[llvm-branch-commits] [llvm] [LoongArch] Select [x]vbitset/[x]vbitrev for or/xor of bitcasted shl(splat(1), y) (PR #241239)
via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Thu Nov 19 00:13:33 PST 2026
https://github.com/heiher updated https://github.com/llvm/llvm-project/pull/241239
>From 89a1690fbe77f30802edf7259ae6a916f5e95839 Mon Sep 17 00:00:00 2001
From: WANG Rui <wangrui at loongson.cn>
Date: Wed, 11 Nov 2026 15:33:36 +0800
Subject: [PATCH 1/2] [LoongArch] Select [x]vbitset/[x]vbitrev for or/xor of
bitcasted shl(splat(1), y)
When the shift of splat(1) is performed in a narrower element type than
the or/xor, e.g.
```llvm
%s = shl <8 x i16> splat (i16 1), %y
%m = bitcast <8 x i16> %s to <2 x i64>
%r = xor <2 x i64> %x, %m
```
the existing [x]vbitset.{b/h/w} and [x]vbitrev.{b/h/w} patterns fail to
match and we emit vrepli + vsll + vor.v/vxor.v.
Fold OP(bitcast(shl(splat(1), y)), x) into
bitcast(OP(shl(splat(1), y), bitcast(x))) for OP in {or, xor} so that
the operation is done in the shift's element type and selected as a
single [x]vbitset or [x]vbitrev.
Assisted-By: LLM
---
.../LoongArch/LoongArchISelLowering.cpp | 39 +++++++++++++++++++
llvm/test/CodeGen/LoongArch/lasx/bitrev.ll | 24 +++---------
llvm/test/CodeGen/LoongArch/lasx/bitset.ll | 24 +++---------
llvm/test/CodeGen/LoongArch/lsx/bitrev.ll | 24 +++---------
llvm/test/CodeGen/LoongArch/lsx/bitset.ll | 24 +++---------
5 files changed, 63 insertions(+), 72 deletions(-)
diff --git a/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp b/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp
index c263ed4570bde0..9fa102ae8abe12 100644
--- a/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp
+++ b/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp
@@ -7415,6 +7415,39 @@ static SDValue performBITCASTCombine(SDNode *N, SelectionDAG &DAG,
return DAG.getBitcast(VT, V);
}
+// OP(bitcast(shl(splat(1), y)), x) -> bitcast(OP(shl(splat(1), y), bitcast(x)))
+// for OP in {or, xor}. Perform the bitwise operation in the element type of
+// the shift so that it can be selected as [x]vbitset.{b/h/w/d} or
+// [x]vbitrev.{b/h/w/d}.
+static SDValue combineVectorBitOpOfBitcastShl(SDNode *N, SelectionDAG &DAG) {
+ EVT VT = N->getValueType(0);
+ const TargetLowering &TLI = DAG.getTargetLoweringInfo();
+ if (!VT.isVector() || !TLI.isTypeLegal(VT))
+ return SDValue();
+
+ for (unsigned I = 0; I < 2; ++I) {
+ SDValue N0 = N->getOperand(I);
+ SDValue N1 = N->getOperand(1 - I);
+ if (N0.getOpcode() != ISD::BITCAST || !N0.hasOneUse())
+ continue;
+ SDValue Shl = N0.getOperand(0);
+ EVT ShlVT = Shl.getValueType();
+ APInt SplatVal;
+ if (Shl.getOpcode() == ISD::SHL && ShlVT.isVector() && ShlVT.isInteger() &&
+ ShlVT != VT && ShlVT.getSizeInBits() == VT.getSizeInBits() &&
+ TLI.isTypeLegal(ShlVT) &&
+ ISD::isConstantSplatVector(Shl.getOperand(0).getNode(), SplatVal) &&
+ SplatVal.isOne()) {
+ SDLoc DL(N);
+ SDValue V = DAG.getNode(N->getOpcode(), DL, ShlVT, Shl,
+ DAG.getBitcast(ShlVT, N1));
+ return DAG.getBitcast(VT, V);
+ }
+ }
+
+ return SDValue();
+}
+
static SDValue performORCombine(SDNode *N, SelectionDAG &DAG,
TargetLowering::DAGCombinerInfo &DCI,
const LoongArchSubtarget &Subtarget) {
@@ -7428,6 +7461,9 @@ static SDValue performORCombine(SDNode *N, SelectionDAG &DAG,
unsigned Shamt;
bool SwapAndRetried = false;
+ if (SDValue V = combineVectorBitOpOfBitcastShl(N, DAG))
+ return V;
+
// BSTRPICK requires the 32S feature.
if (!Subtarget.has32S())
return SDValue();
@@ -7641,6 +7677,9 @@ static SDValue performORCombine(SDNode *N, SelectionDAG &DAG,
static SDValue performXORCombine(SDNode *N, SelectionDAG &DAG,
const LoongArchSubtarget &Subtarget) {
+ if (SDValue V = combineVectorBitOpOfBitcastShl(N, DAG))
+ return V;
+
EVT VT = N->getValueType(0);
// LA64 promotes i32 integer operations to GPR-width operations during
// legalization. Match the idiom before that promotion as well as after
diff --git a/llvm/test/CodeGen/LoongArch/lasx/bitrev.ll b/llvm/test/CodeGen/LoongArch/lasx/bitrev.ll
index 46659f10db998f..c9b9c01f4f70d5 100644
--- a/llvm/test/CodeGen/LoongArch/lasx/bitrev.ll
+++ b/llvm/test/CodeGen/LoongArch/lasx/bitrev.ll
@@ -95,9 +95,7 @@ entry:
define <4 x i64> @xvbitrev_b(<4 x i64> %a, <4 x i64> %b) nounwind {
; CHECK-LABEL: xvbitrev_b:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: xvrepli.b $xr2, 1
-; CHECK-NEXT: xvsll.b $xr1, $xr2, $xr1
-; CHECK-NEXT: xvxor.v $xr0, $xr0, $xr1
+; CHECK-NEXT: xvbitrev.b $xr0, $xr0, $xr1
; CHECK-NEXT: ret
entry:
%0 = bitcast <4 x i64> %b to <32 x i8>
@@ -111,9 +109,7 @@ entry:
define <4 x i64> @xvbitrev_b_nomask(<4 x i64> %a, <4 x i64> %b) nounwind {
; CHECK-LABEL: xvbitrev_b_nomask:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: xvrepli.b $xr2, 1
-; CHECK-NEXT: xvsll.b $xr1, $xr2, $xr1
-; CHECK-NEXT: xvxor.v $xr0, $xr1, $xr0
+; CHECK-NEXT: xvbitrev.b $xr0, $xr0, $xr1
; CHECK-NEXT: ret
entry:
%0 = bitcast <4 x i64> %b to <32 x i8>
@@ -126,9 +122,7 @@ entry:
define <4 x i64> @xvbitrev_h(<4 x i64> %a, <4 x i64> %b) nounwind {
; CHECK-LABEL: xvbitrev_h:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: xvrepli.h $xr2, 1
-; CHECK-NEXT: xvsll.h $xr1, $xr2, $xr1
-; CHECK-NEXT: xvxor.v $xr0, $xr0, $xr1
+; CHECK-NEXT: xvbitrev.h $xr0, $xr0, $xr1
; CHECK-NEXT: ret
entry:
%0 = bitcast <4 x i64> %b to <16 x i16>
@@ -142,9 +136,7 @@ entry:
define <4 x i64> @xvbitrev_h_nomask(<4 x i64> %a, <4 x i64> %b) nounwind {
; CHECK-LABEL: xvbitrev_h_nomask:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: xvrepli.h $xr2, 1
-; CHECK-NEXT: xvsll.h $xr1, $xr2, $xr1
-; CHECK-NEXT: xvxor.v $xr0, $xr1, $xr0
+; CHECK-NEXT: xvbitrev.h $xr0, $xr0, $xr1
; CHECK-NEXT: ret
entry:
%0 = bitcast <4 x i64> %b to <16 x i16>
@@ -157,9 +149,7 @@ entry:
define <4 x i64> @xvbitrev_w(<4 x i64> %a, <4 x i64> %b) nounwind {
; CHECK-LABEL: xvbitrev_w:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: xvrepli.w $xr2, 1
-; CHECK-NEXT: xvsll.w $xr1, $xr2, $xr1
-; CHECK-NEXT: xvxor.v $xr0, $xr0, $xr1
+; CHECK-NEXT: xvbitrev.w $xr0, $xr0, $xr1
; CHECK-NEXT: ret
entry:
%0 = bitcast <4 x i64> %b to <8 x i32>
@@ -173,9 +163,7 @@ entry:
define <4 x i64> @xvbitrev_w_nomask(<4 x i64> %a, <4 x i64> %b) nounwind {
; CHECK-LABEL: xvbitrev_w_nomask:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: xvrepli.w $xr2, 1
-; CHECK-NEXT: xvsll.w $xr1, $xr2, $xr1
-; CHECK-NEXT: xvxor.v $xr0, $xr1, $xr0
+; CHECK-NEXT: xvbitrev.w $xr0, $xr0, $xr1
; CHECK-NEXT: ret
entry:
%0 = bitcast <4 x i64> %b to <8 x i32>
diff --git a/llvm/test/CodeGen/LoongArch/lasx/bitset.ll b/llvm/test/CodeGen/LoongArch/lasx/bitset.ll
index 3539304bfa12bf..6a2defa10ce57c 100644
--- a/llvm/test/CodeGen/LoongArch/lasx/bitset.ll
+++ b/llvm/test/CodeGen/LoongArch/lasx/bitset.ll
@@ -95,9 +95,7 @@ entry:
define <4 x i64> @xvbitset_b(<4 x i64> %a, <4 x i64> %b) nounwind {
; CHECK-LABEL: xvbitset_b:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: xvrepli.b $xr2, 1
-; CHECK-NEXT: xvsll.b $xr1, $xr2, $xr1
-; CHECK-NEXT: xvor.v $xr0, $xr0, $xr1
+; CHECK-NEXT: xvbitset.b $xr0, $xr0, $xr1
; CHECK-NEXT: ret
entry:
%0 = bitcast <4 x i64> %b to <32 x i8>
@@ -111,9 +109,7 @@ entry:
define <4 x i64> @xvbitset_b_nomask(<4 x i64> %a, <4 x i64> %b) nounwind {
; CHECK-LABEL: xvbitset_b_nomask:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: xvrepli.b $xr2, 1
-; CHECK-NEXT: xvsll.b $xr1, $xr2, $xr1
-; CHECK-NEXT: xvor.v $xr0, $xr1, $xr0
+; CHECK-NEXT: xvbitset.b $xr0, $xr0, $xr1
; CHECK-NEXT: ret
entry:
%0 = bitcast <4 x i64> %b to <32 x i8>
@@ -126,9 +122,7 @@ entry:
define <4 x i64> @xvbitset_h(<4 x i64> %a, <4 x i64> %b) nounwind {
; CHECK-LABEL: xvbitset_h:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: xvrepli.h $xr2, 1
-; CHECK-NEXT: xvsll.h $xr1, $xr2, $xr1
-; CHECK-NEXT: xvor.v $xr0, $xr0, $xr1
+; CHECK-NEXT: xvbitset.h $xr0, $xr0, $xr1
; CHECK-NEXT: ret
entry:
%0 = bitcast <4 x i64> %b to <16 x i16>
@@ -142,9 +136,7 @@ entry:
define <4 x i64> @xvbitset_h_nomask(<4 x i64> %a, <4 x i64> %b) nounwind {
; CHECK-LABEL: xvbitset_h_nomask:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: xvrepli.h $xr2, 1
-; CHECK-NEXT: xvsll.h $xr1, $xr2, $xr1
-; CHECK-NEXT: xvor.v $xr0, $xr1, $xr0
+; CHECK-NEXT: xvbitset.h $xr0, $xr0, $xr1
; CHECK-NEXT: ret
entry:
%0 = bitcast <4 x i64> %b to <16 x i16>
@@ -157,9 +149,7 @@ entry:
define <4 x i64> @xvbitset_w(<4 x i64> %a, <4 x i64> %b) nounwind {
; CHECK-LABEL: xvbitset_w:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: xvrepli.w $xr2, 1
-; CHECK-NEXT: xvsll.w $xr1, $xr2, $xr1
-; CHECK-NEXT: xvor.v $xr0, $xr0, $xr1
+; CHECK-NEXT: xvbitset.w $xr0, $xr0, $xr1
; CHECK-NEXT: ret
entry:
%0 = bitcast <4 x i64> %b to <8 x i32>
@@ -173,9 +163,7 @@ entry:
define <4 x i64> @xvbitset_w_nomask(<4 x i64> %a, <4 x i64> %b) nounwind {
; CHECK-LABEL: xvbitset_w_nomask:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: xvrepli.w $xr2, 1
-; CHECK-NEXT: xvsll.w $xr1, $xr2, $xr1
-; CHECK-NEXT: xvor.v $xr0, $xr1, $xr0
+; CHECK-NEXT: xvbitset.w $xr0, $xr0, $xr1
; CHECK-NEXT: ret
entry:
%0 = bitcast <4 x i64> %b to <8 x i32>
diff --git a/llvm/test/CodeGen/LoongArch/lsx/bitrev.ll b/llvm/test/CodeGen/LoongArch/lsx/bitrev.ll
index 0cdc8aa5b59ae3..793aafdea1d558 100644
--- a/llvm/test/CodeGen/LoongArch/lsx/bitrev.ll
+++ b/llvm/test/CodeGen/LoongArch/lsx/bitrev.ll
@@ -95,9 +95,7 @@ entry:
define <2 x i64> @vbitrev_b(<2 x i64> %a, <2 x i64> %b) nounwind {
; CHECK-LABEL: vbitrev_b:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: vrepli.b $vr2, 1
-; CHECK-NEXT: vsll.b $vr1, $vr2, $vr1
-; CHECK-NEXT: vxor.v $vr0, $vr0, $vr1
+; CHECK-NEXT: vbitrev.b $vr0, $vr0, $vr1
; CHECK-NEXT: ret
entry:
%0 = bitcast <2 x i64> %b to <16 x i8>
@@ -111,9 +109,7 @@ entry:
define <2 x i64> @vbitrev_b_nomask(<2 x i64> %a, <2 x i64> %b) nounwind {
; CHECK-LABEL: vbitrev_b_nomask:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: vrepli.b $vr2, 1
-; CHECK-NEXT: vsll.b $vr1, $vr2, $vr1
-; CHECK-NEXT: vxor.v $vr0, $vr1, $vr0
+; CHECK-NEXT: vbitrev.b $vr0, $vr0, $vr1
; CHECK-NEXT: ret
entry:
%0 = bitcast <2 x i64> %b to <16 x i8>
@@ -126,9 +122,7 @@ entry:
define <2 x i64> @vbitrev_h(<2 x i64> %a, <2 x i64> %b) nounwind {
; CHECK-LABEL: vbitrev_h:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: vrepli.h $vr2, 1
-; CHECK-NEXT: vsll.h $vr1, $vr2, $vr1
-; CHECK-NEXT: vxor.v $vr0, $vr0, $vr1
+; CHECK-NEXT: vbitrev.h $vr0, $vr0, $vr1
; CHECK-NEXT: ret
entry:
%0 = bitcast <2 x i64> %b to <8 x i16>
@@ -142,9 +136,7 @@ entry:
define <2 x i64> @vbitrev_h_nomask(<2 x i64> %a, <2 x i64> %b) nounwind {
; CHECK-LABEL: vbitrev_h_nomask:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: vrepli.h $vr2, 1
-; CHECK-NEXT: vsll.h $vr1, $vr2, $vr1
-; CHECK-NEXT: vxor.v $vr0, $vr1, $vr0
+; CHECK-NEXT: vbitrev.h $vr0, $vr0, $vr1
; CHECK-NEXT: ret
entry:
%0 = bitcast <2 x i64> %b to <8 x i16>
@@ -157,9 +149,7 @@ entry:
define <2 x i64> @vbitrev_w(<2 x i64> %a, <2 x i64> %b) nounwind {
; CHECK-LABEL: vbitrev_w:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: vrepli.w $vr2, 1
-; CHECK-NEXT: vsll.w $vr1, $vr2, $vr1
-; CHECK-NEXT: vxor.v $vr0, $vr0, $vr1
+; CHECK-NEXT: vbitrev.w $vr0, $vr0, $vr1
; CHECK-NEXT: ret
entry:
%0 = bitcast <2 x i64> %b to <4 x i32>
@@ -173,9 +163,7 @@ entry:
define <2 x i64> @vbitrev_w_nomask(<2 x i64> %a, <2 x i64> %b) nounwind {
; CHECK-LABEL: vbitrev_w_nomask:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: vrepli.w $vr2, 1
-; CHECK-NEXT: vsll.w $vr1, $vr2, $vr1
-; CHECK-NEXT: vxor.v $vr0, $vr1, $vr0
+; CHECK-NEXT: vbitrev.w $vr0, $vr0, $vr1
; CHECK-NEXT: ret
entry:
%0 = bitcast <2 x i64> %b to <4 x i32>
diff --git a/llvm/test/CodeGen/LoongArch/lsx/bitset.ll b/llvm/test/CodeGen/LoongArch/lsx/bitset.ll
index 17fdaf03aec318..1564264dcbef18 100644
--- a/llvm/test/CodeGen/LoongArch/lsx/bitset.ll
+++ b/llvm/test/CodeGen/LoongArch/lsx/bitset.ll
@@ -95,9 +95,7 @@ entry:
define <2 x i64> @vbitset_b(<2 x i64> %a, <2 x i64> %b) nounwind {
; CHECK-LABEL: vbitset_b:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: vrepli.b $vr2, 1
-; CHECK-NEXT: vsll.b $vr1, $vr2, $vr1
-; CHECK-NEXT: vor.v $vr0, $vr0, $vr1
+; CHECK-NEXT: vbitset.b $vr0, $vr0, $vr1
; CHECK-NEXT: ret
entry:
%0 = bitcast <2 x i64> %b to <16 x i8>
@@ -111,9 +109,7 @@ entry:
define <2 x i64> @vbitset_b_nomask(<2 x i64> %a, <2 x i64> %b) nounwind {
; CHECK-LABEL: vbitset_b_nomask:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: vrepli.b $vr2, 1
-; CHECK-NEXT: vsll.b $vr1, $vr2, $vr1
-; CHECK-NEXT: vor.v $vr0, $vr1, $vr0
+; CHECK-NEXT: vbitset.b $vr0, $vr0, $vr1
; CHECK-NEXT: ret
entry:
%0 = bitcast <2 x i64> %b to <16 x i8>
@@ -126,9 +122,7 @@ entry:
define <2 x i64> @vbitset_h(<2 x i64> %a, <2 x i64> %b) nounwind {
; CHECK-LABEL: vbitset_h:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: vrepli.h $vr2, 1
-; CHECK-NEXT: vsll.h $vr1, $vr2, $vr1
-; CHECK-NEXT: vor.v $vr0, $vr0, $vr1
+; CHECK-NEXT: vbitset.h $vr0, $vr0, $vr1
; CHECK-NEXT: ret
entry:
%0 = bitcast <2 x i64> %b to <8 x i16>
@@ -142,9 +136,7 @@ entry:
define <2 x i64> @vbitset_h_nomask(<2 x i64> %a, <2 x i64> %b) nounwind {
; CHECK-LABEL: vbitset_h_nomask:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: vrepli.h $vr2, 1
-; CHECK-NEXT: vsll.h $vr1, $vr2, $vr1
-; CHECK-NEXT: vor.v $vr0, $vr1, $vr0
+; CHECK-NEXT: vbitset.h $vr0, $vr0, $vr1
; CHECK-NEXT: ret
entry:
%0 = bitcast <2 x i64> %b to <8 x i16>
@@ -157,9 +149,7 @@ entry:
define <2 x i64> @vbitset_w(<2 x i64> %a, <2 x i64> %b) nounwind {
; CHECK-LABEL: vbitset_w:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: vrepli.w $vr2, 1
-; CHECK-NEXT: vsll.w $vr1, $vr2, $vr1
-; CHECK-NEXT: vor.v $vr0, $vr0, $vr1
+; CHECK-NEXT: vbitset.w $vr0, $vr0, $vr1
; CHECK-NEXT: ret
entry:
%0 = bitcast <2 x i64> %b to <4 x i32>
@@ -173,9 +163,7 @@ entry:
define <2 x i64> @vbitset_w_nomask(<2 x i64> %a, <2 x i64> %b) nounwind {
; CHECK-LABEL: vbitset_w_nomask:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: vrepli.w $vr2, 1
-; CHECK-NEXT: vsll.w $vr1, $vr2, $vr1
-; CHECK-NEXT: vor.v $vr0, $vr1, $vr0
+; CHECK-NEXT: vbitset.w $vr0, $vr0, $vr1
; CHECK-NEXT: ret
entry:
%0 = bitcast <2 x i64> %b to <4 x i32>
>From 8d0e258edaac4caeb2b5275ca4f4a8e2552250d6 Mon Sep 17 00:00:00 2001
From: WANG Rui <wangrui at loongson.cn>
Date: Thu, 19 Nov 2026 16:12:30 +0800
Subject: [PATCH 2/2] Fix build breaks after rebase
---
llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp b/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp
index 9fa102ae8abe12..63239787c01bf4 100644
--- a/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp
+++ b/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp
@@ -7436,7 +7436,7 @@ static SDValue combineVectorBitOpOfBitcastShl(SDNode *N, SelectionDAG &DAG) {
if (Shl.getOpcode() == ISD::SHL && ShlVT.isVector() && ShlVT.isInteger() &&
ShlVT != VT && ShlVT.getSizeInBits() == VT.getSizeInBits() &&
TLI.isTypeLegal(ShlVT) &&
- ISD::isConstantSplatVector(Shl.getOperand(0).getNode(), SplatVal) &&
+ ISD::isConstantSplatVector(Shl.getOperand(0), SplatVal) &&
SplatVal.isOne()) {
SDLoc DL(N);
SDValue V = DAG.getNode(N->getOpcode(), DL, ShlVT, Shl,
More information about the llvm-branch-commits
mailing list