Bug Summary

File:src/gnu/usr.bin/clang/libclangCodeGen/../../../llvm/llvm/include/llvm/Support/MathExtras.h
Warning:line 730, column 8
Division by zero

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -cc1 -cc1 -triple amd64-unknown-openbsd7.0 -analyze -disable-free -disable-llvm-verifier -discard-value-names -main-file-name CGBlocks.cpp -analyzer-store=region -analyzer-opt-analyze-nested-blocks -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=cplusplus -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -setup-static-analyzer -mrelocation-model static -mframe-pointer=all -relaxed-aliasing -fno-rounding-math -mconstructor-aliases -munwind-tables -target-cpu x86-64 -tune-cpu generic -debugger-tuning=gdb -fcoverage-compilation-dir=/usr/src/gnu/usr.bin/clang/libclangCodeGen/obj -resource-dir /usr/local/lib/clang/13.0.0 -I /usr/src/gnu/usr.bin/clang/libclangCodeGen/../../../llvm/clang/include -I /usr/src/gnu/usr.bin/clang/libclangCodeGen/../../../llvm/llvm/include -I /usr/src/gnu/usr.bin/clang/libclangCodeGen/../include -I /usr/src/gnu/usr.bin/clang/libclangCodeGen/obj -I /usr/src/gnu/usr.bin/clang/libclangCodeGen/obj/../include -D NDEBUG -D __STDC_LIMIT_MACROS -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D LLVM_PREFIX="/usr" -internal-isystem /usr/include/c++/v1 -internal-isystem /usr/local/lib/clang/13.0.0/include -internal-externc-isystem /usr/include -O2 -Wno-unused-parameter -Wwrite-strings -Wno-missing-field-initializers -Wno-long-long -Wno-comment -std=c++14 -fdeprecated-macro -fdebug-compilation-dir=/usr/src/gnu/usr.bin/clang/libclangCodeGen/obj -ferror-limit 19 -fvisibility-inlines-hidden -fwrapv -stack-protector 2 -fno-rtti -fgnuc-version=4.2.1 -vectorize-loops -vectorize-slp -fno-builtin-malloc -fno-builtin-calloc -fno-builtin-realloc -fno-builtin-valloc -fno-builtin-free -fno-builtin-strdup -fno-builtin-strndup -analyzer-output=html -faddrsig -D__GCC_HAVE_DWARF2_CFI_ASM=1 -o /home/ben/Projects/vmm/scan-build/2022-01-12-194120-40624-1 -x c++ /usr/src/gnu/usr.bin/clang/libclangCodeGen/../../../llvm/clang/lib/CodeGen/CGBlocks.cpp

/usr/src/gnu/usr.bin/clang/libclangCodeGen/../../../llvm/clang/lib/CodeGen/CGBlocks.cpp

1//===--- CGBlocks.cpp - Emit LLVM Code for declarations ---------*- C++ -*-===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This contains code to emit blocks.
10//
11//===----------------------------------------------------------------------===//
12
13#include "CGBlocks.h"
14#include "CGCXXABI.h"
15#include "CGDebugInfo.h"
16#include "CGObjCRuntime.h"
17#include "CGOpenCLRuntime.h"
18#include "CodeGenFunction.h"
19#include "CodeGenModule.h"
20#include "ConstantEmitter.h"
21#include "TargetInfo.h"
22#include "clang/AST/Attr.h"
23#include "clang/AST/DeclObjC.h"
24#include "clang/CodeGen/ConstantInitBuilder.h"
25#include "llvm/ADT/SmallSet.h"
26#include "llvm/IR/DataLayout.h"
27#include "llvm/IR/Module.h"
28#include "llvm/Support/ScopedPrinter.h"
29#include <algorithm>
30#include <cstdio>
31
32using namespace clang;
33using namespace CodeGen;
34
35CGBlockInfo::CGBlockInfo(const BlockDecl *block, StringRef name)
36 : Name(name), CXXThisIndex(0), CanBeGlobal(false), NeedsCopyDispose(false),
37 HasCXXObject(false), UsesStret(false), HasCapturedVariableLayout(false),
38 CapturesNonExternalType(false), LocalAddress(Address::invalid()),
39 StructureType(nullptr), Block(block) {
40
41 // Skip asm prefix, if any. 'name' is usually taken directly from
42 // the mangled name of the enclosing function.
43 if (!name.empty() && name[0] == '\01')
44 name = name.substr(1);
45}
46
47// Anchor the vtable to this translation unit.
48BlockByrefHelpers::~BlockByrefHelpers() {}
49
50/// Build the given block as a global block.
51static llvm::Constant *buildGlobalBlock(CodeGenModule &CGM,
52 const CGBlockInfo &blockInfo,
53 llvm::Constant *blockFn);
54
55/// Build the helper function to copy a block.
56static llvm::Constant *buildCopyHelper(CodeGenModule &CGM,
57 const CGBlockInfo &blockInfo) {
58 return CodeGenFunction(CGM).GenerateCopyHelperFunction(blockInfo);
59}
60
61/// Build the helper function to dispose of a block.
62static llvm::Constant *buildDisposeHelper(CodeGenModule &CGM,
63 const CGBlockInfo &blockInfo) {
64 return CodeGenFunction(CGM).GenerateDestroyHelperFunction(blockInfo);
65}
66
67namespace {
68
69/// Represents a type of copy/destroy operation that should be performed for an
70/// entity that's captured by a block.
71enum class BlockCaptureEntityKind {
72 CXXRecord, // Copy or destroy
73 ARCWeak,
74 ARCStrong,
75 NonTrivialCStruct,
76 BlockObject, // Assign or release
77 None
78};
79
80/// Represents a captured entity that requires extra operations in order for
81/// this entity to be copied or destroyed correctly.
82struct BlockCaptureManagedEntity {
83 BlockCaptureEntityKind CopyKind, DisposeKind;
84 BlockFieldFlags CopyFlags, DisposeFlags;
85 const BlockDecl::Capture *CI;
86 const CGBlockInfo::Capture *Capture;
87
88 BlockCaptureManagedEntity(BlockCaptureEntityKind CopyType,
89 BlockCaptureEntityKind DisposeType,
90 BlockFieldFlags CopyFlags,
91 BlockFieldFlags DisposeFlags,
92 const BlockDecl::Capture &CI,
93 const CGBlockInfo::Capture &Capture)
94 : CopyKind(CopyType), DisposeKind(DisposeType), CopyFlags(CopyFlags),
95 DisposeFlags(DisposeFlags), CI(&CI), Capture(&Capture) {}
96
97 bool operator<(const BlockCaptureManagedEntity &Other) const {
98 return Capture->getOffset() < Other.Capture->getOffset();
99 }
100};
101
102enum class CaptureStrKind {
103 // String for the copy helper.
104 CopyHelper,
105 // String for the dispose helper.
106 DisposeHelper,
107 // Merge the strings for the copy helper and dispose helper.
108 Merged
109};
110
111} // end anonymous namespace
112
113static void findBlockCapturedManagedEntities(
114 const CGBlockInfo &BlockInfo, const LangOptions &LangOpts,
115 SmallVectorImpl<BlockCaptureManagedEntity> &ManagedCaptures);
116
117static std::string getBlockCaptureStr(const BlockCaptureManagedEntity &E,
118 CaptureStrKind StrKind,
119 CharUnits BlockAlignment,
120 CodeGenModule &CGM);
121
122static std::string getBlockDescriptorName(const CGBlockInfo &BlockInfo,
123 CodeGenModule &CGM) {
124 std::string Name = "__block_descriptor_";
125 Name += llvm::to_string(BlockInfo.BlockSize.getQuantity()) + "_";
126
127 if (BlockInfo.needsCopyDisposeHelpers()) {
128 if (CGM.getLangOpts().Exceptions)
129 Name += "e";
130 if (CGM.getCodeGenOpts().ObjCAutoRefCountExceptions)
131 Name += "a";
132 Name += llvm::to_string(BlockInfo.BlockAlign.getQuantity()) + "_";
133
134 SmallVector<BlockCaptureManagedEntity, 4> ManagedCaptures;
135 findBlockCapturedManagedEntities(BlockInfo, CGM.getContext().getLangOpts(),
136 ManagedCaptures);
137
138 for (const BlockCaptureManagedEntity &E : ManagedCaptures) {
139 Name += llvm::to_string(E.Capture->getOffset().getQuantity());
140
141 if (E.CopyKind == E.DisposeKind) {
142 // If CopyKind and DisposeKind are the same, merge the capture
143 // information.
144 assert(E.CopyKind != BlockCaptureEntityKind::None &&((void)0)
145 "shouldn't see BlockCaptureManagedEntity that is None")((void)0);
146 Name += getBlockCaptureStr(E, CaptureStrKind::Merged,
147 BlockInfo.BlockAlign, CGM);
148 } else {
149 // If CopyKind and DisposeKind are not the same, which can happen when
150 // either Kind is None or the captured object is a __strong block,
151 // concatenate the copy and dispose strings.
152 Name += getBlockCaptureStr(E, CaptureStrKind::CopyHelper,
153 BlockInfo.BlockAlign, CGM);
154 Name += getBlockCaptureStr(E, CaptureStrKind::DisposeHelper,
155 BlockInfo.BlockAlign, CGM);
156 }
157 }
158 Name += "_";
159 }
160
161 std::string TypeAtEncoding =
162 CGM.getContext().getObjCEncodingForBlock(BlockInfo.getBlockExpr());
163 /// Replace occurrences of '@' with '\1'. '@' is reserved on ELF platforms as
164 /// a separator between symbol name and symbol version.
165 std::replace(TypeAtEncoding.begin(), TypeAtEncoding.end(), '@', '\1');
166 Name += "e" + llvm::to_string(TypeAtEncoding.size()) + "_" + TypeAtEncoding;
167 Name += "l" + CGM.getObjCRuntime().getRCBlockLayoutStr(CGM, BlockInfo);
168 return Name;
169}
170
171/// buildBlockDescriptor - Build the block descriptor meta-data for a block.
172/// buildBlockDescriptor is accessed from 5th field of the Block_literal
173/// meta-data and contains stationary information about the block literal.
174/// Its definition will have 4 (or optionally 6) words.
175/// \code
176/// struct Block_descriptor {
177/// unsigned long reserved;
178/// unsigned long size; // size of Block_literal metadata in bytes.
179/// void *copy_func_helper_decl; // optional copy helper.
180/// void *destroy_func_decl; // optional destructor helper.
181/// void *block_method_encoding_address; // @encode for block literal signature.
182/// void *block_layout_info; // encoding of captured block variables.
183/// };
184/// \endcode
185static llvm::Constant *buildBlockDescriptor(CodeGenModule &CGM,
186 const CGBlockInfo &blockInfo) {
187 ASTContext &C = CGM.getContext();
188
189 llvm::IntegerType *ulong =
190 cast<llvm::IntegerType>(CGM.getTypes().ConvertType(C.UnsignedLongTy));
191 llvm::PointerType *i8p = nullptr;
192 if (CGM.getLangOpts().OpenCL)
193 i8p =
194 llvm::Type::getInt8PtrTy(
195 CGM.getLLVMContext(), C.getTargetAddressSpace(LangAS::opencl_constant));
196 else
197 i8p = CGM.VoidPtrTy;
198
199 std::string descName;
200
201 // If an equivalent block descriptor global variable exists, return it.
202 if (C.getLangOpts().ObjC &&
203 CGM.getLangOpts().getGC() == LangOptions::NonGC) {
204 descName = getBlockDescriptorName(blockInfo, CGM);
205 if (llvm::GlobalValue *desc = CGM.getModule().getNamedValue(descName))
206 return llvm::ConstantExpr::getBitCast(desc,
207 CGM.getBlockDescriptorType());
208 }
209
210 // If there isn't an equivalent block descriptor global variable, create a new
211 // one.
212 ConstantInitBuilder builder(CGM);
213 auto elements = builder.beginStruct();
214
215 // reserved
216 elements.addInt(ulong, 0);
217
218 // Size
219 // FIXME: What is the right way to say this doesn't fit? We should give
220 // a user diagnostic in that case. Better fix would be to change the
221 // API to size_t.
222 elements.addInt(ulong, blockInfo.BlockSize.getQuantity());
223
224 // Optional copy/dispose helpers.
225 bool hasInternalHelper = false;
226 if (blockInfo.needsCopyDisposeHelpers()) {
227 // copy_func_helper_decl
228 llvm::Constant *copyHelper = buildCopyHelper(CGM, blockInfo);
229 elements.add(copyHelper);
230
231 // destroy_func_decl
232 llvm::Constant *disposeHelper = buildDisposeHelper(CGM, blockInfo);
233 elements.add(disposeHelper);
234
235 if (cast<llvm::Function>(copyHelper->getOperand(0))->hasInternalLinkage() ||
236 cast<llvm::Function>(disposeHelper->getOperand(0))
237 ->hasInternalLinkage())
238 hasInternalHelper = true;
239 }
240
241 // Signature. Mandatory ObjC-style method descriptor @encode sequence.
242 std::string typeAtEncoding =
243 CGM.getContext().getObjCEncodingForBlock(blockInfo.getBlockExpr());
244 elements.add(llvm::ConstantExpr::getBitCast(
245 CGM.GetAddrOfConstantCString(typeAtEncoding).getPointer(), i8p));
246
247 // GC layout.
248 if (C.getLangOpts().ObjC) {
249 if (CGM.getLangOpts().getGC() != LangOptions::NonGC)
250 elements.add(CGM.getObjCRuntime().BuildGCBlockLayout(CGM, blockInfo));
251 else
252 elements.add(CGM.getObjCRuntime().BuildRCBlockLayout(CGM, blockInfo));
253 }
254 else
255 elements.addNullPointer(i8p);
256
257 unsigned AddrSpace = 0;
258 if (C.getLangOpts().OpenCL)
259 AddrSpace = C.getTargetAddressSpace(LangAS::opencl_constant);
260
261 llvm::GlobalValue::LinkageTypes linkage;
262 if (descName.empty()) {
263 linkage = llvm::GlobalValue::InternalLinkage;
264 descName = "__block_descriptor_tmp";
265 } else if (hasInternalHelper) {
266 // If either the copy helper or the dispose helper has internal linkage,
267 // the block descriptor must have internal linkage too.
268 linkage = llvm::GlobalValue::InternalLinkage;
269 } else {
270 linkage = llvm::GlobalValue::LinkOnceODRLinkage;
271 }
272
273 llvm::GlobalVariable *global =
274 elements.finishAndCreateGlobal(descName, CGM.getPointerAlign(),
275 /*constant*/ true, linkage, AddrSpace);
276
277 if (linkage == llvm::GlobalValue::LinkOnceODRLinkage) {
278 if (CGM.supportsCOMDAT())
279 global->setComdat(CGM.getModule().getOrInsertComdat(descName));
280 global->setVisibility(llvm::GlobalValue::HiddenVisibility);
281 global->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global);
282 }
283
284 return llvm::ConstantExpr::getBitCast(global, CGM.getBlockDescriptorType());
285}
286
287/*
288 Purely notional variadic template describing the layout of a block.
289
290 template <class _ResultType, class... _ParamTypes, class... _CaptureTypes>
291 struct Block_literal {
292 /// Initialized to one of:
293 /// extern void *_NSConcreteStackBlock[];
294 /// extern void *_NSConcreteGlobalBlock[];
295 ///
296 /// In theory, we could start one off malloc'ed by setting
297 /// BLOCK_NEEDS_FREE, giving it a refcount of 1, and using
298 /// this isa:
299 /// extern void *_NSConcreteMallocBlock[];
300 struct objc_class *isa;
301
302 /// These are the flags (with corresponding bit number) that the
303 /// compiler is actually supposed to know about.
304 /// 23. BLOCK_IS_NOESCAPE - indicates that the block is non-escaping
305 /// 25. BLOCK_HAS_COPY_DISPOSE - indicates that the block
306 /// descriptor provides copy and dispose helper functions
307 /// 26. BLOCK_HAS_CXX_OBJ - indicates that there's a captured
308 /// object with a nontrivial destructor or copy constructor
309 /// 28. BLOCK_IS_GLOBAL - indicates that the block is allocated
310 /// as global memory
311 /// 29. BLOCK_USE_STRET - indicates that the block function
312 /// uses stret, which objc_msgSend needs to know about
313 /// 30. BLOCK_HAS_SIGNATURE - indicates that the block has an
314 /// @encoded signature string
315 /// And we're not supposed to manipulate these:
316 /// 24. BLOCK_NEEDS_FREE - indicates that the block has been moved
317 /// to malloc'ed memory
318 /// 27. BLOCK_IS_GC - indicates that the block has been moved to
319 /// to GC-allocated memory
320 /// Additionally, the bottom 16 bits are a reference count which
321 /// should be zero on the stack.
322 int flags;
323
324 /// Reserved; should be zero-initialized.
325 int reserved;
326
327 /// Function pointer generated from block literal.
328 _ResultType (*invoke)(Block_literal *, _ParamTypes...);
329
330 /// Block description metadata generated from block literal.
331 struct Block_descriptor *block_descriptor;
332
333 /// Captured values follow.
334 _CapturesTypes captures...;
335 };
336 */
337
338namespace {
339 /// A chunk of data that we actually have to capture in the block.
340 struct BlockLayoutChunk {
341 CharUnits Alignment;
342 CharUnits Size;
343 Qualifiers::ObjCLifetime Lifetime;
344 const BlockDecl::Capture *Capture; // null for 'this'
345 llvm::Type *Type;
346 QualType FieldType;
347
348 BlockLayoutChunk(CharUnits align, CharUnits size,
349 Qualifiers::ObjCLifetime lifetime,
350 const BlockDecl::Capture *capture,
351 llvm::Type *type, QualType fieldType)
352 : Alignment(align), Size(size), Lifetime(lifetime),
353 Capture(capture), Type(type), FieldType(fieldType) {}
354
355 /// Tell the block info that this chunk has the given field index.
356 void setIndex(CGBlockInfo &info, unsigned index, CharUnits offset) {
357 if (!Capture) {
358 info.CXXThisIndex = index;
359 info.CXXThisOffset = offset;
360 } else {
361 auto C = CGBlockInfo::Capture::makeIndex(index, offset, FieldType);
362 info.Captures.insert({Capture->getVariable(), C});
363 }
364 }
365 };
366
367 /// Order by 1) all __strong together 2) next, all byfref together 3) next,
368 /// all __weak together. Preserve descending alignment in all situations.
369 bool operator<(const BlockLayoutChunk &left, const BlockLayoutChunk &right) {
370 if (left.Alignment != right.Alignment)
371 return left.Alignment > right.Alignment;
372
373 auto getPrefOrder = [](const BlockLayoutChunk &chunk) {
374 if (chunk.Capture && chunk.Capture->isByRef())
375 return 1;
376 if (chunk.Lifetime == Qualifiers::OCL_Strong)
377 return 0;
378 if (chunk.Lifetime == Qualifiers::OCL_Weak)
379 return 2;
380 return 3;
381 };
382
383 return getPrefOrder(left) < getPrefOrder(right);
384 }
385} // end anonymous namespace
386
387/// Determines if the given type is safe for constant capture in C++.
388static bool isSafeForCXXConstantCapture(QualType type) {
389 const RecordType *recordType =
390 type->getBaseElementTypeUnsafe()->getAs<RecordType>();
391
392 // Only records can be unsafe.
393 if (!recordType) return true;
394
395 const auto *record = cast<CXXRecordDecl>(recordType->getDecl());
396
397 // Maintain semantics for classes with non-trivial dtors or copy ctors.
398 if (!record->hasTrivialDestructor()) return false;
399 if (record->hasNonTrivialCopyConstructor()) return false;
400
401 // Otherwise, we just have to make sure there aren't any mutable
402 // fields that might have changed since initialization.
403 return !record->hasMutableFields();
404}
405
406/// It is illegal to modify a const object after initialization.
407/// Therefore, if a const object has a constant initializer, we don't
408/// actually need to keep storage for it in the block; we'll just
409/// rematerialize it at the start of the block function. This is
410/// acceptable because we make no promises about address stability of
411/// captured variables.
412static llvm::Constant *tryCaptureAsConstant(CodeGenModule &CGM,
413 CodeGenFunction *CGF,
414 const VarDecl *var) {
415 // Return if this is a function parameter. We shouldn't try to
416 // rematerialize default arguments of function parameters.
417 if (isa<ParmVarDecl>(var))
418 return nullptr;
419
420 QualType type = var->getType();
421
422 // We can only do this if the variable is const.
423 if (!type.isConstQualified()) return nullptr;
424
425 // Furthermore, in C++ we have to worry about mutable fields:
426 // C++ [dcl.type.cv]p4:
427 // Except that any class member declared mutable can be
428 // modified, any attempt to modify a const object during its
429 // lifetime results in undefined behavior.
430 if (CGM.getLangOpts().CPlusPlus && !isSafeForCXXConstantCapture(type))
431 return nullptr;
432
433 // If the variable doesn't have any initializer (shouldn't this be
434 // invalid?), it's not clear what we should do. Maybe capture as
435 // zero?
436 const Expr *init = var->getInit();
437 if (!init) return nullptr;
438
439 return ConstantEmitter(CGM, CGF).tryEmitAbstractForInitializer(*var);
440}
441
442/// Get the low bit of a nonzero character count. This is the
443/// alignment of the nth byte if the 0th byte is universally aligned.
444static CharUnits getLowBit(CharUnits v) {
445 return CharUnits::fromQuantity(v.getQuantity() & (~v.getQuantity() + 1));
446}
447
448static void initializeForBlockHeader(CodeGenModule &CGM, CGBlockInfo &info,
449 SmallVectorImpl<llvm::Type*> &elementTypes) {
450
451 assert(elementTypes.empty())((void)0);
452 if (CGM.getLangOpts().OpenCL) {
453 // The header is basically 'struct { int; int; generic void *;
454 // custom_fields; }'. Assert that struct is packed.
455 auto GenericAS =
456 CGM.getContext().getTargetAddressSpace(LangAS::opencl_generic);
457 auto GenPtrAlign =
458 CharUnits::fromQuantity(CGM.getTarget().getPointerAlign(GenericAS) / 8);
459 auto GenPtrSize =
460 CharUnits::fromQuantity(CGM.getTarget().getPointerWidth(GenericAS) / 8);
461 assert(CGM.getIntSize() <= GenPtrSize)((void)0);
462 assert(CGM.getIntAlign() <= GenPtrAlign)((void)0);
463 assert((2 * CGM.getIntSize()).isMultipleOf(GenPtrAlign))((void)0);
464 elementTypes.push_back(CGM.IntTy); /* total size */
465 elementTypes.push_back(CGM.IntTy); /* align */
466 elementTypes.push_back(
467 CGM.getOpenCLRuntime()
468 .getGenericVoidPointerType()); /* invoke function */
469 unsigned Offset =
470 2 * CGM.getIntSize().getQuantity() + GenPtrSize.getQuantity();
471 unsigned BlockAlign = GenPtrAlign.getQuantity();
472 if (auto *Helper =
473 CGM.getTargetCodeGenInfo().getTargetOpenCLBlockHelper()) {
474 for (auto I : Helper->getCustomFieldTypes()) /* custom fields */ {
475 // TargetOpenCLBlockHelp needs to make sure the struct is packed.
476 // If necessary, add padding fields to the custom fields.
477 unsigned Align = CGM.getDataLayout().getABITypeAlignment(I);
478 if (BlockAlign < Align)
479 BlockAlign = Align;
480 assert(Offset % Align == 0)((void)0);
481 Offset += CGM.getDataLayout().getTypeAllocSize(I);
482 elementTypes.push_back(I);
483 }
484 }
485 info.BlockAlign = CharUnits::fromQuantity(BlockAlign);
486 info.BlockSize = CharUnits::fromQuantity(Offset);
487 } else {
488 // The header is basically 'struct { void *; int; int; void *; void *; }'.
489 // Assert that the struct is packed.
490 assert(CGM.getIntSize() <= CGM.getPointerSize())((void)0);
491 assert(CGM.getIntAlign() <= CGM.getPointerAlign())((void)0);
492 assert((2 * CGM.getIntSize()).isMultipleOf(CGM.getPointerAlign()))((void)0);
493 info.BlockAlign = CGM.getPointerAlign();
494 info.BlockSize = 3 * CGM.getPointerSize() + 2 * CGM.getIntSize();
495 elementTypes.push_back(CGM.VoidPtrTy);
496 elementTypes.push_back(CGM.IntTy);
497 elementTypes.push_back(CGM.IntTy);
498 elementTypes.push_back(CGM.VoidPtrTy);
499 elementTypes.push_back(CGM.getBlockDescriptorType());
500 }
501}
502
503static QualType getCaptureFieldType(const CodeGenFunction &CGF,
504 const BlockDecl::Capture &CI) {
505 const VarDecl *VD = CI.getVariable();
506
507 // If the variable is captured by an enclosing block or lambda expression,
508 // use the type of the capture field.
509 if (CGF.BlockInfo && CI.isNested())
510 return CGF.BlockInfo->getCapture(VD).fieldType();
511 if (auto *FD = CGF.LambdaCaptureFields.lookup(VD))
512 return FD->getType();
513 // If the captured variable is a non-escaping __block variable, the field
514 // type is the reference type. If the variable is a __block variable that
515 // already has a reference type, the field type is the variable's type.
516 return VD->isNonEscapingByref() ?
517 CGF.getContext().getLValueReferenceType(VD->getType()) : VD->getType();
518}
519
520/// Compute the layout of the given block. Attempts to lay the block
521/// out with minimal space requirements.
522static void computeBlockInfo(CodeGenModule &CGM, CodeGenFunction *CGF,
523 CGBlockInfo &info) {
524 ASTContext &C = CGM.getContext();
525 const BlockDecl *block = info.getBlockDecl();
526
527 SmallVector<llvm::Type*, 8> elementTypes;
528 initializeForBlockHeader(CGM, info, elementTypes);
529 bool hasNonConstantCustomFields = false;
530 if (auto *OpenCLHelper =
4
Assuming 'OpenCLHelper' is null
5
Taking false branch
531 CGM.getTargetCodeGenInfo().getTargetOpenCLBlockHelper())
532 hasNonConstantCustomFields =
533 !OpenCLHelper->areAllCustomFieldValuesConstant(info);
534 if (!block->hasCaptures() && !hasNonConstantCustomFields) {
535 info.StructureType =
536 llvm::StructType::get(CGM.getLLVMContext(), elementTypes, true);
537 info.CanBeGlobal = true;
538 return;
539 }
540 else if (C.getLangOpts().ObjC &&
6
Assuming field 'ObjC' is 0
541 CGM.getLangOpts().getGC() == LangOptions::NonGC)
542 info.HasCapturedVariableLayout = true;
543
544 // Collect the layout chunks.
545 SmallVector<BlockLayoutChunk, 16> layout;
546 layout.reserve(block->capturesCXXThis() +
547 (block->capture_end() - block->capture_begin()));
548
549 CharUnits maxFieldAlign;
7
Calling defaulted default constructor for 'CharUnits'
9
Returning from default constructor for 'CharUnits'
550
551 // First, 'this'.
552 if (block->capturesCXXThis()) {
10
Assuming the condition is false
11
Taking false branch
553 assert(CGF && CGF->CurFuncDecl && isa<CXXMethodDecl>(CGF->CurFuncDecl) &&((void)0)
554 "Can't capture 'this' outside a method")((void)0);
555 QualType thisType = cast<CXXMethodDecl>(CGF->CurFuncDecl)->getThisType();
556
557 // Theoretically, this could be in a different address space, so
558 // don't assume standard pointer size/align.
559 llvm::Type *llvmType = CGM.getTypes().ConvertType(thisType);
560 auto TInfo = CGM.getContext().getTypeInfoInChars(thisType);
561 maxFieldAlign = std::max(maxFieldAlign, TInfo.Align);
562
563 layout.push_back(BlockLayoutChunk(TInfo.Align, TInfo.Width,
564 Qualifiers::OCL_None,
565 nullptr, llvmType, thisType));
566 }
567
568 // Next, all the block captures.
569 for (const auto &CI : block->captures()) {
12
Assuming '__begin1' is equal to '__end1'
570 const VarDecl *variable = CI.getVariable();
571
572 if (CI.isEscapingByref()) {
573 // We have to copy/dispose of the __block reference.
574 info.NeedsCopyDispose = true;
575
576 // Just use void* instead of a pointer to the byref type.
577 CharUnits align = CGM.getPointerAlign();
578 maxFieldAlign = std::max(maxFieldAlign, align);
579
580 // Since a __block variable cannot be captured by lambdas, its type and
581 // the capture field type should always match.
582 assert(CGF && getCaptureFieldType(*CGF, CI) == variable->getType() &&((void)0)
583 "capture type differs from the variable type")((void)0);
584 layout.push_back(BlockLayoutChunk(align, CGM.getPointerSize(),
585 Qualifiers::OCL_None, &CI,
586 CGM.VoidPtrTy, variable->getType()));
587 continue;
588 }
589
590 // Otherwise, build a layout chunk with the size and alignment of
591 // the declaration.
592 if (llvm::Constant *constant = tryCaptureAsConstant(CGM, CGF, variable)) {
593 info.Captures[variable] = CGBlockInfo::Capture::makeConstant(constant);
594 continue;
595 }
596
597 QualType VT = getCaptureFieldType(*CGF, CI);
598
599 // If we have a lifetime qualifier, honor it for capture purposes.
600 // That includes *not* copying it if it's __unsafe_unretained.
601 Qualifiers::ObjCLifetime lifetime = VT.getObjCLifetime();
602 if (lifetime) {
603 switch (lifetime) {
604 case Qualifiers::OCL_None: llvm_unreachable("impossible")__builtin_unreachable();
605 case Qualifiers::OCL_ExplicitNone:
606 case Qualifiers::OCL_Autoreleasing:
607 break;
608
609 case Qualifiers::OCL_Strong:
610 case Qualifiers::OCL_Weak:
611 info.NeedsCopyDispose = true;
612 }
613
614 // Block pointers require copy/dispose. So do Objective-C pointers.
615 } else if (VT->isObjCRetainableType()) {
616 // But honor the inert __unsafe_unretained qualifier, which doesn't
617 // actually make it into the type system.
618 if (VT->isObjCInertUnsafeUnretainedType()) {
619 lifetime = Qualifiers::OCL_ExplicitNone;
620 } else {
621 info.NeedsCopyDispose = true;
622 // used for mrr below.
623 lifetime = Qualifiers::OCL_Strong;
624 }
625
626 // So do types that require non-trivial copy construction.
627 } else if (CI.hasCopyExpr()) {
628 info.NeedsCopyDispose = true;
629 info.HasCXXObject = true;
630 if (!VT->getAsCXXRecordDecl()->isExternallyVisible())
631 info.CapturesNonExternalType = true;
632
633 // So do C structs that require non-trivial copy construction or
634 // destruction.
635 } else if (VT.isNonTrivialToPrimitiveCopy() == QualType::PCK_Struct ||
636 VT.isDestructedType() == QualType::DK_nontrivial_c_struct) {
637 info.NeedsCopyDispose = true;
638
639 // And so do types with destructors.
640 } else if (CGM.getLangOpts().CPlusPlus) {
641 if (const CXXRecordDecl *record = VT->getAsCXXRecordDecl()) {
642 if (!record->hasTrivialDestructor()) {
643 info.HasCXXObject = true;
644 info.NeedsCopyDispose = true;
645 if (!record->isExternallyVisible())
646 info.CapturesNonExternalType = true;
647 }
648 }
649 }
650
651 CharUnits size = C.getTypeSizeInChars(VT);
652 CharUnits align = C.getDeclAlign(variable);
653
654 maxFieldAlign = std::max(maxFieldAlign, align);
655
656 llvm::Type *llvmType =
657 CGM.getTypes().ConvertTypeForMem(VT);
658
659 layout.push_back(
660 BlockLayoutChunk(align, size, lifetime, &CI, llvmType, VT));
661 }
662
663 // If that was everything, we're done here.
664 if (layout.empty()) {
13
Taking false branch
665 info.StructureType =
666 llvm::StructType::get(CGM.getLLVMContext(), elementTypes, true);
667 info.CanBeGlobal = true;
668 return;
669 }
670
671 // Sort the layout by alignment. We have to use a stable sort here
672 // to get reproducible results. There should probably be an
673 // llvm::array_pod_stable_sort.
674 llvm::stable_sort(layout);
675
676 // Needed for blocks layout info.
677 info.BlockHeaderForcedGapOffset = info.BlockSize;
678 info.BlockHeaderForcedGapSize = CharUnits::Zero();
679
680 CharUnits &blockSize = info.BlockSize;
681 info.BlockAlign = std::max(maxFieldAlign, info.BlockAlign);
682
683 // Assuming that the first byte in the header is maximally aligned,
684 // get the alignment of the first byte following the header.
685 CharUnits endAlign = getLowBit(blockSize);
686
687 // If the end of the header isn't satisfactorily aligned for the
688 // maximum thing, look for things that are okay with the header-end
689 // alignment, and keep appending them until we get something that's
690 // aligned right. This algorithm is only guaranteed optimal if
691 // that condition is satisfied at some point; otherwise we can get
692 // things like:
693 // header // next byte has alignment 4
694 // something_with_size_5; // next byte has alignment 1
695 // something_with_alignment_8;
696 // which has 7 bytes of padding, as opposed to the naive solution
697 // which might have less (?).
698 if (endAlign < maxFieldAlign) {
14
Taking true branch
699 SmallVectorImpl<BlockLayoutChunk>::iterator
700 li = layout.begin() + 1, le = layout.end();
701
702 // Look for something that the header end is already
703 // satisfactorily aligned for.
704 for (; li != le && endAlign < li->Alignment; ++li)
15
Assuming 'li' is equal to 'le'
705 ;
706
707 // If we found something that's naturally aligned for the end of
708 // the header, keep adding things...
709 if (li != le) {
16
Assuming 'li' is equal to 'le'
17
Taking false branch
710 SmallVectorImpl<BlockLayoutChunk>::iterator first = li;
711 for (; li != le; ++li) {
712 assert(endAlign >= li->Alignment)((void)0);
713
714 li->setIndex(info, elementTypes.size(), blockSize);
715 elementTypes.push_back(li->Type);
716 blockSize += li->Size;
717 endAlign = getLowBit(blockSize);
718
719 // ...until we get to the alignment of the maximum field.
720 if (endAlign >= maxFieldAlign) {
721 break;
722 }
723 }
724 // Don't re-append everything we just appended.
725 layout.erase(first, li);
726 }
727 }
728
729 assert(endAlign == getLowBit(blockSize))((void)0);
730
731 // At this point, we just have to add padding if the end align still
732 // isn't aligned right.
733 if (endAlign < maxFieldAlign) {
18
Taking true branch
734 CharUnits newBlockSize = blockSize.alignTo(maxFieldAlign);
19
Calling 'CharUnits::alignTo'
735 CharUnits padding = newBlockSize - blockSize;
736
737 // If we haven't yet added any fields, remember that there was an
738 // initial gap; this need to go into the block layout bit map.
739 if (blockSize == info.BlockHeaderForcedGapOffset) {
740 info.BlockHeaderForcedGapSize = padding;
741 }
742
743 elementTypes.push_back(llvm::ArrayType::get(CGM.Int8Ty,
744 padding.getQuantity()));
745 blockSize = newBlockSize;
746 endAlign = getLowBit(blockSize); // might be > maxFieldAlign
747 }
748
749 assert(endAlign >= maxFieldAlign)((void)0);
750 assert(endAlign == getLowBit(blockSize))((void)0);
751 // Slam everything else on now. This works because they have
752 // strictly decreasing alignment and we expect that size is always a
753 // multiple of alignment.
754 for (SmallVectorImpl<BlockLayoutChunk>::iterator
755 li = layout.begin(), le = layout.end(); li != le; ++li) {
756 if (endAlign < li->Alignment) {
757 // size may not be multiple of alignment. This can only happen with
758 // an over-aligned variable. We will be adding a padding field to
759 // make the size be multiple of alignment.
760 CharUnits padding = li->Alignment - endAlign;
761 elementTypes.push_back(llvm::ArrayType::get(CGM.Int8Ty,
762 padding.getQuantity()));
763 blockSize += padding;
764 endAlign = getLowBit(blockSize);
765 }
766 assert(endAlign >= li->Alignment)((void)0);
767 li->setIndex(info, elementTypes.size(), blockSize);
768 elementTypes.push_back(li->Type);
769 blockSize += li->Size;
770 endAlign = getLowBit(blockSize);
771 }
772
773 info.StructureType =
774 llvm::StructType::get(CGM.getLLVMContext(), elementTypes, true);
775}
776
777/// Emit a block literal expression in the current function.
778llvm::Value *CodeGenFunction::EmitBlockLiteral(const BlockExpr *blockExpr) {
779 // If the block has no captures, we won't have a pre-computed
780 // layout for it.
781 if (!blockExpr->getBlockDecl()->hasCaptures())
782 // The block literal is emitted as a global variable, and the block invoke
783 // function has to be extracted from its initializer.
784 if (llvm::Constant *Block = CGM.getAddrOfGlobalBlockIfEmitted(blockExpr))
785 return Block;
786
787 CGBlockInfo blockInfo(blockExpr->getBlockDecl(), CurFn->getName());
788 computeBlockInfo(CGM, this, blockInfo);
789 blockInfo.BlockExpression = blockExpr;
790 if (!blockInfo.CanBeGlobal)
791 blockInfo.LocalAddress = CreateTempAlloca(blockInfo.StructureType,
792 blockInfo.BlockAlign, "block");
793 return EmitBlockLiteral(blockInfo);
794}
795
796llvm::Value *CodeGenFunction::EmitBlockLiteral(const CGBlockInfo &blockInfo) {
797 bool IsOpenCL = CGM.getContext().getLangOpts().OpenCL;
798 auto GenVoidPtrTy =
799 IsOpenCL ? CGM.getOpenCLRuntime().getGenericVoidPointerType() : VoidPtrTy;
800 LangAS GenVoidPtrAddr = IsOpenCL ? LangAS::opencl_generic : LangAS::Default;
801 auto GenVoidPtrSize = CharUnits::fromQuantity(
802 CGM.getTarget().getPointerWidth(
803 CGM.getContext().getTargetAddressSpace(GenVoidPtrAddr)) /
804 8);
805 // Using the computed layout, generate the actual block function.
806 bool isLambdaConv = blockInfo.getBlockDecl()->isConversionFromLambda();
807 CodeGenFunction BlockCGF{CGM, true};
808 BlockCGF.SanOpts = SanOpts;
809 auto *InvokeFn = BlockCGF.GenerateBlockFunction(
810 CurGD, blockInfo, LocalDeclMap, isLambdaConv, blockInfo.CanBeGlobal);
811 auto *blockFn = llvm::ConstantExpr::getPointerCast(InvokeFn, GenVoidPtrTy);
812
813 // If there is nothing to capture, we can emit this as a global block.
814 if (blockInfo.CanBeGlobal)
815 return CGM.getAddrOfGlobalBlockIfEmitted(blockInfo.BlockExpression);
816
817 // Otherwise, we have to emit this as a local block.
818
819 Address blockAddr = blockInfo.LocalAddress;
820 assert(blockAddr.isValid() && "block has no address!")((void)0);
821
822 llvm::Constant *isa;
823 llvm::Constant *descriptor;
824 BlockFlags flags;
825 if (!IsOpenCL) {
826 // If the block is non-escaping, set field 'isa 'to NSConcreteGlobalBlock
827 // and set the BLOCK_IS_GLOBAL bit of field 'flags'. Copying a non-escaping
828 // block just returns the original block and releasing it is a no-op.
829 llvm::Constant *blockISA = blockInfo.getBlockDecl()->doesNotEscape()
830 ? CGM.getNSConcreteGlobalBlock()
831 : CGM.getNSConcreteStackBlock();
832 isa = llvm::ConstantExpr::getBitCast(blockISA, VoidPtrTy);
833
834 // Build the block descriptor.
835 descriptor = buildBlockDescriptor(CGM, blockInfo);
836
837 // Compute the initial on-stack block flags.
838 flags = BLOCK_HAS_SIGNATURE;
839 if (blockInfo.HasCapturedVariableLayout)
840 flags |= BLOCK_HAS_EXTENDED_LAYOUT;
841 if (blockInfo.needsCopyDisposeHelpers())
842 flags |= BLOCK_HAS_COPY_DISPOSE;
843 if (blockInfo.HasCXXObject)
844 flags |= BLOCK_HAS_CXX_OBJ;
845 if (blockInfo.UsesStret)
846 flags |= BLOCK_USE_STRET;
847 if (blockInfo.getBlockDecl()->doesNotEscape())
848 flags |= BLOCK_IS_NOESCAPE | BLOCK_IS_GLOBAL;
849 }
850
851 auto projectField = [&](unsigned index, const Twine &name) -> Address {
852 return Builder.CreateStructGEP(blockAddr, index, name);
853 };
854 auto storeField = [&](llvm::Value *value, unsigned index, const Twine &name) {
855 Builder.CreateStore(value, projectField(index, name));
856 };
857
858 // Initialize the block header.
859 {
860 // We assume all the header fields are densely packed.
861 unsigned index = 0;
862 CharUnits offset;
863 auto addHeaderField = [&](llvm::Value *value, CharUnits size,
864 const Twine &name) {
865 storeField(value, index, name);
866 offset += size;
867 index++;
868 };
869
870 if (!IsOpenCL) {
871 addHeaderField(isa, getPointerSize(), "block.isa");
872 addHeaderField(llvm::ConstantInt::get(IntTy, flags.getBitMask()),
873 getIntSize(), "block.flags");
874 addHeaderField(llvm::ConstantInt::get(IntTy, 0), getIntSize(),
875 "block.reserved");
876 } else {
877 addHeaderField(
878 llvm::ConstantInt::get(IntTy, blockInfo.BlockSize.getQuantity()),
879 getIntSize(), "block.size");
880 addHeaderField(
881 llvm::ConstantInt::get(IntTy, blockInfo.BlockAlign.getQuantity()),
882 getIntSize(), "block.align");
883 }
884 addHeaderField(blockFn, GenVoidPtrSize, "block.invoke");
885 if (!IsOpenCL)
886 addHeaderField(descriptor, getPointerSize(), "block.descriptor");
887 else if (auto *Helper =
888 CGM.getTargetCodeGenInfo().getTargetOpenCLBlockHelper()) {
889 for (auto I : Helper->getCustomFieldValues(*this, blockInfo)) {
890 addHeaderField(
891 I.first,
892 CharUnits::fromQuantity(
893 CGM.getDataLayout().getTypeAllocSize(I.first->getType())),
894 I.second);
895 }
896 }
897 }
898
899 // Finally, capture all the values into the block.
900 const BlockDecl *blockDecl = blockInfo.getBlockDecl();
901
902 // First, 'this'.
903 if (blockDecl->capturesCXXThis()) {
904 Address addr =
905 projectField(blockInfo.CXXThisIndex, "block.captured-this.addr");
906 Builder.CreateStore(LoadCXXThis(), addr);
907 }
908
909 // Next, captured variables.
910 for (const auto &CI : blockDecl->captures()) {
911 const VarDecl *variable = CI.getVariable();
912 const CGBlockInfo::Capture &capture = blockInfo.getCapture(variable);
913
914 // Ignore constant captures.
915 if (capture.isConstant()) continue;
916
917 QualType type = capture.fieldType();
918
919 // This will be a [[type]]*, except that a byref entry will just be
920 // an i8**.
921 Address blockField = projectField(capture.getIndex(), "block.captured");
922
923 // Compute the address of the thing we're going to move into the
924 // block literal.
925 Address src = Address::invalid();
926
927 if (blockDecl->isConversionFromLambda()) {
928 // The lambda capture in a lambda's conversion-to-block-pointer is
929 // special; we'll simply emit it directly.
930 src = Address::invalid();
931 } else if (CI.isEscapingByref()) {
932 if (BlockInfo && CI.isNested()) {
933 // We need to use the capture from the enclosing block.
934 const CGBlockInfo::Capture &enclosingCapture =
935 BlockInfo->getCapture(variable);
936
937 // This is a [[type]]*, except that a byref entry will just be an i8**.
938 src = Builder.CreateStructGEP(LoadBlockStruct(),
939 enclosingCapture.getIndex(),
940 "block.capture.addr");
941 } else {
942 auto I = LocalDeclMap.find(variable);
943 assert(I != LocalDeclMap.end())((void)0);
944 src = I->second;
945 }
946 } else {
947 DeclRefExpr declRef(getContext(), const_cast<VarDecl *>(variable),
948 /*RefersToEnclosingVariableOrCapture*/ CI.isNested(),
949 type.getNonReferenceType(), VK_LValue,
950 SourceLocation());
951 src = EmitDeclRefLValue(&declRef).getAddress(*this);
952 };
953
954 // For byrefs, we just write the pointer to the byref struct into
955 // the block field. There's no need to chase the forwarding
956 // pointer at this point, since we're building something that will
957 // live a shorter life than the stack byref anyway.
958 if (CI.isEscapingByref()) {
959 // Get a void* that points to the byref struct.
960 llvm::Value *byrefPointer;
961 if (CI.isNested())
962 byrefPointer = Builder.CreateLoad(src, "byref.capture");
963 else
964 byrefPointer = Builder.CreateBitCast(src.getPointer(), VoidPtrTy);
965
966 // Write that void* into the capture field.
967 Builder.CreateStore(byrefPointer, blockField);
968
969 // If we have a copy constructor, evaluate that into the block field.
970 } else if (const Expr *copyExpr = CI.getCopyExpr()) {
971 if (blockDecl->isConversionFromLambda()) {
972 // If we have a lambda conversion, emit the expression
973 // directly into the block instead.
974 AggValueSlot Slot =
975 AggValueSlot::forAddr(blockField, Qualifiers(),
976 AggValueSlot::IsDestructed,
977 AggValueSlot::DoesNotNeedGCBarriers,
978 AggValueSlot::IsNotAliased,
979 AggValueSlot::DoesNotOverlap);
980 EmitAggExpr(copyExpr, Slot);
981 } else {
982 EmitSynthesizedCXXCopyCtor(blockField, src, copyExpr);
983 }
984
985 // If it's a reference variable, copy the reference into the block field.
986 } else if (type->isReferenceType()) {
987 Builder.CreateStore(src.getPointer(), blockField);
988
989 // If type is const-qualified, copy the value into the block field.
990 } else if (type.isConstQualified() &&
991 type.getObjCLifetime() == Qualifiers::OCL_Strong &&
992 CGM.getCodeGenOpts().OptimizationLevel != 0) {
993 llvm::Value *value = Builder.CreateLoad(src, "captured");
994 Builder.CreateStore(value, blockField);
995
996 // If this is an ARC __strong block-pointer variable, don't do a
997 // block copy.
998 //
999 // TODO: this can be generalized into the normal initialization logic:
1000 // we should never need to do a block-copy when initializing a local
1001 // variable, because the local variable's lifetime should be strictly
1002 // contained within the stack block's.
1003 } else if (type.getObjCLifetime() == Qualifiers::OCL_Strong &&
1004 type->isBlockPointerType()) {
1005 // Load the block and do a simple retain.
1006 llvm::Value *value = Builder.CreateLoad(src, "block.captured_block");
1007 value = EmitARCRetainNonBlock(value);
1008
1009 // Do a primitive store to the block field.
1010 Builder.CreateStore(value, blockField);
1011
1012 // Otherwise, fake up a POD copy into the block field.
1013 } else {
1014 // Fake up a new variable so that EmitScalarInit doesn't think
1015 // we're referring to the variable in its own initializer.
1016 ImplicitParamDecl BlockFieldPseudoVar(getContext(), type,
1017 ImplicitParamDecl::Other);
1018
1019 // We use one of these or the other depending on whether the
1020 // reference is nested.
1021 DeclRefExpr declRef(getContext(), const_cast<VarDecl *>(variable),
1022 /*RefersToEnclosingVariableOrCapture*/ CI.isNested(),
1023 type, VK_LValue, SourceLocation());
1024
1025 ImplicitCastExpr l2r(ImplicitCastExpr::OnStack, type, CK_LValueToRValue,
1026 &declRef, VK_PRValue, FPOptionsOverride());
1027 // FIXME: Pass a specific location for the expr init so that the store is
1028 // attributed to a reasonable location - otherwise it may be attributed to
1029 // locations of subexpressions in the initialization.
1030 EmitExprAsInit(&l2r, &BlockFieldPseudoVar,
1031 MakeAddrLValue(blockField, type, AlignmentSource::Decl),
1032 /*captured by init*/ false);
1033 }
1034
1035 // Push a cleanup for the capture if necessary.
1036 if (!blockInfo.NeedsCopyDispose)
1037 continue;
1038
1039 // Ignore __block captures; there's nothing special in the on-stack block
1040 // that we need to do for them.
1041 if (CI.isByRef())
1042 continue;
1043
1044 // Ignore objects that aren't destructed.
1045 QualType::DestructionKind dtorKind = type.isDestructedType();
1046 if (dtorKind == QualType::DK_none)
1047 continue;
1048
1049 CodeGenFunction::Destroyer *destroyer;
1050
1051 // Block captures count as local values and have imprecise semantics.
1052 // They also can't be arrays, so need to worry about that.
1053 //
1054 // For const-qualified captures, emit clang.arc.use to ensure the captured
1055 // object doesn't get released while we are still depending on its validity
1056 // within the block.
1057 if (type.isConstQualified() &&
1058 type.getObjCLifetime() == Qualifiers::OCL_Strong &&
1059 CGM.getCodeGenOpts().OptimizationLevel != 0) {
1060 assert(CGM.getLangOpts().ObjCAutoRefCount &&((void)0)
1061 "expected ObjC ARC to be enabled")((void)0);
1062 destroyer = emitARCIntrinsicUse;
1063 } else if (dtorKind == QualType::DK_objc_strong_lifetime) {
1064 destroyer = destroyARCStrongImprecise;
1065 } else {
1066 destroyer = getDestroyer(dtorKind);
1067 }
1068
1069 CleanupKind cleanupKind = NormalCleanup;
1070 bool useArrayEHCleanup = needsEHCleanup(dtorKind);
1071 if (useArrayEHCleanup)
1072 cleanupKind = NormalAndEHCleanup;
1073
1074 // Extend the lifetime of the capture to the end of the scope enclosing the
1075 // block expression except when the block decl is in the list of RetExpr's
1076 // cleanup objects, in which case its lifetime ends after the full
1077 // expression.
1078 auto IsBlockDeclInRetExpr = [&]() {
1079 auto *EWC = llvm::dyn_cast_or_null<ExprWithCleanups>(RetExpr);
1080 if (EWC)
1081 for (auto &C : EWC->getObjects())
1082 if (auto *BD = C.dyn_cast<BlockDecl *>())
1083 if (BD == blockDecl)
1084 return true;
1085 return false;
1086 };
1087
1088 if (IsBlockDeclInRetExpr())
1089 pushDestroy(cleanupKind, blockField, type, destroyer, useArrayEHCleanup);
1090 else
1091 pushLifetimeExtendedDestroy(cleanupKind, blockField, type, destroyer,
1092 useArrayEHCleanup);
1093 }
1094
1095 // Cast to the converted block-pointer type, which happens (somewhat
1096 // unfortunately) to be a pointer to function type.
1097 llvm::Value *result = Builder.CreatePointerCast(
1098 blockAddr.getPointer(), ConvertType(blockInfo.getBlockExpr()->getType()));
1099
1100 if (IsOpenCL) {
1101 CGM.getOpenCLRuntime().recordBlockInfo(blockInfo.BlockExpression, InvokeFn,
1102 result);
1103 }
1104
1105 return result;
1106}
1107
1108
1109llvm::Type *CodeGenModule::getBlockDescriptorType() {
1110 if (BlockDescriptorType)
1111 return BlockDescriptorType;
1112
1113 llvm::Type *UnsignedLongTy =
1114 getTypes().ConvertType(getContext().UnsignedLongTy);
1115
1116 // struct __block_descriptor {
1117 // unsigned long reserved;
1118 // unsigned long block_size;
1119 //
1120 // // later, the following will be added
1121 //
1122 // struct {
1123 // void (*copyHelper)();
1124 // void (*copyHelper)();
1125 // } helpers; // !!! optional
1126 //
1127 // const char *signature; // the block signature
1128 // const char *layout; // reserved
1129 // };
1130 BlockDescriptorType = llvm::StructType::create(
1131 "struct.__block_descriptor", UnsignedLongTy, UnsignedLongTy);
1132
1133 // Now form a pointer to that.
1134 unsigned AddrSpace = 0;
1135 if (getLangOpts().OpenCL)
1136 AddrSpace = getContext().getTargetAddressSpace(LangAS::opencl_constant);
1137 BlockDescriptorType = llvm::PointerType::get(BlockDescriptorType, AddrSpace);
1138 return BlockDescriptorType;
1139}
1140
1141llvm::Type *CodeGenModule::getGenericBlockLiteralType() {
1142 if (GenericBlockLiteralType)
1143 return GenericBlockLiteralType;
1144
1145 llvm::Type *BlockDescPtrTy = getBlockDescriptorType();
1146
1147 if (getLangOpts().OpenCL) {
1148 // struct __opencl_block_literal_generic {
1149 // int __size;
1150 // int __align;
1151 // __generic void *__invoke;
1152 // /* custom fields */
1153 // };
1154 SmallVector<llvm::Type *, 8> StructFields(
1155 {IntTy, IntTy, getOpenCLRuntime().getGenericVoidPointerType()});
1156 if (auto *Helper = getTargetCodeGenInfo().getTargetOpenCLBlockHelper()) {
1157 for (auto I : Helper->getCustomFieldTypes())
1158 StructFields.push_back(I);
1159 }
1160 GenericBlockLiteralType = llvm::StructType::create(
1161 StructFields, "struct.__opencl_block_literal_generic");
1162 } else {
1163 // struct __block_literal_generic {
1164 // void *__isa;
1165 // int __flags;
1166 // int __reserved;
1167 // void (*__invoke)(void *);
1168 // struct __block_descriptor *__descriptor;
1169 // };
1170 GenericBlockLiteralType =
1171 llvm::StructType::create("struct.__block_literal_generic", VoidPtrTy,
1172 IntTy, IntTy, VoidPtrTy, BlockDescPtrTy);
1173 }
1174
1175 return GenericBlockLiteralType;
1176}
1177
1178RValue CodeGenFunction::EmitBlockCallExpr(const CallExpr *E,
1179 ReturnValueSlot ReturnValue) {
1180 const auto *BPT = E->getCallee()->getType()->castAs<BlockPointerType>();
1181 llvm::Value *BlockPtr = EmitScalarExpr(E->getCallee());
1182 llvm::Type *GenBlockTy = CGM.getGenericBlockLiteralType();
1183 llvm::Value *Func = nullptr;
1184 QualType FnType = BPT->getPointeeType();
1185 ASTContext &Ctx = getContext();
1186 CallArgList Args;
1187
1188 if (getLangOpts().OpenCL) {
1189 // For OpenCL, BlockPtr is already casted to generic block literal.
1190
1191 // First argument of a block call is a generic block literal casted to
1192 // generic void pointer, i.e. i8 addrspace(4)*
1193 llvm::Type *GenericVoidPtrTy =
1194 CGM.getOpenCLRuntime().getGenericVoidPointerType();
1195 llvm::Value *BlockDescriptor = Builder.CreatePointerCast(
1196 BlockPtr, GenericVoidPtrTy);
1197 QualType VoidPtrQualTy = Ctx.getPointerType(
1198 Ctx.getAddrSpaceQualType(Ctx.VoidTy, LangAS::opencl_generic));
1199 Args.add(RValue::get(BlockDescriptor), VoidPtrQualTy);
1200 // And the rest of the arguments.
1201 EmitCallArgs(Args, FnType->getAs<FunctionProtoType>(), E->arguments());
1202
1203 // We *can* call the block directly unless it is a function argument.
1204 if (!isa<ParmVarDecl>(E->getCalleeDecl()))
1205 Func = CGM.getOpenCLRuntime().getInvokeFunction(E->getCallee());
1206 else {
1207 llvm::Value *FuncPtr = Builder.CreateStructGEP(GenBlockTy, BlockPtr, 2);
1208 Func = Builder.CreateAlignedLoad(GenericVoidPtrTy, FuncPtr,
1209 getPointerAlign());
1210 }
1211 } else {
1212 // Bitcast the block literal to a generic block literal.
1213 BlockPtr = Builder.CreatePointerCast(
1214 BlockPtr, llvm::PointerType::get(GenBlockTy, 0), "block.literal");
1215 // Get pointer to the block invoke function
1216 llvm::Value *FuncPtr = Builder.CreateStructGEP(GenBlockTy, BlockPtr, 3);
1217
1218 // First argument is a block literal casted to a void pointer
1219 BlockPtr = Builder.CreatePointerCast(BlockPtr, VoidPtrTy);
1220 Args.add(RValue::get(BlockPtr), Ctx.VoidPtrTy);
1221 // And the rest of the arguments.
1222 EmitCallArgs(Args, FnType->getAs<FunctionProtoType>(), E->arguments());
1223
1224 // Load the function.
1225 Func = Builder.CreateAlignedLoad(VoidPtrTy, FuncPtr, getPointerAlign());
1226 }
1227
1228 const FunctionType *FuncTy = FnType->castAs<FunctionType>();
1229 const CGFunctionInfo &FnInfo =
1230 CGM.getTypes().arrangeBlockFunctionCall(Args, FuncTy);
1231
1232 // Cast the function pointer to the right type.
1233 llvm::Type *BlockFTy = CGM.getTypes().GetFunctionType(FnInfo);
1234
1235 llvm::Type *BlockFTyPtr = llvm::PointerType::getUnqual(BlockFTy);
1236 Func = Builder.CreatePointerCast(Func, BlockFTyPtr);
1237
1238 // Prepare the callee.
1239 CGCallee Callee(CGCalleeInfo(), Func);
1240
1241 // And call the block.
1242 return EmitCall(FnInfo, Callee, ReturnValue, Args);
1243}
1244
1245Address CodeGenFunction::GetAddrOfBlockDecl(const VarDecl *variable) {
1246 assert(BlockInfo && "evaluating block ref without block information?")((void)0);
1247 const CGBlockInfo::Capture &capture = BlockInfo->getCapture(variable);
1248
1249 // Handle constant captures.
1250 if (capture.isConstant()) return LocalDeclMap.find(variable)->second;
1251
1252 Address addr = Builder.CreateStructGEP(LoadBlockStruct(), capture.getIndex(),
1253 "block.capture.addr");
1254
1255 if (variable->isEscapingByref()) {
1256 // addr should be a void** right now. Load, then cast the result
1257 // to byref*.
1258
1259 auto &byrefInfo = getBlockByrefInfo(variable);
1260 addr = Address(Builder.CreateLoad(addr), byrefInfo.ByrefAlignment);
1261
1262 auto byrefPointerType = llvm::PointerType::get(byrefInfo.Type, 0);
1263 addr = Builder.CreateBitCast(addr, byrefPointerType, "byref.addr");
1264
1265 addr = emitBlockByrefAddress(addr, byrefInfo, /*follow*/ true,
1266 variable->getName());
1267 }
1268
1269 assert((!variable->isNonEscapingByref() ||((void)0)
1270 capture.fieldType()->isReferenceType()) &&((void)0)
1271 "the capture field of a non-escaping variable should have a "((void)0)
1272 "reference type")((void)0);
1273 if (capture.fieldType()->isReferenceType())
1274 addr = EmitLoadOfReference(MakeAddrLValue(addr, capture.fieldType()));
1275
1276 return addr;
1277}
1278
1279void CodeGenModule::setAddrOfGlobalBlock(const BlockExpr *BE,
1280 llvm::Constant *Addr) {
1281 bool Ok = EmittedGlobalBlocks.insert(std::make_pair(BE, Addr)).second;
1282 (void)Ok;
1283 assert(Ok && "Trying to replace an already-existing global block!")((void)0);
1284}
1285
1286llvm::Constant *
1287CodeGenModule::GetAddrOfGlobalBlock(const BlockExpr *BE,
1288 StringRef Name) {
1289 if (llvm::Constant *Block = getAddrOfGlobalBlockIfEmitted(BE))
1
Assuming 'Block' is null
2
Taking false branch
1290 return Block;
1291
1292 CGBlockInfo blockInfo(BE->getBlockDecl(), Name);
1293 blockInfo.BlockExpression = BE;
1294
1295 // Compute information about the layout, etc., of this block.
1296 computeBlockInfo(*this, nullptr, blockInfo);
3
Calling 'computeBlockInfo'
1297
1298 // Using that metadata, generate the actual block function.
1299 {
1300 CodeGenFunction::DeclMapTy LocalDeclMap;
1301 CodeGenFunction(*this).GenerateBlockFunction(
1302 GlobalDecl(), blockInfo, LocalDeclMap,
1303 /*IsLambdaConversionToBlock*/ false, /*BuildGlobalBlock*/ true);
1304 }
1305
1306 return getAddrOfGlobalBlockIfEmitted(BE);
1307}
1308
1309static llvm::Constant *buildGlobalBlock(CodeGenModule &CGM,
1310 const CGBlockInfo &blockInfo,
1311 llvm::Constant *blockFn) {
1312 assert(blockInfo.CanBeGlobal)((void)0);
1313 // Callers should detect this case on their own: calling this function
1314 // generally requires computing layout information, which is a waste of time
1315 // if we've already emitted this block.
1316 assert(!CGM.getAddrOfGlobalBlockIfEmitted(blockInfo.BlockExpression) &&((void)0)
1317 "Refusing to re-emit a global block.")((void)0);
1318
1319 // Generate the constants for the block literal initializer.
1320 ConstantInitBuilder builder(CGM);
1321 auto fields = builder.beginStruct();
1322
1323 bool IsOpenCL = CGM.getLangOpts().OpenCL;
1324 bool IsWindows = CGM.getTarget().getTriple().isOSWindows();
1325 if (!IsOpenCL) {
1326 // isa
1327 if (IsWindows)
1328 fields.addNullPointer(CGM.Int8PtrPtrTy);
1329 else
1330 fields.add(CGM.getNSConcreteGlobalBlock());
1331
1332 // __flags
1333 BlockFlags flags = BLOCK_IS_GLOBAL | BLOCK_HAS_SIGNATURE;
1334 if (blockInfo.UsesStret)
1335 flags |= BLOCK_USE_STRET;
1336
1337 fields.addInt(CGM.IntTy, flags.getBitMask());
1338
1339 // Reserved
1340 fields.addInt(CGM.IntTy, 0);
1341 } else {
1342 fields.addInt(CGM.IntTy, blockInfo.BlockSize.getQuantity());
1343 fields.addInt(CGM.IntTy, blockInfo.BlockAlign.getQuantity());
1344 }
1345
1346 // Function
1347 fields.add(blockFn);
1348
1349 if (!IsOpenCL) {
1350 // Descriptor
1351 fields.add(buildBlockDescriptor(CGM, blockInfo));
1352 } else if (auto *Helper =
1353 CGM.getTargetCodeGenInfo().getTargetOpenCLBlockHelper()) {
1354 for (auto I : Helper->getCustomFieldValues(CGM, blockInfo)) {
1355 fields.add(I);
1356 }
1357 }
1358
1359 unsigned AddrSpace = 0;
1360 if (CGM.getContext().getLangOpts().OpenCL)
1361 AddrSpace = CGM.getContext().getTargetAddressSpace(LangAS::opencl_global);
1362
1363 llvm::GlobalVariable *literal = fields.finishAndCreateGlobal(
1364 "__block_literal_global", blockInfo.BlockAlign,
1365 /*constant*/ !IsWindows, llvm::GlobalVariable::InternalLinkage, AddrSpace);
1366
1367 literal->addAttribute("objc_arc_inert");
1368
1369 // Windows does not allow globals to be initialised to point to globals in
1370 // different DLLs. Any such variables must run code to initialise them.
1371 if (IsWindows) {
1372 auto *Init = llvm::Function::Create(llvm::FunctionType::get(CGM.VoidTy,
1373 {}), llvm::GlobalValue::InternalLinkage, ".block_isa_init",
1374 &CGM.getModule());
1375 llvm::IRBuilder<> b(llvm::BasicBlock::Create(CGM.getLLVMContext(), "entry",
1376 Init));
1377 b.CreateAlignedStore(CGM.getNSConcreteGlobalBlock(),
1378 b.CreateStructGEP(literal->getValueType(), literal, 0),
1379 CGM.getPointerAlign().getAsAlign());
1380 b.CreateRetVoid();
1381 // We can't use the normal LLVM global initialisation array, because we
1382 // need to specify that this runs early in library initialisation.
1383 auto *InitVar = new llvm::GlobalVariable(CGM.getModule(), Init->getType(),
1384 /*isConstant*/true, llvm::GlobalValue::InternalLinkage,
1385 Init, ".block_isa_init_ptr");
1386 InitVar->setSection(".CRT$XCLa");
1387 CGM.addUsedGlobal(InitVar);
1388 }
1389
1390 // Return a constant of the appropriately-casted type.
1391 llvm::Type *RequiredType =
1392 CGM.getTypes().ConvertType(blockInfo.getBlockExpr()->getType());
1393 llvm::Constant *Result =
1394 llvm::ConstantExpr::getPointerCast(literal, RequiredType);
1395 CGM.setAddrOfGlobalBlock(blockInfo.BlockExpression, Result);
1396 if (CGM.getContext().getLangOpts().OpenCL)
1397 CGM.getOpenCLRuntime().recordBlockInfo(
1398 blockInfo.BlockExpression,
1399 cast<llvm::Function>(blockFn->stripPointerCasts()), Result);
1400 return Result;
1401}
1402
1403void CodeGenFunction::setBlockContextParameter(const ImplicitParamDecl *D,
1404 unsigned argNum,
1405 llvm::Value *arg) {
1406 assert(BlockInfo && "not emitting prologue of block invocation function?!")((void)0);
1407
1408 // Allocate a stack slot like for any local variable to guarantee optimal
1409 // debug info at -O0. The mem2reg pass will eliminate it when optimizing.
1410 Address alloc = CreateMemTemp(D->getType(), D->getName() + ".addr");
1411 Builder.CreateStore(arg, alloc);
1412 if (CGDebugInfo *DI = getDebugInfo()) {
1413 if (CGM.getCodeGenOpts().hasReducedDebugInfo()) {
1414 DI->setLocation(D->getLocation());
1415 DI->EmitDeclareOfBlockLiteralArgVariable(
1416 *BlockInfo, D->getName(), argNum,
1417 cast<llvm::AllocaInst>(alloc.getPointer()), Builder);
1418 }
1419 }
1420
1421 SourceLocation StartLoc = BlockInfo->getBlockExpr()->getBody()->getBeginLoc();
1422 ApplyDebugLocation Scope(*this, StartLoc);
1423
1424 // Instead of messing around with LocalDeclMap, just set the value
1425 // directly as BlockPointer.
1426 BlockPointer = Builder.CreatePointerCast(
1427 arg,
1428 BlockInfo->StructureType->getPointerTo(
1429 getContext().getLangOpts().OpenCL
1430 ? getContext().getTargetAddressSpace(LangAS::opencl_generic)
1431 : 0),
1432 "block");
1433}
1434
1435Address CodeGenFunction::LoadBlockStruct() {
1436 assert(BlockInfo && "not in a block invocation function!")((void)0);
1437 assert(BlockPointer && "no block pointer set!")((void)0);
1438 return Address(BlockPointer, BlockInfo->BlockAlign);
1439}
1440
1441llvm::Function *
1442CodeGenFunction::GenerateBlockFunction(GlobalDecl GD,
1443 const CGBlockInfo &blockInfo,
1444 const DeclMapTy &ldm,
1445 bool IsLambdaConversionToBlock,
1446 bool BuildGlobalBlock) {
1447 const BlockDecl *blockDecl = blockInfo.getBlockDecl();
1448
1449 CurGD = GD;
1450
1451 CurEHLocation = blockInfo.getBlockExpr()->getEndLoc();
1452
1453 BlockInfo = &blockInfo;
1454
1455 // Arrange for local static and local extern declarations to appear
1456 // to be local to this function as well, in case they're directly
1457 // referenced in a block.
1458 for (DeclMapTy::const_iterator i = ldm.begin(), e = ldm.end(); i != e; ++i) {
1459 const auto *var = dyn_cast<VarDecl>(i->first);
1460 if (var && !var->hasLocalStorage())
1461 setAddrOfLocalVar(var, i->second);
1462 }
1463
1464 // Begin building the function declaration.
1465
1466 // Build the argument list.
1467 FunctionArgList args;
1468
1469 // The first argument is the block pointer. Just take it as a void*
1470 // and cast it later.
1471 QualType selfTy = getContext().VoidPtrTy;
1472
1473 // For OpenCL passed block pointer can be private AS local variable or
1474 // global AS program scope variable (for the case with and without captures).
1475 // Generic AS is used therefore to be able to accommodate both private and
1476 // generic AS in one implementation.
1477 if (getLangOpts().OpenCL)
1478 selfTy = getContext().getPointerType(getContext().getAddrSpaceQualType(
1479 getContext().VoidTy, LangAS::opencl_generic));
1480
1481 IdentifierInfo *II = &CGM.getContext().Idents.get(".block_descriptor");
1482
1483 ImplicitParamDecl SelfDecl(getContext(), const_cast<BlockDecl *>(blockDecl),
1484 SourceLocation(), II, selfTy,
1485 ImplicitParamDecl::ObjCSelf);
1486 args.push_back(&SelfDecl);
1487
1488 // Now add the rest of the parameters.
1489 args.append(blockDecl->param_begin(), blockDecl->param_end());
1490
1491 // Create the function declaration.
1492 const FunctionProtoType *fnType = blockInfo.getBlockExpr()->getFunctionType();
1493 const CGFunctionInfo &fnInfo =
1494 CGM.getTypes().arrangeBlockFunctionDeclaration(fnType, args);
1495 if (CGM.ReturnSlotInterferesWithArgs(fnInfo))
1496 blockInfo.UsesStret = true;
1497
1498 llvm::FunctionType *fnLLVMType = CGM.getTypes().GetFunctionType(fnInfo);
1499
1500 StringRef name = CGM.getBlockMangledName(GD, blockDecl);
1501 llvm::Function *fn = llvm::Function::Create(
1502 fnLLVMType, llvm::GlobalValue::InternalLinkage, name, &CGM.getModule());
1503 CGM.SetInternalFunctionAttributes(blockDecl, fn, fnInfo);
1504
1505 if (BuildGlobalBlock) {
1506 auto GenVoidPtrTy = getContext().getLangOpts().OpenCL
1507 ? CGM.getOpenCLRuntime().getGenericVoidPointerType()
1508 : VoidPtrTy;
1509 buildGlobalBlock(CGM, blockInfo,
1510 llvm::ConstantExpr::getPointerCast(fn, GenVoidPtrTy));
1511 }
1512
1513 // Begin generating the function.
1514 StartFunction(blockDecl, fnType->getReturnType(), fn, fnInfo, args,
1515 blockDecl->getLocation(),
1516 blockInfo.getBlockExpr()->getBody()->getBeginLoc());
1517
1518 // Okay. Undo some of what StartFunction did.
1519
1520 // At -O0 we generate an explicit alloca for the BlockPointer, so the RA
1521 // won't delete the dbg.declare intrinsics for captured variables.
1522 llvm::Value *BlockPointerDbgLoc = BlockPointer;
1523 if (CGM.getCodeGenOpts().OptimizationLevel == 0) {
1524 // Allocate a stack slot for it, so we can point the debugger to it
1525 Address Alloca = CreateTempAlloca(BlockPointer->getType(),
1526 getPointerAlign(),
1527 "block.addr");
1528 // Set the DebugLocation to empty, so the store is recognized as a
1529 // frame setup instruction by llvm::DwarfDebug::beginFunction().
1530 auto NL = ApplyDebugLocation::CreateEmpty(*this);
1531 Builder.CreateStore(BlockPointer, Alloca);
1532 BlockPointerDbgLoc = Alloca.getPointer();
1533 }
1534
1535 // If we have a C++ 'this' reference, go ahead and force it into
1536 // existence now.
1537 if (blockDecl->capturesCXXThis()) {
1538 Address addr = Builder.CreateStructGEP(
1539 LoadBlockStruct(), blockInfo.CXXThisIndex, "block.captured-this");
1540 CXXThisValue = Builder.CreateLoad(addr, "this");
1541 }
1542
1543 // Also force all the constant captures.
1544 for (const auto &CI : blockDecl->captures()) {
1545 const VarDecl *variable = CI.getVariable();
1546 const CGBlockInfo::Capture &capture = blockInfo.getCapture(variable);
1547 if (!capture.isConstant()) continue;
1548
1549 CharUnits align = getContext().getDeclAlign(variable);
1550 Address alloca =
1551 CreateMemTemp(variable->getType(), align, "block.captured-const");
1552
1553 Builder.CreateStore(capture.getConstant(), alloca);
1554
1555 setAddrOfLocalVar(variable, alloca);
1556 }
1557
1558 // Save a spot to insert the debug information for all the DeclRefExprs.
1559 llvm::BasicBlock *entry = Builder.GetInsertBlock();
1560 llvm::BasicBlock::iterator entry_ptr = Builder.GetInsertPoint();
1561 --entry_ptr;
1562
1563 if (IsLambdaConversionToBlock)
1564 EmitLambdaBlockInvokeBody();
1565 else {
1566 PGO.assignRegionCounters(GlobalDecl(blockDecl), fn);
1567 incrementProfileCounter(blockDecl->getBody());
1568 EmitStmt(blockDecl->getBody());
1569 }
1570
1571 // Remember where we were...
1572 llvm::BasicBlock *resume = Builder.GetInsertBlock();
1573
1574 // Go back to the entry.
1575 ++entry_ptr;
1576 Builder.SetInsertPoint(entry, entry_ptr);
1577
1578 // Emit debug information for all the DeclRefExprs.
1579 // FIXME: also for 'this'
1580 if (CGDebugInfo *DI = getDebugInfo()) {
1581 for (const auto &CI : blockDecl->captures()) {
1582 const VarDecl *variable = CI.getVariable();
1583 DI->EmitLocation(Builder, variable->getLocation());
1584
1585 if (CGM.getCodeGenOpts().hasReducedDebugInfo()) {
1586 const CGBlockInfo::Capture &capture = blockInfo.getCapture(variable);
1587 if (capture.isConstant()) {
1588 auto addr = LocalDeclMap.find(variable)->second;
1589 (void)DI->EmitDeclareOfAutoVariable(variable, addr.getPointer(),
1590 Builder);
1591 continue;
1592 }
1593
1594 DI->EmitDeclareOfBlockDeclRefVariable(
1595 variable, BlockPointerDbgLoc, Builder, blockInfo,
1596 entry_ptr == entry->end() ? nullptr : &*entry_ptr);
1597 }
1598 }
1599 // Recover location if it was changed in the above loop.
1600 DI->EmitLocation(Builder,
1601 cast<CompoundStmt>(blockDecl->getBody())->getRBracLoc());
1602 }
1603
1604 // And resume where we left off.
1605 if (resume == nullptr)
1606 Builder.ClearInsertionPoint();
1607 else
1608 Builder.SetInsertPoint(resume);
1609
1610 FinishFunction(cast<CompoundStmt>(blockDecl->getBody())->getRBracLoc());
1611
1612 return fn;
1613}
1614
1615static std::pair<BlockCaptureEntityKind, BlockFieldFlags>
1616computeCopyInfoForBlockCapture(const BlockDecl::Capture &CI, QualType T,
1617 const LangOptions &LangOpts) {
1618 if (CI.getCopyExpr()) {
1619 assert(!CI.isByRef())((void)0);
1620 // don't bother computing flags
1621 return std::make_pair(BlockCaptureEntityKind::CXXRecord, BlockFieldFlags());
1622 }
1623 BlockFieldFlags Flags;
1624 if (CI.isEscapingByref()) {
1625 Flags = BLOCK_FIELD_IS_BYREF;
1626 if (T.isObjCGCWeak())
1627 Flags |= BLOCK_FIELD_IS_WEAK;
1628 return std::make_pair(BlockCaptureEntityKind::BlockObject, Flags);
1629 }
1630
1631 Flags = BLOCK_FIELD_IS_OBJECT;
1632 bool isBlockPointer = T->isBlockPointerType();
1633 if (isBlockPointer)
1634 Flags = BLOCK_FIELD_IS_BLOCK;
1635
1636 switch (T.isNonTrivialToPrimitiveCopy()) {
1637 case QualType::PCK_Struct:
1638 return std::make_pair(BlockCaptureEntityKind::NonTrivialCStruct,
1639 BlockFieldFlags());
1640 case QualType::PCK_ARCWeak:
1641 // We need to register __weak direct captures with the runtime.
1642 return std::make_pair(BlockCaptureEntityKind::ARCWeak, Flags);
1643 case QualType::PCK_ARCStrong:
1644 // We need to retain the copied value for __strong direct captures.
1645 // If it's a block pointer, we have to copy the block and assign that to
1646 // the destination pointer, so we might as well use _Block_object_assign.
1647 // Otherwise we can avoid that.
1648 return std::make_pair(!isBlockPointer ? BlockCaptureEntityKind::ARCStrong
1649 : BlockCaptureEntityKind::BlockObject,
1650 Flags);
1651 case QualType::PCK_Trivial:
1652 case QualType::PCK_VolatileTrivial: {
1653 if (!T->isObjCRetainableType())
1654 // For all other types, the memcpy is fine.
1655 return std::make_pair(BlockCaptureEntityKind::None, BlockFieldFlags());
1656
1657 // Special rules for ARC captures:
1658 Qualifiers QS = T.getQualifiers();
1659
1660 // Non-ARC captures of retainable pointers are strong and
1661 // therefore require a call to _Block_object_assign.
1662 if (!QS.getObjCLifetime() && !LangOpts.ObjCAutoRefCount)
1663 return std::make_pair(BlockCaptureEntityKind::BlockObject, Flags);
1664
1665 // Otherwise the memcpy is fine.
1666 return std::make_pair(BlockCaptureEntityKind::None, BlockFieldFlags());
1667 }
1668 }
1669 llvm_unreachable("after exhaustive PrimitiveCopyKind switch")__builtin_unreachable();
1670}
1671
1672static std::pair<BlockCaptureEntityKind, BlockFieldFlags>
1673computeDestroyInfoForBlockCapture(const BlockDecl::Capture &CI, QualType T,
1674 const LangOptions &LangOpts);
1675
1676/// Find the set of block captures that need to be explicitly copied or destroy.
1677static void findBlockCapturedManagedEntities(
1678 const CGBlockInfo &BlockInfo, const LangOptions &LangOpts,
1679 SmallVectorImpl<BlockCaptureManagedEntity> &ManagedCaptures) {
1680 for (const auto &CI : BlockInfo.getBlockDecl()->captures()) {
1681 const VarDecl *Variable = CI.getVariable();
1682 const CGBlockInfo::Capture &Capture = BlockInfo.getCapture(Variable);
1683 if (Capture.isConstant())
1684 continue;
1685
1686 QualType VT = Capture.fieldType();
1687 auto CopyInfo = computeCopyInfoForBlockCapture(CI, VT, LangOpts);
1688 auto DisposeInfo = computeDestroyInfoForBlockCapture(CI, VT, LangOpts);
1689 if (CopyInfo.first != BlockCaptureEntityKind::None ||
1690 DisposeInfo.first != BlockCaptureEntityKind::None)
1691 ManagedCaptures.emplace_back(CopyInfo.first, DisposeInfo.first,
1692 CopyInfo.second, DisposeInfo.second, CI,
1693 Capture);
1694 }
1695
1696 // Sort the captures by offset.
1697 llvm::sort(ManagedCaptures);
1698}
1699
1700namespace {
1701/// Release a __block variable.
1702struct CallBlockRelease final : EHScopeStack::Cleanup {
1703 Address Addr;
1704 BlockFieldFlags FieldFlags;
1705 bool LoadBlockVarAddr, CanThrow;
1706
1707 CallBlockRelease(Address Addr, BlockFieldFlags Flags, bool LoadValue,
1708 bool CT)
1709 : Addr(Addr), FieldFlags(Flags), LoadBlockVarAddr(LoadValue),
1710 CanThrow(CT) {}
1711
1712 void Emit(CodeGenFunction &CGF, Flags flags) override {
1713 llvm::Value *BlockVarAddr;
1714 if (LoadBlockVarAddr) {
1715 BlockVarAddr = CGF.Builder.CreateLoad(Addr);
1716 BlockVarAddr = CGF.Builder.CreateBitCast(BlockVarAddr, CGF.VoidPtrTy);
1717 } else {
1718 BlockVarAddr = Addr.getPointer();
1719 }
1720
1721 CGF.BuildBlockRelease(BlockVarAddr, FieldFlags, CanThrow);
1722 }
1723};
1724} // end anonymous namespace
1725
1726/// Check if \p T is a C++ class that has a destructor that can throw.
1727bool CodeGenFunction::cxxDestructorCanThrow(QualType T) {
1728 if (const auto *RD = T->getAsCXXRecordDecl())
1729 if (const CXXDestructorDecl *DD = RD->getDestructor())
1730 return DD->getType()->castAs<FunctionProtoType>()->canThrow();
1731 return false;
1732}
1733
1734// Return a string that has the information about a capture.
1735static std::string getBlockCaptureStr(const BlockCaptureManagedEntity &E,
1736 CaptureStrKind StrKind,
1737 CharUnits BlockAlignment,
1738 CodeGenModule &CGM) {
1739 std::string Str;
1740 ASTContext &Ctx = CGM.getContext();
1741 const BlockDecl::Capture &CI = *E.CI;
1742 QualType CaptureTy = CI.getVariable()->getType();
1743
1744 BlockCaptureEntityKind Kind;
1745 BlockFieldFlags Flags;
1746
1747 // CaptureStrKind::Merged should be passed only when the operations and the
1748 // flags are the same for copy and dispose.
1749 assert((StrKind != CaptureStrKind::Merged ||((void)0)
1750 (E.CopyKind == E.DisposeKind && E.CopyFlags == E.DisposeFlags)) &&((void)0)
1751 "different operations and flags")((void)0);
1752
1753 if (StrKind == CaptureStrKind::DisposeHelper) {
1754 Kind = E.DisposeKind;
1755 Flags = E.DisposeFlags;
1756 } else {
1757 Kind = E.CopyKind;
1758 Flags = E.CopyFlags;
1759 }
1760
1761 switch (Kind) {
1762 case BlockCaptureEntityKind::CXXRecord: {
1763 Str += "c";
1764 SmallString<256> TyStr;
1765 llvm::raw_svector_ostream Out(TyStr);
1766 CGM.getCXXABI().getMangleContext().mangleTypeName(CaptureTy, Out);
1767 Str += llvm::to_string(TyStr.size()) + TyStr.c_str();
1768 break;
1769 }
1770 case BlockCaptureEntityKind::ARCWeak:
1771 Str += "w";
1772 break;
1773 case BlockCaptureEntityKind::ARCStrong:
1774 Str += "s";
1775 break;
1776 case BlockCaptureEntityKind::BlockObject: {
1777 const VarDecl *Var = CI.getVariable();
1778 unsigned F = Flags.getBitMask();
1779 if (F & BLOCK_FIELD_IS_BYREF) {
1780 Str += "r";
1781 if (F & BLOCK_FIELD_IS_WEAK)
1782 Str += "w";
1783 else {
1784 // If CaptureStrKind::Merged is passed, check both the copy expression
1785 // and the destructor.
1786 if (StrKind != CaptureStrKind::DisposeHelper) {
1787 if (Ctx.getBlockVarCopyInit(Var).canThrow())
1788 Str += "c";
1789 }
1790 if (StrKind != CaptureStrKind::CopyHelper) {
1791 if (CodeGenFunction::cxxDestructorCanThrow(CaptureTy))
1792 Str += "d";
1793 }
1794 }
1795 } else {
1796 assert((F & BLOCK_FIELD_IS_OBJECT) && "unexpected flag value")((void)0);
1797 if (F == BLOCK_FIELD_IS_BLOCK)
1798 Str += "b";
1799 else
1800 Str += "o";
1801 }
1802 break;
1803 }
1804 case BlockCaptureEntityKind::NonTrivialCStruct: {
1805 bool IsVolatile = CaptureTy.isVolatileQualified();
1806 CharUnits Alignment =
1807 BlockAlignment.alignmentAtOffset(E.Capture->getOffset());
1808
1809 Str += "n";
1810 std::string FuncStr;
1811 if (StrKind == CaptureStrKind::DisposeHelper)
1812 FuncStr = CodeGenFunction::getNonTrivialDestructorStr(
1813 CaptureTy, Alignment, IsVolatile, Ctx);
1814 else
1815 // If CaptureStrKind::Merged is passed, use the copy constructor string.
1816 // It has all the information that the destructor string has.
1817 FuncStr = CodeGenFunction::getNonTrivialCopyConstructorStr(
1818 CaptureTy, Alignment, IsVolatile, Ctx);
1819 // The underscore is necessary here because non-trivial copy constructor
1820 // and destructor strings can start with a number.
1821 Str += llvm::to_string(FuncStr.size()) + "_" + FuncStr;
1822 break;
1823 }
1824 case BlockCaptureEntityKind::None:
1825 break;
1826 }
1827
1828 return Str;
1829}
1830
1831static std::string getCopyDestroyHelperFuncName(
1832 const SmallVectorImpl<BlockCaptureManagedEntity> &Captures,
1833 CharUnits BlockAlignment, CaptureStrKind StrKind, CodeGenModule &CGM) {
1834 assert((StrKind == CaptureStrKind::CopyHelper ||((void)0)
1835 StrKind == CaptureStrKind::DisposeHelper) &&((void)0)
1836 "unexpected CaptureStrKind")((void)0);
1837 std::string Name = StrKind == CaptureStrKind::CopyHelper
1838 ? "__copy_helper_block_"
1839 : "__destroy_helper_block_";
1840 if (CGM.getLangOpts().Exceptions)
1841 Name += "e";
1842 if (CGM.getCodeGenOpts().ObjCAutoRefCountExceptions)
1843 Name += "a";
1844 Name += llvm::to_string(BlockAlignment.getQuantity()) + "_";
1845
1846 for (const BlockCaptureManagedEntity &E : Captures) {
1847 Name += llvm::to_string(E.Capture->getOffset().getQuantity());
1848 Name += getBlockCaptureStr(E, StrKind, BlockAlignment, CGM);
1849 }
1850
1851 return Name;
1852}
1853
1854static void pushCaptureCleanup(BlockCaptureEntityKind CaptureKind,
1855 Address Field, QualType CaptureType,
1856 BlockFieldFlags Flags, bool ForCopyHelper,
1857 VarDecl *Var, CodeGenFunction &CGF) {
1858 bool EHOnly = ForCopyHelper;
1859
1860 switch (CaptureKind) {
1861 case BlockCaptureEntityKind::CXXRecord:
1862 case BlockCaptureEntityKind::ARCWeak:
1863 case BlockCaptureEntityKind::NonTrivialCStruct:
1864 case BlockCaptureEntityKind::ARCStrong: {
1865 if (CaptureType.isDestructedType() &&
1866 (!EHOnly || CGF.needsEHCleanup(CaptureType.isDestructedType()))) {
1867 CodeGenFunction::Destroyer *Destroyer =
1868 CaptureKind == BlockCaptureEntityKind::ARCStrong
1869 ? CodeGenFunction::destroyARCStrongImprecise
1870 : CGF.getDestroyer(CaptureType.isDestructedType());
1871 CleanupKind Kind =
1872 EHOnly ? EHCleanup
1873 : CGF.getCleanupKind(CaptureType.isDestructedType());
1874 CGF.pushDestroy(Kind, Field, CaptureType, Destroyer, Kind & EHCleanup);
1875 }
1876 break;
1877 }
1878 case BlockCaptureEntityKind::BlockObject: {
1879 if (!EHOnly || CGF.getLangOpts().Exceptions) {
1880 CleanupKind Kind = EHOnly ? EHCleanup : NormalAndEHCleanup;
1881 // Calls to _Block_object_dispose along the EH path in the copy helper
1882 // function don't throw as newly-copied __block variables always have a
1883 // reference count of 2.
1884 bool CanThrow =
1885 !ForCopyHelper && CGF.cxxDestructorCanThrow(CaptureType);
1886 CGF.enterByrefCleanup(Kind, Field, Flags, /*LoadBlockVarAddr*/ true,
1887 CanThrow);
1888 }
1889 break;
1890 }
1891 case BlockCaptureEntityKind::None:
1892 break;
1893 }
1894}
1895
1896static void setBlockHelperAttributesVisibility(bool CapturesNonExternalType,
1897 llvm::Function *Fn,
1898 const CGFunctionInfo &FI,
1899 CodeGenModule &CGM) {
1900 if (CapturesNonExternalType) {
1901 CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, FI);
1902 } else {
1903 Fn->setVisibility(llvm::GlobalValue::HiddenVisibility);
1904 Fn->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global);
1905 CGM.SetLLVMFunctionAttributes(GlobalDecl(), FI, Fn, /*IsThunk=*/false);
1906 CGM.SetLLVMFunctionAttributesForDefinition(nullptr, Fn);
1907 }
1908}
1909/// Generate the copy-helper function for a block closure object:
1910/// static void block_copy_helper(block_t *dst, block_t *src);
1911/// The runtime will have previously initialized 'dst' by doing a
1912/// bit-copy of 'src'.
1913///
1914/// Note that this copies an entire block closure object to the heap;
1915/// it should not be confused with a 'byref copy helper', which moves
1916/// the contents of an individual __block variable to the heap.
1917llvm::Constant *
1918CodeGenFunction::GenerateCopyHelperFunction(const CGBlockInfo &blockInfo) {
1919 SmallVector<BlockCaptureManagedEntity, 4> CopiedCaptures;
1920 findBlockCapturedManagedEntities(blockInfo, getLangOpts(), CopiedCaptures);
1921 std::string FuncName =
1922 getCopyDestroyHelperFuncName(CopiedCaptures, blockInfo.BlockAlign,
1923 CaptureStrKind::CopyHelper, CGM);
1924
1925 if (llvm::GlobalValue *Func = CGM.getModule().getNamedValue(FuncName))
1926 return llvm::ConstantExpr::getBitCast(Func, VoidPtrTy);
1927
1928 ASTContext &C = getContext();
1929
1930 QualType ReturnTy = C.VoidTy;
1931
1932 FunctionArgList args;
1933 ImplicitParamDecl DstDecl(C, C.VoidPtrTy, ImplicitParamDecl::Other);
1934 args.push_back(&DstDecl);
1935 ImplicitParamDecl SrcDecl(C, C.VoidPtrTy, ImplicitParamDecl::Other);
1936 args.push_back(&SrcDecl);
1937
1938 const CGFunctionInfo &FI =
1939 CGM.getTypes().arrangeBuiltinFunctionDeclaration(ReturnTy, args);
1940
1941 // FIXME: it would be nice if these were mergeable with things with
1942 // identical semantics.
1943 llvm::FunctionType *LTy = CGM.getTypes().GetFunctionType(FI);
1944
1945 llvm::Function *Fn =
1946 llvm::Function::Create(LTy, llvm::GlobalValue::LinkOnceODRLinkage,
1947 FuncName, &CGM.getModule());
1948 if (CGM.supportsCOMDAT())
1949 Fn->setComdat(CGM.getModule().getOrInsertComdat(FuncName));
1950
1951 SmallVector<QualType, 2> ArgTys;
1952 ArgTys.push_back(C.VoidPtrTy);
1953 ArgTys.push_back(C.VoidPtrTy);
1954
1955 setBlockHelperAttributesVisibility(blockInfo.CapturesNonExternalType, Fn, FI,
1956 CGM);
1957 StartFunction(GlobalDecl(), ReturnTy, Fn, FI, args);
1958 auto AL = ApplyDebugLocation::CreateArtificial(*this);
1959
1960 llvm::Type *structPtrTy = blockInfo.StructureType->getPointerTo();
1961
1962 Address src = GetAddrOfLocalVar(&SrcDecl);
1963 src = Address(Builder.CreateLoad(src), blockInfo.BlockAlign);
1964 src = Builder.CreateBitCast(src, structPtrTy, "block.source");
1965
1966 Address dst = GetAddrOfLocalVar(&DstDecl);
1967 dst = Address(Builder.CreateLoad(dst), blockInfo.BlockAlign);
1968 dst = Builder.CreateBitCast(dst, structPtrTy, "block.dest");
1969
1970 for (const auto &CopiedCapture : CopiedCaptures) {
1971 const BlockDecl::Capture &CI = *CopiedCapture.CI;
1972 const CGBlockInfo::Capture &capture = *CopiedCapture.Capture;
1973 QualType captureType = CI.getVariable()->getType();
1974 BlockFieldFlags flags = CopiedCapture.CopyFlags;
1975
1976 unsigned index = capture.getIndex();
1977 Address srcField = Builder.CreateStructGEP(src, index);
1978 Address dstField = Builder.CreateStructGEP(dst, index);
1979
1980 switch (CopiedCapture.CopyKind) {
1981 case BlockCaptureEntityKind::CXXRecord:
1982 // If there's an explicit copy expression, we do that.
1983 assert(CI.getCopyExpr() && "copy expression for variable is missing")((void)0);
1984 EmitSynthesizedCXXCopyCtor(dstField, srcField, CI.getCopyExpr());
1985 break;
1986 case BlockCaptureEntityKind::ARCWeak:
1987 EmitARCCopyWeak(dstField, srcField);
1988 break;
1989 case BlockCaptureEntityKind::NonTrivialCStruct: {
1990 // If this is a C struct that requires non-trivial copy construction,
1991 // emit a call to its copy constructor.
1992 QualType varType = CI.getVariable()->getType();
1993 callCStructCopyConstructor(MakeAddrLValue(dstField, varType),
1994 MakeAddrLValue(srcField, varType));
1995 break;
1996 }
1997 case BlockCaptureEntityKind::ARCStrong: {
1998 llvm::Value *srcValue = Builder.CreateLoad(srcField, "blockcopy.src");
1999 // At -O0, store null into the destination field (so that the
2000 // storeStrong doesn't over-release) and then call storeStrong.
2001 // This is a workaround to not having an initStrong call.
2002 if (CGM.getCodeGenOpts().OptimizationLevel == 0) {
2003 auto *ty = cast<llvm::PointerType>(srcValue->getType());
2004 llvm::Value *null = llvm::ConstantPointerNull::get(ty);
2005 Builder.CreateStore(null, dstField);
2006 EmitARCStoreStrongCall(dstField, srcValue, true);
2007
2008 // With optimization enabled, take advantage of the fact that
2009 // the blocks runtime guarantees a memcpy of the block data, and
2010 // just emit a retain of the src field.
2011 } else {
2012 EmitARCRetainNonBlock(srcValue);
2013
2014 // Unless EH cleanup is required, we don't need this anymore, so kill
2015 // it. It's not quite worth the annoyance to avoid creating it in the
2016 // first place.
2017 if (!needsEHCleanup(captureType.isDestructedType()))
2018 cast<llvm::Instruction>(dstField.getPointer())->eraseFromParent();
2019 }
2020 break;
2021 }
2022 case BlockCaptureEntityKind::BlockObject: {
2023 llvm::Value *srcValue = Builder.CreateLoad(srcField, "blockcopy.src");
2024 srcValue = Builder.CreateBitCast(srcValue, VoidPtrTy);
2025 llvm::Value *dstAddr =
2026 Builder.CreateBitCast(dstField.getPointer(), VoidPtrTy);
2027 llvm::Value *args[] = {
2028 dstAddr, srcValue, llvm::ConstantInt::get(Int32Ty, flags.getBitMask())
2029 };
2030
2031 if (CI.isByRef() && C.getBlockVarCopyInit(CI.getVariable()).canThrow())
2032 EmitRuntimeCallOrInvoke(CGM.getBlockObjectAssign(), args);
2033 else
2034 EmitNounwindRuntimeCall(CGM.getBlockObjectAssign(), args);
2035 break;
2036 }
2037 case BlockCaptureEntityKind::None:
2038 continue;
2039 }
2040
2041 // Ensure that we destroy the copied object if an exception is thrown later
2042 // in the helper function.
2043 pushCaptureCleanup(CopiedCapture.CopyKind, dstField, captureType, flags,
2044 /*ForCopyHelper*/ true, CI.getVariable(), *this);
2045 }
2046
2047 FinishFunction();
2048
2049 return llvm::ConstantExpr::getBitCast(Fn, VoidPtrTy);
2050}
2051
2052static BlockFieldFlags
2053getBlockFieldFlagsForObjCObjectPointer(const BlockDecl::Capture &CI,
2054 QualType T) {
2055 BlockFieldFlags Flags = BLOCK_FIELD_IS_OBJECT;
2056 if (T->isBlockPointerType())
2057 Flags = BLOCK_FIELD_IS_BLOCK;
2058 return Flags;
2059}
2060
2061static std::pair<BlockCaptureEntityKind, BlockFieldFlags>
2062computeDestroyInfoForBlockCapture(const BlockDecl::Capture &CI, QualType T,
2063 const LangOptions &LangOpts) {
2064 if (CI.isEscapingByref()) {
2065 BlockFieldFlags Flags = BLOCK_FIELD_IS_BYREF;
2066 if (T.isObjCGCWeak())
2067 Flags |= BLOCK_FIELD_IS_WEAK;
2068 return std::make_pair(BlockCaptureEntityKind::BlockObject, Flags);
2069 }
2070
2071 switch (T.isDestructedType()) {
2072 case QualType::DK_cxx_destructor:
2073 return std::make_pair(BlockCaptureEntityKind::CXXRecord, BlockFieldFlags());
2074 case QualType::DK_objc_strong_lifetime:
2075 // Use objc_storeStrong for __strong direct captures; the
2076 // dynamic tools really like it when we do this.
2077 return std::make_pair(BlockCaptureEntityKind::ARCStrong,
2078 getBlockFieldFlagsForObjCObjectPointer(CI, T));
2079 case QualType::DK_objc_weak_lifetime:
2080 // Support __weak direct captures.
2081 return std::make_pair(BlockCaptureEntityKind::ARCWeak,
2082 getBlockFieldFlagsForObjCObjectPointer(CI, T));
2083 case QualType::DK_nontrivial_c_struct:
2084 return std::make_pair(BlockCaptureEntityKind::NonTrivialCStruct,
2085 BlockFieldFlags());
2086 case QualType::DK_none: {
2087 // Non-ARC captures are strong, and we need to use _Block_object_dispose.
2088 if (T->isObjCRetainableType() && !T.getQualifiers().hasObjCLifetime() &&
2089 !LangOpts.ObjCAutoRefCount)
2090 return std::make_pair(BlockCaptureEntityKind::BlockObject,
2091 getBlockFieldFlagsForObjCObjectPointer(CI, T));
2092 // Otherwise, we have nothing to do.
2093 return std::make_pair(BlockCaptureEntityKind::None, BlockFieldFlags());
2094 }
2095 }
2096 llvm_unreachable("after exhaustive DestructionKind switch")__builtin_unreachable();
2097}
2098
2099/// Generate the destroy-helper function for a block closure object:
2100/// static void block_destroy_helper(block_t *theBlock);
2101///
2102/// Note that this destroys a heap-allocated block closure object;
2103/// it should not be confused with a 'byref destroy helper', which
2104/// destroys the heap-allocated contents of an individual __block
2105/// variable.
2106llvm::Constant *
2107CodeGenFunction::GenerateDestroyHelperFunction(const CGBlockInfo &blockInfo) {
2108 SmallVector<BlockCaptureManagedEntity, 4> DestroyedCaptures;
2109 findBlockCapturedManagedEntities(blockInfo, getLangOpts(), DestroyedCaptures);
2110 std::string FuncName =
2111 getCopyDestroyHelperFuncName(DestroyedCaptures, blockInfo.BlockAlign,
2112 CaptureStrKind::DisposeHelper, CGM);
2113
2114 if (llvm::GlobalValue *Func = CGM.getModule().getNamedValue(FuncName))
2115 return llvm::ConstantExpr::getBitCast(Func, VoidPtrTy);
2116
2117 ASTContext &C = getContext();
2118
2119 QualType ReturnTy = C.VoidTy;
2120
2121 FunctionArgList args;
2122 ImplicitParamDecl SrcDecl(C, C.VoidPtrTy, ImplicitParamDecl::Other);
2123 args.push_back(&SrcDecl);
2124
2125 const CGFunctionInfo &FI =
2126 CGM.getTypes().arrangeBuiltinFunctionDeclaration(ReturnTy, args);
2127
2128 // FIXME: We'd like to put these into a mergable by content, with
2129 // internal linkage.
2130 llvm::FunctionType *LTy = CGM.getTypes().GetFunctionType(FI);
2131
2132 llvm::Function *Fn =
2133 llvm::Function::Create(LTy, llvm::GlobalValue::LinkOnceODRLinkage,
2134 FuncName, &CGM.getModule());
2135 if (CGM.supportsCOMDAT())
2136 Fn->setComdat(CGM.getModule().getOrInsertComdat(FuncName));
2137
2138 SmallVector<QualType, 1> ArgTys;
2139 ArgTys.push_back(C.VoidPtrTy);
2140
2141 setBlockHelperAttributesVisibility(blockInfo.CapturesNonExternalType, Fn, FI,
2142 CGM);
2143 StartFunction(GlobalDecl(), ReturnTy, Fn, FI, args);
2144 markAsIgnoreThreadCheckingAtRuntime(Fn);
2145
2146 auto AL = ApplyDebugLocation::CreateArtificial(*this);
2147
2148 llvm::Type *structPtrTy = blockInfo.StructureType->getPointerTo();
2149
2150 Address src = GetAddrOfLocalVar(&SrcDecl);
2151 src = Address(Builder.CreateLoad(src), blockInfo.BlockAlign);
2152 src = Builder.CreateBitCast(src, structPtrTy, "block");
2153
2154 CodeGenFunction::RunCleanupsScope cleanups(*this);
2155
2156 for (const auto &DestroyedCapture : DestroyedCaptures) {
2157 const BlockDecl::Capture &CI = *DestroyedCapture.CI;
2158 const CGBlockInfo::Capture &capture = *DestroyedCapture.Capture;
2159 BlockFieldFlags flags = DestroyedCapture.DisposeFlags;
2160
2161 Address srcField = Builder.CreateStructGEP(src, capture.getIndex());
2162
2163 pushCaptureCleanup(DestroyedCapture.DisposeKind, srcField,
2164 CI.getVariable()->getType(), flags,
2165 /*ForCopyHelper*/ false, CI.getVariable(), *this);
2166 }
2167
2168 cleanups.ForceCleanup();
2169
2170 FinishFunction();
2171
2172 return llvm::ConstantExpr::getBitCast(Fn, VoidPtrTy);
2173}
2174
2175namespace {
2176
2177/// Emits the copy/dispose helper functions for a __block object of id type.
2178class ObjectByrefHelpers final : public BlockByrefHelpers {
2179 BlockFieldFlags Flags;
2180
2181public:
2182 ObjectByrefHelpers(CharUnits alignment, BlockFieldFlags flags)
2183 : BlockByrefHelpers(alignment), Flags(flags) {}
2184
2185 void emitCopy(CodeGenFunction &CGF, Address destField,
2186 Address srcField) override {
2187 destField = CGF.Builder.CreateBitCast(destField, CGF.VoidPtrTy);
2188
2189 srcField = CGF.Builder.CreateBitCast(srcField, CGF.VoidPtrPtrTy);
2190 llvm::Value *srcValue = CGF.Builder.CreateLoad(srcField);
2191
2192 unsigned flags = (Flags | BLOCK_BYREF_CALLER).getBitMask();
2193
2194 llvm::Value *flagsVal = llvm::ConstantInt::get(CGF.Int32Ty, flags);
2195 llvm::FunctionCallee fn = CGF.CGM.getBlockObjectAssign();
2196
2197 llvm::Value *args[] = { destField.getPointer(), srcValue, flagsVal };
2198 CGF.EmitNounwindRuntimeCall(fn, args);
2199 }
2200
2201 void emitDispose(CodeGenFunction &CGF, Address field) override {
2202 field = CGF.Builder.CreateBitCast(field, CGF.Int8PtrTy->getPointerTo(0));
2203 llvm::Value *value = CGF.Builder.CreateLoad(field);
2204
2205 CGF.BuildBlockRelease(value, Flags | BLOCK_BYREF_CALLER, false);
2206 }
2207
2208 void profileImpl(llvm::FoldingSetNodeID &id) const override {
2209 id.AddInteger(Flags.getBitMask());
2210 }
2211};
2212
2213/// Emits the copy/dispose helpers for an ARC __block __weak variable.
2214class ARCWeakByrefHelpers final : public BlockByrefHelpers {
2215public:
2216 ARCWeakByrefHelpers(CharUnits alignment) : BlockByrefHelpers(alignment) {}
2217
2218 void emitCopy(CodeGenFunction &CGF, Address destField,
2219 Address srcField) override {
2220 CGF.EmitARCMoveWeak(destField, srcField);
2221 }
2222
2223 void emitDispose(CodeGenFunction &CGF, Address field) override {
2224 CGF.EmitARCDestroyWeak(field);
2225 }
2226
2227 void profileImpl(llvm::FoldingSetNodeID &id) const override {
2228 // 0 is distinguishable from all pointers and byref flags
2229 id.AddInteger(0);
2230 }
2231};
2232
2233/// Emits the copy/dispose helpers for an ARC __block __strong variable
2234/// that's not of block-pointer type.
2235class ARCStrongByrefHelpers final : public BlockByrefHelpers {
2236public:
2237 ARCStrongByrefHelpers(CharUnits alignment) : BlockByrefHelpers(alignment) {}
2238
2239 void emitCopy(CodeGenFunction &CGF, Address destField,
2240 Address srcField) override {
2241 // Do a "move" by copying the value and then zeroing out the old
2242 // variable.
2243
2244 llvm::Value *value = CGF.Builder.CreateLoad(srcField);
2245
2246 llvm::Value *null =
2247 llvm::ConstantPointerNull::get(cast<llvm::PointerType>(value->getType()));
2248
2249 if (CGF.CGM.getCodeGenOpts().OptimizationLevel == 0) {
2250 CGF.Builder.CreateStore(null, destField);
2251 CGF.EmitARCStoreStrongCall(destField, value, /*ignored*/ true);
2252 CGF.EmitARCStoreStrongCall(srcField, null, /*ignored*/ true);
2253 return;
2254 }
2255 CGF.Builder.CreateStore(value, destField);
2256 CGF.Builder.CreateStore(null, srcField);
2257 }
2258
2259 void emitDispose(CodeGenFunction &CGF, Address field) override {
2260 CGF.EmitARCDestroyStrong(field, ARCImpreciseLifetime);
2261 }
2262
2263 void profileImpl(llvm::FoldingSetNodeID &id) const override {
2264 // 1 is distinguishable from all pointers and byref flags
2265 id.AddInteger(1);
2266 }
2267};
2268
2269/// Emits the copy/dispose helpers for an ARC __block __strong
2270/// variable that's of block-pointer type.
2271class ARCStrongBlockByrefHelpers final : public BlockByrefHelpers {
2272public:
2273 ARCStrongBlockByrefHelpers(CharUnits alignment)
2274 : BlockByrefHelpers(alignment) {}
2275
2276 void emitCopy(CodeGenFunction &CGF, Address destField,
2277 Address srcField) override {
2278 // Do the copy with objc_retainBlock; that's all that
2279 // _Block_object_assign would do anyway, and we'd have to pass the
2280 // right arguments to make sure it doesn't get no-op'ed.
2281 llvm::Value *oldValue = CGF.Builder.CreateLoad(srcField);
2282 llvm::Value *copy = CGF.EmitARCRetainBlock(oldValue, /*mandatory*/ true);
2283 CGF.Builder.CreateStore(copy, destField);
2284 }
2285
2286 void emitDispose(CodeGenFunction &CGF, Address field) override {
2287 CGF.EmitARCDestroyStrong(field, ARCImpreciseLifetime);
2288 }
2289
2290 void profileImpl(llvm::FoldingSetNodeID &id) const override {
2291 // 2 is distinguishable from all pointers and byref flags
2292 id.AddInteger(2);
2293 }
2294};
2295
2296/// Emits the copy/dispose helpers for a __block variable with a
2297/// nontrivial copy constructor or destructor.
2298class CXXByrefHelpers final : public BlockByrefHelpers {
2299 QualType VarType;
2300 const Expr *CopyExpr;
2301
2302public:
2303 CXXByrefHelpers(CharUnits alignment, QualType type,
2304 const Expr *copyExpr)
2305 : BlockByrefHelpers(alignment), VarType(type), CopyExpr(copyExpr) {}
2306
2307 bool needsCopy() const override { return CopyExpr != nullptr; }
2308 void emitCopy(CodeGenFunction &CGF, Address destField,
2309 Address srcField) override {
2310 if (!CopyExpr) return;
2311 CGF.EmitSynthesizedCXXCopyCtor(destField, srcField, CopyExpr);
2312 }
2313
2314 void emitDispose(CodeGenFunction &CGF, Address field) override {
2315 EHScopeStack::stable_iterator cleanupDepth = CGF.EHStack.stable_begin();
2316 CGF.PushDestructorCleanup(VarType, field);
2317 CGF.PopCleanupBlocks(cleanupDepth);
2318 }
2319
2320 void profileImpl(llvm::FoldingSetNodeID &id) const override {
2321 id.AddPointer(VarType.getCanonicalType().getAsOpaquePtr());
2322 }
2323};
2324
2325/// Emits the copy/dispose helpers for a __block variable that is a non-trivial
2326/// C struct.
2327class NonTrivialCStructByrefHelpers final : public BlockByrefHelpers {
2328 QualType VarType;
2329
2330public:
2331 NonTrivialCStructByrefHelpers(CharUnits alignment, QualType type)
2332 : BlockByrefHelpers(alignment), VarType(type) {}
2333
2334 void emitCopy(CodeGenFunction &CGF, Address destField,
2335 Address srcField) override {
2336 CGF.callCStructMoveConstructor(CGF.MakeAddrLValue(destField, VarType),
2337 CGF.MakeAddrLValue(srcField, VarType));
2338 }
2339
2340 bool needsDispose() const override {
2341 return VarType.isDestructedType();
2342 }
2343
2344 void emitDispose(CodeGenFunction &CGF, Address field) override {
2345 EHScopeStack::stable_iterator cleanupDepth = CGF.EHStack.stable_begin();
2346 CGF.pushDestroy(VarType.isDestructedType(), field, VarType);
2347 CGF.PopCleanupBlocks(cleanupDepth);
2348 }
2349
2350 void profileImpl(llvm::FoldingSetNodeID &id) const override {
2351 id.AddPointer(VarType.getCanonicalType().getAsOpaquePtr());
2352 }
2353};
2354} // end anonymous namespace
2355
2356static llvm::Constant *
2357generateByrefCopyHelper(CodeGenFunction &CGF, const BlockByrefInfo &byrefInfo,
2358 BlockByrefHelpers &generator) {
2359 ASTContext &Context = CGF.getContext();
2360
2361 QualType ReturnTy = Context.VoidTy;
2362
2363 FunctionArgList args;
2364 ImplicitParamDecl Dst(Context, Context.VoidPtrTy, ImplicitParamDecl::Other);
2365 args.push_back(&Dst);
2366
2367 ImplicitParamDecl Src(Context, Context.VoidPtrTy, ImplicitParamDecl::Other);
2368 args.push_back(&Src);
2369
2370 const CGFunctionInfo &FI =
2371 CGF.CGM.getTypes().arrangeBuiltinFunctionDeclaration(ReturnTy, args);
2372
2373 llvm::FunctionType *LTy = CGF.CGM.getTypes().GetFunctionType(FI);
2374
2375 // FIXME: We'd like to put these into a mergable by content, with
2376 // internal linkage.
2377 llvm::Function *Fn =
2378 llvm::Function::Create(LTy, llvm::GlobalValue::InternalLinkage,
2379 "__Block_byref_object_copy_", &CGF.CGM.getModule());
2380
2381 SmallVector<QualType, 2> ArgTys;
2382 ArgTys.push_back(Context.VoidPtrTy);
2383 ArgTys.push_back(Context.VoidPtrTy);
2384
2385 CGF.CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, FI);
2386
2387 CGF.StartFunction(GlobalDecl(), ReturnTy, Fn, FI, args);
2388 // Create a scope with an artificial location for the body of this function.
2389 auto AL = ApplyDebugLocation::CreateArtificial(CGF);
2390
2391 if (generator.needsCopy()) {
2392 llvm::Type *byrefPtrType = byrefInfo.Type->getPointerTo(0);
2393
2394 // dst->x
2395 Address destField = CGF.GetAddrOfLocalVar(&Dst);
2396 destField = Address(CGF.Builder.CreateLoad(destField),
2397 byrefInfo.ByrefAlignment);
2398 destField = CGF.Builder.CreateBitCast(destField, byrefPtrType);
2399 destField = CGF.emitBlockByrefAddress(destField, byrefInfo, false,
2400 "dest-object");
2401
2402 // src->x
2403 Address srcField = CGF.GetAddrOfLocalVar(&Src);
2404 srcField = Address(CGF.Builder.CreateLoad(srcField),
2405 byrefInfo.ByrefAlignment);
2406 srcField = CGF.Builder.CreateBitCast(srcField, byrefPtrType);
2407 srcField = CGF.emitBlockByrefAddress(srcField, byrefInfo, false,
2408 "src-object");
2409
2410 generator.emitCopy(CGF, destField, srcField);
2411 }
2412
2413 CGF.FinishFunction();
2414
2415 return llvm::ConstantExpr::getBitCast(Fn, CGF.Int8PtrTy);
2416}
2417
2418/// Build the copy helper for a __block variable.
2419static llvm::Constant *buildByrefCopyHelper(CodeGenModule &CGM,
2420 const BlockByrefInfo &byrefInfo,
2421 BlockByrefHelpers &generator) {
2422 CodeGenFunction CGF(CGM);
2423 return generateByrefCopyHelper(CGF, byrefInfo, generator);
2424}
2425
2426/// Generate code for a __block variable's dispose helper.
2427static llvm::Constant *
2428generateByrefDisposeHelper(CodeGenFunction &CGF,
2429 const BlockByrefInfo &byrefInfo,
2430 BlockByrefHelpers &generator) {
2431 ASTContext &Context = CGF.getContext();
2432 QualType R = Context.VoidTy;
2433
2434 FunctionArgList args;
2435 ImplicitParamDecl Src(CGF.getContext(), Context.VoidPtrTy,
2436 ImplicitParamDecl::Other);
2437 args.push_back(&Src);
2438
2439 const CGFunctionInfo &FI =
2440 CGF.CGM.getTypes().arrangeBuiltinFunctionDeclaration(R, args);
2441
2442 llvm::FunctionType *LTy = CGF.CGM.getTypes().GetFunctionType(FI);
2443
2444 // FIXME: We'd like to put these into a mergable by content, with
2445 // internal linkage.
2446 llvm::Function *Fn =
2447 llvm::Function::Create(LTy, llvm::GlobalValue::InternalLinkage,
2448 "__Block_byref_object_dispose_",
2449 &CGF.CGM.getModule());
2450
2451 SmallVector<QualType, 1> ArgTys;
2452 ArgTys.push_back(Context.VoidPtrTy);
2453
2454 CGF.CGM.SetInternalFunctionAttributes(GlobalDecl(), Fn, FI);
2455
2456 CGF.StartFunction(GlobalDecl(), R, Fn, FI, args);
2457 // Create a scope with an artificial location for the body of this function.
2458 auto AL = ApplyDebugLocation::CreateArtificial(CGF);
2459
2460 if (generator.needsDispose()) {
2461 Address addr = CGF.GetAddrOfLocalVar(&Src);
2462 addr = Address(CGF.Builder.CreateLoad(addr), byrefInfo.ByrefAlignment);
2463 auto byrefPtrType = byrefInfo.Type->getPointerTo(0);
2464 addr = CGF.Builder.CreateBitCast(addr, byrefPtrType);
2465 addr = CGF.emitBlockByrefAddress(addr, byrefInfo, false, "object");
2466
2467 generator.emitDispose(CGF, addr);
2468 }
2469
2470 CGF.FinishFunction();
2471
2472 return llvm::ConstantExpr::getBitCast(Fn, CGF.Int8PtrTy);
2473}
2474
2475/// Build the dispose helper for a __block variable.
2476static llvm::Constant *buildByrefDisposeHelper(CodeGenModule &CGM,
2477 const BlockByrefInfo &byrefInfo,
2478 BlockByrefHelpers &generator) {
2479 CodeGenFunction CGF(CGM);
2480 return generateByrefDisposeHelper(CGF, byrefInfo, generator);
2481}
2482
2483/// Lazily build the copy and dispose helpers for a __block variable
2484/// with the given information.
2485template <class T>
2486static T *buildByrefHelpers(CodeGenModule &CGM, const BlockByrefInfo &byrefInfo,
2487 T &&generator) {
2488 llvm::FoldingSetNodeID id;
2489 generator.Profile(id);
2490
2491 void *insertPos;
2492 BlockByrefHelpers *node
2493 = CGM.ByrefHelpersCache.FindNodeOrInsertPos(id, insertPos);
2494 if (node) return static_cast<T*>(node);
2495
2496 generator.CopyHelper = buildByrefCopyHelper(CGM, byrefInfo, generator);
2497 generator.DisposeHelper = buildByrefDisposeHelper(CGM, byrefInfo, generator);
2498
2499 T *copy = new (CGM.getContext()) T(std::forward<T>(generator));
2500 CGM.ByrefHelpersCache.InsertNode(copy, insertPos);
2501 return copy;
2502}
2503
2504/// Build the copy and dispose helpers for the given __block variable
2505/// emission. Places the helpers in the global cache. Returns null
2506/// if no helpers are required.
2507BlockByrefHelpers *
2508CodeGenFunction::buildByrefHelpers(llvm::StructType &byrefType,
2509 const AutoVarEmission &emission) {
2510 const VarDecl &var = *emission.Variable;
2511 assert(var.isEscapingByref() &&((void)0)
2512 "only escaping __block variables need byref helpers")((void)0);
2513
2514 QualType type = var.getType();
2515
2516 auto &byrefInfo = getBlockByrefInfo(&var);
2517
2518 // The alignment we care about for the purposes of uniquing byref
2519 // helpers is the alignment of the actual byref value field.
2520 CharUnits valueAlignment =
2521 byrefInfo.ByrefAlignment.alignmentAtOffset(byrefInfo.FieldOffset);
2522
2523 if (const CXXRecordDecl *record = type->getAsCXXRecordDecl()) {
2524 const Expr *copyExpr =
2525 CGM.getContext().getBlockVarCopyInit(&var).getCopyExpr();
2526 if (!copyExpr && record->hasTrivialDestructor()) return nullptr;
2527
2528 return ::buildByrefHelpers(
2529 CGM, byrefInfo, CXXByrefHelpers(valueAlignment, type, copyExpr));
2530 }
2531
2532 // If type is a non-trivial C struct type that is non-trivial to
2533 // destructly move or destroy, build the copy and dispose helpers.
2534 if (type.isNonTrivialToPrimitiveDestructiveMove() == QualType::PCK_Struct ||
2535 type.isDestructedType() == QualType::DK_nontrivial_c_struct)
2536 return ::buildByrefHelpers(
2537 CGM, byrefInfo, NonTrivialCStructByrefHelpers(valueAlignment, type));
2538
2539 // Otherwise, if we don't have a retainable type, there's nothing to do.
2540 // that the runtime does extra copies.
2541 if (!type->isObjCRetainableType()) return nullptr;
2542
2543 Qualifiers qs = type.getQualifiers();
2544
2545 // If we have lifetime, that dominates.
2546 if (Qualifiers::ObjCLifetime lifetime = qs.getObjCLifetime()) {
2547 switch (lifetime) {
2548 case Qualifiers::OCL_None: llvm_unreachable("impossible")__builtin_unreachable();
2549
2550 // These are just bits as far as the runtime is concerned.
2551 case Qualifiers::OCL_ExplicitNone:
2552 case Qualifiers::OCL_Autoreleasing:
2553 return nullptr;
2554
2555 // Tell the runtime that this is ARC __weak, called by the
2556 // byref routines.
2557 case Qualifiers::OCL_Weak:
2558 return ::buildByrefHelpers(CGM, byrefInfo,
2559 ARCWeakByrefHelpers(valueAlignment));
2560
2561 // ARC __strong __block variables need to be retained.
2562 case Qualifiers::OCL_Strong:
2563 // Block pointers need to be copied, and there's no direct
2564 // transfer possible.
2565 if (type->isBlockPointerType()) {
2566 return ::buildByrefHelpers(CGM, byrefInfo,
2567 ARCStrongBlockByrefHelpers(valueAlignment));
2568
2569 // Otherwise, we transfer ownership of the retain from the stack
2570 // to the heap.
2571 } else {
2572 return ::buildByrefHelpers(CGM, byrefInfo,
2573 ARCStrongByrefHelpers(valueAlignment));
2574 }
2575 }
2576 llvm_unreachable("fell out of lifetime switch!")__builtin_unreachable();
2577 }
2578
2579 BlockFieldFlags flags;
2580 if (type->isBlockPointerType()) {
2581 flags |= BLOCK_FIELD_IS_BLOCK;
2582 } else if (CGM.getContext().isObjCNSObjectType(type) ||
2583 type->isObjCObjectPointerType()) {
2584 flags |= BLOCK_FIELD_IS_OBJECT;
2585 } else {
2586 return nullptr;
2587 }
2588
2589 if (type.isObjCGCWeak())
2590 flags |= BLOCK_FIELD_IS_WEAK;
2591
2592 return ::buildByrefHelpers(CGM, byrefInfo,
2593 ObjectByrefHelpers(valueAlignment, flags));
2594}
2595
2596Address CodeGenFunction::emitBlockByrefAddress(Address baseAddr,
2597 const VarDecl *var,
2598 bool followForward) {
2599 auto &info = getBlockByrefInfo(var);
2600 return emitBlockByrefAddress(baseAddr, info, followForward, var->getName());
2601}
2602
2603Address CodeGenFunction::emitBlockByrefAddress(Address baseAddr,
2604 const BlockByrefInfo &info,
2605 bool followForward,
2606 const llvm::Twine &name) {
2607 // Chase the forwarding address if requested.
2608 if (followForward) {
2609 Address forwardingAddr = Builder.CreateStructGEP(baseAddr, 1, "forwarding");
2610 baseAddr = Address(Builder.CreateLoad(forwardingAddr), info.ByrefAlignment);
2611 }
2612
2613 return Builder.CreateStructGEP(baseAddr, info.FieldIndex, name);
2614}
2615
2616/// BuildByrefInfo - This routine changes a __block variable declared as T x
2617/// into:
2618///
2619/// struct {
2620/// void *__isa;
2621/// void *__forwarding;
2622/// int32_t __flags;
2623/// int32_t __size;
2624/// void *__copy_helper; // only if needed
2625/// void *__destroy_helper; // only if needed
2626/// void *__byref_variable_layout;// only if needed
2627/// char padding[X]; // only if needed
2628/// T x;
2629/// } x
2630///
2631const BlockByrefInfo &CodeGenFunction::getBlockByrefInfo(const VarDecl *D) {
2632 auto it = BlockByrefInfos.find(D);
2633 if (it != BlockByrefInfos.end())
2634 return it->second;
2635
2636 llvm::StructType *byrefType =
2637 llvm::StructType::create(getLLVMContext(),
2638 "struct.__block_byref_" + D->getNameAsString());
2639
2640 QualType Ty = D->getType();
2641
2642 CharUnits size;
2643 SmallVector<llvm::Type *, 8> types;
2644
2645 // void *__isa;
2646 types.push_back(Int8PtrTy);
2647 size += getPointerSize();
2648
2649 // void *__forwarding;
2650 types.push_back(llvm::PointerType::getUnqual(byrefType));
2651 size += getPointerSize();
2652
2653 // int32_t __flags;
2654 types.push_back(Int32Ty);
2655 size += CharUnits::fromQuantity(4);
2656
2657 // int32_t __size;
2658 types.push_back(Int32Ty);
2659 size += CharUnits::fromQuantity(4);
2660
2661 // Note that this must match *exactly* the logic in buildByrefHelpers.
2662 bool hasCopyAndDispose = getContext().BlockRequiresCopying(Ty, D);
2663 if (hasCopyAndDispose) {
2664 /// void *__copy_helper;
2665 types.push_back(Int8PtrTy);
2666 size += getPointerSize();
2667
2668 /// void *__destroy_helper;
2669 types.push_back(Int8PtrTy);
2670 size += getPointerSize();
2671 }
2672
2673 bool HasByrefExtendedLayout = false;
2674 Qualifiers::ObjCLifetime Lifetime = Qualifiers::OCL_None;
2675 if (getContext().getByrefLifetime(Ty, Lifetime, HasByrefExtendedLayout) &&
2676 HasByrefExtendedLayout) {
2677 /// void *__byref_variable_layout;
2678 types.push_back(Int8PtrTy);
2679 size += CharUnits::fromQuantity(PointerSizeInBytes);
2680 }
2681
2682 // T x;
2683 llvm::Type *varTy = ConvertTypeForMem(Ty);
2684
2685 bool packed = false;
2686 CharUnits varAlign = getContext().getDeclAlign(D);
2687 CharUnits varOffset = size.alignTo(varAlign);
2688
2689 // We may have to insert padding.
2690 if (varOffset != size) {
2691 llvm::Type *paddingTy =
2692 llvm::ArrayType::get(Int8Ty, (varOffset - size).getQuantity());
2693
2694 types.push_back(paddingTy);
2695 size = varOffset;
2696
2697 // Conversely, we might have to prevent LLVM from inserting padding.
2698 } else if (CGM.getDataLayout().getABITypeAlignment(varTy)
2699 > varAlign.getQuantity()) {
2700 packed = true;
2701 }
2702 types.push_back(varTy);
2703
2704 byrefType->setBody(types, packed);
2705
2706 BlockByrefInfo info;
2707 info.Type = byrefType;
2708 info.FieldIndex = types.size() - 1;
2709 info.FieldOffset = varOffset;
2710 info.ByrefAlignment = std::max(varAlign, getPointerAlign());
2711
2712 auto pair = BlockByrefInfos.insert({D, info});
2713 assert(pair.second && "info was inserted recursively?")((void)0);
2714 return pair.first->second;
2715}
2716
2717/// Initialize the structural components of a __block variable, i.e.
2718/// everything but the actual object.
2719void CodeGenFunction::emitByrefStructureInit(const AutoVarEmission &emission) {
2720 // Find the address of the local.
2721 Address addr = emission.Addr;
2722
2723 // That's an alloca of the byref structure type.
2724 llvm::StructType *byrefType = cast<llvm::StructType>(
2725 cast<llvm::PointerType>(addr.getPointer()->getType())->getElementType());
2726
2727 unsigned nextHeaderIndex = 0;
2728 CharUnits nextHeaderOffset;
2729 auto storeHeaderField = [&](llvm::Value *value, CharUnits fieldSize,
2730 const Twine &name) {
2731 auto fieldAddr = Builder.CreateStructGEP(addr, nextHeaderIndex, name);
2732 Builder.CreateStore(value, fieldAddr);
2733
2734 nextHeaderIndex++;
2735 nextHeaderOffset += fieldSize;
2736 };
2737
2738 // Build the byref helpers if necessary. This is null if we don't need any.
2739 BlockByrefHelpers *helpers = buildByrefHelpers(*byrefType, emission);
2740
2741 const VarDecl &D = *emission.Variable;
2742 QualType type = D.getType();
2743
2744 bool HasByrefExtendedLayout = false;
2745 Qualifiers::ObjCLifetime ByrefLifetime = Qualifiers::OCL_None;
2746 bool ByRefHasLifetime =
2747 getContext().getByrefLifetime(type, ByrefLifetime, HasByrefExtendedLayout);
2748
2749 llvm::Value *V;
2750
2751 // Initialize the 'isa', which is just 0 or 1.
2752 int isa = 0;
2753 if (type.isObjCGCWeak())
2754 isa = 1;
2755 V = Builder.CreateIntToPtr(Builder.getInt32(isa), Int8PtrTy, "isa");
2756 storeHeaderField(V, getPointerSize(), "byref.isa");
2757
2758 // Store the address of the variable into its own forwarding pointer.
2759 storeHeaderField(addr.getPointer(), getPointerSize(), "byref.forwarding");
2760
2761 // Blocks ABI:
2762 // c) the flags field is set to either 0 if no helper functions are
2763 // needed or BLOCK_BYREF_HAS_COPY_DISPOSE if they are,
2764 BlockFlags flags;
2765 if (helpers) flags |= BLOCK_BYREF_HAS_COPY_DISPOSE;
2766 if (ByRefHasLifetime) {
2767 if (HasByrefExtendedLayout) flags |= BLOCK_BYREF_LAYOUT_EXTENDED;
2768 else switch (ByrefLifetime) {
2769 case Qualifiers::OCL_Strong:
2770 flags |= BLOCK_BYREF_LAYOUT_STRONG;
2771 break;
2772 case Qualifiers::OCL_Weak:
2773 flags |= BLOCK_BYREF_LAYOUT_WEAK;
2774 break;
2775 case Qualifiers::OCL_ExplicitNone:
2776 flags |= BLOCK_BYREF_LAYOUT_UNRETAINED;
2777 break;
2778 case Qualifiers::OCL_None:
2779 if (!type->isObjCObjectPointerType() && !type->isBlockPointerType())
2780 flags |= BLOCK_BYREF_LAYOUT_NON_OBJECT;
2781 break;
2782 default:
2783 break;
2784 }
2785 if (CGM.getLangOpts().ObjCGCBitmapPrint) {
2786 printf("\n Inline flag for BYREF variable layout (%d):", flags.getBitMask());
2787 if (flags & BLOCK_BYREF_HAS_COPY_DISPOSE)
2788 printf(" BLOCK_BYREF_HAS_COPY_DISPOSE");
2789 if (flags & BLOCK_BYREF_LAYOUT_MASK) {
2790 BlockFlags ThisFlag(flags.getBitMask() & BLOCK_BYREF_LAYOUT_MASK);
2791 if (ThisFlag == BLOCK_BYREF_LAYOUT_EXTENDED)
2792 printf(" BLOCK_BYREF_LAYOUT_EXTENDED");
2793 if (ThisFlag == BLOCK_BYREF_LAYOUT_STRONG)
2794 printf(" BLOCK_BYREF_LAYOUT_STRONG");
2795 if (ThisFlag == BLOCK_BYREF_LAYOUT_WEAK)
2796 printf(" BLOCK_BYREF_LAYOUT_WEAK");
2797 if (ThisFlag == BLOCK_BYREF_LAYOUT_UNRETAINED)
2798 printf(" BLOCK_BYREF_LAYOUT_UNRETAINED");
2799 if (ThisFlag == BLOCK_BYREF_LAYOUT_NON_OBJECT)
2800 printf(" BLOCK_BYREF_LAYOUT_NON_OBJECT");
2801 }
2802 printf("\n");
2803 }
2804 }
2805 storeHeaderField(llvm::ConstantInt::get(IntTy, flags.getBitMask()),
2806 getIntSize(), "byref.flags");
2807
2808 CharUnits byrefSize = CGM.GetTargetTypeStoreSize(byrefType);
2809 V = llvm::ConstantInt::get(IntTy, byrefSize.getQuantity());
2810 storeHeaderField(V, getIntSize(), "byref.size");
2811
2812 if (helpers) {
2813 storeHeaderField(helpers->CopyHelper, getPointerSize(),
2814 "byref.copyHelper");
2815 storeHeaderField(helpers->DisposeHelper, getPointerSize(),
2816 "byref.disposeHelper");
2817 }
2818
2819 if (ByRefHasLifetime && HasByrefExtendedLayout) {
2820 auto layoutInfo = CGM.getObjCRuntime().BuildByrefLayout(CGM, type);
2821 storeHeaderField(layoutInfo, getPointerSize(), "byref.layout");
2822 }
2823}
2824
2825void CodeGenFunction::BuildBlockRelease(llvm::Value *V, BlockFieldFlags flags,
2826 bool CanThrow) {
2827 llvm::FunctionCallee F = CGM.getBlockObjectDispose();
2828 llvm::Value *args[] = {
2829 Builder.CreateBitCast(V, Int8PtrTy),
2830 llvm::ConstantInt::get(Int32Ty, flags.getBitMask())
2831 };
2832
2833 if (CanThrow)
2834 EmitRuntimeCallOrInvoke(F, args);
2835 else
2836 EmitNounwindRuntimeCall(F, args);
2837}
2838
2839void CodeGenFunction::enterByrefCleanup(CleanupKind Kind, Address Addr,
2840 BlockFieldFlags Flags,
2841 bool LoadBlockVarAddr, bool CanThrow) {
2842 EHStack.pushCleanup<CallBlockRelease>(Kind, Addr, Flags, LoadBlockVarAddr,
2843 CanThrow);
2844}
2845
2846/// Adjust the declaration of something from the blocks API.
2847static void configureBlocksRuntimeObject(CodeGenModule &CGM,
2848 llvm::Constant *C) {
2849 auto *GV = cast<llvm::GlobalValue>(C->stripPointerCasts());
2850
2851 if (CGM.getTarget().getTriple().isOSBinFormatCOFF()) {
2852 IdentifierInfo &II = CGM.getContext().Idents.get(C->getName());
2853 TranslationUnitDecl *TUDecl = CGM.getContext().getTranslationUnitDecl();
2854 DeclContext *DC = TranslationUnitDecl::castToDeclContext(TUDecl);
2855
2856 assert((isa<llvm::Function>(C->stripPointerCasts()) ||((void)0)
2857 isa<llvm::GlobalVariable>(C->stripPointerCasts())) &&((void)0)
2858 "expected Function or GlobalVariable")((void)0);
2859
2860 const NamedDecl *ND = nullptr;
2861 for (const auto *Result : DC->lookup(&II))
2862 if ((ND = dyn_cast<FunctionDecl>(Result)) ||
2863 (ND = dyn_cast<VarDecl>(Result)))
2864 break;
2865
2866 // TODO: support static blocks runtime
2867 if (GV->isDeclaration() && (!ND || !ND->hasAttr<DLLExportAttr>())) {
2868 GV->setDLLStorageClass(llvm::GlobalValue::DLLImportStorageClass);
2869 GV->setLinkage(llvm::GlobalValue::ExternalLinkage);
2870 } else {
2871 GV->setDLLStorageClass(llvm::GlobalValue::DLLExportStorageClass);
2872 GV->setLinkage(llvm::GlobalValue::ExternalLinkage);
2873 }
2874 }
2875
2876 if (CGM.getLangOpts().BlocksRuntimeOptional && GV->isDeclaration() &&
2877 GV->hasExternalLinkage())
2878 GV->setLinkage(llvm::GlobalValue::ExternalWeakLinkage);
2879
2880 CGM.setDSOLocal(GV);
2881}
2882
2883llvm::FunctionCallee CodeGenModule::getBlockObjectDispose() {
2884 if (BlockObjectDispose)
2885 return BlockObjectDispose;
2886
2887 llvm::Type *args[] = { Int8PtrTy, Int32Ty };
2888 llvm::FunctionType *fty
2889 = llvm::FunctionType::get(VoidTy, args, false);
2890 BlockObjectDispose = CreateRuntimeFunction(fty, "_Block_object_dispose");
2891 configureBlocksRuntimeObject(
2892 *this, cast<llvm::Constant>(BlockObjectDispose.getCallee()));
2893 return BlockObjectDispose;
2894}
2895
2896llvm::FunctionCallee CodeGenModule::getBlockObjectAssign() {
2897 if (BlockObjectAssign)
2898 return BlockObjectAssign;
2899
2900 llvm::Type *args[] = { Int8PtrTy, Int8PtrTy, Int32Ty };
2901 llvm::FunctionType *fty
2902 = llvm::FunctionType::get(VoidTy, args, false);
2903 BlockObjectAssign = CreateRuntimeFunction(fty, "_Block_object_assign");
2904 configureBlocksRuntimeObject(
2905 *this, cast<llvm::Constant>(BlockObjectAssign.getCallee()));
2906 return BlockObjectAssign;
2907}
2908
2909llvm::Constant *CodeGenModule::getNSConcreteGlobalBlock() {
2910 if (NSConcreteGlobalBlock)
2911 return NSConcreteGlobalBlock;
2912
2913 NSConcreteGlobalBlock =
2914 GetOrCreateLLVMGlobal("_NSConcreteGlobalBlock", Int8PtrTy, 0, nullptr);
2915 configureBlocksRuntimeObject(*this, NSConcreteGlobalBlock);
2916 return NSConcreteGlobalBlock;
2917}
2918
2919llvm::Constant *CodeGenModule::getNSConcreteStackBlock() {
2920 if (NSConcreteStackBlock)
2921 return NSConcreteStackBlock;
2922
2923 NSConcreteStackBlock =
2924 GetOrCreateLLVMGlobal("_NSConcreteStackBlock", Int8PtrTy, 0, nullptr);
2925 configureBlocksRuntimeObject(*this, NSConcreteStackBlock);
2926 return NSConcreteStackBlock;
2927}

/usr/src/gnu/usr.bin/clang/libclangCodeGen/../../../llvm/clang/include/clang/AST/CharUnits.h

1//===--- CharUnits.h - Character units for sizes and offsets ----*- C++ -*-===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file defines the CharUnits class
10//
11//===----------------------------------------------------------------------===//
12
13#ifndef LLVM_CLANG_AST_CHARUNITS_H
14#define LLVM_CLANG_AST_CHARUNITS_H
15
16#include "llvm/ADT/DenseMapInfo.h"
17#include "llvm/Support/Alignment.h"
18#include "llvm/Support/DataTypes.h"
19#include "llvm/Support/MathExtras.h"
20
21namespace clang {
22
23 /// CharUnits - This is an opaque type for sizes expressed in character units.
24 /// Instances of this type represent a quantity as a multiple of the size
25 /// of the standard C type, char, on the target architecture. As an opaque
26 /// type, CharUnits protects you from accidentally combining operations on
27 /// quantities in bit units and character units.
28 ///
29 /// In both C and C++, an object of type 'char', 'signed char', or 'unsigned
30 /// char' occupies exactly one byte, so 'character unit' and 'byte' refer to
31 /// the same quantity of storage. However, we use the term 'character unit'
32 /// rather than 'byte' to avoid an implication that a character unit is
33 /// exactly 8 bits.
34 ///
35 /// For portability, never assume that a target character is 8 bits wide. Use
36 /// CharUnit values wherever you calculate sizes, offsets, or alignments
37 /// in character units.
38 class CharUnits {
39 public:
40 typedef int64_t QuantityType;
41
42 private:
43 QuantityType Quantity = 0;
8
The value 0 is assigned to 'maxFieldAlign.Quantity'
44
45 explicit CharUnits(QuantityType C) : Quantity(C) {}
46
47 public:
48
49 /// CharUnits - A default constructor.
50 CharUnits() = default;
51
52 /// Zero - Construct a CharUnits quantity of zero.
53 static CharUnits Zero() {
54 return CharUnits(0);
55 }
56
57 /// One - Construct a CharUnits quantity of one.
58 static CharUnits One() {
59 return CharUnits(1);
60 }
61
62 /// fromQuantity - Construct a CharUnits quantity from a raw integer type.
63 static CharUnits fromQuantity(QuantityType Quantity) {
64 return CharUnits(Quantity);
65 }
66
67 // Compound assignment.
68 CharUnits& operator+= (const CharUnits &Other) {
69 Quantity += Other.Quantity;
70 return *this;
71 }
72 CharUnits& operator++ () {
73 ++Quantity;
74 return *this;
75 }
76 CharUnits operator++ (int) {
77 return CharUnits(Quantity++);
78 }
79 CharUnits& operator-= (const CharUnits &Other) {
80 Quantity -= Other.Quantity;
81 return *this;
82 }
83 CharUnits& operator-- () {
84 --Quantity;
85 return *this;
86 }
87 CharUnits operator-- (int) {
88 return CharUnits(Quantity--);
89 }
90
91 // Comparison operators.
92 bool operator== (const CharUnits &Other) const {
93 return Quantity == Other.Quantity;
94 }
95 bool operator!= (const CharUnits &Other) const {
96 return Quantity != Other.Quantity;
97 }
98
99 // Relational operators.
100 bool operator< (const CharUnits &Other) const {
101 return Quantity < Other.Quantity;
102 }
103 bool operator<= (const CharUnits &Other) const {
104 return Quantity <= Other.Quantity;
105 }
106 bool operator> (const CharUnits &Other) const {
107 return Quantity > Other.Quantity;
108 }
109 bool operator>= (const CharUnits &Other) const {
110 return Quantity >= Other.Quantity;
111 }
112
113 // Other predicates.
114
115 /// isZero - Test whether the quantity equals zero.
116 bool isZero() const { return Quantity == 0; }
117
118 /// isOne - Test whether the quantity equals one.
119 bool isOne() const { return Quantity == 1; }
120
121 /// isPositive - Test whether the quantity is greater than zero.
122 bool isPositive() const { return Quantity > 0; }
123
124 /// isNegative - Test whether the quantity is less than zero.
125 bool isNegative() const { return Quantity < 0; }
126
127 /// isPowerOfTwo - Test whether the quantity is a power of two.
128 /// Zero is not a power of two.
129 bool isPowerOfTwo() const {
130 return (Quantity & -Quantity) == Quantity;
131 }
132
133 /// Test whether this is a multiple of the other value.
134 ///
135 /// Among other things, this promises that
136 /// self.alignTo(N) will just return self.
137 bool isMultipleOf(CharUnits N) const {
138 return (*this % N) == 0;
139 }
140
141 // Arithmetic operators.
142 CharUnits operator* (QuantityType N) const {
143 return CharUnits(Quantity * N);
144 }
145 CharUnits &operator*= (QuantityType N) {
146 Quantity *= N;
147 return *this;
148 }
149 CharUnits operator/ (QuantityType N) const {
150 return CharUnits(Quantity / N);
151 }
152 CharUnits &operator/= (QuantityType N) {
153 Quantity /= N;
154 return *this;
155 }
156 QuantityType operator/ (const CharUnits &Other) const {
157 return Quantity / Other.Quantity;
158 }
159 CharUnits operator% (QuantityType N) const {
160 return CharUnits(Quantity % N);
161 }
162 QuantityType operator% (const CharUnits &Other) const {
163 return Quantity % Other.Quantity;
164 }
165 CharUnits operator+ (const CharUnits &Other) const {
166 return CharUnits(Quantity + Other.Quantity);
167 }
168 CharUnits operator- (const CharUnits &Other) const {
169 return CharUnits(Quantity - Other.Quantity);
170 }
171 CharUnits operator- () const {
172 return CharUnits(-Quantity);
173 }
174
175
176 // Conversions.
177
178 /// getQuantity - Get the raw integer representation of this quantity.
179 QuantityType getQuantity() const { return Quantity; }
180
181 /// getAsAlign - Returns Quantity as a valid llvm::Align,
182 /// Beware llvm::Align assumes power of two 8-bit bytes.
183 llvm::Align getAsAlign() const { return llvm::Align(Quantity); }
184
185 /// alignTo - Returns the next integer (mod 2**64) that is
186 /// greater than or equal to this quantity and is a multiple of \p Align.
187 /// Align must be non-zero.
188 CharUnits alignTo(const CharUnits &Align) const {
189 return CharUnits(llvm::alignTo(Quantity, Align.Quantity));
20
Passing the value 0 via 2nd parameter 'Align'
21
Calling 'alignTo'
190 }
191
192 /// Given that this is a non-zero alignment value, what is the
193 /// alignment at the given offset?
194 CharUnits alignmentAtOffset(CharUnits offset) const {
195 assert(Quantity != 0 && "offsetting from unknown alignment?")((void)0);
196 return CharUnits(llvm::MinAlign(Quantity, offset.Quantity));
197 }
198
199 /// Given that this is the alignment of the first element of an
200 /// array, return the minimum alignment of any element in the array.
201 CharUnits alignmentOfArrayElement(CharUnits elementSize) const {
202 // Since we don't track offsetted alignments, the alignment of
203 // the second element (or any odd element) will be minimally
204 // aligned.
205 return alignmentAtOffset(elementSize);
206 }
207
208
209 }; // class CharUnit
210} // namespace clang
211
212inline clang::CharUnits operator* (clang::CharUnits::QuantityType Scale,
213 const clang::CharUnits &CU) {
214 return CU * Scale;
215}
216
217namespace llvm {
218
219template<> struct DenseMapInfo<clang::CharUnits> {
220 static clang::CharUnits getEmptyKey() {
221 clang::CharUnits::QuantityType Quantity =
222 DenseMapInfo<clang::CharUnits::QuantityType>::getEmptyKey();
223
224 return clang::CharUnits::fromQuantity(Quantity);
225 }
226
227 static clang::CharUnits getTombstoneKey() {
228 clang::CharUnits::QuantityType Quantity =
229 DenseMapInfo<clang::CharUnits::QuantityType>::getTombstoneKey();
230
231 return clang::CharUnits::fromQuantity(Quantity);
232 }
233
234 static unsigned getHashValue(const clang::CharUnits &CU) {
235 clang::CharUnits::QuantityType Quantity = CU.getQuantity();
236 return DenseMapInfo<clang::CharUnits::QuantityType>::getHashValue(Quantity);
237 }
238
239 static bool isEqual(const clang::CharUnits &LHS,
240 const clang::CharUnits &RHS) {
241 return LHS == RHS;
242 }
243};
244
245} // end namespace llvm
246
247#endif // LLVM_CLANG_AST_CHARUNITS_H

/usr/src/gnu/usr.bin/clang/libclangCodeGen/../../../llvm/llvm/include/llvm/Support/MathExtras.h

1//===-- llvm/Support/MathExtras.h - Useful math functions -------*- C++ -*-===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file contains some functions that are useful for math stuff.
10//
11//===----------------------------------------------------------------------===//
12
13#ifndef LLVM_SUPPORT_MATHEXTRAS_H
14#define LLVM_SUPPORT_MATHEXTRAS_H
15
16#include "llvm/Support/Compiler.h"
17#include <cassert>
18#include <climits>
19#include <cmath>
20#include <cstdint>
21#include <cstring>
22#include <limits>
23#include <type_traits>
24
25#ifdef __ANDROID_NDK__
26#include <android/api-level.h>
27#endif
28
29#ifdef _MSC_VER
30// Declare these intrinsics manually rather including intrin.h. It's very
31// expensive, and MathExtras.h is popular.
32// #include <intrin.h>
33extern "C" {
34unsigned char _BitScanForward(unsigned long *_Index, unsigned long _Mask);
35unsigned char _BitScanForward64(unsigned long *_Index, unsigned __int64 _Mask);
36unsigned char _BitScanReverse(unsigned long *_Index, unsigned long _Mask);
37unsigned char _BitScanReverse64(unsigned long *_Index, unsigned __int64 _Mask);
38}
39#endif
40
41namespace llvm {
42
43/// The behavior an operation has on an input of 0.
44enum ZeroBehavior {
45 /// The returned value is undefined.
46 ZB_Undefined,
47 /// The returned value is numeric_limits<T>::max()
48 ZB_Max,
49 /// The returned value is numeric_limits<T>::digits
50 ZB_Width
51};
52
53/// Mathematical constants.
54namespace numbers {
55// TODO: Track C++20 std::numbers.
56// TODO: Favor using the hexadecimal FP constants (requires C++17).
57constexpr double e = 2.7182818284590452354, // (0x1.5bf0a8b145749P+1) https://oeis.org/A001113
58 egamma = .57721566490153286061, // (0x1.2788cfc6fb619P-1) https://oeis.org/A001620
59 ln2 = .69314718055994530942, // (0x1.62e42fefa39efP-1) https://oeis.org/A002162
60 ln10 = 2.3025850929940456840, // (0x1.24bb1bbb55516P+1) https://oeis.org/A002392
61 log2e = 1.4426950408889634074, // (0x1.71547652b82feP+0)
62 log10e = .43429448190325182765, // (0x1.bcb7b1526e50eP-2)
63 pi = 3.1415926535897932385, // (0x1.921fb54442d18P+1) https://oeis.org/A000796
64 inv_pi = .31830988618379067154, // (0x1.45f306bc9c883P-2) https://oeis.org/A049541
65 sqrtpi = 1.7724538509055160273, // (0x1.c5bf891b4ef6bP+0) https://oeis.org/A002161
66 inv_sqrtpi = .56418958354775628695, // (0x1.20dd750429b6dP-1) https://oeis.org/A087197
67 sqrt2 = 1.4142135623730950488, // (0x1.6a09e667f3bcdP+0) https://oeis.org/A00219
68 inv_sqrt2 = .70710678118654752440, // (0x1.6a09e667f3bcdP-1)
69 sqrt3 = 1.7320508075688772935, // (0x1.bb67ae8584caaP+0) https://oeis.org/A002194
70 inv_sqrt3 = .57735026918962576451, // (0x1.279a74590331cP-1)
71 phi = 1.6180339887498948482; // (0x1.9e3779b97f4a8P+0) https://oeis.org/A001622
72constexpr float ef = 2.71828183F, // (0x1.5bf0a8P+1) https://oeis.org/A001113
73 egammaf = .577215665F, // (0x1.2788d0P-1) https://oeis.org/A001620
74 ln2f = .693147181F, // (0x1.62e430P-1) https://oeis.org/A002162
75 ln10f = 2.30258509F, // (0x1.26bb1cP+1) https://oeis.org/A002392
76 log2ef = 1.44269504F, // (0x1.715476P+0)
77 log10ef = .434294482F, // (0x1.bcb7b2P-2)
78 pif = 3.14159265F, // (0x1.921fb6P+1) https://oeis.org/A000796
79 inv_pif = .318309886F, // (0x1.45f306P-2) https://oeis.org/A049541
80 sqrtpif = 1.77245385F, // (0x1.c5bf8aP+0) https://oeis.org/A002161
81 inv_sqrtpif = .564189584F, // (0x1.20dd76P-1) https://oeis.org/A087197
82 sqrt2f = 1.41421356F, // (0x1.6a09e6P+0) https://oeis.org/A002193
83 inv_sqrt2f = .707106781F, // (0x1.6a09e6P-1)
84 sqrt3f = 1.73205081F, // (0x1.bb67aeP+0) https://oeis.org/A002194
85 inv_sqrt3f = .577350269F, // (0x1.279a74P-1)
86 phif = 1.61803399F; // (0x1.9e377aP+0) https://oeis.org/A001622
87} // namespace numbers
88
89namespace detail {
90template <typename T, std::size_t SizeOfT> struct TrailingZerosCounter {
91 static unsigned count(T Val, ZeroBehavior) {
92 if (!Val)
93 return std::numeric_limits<T>::digits;
94 if (Val & 0x1)
95 return 0;
96
97 // Bisection method.
98 unsigned ZeroBits = 0;
99 T Shift = std::numeric_limits<T>::digits >> 1;
100 T Mask = std::numeric_limits<T>::max() >> Shift;
101 while (Shift) {
102 if ((Val & Mask) == 0) {
103 Val >>= Shift;
104 ZeroBits |= Shift;
105 }
106 Shift >>= 1;
107 Mask >>= Shift;
108 }
109 return ZeroBits;
110 }
111};
112
113#if defined(__GNUC__4) || defined(_MSC_VER)
114template <typename T> struct TrailingZerosCounter<T, 4> {
115 static unsigned count(T Val, ZeroBehavior ZB) {
116 if (ZB != ZB_Undefined && Val == 0)
117 return 32;
118
119#if __has_builtin(__builtin_ctz)1 || defined(__GNUC__4)
120 return __builtin_ctz(Val);
121#elif defined(_MSC_VER)
122 unsigned long Index;
123 _BitScanForward(&Index, Val);
124 return Index;
125#endif
126 }
127};
128
129#if !defined(_MSC_VER) || defined(_M_X64)
130template <typename T> struct TrailingZerosCounter<T, 8> {
131 static unsigned count(T Val, ZeroBehavior ZB) {
132 if (ZB != ZB_Undefined && Val == 0)
133 return 64;
134
135#if __has_builtin(__builtin_ctzll)1 || defined(__GNUC__4)
136 return __builtin_ctzll(Val);
137#elif defined(_MSC_VER)
138 unsigned long Index;
139 _BitScanForward64(&Index, Val);
140 return Index;
141#endif
142 }
143};
144#endif
145#endif
146} // namespace detail
147
148/// Count number of 0's from the least significant bit to the most
149/// stopping at the first 1.
150///
151/// Only unsigned integral types are allowed.
152///
153/// \param ZB the behavior on an input of 0. Only ZB_Width and ZB_Undefined are
154/// valid arguments.
155template <typename T>
156unsigned countTrailingZeros(T Val, ZeroBehavior ZB = ZB_Width) {
157 static_assert(std::numeric_limits<T>::is_integer &&
158 !std::numeric_limits<T>::is_signed,
159 "Only unsigned integral types are allowed.");
160 return llvm::detail::TrailingZerosCounter<T, sizeof(T)>::count(Val, ZB);
161}
162
163namespace detail {
164template <typename T, std::size_t SizeOfT> struct LeadingZerosCounter {
165 static unsigned count(T Val, ZeroBehavior) {
166 if (!Val)
167 return std::numeric_limits<T>::digits;
168
169 // Bisection method.
170 unsigned ZeroBits = 0;
171 for (T Shift = std::numeric_limits<T>::digits >> 1; Shift; Shift >>= 1) {
172 T Tmp = Val >> Shift;
173 if (Tmp)
174 Val = Tmp;
175 else
176 ZeroBits |= Shift;
177 }
178 return ZeroBits;
179 }
180};
181
182#if defined(__GNUC__4) || defined(_MSC_VER)
183template <typename T> struct LeadingZerosCounter<T, 4> {
184 static unsigned count(T Val, ZeroBehavior ZB) {
185 if (ZB != ZB_Undefined && Val == 0)
186 return 32;
187
188#if __has_builtin(__builtin_clz)1 || defined(__GNUC__4)
189 return __builtin_clz(Val);
190#elif defined(_MSC_VER)
191 unsigned long Index;
192 _BitScanReverse(&Index, Val);
193 return Index ^ 31;
194#endif
195 }
196};
197
198#if !defined(_MSC_VER) || defined(_M_X64)
199template <typename T> struct LeadingZerosCounter<T, 8> {
200 static unsigned count(T Val, ZeroBehavior ZB) {
201 if (ZB != ZB_Undefined && Val == 0)
202 return 64;
203
204#if __has_builtin(__builtin_clzll)1 || defined(__GNUC__4)
205 return __builtin_clzll(Val);
206#elif defined(_MSC_VER)
207 unsigned long Index;
208 _BitScanReverse64(&Index, Val);
209 return Index ^ 63;
210#endif
211 }
212};
213#endif
214#endif
215} // namespace detail
216
217/// Count number of 0's from the most significant bit to the least
218/// stopping at the first 1.
219///
220/// Only unsigned integral types are allowed.
221///
222/// \param ZB the behavior on an input of 0. Only ZB_Width and ZB_Undefined are
223/// valid arguments.
224template <typename T>
225unsigned countLeadingZeros(T Val, ZeroBehavior ZB = ZB_Width) {
226 static_assert(std::numeric_limits<T>::is_integer &&
227 !std::numeric_limits<T>::is_signed,
228 "Only unsigned integral types are allowed.");
229 return llvm::detail::LeadingZerosCounter<T, sizeof(T)>::count(Val, ZB);
230}
231
232/// Get the index of the first set bit starting from the least
233/// significant bit.
234///
235/// Only unsigned integral types are allowed.
236///
237/// \param ZB the behavior on an input of 0. Only ZB_Max and ZB_Undefined are
238/// valid arguments.
239template <typename T> T findFirstSet(T Val, ZeroBehavior ZB = ZB_Max) {
240 if (ZB == ZB_Max && Val == 0)
241 return std::numeric_limits<T>::max();
242
243 return countTrailingZeros(Val, ZB_Undefined);
244}
245
246/// Create a bitmask with the N right-most bits set to 1, and all other
247/// bits set to 0. Only unsigned types are allowed.
248template <typename T> T maskTrailingOnes(unsigned N) {
249 static_assert(std::is_unsigned<T>::value, "Invalid type!");
250 const unsigned Bits = CHAR_BIT8 * sizeof(T);
251 assert(N <= Bits && "Invalid bit index")((void)0);
252 return N == 0 ? 0 : (T(-1) >> (Bits - N));
253}
254
255/// Create a bitmask with the N left-most bits set to 1, and all other
256/// bits set to 0. Only unsigned types are allowed.
257template <typename T> T maskLeadingOnes(unsigned N) {
258 return ~maskTrailingOnes<T>(CHAR_BIT8 * sizeof(T) - N);
259}
260
261/// Create a bitmask with the N right-most bits set to 0, and all other
262/// bits set to 1. Only unsigned types are allowed.
263template <typename T> T maskTrailingZeros(unsigned N) {
264 return maskLeadingOnes<T>(CHAR_BIT8 * sizeof(T) - N);
265}
266
267/// Create a bitmask with the N left-most bits set to 0, and all other
268/// bits set to 1. Only unsigned types are allowed.
269template <typename T> T maskLeadingZeros(unsigned N) {
270 return maskTrailingOnes<T>(CHAR_BIT8 * sizeof(T) - N);
271}
272
273/// Get the index of the last set bit starting from the least
274/// significant bit.
275///
276/// Only unsigned integral types are allowed.
277///
278/// \param ZB the behavior on an input of 0. Only ZB_Max and ZB_Undefined are
279/// valid arguments.
280template <typename T> T findLastSet(T Val, ZeroBehavior ZB = ZB_Max) {
281 if (ZB == ZB_Max && Val == 0)
282 return std::numeric_limits<T>::max();
283
284 // Use ^ instead of - because both gcc and llvm can remove the associated ^
285 // in the __builtin_clz intrinsic on x86.
286 return countLeadingZeros(Val, ZB_Undefined) ^
287 (std::numeric_limits<T>::digits - 1);
288}
289
290/// Macro compressed bit reversal table for 256 bits.
291///
292/// http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable
293static const unsigned char BitReverseTable256[256] = {
294#define R2(n) n, n + 2 * 64, n + 1 * 64, n + 3 * 64
295#define R4(n) R2(n), R2(n + 2 * 16), R2(n + 1 * 16), R2(n + 3 * 16)
296#define R6(n) R4(n), R4(n + 2 * 4), R4(n + 1 * 4), R4(n + 3 * 4)
297 R6(0), R6(2), R6(1), R6(3)
298#undef R2
299#undef R4
300#undef R6
301};
302
303/// Reverse the bits in \p Val.
304template <typename T>
305T reverseBits(T Val) {
306 unsigned char in[sizeof(Val)];
307 unsigned char out[sizeof(Val)];
308 std::memcpy(in, &Val, sizeof(Val));
309 for (unsigned i = 0; i < sizeof(Val); ++i)
310 out[(sizeof(Val) - i) - 1] = BitReverseTable256[in[i]];
311 std::memcpy(&Val, out, sizeof(Val));
312 return Val;
313}
314
315#if __has_builtin(__builtin_bitreverse8)1
316template<>
317inline uint8_t reverseBits<uint8_t>(uint8_t Val) {
318 return __builtin_bitreverse8(Val);
319}
320#endif
321
322#if __has_builtin(__builtin_bitreverse16)1
323template<>
324inline uint16_t reverseBits<uint16_t>(uint16_t Val) {
325 return __builtin_bitreverse16(Val);
326}
327#endif
328
329#if __has_builtin(__builtin_bitreverse32)1
330template<>
331inline uint32_t reverseBits<uint32_t>(uint32_t Val) {
332 return __builtin_bitreverse32(Val);
333}
334#endif
335
336#if __has_builtin(__builtin_bitreverse64)1
337template<>
338inline uint64_t reverseBits<uint64_t>(uint64_t Val) {
339 return __builtin_bitreverse64(Val);
340}
341#endif
342
343// NOTE: The following support functions use the _32/_64 extensions instead of
344// type overloading so that signed and unsigned integers can be used without
345// ambiguity.
346
347/// Return the high 32 bits of a 64 bit value.
348constexpr inline uint32_t Hi_32(uint64_t Value) {
349 return static_cast<uint32_t>(Value >> 32);
350}
351
352/// Return the low 32 bits of a 64 bit value.
353constexpr inline uint32_t Lo_32(uint64_t Value) {
354 return static_cast<uint32_t>(Value);
355}
356
357/// Make a 64-bit integer from a high / low pair of 32-bit integers.
358constexpr inline uint64_t Make_64(uint32_t High, uint32_t Low) {
359 return ((uint64_t)High << 32) | (uint64_t)Low;
360}
361
362/// Checks if an integer fits into the given bit width.
363template <unsigned N> constexpr inline bool isInt(int64_t x) {
364 return N >= 64 || (-(INT64_C(1)1LL<<(N-1)) <= x && x < (INT64_C(1)1LL<<(N-1)));
365}
366// Template specializations to get better code for common cases.
367template <> constexpr inline bool isInt<8>(int64_t x) {
368 return static_cast<int8_t>(x) == x;
369}
370template <> constexpr inline bool isInt<16>(int64_t x) {
371 return static_cast<int16_t>(x) == x;
372}
373template <> constexpr inline bool isInt<32>(int64_t x) {
374 return static_cast<int32_t>(x) == x;
375}
376
377/// Checks if a signed integer is an N bit number shifted left by S.
378template <unsigned N, unsigned S>
379constexpr inline bool isShiftedInt(int64_t x) {
380 static_assert(
381 N > 0, "isShiftedInt<0> doesn't make sense (refers to a 0-bit number.");
382 static_assert(N + S <= 64, "isShiftedInt<N, S> with N + S > 64 is too wide.");
383 return isInt<N + S>(x) && (x % (UINT64_C(1)1ULL << S) == 0);
384}
385
386/// Checks if an unsigned integer fits into the given bit width.
387///
388/// This is written as two functions rather than as simply
389///
390/// return N >= 64 || X < (UINT64_C(1) << N);
391///
392/// to keep MSVC from (incorrectly) warning on isUInt<64> that we're shifting
393/// left too many places.
394template <unsigned N>
395constexpr inline std::enable_if_t<(N < 64), bool> isUInt(uint64_t X) {
396 static_assert(N > 0, "isUInt<0> doesn't make sense");
397 return X < (UINT64_C(1)1ULL << (N));
398}
399template <unsigned N>
400constexpr inline std::enable_if_t<N >= 64, bool> isUInt(uint64_t) {
401 return true;
402}
403
404// Template specializations to get better code for common cases.
405template <> constexpr inline bool isUInt<8>(uint64_t x) {
406 return static_cast<uint8_t>(x) == x;
407}
408template <> constexpr inline bool isUInt<16>(uint64_t x) {
409 return static_cast<uint16_t>(x) == x;
410}
411template <> constexpr inline bool isUInt<32>(uint64_t x) {
412 return static_cast<uint32_t>(x) == x;
413}
414
415/// Checks if a unsigned integer is an N bit number shifted left by S.
416template <unsigned N, unsigned S>
417constexpr inline bool isShiftedUInt(uint64_t x) {
418 static_assert(
419 N > 0, "isShiftedUInt<0> doesn't make sense (refers to a 0-bit number)");
420 static_assert(N + S <= 64,
421 "isShiftedUInt<N, S> with N + S > 64 is too wide.");
422 // Per the two static_asserts above, S must be strictly less than 64. So
423 // 1 << S is not undefined behavior.
424 return isUInt<N + S>(x) && (x % (UINT64_C(1)1ULL << S) == 0);
425}
426
427/// Gets the maximum value for a N-bit unsigned integer.
428inline uint64_t maxUIntN(uint64_t N) {
429 assert(N > 0 && N <= 64 && "integer width out of range")((void)0);
430
431 // uint64_t(1) << 64 is undefined behavior, so we can't do
432 // (uint64_t(1) << N) - 1
433 // without checking first that N != 64. But this works and doesn't have a
434 // branch.
435 return UINT64_MAX0xffffffffffffffffULL >> (64 - N);
436}
437
438/// Gets the minimum value for a N-bit signed integer.
439inline int64_t minIntN(int64_t N) {
440 assert(N > 0 && N <= 64 && "integer width out of range")((void)0);
441
442 return UINT64_C(1)1ULL + ~(UINT64_C(1)1ULL << (N - 1));
443}
444
445/// Gets the maximum value for a N-bit signed integer.
446inline int64_t maxIntN(int64_t N) {
447 assert(N > 0 && N <= 64 && "integer width out of range")((void)0);
448
449 // This relies on two's complement wraparound when N == 64, so we convert to
450 // int64_t only at the very end to avoid UB.
451 return (UINT64_C(1)1ULL << (N - 1)) - 1;
452}
453
454/// Checks if an unsigned integer fits into the given (dynamic) bit width.
455inline bool isUIntN(unsigned N, uint64_t x) {
456 return N >= 64 || x <= maxUIntN(N);
457}
458
459/// Checks if an signed integer fits into the given (dynamic) bit width.
460inline bool isIntN(unsigned N, int64_t x) {
461 return N >= 64 || (minIntN(N) <= x && x <= maxIntN(N));
462}
463
464/// Return true if the argument is a non-empty sequence of ones starting at the
465/// least significant bit with the remainder zero (32 bit version).
466/// Ex. isMask_32(0x0000FFFFU) == true.
467constexpr inline bool isMask_32(uint32_t Value) {
468 return Value && ((Value + 1) & Value) == 0;
469}
470
471/// Return true if the argument is a non-empty sequence of ones starting at the
472/// least significant bit with the remainder zero (64 bit version).
473constexpr inline bool isMask_64(uint64_t Value) {
474 return Value && ((Value + 1) & Value) == 0;
475}
476
477/// Return true if the argument contains a non-empty sequence of ones with the
478/// remainder zero (32 bit version.) Ex. isShiftedMask_32(0x0000FF00U) == true.
479constexpr inline bool isShiftedMask_32(uint32_t Value) {
480 return Value && isMask_32((Value - 1) | Value);
481}
482
483/// Return true if the argument contains a non-empty sequence of ones with the
484/// remainder zero (64 bit version.)
485constexpr inline bool isShiftedMask_64(uint64_t Value) {
486 return Value && isMask_64((Value - 1) | Value);
487}
488
489/// Return true if the argument is a power of two > 0.
490/// Ex. isPowerOf2_32(0x00100000U) == true (32 bit edition.)
491constexpr inline bool isPowerOf2_32(uint32_t Value) {
492 return Value && !(Value & (Value - 1));
493}
494
495/// Return true if the argument is a power of two > 0 (64 bit edition.)
496constexpr inline bool isPowerOf2_64(uint64_t Value) {
497 return Value && !(Value & (Value - 1));
498}
499
500/// Count the number of ones from the most significant bit to the first
501/// zero bit.
502///
503/// Ex. countLeadingOnes(0xFF0FFF00) == 8.
504/// Only unsigned integral types are allowed.
505///
506/// \param ZB the behavior on an input of all ones. Only ZB_Width and
507/// ZB_Undefined are valid arguments.
508template <typename T>
509unsigned countLeadingOnes(T Value, ZeroBehavior ZB = ZB_Width) {
510 static_assert(std::numeric_limits<T>::is_integer &&
511 !std::numeric_limits<T>::is_signed,
512 "Only unsigned integral types are allowed.");
513 return countLeadingZeros<T>(~Value, ZB);
514}
515
516/// Count the number of ones from the least significant bit to the first
517/// zero bit.
518///
519/// Ex. countTrailingOnes(0x00FF00FF) == 8.
520/// Only unsigned integral types are allowed.
521///
522/// \param ZB the behavior on an input of all ones. Only ZB_Width and
523/// ZB_Undefined are valid arguments.
524template <typename T>
525unsigned countTrailingOnes(T Value, ZeroBehavior ZB = ZB_Width) {
526 static_assert(std::numeric_limits<T>::is_integer &&
527 !std::numeric_limits<T>::is_signed,
528 "Only unsigned integral types are allowed.");
529 return countTrailingZeros<T>(~Value, ZB);
530}
531
532namespace detail {
533template <typename T, std::size_t SizeOfT> struct PopulationCounter {
534 static unsigned count(T Value) {
535 // Generic version, forward to 32 bits.
536 static_assert(SizeOfT <= 4, "Not implemented!");
537#if defined(__GNUC__4)
538 return __builtin_popcount(Value);
539#else
540 uint32_t v = Value;
541 v = v - ((v >> 1) & 0x55555555);
542 v = (v & 0x33333333) + ((v >> 2) & 0x33333333);
543 return ((v + (v >> 4) & 0xF0F0F0F) * 0x1010101) >> 24;
544#endif
545 }
546};
547
548template <typename T> struct PopulationCounter<T, 8> {
549 static unsigned count(T Value) {
550#if defined(__GNUC__4)
551 return __builtin_popcountll(Value);
552#else
553 uint64_t v = Value;
554 v = v - ((v >> 1) & 0x5555555555555555ULL);
555 v = (v & 0x3333333333333333ULL) + ((v >> 2) & 0x3333333333333333ULL);
556 v = (v + (v >> 4)) & 0x0F0F0F0F0F0F0F0FULL;
557 return unsigned((uint64_t)(v * 0x0101010101010101ULL) >> 56);
558#endif
559 }
560};
561} // namespace detail
562
563/// Count the number of set bits in a value.
564/// Ex. countPopulation(0xF000F000) = 8
565/// Returns 0 if the word is zero.
566template <typename T>
567inline unsigned countPopulation(T Value) {
568 static_assert(std::numeric_limits<T>::is_integer &&
569 !std::numeric_limits<T>::is_signed,
570 "Only unsigned integral types are allowed.");
571 return detail::PopulationCounter<T, sizeof(T)>::count(Value);
572}
573
574/// Compile time Log2.
575/// Valid only for positive powers of two.
576template <size_t kValue> constexpr inline size_t CTLog2() {
577 static_assert(kValue > 0 && llvm::isPowerOf2_64(kValue),
578 "Value is not a valid power of 2");
579 return 1 + CTLog2<kValue / 2>();
580}
581
582template <> constexpr inline size_t CTLog2<1>() { return 0; }
583
584/// Return the log base 2 of the specified value.
585inline double Log2(double Value) {
586#if defined(__ANDROID_API__) && __ANDROID_API__ < 18
587 return __builtin_log(Value) / __builtin_log(2.0);
588#else
589 return log2(Value);
590#endif
591}
592
593/// Return the floor log base 2 of the specified value, -1 if the value is zero.
594/// (32 bit edition.)
595/// Ex. Log2_32(32) == 5, Log2_32(1) == 0, Log2_32(0) == -1, Log2_32(6) == 2
596inline unsigned Log2_32(uint32_t Value) {
597 return 31 - countLeadingZeros(Value);
598}
599
600/// Return the floor log base 2 of the specified value, -1 if the value is zero.
601/// (64 bit edition.)
602inline unsigned Log2_64(uint64_t Value) {
603 return 63 - countLeadingZeros(Value);
604}
605
606/// Return the ceil log base 2 of the specified value, 32 if the value is zero.
607/// (32 bit edition).
608/// Ex. Log2_32_Ceil(32) == 5, Log2_32_Ceil(1) == 0, Log2_32_Ceil(6) == 3
609inline unsigned Log2_32_Ceil(uint32_t Value) {
610 return 32 - countLeadingZeros(Value - 1);
611}
612
613/// Return the ceil log base 2 of the specified value, 64 if the value is zero.
614/// (64 bit edition.)
615inline unsigned Log2_64_Ceil(uint64_t Value) {
616 return 64 - countLeadingZeros(Value - 1);
617}
618
619/// Return the greatest common divisor of the values using Euclid's algorithm.
620template <typename T>
621inline T greatestCommonDivisor(T A, T B) {
622 while (B) {
623 T Tmp = B;
624 B = A % B;
625 A = Tmp;
626 }
627 return A;
628}
629
630inline uint64_t GreatestCommonDivisor64(uint64_t A, uint64_t B) {
631 return greatestCommonDivisor<uint64_t>(A, B);
632}
633
634/// This function takes a 64-bit integer and returns the bit equivalent double.
635inline double BitsToDouble(uint64_t Bits) {
636 double D;
637 static_assert(sizeof(uint64_t) == sizeof(double), "Unexpected type sizes");
638 memcpy(&D, &Bits, sizeof(Bits));
639 return D;
640}
641
642/// This function takes a 32-bit integer and returns the bit equivalent float.
643inline float BitsToFloat(uint32_t Bits) {
644 float F;
645 static_assert(sizeof(uint32_t) == sizeof(float), "Unexpected type sizes");
646 memcpy(&F, &Bits, sizeof(Bits));
647 return F;
648}
649
650/// This function takes a double and returns the bit equivalent 64-bit integer.
651/// Note that copying doubles around changes the bits of NaNs on some hosts,
652/// notably x86, so this routine cannot be used if these bits are needed.
653inline uint64_t DoubleToBits(double Double) {
654 uint64_t Bits;
655 static_assert(sizeof(uint64_t) == sizeof(double), "Unexpected type sizes");
656 memcpy(&Bits, &Double, sizeof(Double));
657 return Bits;
658}
659
660/// This function takes a float and returns the bit equivalent 32-bit integer.
661/// Note that copying floats around changes the bits of NaNs on some hosts,
662/// notably x86, so this routine cannot be used if these bits are needed.
663inline uint32_t FloatToBits(float Float) {
664 uint32_t Bits;
665 static_assert(sizeof(uint32_t) == sizeof(float), "Unexpected type sizes");
666 memcpy(&Bits, &Float, sizeof(Float));
667 return Bits;
668}
669
670/// A and B are either alignments or offsets. Return the minimum alignment that
671/// may be assumed after adding the two together.
672constexpr inline uint64_t MinAlign(uint64_t A, uint64_t B) {
673 // The largest power of 2 that divides both A and B.
674 //
675 // Replace "-Value" by "1+~Value" in the following commented code to avoid
676 // MSVC warning C4146
677 // return (A | B) & -(A | B);
678 return (A | B) & (1 + ~(A | B));
679}
680
681/// Returns the next power of two (in 64-bits) that is strictly greater than A.
682/// Returns zero on overflow.
683inline uint64_t NextPowerOf2(uint64_t A) {
684 A |= (A >> 1);
685 A |= (A >> 2);
686 A |= (A >> 4);
687 A |= (A >> 8);
688 A |= (A >> 16);
689 A |= (A >> 32);
690 return A + 1;
691}
692
693/// Returns the power of two which is less than or equal to the given value.
694/// Essentially, it is a floor operation across the domain of powers of two.
695inline uint64_t PowerOf2Floor(uint64_t A) {
696 if (!A) return 0;
697 return 1ull << (63 - countLeadingZeros(A, ZB_Undefined));
698}
699
700/// Returns the power of two which is greater than or equal to the given value.
701/// Essentially, it is a ceil operation across the domain of powers of two.
702inline uint64_t PowerOf2Ceil(uint64_t A) {
703 if (!A)
704 return 0;
705 return NextPowerOf2(A - 1);
706}
707
708/// Returns the next integer (mod 2**64) that is greater than or equal to
709/// \p Value and is a multiple of \p Align. \p Align must be non-zero.
710///
711/// If non-zero \p Skew is specified, the return value will be a minimal
712/// integer that is greater than or equal to \p Value and equal to
713/// \p Align * N + \p Skew for some integer N. If \p Skew is larger than
714/// \p Align, its value is adjusted to '\p Skew mod \p Align'.
715///
716/// Examples:
717/// \code
718/// alignTo(5, 8) = 8
719/// alignTo(17, 8) = 24
720/// alignTo(~0LL, 8) = 0
721/// alignTo(321, 255) = 510
722///
723/// alignTo(5, 8, 7) = 7
724/// alignTo(17, 8, 1) = 17
725/// alignTo(~0LL, 8, 3) = 3
726/// alignTo(321, 255, 42) = 552
727/// \endcode
728inline uint64_t alignTo(uint64_t Value, uint64_t Align, uint64_t Skew = 0) {
729 assert(Align != 0u && "Align can't be 0.")((void)0);
730 Skew %= Align;
22
Division by zero
731 return (Value + Align - 1 - Skew) / Align * Align + Skew;
732}
733
734/// Returns the next integer (mod 2**64) that is greater than or equal to
735/// \p Value and is a multiple of \c Align. \c Align must be non-zero.
736template <uint64_t Align> constexpr inline uint64_t alignTo(uint64_t Value) {
737 static_assert(Align != 0u, "Align must be non-zero");
738 return (Value + Align - 1) / Align * Align;
739}
740
741/// Returns the integer ceil(Numerator / Denominator).
742inline uint64_t divideCeil(uint64_t Numerator, uint64_t Denominator) {
743 return alignTo(Numerator, Denominator) / Denominator;
744}
745
746/// Returns the integer nearest(Numerator / Denominator).
747inline uint64_t divideNearest(uint64_t Numerator, uint64_t Denominator) {
748 return (Numerator + (Denominator / 2)) / Denominator;
749}
750
751/// Returns the largest uint64_t less than or equal to \p Value and is
752/// \p Skew mod \p Align. \p Align must be non-zero
753inline uint64_t alignDown(uint64_t Value, uint64_t Align, uint64_t Skew = 0) {
754 assert(Align != 0u && "Align can't be 0.")((void)0);
755 Skew %= Align;
756 return (Value - Skew) / Align * Align + Skew;
757}
758
759/// Sign-extend the number in the bottom B bits of X to a 32-bit integer.
760/// Requires 0 < B <= 32.
761template <unsigned B> constexpr inline int32_t SignExtend32(uint32_t X) {
762 static_assert(B > 0, "Bit width can't be 0.");
763 static_assert(B <= 32, "Bit width out of range.");
764 return int32_t(X << (32 - B)) >> (32 - B);
765}
766
767/// Sign-extend the number in the bottom B bits of X to a 32-bit integer.
768/// Requires 0 < B <= 32.
769inline int32_t SignExtend32(uint32_t X, unsigned B) {
770 assert(B > 0 && "Bit width can't be 0.")((void)0);
771 assert(B <= 32 && "Bit width out of range.")((void)0);
772 return int32_t(X << (32 - B)) >> (32 - B);
773}
774
775/// Sign-extend the number in the bottom B bits of X to a 64-bit integer.
776/// Requires 0 < B <= 64.
777template <unsigned B> constexpr inline int64_t SignExtend64(uint64_t x) {
778 static_assert(B > 0, "Bit width can't be 0.");
779 static_assert(B <= 64, "Bit width out of range.");
780 return int64_t(x << (64 - B)) >> (64 - B);
781}
782
783/// Sign-extend the number in the bottom B bits of X to a 64-bit integer.
784/// Requires 0 < B <= 64.
785inline int64_t SignExtend64(uint64_t X, unsigned B) {
786 assert(B > 0 && "Bit width can't be 0.")((void)0);
787 assert(B <= 64 && "Bit width out of range.")((void)0);
788 return int64_t(X << (64 - B)) >> (64 - B);
789}
790
791/// Subtract two unsigned integers, X and Y, of type T and return the absolute
792/// value of the result.
793template <typename T>
794std::enable_if_t<std::is_unsigned<T>::value, T> AbsoluteDifference(T X, T Y) {
795 return X > Y ? (X - Y) : (Y - X);
796}
797
798/// Add two unsigned integers, X and Y, of type T. Clamp the result to the
799/// maximum representable value of T on overflow. ResultOverflowed indicates if
800/// the result is larger than the maximum representable value of type T.
801template <typename T>
802std::enable_if_t<std::is_unsigned<T>::value, T>
803SaturatingAdd(T X, T Y, bool *ResultOverflowed = nullptr) {
804 bool Dummy;
805 bool &Overflowed = ResultOverflowed ? *ResultOverflowed : Dummy;
806 // Hacker's Delight, p. 29
807 T Z = X + Y;
808 Overflowed = (Z < X || Z < Y);
809 if (Overflowed)
810 return std::numeric_limits<T>::max();
811 else
812 return Z;
813}
814
815/// Multiply two unsigned integers, X and Y, of type T. Clamp the result to the
816/// maximum representable value of T on overflow. ResultOverflowed indicates if
817/// the result is larger than the maximum representable value of type T.
818template <typename T>
819std::enable_if_t<std::is_unsigned<T>::value, T>
820SaturatingMultiply(T X, T Y, bool *ResultOverflowed = nullptr) {
821 bool Dummy;
822 bool &Overflowed = ResultOverflowed ? *ResultOverflowed : Dummy;
823
824 // Hacker's Delight, p. 30 has a different algorithm, but we don't use that
825 // because it fails for uint16_t (where multiplication can have undefined
826 // behavior due to promotion to int), and requires a division in addition
827 // to the multiplication.
828
829 Overflowed = false;
830
831 // Log2(Z) would be either Log2Z or Log2Z + 1.
832 // Special case: if X or Y is 0, Log2_64 gives -1, and Log2Z
833 // will necessarily be less than Log2Max as desired.
834 int Log2Z = Log2_64(X) + Log2_64(Y);
835 const T Max = std::numeric_limits<T>::max();
836 int Log2Max = Log2_64(Max);
837 if (Log2Z < Log2Max) {
838 return X * Y;
839 }
840 if (Log2Z > Log2Max) {
841 Overflowed = true;
842 return Max;
843 }
844
845 // We're going to use the top bit, and maybe overflow one
846 // bit past it. Multiply all but the bottom bit then add
847 // that on at the end.
848 T Z = (X >> 1) * Y;
849 if (Z & ~(Max >> 1)) {
850 Overflowed = true;
851 return Max;
852 }
853 Z <<= 1;
854 if (X & 1)
855 return SaturatingAdd(Z, Y, ResultOverflowed);
856
857 return Z;
858}
859
860/// Multiply two unsigned integers, X and Y, and add the unsigned integer, A to
861/// the product. Clamp the result to the maximum representable value of T on
862/// overflow. ResultOverflowed indicates if the result is larger than the
863/// maximum representable value of type T.
864template <typename T>
865std::enable_if_t<std::is_unsigned<T>::value, T>
866SaturatingMultiplyAdd(T X, T Y, T A, bool *ResultOverflowed = nullptr) {
867 bool Dummy;
868 bool &Overflowed = ResultOverflowed ? *ResultOverflowed : Dummy;
869
870 T Product = SaturatingMultiply(X, Y, &Overflowed);
871 if (Overflowed)
872 return Product;
873
874 return SaturatingAdd(A, Product, &Overflowed);
875}
876
877/// Use this rather than HUGE_VALF; the latter causes warnings on MSVC.
878extern const float huge_valf;
879
880
881/// Add two signed integers, computing the two's complement truncated result,
882/// returning true if overflow occured.
883template <typename T>
884std::enable_if_t<std::is_signed<T>::value, T> AddOverflow(T X, T Y, T &Result) {
885#if __has_builtin(__builtin_add_overflow)1
886 return __builtin_add_overflow(X, Y, &Result);
887#else
888 // Perform the unsigned addition.
889 using U = std::make_unsigned_t<T>;
890 const U UX = static_cast<U>(X);
891 const U UY = static_cast<U>(Y);
892 const U UResult = UX + UY;
893
894 // Convert to signed.
895 Result = static_cast<T>(UResult);
896
897 // Adding two positive numbers should result in a positive number.
898 if (X > 0 && Y > 0)
899 return Result <= 0;
900 // Adding two negatives should result in a negative number.
901 if (X < 0 && Y < 0)
902 return Result >= 0;
903 return false;
904#endif
905}
906
907/// Subtract two signed integers, computing the two's complement truncated
908/// result, returning true if an overflow ocurred.
909template <typename T>
910std::enable_if_t<std::is_signed<T>::value, T> SubOverflow(T X, T Y, T &Result) {
911#if __has_builtin(__builtin_sub_overflow)1
912 return __builtin_sub_overflow(X, Y, &Result);
913#else
914 // Perform the unsigned addition.
915 using U = std::make_unsigned_t<T>;
916 const U UX = static_cast<U>(X);
917 const U UY = static_cast<U>(Y);
918 const U UResult = UX - UY;
919
920 // Convert to signed.
921 Result = static_cast<T>(UResult);
922
923 // Subtracting a positive number from a negative results in a negative number.
924 if (X <= 0 && Y > 0)
925 return Result >= 0;
926 // Subtracting a negative number from a positive results in a positive number.
927 if (X >= 0 && Y < 0)
928 return Result <= 0;
929 return false;
930#endif
931}
932
933/// Multiply two signed integers, computing the two's complement truncated
934/// result, returning true if an overflow ocurred.
935template <typename T>
936std::enable_if_t<std::is_signed<T>::value, T> MulOverflow(T X, T Y, T &Result) {
937 // Perform the unsigned multiplication on absolute values.
938 using U = std::make_unsigned_t<T>;
939 const U UX = X < 0 ? (0 - static_cast<U>(X)) : static_cast<U>(X);
940 const U UY = Y < 0 ? (0 - static_cast<U>(Y)) : static_cast<U>(Y);
941 const U UResult = UX * UY;
942
943 // Convert to signed.
944 const bool IsNegative = (X < 0) ^ (Y < 0);
945 Result = IsNegative ? (0 - UResult) : UResult;
946
947 // If any of the args was 0, result is 0 and no overflow occurs.
948 if (UX == 0 || UY == 0)
949 return false;
950
951 // UX and UY are in [1, 2^n], where n is the number of digits.
952 // Check how the max allowed absolute value (2^n for negative, 2^(n-1) for
953 // positive) divided by an argument compares to the other.
954 if (IsNegative)
955 return UX > (static_cast<U>(std::numeric_limits<T>::max()) + U(1)) / UY;
956 else
957 return UX > (static_cast<U>(std::numeric_limits<T>::max())) / UY;
958}
959
960} // End llvm namespace
961
962#endif