[llvm-branch-commits] [llvm] [LoongArch] Use INSERT_SUBREG of IMPLICIT_DEF to put an FP scalar in a LASX vector (PR #242330)

Jay Foad via llvm-branch-commits llvm-branch-commits at lists.llvm.org
Fri Nov 13 05:55:16 PST 2026


https://github.com/jayfoad created https://github.com/llvm/llvm-project/pull/242330

The patterns for vector_insert, scalar_to_vector and splat of an FP
scalar into a LASX vector used SUBREG_TO_REG to view the FPR as a vector
register. SUBREG_TO_REG means that the instruction defining the FPR also
defined the rest of the vector register, which is not the case. Only
element 0 is used, and scalar_to_vector leaves the other elements
undefined, which is what INSERT_SUBREG into an IMPLICIT_DEF expresses.

This changes the scheduling and register allocation in one test, but not
the number of instructions.

Co-Authored-By: Claude Opus 5.5 <noreply at anthropic.com>

---

<sub>Stack created with <a href="https://github.com/github/gh-stack">GitHub Stacks CLI</a> • <a href="https://gh.io/stacks-feedback">Give Feedback 💬</a></sub>

>From 9b11b3fffbb6bad2f10a9f1c5475a6a65ae1cfba Mon Sep 17 00:00:00 2001
From: Jay Foad <jay.foad at amd.com>
Date: Fri, 13 Nov 2026 13:46:57 +0000
Subject: [PATCH] [LoongArch] Use INSERT_SUBREG of IMPLICIT_DEF to put an FP
 scalar in a LASX vector

The patterns for vector_insert, scalar_to_vector and splat of an FP
scalar into a LASX vector used SUBREG_TO_REG to view the FPR as a vector
register. SUBREG_TO_REG means that the instruction defining the FPR also
defined the rest of the vector register, which is not the case. Only
element 0 is used, and scalar_to_vector leaves the other elements
undefined, which is what INSERT_SUBREG into an IMPLICIT_DEF expresses.

This changes the scheduling and register allocation in one test, but not
the number of instructions.

Co-Authored-By: Claude Opus 5.5 <noreply at anthropic.com>
---
 .../LoongArch/LoongArchLASXInstrInfo.td       | 16 +++++-----
 .../lasx/ir-instruction/extractelement.ll     | 30 +++++++++----------
 2 files changed, 23 insertions(+), 23 deletions(-)

diff --git a/llvm/lib/Target/LoongArch/LoongArchLASXInstrInfo.td b/llvm/lib/Target/LoongArch/LoongArchLASXInstrInfo.td
index 69e8de23cc9c9..d34738f75ca64 100644
--- a/llvm/lib/Target/LoongArch/LoongArchLASXInstrInfo.td
+++ b/llvm/lib/Target/LoongArch/LoongArchLASXInstrInfo.td
@@ -2536,17 +2536,17 @@ def : Pat<(loongarch_xvinsve0 v8f32:$xd, v8f32:$xj, uimm3:$imm),
 def : Pat<(loongarch_xvinsve0 v4f64:$xd, v4f64:$xj, uimm2:$imm),
           (XVINSVE0_D v4f64:$xd, v4f64:$xj, uimm2:$imm)>;
 def : Pat<(vector_insert v8f32:$xd, FPR32:$fj, uimm3:$imm),
-          (XVINSVE0_W v8f32:$xd, (SUBREG_TO_REG FPR32:$fj, sub_32),
-              uimm3:$imm)>;
+          (XVINSVE0_W v8f32:$xd,
+              (INSERT_SUBREG (IMPLICIT_DEF), FPR32:$fj, sub_32), uimm3:$imm)>;
 def : Pat<(vector_insert v4f64:$xd, FPR64:$fj, uimm2:$imm),
-          (XVINSVE0_D v4f64:$xd, (SUBREG_TO_REG FPR64:$fj, sub_64),
-              uimm2:$imm)>;
+          (XVINSVE0_D v4f64:$xd,
+              (INSERT_SUBREG (IMPLICIT_DEF), FPR64:$fj, sub_64), uimm2:$imm)>;
 
 // scalar_to_vector
 def : Pat<(v8f32 (scalar_to_vector FPR32:$fj)),
-          (SUBREG_TO_REG FPR32:$fj, sub_32)>;
+          (INSERT_SUBREG (IMPLICIT_DEF), FPR32:$fj, sub_32)>;
 def : Pat<(v4f64 (scalar_to_vector FPR64:$fj)),
-          (SUBREG_TO_REG FPR64:$fj, sub_64)>;
+          (INSERT_SUBREG (IMPLICIT_DEF), FPR64:$fj, sub_64)>;
 
 // XVPICKVE2GR_W[U]
 def : Pat<(loongarch_vpick_sext_elt v8i32:$xd, uimm3:$imm, i32),
@@ -2740,9 +2740,9 @@ def : Pat<(loongarch_xvreplve0 v8f32:$xj),
 def : Pat<(loongarch_xvreplve0 v4f64:$xj),
           (XVREPLVE0_D v4f64:$xj)>;
 def : Pat<(lasxsplatf32 FPR32:$fj),
-          (XVREPLVE0_W (SUBREG_TO_REG FPR32:$fj, sub_32))>;
+          (XVREPLVE0_W (INSERT_SUBREG (IMPLICIT_DEF), FPR32:$fj, sub_32))>;
 def : Pat<(lasxsplatf64 FPR64:$fj),
-          (XVREPLVE0_D (SUBREG_TO_REG FPR64:$fj, sub_64))>;
+          (XVREPLVE0_D (INSERT_SUBREG (IMPLICIT_DEF), FPR64:$fj, sub_64))>;
 foreach vt = [v32i8, v16i16, v8i32, v4i64, v8f32, v4f64] in
   def : Pat<(vt (loongarch_xvreplve0q LASX256:$xj)),
             (XVREPLVE0_Q LASX256:$xj)>;
diff --git a/llvm/test/CodeGen/LoongArch/lasx/ir-instruction/extractelement.ll b/llvm/test/CodeGen/LoongArch/lasx/ir-instruction/extractelement.ll
index 3511c1d8b7222..9c340bf7a2378 100644
--- a/llvm/test/CodeGen/LoongArch/lasx/ir-instruction/extractelement.ll
+++ b/llvm/test/CodeGen/LoongArch/lasx/ir-instruction/extractelement.ll
@@ -94,10 +94,10 @@ define void @extract_32xi8_idx(ptr %src, ptr %dst, i32 %idx) nounwind {
 ; LA64-LABEL: extract_32xi8_idx:
 ; LA64:       # %bb.0:
 ; LA64-NEXT:    xvld $xr0, $a0, 0
-; LA64-NEXT:    xvpermi.q $xr1, $xr0, 1
 ; LA64-NEXT:    bstrpick.d $a0, $a2, 31, 0
-; LA64-NEXT:    movgr2fr.w $fa2, $a0
-; LA64-NEXT:    xvshuf.b $xr0, $xr1, $xr0, $xr2
+; LA64-NEXT:    movgr2fr.w $fa1, $a0
+; LA64-NEXT:    xvpermi.q $xr2, $xr0, 1
+; LA64-NEXT:    xvshuf.b $xr0, $xr2, $xr0, $xr1
 ; LA64-NEXT:    xvstelm.b $xr0, $a1, 0, 0
 ; LA64-NEXT:    ret
   %v = load volatile <32 x i8>, ptr %src
@@ -119,11 +119,11 @@ define void @extract_16xi16_idx(ptr %src, ptr %dst, i32 %idx) nounwind {
 ; LA64-LABEL: extract_16xi16_idx:
 ; LA64:       # %bb.0:
 ; LA64-NEXT:    xvld $xr0, $a0, 0
-; LA64-NEXT:    xvpermi.q $xr1, $xr0, 1
 ; LA64-NEXT:    bstrpick.d $a0, $a2, 31, 0
-; LA64-NEXT:    movgr2fr.w $fa2, $a0
-; LA64-NEXT:    xvshuf.h $xr2, $xr1, $xr0
-; LA64-NEXT:    xvstelm.h $xr2, $a1, 0, 0
+; LA64-NEXT:    movgr2fr.w $fa1, $a0
+; LA64-NEXT:    xvpermi.q $xr2, $xr0, 1
+; LA64-NEXT:    xvshuf.h $xr1, $xr2, $xr0
+; LA64-NEXT:    xvstelm.h $xr1, $a1, 0, 0
 ; LA64-NEXT:    ret
   %v = load volatile <16 x i16>, ptr %src
   %e = extractelement <16 x i16> %v, i32 %idx
@@ -171,11 +171,11 @@ define void @extract_4xi64_idx(ptr %src, ptr %dst, i32 %idx) nounwind {
 ; LA64-LABEL: extract_4xi64_idx:
 ; LA64:       # %bb.0:
 ; LA64-NEXT:    xvld $xr0, $a0, 0
-; LA64-NEXT:    xvpermi.q $xr1, $xr0, 1
 ; LA64-NEXT:    bstrpick.d $a0, $a2, 31, 0
-; LA64-NEXT:    movgr2fr.w $fa2, $a0
-; LA64-NEXT:    xvshuf.d $xr2, $xr1, $xr0
-; LA64-NEXT:    xvstelm.d $xr2, $a1, 0, 0
+; LA64-NEXT:    movgr2fr.w $fa1, $a0
+; LA64-NEXT:    xvpermi.q $xr2, $xr0, 1
+; LA64-NEXT:    xvshuf.d $xr1, $xr2, $xr0
+; LA64-NEXT:    xvstelm.d $xr1, $a1, 0, 0
 ; LA64-NEXT:    ret
   %v = load volatile <4 x i64>, ptr %src
   %e = extractelement <4 x i64> %v, i32 %idx
@@ -219,11 +219,11 @@ define void @extract_4xdouble_idx(ptr %src, ptr %dst, i32 %idx) nounwind {
 ; LA64-LABEL: extract_4xdouble_idx:
 ; LA64:       # %bb.0:
 ; LA64-NEXT:    xvld $xr0, $a0, 0
-; LA64-NEXT:    xvpermi.q $xr1, $xr0, 1
 ; LA64-NEXT:    bstrpick.d $a0, $a2, 31, 0
-; LA64-NEXT:    movgr2fr.w $fa2, $a0
-; LA64-NEXT:    xvshuf.d $xr2, $xr1, $xr0
-; LA64-NEXT:    xvstelm.d $xr2, $a1, 0, 0
+; LA64-NEXT:    movgr2fr.w $fa1, $a0
+; LA64-NEXT:    xvpermi.q $xr2, $xr0, 1
+; LA64-NEXT:    xvshuf.d $xr1, $xr2, $xr0
+; LA64-NEXT:    xvstelm.d $xr1, $a1, 0, 0
 ; LA64-NEXT:    ret
   %v = load volatile <4 x double>, ptr %src
   %e = extractelement <4 x double> %v, i32 %idx



More information about the llvm-branch-commits mailing list