Skip to content

arm64: Clean up SVE embedded masked codegen #127164

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
dhartglassMSFT merged 12 commits into from
Jun 30, 2026
Merged

arm64: Clean up SVE embedded masked codegen #127164

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
12 commits
Select commit Hold shift + click to select a range
73b0128
Clean up SVE embedded masked codegen
ylpoonlg Jan 22, 2026
4cddda0
Fix optional embedded masked with true mask
ylpoonlg Apr 23, 2026
42b67e6
Refactor RMW register check function
ylpoonlg May 5, 2026
cf72018
Fix SVE opt tests
ylpoonlg May 5, 2026
d144b80
extra checks for unary embedded masked ops in lsra
a74nh Jun 24, 2026
c11e121
Add special cases for ConvertToSingleOddRoundToOdd
a74nh Jun 24, 2026
9dda2b4
Merge remote-tracking branch 'origin/main' into github-movprfx_refact…
ylpoon-arm Jun 24, 2026
235dd4c
Fix formatting
ylpoonlg Jun 24, 2026
3b0e5a9
Merge branch 'main' into github-movprfx_refactor_3
dhartglassMSFT Jun 24, 2026
f4f9c8f
reserve register for MultiplyAddRotateComplex
a74nh Jun 29, 2026
cd44faf
Fix format
ylpoonlg Jun 29, 2026
c981363
fix checkRMWRegisters for FMA embedded ops
a74nh Jun 30, 2026
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
6 changes: 6 additions & 0 deletions src/coreclr/jit/codegen.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1019,6 +1019,12 @@ class CodeGen final : public CodeGenInterface
#endif // defined(TARGET_XARCH)

#ifdef TARGET_ARM64
void genEmbeddedMaskedHWIntrinsic(GenTreeHWIntrinsic* cndSelNode, regNumber targetReg);

#ifdef DEBUG
void checkRMWRegisters(const HWIntrinsic intrin, regNumber targetReg);
#endif

class HWIntrinsicImmOpHelper final
{
public:
Expand Down
5 changes: 4 additions & 1 deletion src/coreclr/jit/emitarm64sve.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19318,7 +19318,10 @@ void emitter::emitInsPairSanityCheck(instrDesc* firstId, instrDesc* secondId)
assert(firstId->idReg2() == secondId->idReg2());

// "predicated using the same governing predicate register and source element size as this instruction."
assert(firstId->idInsOpt() == secondId->idInsOpt());
emitAttr dstSize1 = optGetSveElemsize(firstId->idInsOpt());
emitAttr dstSize2 = insOptsScalableStandard(secondId->idInsOpt()) ? optGetSveElemsize(secondId->idInsOpt())
: optGetDstsize(secondId->idInsOpt());
assert(dstSize1 == dstSize2);
}

// The following instructions cannot use predicated movprfx, else the behaviour will be unpredictable.
Expand Down
1,298 changes: 602 additions & 696 deletions src/coreclr/jit/hwintrinsiccodegenarm64.cpp
View file
Open in desktop

Large diffs are not rendered by default.

21 changes: 12 additions & 9 deletions src/coreclr/jit/lowerarmarch.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3965,23 +3965,26 @@ void Lowering::ContainCheckHWIntrinsic(GenTreeHWIntrinsic* node)
}

// Handle op3
if (op3->IsVectorZero() && op1->IsTrueMask(node->GetSimdBaseType()) && op2->IsEmbMaskOp())
if (op3->IsVectorZero() && op2->IsEmbMaskOp())
{
// When we are merging with zero, we can specialize
// and avoid instantiating the vector constant.
// Do this only if op1 was AllTrueMask
// When we are merging with zero, we can specialize and avoid instantiating the vector constant.
switch (op2->AsHWIntrinsic()->GetHWIntrinsicId())
{
case NI_Sve2_AddPairwise:
case NI_Sve2_MaxNumberPairwise:
case NI_Sve2_MaxPairwise:
case NI_Sve2_MinNumberPairwise:
case NI_Sve2_MinPairwise:
// This is an edge case where these instructions have unpredictable behaviour when
// using predicated movprfx, so the unpredicated variant must be used here. This
// prevents us from performing this optimization as we will need the constant vector
// for masking the result.
break;
case NI_Sve2_ConvertToDoubleOdd:
case NI_Sve2_ConvertToSingleOdd:
case NI_Sve2_ConvertToSingleOddRoundToOdd:
// This is an edge case where these instructions do not support predicated or any movprfx.
if (!op1->IsTrueMask(node->GetSimdBaseType()))
{
// When op1 is not all-true, we will need the constant vector for masking the result.
break;
}
FALLTHROUGH;

default:
MakeSrcContained(node, op3);
Expand Down
5 changes: 4 additions & 1 deletion src/coreclr/jit/lsra.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1436,7 +1436,7 @@ class LinearScan : public RegAllocInterface
FORCEINLINE RefPosition* getNextConsecutiveRefPosition(RefPosition* refPosition);
FORCEINLINE regNumber getNextFPRegWraparound(regNumber reg);
SingleTypeRegSet getOperandCandidates(GenTreeHWIntrinsic* intrinsicTree, HWIntrinsic intrin, size_t opNum);
GenTree* getDelayFreeOperand(GenTreeHWIntrinsic* intrinsicTree, bool embedded = false);
GenTree* getDelayFreeOperand(GenTreeHWIntrinsic* intrinsicTree, GenTreeHWIntrinsic* user = nullptr);
GenTree* getVectorAddrOperand(GenTreeHWIntrinsic* intrinsicTree);
GenTree* getConsecutiveRegistersOperand(const HWIntrinsic intrin, bool* destIsConsecutive);
GenTreeHWIntrinsic* getEmbeddedMaskOperand(const HWIntrinsic intrin);
Expand Down Expand Up @@ -2078,6 +2078,9 @@ class LinearScan : public RegAllocInterface
int BuildConsecutiveRegistersForUse(GenTree* treeNode, GenTree* rmwNode = nullptr);
void BuildConsecutiveRegistersForDef(GenTree* treeNode, int fieldCount);
void BuildHWIntrinsicImmediate(GenTreeHWIntrinsic* intrinsicTree, const HWIntrinsic intrin);
void BuildHWIntrinsicTempRegs(GenTreeHWIntrinsic* intrinsicTree,
const HWIntrinsic intrin,
GenTreeHWIntrinsic* embeddedOp);
int BuildEmbeddedOperandUses(GenTreeHWIntrinsic* embeddedOpNode, GenTree* embeddedDelayFreeOp);
int BuildContainedCselUses(GenTreeHWIntrinsic* containedCselOpNode,
GenTree* delayFreeOp,
Expand Down
106 changes: 83 additions & 23 deletions src/coreclr/jit/lsraarm64.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1430,32 +1430,14 @@ int LinearScan::BuildHWIntrinsic(GenTreeHWIntrinsic* intrinsicTree, int* pDstCou
if (embeddedOp != nullptr)
{
assert(delayFreeOp == nullptr);
delayFreeOp = getDelayFreeOperand(embeddedOp, /* embedded */ true);
delayFreeOp = getDelayFreeOperand(embeddedOp, intrinsicTree);
}

// Build any immediates
BuildHWIntrinsicImmediate(intrinsicTree, intrin);

// Build any additional special cases
switch (intrin.id)
{
case NI_Sve2_GatherVectorInt16SignExtendNonTemporal:
case NI_Sve2_GatherVectorInt32SignExtendNonTemporal:
case NI_Sve2_GatherVectorNonTemporal:
case NI_Sve2_GatherVectorUInt16ZeroExtendNonTemporal:
case NI_Sve2_GatherVectorUInt32ZeroExtendNonTemporal:
case NI_Sve2_Scatter16BitNarrowingNonTemporal:
case NI_Sve2_Scatter32BitNarrowingNonTemporal:
case NI_Sve2_ScatterNonTemporal:
if (!varTypeIsSIMD(intrin.op2->gtType))
{
buildInternalFloatRegisterDefForNode(intrinsicTree, internalFloatRegCandidates());
}
break;

default:
break;
}
// Build any internal temporary registers
BuildHWIntrinsicTempRegs(intrinsicTree, intrin, embeddedOp);

// Build all Operands
for (size_t opNum = 1; opNum <= intrin.numOperands; opNum++)
Expand Down Expand Up @@ -1809,6 +1791,47 @@ void LinearScan::BuildHWIntrinsicImmediate(GenTreeHWIntrinsic* intrinsicTree, co
}
}

//------------------------------------------------------------------------
// BuildHWIntrinsicTempRegs: Build temporary registers used by HWIntrinsic codegen.
//
// Arguments:
// intrinsicTree - Intrinsic tree node for which need to build RefPositions for
// intrin - Underlying intrinsic
// embeddedOp - Embedded mask operand of intrinsicTree, if any
//
void LinearScan::BuildHWIntrinsicTempRegs(GenTreeHWIntrinsic* intrinsicTree,
const HWIntrinsic intrin,
GenTreeHWIntrinsic* embeddedOp)
{
switch (intrin.id)
{
case NI_Sve2_GatherVectorInt16SignExtendNonTemporal:
case NI_Sve2_GatherVectorInt32SignExtendNonTemporal:
case NI_Sve2_GatherVectorNonTemporal:
case NI_Sve2_GatherVectorUInt16ZeroExtendNonTemporal:
case NI_Sve2_GatherVectorUInt32ZeroExtendNonTemporal:
case NI_Sve2_Scatter16BitNarrowingNonTemporal:
case NI_Sve2_Scatter32BitNarrowingNonTemporal:
case NI_Sve2_ScatterNonTemporal:
if (!varTypeIsSIMD(intrin.op2->gtType))
{
buildInternalFloatRegisterDefForNode(intrinsicTree, internalFloatRegCandidates());
}
break;

case NI_Sve_ConditionalSelect:
if ((embeddedOp != nullptr) && embeddedOp->OperIsHWIntrinsic(NI_Sve_MultiplyAddRotateComplex))
{
buildInternalFloatRegisterDefForNode(intrinsicTree, internalFloatRegCandidates());
setInternalRegsDelayFree = true;
}
break;

default:
break;
}
}

//------------------------------------------------------------------------
// BuildEmbeddedOperandUses: Build the uses for an Embedded Mask operand
//
Expand Down Expand Up @@ -2315,17 +2338,18 @@ SingleTypeRegSet LinearScan::getOperandCandidates(GenTreeHWIntrinsic* intrinsicT
//
// Arguments:
// intrinsicTree - Tree to check
// embedded - If this is an embedded operand
// user - The user of intrinsicTree if intrinsicTree is an embedded operand
//
// Return Value:
// The operand that needs to be delay freed
//
GenTree* LinearScan::getDelayFreeOperand(GenTreeHWIntrinsic* intrinsicTree, bool embedded)
GenTree* LinearScan::getDelayFreeOperand(GenTreeHWIntrinsic* intrinsicTree, GenTreeHWIntrinsic* user)
{
bool isRMW = intrinsicTree->isRMWHWIntrinsic(m_compiler);

const NamedIntrinsic intrinsicId = intrinsicTree->GetHWIntrinsicId();
GenTree* delayFreeOp = nullptr;
const bool embedded = user != nullptr;

switch (intrinsicId)
{
Expand Down Expand Up @@ -2381,6 +2405,23 @@ GenTree* LinearScan::getDelayFreeOperand(GenTreeHWIntrinsic* intrinsicTree, bool
assert(delayFreeOp != nullptr);
break;

case NI_Sve2_ConvertToSingleOdd:
case NI_Sve2_ConvertToSingleOddRoundToOdd:
assert(isRMW);
if (embedded && user->OperIsHWIntrinsic(NI_Sve_ConditionalSelect) && user->Op(3)->IsVectorZero() &&
!user->Op(1)->IsTrueMask(user->GetSimdBaseType()))
{
// These instructions cannot use movprfx. Prefer the zero falseOp as the target so codegen can
// preserve inactive lanes by first copying active lanes from op1.
delayFreeOp = user->Op(3);
}
else
{
delayFreeOp = intrinsicTree->Op(1);
}
assert(delayFreeOp != nullptr);
break;

default:
if (isRMW)
{
Expand All @@ -2406,6 +2447,25 @@ GenTree* LinearScan::getDelayFreeOperand(GenTreeHWIntrinsic* intrinsicTree, bool
assert(delayFreeOp != nullptr);
}
}
else if ((intrinsicTree->GetOperandCount() == 1) && embedded &&
!HWIntrinsicInfo::IsReduceOperation(intrinsicId))
{
// Unary embedded masked operations are non-RMW, but also support movprfx.
assert(user->OperIsHWIntrinsic(NI_Sve_ConditionalSelect));

if (user->Op(1)->IsTrueMask(user->GetSimdBaseType()) && user->Op(3)->IsVectorZero())
{
// When all lanes are active, falseOp is irrelevant and the embedded operation can write back to
// op1.
delayFreeOp = intrinsicTree->Op(1);
}
else
{
// Mark the falseOp of the conditional select as delay free.
delayFreeOp = user->Op(3);
}
assert(delayFreeOp != nullptr);
}
break;
}

Expand Down
4 changes: 3 additions & 1 deletion src/tests/JIT/HardwareIntrinsics/Arm/Shared/SveGatherVector.template
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -418,7 +418,9 @@ namespace {Namespace}
[method: MethodImpl(MethodImplOptions.AggressiveInlining)]
private void ConditionalSelectScenario_TrueValue({RetVectorType}<{RetBaseType}> mask, {Op1VectorType}<{Op1BaseType}> op1, {Op2BaseType}* op2, {Op3VectorType}<{Op3BaseType}> op3, {Op1VectorType}<{Op1BaseType}> falseOp)
{
var result = Sve.ConditionalSelect(mask, {Isa}.{Method}(op1, op2, op3), falseOp);
var result = (falseOp == {Op1VectorType}<{Op1BaseType}>.Zero) ?
Sve.ConditionalSelect(mask, {Isa}.{Method}(op1, op2, op3), {Op1VectorType}<{Op1BaseType}>.Zero) :
Sve.ConditionalSelect(mask, {Isa}.{Method}(op1, op2, op3), falseOp);

Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateConditionalSelectResult_TrueValue(mask, op1, op2, op3, falseOp, _dataTable.outArrayPtr);
Expand Down
4 changes: 3 additions & 1 deletion src/tests/JIT/HardwareIntrinsics/Arm/Shared/SveGatherVectorByteOffsetFirstFaulting.template
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -470,7 +470,9 @@ namespace {Namespace}
[method: MethodImpl(MethodImplOptions.AggressiveInlining)]
private void ConditionalSelectScenario_TrueValue({RetVectorType}<{RetBaseType}> mask, {Op1VectorType}<{Op1BaseType}> op1, {Op2BaseType}* op2, {Op3VectorType}<{Op3BaseType}> op3, {Op1VectorType}<{Op1BaseType}> falseOp)
{
var result = Sve.ConditionalSelect(mask, {Isa}.{Method}(op1, op2, op3), falseOp);
var result = (falseOp == {Op1VectorType}<{Op1BaseType}>.Zero) ?
Sve.ConditionalSelect(mask, {Isa}.{Method}(op1, op2, op3), {Op1VectorType}<{Op1BaseType}>.Zero) :
Sve.ConditionalSelect(mask, {Isa}.{Method}(op1, op2, op3), falseOp);

Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateConditionalSelectResult_TrueValue(mask, op1, op2, op3, falseOp, _dataTable.outArrayPtr);
Expand Down
4 changes: 3 additions & 1 deletion src/tests/JIT/HardwareIntrinsics/Arm/Shared/SveGatherVectorByteOffsets.template
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -396,7 +396,9 @@ namespace {Namespace}
[method: MethodImpl(MethodImplOptions.AggressiveInlining)]
private void ConditionalSelectScenario_TrueValue({RetVectorType}<{RetBaseType}> mask, {Op1VectorType}<{Op1BaseType}> op1, {Op2BaseType}* op2, {Op3VectorType}<{Op3BaseType}> op3, {Op1VectorType}<{Op1BaseType}> falseOp)
{
var result = Sve.ConditionalSelect(mask, {Isa}.{Method}(op1, op2, op3), falseOp);
var result = (falseOp == {Op1VectorType}<{Op1BaseType}>.Zero) ?
Sve.ConditionalSelect(mask, {Isa}.{Method}(op1, op2, op3), {Op1VectorType}<{Op1BaseType}>.Zero) :
Sve.ConditionalSelect(mask, {Isa}.{Method}(op1, op2, op3), falseOp);

Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateConditionalSelectResult_TrueValue(mask, op1, op2, op3, falseOp, _dataTable.outArrayPtr);
Expand Down
4 changes: 3 additions & 1 deletion src/tests/JIT/HardwareIntrinsics/Arm/Shared/SveGatherVectorFirstFaultingVectorBases.template
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -470,7 +470,9 @@ namespace {Namespace}
[method: MethodImpl(MethodImplOptions.AggressiveInlining)]
private void ConditionalSelectScenario_TrueValue({RetVectorType}<{RetBaseType}> mask, {Op1VectorType}<{Op1BaseType}> op1, {Op2VectorType}<{Op2BaseType}> op2, {Op1VectorType}<{Op1BaseType}> falseOp)
{
var result = Sve.ConditionalSelect(mask, {Isa}.{Method}(op1, op2), falseOp);
var result = (falseOp == {Op1VectorType}<{Op1BaseType}>.Zero) ?
Sve.ConditionalSelect(mask, {Isa}.{Method}(op1, op2), {Op1VectorType}<{Op1BaseType}>.Zero) :
Sve.ConditionalSelect(mask, {Isa}.{Method}(op1, op2), falseOp);

Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateConditionalSelectResult_TrueValue(mask, op1, op2, falseOp, _dataTable.outArrayPtr);
Expand Down
4 changes: 3 additions & 1 deletion src/tests/JIT/HardwareIntrinsics/Arm/Shared/SveGatherVectorVectorBases.template
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -412,7 +412,9 @@ namespace {Namespace}
[method: MethodImpl(MethodImplOptions.AggressiveInlining)]
private void ConditionalSelectScenario_TrueValue({RetVectorType}<{RetBaseType}> mask, {Op1VectorType}<{Op1BaseType}> op1, {Op2VectorType}<{Op2BaseType}> op2, {Op1VectorType}<{Op1BaseType}> falseOp)
{
var result = Sve.ConditionalSelect(mask, {Isa}.{Method}(op1, op2), falseOp);
var result = (falseOp == {Op1VectorType}<{Op1BaseType}>.Zero) ?
Sve.ConditionalSelect(mask, {Isa}.{Method}(op1, op2), {Op1VectorType}<{Op1BaseType}>.Zero) :
Sve.ConditionalSelect(mask, {Isa}.{Method}(op1, op2), falseOp);

Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateConditionalSelectResult_TrueValue(mask, op1, op2, falseOp, _dataTable.outArrayPtr);
Expand Down
4 changes: 3 additions & 1 deletion src/tests/JIT/HardwareIntrinsics/Arm/Shared/_SveBinaryMaskOpTestTemplate.template
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -345,7 +345,9 @@ namespace {Namespace}
[method: MethodImpl(MethodImplOptions.AggressiveInlining)]
private void ConditionalSelectScenario_TrueValue({Op1VectorType}<{Op1BaseType}> mask, {Op1VectorType}<{Op1BaseType}> op1, {Op1VectorType}<{Op2BaseType}> op2, {Op1VectorType}<{Op1BaseType}> falseOp)
{
var result = Sve.ConditionalSelect(mask, {Isa}.{Method}(op1, op2), falseOp);
var result = (falseOp == {Op1VectorType}<{Op1BaseType}>.Zero) ?
Comment thread
dhartglassMSFT marked this conversation as resolved.
Sve.ConditionalSelect(mask, {Isa}.{Method}(op1, op2), {Op1VectorType}<{Op1BaseType}>.Zero) :
Sve.ConditionalSelect(mask, {Isa}.{Method}(op1, op2), falseOp);

Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateConditionalSelectResult_TrueValue(mask, op1, op2, falseOp, _dataTable.outArrayPtr);
Expand Down
4 changes: 3 additions & 1 deletion ...tests/JIT/HardwareIntrinsics/Arm/Shared/_SveBinaryOpDifferentRetTypeTestTemplate.template
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -403,7 +403,9 @@ namespace {Namespace}
[method: MethodImpl(MethodImplOptions.AggressiveInlining)]
private void ConditionalSelectScenario_TrueValue({RetVectorType}<{RetBaseType}> mask, {Op1VectorType}<{Op1BaseType}> op1, {Op2VectorType}<{Op2BaseType}> op2, {RetVectorType}<{RetBaseType}> falseOp)
{
var result = Sve.ConditionalSelect(mask, {Isa}.{Method}(op1, op2), falseOp);
var result = (falseOp == {RetVectorType}<{RetBaseType}>.Zero) ?
Sve.ConditionalSelect(mask, {Isa}.{Method}(op1, op2), {RetVectorType}<{RetBaseType}>.Zero) :
Sve.ConditionalSelect(mask, {Isa}.{Method}(op1, op2), falseOp);

Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateConditionalSelectResult_TrueValue(mask, op1, op2, falseOp, _dataTable.outArrayPtr);
Expand Down
4 changes: 3 additions & 1 deletion src/tests/JIT/HardwareIntrinsics/Arm/Shared/_SveBinaryOpDifferentTypesTestTemplate.template
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -322,7 +322,9 @@ namespace {Namespace}
[method: MethodImpl(MethodImplOptions.AggressiveInlining)]
private void ConditionalSelectScenario_TrueValue({Op1VectorType}<{Op1BaseType}> mask, {Op1VectorType}<{Op1BaseType}> op1, {Op2VectorType}<{Op2BaseType}> op2, {Op1VectorType}<{Op1BaseType}> falseOp)
{
var result = Sve.ConditionalSelect(mask, {Isa}.{Method}(op1, op2), falseOp);
var result = (falseOp == {Op1VectorType}<{Op1BaseType}>.Zero) ?
Sve.ConditionalSelect(mask, {Isa}.{Method}(op1, op2), {Op1VectorType}<{Op1BaseType}>.Zero) :
Sve.ConditionalSelect(mask, {Isa}.{Method}(op1, op2), falseOp);

Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateConditionalSelectResult_TrueValue(mask, op1, op2, falseOp, _dataTable.outArrayPtr);
Expand Down
4 changes: 3 additions & 1 deletion src/tests/JIT/HardwareIntrinsics/Arm/Shared/_SveBinaryOpTestTemplate.template
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -313,7 +313,9 @@ namespace {Namespace}
[method: MethodImpl(MethodImplOptions.AggressiveInlining)]
private void ConditionalSelectScenario_TrueValue(Vector<{RetBaseType}> mask, Vector<{Op1BaseType}> op1, Vector<{Op2BaseType}> op2, Vector<{RetBaseType}> falseOp)
{
var result = Sve.ConditionalSelect(mask, {Isa}.{Method}(op1, op2), falseOp);
var result = (falseOp == Vector<{RetBaseType}>.Zero) ?
Sve.ConditionalSelect(mask, {Isa}.{Method}(op1, op2), Vector<{RetBaseType}>.Zero) :
Comment thread
dhartglassMSFT marked this conversation as resolved.
Sve.ConditionalSelect(mask, {Isa}.{Method}(op1, op2), falseOp);
ValidateConditionalSelectResult(mask, op1, op2, falseOp, result);
}

Expand Down
4 changes: 3 additions & 1 deletion src/tests/JIT/HardwareIntrinsics/Arm/Shared/_SveBinaryRetMaskOpTestTemplate.template
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -383,7 +383,9 @@ namespace {Namespace}
[method: MethodImpl(MethodImplOptions.AggressiveInlining)]
private void ConditionalSelectScenario_TrueValue({Op1VectorType}<{Op1BaseType}> mask, {Op1VectorType}<{Op1BaseType}> op1, {Op1VectorType}<{Op1BaseType}> op2, {Op1VectorType}<{Op1BaseType}> falseOp)
{
var result = Sve.ConditionalSelect(mask, {Isa}.{Method}(op1, op2), falseOp);
var result = (falseOp == {Op1VectorType}<{Op1BaseType}>.Zero) ?
Sve.ConditionalSelect(mask, {Isa}.{Method}(op1, op2), {Op1VectorType}<{Op1BaseType}>.Zero) :
Sve.ConditionalSelect(mask, {Isa}.{Method}(op1, op2), falseOp);

Unsafe.Write(_dataTable.outArrayPtr, result);
ValidateConditionalSelectResult_TrueValue(mask, op1, op2, falseOp, _dataTable.outArrayPtr);
Expand Down
Loading
Loading