Skip to content

[AArch64] Combine PTEST_FIRST(PTRUE, CONCAT(A, B)) -> PTEST_FIRST(PTRUE, A) #161384

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
kmclaughlin-arm merged 4 commits into from
Oct 2, 2025
Merged

[AArch64] Combine PTEST_FIRST(PTRUE, CONCAT(A, B)) -> PTEST_FIRST(PTRUE, A) #161384

Show file tree
Hide file tree
Changes from 3 commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
a3167b5
[AArch64] Combine PTEST_FIRST(PTRUE, CONCAT(A, B)) -> PTEST_FIRST(PTR…
kmclaughlin-arm Sep 24, 2025
b99f48c
- Add isLane1KnownActive helper
kmclaughlin-arm Oct 1, 2025
c5402af
- Rename isLane1KnownActive -> isLane0KnownActive
kmclaughlin-arm Oct 1, 2025
095b1c3
- Move isLane0KnownActive before looking through reinterpret_cast
kmclaughlin-arm Oct 1, 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
43 changes: 43 additions & 0 deletions llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -27234,6 +27234,21 @@ static bool isLanes1toNKnownZero(SDValue Op) {
}
}

// Return true if the vector operation can guarantee that the first lane of its
// result is active.
static bool isLane0KnownActive(SDValue Op) {
switch (Op.getOpcode()) {
default:
return false;
case AArch64ISD::REINTERPRET_CAST:
return isLane0KnownActive(Op->getOperand(0));
case ISD::SPLAT_VECTOR:
return isOneConstant(Op.getOperand(0));
case AArch64ISD::PTRUE:
return Op.getConstantOperandVal(0) == AArch64SVEPredPattern::all;
};
}

static SDValue removeRedundantInsertVectorElt(SDNode *N) {
assert(N->getOpcode() == ISD::INSERT_VECTOR_ELT && "Unexpected node!");
SDValue InsertVec = N->getOperand(0);
Expand Down Expand Up @@ -27519,6 +27534,32 @@ static SDValue performMULLCombine(SDNode *N,
return SDValue();
}

static SDValue performPTestFirstCombine(SDNode *N,
TargetLowering::DAGCombinerInfo &DCI,
SelectionDAG &DAG) {
if (DCI.isBeforeLegalize())
return SDValue();
Copy link
Collaborator

@paulwalker-arm paulwalker-arm Sep 30, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

You don't need the element type check because AArch64ISD::REINTERPRET_CAST is only allowed to change the element count.


SDLoc DL(N);
auto Mask = N->getOperand(0);
auto Pred = N->getOperand(1);

if (Pred->getOpcode() == AArch64ISD::REINTERPRET_CAST)
Pred = Pred->getOperand(0);

if (!isLane0KnownActive(Mask))
return SDValue();
Comment on lines 27547 to 27548
Copy link
Collaborator

@paulwalker-arm paulwalker-arm Oct 1, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Perhaps move this before looking through Pred's reinterpret_cast because this is what makes that code logically safe.


if (Pred->getOpcode() == ISD::CONCAT_VECTORS) {
Pred = Pred->getOperand(0);
Pred = DAG.getNode(AArch64ISD::REINTERPRET_CAST, DL, MVT::nxv16i1, Pred);
return DAG.getNode(AArch64ISD::PTEST_FIRST, DL, N->getValueType(0), Mask,
Pred);
}

return SDValue();
}

static SDValue
performScalarToVectorCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
SelectionDAG &DAG) {
Expand Down Expand Up @@ -27875,6 +27916,8 @@ SDValue AArch64TargetLowering::PerformDAGCombine(SDNode *N,
case AArch64ISD::UMULL:
case AArch64ISD::PMULL:
return performMULLCombine(N, DCI, DAG);
case AArch64ISD::PTEST_FIRST:
return performPTestFirstCombine(N, DCI, DAG);
case ISD::INTRINSIC_VOID:
case ISD::INTRINSIC_W_CHAIN:
switch (N->getConstantOperandVal(1)) {
Expand Down
4 changes: 4 additions & 0 deletions llvm/lib/Target/AArch64/AArch64InstrInfo.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1503,6 +1503,10 @@ AArch64InstrInfo::canRemovePTestInstr(MachineInstr *PTest, MachineInstr *Mask,
getElementSizeForOpcode(PredOpcode))
return PredOpcode;

if (PTest->getOpcode() == AArch64::PTEST_PP_FIRST &&
Copy link
Collaborator

@paulwalker-arm paulwalker-arm Oct 1, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Perhaps add a comment to match the other cases?

isPTrueOpcode(MaskOpcode) && Mask->getOperand(1).getImm() == 31)
return PredOpcode;

return {};
}

Expand Down
20 changes: 2 additions & 18 deletions llvm/test/CodeGen/AArch64/get-active-lane-mask-extract.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -327,9 +327,6 @@ define void @test_2x8bit_mask_with_extracts_and_ptest(i64 %i, i64 %n) {
; CHECK-SVE2p1-SME2-LABEL: test_2x8bit_mask_with_extracts_and_ptest:
; CHECK-SVE2p1-SME2: // %bb.0: // %entry
; CHECK-SVE2p1-SME2-NEXT: whilelo { p0.h, p1.h }, x0, x1
; CHECK-SVE2p1-SME2-NEXT: ptrue p2.b
; CHECK-SVE2p1-SME2-NEXT: uzp1 p3.b, p0.b, p1.b
; CHECK-SVE2p1-SME2-NEXT: ptest p2, p3.b
; CHECK-SVE2p1-SME2-NEXT: b.pl .LBB11_2
; CHECK-SVE2p1-SME2-NEXT: // %bb.1: // %if.then
; CHECK-SVE2p1-SME2-NEXT: b use
Expand Down Expand Up @@ -368,9 +365,6 @@ define void @test_2x8bit_mask_with_extracts_and_reinterpret_casts(i64 %i, i64 %n
; CHECK-SVE2p1-SME2-LABEL: test_2x8bit_mask_with_extracts_and_reinterpret_casts:
; CHECK-SVE2p1-SME2: // %bb.0: // %entry
; CHECK-SVE2p1-SME2-NEXT: whilelo { p0.s, p1.s }, x0, x1
; CHECK-SVE2p1-SME2-NEXT: ptrue p2.h
; CHECK-SVE2p1-SME2-NEXT: uzp1 p3.h, p0.h, p1.h
; CHECK-SVE2p1-SME2-NEXT: ptest p2, p3.b
; CHECK-SVE2p1-SME2-NEXT: b.pl .LBB12_2
; CHECK-SVE2p1-SME2-NEXT: // %bb.1: // %if.then
; CHECK-SVE2p1-SME2-NEXT: b use
Expand Down Expand Up @@ -413,14 +407,9 @@ define void @test_4x4bit_mask_with_extracts_and_ptest(i64 %i, i64 %n) {
; CHECK-SVE2p1-SME2-NEXT: adds x8, x0, x8
; CHECK-SVE2p1-SME2-NEXT: csinv x8, x8, xzr, lo
; CHECK-SVE2p1-SME2-NEXT: whilelo { p0.s, p1.s }, x0, x1
; CHECK-SVE2p1-SME2-NEXT: whilelo { p2.s, p3.s }, x8, x1
; CHECK-SVE2p1-SME2-NEXT: uzp1 p4.h, p0.h, p1.h
; CHECK-SVE2p1-SME2-NEXT: uzp1 p5.h, p2.h, p3.h
; CHECK-SVE2p1-SME2-NEXT: uzp1 p4.b, p4.b, p5.b
; CHECK-SVE2p1-SME2-NEXT: ptrue p5.b
; CHECK-SVE2p1-SME2-NEXT: ptest p5, p4.b
; CHECK-SVE2p1-SME2-NEXT: b.pl .LBB13_2
; CHECK-SVE2p1-SME2-NEXT: // %bb.1: // %if.then
; CHECK-SVE2p1-SME2-NEXT: whilelo { p2.s, p3.s }, x8, x1
; CHECK-SVE2p1-SME2-NEXT: b use
; CHECK-SVE2p1-SME2-NEXT: .LBB13_2: // %if.end
; CHECK-SVE2p1-SME2-NEXT: ret
Expand Down Expand Up @@ -463,14 +452,9 @@ define void @test_4x2bit_mask_with_extracts_and_reinterpret_casts(i64 %i, i64 %n
; CHECK-SVE2p1-SME2-NEXT: adds x8, x0, x8
; CHECK-SVE2p1-SME2-NEXT: csinv x8, x8, xzr, lo
; CHECK-SVE2p1-SME2-NEXT: whilelo { p0.d, p1.d }, x0, x1
; CHECK-SVE2p1-SME2-NEXT: whilelo { p2.d, p3.d }, x8, x1
; CHECK-SVE2p1-SME2-NEXT: uzp1 p4.s, p0.s, p1.s
; CHECK-SVE2p1-SME2-NEXT: uzp1 p5.s, p2.s, p3.s
; CHECK-SVE2p1-SME2-NEXT: uzp1 p4.h, p4.h, p5.h
; CHECK-SVE2p1-SME2-NEXT: ptrue p5.h
; CHECK-SVE2p1-SME2-NEXT: ptest p5, p4.b
; CHECK-SVE2p1-SME2-NEXT: b.pl .LBB14_2
; CHECK-SVE2p1-SME2-NEXT: // %bb.1: // %if.then
; CHECK-SVE2p1-SME2-NEXT: whilelo { p2.d, p3.d }, x8, x1
; CHECK-SVE2p1-SME2-NEXT: b use
; CHECK-SVE2p1-SME2-NEXT: .LBB14_2: // %if.end
; CHECK-SVE2p1-SME2-NEXT: ret
Expand Down