Skip to content

Commit

Permalink
Reapply [ConstantFold] Remove non-trivial gep-of-gep fold (#93823)
Browse files Browse the repository at this point in the history 
Reapply after #93956, which
changed clang array initialization codegen to avoid size regressions
for unoptimized builds.

-----

This fold is subtly incorrect, because DL-unaware constant folding does
not know the correct index type to use, and just performs the addition
in the type that happens to already be there. This is incorrect, since
sext(X)+sext(Y) is generally not the same as sext(X+Y). See the
`@constexpr_gep_of_gep_with_narrow_type()` for a miscompile with the
current implementation.

One could try to restrict the fold to cases where no overflow occurs,
but I'm not bothering with that here, because the DL-aware constant
folding will take care of this anyway. I've only kept the
straightforward zero-index case, where we just concatenate two GEPs.
  • Loading branch information
@nikic
nikic committed Jun 10, 2024
1 parent e58f830 commit cc158d4
Show file tree
Hide file tree
Showing 5 changed files with 15 additions and 64 deletions.
56 changes: 4 additions & 52 deletions llvm/lib/IR/ConstantFold.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -1416,64 +1416,16 @@ static Constant *foldGEPOfGEP(GEPOperator *GEP, Type *PointeeTy, bool InBounds,
if (GEP->getInRange())
return nullptr;

// Only handle simple case with leading zero index. We cannot perform an
// actual addition as we don't know the correct index type size to use.
Constant *Idx0 = cast<Constant>(Idxs[0]);
if (Idx0->isNullValue()) {
// Handle the simple case of a zero index.
SmallVector<Value*, 16> NewIndices;
NewIndices.reserve(Idxs.size() + GEP->getNumIndices());
NewIndices.append(GEP->idx_begin(), GEP->idx_end());
NewIndices.append(Idxs.begin() + 1, Idxs.end());
return ConstantExpr::getGetElementPtr(
GEP->getSourceElementType(), cast<Constant>(GEP->getPointerOperand()),
NewIndices, InBounds && GEP->isInBounds());
}

gep_type_iterator LastI = gep_type_end(GEP);
for (gep_type_iterator I = gep_type_begin(GEP), E = gep_type_end(GEP);
I != E; ++I)
LastI = I;

// We can't combine GEPs if the last index is a struct type.
if (!LastI.isSequential())
return nullptr;
// We could perform the transform with non-constant index, but prefer leaving
// it as GEP of GEP rather than GEP of add for now.
ConstantInt *CI = dyn_cast<ConstantInt>(Idx0);
if (!CI)
return nullptr;

// TODO: This code may be extended to handle vectors as well.
auto *LastIdx = cast<Constant>(GEP->getOperand(GEP->getNumOperands()-1));
Type *LastIdxTy = LastIdx->getType();
if (LastIdxTy->isVectorTy())
if (!Idx0->isNullValue())
return nullptr;

SmallVector<Value*, 16> NewIndices;
NewIndices.reserve(Idxs.size() + GEP->getNumIndices());
NewIndices.append(GEP->idx_begin(), GEP->idx_end() - 1);

// Add the last index of the source with the first index of the new GEP.
// Make sure to handle the case when they are actually different types.
if (LastIdxTy != Idx0->getType()) {
unsigned CommonExtendedWidth =
std::max(LastIdxTy->getIntegerBitWidth(),
Idx0->getType()->getIntegerBitWidth());
CommonExtendedWidth = std::max(CommonExtendedWidth, 64U);

Type *CommonTy =
Type::getIntNTy(LastIdxTy->getContext(), CommonExtendedWidth);
if (Idx0->getType() != CommonTy)
Idx0 = ConstantFoldCastInstruction(Instruction::SExt, Idx0, CommonTy);
if (LastIdx->getType() != CommonTy)
LastIdx =
ConstantFoldCastInstruction(Instruction::SExt, LastIdx, CommonTy);
if (!Idx0 || !LastIdx)
return nullptr;
}

NewIndices.push_back(ConstantExpr::get(Instruction::Add, Idx0, LastIdx));
NewIndices.append(GEP->idx_begin(), GEP->idx_end());
NewIndices.append(Idxs.begin() + 1, Idxs.end());

return ConstantExpr::getGetElementPtr(
GEP->getSourceElementType(), cast<Constant>(GEP->getPointerOperand()),
NewIndices, InBounds && GEP->isInBounds());
Expand Down
2 changes: 1 addition & 1 deletion llvm/test/Other/constant-fold-gep.ll
View file
Original file line number Diff line number Diff line change
Expand Up @@ -104,7 +104,7 @@

; Fold GEP of a GEP. Very simple cases are folded without targetdata.

; PLAIN: @Y = global ptr getelementptr inbounds ([3 x { i32, i32 }], ptr @ext, i64 2)
; PLAIN: @Y = global ptr getelementptr inbounds ([3 x { i32, i32 }], ptr getelementptr inbounds ([3 x { i32, i32 }], ptr @ext, i64 1), i64 1)
; PLAIN: @Z = global ptr getelementptr inbounds (i32, ptr getelementptr inbounds ([3 x { i32, i32 }], ptr @ext, i64 0, i64 1, i32 0), i64 1)
; OPT: @Y = local_unnamed_addr global ptr getelementptr inbounds (i8, ptr @ext, i64 48)
; OPT: @Z = local_unnamed_addr global ptr getelementptr inbounds (i8, ptr @ext, i64 12)
Expand Down
2 changes: 1 addition & 1 deletion llvm/test/Transforms/InstCombine/gepgep.ll
View file
Original file line number Diff line number Diff line change
Expand Up @@ -10,7 +10,7 @@ declare void @use(ptr)

define void @f() {
; CHECK-LABEL: define void @f() {
; CHECK-NEXT: call void @use(ptr getelementptr (i8, ptr @buffer, i64 add (i64 sub (i64 0, i64 ptrtoint (ptr @buffer to i64)), i64 127)))
; CHECK-NEXT: call void @use(ptr getelementptr (i8, ptr getelementptr (i8, ptr @buffer, i64 add (i64 sub (i64 0, i64 ptrtoint (ptr @buffer to i64)), i64 63)), i64 64))
; CHECK-NEXT: ret void
;
call void @use(ptr getelementptr (i8, ptr getelementptr (i8, ptr @buffer, i64 add (i64 sub (i64 0, i64 ptrtoint (ptr @buffer to i64)), i64 63)), i64 64))
Expand Down
3 changes: 1 addition & 2 deletions llvm/test/Transforms/InstCombine/getelementptr.ll
View file
Original file line number Diff line number Diff line change
Expand Up @@ -1716,10 +1716,9 @@ if.else:

@g = external global i8

; FIXME: This is a miscompile
define ptr @constexpr_gep_of_gep_with_narrow_type() {
; CHECK-LABEL: @constexpr_gep_of_gep_with_narrow_type(
; CHECK-NEXT: ret ptr getelementptr (i8, ptr @g, i64 -2)
; CHECK-NEXT: ret ptr getelementptr (i8, ptr @g, i64 254)
;
ret ptr getelementptr (i8, ptr getelementptr (i8, ptr @g, i8 127), i8 127)
}
Expand Down
16 changes: 8 additions & 8 deletions llvm/test/Transforms/InstCombine/ptrtoint-nullgep.ll
View file
Original file line number Diff line number Diff line change
Expand Up @@ -68,10 +68,10 @@ define void @constant_fold_ptrtoint_of_gep_of_nullgep() {
; LLPARSER-NEXT: call void @use_i64(i64 ptrtoint (ptr addrspace(1) getelementptr inbounds (i8, ptr addrspace(1) null, i64 1234) to i64))
; LLPARSER-NEXT: call void @use_i64(i64 ptrtoint (ptr addrspace(1) getelementptr (i8, ptr addrspace(1) null, i64 1234) to i64))
; LLPARSER-NEXT: call void @use_i64(i64 ptrtoint (ptr addrspace(1) getelementptr (i8, ptr addrspace(1) null, i64 1234) to i64))
; LLPARSER-NEXT: call void @use_i64(i64 0)
; LLPARSER-NEXT: call void @use_i64(i64 0)
; LLPARSER-NEXT: call void @use_i64(i64 0)
; LLPARSER-NEXT: call void @use_i64(i64 0)
; LLPARSER-NEXT: call void @use_i64(i64 ptrtoint (ptr addrspace(1) getelementptr inbounds (i8, ptr addrspace(1) getelementptr inbounds (i8, ptr addrspace(1) null, i64 -1), i64 1) to i64))
; LLPARSER-NEXT: call void @use_i64(i64 ptrtoint (ptr addrspace(1) getelementptr (i8, ptr addrspace(1) getelementptr inbounds (i8, ptr addrspace(1) null, i64 -1), i64 1) to i64))
; LLPARSER-NEXT: call void @use_i64(i64 ptrtoint (ptr addrspace(1) getelementptr inbounds (i8, ptr addrspace(1) getelementptr (i8, ptr addrspace(1) null, i64 -1), i64 1) to i64))
; LLPARSER-NEXT: call void @use_i64(i64 ptrtoint (ptr addrspace(1) getelementptr (i8, ptr addrspace(1) getelementptr (i8, ptr addrspace(1) null, i64 -1), i64 1) to i64))
; LLPARSER-NEXT: ret void
;
; INSTSIMPLIFY-LABEL: define {{[^@]+}}@constant_fold_ptrtoint_of_gep_of_nullgep() {
Expand All @@ -83,10 +83,10 @@ define void @constant_fold_ptrtoint_of_gep_of_nullgep() {
; INSTSIMPLIFY-NEXT: call void @use_i64(i64 ptrtoint (ptr addrspace(1) getelementptr inbounds (i8, ptr addrspace(1) null, i64 1234) to i64))
; INSTSIMPLIFY-NEXT: call void @use_i64(i64 ptrtoint (ptr addrspace(1) getelementptr (i8, ptr addrspace(1) null, i64 1234) to i64))
; INSTSIMPLIFY-NEXT: call void @use_i64(i64 ptrtoint (ptr addrspace(1) getelementptr (i8, ptr addrspace(1) null, i64 1234) to i64))
; INSTSIMPLIFY-NEXT: call void @use_i64(i64 0)
; INSTSIMPLIFY-NEXT: call void @use_i64(i64 0)
; INSTSIMPLIFY-NEXT: call void @use_i64(i64 0)
; INSTSIMPLIFY-NEXT: call void @use_i64(i64 0)
; INSTSIMPLIFY-NEXT: call void @use_i64(i64 ptrtoint (ptr addrspace(1) getelementptr inbounds (i8, ptr addrspace(1) getelementptr inbounds (i8, ptr addrspace(1) null, i64 -1), i64 1) to i64))
; INSTSIMPLIFY-NEXT: call void @use_i64(i64 ptrtoint (ptr addrspace(1) getelementptr (i8, ptr addrspace(1) getelementptr inbounds (i8, ptr addrspace(1) null, i64 -1), i64 1) to i64))
; INSTSIMPLIFY-NEXT: call void @use_i64(i64 ptrtoint (ptr addrspace(1) getelementptr inbounds (i8, ptr addrspace(1) getelementptr (i8, ptr addrspace(1) null, i64 -1), i64 1) to i64))
; INSTSIMPLIFY-NEXT: call void @use_i64(i64 ptrtoint (ptr addrspace(1) getelementptr (i8, ptr addrspace(1) getelementptr (i8, ptr addrspace(1) null, i64 -1), i64 1) to i64))
; INSTSIMPLIFY-NEXT: ret void
;
; INSTCOMBINE-LABEL: define {{[^@]+}}@constant_fold_ptrtoint_of_gep_of_nullgep() {
Expand Down

0 comments on commit cc158d4

Please sign in to comment.