Skip to content

Commit 7fb66d4

Browse files
frasercrmckfrasercrmck
committed
[MemCpyOpt] Fix a variety of scalable-type crashes
This patch fixes a variety of crashes resulting from the `MemCpyOptPass` casting `TypeSize` to a constant integer, whether implicitly or explicitly. Since the `MemsetRanges` requires a constant size to work, all but one of the fixes in this patch simply involve skipping the various optimizations for scalable types as cleanly as possible. The optimization of `byval` parameters, however, has been updated to work on scalable types in theory. In practice, this optimization is only valid when the length of the `memcpy` is known to be larger than the scalable type size, which is currently never the case. This could perhaps be done in the future using the `vscale_range` attribute. Some implicit casts have been left as they were, under the knowledge they are only called on aggregate types. These should never be scalably-sized. Reviewed By: nikic, tra Differential Revision: https://reviews.llvm.org/D109329
1 parent caabf2a commit 7fb66d4

File tree

3 files changed

+128
-11
lines changed

3 files changed

+128
-11
lines changed

‎llvm/include/llvm/Transforms/Scalar/MemCpyOptimizer.h‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -62,7 +62,7 @@ class MemCpyOptPass : public PassInfoMixin<MemCpyOptPass> {
6262
bool processMemCpy(MemCpyInst *M, BasicBlock::iterator &BBI);
6363
bool processMemMove(MemMoveInst *M);
6464
bool performCallSlotOptzn(Instruction *cpyLoad, Instruction *cpyStore,
65-
Value *cpyDst, Value *cpySrc, uint64_t cpyLen,
65+
Value *cpyDst, Value *cpySrc, TypeSize cpyLen,
6666
Align cpyAlign, CallInst *C);
6767
bool processMemCpyMemCpyDependence(MemCpyInst *M, MemCpyInst *MDep);
6868
bool processMemSetMemCpyDependence(MemCpyInst *MemCpy, MemSetInst *MemSet);

‎llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp‎

Lines changed: 26 additions & 10 deletions
Original file line numberDiff line numberDiff line change
@@ -178,9 +178,9 @@ class MemsetRanges {
178178
}
179179

180180
void addStore(int64_t OffsetFromFirst, StoreInst *SI) {
181-
int64_t StoreSize = DL.getTypeStoreSize(SI->getOperand(0)->getType());
182-
183-
addRange(OffsetFromFirst, StoreSize, SI->getPointerOperand(),
181+
TypeSize StoreSize = DL.getTypeStoreSize(SI->getOperand(0)->getType());
182+
assert(!StoreSize.isScalable() && "Can't track scalable-typed stores");
183+
addRange(OffsetFromFirst, StoreSize.getFixedSize(), SI->getPointerOperand(),
184184
SI->getAlign().value(), SI);
185185
}
186186

@@ -363,6 +363,11 @@ Instruction *MemCpyOptPass::tryMergingIntoMemset(Instruction *StartInst,
363363
Value *ByteVal) {
364364
const DataLayout &DL = StartInst->getModule()->getDataLayout();
365365

366+
// We can't track scalable types
367+
if (StoreInst *SI = dyn_cast<StoreInst>(StartInst))
368+
if (DL.getTypeStoreSize(SI->getOperand(0)->getType()).isScalable())
369+
return nullptr;
370+
366371
// Okay, so we now have a single store that can be splatable. Scan to find
367372
// all subsequent stores of the same value to offset from the same pointer.
368373
// Join these together into ranges, so we can decide whether contiguous blocks
@@ -416,6 +421,10 @@ Instruction *MemCpyOptPass::tryMergingIntoMemset(Instruction *StartInst,
416421
if (DL.isNonIntegralPointerType(StoredVal->getType()->getScalarType()))
417422
break;
418423

424+
// We can't track ranges involving scalable types.
425+
if (DL.getTypeStoreSize(StoredVal->getType()).isScalable())
426+
break;
427+
419428
// Check to see if this stored value is of the same byte-splattable value.
420429
Value *StoredByte = isBytewiseValue(StoredVal, DL);
421430
if (isa<UndefValue>(ByteVal) && StoredByte)
@@ -836,7 +845,7 @@ bool MemCpyOptPass::processMemSet(MemSetInst *MSI, BasicBlock::iterator &BBI) {
836845
/// the call write its result directly into the destination of the memcpy.
837846
bool MemCpyOptPass::performCallSlotOptzn(Instruction *cpyLoad,
838847
Instruction *cpyStore, Value *cpyDest,
839-
Value *cpySrc, uint64_t cpyLen,
848+
Value *cpySrc, TypeSize cpySize,
840849
Align cpyAlign, CallInst *C) {
841850
// The general transformation to keep in mind is
842851
//
@@ -852,6 +861,10 @@ bool MemCpyOptPass::performCallSlotOptzn(Instruction *cpyLoad,
852861
// src only holds uninitialized values at the moment of the call, meaning that
853862
// the memcpy can be discarded rather than moved.
854863

864+
// We can't optimize scalable types.
865+
if (cpySize.isScalable())
866+
return false;
867+
855868
// Lifetime marks shouldn't be operated on.
856869
if (Function *F = C->getCalledFunction())
857870
if (F->isIntrinsic() && F->getIntrinsicID() == Intrinsic::lifetime_start)
@@ -870,13 +883,13 @@ bool MemCpyOptPass::performCallSlotOptzn(Instruction *cpyLoad,
870883
uint64_t srcSize = DL.getTypeAllocSize(srcAlloca->getAllocatedType()) *
871884
srcArraySize->getZExtValue();
872885

873-
if (cpyLen < srcSize)
886+
if (cpySize < srcSize)
874887
return false;
875888

876889
// Check that accessing the first srcSize bytes of dest will not cause a
877890
// trap. Otherwise the transform is invalid since it might cause a trap
878891
// to occur earlier than it otherwise would.
879-
if (!isDereferenceableAndAlignedPointer(cpyDest, Align(1), APInt(64, cpyLen),
892+
if (!isDereferenceableAndAlignedPointer(cpyDest, Align(1), APInt(64, cpySize),
880893
DL, C, DT))
881894
return false;
882895

@@ -1370,8 +1383,10 @@ bool MemCpyOptPass::processMemCpy(MemCpyInst *M, BasicBlock::iterator &BBI) {
13701383
// of conservatively taking the minimum?
13711384
Align Alignment = std::min(M->getDestAlign().valueOrOne(),
13721385
M->getSourceAlign().valueOrOne());
1373-
if (performCallSlotOptzn(M, M, M->getDest(), M->getSource(),
1374-
CopySize->getZExtValue(), Alignment, C)) {
1386+
if (performCallSlotOptzn(
1387+
M, M, M->getDest(), M->getSource(),
1388+
TypeSize::getFixed(CopySize->getZExtValue()), Alignment,
1389+
C)) {
13751390
LLVM_DEBUG(dbgs() << "Performed call slot optimization:\n"
13761391
<< " call: " << *C << "\n"
13771392
<< " memcpy: " << *M << "\n");
@@ -1435,7 +1450,7 @@ bool MemCpyOptPass::processByValArgument(CallBase &CB, unsigned ArgNo) {
14351450
// Find out what feeds this byval argument.
14361451
Value *ByValArg = CB.getArgOperand(ArgNo);
14371452
Type *ByValTy = CB.getParamByValType(ArgNo);
1438-
uint64_t ByValSize = DL.getTypeAllocSize(ByValTy);
1453+
TypeSize ByValSize = DL.getTypeAllocSize(ByValTy);
14391454
MemoryLocation Loc(ByValArg, LocationSize::precise(ByValSize));
14401455
MemoryUseOrDef *CallAccess = MSSA->getMemoryAccess(&CB);
14411456
if (!CallAccess)
@@ -1455,7 +1470,8 @@ bool MemCpyOptPass::processByValArgument(CallBase &CB, unsigned ArgNo) {
14551470

14561471
// The length of the memcpy must be larger or equal to the size of the byval.
14571472
ConstantInt *C1 = dyn_cast<ConstantInt>(MDep->getLength());
1458-
if (!C1 || C1->getValue().getZExtValue() < ByValSize)
1473+
if (!C1 || !TypeSize::isKnownGE(
1474+
TypeSize::getFixed(C1->getValue().getZExtValue()), ByValSize))
14591475
return false;
14601476

14611477
// Get the alignment of the byval. If the call doesn't specify the alignment,

‎llvm/test/Transforms/MemCpyOpt/vscale-crashes.ll‎

Lines changed: 101 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,101 @@
1+
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2+
; RUN: opt < %s -memcpyopt -S -verify-memoryssa | FileCheck %s
3+
4+
; Check that a call featuring a scalable-vector byval argument fed by a memcpy
5+
; doesn't crash the compiler. It previously assumed the byval type's size could
6+
; be represented as a known constant amount.
7+
define void @byval_caller(i8 *%P) {
8+
; CHECK-LABEL: @byval_caller(
9+
; CHECK-NEXT: [[A:%.*]] = alloca i8, align 1
10+
; CHECK-NEXT: call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 [[A]], i8* align 4 [[P:%.*]], i64 8, i1 false)
11+
; CHECK-NEXT: [[VA:%.*]] = bitcast i8* [[A]] to <vscale x 1 x i8>*
12+
; CHECK-NEXT: call void @byval_callee(<vscale x 1 x i8>* byval(<vscale x 1 x i8>) align 1 [[VA]])
13+
; CHECK-NEXT: ret void
14+
;
15+
%a = alloca i8
16+
call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %a, i8* align 4 %P, i64 8, i1 false)
17+
%va = bitcast i8* %a to <vscale x 1 x i8>*
18+
call void @byval_callee(<vscale x 1 x i8>* align 1 byval(<vscale x 1 x i8>) %va)
19+
ret void
20+
}
21+
22+
declare void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4, i8* align 4, i64, i1)
23+
declare void @byval_callee(<vscale x 1 x i8>* align 1 byval(<vscale x 1 x i8>))
24+
25+
; Check that two scalable-vector stores (overlapping, with a constant offset)
26+
; do not crash the compiler when checked whether or not they can be merged into
27+
; a single memset. There was previously an assumption that the stored values'
28+
; sizes could be represented by a known constant amount.
29+
define void @merge_stores_both_scalable(<vscale x 1 x i8>* %ptr) {
30+
; CHECK-LABEL: @merge_stores_both_scalable(
31+
; CHECK-NEXT: store <vscale x 1 x i8> zeroinitializer, <vscale x 1 x i8>* [[PTR:%.*]], align 1
32+
; CHECK-NEXT: [[PTRI8:%.*]] = bitcast <vscale x 1 x i8>* [[PTR]] to i8*
33+
; CHECK-NEXT: [[PTR_NEXT:%.*]] = getelementptr i8, i8* [[PTRI8]], i64 1
34+
; CHECK-NEXT: [[PTR_NEXT_2:%.*]] = bitcast i8* [[PTR_NEXT]] to <vscale x 1 x i8>*
35+
; CHECK-NEXT: store <vscale x 1 x i8> zeroinitializer, <vscale x 1 x i8>* [[PTR_NEXT_2]], align 1
36+
; CHECK-NEXT: ret void
37+
;
38+
store <vscale x 1 x i8> zeroinitializer, <vscale x 1 x i8>* %ptr
39+
%ptri8 = bitcast <vscale x 1 x i8>* %ptr to i8*
40+
%ptr.next = getelementptr i8, i8* %ptri8, i64 1
41+
%ptr.next.2 = bitcast i8* %ptr.next to <vscale x 1 x i8>*
42+
store <vscale x 1 x i8> zeroinitializer, <vscale x 1 x i8>* %ptr.next.2
43+
ret void
44+
}
45+
46+
; As above, but where the base is scalable but the subsequent store(s) are not.
47+
define void @merge_stores_first_scalable(<vscale x 1 x i8>* %ptr) {
48+
; CHECK-LABEL: @merge_stores_first_scalable(
49+
; CHECK-NEXT: store <vscale x 1 x i8> zeroinitializer, <vscale x 1 x i8>* [[PTR:%.*]], align 1
50+
; CHECK-NEXT: [[PTRI8:%.*]] = bitcast <vscale x 1 x i8>* [[PTR]] to i8*
51+
; CHECK-NEXT: [[PTR_NEXT:%.*]] = getelementptr i8, i8* [[PTRI8]], i64 1
52+
; CHECK-NEXT: store i8 0, i8* [[PTR_NEXT]], align 1
53+
; CHECK-NEXT: ret void
54+
;
55+
store <vscale x 1 x i8> zeroinitializer, <vscale x 1 x i8>* %ptr
56+
%ptri8 = bitcast <vscale x 1 x i8>* %ptr to i8*
57+
%ptr.next = getelementptr i8, i8* %ptri8, i64 1
58+
store i8 zeroinitializer, i8* %ptr.next
59+
ret void
60+
}
61+
62+
; As above, but where the base is not scalable but the subsequent store(s) are.
63+
define void @merge_stores_second_scalable(i8* %ptr) {
64+
; CHECK-LABEL: @merge_stores_second_scalable(
65+
; CHECK-NEXT: store i8 0, i8* [[PTR:%.*]], align 1
66+
; CHECK-NEXT: [[PTR_NEXT:%.*]] = getelementptr i8, i8* [[PTR]], i64 1
67+
; CHECK-NEXT: [[PTR_NEXT_2:%.*]] = bitcast i8* [[PTR_NEXT]] to <vscale x 1 x i8>*
68+
; CHECK-NEXT: store <vscale x 1 x i8> zeroinitializer, <vscale x 1 x i8>* [[PTR_NEXT_2]], align 1
69+
; CHECK-NEXT: ret void
70+
;
71+
store i8 zeroinitializer, i8* %ptr
72+
%ptr.next = getelementptr i8, i8* %ptr, i64 1
73+
%ptr.next.2 = bitcast i8* %ptr.next to <vscale x 1 x i8>*
74+
store <vscale x 1 x i8> zeroinitializer, <vscale x 1 x i8>* %ptr.next.2
75+
ret void
76+
}
77+
78+
; Check that the call-slot optimization doesn't crash when encountering scalable types.
79+
define void @callslotoptzn(<vscale x 4 x float> %val, <vscale x 4 x float>* %out) {
80+
; CHECK-LABEL: @callslotoptzn(
81+
; CHECK-NEXT: [[ALLOC:%.*]] = alloca <vscale x 4 x float>, align 16
82+
; CHECK-NEXT: [[IDX:%.*]] = tail call <vscale x 4 x i32> @llvm.experimental.stepvector.nxv4i32()
83+
; CHECK-NEXT: [[BALLOC:%.*]] = getelementptr inbounds <vscale x 4 x float>, <vscale x 4 x float>* [[ALLOC]], i64 0, i64 0
84+
; CHECK-NEXT: [[STRIDE:%.*]] = getelementptr inbounds float, float* [[BALLOC]], <vscale x 4 x i32> [[IDX]]
85+
; CHECK-NEXT: call void @llvm.masked.scatter.nxv4f32.nxv4p0f32(<vscale x 4 x float> [[VAL:%.*]], <vscale x 4 x float*> [[STRIDE]], i32 4, <vscale x 4 x i1> shufflevector (<vscale x 4 x i1> insertelement (<vscale x 4 x i1> poison, i1 true, i32 0), <vscale x 4 x i1> poison, <vscale x 4 x i32> zeroinitializer))
86+
; CHECK-NEXT: [[LI:%.*]] = load <vscale x 4 x float>, <vscale x 4 x float>* [[ALLOC]], align 4
87+
; CHECK-NEXT: store <vscale x 4 x float> [[LI]], <vscale x 4 x float>* [[OUT:%.*]], align 4
88+
; CHECK-NEXT: ret void
89+
;
90+
%alloc = alloca <vscale x 4 x float>, align 16
91+
%idx = tail call <vscale x 4 x i32> @llvm.experimental.stepvector.nxv4i32()
92+
%balloc = getelementptr inbounds <vscale x 4 x float>, <vscale x 4 x float>* %alloc, i64 0, i64 0
93+
%stride = getelementptr inbounds float, float* %balloc, <vscale x 4 x i32> %idx
94+
call void @llvm.masked.scatter.nxv4f32.nxv4p0f32(<vscale x 4 x float> %val, <vscale x 4 x float*> %stride, i32 4, <vscale x 4 x i1> shufflevector (<vscale x 4 x i1> insertelement (<vscale x 4 x i1> poison, i1 true, i32 0), <vscale x 4 x i1> poison, <vscale x 4 x i32> zeroinitializer))
95+
%li = load <vscale x 4 x float>, <vscale x 4 x float>* %alloc, align 4
96+
store <vscale x 4 x float> %li, <vscale x 4 x float>* %out, align 4
97+
ret void
98+
}
99+
100+
declare <vscale x 4 x i32> @llvm.experimental.stepvector.nxv4i32()
101+
declare void @llvm.masked.scatter.nxv4f32.nxv4p0f32(<vscale x 4 x float> , <vscale x 4 x float*> , i32, <vscale x 4 x i1>)

0 commit comments

Comments
 (0)