[llvm-branch-commits] [llvm] [LoongArch] Select [X]VBITSELI_B for byte splat bitsel on wider element types (PR #242143)

via llvm-branch-commits llvm-branch-commits at lists.llvm.org
Thu Nov 12 19:09:29 PST 2026


https://github.com/heiher created https://github.com/llvm/llvm-project/pull/242143

The VBITSELI_B/XVBITSELI_B patterns only matched v16i8/v32i8. When the bitwise select idiom `(a & imm) | (~a & b)` appears on v8i16/v4i32/v2i64 (or the LASX equivalents) with a byte splat constant, e.g. from portable SIMD implementations of the bitseli intrinsics, we emitted:

```
  vrepli.b  $vr2, 127
  vbitsel.v $vr0, $vr1, $vr2, $vr0
```

Use vsplat_i8_uimm8, as VANDI_B/VORI_B/etc. already do, so that we now emit:

```
  vbitseli.b $vr0, $vr1, 127
```

Assisted-By: LLM

>From 8d193a06c0e444eb71aad23a67ae819bb45d23ee Mon Sep 17 00:00:00 2001
From: WANG Rui <wangrui at loongson.cn>
Date: Fri, 13 Nov 2026 11:04:55 +0800
Subject: [PATCH] [LoongArch] Select [X]VBITSELI_B for byte splat bitsel on
 wider element types

The VBITSELI_B/XVBITSELI_B patterns only matched v16i8/v32i8. When the
bitwise select idiom `(a & imm) | (~a & b)` appears on v8i16/v4i32/v2i64
(or the LASX equivalents) with a byte splat constant, e.g. from portable
SIMD implementations of the bitseli intrinsics, we emitted:

```
  vrepli.b  $vr2, 127
  vbitsel.v $vr0, $vr1, $vr2, $vr0
```

Use vsplat_i8_uimm8, as VANDI_B/VORI_B/etc. already do, so that we now
emit:

```
  vbitseli.b $vr0, $vr1, 127
```

Assisted-By: LLM
---
 llvm/lib/Target/LoongArch/LoongArchLASXInstrInfo.td | 4 ++++
 llvm/lib/Target/LoongArch/LoongArchLSXInstrInfo.td  | 4 ++++
 llvm/test/CodeGen/LoongArch/lasx/bitsel.ll          | 9 +++------
 llvm/test/CodeGen/LoongArch/lsx/bitsel.ll           | 9 +++------
 4 files changed, 14 insertions(+), 12 deletions(-)

diff --git a/llvm/lib/Target/LoongArch/LoongArchLASXInstrInfo.td b/llvm/lib/Target/LoongArch/LoongArchLASXInstrInfo.td
index 69e8de23cc9c9..6a806ac4a6405 100644
--- a/llvm/lib/Target/LoongArch/LoongArchLASXInstrInfo.td
+++ b/llvm/lib/Target/LoongArch/LoongArchLASXInstrInfo.td
@@ -2341,6 +2341,10 @@ foreach vt = [v32i8, v16i16, v8i32, v4i64] in
 def : Pat<(v32i8 (or (and LASX256:$vd, (v32i8 (SplatPat_uimm8 uimm8:$imm))),
                      (loongarch_vandn LASX256:$vd, LASX256:$vj))),
           (XVBITSELI_B LASX256:$vd, LASX256:$vj, uimm8:$imm)>;
+foreach vt = [v16i16, v8i32, v4i64] in
+def : Pat<(vt (or (and LASX256:$vd, (vt (vsplat_i8_uimm8 grlenimm:$imm))),
+                  (loongarch_vandn LASX256:$vd, LASX256:$vj))),
+          (XVBITSELI_B LASX256:$vd, LASX256:$vj, grlenimm:$imm)>;
 
 // Vector bswaps
 def : Pat<(bswap (v16i16 LASX256:$xj)), (XVSHUF4I_B LASX256:$xj, 0b10110001)>;
diff --git a/llvm/lib/Target/LoongArch/LoongArchLSXInstrInfo.td b/llvm/lib/Target/LoongArch/LoongArchLSXInstrInfo.td
index 25ead5d1deb30..3e209f04aa800 100644
--- a/llvm/lib/Target/LoongArch/LoongArchLSXInstrInfo.td
+++ b/llvm/lib/Target/LoongArch/LoongArchLSXInstrInfo.td
@@ -2585,6 +2585,10 @@ foreach vt = [v16i8, v8i16, v4i32, v2i64] in
 def : Pat<(v16i8 (or (and LSX128:$vd, (v16i8 (SplatPat_uimm8 uimm8:$imm))),
                      (loongarch_vandn LSX128:$vd, LSX128:$vj))),
           (VBITSELI_B LSX128:$vd, LSX128:$vj, uimm8:$imm)>;
+foreach vt = [v8i16, v4i32, v2i64] in
+def : Pat<(vt (or (and LSX128:$vd, (vt (vsplat_i8_uimm8 grlenimm:$imm))),
+                  (loongarch_vandn LSX128:$vd, LSX128:$vj))),
+          (VBITSELI_B LSX128:$vd, LSX128:$vj, grlenimm:$imm)>;
 
 // Vector bswaps
 def : Pat<(bswap (v8i16 LSX128:$vj)), (VSHUF4I_B LSX128:$vj, 0b10110001)>;
diff --git a/llvm/test/CodeGen/LoongArch/lasx/bitsel.ll b/llvm/test/CodeGen/LoongArch/lasx/bitsel.ll
index fd1b1497af554..95881eeed9f1d 100644
--- a/llvm/test/CodeGen/LoongArch/lasx/bitsel.ll
+++ b/llvm/test/CodeGen/LoongArch/lasx/bitsel.ll
@@ -70,8 +70,7 @@ entry:
 define <16 x i16> @bitseli_v16i16(<16 x i16> %a, <16 x i16> %b) nounwind {
 ; CHECK-LABEL: bitseli_v16i16:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    xvrepli.b $xr2, 127
-; CHECK-NEXT:    xvbitsel.v $xr0, $xr1, $xr2, $xr0
+; CHECK-NEXT:    xvbitseli.b $xr0, $xr1, 127
 ; CHECK-NEXT:    ret
 entry:
   %0 = xor <16 x i16> %a, splat (i16 -1)
@@ -84,8 +83,7 @@ entry:
 define <8 x i32> @bitseli_v8i32(<8 x i32> %a, <8 x i32> %b) nounwind {
 ; CHECK-LABEL: bitseli_v8i32:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    xvrepli.b $xr2, 127
-; CHECK-NEXT:    xvbitsel.v $xr0, $xr1, $xr2, $xr0
+; CHECK-NEXT:    xvbitseli.b $xr0, $xr1, 127
 ; CHECK-NEXT:    ret
 entry:
   %0 = xor <8 x i32> %a, splat (i32 -1)
@@ -98,8 +96,7 @@ entry:
 define <4 x i64> @bitseli_v4i64(<4 x i64> %a, <4 x i64> %b) nounwind {
 ; CHECK-LABEL: bitseli_v4i64:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    xvrepli.b $xr2, 127
-; CHECK-NEXT:    xvbitsel.v $xr0, $xr1, $xr2, $xr0
+; CHECK-NEXT:    xvbitseli.b $xr0, $xr1, 127
 ; CHECK-NEXT:    ret
 entry:
   %0 = xor <4 x i64> %a, splat (i64 -1)
diff --git a/llvm/test/CodeGen/LoongArch/lsx/bitsel.ll b/llvm/test/CodeGen/LoongArch/lsx/bitsel.ll
index f1f598dc459b2..25bd450be2cfa 100644
--- a/llvm/test/CodeGen/LoongArch/lsx/bitsel.ll
+++ b/llvm/test/CodeGen/LoongArch/lsx/bitsel.ll
@@ -70,8 +70,7 @@ entry:
 define <8 x i16> @bitseli_v8i16(<8 x i16> %a, <8 x i16> %b) nounwind {
 ; CHECK-LABEL: bitseli_v8i16:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    vrepli.b $vr2, 127
-; CHECK-NEXT:    vbitsel.v $vr0, $vr1, $vr2, $vr0
+; CHECK-NEXT:    vbitseli.b $vr0, $vr1, 127
 ; CHECK-NEXT:    ret
 entry:
   %0 = xor <8 x i16> %a, splat (i16 -1)
@@ -84,8 +83,7 @@ entry:
 define <4 x i32> @bitseli_v4i32(<4 x i32> %a, <4 x i32> %b) nounwind {
 ; CHECK-LABEL: bitseli_v4i32:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    vrepli.b $vr2, 127
-; CHECK-NEXT:    vbitsel.v $vr0, $vr1, $vr2, $vr0
+; CHECK-NEXT:    vbitseli.b $vr0, $vr1, 127
 ; CHECK-NEXT:    ret
 entry:
   %0 = xor <4 x i32> %a, splat (i32 -1)
@@ -98,8 +96,7 @@ entry:
 define <2 x i64> @bitseli_v2i64(<2 x i64> %a, <2 x i64> %b) nounwind {
 ; CHECK-LABEL: bitseli_v2i64:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    vrepli.b $vr2, 127
-; CHECK-NEXT:    vbitsel.v $vr0, $vr1, $vr2, $vr0
+; CHECK-NEXT:    vbitseli.b $vr0, $vr1, 127
 ; CHECK-NEXT:    ret
 entry:
   %0 = xor <2 x i64> %a, splat (i64 -1)



More information about the llvm-branch-commits mailing list