Bug Summary

File:llvm/lib/Transforms/IPO/GlobalOpt.cpp
Warning:line 1885, column 27
Called C++ object pointer is null

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -cc1 -cc1 -triple x86_64-pc-linux-gnu -analyze -disable-free -disable-llvm-verifier -discard-value-names -main-file-name GlobalOpt.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 -analyzer-config-compatibility-mode=true -mrelocation-model pic -pic-level 2 -mframe-pointer=none -fmath-errno -fno-rounding-math -mconstructor-aliases -munwind-tables -target-cpu x86-64 -tune-cpu generic -debugger-tuning=gdb -ffunction-sections -fdata-sections -fcoverage-compilation-dir=/build/llvm-toolchain-snapshot-14~++20210903100615+fd66b44ec19e/build-llvm/lib/Transforms/IPO -resource-dir /usr/lib/llvm-14/lib/clang/14.0.0 -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-14~++20210903100615+fd66b44ec19e/build-llvm/lib/Transforms/IPO -I /build/llvm-toolchain-snapshot-14~++20210903100615+fd66b44ec19e/llvm/lib/Transforms/IPO -I /build/llvm-toolchain-snapshot-14~++20210903100615+fd66b44ec19e/build-llvm/include -I /build/llvm-toolchain-snapshot-14~++20210903100615+fd66b44ec19e/llvm/include -D NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/c++/10 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/x86_64-linux-gnu/c++/10 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/c++/10/backward -internal-isystem /usr/lib/llvm-14/lib/clang/14.0.0/include -internal-isystem /usr/local/include -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../x86_64-linux-gnu/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -O2 -Wno-unused-parameter -Wwrite-strings -Wno-missing-field-initializers -Wno-long-long -Wno-maybe-uninitialized -Wno-class-memaccess -Wno-redundant-move -Wno-pessimizing-move -Wno-noexcept-type -Wno-comment -std=c++14 -fdeprecated-macro -fdebug-compilation-dir=/build/llvm-toolchain-snapshot-14~++20210903100615+fd66b44ec19e/build-llvm/lib/Transforms/IPO -fdebug-prefix-map=/build/llvm-toolchain-snapshot-14~++20210903100615+fd66b44ec19e=. -ferror-limit 19 -fvisibility-inlines-hidden -stack-protector 2 -fgnuc-version=4.2.1 -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -faddrsig -D__GCC_HAVE_DWARF2_CFI_ASM=1 -o /tmp/scan-build-2021-09-04-040900-46481-1 -x c++ /build/llvm-toolchain-snapshot-14~++20210903100615+fd66b44ec19e/llvm/lib/Transforms/IPO/GlobalOpt.cpp
1//===- GlobalOpt.cpp - Optimize Global Variables --------------------------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This pass transforms simple global variables that never have their address
10// taken. If obviously true, it marks read/write globals as constant, deletes
11// variables only stored to, etc.
12//
13//===----------------------------------------------------------------------===//
14
15#include "llvm/Transforms/IPO/GlobalOpt.h"
16#include "llvm/ADT/DenseMap.h"
17#include "llvm/ADT/STLExtras.h"
18#include "llvm/ADT/SmallPtrSet.h"
19#include "llvm/ADT/SmallVector.h"
20#include "llvm/ADT/Statistic.h"
21#include "llvm/ADT/Twine.h"
22#include "llvm/ADT/iterator_range.h"
23#include "llvm/Analysis/BlockFrequencyInfo.h"
24#include "llvm/Analysis/ConstantFolding.h"
25#include "llvm/Analysis/MemoryBuiltins.h"
26#include "llvm/Analysis/TargetLibraryInfo.h"
27#include "llvm/Analysis/TargetTransformInfo.h"
28#include "llvm/BinaryFormat/Dwarf.h"
29#include "llvm/IR/Attributes.h"
30#include "llvm/IR/BasicBlock.h"
31#include "llvm/IR/CallingConv.h"
32#include "llvm/IR/Constant.h"
33#include "llvm/IR/Constants.h"
34#include "llvm/IR/DataLayout.h"
35#include "llvm/IR/DebugInfoMetadata.h"
36#include "llvm/IR/DerivedTypes.h"
37#include "llvm/IR/Dominators.h"
38#include "llvm/IR/Function.h"
39#include "llvm/IR/GetElementPtrTypeIterator.h"
40#include "llvm/IR/GlobalAlias.h"
41#include "llvm/IR/GlobalValue.h"
42#include "llvm/IR/GlobalVariable.h"
43#include "llvm/IR/IRBuilder.h"
44#include "llvm/IR/InstrTypes.h"
45#include "llvm/IR/Instruction.h"
46#include "llvm/IR/Instructions.h"
47#include "llvm/IR/IntrinsicInst.h"
48#include "llvm/IR/Module.h"
49#include "llvm/IR/Operator.h"
50#include "llvm/IR/Type.h"
51#include "llvm/IR/Use.h"
52#include "llvm/IR/User.h"
53#include "llvm/IR/Value.h"
54#include "llvm/IR/ValueHandle.h"
55#include "llvm/InitializePasses.h"
56#include "llvm/Pass.h"
57#include "llvm/Support/AtomicOrdering.h"
58#include "llvm/Support/Casting.h"
59#include "llvm/Support/CommandLine.h"
60#include "llvm/Support/Debug.h"
61#include "llvm/Support/ErrorHandling.h"
62#include "llvm/Support/MathExtras.h"
63#include "llvm/Support/raw_ostream.h"
64#include "llvm/Transforms/IPO.h"
65#include "llvm/Transforms/Utils/CtorUtils.h"
66#include "llvm/Transforms/Utils/Evaluator.h"
67#include "llvm/Transforms/Utils/GlobalStatus.h"
68#include "llvm/Transforms/Utils/Local.h"
69#include <cassert>
70#include <cstdint>
71#include <utility>
72#include <vector>
73
74using namespace llvm;
75
76#define DEBUG_TYPE"globalopt" "globalopt"
77
78STATISTIC(NumMarked , "Number of globals marked constant")static llvm::Statistic NumMarked = {"globalopt", "NumMarked",
"Number of globals marked constant"}
;
79STATISTIC(NumUnnamed , "Number of globals marked unnamed_addr")static llvm::Statistic NumUnnamed = {"globalopt", "NumUnnamed"
, "Number of globals marked unnamed_addr"}
;
80STATISTIC(NumSRA , "Number of aggregate globals broken into scalars")static llvm::Statistic NumSRA = {"globalopt", "NumSRA", "Number of aggregate globals broken into scalars"
}
;
81STATISTIC(NumSubstitute,"Number of globals with initializers stored into them")static llvm::Statistic NumSubstitute = {"globalopt", "NumSubstitute"
, "Number of globals with initializers stored into them"}
;
82STATISTIC(NumDeleted , "Number of globals deleted")static llvm::Statistic NumDeleted = {"globalopt", "NumDeleted"
, "Number of globals deleted"}
;
83STATISTIC(NumGlobUses , "Number of global uses devirtualized")static llvm::Statistic NumGlobUses = {"globalopt", "NumGlobUses"
, "Number of global uses devirtualized"}
;
84STATISTIC(NumLocalized , "Number of globals localized")static llvm::Statistic NumLocalized = {"globalopt", "NumLocalized"
, "Number of globals localized"}
;
85STATISTIC(NumShrunkToBool , "Number of global vars shrunk to booleans")static llvm::Statistic NumShrunkToBool = {"globalopt", "NumShrunkToBool"
, "Number of global vars shrunk to booleans"}
;
86STATISTIC(NumFastCallFns , "Number of functions converted to fastcc")static llvm::Statistic NumFastCallFns = {"globalopt", "NumFastCallFns"
, "Number of functions converted to fastcc"}
;
87STATISTIC(NumCtorsEvaluated, "Number of static ctors evaluated")static llvm::Statistic NumCtorsEvaluated = {"globalopt", "NumCtorsEvaluated"
, "Number of static ctors evaluated"}
;
88STATISTIC(NumNestRemoved , "Number of nest attributes removed")static llvm::Statistic NumNestRemoved = {"globalopt", "NumNestRemoved"
, "Number of nest attributes removed"}
;
89STATISTIC(NumAliasesResolved, "Number of global aliases resolved")static llvm::Statistic NumAliasesResolved = {"globalopt", "NumAliasesResolved"
, "Number of global aliases resolved"}
;
90STATISTIC(NumAliasesRemoved, "Number of global aliases eliminated")static llvm::Statistic NumAliasesRemoved = {"globalopt", "NumAliasesRemoved"
, "Number of global aliases eliminated"}
;
91STATISTIC(NumCXXDtorsRemoved, "Number of global C++ destructors removed")static llvm::Statistic NumCXXDtorsRemoved = {"globalopt", "NumCXXDtorsRemoved"
, "Number of global C++ destructors removed"}
;
92STATISTIC(NumInternalFunc, "Number of internal functions")static llvm::Statistic NumInternalFunc = {"globalopt", "NumInternalFunc"
, "Number of internal functions"}
;
93STATISTIC(NumColdCC, "Number of functions marked coldcc")static llvm::Statistic NumColdCC = {"globalopt", "NumColdCC",
"Number of functions marked coldcc"}
;
94
95static cl::opt<bool>
96 EnableColdCCStressTest("enable-coldcc-stress-test",
97 cl::desc("Enable stress test of coldcc by adding "
98 "calling conv to all internal functions."),
99 cl::init(false), cl::Hidden);
100
101static cl::opt<int> ColdCCRelFreq(
102 "coldcc-rel-freq", cl::Hidden, cl::init(2), cl::ZeroOrMore,
103 cl::desc(
104 "Maximum block frequency, expressed as a percentage of caller's "
105 "entry frequency, for a call site to be considered cold for enabling"
106 "coldcc"));
107
108/// Is this global variable possibly used by a leak checker as a root? If so,
109/// we might not really want to eliminate the stores to it.
110static bool isLeakCheckerRoot(GlobalVariable *GV) {
111 // A global variable is a root if it is a pointer, or could plausibly contain
112 // a pointer. There are two challenges; one is that we could have a struct
113 // the has an inner member which is a pointer. We recurse through the type to
114 // detect these (up to a point). The other is that we may actually be a union
115 // of a pointer and another type, and so our LLVM type is an integer which
116 // gets converted into a pointer, or our type is an [i8 x #] with a pointer
117 // potentially contained here.
118
119 if (GV->hasPrivateLinkage())
120 return false;
121
122 SmallVector<Type *, 4> Types;
123 Types.push_back(GV->getValueType());
124
125 unsigned Limit = 20;
126 do {
127 Type *Ty = Types.pop_back_val();
128 switch (Ty->getTypeID()) {
129 default: break;
130 case Type::PointerTyID:
131 return true;
132 case Type::FixedVectorTyID:
133 case Type::ScalableVectorTyID:
134 if (cast<VectorType>(Ty)->getElementType()->isPointerTy())
135 return true;
136 break;
137 case Type::ArrayTyID:
138 Types.push_back(cast<ArrayType>(Ty)->getElementType());
139 break;
140 case Type::StructTyID: {
141 StructType *STy = cast<StructType>(Ty);
142 if (STy->isOpaque()) return true;
143 for (StructType::element_iterator I = STy->element_begin(),
144 E = STy->element_end(); I != E; ++I) {
145 Type *InnerTy = *I;
146 if (isa<PointerType>(InnerTy)) return true;
147 if (isa<StructType>(InnerTy) || isa<ArrayType>(InnerTy) ||
148 isa<VectorType>(InnerTy))
149 Types.push_back(InnerTy);
150 }
151 break;
152 }
153 }
154 if (--Limit == 0) return true;
155 } while (!Types.empty());
156 return false;
157}
158
159/// Given a value that is stored to a global but never read, determine whether
160/// it's safe to remove the store and the chain of computation that feeds the
161/// store.
162static bool IsSafeComputationToRemove(
163 Value *V, function_ref<TargetLibraryInfo &(Function &)> GetTLI) {
164 do {
165 if (isa<Constant>(V))
166 return true;
167 if (!V->hasOneUse())
168 return false;
169 if (isa<LoadInst>(V) || isa<InvokeInst>(V) || isa<Argument>(V) ||
170 isa<GlobalValue>(V))
171 return false;
172 if (isAllocationFn(V, GetTLI))
173 return true;
174
175 Instruction *I = cast<Instruction>(V);
176 if (I->mayHaveSideEffects())
177 return false;
178 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
179 if (!GEP->hasAllConstantIndices())
180 return false;
181 } else if (I->getNumOperands() != 1) {
182 return false;
183 }
184
185 V = I->getOperand(0);
186 } while (true);
187}
188
189/// This GV is a pointer root. Loop over all users of the global and clean up
190/// any that obviously don't assign the global a value that isn't dynamically
191/// allocated.
192static bool
193CleanupPointerRootUsers(GlobalVariable *GV,
194 function_ref<TargetLibraryInfo &(Function &)> GetTLI) {
195 // A brief explanation of leak checkers. The goal is to find bugs where
196 // pointers are forgotten, causing an accumulating growth in memory
197 // usage over time. The common strategy for leak checkers is to explicitly
198 // allow the memory pointed to by globals at exit. This is popular because it
199 // also solves another problem where the main thread of a C++ program may shut
200 // down before other threads that are still expecting to use those globals. To
201 // handle that case, we expect the program may create a singleton and never
202 // destroy it.
203
204 bool Changed = false;
205
206 // If Dead[n].first is the only use of a malloc result, we can delete its
207 // chain of computation and the store to the global in Dead[n].second.
208 SmallVector<std::pair<Instruction *, Instruction *>, 32> Dead;
209
210 // Constants can't be pointers to dynamically allocated memory.
211 for (Value::user_iterator UI = GV->user_begin(), E = GV->user_end();
212 UI != E;) {
213 User *U = *UI++;
214 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
215 Value *V = SI->getValueOperand();
216 if (isa<Constant>(V)) {
217 Changed = true;
218 SI->eraseFromParent();
219 } else if (Instruction *I = dyn_cast<Instruction>(V)) {
220 if (I->hasOneUse())
221 Dead.push_back(std::make_pair(I, SI));
222 }
223 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(U)) {
224 if (isa<Constant>(MSI->getValue())) {
225 Changed = true;
226 MSI->eraseFromParent();
227 } else if (Instruction *I = dyn_cast<Instruction>(MSI->getValue())) {
228 if (I->hasOneUse())
229 Dead.push_back(std::make_pair(I, MSI));
230 }
231 } else if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(U)) {
232 GlobalVariable *MemSrc = dyn_cast<GlobalVariable>(MTI->getSource());
233 if (MemSrc && MemSrc->isConstant()) {
234 Changed = true;
235 MTI->eraseFromParent();
236 } else if (Instruction *I = dyn_cast<Instruction>(MemSrc)) {
237 if (I->hasOneUse())
238 Dead.push_back(std::make_pair(I, MTI));
239 }
240 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
241 if (CE->use_empty()) {
242 CE->destroyConstant();
243 Changed = true;
244 }
245 } else if (Constant *C = dyn_cast<Constant>(U)) {
246 if (isSafeToDestroyConstant(C)) {
247 C->destroyConstant();
248 // This could have invalidated UI, start over from scratch.
249 Dead.clear();
250 CleanupPointerRootUsers(GV, GetTLI);
251 return true;
252 }
253 }
254 }
255
256 for (int i = 0, e = Dead.size(); i != e; ++i) {
257 if (IsSafeComputationToRemove(Dead[i].first, GetTLI)) {
258 Dead[i].second->eraseFromParent();
259 Instruction *I = Dead[i].first;
260 do {
261 if (isAllocationFn(I, GetTLI))
262 break;
263 Instruction *J = dyn_cast<Instruction>(I->getOperand(0));
264 if (!J)
265 break;
266 I->eraseFromParent();
267 I = J;
268 } while (true);
269 I->eraseFromParent();
270 Changed = true;
271 }
272 }
273
274 return Changed;
275}
276
277/// We just marked GV constant. Loop over all users of the global, cleaning up
278/// the obvious ones. This is largely just a quick scan over the use list to
279/// clean up the easy and obvious cruft. This returns true if it made a change.
280static bool CleanupConstantGlobalUsers(
281 Value *V, Constant *Init, const DataLayout &DL,
282 function_ref<TargetLibraryInfo &(Function &)> GetTLI) {
283 bool Changed = false;
284 // Note that we need to use a weak value handle for the worklist items. When
285 // we delete a constant array, we may also be holding pointer to one of its
286 // elements (or an element of one of its elements if we're dealing with an
287 // array of arrays) in the worklist.
288 SmallVector<WeakTrackingVH, 8> WorkList(V->users());
289 while (!WorkList.empty()) {
290 Value *UV = WorkList.pop_back_val();
291 if (!UV)
292 continue;
293
294 User *U = cast<User>(UV);
295
296 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
297 if (Init) {
298 if (auto *Casted =
299 ConstantFoldLoadThroughBitcast(Init, LI->getType(), DL)) {
300 // Replace the load with the initializer.
301 LI->replaceAllUsesWith(Casted);
302 LI->eraseFromParent();
303 Changed = true;
304 }
305 }
306 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
307 // Store must be unreachable or storing Init into the global.
308 SI->eraseFromParent();
309 Changed = true;
310 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U)) {
311 if (CE->getOpcode() == Instruction::GetElementPtr) {
312 Constant *SubInit = nullptr;
313 if (Init)
314 SubInit = ConstantFoldLoadThroughGEPConstantExpr(
315 Init, CE, V->getType()->getPointerElementType(), DL);
316 Changed |= CleanupConstantGlobalUsers(CE, SubInit, DL, GetTLI);
317 } else if ((CE->getOpcode() == Instruction::BitCast &&
318 CE->getType()->isPointerTy()) ||
319 CE->getOpcode() == Instruction::AddrSpaceCast) {
320 // Pointer cast, delete any stores and memsets to the global.
321 Changed |= CleanupConstantGlobalUsers(CE, nullptr, DL, GetTLI);
322 }
323
324 if (CE->use_empty()) {
325 CE->destroyConstant();
326 Changed = true;
327 }
328 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
329 // Do not transform "gepinst (gep constexpr (GV))" here, because forming
330 // "gepconstexpr (gep constexpr (GV))" will cause the two gep's to fold
331 // and will invalidate our notion of what Init is.
332 Constant *SubInit = nullptr;
333 if (!isa<ConstantExpr>(GEP->getOperand(0))) {
334 ConstantExpr *CE = dyn_cast_or_null<ConstantExpr>(
335 ConstantFoldInstruction(GEP, DL, &GetTLI(*GEP->getFunction())));
336 if (Init && CE && CE->getOpcode() == Instruction::GetElementPtr)
337 SubInit = ConstantFoldLoadThroughGEPConstantExpr(
338 Init, CE, V->getType()->getPointerElementType(), DL);
339
340 // If the initializer is an all-null value and we have an inbounds GEP,
341 // we already know what the result of any load from that GEP is.
342 // TODO: Handle splats.
343 if (Init && isa<ConstantAggregateZero>(Init) && GEP->isInBounds())
344 SubInit = Constant::getNullValue(GEP->getResultElementType());
345 }
346 Changed |= CleanupConstantGlobalUsers(GEP, SubInit, DL, GetTLI);
347
348 if (GEP->use_empty()) {
349 GEP->eraseFromParent();
350 Changed = true;
351 }
352 } else if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(U)) { // memset/cpy/mv
353 if (MI->getRawDest() == V) {
354 MI->eraseFromParent();
355 Changed = true;
356 }
357
358 } else if (Constant *C = dyn_cast<Constant>(U)) {
359 // If we have a chain of dead constantexprs or other things dangling from
360 // us, and if they are all dead, nuke them without remorse.
361 if (isSafeToDestroyConstant(C)) {
362 C->destroyConstant();
363 CleanupConstantGlobalUsers(V, Init, DL, GetTLI);
364 return true;
365 }
366 }
367 }
368 return Changed;
369}
370
371static bool isSafeSROAElementUse(Value *V);
372
373/// Return true if the specified GEP is a safe user of a derived
374/// expression from a global that we want to SROA.
375static bool isSafeSROAGEP(User *U) {
376 // Check to see if this ConstantExpr GEP is SRA'able. In particular, we
377 // don't like < 3 operand CE's, and we don't like non-constant integer
378 // indices. This enforces that all uses are 'gep GV, 0, C, ...' for some
379 // value of C.
380 if (U->getNumOperands() < 3 || !isa<Constant>(U->getOperand(1)) ||
381 !cast<Constant>(U->getOperand(1))->isNullValue())
382 return false;
383
384 gep_type_iterator GEPI = gep_type_begin(U), E = gep_type_end(U);
385 ++GEPI; // Skip over the pointer index.
386
387 // For all other level we require that the indices are constant and inrange.
388 // In particular, consider: A[0][i]. We cannot know that the user isn't doing
389 // invalid things like allowing i to index an out-of-range subscript that
390 // accesses A[1]. This can also happen between different members of a struct
391 // in llvm IR.
392 for (; GEPI != E; ++GEPI) {
393 if (GEPI.isStruct())
394 continue;
395
396 ConstantInt *IdxVal = dyn_cast<ConstantInt>(GEPI.getOperand());
397 if (!IdxVal || (GEPI.isBoundedSequential() &&
398 IdxVal->getZExtValue() >= GEPI.getSequentialNumElements()))
399 return false;
400 }
401
402 return llvm::all_of(U->users(),
403 [](User *UU) { return isSafeSROAElementUse(UU); });
404}
405
406/// Return true if the specified instruction is a safe user of a derived
407/// expression from a global that we want to SROA.
408static bool isSafeSROAElementUse(Value *V) {
409 // We might have a dead and dangling constant hanging off of here.
410 if (Constant *C = dyn_cast<Constant>(V))
411 return isSafeToDestroyConstant(C);
412
413 Instruction *I = dyn_cast<Instruction>(V);
414 if (!I) return false;
415
416 // Loads are ok.
417 if (isa<LoadInst>(I)) return true;
418
419 // Stores *to* the pointer are ok.
420 if (StoreInst *SI = dyn_cast<StoreInst>(I))
421 return SI->getOperand(0) != V;
422
423 // Otherwise, it must be a GEP. Check it and its users are safe to SRA.
424 return isa<GetElementPtrInst>(I) && isSafeSROAGEP(I);
425}
426
427/// Look at all uses of the global and decide whether it is safe for us to
428/// perform this transformation.
429static bool GlobalUsersSafeToSRA(GlobalValue *GV) {
430 for (User *U : GV->users()) {
431 // The user of the global must be a GEP Inst or a ConstantExpr GEP.
432 if (!isa<GetElementPtrInst>(U) &&
433 (!isa<ConstantExpr>(U) ||
434 cast<ConstantExpr>(U)->getOpcode() != Instruction::GetElementPtr))
435 return false;
436
437 // Check the gep and it's users are safe to SRA
438 if (!isSafeSROAGEP(U))
439 return false;
440 }
441
442 return true;
443}
444
445static bool IsSRASequential(Type *T) {
446 return isa<ArrayType>(T) || isa<VectorType>(T);
447}
448static uint64_t GetSRASequentialNumElements(Type *T) {
449 if (ArrayType *AT = dyn_cast<ArrayType>(T))
450 return AT->getNumElements();
451 return cast<FixedVectorType>(T)->getNumElements();
452}
453static Type *GetSRASequentialElementType(Type *T) {
454 if (ArrayType *AT = dyn_cast<ArrayType>(T))
455 return AT->getElementType();
456 return cast<VectorType>(T)->getElementType();
457}
458static bool CanDoGlobalSRA(GlobalVariable *GV) {
459 Constant *Init = GV->getInitializer();
460
461 if (isa<StructType>(Init->getType())) {
462 // nothing to check
463 } else if (IsSRASequential(Init->getType())) {
464 if (GetSRASequentialNumElements(Init->getType()) > 16 &&
465 GV->hasNUsesOrMore(16))
466 return false; // It's not worth it.
467 } else
468 return false;
469
470 return GlobalUsersSafeToSRA(GV);
471}
472
473/// Copy over the debug info for a variable to its SRA replacements.
474static void transferSRADebugInfo(GlobalVariable *GV, GlobalVariable *NGV,
475 uint64_t FragmentOffsetInBits,
476 uint64_t FragmentSizeInBits,
477 uint64_t VarSize) {
478 SmallVector<DIGlobalVariableExpression *, 1> GVs;
479 GV->getDebugInfo(GVs);
480 for (auto *GVE : GVs) {
481 DIVariable *Var = GVE->getVariable();
482 DIExpression *Expr = GVE->getExpression();
483 // If the FragmentSize is smaller than the variable,
484 // emit a fragment expression.
485 if (FragmentSizeInBits < VarSize) {
486 if (auto E = DIExpression::createFragmentExpression(
487 Expr, FragmentOffsetInBits, FragmentSizeInBits))
488 Expr = *E;
489 else
490 return;
491 }
492 auto *NGVE = DIGlobalVariableExpression::get(GVE->getContext(), Var, Expr);
493 NGV->addDebugInfo(NGVE);
494 }
495}
496
497/// Perform scalar replacement of aggregates on the specified global variable.
498/// This opens the door for other optimizations by exposing the behavior of the
499/// program in a more fine-grained way. We have determined that this
500/// transformation is safe already. We return the first global variable we
501/// insert so that the caller can reprocess it.
502static GlobalVariable *SRAGlobal(GlobalVariable *GV, const DataLayout &DL) {
503 // Make sure this global only has simple uses that we can SRA.
504 if (!CanDoGlobalSRA(GV))
505 return nullptr;
506
507 assert(GV->hasLocalLinkage())(static_cast<void> (0));
508 Constant *Init = GV->getInitializer();
509 Type *Ty = Init->getType();
510 uint64_t VarSize = DL.getTypeSizeInBits(Ty);
511
512 std::map<unsigned, GlobalVariable *> NewGlobals;
513
514 // Get the alignment of the global, either explicit or target-specific.
515 Align StartAlignment =
516 DL.getValueOrABITypeAlignment(GV->getAlign(), GV->getType());
517
518 // Loop over all users and create replacement variables for used aggregate
519 // elements.
520 for (User *GEP : GV->users()) {
521 assert(((isa<ConstantExpr>(GEP) && cast<ConstantExpr>(GEP)->getOpcode() ==(static_cast<void> (0))
522 Instruction::GetElementPtr) ||(static_cast<void> (0))
523 isa<GetElementPtrInst>(GEP)) &&(static_cast<void> (0))
524 "NonGEP CE's are not SRAable!")(static_cast<void> (0));
525
526 // Ignore the 1th operand, which has to be zero or else the program is quite
527 // broken (undefined). Get the 2nd operand, which is the structure or array
528 // index.
529 unsigned ElementIdx = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue();
530 if (NewGlobals.count(ElementIdx) == 1)
531 continue; // we`ve already created replacement variable
532 assert(NewGlobals.count(ElementIdx) == 0)(static_cast<void> (0));
533
534 Type *ElTy = nullptr;
535 if (StructType *STy = dyn_cast<StructType>(Ty))
536 ElTy = STy->getElementType(ElementIdx);
537 else
538 ElTy = GetSRASequentialElementType(Ty);
539 assert(ElTy)(static_cast<void> (0));
540
541 Constant *In = Init->getAggregateElement(ElementIdx);
542 assert(In && "Couldn't get element of initializer?")(static_cast<void> (0));
543
544 GlobalVariable *NGV = new GlobalVariable(
545 ElTy, false, GlobalVariable::InternalLinkage, In,
546 GV->getName() + "." + Twine(ElementIdx), GV->getThreadLocalMode(),
547 GV->getType()->getAddressSpace());
548 NGV->setExternallyInitialized(GV->isExternallyInitialized());
549 NGV->copyAttributesFrom(GV);
550 NewGlobals.insert(std::make_pair(ElementIdx, NGV));
551
552 if (StructType *STy = dyn_cast<StructType>(Ty)) {
553 const StructLayout &Layout = *DL.getStructLayout(STy);
554
555 // Calculate the known alignment of the field. If the original aggregate
556 // had 256 byte alignment for example, something might depend on that:
557 // propagate info to each field.
558 uint64_t FieldOffset = Layout.getElementOffset(ElementIdx);
559 Align NewAlign = commonAlignment(StartAlignment, FieldOffset);
560 if (NewAlign > DL.getABITypeAlign(STy->getElementType(ElementIdx)))
561 NGV->setAlignment(NewAlign);
562
563 // Copy over the debug info for the variable.
564 uint64_t Size = DL.getTypeAllocSizeInBits(NGV->getValueType());
565 uint64_t FragmentOffsetInBits = Layout.getElementOffsetInBits(ElementIdx);
566 transferSRADebugInfo(GV, NGV, FragmentOffsetInBits, Size, VarSize);
567 } else {
568 uint64_t EltSize = DL.getTypeAllocSize(ElTy);
569 Align EltAlign = DL.getABITypeAlign(ElTy);
570 uint64_t FragmentSizeInBits = DL.getTypeAllocSizeInBits(ElTy);
571
572 // Calculate the known alignment of the field. If the original aggregate
573 // had 256 byte alignment for example, something might depend on that:
574 // propagate info to each field.
575 Align NewAlign = commonAlignment(StartAlignment, EltSize * ElementIdx);
576 if (NewAlign > EltAlign)
577 NGV->setAlignment(NewAlign);
578 transferSRADebugInfo(GV, NGV, FragmentSizeInBits * ElementIdx,
579 FragmentSizeInBits, VarSize);
580 }
581 }
582
583 if (NewGlobals.empty())
584 return nullptr;
585
586 Module::GlobalListType &Globals = GV->getParent()->getGlobalList();
587 for (auto NewGlobalVar : NewGlobals)
588 Globals.push_back(NewGlobalVar.second);
589
590 LLVM_DEBUG(dbgs() << "PERFORMING GLOBAL SRA ON: " << *GV << "\n")do { } while (false);
591
592 Constant *NullInt =Constant::getNullValue(Type::getInt32Ty(GV->getContext()));
593
594 // Loop over all of the uses of the global, replacing the constantexpr geps,
595 // with smaller constantexpr geps or direct references.
596 while (!GV->use_empty()) {
597 User *GEP = GV->user_back();
598 assert(((isa<ConstantExpr>(GEP) &&(static_cast<void> (0))
599 cast<ConstantExpr>(GEP)->getOpcode()==Instruction::GetElementPtr)||(static_cast<void> (0))
600 isa<GetElementPtrInst>(GEP)) && "NonGEP CE's are not SRAable!")(static_cast<void> (0));
601
602 // Ignore the 1th operand, which has to be zero or else the program is quite
603 // broken (undefined). Get the 2nd operand, which is the structure or array
604 // index.
605 unsigned ElementIdx = cast<ConstantInt>(GEP->getOperand(2))->getZExtValue();
606 assert(NewGlobals.count(ElementIdx) == 1)(static_cast<void> (0));
607
608 Value *NewPtr = NewGlobals[ElementIdx];
609 Type *NewTy = NewGlobals[ElementIdx]->getValueType();
610
611 // Form a shorter GEP if needed.
612 if (GEP->getNumOperands() > 3) {
613 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(GEP)) {
614 SmallVector<Constant*, 8> Idxs;
615 Idxs.push_back(NullInt);
616 for (unsigned i = 3, e = CE->getNumOperands(); i != e; ++i)
617 Idxs.push_back(CE->getOperand(i));
618 NewPtr =
619 ConstantExpr::getGetElementPtr(NewTy, cast<Constant>(NewPtr), Idxs);
620 } else {
621 GetElementPtrInst *GEPI = cast<GetElementPtrInst>(GEP);
622 SmallVector<Value*, 8> Idxs;
623 Idxs.push_back(NullInt);
624 for (unsigned i = 3, e = GEPI->getNumOperands(); i != e; ++i)
625 Idxs.push_back(GEPI->getOperand(i));
626 NewPtr = GetElementPtrInst::Create(
627 NewTy, NewPtr, Idxs, GEPI->getName() + "." + Twine(ElementIdx),
628 GEPI);
629 }
630 }
631 GEP->replaceAllUsesWith(NewPtr);
632
633 // We changed the pointer of any memory access user. Recalculate alignments.
634 for (User *U : NewPtr->users()) {
635 if (auto *Load = dyn_cast<LoadInst>(U)) {
636 Align PrefAlign = DL.getPrefTypeAlign(Load->getType());
637 Align NewAlign = getOrEnforceKnownAlignment(Load->getPointerOperand(),
638 PrefAlign, DL, Load);
639 Load->setAlignment(NewAlign);
640 }
641 if (auto *Store = dyn_cast<StoreInst>(U)) {
642 Align PrefAlign =
643 DL.getPrefTypeAlign(Store->getValueOperand()->getType());
644 Align NewAlign = getOrEnforceKnownAlignment(Store->getPointerOperand(),
645 PrefAlign, DL, Store);
646 Store->setAlignment(NewAlign);
647 }
648 }
649
650 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(GEP))
651 GEPI->eraseFromParent();
652 else
653 cast<ConstantExpr>(GEP)->destroyConstant();
654 }
655
656 // Delete the old global, now that it is dead.
657 Globals.erase(GV);
658 ++NumSRA;
659
660 assert(NewGlobals.size() > 0)(static_cast<void> (0));
661 return NewGlobals.begin()->second;
662}
663
664/// Return true if all users of the specified value will trap if the value is
665/// dynamically null. PHIs keeps track of any phi nodes we've seen to avoid
666/// reprocessing them.
667static bool AllUsesOfValueWillTrapIfNull(const Value *V,
668 SmallPtrSetImpl<const PHINode*> &PHIs) {
669 for (const User *U : V->users()) {
670 if (const Instruction *I = dyn_cast<Instruction>(U)) {
671 // If null pointer is considered valid, then all uses are non-trapping.
672 // Non address-space 0 globals have already been pruned by the caller.
673 if (NullPointerIsDefined(I->getFunction()))
674 return false;
675 }
676 if (isa<LoadInst>(U)) {
677 // Will trap.
678 } else if (const StoreInst *SI = dyn_cast<StoreInst>(U)) {
679 if (SI->getOperand(0) == V) {
680 //cerr << "NONTRAPPING USE: " << *U;
681 return false; // Storing the value.
682 }
683 } else if (const CallInst *CI = dyn_cast<CallInst>(U)) {
684 if (CI->getCalledOperand() != V) {
685 //cerr << "NONTRAPPING USE: " << *U;
686 return false; // Not calling the ptr
687 }
688 } else if (const InvokeInst *II = dyn_cast<InvokeInst>(U)) {
689 if (II->getCalledOperand() != V) {
690 //cerr << "NONTRAPPING USE: " << *U;
691 return false; // Not calling the ptr
692 }
693 } else if (const BitCastInst *CI = dyn_cast<BitCastInst>(U)) {
694 if (!AllUsesOfValueWillTrapIfNull(CI, PHIs)) return false;
695 } else if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) {
696 if (!AllUsesOfValueWillTrapIfNull(GEPI, PHIs)) return false;
697 } else if (const PHINode *PN = dyn_cast<PHINode>(U)) {
698 // If we've already seen this phi node, ignore it, it has already been
699 // checked.
700 if (PHIs.insert(PN).second && !AllUsesOfValueWillTrapIfNull(PN, PHIs))
701 return false;
702 } else if (isa<ICmpInst>(U) &&
703 !ICmpInst::isSigned(cast<ICmpInst>(U)->getPredicate()) &&
704 isa<LoadInst>(U->getOperand(0)) &&
705 isa<ConstantPointerNull>(U->getOperand(1))) {
706 assert(isa<GlobalValue>(cast<LoadInst>(U->getOperand(0))(static_cast<void> (0))
707 ->getPointerOperand()(static_cast<void> (0))
708 ->stripPointerCasts()) &&(static_cast<void> (0))
709 "Should be GlobalVariable")(static_cast<void> (0));
710 // This and only this kind of non-signed ICmpInst is to be replaced with
711 // the comparing of the value of the created global init bool later in
712 // optimizeGlobalAddressOfMalloc for the global variable.
713 } else {
714 //cerr << "NONTRAPPING USE: " << *U;
715 return false;
716 }
717 }
718 return true;
719}
720
721/// Return true if all uses of any loads from GV will trap if the loaded value
722/// is null. Note that this also permits comparisons of the loaded value
723/// against null, as a special case.
724static bool allUsesOfLoadedValueWillTrapIfNull(const GlobalVariable *GV) {
725 SmallVector<const Value *, 4> Worklist;
726 Worklist.push_back(GV);
727 while (!Worklist.empty()) {
728 const Value *P = Worklist.pop_back_val();
729 for (auto *U : P->users()) {
730 if (auto *LI = dyn_cast<LoadInst>(U)) {
731 SmallPtrSet<const PHINode *, 8> PHIs;
732 if (!AllUsesOfValueWillTrapIfNull(LI, PHIs))
733 return false;
734 } else if (auto *SI = dyn_cast<StoreInst>(U)) {
735 // Ignore stores to the global.
736 if (SI->getPointerOperand() != P)
737 return false;
738 } else if (auto *CE = dyn_cast<ConstantExpr>(U)) {
739 if (CE->stripPointerCasts() != GV)
740 return false;
741 // Check further the ConstantExpr.
742 Worklist.push_back(CE);
743 } else {
744 // We don't know or understand this user, bail out.
745 return false;
746 }
747 }
748 }
749
750 return true;
751}
752
753/// Get all the loads/store uses for global variable \p GV.
754static void allUsesOfLoadAndStores(GlobalVariable *GV,
755 SmallVector<Value *, 4> &Uses) {
756 SmallVector<Value *, 4> Worklist;
757 Worklist.push_back(GV);
758 while (!Worklist.empty()) {
759 auto *P = Worklist.pop_back_val();
760 for (auto *U : P->users()) {
761 if (auto *CE = dyn_cast<ConstantExpr>(U)) {
762 Worklist.push_back(CE);
763 continue;
764 }
765
766 assert((isa<LoadInst>(U) || isa<StoreInst>(U)) &&(static_cast<void> (0))
767 "Expect only load or store instructions")(static_cast<void> (0));
768 Uses.push_back(U);
769 }
770 }
771}
772
773static bool OptimizeAwayTrappingUsesOfValue(Value *V, Constant *NewV) {
774 bool Changed = false;
775 for (auto UI = V->user_begin(), E = V->user_end(); UI != E; ) {
776 Instruction *I = cast<Instruction>(*UI++);
777 // Uses are non-trapping if null pointer is considered valid.
778 // Non address-space 0 globals are already pruned by the caller.
779 if (NullPointerIsDefined(I->getFunction()))
780 return false;
781 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
782 LI->setOperand(0, NewV);
783 Changed = true;
784 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
785 if (SI->getOperand(1) == V) {
786 SI->setOperand(1, NewV);
787 Changed = true;
788 }
789 } else if (isa<CallInst>(I) || isa<InvokeInst>(I)) {
790 CallBase *CB = cast<CallBase>(I);
791 if (CB->getCalledOperand() == V) {
792 // Calling through the pointer! Turn into a direct call, but be careful
793 // that the pointer is not also being passed as an argument.
794 CB->setCalledOperand(NewV);
795 Changed = true;
796 bool PassedAsArg = false;
797 for (unsigned i = 0, e = CB->arg_size(); i != e; ++i)
798 if (CB->getArgOperand(i) == V) {
799 PassedAsArg = true;
800 CB->setArgOperand(i, NewV);
801 }
802
803 if (PassedAsArg) {
804 // Being passed as an argument also. Be careful to not invalidate UI!
805 UI = V->user_begin();
806 }
807 }
808 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
809 Changed |= OptimizeAwayTrappingUsesOfValue(CI,
810 ConstantExpr::getCast(CI->getOpcode(),
811 NewV, CI->getType()));
812 if (CI->use_empty()) {
813 Changed = true;
814 CI->eraseFromParent();
815 }
816 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
817 // Should handle GEP here.
818 SmallVector<Constant*, 8> Idxs;
819 Idxs.reserve(GEPI->getNumOperands()-1);
820 for (User::op_iterator i = GEPI->op_begin() + 1, e = GEPI->op_end();
821 i != e; ++i)
822 if (Constant *C = dyn_cast<Constant>(*i))
823 Idxs.push_back(C);
824 else
825 break;
826 if (Idxs.size() == GEPI->getNumOperands()-1)
827 Changed |= OptimizeAwayTrappingUsesOfValue(
828 GEPI, ConstantExpr::getGetElementPtr(GEPI->getSourceElementType(),
829 NewV, Idxs));
830 if (GEPI->use_empty()) {
831 Changed = true;
832 GEPI->eraseFromParent();
833 }
834 }
835 }
836
837 return Changed;
838}
839
840/// The specified global has only one non-null value stored into it. If there
841/// are uses of the loaded value that would trap if the loaded value is
842/// dynamically null, then we know that they cannot be reachable with a null
843/// optimize away the load.
844static bool OptimizeAwayTrappingUsesOfLoads(
845 GlobalVariable *GV, Constant *LV, const DataLayout &DL,
846 function_ref<TargetLibraryInfo &(Function &)> GetTLI) {
847 bool Changed = false;
848
849 // Keep track of whether we are able to remove all the uses of the global
850 // other than the store that defines it.
851 bool AllNonStoreUsesGone = true;
852
853 // Replace all uses of loads with uses of uses of the stored value.
854 for (Value::user_iterator GUI = GV->user_begin(), E = GV->user_end(); GUI != E;){
855 User *GlobalUser = *GUI++;
856 if (LoadInst *LI = dyn_cast<LoadInst>(GlobalUser)) {
857 Changed |= OptimizeAwayTrappingUsesOfValue(LI, LV);
858 // If we were able to delete all uses of the loads
859 if (LI->use_empty()) {
860 LI->eraseFromParent();
861 Changed = true;
862 } else {
863 AllNonStoreUsesGone = false;
864 }
865 } else if (isa<StoreInst>(GlobalUser)) {
866 // Ignore the store that stores "LV" to the global.
867 assert(GlobalUser->getOperand(1) == GV &&(static_cast<void> (0))
868 "Must be storing *to* the global")(static_cast<void> (0));
869 } else {
870 AllNonStoreUsesGone = false;
871
872 // If we get here we could have other crazy uses that are transitively
873 // loaded.
874 assert((isa<PHINode>(GlobalUser) || isa<SelectInst>(GlobalUser) ||(static_cast<void> (0))
875 isa<ConstantExpr>(GlobalUser) || isa<CmpInst>(GlobalUser) ||(static_cast<void> (0))
876 isa<BitCastInst>(GlobalUser) ||(static_cast<void> (0))
877 isa<GetElementPtrInst>(GlobalUser)) &&(static_cast<void> (0))
878 "Only expect load and stores!")(static_cast<void> (0));
879 }
880 }
881
882 if (Changed) {
883 LLVM_DEBUG(dbgs() << "OPTIMIZED LOADS FROM STORED ONCE POINTER: " << *GVdo { } while (false)
884 << "\n")do { } while (false);
885 ++NumGlobUses;
886 }
887
888 // If we nuked all of the loads, then none of the stores are needed either,
889 // nor is the global.
890 if (AllNonStoreUsesGone) {
891 if (isLeakCheckerRoot(GV)) {
892 Changed |= CleanupPointerRootUsers(GV, GetTLI);
893 } else {
894 Changed = true;
895 CleanupConstantGlobalUsers(GV, nullptr, DL, GetTLI);
896 }
897 if (GV->use_empty()) {
898 LLVM_DEBUG(dbgs() << " *** GLOBAL NOW DEAD!\n")do { } while (false);
899 Changed = true;
900 GV->eraseFromParent();
901 ++NumDeleted;
902 }
903 }
904 return Changed;
905}
906
907/// Walk the use list of V, constant folding all of the instructions that are
908/// foldable.
909static void ConstantPropUsersOf(Value *V, const DataLayout &DL,
910 TargetLibraryInfo *TLI) {
911 for (Value::user_iterator UI = V->user_begin(), E = V->user_end(); UI != E; )
912 if (Instruction *I = dyn_cast<Instruction>(*UI++))
913 if (Constant *NewC = ConstantFoldInstruction(I, DL, TLI)) {
914 I->replaceAllUsesWith(NewC);
915
916 // Advance UI to the next non-I use to avoid invalidating it!
917 // Instructions could multiply use V.
918 while (UI != E && *UI == I)
919 ++UI;
920 if (isInstructionTriviallyDead(I, TLI))
921 I->eraseFromParent();
922 }
923}
924
925/// This function takes the specified global variable, and transforms the
926/// program as if it always contained the result of the specified malloc.
927/// Because it is always the result of the specified malloc, there is no reason
928/// to actually DO the malloc. Instead, turn the malloc into a global, and any
929/// loads of GV as uses of the new global.
930static GlobalVariable *
931OptimizeGlobalAddressOfMalloc(GlobalVariable *GV, CallInst *CI, Type *AllocTy,
932 ConstantInt *NElements, const DataLayout &DL,
933 TargetLibraryInfo *TLI) {
934 LLVM_DEBUG(errs() << "PROMOTING GLOBAL: " << *GV << " CALL = " << *CIdo { } while (false)
935 << '\n')do { } while (false);
936
937 Type *GlobalType;
938 if (NElements->getZExtValue() == 1)
939 GlobalType = AllocTy;
940 else
941 // If we have an array allocation, the global variable is of an array.
942 GlobalType = ArrayType::get(AllocTy, NElements->getZExtValue());
943
944 // Create the new global variable. The contents of the malloc'd memory is
945 // undefined, so initialize with an undef value.
946 GlobalVariable *NewGV = new GlobalVariable(
947 *GV->getParent(), GlobalType, false, GlobalValue::InternalLinkage,
948 UndefValue::get(GlobalType), GV->getName() + ".body", nullptr,
949 GV->getThreadLocalMode());
950
951 // If there are bitcast users of the malloc (which is typical, usually we have
952 // a malloc + bitcast) then replace them with uses of the new global. Update
953 // other users to use the global as well.
954 BitCastInst *TheBC = nullptr;
955 while (!CI->use_empty()) {
956 Instruction *User = cast<Instruction>(CI->user_back());
957 if (BitCastInst *BCI = dyn_cast<BitCastInst>(User)) {
958 if (BCI->getType() == NewGV->getType()) {
959 BCI->replaceAllUsesWith(NewGV);
960 BCI->eraseFromParent();
961 } else {
962 BCI->setOperand(0, NewGV);
963 }
964 } else {
965 if (!TheBC)
966 TheBC = new BitCastInst(NewGV, CI->getType(), "newgv", CI);
967 User->replaceUsesOfWith(CI, TheBC);
968 }
969 }
970
971 SmallPtrSet<Constant *, 1> RepValues;
972 RepValues.insert(NewGV);
973
974 // If there is a comparison against null, we will insert a global bool to
975 // keep track of whether the global was initialized yet or not.
976 GlobalVariable *InitBool =
977 new GlobalVariable(Type::getInt1Ty(GV->getContext()), false,
978 GlobalValue::InternalLinkage,
979 ConstantInt::getFalse(GV->getContext()),
980 GV->getName()+".init", GV->getThreadLocalMode());
981 bool InitBoolUsed = false;
982
983 // Loop over all instruction uses of GV, processing them in turn.
984 SmallVector<Value *, 4> Guses;
985 allUsesOfLoadAndStores(GV, Guses);
986 for (auto *U : Guses) {
987 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
988 // The global is initialized when the store to it occurs. If the stored
989 // value is null value, the global bool is set to false, otherwise true.
990 new StoreInst(ConstantInt::getBool(
991 GV->getContext(),
992 !isa<ConstantPointerNull>(SI->getValueOperand())),
993 InitBool, false, Align(1), SI->getOrdering(),
994 SI->getSyncScopeID(), SI);
995 SI->eraseFromParent();
996 continue;
997 }
998
999 LoadInst *LI = cast<LoadInst>(U);
1000 while (!LI->use_empty()) {
1001 Use &LoadUse = *LI->use_begin();
1002 ICmpInst *ICI = dyn_cast<ICmpInst>(LoadUse.getUser());
1003 if (!ICI) {
1004 auto *CE = ConstantExpr::getBitCast(NewGV, LI->getType());
1005 RepValues.insert(CE);
1006 LoadUse.set(CE);
1007 continue;
1008 }
1009
1010 // Replace the cmp X, 0 with a use of the bool value.
1011 Value *LV = new LoadInst(InitBool->getValueType(), InitBool,
1012 InitBool->getName() + ".val", false, Align(1),
1013 LI->getOrdering(), LI->getSyncScopeID(), LI);
1014 InitBoolUsed = true;
1015 switch (ICI->getPredicate()) {
1016 default: llvm_unreachable("Unknown ICmp Predicate!")__builtin_unreachable();
1017 case ICmpInst::ICMP_ULT: // X < null -> always false
1018 LV = ConstantInt::getFalse(GV->getContext());
1019 break;
1020 case ICmpInst::ICMP_UGE: // X >= null -> always true
1021 LV = ConstantInt::getTrue(GV->getContext());
1022 break;
1023 case ICmpInst::ICMP_ULE:
1024 case ICmpInst::ICMP_EQ:
1025 LV = BinaryOperator::CreateNot(LV, "notinit", ICI);
1026 break;
1027 case ICmpInst::ICMP_NE:
1028 case ICmpInst::ICMP_UGT:
1029 break; // no change.
1030 }
1031 ICI->replaceAllUsesWith(LV);
1032 ICI->eraseFromParent();
1033 }
1034 LI->eraseFromParent();
1035 }
1036
1037 // If the initialization boolean was used, insert it, otherwise delete it.
1038 if (!InitBoolUsed) {
1039 while (!InitBool->use_empty()) // Delete initializations
1040 cast<StoreInst>(InitBool->user_back())->eraseFromParent();
1041 delete InitBool;
1042 } else
1043 GV->getParent()->getGlobalList().insert(GV->getIterator(), InitBool);
1044
1045 // Now the GV is dead, nuke it and the malloc..
1046 GV->eraseFromParent();
1047 CI->eraseFromParent();
1048
1049 // To further other optimizations, loop over all users of NewGV and try to
1050 // constant prop them. This will promote GEP instructions with constant
1051 // indices into GEP constant-exprs, which will allow global-opt to hack on it.
1052 for (auto *CE : RepValues)
1053 ConstantPropUsersOf(CE, DL, TLI);
1054
1055 return NewGV;
1056}
1057
1058/// Scan the use-list of GV checking to make sure that there are no complex uses
1059/// of GV. We permit simple things like dereferencing the pointer, but not
1060/// storing through the address, unless it is to the specified global.
1061static bool
1062valueIsOnlyUsedLocallyOrStoredToOneGlobal(const CallInst *CI,
1063 const GlobalVariable *GV) {
1064 SmallPtrSet<const Value *, 4> Visited;
1065 SmallVector<const Value *, 4> Worklist;
1066 Worklist.push_back(CI);
1067
1068 while (!Worklist.empty()) {
1069 const Value *V = Worklist.pop_back_val();
1070 if (!Visited.insert(V).second)
1071 continue;
1072
1073 for (const Use &VUse : V->uses()) {
1074 const User *U = VUse.getUser();
1075 if (isa<LoadInst>(U) || isa<CmpInst>(U))
1076 continue; // Fine, ignore.
1077
1078 if (auto *SI = dyn_cast<StoreInst>(U)) {
1079 if (SI->getValueOperand() == V &&
1080 SI->getPointerOperand()->stripPointerCasts() != GV)
1081 return false; // Storing the pointer not into GV... bad.
1082 continue; // Otherwise, storing through it, or storing into GV... fine.
1083 }
1084
1085 if (auto *BCI = dyn_cast<BitCastInst>(U)) {
1086 Worklist.push_back(BCI);
1087 continue;
1088 }
1089
1090 if (auto *GEPI = dyn_cast<GetElementPtrInst>(U)) {
1091 Worklist.push_back(GEPI);
1092 continue;
1093 }
1094
1095 return false;
1096 }
1097 }
1098
1099 return true;
1100}
1101
1102/// This function is called when we see a pointer global variable with a single
1103/// value stored it that is a malloc or cast of malloc.
1104static bool tryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV, CallInst *CI,
1105 Type *AllocTy,
1106 AtomicOrdering Ordering,
1107 const DataLayout &DL,
1108 TargetLibraryInfo *TLI) {
1109 // If this is a malloc of an abstract type, don't touch it.
1110 if (!AllocTy->isSized())
1111 return false;
1112
1113 // We can't optimize this global unless all uses of it are *known* to be
1114 // of the malloc value, not of the null initializer value (consider a use
1115 // that compares the global's value against zero to see if the malloc has
1116 // been reached). To do this, we check to see if all uses of the global
1117 // would trap if the global were null: this proves that they must all
1118 // happen after the malloc.
1119 if (!allUsesOfLoadedValueWillTrapIfNull(GV))
1120 return false;
1121
1122 // We can't optimize this if the malloc itself is used in a complex way,
1123 // for example, being stored into multiple globals. This allows the
1124 // malloc to be stored into the specified global, loaded, gep, icmp'd.
1125 // These are all things we could transform to using the global for.
1126 if (!valueIsOnlyUsedLocallyOrStoredToOneGlobal(CI, GV))
1127 return false;
1128
1129 // If we have a global that is only initialized with a fixed size malloc,
1130 // transform the program to use global memory instead of malloc'd memory.
1131 // This eliminates dynamic allocation, avoids an indirection accessing the
1132 // data, and exposes the resultant global to further GlobalOpt.
1133 // We cannot optimize the malloc if we cannot determine malloc array size.
1134 Value *NElems = getMallocArraySize(CI, DL, TLI, true);
1135 if (!NElems)
1136 return false;
1137
1138 if (ConstantInt *NElements = dyn_cast<ConstantInt>(NElems))
1139 // Restrict this transformation to only working on small allocations
1140 // (2048 bytes currently), as we don't want to introduce a 16M global or
1141 // something.
1142 if (NElements->getZExtValue() * DL.getTypeAllocSize(AllocTy) < 2048) {
1143 OptimizeGlobalAddressOfMalloc(GV, CI, AllocTy, NElements, DL, TLI);
1144 return true;
1145 }
1146
1147 return false;
1148}
1149
1150// Try to optimize globals based on the knowledge that only one value (besides
1151// its initializer) is ever stored to the global.
1152static bool
1153optimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal,
1154 AtomicOrdering Ordering, const DataLayout &DL,
1155 function_ref<TargetLibraryInfo &(Function &)> GetTLI) {
1156 // Ignore no-op GEPs and bitcasts.
1157 StoredOnceVal = StoredOnceVal->stripPointerCasts();
1158
1159 // If we are dealing with a pointer global that is initialized to null and
1160 // only has one (non-null) value stored into it, then we can optimize any
1161 // users of the loaded value (often calls and loads) that would trap if the
1162 // value was null.
1163 if (GV->getInitializer()->getType()->isPointerTy() &&
1164 GV->getInitializer()->isNullValue() &&
1165 StoredOnceVal->getType()->isPointerTy() &&
1166 !NullPointerIsDefined(
1167 nullptr /* F */,
1168 GV->getInitializer()->getType()->getPointerAddressSpace())) {
1169 if (Constant *SOVC = dyn_cast<Constant>(StoredOnceVal)) {
1170 if (GV->getInitializer()->getType() != SOVC->getType())
1171 SOVC = ConstantExpr::getBitCast(SOVC, GV->getInitializer()->getType());
1172
1173 // Optimize away any trapping uses of the loaded value.
1174 if (OptimizeAwayTrappingUsesOfLoads(GV, SOVC, DL, GetTLI))
1175 return true;
1176 } else if (CallInst *CI = extractMallocCall(StoredOnceVal, GetTLI)) {
1177 auto *TLI = &GetTLI(*CI->getFunction());
1178 Type *MallocType = getMallocAllocatedType(CI, TLI);
1179 if (MallocType && tryToOptimizeStoreOfMallocToGlobal(GV, CI, MallocType,
1180 Ordering, DL, TLI))
1181 return true;
1182 }
1183 }
1184
1185 return false;
1186}
1187
1188/// At this point, we have learned that the only two values ever stored into GV
1189/// are its initializer and OtherVal. See if we can shrink the global into a
1190/// boolean and select between the two values whenever it is used. This exposes
1191/// the values to other scalar optimizations.
1192static bool TryToShrinkGlobalToBoolean(GlobalVariable *GV, Constant *OtherVal) {
1193 Type *GVElType = GV->getValueType();
1194
1195 // If GVElType is already i1, it is already shrunk. If the type of the GV is
1196 // an FP value, pointer or vector, don't do this optimization because a select
1197 // between them is very expensive and unlikely to lead to later
1198 // simplification. In these cases, we typically end up with "cond ? v1 : v2"
1199 // where v1 and v2 both require constant pool loads, a big loss.
1200 if (GVElType == Type::getInt1Ty(GV->getContext()) ||
1201 GVElType->isFloatingPointTy() ||
1202 GVElType->isPointerTy() || GVElType->isVectorTy())
1203 return false;
1204
1205 // Walk the use list of the global seeing if all the uses are load or store.
1206 // If there is anything else, bail out.
1207 for (User *U : GV->users())
1208 if (!isa<LoadInst>(U) && !isa<StoreInst>(U))
1209 return false;
1210
1211 LLVM_DEBUG(dbgs() << " *** SHRINKING TO BOOL: " << *GV << "\n")do { } while (false);
1212
1213 // Create the new global, initializing it to false.
1214 GlobalVariable *NewGV = new GlobalVariable(Type::getInt1Ty(GV->getContext()),
1215 false,
1216 GlobalValue::InternalLinkage,
1217 ConstantInt::getFalse(GV->getContext()),
1218 GV->getName()+".b",
1219 GV->getThreadLocalMode(),
1220 GV->getType()->getAddressSpace());
1221 NewGV->copyAttributesFrom(GV);
1222 GV->getParent()->getGlobalList().insert(GV->getIterator(), NewGV);
1223
1224 Constant *InitVal = GV->getInitializer();
1225 assert(InitVal->getType() != Type::getInt1Ty(GV->getContext()) &&(static_cast<void> (0))
1226 "No reason to shrink to bool!")(static_cast<void> (0));
1227
1228 SmallVector<DIGlobalVariableExpression *, 1> GVs;
1229 GV->getDebugInfo(GVs);
1230
1231 // If initialized to zero and storing one into the global, we can use a cast
1232 // instead of a select to synthesize the desired value.
1233 bool IsOneZero = false;
1234 bool EmitOneOrZero = true;
1235 auto *CI = dyn_cast<ConstantInt>(OtherVal);
1236 if (CI && CI->getValue().getActiveBits() <= 64) {
1237 IsOneZero = InitVal->isNullValue() && CI->isOne();
1238
1239 auto *CIInit = dyn_cast<ConstantInt>(GV->getInitializer());
1240 if (CIInit && CIInit->getValue().getActiveBits() <= 64) {
1241 uint64_t ValInit = CIInit->getZExtValue();
1242 uint64_t ValOther = CI->getZExtValue();
1243 uint64_t ValMinus = ValOther - ValInit;
1244
1245 for(auto *GVe : GVs){
1246 DIGlobalVariable *DGV = GVe->getVariable();
1247 DIExpression *E = GVe->getExpression();
1248 const DataLayout &DL = GV->getParent()->getDataLayout();
1249 unsigned SizeInOctets =
1250 DL.getTypeAllocSizeInBits(NewGV->getValueType()) / 8;
1251
1252 // It is expected that the address of global optimized variable is on
1253 // top of the stack. After optimization, value of that variable will
1254 // be ether 0 for initial value or 1 for other value. The following
1255 // expression should return constant integer value depending on the
1256 // value at global object address:
1257 // val * (ValOther - ValInit) + ValInit:
1258 // DW_OP_deref DW_OP_constu <ValMinus>
1259 // DW_OP_mul DW_OP_constu <ValInit> DW_OP_plus DW_OP_stack_value
1260 SmallVector<uint64_t, 12> Ops = {
1261 dwarf::DW_OP_deref_size, SizeInOctets,
1262 dwarf::DW_OP_constu, ValMinus,
1263 dwarf::DW_OP_mul, dwarf::DW_OP_constu, ValInit,
1264 dwarf::DW_OP_plus};
1265 bool WithStackValue = true;
1266 E = DIExpression::prependOpcodes(E, Ops, WithStackValue);
1267 DIGlobalVariableExpression *DGVE =
1268 DIGlobalVariableExpression::get(NewGV->getContext(), DGV, E);
1269 NewGV->addDebugInfo(DGVE);
1270 }
1271 EmitOneOrZero = false;
1272 }
1273 }
1274
1275 if (EmitOneOrZero) {
1276 // FIXME: This will only emit address for debugger on which will
1277 // be written only 0 or 1.
1278 for(auto *GV : GVs)
1279 NewGV->addDebugInfo(GV);
1280 }
1281
1282 while (!GV->use_empty()) {
1283 Instruction *UI = cast<Instruction>(GV->user_back());
1284 if (StoreInst *SI = dyn_cast<StoreInst>(UI)) {
1285 // Change the store into a boolean store.
1286 bool StoringOther = SI->getOperand(0) == OtherVal;
1287 // Only do this if we weren't storing a loaded value.
1288 Value *StoreVal;
1289 if (StoringOther || SI->getOperand(0) == InitVal) {
1290 StoreVal = ConstantInt::get(Type::getInt1Ty(GV->getContext()),
1291 StoringOther);
1292 } else {
1293 // Otherwise, we are storing a previously loaded copy. To do this,
1294 // change the copy from copying the original value to just copying the
1295 // bool.
1296 Instruction *StoredVal = cast<Instruction>(SI->getOperand(0));
1297
1298 // If we've already replaced the input, StoredVal will be a cast or
1299 // select instruction. If not, it will be a load of the original
1300 // global.
1301 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
1302 assert(LI->getOperand(0) == GV && "Not a copy!")(static_cast<void> (0));
1303 // Insert a new load, to preserve the saved value.
1304 StoreVal = new LoadInst(NewGV->getValueType(), NewGV,
1305 LI->getName() + ".b", false, Align(1),
1306 LI->getOrdering(), LI->getSyncScopeID(), LI);
1307 } else {
1308 assert((isa<CastInst>(StoredVal) || isa<SelectInst>(StoredVal)) &&(static_cast<void> (0))
1309 "This is not a form that we understand!")(static_cast<void> (0));
1310 StoreVal = StoredVal->getOperand(0);
1311 assert(isa<LoadInst>(StoreVal) && "Not a load of NewGV!")(static_cast<void> (0));
1312 }
1313 }
1314 StoreInst *NSI =
1315 new StoreInst(StoreVal, NewGV, false, Align(1), SI->getOrdering(),
1316 SI->getSyncScopeID(), SI);
1317 NSI->setDebugLoc(SI->getDebugLoc());
1318 } else {
1319 // Change the load into a load of bool then a select.
1320 LoadInst *LI = cast<LoadInst>(UI);
1321 LoadInst *NLI = new LoadInst(NewGV->getValueType(), NewGV,
1322 LI->getName() + ".b", false, Align(1),
1323 LI->getOrdering(), LI->getSyncScopeID(), LI);
1324 Instruction *NSI;
1325 if (IsOneZero)
1326 NSI = new ZExtInst(NLI, LI->getType(), "", LI);
1327 else
1328 NSI = SelectInst::Create(NLI, OtherVal, InitVal, "", LI);
1329 NSI->takeName(LI);
1330 // Since LI is split into two instructions, NLI and NSI both inherit the
1331 // same DebugLoc
1332 NLI->setDebugLoc(LI->getDebugLoc());
1333 NSI->setDebugLoc(LI->getDebugLoc());
1334 LI->replaceAllUsesWith(NSI);
1335 }
1336 UI->eraseFromParent();
1337 }
1338
1339 // Retain the name of the old global variable. People who are debugging their
1340 // programs may expect these variables to be named the same.
1341 NewGV->takeName(GV);
1342 GV->eraseFromParent();
1343 return true;
1344}
1345
1346static bool deleteIfDead(
1347 GlobalValue &GV, SmallPtrSetImpl<const Comdat *> &NotDiscardableComdats) {
1348 GV.removeDeadConstantUsers();
1349
1350 if (!GV.isDiscardableIfUnused() && !GV.isDeclaration())
1351 return false;
1352
1353 if (const Comdat *C = GV.getComdat())
1354 if (!GV.hasLocalLinkage() && NotDiscardableComdats.count(C))
1355 return false;
1356
1357 bool Dead;
1358 if (auto *F = dyn_cast<Function>(&GV))
1359 Dead = (F->isDeclaration() && F->use_empty()) || F->isDefTriviallyDead();
1360 else
1361 Dead = GV.use_empty();
1362 if (!Dead)
1363 return false;
1364
1365 LLVM_DEBUG(dbgs() << "GLOBAL DEAD: " << GV << "\n")do { } while (false);
1366 GV.eraseFromParent();
1367 ++NumDeleted;
1368 return true;
1369}
1370
1371static bool isPointerValueDeadOnEntryToFunction(
1372 const Function *F, GlobalValue *GV,
1373 function_ref<DominatorTree &(Function &)> LookupDomTree) {
1374 // Find all uses of GV. We expect them all to be in F, and if we can't
1375 // identify any of the uses we bail out.
1376 //
1377 // On each of these uses, identify if the memory that GV points to is
1378 // used/required/live at the start of the function. If it is not, for example
1379 // if the first thing the function does is store to the GV, the GV can
1380 // possibly be demoted.
1381 //
1382 // We don't do an exhaustive search for memory operations - simply look
1383 // through bitcasts as they're quite common and benign.
1384 const DataLayout &DL = GV->getParent()->getDataLayout();
1385 SmallVector<LoadInst *, 4> Loads;
1386 SmallVector<StoreInst *, 4> Stores;
1387 for (auto *U : GV->users()) {
1388 if (Operator::getOpcode(U) == Instruction::BitCast) {
1389 for (auto *UU : U->users()) {
1390 if (auto *LI = dyn_cast<LoadInst>(UU))
1391 Loads.push_back(LI);
1392 else if (auto *SI = dyn_cast<StoreInst>(UU))
1393 Stores.push_back(SI);
1394 else
1395 return false;
1396 }
1397 continue;
1398 }
1399
1400 Instruction *I = dyn_cast<Instruction>(U);
1401 if (!I)
1402 return false;
1403 assert(I->getParent()->getParent() == F)(static_cast<void> (0));
1404
1405 if (auto *LI = dyn_cast<LoadInst>(I))
1406 Loads.push_back(LI);
1407 else if (auto *SI = dyn_cast<StoreInst>(I))
1408 Stores.push_back(SI);
1409 else
1410 return false;
1411 }
1412
1413 // We have identified all uses of GV into loads and stores. Now check if all
1414 // of them are known not to depend on the value of the global at the function
1415 // entry point. We do this by ensuring that every load is dominated by at
1416 // least one store.
1417 auto &DT = LookupDomTree(*const_cast<Function *>(F));
1418
1419 // The below check is quadratic. Check we're not going to do too many tests.
1420 // FIXME: Even though this will always have worst-case quadratic time, we
1421 // could put effort into minimizing the average time by putting stores that
1422 // have been shown to dominate at least one load at the beginning of the
1423 // Stores array, making subsequent dominance checks more likely to succeed
1424 // early.
1425 //
1426 // The threshold here is fairly large because global->local demotion is a
1427 // very powerful optimization should it fire.
1428 const unsigned Threshold = 100;
1429 if (Loads.size() * Stores.size() > Threshold)
1430 return false;
1431
1432 for (auto *L : Loads) {
1433 auto *LTy = L->getType();
1434 if (none_of(Stores, [&](const StoreInst *S) {
1435 auto *STy = S->getValueOperand()->getType();
1436 // The load is only dominated by the store if DomTree says so
1437 // and the number of bits loaded in L is less than or equal to
1438 // the number of bits stored in S.
1439 return DT.dominates(S, L) &&
1440 DL.getTypeStoreSize(LTy).getFixedSize() <=
1441 DL.getTypeStoreSize(STy).getFixedSize();
1442 }))
1443 return false;
1444 }
1445 // All loads have known dependences inside F, so the global can be localized.
1446 return true;
1447}
1448
1449/// C may have non-instruction users. Can all of those users be turned into
1450/// instructions?
1451static bool allNonInstructionUsersCanBeMadeInstructions(Constant *C) {
1452 // We don't do this exhaustively. The most common pattern that we really need
1453 // to care about is a constant GEP or constant bitcast - so just looking
1454 // through one single ConstantExpr.
1455 //
1456 // The set of constants that this function returns true for must be able to be
1457 // handled by makeAllConstantUsesInstructions.
1458 for (auto *U : C->users()) {
1459 if (isa<Instruction>(U))
1460 continue;
1461 if (!isa<ConstantExpr>(U))
1462 // Non instruction, non-constantexpr user; cannot convert this.
1463 return false;
1464 for (auto *UU : U->users())
1465 if (!isa<Instruction>(UU))
1466 // A constantexpr used by another constant. We don't try and recurse any
1467 // further but just bail out at this point.
1468 return false;
1469 }
1470
1471 return true;
1472}
1473
1474/// C may have non-instruction users, and
1475/// allNonInstructionUsersCanBeMadeInstructions has returned true. Convert the
1476/// non-instruction users to instructions.
1477static void makeAllConstantUsesInstructions(Constant *C) {
1478 SmallVector<ConstantExpr*,4> Users;
1479 for (auto *U : C->users()) {
1480 if (isa<ConstantExpr>(U))
1481 Users.push_back(cast<ConstantExpr>(U));
1482 else
1483 // We should never get here; allNonInstructionUsersCanBeMadeInstructions
1484 // should not have returned true for C.
1485 assert((static_cast<void> (0))
1486 isa<Instruction>(U) &&(static_cast<void> (0))
1487 "Can't transform non-constantexpr non-instruction to instruction!")(static_cast<void> (0));
1488 }
1489
1490 SmallVector<Value*,4> UUsers;
1491 for (auto *U : Users) {
1492 UUsers.clear();
1493 append_range(UUsers, U->users());
1494 for (auto *UU : UUsers) {
1495 Instruction *UI = cast<Instruction>(UU);
1496 Instruction *NewU = U->getAsInstruction();
1497 NewU->insertBefore(UI);
1498 UI->replaceUsesOfWith(U, NewU);
1499 }
1500 // We've replaced all the uses, so destroy the constant. (destroyConstant
1501 // will update value handles and metadata.)
1502 U->destroyConstant();
1503 }
1504}
1505
1506/// Analyze the specified global variable and optimize
1507/// it if possible. If we make a change, return true.
1508static bool
1509processInternalGlobal(GlobalVariable *GV, const GlobalStatus &GS,
1510 function_ref<TargetLibraryInfo &(Function &)> GetTLI,
1511 function_ref<DominatorTree &(Function &)> LookupDomTree) {
1512 auto &DL = GV->getParent()->getDataLayout();
1513 // If this is a first class global and has only one accessing function and
1514 // this function is non-recursive, we replace the global with a local alloca
1515 // in this function.
1516 //
1517 // NOTE: It doesn't make sense to promote non-single-value types since we
1518 // are just replacing static memory to stack memory.
1519 //
1520 // If the global is in different address space, don't bring it to stack.
1521 if (!GS.HasMultipleAccessingFunctions &&
1522 GS.AccessingFunction &&
1523 GV->getValueType()->isSingleValueType() &&
1524 GV->getType()->getAddressSpace() == 0 &&
1525 !GV->isExternallyInitialized() &&
1526 allNonInstructionUsersCanBeMadeInstructions(GV) &&
1527 GS.AccessingFunction->doesNotRecurse() &&
1528 isPointerValueDeadOnEntryToFunction(GS.AccessingFunction, GV,
1529 LookupDomTree)) {
1530 const DataLayout &DL = GV->getParent()->getDataLayout();
1531
1532 LLVM_DEBUG(dbgs() << "LOCALIZING GLOBAL: " << *GV << "\n")do { } while (false);
1533 Instruction &FirstI = const_cast<Instruction&>(*GS.AccessingFunction
1534 ->getEntryBlock().begin());
1535 Type *ElemTy = GV->getValueType();
1536 // FIXME: Pass Global's alignment when globals have alignment
1537 AllocaInst *Alloca = new AllocaInst(ElemTy, DL.getAllocaAddrSpace(), nullptr,
1538 GV->getName(), &FirstI);
1539 if (!isa<UndefValue>(GV->getInitializer()))
1540 new StoreInst(GV->getInitializer(), Alloca, &FirstI);
1541
1542 makeAllConstantUsesInstructions(GV);
1543
1544 GV->replaceAllUsesWith(Alloca);
1545 GV->eraseFromParent();
1546 ++NumLocalized;
1547 return true;
1548 }
1549
1550 bool Changed = false;
1551
1552 // If the global is never loaded (but may be stored to), it is dead.
1553 // Delete it now.
1554 if (!GS.IsLoaded) {
1555 LLVM_DEBUG(dbgs() << "GLOBAL NEVER LOADED: " << *GV << "\n")do { } while (false);
1556
1557 if (isLeakCheckerRoot(GV)) {
1558 // Delete any constant stores to the global.
1559 Changed = CleanupPointerRootUsers(GV, GetTLI);
1560 } else {
1561 // Delete any stores we can find to the global. We may not be able to
1562 // make it completely dead though.
1563 Changed =
1564 CleanupConstantGlobalUsers(GV, GV->getInitializer(), DL, GetTLI);
1565 }
1566
1567 // If the global is dead now, delete it.
1568 if (GV->use_empty()) {
1569 GV->eraseFromParent();
1570 ++NumDeleted;
1571 Changed = true;
1572 }
1573 return Changed;
1574
1575 }
1576 if (GS.StoredType <= GlobalStatus::InitializerStored) {
1577 LLVM_DEBUG(dbgs() << "MARKING CONSTANT: " << *GV << "\n")do { } while (false);
1578
1579 // Don't actually mark a global constant if it's atomic because atomic loads
1580 // are implemented by a trivial cmpxchg in some edge-cases and that usually
1581 // requires write access to the variable even if it's not actually changed.
1582 if (GS.Ordering == AtomicOrdering::NotAtomic) {
1583 assert(!GV->isConstant() && "Expected a non-constant global")(static_cast<void> (0));
1584 GV->setConstant(true);
1585 Changed = true;
1586 }
1587
1588 // Clean up any obviously simplifiable users now.
1589 Changed |= CleanupConstantGlobalUsers(GV, GV->getInitializer(), DL, GetTLI);
1590
1591 // If the global is dead now, just nuke it.
1592 if (GV->use_empty()) {
1593 LLVM_DEBUG(dbgs() << " *** Marking constant allowed us to simplify "do { } while (false)
1594 << "all users and delete global!\n")do { } while (false);
1595 GV->eraseFromParent();
1596 ++NumDeleted;
1597 return true;
1598 }
1599
1600 // Fall through to the next check; see if we can optimize further.
1601 ++NumMarked;
1602 }
1603 if (!GV->getInitializer()->getType()->isSingleValueType()) {
1604 const DataLayout &DL = GV->getParent()->getDataLayout();
1605 if (SRAGlobal(GV, DL))
1606 return true;
1607 }
1608 if (GS.StoredType == GlobalStatus::StoredOnce && GS.StoredOnceValue) {
1609 // Avoid speculating constant expressions that might trap (div/rem).
1610 auto *SOVConstant = dyn_cast<Constant>(GS.StoredOnceValue);
1611 if (SOVConstant && SOVConstant->canTrap())
1612 return Changed;
1613
1614 // If the initial value for the global was an undef value, and if only
1615 // one other value was stored into it, we can just change the
1616 // initializer to be the stored value, then delete all stores to the
1617 // global. This allows us to mark it constant.
1618 if (SOVConstant && SOVConstant->getType() == GV->getValueType() &&
1619 isa<UndefValue>(GV->getInitializer())) {
1620 // Change the initial value here.
1621 GV->setInitializer(SOVConstant);
1622
1623 // Clean up any obviously simplifiable users now.
1624 CleanupConstantGlobalUsers(GV, GV->getInitializer(), DL, GetTLI);
1625
1626 if (GV->use_empty()) {
1627 LLVM_DEBUG(dbgs() << " *** Substituting initializer allowed us to "do { } while (false)
1628 << "simplify all users and delete global!\n")do { } while (false);
1629 GV->eraseFromParent();
1630 ++NumDeleted;
1631 }
1632 ++NumSubstitute;
1633 return true;
1634 }
1635
1636 // Try to optimize globals based on the knowledge that only one value
1637 // (besides its initializer) is ever stored to the global.
1638 if (optimizeOnceStoredGlobal(GV, GS.StoredOnceValue, GS.Ordering, DL,
1639 GetTLI))
1640 return true;
1641
1642 // Otherwise, if the global was not a boolean, we can shrink it to be a
1643 // boolean.
1644 if (SOVConstant && GS.Ordering == AtomicOrdering::NotAtomic) {
1645 if (TryToShrinkGlobalToBoolean(GV, SOVConstant)) {
1646 ++NumShrunkToBool;
1647 return true;
1648 }
1649 }
1650 }
1651
1652 return Changed;
1653}
1654
1655/// Analyze the specified global variable and optimize it if possible. If we
1656/// make a change, return true.
1657static bool
1658processGlobal(GlobalValue &GV,
1659 function_ref<TargetLibraryInfo &(Function &)> GetTLI,
1660 function_ref<DominatorTree &(Function &)> LookupDomTree) {
1661 if (GV.getName().startswith("llvm."))
1662 return false;
1663
1664 GlobalStatus GS;
1665
1666 if (GlobalStatus::analyzeGlobal(&GV, GS))
1667 return false;
1668
1669 bool Changed = false;
1670 if (!GS.IsCompared && !GV.hasGlobalUnnamedAddr()) {
1671 auto NewUnnamedAddr = GV.hasLocalLinkage() ? GlobalValue::UnnamedAddr::Global
1672 : GlobalValue::UnnamedAddr::Local;
1673 if (NewUnnamedAddr != GV.getUnnamedAddr()) {
1674 GV.setUnnamedAddr(NewUnnamedAddr);
1675 NumUnnamed++;
1676 Changed = true;
1677 }
1678 }
1679
1680 // Do more involved optimizations if the global is internal.
1681 if (!GV.hasLocalLinkage())
1682 return Changed;
1683
1684 auto *GVar = dyn_cast<GlobalVariable>(&GV);
1685 if (!GVar)
1686 return Changed;
1687
1688 if (GVar->isConstant() || !GVar->hasInitializer())
1689 return Changed;
1690
1691 return processInternalGlobal(GVar, GS, GetTLI, LookupDomTree) || Changed;
1692}
1693
1694/// Walk all of the direct calls of the specified function, changing them to
1695/// FastCC.
1696static void ChangeCalleesToFastCall(Function *F) {
1697 for (User *U : F->users()) {
1698 if (isa<BlockAddress>(U))
1699 continue;
1700 cast<CallBase>(U)->setCallingConv(CallingConv::Fast);
1701 }
1702}
1703
1704static AttributeList StripAttr(LLVMContext &C, AttributeList Attrs,
1705 Attribute::AttrKind A) {
1706 unsigned AttrIndex;
1707 if (Attrs.hasAttrSomewhere(A, &AttrIndex))
1708 return Attrs.removeAttributeAtIndex(C, AttrIndex, A);
1709 return Attrs;
1710}
1711
1712static void RemoveAttribute(Function *F, Attribute::AttrKind A) {
1713 F->setAttributes(StripAttr(F->getContext(), F->getAttributes(), A));
1714 for (User *U : F->users()) {
1715 if (isa<BlockAddress>(U))
1716 continue;
1717 CallBase *CB = cast<CallBase>(U);
1718 CB->setAttributes(StripAttr(F->getContext(), CB->getAttributes(), A));
1719 }
1720}
1721
1722/// Return true if this is a calling convention that we'd like to change. The
1723/// idea here is that we don't want to mess with the convention if the user
1724/// explicitly requested something with performance implications like coldcc,
1725/// GHC, or anyregcc.
1726static bool hasChangeableCC(Function *F) {
1727 CallingConv::ID CC = F->getCallingConv();
1728
1729 // FIXME: Is it worth transforming x86_stdcallcc and x86_fastcallcc?
1730 if (CC != CallingConv::C && CC != CallingConv::X86_ThisCall)
1731 return false;
1732
1733 // FIXME: Change CC for the whole chain of musttail calls when possible.
1734 //
1735 // Can't change CC of the function that either has musttail calls, or is a
1736 // musttail callee itself
1737 for (User *U : F->users()) {
1738 if (isa<BlockAddress>(U))
1739 continue;
1740 CallInst* CI = dyn_cast<CallInst>(U);
1741 if (!CI)
1742 continue;
1743
1744 if (CI->isMustTailCall())
1745 return false;
1746 }
1747
1748 for (BasicBlock &BB : *F)
1749 if (BB.getTerminatingMustTailCall())
1750 return false;
1751
1752 return true;
1753}
1754
1755/// Return true if the block containing the call site has a BlockFrequency of
1756/// less than ColdCCRelFreq% of the entry block.
1757static bool isColdCallSite(CallBase &CB, BlockFrequencyInfo &CallerBFI) {
1758 const BranchProbability ColdProb(ColdCCRelFreq, 100);
1759 auto *CallSiteBB = CB.getParent();
1760 auto CallSiteFreq = CallerBFI.getBlockFreq(CallSiteBB);
1761 auto CallerEntryFreq =
1762 CallerBFI.getBlockFreq(&(CB.getCaller()->getEntryBlock()));
1763 return CallSiteFreq < CallerEntryFreq * ColdProb;
1764}
1765
1766// This function checks if the input function F is cold at all call sites. It
1767// also looks each call site's containing function, returning false if the
1768// caller function contains other non cold calls. The input vector AllCallsCold
1769// contains a list of functions that only have call sites in cold blocks.
1770static bool
1771isValidCandidateForColdCC(Function &F,
1772 function_ref<BlockFrequencyInfo &(Function &)> GetBFI,
1773 const std::vector<Function *> &AllCallsCold) {
1774
1775 if (F.user_empty())
1776 return false;
1777
1778 for (User *U : F.users()) {
1779 if (isa<BlockAddress>(U))
1780 continue;
1781
1782 CallBase &CB = cast<CallBase>(*U);
1783 Function *CallerFunc = CB.getParent()->getParent();
1784 BlockFrequencyInfo &CallerBFI = GetBFI(*CallerFunc);
1785 if (!isColdCallSite(CB, CallerBFI))
1786 return false;
1787 if (!llvm::is_contained(AllCallsCold, CallerFunc))
1788 return false;
1789 }
1790 return true;
1791}
1792
1793static void changeCallSitesToColdCC(Function *F) {
1794 for (User *U : F->users()) {
1795 if (isa<BlockAddress>(U))
1796 continue;
1797 cast<CallBase>(U)->setCallingConv(CallingConv::Cold);
1798 }
1799}
1800
1801// This function iterates over all the call instructions in the input Function
1802// and checks that all call sites are in cold blocks and are allowed to use the
1803// coldcc calling convention.
1804static bool
1805hasOnlyColdCalls(Function &F,
1806 function_ref<BlockFrequencyInfo &(Function &)> GetBFI) {
1807 for (BasicBlock &BB : F) {
1808 for (Instruction &I : BB) {
1809 if (CallInst *CI = dyn_cast<CallInst>(&I)) {
1810 // Skip over isline asm instructions since they aren't function calls.
1811 if (CI->isInlineAsm())
1812 continue;
1813 Function *CalledFn = CI->getCalledFunction();
1814 if (!CalledFn)
1815 return false;
1816 if (!CalledFn->hasLocalLinkage())
1817 return false;
1818 // Skip over instrinsics since they won't remain as function calls.
1819 if (CalledFn->getIntrinsicID() != Intrinsic::not_intrinsic)
1820 continue;
1821 // Check if it's valid to use coldcc calling convention.
1822 if (!hasChangeableCC(CalledFn) || CalledFn->isVarArg() ||
1823 CalledFn->hasAddressTaken())
1824 return false;
1825 BlockFrequencyInfo &CallerBFI = GetBFI(F);
1826 if (!isColdCallSite(*CI, CallerBFI))
1827 return false;
1828 }
1829 }
1830 }
1831 return true;
1832}
1833
1834static bool hasMustTailCallers(Function *F) {
1835 for (User *U : F->users()) {
1836 CallBase *CB = dyn_cast<CallBase>(U);
1837 if (!CB) {
1838 assert(isa<BlockAddress>(U) &&(static_cast<void> (0))
1839 "Expected either CallBase or BlockAddress")(static_cast<void> (0));
1840 continue;
1841 }
1842 if (CB->isMustTailCall())
1843 return true;
1844 }
1845 return false;
1846}
1847
1848static bool hasInvokeCallers(Function *F) {
1849 for (User *U : F->users())
1850 if (isa<InvokeInst>(U))
1851 return true;
1852 return false;
1853}
1854
1855static void RemovePreallocated(Function *F) {
1856 RemoveAttribute(F, Attribute::Preallocated);
1857
1858 auto *M = F->getParent();
1859
1860 IRBuilder<> Builder(M->getContext());
1861
1862 // Cannot modify users() while iterating over it, so make a copy.
1863 SmallVector<User *, 4> PreallocatedCalls(F->users());
1864 for (User *U : PreallocatedCalls) {
20
Assuming '__begin1' is not equal to '__end1'
1865 CallBase *CB = dyn_cast<CallBase>(U);
21
Assuming 'U' is a 'CallBase'
1866 if (!CB
21.1
'CB' is non-null
)
22
Taking false branch
1867 continue;
1868
1869 assert((static_cast<void> (0))
1870 !CB->isMustTailCall() &&(static_cast<void> (0))
1871 "Shouldn't call RemotePreallocated() on a musttail preallocated call")(static_cast<void> (0));
1872 // Create copy of call without "preallocated" operand bundle.
1873 SmallVector<OperandBundleDef, 1> OpBundles;
1874 CB->getOperandBundlesAsDefs(OpBundles);
1875 CallBase *PreallocatedSetup = nullptr;
23
'PreallocatedSetup' initialized to a null pointer value
1876 for (auto *It = OpBundles.begin(); It != OpBundles.end(); ++It) {
24
Assuming the condition is false
25
Loop condition is false. Execution continues on line 1883
1877 if (It->getTag() == "preallocated") {
1878 PreallocatedSetup = cast<CallBase>(*It->input_begin());
1879 OpBundles.erase(It);
1880 break;
1881 }
1882 }
1883 assert(PreallocatedSetup && "Did not find preallocated bundle")(static_cast<void> (0));
1884 uint64_t ArgCount =
1885 cast<ConstantInt>(PreallocatedSetup->getArgOperand(0))->getZExtValue();
26
Called C++ object pointer is null
1886
1887 assert((isa<CallInst>(CB) || isa<InvokeInst>(CB)) &&(static_cast<void> (0))
1888 "Unknown indirect call type")(static_cast<void> (0));
1889 CallBase *NewCB = CallBase::Create(CB, OpBundles, CB);
1890 CB->replaceAllUsesWith(NewCB);
1891 NewCB->takeName(CB);
1892 CB->eraseFromParent();
1893
1894 Builder.SetInsertPoint(PreallocatedSetup);
1895 auto *StackSave =
1896 Builder.CreateCall(Intrinsic::getDeclaration(M, Intrinsic::stacksave));
1897
1898 Builder.SetInsertPoint(NewCB->getNextNonDebugInstruction());
1899 Builder.CreateCall(Intrinsic::getDeclaration(M, Intrinsic::stackrestore),
1900 StackSave);
1901
1902 // Replace @llvm.call.preallocated.arg() with alloca.
1903 // Cannot modify users() while iterating over it, so make a copy.
1904 // @llvm.call.preallocated.arg() can be called with the same index multiple
1905 // times. So for each @llvm.call.preallocated.arg(), we see if we have
1906 // already created a Value* for the index, and if not, create an alloca and
1907 // bitcast right after the @llvm.call.preallocated.setup() so that it
1908 // dominates all uses.
1909 SmallVector<Value *, 2> ArgAllocas(ArgCount);
1910 SmallVector<User *, 2> PreallocatedArgs(PreallocatedSetup->users());
1911 for (auto *User : PreallocatedArgs) {
1912 auto *UseCall = cast<CallBase>(User);
1913 assert(UseCall->getCalledFunction()->getIntrinsicID() ==(static_cast<void> (0))
1914 Intrinsic::call_preallocated_arg &&(static_cast<void> (0))
1915 "preallocated token use was not a llvm.call.preallocated.arg")(static_cast<void> (0));
1916 uint64_t AllocArgIndex =
1917 cast<ConstantInt>(UseCall->getArgOperand(1))->getZExtValue();
1918 Value *AllocaReplacement = ArgAllocas[AllocArgIndex];
1919 if (!AllocaReplacement) {
1920 auto AddressSpace = UseCall->getType()->getPointerAddressSpace();
1921 auto *ArgType =
1922 UseCall->getFnAttr(Attribute::Preallocated).getValueAsType();
1923 auto *InsertBefore = PreallocatedSetup->getNextNonDebugInstruction();
1924 Builder.SetInsertPoint(InsertBefore);
1925 auto *Alloca =
1926 Builder.CreateAlloca(ArgType, AddressSpace, nullptr, "paarg");
1927 auto *BitCast = Builder.CreateBitCast(
1928 Alloca, Type::getInt8PtrTy(M->getContext()), UseCall->getName());
1929 ArgAllocas[AllocArgIndex] = BitCast;
1930 AllocaReplacement = BitCast;
1931 }
1932
1933 UseCall->replaceAllUsesWith(AllocaReplacement);
1934 UseCall->eraseFromParent();
1935 }
1936 // Remove @llvm.call.preallocated.setup().
1937 cast<Instruction>(PreallocatedSetup)->eraseFromParent();
1938 }
1939}
1940
1941static bool
1942OptimizeFunctions(Module &M,
1943 function_ref<TargetLibraryInfo &(Function &)> GetTLI,
1944 function_ref<TargetTransformInfo &(Function &)> GetTTI,
1945 function_ref<BlockFrequencyInfo &(Function &)> GetBFI,
1946 function_ref<DominatorTree &(Function &)> LookupDomTree,
1947 SmallPtrSetImpl<const Comdat *> &NotDiscardableComdats) {
1948
1949 bool Changed = false;
1950
1951 std::vector<Function *> AllCallsCold;
1952 for (Module::iterator FI = M.begin(), E = M.end(); FI != E;) {
6
Loop condition is false. Execution continues on line 1959
1953 Function *F = &*FI++;
1954 if (hasOnlyColdCalls(*F, GetBFI))
1955 AllCallsCold.push_back(F);
1956 }
1957
1958 // Optimize functions.
1959 for (Module::iterator FI = M.begin(), E = M.end(); FI != E; ) {
7
Loop condition is true. Entering loop body
1960 Function *F = &*FI++;
1961
1962 // Don't perform global opt pass on naked functions; we don't want fast
1963 // calling conventions for naked functions.
1964 if (F->hasFnAttribute(Attribute::Naked))
8
Assuming the condition is false
1965 continue;
1966
1967 // Functions without names cannot be referenced outside this module.
1968 if (!F->hasName() && !F->isDeclaration() && !F->hasLocalLinkage())
9
Assuming the condition is false
1969 F->setLinkage(GlobalValue::InternalLinkage);
1970
1971 if (deleteIfDead(*F, NotDiscardableComdats)) {
10
Taking false branch
1972 Changed = true;
1973 continue;
1974 }
1975
1976 // LLVM's definition of dominance allows instructions that are cyclic
1977 // in unreachable blocks, e.g.:
1978 // %pat = select i1 %condition, @global, i16* %pat
1979 // because any instruction dominates an instruction in a block that's
1980 // not reachable from entry.
1981 // So, remove unreachable blocks from the function, because a) there's
1982 // no point in analyzing them and b) GlobalOpt should otherwise grow
1983 // some more complicated logic to break these cycles.
1984 // Removing unreachable blocks might invalidate the dominator so we
1985 // recalculate it.
1986 if (!F->isDeclaration()) {
11
Assuming the condition is false
12
Taking false branch
1987 if (removeUnreachableBlocks(*F)) {
1988 auto &DT = LookupDomTree(*F);
1989 DT.recalculate(*F);
1990 Changed = true;
1991 }
1992 }
1993
1994 Changed |= processGlobal(*F, GetTLI, LookupDomTree);
1995
1996 if (!F->hasLocalLinkage())
1997 continue;
1998
1999 // If we have an inalloca parameter that we can safely remove the
2000 // inalloca attribute from, do so. This unlocks optimizations that
2001 // wouldn't be safe in the presence of inalloca.
2002 // FIXME: We should also hoist alloca affected by this to the entry
2003 // block if possible.
2004 if (F->getAttributes().hasAttrSomewhere(Attribute::InAlloca) &&
13
Assuming the condition is false
14
Taking false branch
2005 !F->hasAddressTaken() && !hasMustTailCallers(F)) {
2006 RemoveAttribute(F, Attribute::InAlloca);
2007 Changed = true;
2008 }
2009
2010 // FIXME: handle invokes
2011 // FIXME: handle musttail
2012 if (F->getAttributes().hasAttrSomewhere(Attribute::Preallocated)) {
15
Assuming the condition is true
2013 if (!F->hasAddressTaken() && !hasMustTailCallers(F) &&
16
Assuming the condition is true
18
Taking true branch
2014 !hasInvokeCallers(F)) {
17
Assuming the condition is true
2015 RemovePreallocated(F);
19
Calling 'RemovePreallocated'
2016 Changed = true;
2017 }
2018 continue;
2019 }
2020
2021 if (hasChangeableCC(F) && !F->isVarArg() && !F->hasAddressTaken()) {
2022 NumInternalFunc++;
2023 TargetTransformInfo &TTI = GetTTI(*F);
2024 // Change the calling convention to coldcc if either stress testing is
2025 // enabled or the target would like to use coldcc on functions which are
2026 // cold at all call sites and the callers contain no other non coldcc
2027 // calls.
2028 if (EnableColdCCStressTest ||
2029 (TTI.useColdCCForColdCall(*F) &&
2030 isValidCandidateForColdCC(*F, GetBFI, AllCallsCold))) {
2031 F->setCallingConv(CallingConv::Cold);
2032 changeCallSitesToColdCC(F);
2033 Changed = true;
2034 NumColdCC++;
2035 }
2036 }
2037
2038 if (hasChangeableCC(F) && !F->isVarArg() &&
2039 !F->hasAddressTaken()) {
2040 // If this function has a calling convention worth changing, is not a
2041 // varargs function, and is only called directly, promote it to use the
2042 // Fast calling convention.
2043 F->setCallingConv(CallingConv::Fast);
2044 ChangeCalleesToFastCall(F);
2045 ++NumFastCallFns;
2046 Changed = true;
2047 }
2048
2049 if (F->getAttributes().hasAttrSomewhere(Attribute::Nest) &&
2050 !F->hasAddressTaken()) {
2051 // The function is not used by a trampoline intrinsic, so it is safe
2052 // to remove the 'nest' attribute.
2053 RemoveAttribute(F, Attribute::Nest);
2054 ++NumNestRemoved;
2055 Changed = true;
2056 }
2057 }
2058 return Changed;
2059}
2060
2061static bool
2062OptimizeGlobalVars(Module &M,
2063 function_ref<TargetLibraryInfo &(Function &)> GetTLI,
2064 function_ref<DominatorTree &(Function &)> LookupDomTree,
2065 SmallPtrSetImpl<const Comdat *> &NotDiscardableComdats) {
2066 bool Changed = false;
2067
2068 for (Module::global_iterator GVI = M.global_begin(), E = M.global_end();
2069 GVI != E; ) {
2070 GlobalVariable *GV = &*GVI++;
2071 // Global variables without names cannot be referenced outside this module.
2072 if (!GV->hasName() && !GV->isDeclaration() && !GV->hasLocalLinkage())
2073 GV->setLinkage(GlobalValue::InternalLinkage);
2074 // Simplify the initializer.
2075 if (GV->hasInitializer())
2076 if (auto *C = dyn_cast<Constant>(GV->getInitializer())) {
2077 auto &DL = M.getDataLayout();
2078 // TLI is not used in the case of a Constant, so use default nullptr
2079 // for that optional parameter, since we don't have a Function to
2080 // provide GetTLI anyway.
2081 Constant *New = ConstantFoldConstant(C, DL, /*TLI*/ nullptr);
2082 if (New != C)
2083 GV->setInitializer(New);
2084 }
2085
2086 if (deleteIfDead(*GV, NotDiscardableComdats)) {
2087 Changed = true;
2088 continue;
2089 }
2090
2091 Changed |= processGlobal(*GV, GetTLI, LookupDomTree);
2092 }
2093 return Changed;
2094}
2095
2096/// Evaluate a piece of a constantexpr store into a global initializer. This
2097/// returns 'Init' modified to reflect 'Val' stored into it. At this point, the
2098/// GEP operands of Addr [0, OpNo) have been stepped into.
2099static Constant *EvaluateStoreInto(Constant *Init, Constant *Val,
2100 ConstantExpr *Addr, unsigned OpNo) {
2101 // Base case of the recursion.
2102 if (OpNo == Addr->getNumOperands()) {
2103 assert(Val->getType() == Init->getType() && "Type mismatch!")(static_cast<void> (0));
2104 return Val;
2105 }
2106
2107 SmallVector<Constant*, 32> Elts;
2108 if (StructType *STy = dyn_cast<StructType>(Init->getType())) {
2109 // Break up the constant into its elements.
2110 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
2111 Elts.push_back(Init->getAggregateElement(i));
2112
2113 // Replace the element that we are supposed to.
2114 ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo));
2115 unsigned Idx = CU->getZExtValue();
2116 assert(Idx < STy->getNumElements() && "Struct index out of range!")(static_cast<void> (0));
2117 Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1);
2118
2119 // Return the modified struct.
2120 return ConstantStruct::get(STy, Elts);
2121 }
2122
2123 ConstantInt *CI = cast<ConstantInt>(Addr->getOperand(OpNo));
2124 uint64_t NumElts;
2125 if (ArrayType *ATy = dyn_cast<ArrayType>(Init->getType()))
2126 NumElts = ATy->getNumElements();
2127 else
2128 NumElts = cast<FixedVectorType>(Init->getType())->getNumElements();
2129
2130 // Break up the array into elements.
2131 for (uint64_t i = 0, e = NumElts; i != e; ++i)
2132 Elts.push_back(Init->getAggregateElement(i));
2133
2134 assert(CI->getZExtValue() < NumElts)(static_cast<void> (0));
2135 Elts[CI->getZExtValue()] =
2136 EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1);
2137
2138 if (Init->getType()->isArrayTy())
2139 return ConstantArray::get(cast<ArrayType>(Init->getType()), Elts);
2140 return ConstantVector::get(Elts);
2141}
2142
2143/// We have decided that Addr (which satisfies the predicate
2144/// isSimpleEnoughPointerToCommit) should get Val as its value. Make it happen.
2145static void CommitValueTo(Constant *Val, Constant *Addr) {
2146 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Addr)) {
2147 assert(GV->hasInitializer())(static_cast<void> (0));
2148 GV->setInitializer(Val);
2149 return;
2150 }
2151
2152 ConstantExpr *CE = cast<ConstantExpr>(Addr);
2153 GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
2154 GV->setInitializer(EvaluateStoreInto(GV->getInitializer(), Val, CE, 2));
2155}
2156
2157/// Given a map of address -> value, where addresses are expected to be some form
2158/// of either a global or a constant GEP, set the initializer for the address to
2159/// be the value. This performs mostly the same function as CommitValueTo()
2160/// and EvaluateStoreInto() but is optimized to be more efficient for the common
2161/// case where the set of addresses are GEPs sharing the same underlying global,
2162/// processing the GEPs in batches rather than individually.
2163///
2164/// To give an example, consider the following C++ code adapted from the clang
2165/// regression tests:
2166/// struct S {
2167/// int n = 10;
2168/// int m = 2 * n;
2169/// S(int a) : n(a) {}
2170/// };
2171///
2172/// template<typename T>
2173/// struct U {
2174/// T *r = &q;
2175/// T q = 42;
2176/// U *p = this;
2177/// };
2178///
2179/// U<S> e;
2180///
2181/// The global static constructor for 'e' will need to initialize 'r' and 'p' of
2182/// the outer struct, while also initializing the inner 'q' structs 'n' and 'm'
2183/// members. This batch algorithm will simply use general CommitValueTo() method
2184/// to handle the complex nested S struct initialization of 'q', before
2185/// processing the outermost members in a single batch. Using CommitValueTo() to
2186/// handle member in the outer struct is inefficient when the struct/array is
2187/// very large as we end up creating and destroy constant arrays for each
2188/// initialization.
2189/// For the above case, we expect the following IR to be generated:
2190///
2191/// %struct.U = type { %struct.S*, %struct.S, %struct.U* }
2192/// %struct.S = type { i32, i32 }
2193/// @e = global %struct.U { %struct.S* gep inbounds (%struct.U, %struct.U* @e,
2194/// i64 0, i32 1),
2195/// %struct.S { i32 42, i32 84 }, %struct.U* @e }
2196/// The %struct.S { i32 42, i32 84 } inner initializer is treated as a complex
2197/// constant expression, while the other two elements of @e are "simple".
2198static void BatchCommitValueTo(const DenseMap<Constant*, Constant*> &Mem) {
2199 SmallVector<std::pair<GlobalVariable*, Constant*>, 32> GVs;
2200 SmallVector<std::pair<ConstantExpr*, Constant*>, 32> ComplexCEs;
2201 SmallVector<std::pair<ConstantExpr*, Constant*>, 32> SimpleCEs;
2202 SimpleCEs.reserve(Mem.size());
2203
2204 for (const auto &I : Mem) {
2205 if (auto *GV = dyn_cast<GlobalVariable>(I.first)) {
2206 GVs.push_back(std::make_pair(GV, I.second));
2207 } else {
2208 ConstantExpr *GEP = cast<ConstantExpr>(I.first);
2209 // We don't handle the deeply recursive case using the batch method.
2210 if (GEP->getNumOperands() > 3)
2211 ComplexCEs.push_back(std::make_pair(GEP, I.second));
2212 else
2213 SimpleCEs.push_back(std::make_pair(GEP, I.second));
2214 }
2215 }
2216
2217 // The algorithm below doesn't handle cases like nested structs, so use the
2218 // slower fully general method if we have to.
2219 for (auto ComplexCE : ComplexCEs)
2220 CommitValueTo(ComplexCE.second, ComplexCE.first);
2221
2222 for (auto GVPair : GVs) {
2223 assert(GVPair.first->hasInitializer())(static_cast<void> (0));
2224 GVPair.first->setInitializer(GVPair.second);
2225 }
2226
2227 if (SimpleCEs.empty())
2228 return;
2229
2230 // We cache a single global's initializer elements in the case where the
2231 // subsequent address/val pair uses the same one. This avoids throwing away and
2232 // rebuilding the constant struct/vector/array just because one element is
2233 // modified at a time.
2234 SmallVector<Constant *, 32> Elts;
2235 Elts.reserve(SimpleCEs.size());
2236 GlobalVariable *CurrentGV = nullptr;
2237
2238 auto commitAndSetupCache = [&](GlobalVariable *GV, bool Update) {
2239 Constant *Init = GV->getInitializer();
2240 Type *Ty = Init->getType();
2241 if (Update) {
2242 if (CurrentGV) {
2243 assert(CurrentGV && "Expected a GV to commit to!")(static_cast<void> (0));
2244 Type *CurrentInitTy = CurrentGV->getInitializer()->getType();
2245 // We have a valid cache that needs to be committed.
2246 if (StructType *STy = dyn_cast<StructType>(CurrentInitTy))
2247 CurrentGV->setInitializer(ConstantStruct::get(STy, Elts));
2248 else if (ArrayType *ArrTy = dyn_cast<ArrayType>(CurrentInitTy))
2249 CurrentGV->setInitializer(ConstantArray::get(ArrTy, Elts));
2250 else
2251 CurrentGV->setInitializer(ConstantVector::get(Elts));
2252 }
2253 if (CurrentGV == GV)
2254 return;
2255 // Need to clear and set up cache for new initializer.
2256 CurrentGV = GV;
2257 Elts.clear();
2258 unsigned NumElts;
2259 if (auto *STy = dyn_cast<StructType>(Ty))
2260 NumElts = STy->getNumElements();
2261 else if (auto *ATy = dyn_cast<ArrayType>(Ty))
2262 NumElts = ATy->getNumElements();
2263 else
2264 NumElts = cast<FixedVectorType>(Ty)->getNumElements();
2265 for (unsigned i = 0, e = NumElts; i != e; ++i)
2266 Elts.push_back(Init->getAggregateElement(i));
2267 }
2268 };
2269
2270 for (auto CEPair : SimpleCEs) {
2271 ConstantExpr *GEP = CEPair.first;
2272 Constant *Val = CEPair.second;
2273
2274 GlobalVariable *GV = cast<GlobalVariable>(GEP->getOperand(0));
2275 commitAndSetupCache(GV, GV != CurrentGV);
2276 ConstantInt *CI = cast<ConstantInt>(GEP->getOperand(2));
2277 Elts[CI->getZExtValue()] = Val;
2278 }
2279 // The last initializer in the list needs to be committed, others
2280 // will be committed on a new initializer being processed.
2281 commitAndSetupCache(CurrentGV, true);
2282}
2283
2284/// Evaluate static constructors in the function, if we can. Return true if we
2285/// can, false otherwise.
2286static bool EvaluateStaticConstructor(Function *F, const DataLayout &DL,
2287 TargetLibraryInfo *TLI) {
2288 // Call the function.
2289 Evaluator Eval(DL, TLI);
2290 Constant *RetValDummy;
2291 bool EvalSuccess = Eval.EvaluateFunction(F, RetValDummy,
2292 SmallVector<Constant*, 0>());
2293
2294 if (EvalSuccess) {
2295 ++NumCtorsEvaluated;
2296
2297 // We succeeded at evaluation: commit the result.
2298 LLVM_DEBUG(dbgs() << "FULLY EVALUATED GLOBAL CTOR FUNCTION '"do { } while (false)
2299 << F->getName() << "' to "do { } while (false)
2300 << Eval.getMutatedMemory().size() << " stores.\n")do { } while (false);
2301 BatchCommitValueTo(Eval.getMutatedMemory());
2302 for (GlobalVariable *GV : Eval.getInvariants())
2303 GV->setConstant(true);
2304 }
2305
2306 return EvalSuccess;
2307}
2308
2309static int compareNames(Constant *const *A, Constant *const *B) {
2310 Value *AStripped = (*A)->stripPointerCasts();
2311 Value *BStripped = (*B)->stripPointerCasts();
2312 return AStripped->getName().compare(BStripped->getName());
2313}
2314
2315static void setUsedInitializer(GlobalVariable &V,
2316 const SmallPtrSetImpl<GlobalValue *> &Init) {
2317 if (Init.empty()) {
2318 V.eraseFromParent();
2319 return;
2320 }
2321
2322 // Type of pointer to the array of pointers.
2323 PointerType *Int8PtrTy = Type::getInt8PtrTy(V.getContext(), 0);
2324
2325 SmallVector<Constant *, 8> UsedArray;
2326 for (GlobalValue *GV : Init) {
2327 Constant *Cast
2328 = ConstantExpr::getPointerBitCastOrAddrSpaceCast(GV, Int8PtrTy);
2329 UsedArray.push_back(Cast);
2330 }
2331 // Sort to get deterministic order.
2332 array_pod_sort(UsedArray.begin(), UsedArray.end(), compareNames);
2333 ArrayType *ATy = ArrayType::get(Int8PtrTy, UsedArray.size());
2334
2335 Module *M = V.getParent();
2336 V.removeFromParent();
2337 GlobalVariable *NV =
2338 new GlobalVariable(*M, ATy, false, GlobalValue::AppendingLinkage,
2339 ConstantArray::get(ATy, UsedArray), "");
2340 NV->takeName(&V);
2341 NV->setSection("llvm.metadata");
2342 delete &V;
2343}
2344
2345namespace {
2346
2347/// An easy to access representation of llvm.used and llvm.compiler.used.
2348class LLVMUsed {
2349 SmallPtrSet<GlobalValue *, 4> Used;
2350 SmallPtrSet<GlobalValue *, 4> CompilerUsed;
2351 GlobalVariable *UsedV;
2352 GlobalVariable *CompilerUsedV;
2353
2354public:
2355 LLVMUsed(Module &M) {
2356 SmallVector<GlobalValue *, 4> Vec;
2357 UsedV = collectUsedGlobalVariables(M, Vec, false);
2358 Used = {Vec.begin(), Vec.end()};
2359 Vec.clear();
2360 CompilerUsedV = collectUsedGlobalVariables(M, Vec, true);
2361 CompilerUsed = {Vec.begin(), Vec.end()};
2362 }
2363
2364 using iterator = SmallPtrSet<GlobalValue *, 4>::iterator;
2365 using used_iterator_range = iterator_range<iterator>;
2366
2367 iterator usedBegin() { return Used.begin(); }
2368 iterator usedEnd() { return Used.end(); }
2369
2370 used_iterator_range used() {
2371 return used_iterator_range(usedBegin(), usedEnd());
2372 }
2373
2374 iterator compilerUsedBegin() { return CompilerUsed.begin(); }
2375 iterator compilerUsedEnd() { return CompilerUsed.end(); }
2376
2377 used_iterator_range compilerUsed() {
2378 return used_iterator_range(compilerUsedBegin(), compilerUsedEnd());
2379 }
2380
2381 bool usedCount(GlobalValue *GV) const { return Used.count(GV); }
2382
2383 bool compilerUsedCount(GlobalValue *GV) const {
2384 return CompilerUsed.count(GV);
2385 }
2386
2387 bool usedErase(GlobalValue *GV) { return Used.erase(GV); }
2388 bool compilerUsedErase(GlobalValue *GV) { return CompilerUsed.erase(GV); }
2389 bool usedInsert(GlobalValue *GV) { return Used.insert(GV).second; }
2390
2391 bool compilerUsedInsert(GlobalValue *GV) {
2392 return CompilerUsed.insert(GV).second;
2393 }
2394
2395 void syncVariablesAndSets() {
2396 if (UsedV)
2397 setUsedInitializer(*UsedV, Used);
2398 if (CompilerUsedV)
2399 setUsedInitializer(*CompilerUsedV, CompilerUsed);
2400 }
2401};
2402
2403} // end anonymous namespace
2404
2405static bool hasUseOtherThanLLVMUsed(GlobalAlias &GA, const LLVMUsed &U) {
2406 if (GA.use_empty()) // No use at all.
2407 return false;
2408
2409 assert((!U.usedCount(&GA) || !U.compilerUsedCount(&GA)) &&(static_cast<void> (0))
2410 "We should have removed the duplicated "(static_cast<void> (0))
2411 "element from llvm.compiler.used")(static_cast<void> (0));
2412 if (!GA.hasOneUse())
2413 // Strictly more than one use. So at least one is not in llvm.used and
2414 // llvm.compiler.used.
2415 return true;
2416
2417 // Exactly one use. Check if it is in llvm.used or llvm.compiler.used.
2418 return !U.usedCount(&GA) && !U.compilerUsedCount(&GA);
2419}
2420
2421static bool hasMoreThanOneUseOtherThanLLVMUsed(GlobalValue &V,
2422 const LLVMUsed &U) {
2423 unsigned N = 2;
2424 assert((!U.usedCount(&V) || !U.compilerUsedCount(&V)) &&(static_cast<void> (0))
2425 "We should have removed the duplicated "(static_cast<void> (0))
2426 "element from llvm.compiler.used")(static_cast<void> (0));
2427 if (U.usedCount(&V) || U.compilerUsedCount(&V))
2428 ++N;
2429 return V.hasNUsesOrMore(N);
2430}
2431
2432static bool mayHaveOtherReferences(GlobalAlias &GA, const LLVMUsed &U) {
2433 if (!GA.hasLocalLinkage())
2434 return true;
2435
2436 return U.usedCount(&GA) || U.compilerUsedCount(&GA);
2437}
2438
2439static bool hasUsesToReplace(GlobalAlias &GA, const LLVMUsed &U,
2440 bool &RenameTarget) {
2441 RenameTarget = false;
2442 bool Ret = false;
2443 if (hasUseOtherThanLLVMUsed(GA, U))
2444 Ret = true;
2445
2446 // If the alias is externally visible, we may still be able to simplify it.
2447 if (!mayHaveOtherReferences(GA, U))
2448 return Ret;
2449
2450 // If the aliasee has internal linkage, give it the name and linkage
2451 // of the alias, and delete the alias. This turns:
2452 // define internal ... @f(...)
2453 // @a = alias ... @f
2454 // into:
2455 // define ... @a(...)
2456 Constant *Aliasee = GA.getAliasee();
2457 GlobalValue *Target = cast<GlobalValue>(Aliasee->stripPointerCasts());
2458 if (!Target->hasLocalLinkage())
2459 return Ret;
2460
2461 // Do not perform the transform if multiple aliases potentially target the
2462 // aliasee. This check also ensures that it is safe to replace the section
2463 // and other attributes of the aliasee with those of the alias.
2464 if (hasMoreThanOneUseOtherThanLLVMUsed(*Target, U))
2465 return Ret;
2466
2467 RenameTarget = true;
2468 return true;
2469}
2470
2471static bool
2472OptimizeGlobalAliases(Module &M,
2473 SmallPtrSetImpl<const Comdat *> &NotDiscardableComdats) {
2474 bool Changed = false;
2475 LLVMUsed Used(M);
2476
2477 for (GlobalValue *GV : Used.used())
2478 Used.compilerUsedErase(GV);
2479
2480 for (Module::alias_iterator I = M.alias_begin(), E = M.alias_end();
2481 I != E;) {
2482 GlobalAlias *J = &*I++;
2483
2484 // Aliases without names cannot be referenced outside this module.
2485 if (!J->hasName() && !J->isDeclaration() && !J->hasLocalLinkage())
2486 J->setLinkage(GlobalValue::InternalLinkage);
2487
2488 if (deleteIfDead(*J, NotDiscardableComdats)) {
2489 Changed = true;
2490 continue;
2491 }
2492
2493 // If the alias can change at link time, nothing can be done - bail out.
2494 if (J->isInterposable())
2495 continue;
2496
2497 Constant *Aliasee = J->getAliasee();
2498 GlobalValue *Target = dyn_cast<GlobalValue>(Aliasee->stripPointerCasts());
2499 // We can't trivially replace the alias with the aliasee if the aliasee is
2500 // non-trivial in some way. We also can't replace the alias with the aliasee
2501 // if the aliasee is interposable because aliases point to the local
2502 // definition.
2503 // TODO: Try to handle non-zero GEPs of local aliasees.
2504 if (!Target || Target->isInterposable())
2505 continue;
2506 Target->removeDeadConstantUsers();
2507
2508 // Make all users of the alias use the aliasee instead.
2509 bool RenameTarget;
2510 if (!hasUsesToReplace(*J, Used, RenameTarget))
2511 continue;
2512
2513 J->replaceAllUsesWith(ConstantExpr::getBitCast(Aliasee, J->getType()));
2514 ++NumAliasesResolved;
2515 Changed = true;
2516
2517 if (RenameTarget) {
2518 // Give the aliasee the name, linkage and other attributes of the alias.
2519 Target->takeName(&*J);
2520 Target->setLinkage(J->getLinkage());
2521 Target->setDSOLocal(J->isDSOLocal());
2522 Target->setVisibility(J->getVisibility());
2523 Target->setDLLStorageClass(J->getDLLStorageClass());
2524
2525 if (Used.usedErase(&*J))
2526 Used.usedInsert(Target);
2527
2528 if (Used.compilerUsedErase(&*J))
2529 Used.compilerUsedInsert(Target);
2530 } else if (mayHaveOtherReferences(*J, Used))
2531 continue;
2532
2533 // Delete the alias.
2534 M.getAliasList().erase(J);
2535 ++NumAliasesRemoved;
2536 Changed = true;
2537 }
2538
2539 Used.syncVariablesAndSets();
2540
2541 return Changed;
2542}
2543
2544static Function *
2545FindCXAAtExit(Module &M, function_ref<TargetLibraryInfo &(Function &)> GetTLI) {
2546 // Hack to get a default TLI before we have actual Function.
2547 auto FuncIter = M.begin();
2548 if (FuncIter == M.end())
2549 return nullptr;
2550 auto *TLI = &GetTLI(*FuncIter);
2551
2552 LibFunc F = LibFunc_cxa_atexit;
2553 if (!TLI->has(F))
2554 return nullptr;
2555
2556 Function *Fn = M.getFunction(TLI->getName(F));
2557 if (!Fn)
2558 return nullptr;
2559
2560 // Now get the actual TLI for Fn.
2561 TLI = &GetTLI(*Fn);
2562
2563 // Make sure that the function has the correct prototype.
2564 if (!TLI->getLibFunc(*Fn, F) || F != LibFunc_cxa_atexit)
2565 return nullptr;
2566
2567 return Fn;
2568}
2569
2570/// Returns whether the given function is an empty C++ destructor and can
2571/// therefore be eliminated.
2572/// Note that we assume that other optimization passes have already simplified
2573/// the code so we simply check for 'ret'.
2574static bool cxxDtorIsEmpty(const Function &Fn) {
2575 // FIXME: We could eliminate C++ destructors if they're readonly/readnone and
2576 // nounwind, but that doesn't seem worth doing.
2577 if (Fn.isDeclaration())
2578 return false;
2579
2580 for (auto &I : Fn.getEntryBlock()) {
2581 if (isa<DbgInfoIntrinsic>(I))
2582 continue;
2583 if (isa<ReturnInst>(I))
2584 return true;
2585 break;
2586 }
2587 return false;
2588}
2589
2590static bool OptimizeEmptyGlobalCXXDtors(Function *CXAAtExitFn) {
2591 /// Itanium C++ ABI p3.3.5:
2592 ///
2593 /// After constructing a global (or local static) object, that will require
2594 /// destruction on exit, a termination function is registered as follows:
2595 ///
2596 /// extern "C" int __cxa_atexit ( void (*f)(void *), void *p, void *d );
2597 ///
2598 /// This registration, e.g. __cxa_atexit(f,p,d), is intended to cause the
2599 /// call f(p) when DSO d is unloaded, before all such termination calls
2600 /// registered before this one. It returns zero if registration is
2601 /// successful, nonzero on failure.
2602
2603 // This pass will look for calls to __cxa_atexit where the function is trivial
2604 // and remove them.
2605 bool Changed = false;
2606
2607 for (auto I = CXAAtExitFn->user_begin(), E = CXAAtExitFn->user_end();
2608 I != E;) {
2609 // We're only interested in calls. Theoretically, we could handle invoke
2610 // instructions as well, but neither llvm-gcc nor clang generate invokes
2611 // to __cxa_atexit.
2612 CallInst *CI = dyn_cast<CallInst>(*I++);
2613 if (!CI)
2614 continue;
2615
2616 Function *DtorFn =
2617 dyn_cast<Function>(CI->getArgOperand(0)->stripPointerCasts());
2618 if (!DtorFn || !cxxDtorIsEmpty(*DtorFn))
2619 continue;
2620
2621 // Just remove the call.
2622 CI->replaceAllUsesWith(Constant::getNullValue(CI->getType()));
2623 CI->eraseFromParent();
2624
2625 ++NumCXXDtorsRemoved;
2626
2627 Changed |= true;
2628 }
2629
2630 return Changed;
2631}
2632
2633static bool optimizeGlobalsInModule(
2634 Module &M, const DataLayout &DL,
2635 function_ref<TargetLibraryInfo &(Function &)> GetTLI,
2636 function_ref<TargetTransformInfo &(Function &)> GetTTI,
2637 function_ref<BlockFrequencyInfo &(Function &)> GetBFI,
2638 function_ref<DominatorTree &(Function &)> LookupDomTree) {
2639 SmallPtrSet<const Comdat *, 8> NotDiscardableComdats;
2640 bool Changed = false;
2641 bool LocalChange = true;
2642 while (LocalChange) {
4
Loop condition is true. Entering loop body
2643 LocalChange = false;
2644
2645 NotDiscardableComdats.clear();
2646 for (const GlobalVariable &GV : M.globals())
2647 if (const Comdat *C = GV.getComdat())
2648 if (!GV.isDiscardableIfUnused() || !GV.use_empty())
2649 NotDiscardableComdats.insert(C);
2650 for (Function &F : M)
2651 if (const Comdat *C = F.getComdat())
2652 if (!F.isDefTriviallyDead())
2653 NotDiscardableComdats.insert(C);
2654 for (GlobalAlias &GA : M.aliases())
2655 if (const Comdat *C = GA.getComdat())
2656 if (!GA.isDiscardableIfUnused() || !GA.use_empty())
2657 NotDiscardableComdats.insert(C);
2658
2659 // Delete functions that are trivially dead, ccc -> fastcc
2660 LocalChange |= OptimizeFunctions(M, GetTLI, GetTTI, GetBFI, LookupDomTree,
5
Calling 'OptimizeFunctions'
2661 NotDiscardableComdats);
2662
2663 // Optimize global_ctors list.
2664 LocalChange |= optimizeGlobalCtorsList(M, [&](Function *F) {
2665 return EvaluateStaticConstructor(F, DL, &GetTLI(*F));
2666 });
2667
2668 // Optimize non-address-taken globals.
2669 LocalChange |=
2670 OptimizeGlobalVars(M, GetTLI, LookupDomTree, NotDiscardableComdats);
2671
2672 // Resolve aliases, when possible.
2673 LocalChange |= OptimizeGlobalAliases(M, NotDiscardableComdats);
2674
2675 // Try to remove trivial global destructors if they are not removed
2676 // already.
2677 Function *CXAAtExitFn = FindCXAAtExit(M, GetTLI);
2678 if (CXAAtExitFn)
2679 LocalChange |= OptimizeEmptyGlobalCXXDtors(CXAAtExitFn);
2680
2681 Changed |= LocalChange;
2682 }
2683
2684 // TODO: Move all global ctors functions to the end of the module for code
2685 // layout.
2686
2687 return Changed;
2688}
2689
2690PreservedAnalyses GlobalOptPass::run(Module &M, ModuleAnalysisManager &AM) {
2691 auto &DL = M.getDataLayout();
2692 auto &FAM =
2693 AM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager();
2694 auto LookupDomTree = [&FAM](Function &F) -> DominatorTree &{
2695 return FAM.getResult<DominatorTreeAnalysis>(F);
2696 };
2697 auto GetTLI = [&FAM](Function &F) -> TargetLibraryInfo & {
2698 return FAM.getResult<TargetLibraryAnalysis>(F);
2699 };
2700 auto GetTTI = [&FAM](Function &F) -> TargetTransformInfo & {
2701 return FAM.getResult<TargetIRAnalysis>(F);
2702 };
2703
2704 auto GetBFI = [&FAM](Function &F) -> BlockFrequencyInfo & {
2705 return FAM.getResult<BlockFrequencyAnalysis>(F);
2706 };
2707
2708 if (!optimizeGlobalsInModule(M, DL, GetTLI, GetTTI, GetBFI, LookupDomTree))
2709 return PreservedAnalyses::all();
2710 return PreservedAnalyses::none();
2711}
2712
2713namespace {
2714
2715struct GlobalOptLegacyPass : public ModulePass {
2716 static char ID; // Pass identification, replacement for typeid
2717
2718 GlobalOptLegacyPass() : ModulePass(ID) {
2719 initializeGlobalOptLegacyPassPass(*PassRegistry::getPassRegistry());
2720 }
2721
2722 bool runOnModule(Module &M) override {
2723 if (skipModule(M))
1
Assuming the condition is false
2
Taking false branch
2724 return false;
2725
2726 auto &DL = M.getDataLayout();
2727 auto LookupDomTree = [this](Function &F) -> DominatorTree & {
2728 return this->getAnalysis<DominatorTreeWrapperPass>(F).getDomTree();
2729 };
2730 auto GetTLI = [this](Function &F) -> TargetLibraryInfo & {
2731 return this->getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
2732 };
2733 auto GetTTI = [this](Function &F) -> TargetTransformInfo & {
2734 return this->getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
2735 };
2736
2737 auto GetBFI = [this](Function &F) -> BlockFrequencyInfo & {
2738 return this->getAnalysis<BlockFrequencyInfoWrapperPass>(F).getBFI();
2739 };
2740
2741 return optimizeGlobalsInModule(M, DL, GetTLI, GetTTI, GetBFI,
3
Calling 'optimizeGlobalsInModule'
2742 LookupDomTree);
2743 }
2744
2745 void getAnalysisUsage(AnalysisUsage &AU) const override {
2746 AU.addRequired<TargetLibraryInfoWrapperPass>();
2747 AU.addRequired<TargetTransformInfoWrapperPass>();
2748 AU.addRequired<DominatorTreeWrapperPass>();
2749 AU.addRequired<BlockFrequencyInfoWrapperPass>();
2750 }
2751};
2752
2753} // end anonymous namespace
2754
2755char GlobalOptLegacyPass::ID = 0;
2756
2757INITIALIZE_PASS_BEGIN(GlobalOptLegacyPass, "globalopt",static void *initializeGlobalOptLegacyPassPassOnce(PassRegistry
&Registry) {
2758 "Global Variable Optimizer", false, false)static void *initializeGlobalOptLegacyPassPassOnce(PassRegistry
&Registry) {
2759INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)initializeTargetLibraryInfoWrapperPassPass(Registry);
2760INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)initializeTargetTransformInfoWrapperPassPass(Registry);
2761INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)initializeBlockFrequencyInfoWrapperPassPass(Registry);
2762INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)initializeDominatorTreeWrapperPassPass(Registry);
2763INITIALIZE_PASS_END(GlobalOptLegacyPass, "globalopt",PassInfo *PI = new PassInfo( "Global Variable Optimizer", "globalopt"
, &GlobalOptLegacyPass::ID, PassInfo::NormalCtor_t(callDefaultCtor
<GlobalOptLegacyPass>), false, false); Registry.registerPass
(*PI, true); return PI; } static llvm::once_flag InitializeGlobalOptLegacyPassPassFlag
; void llvm::initializeGlobalOptLegacyPassPass(PassRegistry &
Registry) { llvm::call_once(InitializeGlobalOptLegacyPassPassFlag
, initializeGlobalOptLegacyPassPassOnce, std::ref(Registry));
}
2764 "Global Variable Optimizer", false, false)PassInfo *PI = new PassInfo( "Global Variable Optimizer", "globalopt"
, &GlobalOptLegacyPass::ID, PassInfo::NormalCtor_t(callDefaultCtor
<GlobalOptLegacyPass>), false, false); Registry.registerPass
(*PI, true); return PI; } static llvm::once_flag InitializeGlobalOptLegacyPassPassFlag
; void llvm::initializeGlobalOptLegacyPassPass(PassRegistry &
Registry) { llvm::call_once(InitializeGlobalOptLegacyPassPassFlag
, initializeGlobalOptLegacyPassPassOnce, std::ref(Registry));
}
2765
2766ModulePass *llvm::createGlobalOptimizerPass() {
2767 return new GlobalOptLegacyPass();
2768}