Skip to content

Arm64: re-enable use of predicate variants #117313

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
amanasifkhalid merged 9 commits into from
Aug 11, 2025
Merged

Arm64: re-enable use of predicate variants #117313

Show file tree
Hide file tree
Changes from 8 commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
09cbb85
Arm64: re-enable use of predicate variants
a74nh Jul 3, 2025
af485a9
Add constant folding for ConditionalSelect_Predicates
a74nh Jul 4, 2025
c5c1f5d
don't return after folding to predicate variant
a74nh Jul 4, 2025
9e8b728
update endif
a74nh Jul 16, 2025
0e0dd38
Add simd16_t mask
a74nh Jul 29, 2025
8f3ff2e
Add header comment for gtFoldExprHWIntrinsic
a74nh Jul 29, 2025
a980e07
ensure op1 etc is set
a74nh Jul 29, 2025
7581435
make canFold checks more explicit
a74nh Aug 6, 2025
d5b3c7b
ifdef
a74nh Aug 11, 2025
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
115 changes: 109 additions & 6 deletions src/coreclr/jit/gentree.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -32109,6 +32109,15 @@ bool GenTree::CanDivOrModPossiblyOverflow(Compiler* comp) const
return true;
}

//------------------------------------------------------------------------
// gtFoldExprHWIntrinsic: Attempt to fold a HWIntrinsic
//
// Arguments:
// tree - HWIntrinsic to fold
//
// Return Value:
// folded expression if it could be folded, else the original tree
//
#if defined(FEATURE_HW_INTRINSICS)
GenTree* Compiler::gtFoldExprHWIntrinsic(GenTreeHWIntrinsic* tree)
{
Expand Down Expand Up @@ -32252,7 +32261,8 @@ GenTree* Compiler::gtFoldExprHWIntrinsic(GenTreeHWIntrinsic* tree)
// We shouldn't find AND_NOT nodes since it should only be produced in lowering
assert(oper != GT_AND_NOT);

#if defined(FEATURE_MASKED_HW_INTRINSICS) && defined(TARGET_XARCH)
#if defined(FEATURE_MASKED_HW_INTRINSICS)
#if defined(TARGET_XARCH)
if (GenTreeHWIntrinsic::OperIsBitwiseHWIntrinsic(oper))
{
// Comparisons that produce masks lead to more verbose trees than
Expand Down Expand Up @@ -32370,7 +32380,75 @@ GenTree* Compiler::gtFoldExprHWIntrinsic(GenTreeHWIntrinsic* tree)
}
}
}
#endif // FEATURE_MASKED_HW_INTRINSICS && TARGET_XARCH
#elif defined(TARGET_ARM64)
// Check if the tree can be folded into a mask variant
if (HWIntrinsicInfo::HasAllMaskVariant(tree->GetHWIntrinsicId()))
{
NamedIntrinsic maskVariant = HWIntrinsicInfo::GetMaskVariant(tree->GetHWIntrinsicId());

assert(opCount == (size_t)HWIntrinsicInfo::lookupNumArgs(maskVariant));

// Check all operands are valid
bool canFold = true;
if (ni == NI_Sve_ConditionalSelect)
{
assert(varTypeIsMask(op1));
canFold = (op2->OperIsConvertMaskToVector() && op3->OperIsConvertMaskToVector());
}
else
{
for (size_t i = 1; i <= opCount && canFold; i++)
{
canFold &= tree->Op(i)->OperIsConvertMaskToVector();
}
}

if (canFold)
{
// Convert all the operands to masks
for (size_t i = 1; i <= opCount; i++)
{
if (tree->Op(i)->OperIsConvertMaskToVector())
{
// Replace with op1.
tree->Op(i) = tree->Op(i)->AsHWIntrinsic()->Op(1);
}
else if (tree->Op(i)->IsVectorZero())
{
// Replace the vector of zeroes with a mask of zeroes.
tree->Op(i) = gtNewSimdFalseMaskByteNode();
tree->Op(i)->SetMorphed(this);
}
assert(varTypeIsMask(tree->Op(i)));
}

// Switch to the mask variant
switch (opCount)
{
case 1:
tree->ResetHWIntrinsicId(maskVariant, tree->Op(1));
break;
case 2:
tree->ResetHWIntrinsicId(maskVariant, tree->Op(1), tree->Op(2));
break;
case 3:
tree->ResetHWIntrinsicId(maskVariant, this, tree->Op(1), tree->Op(2), tree->Op(3));
break;
default:
unreached();
}

tree->gtType = TYP_MASK;
tree->SetMorphed(this);
tree = gtNewSimdCvtMaskToVectorNode(retType, tree, simdBaseJitType, simdSize)->AsHWIntrinsic();
tree->SetMorphed(this);
op1 = tree->Op(1);
op2 = nullptr;
op3 = nullptr;
}
}
#endif // TARGET_ARM64
#endif // FEATURE_MASKED_HW_INTRINSICS

switch (ni)
{
Expand Down Expand Up @@ -33559,7 +33637,7 @@ GenTree* Compiler::gtFoldExprHWIntrinsic(GenTreeHWIntrinsic* tree)
// op2 = op2 & op1
op2->AsVecCon()->EvaluateBinaryInPlace(GT_AND, false, simdBaseType, op1->AsVecCon());

// op3 = op2 & ~op1
// op3 = op3 & ~op1
op3->AsVecCon()->EvaluateBinaryInPlace(GT_AND_NOT, false, simdBaseType, op1->AsVecCon());

// op2 = op2 | op3
Expand All @@ -33572,8 +33650,8 @@ GenTree* Compiler::gtFoldExprHWIntrinsic(GenTreeHWIntrinsic* tree)

#if defined(TARGET_ARM64)
case NI_Sve_ConditionalSelect:
case NI_Sve_ConditionalSelect_Predicates:
{
assert(!varTypeIsMask(retType));
assert(varTypeIsMask(op1));

if (cnsNode != op1)
Expand Down Expand Up @@ -33602,10 +33680,11 @@ GenTree* Compiler::gtFoldExprHWIntrinsic(GenTreeHWIntrinsic* tree)

if (op2->IsCnsVec() && op3->IsCnsVec())
{
assert(ni == NI_Sve_ConditionalSelect);
assert(op2->gtType == TYP_SIMD16);
assert(op3->gtType == TYP_SIMD16);

simd16_t op1SimdVal;
simd16_t op1SimdVal = {};
EvaluateSimdCvtMaskToVector<simd16_t>(simdBaseType, &op1SimdVal, op1->AsMskCon()->gtSimdMaskVal);

// op2 = op2 & op1
Expand All @@ -33614,7 +33693,7 @@ GenTree* Compiler::gtFoldExprHWIntrinsic(GenTreeHWIntrinsic* tree)
op1SimdVal);
op2->AsVecCon()->gtSimd16Val = result;

// op3 = op2 & ~op1
// op3 = op3 & ~op1
result = {};
EvaluateBinarySimd<simd16_t>(GT_AND_NOT, false, simdBaseType, &result, op3->AsVecCon()->gtSimd16Val,
op1SimdVal);
Expand All @@ -33625,6 +33704,30 @@ GenTree* Compiler::gtFoldExprHWIntrinsic(GenTreeHWIntrinsic* tree)

resultNode = op2;
}
else if (op2->IsCnsMsk() && op3->IsCnsMsk())
{
assert(ni == NI_Sve_ConditionalSelect_Predicates);

// op2 = op2 & op1
simdmask_t result = {};
EvaluateBinaryMask<simd16_t>(GT_AND, false, simdBaseType, &result, op2->AsMskCon()->gtSimdMaskVal,
op1->AsMskCon()->gtSimdMaskVal);
op2->AsMskCon()->gtSimdMaskVal = result;

// op3 = op3 & ~op1
result = {};
EvaluateBinaryMask<simd16_t>(GT_AND_NOT, false, simdBaseType, &result,
op3->AsMskCon()->gtSimdMaskVal, op1->AsMskCon()->gtSimdMaskVal);
op3->AsMskCon()->gtSimdMaskVal = result;

// op2 = op2 | op3
result = {};
EvaluateBinaryMask<simd16_t>(GT_OR, false, simdBaseType, &result, op2->AsMskCon()->gtSimdMaskVal,
op3->AsMskCon()->gtSimdMaskVal);
op2->AsMskCon()->gtSimdMaskVal = result;

resultNode = op2;
}
break;
}
#endif // TARGET_ARM64
Expand Down
6 changes: 6 additions & 0 deletions src/coreclr/jit/simd.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1090,6 +1090,12 @@ void EvaluateBinaryMask(
break;
}

case 16:
{
bitMask = 0x0001000100010001;
break;
}

default:
{
unreached();
Expand Down
Loading
Loading