResultReg); |
1467 | |
1468 | MI.eraseFromParent(); |
1469 | return Legalized; |
1470 | } |
1471 | |
1472 | // Unmerge the original values to the GCD type, and recombine to the next |
1473 | // multiple greater than the original type. |
1474 | // |
1475 | // %3:_(s12) = G_MERGE_VALUES %0:_(s4), %1:_(s4), %2:_(s4) -> s6 |
1476 | // %4:_(s2), %5:_(s2) = G_UNMERGE_VALUES %0 |
1477 | // %6:_(s2), %7:_(s2) = G_UNMERGE_VALUES %1 |
1478 | // %8:_(s2), %9:_(s2) = G_UNMERGE_VALUES %2 |
1479 | // %10:_(s6) = G_MERGE_VALUES %4, %5, %6 |
1480 | // %11:_(s6) = G_MERGE_VALUES %7, %8, %9 |
1481 | // %12:_(s12) = G_MERGE_VALUES %10, %11 |
1482 | // |
1483 | // Padding with undef if necessary: |
1484 | // |
1485 | // %2:_(s8) = G_MERGE_VALUES %0:_(s4), %1:_(s4) -> s6 |
1486 | // %3:_(s2), %4:_(s2) = G_UNMERGE_VALUES %0 |
1487 | // %5:_(s2), %6:_(s2) = G_UNMERGE_VALUES %1 |
1488 | // %7:_(s2) = G_IMPLICIT_DEF |
1489 | // %8:_(s6) = G_MERGE_VALUES %3, %4, %5 |
1490 | // %9:_(s6) = G_MERGE_VALUES %6, %7, %7 |
1491 | // %10:_(s12) = G_MERGE_VALUES %8, %9 |
1492 | |
1493 | const int GCD = greatestCommonDivisor(SrcSize, WideSize); |
1494 | LLT GCDTy = LLT::scalar(GCD); |
1495 | |
1496 | SmallVector<Register, 8> Parts; |
1497 | SmallVector<Register, 8> NewMergeRegs; |
1498 | SmallVector<Register, 8> Unmerges; |
1499 | LLT WideDstTy = LLT::scalar(NumMerge * WideSize); |
1500 | |
1501 | // Decompose the original operands if they don't evenly divide. |
1502 | for (int I = 1, E = MI.getNumOperands(); I != E; ++I) { |
1503 | Register SrcReg = MI.getOperand(I).getReg(); |
1504 | if (GCD == SrcSize) { |
1505 | Unmerges.push_back(SrcReg); |
1506 | } else { |
1507 | auto Unmerge = MIRBuilder.buildUnmerge(GCDTy, SrcReg); |
1508 | for (int J = 0, JE = Unmerge->getNumOperands() - 1; J != JE; ++J) |
1509 | Unmerges.push_back(Unmerge.getReg(J)); |
1510 | } |
1511 | } |
1512 | |
1513 | // Pad with undef to the next size that is a multiple of the requested size. |
1514 | if (static_cast<int>(Unmerges.size()) != NumMerge * WideSize) { |
1515 | Register UndefReg = MIRBuilder.buildUndef(GCDTy).getReg(0); |
1516 | for (int I = Unmerges.size(); I != NumMerge * WideSize; ++I) |
1517 | Unmerges.push_back(UndefReg); |
1518 | } |
1519 | |
1520 | const int PartsPerGCD = WideSize / GCD; |
1521 | |
1522 | // Build merges of each piece. |
1523 | ArrayRef<Register> Slicer(Unmerges); |
1524 | for (int I = 0; I != NumMerge; ++I, Slicer = Slicer.drop_front(PartsPerGCD)) { |
1525 | auto Merge = MIRBuilder.buildMerge(WideTy, Slicer.take_front(PartsPerGCD)); |
1526 | NewMergeRegs.push_back(Merge.getReg(0)); |
1527 | } |
1528 | |
1529 | // A truncate may be necessary if the requested type doesn't evenly divide the |
1530 | // original result type. |
1531 | if (DstTy.getSizeInBits() == WideDstTy.getSizeInBits()) { |
1532 | MIRBuilder.buildMerge(DstReg, NewMergeRegs); |
1533 | } else { |
1534 | auto FinalMerge = MIRBuilder.buildMerge(WideDstTy, NewMergeRegs); |
1535 | MIRBuilder.buildTrunc(DstReg, FinalMerge.getReg(0)); |
1536 | } |
1537 | |
1538 | MI.eraseFromParent(); |
1539 | return Legalized; |
1540 | } |
1541 | |
1542 | Register LegalizerHelper::widenWithUnmerge(LLT WideTy, Register OrigReg) { |
1543 | Register WideReg = MRI.createGenericVirtualRegister(WideTy); |
1544 | LLT OrigTy = MRI.getType(OrigReg); |
1545 | LLT LCMTy = getLCMType(WideTy, OrigTy); |
1546 | |
1547 | const int NumMergeParts = LCMTy.getSizeInBits() / WideTy.getSizeInBits(); |
1548 | const int NumUnmergeParts = LCMTy.getSizeInBits() / OrigTy.getSizeInBits(); |
1549 | |
1550 | Register UnmergeSrc = WideReg; |
1551 | |
1552 | // Create a merge to the LCM type, padding with undef |
1553 | // %0:_(<3 x s32>) = G_FOO => <4 x s32> |
1554 | // => |
1555 | // %1:_(<4 x s32>) = G_FOO |
1556 | // %2:_(<4 x s32>) = G_IMPLICIT_DEF |
1557 | // %3:_(<12 x s32>) = G_CONCAT_VECTORS %1, %2, %2 |
1558 | // %0:_(<3 x s32>), %4:_, %5:_, %6:_ = G_UNMERGE_VALUES %3 |
1559 | if (NumMergeParts > 1) { |
1560 | Register Undef = MIRBuilder.buildUndef(WideTy).getReg(0); |
1561 | SmallVector<Register, 8> MergeParts(NumMergeParts, Undef); |
1562 | MergeParts[0] = WideReg; |
1563 | UnmergeSrc = MIRBuilder.buildMerge(LCMTy, MergeParts).getReg(0); |
1564 | } |
1565 | |
1566 | // Unmerge to the original register and pad with dead defs. |
1567 | SmallVector<Register, 8> UnmergeResults(NumUnmergeParts); |
1568 | UnmergeResults[0] = OrigReg; |
1569 | for (int I = 1; I != NumUnmergeParts; ++I) |
1570 | UnmergeResults[I] = MRI.createGenericVirtualRegister(OrigTy); |
1571 | |
1572 | MIRBuilder.buildUnmerge(UnmergeResults, UnmergeSrc); |
1573 | return WideReg; |
1574 | } |
1575 | |
1576 | LegalizerHelper::LegalizeResult |
1577 | LegalizerHelper::widenScalarUnmergeValues(MachineInstr &MI, unsigned TypeIdx, |
1578 | LLT WideTy) { |
1579 | if (TypeIdx != 0) |
1580 | return UnableToLegalize; |
1581 | |
1582 | int NumDst = MI.getNumOperands() - 1; |
1583 | Register SrcReg = MI.getOperand(NumDst).getReg(); |
1584 | LLT SrcTy = MRI.getType(SrcReg); |
1585 | if (SrcTy.isVector()) |
1586 | return UnableToLegalize; |
1587 | |
1588 | Register Dst0Reg = MI.getOperand(0).getReg(); |
1589 | LLT DstTy = MRI.getType(Dst0Reg); |
1590 | if (!DstTy.isScalar()) |
1591 | return UnableToLegalize; |
1592 | |
1593 | if (WideTy.getSizeInBits() >= SrcTy.getSizeInBits()) { |
1594 | if (SrcTy.isPointer()) { |
1595 | const DataLayout &DL = MIRBuilder.getDataLayout(); |
1596 | if (DL.isNonIntegralAddressSpace(SrcTy.getAddressSpace())) { |
1597 | LLVM_DEBUG(do { } while (false) |
1598 | dbgs() << "Not casting non-integral address space integer\n")do { } while (false); |
1599 | return UnableToLegalize; |
1600 | } |
1601 | |
1602 | SrcTy = LLT::scalar(SrcTy.getSizeInBits()); |
1603 | SrcReg = MIRBuilder.buildPtrToInt(SrcTy, SrcReg).getReg(0); |
1604 | } |
1605 | |
1606 | // Widen SrcTy to WideTy. This does not affect the result, but since the |
1607 | // user requested this size, it is probably better handled than SrcTy and |
1608 | // should reduce the total number of legalization artifacts |
1609 | if (WideTy.getSizeInBits() > SrcTy.getSizeInBits()) { |
1610 | SrcTy = WideTy; |
1611 | SrcReg = MIRBuilder.buildAnyExt(WideTy, SrcReg).getReg(0); |
1612 | } |
1613 | |
1614 | // Theres no unmerge type to target. Directly extract the bits from the |
1615 | // source type |
1616 | unsigned DstSize = DstTy.getSizeInBits(); |
1617 | |
1618 | MIRBuilder.buildTrunc(Dst0Reg, SrcReg); |
1619 | for (int I = 1; I != NumDst; ++I) { |
1620 | auto ShiftAmt = MIRBuilder.buildConstant(SrcTy, DstSize * I); |
1621 | auto Shr = MIRBuilder.buildLShr(SrcTy, SrcReg, ShiftAmt); |
1622 | MIRBuilder.buildTrunc(MI.getOperand(I), Shr); |
1623 | } |
1624 | |
1625 | MI.eraseFromParent(); |
1626 | return Legalized; |
1627 | } |
1628 | |
1629 | // Extend the source to a wider type. |
1630 | LLT LCMTy = getLCMType(SrcTy, WideTy); |
1631 | |
1632 | Register WideSrc = SrcReg; |
1633 | if (LCMTy.getSizeInBits() != SrcTy.getSizeInBits()) { |
1634 | // TODO: If this is an integral address space, cast to integer and anyext. |
1635 | if (SrcTy.isPointer()) { |
1636 | LLVM_DEBUG(dbgs() << "Widening pointer source types not implemented\n")do { } while (false); |
1637 | return UnableToLegalize; |
1638 | } |
1639 | |
1640 | WideSrc = MIRBuilder.buildAnyExt(LCMTy, WideSrc).getReg(0); |
1641 | } |
1642 | |
1643 | auto Unmerge = MIRBuilder.buildUnmerge(WideTy, WideSrc); |
1644 | |
1645 | // Create a sequence of unmerges and merges to the original results. Since we |
1646 | // may have widened the source, we will need to pad the results with dead defs |
1647 | // to cover the source register. |
1648 | // e.g. widen s48 to s64: |
1649 | // %1:_(s48), %2:_(s48) = G_UNMERGE_VALUES %0:_(s96) |
1650 | // |
1651 | // => |
1652 | // %4:_(s192) = G_ANYEXT %0:_(s96) |
1653 | // %5:_(s64), %6, %7 = G_UNMERGE_VALUES %4 ; Requested unmerge |
1654 | // ; unpack to GCD type, with extra dead defs |
1655 | // %8:_(s16), %9, %10, %11 = G_UNMERGE_VALUES %5:_(s64) |
1656 | // %12:_(s16), %13, dead %14, dead %15 = G_UNMERGE_VALUES %6:_(s64) |
1657 | // dead %16:_(s16), dead %17, dead %18, dead %18 = G_UNMERGE_VALUES %7:_(s64) |
1658 | // %1:_(s48) = G_MERGE_VALUES %8:_(s16), %9, %10 ; Remerge to destination |
1659 | // %2:_(s48) = G_MERGE_VALUES %11:_(s16), %12, %13 ; Remerge to destination |
1660 | const LLT GCDTy = getGCDType(WideTy, DstTy); |
1661 | const int NumUnmerge = Unmerge->getNumOperands() - 1; |
1662 | const int PartsPerRemerge = DstTy.getSizeInBits() / GCDTy.getSizeInBits(); |
1663 | |
1664 | // Directly unmerge to the destination without going through a GCD type |
1665 | // if possible |
1666 | if (PartsPerRemerge == 1) { |
1667 | const int PartsPerUnmerge = WideTy.getSizeInBits() / DstTy.getSizeInBits(); |
1668 | |
1669 | for (int I = 0; I != NumUnmerge; ++I) { |
1670 | auto MIB = MIRBuilder.buildInstr(TargetOpcode::G_UNMERGE_VALUES); |
1671 | |
1672 | for (int J = 0; J != PartsPerUnmerge; ++J) { |
1673 | int Idx = I * PartsPerUnmerge + J; |
1674 | if (Idx < NumDst) |
1675 | MIB.addDef(MI.getOperand(Idx).getReg()); |
1676 | else { |
1677 | // Create dead def for excess components. |
1678 | MIB.addDef(MRI.createGenericVirtualRegister(DstTy)); |
1679 | } |
1680 | } |
1681 | |
1682 | MIB.addUse(Unmerge.getReg(I)); |
1683 | } |
1684 | } else { |
1685 | SmallVector<Register, 16> Parts; |
1686 | for (int J = 0; J != NumUnmerge; ++J) |
1687 | extractGCDType(Parts, GCDTy, Unmerge.getReg(J)); |
1688 | |
1689 | SmallVector<Register, 8> RemergeParts; |
1690 | for (int I = 0; I != NumDst; ++I) { |
1691 | for (int J = 0; J < PartsPerRemerge; ++J) { |
1692 | const int Idx = I * PartsPerRemerge + J; |
1693 | RemergeParts.emplace_back(Parts[Idx]); |
1694 | } |
1695 | |
1696 | MIRBuilder.buildMerge(MI.getOperand(I).getReg(), RemergeParts); |
1697 | RemergeParts.clear(); |
1698 | } |
1699 | } |
1700 | |
1701 | MI.eraseFromParent(); |
1702 | return Legalized; |
1703 | } |
1704 | |
1705 | LegalizerHelper::LegalizeResult |
1706 | LegalizerHelper::widenScalarExtract(MachineInstr &MI, unsigned TypeIdx, |
1707 | LLT WideTy) { |
1708 | Register DstReg = MI.getOperand(0).getReg(); |
1709 | Register SrcReg = MI.getOperand(1).getReg(); |
1710 | LLT SrcTy = MRI.getType(SrcReg); |
1711 | |
1712 | LLT DstTy = MRI.getType(DstReg); |
1713 | unsigned Offset = MI.getOperand(2).getImm(); |
1714 | |
1715 | if (TypeIdx == 0) { |
1716 | if (SrcTy.isVector() || DstTy.isVector()) |
1717 | return UnableToLegalize; |
1718 | |
1719 | SrcOp Src(SrcReg); |
1720 | if (SrcTy.isPointer()) { |
1721 | // Extracts from pointers can be handled only if they are really just |
1722 | // simple integers. |
1723 | const DataLayout &DL = MIRBuilder.getDataLayout(); |
1724 | if (DL.isNonIntegralAddressSpace(SrcTy.getAddressSpace())) |
1725 | return UnableToLegalize; |
1726 | |
1727 | LLT SrcAsIntTy = LLT::scalar(SrcTy.getSizeInBits()); |
1728 | Src = MIRBuilder.buildPtrToInt(SrcAsIntTy, Src); |
1729 | SrcTy = SrcAsIntTy; |
1730 | } |
1731 | |
1732 | if (DstTy.isPointer()) |
1733 | return UnableToLegalize; |
1734 | |
1735 | if (Offset == 0) { |
1736 | // Avoid a shift in the degenerate case. |
1737 | MIRBuilder.buildTrunc(DstReg, |
1738 | MIRBuilder.buildAnyExtOrTrunc(WideTy, Src)); |
1739 | MI.eraseFromParent(); |
1740 | return Legalized; |
1741 | } |
1742 | |
1743 | // Do a shift in the source type. |
1744 | LLT ShiftTy = SrcTy; |
1745 | if (WideTy.getSizeInBits() > SrcTy.getSizeInBits()) { |
1746 | Src = MIRBuilder.buildAnyExt(WideTy, Src); |
1747 | ShiftTy = WideTy; |
1748 | } |
1749 | |
1750 | auto LShr = MIRBuilder.buildLShr( |
1751 | ShiftTy, Src, MIRBuilder.buildConstant(ShiftTy, Offset)); |
1752 | MIRBuilder.buildTrunc(DstReg, LShr); |
1753 | MI.eraseFromParent(); |
1754 | return Legalized; |
1755 | } |
1756 | |
1757 | if (SrcTy.isScalar()) { |
1758 | Observer.changingInstr(MI); |
1759 | widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT); |
1760 | Observer.changedInstr(MI); |
1761 | return Legalized; |
1762 | } |
1763 | |
1764 | if (!SrcTy.isVector()) |
1765 | return UnableToLegalize; |
1766 | |
1767 | if (DstTy != SrcTy.getElementType()) |
1768 | return UnableToLegalize; |
1769 | |
1770 | if (Offset % SrcTy.getScalarSizeInBits() != 0) |
1771 | return UnableToLegalize; |
1772 | |
1773 | Observer.changingInstr(MI); |
1774 | widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT); |
1775 | |
1776 | MI.getOperand(2).setImm((WideTy.getSizeInBits() / SrcTy.getSizeInBits()) * |
1777 | Offset); |
1778 | widenScalarDst(MI, WideTy.getScalarType(), 0); |
1779 | Observer.changedInstr(MI); |
1780 | return Legalized; |
1781 | } |
1782 | |
1783 | LegalizerHelper::LegalizeResult |
1784 | LegalizerHelper::widenScalarInsert(MachineInstr &MI, unsigned TypeIdx, |
1785 | LLT WideTy) { |
1786 | if (TypeIdx != 0 || WideTy.isVector()) |
1787 | return UnableToLegalize; |
1788 | Observer.changingInstr(MI); |
1789 | widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT); |
1790 | widenScalarDst(MI, WideTy); |
1791 | Observer.changedInstr(MI); |
1792 | return Legalized; |
1793 | } |
1794 | |
1795 | LegalizerHelper::LegalizeResult |
1796 | LegalizerHelper::widenScalarAddSubOverflow(MachineInstr &MI, unsigned TypeIdx, |
1797 | LLT WideTy) { |
1798 | if (TypeIdx == 1) |
1799 | return UnableToLegalize; // TODO |
1800 | |
1801 | unsigned Opcode; |
1802 | unsigned ExtOpcode; |
1803 | Optional<Register> CarryIn = None; |
1804 | switch (MI.getOpcode()) { |
1805 | default: |
1806 | llvm_unreachable("Unexpected opcode!")__builtin_unreachable(); |
1807 | case TargetOpcode::G_SADDO: |
1808 | Opcode = TargetOpcode::G_ADD; |
1809 | ExtOpcode = TargetOpcode::G_SEXT; |
1810 | break; |
1811 | case TargetOpcode::G_SSUBO: |
1812 | Opcode = TargetOpcode::G_SUB; |
1813 | ExtOpcode = TargetOpcode::G_SEXT; |
1814 | break; |
1815 | case TargetOpcode::G_UADDO: |
1816 | Opcode = TargetOpcode::G_ADD; |
1817 | ExtOpcode = TargetOpcode::G_ZEXT; |
1818 | break; |
1819 | case TargetOpcode::G_USUBO: |
1820 | Opcode = TargetOpcode::G_SUB; |
1821 | ExtOpcode = TargetOpcode::G_ZEXT; |
1822 | break; |
1823 | case TargetOpcode::G_SADDE: |
1824 | Opcode = TargetOpcode::G_UADDE; |
1825 | ExtOpcode = TargetOpcode::G_SEXT; |
1826 | CarryIn = MI.getOperand(4).getReg(); |
1827 | break; |
1828 | case TargetOpcode::G_SSUBE: |
1829 | Opcode = TargetOpcode::G_USUBE; |
1830 | ExtOpcode = TargetOpcode::G_SEXT; |
1831 | CarryIn = MI.getOperand(4).getReg(); |
1832 | break; |
1833 | case TargetOpcode::G_UADDE: |
1834 | Opcode = TargetOpcode::G_UADDE; |
1835 | ExtOpcode = TargetOpcode::G_ZEXT; |
1836 | CarryIn = MI.getOperand(4).getReg(); |
1837 | break; |
1838 | case TargetOpcode::G_USUBE: |
1839 | Opcode = TargetOpcode::G_USUBE; |
1840 | ExtOpcode = TargetOpcode::G_ZEXT; |
1841 | CarryIn = MI.getOperand(4).getReg(); |
1842 | break; |
1843 | } |
1844 | |
1845 | auto LHSExt = MIRBuilder.buildInstr(ExtOpcode, {WideTy}, {MI.getOperand(2)}); |
1846 | auto RHSExt = MIRBuilder.buildInstr(ExtOpcode, {WideTy}, {MI.getOperand(3)}); |
1847 | // Do the arithmetic in the larger type. |
1848 | Register NewOp; |
1849 | if (CarryIn) { |
1850 | LLT CarryOutTy = MRI.getType(MI.getOperand(1).getReg()); |
1851 | NewOp = MIRBuilder |
1852 | .buildInstr(Opcode, {WideTy, CarryOutTy}, |
1853 | {LHSExt, RHSExt, *CarryIn}) |
1854 | .getReg(0); |
1855 | } else { |
1856 | NewOp = MIRBuilder.buildInstr(Opcode, {WideTy}, {LHSExt, RHSExt}).getReg(0); |
1857 | } |
1858 | LLT OrigTy = MRI.getType(MI.getOperand(0).getReg()); |
1859 | auto TruncOp = MIRBuilder.buildTrunc(OrigTy, NewOp); |
1860 | auto ExtOp = MIRBuilder.buildInstr(ExtOpcode, {WideTy}, {TruncOp}); |
1861 | // There is no overflow if the ExtOp is the same as NewOp. |
1862 | MIRBuilder.buildICmp(CmpInst::ICMP_NE, MI.getOperand(1), NewOp, ExtOp); |
1863 | // Now trunc the NewOp to the original result. |
1864 | MIRBuilder.buildTrunc(MI.getOperand(0), NewOp); |
1865 | MI.eraseFromParent(); |
1866 | return Legalized; |
1867 | } |
1868 | |
1869 | LegalizerHelper::LegalizeResult |
1870 | LegalizerHelper::widenScalarAddSubShlSat(MachineInstr &MI, unsigned TypeIdx, |
1871 | LLT WideTy) { |
1872 | bool IsSigned = MI.getOpcode() == TargetOpcode::G_SADDSAT || |
1873 | MI.getOpcode() == TargetOpcode::G_SSUBSAT || |
1874 | MI.getOpcode() == TargetOpcode::G_SSHLSAT; |
1875 | bool IsShift = MI.getOpcode() == TargetOpcode::G_SSHLSAT || |
1876 | MI.getOpcode() == TargetOpcode::G_USHLSAT; |
1877 | // We can convert this to: |
1878 | // 1. Any extend iN to iM |
1879 | // 2. SHL by M-N |
1880 | // 3. [US][ADD|SUB|SHL]SAT |
1881 | // 4. L/ASHR by M-N |
1882 | // |
1883 | // It may be more efficient to lower this to a min and a max operation in |
1884 | // the higher precision arithmetic if the promoted operation isn't legal, |
1885 | // but this decision is up to the target's lowering request. |
1886 | Register DstReg = MI.getOperand(0).getReg(); |
1887 | |
1888 | unsigned NewBits = WideTy.getScalarSizeInBits(); |
1889 | unsigned SHLAmount = NewBits - MRI.getType(DstReg).getScalarSizeInBits(); |
1890 | |
1891 | // Shifts must zero-extend the RHS to preserve the unsigned quantity, and |
1892 | // must not left shift the RHS to preserve the shift amount. |
1893 | auto LHS = MIRBuilder.buildAnyExt(WideTy, MI.getOperand(1)); |
1894 | auto RHS = IsShift ? MIRBuilder.buildZExt(WideTy, MI.getOperand(2)) |
1895 | : MIRBuilder.buildAnyExt(WideTy, MI.getOperand(2)); |
1896 | auto ShiftK = MIRBuilder.buildConstant(WideTy, SHLAmount); |
1897 | auto ShiftL = MIRBuilder.buildShl(WideTy, LHS, ShiftK); |
1898 | auto ShiftR = IsShift ? RHS : MIRBuilder.buildShl(WideTy, RHS, ShiftK); |
1899 | |
1900 | auto WideInst = MIRBuilder.buildInstr(MI.getOpcode(), {WideTy}, |
1901 | {ShiftL, ShiftR}, MI.getFlags()); |
1902 | |
1903 | // Use a shift that will preserve the number of sign bits when the trunc is |
1904 | // folded away. |
1905 | auto Result = IsSigned ? MIRBuilder.buildAShr(WideTy, WideInst, ShiftK) |
1906 | : MIRBuilder.buildLShr(WideTy, WideInst, ShiftK); |
1907 | |
1908 | MIRBuilder.buildTrunc(DstReg, Result); |
1909 | MI.eraseFromParent(); |
1910 | return Legalized; |
1911 | } |
1912 | |
1913 | LegalizerHelper::LegalizeResult |
1914 | LegalizerHelper::widenScalarMulo(MachineInstr &MI, unsigned TypeIdx, |
1915 | LLT WideTy) { |
1916 | if (TypeIdx == 1) |
1917 | return UnableToLegalize; |
1918 | |
1919 | bool IsSigned = MI.getOpcode() == TargetOpcode::G_SMULO; |
1920 | Register Result = MI.getOperand(0).getReg(); |
1921 | Register OriginalOverflow = MI.getOperand(1).getReg(); |
1922 | Register LHS = MI.getOperand(2).getReg(); |
1923 | Register RHS = MI.getOperand(3).getReg(); |
1924 | LLT SrcTy = MRI.getType(LHS); |
1925 | LLT OverflowTy = MRI.getType(OriginalOverflow); |
1926 | unsigned SrcBitWidth = SrcTy.getScalarSizeInBits(); |
1927 | |
1928 | // To determine if the result overflowed in the larger type, we extend the |
1929 | // input to the larger type, do the multiply (checking if it overflows), |
1930 | // then also check the high bits of the result to see if overflow happened |
1931 | // there. |
1932 | unsigned ExtOp = IsSigned ? TargetOpcode::G_SEXT : TargetOpcode::G_ZEXT; |
1933 | auto LeftOperand = MIRBuilder.buildInstr(ExtOp, {WideTy}, {LHS}); |
1934 | auto RightOperand = MIRBuilder.buildInstr(ExtOp, {WideTy}, {RHS}); |
1935 | |
1936 | auto Mulo = MIRBuilder.buildInstr(MI.getOpcode(), {WideTy, OverflowTy}, |
1937 | {LeftOperand, RightOperand}); |
1938 | auto Mul = Mulo->getOperand(0); |
1939 | MIRBuilder.buildTrunc(Result, Mul); |
1940 | |
1941 | MachineInstrBuilder ExtResult; |
1942 | // Overflow occurred if it occurred in the larger type, or if the high part |
1943 | // of the result does not zero/sign-extend the low part. Check this second |
1944 | // possibility first. |
1945 | if (IsSigned) { |
1946 | // For signed, overflow occurred when the high part does not sign-extend |
1947 | // the low part. |
1948 | ExtResult = MIRBuilder.buildSExtInReg(WideTy, Mul, SrcBitWidth); |
1949 | } else { |
1950 | // Unsigned overflow occurred when the high part does not zero-extend the |
1951 | // low part. |
1952 | ExtResult = MIRBuilder.buildZExtInReg(WideTy, Mul, SrcBitWidth); |
1953 | } |
1954 | |
1955 | // Multiplication cannot overflow if the WideTy is >= 2 * original width, |
1956 | // so we don't need to check the overflow result of larger type Mulo. |
1957 | if (WideTy.getScalarSizeInBits() < 2 * SrcBitWidth) { |
1958 | auto Overflow = |
1959 | MIRBuilder.buildICmp(CmpInst::ICMP_NE, OverflowTy, Mul, ExtResult); |
1960 | // Finally check if the multiplication in the larger type itself overflowed. |
1961 | MIRBuilder.buildOr(OriginalOverflow, Mulo->getOperand(1), Overflow); |
1962 | } else { |
1963 | MIRBuilder.buildICmp(CmpInst::ICMP_NE, OriginalOverflow, Mul, ExtResult); |
1964 | } |
1965 | MI.eraseFromParent(); |
1966 | return Legalized; |
1967 | } |
1968 | |
1969 | LegalizerHelper::LegalizeResult |
1970 | LegalizerHelper::widenScalar(MachineInstr &MI, unsigned TypeIdx, LLT WideTy) { |
1971 | switch (MI.getOpcode()) { |
1972 | default: |
1973 | return UnableToLegalize; |
1974 | case TargetOpcode::G_ATOMICRMW_XCHG: |
1975 | case TargetOpcode::G_ATOMICRMW_ADD: |
1976 | case TargetOpcode::G_ATOMICRMW_SUB: |
1977 | case TargetOpcode::G_ATOMICRMW_AND: |
1978 | case TargetOpcode::G_ATOMICRMW_OR: |
1979 | case TargetOpcode::G_ATOMICRMW_XOR: |
1980 | case TargetOpcode::G_ATOMICRMW_MIN: |
1981 | case TargetOpcode::G_ATOMICRMW_MAX: |
1982 | case TargetOpcode::G_ATOMICRMW_UMIN: |
1983 | case TargetOpcode::G_ATOMICRMW_UMAX: |
1984 | assert(TypeIdx == 0 && "atomicrmw with second scalar type")((void)0); |
1985 | Observer.changingInstr(MI); |
1986 | widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ANYEXT); |
1987 | widenScalarDst(MI, WideTy, 0); |
1988 | Observer.changedInstr(MI); |
1989 | return Legalized; |
1990 | case TargetOpcode::G_ATOMIC_CMPXCHG: |
1991 | assert(TypeIdx == 0 && "G_ATOMIC_CMPXCHG with second scalar type")((void)0); |
1992 | Observer.changingInstr(MI); |
1993 | widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ANYEXT); |
1994 | widenScalarSrc(MI, WideTy, 3, TargetOpcode::G_ANYEXT); |
1995 | widenScalarDst(MI, WideTy, 0); |
1996 | Observer.changedInstr(MI); |
1997 | return Legalized; |
1998 | case TargetOpcode::G_ATOMIC_CMPXCHG_WITH_SUCCESS: |
1999 | if (TypeIdx == 0) { |
2000 | Observer.changingInstr(MI); |
2001 | widenScalarSrc(MI, WideTy, 3, TargetOpcode::G_ANYEXT); |
2002 | widenScalarSrc(MI, WideTy, 4, TargetOpcode::G_ANYEXT); |
2003 | widenScalarDst(MI, WideTy, 0); |
2004 | Observer.changedInstr(MI); |
2005 | return Legalized; |
2006 | } |
2007 | assert(TypeIdx == 1 &&((void)0) |
2008 | "G_ATOMIC_CMPXCHG_WITH_SUCCESS with third scalar type")((void)0); |
2009 | Observer.changingInstr(MI); |
2010 | widenScalarDst(MI, WideTy, 1); |
2011 | Observer.changedInstr(MI); |
2012 | return Legalized; |
2013 | case TargetOpcode::G_EXTRACT: |
2014 | return widenScalarExtract(MI, TypeIdx, WideTy); |
2015 | case TargetOpcode::G_INSERT: |
2016 | return widenScalarInsert(MI, TypeIdx, WideTy); |
2017 | case TargetOpcode::G_MERGE_VALUES: |
2018 | return widenScalarMergeValues(MI, TypeIdx, WideTy); |
2019 | case TargetOpcode::G_UNMERGE_VALUES: |
2020 | return widenScalarUnmergeValues(MI, TypeIdx, WideTy); |
2021 | case TargetOpcode::G_SADDO: |
2022 | case TargetOpcode::G_SSUBO: |
2023 | case TargetOpcode::G_UADDO: |
2024 | case TargetOpcode::G_USUBO: |
2025 | case TargetOpcode::G_SADDE: |
2026 | case TargetOpcode::G_SSUBE: |
2027 | case TargetOpcode::G_UADDE: |
2028 | case TargetOpcode::G_USUBE: |
2029 | return widenScalarAddSubOverflow(MI, TypeIdx, WideTy); |
2030 | case TargetOpcode::G_UMULO: |
2031 | case TargetOpcode::G_SMULO: |
2032 | return widenScalarMulo(MI, TypeIdx, WideTy); |
2033 | case TargetOpcode::G_SADDSAT: |
2034 | case TargetOpcode::G_SSUBSAT: |
2035 | case TargetOpcode::G_SSHLSAT: |
2036 | case TargetOpcode::G_UADDSAT: |
2037 | case TargetOpcode::G_USUBSAT: |
2038 | case TargetOpcode::G_USHLSAT: |
2039 | return widenScalarAddSubShlSat(MI, TypeIdx, WideTy); |
2040 | case TargetOpcode::G_CTTZ: |
2041 | case TargetOpcode::G_CTTZ_ZERO_UNDEF: |
2042 | case TargetOpcode::G_CTLZ: |
2043 | case TargetOpcode::G_CTLZ_ZERO_UNDEF: |
2044 | case TargetOpcode::G_CTPOP: { |
2045 | if (TypeIdx == 0) { |
2046 | Observer.changingInstr(MI); |
2047 | widenScalarDst(MI, WideTy, 0); |
2048 | Observer.changedInstr(MI); |
2049 | return Legalized; |
2050 | } |
2051 | |
2052 | Register SrcReg = MI.getOperand(1).getReg(); |
2053 | |
2054 | // First ZEXT the input. |
2055 | auto MIBSrc = MIRBuilder.buildZExt(WideTy, SrcReg); |
2056 | LLT CurTy = MRI.getType(SrcReg); |
2057 | if (MI.getOpcode() == TargetOpcode::G_CTTZ) { |
2058 | // The count is the same in the larger type except if the original |
2059 | // value was zero. This can be handled by setting the bit just off |
2060 | // the top of the original type. |
2061 | auto TopBit = |
2062 | APInt::getOneBitSet(WideTy.getSizeInBits(), CurTy.getSizeInBits()); |
2063 | MIBSrc = MIRBuilder.buildOr( |
2064 | WideTy, MIBSrc, MIRBuilder.buildConstant(WideTy, TopBit)); |
2065 | } |
2066 | |
2067 | // Perform the operation at the larger size. |
2068 | auto MIBNewOp = MIRBuilder.buildInstr(MI.getOpcode(), {WideTy}, {MIBSrc}); |
2069 | // This is already the correct result for CTPOP and CTTZs |
2070 | if (MI.getOpcode() == TargetOpcode::G_CTLZ || |
2071 | MI.getOpcode() == TargetOpcode::G_CTLZ_ZERO_UNDEF) { |
2072 | // The correct result is NewOp - (Difference in widety and current ty). |
2073 | unsigned SizeDiff = WideTy.getSizeInBits() - CurTy.getSizeInBits(); |
2074 | MIBNewOp = MIRBuilder.buildSub( |
2075 | WideTy, MIBNewOp, MIRBuilder.buildConstant(WideTy, SizeDiff)); |
2076 | } |
2077 | |
2078 | MIRBuilder.buildZExtOrTrunc(MI.getOperand(0), MIBNewOp); |
2079 | MI.eraseFromParent(); |
2080 | return Legalized; |
2081 | } |
2082 | case TargetOpcode::G_BSWAP: { |
2083 | Observer.changingInstr(MI); |
2084 | Register DstReg = MI.getOperand(0).getReg(); |
2085 | |
2086 | Register ShrReg = MRI.createGenericVirtualRegister(WideTy); |
2087 | Register DstExt = MRI.createGenericVirtualRegister(WideTy); |
2088 | Register ShiftAmtReg = MRI.createGenericVirtualRegister(WideTy); |
2089 | widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT); |
2090 | |
2091 | MI.getOperand(0).setReg(DstExt); |
2092 | |
2093 | MIRBuilder.setInsertPt(MIRBuilder.getMBB(), ++MIRBuilder.getInsertPt()); |
2094 | |
2095 | LLT Ty = MRI.getType(DstReg); |
2096 | unsigned DiffBits = WideTy.getScalarSizeInBits() - Ty.getScalarSizeInBits(); |
2097 | MIRBuilder.buildConstant(ShiftAmtReg, DiffBits); |
2098 | MIRBuilder.buildLShr(ShrReg, DstExt, ShiftAmtReg); |
2099 | |
2100 | MIRBuilder.buildTrunc(DstReg, ShrReg); |
2101 | Observer.changedInstr(MI); |
2102 | return Legalized; |
2103 | } |
2104 | case TargetOpcode::G_BITREVERSE: { |
2105 | Observer.changingInstr(MI); |
2106 | |
2107 | Register DstReg = MI.getOperand(0).getReg(); |
2108 | LLT Ty = MRI.getType(DstReg); |
2109 | unsigned DiffBits = WideTy.getScalarSizeInBits() - Ty.getScalarSizeInBits(); |
2110 | |
2111 | Register DstExt = MRI.createGenericVirtualRegister(WideTy); |
2112 | widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT); |
2113 | MI.getOperand(0).setReg(DstExt); |
2114 | MIRBuilder.setInsertPt(MIRBuilder.getMBB(), ++MIRBuilder.getInsertPt()); |
2115 | |
2116 | auto ShiftAmt = MIRBuilder.buildConstant(WideTy, DiffBits); |
2117 | auto Shift = MIRBuilder.buildLShr(WideTy, DstExt, ShiftAmt); |
2118 | MIRBuilder.buildTrunc(DstReg, Shift); |
2119 | Observer.changedInstr(MI); |
2120 | return Legalized; |
2121 | } |
2122 | case TargetOpcode::G_FREEZE: |
2123 | Observer.changingInstr(MI); |
2124 | widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT); |
2125 | widenScalarDst(MI, WideTy); |
2126 | Observer.changedInstr(MI); |
2127 | return Legalized; |
2128 | |
2129 | case TargetOpcode::G_ABS: |
2130 | Observer.changingInstr(MI); |
2131 | widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_SEXT); |
2132 | widenScalarDst(MI, WideTy); |
2133 | Observer.changedInstr(MI); |
2134 | return Legalized; |
2135 | |
2136 | case TargetOpcode::G_ADD: |
2137 | case TargetOpcode::G_AND: |
2138 | case TargetOpcode::G_MUL: |
2139 | case TargetOpcode::G_OR: |
2140 | case TargetOpcode::G_XOR: |
2141 | case TargetOpcode::G_SUB: |
2142 | // Perform operation at larger width (any extension is fines here, high bits |
2143 | // don't affect the result) and then truncate the result back to the |
2144 | // original type. |
2145 | Observer.changingInstr(MI); |
2146 | widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT); |
2147 | widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ANYEXT); |
2148 | widenScalarDst(MI, WideTy); |
2149 | Observer.changedInstr(MI); |
2150 | return Legalized; |
2151 | |
2152 | case TargetOpcode::G_SBFX: |
2153 | case TargetOpcode::G_UBFX: |
2154 | Observer.changingInstr(MI); |
2155 | |
2156 | if (TypeIdx == 0) { |
2157 | widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT); |
2158 | widenScalarDst(MI, WideTy); |
2159 | } else { |
2160 | widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ZEXT); |
2161 | widenScalarSrc(MI, WideTy, 3, TargetOpcode::G_ZEXT); |
2162 | } |
2163 | |
2164 | Observer.changedInstr(MI); |
2165 | return Legalized; |
2166 | |
2167 | case TargetOpcode::G_SHL: |
2168 | Observer.changingInstr(MI); |
2169 | |
2170 | if (TypeIdx == 0) { |
2171 | widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT); |
2172 | widenScalarDst(MI, WideTy); |
2173 | } else { |
2174 | assert(TypeIdx == 1)((void)0); |
2175 | // The "number of bits to shift" operand must preserve its value as an |
2176 | // unsigned integer: |
2177 | widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ZEXT); |
2178 | } |
2179 | |
2180 | Observer.changedInstr(MI); |
2181 | return Legalized; |
2182 | |
2183 | case TargetOpcode::G_SDIV: |
2184 | case TargetOpcode::G_SREM: |
2185 | case TargetOpcode::G_SMIN: |
2186 | case TargetOpcode::G_SMAX: |
2187 | Observer.changingInstr(MI); |
2188 | widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_SEXT); |
2189 | widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_SEXT); |
2190 | widenScalarDst(MI, WideTy); |
2191 | Observer.changedInstr(MI); |
2192 | return Legalized; |
2193 | |
2194 | case TargetOpcode::G_SDIVREM: |
2195 | Observer.changingInstr(MI); |
2196 | widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_SEXT); |
2197 | widenScalarSrc(MI, WideTy, 3, TargetOpcode::G_SEXT); |
2198 | widenScalarDst(MI, WideTy); |
2199 | widenScalarDst(MI, WideTy, 1); |
2200 | Observer.changedInstr(MI); |
2201 | return Legalized; |
2202 | |
2203 | case TargetOpcode::G_ASHR: |
2204 | case TargetOpcode::G_LSHR: |
2205 | Observer.changingInstr(MI); |
2206 | |
2207 | if (TypeIdx == 0) { |
2208 | unsigned CvtOp = MI.getOpcode() == TargetOpcode::G_ASHR ? |
2209 | TargetOpcode::G_SEXT : TargetOpcode::G_ZEXT; |
2210 | |
2211 | widenScalarSrc(MI, WideTy, 1, CvtOp); |
2212 | widenScalarDst(MI, WideTy); |
2213 | } else { |
2214 | assert(TypeIdx == 1)((void)0); |
2215 | // The "number of bits to shift" operand must preserve its value as an |
2216 | // unsigned integer: |
2217 | widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ZEXT); |
2218 | } |
2219 | |
2220 | Observer.changedInstr(MI); |
2221 | return Legalized; |
2222 | case TargetOpcode::G_UDIV: |
2223 | case TargetOpcode::G_UREM: |
2224 | case TargetOpcode::G_UMIN: |
2225 | case TargetOpcode::G_UMAX: |
2226 | Observer.changingInstr(MI); |
2227 | widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ZEXT); |
2228 | widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ZEXT); |
2229 | widenScalarDst(MI, WideTy); |
2230 | Observer.changedInstr(MI); |
2231 | return Legalized; |
2232 | |
2233 | case TargetOpcode::G_UDIVREM: |
2234 | Observer.changingInstr(MI); |
2235 | widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ZEXT); |
2236 | widenScalarSrc(MI, WideTy, 3, TargetOpcode::G_ZEXT); |
2237 | widenScalarDst(MI, WideTy); |
2238 | widenScalarDst(MI, WideTy, 1); |
2239 | Observer.changedInstr(MI); |
2240 | return Legalized; |
2241 | |
2242 | case TargetOpcode::G_SELECT: |
2243 | Observer.changingInstr(MI); |
2244 | if (TypeIdx == 0) { |
2245 | // Perform operation at larger width (any extension is fine here, high |
2246 | // bits don't affect the result) and then truncate the result back to the |
2247 | // original type. |
2248 | widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ANYEXT); |
2249 | widenScalarSrc(MI, WideTy, 3, TargetOpcode::G_ANYEXT); |
2250 | widenScalarDst(MI, WideTy); |
2251 | } else { |
2252 | bool IsVec = MRI.getType(MI.getOperand(1).getReg()).isVector(); |
2253 | // Explicit extension is required here since high bits affect the result. |
2254 | widenScalarSrc(MI, WideTy, 1, MIRBuilder.getBoolExtOp(IsVec, false)); |
2255 | } |
2256 | Observer.changedInstr(MI); |
2257 | return Legalized; |
2258 | |
2259 | case TargetOpcode::G_FPTOSI: |
2260 | case TargetOpcode::G_FPTOUI: |
2261 | Observer.changingInstr(MI); |
2262 | |
2263 | if (TypeIdx == 0) |
2264 | widenScalarDst(MI, WideTy); |
2265 | else |
2266 | widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_FPEXT); |
2267 | |
2268 | Observer.changedInstr(MI); |
2269 | return Legalized; |
2270 | case TargetOpcode::G_SITOFP: |
2271 | Observer.changingInstr(MI); |
2272 | |
2273 | if (TypeIdx == 0) |
2274 | widenScalarDst(MI, WideTy, 0, TargetOpcode::G_FPTRUNC); |
2275 | else |
2276 | widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_SEXT); |
2277 | |
2278 | Observer.changedInstr(MI); |
2279 | return Legalized; |
2280 | case TargetOpcode::G_UITOFP: |
2281 | Observer.changingInstr(MI); |
2282 | |
2283 | if (TypeIdx == 0) |
2284 | widenScalarDst(MI, WideTy, 0, TargetOpcode::G_FPTRUNC); |
2285 | else |
2286 | widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ZEXT); |
2287 | |
2288 | Observer.changedInstr(MI); |
2289 | return Legalized; |
2290 | case TargetOpcode::G_LOAD: |
2291 | case TargetOpcode::G_SEXTLOAD: |
2292 | case TargetOpcode::G_ZEXTLOAD: |
2293 | Observer.changingInstr(MI); |
2294 | widenScalarDst(MI, WideTy); |
2295 | Observer.changedInstr(MI); |
2296 | return Legalized; |
2297 | |
2298 | case TargetOpcode::G_STORE: { |
2299 | if (TypeIdx != 0) |
2300 | return UnableToLegalize; |
2301 | |
2302 | LLT Ty = MRI.getType(MI.getOperand(0).getReg()); |
2303 | if (!Ty.isScalar()) |
2304 | return UnableToLegalize; |
2305 | |
2306 | Observer.changingInstr(MI); |
2307 | |
2308 | unsigned ExtType = Ty.getScalarSizeInBits() == 1 ? |
2309 | TargetOpcode::G_ZEXT : TargetOpcode::G_ANYEXT; |
2310 | widenScalarSrc(MI, WideTy, 0, ExtType); |
2311 | |
2312 | Observer.changedInstr(MI); |
2313 | return Legalized; |
2314 | } |
2315 | case TargetOpcode::G_CONSTANT: { |
2316 | MachineOperand &SrcMO = MI.getOperand(1); |
2317 | LLVMContext &Ctx = MIRBuilder.getMF().getFunction().getContext(); |
2318 | unsigned ExtOpc = LI.getExtOpcodeForWideningConstant( |
2319 | MRI.getType(MI.getOperand(0).getReg())); |
2320 | assert((ExtOpc == TargetOpcode::G_ZEXT || ExtOpc == TargetOpcode::G_SEXT ||((void)0) |
2321 | ExtOpc == TargetOpcode::G_ANYEXT) &&((void)0) |
2322 | "Illegal Extend")((void)0); |
2323 | const APInt &SrcVal = SrcMO.getCImm()->getValue(); |
2324 | const APInt &Val = (ExtOpc == TargetOpcode::G_SEXT) |
2325 | ? SrcVal.sext(WideTy.getSizeInBits()) |
2326 | : SrcVal.zext(WideTy.getSizeInBits()); |
2327 | Observer.changingInstr(MI); |
2328 | SrcMO.setCImm(ConstantInt::get(Ctx, Val)); |
2329 | |
2330 | widenScalarDst(MI, WideTy); |
2331 | Observer.changedInstr(MI); |
2332 | return Legalized; |
2333 | } |
2334 | case TargetOpcode::G_FCONSTANT: { |
2335 | MachineOperand &SrcMO = MI.getOperand(1); |
2336 | LLVMContext &Ctx = MIRBuilder.getMF().getFunction().getContext(); |
2337 | APFloat Val = SrcMO.getFPImm()->getValueAPF(); |
2338 | bool LosesInfo; |
2339 | switch (WideTy.getSizeInBits()) { |
2340 | case 32: |
2341 | Val.convert(APFloat::IEEEsingle(), APFloat::rmNearestTiesToEven, |
2342 | &LosesInfo); |
2343 | break; |
2344 | case 64: |
2345 | Val.convert(APFloat::IEEEdouble(), APFloat::rmNearestTiesToEven, |
2346 | &LosesInfo); |
2347 | break; |
2348 | default: |
2349 | return UnableToLegalize; |
2350 | } |
2351 | |
2352 | assert(!LosesInfo && "extend should always be lossless")((void)0); |
2353 | |
2354 | Observer.changingInstr(MI); |
2355 | SrcMO.setFPImm(ConstantFP::get(Ctx, Val)); |
2356 | |
2357 | widenScalarDst(MI, WideTy, 0, TargetOpcode::G_FPTRUNC); |
2358 | Observer.changedInstr(MI); |
2359 | return Legalized; |
2360 | } |
2361 | case TargetOpcode::G_IMPLICIT_DEF: { |
2362 | Observer.changingInstr(MI); |
2363 | widenScalarDst(MI, WideTy); |
2364 | Observer.changedInstr(MI); |
2365 | return Legalized; |
2366 | } |
2367 | case TargetOpcode::G_BRCOND: |
2368 | Observer.changingInstr(MI); |
2369 | widenScalarSrc(MI, WideTy, 0, MIRBuilder.getBoolExtOp(false, false)); |
2370 | Observer.changedInstr(MI); |
2371 | return Legalized; |
2372 | |
2373 | case TargetOpcode::G_FCMP: |
2374 | Observer.changingInstr(MI); |
2375 | if (TypeIdx == 0) |
2376 | widenScalarDst(MI, WideTy); |
2377 | else { |
2378 | widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_FPEXT); |
2379 | widenScalarSrc(MI, WideTy, 3, TargetOpcode::G_FPEXT); |
2380 | } |
2381 | Observer.changedInstr(MI); |
2382 | return Legalized; |
2383 | |
2384 | case TargetOpcode::G_ICMP: |
2385 | Observer.changingInstr(MI); |
2386 | if (TypeIdx == 0) |
2387 | widenScalarDst(MI, WideTy); |
2388 | else { |
2389 | unsigned ExtOpcode = CmpInst::isSigned(static_cast<CmpInst::Predicate>( |
2390 | MI.getOperand(1).getPredicate())) |
2391 | ? TargetOpcode::G_SEXT |
2392 | : TargetOpcode::G_ZEXT; |
2393 | widenScalarSrc(MI, WideTy, 2, ExtOpcode); |
2394 | widenScalarSrc(MI, WideTy, 3, ExtOpcode); |
2395 | } |
2396 | Observer.changedInstr(MI); |
2397 | return Legalized; |
2398 | |
2399 | case TargetOpcode::G_PTR_ADD: |
2400 | assert(TypeIdx == 1 && "unable to legalize pointer of G_PTR_ADD")((void)0); |
2401 | Observer.changingInstr(MI); |
2402 | widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_SEXT); |
2403 | Observer.changedInstr(MI); |
2404 | return Legalized; |
2405 | |
2406 | case TargetOpcode::G_PHI: { |
2407 | assert(TypeIdx == 0 && "Expecting only Idx 0")((void)0); |
2408 | |
2409 | Observer.changingInstr(MI); |
2410 | for (unsigned I = 1; I < MI.getNumOperands(); I += 2) { |
2411 | MachineBasicBlock &OpMBB = *MI.getOperand(I + 1).getMBB(); |
2412 | MIRBuilder.setInsertPt(OpMBB, OpMBB.getFirstTerminator()); |
2413 | widenScalarSrc(MI, WideTy, I, TargetOpcode::G_ANYEXT); |
2414 | } |
2415 | |
2416 | MachineBasicBlock &MBB = *MI.getParent(); |
2417 | MIRBuilder.setInsertPt(MBB, --MBB.getFirstNonPHI()); |
2418 | widenScalarDst(MI, WideTy); |
2419 | Observer.changedInstr(MI); |
2420 | return Legalized; |
2421 | } |
2422 | case TargetOpcode::G_EXTRACT_VECTOR_ELT: { |
2423 | if (TypeIdx == 0) { |
2424 | Register VecReg = MI.getOperand(1).getReg(); |
2425 | LLT VecTy = MRI.getType(VecReg); |
2426 | Observer.changingInstr(MI); |
2427 | |
2428 | widenScalarSrc( |
2429 | MI, LLT::vector(VecTy.getElementCount(), WideTy.getSizeInBits()), 1, |
2430 | TargetOpcode::G_SEXT); |
2431 | |
2432 | widenScalarDst(MI, WideTy, 0); |
2433 | Observer.changedInstr(MI); |
2434 | return Legalized; |
2435 | } |
2436 | |
2437 | if (TypeIdx != 2) |
2438 | return UnableToLegalize; |
2439 | Observer.changingInstr(MI); |
2440 | // TODO: Probably should be zext |
2441 | widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_SEXT); |
2442 | Observer.changedInstr(MI); |
2443 | return Legalized; |
2444 | } |
2445 | case TargetOpcode::G_INSERT_VECTOR_ELT: { |
2446 | if (TypeIdx == 1) { |
2447 | Observer.changingInstr(MI); |
2448 | |
2449 | Register VecReg = MI.getOperand(1).getReg(); |
2450 | LLT VecTy = MRI.getType(VecReg); |
2451 | LLT WideVecTy = LLT::vector(VecTy.getElementCount(), WideTy); |
2452 | |
2453 | widenScalarSrc(MI, WideVecTy, 1, TargetOpcode::G_ANYEXT); |
2454 | widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ANYEXT); |
2455 | widenScalarDst(MI, WideVecTy, 0); |
2456 | Observer.changedInstr(MI); |
2457 | return Legalized; |
2458 | } |
2459 | |
2460 | if (TypeIdx == 2) { |
2461 | Observer.changingInstr(MI); |
2462 | // TODO: Probably should be zext |
2463 | widenScalarSrc(MI, WideTy, 3, TargetOpcode::G_SEXT); |
2464 | Observer.changedInstr(MI); |
2465 | return Legalized; |
2466 | } |
2467 | |
2468 | return UnableToLegalize; |
2469 | } |
2470 | case TargetOpcode::G_FADD: |
2471 | case TargetOpcode::G_FMUL: |
2472 | case TargetOpcode::G_FSUB: |
2473 | case TargetOpcode::G_FMA: |
2474 | case TargetOpcode::G_FMAD: |
2475 | case TargetOpcode::G_FNEG: |
2476 | case TargetOpcode::G_FABS: |
2477 | case TargetOpcode::G_FCANONICALIZE: |
2478 | case TargetOpcode::G_FMINNUM: |
2479 | case TargetOpcode::G_FMAXNUM: |
2480 | case TargetOpcode::G_FMINNUM_IEEE: |
2481 | case TargetOpcode::G_FMAXNUM_IEEE: |
2482 | case TargetOpcode::G_FMINIMUM: |
2483 | case TargetOpcode::G_FMAXIMUM: |
2484 | case TargetOpcode::G_FDIV: |
2485 | case TargetOpcode::G_FREM: |
2486 | case TargetOpcode::G_FCEIL: |
2487 | case TargetOpcode::G_FFLOOR: |
2488 | case TargetOpcode::G_FCOS: |
2489 | case TargetOpcode::G_FSIN: |
2490 | case TargetOpcode::G_FLOG10: |
2491 | case TargetOpcode::G_FLOG: |
2492 | case TargetOpcode::G_FLOG2: |
2493 | case TargetOpcode::G_FRINT: |
2494 | case TargetOpcode::G_FNEARBYINT: |
2495 | case TargetOpcode::G_FSQRT: |
2496 | case TargetOpcode::G_FEXP: |
2497 | case TargetOpcode::G_FEXP2: |
2498 | case TargetOpcode::G_FPOW: |
2499 | case TargetOpcode::G_INTRINSIC_TRUNC: |
2500 | case TargetOpcode::G_INTRINSIC_ROUND: |
2501 | case TargetOpcode::G_INTRINSIC_ROUNDEVEN: |
2502 | assert(TypeIdx == 0)((void)0); |
2503 | Observer.changingInstr(MI); |
2504 | |
2505 | for (unsigned I = 1, E = MI.getNumOperands(); I != E; ++I) |
2506 | widenScalarSrc(MI, WideTy, I, TargetOpcode::G_FPEXT); |
2507 | |
2508 | widenScalarDst(MI, WideTy, 0, TargetOpcode::G_FPTRUNC); |
2509 | Observer.changedInstr(MI); |
2510 | return Legalized; |
2511 | case TargetOpcode::G_FPOWI: { |
2512 | if (TypeIdx != 0) |
2513 | return UnableToLegalize; |
2514 | Observer.changingInstr(MI); |
2515 | widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_FPEXT); |
2516 | widenScalarDst(MI, WideTy, 0, TargetOpcode::G_FPTRUNC); |
2517 | Observer.changedInstr(MI); |
2518 | return Legalized; |
2519 | } |
2520 | case TargetOpcode::G_INTTOPTR: |
2521 | if (TypeIdx != 1) |
2522 | return UnableToLegalize; |
2523 | |
2524 | Observer.changingInstr(MI); |
2525 | widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ZEXT); |
2526 | Observer.changedInstr(MI); |
2527 | return Legalized; |
2528 | case TargetOpcode::G_PTRTOINT: |
2529 | if (TypeIdx != 0) |
2530 | return UnableToLegalize; |
2531 | |
2532 | Observer.changingInstr(MI); |
2533 | widenScalarDst(MI, WideTy, 0); |
2534 | Observer.changedInstr(MI); |
2535 | return Legalized; |
2536 | case TargetOpcode::G_BUILD_VECTOR: { |
2537 | Observer.changingInstr(MI); |
2538 | |
2539 | const LLT WideEltTy = TypeIdx == 1 ? WideTy : WideTy.getElementType(); |
2540 | for (int I = 1, E = MI.getNumOperands(); I != E; ++I) |
2541 | widenScalarSrc(MI, WideEltTy, I, TargetOpcode::G_ANYEXT); |
2542 | |
2543 | // Avoid changing the result vector type if the source element type was |
2544 | // requested. |
2545 | if (TypeIdx == 1) { |
2546 | MI.setDesc(MIRBuilder.getTII().get(TargetOpcode::G_BUILD_VECTOR_TRUNC)); |
2547 | } else { |
2548 | widenScalarDst(MI, WideTy, 0); |
2549 | } |
2550 | |
2551 | Observer.changedInstr(MI); |
2552 | return Legalized; |
2553 | } |
2554 | case TargetOpcode::G_SEXT_INREG: |
2555 | if (TypeIdx != 0) |
2556 | return UnableToLegalize; |
2557 | |
2558 | Observer.changingInstr(MI); |
2559 | widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT); |
2560 | widenScalarDst(MI, WideTy, 0, TargetOpcode::G_TRUNC); |
2561 | Observer.changedInstr(MI); |
2562 | return Legalized; |
2563 | case TargetOpcode::G_PTRMASK: { |
2564 | if (TypeIdx != 1) |
2565 | return UnableToLegalize; |
2566 | Observer.changingInstr(MI); |
2567 | widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ZEXT); |
2568 | Observer.changedInstr(MI); |
2569 | return Legalized; |
2570 | } |
2571 | } |
2572 | } |
2573 | |
2574 | static void getUnmergePieces(SmallVectorImpl<Register> &Pieces, |
2575 | MachineIRBuilder &B, Register Src, LLT Ty) { |
2576 | auto Unmerge = B.buildUnmerge(Ty, Src); |
2577 | for (int I = 0, E = Unmerge->getNumOperands() - 1; I != E; ++I) |
2578 | Pieces.push_back(Unmerge.getReg(I)); |
2579 | } |
2580 | |
2581 | LegalizerHelper::LegalizeResult |
2582 | LegalizerHelper::lowerBitcast(MachineInstr &MI) { |
2583 | Register Dst = MI.getOperand(0).getReg(); |
2584 | Register Src = MI.getOperand(1).getReg(); |
2585 | LLT DstTy = MRI.getType(Dst); |
2586 | LLT SrcTy = MRI.getType(Src); |
2587 | |
2588 | if (SrcTy.isVector()) { |
2589 | LLT SrcEltTy = SrcTy.getElementType(); |
2590 | SmallVector<Register, 8> SrcRegs; |
2591 | |
2592 | if (DstTy.isVector()) { |
2593 | int NumDstElt = DstTy.getNumElements(); |
2594 | int NumSrcElt = SrcTy.getNumElements(); |
2595 | |
2596 | LLT DstEltTy = DstTy.getElementType(); |
2597 | LLT DstCastTy = DstEltTy; // Intermediate bitcast result type |
2598 | LLT SrcPartTy = SrcEltTy; // Original unmerge result type. |
2599 | |
2600 | // If there's an element size mismatch, insert intermediate casts to match |
2601 | // the result element type. |
2602 | if (NumSrcElt < NumDstElt) { // Source element type is larger. |
2603 | // %1:_(<4 x s8>) = G_BITCAST %0:_(<2 x s16>) |
2604 | // |
2605 | // => |
2606 | // |
2607 | // %2:_(s16), %3:_(s16) = G_UNMERGE_VALUES %0 |
2608 | // %3:_(<2 x s8>) = G_BITCAST %2 |
2609 | // %4:_(<2 x s8>) = G_BITCAST %3 |
2610 | // %1:_(<4 x s16>) = G_CONCAT_VECTORS %3, %4 |
2611 | DstCastTy = LLT::fixed_vector(NumDstElt / NumSrcElt, DstEltTy); |
2612 | SrcPartTy = SrcEltTy; |
2613 | } else if (NumSrcElt > NumDstElt) { // Source element type is smaller. |
2614 | // |
2615 | // %1:_(<2 x s16>) = G_BITCAST %0:_(<4 x s8>) |
2616 | // |
2617 | // => |
2618 | // |
2619 | // %2:_(<2 x s8>), %3:_(<2 x s8>) = G_UNMERGE_VALUES %0 |
2620 | // %3:_(s16) = G_BITCAST %2 |
2621 | // %4:_(s16) = G_BITCAST %3 |
2622 | // %1:_(<2 x s16>) = G_BUILD_VECTOR %3, %4 |
2623 | SrcPartTy = LLT::fixed_vector(NumSrcElt / NumDstElt, SrcEltTy); |
2624 | DstCastTy = DstEltTy; |
2625 | } |
2626 | |
2627 | getUnmergePieces(SrcRegs, MIRBuilder, Src, SrcPartTy); |
2628 | for (Register &SrcReg : SrcRegs) |
2629 | SrcReg = MIRBuilder.buildBitcast(DstCastTy, SrcReg).getReg(0); |
2630 | } else |
2631 | getUnmergePieces(SrcRegs, MIRBuilder, Src, SrcEltTy); |
2632 | |
2633 | MIRBuilder.buildMerge(Dst, SrcRegs); |
2634 | MI.eraseFromParent(); |
2635 | return Legalized; |
2636 | } |
2637 | |
2638 | if (DstTy.isVector()) { |
2639 | SmallVector<Register, 8> SrcRegs; |
2640 | getUnmergePieces(SrcRegs, MIRBuilder, Src, DstTy.getElementType()); |
2641 | MIRBuilder.buildMerge(Dst, SrcRegs); |
2642 | MI.eraseFromParent(); |
2643 | return Legalized; |
2644 | } |
2645 | |
2646 | return UnableToLegalize; |
2647 | } |
2648 | |
2649 | /// Figure out the bit offset into a register when coercing a vector index for |
2650 | /// the wide element type. This is only for the case when promoting vector to |
2651 | /// one with larger elements. |
2652 | // |
2653 | /// |
2654 | /// %offset_idx = G_AND %idx, ~(-1 << Log2(DstEltSize / SrcEltSize)) |
2655 | /// %offset_bits = G_SHL %offset_idx, Log2(SrcEltSize) |
2656 | static Register getBitcastWiderVectorElementOffset(MachineIRBuilder &B, |
2657 | Register Idx, |
2658 | unsigned NewEltSize, |
2659 | unsigned OldEltSize) { |
2660 | const unsigned Log2EltRatio = Log2_32(NewEltSize / OldEltSize); |
2661 | LLT IdxTy = B.getMRI()->getType(Idx); |
2662 | |
2663 | // Now figure out the amount we need to shift to get the target bits. |
2664 | auto OffsetMask = B.buildConstant( |
2665 | IdxTy, ~(APInt::getAllOnesValue(IdxTy.getSizeInBits()) << Log2EltRatio)); |
2666 | auto OffsetIdx = B.buildAnd(IdxTy, Idx, OffsetMask); |
2667 | return B.buildShl(IdxTy, OffsetIdx, |
2668 | B.buildConstant(IdxTy, Log2_32(OldEltSize))).getReg(0); |
2669 | } |
2670 | |
2671 | /// Perform a G_EXTRACT_VECTOR_ELT in a different sized vector element. If this |
2672 | /// is casting to a vector with a smaller element size, perform multiple element |
2673 | /// extracts and merge the results. If this is coercing to a vector with larger |
2674 | /// elements, index the bitcasted vector and extract the target element with bit |
2675 | /// operations. This is intended to force the indexing in the native register |
2676 | /// size for architectures that can dynamically index the register file. |
2677 | LegalizerHelper::LegalizeResult |
2678 | LegalizerHelper::bitcastExtractVectorElt(MachineInstr &MI, unsigned TypeIdx, |
2679 | LLT CastTy) { |
2680 | if (TypeIdx != 1) |
2681 | return UnableToLegalize; |
2682 | |
2683 | Register Dst = MI.getOperand(0).getReg(); |
2684 | Register SrcVec = MI.getOperand(1).getReg(); |
2685 | Register Idx = MI.getOperand(2).getReg(); |
2686 | LLT SrcVecTy = MRI.getType(SrcVec); |
2687 | LLT IdxTy = MRI.getType(Idx); |
2688 | |
2689 | LLT SrcEltTy = SrcVecTy.getElementType(); |
2690 | unsigned NewNumElts = CastTy.isVector() ? CastTy.getNumElements() : 1; |
2691 | unsigned OldNumElts = SrcVecTy.getNumElements(); |
2692 | |
2693 | LLT NewEltTy = CastTy.isVector() ? CastTy.getElementType() : CastTy; |
2694 | Register CastVec = MIRBuilder.buildBitcast(CastTy, SrcVec).getReg(0); |
2695 | |
2696 | const unsigned NewEltSize = NewEltTy.getSizeInBits(); |
2697 | const unsigned OldEltSize = SrcEltTy.getSizeInBits(); |
2698 | if (NewNumElts > OldNumElts) { |
2699 | // Decreasing the vector element size |
2700 | // |
2701 | // e.g. i64 = extract_vector_elt x:v2i64, y:i32 |
2702 | // => |
2703 | // v4i32:castx = bitcast x:v2i64 |
2704 | // |
2705 | // i64 = bitcast |
2706 | // (v2i32 build_vector (i32 (extract_vector_elt castx, (2 * y))), |
2707 | // (i32 (extract_vector_elt castx, (2 * y + 1))) |
2708 | // |
2709 | if (NewNumElts % OldNumElts != 0) |
2710 | return UnableToLegalize; |
2711 | |
2712 | // Type of the intermediate result vector. |
2713 | const unsigned NewEltsPerOldElt = NewNumElts / OldNumElts; |
2714 | LLT MidTy = |
2715 | LLT::scalarOrVector(ElementCount::getFixed(NewEltsPerOldElt), NewEltTy); |
2716 | |
2717 | auto NewEltsPerOldEltK = MIRBuilder.buildConstant(IdxTy, NewEltsPerOldElt); |
2718 | |
2719 | SmallVector<Register, 8> NewOps(NewEltsPerOldElt); |
2720 | auto NewBaseIdx = MIRBuilder.buildMul(IdxTy, Idx, NewEltsPerOldEltK); |
2721 | |
2722 | for (unsigned I = 0; I < NewEltsPerOldElt; ++I) { |
2723 | auto IdxOffset = MIRBuilder.buildConstant(IdxTy, I); |
2724 | auto TmpIdx = MIRBuilder.buildAdd(IdxTy, NewBaseIdx, IdxOffset); |
2725 | auto Elt = MIRBuilder.buildExtractVectorElement(NewEltTy, CastVec, TmpIdx); |
2726 | NewOps[I] = Elt.getReg(0); |
2727 | } |
2728 | |
2729 | auto NewVec = MIRBuilder.buildBuildVector(MidTy, NewOps); |
2730 | MIRBuilder.buildBitcast(Dst, NewVec); |
2731 | MI.eraseFromParent(); |
2732 | return Legalized; |
2733 | } |
2734 | |
2735 | if (NewNumElts < OldNumElts) { |
2736 | if (NewEltSize % OldEltSize != 0) |
2737 | return UnableToLegalize; |
2738 | |
2739 | // This only depends on powers of 2 because we use bit tricks to figure out |
2740 | // the bit offset we need to shift to get the target element. A general |
2741 | // expansion could emit division/multiply. |
2742 | if (!isPowerOf2_32(NewEltSize / OldEltSize)) |
2743 | return UnableToLegalize; |
2744 | |
2745 | // Increasing the vector element size. |
2746 | // %elt:_(small_elt) = G_EXTRACT_VECTOR_ELT %vec:_(<N x small_elt>), %idx |
2747 | // |
2748 | // => |
2749 | // |
2750 | // %cast = G_BITCAST %vec |
2751 | // %scaled_idx = G_LSHR %idx, Log2(DstEltSize / SrcEltSize) |
2752 | // %wide_elt = G_EXTRACT_VECTOR_ELT %cast, %scaled_idx |
2753 | // %offset_idx = G_AND %idx, ~(-1 << Log2(DstEltSize / SrcEltSize)) |
2754 | // %offset_bits = G_SHL %offset_idx, Log2(SrcEltSize) |
2755 | // %elt_bits = G_LSHR %wide_elt, %offset_bits |
2756 | // %elt = G_TRUNC %elt_bits |
2757 | |
2758 | const unsigned Log2EltRatio = Log2_32(NewEltSize / OldEltSize); |
2759 | auto Log2Ratio = MIRBuilder.buildConstant(IdxTy, Log2EltRatio); |
2760 | |
2761 | // Divide to get the index in the wider element type. |
2762 | auto ScaledIdx = MIRBuilder.buildLShr(IdxTy, Idx, Log2Ratio); |
2763 | |
2764 | Register WideElt = CastVec; |
2765 | if (CastTy.isVector()) { |
2766 | WideElt = MIRBuilder.buildExtractVectorElement(NewEltTy, CastVec, |
2767 | ScaledIdx).getReg(0); |
2768 | } |
2769 | |
2770 | // Compute the bit offset into the register of the target element. |
2771 | Register OffsetBits = getBitcastWiderVectorElementOffset( |
2772 | MIRBuilder, Idx, NewEltSize, OldEltSize); |
2773 | |
2774 | // Shift the wide element to get the target element. |
2775 | auto ExtractedBits = MIRBuilder.buildLShr(NewEltTy, WideElt, OffsetBits); |
2776 | MIRBuilder.buildTrunc(Dst, ExtractedBits); |
2777 | MI.eraseFromParent(); |
2778 | return Legalized; |
2779 | } |
2780 | |
2781 | return UnableToLegalize; |
2782 | } |
2783 | |
2784 | /// Emit code to insert \p InsertReg into \p TargetRet at \p OffsetBits in \p |
2785 | /// TargetReg, while preserving other bits in \p TargetReg. |
2786 | /// |
2787 | /// (InsertReg << Offset) | (TargetReg & ~(-1 >> InsertReg.size()) << Offset) |
2788 | static Register buildBitFieldInsert(MachineIRBuilder &B, |
2789 | Register TargetReg, Register InsertReg, |
2790 | Register OffsetBits) { |
2791 | LLT TargetTy = B.getMRI()->getType(TargetReg); |
2792 | LLT InsertTy = B.getMRI()->getType(InsertReg); |
2793 | auto ZextVal = B.buildZExt(TargetTy, InsertReg); |
2794 | auto ShiftedInsertVal = B.buildShl(TargetTy, ZextVal, OffsetBits); |
2795 | |
2796 | // Produce a bitmask of the value to insert |
2797 | auto EltMask = B.buildConstant( |
2798 | TargetTy, APInt::getLowBitsSet(TargetTy.getSizeInBits(), |
2799 | InsertTy.getSizeInBits())); |
2800 | // Shift it into position |
2801 | auto ShiftedMask = B.buildShl(TargetTy, EltMask, OffsetBits); |
2802 | auto InvShiftedMask = B.buildNot(TargetTy, ShiftedMask); |
2803 | |
2804 | // Clear out the bits in the wide element |
2805 | auto MaskedOldElt = B.buildAnd(TargetTy, TargetReg, InvShiftedMask); |
2806 | |
2807 | // The value to insert has all zeros already, so stick it into the masked |
2808 | // wide element. |
2809 | return B.buildOr(TargetTy, MaskedOldElt, ShiftedInsertVal).getReg(0); |
2810 | } |
2811 | |
2812 | /// Perform a G_INSERT_VECTOR_ELT in a different sized vector element. If this |
2813 | /// is increasing the element size, perform the indexing in the target element |
2814 | /// type, and use bit operations to insert at the element position. This is |
2815 | /// intended for architectures that can dynamically index the register file and |
2816 | /// want to force indexing in the native register size. |
2817 | LegalizerHelper::LegalizeResult |
2818 | LegalizerHelper::bitcastInsertVectorElt(MachineInstr &MI, unsigned TypeIdx, |
2819 | LLT CastTy) { |
2820 | if (TypeIdx != 0) |
2821 | return UnableToLegalize; |
2822 | |
2823 | Register Dst = MI.getOperand(0).getReg(); |
2824 | Register SrcVec = MI.getOperand(1).getReg(); |
2825 | Register Val = MI.getOperand(2).getReg(); |
2826 | Register Idx = MI.getOperand(3).getReg(); |
2827 | |
2828 | LLT VecTy = MRI.getType(Dst); |
2829 | LLT IdxTy = MRI.getType(Idx); |
2830 | |
2831 | LLT VecEltTy = VecTy.getElementType(); |
2832 | LLT NewEltTy = CastTy.isVector() ? CastTy.getElementType() : CastTy; |
2833 | const unsigned NewEltSize = NewEltTy.getSizeInBits(); |
2834 | const unsigned OldEltSize = VecEltTy.getSizeInBits(); |
2835 | |
2836 | unsigned NewNumElts = CastTy.isVector() ? CastTy.getNumElements() : 1; |
2837 | unsigned OldNumElts = VecTy.getNumElements(); |
2838 | |
2839 | Register CastVec = MIRBuilder.buildBitcast(CastTy, SrcVec).getReg(0); |
2840 | if (NewNumElts < OldNumElts) { |
2841 | if (NewEltSize % OldEltSize != 0) |
2842 | return UnableToLegalize; |
2843 | |
2844 | // This only depends on powers of 2 because we use bit tricks to figure out |
2845 | // the bit offset we need to shift to get the target element. A general |
2846 | // expansion could emit division/multiply. |
2847 | if (!isPowerOf2_32(NewEltSize / OldEltSize)) |
2848 | return UnableToLegalize; |
2849 | |
2850 | const unsigned Log2EltRatio = Log2_32(NewEltSize / OldEltSize); |
2851 | auto Log2Ratio = MIRBuilder.buildConstant(IdxTy, Log2EltRatio); |
2852 | |
2853 | // Divide to get the index in the wider element type. |
2854 | auto ScaledIdx = MIRBuilder.buildLShr(IdxTy, Idx, Log2Ratio); |
2855 | |
2856 | Register ExtractedElt = CastVec; |
2857 | if (CastTy.isVector()) { |
2858 | ExtractedElt = MIRBuilder.buildExtractVectorElement(NewEltTy, CastVec, |
2859 | ScaledIdx).getReg(0); |
2860 | } |
2861 | |
2862 | // Compute the bit offset into the register of the target element. |
2863 | Register OffsetBits = getBitcastWiderVectorElementOffset( |
2864 | MIRBuilder, Idx, NewEltSize, OldEltSize); |
2865 | |
2866 | Register InsertedElt = buildBitFieldInsert(MIRBuilder, ExtractedElt, |
2867 | Val, OffsetBits); |
2868 | if (CastTy.isVector()) { |
2869 | InsertedElt = MIRBuilder.buildInsertVectorElement( |
2870 | CastTy, CastVec, InsertedElt, ScaledIdx).getReg(0); |
2871 | } |
2872 | |
2873 | MIRBuilder.buildBitcast(Dst, InsertedElt); |
2874 | MI.eraseFromParent(); |
2875 | return Legalized; |
2876 | } |
2877 | |
2878 | return UnableToLegalize; |
2879 | } |
2880 | |
2881 | LegalizerHelper::LegalizeResult LegalizerHelper::lowerLoad(GAnyLoad &LoadMI) { |
2882 | // Lower to a memory-width G_LOAD and a G_SEXT/G_ZEXT/G_ANYEXT |
2883 | Register DstReg = LoadMI.getDstReg(); |
2884 | Register PtrReg = LoadMI.getPointerReg(); |
2885 | LLT DstTy = MRI.getType(DstReg); |
2886 | MachineMemOperand &MMO = LoadMI.getMMO(); |
2887 | LLT MemTy = MMO.getMemoryType(); |
2888 | MachineFunction &MF = MIRBuilder.getMF(); |
2889 | if (MemTy.isVector()) |
2890 | return UnableToLegalize; |
2891 | |
2892 | unsigned MemSizeInBits = MemTy.getSizeInBits(); |
2893 | unsigned MemStoreSizeInBits = 8 * MemTy.getSizeInBytes(); |
2894 | |
2895 | if (MemSizeInBits != MemStoreSizeInBits) { |
2896 | // Promote to a byte-sized load if not loading an integral number of |
2897 | // bytes. For example, promote EXTLOAD:i20 -> EXTLOAD:i24. |
2898 | LLT WideMemTy = LLT::scalar(MemStoreSizeInBits); |
2899 | MachineMemOperand *NewMMO = |
2900 | MF.getMachineMemOperand(&MMO, MMO.getPointerInfo(), WideMemTy); |
2901 | |
2902 | Register LoadReg = DstReg; |
2903 | LLT LoadTy = DstTy; |
2904 | |
2905 | // If this wasn't already an extending load, we need to widen the result |
2906 | // register to avoid creating a load with a narrower result than the source. |
2907 | if (MemStoreSizeInBits > DstTy.getSizeInBits()) { |
2908 | LoadTy = WideMemTy; |
2909 | LoadReg = MRI.createGenericVirtualRegister(WideMemTy); |
2910 | } |
2911 | |
2912 | if (isa<GSExtLoad>(LoadMI)) { |
2913 | auto NewLoad = MIRBuilder.buildLoad(LoadTy, PtrReg, *NewMMO); |
2914 | MIRBuilder.buildSExtInReg(LoadReg, NewLoad, MemSizeInBits); |
2915 | } else if (isa<GZExtLoad>(LoadMI) || WideMemTy == DstTy) { |
2916 | auto NewLoad = MIRBuilder.buildLoad(LoadTy, PtrReg, *NewMMO); |
2917 | // The extra bits are guaranteed to be zero, since we stored them that |
2918 | // way. A zext load from Wide thus automatically gives zext from MemVT. |
2919 | MIRBuilder.buildAssertZExt(LoadReg, NewLoad, MemSizeInBits); |
2920 | } else { |
2921 | MIRBuilder.buildLoad(LoadReg, PtrReg, *NewMMO); |
2922 | } |
2923 | |
2924 | if (DstTy != LoadTy) |
2925 | MIRBuilder.buildTrunc(DstReg, LoadReg); |
2926 | |
2927 | LoadMI.eraseFromParent(); |
2928 | return Legalized; |
2929 | } |
2930 | |
2931 | // This load needs splitting into power of 2 sized loads. |
2932 | if (DstTy.isVector()) |
2933 | return UnableToLegalize; |
2934 | if (isPowerOf2_32(MemSizeInBits)) |
2935 | return UnableToLegalize; // Don't know what we're being asked to do. |
2936 | |
2937 | // Big endian lowering not implemented. |
2938 | if (MIRBuilder.getDataLayout().isBigEndian()) |
2939 | return UnableToLegalize; |
2940 | |
2941 | // Our strategy here is to generate anyextending loads for the smaller |
2942 | // types up to next power-2 result type, and then combine the two larger |
2943 | // result values together, before truncating back down to the non-pow-2 |
2944 | // type. |
2945 | // E.g. v1 = i24 load => |
2946 | // v2 = i32 zextload (2 byte) |
2947 | // v3 = i32 load (1 byte) |
2948 | // v4 = i32 shl v3, 16 |
2949 | // v5 = i32 or v4, v2 |
2950 | // v1 = i24 trunc v5 |
2951 | // By doing this we generate the correct truncate which should get |
2952 | // combined away as an artifact with a matching extend. |
2953 | uint64_t LargeSplitSize = PowerOf2Floor(MemSizeInBits); |
2954 | uint64_t SmallSplitSize = MemSizeInBits - LargeSplitSize; |
2955 | |
2956 | MachineMemOperand *LargeMMO = |
2957 | MF.getMachineMemOperand(&MMO, 0, LargeSplitSize / 8); |
2958 | MachineMemOperand *SmallMMO = |
2959 | MF.getMachineMemOperand(&MMO, LargeSplitSize / 8, SmallSplitSize / 8); |
2960 | |
2961 | LLT PtrTy = MRI.getType(PtrReg); |
2962 | unsigned AnyExtSize = PowerOf2Ceil(DstTy.getSizeInBits()); |
2963 | LLT AnyExtTy = LLT::scalar(AnyExtSize); |
2964 | auto LargeLoad = MIRBuilder.buildLoadInstr(TargetOpcode::G_ZEXTLOAD, AnyExtTy, |
2965 | PtrReg, *LargeMMO); |
2966 | |
2967 | auto OffsetCst = MIRBuilder.buildConstant(LLT::scalar(PtrTy.getSizeInBits()), |
2968 | LargeSplitSize / 8); |
2969 | Register PtrAddReg = MRI.createGenericVirtualRegister(PtrTy); |
2970 | auto SmallPtr = MIRBuilder.buildPtrAdd(PtrAddReg, PtrReg, OffsetCst); |
2971 | auto SmallLoad = MIRBuilder.buildLoadInstr(LoadMI.getOpcode(), AnyExtTy, |
2972 | SmallPtr, *SmallMMO); |
2973 | |
2974 | auto ShiftAmt = MIRBuilder.buildConstant(AnyExtTy, LargeSplitSize); |
2975 | auto Shift = MIRBuilder.buildShl(AnyExtTy, SmallLoad, ShiftAmt); |
2976 | |
2977 | if (AnyExtTy == DstTy) |
2978 | MIRBuilder.buildOr(DstReg, Shift, LargeLoad); |
2979 | else { |
2980 | auto Or = MIRBuilder.buildOr(AnyExtTy, Shift, LargeLoad); |
2981 | MIRBuilder.buildTrunc(DstReg, {Or}); |
2982 | } |
2983 | |
2984 | LoadMI.eraseFromParent(); |
2985 | return Legalized; |
2986 | } |
2987 | |
2988 | LegalizerHelper::LegalizeResult LegalizerHelper::lowerStore(GStore &StoreMI) { |
2989 | // Lower a non-power of 2 store into multiple pow-2 stores. |
2990 | // E.g. split an i24 store into an i16 store + i8 store. |
2991 | // We do this by first extending the stored value to the next largest power |
2992 | // of 2 type, and then using truncating stores to store the components. |
2993 | // By doing this, likewise with G_LOAD, generate an extend that can be |
2994 | // artifact-combined away instead of leaving behind extracts. |
2995 | Register SrcReg = StoreMI.getValueReg(); |
2996 | Register PtrReg = StoreMI.getPointerReg(); |
2997 | LLT SrcTy = MRI.getType(SrcReg); |
2998 | MachineFunction &MF = MIRBuilder.getMF(); |
2999 | MachineMemOperand &MMO = **StoreMI.memoperands_begin(); |
3000 | LLT MemTy = MMO.getMemoryType(); |
3001 | |
3002 | if (SrcTy.isVector()) |
3003 | return UnableToLegalize; |
3004 | |
3005 | unsigned StoreWidth = MemTy.getSizeInBits(); |
3006 | unsigned StoreSizeInBits = 8 * MemTy.getSizeInBytes(); |
3007 | |
3008 | if (StoreWidth != StoreSizeInBits) { |
3009 | // Promote to a byte-sized store with upper bits zero if not |
3010 | // storing an integral number of bytes. For example, promote |
3011 | // TRUNCSTORE:i1 X -> TRUNCSTORE:i8 (and X, 1) |
3012 | LLT WideTy = LLT::scalar(StoreSizeInBits); |
3013 | |
3014 | if (StoreSizeInBits > SrcTy.getSizeInBits()) { |
3015 | // Avoid creating a store with a narrower source than result. |
3016 | SrcReg = MIRBuilder.buildAnyExt(WideTy, SrcReg).getReg(0); |
3017 | SrcTy = WideTy; |
3018 | } |
3019 | |
3020 | auto ZextInReg = MIRBuilder.buildZExtInReg(SrcTy, SrcReg, StoreWidth); |
3021 | |
3022 | MachineMemOperand *NewMMO = |
3023 | MF.getMachineMemOperand(&MMO, MMO.getPointerInfo(), WideTy); |
3024 | MIRBuilder.buildStore(ZextInReg, PtrReg, *NewMMO); |
3025 | StoreMI.eraseFromParent(); |
3026 | return Legalized; |
3027 | } |
3028 | |
3029 | if (isPowerOf2_32(MemTy.getSizeInBits())) |
3030 | return UnableToLegalize; // Don't know what we're being asked to do. |
3031 | |
3032 | // Extend to the next pow-2. If this store was itself the result of lowering, |
3033 | // e.g. an s56 store being broken into s32 + s24, we might have a stored type |
3034 | // that's wider the stored size. |
3035 | const LLT NewSrcTy = LLT::scalar(NextPowerOf2(MemTy.getSizeInBits())); |
3036 | auto ExtVal = MIRBuilder.buildAnyExtOrTrunc(NewSrcTy, SrcReg); |
3037 | |
3038 | // Obtain the smaller value by shifting away the larger value. |
3039 | uint64_t LargeSplitSize = PowerOf2Floor(MemTy.getSizeInBits()); |
3040 | uint64_t SmallSplitSize = MemTy.getSizeInBits() - LargeSplitSize; |
3041 | auto ShiftAmt = MIRBuilder.buildConstant(NewSrcTy, LargeSplitSize); |
3042 | auto SmallVal = MIRBuilder.buildLShr(NewSrcTy, ExtVal, ShiftAmt); |
3043 | |
3044 | // Generate the PtrAdd and truncating stores. |
3045 | LLT PtrTy = MRI.getType(PtrReg); |
3046 | auto OffsetCst = MIRBuilder.buildConstant( |
3047 | LLT::scalar(PtrTy.getSizeInBits()), LargeSplitSize / 8); |
3048 | Register PtrAddReg = MRI.createGenericVirtualRegister(PtrTy); |
3049 | auto SmallPtr = |
3050 | MIRBuilder.buildPtrAdd(PtrAddReg, PtrReg, OffsetCst); |
3051 | |
3052 | MachineMemOperand *LargeMMO = |
3053 | MF.getMachineMemOperand(&MMO, 0, LargeSplitSize / 8); |
3054 | MachineMemOperand *SmallMMO = |
3055 | MF.getMachineMemOperand(&MMO, LargeSplitSize / 8, SmallSplitSize / 8); |
3056 | MIRBuilder.buildStore(ExtVal, PtrReg, *LargeMMO); |
3057 | MIRBuilder.buildStore(SmallVal, SmallPtr, *SmallMMO); |
3058 | StoreMI.eraseFromParent(); |
3059 | return Legalized; |
3060 | } |
3061 | |
3062 | LegalizerHelper::LegalizeResult |
3063 | LegalizerHelper::bitcast(MachineInstr &MI, unsigned TypeIdx, LLT CastTy) { |
3064 | switch (MI.getOpcode()) { |
3065 | case TargetOpcode::G_LOAD: { |
3066 | if (TypeIdx != 0) |
3067 | return UnableToLegalize; |
3068 | MachineMemOperand &MMO = **MI.memoperands_begin(); |
3069 | |
3070 | // Not sure how to interpret a bitcast of an extending load. |
3071 | if (MMO.getMemoryType().getSizeInBits() != CastTy.getSizeInBits()) |
3072 | return UnableToLegalize; |
3073 | |
3074 | Observer.changingInstr(MI); |
3075 | bitcastDst(MI, CastTy, 0); |
3076 | MMO.setType(CastTy); |
3077 | Observer.changedInstr(MI); |
3078 | return Legalized; |
3079 | } |
3080 | case TargetOpcode::G_STORE: { |
3081 | if (TypeIdx != 0) |
3082 | return UnableToLegalize; |
3083 | |
3084 | MachineMemOperand &MMO = **MI.memoperands_begin(); |
3085 | |
3086 | // Not sure how to interpret a bitcast of a truncating store. |
3087 | if (MMO.getMemoryType().getSizeInBits() != CastTy.getSizeInBits()) |
3088 | return UnableToLegalize; |
3089 | |
3090 | Observer.changingInstr(MI); |
3091 | bitcastSrc(MI, CastTy, 0); |
3092 | MMO.setType(CastTy); |
3093 | Observer.changedInstr(MI); |
3094 | return Legalized; |
3095 | } |
3096 | case TargetOpcode::G_SELECT: { |
3097 | if (TypeIdx != 0) |
3098 | return UnableToLegalize; |
3099 | |
3100 | if (MRI.getType(MI.getOperand(1).getReg()).isVector()) { |
3101 | LLVM_DEBUG(do { } while (false) |
3102 | dbgs() << "bitcast action not implemented for vector select\n")do { } while (false); |
3103 | return UnableToLegalize; |
3104 | } |
3105 | |
3106 | Observer.changingInstr(MI); |
3107 | bitcastSrc(MI, CastTy, 2); |
3108 | bitcastSrc(MI, CastTy, 3); |
3109 | bitcastDst(MI, CastTy, 0); |
3110 | Observer.changedInstr(MI); |
3111 | return Legalized; |
3112 | } |
3113 | case TargetOpcode::G_AND: |
3114 | case TargetOpcode::G_OR: |
3115 | case TargetOpcode::G_XOR: { |
3116 | Observer.changingInstr(MI); |
3117 | bitcastSrc(MI, CastTy, 1); |
3118 | bitcastSrc(MI, CastTy, 2); |
3119 | bitcastDst(MI, CastTy, 0); |
3120 | Observer.changedInstr(MI); |
3121 | return Legalized; |
3122 | } |
3123 | case TargetOpcode::G_EXTRACT_VECTOR_ELT: |
3124 | return bitcastExtractVectorElt(MI, TypeIdx, CastTy); |
3125 | case TargetOpcode::G_INSERT_VECTOR_ELT: |
3126 | return bitcastInsertVectorElt(MI, TypeIdx, CastTy); |
3127 | default: |
3128 | return UnableToLegalize; |
3129 | } |
3130 | } |
3131 | |
3132 | // Legalize an instruction by changing the opcode in place. |
3133 | void LegalizerHelper::changeOpcode(MachineInstr &MI, unsigned NewOpcode) { |
3134 | Observer.changingInstr(MI); |
3135 | MI.setDesc(MIRBuilder.getTII().get(NewOpcode)); |
3136 | Observer.changedInstr(MI); |
3137 | } |
3138 | |
3139 | LegalizerHelper::LegalizeResult |
3140 | LegalizerHelper::lower(MachineInstr &MI, unsigned TypeIdx, LLT LowerHintTy) { |
3141 | using namespace TargetOpcode; |
3142 | |
3143 | switch(MI.getOpcode()) { |
3144 | default: |
3145 | return UnableToLegalize; |
3146 | case TargetOpcode::G_BITCAST: |
3147 | return lowerBitcast(MI); |
3148 | case TargetOpcode::G_SREM: |
3149 | case TargetOpcode::G_UREM: { |
3150 | LLT Ty = MRI.getType(MI.getOperand(0).getReg()); |
3151 | auto Quot = |
3152 | MIRBuilder.buildInstr(MI.getOpcode() == G_SREM ? G_SDIV : G_UDIV, {Ty}, |
3153 | {MI.getOperand(1), MI.getOperand(2)}); |
3154 | |
3155 | auto Prod = MIRBuilder.buildMul(Ty, Quot, MI.getOperand(2)); |
3156 | MIRBuilder.buildSub(MI.getOperand(0), MI.getOperand(1), Prod); |
3157 | MI.eraseFromParent(); |
3158 | return Legalized; |
3159 | } |
3160 | case TargetOpcode::G_SADDO: |
3161 | case TargetOpcode::G_SSUBO: |
3162 | return lowerSADDO_SSUBO(MI); |
3163 | case TargetOpcode::G_UMULH: |
3164 | case TargetOpcode::G_SMULH: |
3165 | return lowerSMULH_UMULH(MI); |
3166 | case TargetOpcode::G_SMULO: |
3167 | case TargetOpcode::G_UMULO: { |
3168 | // Generate G_UMULH/G_SMULH to check for overflow and a normal G_MUL for the |
3169 | // result. |
3170 | Register Res = MI.getOperand(0).getReg(); |
3171 | Register Overflow = MI.getOperand(1).getReg(); |
3172 | Register LHS = MI.getOperand(2).getReg(); |
3173 | Register RHS = MI.getOperand(3).getReg(); |
3174 | LLT Ty = MRI.getType(Res); |
3175 | |
3176 | unsigned Opcode = MI.getOpcode() == TargetOpcode::G_SMULO |
3177 | ? TargetOpcode::G_SMULH |
3178 | : TargetOpcode::G_UMULH; |
3179 | |
3180 | Observer.changingInstr(MI); |
3181 | const auto &TII = MIRBuilder.getTII(); |
3182 | MI.setDesc(TII.get(TargetOpcode::G_MUL)); |
3183 | MI.RemoveOperand(1); |
3184 | Observer.changedInstr(MI); |
3185 | |
3186 | auto HiPart = MIRBuilder.buildInstr(Opcode, {Ty}, {LHS, RHS}); |
3187 | auto Zero = MIRBuilder.buildConstant(Ty, 0); |
3188 | |
3189 | // Move insert point forward so we can use the Res register if needed. |
3190 | MIRBuilder.setInsertPt(MIRBuilder.getMBB(), ++MIRBuilder.getInsertPt()); |
3191 | |
3192 | // For *signed* multiply, overflow is detected by checking: |
3193 | // (hi != (lo >> bitwidth-1)) |
3194 | if (Opcode == TargetOpcode::G_SMULH) { |
3195 | auto ShiftAmt = MIRBuilder.buildConstant(Ty, Ty.getSizeInBits() - 1); |
3196 | auto Shifted = MIRBuilder.buildAShr(Ty, Res, ShiftAmt); |
3197 | MIRBuilder.buildICmp(CmpInst::ICMP_NE, Overflow, HiPart, Shifted); |
3198 | } else { |
3199 | MIRBuilder.buildICmp(CmpInst::ICMP_NE, Overflow, HiPart, Zero); |
3200 | } |
3201 | return Legalized; |
3202 | } |
3203 | case TargetOpcode::G_FNEG: { |
3204 | Register Res = MI.getOperand(0).getReg(); |
3205 | LLT Ty = MRI.getType(Res); |
3206 | |
3207 | // TODO: Handle vector types once we are able to |
3208 | // represent them. |
3209 | if (Ty.isVector()) |
3210 | return UnableToLegalize; |
3211 | auto SignMask = |
3212 | MIRBuilder.buildConstant(Ty, APInt::getSignMask(Ty.getSizeInBits())); |
3213 | Register SubByReg = MI.getOperand(1).getReg(); |
3214 | MIRBuilder.buildXor(Res, SubByReg, SignMask); |
3215 | MI.eraseFromParent(); |
3216 | return Legalized; |
3217 | } |
3218 | case TargetOpcode::G_FSUB: { |
3219 | Register Res = MI.getOperand(0).getReg(); |
3220 | LLT Ty = MRI.getType(Res); |
3221 | |
3222 | // Lower (G_FSUB LHS, RHS) to (G_FADD LHS, (G_FNEG RHS)). |
3223 | // First, check if G_FNEG is marked as Lower. If so, we may |
3224 | // end up with an infinite loop as G_FSUB is used to legalize G_FNEG. |
3225 | if (LI.getAction({G_FNEG, {Ty}}).Action == Lower) |
3226 | return UnableToLegalize; |
3227 | Register LHS = MI.getOperand(1).getReg(); |
3228 | Register RHS = MI.getOperand(2).getReg(); |
3229 | Register Neg = MRI.createGenericVirtualRegister(Ty); |
3230 | MIRBuilder.buildFNeg(Neg, RHS); |
3231 | MIRBuilder.buildFAdd(Res, LHS, Neg, MI.getFlags()); |
3232 | MI.eraseFromParent(); |
3233 | return Legalized; |
3234 | } |
3235 | case TargetOpcode::G_FMAD: |
3236 | return lowerFMad(MI); |
3237 | case TargetOpcode::G_FFLOOR: |
3238 | return lowerFFloor(MI); |
3239 | case TargetOpcode::G_INTRINSIC_ROUND: |
3240 | return lowerIntrinsicRound(MI); |
3241 | case TargetOpcode::G_INTRINSIC_ROUNDEVEN: { |
3242 | // Since round even is the assumed rounding mode for unconstrained FP |
3243 | // operations, rint and roundeven are the same operation. |
3244 | changeOpcode(MI, TargetOpcode::G_FRINT); |
3245 | return Legalized; |
3246 | } |
3247 | case TargetOpcode::G_ATOMIC_CMPXCHG_WITH_SUCCESS: { |
3248 | Register OldValRes = MI.getOperand(0).getReg(); |
3249 | Register SuccessRes = MI.getOperand(1).getReg(); |
3250 | Register Addr = MI.getOperand(2).getReg(); |
3251 | Register CmpVal = MI.getOperand(3).getReg(); |
3252 | Register NewVal = MI.getOperand(4).getReg(); |
3253 | MIRBuilder.buildAtomicCmpXchg(OldValRes, Addr, CmpVal, NewVal, |
3254 | **MI.memoperands_begin()); |
3255 | MIRBuilder.buildICmp(CmpInst::ICMP_EQ, SuccessRes, OldValRes, CmpVal); |
3256 | MI.eraseFromParent(); |
3257 | return Legalized; |
3258 | } |
3259 | case TargetOpcode::G_LOAD: |
3260 | case TargetOpcode::G_SEXTLOAD: |
3261 | case TargetOpcode::G_ZEXTLOAD: |
3262 | return lowerLoad(cast<GAnyLoad>(MI)); |
3263 | case TargetOpcode::G_STORE: |
3264 | return lowerStore(cast<GStore>(MI)); |
3265 | case TargetOpcode::G_CTLZ_ZERO_UNDEF: |
3266 | case TargetOpcode::G_CTTZ_ZERO_UNDEF: |
3267 | case TargetOpcode::G_CTLZ: |
3268 | case TargetOpcode::G_CTTZ: |
3269 | case TargetOpcode::G_CTPOP: |
3270 | return lowerBitCount(MI); |
3271 | case G_UADDO: { |
3272 | Register Res = MI.getOperand(0).getReg(); |
3273 | Register CarryOut = MI.getOperand(1).getReg(); |
3274 | Register LHS = MI.getOperand(2).getReg(); |
3275 | Register RHS = MI.getOperand(3).getReg(); |
3276 | |
3277 | MIRBuilder.buildAdd(Res, LHS, RHS); |
3278 | MIRBuilder.buildICmp(CmpInst::ICMP_ULT, CarryOut, Res, RHS); |
3279 | |
3280 | MI.eraseFromParent(); |
3281 | return Legalized; |
3282 | } |
3283 | case G_UADDE: { |
3284 | Register Res = MI.getOperand(0).getReg(); |
3285 | Register CarryOut = MI.getOperand(1).getReg(); |
3286 | Register LHS = MI.getOperand(2).getReg(); |
3287 | Register RHS = MI.getOperand(3).getReg(); |
3288 | Register CarryIn = MI.getOperand(4).getReg(); |
3289 | LLT Ty = MRI.getType(Res); |
3290 | |
3291 | auto TmpRes = MIRBuilder.buildAdd(Ty, LHS, RHS); |
3292 | auto ZExtCarryIn = MIRBuilder.buildZExt(Ty, CarryIn); |
3293 | MIRBuilder.buildAdd(Res, TmpRes, ZExtCarryIn); |
3294 | MIRBuilder.buildICmp(CmpInst::ICMP_ULT, CarryOut, Res, LHS); |
3295 | |
3296 | MI.eraseFromParent(); |
3297 | return Legalized; |
3298 | } |
3299 | case G_USUBO: { |
3300 | Register Res = MI.getOperand(0).getReg(); |
3301 | Register BorrowOut = MI.getOperand(1).getReg(); |
3302 | Register LHS = MI.getOperand(2).getReg(); |
3303 | Register RHS = MI.getOperand(3).getReg(); |
3304 | |
3305 | MIRBuilder.buildSub(Res, LHS, RHS); |
3306 | MIRBuilder.buildICmp(CmpInst::ICMP_ULT, BorrowOut, LHS, RHS); |
3307 | |
3308 | MI.eraseFromParent(); |
3309 | return Legalized; |
3310 | } |
3311 | case G_USUBE: { |
3312 | Register Res = MI.getOperand(0).getReg(); |
3313 | Register BorrowOut = MI.getOperand(1).getReg(); |
3314 | Register LHS = MI.getOperand(2).getReg(); |
3315 | Register RHS = MI.getOperand(3).getReg(); |
3316 | Register BorrowIn = MI.getOperand(4).getReg(); |
3317 | const LLT CondTy = MRI.getType(BorrowOut); |
3318 | const LLT Ty = MRI.getType(Res); |
3319 | |
3320 | auto TmpRes = MIRBuilder.buildSub(Ty, LHS, RHS); |
3321 | auto ZExtBorrowIn = MIRBuilder.buildZExt(Ty, BorrowIn); |
3322 | MIRBuilder.buildSub(Res, TmpRes, ZExtBorrowIn); |
3323 | |
3324 | auto LHS_EQ_RHS = MIRBuilder.buildICmp(CmpInst::ICMP_EQ, CondTy, LHS, RHS); |
3325 | auto LHS_ULT_RHS = MIRBuilder.buildICmp(CmpInst::ICMP_ULT, CondTy, LHS, RHS); |
3326 | MIRBuilder.buildSelect(BorrowOut, LHS_EQ_RHS, BorrowIn, LHS_ULT_RHS); |
3327 | |
3328 | MI.eraseFromParent(); |
3329 | return Legalized; |
3330 | } |
3331 | case G_UITOFP: |
3332 | return lowerUITOFP(MI); |
3333 | case G_SITOFP: |
3334 | return lowerSITOFP(MI); |
3335 | case G_FPTOUI: |
3336 | return lowerFPTOUI(MI); |
3337 | case G_FPTOSI: |
3338 | return lowerFPTOSI(MI); |
3339 | case G_FPTRUNC: |
3340 | return lowerFPTRUNC(MI); |
3341 | case G_FPOWI: |
3342 | return lowerFPOWI(MI); |
3343 | case G_SMIN: |
3344 | case G_SMAX: |
3345 | case G_UMIN: |
3346 | case G_UMAX: |
3347 | return lowerMinMax(MI); |
3348 | case G_FCOPYSIGN: |
3349 | return lowerFCopySign(MI); |
3350 | case G_FMINNUM: |
3351 | case G_FMAXNUM: |
3352 | return lowerFMinNumMaxNum(MI); |
3353 | case G_MERGE_VALUES: |
3354 | return lowerMergeValues(MI); |
3355 | case G_UNMERGE_VALUES: |
3356 | return lowerUnmergeValues(MI); |
3357 | case TargetOpcode::G_SEXT_INREG: { |
3358 | assert(MI.getOperand(2).isImm() && "Expected immediate")((void)0); |
3359 | int64_t SizeInBits = MI.getOperand(2).getImm(); |
3360 | |
3361 | Register DstReg = MI.getOperand(0).getReg(); |
3362 | Register SrcReg = MI.getOperand(1).getReg(); |
3363 | LLT DstTy = MRI.getType(DstReg); |
3364 | Register TmpRes = MRI.createGenericVirtualRegister(DstTy); |
3365 | |
3366 | auto MIBSz = MIRBuilder.buildConstant(DstTy, DstTy.getScalarSizeInBits() - SizeInBits); |
3367 | MIRBuilder.buildShl(TmpRes, SrcReg, MIBSz->getOperand(0)); |
3368 | MIRBuilder.buildAShr(DstReg, TmpRes, MIBSz->getOperand(0)); |
3369 | MI.eraseFromParent(); |
3370 | return Legalized; |
3371 | } |
3372 | case G_EXTRACT_VECTOR_ELT: |
3373 | case G_INSERT_VECTOR_ELT: |
3374 | return lowerExtractInsertVectorElt(MI); |
3375 | case G_SHUFFLE_VECTOR: |
3376 | return lowerShuffleVector(MI); |
3377 | case G_DYN_STACKALLOC: |
3378 | return lowerDynStackAlloc(MI); |
3379 | case G_EXTRACT: |
3380 | return lowerExtract(MI); |
3381 | case G_INSERT: |
3382 | return lowerInsert(MI); |
3383 | case G_BSWAP: |
3384 | return lowerBswap(MI); |
3385 | case G_BITREVERSE: |
3386 | return lowerBitreverse(MI); |
3387 | case G_READ_REGISTER: |
3388 | case G_WRITE_REGISTER: |
3389 | return lowerReadWriteRegister(MI); |
3390 | case G_UADDSAT: |
3391 | case G_USUBSAT: { |
3392 | // Try to make a reasonable guess about which lowering strategy to use. The |
3393 | // target can override this with custom lowering and calling the |
3394 | // implementation functions. |
3395 | LLT Ty = MRI.getType(MI.getOperand(0).getReg()); |
3396 | if (LI.isLegalOrCustom({G_UMIN, Ty})) |
3397 | return lowerAddSubSatToMinMax(MI); |
3398 | return lowerAddSubSatToAddoSubo(MI); |
3399 | } |
3400 | case G_SADDSAT: |
3401 | case G_SSUBSAT: { |
3402 | LLT Ty = MRI.getType(MI.getOperand(0).getReg()); |
3403 | |
3404 | // FIXME: It would probably make more sense to see if G_SADDO is preferred, |
3405 | // since it's a shorter expansion. However, we would need to figure out the |
3406 | // preferred boolean type for the carry out for the query. |
3407 | if (LI.isLegalOrCustom({G_SMIN, Ty}) && LI.isLegalOrCustom({G_SMAX, Ty})) |
3408 | return lowerAddSubSatToMinMax(MI); |
3409 | return lowerAddSubSatToAddoSubo(MI); |
3410 | } |
3411 | case G_SSHLSAT: |
3412 | case G_USHLSAT: |
3413 | return lowerShlSat(MI); |
3414 | case G_ABS: |
3415 | return lowerAbsToAddXor(MI); |
3416 | case G_SELECT: |
3417 | return lowerSelect(MI); |
3418 | case G_SDIVREM: |
3419 | case G_UDIVREM: |
3420 | return lowerDIVREM(MI); |
3421 | case G_FSHL: |
3422 | case G_FSHR: |
3423 | return lowerFunnelShift(MI); |
3424 | case G_ROTL: |
3425 | case G_ROTR: |
3426 | return lowerRotate(MI); |
3427 | } |
3428 | } |
3429 | |
3430 | Align LegalizerHelper::getStackTemporaryAlignment(LLT Ty, |
3431 | Align MinAlign) const { |
3432 | // FIXME: We're missing a way to go back from LLT to llvm::Type to query the |
3433 | // datalayout for the preferred alignment. Also there should be a target hook |
3434 | // for this to allow targets to reduce the alignment and ignore the |
3435 | // datalayout. e.g. AMDGPU should always use a 4-byte alignment, regardless of |
3436 | // the type. |
3437 | return std::max(Align(PowerOf2Ceil(Ty.getSizeInBytes())), MinAlign); |
3438 | } |
3439 | |
3440 | MachineInstrBuilder |
3441 | LegalizerHelper::createStackTemporary(TypeSize Bytes, Align Alignment, |
3442 | MachinePointerInfo &PtrInfo) { |
3443 | MachineFunction &MF = MIRBuilder.getMF(); |
3444 | const DataLayout &DL = MIRBuilder.getDataLayout(); |
3445 | int FrameIdx = MF.getFrameInfo().CreateStackObject(Bytes, Alignment, false); |
3446 | |
3447 | unsigned AddrSpace = DL.getAllocaAddrSpace(); |
3448 | LLT FramePtrTy = LLT::pointer(AddrSpace, DL.getPointerSizeInBits(AddrSpace)); |
3449 | |
3450 | PtrInfo = MachinePointerInfo::getFixedStack(MF, FrameIdx); |
3451 | return MIRBuilder.buildFrameIndex(FramePtrTy, FrameIdx); |
3452 | } |
3453 | |
3454 | static Register clampDynamicVectorIndex(MachineIRBuilder &B, Register IdxReg, |
3455 | LLT VecTy) { |
3456 | int64_t IdxVal; |
3457 | if (mi_match(IdxReg, *B.getMRI(), m_ICst(IdxVal))) |
3458 | return IdxReg; |
3459 | |
3460 | LLT IdxTy = B.getMRI()->getType(IdxReg); |
3461 | unsigned NElts = VecTy.getNumElements(); |
3462 | if (isPowerOf2_32(NElts)) { |
3463 | APInt Imm = APInt::getLowBitsSet(IdxTy.getSizeInBits(), Log2_32(NElts)); |
3464 | return B.buildAnd(IdxTy, IdxReg, B.buildConstant(IdxTy, Imm)).getReg(0); |
3465 | } |
3466 | |
3467 | return B.buildUMin(IdxTy, IdxReg, B.buildConstant(IdxTy, NElts - 1)) |
3468 | .getReg(0); |
3469 | } |
3470 | |
3471 | Register LegalizerHelper::getVectorElementPointer(Register VecPtr, LLT VecTy, |
3472 | Register Index) { |
3473 | LLT EltTy = VecTy.getElementType(); |
3474 | |
3475 | // Calculate the element offset and add it to the pointer. |
3476 | unsigned EltSize = EltTy.getSizeInBits() / 8; // FIXME: should be ABI size. |
3477 | assert(EltSize * 8 == EltTy.getSizeInBits() &&((void)0) |
3478 | "Converting bits to bytes lost precision")((void)0); |
3479 | |
3480 | Index = clampDynamicVectorIndex(MIRBuilder, Index, VecTy); |
3481 | |
3482 | LLT IdxTy = MRI.getType(Index); |
3483 | auto Mul = MIRBuilder.buildMul(IdxTy, Index, |
3484 | MIRBuilder.buildConstant(IdxTy, EltSize)); |
3485 | |
3486 | LLT PtrTy = MRI.getType(VecPtr); |
3487 | return MIRBuilder.buildPtrAdd(PtrTy, VecPtr, Mul).getReg(0); |
3488 | } |
3489 | |
3490 | LegalizerHelper::LegalizeResult LegalizerHelper::fewerElementsVectorImplicitDef( |
3491 | MachineInstr &MI, unsigned TypeIdx, LLT NarrowTy) { |
3492 | Register DstReg = MI.getOperand(0).getReg(); |
3493 | LLT DstTy = MRI.getType(DstReg); |
3494 | LLT LCMTy = getLCMType(DstTy, NarrowTy); |
3495 | |
3496 | unsigned NumParts = LCMTy.getSizeInBits() / NarrowTy.getSizeInBits(); |
3497 | |
3498 | auto NewUndef = MIRBuilder.buildUndef(NarrowTy); |
3499 | SmallVector<Register, 8> Parts(NumParts, NewUndef.getReg(0)); |
3500 | |
3501 | buildWidenedRemergeToDst(DstReg, LCMTy, Parts); |
3502 | MI.eraseFromParent(); |
3503 | return Legalized; |
3504 | } |
3505 | |
3506 | // Handle splitting vector operations which need to have the same number of |
3507 | // elements in each type index, but each type index may have a different element |
3508 | // type. |
3509 | // |
3510 | // e.g. <4 x s64> = G_SHL <4 x s64>, <4 x s32> -> |
3511 | // <2 x s64> = G_SHL <2 x s64>, <2 x s32> |
3512 | // <2 x s64> = G_SHL <2 x s64>, <2 x s32> |
3513 | // |
3514 | // Also handles some irregular breakdown cases, e.g. |
3515 | // e.g. <3 x s64> = G_SHL <3 x s64>, <3 x s32> -> |
3516 | // <2 x s64> = G_SHL <2 x s64>, <2 x s32> |
3517 | // s64 = G_SHL s64, s32 |
3518 | LegalizerHelper::LegalizeResult |
3519 | LegalizerHelper::fewerElementsVectorMultiEltType( |
3520 | MachineInstr &MI, unsigned TypeIdx, LLT NarrowTyArg) { |
3521 | if (TypeIdx != 0) |
3522 | return UnableToLegalize; |
3523 | |
3524 | const LLT NarrowTy0 = NarrowTyArg; |
3525 | const Register DstReg = MI.getOperand(0).getReg(); |
3526 | LLT DstTy = MRI.getType(DstReg); |
3527 | LLT LeftoverTy0; |
3528 | |
3529 | // All of the operands need to have the same number of elements, so if we can |
3530 | // determine a type breakdown for the result type, we can for all of the |
3531 | // source types. |
3532 | int NumParts = getNarrowTypeBreakDown(DstTy, NarrowTy0, LeftoverTy0).first; |
3533 | if (NumParts < 0) |
3534 | return UnableToLegalize; |
3535 | |
3536 | SmallVector<MachineInstrBuilder, 4> NewInsts; |
3537 | |
3538 | SmallVector<Register, 4> DstRegs, LeftoverDstRegs; |
3539 | SmallVector<Register, 4> PartRegs, LeftoverRegs; |
3540 | |
3541 | for (unsigned I = 1, E = MI.getNumOperands(); I != E; ++I) { |
3542 | Register SrcReg = MI.getOperand(I).getReg(); |
3543 | LLT SrcTyI = MRI.getType(SrcReg); |
3544 | const auto NewEC = NarrowTy0.isVector() ? NarrowTy0.getElementCount() |
3545 | : ElementCount::getFixed(1); |
3546 | LLT NarrowTyI = LLT::scalarOrVector(NewEC, SrcTyI.getScalarType()); |
3547 | LLT LeftoverTyI; |
3548 | |
3549 | // Split this operand into the requested typed registers, and any leftover |
3550 | // required to reproduce the original type. |
3551 | if (!extractParts(SrcReg, SrcTyI, NarrowTyI, LeftoverTyI, PartRegs, |
3552 | LeftoverRegs)) |
3553 | return UnableToLegalize; |
3554 | |
3555 | if (I == 1) { |
3556 | // For the first operand, create an instruction for each part and setup |
3557 | // the result. |
3558 | for (Register PartReg : PartRegs) { |
3559 | Register PartDstReg = MRI.createGenericVirtualRegister(NarrowTy0); |
3560 | NewInsts.push_back(MIRBuilder.buildInstrNoInsert(MI.getOpcode()) |
3561 | .addDef(PartDstReg) |
3562 | .addUse(PartReg)); |
3563 | DstRegs.push_back(PartDstReg); |
3564 | } |
3565 | |
3566 | for (Register LeftoverReg : LeftoverRegs) { |
3567 | Register PartDstReg = MRI.createGenericVirtualRegister(LeftoverTy0); |
3568 | NewInsts.push_back(MIRBuilder.buildInstrNoInsert(MI.getOpcode()) |
3569 | .addDef(PartDstReg) |
3570 | .addUse(LeftoverReg)); |
3571 | LeftoverDstRegs.push_back(PartDstReg); |
3572 | } |
3573 | } else { |
3574 | assert(NewInsts.size() == PartRegs.size() + LeftoverRegs.size())((void)0); |
3575 | |
3576 | // Add the newly created operand splits to the existing instructions. The |
3577 | // odd-sized pieces are ordered after the requested NarrowTyArg sized |
3578 | // pieces. |
3579 | unsigned InstCount = 0; |
3580 | for (unsigned J = 0, JE = PartRegs.size(); J != JE; ++J) |
3581 | NewInsts[InstCount++].addUse(PartRegs[J]); |
3582 | for (unsigned J = 0, JE = LeftoverRegs.size(); J != JE; ++J) |
3583 | NewInsts[InstCount++].addUse(LeftoverRegs[J]); |
3584 | } |
3585 | |
3586 | PartRegs.clear(); |
3587 | LeftoverRegs.clear(); |
3588 | } |
3589 | |
3590 | // Insert the newly built operations and rebuild the result register. |
3591 | for (auto &MIB : NewInsts) |
3592 | MIRBuilder.insertInstr(MIB); |
3593 | |
3594 | insertParts(DstReg, DstTy, NarrowTy0, DstRegs, LeftoverTy0, LeftoverDstRegs); |
3595 | |
3596 | MI.eraseFromParent(); |
3597 | return Legalized; |
3598 | } |
3599 | |
3600 | LegalizerHelper::LegalizeResult |
3601 | LegalizerHelper::fewerElementsVectorCasts(MachineInstr &MI, unsigned TypeIdx, |
3602 | LLT NarrowTy) { |
3603 | if (TypeIdx != 0) |
3604 | return UnableToLegalize; |
3605 | |
3606 | Register DstReg = MI.getOperand(0).getReg(); |
3607 | Register SrcReg = MI.getOperand(1).getReg(); |
3608 | LLT DstTy = MRI.getType(DstReg); |
3609 | LLT SrcTy = MRI.getType(SrcReg); |
3610 | |
3611 | LLT NarrowTy0 = NarrowTy; |
3612 | LLT NarrowTy1; |
3613 | unsigned NumParts; |
3614 | |
3615 | if (NarrowTy.isVector()) { |
3616 | // Uneven breakdown not handled. |
3617 | NumParts = DstTy.getNumElements() / NarrowTy.getNumElements(); |
3618 | if (NumParts * NarrowTy.getNumElements() != DstTy.getNumElements()) |
3619 | return UnableToLegalize; |
3620 | |
3621 | NarrowTy1 = LLT::vector(NarrowTy.getElementCount(), SrcTy.getElementType()); |
3622 | } else { |
3623 | NumParts = DstTy.getNumElements(); |
3624 | NarrowTy1 = SrcTy.getElementType(); |
3625 | } |
3626 | |
3627 | SmallVector<Register, 4> SrcRegs, DstRegs; |
3628 | extractParts(SrcReg, NarrowTy1, NumParts, SrcRegs); |
3629 | |
3630 | for (unsigned I = 0; I < NumParts; ++I) { |
3631 | Register DstReg = MRI.createGenericVirtualRegister(NarrowTy0); |
3632 | MachineInstr *NewInst = |
3633 | MIRBuilder.buildInstr(MI.getOpcode(), {DstReg}, {SrcRegs[I]}); |
3634 | |
3635 | NewInst->setFlags(MI.getFlags()); |
3636 | DstRegs.push_back(DstReg); |
3637 | } |
3638 | |
3639 | if (NarrowTy.isVector()) |
3640 | MIRBuilder.buildConcatVectors(DstReg, DstRegs); |
3641 | else |
3642 | MIRBuilder.buildBuildVector(DstReg, DstRegs); |
3643 | |
3644 | MI.eraseFromParent(); |
3645 | return Legalized; |
3646 | } |
3647 | |
3648 | LegalizerHelper::LegalizeResult |
3649 | LegalizerHelper::fewerElementsVectorCmp(MachineInstr &MI, unsigned TypeIdx, |
3650 | LLT NarrowTy) { |
3651 | Register DstReg = MI.getOperand(0).getReg(); |
3652 | Register Src0Reg = MI.getOperand(2).getReg(); |
3653 | LLT DstTy = MRI.getType(DstReg); |
3654 | LLT SrcTy = MRI.getType(Src0Reg); |
3655 | |
3656 | unsigned NumParts; |
3657 | LLT NarrowTy0, NarrowTy1; |
3658 | |
3659 | if (TypeIdx == 0) { |
3660 | unsigned NewElts = NarrowTy.isVector() ? NarrowTy.getNumElements() : 1; |
3661 | unsigned OldElts = DstTy.getNumElements(); |
3662 | |
3663 | NarrowTy0 = NarrowTy; |
3664 | NumParts = NarrowTy.isVector() ? (OldElts / NewElts) : DstTy.getNumElements(); |
3665 | NarrowTy1 = NarrowTy.isVector() ? LLT::vector(NarrowTy.getElementCount(), |
3666 | SrcTy.getScalarSizeInBits()) |
3667 | : SrcTy.getElementType(); |
3668 | |
3669 | } else { |
3670 | unsigned NewElts = NarrowTy.isVector() ? NarrowTy.getNumElements() : 1; |
3671 | unsigned OldElts = SrcTy.getNumElements(); |
3672 | |
3673 | NumParts = NarrowTy.isVector() ? (OldElts / NewElts) : |
3674 | NarrowTy.getNumElements(); |
3675 | NarrowTy0 = |
3676 | LLT::vector(NarrowTy.getElementCount(), DstTy.getScalarSizeInBits()); |
3677 | NarrowTy1 = NarrowTy; |
3678 | } |
3679 | |
3680 | // FIXME: Don't know how to handle the situation where the small vectors |
3681 | // aren't all the same size yet. |
3682 | if (NarrowTy1.isVector() && |
3683 | NarrowTy1.getNumElements() * NumParts != DstTy.getNumElements()) |
3684 | return UnableToLegalize; |
3685 | |
3686 | CmpInst::Predicate Pred |
3687 | = static_cast<CmpInst::Predicate>(MI.getOperand(1).getPredicate()); |
3688 | |
3689 | SmallVector<Register, 2> Src1Regs, Src2Regs, DstRegs; |
3690 | extractParts(MI.getOperand(2).getReg(), NarrowTy1, NumParts, Src1Regs); |
3691 | extractParts(MI.getOperand(3).getReg(), NarrowTy1, NumParts, Src2Regs); |
3692 | |
3693 | for (unsigned I = 0; I < NumParts; ++I) { |
3694 | Register DstReg = MRI.createGenericVirtualRegister(NarrowTy0); |
3695 | DstRegs.push_back(DstReg); |
3696 | |
3697 | if (MI.getOpcode() == TargetOpcode::G_ICMP) |
3698 | MIRBuilder.buildICmp(Pred, DstReg, Src1Regs[I], Src2Regs[I]); |
3699 | else { |
3700 | MachineInstr *NewCmp |
3701 | = MIRBuilder.buildFCmp(Pred, DstReg, Src1Regs[I], Src2Regs[I]); |
3702 | NewCmp->setFlags(MI.getFlags()); |
3703 | } |
3704 | } |
3705 | |
3706 | if (NarrowTy1.isVector()) |
3707 | MIRBuilder.buildConcatVectors(DstReg, DstRegs); |
3708 | else |
3709 | MIRBuilder.buildBuildVector(DstReg, DstRegs); |
3710 | |
3711 | MI.eraseFromParent(); |
3712 | return Legalized; |
3713 | } |
3714 | |
3715 | LegalizerHelper::LegalizeResult |
3716 | LegalizerHelper::fewerElementsVectorSelect(MachineInstr &MI, unsigned TypeIdx, |
3717 | LLT NarrowTy) { |
3718 | Register DstReg = MI.getOperand(0).getReg(); |
3719 | Register CondReg = MI.getOperand(1).getReg(); |
3720 | |
3721 | unsigned NumParts = 0; |
3722 | LLT NarrowTy0, NarrowTy1; |
3723 | |
3724 | LLT DstTy = MRI.getType(DstReg); |
3725 | LLT CondTy = MRI.getType(CondReg); |
3726 | unsigned Size = DstTy.getSizeInBits(); |
3727 | |
3728 | assert(TypeIdx == 0 || CondTy.isVector())((void)0); |
3729 | |
3730 | if (TypeIdx == 0) { |
3731 | NarrowTy0 = NarrowTy; |
3732 | NarrowTy1 = CondTy; |
3733 | |
3734 | unsigned NarrowSize = NarrowTy0.getSizeInBits(); |
3735 | // FIXME: Don't know how to handle the situation where the small vectors |
3736 | // aren't all the same size yet. |
3737 | if (Size % NarrowSize != 0) |
3738 | return UnableToLegalize; |
3739 | |
3740 | NumParts = Size / NarrowSize; |
3741 | |
3742 | // Need to break down the condition type |
3743 | if (CondTy.isVector()) { |
3744 | if (CondTy.getNumElements() == NumParts) |
3745 | NarrowTy1 = CondTy.getElementType(); |
3746 | else |
3747 | NarrowTy1 = |
3748 | LLT::vector(CondTy.getElementCount().divideCoefficientBy(NumParts), |
3749 | CondTy.getScalarSizeInBits()); |
3750 | } |
3751 | } else { |
3752 | NumParts = CondTy.getNumElements(); |
3753 | if (NarrowTy.isVector()) { |
3754 | // TODO: Handle uneven breakdown. |
3755 | if (NumParts * NarrowTy.getNumElements() != CondTy.getNumElements()) |
3756 | return UnableToLegalize; |
3757 | |
3758 | return UnableToLegalize; |
3759 | } else { |
3760 | NarrowTy0 = DstTy.getElementType(); |
3761 | NarrowTy1 = NarrowTy; |
3762 | } |
3763 | } |
3764 | |
3765 | SmallVector<Register, 2> DstRegs, Src0Regs, Src1Regs, Src2Regs; |
3766 | if (CondTy.isVector()) |
3767 | extractParts(MI.getOperand(1).getReg(), NarrowTy1, NumParts, Src0Regs); |
3768 | |
3769 | extractParts(MI.getOperand(2).getReg(), NarrowTy0, NumParts, Src1Regs); |
3770 | extractParts(MI.getOperand(3).getReg(), NarrowTy0, NumParts, Src2Regs); |
3771 | |
3772 | for (unsigned i = 0; i < NumParts; ++i) { |
3773 | Register DstReg = MRI.createGenericVirtualRegister(NarrowTy0); |
3774 | MIRBuilder.buildSelect(DstReg, CondTy.isVector() ? Src0Regs[i] : CondReg, |
3775 | Src1Regs[i], Src2Regs[i]); |
3776 | DstRegs.push_back(DstReg); |
3777 | } |
3778 | |
3779 | if (NarrowTy0.isVector()) |
3780 | MIRBuilder.buildConcatVectors(DstReg, DstRegs); |
3781 | else |
3782 | MIRBuilder.buildBuildVector(DstReg, DstRegs); |
3783 | |
3784 | MI.eraseFromParent(); |
3785 | return Legalized; |
3786 | } |
3787 | |
3788 | LegalizerHelper::LegalizeResult |
3789 | LegalizerHelper::fewerElementsVectorPhi(MachineInstr &MI, unsigned TypeIdx, |
3790 | LLT NarrowTy) { |
3791 | const Register DstReg = MI.getOperand(0).getReg(); |
3792 | LLT PhiTy = MRI.getType(DstReg); |
3793 | LLT LeftoverTy; |
3794 | |
3795 | // All of the operands need to have the same number of elements, so if we can |
3796 | // determine a type breakdown for the result type, we can for all of the |
3797 | // source types. |
3798 | int NumParts, NumLeftover; |
3799 | std::tie(NumParts, NumLeftover) |
3800 | = getNarrowTypeBreakDown(PhiTy, NarrowTy, LeftoverTy); |
3801 | if (NumParts < 0) |
3802 | return UnableToLegalize; |
3803 | |
3804 | SmallVector<Register, 4> DstRegs, LeftoverDstRegs; |
3805 | SmallVector<MachineInstrBuilder, 4> NewInsts; |
3806 | |
3807 | const int TotalNumParts = NumParts + NumLeftover; |
3808 | |
3809 | // Insert the new phis in the result block first. |
3810 | for (int I = 0; I != TotalNumParts; ++I) { |
3811 | LLT Ty = I < NumParts ? NarrowTy : LeftoverTy; |
3812 | Register PartDstReg = MRI.createGenericVirtualRegister(Ty); |
3813 | NewInsts.push_back(MIRBuilder.buildInstr(TargetOpcode::G_PHI) |
3814 | .addDef(PartDstReg)); |
3815 | if (I < NumParts) |
3816 | DstRegs.push_back(PartDstReg); |
3817 | else |
3818 | LeftoverDstRegs.push_back(PartDstReg); |
3819 | } |
3820 | |
3821 | MachineBasicBlock *MBB = MI.getParent(); |
3822 | MIRBuilder.setInsertPt(*MBB, MBB->getFirstNonPHI()); |
3823 | insertParts(DstReg, PhiTy, NarrowTy, DstRegs, LeftoverTy, LeftoverDstRegs); |
3824 | |
3825 | SmallVector<Register, 4> PartRegs, LeftoverRegs; |
3826 | |
3827 | // Insert code to extract the incoming values in each predecessor block. |
3828 | for (unsigned I = 1, E = MI.getNumOperands(); I != E; I += 2) { |
3829 | PartRegs.clear(); |
3830 | LeftoverRegs.clear(); |
3831 | |
3832 | Register SrcReg = MI.getOperand(I).getReg(); |
3833 | MachineBasicBlock &OpMBB = *MI.getOperand(I + 1).getMBB(); |
3834 | MIRBuilder.setInsertPt(OpMBB, OpMBB.getFirstTerminator()); |
3835 | |
3836 | LLT Unused; |
3837 | if (!extractParts(SrcReg, PhiTy, NarrowTy, Unused, PartRegs, |
3838 | LeftoverRegs)) |
3839 | return UnableToLegalize; |
3840 | |
3841 | // Add the newly created operand splits to the existing instructions. The |
3842 | // odd-sized pieces are ordered after the requested NarrowTyArg sized |
3843 | // pieces. |
3844 | for (int J = 0; J != TotalNumParts; ++J) { |
3845 | MachineInstrBuilder MIB = NewInsts[J]; |
3846 | MIB.addUse(J < NumParts ? PartRegs[J] : LeftoverRegs[J - NumParts]); |
3847 | MIB.addMBB(&OpMBB); |
3848 | } |
3849 | } |
3850 | |
3851 | MI.eraseFromParent(); |
3852 | return Legalized; |
3853 | } |
3854 | |
3855 | LegalizerHelper::LegalizeResult |
3856 | LegalizerHelper::fewerElementsVectorUnmergeValues(MachineInstr &MI, |
3857 | unsigned TypeIdx, |
3858 | LLT NarrowTy) { |
3859 | if (TypeIdx != 1) |
3860 | return UnableToLegalize; |
3861 | |
3862 | const int NumDst = MI.getNumOperands() - 1; |
3863 | const Register SrcReg = MI.getOperand(NumDst).getReg(); |
3864 | LLT SrcTy = MRI.getType(SrcReg); |
3865 | |
3866 | LLT DstTy = MRI.getType(MI.getOperand(0).getReg()); |
3867 | |
3868 | // TODO: Create sequence of extracts. |
3869 | if (DstTy == NarrowTy) |
3870 | return UnableToLegalize; |
3871 | |
3872 | LLT GCDTy = getGCDType(SrcTy, NarrowTy); |
3873 | if (DstTy == GCDTy) { |
3874 | // This would just be a copy of the same unmerge. |
3875 | // TODO: Create extracts, pad with undef and create intermediate merges. |
3876 | return UnableToLegalize; |
3877 | } |
3878 | |
3879 | auto Unmerge = MIRBuilder.buildUnmerge(GCDTy, SrcReg); |
3880 | const int NumUnmerge = Unmerge->getNumOperands() - 1; |
3881 | const int PartsPerUnmerge = NumDst / NumUnmerge; |
3882 | |
3883 | for (int I = 0; I != NumUnmerge; ++I) { |
3884 | auto MIB = MIRBuilder.buildInstr(TargetOpcode::G_UNMERGE_VALUES); |
3885 | |
3886 | for (int J = 0; J != PartsPerUnmerge; ++J) |
3887 | MIB.addDef(MI.getOperand(I * PartsPerUnmerge + J).getReg()); |
3888 | MIB.addUse(Unmerge.getReg(I)); |
3889 | } |
3890 | |
3891 | MI.eraseFromParent(); |
3892 | return Legalized; |
3893 | } |
3894 | |
3895 | LegalizerHelper::LegalizeResult |
3896 | LegalizerHelper::fewerElementsVectorMulo(MachineInstr &MI, unsigned TypeIdx, |
3897 | LLT NarrowTy) { |
3898 | Register Result = MI.getOperand(0).getReg(); |
3899 | Register Overflow = MI.getOperand(1).getReg(); |
3900 | Register LHS = MI.getOperand(2).getReg(); |
3901 | Register RHS = MI.getOperand(3).getReg(); |
3902 | |
3903 | LLT SrcTy = MRI.getType(LHS); |
3904 | if (!SrcTy.isVector()) |
3905 | return UnableToLegalize; |
3906 | |
3907 | LLT ElementType = SrcTy.getElementType(); |
3908 | LLT OverflowElementTy = MRI.getType(Overflow).getElementType(); |
3909 | const ElementCount NumResult = SrcTy.getElementCount(); |
3910 | LLT GCDTy = getGCDType(SrcTy, NarrowTy); |
3911 | |
3912 | // Unmerge the operands to smaller parts of GCD type. |
3913 | auto UnmergeLHS = MIRBuilder.buildUnmerge(GCDTy, LHS); |
3914 | auto UnmergeRHS = MIRBuilder.buildUnmerge(GCDTy, RHS); |
3915 | |
3916 | const int NumOps = UnmergeLHS->getNumOperands() - 1; |
3917 | const ElementCount PartsPerUnmerge = NumResult.divideCoefficientBy(NumOps); |
3918 | LLT OverflowTy = LLT::scalarOrVector(PartsPerUnmerge, OverflowElementTy); |
3919 | LLT ResultTy = LLT::scalarOrVector(PartsPerUnmerge, ElementType); |
3920 | |
3921 | // Perform the operation over unmerged parts. |
3922 | SmallVector<Register, 8> ResultParts; |
3923 | SmallVector<Register, 8> OverflowParts; |
3924 | for (int I = 0; I != NumOps; ++I) { |
3925 | Register Operand1 = UnmergeLHS->getOperand(I).getReg(); |
3926 | Register Operand2 = UnmergeRHS->getOperand(I).getReg(); |
3927 | auto PartMul = MIRBuilder.buildInstr(MI.getOpcode(), {ResultTy, OverflowTy}, |
3928 | {Operand1, Operand2}); |
3929 | ResultParts.push_back(PartMul->getOperand(0).getReg()); |
3930 | OverflowParts.push_back(PartMul->getOperand(1).getReg()); |
3931 | } |
3932 | |
3933 | LLT ResultLCMTy = buildLCMMergePieces(SrcTy, NarrowTy, GCDTy, ResultParts); |
3934 | LLT OverflowLCMTy = |
3935 | LLT::scalarOrVector(ResultLCMTy.getElementCount(), OverflowElementTy); |
3936 | |
3937 | // Recombine the pieces to the original result and overflow registers. |
3938 | buildWidenedRemergeToDst(Result, ResultLCMTy, ResultParts); |
3939 | buildWidenedRemergeToDst(Overflow, OverflowLCMTy, OverflowParts); |
3940 | MI.eraseFromParent(); |
3941 | return Legalized; |
3942 | } |
3943 | |
3944 | // Handle FewerElementsVector a G_BUILD_VECTOR or G_CONCAT_VECTORS that produces |
3945 | // a vector |
3946 | // |
3947 | // Create a G_BUILD_VECTOR or G_CONCAT_VECTORS of NarrowTy pieces, padding with |
3948 | // undef as necessary. |
3949 | // |
3950 | // %3:_(<3 x s16>) = G_BUILD_VECTOR %0, %1, %2 |
3951 | // -> <2 x s16> |
3952 | // |
3953 | // %4:_(s16) = G_IMPLICIT_DEF |
3954 | // %5:_(<2 x s16>) = G_BUILD_VECTOR %0, %1 |
3955 | // %6:_(<2 x s16>) = G_BUILD_VECTOR %2, %4 |
3956 | // %7:_(<2 x s16>) = G_IMPLICIT_DEF |
3957 | // %8:_(<6 x s16>) = G_CONCAT_VECTORS %5, %6, %7 |
3958 | // %3:_(<3 x s16>), %8:_(<3 x s16>) = G_UNMERGE_VALUES %8 |
3959 | LegalizerHelper::LegalizeResult |
3960 | LegalizerHelper::fewerElementsVectorMerge(MachineInstr &MI, unsigned TypeIdx, |
3961 | LLT NarrowTy) { |
3962 | Register DstReg = MI.getOperand(0).getReg(); |
3963 | LLT DstTy = MRI.getType(DstReg); |
3964 | LLT SrcTy = MRI.getType(MI.getOperand(1).getReg()); |
3965 | LLT GCDTy = getGCDType(getGCDType(SrcTy, NarrowTy), DstTy); |
3966 | |
3967 | // Break into a common type |
3968 | SmallVector<Register, 16> Parts; |
3969 | for (unsigned I = 1, E = MI.getNumOperands(); I != E; ++I) |
3970 | extractGCDType(Parts, GCDTy, MI.getOperand(I).getReg()); |
3971 | |
3972 | // Build the requested new merge, padding with undef. |
3973 | LLT LCMTy = buildLCMMergePieces(DstTy, NarrowTy, GCDTy, Parts, |
3974 | TargetOpcode::G_ANYEXT); |
3975 | |
3976 | // Pack into the original result register. |
3977 | buildWidenedRemergeToDst(DstReg, LCMTy, Parts); |
3978 | |
3979 | MI.eraseFromParent(); |
3980 | return Legalized; |
3981 | } |
3982 | |
3983 | LegalizerHelper::LegalizeResult |
3984 | LegalizerHelper::fewerElementsVectorExtractInsertVectorElt(MachineInstr &MI, |
3985 | unsigned TypeIdx, |
3986 | LLT NarrowVecTy) { |
3987 | Register DstReg = MI.getOperand(0).getReg(); |
3988 | Register SrcVec = MI.getOperand(1).getReg(); |
3989 | Register InsertVal; |
3990 | bool IsInsert = MI.getOpcode() == TargetOpcode::G_INSERT_VECTOR_ELT; |
3991 | |
3992 | assert((IsInsert ? TypeIdx == 0 : TypeIdx == 1) && "not a vector type index")((void)0); |
3993 | if (IsInsert) |
3994 | InsertVal = MI.getOperand(2).getReg(); |
3995 | |
3996 | Register Idx = MI.getOperand(MI.getNumOperands() - 1).getReg(); |
3997 | |
3998 | // TODO: Handle total scalarization case. |
3999 | if (!NarrowVecTy.isVector()) |
4000 | return UnableToLegalize; |
4001 | |
4002 | LLT VecTy = MRI.getType(SrcVec); |
4003 | |
4004 | // If the index is a constant, we can really break this down as you would |
4005 | // expect, and index into the target size pieces. |
4006 | int64_t IdxVal; |
4007 | auto MaybeCst = |
4008 | getConstantVRegValWithLookThrough(Idx, MRI, /*LookThroughInstrs*/ true, |
4009 | /*HandleFConstants*/ false); |
4010 | if (MaybeCst) { |
4011 | IdxVal = MaybeCst->Value.getSExtValue(); |
4012 | // Avoid out of bounds indexing the pieces. |
4013 | if (IdxVal >= VecTy.getNumElements()) { |
4014 | MIRBuilder.buildUndef(DstReg); |
4015 | MI.eraseFromParent(); |
4016 | return Legalized; |
4017 | } |
4018 | |
4019 | SmallVector<Register, 8> VecParts; |
4020 | LLT GCDTy = extractGCDType(VecParts, VecTy, NarrowVecTy, SrcVec); |
4021 | |
4022 | // Build a sequence of NarrowTy pieces in VecParts for this operand. |
4023 | LLT LCMTy = buildLCMMergePieces(VecTy, NarrowVecTy, GCDTy, VecParts, |
4024 | TargetOpcode::G_ANYEXT); |
4025 | |
4026 | unsigned NewNumElts = NarrowVecTy.getNumElements(); |
4027 | |
4028 | LLT IdxTy = MRI.getType(Idx); |
4029 | int64_t PartIdx = IdxVal / NewNumElts; |
4030 | auto NewIdx = |
4031 | MIRBuilder.buildConstant(IdxTy, IdxVal - NewNumElts * PartIdx); |
4032 | |
4033 | if (IsInsert) { |
4034 | LLT PartTy = MRI.getType(VecParts[PartIdx]); |
4035 | |
4036 | // Use the adjusted index to insert into one of the subvectors. |
4037 | auto InsertPart = MIRBuilder.buildInsertVectorElement( |
4038 | PartTy, VecParts[PartIdx], InsertVal, NewIdx); |
4039 | VecParts[PartIdx] = InsertPart.getReg(0); |
4040 | |
4041 | // Recombine the inserted subvector with the others to reform the result |
4042 | // vector. |
4043 | buildWidenedRemergeToDst(DstReg, LCMTy, VecParts); |
4044 | } else { |
4045 | MIRBuilder.buildExtractVectorElement(DstReg, VecParts[PartIdx], NewIdx); |
4046 | } |
4047 | |
4048 | MI.eraseFromParent(); |
4049 | return Legalized; |
4050 | } |
4051 | |
4052 | // With a variable index, we can't perform the operation in a smaller type, so |
4053 | // we're forced to expand this. |
4054 | // |
4055 | // TODO: We could emit a chain of compare/select to figure out which piece to |
4056 | // index. |
4057 | return lowerExtractInsertVectorElt(MI); |
4058 | } |
4059 | |
4060 | LegalizerHelper::LegalizeResult |
4061 | LegalizerHelper::reduceLoadStoreWidth(GLoadStore &LdStMI, unsigned TypeIdx, |
4062 | LLT NarrowTy) { |
4063 | // FIXME: Don't know how to handle secondary types yet. |
4064 | if (TypeIdx != 0) |
4065 | return UnableToLegalize; |
4066 | |
4067 | // This implementation doesn't work for atomics. Give up instead of doing |
4068 | // something invalid. |
4069 | if (LdStMI.isAtomic()) |
4070 | return UnableToLegalize; |
4071 | |
4072 | bool IsLoad = isa<GLoad>(LdStMI); |
4073 | Register ValReg = LdStMI.getReg(0); |
4074 | Register AddrReg = LdStMI.getPointerReg(); |
4075 | LLT ValTy = MRI.getType(ValReg); |
4076 | |
4077 | // FIXME: Do we need a distinct NarrowMemory legalize action? |
4078 | if (ValTy.getSizeInBits() != 8 * LdStMI.getMemSize()) { |
4079 | LLVM_DEBUG(dbgs() << "Can't narrow extload/truncstore\n")do { } while (false); |
4080 | return UnableToLegalize; |
4081 | } |
4082 | |
4083 | int NumParts = -1; |
4084 | int NumLeftover = -1; |
4085 | LLT LeftoverTy; |
4086 | SmallVector<Register, 8> NarrowRegs, NarrowLeftoverRegs; |
4087 | if (IsLoad) { |
4088 | std::tie(NumParts, NumLeftover) = getNarrowTypeBreakDown(ValTy, NarrowTy, LeftoverTy); |
4089 | } else { |
4090 | if (extractParts(ValReg, ValTy, NarrowTy, LeftoverTy, NarrowRegs, |
4091 | NarrowLeftoverRegs)) { |
4092 | NumParts = NarrowRegs.size(); |
4093 | NumLeftover = NarrowLeftoverRegs.size(); |
Value stored to 'NumLeftover' is never read | |
4094 | } |
4095 | } |
4096 | |
4097 | if (NumParts == -1) |
4098 | return UnableToLegalize; |
4099 | |
4100 | LLT PtrTy = MRI.getType(AddrReg); |
4101 | const LLT OffsetTy = LLT::scalar(PtrTy.getSizeInBits()); |
4102 | |
4103 | unsigned TotalSize = ValTy.getSizeInBits(); |
4104 | |
4105 | // Split the load/store into PartTy sized pieces starting at Offset. If this |
4106 | // is a load, return the new registers in ValRegs. For a store, each elements |
4107 | // of ValRegs should be PartTy. Returns the next offset that needs to be |
4108 | // handled. |
4109 | auto MMO = LdStMI.getMMO(); |
4110 | auto splitTypePieces = [=](LLT PartTy, SmallVectorImpl<Register> &ValRegs, |
4111 | unsigned Offset) -> unsigned { |
4112 | MachineFunction &MF = MIRBuilder.getMF(); |
4113 | unsigned PartSize = PartTy.getSizeInBits(); |
4114 | for (unsigned Idx = 0, E = NumParts; Idx != E && Offset < TotalSize; |
4115 | Offset += PartSize, ++Idx) { |
4116 | unsigned ByteOffset = Offset / 8; |
4117 | Register NewAddrReg; |
4118 | |
4119 | MIRBuilder.materializePtrAdd(NewAddrReg, AddrReg, OffsetTy, ByteOffset); |
4120 | |
4121 | MachineMemOperand *NewMMO = |
4122 | MF.getMachineMemOperand(&MMO, ByteOffset, PartTy); |
4123 | |
4124 | if (IsLoad) { |
4125 | Register Dst = MRI.createGenericVirtualRegister(PartTy); |
4126 | ValRegs.push_back(Dst); |
4127 | MIRBuilder.buildLoad(Dst, NewAddrReg, *NewMMO); |
4128 | } else { |
4129 | MIRBuilder.buildStore(ValRegs[Idx], NewAddrReg, *NewMMO); |
4130 | } |
4131 | } |
4132 | |
4133 | return Offset; |
4134 | }; |
4135 | |
4136 | unsigned HandledOffset = splitTypePieces(NarrowTy, NarrowRegs, 0); |
4137 | |
4138 | // Handle the rest of the register if this isn't an even type breakdown. |
4139 | if (LeftoverTy.isValid()) |
4140 | splitTypePieces(LeftoverTy, NarrowLeftoverRegs, HandledOffset); |
4141 | |
4142 | if (IsLoad) { |
4143 | insertParts(ValReg, ValTy, NarrowTy, NarrowRegs, |
4144 | LeftoverTy, NarrowLeftoverRegs); |
4145 | } |
4146 | |
4147 | LdStMI.eraseFromParent(); |
4148 | return Legalized; |
4149 | } |
4150 | |
4151 | LegalizerHelper::LegalizeResult |
4152 | LegalizerHelper::reduceOperationWidth(MachineInstr &MI, unsigned int TypeIdx, |
4153 | LLT NarrowTy) { |
4154 | assert(TypeIdx == 0 && "only one type index expected")((void)0); |
4155 | |
4156 | const unsigned Opc = MI.getOpcode(); |
4157 | const int NumDefOps = MI.getNumExplicitDefs(); |
4158 | const int NumSrcOps = MI.getNumOperands() - NumDefOps; |
4159 | const unsigned Flags = MI.getFlags(); |
4160 | const unsigned NarrowSize = NarrowTy.getSizeInBits(); |
4161 | const LLT NarrowScalarTy = LLT::scalar(NarrowSize); |
4162 | |
4163 | assert(MI.getNumOperands() <= 4 && "expected instruction with either 1 "((void)0) |
4164 | "result and 1-3 sources or 2 results and "((void)0) |
4165 | "1-2 sources")((void)0); |
4166 | |
4167 | SmallVector<Register, 2> DstRegs; |
4168 | for (int I = 0; I < NumDefOps; ++I) |
4169 | DstRegs.push_back(MI.getOperand(I).getReg()); |
4170 | |
4171 | // First of all check whether we are narrowing (changing the element type) |
4172 | // or reducing the vector elements |
4173 | const LLT DstTy = MRI.getType(DstRegs[0]); |
4174 | const bool IsNarrow = NarrowTy.getScalarType() != DstTy.getScalarType(); |
4175 | |
4176 | SmallVector<Register, 8> ExtractedRegs[3]; |
4177 | SmallVector<Register, 8> Parts; |
4178 | |
4179 | // Break down all the sources into NarrowTy pieces we can operate on. This may |
4180 | // involve creating merges to a wider type, padded with undef. |
4181 | for (int I = 0; I != NumSrcOps; ++I) { |
4182 | Register SrcReg = MI.getOperand(I + NumDefOps).getReg(); |
4183 | LLT SrcTy = MRI.getType(SrcReg); |
4184 | |
4185 | // The type to narrow SrcReg to. For narrowing, this is a smaller scalar. |
4186 | // For fewerElements, this is a smaller vector with the same element type. |
4187 | LLT OpNarrowTy; |
4188 | if (IsNarrow) { |
4189 | OpNarrowTy = NarrowScalarTy; |
4190 | |
4191 | // In case of narrowing, we need to cast vectors to scalars for this to |
4192 | // work properly |
4193 | // FIXME: Can we do without the bitcast here if we're narrowing? |
4194 | if (SrcTy.isVector()) { |
4195 | SrcTy = LLT::scalar(SrcTy.getSizeInBits()); |
4196 | SrcReg = MIRBuilder.buildBitcast(SrcTy, SrcReg).getReg(0); |
4197 | } |
4198 | } else { |
4199 | auto NarrowEC = NarrowTy.isVector() ? NarrowTy.getElementCount() |
4200 | : ElementCount::getFixed(1); |
4201 | OpNarrowTy = LLT::scalarOrVector(NarrowEC, SrcTy.getScalarType()); |
4202 | } |
4203 | |
4204 | LLT GCDTy = extractGCDType(ExtractedRegs[I], SrcTy, OpNarrowTy, SrcReg); |
4205 | |
4206 | // Build a sequence of NarrowTy pieces in ExtractedRegs for this operand. |
4207 | buildLCMMergePieces(SrcTy, OpNarrowTy, GCDTy, ExtractedRegs[I], |
4208 | TargetOpcode::G_ANYEXT); |
4209 | } |
4210 | |
4211 | SmallVector<Register, 8> ResultRegs[2]; |
4212 | |
4213 | // Input operands for each sub-instruction. |
4214 | SmallVector<SrcOp, 4> InputRegs(NumSrcOps, Register()); |
4215 | |
4216 | int NumParts = ExtractedRegs[0].size(); |
4217 | const unsigned DstSize = DstTy.getSizeInBits(); |
4218 | const LLT DstScalarTy = LLT::scalar(DstSize); |
4219 | |
4220 | // Narrowing needs to use scalar types |
4221 | LLT DstLCMTy, NarrowDstTy; |
4222 | if (IsNarrow) { |
4223 | DstLCMTy = getLCMType(DstScalarTy, NarrowScalarTy); |
4224 | NarrowDstTy = NarrowScalarTy; |
4225 | } else { |
4226 | DstLCMTy = getLCMType(DstTy, NarrowTy); |
4227 | NarrowDstTy = NarrowTy; |
4228 | } |
4229 | |
4230 | // We widened the source registers to satisfy merge/unmerge size |
4231 | // constraints. We'll have some extra fully undef parts. |
4232 | const int NumRealParts = (DstSize + NarrowSize - 1) / NarrowSize; |
4233 | |
4234 | for (int I = 0; I != NumRealParts; ++I) { |
4235 | // Emit this instruction on each of the split pieces. |
4236 | for (int J = 0; J != NumSrcOps; ++J) |
4237 | InputRegs[J] = ExtractedRegs[J][I]; |
4238 | |
4239 | MachineInstrBuilder Inst; |
4240 | if (NumDefOps == 1) |
4241 | Inst = MIRBuilder.buildInstr(Opc, {NarrowDstTy}, InputRegs, Flags); |
4242 | else |
4243 | Inst = MIRBuilder.buildInstr(Opc, {NarrowDstTy, NarrowDstTy}, InputRegs, |
4244 | Flags); |
4245 | |
4246 | for (int J = 0; J != NumDefOps; ++J) |
4247 | ResultRegs[J].push_back(Inst.getReg(J)); |
4248 | } |
4249 | |
4250 | // Fill out the widened result with undef instead of creating instructions |
4251 | // with undef inputs. |
4252 | int NumUndefParts = NumParts - NumRealParts; |
4253 | if (NumUndefParts != 0) { |
4254 | Register Undef = MIRBuilder.buildUndef(NarrowDstTy).getReg(0); |
4255 | for (int I = 0; I != NumDefOps; ++I) |
4256 | ResultRegs[I].append(NumUndefParts, Undef); |
4257 | } |
4258 | |
4259 | // Extract the possibly padded result. Use a scratch register if we need to do |
4260 | // a final bitcast, otherwise use the original result register. |
4261 | Register MergeDstReg; |
4262 | for (int I = 0; I != NumDefOps; ++I) { |
4263 | if (IsNarrow && DstTy.isVector()) |
4264 | MergeDstReg = MRI.createGenericVirtualRegister(DstScalarTy); |
4265 | else |
4266 | MergeDstReg = DstRegs[I]; |
4267 | |
4268 | buildWidenedRemergeToDst(MergeDstReg, DstLCMTy, ResultRegs[I]); |
4269 | |
4270 | // Recast to vector if we narrowed a vector |
4271 | if (IsNarrow && DstTy.isVector()) |
4272 | MIRBuilder.buildBitcast(DstRegs[I], MergeDstReg); |
4273 | } |
4274 | |
4275 | MI.eraseFromParent(); |
4276 | return Legalized; |
4277 | } |
4278 | |
4279 | LegalizerHelper::LegalizeResult |
4280 | LegalizerHelper::fewerElementsVectorSextInReg(MachineInstr &MI, unsigned TypeIdx, |
4281 | LLT NarrowTy) { |
4282 | Register DstReg = MI.getOperand(0).getReg(); |
4283 | Register SrcReg = MI.getOperand(1).getReg(); |
4284 | int64_t Imm = MI.getOperand(2).getImm(); |
4285 | |
4286 | LLT DstTy = MRI.getType(DstReg); |
4287 | |
4288 | SmallVector<Register, 8> Parts; |
4289 | LLT GCDTy = extractGCDType(Parts, DstTy, NarrowTy, SrcReg); |
4290 | LLT LCMTy = buildLCMMergePieces(DstTy, NarrowTy, GCDTy, Parts); |
4291 | |
4292 | for (Register &R : Parts) |
4293 | R = MIRBuilder.buildSExtInReg(NarrowTy, R, Imm).getReg(0); |
4294 | |
4295 | buildWidenedRemergeToDst(DstReg, LCMTy, Parts); |
4296 | |
4297 | MI.eraseFromParent(); |
4298 | return Legalized; |
4299 | } |
4300 | |
4301 | LegalizerHelper::LegalizeResult |
4302 | LegalizerHelper::fewerElementsVector(MachineInstr &MI, unsigned TypeIdx, |
4303 | LLT NarrowTy) { |
4304 | using namespace TargetOpcode; |
4305 | |
4306 | switch (MI.getOpcode()) { |
4307 | case G_IMPLICIT_DEF: |
4308 | return fewerElementsVectorImplicitDef(MI, TypeIdx, NarrowTy); |
4309 | case G_TRUNC: |
4310 | case G_AND: |
4311 | case G_OR: |
4312 | case G_XOR: |
4313 | case G_ADD: |
4314 | case G_SUB: |
4315 | case G_MUL: |
4316 | case G_PTR_ADD: |
4317 | case G_SMULH: |
4318 | case G_UMULH: |
4319 | case G_FADD: |
4320 | case G_FMUL: |
4321 | case G_FSUB: |
4322 | case G_FNEG: |
4323 | case G_FABS: |
4324 | case G_FCANONICALIZE: |
4325 | case G_FDIV: |
4326 | case G_FREM: |
4327 | case G_FMA: |
4328 | case G_FMAD: |
4329 | case G_FPOW: |
4330 | case G_FEXP: |
4331 | case G_FEXP2: |
4332 | case G_FLOG: |
4333 | case G_FLOG2: |
4334 | case G_FLOG10: |
4335 | case G_FNEARBYINT: |
4336 | case G_FCEIL: |
4337 | case G_FFLOOR: |
4338 | case G_FRINT: |
4339 | case G_INTRINSIC_ROUND: |
4340 | case G_INTRINSIC_ROUNDEVEN: |
4341 | case G_INTRINSIC_TRUNC: |
4342 | case G_FCOS: |
4343 | case G_FSIN: |
4344 | case G_FSQRT: |
4345 | case G_BSWAP: |
4346 | case G_BITREVERSE: |
4347 | case G_SDIV: |
4348 | case G_UDIV: |
4349 | case G_SREM: |
4350 | case G_UREM: |
4351 | case G_SDIVREM: |
4352 | case G_UDIVREM: |
4353 | case G_SMIN: |
4354 | case G_SMAX: |
4355 | case G_UMIN: |
4356 | case G_UMAX: |
4357 | case G_ABS: |
4358 | case G_FMINNUM: |
4359 | case G_FMAXNUM: |
4360 | case G_FMINNUM_IEEE: |
4361 | case G_FMAXNUM_IEEE: |
4362 | case G_FMINIMUM: |
4363 | case G_FMAXIMUM: |
4364 | case G_FSHL: |
4365 | case G_FSHR: |
4366 | case G_FREEZE: |
4367 | case G_SADDSAT: |
4368 | case G_SSUBSAT: |
4369 | case G_UADDSAT: |
4370 | case G_USUBSAT: |
4371 | return reduceOperationWidth(MI, TypeIdx, NarrowTy); |
4372 | case G_UMULO: |
4373 | case G_SMULO: |
4374 | return fewerElementsVectorMulo(MI, TypeIdx, NarrowTy); |
4375 | case G_SHL: |
4376 | case G_LSHR: |
4377 | case G_ASHR: |
4378 | case G_SSHLSAT: |
4379 | case G_USHLSAT: |
4380 | case G_CTLZ: |
4381 | case G_CTLZ_ZERO_UNDEF: |
4382 | case G_CTTZ: |
4383 | case G_CTTZ_ZERO_UNDEF: |
4384 | case G_CTPOP: |
4385 | case G_FCOPYSIGN: |
4386 | return fewerElementsVectorMultiEltType(MI, TypeIdx, NarrowTy); |
4387 | case G_ZEXT: |
4388 | case G_SEXT: |
4389 | case G_ANYEXT: |
4390 | case G_FPEXT: |
4391 | case G_FPTRUNC: |
4392 | case G_SITOFP: |
4393 | case G_UITOFP: |
4394 | case G_FPTOSI: |
4395 | case G_FPTOUI: |
4396 | case G_INTTOPTR: |
4397 | case G_PTRTOINT: |
4398 | case G_ADDRSPACE_CAST: |
4399 | return fewerElementsVectorCasts(MI, TypeIdx, NarrowTy); |
4400 | case G_ICMP: |
4401 | case G_FCMP: |
4402 | return fewerElementsVectorCmp(MI, TypeIdx, NarrowTy); |
4403 | case G_SELECT: |
4404 | return fewerElementsVectorSelect(MI, TypeIdx, NarrowTy); |
4405 | case G_PHI: |
4406 | return fewerElementsVectorPhi(MI, TypeIdx, NarrowTy); |
4407 | case G_UNMERGE_VALUES: |
4408 | return fewerElementsVectorUnmergeValues(MI, TypeIdx, NarrowTy); |
4409 | case G_BUILD_VECTOR: |
4410 | assert(TypeIdx == 0 && "not a vector type index")((void)0); |
4411 | return fewerElementsVectorMerge(MI, TypeIdx, NarrowTy); |
4412 | case G_CONCAT_VECTORS: |
4413 | if (TypeIdx != 1) // TODO: This probably does work as expected already. |
4414 | return UnableToLegalize; |
4415 | return fewerElementsVectorMerge(MI, TypeIdx, NarrowTy); |
4416 | case G_EXTRACT_VECTOR_ELT: |
4417 | case G_INSERT_VECTOR_ELT: |
4418 | return fewerElementsVectorExtractInsertVectorElt(MI, TypeIdx, NarrowTy); |
4419 | case G_LOAD: |
4420 | case G_STORE: |
4421 | return reduceLoadStoreWidth(cast<GLoadStore>(MI), TypeIdx, NarrowTy); |
4422 | case G_SEXT_INREG: |
4423 | return fewerElementsVectorSextInReg(MI, TypeIdx, NarrowTy); |
4424 | GISEL_VECREDUCE_CASES_NONSEQcase TargetOpcode::G_VECREDUCE_FADD: case TargetOpcode::G_VECREDUCE_FMUL : case TargetOpcode::G_VECREDUCE_FMAX: case TargetOpcode::G_VECREDUCE_FMIN : case TargetOpcode::G_VECREDUCE_ADD: case TargetOpcode::G_VECREDUCE_MUL : case TargetOpcode::G_VECREDUCE_AND: case TargetOpcode::G_VECREDUCE_OR : case TargetOpcode::G_VECREDUCE_XOR: case TargetOpcode::G_VECREDUCE_SMAX : case TargetOpcode::G_VECREDUCE_SMIN: case TargetOpcode::G_VECREDUCE_UMAX : case TargetOpcode::G_VECREDUCE_UMIN: |
4425 | return fewerElementsVectorReductions(MI, TypeIdx, NarrowTy); |
4426 | case G_SHUFFLE_VECTOR: |
4427 | return fewerElementsVectorShuffle(MI, TypeIdx, NarrowTy); |
4428 | default: |
4429 | return UnableToLegalize; |
4430 | } |
4431 | } |
4432 | |
4433 | LegalizerHelper::LegalizeResult LegalizerHelper::fewerElementsVectorShuffle( |
4434 | MachineInstr &MI, unsigned int TypeIdx, LLT NarrowTy) { |
4435 | assert(MI.getOpcode() == TargetOpcode::G_SHUFFLE_VECTOR)((void)0); |
4436 | if (TypeIdx != 0) |
4437 | return UnableToLegalize; |
4438 | |
4439 | Register DstReg = MI.getOperand(0).getReg(); |
4440 | Register Src1Reg = MI.getOperand(1).getReg(); |
4441 | Register Src2Reg = MI.getOperand(2).getReg(); |
4442 | ArrayRef<int> Mask = MI.getOperand(3).getShuffleMask(); |
4443 | LLT DstTy = MRI.getType(DstReg); |
4444 | LLT Src1Ty = MRI.getType(Src1Reg); |
4445 | LLT Src2Ty = MRI.getType(Src2Reg); |
4446 | // The shuffle should be canonicalized by now. |
4447 | if (DstTy != Src1Ty) |
4448 | return UnableToLegalize; |
4449 | if (DstTy != Src2Ty) |
4450 | return UnableToLegalize; |
4451 | |
4452 | if (!isPowerOf2_32(DstTy.getNumElements())) |
4453 | return UnableToLegalize; |
4454 | |
4455 | // We only support splitting a shuffle into 2, so adjust NarrowTy accordingly. |
4456 | // Further legalization attempts will be needed to do split further. |
4457 | NarrowTy = |
4458 | DstTy.changeElementCount(DstTy.getElementCount().divideCoefficientBy(2)); |
4459 | unsigned NewElts = NarrowTy.getNumElements(); |
4460 | |
4461 | SmallVector<Register> SplitSrc1Regs, SplitSrc2Regs; |
4462 | extractParts(Src1Reg, NarrowTy, 2, SplitSrc1Regs); |
4463 | extractParts(Src2Reg, NarrowTy, 2, SplitSrc2Regs); |
4464 | Register Inputs[4] = {SplitSrc1Regs[0], SplitSrc1Regs[1], SplitSrc2Regs[0], |
4465 | SplitSrc2Regs[1]}; |
4466 | |
4467 | Register Hi, Lo; |
4468 | |
4469 | // If Lo or Hi uses elements from at most two of the four input vectors, then |
4470 | // express it as a vector shuffle of those two inputs. Otherwise extract the |
4471 | // input elements by hand and construct the Lo/Hi output using a BUILD_VECTOR. |
4472 | SmallVector<int, 16> Ops; |
4473 | for (unsigned High = 0; High < 2; ++High) { |
4474 | Register &Output = High ? Hi : Lo; |
4475 | |
4476 | // Build a shuffle mask for the output, discovering on the fly which |
4477 | // input vectors to use as shuffle operands (recorded in InputUsed). |
4478 | // If building a suitable shuffle vector proves too hard, then bail |
4479 | // out with useBuildVector set. |
4480 | unsigned InputUsed[2] = {-1U, -1U}; // Not yet discovered. |
4481 | unsigned FirstMaskIdx = High * NewElts; |
4482 | bool UseBuildVector = false; |
4483 | for (unsigned MaskOffset = 0; MaskOffset < NewElts; ++MaskOffset) { |
4484 | // The mask element. This indexes into the input. |
4485 | int Idx = Mask[FirstMaskIdx + MaskOffset]; |
4486 | |
4487 | // The input vector this mask element indexes into. |
4488 | unsigned Input = (unsigned)Idx / NewElts; |
4489 | |
4490 | if (Input >= array_lengthof(Inputs)) { |
4491 | // The mask element does not index into any input vector. |
4492 | Ops.push_back(-1); |
4493 | continue; |
4494 | } |
4495 | |
4496 | // Turn the index into an offset from the start of the input vector. |
4497 | Idx -= Input * NewElts; |
4498 | |
4499 | // Find or create a shuffle vector operand to hold this input. |
4500 | unsigned OpNo; |
4501 | for (OpNo = 0; OpNo < array_lengthof(InputUsed); ++OpNo) { |
4502 | if (InputUsed[OpNo] == Input) { |
4503 | // This input vector is already an operand. |
4504 | break; |
4505 | } else if (InputUsed[OpNo] == -1U) { |
4506 | // Create a new operand for this input vector. |
4507 | InputUsed[OpNo] = Input; |
4508 | break; |
4509 | } |
4510 | } |
4511 | |
4512 | if (OpNo >= array_lengthof(InputUsed)) { |
4513 | // More than two input vectors used! Give up on trying to create a |
4514 | // shuffle vector. Insert all elements into a BUILD_VECTOR instead. |
4515 | UseBuildVector = true; |
4516 | break; |
4517 | } |
4518 | |
4519 | // Add the mask index for the new shuffle vector. |
4520 | Ops.push_back(Idx + OpNo * NewElts); |
4521 | } |
4522 | |
4523 | if (UseBuildVector) { |
4524 | LLT EltTy = NarrowTy.getElementType(); |
4525 | SmallVector<Register, 16> SVOps; |
4526 | |
4527 | // Extract the input elements by hand. |
4528 | for (unsigned MaskOffset = 0; MaskOffset < NewElts; ++MaskOffset) { |
4529 | // The mask element. This indexes into the input. |
4530 | int Idx = Mask[FirstMaskIdx + MaskOffset]; |
4531 | |
4532 | // The input vector this mask element indexes into. |
4533 | unsigned Input = (unsigned)Idx / NewElts; |
4534 | |
4535 | if (Input >= array_lengthof(Inputs)) { |
4536 | // The mask element is "undef" or indexes off the end of the input. |
4537 | SVOps.push_back(MIRBuilder.buildUndef(EltTy).getReg(0)); |
4538 | continue; |
4539 | } |
4540 | |
4541 | // Turn the index into an offset from the start of the input vector. |
4542 | Idx -= Input * NewElts; |
4543 | |
4544 | // Extract the vector element by hand. |
4545 | SVOps.push_back(MIRBuilder |
4546 | .buildExtractVectorElement( |
4547 | EltTy, Inputs[Input], |
4548 | MIRBuilder.buildConstant(LLT::scalar(32), Idx)) |
4549 | .getReg(0)); |
4550 | } |
4551 | |
4552 | // Construct the Lo/Hi output using a G_BUILD_VECTOR. |
4553 | Output = MIRBuilder.buildBuildVector(NarrowTy, SVOps).getReg(0); |
4554 | } else if (InputUsed[0] == -1U) { |
4555 | // No input vectors were used! The result is undefined. |
4556 | Output = MIRBuilder.buildUndef(NarrowTy).getReg(0); |
4557 | } else { |
4558 | Register Op0 = Inputs[InputUsed[0]]; |
4559 | // If only one input was used, use an undefined vector for the other. |
4560 | Register Op1 = InputUsed[1] == -1U |
4561 | ? MIRBuilder.buildUndef(NarrowTy).getReg(0) |
4562 | : Inputs[InputUsed[1]]; |
4563 | // At least one input vector was used. Create a new shuffle vector. |
4564 | Output = MIRBuilder.buildShuffleVector(NarrowTy, Op0, Op1, Ops).getReg(0); |
4565 | } |
4566 | |
4567 | Ops.clear(); |
4568 | } |
4569 | |
4570 | MIRBuilder.buildConcatVectors(DstReg, {Lo, Hi}); |
4571 | MI.eraseFromParent(); |
4572 | return Legalized; |
4573 | } |
4574 | |
4575 | LegalizerHelper::LegalizeResult LegalizerHelper::fewerElementsVectorReductions( |
4576 | MachineInstr &MI, unsigned int TypeIdx, LLT NarrowTy) { |
4577 | unsigned Opc = MI.getOpcode(); |
4578 | assert(Opc != TargetOpcode::G_VECREDUCE_SEQ_FADD &&((void)0) |
4579 | Opc != TargetOpcode::G_VECREDUCE_SEQ_FMUL &&((void)0) |
4580 | "Sequential reductions not expected")((void)0); |
4581 | |
4582 | if (TypeIdx != 1) |
4583 | return UnableToLegalize; |
4584 | |
4585 | // The semantics of the normal non-sequential reductions allow us to freely |
4586 | // re-associate the operation. |
4587 | Register SrcReg = MI.getOperand(1).getReg(); |
4588 | LLT SrcTy = MRI.getType(SrcReg); |
4589 | Register DstReg = MI.getOperand(0).getReg(); |
4590 | LLT DstTy = MRI.getType(DstReg); |
4591 | |
4592 | if (SrcTy.getNumElements() % NarrowTy.getNumElements() != 0) |
4593 | return UnableToLegalize; |
4594 | |
4595 | SmallVector<Register> SplitSrcs; |
4596 | const unsigned NumParts = SrcTy.getNumElements() / NarrowTy.getNumElements(); |
4597 | extractParts(SrcReg, NarrowTy, NumParts, SplitSrcs); |
4598 | SmallVector<Register> PartialReductions; |
4599 | for (unsigned Part = 0; Part < NumParts; ++Part) { |
4600 | PartialReductions.push_back( |
4601 | MIRBuilder.buildInstr(Opc, {DstTy}, {SplitSrcs[Part]}).getReg(0)); |
4602 | } |
4603 | |
4604 | unsigned ScalarOpc; |
4605 | switch (Opc) { |
4606 | case TargetOpcode::G_VECREDUCE_FADD: |
4607 | ScalarOpc = TargetOpcode::G_FADD; |
4608 | break; |
4609 | case TargetOpcode::G_VECREDUCE_FMUL: |
4610 | ScalarOpc = TargetOpcode::G_FMUL; |
4611 | break; |
4612 | case TargetOpcode::G_VECREDUCE_FMAX: |
4613 | ScalarOpc = TargetOpcode::G_FMAXNUM; |
4614 | break; |
4615 | case TargetOpcode::G_VECREDUCE_FMIN: |
4616 | ScalarOpc = TargetOpcode::G_FMINNUM; |
4617 | break; |
4618 | case TargetOpcode::G_VECREDUCE_ADD: |
4619 | ScalarOpc = TargetOpcode::G_ADD; |
4620 | break; |
4621 | case TargetOpcode::G_VECREDUCE_MUL: |
4622 | ScalarOpc = TargetOpcode::G_MUL; |
4623 | break; |
4624 | case TargetOpcode::G_VECREDUCE_AND: |
4625 | ScalarOpc = TargetOpcode::G_AND; |
4626 | break; |
4627 | case TargetOpcode::G_VECREDUCE_OR: |
4628 | ScalarOpc = TargetOpcode::G_OR; |
4629 | break; |
4630 | case TargetOpcode::G_VECREDUCE_XOR: |
4631 | ScalarOpc = TargetOpcode::G_XOR; |
4632 | break; |
4633 | case TargetOpcode::G_VECREDUCE_SMAX: |
4634 | ScalarOpc = TargetOpcode::G_SMAX; |
4635 | break; |
4636 | case TargetOpcode::G_VECREDUCE_SMIN: |
4637 | ScalarOpc = TargetOpcode::G_SMIN; |
4638 | break; |
4639 | case TargetOpcode::G_VECREDUCE_UMAX: |
4640 | ScalarOpc = TargetOpcode::G_UMAX; |
4641 | break; |
4642 | case TargetOpcode::G_VECREDUCE_UMIN: |
4643 | ScalarOpc = TargetOpcode::G_UMIN; |
4644 | break; |
4645 | default: |
4646 | LLVM_DEBUG(dbgs() << "Can't legalize: unknown reduction kind.\n")do { } while (false); |
4647 | return UnableToLegalize; |
4648 | } |
4649 | |
4650 | // If the types involved are powers of 2, we can generate intermediate vector |
4651 | // ops, before generating a final reduction operation. |
4652 | if (isPowerOf2_32(SrcTy.getNumElements()) && |
4653 | isPowerOf2_32(NarrowTy.getNumElements())) { |
4654 | return tryNarrowPow2Reduction(MI, SrcReg, SrcTy, NarrowTy, ScalarOpc); |
4655 | } |
4656 | |
4657 | Register Acc = PartialReductions[0]; |
4658 | for (unsigned Part = 1; Part < NumParts; ++Part) { |
4659 | if (Part == NumParts - 1) { |
4660 | MIRBuilder.buildInstr(ScalarOpc, {DstReg}, |
4661 | {Acc, PartialReductions[Part]}); |
4662 | } else { |
4663 | Acc = MIRBuilder |
4664 | .buildInstr(ScalarOpc, {DstTy}, {Acc, PartialReductions[Part]}) |
4665 | .getReg(0); |
4666 | } |
4667 | } |
4668 | MI.eraseFromParent(); |
4669 | return Legalized; |
4670 | } |
4671 | |
4672 | LegalizerHelper::LegalizeResult |
4673 | LegalizerHelper::tryNarrowPow2Reduction(MachineInstr &MI, Register SrcReg, |
4674 | LLT SrcTy, LLT NarrowTy, |
4675 | unsigned ScalarOpc) { |
4676 | SmallVector<Register> SplitSrcs; |
4677 | // Split the sources into NarrowTy size pieces. |
4678 | extractParts(SrcReg, NarrowTy, |
4679 | SrcTy.getNumElements() / NarrowTy.getNumElements(), SplitSrcs); |
4680 | // We're going to do a tree reduction using vector operations until we have |
4681 | // one NarrowTy size value left. |
4682 | while (SplitSrcs.size() > 1) { |
4683 | SmallVector<Register> PartialRdxs; |
4684 | for (unsigned Idx = 0; Idx < SplitSrcs.size()-1; Idx += 2) { |
4685 | Register LHS = SplitSrcs[Idx]; |
4686 | Register RHS = SplitSrcs[Idx + 1]; |
4687 | // Create the intermediate vector op. |
4688 | Register Res = |
4689 | MIRBuilder.buildInstr(ScalarOpc, {NarrowTy}, {LHS, RHS}).getReg(0); |
4690 | PartialRdxs.push_back(Res); |
4691 | } |
4692 | SplitSrcs = std::move(PartialRdxs); |
4693 | } |
4694 | // Finally generate the requested NarrowTy based reduction. |
4695 | Observer.changingInstr(MI); |
4696 | MI.getOperand(1).setReg(SplitSrcs[0]); |
4697 | Observer.changedInstr(MI); |
4698 | return Legalized; |
4699 | } |
4700 | |
4701 | LegalizerHelper::LegalizeResult |
4702 | LegalizerHelper::narrowScalarShiftByConstant(MachineInstr &MI, const APInt &Amt, |
4703 | const LLT HalfTy, const LLT AmtTy) { |
4704 | |
4705 | Register InL = MRI.createGenericVirtualRegister(HalfTy); |
4706 | Register InH = MRI.createGenericVirtualRegister(HalfTy); |
4707 | MIRBuilder.buildUnmerge({InL, InH}, MI.getOperand(1)); |
4708 | |
4709 | if (Amt.isNullValue()) { |
4710 | MIRBuilder.buildMerge(MI.getOperand(0), {InL, InH}); |
4711 | MI.eraseFromParent(); |
4712 | return Legalized; |
4713 | } |
4714 | |
4715 | LLT NVT = HalfTy; |
4716 | unsigned NVTBits = HalfTy.getSizeInBits(); |
4717 | unsigned VTBits = 2 * NVTBits; |
4718 | |
4719 | SrcOp Lo(Register(0)), Hi(Register(0)); |
4720 | if (MI.getOpcode() == TargetOpcode::G_SHL) { |
4721 | if (Amt.ugt(VTBits)) { |
4722 | Lo = Hi = MIRBuilder.buildConstant(NVT, 0); |
4723 | } else if (Amt.ugt(NVTBits)) { |
4724 | Lo = MIRBuilder.buildConstant(NVT, 0); |
4725 | Hi = MIRBuilder.buildShl(NVT, InL, |
4726 | MIRBuilder.buildConstant(AmtTy, Amt - NVTBits)); |
4727 | } else if (Amt == NVTBits) { |
4728 | Lo = MIRBuilder.buildConstant(NVT, 0); |
4729 | Hi = InL; |
4730 | } else { |
4731 | Lo = MIRBuilder.buildShl(NVT, InL, MIRBuilder.buildConstant(AmtTy, Amt)); |
4732 | auto OrLHS = |
4733 | MIRBuilder.buildShl(NVT, InH, MIRBuilder.buildConstant(AmtTy, Amt)); |
4734 | auto OrRHS = MIRBuilder.buildLShr( |
4735 | NVT, InL, MIRBuilder.buildConstant(AmtTy, -Amt + NVTBits)); |
4736 | Hi = MIRBuilder.buildOr(NVT, OrLHS, OrRHS); |
4737 | } |
4738 | } else if (MI.getOpcode() == TargetOpcode::G_LSHR) { |
4739 | if (Amt.ugt(VTBits)) { |
4740 | Lo = Hi = MIRBuilder.buildConstant(NVT, 0); |
4741 | } else if (Amt.ugt(NVTBits)) { |
4742 | Lo = MIRBuilder.buildLShr(NVT, InH, |
4743 | MIRBuilder.buildConstant(AmtTy, Amt - NVTBits)); |
4744 | Hi = MIRBuilder.buildConstant(NVT, 0); |
4745 | } else if (Amt == NVTBits) { |
4746 | Lo = InH; |
4747 | Hi = MIRBuilder.buildConstant(NVT, 0); |
4748 | } else { |
4749 | auto ShiftAmtConst = MIRBuilder.buildConstant(AmtTy, Amt); |
4750 | |
4751 | auto OrLHS = MIRBuilder.buildLShr(NVT, InL, ShiftAmtConst); |
4752 | auto OrRHS = MIRBuilder.buildShl( |
4753 | NVT, InH, MIRBuilder.buildConstant(AmtTy, -Amt + NVTBits)); |
4754 | |
4755 | Lo = MIRBuilder.buildOr(NVT, OrLHS, OrRHS); |
4756 | Hi = MIRBuilder.buildLShr(NVT, InH, ShiftAmtConst); |
4757 | } |
4758 | } else { |
4759 | if (Amt.ugt(VTBits)) { |
4760 | Hi = Lo = MIRBuilder.buildAShr( |
4761 | NVT, InH, MIRBuilder.buildConstant(AmtTy, NVTBits - 1)); |
4762 | } else if (Amt.ugt(NVTBits)) { |
4763 | Lo = MIRBuilder.buildAShr(NVT, InH, |
4764 | MIRBuilder.buildConstant(AmtTy, Amt - NVTBits)); |
4765 | Hi = MIRBuilder.buildAShr(NVT, InH, |
4766 | MIRBuilder.buildConstant(AmtTy, NVTBits - 1)); |
4767 | } else if (Amt == NVTBits) { |
4768 | Lo = InH; |
4769 | Hi = MIRBuilder.buildAShr(NVT, InH, |
4770 | MIRBuilder.buildConstant(AmtTy, NVTBits - 1)); |
4771 | } else { |
4772 | auto ShiftAmtConst = MIRBuilder.buildConstant(AmtTy, Amt); |
4773 | |
4774 | auto OrLHS = MIRBuilder.buildLShr(NVT, InL, ShiftAmtConst); |
4775 | auto OrRHS = MIRBuilder.buildShl( |
4776 | NVT, InH, MIRBuilder.buildConstant(AmtTy, -Amt + NVTBits)); |
4777 | |
4778 | Lo = MIRBuilder.buildOr(NVT, OrLHS, OrRHS); |
4779 | Hi = MIRBuilder.buildAShr(NVT, InH, ShiftAmtConst); |
4780 | } |
4781 | } |
4782 | |
4783 | MIRBuilder.buildMerge(MI.getOperand(0), {Lo, Hi}); |
4784 | MI.eraseFromParent(); |
4785 | |
4786 | return Legalized; |
4787 | } |
4788 | |
4789 | // TODO: Optimize if constant shift amount. |
4790 | LegalizerHelper::LegalizeResult |
4791 | LegalizerHelper::narrowScalarShift(MachineInstr &MI, unsigned TypeIdx, |
4792 | LLT RequestedTy) { |
4793 | if (TypeIdx == 1) { |
4794 | Observer.changingInstr(MI); |
4795 | narrowScalarSrc(MI, RequestedTy, 2); |
4796 | Observer.changedInstr(MI); |
4797 | return Legalized; |
4798 | } |
4799 | |
4800 | Register DstReg = MI.getOperand(0).getReg(); |
4801 | LLT DstTy = MRI.getType(DstReg); |
4802 | if (DstTy.isVector()) |
4803 | return UnableToLegalize; |
4804 | |
4805 | Register Amt = MI.getOperand(2).getReg(); |
4806 | LLT ShiftAmtTy = MRI.getType(Amt); |
4807 | const unsigned DstEltSize = DstTy.getScalarSizeInBits(); |
4808 | if (DstEltSize % 2 != 0) |
4809 | return UnableToLegalize; |
4810 | |
4811 | // Ignore the input type. We can only go to exactly half the size of the |
4812 | // input. If that isn't small enough, the resulting pieces will be further |
4813 | // legalized. |
4814 | const unsigned NewBitSize = DstEltSize / 2; |
4815 | const LLT HalfTy = LLT::scalar(NewBitSize); |
4816 | const LLT CondTy = LLT::scalar(1); |
4817 | |
4818 | if (const MachineInstr *KShiftAmt = |
4819 | getOpcodeDef(TargetOpcode::G_CONSTANT, Amt, MRI)) { |
4820 | return narrowScalarShiftByConstant( |
4821 | MI, KShiftAmt->getOperand(1).getCImm()->getValue(), HalfTy, ShiftAmtTy); |
4822 | } |
4823 | |
4824 | // TODO: Expand with known bits. |
4825 | |
4826 | // Handle the fully general expansion by an unknown amount. |
4827 | auto NewBits = MIRBuilder.buildConstant(ShiftAmtTy, NewBitSize); |
4828 | |
4829 | Register InL = MRI.createGenericVirtualRegister(HalfTy); |
4830 | Register InH = MRI.createGenericVirtualRegister(HalfTy); |
4831 | MIRBuilder.buildUnmerge({InL, InH}, MI.getOperand(1)); |
4832 | |
4833 | auto AmtExcess = MIRBuilder.buildSub(ShiftAmtTy, Amt, NewBits); |
4834 | auto AmtLack = MIRBuilder.buildSub(ShiftAmtTy, NewBits, Amt); |
4835 | |
4836 | auto Zero = MIRBuilder.buildConstant(ShiftAmtTy, 0); |
4837 | auto IsShort = MIRBuilder.buildICmp(ICmpInst::ICMP_ULT, CondTy, Amt, NewBits); |
4838 | auto IsZero = MIRBuilder.buildICmp(ICmpInst::ICMP_EQ, CondTy, Amt, Zero); |
4839 | |
4840 | Register ResultRegs[2]; |
4841 | switch (MI.getOpcode()) { |
4842 | case TargetOpcode::G_SHL: { |
4843 | // Short: ShAmt < NewBitSize |
4844 | auto LoS = MIRBuilder.buildShl(HalfTy, InL, Amt); |
4845 | |
4846 | auto LoOr = MIRBuilder.buildLShr(HalfTy, InL, AmtLack); |
4847 | auto HiOr = MIRBuilder.buildShl(HalfTy, InH, Amt); |
4848 | auto HiS = MIRBuilder.buildOr(HalfTy, LoOr, HiOr); |
4849 | |
4850 | // Long: ShAmt >= NewBitSize |
4851 | auto LoL = MIRBuilder.buildConstant(HalfTy, 0); // Lo part is zero. |
4852 | auto HiL = MIRBuilder.buildShl(HalfTy, InL, AmtExcess); // Hi from Lo part. |
4853 | |
4854 | auto Lo = MIRBuilder.buildSelect(HalfTy, IsShort, LoS, LoL); |
4855 | auto Hi = MIRBuilder.buildSelect( |
4856 | HalfTy, IsZero, InH, MIRBuilder.buildSelect(HalfTy, IsShort, HiS, HiL)); |
4857 | |
4858 | ResultRegs[0] = Lo.getReg(0); |
4859 | ResultRegs[1] = Hi.getReg(0); |
4860 | break; |
4861 | } |
4862 | case TargetOpcode::G_LSHR: |
4863 | case TargetOpcode::G_ASHR: { |
4864 | // Short: ShAmt < NewBitSize |
4865 | auto HiS = MIRBuilder.buildInstr(MI.getOpcode(), {HalfTy}, {InH, Amt}); |
4866 | |
4867 | auto LoOr = MIRBuilder.buildLShr(HalfTy, InL, Amt); |
4868 | auto HiOr = MIRBuilder.buildShl(HalfTy, InH, AmtLack); |
4869 | auto LoS = MIRBuilder.buildOr(HalfTy, LoOr, HiOr); |
4870 | |
4871 | // Long: ShAmt >= NewBitSize |
4872 | MachineInstrBuilder HiL; |
4873 | if (MI.getOpcode() == TargetOpcode::G_LSHR) { |
4874 | HiL = MIRBuilder.buildConstant(HalfTy, 0); // Hi part is zero. |
4875 | } else { |
4876 | auto ShiftAmt = MIRBuilder.buildConstant(ShiftAmtTy, NewBitSize - 1); |
4877 | HiL = MIRBuilder.buildAShr(HalfTy, InH, ShiftAmt); // Sign of Hi part. |
4878 | } |
4879 | auto LoL = MIRBuilder.buildInstr(MI.getOpcode(), {HalfTy}, |
4880 | {InH, AmtExcess}); // Lo from Hi part. |
4881 | |
4882 | auto Lo = MIRBuilder.buildSelect( |
4883 | HalfTy, IsZero, InL, MIRBuilder.buildSelect(HalfTy, IsShort, LoS, LoL)); |
4884 | |
4885 | auto Hi = MIRBuilder.buildSelect(HalfTy, IsShort, HiS, HiL); |
4886 | |
4887 | ResultRegs[0] = Lo.getReg(0); |
4888 | ResultRegs[1] = Hi.getReg(0); |
4889 | break; |
4890 | } |
4891 | default: |
4892 | llvm_unreachable("not a shift")__builtin_unreachable(); |
4893 | } |
4894 | |
4895 | MIRBuilder.buildMerge(DstReg, ResultRegs); |
4896 | MI.eraseFromParent(); |
4897 | return Legalized; |
4898 | } |
4899 | |
4900 | LegalizerHelper::LegalizeResult |
4901 | LegalizerHelper::moreElementsVectorPhi(MachineInstr &MI, unsigned TypeIdx, |
4902 | LLT MoreTy) { |
4903 | assert(TypeIdx == 0 && "Expecting only Idx 0")((void)0); |
4904 | |
4905 | Observer.changingInstr(MI); |
4906 | for (unsigned I = 1, E = MI.getNumOperands(); I != E; I += 2) { |
4907 | MachineBasicBlock &OpMBB = *MI.getOperand(I + 1).getMBB(); |
4908 | MIRBuilder.setInsertPt(OpMBB, OpMBB.getFirstTerminator()); |
4909 | moreElementsVectorSrc(MI, MoreTy, I); |
4910 | } |
4911 | |
4912 | MachineBasicBlock &MBB = *MI.getParent(); |
4913 | MIRBuilder.setInsertPt(MBB, --MBB.getFirstNonPHI()); |
4914 | moreElementsVectorDst(MI, MoreTy, 0); |
4915 | Observer.changedInstr(MI); |
4916 | return Legalized; |
4917 | } |
4918 | |
4919 | LegalizerHelper::LegalizeResult |
4920 | LegalizerHelper::moreElementsVector(MachineInstr &MI, unsigned TypeIdx, |
4921 | LLT MoreTy) { |
4922 | unsigned Opc = MI.getOpcode(); |
4923 | switch (Opc) { |
4924 | case TargetOpcode::G_IMPLICIT_DEF: |
4925 | case TargetOpcode::G_LOAD: { |
4926 | if (TypeIdx != 0) |
4927 | return UnableToLegalize; |
4928 | Observer.changingInstr(MI); |
4929 | moreElementsVectorDst(MI, MoreTy, 0); |
4930 | Observer.changedInstr(MI); |
4931 | return Legalized; |
4932 | } |
4933 | case TargetOpcode::G_STORE: |
4934 | if (TypeIdx != 0) |
4935 | return UnableToLegalize; |
4936 | Observer.changingInstr(MI); |
4937 | moreElementsVectorSrc(MI, MoreTy, 0); |
4938 | Observer.changedInstr(MI); |
4939 | return Legalized; |
4940 | case TargetOpcode::G_AND: |
4941 | case TargetOpcode::G_OR: |
4942 | case TargetOpcode::G_XOR: |
4943 | case TargetOpcode::G_SMIN: |
4944 | case TargetOpcode::G_SMAX: |
4945 | case TargetOpcode::G_UMIN: |
4946 | case TargetOpcode::G_UMAX: |
4947 | case TargetOpcode::G_FMINNUM: |
4948 | case TargetOpcode::G_FMAXNUM: |
4949 | case TargetOpcode::G_FMINNUM_IEEE: |
4950 | case TargetOpcode::G_FMAXNUM_IEEE: |
4951 | case TargetOpcode::G_FMINIMUM: |
4952 | case TargetOpcode::G_FMAXIMUM: { |
4953 | Observer.changingInstr(MI); |
4954 | moreElementsVectorSrc(MI, MoreTy, 1); |
4955 | moreElementsVectorSrc(MI, MoreTy, 2); |
4956 | moreElementsVectorDst(MI, MoreTy, 0); |
4957 | Observer.changedInstr(MI); |
4958 | return Legalized; |
4959 | } |
4960 | case TargetOpcode::G_EXTRACT: |
4961 | if (TypeIdx != 1) |
4962 | return UnableToLegalize; |
4963 | Observer.changingInstr(MI); |
4964 | moreElementsVectorSrc(MI, MoreTy, 1); |
4965 | Observer.changedInstr(MI); |
4966 | return Legalized; |
4967 | case TargetOpcode::G_INSERT: |
4968 | case TargetOpcode::G_FREEZE: |
4969 | if (TypeIdx != 0) |
4970 | return UnableToLegalize; |
4971 | Observer.changingInstr(MI); |
4972 | moreElementsVectorSrc(MI, MoreTy, 1); |
4973 | moreElementsVectorDst(MI, MoreTy, 0); |
4974 | Observer.changedInstr(MI); |
4975 | return Legalized; |
4976 | case TargetOpcode::G_SELECT: |
4977 | if (TypeIdx != 0) |
4978 | return UnableToLegalize; |
4979 | if (MRI.getType(MI.getOperand(1).getReg()).isVector()) |
4980 | return UnableToLegalize; |
4981 | |
4982 | Observer.changingInstr(MI); |
4983 | moreElementsVectorSrc(MI, MoreTy, 2); |
4984 | moreElementsVectorSrc(MI, MoreTy, 3); |
4985 | moreElementsVectorDst(MI, MoreTy, 0); |
4986 | Observer.changedInstr(MI); |
4987 | return Legalized; |
4988 | case TargetOpcode::G_UNMERGE_VALUES: { |
4989 | if (TypeIdx != 1) |
4990 | return UnableToLegalize; |
4991 | |
4992 | LLT DstTy = MRI.getType(MI.getOperand(0).getReg()); |
4993 | int NumDst = MI.getNumOperands() - 1; |
4994 | moreElementsVectorSrc(MI, MoreTy, NumDst); |
4995 | |
4996 | auto MIB = MIRBuilder.buildInstr(TargetOpcode::G_UNMERGE_VALUES); |
4997 | for (int I = 0; I != NumDst; ++I) |
4998 | MIB.addDef(MI.getOperand(I).getReg()); |
4999 | |
5000 | int NewNumDst = MoreTy.getSizeInBits() / DstTy.getSizeInBits(); |
5001 | for (int I = NumDst; I != NewNumDst; ++I) |
5002 | MIB.addDef(MRI.createGenericVirtualRegister(DstTy)); |
5003 | |
5004 | MIB.addUse(MI.getOperand(NumDst).getReg()); |
5005 | MI.eraseFromParent(); |
5006 | return Legalized; |
5007 | } |
5008 | case TargetOpcode::G_PHI: |
5009 | return moreElementsVectorPhi(MI, TypeIdx, MoreTy); |
5010 | case TargetOpcode::G_SHUFFLE_VECTOR: |
5011 | return moreElementsVectorShuffle(MI, TypeIdx, MoreTy); |
5012 | default: |
5013 | return UnableToLegalize; |
5014 | } |
5015 | } |
5016 | |
5017 | LegalizerHelper::LegalizeResult |
5018 | LegalizerHelper::moreElementsVectorShuffle(MachineInstr &MI, |
5019 | unsigned int TypeIdx, LLT MoreTy) { |
5020 | if (TypeIdx != 0) |
5021 | return UnableToLegalize; |
5022 | |
5023 | Register DstReg = MI.getOperand(0).getReg(); |
5024 | Register Src1Reg = MI.getOperand(1).getReg(); |
5025 | Register Src2Reg = MI.getOperand(2).getReg(); |
5026 | ArrayRef<int> Mask = MI.getOperand(3).getShuffleMask(); |
5027 | LLT DstTy = MRI.getType(DstReg); |
5028 | LLT Src1Ty = MRI.getType(Src1Reg); |
5029 | LLT Src2Ty = MRI.getType(Src2Reg); |
5030 | unsigned NumElts = DstTy.getNumElements(); |
5031 | unsigned WidenNumElts = MoreTy.getNumElements(); |
5032 | |
5033 | // Expect a canonicalized shuffle. |
5034 | if (DstTy != Src1Ty || DstTy != Src2Ty) |
5035 | return UnableToLegalize; |
5036 | |
5037 | moreElementsVectorSrc(MI, MoreTy, 1); |
5038 | moreElementsVectorSrc(MI, MoreTy, 2); |
5039 | |
5040 | // Adjust mask based on new input vector length. |
5041 | SmallVector<int, 16> NewMask; |
5042 | for (unsigned I = 0; I != NumElts; ++I) { |
5043 | int Idx = Mask[I]; |
5044 | if (Idx < static_cast<int>(NumElts)) |
5045 | NewMask.push_back(Idx); |
5046 | else |
5047 | NewMask.push_back(Idx - NumElts + WidenNumElts); |
5048 | } |
5049 | for (unsigned I = NumElts; I != WidenNumElts; ++I) |
5050 | NewMask.push_back(-1); |
5051 | moreElementsVectorDst(MI, MoreTy, 0); |
5052 | MIRBuilder.setInstrAndDebugLoc(MI); |
5053 | MIRBuilder.buildShuffleVector(MI.getOperand(0).getReg(), |
5054 | MI.getOperand(1).getReg(), |
5055 | MI.getOperand(2).getReg(), NewMask); |
5056 | MI.eraseFromParent(); |
5057 | return Legalized; |
5058 | } |
5059 | |
5060 | void LegalizerHelper::multiplyRegisters(SmallVectorImpl<Register> &DstRegs, |
5061 | ArrayRef<Register> Src1Regs, |
5062 | ArrayRef<Register> Src2Regs, |
5063 | LLT NarrowTy) { |
5064 | MachineIRBuilder &B = MIRBuilder; |
5065 | unsigned SrcParts = Src1Regs.size(); |
5066 | unsigned DstParts = DstRegs.size(); |
5067 | |
5068 | unsigned DstIdx = 0; // Low bits of the result. |
5069 | Register FactorSum = |
5070 | B.buildMul(NarrowTy, Src1Regs[DstIdx], Src2Regs[DstIdx]).getReg(0); |
5071 | DstRegs[DstIdx] = FactorSum; |
5072 | |
5073 | unsigned CarrySumPrevDstIdx; |
5074 | SmallVector<Register, 4> Factors; |
5075 | |
5076 | for (DstIdx = 1; DstIdx < DstParts; DstIdx++) { |
5077 | // Collect low parts of muls for DstIdx. |
5078 | for (unsigned i = DstIdx + 1 < SrcParts ? 0 : DstIdx - SrcParts + 1; |
5079 | i <= std::min(DstIdx, SrcParts - 1); ++i) { |
5080 | MachineInstrBuilder Mul = |
5081 | B.buildMul(NarrowTy, Src1Regs[DstIdx - i], Src2Regs[i]); |
5082 | Factors.push_back(Mul.getReg(0)); |
5083 | } |
5084 | // Collect high parts of muls from previous DstIdx. |
5085 | for (unsigned i = DstIdx < SrcParts ? 0 : DstIdx - SrcParts; |
5086 | i <= std::min(DstIdx - 1, SrcParts - 1); ++i) { |
5087 | MachineInstrBuilder Umulh = |
5088 | B.buildUMulH(NarrowTy, Src1Regs[DstIdx - 1 - i], Src2Regs[i]); |
5089 | Factors.push_back(Umulh.getReg(0)); |
5090 | } |
5091 | // Add CarrySum from additions calculated for previous DstIdx. |
5092 | if (DstIdx != 1) { |
5093 | Factors.push_back(CarrySumPrevDstIdx); |
5094 | } |
5095 | |
5096 | Register CarrySum; |
5097 | // Add all factors and accumulate all carries into CarrySum. |
5098 | if (DstIdx != DstParts - 1) { |
5099 | MachineInstrBuilder Uaddo = |
5100 | B.buildUAddo(NarrowTy, LLT::scalar(1), Factors[0], Factors[1]); |
5101 | FactorSum = Uaddo.getReg(0); |
5102 | CarrySum = B.buildZExt(NarrowTy, Uaddo.getReg(1)).getReg(0); |
5103 | for (unsigned i = 2; i < Factors.size(); ++i) { |
5104 | MachineInstrBuilder Uaddo = |
5105 | B.buildUAddo(NarrowTy, LLT::scalar(1), FactorSum, Factors[i]); |
5106 | FactorSum = Uaddo.getReg(0); |
5107 | MachineInstrBuilder Carry = B.buildZExt(NarrowTy, Uaddo.getReg(1)); |
5108 | CarrySum = B.buildAdd(NarrowTy, CarrySum, Carry).getReg(0); |
5109 | } |
5110 | } else { |
5111 | // Since value for the next index is not calculated, neither is CarrySum. |
5112 | FactorSum = B.buildAdd(NarrowTy, Factors[0], Factors[1]).getReg(0); |
5113 | for (unsigned i = 2; i < Factors.size(); ++i) |
5114 | FactorSum = B.buildAdd(NarrowTy, FactorSum, Factors[i]).getReg(0); |
5115 | } |
5116 | |
5117 | CarrySumPrevDstIdx = CarrySum; |
5118 | DstRegs[DstIdx] = FactorSum; |
5119 | Factors.clear(); |
5120 | } |
5121 | } |
5122 | |
5123 | LegalizerHelper::LegalizeResult |
5124 | LegalizerHelper::narrowScalarAddSub(MachineInstr &MI, unsigned TypeIdx, |
5125 | LLT NarrowTy) { |
5126 | if (TypeIdx != 0) |
5127 | return UnableToLegalize; |
5128 | |
5129 | Register DstReg = MI.getOperand(0).getReg(); |
5130 | LLT DstType = MRI.getType(DstReg); |
5131 | // FIXME: add support for vector types |
5132 | if (DstType.isVector()) |
5133 | return UnableToLegalize; |
5134 | |
5135 | unsigned Opcode = MI.getOpcode(); |
5136 | unsigned OpO, OpE, OpF; |
5137 | switch (Opcode) { |
5138 | case TargetOpcode::G_SADDO: |
5139 | case TargetOpcode::G_SADDE: |
5140 | case TargetOpcode::G_UADDO: |
5141 | case TargetOpcode::G_UADDE: |
5142 | case TargetOpcode::G_ADD: |
5143 | OpO = TargetOpcode::G_UADDO; |
5144 | OpE = TargetOpcode::G_UADDE; |
5145 | OpF = TargetOpcode::G_UADDE; |
5146 | if (Opcode == TargetOpcode::G_SADDO || Opcode == TargetOpcode::G_SADDE) |
5147 | OpF = TargetOpcode::G_SADDE; |
5148 | break; |
5149 | case TargetOpcode::G_SSUBO: |
5150 | case TargetOpcode::G_SSUBE: |
5151 | case TargetOpcode::G_USUBO: |
5152 | case TargetOpcode::G_USUBE: |
5153 | case TargetOpcode::G_SUB: |
5154 | OpO = TargetOpcode::G_USUBO; |
5155 | OpE = TargetOpcode::G_USUBE; |
5156 | OpF = TargetOpcode::G_USUBE; |
5157 | if (Opcode == TargetOpcode::G_SSUBO || Opcode == TargetOpcode::G_SSUBE) |
5158 | OpF = TargetOpcode::G_SSUBE; |
5159 | break; |
5160 | default: |
5161 | llvm_unreachable("Unexpected add/sub opcode!")__builtin_unreachable(); |
5162 | } |
5163 | |
5164 | // 1 for a plain add/sub, 2 if this is an operation with a carry-out. |
5165 | unsigned NumDefs = MI.getNumExplicitDefs(); |
5166 | Register Src1 = MI.getOperand(NumDefs).getReg(); |
5167 | Register Src2 = MI.getOperand(NumDefs + 1).getReg(); |
5168 | Register CarryDst, CarryIn; |
5169 | if (NumDefs == 2) |
5170 | CarryDst = MI.getOperand(1).getReg(); |
5171 | if (MI.getNumOperands() == NumDefs + 3) |
5172 | CarryIn = MI.getOperand(NumDefs + 2).getReg(); |
5173 | |
5174 | LLT RegTy = MRI.getType(MI.getOperand(0).getReg()); |
5175 | LLT LeftoverTy, DummyTy; |
5176 | SmallVector<Register, 2> Src1Regs, Src2Regs, Src1Left, Src2Left, DstRegs; |
5177 | extractParts(Src1, RegTy, NarrowTy, LeftoverTy, Src1Regs, Src1Left); |
5178 | extractParts(Src2, RegTy, NarrowTy, DummyTy, Src2Regs, Src2Left); |
5179 | |
5180 | int NarrowParts = Src1Regs.size(); |
5181 | for (int I = 0, E = Src1Left.size(); I != E; ++I) { |
5182 | Src1Regs.push_back(Src1Left[I]); |
5183 | Src2Regs.push_back(Src2Left[I]); |
5184 | } |
5185 | DstRegs.reserve(Src1Regs.size()); |
5186 | |
5187 | for (int i = 0, e = Src1Regs.size(); i != e; ++i) { |
5188 | Register DstReg = |
5189 | MRI.createGenericVirtualRegister(MRI.getType(Src1Regs[i])); |
5190 | Register CarryOut = MRI.createGenericVirtualRegister(LLT::scalar(1)); |
5191 | // Forward the final carry-out to the destination register |
5192 | if (i == e - 1 && CarryDst) |
5193 | CarryOut = CarryDst; |
5194 | |
5195 | if (!CarryIn) { |
5196 | MIRBuilder.buildInstr(OpO, {DstReg, CarryOut}, |
5197 | {Src1Regs[i], Src2Regs[i]}); |
5198 | } else if (i == e - 1) { |
5199 | MIRBuilder.buildInstr(OpF, {DstReg, CarryOut}, |
5200 | {Src1Regs[i], Src2Regs[i], CarryIn}); |
5201 | } else { |
5202 | MIRBuilder.buildInstr(OpE, {DstReg, CarryOut}, |
5203 | {Src1Regs[i], Src2Regs[i], CarryIn}); |
5204 | } |
5205 | |
5206 | DstRegs.push_back(DstReg); |
5207 | CarryIn = CarryOut; |
5208 | } |
5209 | insertParts(MI.getOperand(0).getReg(), RegTy, NarrowTy, |
5210 | makeArrayRef(DstRegs).take_front(NarrowParts), LeftoverTy, |
5211 | makeArrayRef(DstRegs).drop_front(NarrowParts)); |
5212 | |
5213 | MI.eraseFromParent(); |
5214 | return Legalized; |
5215 | } |
5216 | |
5217 | LegalizerHelper::LegalizeResult |
5218 | LegalizerHelper::narrowScalarMul(MachineInstr &MI, LLT NarrowTy) { |
5219 | Register DstReg = MI.getOperand(0).getReg(); |
5220 | Register Src1 = MI.getOperand(1).getReg(); |
5221 | Register Src2 = MI.getOperand(2).getReg(); |
5222 | |
5223 | LLT Ty = MRI.getType(DstReg); |
5224 | if (Ty.isVector()) |
5225 | return UnableToLegalize; |
5226 | |
5227 | unsigned SrcSize = MRI.getType(Src1).getSizeInBits(); |
5228 | unsigned DstSize = Ty.getSizeInBits(); |
5229 | unsigned NarrowSize = NarrowTy.getSizeInBits(); |
5230 | if (DstSize % NarrowSize != 0 || SrcSize % NarrowSize != 0) |
5231 | return UnableToLegalize; |
5232 | |
5233 | unsigned NumDstParts = DstSize / NarrowSize; |
5234 | unsigned NumSrcParts = SrcSize / NarrowSize; |
5235 | bool IsMulHigh = MI.getOpcode() == TargetOpcode::G_UMULH; |
5236 | unsigned DstTmpParts = NumDstParts * (IsMulHigh ? 2 : 1); |
5237 | |
5238 | SmallVector<Register, 2> Src1Parts, Src2Parts; |
5239 | SmallVector<Register, 2> DstTmpRegs(DstTmpParts); |
5240 | extractParts(Src1, NarrowTy, NumSrcParts, Src1Parts); |
5241 | extractParts(Src2, NarrowTy, NumSrcParts, Src2Parts); |
5242 | multiplyRegisters(DstTmpRegs, Src1Parts, Src2Parts, NarrowTy); |
5243 | |
5244 | // Take only high half of registers if this is high mul. |
5245 | ArrayRef<Register> DstRegs( |
5246 | IsMulHigh ? &DstTmpRegs[DstTmpParts / 2] : &DstTmpRegs[0], NumDstParts); |
5247 | MIRBuilder.buildMerge(DstReg, DstRegs); |
5248 | MI.eraseFromParent(); |
5249 | return Legalized; |
5250 | } |
5251 | |
5252 | LegalizerHelper::LegalizeResult |
5253 | LegalizerHelper::narrowScalarFPTOI(MachineInstr &MI, unsigned TypeIdx, |
5254 | LLT NarrowTy) { |
5255 | if (TypeIdx != 0) |
5256 | return UnableToLegalize; |
5257 | |
5258 | bool IsSigned = MI.getOpcode() == TargetOpcode::G_FPTOSI; |
5259 | |
5260 | Register Src = MI.getOperand(1).getReg(); |
5261 | LLT SrcTy = MRI.getType(Src); |
5262 | |
5263 | // If all finite floats fit into the narrowed integer type, we can just swap |
5264 | // out the result type. This is practically only useful for conversions from |
5265 | // half to at least 16-bits, so just handle the one case. |
5266 | if (SrcTy.getScalarType() != LLT::scalar(16) || |
5267 | NarrowTy.getScalarSizeInBits() < (IsSigned ? 17u : 16u)) |
5268 | return UnableToLegalize; |
5269 | |
5270 | Observer.changingInstr(MI); |
5271 | narrowScalarDst(MI, NarrowTy, 0, |
5272 | IsSigned ? TargetOpcode::G_SEXT : TargetOpcode::G_ZEXT); |
5273 | Observer.changedInstr(MI); |
5274 | return Legalized; |
5275 | } |
5276 | |
5277 | LegalizerHelper::LegalizeResult |
5278 | LegalizerHelper::narrowScalarExtract(MachineInstr &MI, unsigned TypeIdx, |
5279 | LLT NarrowTy) { |
5280 | if (TypeIdx != 1) |
5281 | return UnableToLegalize; |
5282 | |
5283 | uint64_t NarrowSize = NarrowTy.getSizeInBits(); |
5284 | |
5285 | int64_t SizeOp1 = MRI.getType(MI.getOperand(1).getReg()).getSizeInBits(); |
5286 | // FIXME: add support for when SizeOp1 isn't an exact multiple of |
5287 | // NarrowSize. |
5288 | if (SizeOp1 % NarrowSize != 0) |
5289 | return UnableToLegalize; |
5290 | int NumParts = SizeOp1 / NarrowSize; |
5291 | |
5292 | SmallVector<Register, 2> SrcRegs, DstRegs; |
5293 | SmallVector<uint64_t, 2> Indexes; |
5294 | extractParts(MI.getOperand(1).getReg(), NarrowTy, NumParts, SrcRegs); |
5295 | |
5296 | Register OpReg = MI.getOperand(0).getReg(); |
5297 | uint64_t OpStart = MI.getOperand(2).getImm(); |
5298 | uint64_t OpSize = MRI.getType(OpReg).getSizeInBits(); |
5299 | for (int i = 0; i < NumParts; ++i) { |
5300 | unsigned SrcStart = i * NarrowSize; |
5301 | |
5302 | if (SrcStart + NarrowSize <= OpStart || SrcStart >= OpStart + OpSize) { |
5303 | // No part of the extract uses this subregister, ignore it. |
5304 | continue; |
5305 | } else if (SrcStart == OpStart && NarrowTy == MRI.getType(OpReg)) { |
5306 | // The entire subregister is extracted, forward the value. |
5307 | DstRegs.push_back(SrcRegs[i]); |
5308 | continue; |
5309 | } |
5310 | |
5311 | // OpSegStart is where this destination segment would start in OpReg if it |
5312 | // extended infinitely in both directions. |
5313 | int64_t ExtractOffset; |
5314 | uint64_t SegSize; |
5315 | if (OpStart < SrcStart) { |
5316 | ExtractOffset = 0; |
5317 | SegSize = std::min(NarrowSize, OpStart + OpSize - SrcStart); |
5318 | } else { |
5319 | ExtractOffset = OpStart - SrcStart; |
5320 | SegSize = std::min(SrcStart + NarrowSize - OpStart, OpSize); |
5321 | } |
5322 | |
5323 | Register SegReg = SrcRegs[i]; |
5324 | if (ExtractOffset != 0 || SegSize != NarrowSize) { |
5325 | // A genuine extract is needed. |
5326 | SegReg = MRI.createGenericVirtualRegister(LLT::scalar(SegSize)); |
5327 | MIRBuilder.buildExtract(SegReg, SrcRegs[i], ExtractOffset); |
5328 | } |
5329 | |
5330 | DstRegs.push_back(SegReg); |
5331 | } |
5332 | |
5333 | Register DstReg = MI.getOperand(0).getReg(); |
5334 | if (MRI.getType(DstReg).isVector()) |
5335 | MIRBuilder.buildBuildVector(DstReg, DstRegs); |
5336 | else if (DstRegs.size() > 1) |
5337 | MIRBuilder.buildMerge(DstReg, DstRegs); |
5338 | else |
5339 | MIRBuilder.buildCopy(DstReg, DstRegs[0]); |
5340 | MI.eraseFromParent(); |
5341 | return Legalized; |
5342 | } |
5343 | |
5344 | LegalizerHelper::LegalizeResult |
5345 | LegalizerHelper::narrowScalarInsert(MachineInstr &MI, unsigned TypeIdx, |
5346 | LLT NarrowTy) { |
5347 | // FIXME: Don't know how to handle secondary types yet. |
5348 | if (TypeIdx != 0) |
5349 | return UnableToLegalize; |
5350 | |
5351 | SmallVector<Register, 2> SrcRegs, LeftoverRegs, DstRegs; |
5352 | SmallVector<uint64_t, 2> Indexes; |
5353 | LLT RegTy = MRI.getType(MI.getOperand(0).getReg()); |
5354 | LLT LeftoverTy; |
5355 | extractParts(MI.getOperand(1).getReg(), RegTy, NarrowTy, LeftoverTy, SrcRegs, |
5356 | LeftoverRegs); |
5357 | |
5358 | for (Register Reg : LeftoverRegs) |
5359 | SrcRegs.push_back(Reg); |
5360 | |
5361 | uint64_t NarrowSize = NarrowTy.getSizeInBits(); |
5362 | Register OpReg = MI.getOperand(2).getReg(); |
5363 | uint64_t OpStart = MI.getOperand(3).getImm(); |
5364 | uint64_t OpSize = MRI.getType(OpReg).getSizeInBits(); |
5365 | for (int I = 0, E = SrcRegs.size(); I != E; ++I) { |
5366 | unsigned DstStart = I * NarrowSize; |
5367 | |
5368 | if (DstStart == OpStart && NarrowTy == MRI.getType(OpReg)) { |
5369 | // The entire subregister is defined by this insert, forward the new |
5370 | // value. |
5371 | DstRegs.push_back(OpReg); |
5372 | continue; |
5373 | } |
5374 | |
5375 | Register SrcReg = SrcRegs[I]; |
5376 | if (MRI.getType(SrcRegs[I]) == LeftoverTy) { |
5377 | // The leftover reg is smaller than NarrowTy, so we need to extend it. |
5378 | SrcReg = MRI.createGenericVirtualRegister(NarrowTy); |
5379 | MIRBuilder.buildAnyExt(SrcReg, SrcRegs[I]); |
5380 | } |
5381 | |
5382 | if (DstStart + NarrowSize <= OpStart || DstStart >= OpStart + OpSize) { |
5383 | // No part of the insert affects this subregister, forward the original. |
5384 | DstRegs.push_back(SrcReg); |
5385 | continue; |
5386 | } |
5387 | |
5388 | // OpSegStart is where this destination segment would start in OpReg if it |
5389 | // extended infinitely in both directions. |
5390 | int64_t ExtractOffset, InsertOffset; |
5391 | uint64_t SegSize; |
5392 | if (OpStart < DstStart) { |
5393 | InsertOffset = 0; |
5394 | ExtractOffset = DstStart - OpStart; |
5395 | SegSize = std::min(NarrowSize, OpStart + OpSize - DstStart); |
5396 | } else { |
5397 | InsertOffset = OpStart - DstStart; |
5398 | ExtractOffset = 0; |
5399 | SegSize = |
5400 | std::min(NarrowSize - InsertOffset, OpStart + OpSize - DstStart); |
5401 | } |
5402 | |
5403 | Register SegReg = OpReg; |
5404 | if (ExtractOffset != 0 || SegSize != OpSize) { |
5405 | // A genuine extract is needed. |
5406 | SegReg = MRI.createGenericVirtualRegister(LLT::scalar(SegSize)); |
5407 | MIRBuilder.buildExtract(SegReg, OpReg, ExtractOffset); |
5408 | } |
5409 | |
5410 | Register DstReg = MRI.createGenericVirtualRegister(NarrowTy); |
5411 | MIRBuilder.buildInsert(DstReg, SrcReg, SegReg, InsertOffset); |
5412 | DstRegs.push_back(DstReg); |
5413 | } |
5414 | |
5415 | uint64_t WideSize = DstRegs.size() * NarrowSize; |
5416 | Register DstReg = MI.getOperand(0).getReg(); |
5417 | if (WideSize > RegTy.getSizeInBits()) { |
5418 | Register MergeReg = MRI.createGenericVirtualRegister(LLT::scalar(WideSize)); |
5419 | MIRBuilder.buildMerge(MergeReg, DstRegs); |
5420 | MIRBuilder.buildTrunc(DstReg, MergeReg); |
5421 | } else |
5422 | MIRBuilder.buildMerge(DstReg, DstRegs); |
5423 | |
5424 | MI.eraseFromParent(); |
5425 | return Legalized; |
5426 | } |
5427 | |
5428 | LegalizerHelper::LegalizeResult |
5429 | LegalizerHelper::narrowScalarBasic(MachineInstr &MI, unsigned TypeIdx, |
5430 | LLT NarrowTy) { |
5431 | Register DstReg = MI.getOperand(0).getReg(); |
5432 | LLT DstTy = MRI.getType(DstReg); |
5433 | |
5434 | assert(MI.getNumOperands() == 3 && TypeIdx == 0)((void)0); |
5435 | |
5436 | SmallVector<Register, 4> DstRegs, DstLeftoverRegs; |
5437 | SmallVector<Register, 4> Src0Regs, Src0LeftoverRegs; |
5438 | SmallVector<Register, 4> Src1Regs, Src1LeftoverRegs; |
5439 | LLT LeftoverTy; |
5440 | if (!extractParts(MI.getOperand(1).getReg(), DstTy, NarrowTy, LeftoverTy, |
5441 | Src0Regs, Src0LeftoverRegs)) |
5442 | return UnableToLegalize; |
5443 | |
5444 | LLT Unused; |
5445 | if (!extractParts(MI.getOperand(2).getReg(), DstTy, NarrowTy, Unused, |
5446 | Src1Regs, Src1LeftoverRegs)) |
5447 | llvm_unreachable("inconsistent extractParts result")__builtin_unreachable(); |
5448 | |
5449 | for (unsigned I = 0, E = Src1Regs.size(); I != E; ++I) { |
5450 | auto Inst = MIRBuilder.buildInstr(MI.getOpcode(), {NarrowTy}, |
5451 | {Src0Regs[I], Src1Regs[I]}); |
5452 | DstRegs.push_back(Inst.getReg(0)); |
5453 | } |
5454 | |
5455 | for (unsigned I = 0, E = Src1LeftoverRegs.size(); I != E; ++I) { |
5456 | auto Inst = MIRBuilder.buildInstr( |
5457 | MI.getOpcode(), |
5458 | {LeftoverTy}, {Src0LeftoverRegs[I], Src1LeftoverRegs[I]}); |
5459 | DstLeftoverRegs.push_back(Inst.getReg(0)); |
5460 | } |
5461 | |
5462 | insertParts(DstReg, DstTy, NarrowTy, DstRegs, |
5463 | LeftoverTy, DstLeftoverRegs); |
5464 | |
5465 | MI.eraseFromParent(); |
5466 | return Legalized; |
5467 | } |
5468 | |
5469 | LegalizerHelper::LegalizeResult |
5470 | LegalizerHelper::narrowScalarExt(MachineInstr &MI, unsigned TypeIdx, |
5471 | LLT NarrowTy) { |
5472 | if (TypeIdx != 0) |
5473 | return UnableToLegalize; |
5474 | |
5475 | Register DstReg = MI.getOperand(0).getReg(); |
5476 | Register SrcReg = MI.getOperand(1).getReg(); |
5477 | |
5478 | LLT DstTy = MRI.getType(DstReg); |
5479 | if (DstTy.isVector()) |
5480 | return UnableToLegalize; |
5481 | |
5482 | SmallVector<Register, 8> Parts; |
5483 | LLT GCDTy = extractGCDType(Parts, DstTy, NarrowTy, SrcReg); |
5484 | LLT LCMTy = buildLCMMergePieces(DstTy, NarrowTy, GCDTy, Parts, MI.getOpcode()); |
5485 | buildWidenedRemergeToDst(DstReg, LCMTy, Parts); |
5486 | |
5487 | MI.eraseFromParent(); |
5488 | return Legalized; |
5489 | } |
5490 | |
5491 | LegalizerHelper::LegalizeResult |
5492 | LegalizerHelper::narrowScalarSelect(MachineInstr &MI, unsigned TypeIdx, |
5493 | LLT NarrowTy) { |
5494 | if (TypeIdx != 0) |
5495 | return UnableToLegalize; |
5496 | |
5497 | Register CondReg = MI.getOperand(1).getReg(); |
5498 | LLT CondTy = MRI.getType(CondReg); |
5499 | if (CondTy.isVector()) // TODO: Handle vselect |
5500 | return UnableToLegalize; |
5501 | |
5502 | Register DstReg = MI.getOperand(0).getReg(); |
5503 | LLT DstTy = MRI.getType(DstReg); |
5504 | |
5505 | SmallVector<Register, 4> DstRegs, DstLeftoverRegs; |
5506 | SmallVector<Register, 4> Src1Regs, Src1LeftoverRegs; |
5507 | SmallVector<Register, 4> Src2Regs, Src2LeftoverRegs; |
5508 | LLT LeftoverTy; |
5509 | if (!extractParts(MI.getOperand(2).getReg(), DstTy, NarrowTy, LeftoverTy, |
5510 | Src1Regs, Src1LeftoverRegs)) |
5511 | return UnableToLegalize; |
5512 | |
5513 | LLT Unused; |
5514 | if (!extractParts(MI.getOperand(3).getReg(), DstTy, NarrowTy, Unused, |
5515 | Src2Regs, Src2LeftoverRegs)) |
5516 | llvm_unreachable("inconsistent extractParts result")__builtin_unreachable(); |
5517 | |
5518 | for (unsigned I = 0, E = Src1Regs.size(); I != E; ++I) { |
5519 | auto Select = MIRBuilder.buildSelect(NarrowTy, |
5520 | CondReg, Src1Regs[I], Src2Regs[I]); |
5521 | DstRegs.push_back(Select.getReg(0)); |
5522 | } |
5523 | |
5524 | for (unsigned I = 0, E = Src1LeftoverRegs.size(); I != E; ++I) { |
5525 | auto Select = MIRBuilder.buildSelect( |
5526 | LeftoverTy, CondReg, Src1LeftoverRegs[I], Src2LeftoverRegs[I]); |
5527 | DstLeftoverRegs.push_back(Select.getReg(0)); |
5528 | } |
5529 | |
5530 | insertParts(DstReg, DstTy, NarrowTy, DstRegs, |
5531 | LeftoverTy, DstLeftoverRegs); |
5532 | |
5533 | MI.eraseFromParent(); |
5534 | return Legalized; |
5535 | } |
5536 | |
5537 | LegalizerHelper::LegalizeResult |
5538 | LegalizerHelper::narrowScalarCTLZ(MachineInstr &MI, unsigned TypeIdx, |
5539 | LLT NarrowTy) { |
5540 | if (TypeIdx != 1) |
5541 | return UnableToLegalize; |
5542 | |
5543 | Register DstReg = MI.getOperand(0).getReg(); |
5544 | Register SrcReg = MI.getOperand(1).getReg(); |
5545 | LLT DstTy = MRI.getType(DstReg); |
5546 | LLT SrcTy = MRI.getType(SrcReg); |
5547 | unsigned NarrowSize = NarrowTy.getSizeInBits(); |
5548 | |
5549 | if (SrcTy.isScalar() && SrcTy.getSizeInBits() == 2 * NarrowSize) { |
5550 | const bool IsUndef = MI.getOpcode() == TargetOpcode::G_CTLZ_ZERO_UNDEF; |
5551 | |
5552 | MachineIRBuilder &B = MIRBuilder; |
5553 | auto UnmergeSrc = B.buildUnmerge(NarrowTy, SrcReg); |
5554 | // ctlz(Hi:Lo) -> Hi == 0 ? (NarrowSize + ctlz(Lo)) : ctlz(Hi) |
5555 | auto C_0 = B.buildConstant(NarrowTy, 0); |
5556 | auto HiIsZero = B.buildICmp(CmpInst::ICMP_EQ, LLT::scalar(1), |
5557 | UnmergeSrc.getReg(1), C_0); |
5558 | auto LoCTLZ = IsUndef ? |
5559 | B.buildCTLZ_ZERO_UNDEF(DstTy, UnmergeSrc.getReg(0)) : |
5560 | B.buildCTLZ(DstTy, UnmergeSrc.getReg(0)); |
5561 | auto C_NarrowSize = B.buildConstant(DstTy, NarrowSize); |
5562 | auto HiIsZeroCTLZ = B.buildAdd(DstTy, LoCTLZ, C_NarrowSize); |
5563 | auto HiCTLZ = B.buildCTLZ_ZERO_UNDEF(DstTy, UnmergeSrc.getReg(1)); |
5564 | B.buildSelect(DstReg, HiIsZero, HiIsZeroCTLZ, HiCTLZ); |
5565 | |
5566 | MI.eraseFromParent(); |
5567 | return Legalized; |
5568 | } |
5569 | |
5570 | return UnableToLegalize; |
5571 | } |
5572 | |
5573 | LegalizerHelper::LegalizeResult |
5574 | LegalizerHelper::narrowScalarCTTZ(MachineInstr &MI, unsigned TypeIdx, |
5575 | LLT NarrowTy) { |
5576 | if (TypeIdx != 1) |
5577 | return UnableToLegalize; |
5578 | |
5579 | Register DstReg = MI.getOperand(0).getReg(); |
5580 | Register SrcReg = MI.getOperand(1).getReg(); |
5581 | LLT DstTy = MRI.getType(DstReg); |
5582 | LLT SrcTy = MRI.getType(SrcReg); |
5583 | unsigned NarrowSize = NarrowTy.getSizeInBits(); |
5584 | |
5585 | if (SrcTy.isScalar() && SrcTy.getSizeInBits() == 2 * NarrowSize) { |
5586 | const bool IsUndef = MI.getOpcode() == TargetOpcode::G_CTTZ_ZERO_UNDEF; |
5587 | |
5588 | MachineIRBuilder &B = MIRBuilder; |
5589 | auto UnmergeSrc = B.buildUnmerge(NarrowTy, SrcReg); |
5590 | // cttz(Hi:Lo) -> Lo == 0 ? (cttz(Hi) + NarrowSize) : cttz(Lo) |
5591 | auto C_0 = B.buildConstant(NarrowTy, 0); |
5592 | auto LoIsZero = B.buildICmp(CmpInst::ICMP_EQ, LLT::scalar(1), |
5593 | UnmergeSrc.getReg(0), C_0); |
5594 | auto HiCTTZ = IsUndef ? |
5595 | B.buildCTTZ_ZERO_UNDEF(DstTy, UnmergeSrc.getReg(1)) : |
5596 | B.buildCTTZ(DstTy, UnmergeSrc.getReg(1)); |
5597 | auto C_NarrowSize = B.buildConstant(DstTy, NarrowSize); |
5598 | auto LoIsZeroCTTZ = B.buildAdd(DstTy, HiCTTZ, C_NarrowSize); |
5599 | auto LoCTTZ = B.buildCTTZ_ZERO_UNDEF(DstTy, UnmergeSrc.getReg(0)); |
5600 | B.buildSelect(DstReg, LoIsZero, LoIsZeroCTTZ, LoCTTZ); |
5601 | |
5602 | MI.eraseFromParent(); |
5603 | return Legalized; |
5604 | } |
5605 | |
5606 | return UnableToLegalize; |
5607 | } |
5608 | |
5609 | LegalizerHelper::LegalizeResult |
5610 | LegalizerHelper::narrowScalarCTPOP(MachineInstr &MI, unsigned TypeIdx, |
5611 | LLT NarrowTy) { |
5612 | if (TypeIdx != 1) |
5613 | return UnableToLegalize; |
5614 | |
5615 | Register DstReg = MI.getOperand(0).getReg(); |
5616 | LLT DstTy = MRI.getType(DstReg); |
5617 | LLT SrcTy = MRI.getType(MI.getOperand(1).getReg()); |
5618 | unsigned NarrowSize = NarrowTy.getSizeInBits(); |
5619 | |
5620 | if (SrcTy.isScalar() && SrcTy.getSizeInBits() == 2 * NarrowSize) { |
5621 | auto UnmergeSrc = MIRBuilder.buildUnmerge(NarrowTy, MI.getOperand(1)); |
5622 | |
5623 | auto LoCTPOP = MIRBuilder.buildCTPOP(DstTy, UnmergeSrc.getReg(0)); |
5624 | auto HiCTPOP = MIRBuilder.buildCTPOP(DstTy, UnmergeSrc.getReg(1)); |
5625 | MIRBuilder.buildAdd(DstReg, HiCTPOP, LoCTPOP); |
5626 | |
5627 | MI.eraseFromParent(); |
5628 | return Legalized; |
5629 | } |
5630 | |
5631 | return UnableToLegalize; |
5632 | } |
5633 | |
5634 | LegalizerHelper::LegalizeResult |
5635 | LegalizerHelper::lowerBitCount(MachineInstr &MI) { |
5636 | unsigned Opc = MI.getOpcode(); |
5637 | const auto &TII = MIRBuilder.getTII(); |
5638 | auto isSupported = [this](const LegalityQuery &Q) { |
5639 | auto QAction = LI.getAction(Q).Action; |
5640 | return QAction == Legal || QAction == Libcall || QAction == Custom; |
5641 | }; |
5642 | switch (Opc) { |
5643 | default: |
5644 | return UnableToLegalize; |
5645 | case TargetOpcode::G_CTLZ_ZERO_UNDEF: { |
5646 | // This trivially expands to CTLZ. |
5647 | Observer.changingInstr(MI); |
5648 | MI.setDesc(TII.get(TargetOpcode::G_CTLZ)); |
5649 | Observer.changedInstr(MI); |
5650 | return Legalized; |
5651 | } |
5652 | case TargetOpcode::G_CTLZ: { |
5653 | Register DstReg = MI.getOperand(0).getReg(); |
5654 | Register SrcReg = MI.getOperand(1).getReg(); |
5655 | LLT DstTy = MRI.getType(DstReg); |
5656 | LLT SrcTy = MRI.getType(SrcReg); |
5657 | unsigned Len = SrcTy.getSizeInBits(); |
5658 | |
5659 | if (isSupported({TargetOpcode::G_CTLZ_ZERO_UNDEF, {DstTy, SrcTy}})) { |
5660 | // If CTLZ_ZERO_UNDEF is supported, emit that and a select for zero. |
5661 | auto CtlzZU = MIRBuilder.buildCTLZ_ZERO_UNDEF(DstTy, SrcReg); |
5662 | auto ZeroSrc = MIRBuilder.buildConstant(SrcTy, 0); |
5663 | auto ICmp = MIRBuilder.buildICmp( |
5664 | CmpInst::ICMP_EQ, SrcTy.changeElementSize(1), SrcReg, ZeroSrc); |
5665 | auto LenConst = MIRBuilder.buildConstant(DstTy, Len); |
5666 | MIRBuilder.buildSelect(DstReg, ICmp, LenConst, CtlzZU); |
5667 | MI.eraseFromParent(); |
5668 | return Legalized; |
5669 | } |
5670 | // for now, we do this: |
5671 | // NewLen = NextPowerOf2(Len); |
5672 | // x = x | (x >> 1); |
5673 | // x = x | (x >> 2); |
5674 | // ... |
5675 | // x = x | (x >>16); |
5676 | // x = x | (x >>32); // for 64-bit input |
5677 | // Upto NewLen/2 |
5678 | // return Len - popcount(x); |
5679 | // |
5680 | // Ref: "Hacker's Delight" by Henry Warren |
5681 | Register Op = SrcReg; |
5682 | unsigned NewLen = PowerOf2Ceil(Len); |
5683 | for (unsigned i = 0; (1U << i) <= (NewLen / 2); ++i) { |
5684 | auto MIBShiftAmt = MIRBuilder.buildConstant(SrcTy, 1ULL << i); |
5685 | auto MIBOp = MIRBuilder.buildOr( |
5686 | SrcTy, Op, MIRBuilder.buildLShr(SrcTy, Op, MIBShiftAmt)); |
5687 | Op = MIBOp.getReg(0); |
5688 | } |
5689 | auto MIBPop = MIRBuilder.buildCTPOP(DstTy, Op); |
5690 | MIRBuilder.buildSub(MI.getOperand(0), MIRBuilder.buildConstant(DstTy, Len), |
5691 | MIBPop); |
5692 | MI.eraseFromParent(); |
5693 | return Legalized; |
5694 | } |
5695 | case TargetOpcode::G_CTTZ_ZERO_UNDEF: { |
5696 | // This trivially expands to CTTZ. |
5697 | Observer.changingInstr(MI); |
5698 | MI.setDesc(TII.get(TargetOpcode::G_CTTZ)); |
5699 | Observer.changedInstr(MI); |
5700 | return Legalized; |
5701 | } |
5702 | case TargetOpcode::G_CTTZ: { |
5703 | Register DstReg = MI.getOperand(0).getReg(); |
5704 | Register SrcReg = MI.getOperand(1).getReg(); |
5705 | LLT DstTy = MRI.getType(DstReg); |
5706 | LLT SrcTy = MRI.getType(SrcReg); |
5707 | |
5708 | unsigned Len = SrcTy.getSizeInBits(); |
5709 | if (isSupported({TargetOpcode::G_CTTZ_ZERO_UNDEF, {DstTy, SrcTy}})) { |
5710 | // If CTTZ_ZERO_UNDEF is legal or custom, emit that and a select with |
5711 | // zero. |
5712 | auto CttzZU = MIRBuilder.buildCTTZ_ZERO_UNDEF(DstTy, SrcReg); |
5713 | auto Zero = MIRBuilder.buildConstant(SrcTy, 0); |
5714 | auto ICmp = MIRBuilder.buildICmp( |
5715 | CmpInst::ICMP_EQ, DstTy.changeElementSize(1), SrcReg, Zero); |
5716 | auto LenConst = MIRBuilder.buildConstant(DstTy, Len); |
5717 | MIRBuilder.buildSelect(DstReg, ICmp, LenConst, CttzZU); |
5718 | MI.eraseFromParent(); |
5719 | return Legalized; |
5720 | } |
5721 | // for now, we use: { return popcount(~x & (x - 1)); } |
5722 | // unless the target has ctlz but not ctpop, in which case we use: |
5723 | // { return 32 - nlz(~x & (x-1)); } |
5724 | // Ref: "Hacker's Delight" by Henry Warren |
5725 | auto MIBCstNeg1 = MIRBuilder.buildConstant(SrcTy, -1); |
5726 | auto MIBNot = MIRBuilder.buildXor(SrcTy, SrcReg, MIBCstNeg1); |
5727 | auto MIBTmp = MIRBuilder.buildAnd( |
5728 | SrcTy, MIBNot, MIRBuilder.buildAdd(SrcTy, SrcReg, MIBCstNeg1)); |
5729 | if (!isSupported({TargetOpcode::G_CTPOP, {SrcTy, SrcTy}}) && |
5730 | isSupported({TargetOpcode::G_CTLZ, {SrcTy, SrcTy}})) { |
5731 | auto MIBCstLen = MIRBuilder.buildConstant(SrcTy, Len); |
5732 | MIRBuilder.buildSub(MI.getOperand(0), MIBCstLen, |
5733 | MIRBuilder.buildCTLZ(SrcTy, MIBTmp)); |
5734 | MI.eraseFromParent(); |
5735 | return Legalized; |
5736 | } |
5737 | MI.setDesc(TII.get(TargetOpcode::G_CTPOP)); |
5738 | MI.getOperand(1).setReg(MIBTmp.getReg(0)); |
5739 | return Legalized; |
5740 | } |
5741 | case TargetOpcode::G_CTPOP: { |
5742 | Register SrcReg = MI.getOperand(1).getReg(); |
5743 | LLT Ty = MRI.getType(SrcReg); |
5744 | unsigned Size = Ty.getSizeInBits(); |
5745 | MachineIRBuilder &B = MIRBuilder; |
5746 | |
5747 | // Count set bits in blocks of 2 bits. Default approach would be |
5748 | // B2Count = { val & 0x55555555 } + { (val >> 1) & 0x55555555 } |
5749 | // We use following formula instead: |
5750 | // B2Count = val - { (val >> 1) & 0x55555555 } |
5751 | // since it gives same result in blocks of 2 with one instruction less. |
5752 | auto C_1 = B.buildConstant(Ty, 1); |
5753 | auto B2Set1LoTo1Hi = B.buildLShr(Ty, SrcReg, C_1); |
5754 | APInt B2Mask1HiTo0 = APInt::getSplat(Size, APInt(8, 0x55)); |
5755 | auto C_B2Mask1HiTo0 = B.buildConstant(Ty, B2Mask1HiTo0); |
5756 | auto B2Count1Hi = B.buildAnd(Ty, B2Set1LoTo1Hi, C_B2Mask1HiTo0); |
5757 | auto B2Count = B.buildSub(Ty, SrcReg, B2Count1Hi); |
5758 | |
5759 | // In order to get count in blocks of 4 add values from adjacent block of 2. |
5760 | // B4Count = { B2Count & 0x33333333 } + { (B2Count >> 2) & 0x33333333 } |
5761 | auto C_2 = B.buildConstant(Ty, 2); |
5762 | auto B4Set2LoTo2Hi = B.buildLShr(Ty, B2Count, C_2); |
5763 | APInt B4Mask2HiTo0 = APInt::getSplat(Size, APInt(8, 0x33)); |
5764 | auto C_B4Mask2HiTo0 = B.buildConstant(Ty, B4Mask2HiTo0); |
5765 | auto B4HiB2Count = B.buildAnd(Ty, B4Set2LoTo2Hi, C_B4Mask2HiTo0); |
5766 | auto B4LoB2Count = B.buildAnd(Ty, B2Count, C_B4Mask2HiTo0); |
5767 | auto B4Count = B.buildAdd(Ty, B4HiB2Count, B4LoB2Count); |
5768 | |
5769 | // For count in blocks of 8 bits we don't have to mask high 4 bits before |
5770 | // addition since count value sits in range {0,...,8} and 4 bits are enough |
5771 | // to hold such binary values. After addition high 4 bits still hold count |
5772 | // of set bits in high 4 bit block, set them to zero and get 8 bit result. |
5773 | // B8Count = { B4Count + (B4Count >> 4) } & 0x0F0F0F0F |
5774 | auto C_4 = B.buildConstant(Ty, 4); |
5775 | auto B8HiB4Count = B.buildLShr(Ty, B4Count, C_4); |
5776 | auto B8CountDirty4Hi = B.buildAdd(Ty, B8HiB4Count, B4Count); |
5777 | APInt B8Mask4HiTo0 = APInt::getSplat(Size, APInt(8, 0x0F)); |
5778 | auto C_B8Mask4HiTo0 = B.buildConstant(Ty, B8Mask4HiTo0); |
5779 | auto B8Count = B.buildAnd(Ty, B8CountDirty4Hi, C_B8Mask4HiTo0); |
5780 | |
5781 | assert(Size<=128 && "Scalar size is too large for CTPOP lower algorithm")((void)0); |
5782 | // 8 bits can hold CTPOP result of 128 bit int or smaller. Mul with this |
5783 | // bitmask will set 8 msb in ResTmp to sum of all B8Counts in 8 bit blocks. |
5784 | auto MulMask = B.buildConstant(Ty, APInt::getSplat(Size, APInt(8, 0x01))); |
5785 | auto ResTmp = B.buildMul(Ty, B8Count, MulMask); |
5786 | |
5787 | // Shift count result from 8 high bits to low bits. |
5788 | auto C_SizeM8 = B.buildConstant(Ty, Size - 8); |
5789 | B.buildLShr(MI.getOperand(0).getReg(), ResTmp, C_SizeM8); |
5790 | |
5791 | MI.eraseFromParent(); |
5792 | return Legalized; |
5793 | } |
5794 | } |
5795 | } |
5796 | |
5797 | // Check that (every element of) Reg is undef or not an exact multiple of BW. |
5798 | static bool isNonZeroModBitWidthOrUndef(const MachineRegisterInfo &MRI, |
5799 | Register Reg, unsigned BW) { |
5800 | return matchUnaryPredicate( |
5801 | MRI, Reg, |
5802 | [=](const Constant *C) { |
5803 | // Null constant here means an undef. |
5804 | const ConstantInt *CI = dyn_cast_or_null<ConstantInt>(C); |
5805 | return !CI || CI->getValue().urem(BW) != 0; |
5806 | }, |
5807 | /*AllowUndefs*/ true); |
5808 | } |
5809 | |
5810 | LegalizerHelper::LegalizeResult |
5811 | LegalizerHelper::lowerFunnelShiftWithInverse(MachineInstr &MI) { |
5812 | Register Dst = MI.getOperand(0).getReg(); |
5813 | Register X = MI.getOperand(1).getReg(); |
5814 | Register Y = MI.getOperand(2).getReg(); |
5815 | Register Z = MI.getOperand(3).getReg(); |
5816 | LLT Ty = MRI.getType(Dst); |
5817 | LLT ShTy = MRI.getType(Z); |
5818 | |
5819 | unsigned BW = Ty.getScalarSizeInBits(); |
5820 | |
5821 | if (!isPowerOf2_32(BW)) |
5822 | return UnableToLegalize; |
5823 | |
5824 | const bool IsFSHL = MI.getOpcode() == TargetOpcode::G_FSHL; |
5825 | unsigned RevOpcode = IsFSHL ? TargetOpcode::G_FSHR : TargetOpcode::G_FSHL; |
5826 | |
5827 | if (isNonZeroModBitWidthOrUndef(MRI, Z, BW)) { |
5828 | // fshl X, Y, Z -> fshr X, Y, -Z |
5829 | // fshr X, Y, Z -> fshl X, Y, -Z |
5830 | auto Zero = MIRBuilder.buildConstant(ShTy, 0); |
5831 | Z = MIRBuilder.buildSub(Ty, Zero, Z).getReg(0); |
5832 | } else { |
5833 | // fshl X, Y, Z -> fshr (srl X, 1), (fshr X, Y, 1), ~Z |
5834 | // fshr X, Y, Z -> fshl (fshl X, Y, 1), (shl Y, 1), ~Z |
5835 | auto One = MIRBuilder.buildConstant(ShTy, 1); |
5836 | if (IsFSHL) { |
5837 | Y = MIRBuilder.buildInstr(RevOpcode, {Ty}, {X, Y, One}).getReg(0); |
5838 | X = MIRBuilder.buildLShr(Ty, X, One).getReg(0); |
5839 | } else { |
5840 | X = MIRBuilder.buildInstr(RevOpcode, {Ty}, {X, Y, One}).getReg(0); |
5841 | Y = MIRBuilder.buildShl(Ty, Y, One).getReg(0); |
5842 | } |
5843 | |
5844 | Z = MIRBuilder.buildNot(ShTy, Z).getReg(0); |
5845 | } |
5846 | |
5847 | MIRBuilder.buildInstr(RevOpcode, {Dst}, {X, Y, Z}); |
5848 | MI.eraseFromParent(); |
5849 | return Legalized; |
5850 | } |
5851 | |
5852 | LegalizerHelper::LegalizeResult |
5853 | LegalizerHelper::lowerFunnelShiftAsShifts(MachineInstr &MI) { |
5854 | Register Dst = MI.getOperand(0).getReg(); |
5855 | Register X = MI.getOperand(1).getReg(); |
5856 | Register Y = MI.getOperand(2).getReg(); |
5857 | Register Z = MI.getOperand(3).getReg(); |
5858 | LLT Ty = MRI.getType(Dst); |
5859 | LLT ShTy = MRI.getType(Z); |
5860 | |
5861 | const unsigned BW = Ty.getScalarSizeInBits(); |
5862 | const bool IsFSHL = MI.getOpcode() == TargetOpcode::G_FSHL; |
5863 | |
5864 | Register ShX, ShY; |
5865 | Register ShAmt, InvShAmt; |
5866 | |
5867 | // FIXME: Emit optimized urem by constant instead of letting it expand later. |
5868 | if (isNonZeroModBitWidthOrUndef(MRI, Z, BW)) { |
5869 | // fshl: X << C | Y >> (BW - C) |
5870 | // fshr: X << (BW - C) | Y >> C |
5871 | // where C = Z % BW is not zero |
5872 | auto BitWidthC = MIRBuilder.buildConstant(ShTy, BW); |
5873 | ShAmt = MIRBuilder.buildURem(ShTy, Z, BitWidthC).getReg(0); |
5874 | InvShAmt = MIRBuilder.buildSub(ShTy, BitWidthC, ShAmt).getReg(0); |
5875 | ShX = MIRBuilder.buildShl(Ty, X, IsFSHL ? ShAmt : InvShAmt).getReg(0); |
5876 | ShY = MIRBuilder.buildLShr(Ty, Y, IsFSHL ? InvShAmt : ShAmt).getReg(0); |
5877 | } else { |
5878 | // fshl: X << (Z % BW) | Y >> 1 >> (BW - 1 - (Z % BW)) |
5879 | // fshr: X << 1 << (BW - 1 - (Z % BW)) | Y >> (Z % BW) |
5880 | auto Mask = MIRBuilder.buildConstant(ShTy, BW - 1); |
5881 | if (isPowerOf2_32(BW)) { |
5882 | // Z % BW -> Z & (BW - 1) |
5883 | ShAmt = MIRBuilder.buildAnd(ShTy, Z, Mask).getReg(0); |
5884 | // (BW - 1) - (Z % BW) -> ~Z & (BW - 1) |
5885 | auto NotZ = MIRBuilder.buildNot(ShTy, Z); |
5886 | InvShAmt = MIRBuilder.buildAnd(ShTy, NotZ, Mask).getReg(0); |
5887 | } else { |
5888 | auto BitWidthC = MIRBuilder.buildConstant(ShTy, BW); |
5889 | ShAmt = MIRBuilder.buildURem(ShTy, Z, BitWidthC).getReg(0); |
5890 | InvShAmt = MIRBuilder.buildSub(ShTy, Mask, ShAmt).getReg(0); |
5891 | } |
5892 | |
5893 | auto One = MIRBuilder.buildConstant(ShTy, 1); |
5894 | if (IsFSHL) { |
5895 | ShX = MIRBuilder.buildShl(Ty, X, ShAmt).getReg(0); |
5896 | auto ShY1 = MIRBuilder.buildLShr(Ty, Y, One); |
5897 | ShY = MIRBuilder.buildLShr(Ty, ShY1, InvShAmt).getReg(0); |
5898 | } else { |
5899 | auto ShX1 = MIRBuilder.buildShl(Ty, X, One); |
5900 | ShX = MIRBuilder.buildShl(Ty, ShX1, InvShAmt).getReg(0); |
5901 | ShY = MIRBuilder.buildLShr(Ty, Y, ShAmt).getReg(0); |
5902 | } |
5903 | } |
5904 | |
5905 | MIRBuilder.buildOr(Dst, ShX, ShY); |
5906 | MI.eraseFromParent(); |
5907 | return Legalized; |
5908 | } |
5909 | |
5910 | LegalizerHelper::LegalizeResult |
5911 | LegalizerHelper::lowerFunnelShift(MachineInstr &MI) { |
5912 | // These operations approximately do the following (while avoiding undefined |
5913 | // shifts by BW): |
5914 | // G_FSHL: (X << (Z % BW)) | (Y >> (BW - (Z % BW))) |
5915 | // G_FSHR: (X << (BW - (Z % BW))) | (Y >> (Z % BW)) |
5916 | Register Dst = MI.getOperand(0).getReg(); |
5917 | LLT Ty = MRI.getType(Dst); |
5918 | LLT ShTy = MRI.getType(MI.getOperand(3).getReg()); |
5919 | |
5920 | bool IsFSHL = MI.getOpcode() == TargetOpcode::G_FSHL; |
5921 | unsigned RevOpcode = IsFSHL ? TargetOpcode::G_FSHR : TargetOpcode::G_FSHL; |
5922 | |
5923 | // TODO: Use smarter heuristic that accounts for vector legalization. |
5924 | if (LI.getAction({RevOpcode, {Ty, ShTy}}).Action == Lower) |
5925 | return lowerFunnelShiftAsShifts(MI); |
5926 | |
5927 | // This only works for powers of 2, fallback to shifts if it fails. |
5928 | LegalizerHelper::LegalizeResult Result = lowerFunnelShiftWithInverse(MI); |
5929 | if (Result == UnableToLegalize) |
5930 | return lowerFunnelShiftAsShifts(MI); |
5931 | return Result; |
5932 | } |
5933 | |
5934 | LegalizerHelper::LegalizeResult |
5935 | LegalizerHelper::lowerRotateWithReverseRotate(MachineInstr &MI) { |
5936 | Register Dst = MI.getOperand(0).getReg(); |
5937 | Register Src = MI.getOperand(1).getReg(); |
5938 | Register Amt = MI.getOperand(2).getReg(); |
5939 | LLT AmtTy = MRI.getType(Amt); |
5940 | auto Zero = MIRBuilder.buildConstant(AmtTy, 0); |
5941 | bool IsLeft = MI.getOpcode() == TargetOpcode::G_ROTL; |
5942 | unsigned RevRot = IsLeft ? TargetOpcode::G_ROTR : TargetOpcode::G_ROTL; |
5943 | auto Neg = MIRBuilder.buildSub(AmtTy, Zero, Amt); |
5944 | MIRBuilder.buildInstr(RevRot, {Dst}, {Src, Neg}); |
5945 | MI.eraseFromParent(); |
5946 | return Legalized; |
5947 | } |
5948 | |
5949 | LegalizerHelper::LegalizeResult LegalizerHelper::lowerRotate(MachineInstr &MI) { |
5950 | Register Dst = MI.getOperand(0).getReg(); |
5951 | Register Src = MI.getOperand(1).getReg(); |
5952 | Register Amt = MI.getOperand(2).getReg(); |
5953 | LLT DstTy = MRI.getType(Dst); |
5954 | LLT SrcTy = MRI.getType(Dst); |
5955 | LLT AmtTy = MRI.getType(Amt); |
5956 | |
5957 | unsigned EltSizeInBits = DstTy.getScalarSizeInBits(); |
5958 | bool IsLeft = MI.getOpcode() == TargetOpcode::G_ROTL; |
5959 | |
5960 | MIRBuilder.setInstrAndDebugLoc(MI); |
5961 | |
5962 | // If a rotate in the other direction is supported, use it. |
5963 | unsigned RevRot = IsLeft ? TargetOpcode::G_ROTR : TargetOpcode::G_ROTL; |
5964 | if (LI.isLegalOrCustom({RevRot, {DstTy, SrcTy}}) && |
5965 | isPowerOf2_32(EltSizeInBits)) |
5966 | return lowerRotateWithReverseRotate(MI); |
5967 | |
5968 | auto Zero = MIRBuilder.buildConstant(AmtTy, 0); |
5969 | unsigned ShOpc = IsLeft ? TargetOpcode::G_SHL : TargetOpcode::G_LSHR; |
5970 | unsigned RevShiftOpc = IsLeft ? TargetOpcode::G_LSHR : TargetOpcode::G_SHL; |
5971 | auto BitWidthMinusOneC = MIRBuilder.buildConstant(AmtTy, EltSizeInBits - 1); |
5972 | Register ShVal; |
5973 | Register RevShiftVal; |
5974 | if (isPowerOf2_32(EltSizeInBits)) { |
5975 | // (rotl x, c) -> x << (c & (w - 1)) | x >> (-c & (w - 1)) |
5976 | // (rotr x, c) -> x >> (c & (w - 1)) | x << (-c & (w - 1)) |
5977 | auto NegAmt = MIRBuilder.buildSub(AmtTy, Zero, Amt); |
5978 | auto ShAmt = MIRBuilder.buildAnd(AmtTy, Amt, BitWidthMinusOneC); |
5979 | ShVal = MIRBuilder.buildInstr(ShOpc, {DstTy}, {Src, ShAmt}).getReg(0); |
5980 | auto RevAmt = MIRBuilder.buildAnd(AmtTy, NegAmt, BitWidthMinusOneC); |
5981 | RevShiftVal = |
5982 | MIRBuilder.buildInstr(RevShiftOpc, {DstTy}, {Src, RevAmt}).getReg(0); |
5983 | } else { |
5984 | // (rotl x, c) -> x << (c % w) | x >> 1 >> (w - 1 - (c % w)) |
5985 | // (rotr x, c) -> x >> (c % w) | x << 1 << (w - 1 - (c % w)) |
5986 | auto BitWidthC = MIRBuilder.buildConstant(AmtTy, EltSizeInBits); |
5987 | auto ShAmt = MIRBuilder.buildURem(AmtTy, Amt, BitWidthC); |
5988 | ShVal = MIRBuilder.buildInstr(ShOpc, {DstTy}, {Src, ShAmt}).getReg(0); |
5989 | auto RevAmt = MIRBuilder.buildSub(AmtTy, BitWidthMinusOneC, ShAmt); |
5990 | auto One = MIRBuilder.buildConstant(AmtTy, 1); |
5991 | auto Inner = MIRBuilder.buildInstr(RevShiftOpc, {DstTy}, {Src, One}); |
5992 | RevShiftVal = |
5993 | MIRBuilder.buildInstr(RevShiftOpc, {DstTy}, {Inner, RevAmt}).getReg(0); |
5994 | } |
5995 | MIRBuilder.buildOr(Dst, ShVal, RevShiftVal); |
5996 | MI.eraseFromParent(); |
5997 | return Legalized; |
5998 | } |
5999 | |
6000 | // Expand s32 = G_UITOFP s64 using bit operations to an IEEE float |
6001 | // representation. |
6002 | LegalizerHelper::LegalizeResult |
6003 | LegalizerHelper::lowerU64ToF32BitOps(MachineInstr &MI) { |
6004 | Register Dst = MI.getOperand(0).getReg(); |
6005 | Register Src = MI.getOperand(1).getReg(); |
6006 | const LLT S64 = LLT::scalar(64); |
6007 | const LLT S32 = LLT::scalar(32); |
6008 | const LLT S1 = LLT::scalar(1); |
6009 | |
6010 | assert(MRI.getType(Src) == S64 && MRI.getType(Dst) == S32)((void)0); |
6011 | |
6012 | // unsigned cul2f(ulong u) { |
6013 | // uint lz = clz(u); |
6014 | // uint e = (u != 0) ? 127U + 63U - lz : 0; |
6015 | // u = (u << lz) & 0x7fffffffffffffffUL; |
6016 | // ulong t = u & 0xffffffffffUL; |
6017 | // uint v = (e << 23) | (uint)(u >> 40); |
6018 | // uint r = t > 0x8000000000UL ? 1U : (t == 0x8000000000UL ? v & 1U : 0U); |
6019 | // return as_float(v + r); |
6020 | // } |
6021 | |
6022 | auto Zero32 = MIRBuilder.buildConstant(S32, 0); |
6023 | auto Zero64 = MIRBuilder.buildConstant(S64, 0); |
6024 | |
6025 | auto LZ = MIRBuilder.buildCTLZ_ZERO_UNDEF(S32, Src); |
6026 | |
6027 | auto K = MIRBuilder.buildConstant(S32, 127U + 63U); |
6028 | auto Sub = MIRBuilder.buildSub(S32, K, LZ); |
6029 | |
6030 | auto NotZero = MIRBuilder.buildICmp(CmpInst::ICMP_NE, S1, Src, Zero64); |
6031 | auto E = MIRBuilder.buildSelect(S32, NotZero, Sub, Zero32); |
6032 | |
6033 | auto Mask0 = MIRBuilder.buildConstant(S64, (-1ULL) >> 1); |
6034 | auto ShlLZ = MIRBuilder.buildShl(S64, Src, LZ); |
6035 | |
6036 | auto U = MIRBuilder.buildAnd(S64, ShlLZ, Mask0); |
6037 | |
6038 | auto Mask1 = MIRBuilder.buildConstant(S64, 0xffffffffffULL); |
6039 | auto T = MIRBuilder.buildAnd(S64, U, Mask1); |
6040 | |
6041 | auto UShl = MIRBuilder.buildLShr(S64, U, MIRBuilder.buildConstant(S64, 40)); |
6042 | auto ShlE = MIRBuilder.buildShl(S32, E, MIRBuilder.buildConstant(S32, 23)); |
6043 | auto V = MIRBuilder.buildOr(S32, ShlE, MIRBuilder.buildTrunc(S32, UShl)); |
6044 | |
6045 | auto C = MIRBuilder.buildConstant(S64, 0x8000000000ULL); |
6046 | auto RCmp = MIRBuilder.buildICmp(CmpInst::ICMP_UGT, S1, T, C); |
6047 | auto TCmp = MIRBuilder.buildICmp(CmpInst::ICMP_EQ, S1, T, C); |
6048 | auto One = MIRBuilder.buildConstant(S32, 1); |
6049 | |
6050 | auto VTrunc1 = MIRBuilder.buildAnd(S32, V, One); |
6051 | auto Select0 = MIRBuilder.buildSelect(S32, TCmp, VTrunc1, Zero32); |
6052 | auto R = MIRBuilder.buildSelect(S32, RCmp, One, Select0); |
6053 | MIRBuilder.buildAdd(Dst, V, R); |
6054 | |
6055 | MI.eraseFromParent(); |
6056 | return Legalized; |
6057 | } |
6058 | |
6059 | LegalizerHelper::LegalizeResult LegalizerHelper::lowerUITOFP(MachineInstr &MI) { |
6060 | Register Dst = MI.getOperand(0).getReg(); |
6061 | Register Src = MI.getOperand(1).getReg(); |
6062 | LLT DstTy = MRI.getType(Dst); |
6063 | LLT SrcTy = MRI.getType(Src); |
6064 | |
6065 | if (SrcTy == LLT::scalar(1)) { |
6066 | auto True = MIRBuilder.buildFConstant(DstTy, 1.0); |
6067 | auto False = MIRBuilder.buildFConstant(DstTy, 0.0); |
6068 | MIRBuilder.buildSelect(Dst, Src, True, False); |
6069 | MI.eraseFromParent(); |
6070 | return Legalized; |
6071 | } |
6072 | |
6073 | if (SrcTy != LLT::scalar(64)) |
6074 | return UnableToLegalize; |
6075 | |
6076 | if (DstTy == LLT::scalar(32)) { |
6077 | // TODO: SelectionDAG has several alternative expansions to port which may |
6078 | // be more reasonble depending on the available instructions. If a target |
6079 | // has sitofp, does not have CTLZ, or can efficiently use f64 as an |
6080 | // intermediate type, this is probably worse. |
6081 | return lowerU64ToF32BitOps(MI); |
6082 | } |
6083 | |
6084 | return UnableToLegalize; |
6085 | } |
6086 | |
6087 | LegalizerHelper::LegalizeResult LegalizerHelper::lowerSITOFP(MachineInstr &MI) { |
6088 | Register Dst = MI.getOperand(0).getReg(); |
6089 | Register Src = MI.getOperand(1).getReg(); |
6090 | LLT DstTy = MRI.getType(Dst); |
6091 | LLT SrcTy = MRI.getType(Src); |
6092 | |
6093 | const LLT S64 = LLT::scalar(64); |
6094 | const LLT S32 = LLT::scalar(32); |
6095 | const LLT S1 = LLT::scalar(1); |
6096 | |
6097 | if (SrcTy == S1) { |
6098 | auto True = MIRBuilder.buildFConstant(DstTy, -1.0); |
6099 | auto False = MIRBuilder.buildFConstant(DstTy, 0.0); |
6100 | MIRBuilder.buildSelect(Dst, Src, True, False); |
6101 | MI.eraseFromParent(); |
6102 | return Legalized; |
6103 | } |
6104 | |
6105 | if (SrcTy != S64) |
6106 | return UnableToLegalize; |
6107 | |
6108 | if (DstTy == S32) { |
6109 | // signed cl2f(long l) { |
6110 | // long s = l >> 63; |
6111 | // float r = cul2f((l + s) ^ s); |
6112 | // return s ? -r : r; |
6113 | // } |
6114 | Register L = Src; |
6115 | auto SignBit = MIRBuilder.buildConstant(S64, 63); |
6116 | auto S = MIRBuilder.buildAShr(S64, L, SignBit); |
6117 | |
6118 | auto LPlusS = MIRBuilder.buildAdd(S64, L, S); |
6119 | auto Xor = MIRBuilder.buildXor(S64, LPlusS, S); |
6120 | auto R = MIRBuilder.buildUITOFP(S32, Xor); |
6121 | |
6122 | auto RNeg = MIRBuilder.buildFNeg(S32, R); |
6123 | auto SignNotZero = MIRBuilder.buildICmp(CmpInst::ICMP_NE, S1, S, |
6124 | MIRBuilder.buildConstant(S64, 0)); |
6125 | MIRBuilder.buildSelect(Dst, SignNotZero, RNeg, R); |
6126 | MI.eraseFromParent(); |
6127 | return Legalized; |
6128 | } |
6129 | |
6130 | return UnableToLegalize; |
6131 | } |
6132 | |
6133 | LegalizerHelper::LegalizeResult LegalizerHelper::lowerFPTOUI(MachineInstr &MI) { |
6134 | Register Dst = MI.getOperand(0).getReg(); |
6135 | Register Src = MI.getOperand(1).getReg(); |
6136 | LLT DstTy = MRI.getType(Dst); |
6137 | LLT SrcTy = MRI.getType(Src); |
6138 | const LLT S64 = LLT::scalar(64); |
6139 | const LLT S32 = LLT::scalar(32); |
6140 | |
6141 | if (SrcTy != S64 && SrcTy != S32) |
6142 | return UnableToLegalize; |
6143 | if (DstTy != S32 && DstTy != S64) |
6144 | return UnableToLegalize; |
6145 | |
6146 | // FPTOSI gives same result as FPTOUI for positive signed integers. |
6147 | // FPTOUI needs to deal with fp values that convert to unsigned integers |
6148 | // greater or equal to 2^31 for float or 2^63 for double. For brevity 2^Exp. |
6149 | |
6150 | APInt TwoPExpInt = APInt::getSignMask(DstTy.getSizeInBits()); |
6151 | APFloat TwoPExpFP(SrcTy.getSizeInBits() == 32 ? APFloat::IEEEsingle() |
6152 | : APFloat::IEEEdouble(), |
6153 | APInt::getNullValue(SrcTy.getSizeInBits())); |
6154 | TwoPExpFP.convertFromAPInt(TwoPExpInt, false, APFloat::rmNearestTiesToEven); |
6155 | |
6156 | MachineInstrBuilder FPTOSI = MIRBuilder.buildFPTOSI(DstTy, Src); |
6157 | |
6158 | MachineInstrBuilder Threshold = MIRBuilder.buildFConstant(SrcTy, TwoPExpFP); |
6159 | // For fp Value greater or equal to Threshold(2^Exp), we use FPTOSI on |
6160 | // (Value - 2^Exp) and add 2^Exp by setting highest bit in result to 1. |
6161 | MachineInstrBuilder FSub = MIRBuilder.buildFSub(SrcTy, Src, Threshold); |
6162 | MachineInstrBuilder ResLowBits = MIRBuilder.buildFPTOSI(DstTy, FSub); |
6163 | MachineInstrBuilder ResHighBit = MIRBuilder.buildConstant(DstTy, TwoPExpInt); |
6164 | MachineInstrBuilder Res = MIRBuilder.buildXor(DstTy, ResLowBits, ResHighBit); |
6165 | |
6166 | const LLT S1 = LLT::scalar(1); |
6167 | |
6168 | MachineInstrBuilder FCMP = |
6169 | MIRBuilder.buildFCmp(CmpInst::FCMP_ULT, S1, Src, Threshold); |
6170 | MIRBuilder.buildSelect(Dst, FCMP, FPTOSI, Res); |
6171 | |
6172 | MI.eraseFromParent(); |
6173 | return Legalized; |
6174 | } |
6175 | |
6176 | LegalizerHelper::LegalizeResult LegalizerHelper::lowerFPTOSI(MachineInstr &MI) { |
6177 | Register Dst = MI.getOperand(0).getReg(); |
6178 | Register Src = MI.getOperand(1).getReg(); |
6179 | LLT DstTy = MRI.getType(Dst); |
6180 | LLT SrcTy = MRI.getType(Src); |
6181 | const LLT S64 = LLT::scalar(64); |
6182 | const LLT S32 = LLT::scalar(32); |
6183 | |
6184 | // FIXME: Only f32 to i64 conversions are supported. |
6185 | if (SrcTy.getScalarType() != S32 || DstTy.getScalarType() != S64) |
6186 | return UnableToLegalize; |
6187 | |
6188 | // Expand f32 -> i64 conversion |
6189 | // This algorithm comes from compiler-rt's implementation of fixsfdi: |
6190 | // https://github.com/llvm/llvm-project/blob/main/compiler-rt/lib/builtins/fixsfdi.c |
6191 | |
6192 | unsigned SrcEltBits = SrcTy.getScalarSizeInBits(); |
6193 | |
6194 | auto ExponentMask = MIRBuilder.buildConstant(SrcTy, 0x7F800000); |
6195 | auto ExponentLoBit = MIRBuilder.buildConstant(SrcTy, 23); |
6196 | |
6197 | auto AndExpMask = MIRBuilder.buildAnd(SrcTy, Src, ExponentMask); |
6198 | auto ExponentBits = MIRBuilder.buildLShr(SrcTy, AndExpMask, ExponentLoBit); |
6199 | |
6200 | auto SignMask = MIRBuilder.buildConstant(SrcTy, |
6201 | APInt::getSignMask(SrcEltBits)); |
6202 | auto AndSignMask = MIRBuilder.buildAnd(SrcTy, Src, SignMask); |
6203 | auto SignLowBit = MIRBuilder.buildConstant(SrcTy, SrcEltBits - 1); |
6204 | auto Sign = MIRBuilder.buildAShr(SrcTy, AndSignMask, SignLowBit); |
6205 | Sign = MIRBuilder.buildSExt(DstTy, Sign); |
6206 | |
6207 | auto MantissaMask = MIRBuilder.buildConstant(SrcTy, 0x007FFFFF); |
6208 | auto AndMantissaMask = MIRBuilder.buildAnd(SrcTy, Src, MantissaMask); |
6209 | auto K = MIRBuilder.buildConstant(SrcTy, 0x00800000); |
6210 | |
6211 | auto R = MIRBuilder.buildOr(SrcTy, AndMantissaMask, K); |
6212 | R = MIRBuilder.buildZExt(DstTy, R); |
6213 | |
6214 | auto Bias = MIRBuilder.buildConstant(SrcTy, 127); |
6215 | auto Exponent = MIRBuilder.buildSub(SrcTy, ExponentBits, Bias); |
6216 | auto SubExponent = MIRBuilder.buildSub(SrcTy, Exponent, ExponentLoBit); |
6217 | auto ExponentSub = MIRBuilder.buildSub(SrcTy, ExponentLoBit, Exponent); |
6218 | |
6219 | auto Shl = MIRBuilder.buildShl(DstTy, R, SubExponent); |
6220 | auto Srl = MIRBuilder.buildLShr(DstTy, R, ExponentSub); |
6221 | |
6222 | const LLT S1 = LLT::scalar(1); |
6223 | auto CmpGt = MIRBuilder.buildICmp(CmpInst::ICMP_SGT, |
6224 | S1, Exponent, ExponentLoBit); |
6225 | |
6226 | R = MIRBuilder.buildSelect(DstTy, CmpGt, Shl, Srl); |
6227 | |
6228 | auto XorSign = MIRBuilder.buildXor(DstTy, R, Sign); |
6229 | auto Ret = MIRBuilder.buildSub(DstTy, XorSign, Sign); |
6230 | |
6231 | auto ZeroSrcTy = MIRBuilder.buildConstant(SrcTy, 0); |
6232 | |
6233 | auto ExponentLt0 = MIRBuilder.buildICmp(CmpInst::ICMP_SLT, |
6234 | S1, Exponent, ZeroSrcTy); |
6235 | |
6236 | auto ZeroDstTy = MIRBuilder.buildConstant(DstTy, 0); |
6237 | MIRBuilder.buildSelect(Dst, ExponentLt0, ZeroDstTy, Ret); |
6238 | |
6239 | MI.eraseFromParent(); |
6240 | return Legalized; |
6241 | } |
6242 | |
6243 | // f64 -> f16 conversion using round-to-nearest-even rounding mode. |
6244 | LegalizerHelper::LegalizeResult |
6245 | LegalizerHelper::lowerFPTRUNC_F64_TO_F16(MachineInstr &MI) { |
6246 | Register Dst = MI.getOperand(0).getReg(); |
6247 | Register Src = MI.getOperand(1).getReg(); |
6248 | |
6249 | if (MRI.getType(Src).isVector()) // TODO: Handle vectors directly. |
6250 | return UnableToLegalize; |
6251 | |
6252 | const unsigned ExpMask = 0x7ff; |
6253 | const unsigned ExpBiasf64 = 1023; |
6254 | const unsigned ExpBiasf16 = 15; |
6255 | const LLT S32 = LLT::scalar(32); |
6256 | const LLT S1 = LLT::scalar(1); |
6257 | |
6258 | auto Unmerge = MIRBuilder.buildUnmerge(S32, Src); |
6259 | Register U = Unmerge.getReg(0); |
6260 | Register UH = Unmerge.getReg(1); |
6261 | |
6262 | auto E = MIRBuilder.buildLShr(S32, UH, MIRBuilder.buildConstant(S32, 20)); |
6263 | E = MIRBuilder.buildAnd(S32, E, MIRBuilder.buildConstant(S32, ExpMask)); |
6264 | |
6265 | // Subtract the fp64 exponent bias (1023) to get the real exponent and |
6266 | // add the f16 bias (15) to get the biased exponent for the f16 format. |
6267 | E = MIRBuilder.buildAdd( |
6268 | S32, E, MIRBuilder.buildConstant(S32, -ExpBiasf64 + ExpBiasf16)); |
6269 | |
6270 | auto M = MIRBuilder.buildLShr(S32, UH, MIRBuilder.buildConstant(S32, 8)); |
6271 | M = MIRBuilder.buildAnd(S32, M, MIRBuilder.buildConstant(S32, 0xffe)); |
6272 | |
6273 | auto MaskedSig = MIRBuilder.buildAnd(S32, UH, |
6274 | MIRBuilder.buildConstant(S32, 0x1ff)); |
6275 | MaskedSig = MIRBuilder.buildOr(S32, MaskedSig, U); |
6276 | |
6277 | auto Zero = MIRBuilder.buildConstant(S32, 0); |
6278 | auto SigCmpNE0 = MIRBuilder.buildICmp(CmpInst::ICMP_NE, S1, MaskedSig, Zero); |
6279 | auto Lo40Set = MIRBuilder.buildZExt(S32, SigCmpNE0); |
6280 | M = MIRBuilder.buildOr(S32, M, Lo40Set); |
6281 | |
6282 | // (M != 0 ? 0x0200 : 0) | 0x7c00; |
6283 | auto Bits0x200 = MIRBuilder.buildConstant(S32, 0x0200); |
6284 | auto CmpM_NE0 = MIRBuilder.buildICmp(CmpInst::ICMP_NE, S1, M, Zero); |
6285 | auto SelectCC = MIRBuilder.buildSelect(S32, CmpM_NE0, Bits0x200, Zero); |
6286 | |
6287 | auto Bits0x7c00 = MIRBuilder.buildConstant(S32, 0x7c00); |
6288 | auto I = MIRBuilder.buildOr(S32, SelectCC, Bits0x7c00); |
6289 | |
6290 | // N = M | (E << 12); |
6291 | auto EShl12 = MIRBuilder.buildShl(S32, E, MIRBuilder.buildConstant(S32, 12)); |
6292 | auto N = MIRBuilder.buildOr(S32, M, EShl12); |
6293 | |
6294 | // B = clamp(1-E, 0, 13); |
6295 | auto One = MIRBuilder.buildConstant(S32, 1); |
6296 | auto OneSubExp = MIRBuilder.buildSub(S32, One, E); |
6297 | auto B = MIRBuilder.buildSMax(S32, OneSubExp, Zero); |
6298 | B = MIRBuilder.buildSMin(S32, B, MIRBuilder.buildConstant(S32, 13)); |
6299 | |
6300 | auto SigSetHigh = MIRBuilder.buildOr(S32, M, |
6301 | MIRBuilder.buildConstant(S32, 0x1000)); |
6302 | |
6303 | auto D = MIRBuilder.buildLShr(S32, SigSetHigh, B); |
6304 | auto D0 = MIRBuilder.buildShl(S32, D, B); |
6305 | |
6306 | auto D0_NE_SigSetHigh = MIRBuilder.buildICmp(CmpInst::ICMP_NE, S1, |
6307 | D0, SigSetHigh); |
6308 | auto D1 = MIRBuilder.buildZExt(S32, D0_NE_SigSetHigh); |
6309 | D = MIRBuilder.buildOr(S32, D, D1); |
6310 | |
6311 | auto CmpELtOne = MIRBuilder.buildICmp(CmpInst::ICMP_SLT, S1, E, One); |
6312 | auto V = MIRBuilder.buildSelect(S32, CmpELtOne, D, N); |
6313 | |
6314 | auto VLow3 = MIRBuilder.buildAnd(S32, V, MIRBuilder.buildConstant(S32, 7)); |
6315 | V = MIRBuilder.buildLShr(S32, V, MIRBuilder.buildConstant(S32, 2)); |
6316 | |
6317 | auto VLow3Eq3 = MIRBuilder.buildICmp(CmpInst::ICMP_EQ, S1, VLow3, |
6318 | MIRBuilder.buildConstant(S32, 3)); |
6319 | auto V0 = MIRBuilder.buildZExt(S32, VLow3Eq3); |
6320 | |
6321 | auto VLow3Gt5 = MIRBuilder.buildICmp(CmpInst::ICMP_SGT, S1, VLow3, |
6322 | MIRBuilder.buildConstant(S32, 5)); |
6323 | auto V1 = MIRBuilder.buildZExt(S32, VLow3Gt5); |
6324 | |
6325 | V1 = MIRBuilder.buildOr(S32, V0, V1); |
6326 | V = MIRBuilder.buildAdd(S32, V, V1); |
6327 | |
6328 | auto CmpEGt30 = MIRBuilder.buildICmp(CmpInst::ICMP_SGT, S1, |
6329 | E, MIRBuilder.buildConstant(S32, 30)); |
6330 | V = MIRBuilder.buildSelect(S32, CmpEGt30, |
6331 | MIRBuilder.buildConstant(S32, 0x7c00), V); |
6332 | |
6333 | auto CmpEGt1039 = MIRBuilder.buildICmp(CmpInst::ICMP_EQ, S1, |
6334 | E, MIRBuilder.buildConstant(S32, 1039)); |
6335 | V = MIRBuilder.buildSelect(S32, CmpEGt1039, I, V); |
6336 | |
6337 | // Extract the sign bit. |
6338 | auto Sign = MIRBuilder.buildLShr(S32, UH, MIRBuilder.buildConstant(S32, 16)); |
6339 | Sign = MIRBuilder.buildAnd(S32, Sign, MIRBuilder.buildConstant(S32, 0x8000)); |
6340 | |
6341 | // Insert the sign bit |
6342 | V = MIRBuilder.buildOr(S32, Sign, V); |
6343 | |
6344 | MIRBuilder.buildTrunc(Dst, V); |
6345 | MI.eraseFromParent(); |
6346 | return Legalized; |
6347 | } |
6348 | |
6349 | LegalizerHelper::LegalizeResult |
6350 | LegalizerHelper::lowerFPTRUNC(MachineInstr &MI) { |
6351 | Register Dst = MI.getOperand(0).getReg(); |
6352 | Register Src = MI.getOperand(1).getReg(); |
6353 | |
6354 | LLT DstTy = MRI.getType(Dst); |
6355 | LLT SrcTy = MRI.getType(Src); |
6356 | const LLT S64 = LLT::scalar(64); |
6357 | const LLT S16 = LLT::scalar(16); |
6358 | |
6359 | if (DstTy.getScalarType() == S16 && SrcTy.getScalarType() == S64) |
6360 | return lowerFPTRUNC_F64_TO_F16(MI); |
6361 | |
6362 | return UnableToLegalize; |
6363 | } |
6364 | |
6365 | // TODO: If RHS is a constant SelectionDAGBuilder expands this into a |
6366 | // multiplication tree. |
6367 | LegalizerHelper::LegalizeResult LegalizerHelper::lowerFPOWI(MachineInstr &MI) { |
6368 | Register Dst = MI.getOperand(0).getReg(); |
6369 | Register Src0 = MI.getOperand(1).getReg(); |
6370 | Register Src1 = MI.getOperand(2).getReg(); |
6371 | LLT Ty = MRI.getType(Dst); |
6372 | |
6373 | auto CvtSrc1 = MIRBuilder.buildSITOFP(Ty, Src1); |
6374 | MIRBuilder.buildFPow(Dst, Src0, CvtSrc1, MI.getFlags()); |
6375 | MI.eraseFromParent(); |
6376 | return Legalized; |
6377 | } |
6378 | |
6379 | static CmpInst::Predicate minMaxToCompare(unsigned Opc) { |
6380 | switch (Opc) { |
6381 | case TargetOpcode::G_SMIN: |
6382 | return CmpInst::ICMP_SLT; |
6383 | case TargetOpcode::G_SMAX: |
6384 | return CmpInst::ICMP_SGT; |
6385 | case TargetOpcode::G_UMIN: |
6386 | return CmpInst::ICMP_ULT; |
6387 | case TargetOpcode::G_UMAX: |
6388 | return CmpInst::ICMP_UGT; |
6389 | default: |
6390 | llvm_unreachable("not in integer min/max")__builtin_unreachable(); |
6391 | } |
6392 | } |
6393 | |
6394 | LegalizerHelper::LegalizeResult LegalizerHelper::lowerMinMax(MachineInstr &MI) { |
6395 | Register Dst = MI.getOperand(0).getReg(); |
6396 | Register Src0 = MI.getOperand(1).getReg(); |
6397 | Register Src1 = MI.getOperand(2).getReg(); |
6398 | |
6399 | const CmpInst::Predicate Pred = minMaxToCompare(MI.getOpcode()); |
6400 | LLT CmpType = MRI.getType(Dst).changeElementSize(1); |
6401 | |
6402 | auto Cmp = MIRBuilder.buildICmp(Pred, CmpType, Src0, Src1); |
6403 | MIRBuilder.buildSelect(Dst, Cmp, Src0, Src1); |
6404 | |
6405 | MI.eraseFromParent(); |
6406 | return Legalized; |
6407 | } |
6408 | |
6409 | LegalizerHelper::LegalizeResult |
6410 | LegalizerHelper::lowerFCopySign(MachineInstr &MI) { |
6411 | Register Dst = MI.getOperand(0).getReg(); |
6412 | Register Src0 = MI.getOperand(1).getReg(); |
6413 | Register Src1 = MI.getOperand(2).getReg(); |
6414 | |
6415 | const LLT Src0Ty = MRI.getType(Src0); |
6416 | const LLT Src1Ty = MRI.getType(Src1); |
6417 | |
6418 | const int Src0Size = Src0Ty.getScalarSizeInBits(); |
6419 | const int Src1Size = Src1Ty.getScalarSizeInBits(); |
6420 | |
6421 | auto SignBitMask = MIRBuilder.buildConstant( |
6422 | Src0Ty, APInt::getSignMask(Src0Size)); |
6423 | |
6424 | auto NotSignBitMask = MIRBuilder.buildConstant( |
6425 | Src0Ty, APInt::getLowBitsSet(Src0Size, Src0Size - 1)); |
6426 | |
6427 | Register And0 = MIRBuilder.buildAnd(Src0Ty, Src0, NotSignBitMask).getReg(0); |
6428 | Register And1; |
6429 | if (Src0Ty == Src1Ty) { |
6430 | And1 = MIRBuilder.buildAnd(Src1Ty, Src1, SignBitMask).getReg(0); |
6431 | } else if (Src0Size > Src1Size) { |
6432 | auto ShiftAmt = MIRBuilder.buildConstant(Src0Ty, Src0Size - Src1Size); |
6433 | auto Zext = MIRBuilder.buildZExt(Src0Ty, Src1); |
6434 | auto Shift = MIRBuilder.buildShl(Src0Ty, Zext, ShiftAmt); |
6435 | And1 = MIRBuilder.buildAnd(Src0Ty, Shift, SignBitMask).getReg(0); |
6436 | } else { |
6437 | auto ShiftAmt = MIRBuilder.buildConstant(Src1Ty, Src1Size - Src0Size); |
6438 | auto Shift = MIRBuilder.buildLShr(Src1Ty, Src1, ShiftAmt); |
6439 | auto Trunc = MIRBuilder.buildTrunc(Src0Ty, Shift); |
6440 | And1 = MIRBuilder.buildAnd(Src0Ty, Trunc, SignBitMask).getReg(0); |
6441 | } |
6442 | |
6443 | // Be careful about setting nsz/nnan/ninf on every instruction, since the |
6444 | // constants are a nan and -0.0, but the final result should preserve |
6445 | // everything. |
6446 | unsigned Flags = MI.getFlags(); |
6447 | MIRBuilder.buildOr(Dst, And0, And1, Flags); |
6448 | |
6449 | MI.eraseFromParent(); |
6450 | return Legalized; |
6451 | } |
6452 | |
6453 | LegalizerHelper::LegalizeResult |
6454 | LegalizerHelper::lowerFMinNumMaxNum(MachineInstr &MI) { |
6455 | unsigned NewOp = MI.getOpcode() == TargetOpcode::G_FMINNUM ? |
6456 | TargetOpcode::G_FMINNUM_IEEE : TargetOpcode::G_FMAXNUM_IEEE; |
6457 | |
6458 | Register Dst = MI.getOperand(0).getReg(); |
6459 | Register Src0 = MI.getOperand(1).getReg(); |
6460 | Register Src1 = MI.getOperand(2).getReg(); |
6461 | LLT Ty = MRI.getType(Dst); |
6462 | |
6463 | if (!MI.getFlag(MachineInstr::FmNoNans)) { |
6464 | // Insert canonicalizes if it's possible we need to quiet to get correct |
6465 | // sNaN behavior. |
6466 | |
6467 | // Note this must be done here, and not as an optimization combine in the |
6468 | // absence of a dedicate quiet-snan instruction as we're using an |
6469 | // omni-purpose G_FCANONICALIZE. |
6470 | if (!isKnownNeverSNaN(Src0, MRI)) |
6471 | Src0 = MIRBuilder.buildFCanonicalize(Ty, Src0, MI.getFlags()).getReg(0); |
6472 | |
6473 | if (!isKnownNeverSNaN(Src1, MRI)) |
6474 | Src1 = MIRBuilder.buildFCanonicalize(Ty, Src1, MI.getFlags()).getReg(0); |
6475 | } |
6476 | |
6477 | // If there are no nans, it's safe to simply replace this with the non-IEEE |
6478 | // version. |
6479 | MIRBuilder.buildInstr(NewOp, {Dst}, {Src0, Src1}, MI.getFlags()); |
6480 | MI.eraseFromParent(); |
6481 | return Legalized; |
6482 | } |
6483 | |
6484 | LegalizerHelper::LegalizeResult LegalizerHelper::lowerFMad(MachineInstr &MI) { |
6485 | // Expand G_FMAD a, b, c -> G_FADD (G_FMUL a, b), c |
6486 | Register DstReg = MI.getOperand(0).getReg(); |
6487 | LLT Ty = MRI.getType(DstReg); |
6488 | unsigned Flags = MI.getFlags(); |
6489 | |
6490 | auto Mul = MIRBuilder.buildFMul(Ty, MI.getOperand(1), MI.getOperand(2), |
6491 | Flags); |
6492 | MIRBuilder.buildFAdd(DstReg, Mul, MI.getOperand(3), Flags); |
6493 | MI.eraseFromParent(); |
6494 | return Legalized; |
6495 | } |
6496 | |
6497 | LegalizerHelper::LegalizeResult |
6498 | LegalizerHelper::lowerIntrinsicRound(MachineInstr &MI) { |
6499 | Register DstReg = MI.getOperand(0).getReg(); |
6500 | Register X = MI.getOperand(1).getReg(); |
6501 | const unsigned Flags = MI.getFlags(); |
6502 | const LLT Ty = MRI.getType(DstReg); |
6503 | const LLT CondTy = Ty.changeElementSize(1); |
6504 | |
6505 | // round(x) => |
6506 | // t = trunc(x); |
6507 | // d = fabs(x - t); |
6508 | // o = copysign(1.0f, x); |
6509 | // return t + (d >= 0.5 ? o : 0.0); |
6510 | |
6511 | auto T = MIRBuilder.buildIntrinsicTrunc(Ty, X, Flags); |
6512 | |
6513 | auto Diff = MIRBuilder.buildFSub(Ty, X, T, Flags); |
6514 | auto AbsDiff = MIRBuilder.buildFAbs(Ty, Diff, Flags); |
6515 | auto Zero = MIRBuilder.buildFConstant(Ty, 0.0); |
6516 | auto One = MIRBuilder.buildFConstant(Ty, 1.0); |
6517 | auto Half = MIRBuilder.buildFConstant(Ty, 0.5); |
6518 | auto SignOne = MIRBuilder.buildFCopysign(Ty, One, X); |
6519 | |
6520 | auto Cmp = MIRBuilder.buildFCmp(CmpInst::FCMP_OGE, CondTy, AbsDiff, Half, |
6521 | Flags); |
6522 | auto Sel = MIRBuilder.buildSelect(Ty, Cmp, SignOne, Zero, Flags); |
6523 | |
6524 | MIRBuilder.buildFAdd(DstReg, T, Sel, Flags); |
6525 | |
6526 | MI.eraseFromParent(); |
6527 | return Legalized; |
6528 | } |
6529 | |
6530 | LegalizerHelper::LegalizeResult |
6531 | LegalizerHelper::lowerFFloor(MachineInstr &MI) { |
6532 | Register DstReg = MI.getOperand(0).getReg(); |
6533 | Register SrcReg = MI.getOperand(1).getReg(); |
6534 | unsigned Flags = MI.getFlags(); |
6535 | LLT Ty = MRI.getType(DstReg); |
6536 | const LLT CondTy = Ty.changeElementSize(1); |
6537 | |
6538 | // result = trunc(src); |
6539 | // if (src < 0.0 && src != result) |
6540 | // result += -1.0. |
6541 | |
6542 | auto Trunc = MIRBuilder.buildIntrinsicTrunc(Ty, SrcReg, Flags); |
6543 | auto Zero = MIRBuilder.buildFConstant(Ty, 0.0); |
6544 | |
6545 | auto Lt0 = MIRBuilder.buildFCmp(CmpInst::FCMP_OLT, CondTy, |
6546 | SrcReg, Zero, Flags); |
6547 | auto NeTrunc = MIRBuilder.buildFCmp(CmpInst::FCMP_ONE, CondTy, |
6548 | SrcReg, Trunc, Flags); |
6549 | auto And = MIRBuilder.buildAnd(CondTy, Lt0, NeTrunc); |
6550 | auto AddVal = MIRBuilder.buildSITOFP(Ty, And); |
6551 | |
6552 | MIRBuilder.buildFAdd(DstReg, Trunc, AddVal, Flags); |
6553 | MI.eraseFromParent(); |
6554 | return Legalized; |
6555 | } |
6556 | |
6557 | LegalizerHelper::LegalizeResult |
6558 | LegalizerHelper::lowerMergeValues(MachineInstr &MI) { |
6559 | const unsigned NumOps = MI.getNumOperands(); |
6560 | Register DstReg = MI.getOperand(0).getReg(); |
6561 | Register Src0Reg = MI.getOperand(1).getReg(); |
6562 | LLT DstTy = MRI.getType(DstReg); |
6563 | LLT SrcTy = MRI.getType(Src0Reg); |
6564 | unsigned PartSize = SrcTy.getSizeInBits(); |
6565 | |
6566 | LLT WideTy = LLT::scalar(DstTy.getSizeInBits()); |
6567 | Register ResultReg = MIRBuilder.buildZExt(WideTy, Src0Reg).getReg(0); |
6568 | |
6569 | for (unsigned I = 2; I != NumOps; ++I) { |
6570 | const unsigned Offset = (I - 1) * PartSize; |
6571 | |
6572 | Register SrcReg = MI.getOperand(I).getReg(); |
6573 | auto ZextInput = MIRBuilder.buildZExt(WideTy, SrcReg); |
6574 | |
6575 | Register NextResult = I + 1 == NumOps && WideTy == DstTy ? DstReg : |
6576 | MRI.createGenericVirtualRegister(WideTy); |
6577 | |
6578 | auto ShiftAmt = MIRBuilder.buildConstant(WideTy, Offset); |
6579 | auto Shl = MIRBuilder.buildShl(WideTy, ZextInput, ShiftAmt); |
6580 | MIRBuilder.buildOr(NextResult, ResultReg, Shl); |
6581 | ResultReg = NextResult; |
6582 | } |
6583 | |
6584 | if (DstTy.isPointer()) { |
6585 | if (MIRBuilder.getDataLayout().isNonIntegralAddressSpace( |
6586 | DstTy.getAddressSpace())) { |
6587 | LLVM_DEBUG(dbgs() << "Not casting nonintegral address space\n")do { } while (false); |
6588 | return UnableToLegalize; |
6589 | } |
6590 | |
6591 | MIRBuilder.buildIntToPtr(DstReg, ResultReg); |
6592 | } |
6593 | |
6594 | MI.eraseFromParent(); |
6595 | return Legalized; |
6596 | } |
6597 | |
6598 | LegalizerHelper::LegalizeResult |
6599 | LegalizerHelper::lowerUnmergeValues(MachineInstr &MI) { |
6600 | const unsigned NumDst = MI.getNumOperands() - 1; |
6601 | Register SrcReg = MI.getOperand(NumDst).getReg(); |
6602 | Register Dst0Reg = MI.getOperand(0).getReg(); |
6603 | LLT DstTy = MRI.getType(Dst0Reg); |
6604 | if (DstTy.isPointer()) |
6605 | return UnableToLegalize; // TODO |
6606 | |
6607 | SrcReg = coerceToScalar(SrcReg); |
6608 | if (!SrcReg) |
6609 | return UnableToLegalize; |
6610 | |
6611 | // Expand scalarizing unmerge as bitcast to integer and shift. |
6612 | LLT IntTy = MRI.getType(SrcReg); |
6613 | |
6614 | MIRBuilder.buildTrunc(Dst0Reg, SrcReg); |
6615 | |
6616 | const unsigned DstSize = DstTy.getSizeInBits(); |
6617 | unsigned Offset = DstSize; |
6618 | for (unsigned I = 1; I != NumDst; ++I, Offset += DstSize) { |
6619 | auto ShiftAmt = MIRBuilder.buildConstant(IntTy, Offset); |
6620 | auto Shift = MIRBuilder.buildLShr(IntTy, SrcReg, ShiftAmt); |
6621 | MIRBuilder.buildTrunc(MI.getOperand(I), Shift); |
6622 | } |
6623 | |
6624 | MI.eraseFromParent(); |
6625 | return Legalized; |
6626 | } |
6627 | |
6628 | /// Lower a vector extract or insert by writing the vector to a stack temporary |
6629 | /// and reloading the element or vector. |
6630 | /// |
6631 | /// %dst = G_EXTRACT_VECTOR_ELT %vec, %idx |
6632 | /// => |
6633 | /// %stack_temp = G_FRAME_INDEX |
6634 | /// G_STORE %vec, %stack_temp |
6635 | /// %idx = clamp(%idx, %vec.getNumElements()) |
6636 | /// %element_ptr = G_PTR_ADD %stack_temp, %idx |
6637 | /// %dst = G_LOAD %element_ptr |
6638 | LegalizerHelper::LegalizeResult |
6639 | LegalizerHelper::lowerExtractInsertVectorElt(MachineInstr &MI) { |
6640 | Register DstReg = MI.getOperand(0).getReg(); |
6641 | Register SrcVec = MI.getOperand(1).getReg(); |
6642 | Register InsertVal; |
6643 | if (MI.getOpcode() == TargetOpcode::G_INSERT_VECTOR_ELT) |
6644 | InsertVal = MI.getOperand(2).getReg(); |
6645 | |
6646 | Register Idx = MI.getOperand(MI.getNumOperands() - 1).getReg(); |
6647 | |
6648 | LLT VecTy = MRI.getType(SrcVec); |
6649 | LLT EltTy = VecTy.getElementType(); |
6650 | if (!EltTy.isByteSized()) { // Not implemented. |
6651 | LLVM_DEBUG(dbgs() << "Can't handle non-byte element vectors yet\n")do { } while (false); |
6652 | return UnableToLegalize; |
6653 | } |
6654 | |
6655 | unsigned EltBytes = EltTy.getSizeInBytes(); |
6656 | Align VecAlign = getStackTemporaryAlignment(VecTy); |
6657 | Align EltAlign; |
6658 | |
6659 | MachinePointerInfo PtrInfo; |
6660 | auto StackTemp = createStackTemporary(TypeSize::Fixed(VecTy.getSizeInBytes()), |
6661 | VecAlign, PtrInfo); |
6662 | MIRBuilder.buildStore(SrcVec, StackTemp, PtrInfo, VecAlign); |
6663 | |
6664 | // Get the pointer to the element, and be sure not to hit undefined behavior |
6665 | // if the index is out of bounds. |
6666 | Register EltPtr = getVectorElementPointer(StackTemp.getReg(0), VecTy, Idx); |
6667 | |
6668 | int64_t IdxVal; |
6669 | if (mi_match(Idx, MRI, m_ICst(IdxVal))) { |
6670 | int64_t Offset = IdxVal * EltBytes; |
6671 | PtrInfo = PtrInfo.getWithOffset(Offset); |
6672 | EltAlign = commonAlignment(VecAlign, Offset); |
6673 | } else { |
6674 | // We lose information with a variable offset. |
6675 | EltAlign = getStackTemporaryAlignment(EltTy); |
6676 | PtrInfo = MachinePointerInfo(MRI.getType(EltPtr).getAddressSpace()); |
6677 | } |
6678 | |
6679 | if (InsertVal) { |
6680 | // Write the inserted element |
6681 | MIRBuilder.buildStore(InsertVal, EltPtr, PtrInfo, EltAlign); |
6682 | |
6683 | // Reload the whole vector. |
6684 | MIRBuilder.buildLoad(DstReg, StackTemp, PtrInfo, VecAlign); |
6685 | } else { |
6686 | MIRBuilder.buildLoad(DstReg, EltPtr, PtrInfo, EltAlign); |
6687 | } |
6688 | |
6689 | MI.eraseFromParent(); |
6690 | return Legalized; |
6691 | } |
6692 | |
6693 | LegalizerHelper::LegalizeResult |
6694 | LegalizerHelper::lowerShuffleVector(MachineInstr &MI) { |
6695 | Register DstReg = MI.getOperand(0).getReg(); |
6696 | Register Src0Reg = MI.getOperand(1).getReg(); |
6697 | Register Src1Reg = MI.getOperand(2).getReg(); |
6698 | LLT Src0Ty = MRI.getType(Src0Reg); |
6699 | LLT DstTy = MRI.getType(DstReg); |
6700 | LLT IdxTy = LLT::scalar(32); |
6701 | |
6702 | ArrayRef<int> Mask = MI.getOperand(3).getShuffleMask(); |
6703 | |
6704 | if (DstTy.isScalar()) { |
6705 | if (Src0Ty.isVector()) |
6706 | return UnableToLegalize; |
6707 | |
6708 | // This is just a SELECT. |
6709 | assert(Mask.size() == 1 && "Expected a single mask element")((void)0); |
6710 | Register Val; |
6711 | if (Mask[0] < 0 || Mask[0] > 1) |
6712 | Val = MIRBuilder.buildUndef(DstTy).getReg(0); |
6713 | else |
6714 | Val = Mask[0] == 0 ? Src0Reg : Src1Reg; |
6715 | MIRBuilder.buildCopy(DstReg, Val); |
6716 | MI.eraseFromParent(); |
6717 | return Legalized; |
6718 | } |
6719 | |
6720 | Register Undef; |
6721 | SmallVector<Register, 32> BuildVec; |
6722 | LLT EltTy = DstTy.getElementType(); |
6723 | |
6724 | for (int Idx : Mask) { |
6725 | if (Idx < 0) { |
6726 | if (!Undef.isValid()) |
6727 | Undef = MIRBuilder.buildUndef(EltTy).getReg(0); |
6728 | BuildVec.push_back(Undef); |
6729 | continue; |
6730 | } |
6731 | |
6732 | if (Src0Ty.isScalar()) { |
6733 | BuildVec.push_back(Idx == 0 ? Src0Reg : Src1Reg); |
6734 | } else { |
6735 | int NumElts = Src0Ty.getNumElements(); |
6736 | Register SrcVec = Idx < NumElts ? Src0Reg : Src1Reg; |
6737 | int ExtractIdx = Idx < NumElts ? Idx : Idx - NumElts; |
6738 | auto IdxK = MIRBuilder.buildConstant(IdxTy, ExtractIdx); |
6739 | auto Extract = MIRBuilder.buildExtractVectorElement(EltTy, SrcVec, IdxK); |
6740 | BuildVec.push_back(Extract.getReg(0)); |
6741 | } |
6742 | } |
6743 | |
6744 | MIRBuilder.buildBuildVector(DstReg, BuildVec); |
6745 | MI.eraseFromParent(); |
6746 | return Legalized; |
6747 | } |
6748 | |
6749 | LegalizerHelper::LegalizeResult |
6750 | LegalizerHelper::lowerDynStackAlloc(MachineInstr &MI) { |
6751 | const auto &MF = *MI.getMF(); |
6752 | const auto &TFI = *MF.getSubtarget().getFrameLowering(); |
6753 | if (TFI.getStackGrowthDirection() == TargetFrameLowering::StackGrowsUp) |
6754 | return UnableToLegalize; |
6755 | |
6756 | Register Dst = MI.getOperand(0).getReg(); |
6757 | Register AllocSize = MI.getOperand(1).getReg(); |
6758 | Align Alignment = assumeAligned(MI.getOperand(2).getImm()); |
6759 | |
6760 | LLT PtrTy = MRI.getType(Dst); |
6761 | LLT IntPtrTy = LLT::scalar(PtrTy.getSizeInBits()); |
6762 | |
6763 | Register SPReg = TLI.getStackPointerRegisterToSaveRestore(); |
6764 | auto SPTmp = MIRBuilder.buildCopy(PtrTy, SPReg); |
6765 | SPTmp = MIRBuilder.buildCast(IntPtrTy, SPTmp); |
6766 | |
6767 | // Subtract the final alloc from the SP. We use G_PTRTOINT here so we don't |
6768 | // have to generate an extra instruction to negate the alloc and then use |
6769 | // G_PTR_ADD to add the negative offset. |
6770 | auto Alloc = MIRBuilder.buildSub(IntPtrTy, SPTmp, AllocSize); |
6771 | if (Alignment > Align(1)) { |
6772 | APInt AlignMask(IntPtrTy.getSizeInBits(), Alignment.value(), true); |
6773 | AlignMask.negate(); |
6774 | auto AlignCst = MIRBuilder.buildConstant(IntPtrTy, AlignMask); |
6775 | Alloc = MIRBuilder.buildAnd(IntPtrTy, Alloc, AlignCst); |
6776 | } |
6777 | |
6778 | SPTmp = MIRBuilder.buildCast(PtrTy, Alloc); |
6779 | MIRBuilder.buildCopy(SPReg, SPTmp); |
6780 | MIRBuilder.buildCopy(Dst, SPTmp); |
6781 | |
6782 | MI.eraseFromParent(); |
6783 | return Legalized; |
6784 | } |
6785 | |
6786 | LegalizerHelper::LegalizeResult |
6787 | LegalizerHelper::lowerExtract(MachineInstr &MI) { |
6788 | Register Dst = MI.getOperand(0).getReg(); |
6789 | Register Src = MI.getOperand(1).getReg(); |
6790 | unsigned Offset = MI.getOperand(2).getImm(); |
6791 | |
6792 | LLT DstTy = MRI.getType(Dst); |
6793 | LLT SrcTy = MRI.getType(Src); |
6794 | |
6795 | if (DstTy.isScalar() && |
6796 | (SrcTy.isScalar() || |
6797 | (SrcTy.isVector() && DstTy == SrcTy.getElementType()))) { |
6798 | LLT SrcIntTy = SrcTy; |
6799 | if (!SrcTy.isScalar()) { |
6800 | SrcIntTy = LLT::scalar(SrcTy.getSizeInBits()); |
6801 | Src = MIRBuilder.buildBitcast(SrcIntTy, Src).getReg(0); |
6802 | } |
6803 | |
6804 | if (Offset == 0) |
6805 | MIRBuilder.buildTrunc(Dst, Src); |
6806 | else { |
6807 | auto ShiftAmt = MIRBuilder.buildConstant(SrcIntTy, Offset); |
6808 | auto Shr = MIRBuilder.buildLShr(SrcIntTy, Src, ShiftAmt); |
6809 | MIRBuilder.buildTrunc(Dst, Shr); |
6810 | } |
6811 | |
6812 | MI.eraseFromParent(); |
6813 | return Legalized; |
6814 | } |
6815 | |
6816 | return UnableToLegalize; |
6817 | } |
6818 | |
6819 | LegalizerHelper::LegalizeResult LegalizerHelper::lowerInsert(MachineInstr &MI) { |
6820 | Register Dst = MI.getOperand(0).getReg(); |
6821 | Register Src = MI.getOperand(1).getReg(); |
6822 | Register InsertSrc = MI.getOperand(2).getReg(); |
6823 | uint64_t Offset = MI.getOperand(3).getImm(); |
6824 | |
6825 | LLT DstTy = MRI.getType(Src); |
6826 | LLT InsertTy = MRI.getType(InsertSrc); |
6827 | |
6828 | if (InsertTy.isVector() || |
6829 | (DstTy.isVector() && DstTy.getElementType() != InsertTy)) |
6830 | return UnableToLegalize; |
6831 | |
6832 | const DataLayout &DL = MIRBuilder.getDataLayout(); |
6833 | if ((DstTy.isPointer() && |
6834 | DL.isNonIntegralAddressSpace(DstTy.getAddressSpace())) || |
6835 | (InsertTy.isPointer() && |
6836 | DL.isNonIntegralAddressSpace(InsertTy.getAddressSpace()))) { |
6837 | LLVM_DEBUG(dbgs() << "Not casting non-integral address space integer\n")do { } while (false); |
6838 | return UnableToLegalize; |
6839 | } |
6840 | |
6841 | LLT IntDstTy = DstTy; |
6842 | |
6843 | if (!DstTy.isScalar()) { |
6844 | IntDstTy = LLT::scalar(DstTy.getSizeInBits()); |
6845 | Src = MIRBuilder.buildCast(IntDstTy, Src).getReg(0); |
6846 | } |
6847 | |
6848 | if (!InsertTy.isScalar()) { |
6849 | const LLT IntInsertTy = LLT::scalar(InsertTy.getSizeInBits()); |
6850 | InsertSrc = MIRBuilder.buildPtrToInt(IntInsertTy, InsertSrc).getReg(0); |
6851 | } |
6852 | |
6853 | Register ExtInsSrc = MIRBuilder.buildZExt(IntDstTy, InsertSrc).getReg(0); |
6854 | if (Offset != 0) { |
6855 | auto ShiftAmt = MIRBuilder.buildConstant(IntDstTy, Offset); |
6856 | ExtInsSrc = MIRBuilder.buildShl(IntDstTy, ExtInsSrc, ShiftAmt).getReg(0); |
6857 | } |
6858 | |
6859 | APInt MaskVal = APInt::getBitsSetWithWrap( |
6860 | DstTy.getSizeInBits(), Offset + InsertTy.getSizeInBits(), Offset); |
6861 | |
6862 | auto Mask = MIRBuilder.buildConstant(IntDstTy, MaskVal); |
6863 | auto MaskedSrc = MIRBuilder.buildAnd(IntDstTy, Src, Mask); |
6864 | auto Or = MIRBuilder.buildOr(IntDstTy, MaskedSrc, ExtInsSrc); |
6865 | |
6866 | MIRBuilder.buildCast(Dst, Or); |
6867 | MI.eraseFromParent(); |
6868 | return Legalized; |
6869 | } |
6870 | |
6871 | LegalizerHelper::LegalizeResult |
6872 | LegalizerHelper::lowerSADDO_SSUBO(MachineInstr &MI) { |
6873 | Register Dst0 = MI.getOperand(0).getReg(); |
6874 | Register Dst1 = MI.getOperand(1).getReg(); |
6875 | Register LHS = MI.getOperand(2).getReg(); |
6876 | Register RHS = MI.getOperand(3).getReg(); |
6877 | const bool IsAdd = MI.getOpcode() == TargetOpcode::G_SADDO; |
6878 | |
6879 | LLT Ty = MRI.getType(Dst0); |
6880 | LLT BoolTy = MRI.getType(Dst1); |
6881 | |
6882 | if (IsAdd) |
6883 | MIRBuilder.buildAdd(Dst0, LHS, RHS); |
6884 | else |
6885 | MIRBuilder.buildSub(Dst0, LHS, RHS); |
6886 | |
6887 | // TODO: If SADDSAT/SSUBSAT is legal, compare results to detect overflow. |
6888 | |
6889 | auto Zero = MIRBuilder.buildConstant(Ty, 0); |
6890 | |
6891 | // For an addition, the result should be less than one of the operands (LHS) |
6892 | // if and only if the other operand (RHS) is negative, otherwise there will |
6893 | // be overflow. |
6894 | // For a subtraction, the result should be less than one of the operands |
6895 | // (LHS) if and only if the other operand (RHS) is (non-zero) positive, |
6896 | // otherwise there will be overflow. |
6897 | auto ResultLowerThanLHS = |
6898 | MIRBuilder.buildICmp(CmpInst::ICMP_SLT, BoolTy, Dst0, LHS); |
6899 | auto ConditionRHS = MIRBuilder.buildICmp( |
6900 | IsAdd ? CmpInst::ICMP_SLT : CmpInst::ICMP_SGT, BoolTy, RHS, Zero); |
6901 | |
6902 | MIRBuilder.buildXor(Dst1, ConditionRHS, ResultLowerThanLHS); |
6903 | MI.eraseFromParent(); |
6904 | return Legalized; |
6905 | } |
6906 | |
6907 | LegalizerHelper::LegalizeResult |
6908 | LegalizerHelper::lowerAddSubSatToMinMax(MachineInstr &MI) { |
6909 | Register Res = MI.getOperand(0).getReg(); |
6910 | Register LHS = MI.getOperand(1).getReg(); |
6911 | Register RHS = MI.getOperand(2).getReg(); |
6912 | LLT Ty = MRI.getType(Res); |
6913 | bool IsSigned; |
6914 | bool IsAdd; |
6915 | unsigned BaseOp; |
6916 | switch (MI.getOpcode()) { |
6917 | default: |
6918 | llvm_unreachable("unexpected addsat/subsat opcode")__builtin_unreachable(); |
6919 | case TargetOpcode::G_UADDSAT: |
6920 | IsSigned = false; |
6921 | IsAdd = true; |
6922 | BaseOp = TargetOpcode::G_ADD; |
6923 | break; |
6924 | case TargetOpcode::G_SADDSAT: |
6925 | IsSigned = true; |
6926 | IsAdd = true; |
6927 | BaseOp = TargetOpcode::G_ADD; |
6928 | break; |
6929 | case TargetOpcode::G_USUBSAT: |
6930 | IsSigned = false; |
6931 | IsAdd = false; |
6932 | BaseOp = TargetOpcode::G_SUB; |
6933 | break; |
6934 | case TargetOpcode::G_SSUBSAT: |
6935 | IsSigned = true; |
6936 | IsAdd = false; |
6937 | BaseOp = TargetOpcode::G_SUB; |
6938 | break; |
6939 | } |
6940 | |
6941 | if (IsSigned) { |
6942 | // sadd.sat(a, b) -> |
6943 | // hi = 0x7fffffff - smax(a, 0) |
6944 | // lo = 0x80000000 - smin(a, 0) |
6945 | // a + smin(smax(lo, b), hi) |
6946 | // ssub.sat(a, b) -> |
6947 | // lo = smax(a, -1) - 0x7fffffff |
6948 | // hi = smin(a, -1) - 0x80000000 |
6949 | // a - smin(smax(lo, b), hi) |
6950 | // TODO: AMDGPU can use a "median of 3" instruction here: |
6951 | // a +/- med3(lo, b, hi) |
6952 | uint64_t NumBits = Ty.getScalarSizeInBits(); |
6953 | auto MaxVal = |
6954 | MIRBuilder.buildConstant(Ty, APInt::getSignedMaxValue(NumBits)); |
6955 | auto MinVal = |
6956 | MIRBuilder.buildConstant(Ty, APInt::getSignedMinValue(NumBits)); |
6957 | MachineInstrBuilder Hi, Lo; |
6958 | if (IsAdd) { |
6959 | auto Zero = MIRBuilder.buildConstant(Ty, 0); |
6960 | Hi = MIRBuilder.buildSub(Ty, MaxVal, MIRBuilder.buildSMax(Ty, LHS, Zero)); |
6961 | Lo = MIRBuilder.buildSub(Ty, MinVal, MIRBuilder.buildSMin(Ty, LHS, Zero)); |
6962 | } else { |
6963 | auto NegOne = MIRBuilder.buildConstant(Ty, -1); |
6964 | Lo = MIRBuilder.buildSub(Ty, MIRBuilder.buildSMax(Ty, LHS, NegOne), |
6965 | MaxVal); |
6966 | Hi = MIRBuilder.buildSub(Ty, MIRBuilder.buildSMin(Ty, LHS, NegOne), |
6967 | MinVal); |
6968 | } |
6969 | auto RHSClamped = |
6970 | MIRBuilder.buildSMin(Ty, MIRBuilder.buildSMax(Ty, Lo, RHS), Hi); |
6971 | MIRBuilder.buildInstr(BaseOp, {Res}, {LHS, RHSClamped}); |
6972 | } else { |
6973 | // uadd.sat(a, b) -> a + umin(~a, b) |
6974 | // usub.sat(a, b) -> a - umin(a, b) |
6975 | Register Not = IsAdd ? MIRBuilder.buildNot(Ty, LHS).getReg(0) : LHS; |
6976 | auto Min = MIRBuilder.buildUMin(Ty, Not, RHS); |
6977 | MIRBuilder.buildInstr(BaseOp, {Res}, {LHS, Min}); |
6978 | } |
6979 | |
6980 | MI.eraseFromParent(); |
6981 | return Legalized; |
6982 | } |
6983 | |
6984 | LegalizerHelper::LegalizeResult |
6985 | LegalizerHelper::lowerAddSubSatToAddoSubo(MachineInstr &MI) { |
6986 | Register Res = MI.getOperand(0).getReg(); |
6987 | Register LHS = MI.getOperand(1).getReg(); |
6988 | Register RHS = MI.getOperand(2).getReg(); |
6989 | LLT Ty = MRI.getType(Res); |
6990 | LLT BoolTy = Ty.changeElementSize(1); |
6991 | bool IsSigned; |
6992 | bool IsAdd; |
6993 | unsigned OverflowOp; |
6994 | switch (MI.getOpcode()) { |
6995 | default: |
6996 | llvm_unreachable("unexpected addsat/subsat opcode")__builtin_unreachable(); |
6997 | case TargetOpcode::G_UADDSAT: |
6998 | IsSigned = false; |
6999 | IsAdd = true; |
7000 | OverflowOp = TargetOpcode::G_UADDO; |
7001 | break; |
7002 | case TargetOpcode::G_SADDSAT: |
7003 | IsSigned = true; |
7004 | IsAdd = true; |
7005 | OverflowOp = TargetOpcode::G_SADDO; |
7006 | break; |
7007 | case TargetOpcode::G_USUBSAT: |
7008 | IsSigned = false; |
7009 | IsAdd = false; |
7010 | OverflowOp = TargetOpcode::G_USUBO; |
7011 | break; |
7012 | case TargetOpcode::G_SSUBSAT: |
7013 | IsSigned = true; |
7014 | IsAdd = false; |
7015 | OverflowOp = TargetOpcode::G_SSUBO; |
7016 | break; |
7017 | } |
7018 | |
7019 | auto OverflowRes = |
7020 | MIRBuilder.buildInstr(OverflowOp, {Ty, BoolTy}, {LHS, RHS}); |
7021 | Register Tmp = OverflowRes.getReg(0); |
7022 | Register Ov = OverflowRes.getReg(1); |
7023 | MachineInstrBuilder Clamp; |
7024 | if (IsSigned) { |
7025 | // sadd.sat(a, b) -> |
7026 | // {tmp, ov} = saddo(a, b) |
7027 | // ov ? (tmp >>s 31) + 0x80000000 : r |
7028 | // ssub.sat(a, b) -> |
7029 | // {tmp, ov} = ssubo(a, b) |
7030 | // ov ? (tmp >>s 31) + 0x80000000 : r |
7031 | uint64_t NumBits = Ty.getScalarSizeInBits(); |
7032 | auto ShiftAmount = MIRBuilder.buildConstant(Ty, NumBits - 1); |
7033 | auto Sign = MIRBuilder.buildAShr(Ty, Tmp, ShiftAmount); |
7034 | auto MinVal = |
7035 | MIRBuilder.buildConstant(Ty, APInt::getSignedMinValue(NumBits)); |
7036 | Clamp = MIRBuilder.buildAdd(Ty, Sign, MinVal); |
7037 | } else { |
7038 | // uadd.sat(a, b) -> |
7039 | // {tmp, ov} = uaddo(a, b) |
7040 | // ov ? 0xffffffff : tmp |
7041 | // usub.sat(a, b) -> |
7042 | // {tmp, ov} = usubo(a, b) |
7043 | // ov ? 0 : tmp |
7044 | Clamp = MIRBuilder.buildConstant(Ty, IsAdd ? -1 : 0); |
7045 | } |
7046 | MIRBuilder.buildSelect(Res, Ov, Clamp, Tmp); |
7047 | |
7048 | MI.eraseFromParent(); |
7049 | return Legalized; |
7050 | } |
7051 | |
7052 | LegalizerHelper::LegalizeResult |
7053 | LegalizerHelper::lowerShlSat(MachineInstr &MI) { |
7054 | assert((MI.getOpcode() == TargetOpcode::G_SSHLSAT ||((void)0) |
7055 | MI.getOpcode() == TargetOpcode::G_USHLSAT) &&((void)0) |
7056 | "Expected shlsat opcode!")((void)0); |
7057 | bool IsSigned = MI.getOpcode() == TargetOpcode::G_SSHLSAT; |
7058 | Register Res = MI.getOperand(0).getReg(); |
7059 | Register LHS = MI.getOperand(1).getReg(); |
7060 | Register RHS = MI.getOperand(2).getReg(); |
7061 | LLT Ty = MRI.getType(Res); |
7062 | LLT BoolTy = Ty.changeElementSize(1); |
7063 | |
7064 | unsigned BW = Ty.getScalarSizeInBits(); |
7065 | auto Result = MIRBuilder.buildShl(Ty, LHS, RHS); |
7066 | auto Orig = IsSigned ? MIRBuilder.buildAShr(Ty, Result, RHS) |
7067 | : MIRBuilder.buildLShr(Ty, Result, RHS); |
7068 | |
7069 | MachineInstrBuilder SatVal; |
7070 | if (IsSigned) { |
7071 | auto SatMin = MIRBuilder.buildConstant(Ty, APInt::getSignedMinValue(BW)); |
7072 | auto SatMax = MIRBuilder.buildConstant(Ty, APInt::getSignedMaxValue(BW)); |
7073 | auto Cmp = MIRBuilder.buildICmp(CmpInst::ICMP_SLT, BoolTy, LHS, |
7074 | MIRBuilder.buildConstant(Ty, 0)); |
7075 | SatVal = MIRBuilder.buildSelect(Ty, Cmp, SatMin, SatMax); |
7076 | } else { |
7077 | SatVal = MIRBuilder.buildConstant(Ty, APInt::getMaxValue(BW)); |
7078 | } |
7079 | auto Ov = MIRBuilder.buildICmp(CmpInst::ICMP_NE, BoolTy, LHS, Orig); |
7080 | MIRBuilder.buildSelect(Res, Ov, SatVal, Result); |
7081 | |
7082 | MI.eraseFromParent(); |
7083 | return Legalized; |
7084 | } |
7085 | |
7086 | LegalizerHelper::LegalizeResult |
7087 | LegalizerHelper::lowerBswap(MachineInstr &MI) { |
7088 | Register Dst = MI.getOperand(0).getReg(); |
7089 | Register Src = MI.getOperand(1).getReg(); |
7090 | const LLT Ty = MRI.getType(Src); |
7091 | unsigned SizeInBytes = (Ty.getScalarSizeInBits() + 7) / 8; |
7092 | unsigned BaseShiftAmt = (SizeInBytes - 1) * 8; |
7093 | |
7094 | // Swap most and least significant byte, set remaining bytes in Res to zero. |
7095 | auto ShiftAmt = MIRBuilder.buildConstant(Ty, BaseShiftAmt); |
7096 | auto LSByteShiftedLeft = MIRBuilder.buildShl(Ty, Src, ShiftAmt); |
7097 | auto MSByteShiftedRight = MIRBuilder.buildLShr(Ty, Src, ShiftAmt); |
7098 | auto Res = MIRBuilder.buildOr(Ty, MSByteShiftedRight, LSByteShiftedLeft); |
7099 | |
7100 | // Set i-th high/low byte in Res to i-th low/high byte from Src. |
7101 | for (unsigned i = 1; i < SizeInBytes / 2; ++i) { |
7102 | // AND with Mask leaves byte i unchanged and sets remaining bytes to 0. |
7103 | APInt APMask(SizeInBytes * 8, 0xFF << (i * 8)); |
7104 | auto Mask = MIRBuilder.buildConstant(Ty, APMask); |
7105 | auto ShiftAmt = MIRBuilder.buildConstant(Ty, BaseShiftAmt - 16 * i); |
7106 | // Low byte shifted left to place of high byte: (Src & Mask) << ShiftAmt. |
7107 | auto LoByte = MIRBuilder.buildAnd(Ty, Src, Mask); |
7108 | auto LoShiftedLeft = MIRBuilder.buildShl(Ty, LoByte, ShiftAmt); |
7109 | Res = MIRBuilder.buildOr(Ty, Res, LoShiftedLeft); |
7110 | // High byte shifted right to place of low byte: (Src >> ShiftAmt) & Mask. |
7111 | auto SrcShiftedRight = MIRBuilder.buildLShr(Ty, Src, ShiftAmt); |
7112 | auto HiShiftedRight = MIRBuilder.buildAnd(Ty, SrcShiftedRight, Mask); |
7113 | Res = MIRBuilder.buildOr(Ty, Res, HiShiftedRight); |
7114 | } |
7115 | Res.getInstr()->getOperand(0).setReg(Dst); |
7116 | |
7117 | MI.eraseFromParent(); |
7118 | return Legalized; |
7119 | } |
7120 | |
7121 | //{ (Src & Mask) >> N } | { (Src << N) & Mask } |
7122 | static MachineInstrBuilder SwapN(unsigned N, DstOp Dst, MachineIRBuilder &B, |
7123 | MachineInstrBuilder Src, APInt Mask) { |
7124 | const LLT Ty = Dst.getLLTTy(*B.getMRI()); |
7125 | MachineInstrBuilder C_N = B.buildConstant(Ty, N); |
7126 | MachineInstrBuilder MaskLoNTo0 = B.buildConstant(Ty, Mask); |
7127 | auto LHS = B.buildLShr(Ty, B.buildAnd(Ty, Src, MaskLoNTo0), C_N); |
7128 | auto RHS = B.buildAnd(Ty, B.buildShl(Ty, Src, C_N), MaskLoNTo0); |
7129 | return B.buildOr(Dst, LHS, RHS); |
7130 | } |
7131 | |
7132 | LegalizerHelper::LegalizeResult |
7133 | LegalizerHelper::lowerBitreverse(MachineInstr &MI) { |
7134 | Register Dst = MI.getOperand(0).getReg(); |
7135 | Register Src = MI.getOperand(1).getReg(); |
7136 | const LLT Ty = MRI.getType(Src); |
7137 | unsigned Size = Ty.getSizeInBits(); |
7138 | |
7139 | MachineInstrBuilder BSWAP = |
7140 | MIRBuilder.buildInstr(TargetOpcode::G_BSWAP, {Ty}, {Src}); |
7141 | |
7142 | // swap high and low 4 bits in 8 bit blocks 7654|3210 -> 3210|7654 |
7143 | // [(val & 0xF0F0F0F0) >> 4] | [(val & 0x0F0F0F0F) << 4] |
7144 | // -> [(val & 0xF0F0F0F0) >> 4] | [(val << 4) & 0xF0F0F0F0] |
7145 | MachineInstrBuilder Swap4 = |
7146 | SwapN(4, Ty, MIRBuilder, BSWAP, APInt::getSplat(Size, APInt(8, 0xF0))); |
7147 | |
7148 | // swap high and low 2 bits in 4 bit blocks 32|10 76|54 -> 10|32 54|76 |
7149 | // [(val & 0xCCCCCCCC) >> 2] & [(val & 0x33333333) << 2] |
7150 | // -> [(val & 0xCCCCCCCC) >> 2] & [(val << 2) & 0xCCCCCCCC] |
7151 | MachineInstrBuilder Swap2 = |
7152 | SwapN(2, Ty, MIRBuilder, Swap4, APInt::getSplat(Size, APInt(8, 0xCC))); |
7153 | |
7154 | // swap high and low 1 bit in 2 bit blocks 1|0 3|2 5|4 7|6 -> 0|1 2|3 4|5 6|7 |
7155 | // [(val & 0xAAAAAAAA) >> 1] & [(val & 0x55555555) << 1] |
7156 | // -> [(val & 0xAAAAAAAA) >> 1] & [(val << 1) & 0xAAAAAAAA] |
7157 | SwapN(1, Dst, MIRBuilder, Swap2, APInt::getSplat(Size, APInt(8, 0xAA))); |
7158 | |
7159 | MI.eraseFromParent(); |
7160 | return Legalized; |
7161 | } |
7162 | |
7163 | LegalizerHelper::LegalizeResult |
7164 | LegalizerHelper::lowerReadWriteRegister(MachineInstr &MI) { |
7165 | MachineFunction &MF = MIRBuilder.getMF(); |
7166 | |
7167 | bool IsRead = MI.getOpcode() == TargetOpcode::G_READ_REGISTER; |
7168 | int NameOpIdx = IsRead ? 1 : 0; |
7169 | int ValRegIndex = IsRead ? 0 : 1; |
7170 | |
7171 | Register ValReg = MI.getOperand(ValRegIndex).getReg(); |
7172 | const LLT Ty = MRI.getType(ValReg); |
7173 | const MDString *RegStr = cast<MDString>( |
7174 | cast<MDNode>(MI.getOperand(NameOpIdx).getMetadata())->getOperand(0)); |
7175 | |
7176 | Register PhysReg = TLI.getRegisterByName(RegStr->getString().data(), Ty, MF); |
7177 | if (!PhysReg.isValid()) |
7178 | return UnableToLegalize; |
7179 | |
7180 | if (IsRead) |
7181 | MIRBuilder.buildCopy(ValReg, PhysReg); |
7182 | else |
7183 | MIRBuilder.buildCopy(PhysReg, ValReg); |
7184 | |
7185 | MI.eraseFromParent(); |
7186 | return Legalized; |
7187 | } |
7188 | |
7189 | LegalizerHelper::LegalizeResult |
7190 | LegalizerHelper::lowerSMULH_UMULH(MachineInstr &MI) { |
7191 | bool IsSigned = MI.getOpcode() == TargetOpcode::G_SMULH; |
7192 | unsigned ExtOp = IsSigned ? TargetOpcode::G_SEXT : TargetOpcode::G_ZEXT; |
7193 | Register Result = MI.getOperand(0).getReg(); |
7194 | LLT OrigTy = MRI.getType(Result); |
7195 | auto SizeInBits = OrigTy.getScalarSizeInBits(); |
7196 | LLT WideTy = OrigTy.changeElementSize(SizeInBits * 2); |
7197 | |
7198 | auto LHS = MIRBuilder.buildInstr(ExtOp, {WideTy}, {MI.getOperand(1)}); |
7199 | auto RHS = MIRBuilder.buildInstr(ExtOp, {WideTy}, {MI.getOperand(2)}); |
7200 | auto Mul = MIRBuilder.buildMul(WideTy, LHS, RHS); |
7201 | unsigned ShiftOp = IsSigned ? TargetOpcode::G_ASHR : TargetOpcode::G_LSHR; |
7202 | |
7203 | auto ShiftAmt = MIRBuilder.buildConstant(WideTy, SizeInBits); |
7204 | auto Shifted = MIRBuilder.buildInstr(ShiftOp, {WideTy}, {Mul, ShiftAmt}); |
7205 | MIRBuilder.buildTrunc(Result, Shifted); |
7206 | |
7207 | MI.eraseFromParent(); |
7208 | return Legalized; |
7209 | } |
7210 | |
7211 | LegalizerHelper::LegalizeResult LegalizerHelper::lowerSelect(MachineInstr &MI) { |
7212 | // Implement vector G_SELECT in terms of XOR, AND, OR. |
7213 | Register DstReg = MI.getOperand(0).getReg(); |
7214 | Register MaskReg = MI.getOperand(1).getReg(); |
7215 | Register Op1Reg = MI.getOperand(2).getReg(); |
7216 | Register Op2Reg = MI.getOperand(3).getReg(); |
7217 | LLT DstTy = MRI.getType(DstReg); |
7218 | LLT MaskTy = MRI.getType(MaskReg); |
7219 | LLT Op1Ty = MRI.getType(Op1Reg); |
7220 | if (!DstTy.isVector()) |
7221 | return UnableToLegalize; |
7222 | |
7223 | // Vector selects can have a scalar predicate. If so, splat into a vector and |
7224 | // finish for later legalization attempts to try again. |
7225 | if (MaskTy.isScalar()) { |
7226 | Register MaskElt = MaskReg; |
7227 | if (MaskTy.getSizeInBits() < DstTy.getScalarSizeInBits()) |
7228 | MaskElt = MIRBuilder.buildSExt(DstTy.getElementType(), MaskElt).getReg(0); |
7229 | // Generate a vector splat idiom to be pattern matched later. |
7230 | auto ShufSplat = MIRBuilder.buildShuffleSplat(DstTy, MaskElt); |
7231 | Observer.changingInstr(MI); |
7232 | MI.getOperand(1).setReg(ShufSplat.getReg(0)); |
7233 | Observer.changedInstr(MI); |
7234 | return Legalized; |
7235 | } |
7236 | |
7237 | if (MaskTy.getSizeInBits() != Op1Ty.getSizeInBits()) { |
7238 | return UnableToLegalize; |
7239 | } |
7240 | |
7241 | auto NotMask = MIRBuilder.buildNot(MaskTy, MaskReg); |
7242 | auto NewOp1 = MIRBuilder.buildAnd(MaskTy, Op1Reg, MaskReg); |
7243 | auto NewOp2 = MIRBuilder.buildAnd(MaskTy, Op2Reg, NotMask); |
7244 | MIRBuilder.buildOr(DstReg, NewOp1, NewOp2); |
7245 | MI.eraseFromParent(); |
7246 | return Legalized; |
7247 | } |
7248 | |
7249 | LegalizerHelper::LegalizeResult LegalizerHelper::lowerDIVREM(MachineInstr &MI) { |
7250 | // Split DIVREM into individual instructions. |
7251 | unsigned Opcode = MI.getOpcode(); |
7252 | |
7253 | MIRBuilder.buildInstr( |
7254 | Opcode == TargetOpcode::G_SDIVREM ? TargetOpcode::G_SDIV |
7255 | : TargetOpcode::G_UDIV, |
7256 | {MI.getOperand(0).getReg()}, {MI.getOperand(2), MI.getOperand(3)}); |
7257 | MIRBuilder.buildInstr( |
7258 | Opcode == TargetOpcode::G_SDIVREM ? TargetOpcode::G_SREM |
7259 | : TargetOpcode::G_UREM, |
7260 | {MI.getOperand(1).getReg()}, {MI.getOperand(2), MI.getOperand(3)}); |
7261 | MI.eraseFromParent(); |
7262 | return Legalized; |
7263 | } |
7264 | |
7265 | LegalizerHelper::LegalizeResult |
7266 | LegalizerHelper::lowerAbsToAddXor(MachineInstr &MI) { |
7267 | // Expand %res = G_ABS %a into: |
7268 | // %v1 = G_ASHR %a, scalar_size-1 |
7269 | // %v2 = G_ADD %a, %v1 |
7270 | // %res = G_XOR %v2, %v1 |
7271 | LLT DstTy = MRI.getType(MI.getOperand(0).getReg()); |
7272 | Register OpReg = MI.getOperand(1).getReg(); |
7273 | auto ShiftAmt = |
7274 | MIRBuilder.buildConstant(DstTy, DstTy.getScalarSizeInBits() - 1); |
7275 | auto Shift = MIRBuilder.buildAShr(DstTy, OpReg, ShiftAmt); |
7276 | auto Add = MIRBuilder.buildAdd(DstTy, OpReg, Shift); |
7277 | MIRBuilder.buildXor(MI.getOperand(0).getReg(), Add, Shift); |
7278 | MI.eraseFromParent(); |
7279 | return Legalized; |
7280 | } |
7281 | |
7282 | LegalizerHelper::LegalizeResult |
7283 | LegalizerHelper::lowerAbsToMaxNeg(MachineInstr &MI) { |
7284 | // Expand %res = G_ABS %a into: |
7285 | // %v1 = G_CONSTANT 0 |
7286 | // %v2 = G_SUB %v1, %a |
7287 | // %res = G_SMAX %a, %v2 |
7288 | Register SrcReg = MI.getOperand(1).getReg(); |
7289 | LLT Ty = MRI.getType(SrcReg); |
7290 | auto Zero = MIRBuilder.buildConstant(Ty, 0).getReg(0); |
7291 | auto Sub = MIRBuilder.buildSub(Ty, Zero, SrcReg).getReg(0); |
7292 | MIRBuilder.buildSMax(MI.getOperand(0), SrcReg, Sub); |
7293 | MI.eraseFromParent(); |
7294 | return Legalized; |
7295 | } |