Skip to content

[msan] Implement shadow propagation for _mm_dp_pd, _mm_dp_ps, _mm256_dp_ps #94875

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 7 commits into from
Jun 12, 2024
Merged

[msan] Implement shadow propagation for _mm_dp_pd, _mm_dp_ps, _mm256_dp_ps #94875

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
316c1b7
[𝘀𝗽𝗿] changes to main this commit is based on
vitalybuka Jun 8, 2024
ea7ea8a
[𝘀𝗽𝗿] initial version
vitalybuka Jun 8, 2024
4984fbc
[𝘀𝗽𝗿] changes introduced through rebase
vitalybuka Jun 10, 2024
ce0d619
rebase
vitalybuka Jun 10, 2024
2689604
[𝘀𝗽𝗿] changes introduced through rebase
vitalybuka Jun 10, 2024
b4ae901
@fmayer comments
vitalybuka Jun 10, 2024
189fec3
rebase
vitalybuka Jun 10, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
77 changes: 76 additions & 1 deletion llvm/lib/Transforms/Instrumentation/MemorySanitizer.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3287,6 +3287,75 @@ struct MemorySanitizerVisitor : public InstVisitor<MemorySanitizerVisitor> {
setOriginForNaryOp(I);
}

// Convert `Mask` into `<n x i1>`.
Constant *createDppMask(unsigned Width, unsigned Mask) {
SmallVector<Constant *, 4> R(Width);
for (auto &M : R) {
M = ConstantInt::getBool(F.getContext(), Mask & 1);
Mask >>= 1;
}
return ConstantVector::get(R);
}

// Calculate output shadow as array of booleans `<n x i1>`, assuming if any
// arg is poisoned, entire dot product is poisoned.
Value *findDppPoisonedOutput(IRBuilder<> &IRB, Value *S, unsigned SrcMask,
unsigned DstMask) {
const unsigned Width =
cast<FixedVectorType>(S->getType())->getNumElements();

S = IRB.CreateSelect(createDppMask(Width, SrcMask), S,
Constant::getNullValue(S->getType()));
Value *SElem = IRB.CreateOrReduce(S);
Value *IsClean = IRB.CreateIsNull(SElem, "_msdpp");
Value *DstMaskV = createDppMask(Width, DstMask);

return IRB.CreateSelect(
IsClean, Constant::getNullValue(DstMaskV->getType()), DstMaskV);
}

// See `Intel Intrinsics Guide` for `_dp_p*` instructions.
//
// 2 and 4 element versions produce single scalar of dot product, and then
// puts it into elements of output vector, selected by 4 lowest bits of the
// mask. Top 4 bits of the mask control which elements of input to use for dot
// product.
//
// 8 element version mask still has only 4 bit for input, and 4 bit for output
// mask. According to the spec it just operates as 4 element version on first
// 4 elements of inputs and output, and then on last 4 elements of inputs and
// output.
void handleDppIntrinsic(IntrinsicInst &I) {
IRBuilder<> IRB(&I);

Value *S0 = getShadow(&I, 0);
Value *S1 = getShadow(&I, 1);
Value *S = IRB.CreateOr(S0, S1);

const unsigned Width =
cast<FixedVectorType>(S->getType())->getNumElements();
assert(Width == 2 || Width == 4 || Width == 8);

const unsigned Mask = cast<ConstantInt>(I.getArgOperand(2))->getZExtValue();
const unsigned SrcMask = Mask >> 4;
const unsigned DstMask = Mask & 0xf;

// Calculate shadow as `<n x i1>`.
Value *SI1 = findDppPoisonedOutput(IRB, S, SrcMask, DstMask);
if (Width == 8) {
// First 4 elements of shadow are already calculated. `makeDppShadow`
// operats on 32 bit masks, so we can just shift masks, and repeat.
SI1 = IRB.CreateOr(
SI1, findDppPoisonedOutput(IRB, S, SrcMask << 4, DstMask << 4));
}
// Extend to real size of shadow, poisoning either all or none bits of an
// element.
S = IRB.CreateSExt(SI1, S->getType(), "_msdpp");

setShadow(&I, S);
setOriginForNaryOp(I);
}

// Instrument sum-of-absolute-differences intrinsic.
void handleVectorSadIntrinsic(IntrinsicInst &I) {
const unsigned SignificantBitsPerResultElement = 16;
Expand Down Expand Up @@ -3642,7 +3711,7 @@ struct MemorySanitizerVisitor : public InstVisitor<MemorySanitizerVisitor> {
setOriginForNaryOp(I);
}

SmallVector<int, 8> getPclmulMask(unsigned Width, bool OddElements) {
static SmallVector<int, 8> getPclmulMask(unsigned Width, bool OddElements) {
SmallVector<int, 8> Mask;
for (unsigned X = OddElements ? 1 : 0; X < Width; X += 2) {
Mask.append(2, X);
Expand Down Expand Up @@ -3958,6 +4027,12 @@ struct MemorySanitizerVisitor : public InstVisitor<MemorySanitizerVisitor> {
handleVectorPackIntrinsic(I);
break;

case Intrinsic::x86_avx_dp_ps_256:
case Intrinsic::x86_sse41_dppd:
case Intrinsic::x86_sse41_dpps:
handleDppIntrinsic(I);
break;

case Intrinsic::x86_mmx_packsswb:
case Intrinsic::x86_mmx_packuswb:
handleVectorPackIntrinsic(I, 16);
Expand Down
23 changes: 12 additions & 11 deletions llvm/test/Instrumentation/MemorySanitizer/X86/avx-intrinsics-x86.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -389,18 +389,19 @@ define <8 x float> @test_x86_avx_dp_ps_256(<8 x float> %a0, <8 x float> %a1) #0
; CHECK-NEXT: [[TMP1:%.*]] = load <8 x i32>, ptr @__msan_param_tls, align 8
; CHECK-NEXT: [[TMP2:%.*]] = load <8 x i32>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 32) to ptr), align 8
; CHECK-NEXT: call void @llvm.donothing()
; CHECK-NEXT: [[TMP3:%.*]] = bitcast <8 x i32> [[TMP1]] to i256
; CHECK-NEXT: [[_MSCMP:%.*]] = icmp ne i256 [[TMP3]], 0
; CHECK-NEXT: [[TMP4:%.*]] = bitcast <8 x i32> [[TMP2]] to i256
; CHECK-NEXT: [[_MSCMP1:%.*]] = icmp ne i256 [[TMP4]], 0
; CHECK-NEXT: [[_MSOR:%.*]] = or i1 [[_MSCMP]], [[_MSCMP1]]
; CHECK-NEXT: br i1 [[_MSOR]], label [[TMP5:%.*]], label [[TMP6:%.*]], !prof [[PROF0]]
; CHECK: 5:
; CHECK-NEXT: call void @__msan_warning_noreturn()
; CHECK-NEXT: unreachable
; CHECK: 6:
; CHECK-NEXT: [[TMP3:%.*]] = or <8 x i32> [[TMP1]], [[TMP2]]
; CHECK-NEXT: [[TMP4:%.*]] = select <8 x i1> <i1 false, i1 true, i1 true, i1 true, i1 false, i1 false, i1 false, i1 false>, <8 x i32> [[TMP3]], <8 x i32> zeroinitializer
; CHECK-NEXT: [[TMP5:%.*]] = call i32 @llvm.vector.reduce.or.v8i32(<8 x i32> [[TMP4]])
; CHECK-NEXT: [[_MSDPP:%.*]] = icmp eq i32 [[TMP5]], 0
; CHECK-NEXT: [[TMP6:%.*]] = select i1 [[_MSDPP]], <8 x i1> zeroinitializer, <8 x i1> <i1 false, i1 true, i1 true, i1 true, i1 false, i1 false, i1 false, i1 false>
; CHECK-NEXT: [[TMP7:%.*]] = select <8 x i1> <i1 false, i1 false, i1 false, i1 false, i1 false, i1 true, i1 true, i1 true>, <8 x i32> [[TMP3]], <8 x i32> zeroinitializer
; CHECK-NEXT: [[TMP8:%.*]] = call i32 @llvm.vector.reduce.or.v8i32(<8 x i32> [[TMP7]])
; CHECK-NEXT: [[_MSDPP1:%.*]] = icmp eq i32 [[TMP8]], 0
; CHECK-NEXT: [[TMP9:%.*]] = select i1 [[_MSDPP1]], <8 x i1> zeroinitializer, <8 x i1> <i1 false, i1 false, i1 false, i1 false, i1 false, i1 true, i1 true, i1 true>
; CHECK-NEXT: [[TMP10:%.*]] = or <8 x i1> [[TMP6]], [[TMP9]]
; CHECK-NEXT: [[_MSDPP2:%.*]] = sext <8 x i1> [[TMP10]] to <8 x i32>
; CHECK-NEXT: [[RES:%.*]] = call <8 x float> @llvm.x86.avx.dp.ps.256(<8 x float> [[A0:%.*]], <8 x float> [[A1:%.*]], i8 -18)
; CHECK-NEXT: store <8 x i32> zeroinitializer, ptr @__msan_retval_tls, align 8
; CHECK-NEXT: store <8 x i32> [[_MSDPP2]], ptr @__msan_retval_tls, align 8
; CHECK-NEXT: ret <8 x float> [[RES]]
;
%res = call <8 x float> @llvm.x86.avx.dp.ps.256(<8 x float> %a0, <8 x float> %a1, i8 -18) ; <<8 x float>> [#uses=1]
Expand Down
38 changes: 15 additions & 23 deletions llvm/test/Instrumentation/MemorySanitizer/X86/sse41-intrinsics-x86.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -45,18 +45,14 @@ define <2 x double> @test_x86_sse41_dppd(<2 x double> %a0, <2 x double> %a1) #0
; CHECK-NEXT: [[TMP1:%.*]] = load <2 x i64>, ptr @__msan_param_tls, align 8
; CHECK-NEXT: [[TMP2:%.*]] = load <2 x i64>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
; CHECK-NEXT: call void @llvm.donothing()
; CHECK-NEXT: [[TMP3:%.*]] = bitcast <2 x i64> [[TMP1]] to i128
; CHECK-NEXT: [[_MSCMP:%.*]] = icmp ne i128 [[TMP3]], 0
; CHECK-NEXT: [[TMP4:%.*]] = bitcast <2 x i64> [[TMP2]] to i128
; CHECK-NEXT: [[_MSCMP1:%.*]] = icmp ne i128 [[TMP4]], 0
; CHECK-NEXT: [[_MSOR:%.*]] = or i1 [[_MSCMP]], [[_MSCMP1]]
; CHECK-NEXT: br i1 [[_MSOR]], label [[TMP5:%.*]], label [[TMP6:%.*]], !prof [[PROF0:![0-9]+]]
; CHECK: 5:
; CHECK-NEXT: call void @__msan_warning_noreturn()
; CHECK-NEXT: unreachable
; CHECK: 6:
; CHECK-NEXT: [[TMP3:%.*]] = or <2 x i64> [[TMP1]], [[TMP2]]
; CHECK-NEXT: [[TMP4:%.*]] = select <2 x i1> <i1 false, i1 true>, <2 x i64> [[TMP3]], <2 x i64> zeroinitializer
; CHECK-NEXT: [[TMP5:%.*]] = call i64 @llvm.vector.reduce.or.v2i64(<2 x i64> [[TMP4]])
; CHECK-NEXT: [[_MSDPP:%.*]] = icmp eq i64 [[TMP5]], 0
; CHECK-NEXT: [[TMP6:%.*]] = select i1 [[_MSDPP]], <2 x i1> zeroinitializer, <2 x i1> <i1 false, i1 true>
; CHECK-NEXT: [[_MSDPP1:%.*]] = sext <2 x i1> [[TMP6]] to <2 x i64>
; CHECK-NEXT: [[RES:%.*]] = call <2 x double> @llvm.x86.sse41.dppd(<2 x double> [[A0:%.*]], <2 x double> [[A1:%.*]], i8 -18)
; CHECK-NEXT: store <2 x i64> zeroinitializer, ptr @__msan_retval_tls, align 8
; CHECK-NEXT: store <2 x i64> [[_MSDPP1]], ptr @__msan_retval_tls, align 8
; CHECK-NEXT: ret <2 x double> [[RES]]
;
%res = call <2 x double> @llvm.x86.sse41.dppd(<2 x double> %a0, <2 x double> %a1, i8 -18) ; <<2 x double>> [#uses=1]
Expand All @@ -70,18 +66,14 @@ define <4 x float> @test_x86_sse41_dpps(<4 x float> %a0, <4 x float> %a1) #0 {
; CHECK-NEXT: [[TMP1:%.*]] = load <4 x i32>, ptr @__msan_param_tls, align 8
; CHECK-NEXT: [[TMP2:%.*]] = load <4 x i32>, ptr inttoptr (i64 add (i64 ptrtoint (ptr @__msan_param_tls to i64), i64 16) to ptr), align 8
; CHECK-NEXT: call void @llvm.donothing()
; CHECK-NEXT: [[TMP3:%.*]] = bitcast <4 x i32> [[TMP1]] to i128
; CHECK-NEXT: [[_MSCMP:%.*]] = icmp ne i128 [[TMP3]], 0
; CHECK-NEXT: [[TMP4:%.*]] = bitcast <4 x i32> [[TMP2]] to i128
; CHECK-NEXT: [[_MSCMP1:%.*]] = icmp ne i128 [[TMP4]], 0
; CHECK-NEXT: [[_MSOR:%.*]] = or i1 [[_MSCMP]], [[_MSCMP1]]
; CHECK-NEXT: br i1 [[_MSOR]], label [[TMP5:%.*]], label [[TMP6:%.*]], !prof [[PROF0]]
; CHECK: 5:
; CHECK-NEXT: call void @__msan_warning_noreturn()
; CHECK-NEXT: unreachable
; CHECK: 6:
; CHECK-NEXT: [[TMP3:%.*]] = or <4 x i32> [[TMP1]], [[TMP2]]
; CHECK-NEXT: [[TMP4:%.*]] = select <4 x i1> <i1 false, i1 true, i1 true, i1 true>, <4 x i32> [[TMP3]], <4 x i32> zeroinitializer
; CHECK-NEXT: [[TMP5:%.*]] = call i32 @llvm.vector.reduce.or.v4i32(<4 x i32> [[TMP4]])
; CHECK-NEXT: [[_MSDPP:%.*]] = icmp eq i32 [[TMP5]], 0
; CHECK-NEXT: [[TMP6:%.*]] = select i1 [[_MSDPP]], <4 x i1> zeroinitializer, <4 x i1> <i1 false, i1 true, i1 true, i1 true>
; CHECK-NEXT: [[_MSDPP1:%.*]] = sext <4 x i1> [[TMP6]] to <4 x i32>
; CHECK-NEXT: [[RES:%.*]] = call <4 x float> @llvm.x86.sse41.dpps(<4 x float> [[A0:%.*]], <4 x float> [[A1:%.*]], i8 -18)
; CHECK-NEXT: store <4 x i32> zeroinitializer, ptr @__msan_retval_tls, align 8
; CHECK-NEXT: store <4 x i32> [[_MSDPP1]], ptr @__msan_retval_tls, align 8
; CHECK-NEXT: ret <4 x float> [[RES]]
;
%res = call <4 x float> @llvm.x86.sse41.dpps(<4 x float> %a0, <4 x float> %a1, i8 -18) ; <<4 x float>> [#uses=1]
Expand All @@ -100,7 +92,7 @@ define <4 x float> @test_x86_sse41_insertps(<4 x float> %a0, <4 x float> %a1) #0
; CHECK-NEXT: [[TMP4:%.*]] = bitcast <4 x i32> [[TMP2]] to i128
; CHECK-NEXT: [[_MSCMP1:%.*]] = icmp ne i128 [[TMP4]], 0
; CHECK-NEXT: [[_MSOR:%.*]] = or i1 [[_MSCMP]], [[_MSCMP1]]
; CHECK-NEXT: br i1 [[_MSOR]], label [[TMP5:%.*]], label [[TMP6:%.*]], !prof [[PROF0]]
; CHECK-NEXT: br i1 [[_MSOR]], label [[TMP5:%.*]], label [[TMP6:%.*]], !prof [[PROF0:![0-9]+]]
; CHECK: 5:
; CHECK-NEXT: call void @__msan_warning_noreturn()
; CHECK-NEXT: unreachable
Expand Down
Loading