Skip to content

Commit 22ec44f

Browse files
david-armdavid-arm
authored
[DAGCombiner] Add support for scalarising extracts of a vector setcc (llvm#116031)
For IR like this: %icmp = icmp ult <4 x i32> %a, splat (i32 5) %res = extractelement <4 x i1> %icmp, i32 1 where there is only one use of %icmp we can take a similar approach to what we already do for binary ops such add, sub, etc. and convert this into %ext = extractelement <4 x i32> %a, i32 1 %res = icmp ult i32 %ext, 5 For AArch64 targets at least the scalar boolean result will almost certainly need to be in a GPR anyway, since it will probably be used by branches for control flow. I've tried to reuse existing code in scalarizeExtractedBinop to also work for setcc. NOTE: The optimisations don't apply for tests such as extract_icmp_v4i32_splat_rhs in the file CodeGen/AArch64/extract-vector-cmp.ll because scalarizeExtractedBinOp only works if one of the input operands is a constant.
1 parent e5faeb6 commit 22ec44f

File tree

7 files changed

+281
-41
lines changed

7 files changed

+281
-41
lines changed

‎llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp‎

Lines changed: 25 additions & 16 deletions
Original file line numberDiff line numberDiff line change
@@ -22746,16 +22746,21 @@ SDValue DAGCombiner::scalarizeExtractedVectorLoad(SDNode *EVE, EVT InVecVT,
2274622746

2274722747
/// Transform a vector binary operation into a scalar binary operation by moving
2274822748
/// the math/logic after an extract element of a vector.
22749-
static SDValue scalarizeExtractedBinop(SDNode *ExtElt, SelectionDAG &DAG,
22750-
const SDLoc &DL, bool LegalOperations) {
22749+
static SDValue scalarizeExtractedBinOp(SDNode *ExtElt, SelectionDAG &DAG,
22750+
const SDLoc &DL) {
2275122751
const TargetLowering &TLI = DAG.getTargetLoweringInfo();
2275222752
SDValue Vec = ExtElt->getOperand(0);
2275322753
SDValue Index = ExtElt->getOperand(1);
2275422754
auto *IndexC = dyn_cast<ConstantSDNode>(Index);
22755-
if (!IndexC || !TLI.isBinOp(Vec.getOpcode()) || !Vec.hasOneUse() ||
22755+
unsigned Opc = Vec.getOpcode();
22756+
if (!IndexC || (!TLI.isBinOp(Opc) && Opc != ISD::SETCC) || !Vec.hasOneUse() ||
2275622757
Vec->getNumValues() != 1)
2275722758
return SDValue();
2275822759

22760+
EVT ResVT = ExtElt->getValueType(0);
22761+
if (Opc == ISD::SETCC && ResVT != Vec.getValueType().getVectorElementType())
22762+
return SDValue();
22763+
2275922764
// Targets may want to avoid this to prevent an expensive register transfer.
2276022765
if (!TLI.shouldScalarizeBinop(Vec))
2276122766
return SDValue();
@@ -22766,19 +22771,23 @@ static SDValue scalarizeExtractedBinop(SDNode *ExtElt, SelectionDAG &DAG,
2276622771
SDValue Op0 = Vec.getOperand(0);
2276722772
SDValue Op1 = Vec.getOperand(1);
2276822773
APInt SplatVal;
22769-
if (isAnyConstantBuildVector(Op0, true) ||
22770-
ISD::isConstantSplatVector(Op0.getNode(), SplatVal) ||
22771-
isAnyConstantBuildVector(Op1, true) ||
22772-
ISD::isConstantSplatVector(Op1.getNode(), SplatVal)) {
22773-
// extractelt (binop X, C), IndexC --> binop (extractelt X, IndexC), C'
22774-
// extractelt (binop C, X), IndexC --> binop C', (extractelt X, IndexC)
22775-
EVT VT = ExtElt->getValueType(0);
22776-
SDValue Ext0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, Op0, Index);
22777-
SDValue Ext1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, Op1, Index);
22778-
return DAG.getNode(Vec.getOpcode(), DL, VT, Ext0, Ext1);
22779-
}
22774+
if (!isAnyConstantBuildVector(Op0, true) &&
22775+
!ISD::isConstantSplatVector(Op0.getNode(), SplatVal) &&
22776+
!isAnyConstantBuildVector(Op1, true) &&
22777+
!ISD::isConstantSplatVector(Op1.getNode(), SplatVal))
22778+
return SDValue();
2278022779

22781-
return SDValue();
22780+
// extractelt (op X, C), IndexC --> op (extractelt X, IndexC), C'
22781+
// extractelt (op C, X), IndexC --> op C', (extractelt X, IndexC)
22782+
EVT OpVT = Op0->getValueType(0).getVectorElementType();
22783+
Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, OpVT, Op0, Index);
22784+
Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, OpVT, Op1, Index);
22785+
22786+
if (Opc == ISD::SETCC)
22787+
return DAG.getSetCC(DL, ResVT, Op0, Op1,
22788+
cast<CondCodeSDNode>(Vec->getOperand(2))->get());
22789+
else
22790+
return DAG.getNode(Opc, DL, ResVT, Op0, Op1);
2278222791
}
2278322792

2278422793
// Given a ISD::EXTRACT_VECTOR_ELT, which is a glorified bit sequence extract,
@@ -23011,7 +23020,7 @@ SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) {
2301123020
}
2301223021
}
2301323022

23014-
if (SDValue BO = scalarizeExtractedBinop(N, DAG, DL, LegalOperations))
23023+
if (SDValue BO = scalarizeExtractedBinOp(N, DAG, DL))
2301523024
return BO;
2301623025

2301723026
if (VecVT.isScalableVector())

‎llvm/lib/Target/AArch64/AArch64ISelLowering.h‎

Lines changed: 4 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1348,6 +1348,10 @@ class AArch64TargetLowering : public TargetLowering {
13481348
unsigned getMinimumJumpTableEntries() const override;
13491349

13501350
bool softPromoteHalfType() const override { return true; }
1351+
1352+
bool shouldScalarizeBinop(SDValue VecOp) const override {
1353+
return VecOp.getOpcode() == ISD::SETCC;
1354+
}
13511355
};
13521356

13531357
namespace AArch64 {

‎llvm/lib/Target/RISCV/RISCVISelLowering.cpp‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -2093,7 +2093,7 @@ bool RISCVTargetLowering::shouldScalarizeBinop(SDValue VecOp) const {
20932093

20942094
// Assume target opcodes can't be scalarized.
20952095
// TODO - do we have any exceptions?
2096-
if (Opc >= ISD::BUILTIN_OP_END)
2096+
if (Opc >= ISD::BUILTIN_OP_END || !isBinOp(Opc))
20972097
return false;
20982098

20992099
// If the vector op is not supported, try to convert to scalar.

‎llvm/lib/Target/WebAssembly/WebAssemblyISelLowering.cpp‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -429,7 +429,7 @@ bool WebAssemblyTargetLowering::shouldScalarizeBinop(SDValue VecOp) const {
429429

430430
// Assume target opcodes can't be scalarized.
431431
// TODO - do we have any exceptions?
432-
if (Opc >= ISD::BUILTIN_OP_END)
432+
if (Opc >= ISD::BUILTIN_OP_END || !isBinOp(Opc))
433433
return false;
434434

435435
// If the vector op is not supported, try to convert to scalar.

‎llvm/lib/Target/X86/X86ISelLowering.cpp‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -3300,7 +3300,7 @@ bool X86TargetLowering::shouldScalarizeBinop(SDValue VecOp) const {
33003300

33013301
// Assume target opcodes can't be scalarized.
33023302
// TODO - do we have any exceptions?
3303-
if (Opc >= ISD::BUILTIN_OP_END)
3303+
if (Opc >= ISD::BUILTIN_OP_END || !isBinOp(Opc))
33043304
return false;
33053305

33063306
// If the vector op is not supported, try to convert to scalar.

‎llvm/test/CodeGen/AArch64/dag-combine-concat-vectors.ll‎

Lines changed: 24 additions & 22 deletions
Original file line numberDiff line numberDiff line change
@@ -5,7 +5,7 @@
55

66
declare void @llvm.masked.scatter.nxv16i8.nxv16p0(<vscale x 16 x i8>, <vscale x 16 x ptr>, i32 immarg, <vscale x 16 x i1>)
77

8-
define fastcc i8 @allocno_reload_assign() {
8+
define fastcc i8 @allocno_reload_assign(ptr %p) {
99
; CHECK-LABEL: allocno_reload_assign:
1010
; CHECK: // %bb.0:
1111
; CHECK-NEXT: fmov d0, xzr
@@ -14,8 +14,8 @@ define fastcc i8 @allocno_reload_assign() {
1414
; CHECK-NEXT: cmpeq p0.d, p0/z, z0.d, #0
1515
; CHECK-NEXT: uzp1 p0.s, p0.s, p0.s
1616
; CHECK-NEXT: uzp1 p0.h, p0.h, p0.h
17-
; CHECK-NEXT: uzp1 p0.b, p0.b, p0.b
18-
; CHECK-NEXT: mov z0.b, p0/z, #1 // =0x1
17+
; CHECK-NEXT: uzp1 p8.b, p0.b, p0.b
18+
; CHECK-NEXT: mov z0.b, p8/z, #1 // =0x1
1919
; CHECK-NEXT: fmov w8, s0
2020
; CHECK-NEXT: mov z0.b, #0 // =0x0
2121
; CHECK-NEXT: uunpklo z1.h, z0.b
@@ -30,34 +30,35 @@ define fastcc i8 @allocno_reload_assign() {
3030
; CHECK-NEXT: punpklo p1.h, p0.b
3131
; CHECK-NEXT: punpkhi p0.h, p0.b
3232
; CHECK-NEXT: punpklo p2.h, p1.b
33-
; CHECK-NEXT: punpkhi p3.h, p1.b
33+
; CHECK-NEXT: punpkhi p4.h, p1.b
3434
; CHECK-NEXT: uunpklo z0.d, z2.s
3535
; CHECK-NEXT: uunpkhi z1.d, z2.s
36-
; CHECK-NEXT: punpklo p5.h, p0.b
36+
; CHECK-NEXT: punpklo p6.h, p0.b
3737
; CHECK-NEXT: uunpklo z2.d, z3.s
3838
; CHECK-NEXT: uunpkhi z3.d, z3.s
39-
; CHECK-NEXT: punpkhi p7.h, p0.b
39+
; CHECK-NEXT: punpkhi p0.h, p0.b
4040
; CHECK-NEXT: uunpklo z4.d, z5.s
4141
; CHECK-NEXT: uunpkhi z5.d, z5.s
4242
; CHECK-NEXT: uunpklo z6.d, z7.s
4343
; CHECK-NEXT: uunpkhi z7.d, z7.s
44-
; CHECK-NEXT: punpklo p0.h, p2.b
45-
; CHECK-NEXT: punpkhi p1.h, p2.b
46-
; CHECK-NEXT: punpklo p2.h, p3.b
47-
; CHECK-NEXT: punpkhi p3.h, p3.b
48-
; CHECK-NEXT: punpklo p4.h, p5.b
49-
; CHECK-NEXT: punpkhi p5.h, p5.b
50-
; CHECK-NEXT: punpklo p6.h, p7.b
51-
; CHECK-NEXT: punpkhi p7.h, p7.b
44+
; CHECK-NEXT: punpklo p1.h, p2.b
45+
; CHECK-NEXT: punpkhi p2.h, p2.b
46+
; CHECK-NEXT: punpklo p3.h, p4.b
47+
; CHECK-NEXT: punpkhi p4.h, p4.b
48+
; CHECK-NEXT: punpklo p5.h, p6.b
49+
; CHECK-NEXT: punpkhi p6.h, p6.b
50+
; CHECK-NEXT: punpklo p7.h, p0.b
51+
; CHECK-NEXT: punpkhi p0.h, p0.b
5252
; CHECK-NEXT: .LBB0_1: // =>This Inner Loop Header: Depth=1
53-
; CHECK-NEXT: st1b { z0.d }, p0, [z16.d]
54-
; CHECK-NEXT: st1b { z1.d }, p1, [z16.d]
55-
; CHECK-NEXT: st1b { z2.d }, p2, [z16.d]
56-
; CHECK-NEXT: st1b { z3.d }, p3, [z16.d]
57-
; CHECK-NEXT: st1b { z4.d }, p4, [z16.d]
58-
; CHECK-NEXT: st1b { z5.d }, p5, [z16.d]
59-
; CHECK-NEXT: st1b { z6.d }, p6, [z16.d]
60-
; CHECK-NEXT: st1b { z7.d }, p7, [z16.d]
53+
; CHECK-NEXT: st1b { z0.d }, p1, [z16.d]
54+
; CHECK-NEXT: st1b { z1.d }, p2, [z16.d]
55+
; CHECK-NEXT: st1b { z2.d }, p3, [z16.d]
56+
; CHECK-NEXT: st1b { z3.d }, p4, [z16.d]
57+
; CHECK-NEXT: st1b { z4.d }, p5, [z16.d]
58+
; CHECK-NEXT: st1b { z5.d }, p6, [z16.d]
59+
; CHECK-NEXT: st1b { z6.d }, p7, [z16.d]
60+
; CHECK-NEXT: st1b { z7.d }, p0, [z16.d]
61+
; CHECK-NEXT: str p8, [x0]
6162
; CHECK-NEXT: b .LBB0_1
6263
br label %1
6364

@@ -66,6 +67,7 @@ define fastcc i8 @allocno_reload_assign() {
6667
%constexpr1 = shufflevector <vscale x 16 x i1> %constexpr, <vscale x 16 x i1> poison, <vscale x 16 x i32> zeroinitializer
6768
%constexpr2 = xor <vscale x 16 x i1> %constexpr1, shufflevector (<vscale x 16 x i1> insertelement (<vscale x 16 x i1> poison, i1 true, i64 0), <vscale x 16 x i1> poison, <vscale x 16 x i32> zeroinitializer)
6869
call void @llvm.masked.scatter.nxv16i8.nxv16p0(<vscale x 16 x i8> zeroinitializer, <vscale x 16 x ptr> zeroinitializer, i32 0, <vscale x 16 x i1> %constexpr2)
70+
store <vscale x 16 x i1> %constexpr, ptr %p, align 16
6971
br label %1
7072
}
7173

0 commit comments

Comments
 (0)