[llvm-branch-commits] [llvm] [LoongArch] Select [x]vbitclr for and-not of bitcasted shl(splat(1), y) (PR #241179)
via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Wed Nov 11 01:01:35 PST 2026
llvmorg-github-actions[bot] wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-backend-loongarch
Author: hev (heiher)
<details>
<summary>Changes</summary>
When the shift of splat(1) is performed in a narrower element type than the and-not, e.g.
```llvm
%s = shl <8 x i16> splat (i16 1), %y
%m = bitcast <8 x i16> %s to <2 x i64>
%n = xor <2 x i64> %m, splat (i64 -1)
%r = and <2 x i64> %x, %n
```
the VANDN node is created in v2i64 with a bitcast operand, so the existing [x]vbitclr.{b/h/w} patterns fail to match and we emit vrepli + vsll + vandn.v.
Fold VANDN(bitcast(shl(splat(1), y)), x) into
bitcast(VANDN(shl(splat(1), y), bitcast(x))) so that the operation is done in the shift's element type and selected as a single [x]vbitclr.
Assisted-By: LLM
---
Full diff: https://github.com/llvm/llvm-project/pull/241179.diff
3 Files Affected:
- (modified) llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp (+19)
- (modified) llvm/test/CodeGen/LoongArch/lasx/bitclr.ll (+6-18)
- (modified) llvm/test/CodeGen/LoongArch/lsx/bitclr.ll (+6-18)
``````````diff
diff --git a/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp b/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp
index b5f76499278cb..9ef5ccf8f1aac 100644
--- a/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp
+++ b/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp
@@ -8673,6 +8673,25 @@ static SDValue performVANDNCombine(SDNode *N, SelectionDAG &DAG,
if (SDValue Not = isNOT(N0, DAG))
return DAG.getNode(ISD::AND, DL, VT, DAG.getBitcast(VT, Not), N1);
+ // VANDN(bitcast(shl(splat(1), y)), x)
+ // -> bitcast(VANDN(shl(splat(1), y), bitcast(x)))
+ // Perform the VANDN in the element type of the shift so that it can be
+ // selected as [x]vbitclr.{b/h/w/d}.
+ if (N0.getOpcode() == ISD::BITCAST && N0.hasOneUse()) {
+ 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() &&
+ DAG.getTargetLoweringInfo().isTypeLegal(ShlVT) &&
+ ISD::isConstantSplatVector(Shl.getOperand(0).getNode(), SplatVal) &&
+ SplatVal.isOne()) {
+ SDValue V = DAG.getNode(LoongArchISD::VANDN, DL, ShlVT, Shl,
+ DAG.getBitcast(ShlVT, N1));
+ return DAG.getBitcast(VT, V);
+ }
+ }
+
// Folds for better commutativity:
if (N1->hasOneUse()) {
// VANDN(x,NOT(y)) -> AND(NOT(x),NOT(y)) -> NOT(OR(X,Y)).
diff --git a/llvm/test/CodeGen/LoongArch/lasx/bitclr.ll b/llvm/test/CodeGen/LoongArch/lasx/bitclr.ll
index d18ec2795899a..ee54a3e1a47c8 100644
--- a/llvm/test/CodeGen/LoongArch/lasx/bitclr.ll
+++ b/llvm/test/CodeGen/LoongArch/lasx/bitclr.ll
@@ -95,9 +95,7 @@ entry:
define <4 x i64> @xvbitclr_b(<4 x i64> %a, <4 x i64> %b) nounwind {
; CHECK-LABEL: xvbitclr_b:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: xvrepli.b $xr2, 1
-; CHECK-NEXT: xvsll.b $xr1, $xr2, $xr1
-; CHECK-NEXT: xvandn.v $xr0, $xr1, $xr0
+; CHECK-NEXT: xvbitclr.b $xr0, $xr0, $xr1
; CHECK-NEXT: ret
entry:
%0 = bitcast <4 x i64> %b to <32 x i8>
@@ -112,9 +110,7 @@ entry:
define <4 x i64> @xvbitclr_b_nomask(<4 x i64> %a, <4 x i64> %b) nounwind {
; CHECK-LABEL: xvbitclr_b_nomask:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: xvrepli.b $xr2, 1
-; CHECK-NEXT: xvsll.b $xr1, $xr2, $xr1
-; CHECK-NEXT: xvandn.v $xr0, $xr1, $xr0
+; CHECK-NEXT: xvbitclr.b $xr0, $xr0, $xr1
; CHECK-NEXT: ret
entry:
%0 = bitcast <4 x i64> %b to <32 x i8>
@@ -128,9 +124,7 @@ entry:
define <4 x i64> @xvbitclr_h(<4 x i64> %a, <4 x i64> %b) nounwind {
; CHECK-LABEL: xvbitclr_h:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: xvrepli.h $xr2, 1
-; CHECK-NEXT: xvsll.h $xr1, $xr2, $xr1
-; CHECK-NEXT: xvandn.v $xr0, $xr1, $xr0
+; CHECK-NEXT: xvbitclr.h $xr0, $xr0, $xr1
; CHECK-NEXT: ret
entry:
%0 = bitcast <4 x i64> %b to <16 x i16>
@@ -145,9 +139,7 @@ entry:
define <4 x i64> @xvbitclr_h_nomask(<4 x i64> %a, <4 x i64> %b) nounwind {
; CHECK-LABEL: xvbitclr_h_nomask:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: xvrepli.h $xr2, 1
-; CHECK-NEXT: xvsll.h $xr1, $xr2, $xr1
-; CHECK-NEXT: xvandn.v $xr0, $xr1, $xr0
+; CHECK-NEXT: xvbitclr.h $xr0, $xr0, $xr1
; CHECK-NEXT: ret
entry:
%0 = bitcast <4 x i64> %b to <16 x i16>
@@ -161,9 +153,7 @@ entry:
define <4 x i64> @xvbitclr_w(<4 x i64> %a, <4 x i64> %b) nounwind {
; CHECK-LABEL: xvbitclr_w:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: xvrepli.w $xr2, 1
-; CHECK-NEXT: xvsll.w $xr1, $xr2, $xr1
-; CHECK-NEXT: xvandn.v $xr0, $xr1, $xr0
+; CHECK-NEXT: xvbitclr.w $xr0, $xr0, $xr1
; CHECK-NEXT: ret
entry:
%0 = bitcast <4 x i64> %b to <8 x i32>
@@ -178,9 +168,7 @@ entry:
define <4 x i64> @xvbitclr_w_nomask(<4 x i64> %a, <4 x i64> %b) nounwind {
; CHECK-LABEL: xvbitclr_w_nomask:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: xvrepli.w $xr2, 1
-; CHECK-NEXT: xvsll.w $xr1, $xr2, $xr1
-; CHECK-NEXT: xvandn.v $xr0, $xr1, $xr0
+; CHECK-NEXT: xvbitclr.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/bitclr.ll b/llvm/test/CodeGen/LoongArch/lsx/bitclr.ll
index 5e3d61efaf609..050d32626254e 100644
--- a/llvm/test/CodeGen/LoongArch/lsx/bitclr.ll
+++ b/llvm/test/CodeGen/LoongArch/lsx/bitclr.ll
@@ -95,9 +95,7 @@ entry:
define <2 x i64> @vbitclr_b(<2 x i64> %a, <2 x i64> %b) nounwind {
; CHECK-LABEL: vbitclr_b:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: vrepli.b $vr2, 1
-; CHECK-NEXT: vsll.b $vr1, $vr2, $vr1
-; CHECK-NEXT: vandn.v $vr0, $vr1, $vr0
+; CHECK-NEXT: vbitclr.b $vr0, $vr0, $vr1
; CHECK-NEXT: ret
entry:
%0 = bitcast <2 x i64> %b to <16 x i8>
@@ -112,9 +110,7 @@ entry:
define <2 x i64> @vbitclr_b_nomask(<2 x i64> %a, <2 x i64> %b) nounwind {
; CHECK-LABEL: vbitclr_b_nomask:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: vrepli.b $vr2, 1
-; CHECK-NEXT: vsll.b $vr1, $vr2, $vr1
-; CHECK-NEXT: vandn.v $vr0, $vr1, $vr0
+; CHECK-NEXT: vbitclr.b $vr0, $vr0, $vr1
; CHECK-NEXT: ret
entry:
%0 = bitcast <2 x i64> %b to <16 x i8>
@@ -128,9 +124,7 @@ entry:
define <2 x i64> @vbitclr_h(<2 x i64> %a, <2 x i64> %b) nounwind {
; CHECK-LABEL: vbitclr_h:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: vrepli.h $vr2, 1
-; CHECK-NEXT: vsll.h $vr1, $vr2, $vr1
-; CHECK-NEXT: vandn.v $vr0, $vr1, $vr0
+; CHECK-NEXT: vbitclr.h $vr0, $vr0, $vr1
; CHECK-NEXT: ret
entry:
%0 = bitcast <2 x i64> %b to <8 x i16>
@@ -145,9 +139,7 @@ entry:
define <2 x i64> @vbitclr_h_nomask(<2 x i64> %a, <2 x i64> %b) nounwind {
; CHECK-LABEL: vbitclr_h_nomask:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: vrepli.h $vr2, 1
-; CHECK-NEXT: vsll.h $vr1, $vr2, $vr1
-; CHECK-NEXT: vandn.v $vr0, $vr1, $vr0
+; CHECK-NEXT: vbitclr.h $vr0, $vr0, $vr1
; CHECK-NEXT: ret
entry:
%0 = bitcast <2 x i64> %b to <8 x i16>
@@ -161,9 +153,7 @@ entry:
define <2 x i64> @vbitclr_w(<2 x i64> %a, <2 x i64> %b) nounwind {
; CHECK-LABEL: vbitclr_w:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: vrepli.w $vr2, 1
-; CHECK-NEXT: vsll.w $vr1, $vr2, $vr1
-; CHECK-NEXT: vandn.v $vr0, $vr1, $vr0
+; CHECK-NEXT: vbitclr.w $vr0, $vr0, $vr1
; CHECK-NEXT: ret
entry:
%0 = bitcast <2 x i64> %b to <4 x i32>
@@ -178,9 +168,7 @@ entry:
define <2 x i64> @vbitclr_w_nomask(<2 x i64> %a, <2 x i64> %b) nounwind {
; CHECK-LABEL: vbitclr_w_nomask:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: vrepli.w $vr2, 1
-; CHECK-NEXT: vsll.w $vr1, $vr2, $vr1
-; CHECK-NEXT: vandn.v $vr0, $vr1, $vr0
+; CHECK-NEXT: vbitclr.w $vr0, $vr0, $vr1
; CHECK-NEXT: ret
entry:
%0 = bitcast <2 x i64> %b to <4 x i32>
``````````
</details>
https://github.com/llvm/llvm-project/pull/241179
More information about the llvm-branch-commits
mailing list