forked from OSchip/llvm-project
Fix a bug where we would 'promote' an allocation from one type to another
where the second has less alignment required. If we had explicit alignment support in the IR, we could handle this case, but we can't until we do. llvm-svn: 23960
This commit is contained in:
parent
ac87beb03a
commit
355ecc09f8
llvm/lib/Transforms/Scalar
|
@ -3795,10 +3795,14 @@ Instruction *InstCombiner::PromoteCastOfAllocation(CastInst &CI,
|
|||
const Type *AllocElTy = AI.getAllocatedType();
|
||||
const Type *CastElTy = PTy->getElementType();
|
||||
if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
|
||||
|
||||
|
||||
unsigned AllocElTyAlign = TD->getTypeSize(AllocElTy);
|
||||
unsigned CastElTyAlign = TD->getTypeSize(CastElTy);
|
||||
if (CastElTyAlign < AllocElTyAlign) return 0;
|
||||
|
||||
uint64_t AllocElTySize = TD->getTypeSize(AllocElTy);
|
||||
uint64_t CastElTySize = TD->getTypeSize(CastElTy);
|
||||
|
||||
|
||||
// If the allocation is for an even multiple of the cast type size
|
||||
if (CastElTySize == 0 || AllocElTySize % CastElTySize != 0)
|
||||
return 0;
|
||||
|
|
Loading…
Reference in New Issue