This is an archive of the discontinued LLVM Phabricator instance.

Differential D102257

[GlobalOpt] Remove heap SROA
ClosedPublic

Authored by MaskRay on May 11 2021, 10:24 AM.

Details

Reviewers
aeubanks
evgeny777
reames
Commits
rG129f466e222e: [GlobalOpt] Remove heap SROA
Summary

GlobalOpt implements a heap SROA (SROA for an malloc allocatated struct or array
of structs) which is largely undertested (heap-sra-[1234].ll are basically the
same test with very little difference) and does not trigger at all when
bootstrapping clang (it only supports the case of one single store).

The heap SROA implementation causes PR50027 (GEP is not properly handled; crash or miscompile).
Just drop the implementation. I have deleted some obviously duplicated tests
but kept heap-sra-[12]{,-no-nullopt}.ll.

Diff Detail

Event Timeline

MaskRay created this revision.May 11 2021, 10:24 AM
Herald added subscribers: jfb, hiraditya. · View Herald TranscriptMay 11 2021, 10:24 AM
MaskRay requested review of this revision.May 11 2021, 10:24 AM
Herald added a project: Restricted Project. · View Herald TranscriptMay 11 2021, 10:24 AM
Herald added a subscriber: llvm-commits. · View Herald Transcript
MaskRay mentioned this in D101245: [GlobalOpt] Disable heap SROA when GEP of the only storer is used.May 11 2021, 10:25 AM
MaskRay updated this revision to Diff 344470.May 11 2021, 10:32 AM

Delete more dead code

aeubanks added a comment.May 11 2021, 10:42 AM

seems fine, small nits

llvm/lib/Transforms/IPO/GlobalOpt.cpp
1015–1026

these could still be used to get to OptimizeGlobalAddressOfMalloc, why are these being deleted?

if you are deleting the Phi part, PHIs is no longer used

MaskRay edited the summary of this revision. (Show Details)May 11 2021, 10:50 AM
MaskRay updated this revision to Diff 344484.May 11 2021, 10:50 AM

Delete PHIS - unused after removal of heap SROA.

aeubanks accepted this revision.May 11 2021, 11:07 AM

lgtm

This revision is now accepted and ready to land.May 11 2021, 11:07 AM
MaskRay updated this revision to Diff 344509.May 11 2021, 11:34 AM

Fix the case of the function name while updating its signature.

This revision was landed with ongoing or failed builds.May 11 2021, 11:34 AM
Closed by commit rG129f466e222e: [GlobalOpt] Remove heap SROA (authored by MaskRay). · Explain Why
This revision was automatically updated to reflect the committed changes.
MaskRay added a commit: rG129f466e222e: [GlobalOpt] Remove heap SROA.
Harbormaster completed remote builds in B103819: Diff 344509.May 11 2021, 1:07 PM

Revision Contents

Diff 344484

llvm/lib/Transforms/IPO/GlobalOpt.cpp

Show First 20 Lines • Show All 72 Lines • ▼ Show 20 Lines
using namespace llvm; using namespace llvm;
#define DEBUG_TYPE "globalopt" #define DEBUG_TYPE "globalopt"
STATISTIC(NumMarked , "Number of globals marked constant"); STATISTIC(NumMarked , "Number of globals marked constant");
STATISTIC(NumUnnamed , "Number of globals marked unnamed_addr"); STATISTIC(NumUnnamed , "Number of globals marked unnamed_addr");
STATISTIC(NumSRA , "Number of aggregate globals broken into scalars"); STATISTIC(NumSRA , "Number of aggregate globals broken into scalars");
STATISTIC(NumHeapSRA , "Number of heap objects SRA'd");
STATISTIC(NumSubstitute,"Number of globals with initializers stored into them"); STATISTIC(NumSubstitute,"Number of globals with initializers stored into them");
STATISTIC(NumDeleted , "Number of globals deleted"); STATISTIC(NumDeleted , "Number of globals deleted");
STATISTIC(NumGlobUses , "Number of global uses devirtualized"); STATISTIC(NumGlobUses , "Number of global uses devirtualized");
STATISTIC(NumLocalized , "Number of globals localized"); STATISTIC(NumLocalized , "Number of globals localized");
STATISTIC(NumShrunkToBool , "Number of global vars shrunk to booleans"); STATISTIC(NumShrunkToBool , "Number of global vars shrunk to booleans");
STATISTIC(NumFastCallFns , "Number of functions converted to fastcc"); STATISTIC(NumFastCallFns , "Number of functions converted to fastcc");
STATISTIC(NumCtorsEvaluated, "Number of static ctors evaluated"); STATISTIC(NumCtorsEvaluated, "Number of static ctors evaluated");
STATISTIC(NumNestRemoved , "Number of nest attributes removed"); STATISTIC(NumNestRemoved , "Number of nest attributes removed");
▲ Show 20 Lines • Show All 901 Lines • ▼ Show 20 Lines if (RepValue != NewGV)
ConstantPropUsersOf(RepValue, DL, TLI); ConstantPropUsersOf(RepValue, DL, TLI);
return NewGV; return NewGV;
} }
/// Scan the use-list of V checking to make sure that there are no complex uses /// Scan the use-list of V checking to make sure that there are no complex uses
/// of V. We permit simple things like dereferencing the pointer, but not /// of V. We permit simple things like dereferencing the pointer, but not
/// storing through the address, unless it is to the specified global. /// storing through the address, unless it is to the specified global.
static bool ValueIsOnlyUsedLocallyOrStoredToOneGlobal(const Instruction *V, static bool
const GlobalVariable *GV, ValueIsOnlyUsedLocallyOrStoredToOneGlobal(const Instruction *V,
Lint: Pre-merge checks
Inline Actions

clang-tidy: warning: invalid case style for function 'ValueIsOnlyUsedLocallyOrStoredToOneGlobal' [readability-identifier-naming]
not useful

Lint: Pre-merge checks: clang-tidy: warning: invalid case style for function…
SmallPtrSetImpl<const PHINode*> &PHIs) { const GlobalVariable *GV) {
for (const User *U : V->users()) { for (const User *U : V->users()) {
const Instruction *Inst = cast<Instruction>(U); const Instruction *Inst = cast<Instruction>(U);
if (isa<LoadInst>(Inst) || isa<CmpInst>(Inst)) { if (isa<LoadInst>(Inst) || isa<CmpInst>(Inst)) {
continue; // Fine, ignore. continue; // Fine, ignore.
} }
if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) { if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
if (SI->getOperand(0) == V && SI->getOperand(1) != GV) if (SI->getOperand(0) == V && SI->getOperand(1) != GV)
return false; // Storing the pointer itself... bad. return false; // Storing the pointer itself... bad.
continue; // Otherwise, storing through it, or storing into GV... fine. continue; // Otherwise, storing through it, or storing into GV... fine.
} }
// Must index into the array and into the struct.
if (isa<GetElementPtrInst>(Inst) && Inst->getNumOperands() >= 3) {
if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(Inst, GV, PHIs))
return false;
continue;
}
if (const PHINode *PN = dyn_cast<PHINode>(Inst)) {
// PHIs are ok if all uses are ok. Don't infinitely recurse through PHI
// cycles.
if (PHIs.insert(PN).second)
if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(PN, GV, PHIs))
aeubanksUnsubmitted
Not Done
ReplyInline Actions

these could still be used to get to OptimizeGlobalAddressOfMalloc, why are these being deleted?

if you are deleting the Phi part, PHIs is no longer used

aeubanks: these could still be used to get to `OptimizeGlobalAddressOfMalloc`, why are these being…
return false;
continue;
}
if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Inst)) { if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Inst)) {
if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(BCI, GV, PHIs)) if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(BCI, GV))
return false;
continue;
}
return false;
}
return true;
}
/// The Alloc pointer is stored into GV somewhere. Transform all uses of the
/// allocation into loads from the global and uses of the resultant pointer.
/// Further, delete the store into GV. This assumes that these value pass the
/// 'ValueIsOnlyUsedLocallyOrStoredToOneGlobal' predicate.
static void ReplaceUsesOfMallocWithGlobal(Instruction *Alloc,
GlobalVariable *GV) {
while (!Alloc->use_empty()) {
Instruction *U = cast<Instruction>(*Alloc->user_begin());
Instruction *InsertPt = U;
if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
// If this is the store of the allocation into the global, remove it.
if (SI->getOperand(1) == GV) {
SI->eraseFromParent();
continue;
}
} else if (PHINode *PN = dyn_cast<PHINode>(U)) {
// Insert the load in the corresponding predecessor, not right before the
// PHI.
InsertPt = PN->getIncomingBlock(*Alloc->use_begin())->getTerminator();
} else if (isa<BitCastInst>(U)) {
// Must be bitcast between the malloc and store to initialize the global.
ReplaceUsesOfMallocWithGlobal(U, GV);
U->eraseFromParent();
continue;
} else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(U)) {
// If this is a "GEP bitcast" and the user is a store to the global, then
// just process it as a bitcast.
if (GEPI->hasAllZeroIndices() && GEPI->hasOneUse())
if (StoreInst *SI = dyn_cast<StoreInst>(GEPI->user_back()))
if (SI->getOperand(1) == GV) {
// Must be bitcast GEP between the malloc and store to initialize
// the global.
ReplaceUsesOfMallocWithGlobal(GEPI, GV);
GEPI->eraseFromParent();
continue;
}
}
// Insert a load from the global, and use it instead of the malloc.
Value *NL =
new LoadInst(GV->getValueType(), GV, GV->getName() + ".val", InsertPt);
U->replaceUsesOfWith(Alloc, NL);
}
}
/// Verify that all uses of V (a load, or a phi of a load) are simple enough to
/// perform heap SRA on. This permits GEP's that index through the array and
/// struct field, icmps of null, and PHIs.
static bool LoadUsesSimpleEnoughForHeapSRA(const Value *V,
SmallPtrSetImpl<const PHINode*> &LoadUsingPHIs,
SmallPtrSetImpl<const PHINode*> &LoadUsingPHIsPerLoad) {
// We permit two users of the load: setcc comparing against the null
// pointer, and a getelementptr of a specific form.
for (const User *U : V->users()) {
const Instruction *UI = cast<Instruction>(U);
// Comparison against null is ok.
if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UI)) {
if (!isa<ConstantPointerNull>(ICI->getOperand(1)))
return false; return false;
continue; continue;
} }
// getelementptr is also ok, but only a simple form.
if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(UI)) {
// Must index into the array and into the struct.
if (GEPI->getNumOperands() < 3)
return false; return false;
// Otherwise the GEP is ok.
continue;
}
if (const PHINode *PN = dyn_cast<PHINode>(UI)) {
if (!LoadUsingPHIsPerLoad.insert(PN).second)
// This means some phi nodes are dependent on each other.
// Avoid infinite looping!
return false;
if (!LoadUsingPHIs.insert(PN).second)
// If we have already analyzed this PHI, then it is safe.
continue;
// Make sure all uses of the PHI are simple enough to transform.
if (!LoadUsesSimpleEnoughForHeapSRA(PN,
LoadUsingPHIs, LoadUsingPHIsPerLoad))
return false;
continue;
}
// Otherwise we don't know what this is, not ok.
return false;
}
return true;
}
/// If all users of values loaded from GV are simple enough to perform HeapSRA,
/// return true.
static bool AllGlobalLoadUsesSimpleEnoughForHeapSRA(const GlobalVariable *GV,
Instruction *StoredVal) {
SmallPtrSet<const PHINode*, 32> LoadUsingPHIs;
SmallPtrSet<const PHINode*, 32> LoadUsingPHIsPerLoad;
for (const User *U : GV->users())
if (const LoadInst *LI = dyn_cast<LoadInst>(U)) {
if (!LoadUsesSimpleEnoughForHeapSRA(LI, LoadUsingPHIs,
LoadUsingPHIsPerLoad))
return false;
LoadUsingPHIsPerLoad.clear();
}
// If we reach here, we know that all uses of the loads and transitive uses
// (through PHI nodes) are simple enough to transform. However, we don't know
// that all inputs the to the PHI nodes are in the same equivalence sets.
// Check to verify that all operands of the PHIs are either PHIS that can be
// transformed, loads from GV, or MI itself.
for (const PHINode *PN : LoadUsingPHIs) {
for (unsigned op = 0, e = PN->getNumIncomingValues(); op != e; ++op) {
Value *InVal = PN->getIncomingValue(op);
// PHI of the stored value itself is ok.
if (InVal == StoredVal) continue;
if (const PHINode *InPN = dyn_cast<PHINode>(InVal)) {
// One of the PHIs in our set is (optimistically) ok.
if (LoadUsingPHIs.count(InPN))
continue;
return false;
}
// Load from GV is ok.
if (const LoadInst *LI = dyn_cast<LoadInst>(InVal))
if (LI->getOperand(0) == GV)
continue;
// UNDEF? NULL?
// Anything else is rejected.
return false;
}
} }
return true; return true;
} }
static Value *GetHeapSROAValue(Value *V, unsigned FieldNo,
DenseMap<Value *, std::vector<Value *>> &InsertedScalarizedValues,
std::vector<std::pair<PHINode *, unsigned>> &PHIsToRewrite) {
std::vector<Value *> &FieldVals = InsertedScalarizedValues[V];
if (FieldNo >= FieldVals.size())
FieldVals.resize(FieldNo+1);
// If we already have this value, just reuse the previously scalarized
// version.
if (Value *FieldVal = FieldVals[FieldNo])
return FieldVal;
// Depending on what instruction this is, we have several cases.
Value *Result;
if (LoadInst *LI = dyn_cast<LoadInst>(V)) {
// This is a scalarized version of the load from the global. Just create
// a new Load of the scalarized global.
Value *V = GetHeapSROAValue(LI->getOperand(0), FieldNo,
InsertedScalarizedValues, PHIsToRewrite);
Result = new LoadInst(V->getType()->getPointerElementType(), V,
LI->getName() + ".f" + Twine(FieldNo), LI);
} else {
PHINode *PN = cast<PHINode>(V);
// PN's type is pointer to struct. Make a new PHI of pointer to struct
// field.
PointerType *PTy = cast<PointerType>(PN->getType());
StructType *ST = cast<StructType>(PTy->getElementType());
unsigned AS = PTy->getAddressSpace();
PHINode *NewPN =
PHINode::Create(PointerType::get(ST->getElementType(FieldNo), AS),
PN->getNumIncomingValues(),
PN->getName()+".f"+Twine(FieldNo), PN);
Result = NewPN;
PHIsToRewrite.push_back(std::make_pair(PN, FieldNo));
}
return FieldVals[FieldNo] = Result;
}
/// Given a load instruction and a value derived from the load, rewrite the
/// derived value to use the HeapSRoA'd load.
static void RewriteHeapSROALoadUser(Instruction *LoadUser,
DenseMap<Value *, std::vector<Value *>> &InsertedScalarizedValues,
std::vector<std::pair<PHINode *, unsigned>> &PHIsToRewrite) {
// If this is a comparison against null, handle it.
if (ICmpInst *SCI = dyn_cast<ICmpInst>(LoadUser)) {
assert(isa<ConstantPointerNull>(SCI->getOperand(1)));
// If we have a setcc of the loaded pointer, we can use a setcc of any
// field.
Value *NPtr = GetHeapSROAValue(SCI->getOperand(0), 0,
InsertedScalarizedValues, PHIsToRewrite);
Value *New = new ICmpInst(SCI, SCI->getPredicate(), NPtr,
Constant::getNullValue(NPtr->getType()),
SCI->getName());
SCI->replaceAllUsesWith(New);
SCI->eraseFromParent();
return;
}
// Handle 'getelementptr Ptr, Idx, i32 FieldNo ...'
if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(LoadUser)) {
assert(GEPI->getNumOperands() >= 3 && isa<ConstantInt>(GEPI->getOperand(2))
&& "Unexpected GEPI!");
// Load the pointer for this field.
unsigned FieldNo = cast<ConstantInt>(GEPI->getOperand(2))->getZExtValue();
Value *NewPtr = GetHeapSROAValue(GEPI->getOperand(0), FieldNo,
InsertedScalarizedValues, PHIsToRewrite);
// Create the new GEP idx vector.
SmallVector<Value*, 8> GEPIdx;
GEPIdx.push_back(GEPI->getOperand(1));
GEPIdx.append(GEPI->op_begin()+3, GEPI->op_end());
Value *NGEPI = GetElementPtrInst::Create(GEPI->getResultElementType(), NewPtr, GEPIdx,
GEPI->getName(), GEPI);
GEPI->replaceAllUsesWith(NGEPI);
GEPI->eraseFromParent();
return;
}
// Recursively transform the users of PHI nodes. This will lazily create the
// PHIs that are needed for individual elements. Keep track of what PHIs we
// see in InsertedScalarizedValues so that we don't get infinite loops (very
// antisocial). If the PHI is already in InsertedScalarizedValues, it has
// already been seen first by another load, so its uses have already been
// processed.
PHINode *PN = cast<PHINode>(LoadUser);
if (!InsertedScalarizedValues.insert(std::make_pair(PN,
std::vector<Value *>())).second)
return;
// If this is the first time we've seen this PHI, recursively process all
// users.
for (auto UI = PN->user_begin(), E = PN->user_end(); UI != E;) {
Instruction *User = cast<Instruction>(*UI++);
RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite);
}
}
/// We are performing Heap SRoA on a global. Ptr is a value loaded from the
/// global. Eliminate all uses of Ptr, making them use FieldGlobals instead.
/// All uses of loaded values satisfy AllGlobalLoadUsesSimpleEnoughForHeapSRA.
static void RewriteUsesOfLoadForHeapSRoA(LoadInst *Load,
DenseMap<Value *, std::vector<Value *>> &InsertedScalarizedValues,
std::vector<std::pair<PHINode *, unsigned> > &PHIsToRewrite) {
for (auto UI = Load->user_begin(), E = Load->user_end(); UI != E;) {
Instruction *User = cast<Instruction>(*UI++);
RewriteHeapSROALoadUser(User, InsertedScalarizedValues, PHIsToRewrite);
}
if (Load->use_empty()) {
Load->eraseFromParent();
InsertedScalarizedValues.erase(Load);
}
}
/// CI is an allocation of an array of structures. Break it up into multiple
/// allocations of arrays of the fields.
static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV, CallInst *CI,
Value *NElems, const DataLayout &DL,
const TargetLibraryInfo *TLI) {
LLVM_DEBUG(dbgs() << "SROA HEAP ALLOC: " << *GV << " MALLOC = " << *CI
<< '\n');
Type *MAT = getMallocAllocatedType(CI, TLI);
StructType *STy = cast<StructType>(MAT);
// There is guaranteed to be at least one use of the malloc (storing
// it into GV). If there are other uses, change them to be uses of
// the global to simplify later code. This also deletes the store
// into GV.
ReplaceUsesOfMallocWithGlobal(CI, GV);
// Okay, at this point, there are no users of the malloc. Insert N
// new mallocs at the same place as CI, and N globals.
std::vector<Value *> FieldGlobals;
std::vector<Value *> FieldMallocs;
SmallVector<OperandBundleDef, 1> OpBundles;
CI->getOperandBundlesAsDefs(OpBundles);
unsigned AS = GV->getType()->getPointerAddressSpace();
for (unsigned FieldNo = 0, e = STy->getNumElements(); FieldNo != e;++FieldNo){
Type *FieldTy = STy->getElementType(FieldNo);
PointerType *PFieldTy = PointerType::get(FieldTy, AS);
GlobalVariable *NGV = new GlobalVariable(
*GV->getParent(), PFieldTy, false, GlobalValue::InternalLinkage,
Constant::getNullValue(PFieldTy), GV->getName() + ".f" + Twine(FieldNo),
nullptr, GV->getThreadLocalMode());
NGV->copyAttributesFrom(GV);
FieldGlobals.push_back(NGV);
unsigned TypeSize = DL.getTypeAllocSize(FieldTy);
if (StructType *ST = dyn_cast<StructType>(FieldTy))
TypeSize = DL.getStructLayout(ST)->getSizeInBytes();
Type *IntPtrTy = DL.getIntPtrType(CI->getType());
Value *NMI = CallInst::CreateMalloc(CI, IntPtrTy, FieldTy,
ConstantInt::get(IntPtrTy, TypeSize),
NElems, OpBundles, nullptr,
CI->getName() + ".f" + Twine(FieldNo));
FieldMallocs.push_back(NMI);
new StoreInst(NMI, NGV, CI);
}
// The tricky aspect of this transformation is handling the case when malloc
// fails. In the original code, malloc failing would set the result pointer
// of malloc to null. In this case, some mallocs could succeed and others
// could fail. As such, we emit code that looks like this:
// F0 = malloc(field0)
// F1 = malloc(field1)
// F2 = malloc(field2)
// if (F0 == 0 || F1 == 0 || F2 == 0) {
// if (F0) { free(F0); F0 = 0; }
// if (F1) { free(F1); F1 = 0; }
// if (F2) { free(F2); F2 = 0; }
// }
// The malloc can also fail if its argument is too large.
Constant *ConstantZero = ConstantInt::get(CI->getArgOperand(0)->getType(), 0);
Value *RunningOr = new ICmpInst(CI, ICmpInst::ICMP_SLT, CI->getArgOperand(0),
ConstantZero, "isneg");
for (unsigned i = 0, e = FieldMallocs.size(); i != e; ++i) {
Value *Cond = new ICmpInst(CI, ICmpInst::ICMP_EQ, FieldMallocs[i],
Constant::getNullValue(FieldMallocs[i]->getType()),
"isnull");
RunningOr = BinaryOperator::CreateOr(RunningOr, Cond, "tmp", CI);
}
// Split the basic block at the old malloc.
BasicBlock *OrigBB = CI->getParent();
BasicBlock *ContBB =
OrigBB->splitBasicBlock(CI->getIterator(), "malloc_cont");
// Create the block to check the first condition. Put all these blocks at the
// end of the function as they are unlikely to be executed.
BasicBlock *NullPtrBlock = BasicBlock::Create(OrigBB->getContext(),
"malloc_ret_null",
OrigBB->getParent());
// Remove the uncond branch from OrigBB to ContBB, turning it into a cond
// branch on RunningOr.
OrigBB->getTerminator()->eraseFromParent();
BranchInst::Create(NullPtrBlock, ContBB, RunningOr, OrigBB);
// Within the NullPtrBlock, we need to emit a comparison and branch for each
// pointer, because some may be null while others are not.
for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
Value *GVVal =
new LoadInst(cast<GlobalVariable>(FieldGlobals[i])->getValueType(),
FieldGlobals[i], "tmp", NullPtrBlock);
Value *Cmp = new ICmpInst(*NullPtrBlock, ICmpInst::ICMP_NE, GVVal,
Constant::getNullValue(GVVal->getType()));
BasicBlock *FreeBlock = BasicBlock::Create(Cmp->getContext(), "free_it",
OrigBB->getParent());
BasicBlock *NextBlock = BasicBlock::Create(Cmp->getContext(), "next",
OrigBB->getParent());
Instruction *BI = BranchInst::Create(FreeBlock, NextBlock,
Cmp, NullPtrBlock);
// Fill in FreeBlock.
CallInst::CreateFree(GVVal, OpBundles, BI);
new StoreInst(Constant::getNullValue(GVVal->getType()), FieldGlobals[i],
FreeBlock);
BranchInst::Create(NextBlock, FreeBlock);
NullPtrBlock = NextBlock;
}
BranchInst::Create(ContBB, NullPtrBlock);
// CI is no longer needed, remove it.
CI->eraseFromParent();
/// As we process loads, if we can't immediately update all uses of the load,
/// keep track of what scalarized loads are inserted for a given load.
DenseMap<Value *, std::vector<Value *>> InsertedScalarizedValues;
InsertedScalarizedValues[GV] = FieldGlobals;
std::vector<std::pair<PHINode *, unsigned>> PHIsToRewrite;
// Okay, the malloc site is completely handled. All of the uses of GV are now
// loads, and all uses of those loads are simple. Rewrite them to use loads
// of the per-field globals instead.
for (auto UI = GV->user_begin(), E = GV->user_end(); UI != E;) {
Instruction *User = cast<Instruction>(*UI++);
if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
RewriteUsesOfLoadForHeapSRoA(LI, InsertedScalarizedValues, PHIsToRewrite);
continue;
}
// Must be a store of null.
StoreInst *SI = cast<StoreInst>(User);
assert(isa<ConstantPointerNull>(SI->getOperand(0)) &&
"Unexpected heap-sra user!");
// Insert a store of null into each global.
for (unsigned i = 0, e = FieldGlobals.size(); i != e; ++i) {
Type *ValTy = cast<GlobalValue>(FieldGlobals[i])->getValueType();
Constant *Null = Constant::getNullValue(ValTy);
new StoreInst(Null, FieldGlobals[i], SI);
}
// Erase the original store.
SI->eraseFromParent();
}
// While we have PHIs that are interesting to rewrite, do it.
while (!PHIsToRewrite.empty()) {
PHINode *PN = PHIsToRewrite.back().first;
unsigned FieldNo = PHIsToRewrite.back().second;
PHIsToRewrite.pop_back();
PHINode *FieldPN = cast<PHINode>(InsertedScalarizedValues[PN][FieldNo]);
assert(FieldPN->getNumIncomingValues() == 0 &&"Already processed this phi");
// Add all the incoming values. This can materialize more phis.
for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
Value *InVal = PN->getIncomingValue(i);
InVal = GetHeapSROAValue(InVal, FieldNo, InsertedScalarizedValues,
PHIsToRewrite);
FieldPN->addIncoming(InVal, PN->getIncomingBlock(i));
}
}
// Drop all inter-phi links and any loads that made it this far.
for (DenseMap<Value *, std::vector<Value *>>::iterator
I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end();
I != E; ++I) {
if (PHINode *PN = dyn_cast<PHINode>(I->first))
PN->dropAllReferences();
else if (LoadInst *LI = dyn_cast<LoadInst>(I->first))
LI->dropAllReferences();
}
// Delete all the phis and loads now that inter-references are dead.
for (DenseMap<Value *, std::vector<Value *>>::iterator
I = InsertedScalarizedValues.begin(), E = InsertedScalarizedValues.end();
I != E; ++I) {
if (PHINode *PN = dyn_cast<PHINode>(I->first))
PN->eraseFromParent();
else if (LoadInst *LI = dyn_cast<LoadInst>(I->first))
LI->eraseFromParent();
}
// The old global is now dead, remove it.
GV->eraseFromParent();
++NumHeapSRA;
return cast<GlobalVariable>(FieldGlobals[0]);
}
/// This function is called when we see a pointer global variable with a single /// This function is called when we see a pointer global variable with a single
/// value stored it that is a malloc or cast of malloc. /// value stored it that is a malloc or cast of malloc.
static bool tryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV, CallInst *CI, static bool tryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV, CallInst *CI,
Type *AllocTy, Type *AllocTy,
AtomicOrdering Ordering, AtomicOrdering Ordering,
const DataLayout &DL, const DataLayout &DL,
TargetLibraryInfo *TLI) { TargetLibraryInfo *TLI) {
// If this is a malloc of an abstract type, don't touch it. // If this is a malloc of an abstract type, don't touch it.
if (!AllocTy->isSized()) if (!AllocTy->isSized())
return false; return false;
// We can't optimize this global unless all uses of it are *known* to be // We can't optimize this global unless all uses of it are *known* to be
// of the malloc value, not of the null initializer value (consider a use // of the malloc value, not of the null initializer value (consider a use
// that compares the global's value against zero to see if the malloc has // that compares the global's value against zero to see if the malloc has
// been reached). To do this, we check to see if all uses of the global // been reached). To do this, we check to see if all uses of the global
// would trap if the global were null: this proves that they must all // would trap if the global were null: this proves that they must all
// happen after the malloc. // happen after the malloc.
if (!AllUsesOfLoadedValueWillTrapIfNull(GV)) if (!AllUsesOfLoadedValueWillTrapIfNull(GV))
return false; return false;
// We can't optimize this if the malloc itself is used in a complex way, // We can't optimize this if the malloc itself is used in a complex way,
// for example, being stored into multiple globals. This allows the // for example, being stored into multiple globals. This allows the
// malloc to be stored into the specified global, loaded icmp'd, and // malloc to be stored into the specified global, loaded icmp'd.
// GEP'd. These are all things we could transform to using the global // These are all things we could transform to using the global for.
// for. if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(CI, GV))
SmallPtrSet<const PHINode*, 8> PHIs;
if (!ValueIsOnlyUsedLocallyOrStoredToOneGlobal(CI, GV, PHIs))
return false; return false;
// If we have a global that is only initialized with a fixed size malloc, // If we have a global that is only initialized with a fixed size malloc,
// transform the program to use global memory instead of malloc'd memory. // transform the program to use global memory instead of malloc'd memory.
// This eliminates dynamic allocation, avoids an indirection accessing the // This eliminates dynamic allocation, avoids an indirection accessing the
// data, and exposes the resultant global to further GlobalOpt. // data, and exposes the resultant global to further GlobalOpt.
// We cannot optimize the malloc if we cannot determine malloc array size. // We cannot optimize the malloc if we cannot determine malloc array size.
Value *NElems = getMallocArraySize(CI, DL, TLI, true); Value *NElems = getMallocArraySize(CI, DL, TLI, true);
if (!NElems) if (!NElems)
return false; return false;
if (ConstantInt *NElements = dyn_cast<ConstantInt>(NElems)) if (ConstantInt *NElements = dyn_cast<ConstantInt>(NElems))
// Restrict this transformation to only working on small allocations // Restrict this transformation to only working on small allocations
// (2048 bytes currently), as we don't want to introduce a 16M global or // (2048 bytes currently), as we don't want to introduce a 16M global or
// something. // something.
if (NElements->getZExtValue() * DL.getTypeAllocSize(AllocTy) < 2048) { if (NElements->getZExtValue() * DL.getTypeAllocSize(AllocTy) < 2048) {
OptimizeGlobalAddressOfMalloc(GV, CI, AllocTy, NElements, DL, TLI); OptimizeGlobalAddressOfMalloc(GV, CI, AllocTy, NElements, DL, TLI);
return true; return true;
} }
// If the allocation is an array of structures, consider transforming this
// into multiple malloc'd arrays, one for each field. This is basically
// SRoA for malloc'd memory.
if (Ordering != AtomicOrdering::NotAtomic)
return false;
// If this is an allocation of a fixed size array of structs, analyze as a
// variable size array. malloc [100 x struct],1 -> malloc struct, 100
if (NElems == ConstantInt::get(CI->getArgOperand(0)->getType(), 1))
if (ArrayType *AT = dyn_cast<ArrayType>(AllocTy))
AllocTy = AT->getElementType();
StructType *AllocSTy = dyn_cast<StructType>(AllocTy);
if (!AllocSTy)
return false;
// This the structure has an unreasonable number of fields, leave it
// alone.
if (AllocSTy->getNumElements() <= 16 && AllocSTy->getNumElements() != 0 &&
AllGlobalLoadUsesSimpleEnoughForHeapSRA(GV, CI)) {
// If this is a fixed size array, transform the Malloc to be an alloc of
// structs. malloc [100 x struct],1 -> malloc struct, 100
if (ArrayType *AT = dyn_cast<ArrayType>(getMallocAllocatedType(CI, TLI))) {
Type *IntPtrTy = DL.getIntPtrType(CI->getType());
unsigned TypeSize = DL.getStructLayout(AllocSTy)->getSizeInBytes();
Value *AllocSize = ConstantInt::get(IntPtrTy, TypeSize);
Value *NumElements = ConstantInt::get(IntPtrTy, AT->getNumElements());
SmallVector<OperandBundleDef, 1> OpBundles;
CI->getOperandBundlesAsDefs(OpBundles);
Instruction *Malloc =
CallInst::CreateMalloc(CI, IntPtrTy, AllocSTy, AllocSize, NumElements,
OpBundles, nullptr, CI->getName());
Instruction *Cast = new BitCastInst(Malloc, CI->getType(), "tmp", CI);
CI->replaceAllUsesWith(Cast);
CI->eraseFromParent();
if (BitCastInst *BCI = dyn_cast<BitCastInst>(Malloc))
CI = cast<CallInst>(BCI->getOperand(0));
else
CI = cast<CallInst>(Malloc);
}
PerformHeapAllocSRoA(GV, CI, getMallocArraySize(CI, DL, TLI, true), DL,
TLI);
return true;
}
return false; return false;
} }
// Try to optimize globals based on the knowledge that only one value (besides // Try to optimize globals based on the knowledge that only one value (besides
// its initializer) is ever stored to the global. // its initializer) is ever stored to the global.
static bool static bool
optimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal, optimizeOnceStoredGlobal(GlobalVariable *GV, Value *StoredOnceVal,
AtomicOrdering Ordering, const DataLayout &DL, AtomicOrdering Ordering, const DataLayout &DL,
▲ Show 20 Lines • Show All 1,612 Lines • Show Last 20 Lines

llvm/test/Transforms/GlobalOpt/MallocSROA-section.ll

; RUN: opt -globalopt -S < %s | FileCheck %s ; RUN: opt -globalopt -S < %s | FileCheck %s
; CHECK: @Y.f0 ; CHECK: @Y = {{.*}} section ".foo"
; CHECK: section ".foo"
; CHECK: @Y.f1
; CHECK: section ".foo"
%struct.xyz = type { double, i32 } %struct.xyz = type { double, i32 }
@Y = internal global %struct.xyz* null ,section ".foo" ; <%struct.xyz**> [#uses=2] @Y = internal global %struct.xyz* null ,section ".foo" ; <%struct.xyz**> [#uses=2]
@numf2s = external global i32 ; <i32*> [#uses=1] @numf2s = external global i32 ; <i32*> [#uses=1]
define void @init_net() { define void @init_net() {
entry: entry:
Show All 15 Lines

llvm/test/Transforms/GlobalOpt/heap-sra-1.ll

; RUN: opt < %s -globalopt -S | FileCheck %s ; RUN: opt < %s -globalopt -S | FileCheck %s
target datalayout = "E-p:64:64:64-a0:0:8-f32:32:32-f64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-v64:64:64-v128:128:128" target datalayout = "E-p:64:64:64-a0:0:8-f32:32:32-f64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-v64:64:64-v128:128:128"
;; Heap SROA has been removed. This tests we don't perform heap SROA.
; CHECK: @X =
%struct.foo = type { i32, i32 } %struct.foo = type { i32, i32 }
@X = internal global %struct.foo* null @X = internal global %struct.foo* null
; CHECK: @X.f0
; CHECK: @X.f1
define void @bar(i64 %Size) nounwind noinline { define void @bar(i64 %Size) nounwind noinline {
entry: entry:
%mallocsize = mul i64 %Size, 8 ; <i64> [#uses=1] %mallocsize = mul i64 %Size, 8 ; <i64> [#uses=1]
%malloccall = tail call i8* @malloc(i64 %mallocsize) ; <i8*> [#uses=1] %malloccall = tail call i8* @malloc(i64 %mallocsize) ; <i8*> [#uses=1]
%.sub = bitcast i8* %malloccall to %struct.foo* ; <%struct.foo*> [#uses=1] %.sub = bitcast i8* %malloccall to %struct.foo* ; <%struct.foo*> [#uses=1]
store %struct.foo* %.sub, %struct.foo** @X, align 4 store %struct.foo* %.sub, %struct.foo** @X, align 4
ret void ret void
Show All 30 Lines

llvm/test/Transforms/GlobalOpt/heap-sra-2.ll

; RUN: opt < %s -globalopt -S | FileCheck %s ; RUN: opt < %s -globalopt -S | FileCheck %s
target datalayout = "E-p:64:64:64-a0:0:8-f32:32:32-f64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-v64:64:64-v128:128:128" target datalayout = "E-p:64:64:64-a0:0:8-f32:32:32-f64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-v64:64:64-v128:128:128"
;; Heap SROA has been removed. This tests we don't perform heap SROA.
; CHECK: @X =
%struct.foo = type { i32, i32 } %struct.foo = type { i32, i32 }
@X = internal global %struct.foo* null ; <%struct.foo**> [#uses=2] @X = internal global %struct.foo* null ; <%struct.foo**> [#uses=2]
; CHECK: @X.f0
; CHECK: @X.f1
define void @bar(i32 %Size) nounwind noinline { define void @bar(i32 %Size) nounwind noinline {
entry: entry:
%malloccall = tail call i8* @malloc(i64 8000000) ; <i8*> [#uses=1] %malloccall = tail call i8* @malloc(i64 8000000) ; <i8*> [#uses=1]
%0 = bitcast i8* %malloccall to [1000000 x %struct.foo]* ; <[1000000 x %struct.foo]*> [#uses=1] %0 = bitcast i8* %malloccall to [1000000 x %struct.foo]* ; <[1000000 x %struct.foo]*> [#uses=1]
%.sub = getelementptr [1000000 x %struct.foo], [1000000 x %struct.foo]* %0, i32 0, i32 0 ; <%struct.foo*> [#uses=1] %.sub = getelementptr [1000000 x %struct.foo], [1000000 x %struct.foo]* %0, i32 0, i32 0 ; <%struct.foo*> [#uses=1]
store %struct.foo* %.sub, %struct.foo** @X, align 4 store %struct.foo* %.sub, %struct.foo** @X, align 4
ret void ret void
Show All 30 Lines

llvm/test/Transforms/GlobalOpt/heap-sra-3-no-null-opt.ll

  • This file was deleted.
; RUN: opt < %s -globalopt -S | FileCheck %s
target datalayout = "E-p:64:64:64-a0:0:8-f32:32:32-f64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-v64:64:64-v128:128:128"
%struct.foo = type { i32, i32 }
@X = internal global %struct.foo* null
; CHECK: @X
; CHECK-NOT: @X.f0
define void @bar(i64 %Size) nounwind noinline #0 {
entry:
%mallocsize = mul i64 8, %Size ; <i64> [#uses=1]
; CHECK: mul i64 8, %Size
%malloccall = tail call i8* @malloc(i64 %mallocsize) ; <i8*> [#uses=1]
%.sub = bitcast i8* %malloccall to %struct.foo* ; <%struct.foo*> [#uses=1]
store %struct.foo* %.sub, %struct.foo** @X, align 4
ret void
}
declare noalias i8* @malloc(i64)
define i32 @baz() nounwind readonly noinline #0 {
bb1.thread:
; CHECK: load %struct.foo*, %struct.foo** @X, align 4
%0 = load %struct.foo*, %struct.foo** @X, align 4
br label %bb1
bb1: ; preds = %bb1, %bb1.thread
%i.0.reg2mem.0 = phi i32 [ 0, %bb1.thread ], [ %indvar.next, %bb1 ]
%sum.0.reg2mem.0 = phi i32 [ 0, %bb1.thread ], [ %3, %bb1 ]
%1 = getelementptr %struct.foo, %struct.foo* %0, i32 %i.0.reg2mem.0, i32 0
%2 = load i32, i32* %1, align 4
%3 = add i32 %2, %sum.0.reg2mem.0
%indvar.next = add i32 %i.0.reg2mem.0, 1
%exitcond = icmp eq i32 %indvar.next, 1200
br i1 %exitcond, label %bb2, label %bb1
bb2: ; preds = %bb1
ret i32 %3
}
attributes #0 = { null_pointer_is_valid }

llvm/test/Transforms/GlobalOpt/heap-sra-3.ll

  • This file was deleted.
; RUN: opt < %s -globalopt -S | FileCheck %s
target datalayout = "E-p:64:64:64-a0:0:8-f32:32:32-f64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-v64:64:64-v128:128:128"
%struct.foo = type { i32, i32 }
@X = internal global %struct.foo* null
; CHECK: @X.f0
; CHECK: @X.f1
define void @bar(i64 %Size) nounwind noinline {
entry:
%mallocsize = mul i64 8, %Size ; <i64> [#uses=1]
; CHECK: mul i64 %Size, 4
%malloccall = tail call i8* @malloc(i64 %mallocsize) ; <i8*> [#uses=1]
%.sub = bitcast i8* %malloccall to %struct.foo* ; <%struct.foo*> [#uses=1]
store %struct.foo* %.sub, %struct.foo** @X, align 4
ret void
}
declare noalias i8* @malloc(i64)
define i32 @baz() nounwind readonly noinline {
bb1.thread:
%0 = load %struct.foo*, %struct.foo** @X, align 4
br label %bb1
bb1: ; preds = %bb1, %bb1.thread
%i.0.reg2mem.0 = phi i32 [ 0, %bb1.thread ], [ %indvar.next, %bb1 ]
%sum.0.reg2mem.0 = phi i32 [ 0, %bb1.thread ], [ %3, %bb1 ]
%1 = getelementptr %struct.foo, %struct.foo* %0, i32 %i.0.reg2mem.0, i32 0
%2 = load i32, i32* %1, align 4
%3 = add i32 %2, %sum.0.reg2mem.0
%indvar.next = add i32 %i.0.reg2mem.0, 1
%exitcond = icmp eq i32 %indvar.next, 1200
br i1 %exitcond, label %bb2, label %bb1
bb2: ; preds = %bb1
ret i32 %3
}
define void @bam(i64 %Size) nounwind noinline #0 {
entry:
%0 = load %struct.foo*, %struct.foo** @X, align 4
ret void
}
attributes #0 = { null_pointer_is_valid }

llvm/test/Transforms/GlobalOpt/heap-sra-4-no-null-opt.ll

  • This file was deleted.
; RUN: opt < %s -globalopt -S | FileCheck %s
target datalayout = "E-p:64:64:64-a0:0:8-f32:32:32-f64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-v64:64:64-v128:128:128"
%struct.foo = type { i32, i32 }
@X = internal global %struct.foo* null
; CHECK: @X
; CHECK-NOT: @X.f0
define void @bar(i64 %Size) nounwind noinline #0 {
entry:
%mallocsize = shl i64 %Size, 3 ; <i64> [#uses=1]
%malloccall = tail call i8* @malloc(i64 %mallocsize) ; <i8*> [#uses=1]
; CHECK: shl i64 %Size, 3
%.sub = bitcast i8* %malloccall to %struct.foo* ; <%struct.foo*> [#uses=1]
store %struct.foo* %.sub, %struct.foo** @X, align 4
ret void
}
declare noalias i8* @malloc(i64)
define i32 @baz() nounwind readonly noinline #0 {
; CHECK-LABEL: @baz(
bb1.thread:
; CHECK: load %struct.foo*, %struct.foo** @X, align 4
%0 = load %struct.foo*, %struct.foo** @X, align 4
br label %bb1
bb1: ; preds = %bb1, %bb1.thread
%i.0.reg2mem.0 = phi i32 [ 0, %bb1.thread ], [ %indvar.next, %bb1 ]
%sum.0.reg2mem.0 = phi i32 [ 0, %bb1.thread ], [ %3, %bb1 ]
%1 = getelementptr %struct.foo, %struct.foo* %0, i32 %i.0.reg2mem.0, i32 0
%2 = load i32, i32* %1, align 4
; CHECK: load i32, i32* %1, align 4
%3 = add i32 %2, %sum.0.reg2mem.0
%indvar.next = add i32 %i.0.reg2mem.0, 1
%exitcond = icmp eq i32 %indvar.next, 1200
br i1 %exitcond, label %bb2, label %bb1
bb2: ; preds = %bb1
ret i32 %3
}
attributes #0 = { null_pointer_is_valid }

llvm/test/Transforms/GlobalOpt/heap-sra-4.ll

  • This file was deleted.
; RUN: opt < %s -globalopt -S | FileCheck %s
target datalayout = "E-p:64:64:64-a0:0:8-f32:32:32-f64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-v64:64:64-v128:128:128"
%struct.foo = type { i32, i32 }
@X = internal global %struct.foo* null
; CHECK: @X.f0
; CHECK: @X.f1
define void @bar(i64 %Size) nounwind noinline {
entry:
%mallocsize = shl i64 %Size, 3 ; <i64> [#uses=1]
%malloccall = tail call i8* @malloc(i64 %mallocsize) ; <i8*> [#uses=1]
; CHECK: mul i64 %Size, 4
%.sub = bitcast i8* %malloccall to %struct.foo* ; <%struct.foo*> [#uses=1]
store %struct.foo* %.sub, %struct.foo** @X, align 4
ret void
}
declare noalias i8* @malloc(i64)
define i32 @baz() nounwind readonly noinline {
bb1.thread:
%0 = load %struct.foo*, %struct.foo** @X, align 4
br label %bb1
bb1: ; preds = %bb1, %bb1.thread
%i.0.reg2mem.0 = phi i32 [ 0, %bb1.thread ], [ %indvar.next, %bb1 ]
%sum.0.reg2mem.0 = phi i32 [ 0, %bb1.thread ], [ %3, %bb1 ]
%1 = getelementptr %struct.foo, %struct.foo* %0, i32 %i.0.reg2mem.0, i32 0
%2 = load i32, i32* %1, align 4
%3 = add i32 %2, %sum.0.reg2mem.0
%indvar.next = add i32 %i.0.reg2mem.0, 1
%exitcond = icmp eq i32 %indvar.next, 1200
br i1 %exitcond, label %bb2, label %bb1
bb2: ; preds = %bb1
ret i32 %3
}
define void @bam(i64 %Size) nounwind noinline #0 {
entry:
%0 = load %struct.foo*, %struct.foo** @X, align 4
ret void
}
attributes #0 = { null_pointer_is_valid }

llvm/test/Transforms/GlobalOpt/heap-sra-phi.ll

; RUN: opt < %s -globalopt -S | FileCheck %s ; RUN: opt < %s -globalopt -S | FileCheck %s
; CHECK: tmp.f1 = phi i32*
; CHECK: tmp.f0 = phi i32*
target datalayout = "E-p:64:64:64-a0:0:8-f32:32:32-f64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-v64:64:64-v128:128:128" target datalayout = "E-p:64:64:64-a0:0:8-f32:32:32-f64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-v64:64:64-v128:128:128"
;; Heap SROA has been removed. This tests we don't perform heap SROA.
; CHECK: @X =
%struct.foo = type { i32, i32 } %struct.foo = type { i32, i32 }
@X = internal global %struct.foo* null ; <%struct.foo**> [#uses=2] @X = internal global %struct.foo* null ; <%struct.foo**> [#uses=2]
define void @bar(i32 %Size) nounwind noinline { define void @bar(i32 %Size) nounwind noinline {
entry: entry:
%malloccall = tail call i8* @malloc(i64 8000000) ; <i8*> [#uses=1] %malloccall = tail call i8* @malloc(i64 8000000) ; <i8*> [#uses=1]
%tmp = bitcast i8* %malloccall to [1000000 x %struct.foo]* ; <[1000000 x %struct.foo]*> [#uses=1] %tmp = bitcast i8* %malloccall to [1000000 x %struct.foo]* ; <[1000000 x %struct.foo]*> [#uses=1]
%.sub = getelementptr [1000000 x %struct.foo], [1000000 x %struct.foo]* %tmp, i32 0, i32 0 ; <%struct.foo*> [#uses=1] %.sub = getelementptr [1000000 x %struct.foo], [1000000 x %struct.foo]* %tmp, i32 0, i32 0 ; <%struct.foo*> [#uses=1]
Show All 39 Lines
llvm/test/Transforms/GlobalOpt/heap-sra-3-no-null-opt.ll