[llvm-branch-commits] [llvm] [LoongArch] Use INSERT_SUBREG of IMPLICIT_DEF for scalar FP rounding via LSX. NFC (PR #242329)
Jay Foad via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Fri Nov 13 05:55:08 PST 2026
https://github.com/jayfoad created https://github.com/llvm/llvm-project/pull/242329
fceil, ffloor, ftrunc and froundeven of an FP scalar are selected to a
splat of element 0 of the FPR viewed as an LSX vector, a vector rounding
instruction, and an extract of element 0. Only element 0 is used, so use
INSERT_SUBREG into an IMPLICIT_DEF instead of SUBREG_TO_REG, which means
that the instruction defining the FPR also defined the rest of the
vector register.
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 145177498bcfe35091b7bae0731e5436167134e3 Mon Sep 17 00:00:00 2001
From: Jay Foad <jay.foad at amd.com>
Date: Fri, 13 Nov 2026 13:46:05 +0000
Subject: [PATCH] [LoongArch] Use INSERT_SUBREG of IMPLICIT_DEF for scalar FP
rounding via LSX. NFC
fceil, ffloor, ftrunc and froundeven of an FP scalar are selected to a
splat of element 0 of the FPR viewed as an LSX vector, a vector rounding
instruction, and an extract of element 0. Only element 0 is used, so use
INSERT_SUBREG into an IMPLICIT_DEF instead of SUBREG_TO_REG, which means
that the instruction defining the FPR also defined the rest of the
vector register.
Co-Authored-By: Claude Opus 5.5 <noreply at anthropic.com>
---
.../Target/LoongArch/LoongArchLSXInstrInfo.td | 24 ++++++++++++-------
1 file changed, 16 insertions(+), 8 deletions(-)
diff --git a/llvm/lib/Target/LoongArch/LoongArchLSXInstrInfo.td b/llvm/lib/Target/LoongArch/LoongArchLSXInstrInfo.td
index 42599cd6f2cda..c4ccd06ed6ca6 100644
--- a/llvm/lib/Target/LoongArch/LoongArchLSXInstrInfo.td
+++ b/llvm/lib/Target/LoongArch/LoongArchLSXInstrInfo.td
@@ -3508,28 +3508,36 @@ def : Pat<(loongarch_vfrsqrte v2f64:$src),
// Vector floating-point conversion
def : Pat<(f32 (fceil FPR32:$fj)),
(f32 (EXTRACT_SUBREG (VFRINTRP_S (VREPLVEI_W
- (SUBREG_TO_REG FPR32:$fj, sub_32), 0)), sub_32))>;
+ (INSERT_SUBREG (IMPLICIT_DEF), FPR32:$fj, sub_32), 0)),
+ sub_32))>;
def : Pat<(f64 (fceil FPR64:$fj)),
(f64 (EXTRACT_SUBREG (VFRINTRP_D (VREPLVEI_D
- (SUBREG_TO_REG FPR64:$fj, sub_64), 0)), sub_64))>;
+ (INSERT_SUBREG (IMPLICIT_DEF), FPR64:$fj, sub_64), 0)),
+ sub_64))>;
def : Pat<(f32 (ffloor FPR32:$fj)),
(f32 (EXTRACT_SUBREG (VFRINTRM_S (VREPLVEI_W
- (SUBREG_TO_REG FPR32:$fj, sub_32), 0)), sub_32))>;
+ (INSERT_SUBREG (IMPLICIT_DEF), FPR32:$fj, sub_32), 0)),
+ sub_32))>;
def : Pat<(f64 (ffloor FPR64:$fj)),
(f64 (EXTRACT_SUBREG (VFRINTRM_D (VREPLVEI_D
- (SUBREG_TO_REG FPR64:$fj, sub_64), 0)), sub_64))>;
+ (INSERT_SUBREG (IMPLICIT_DEF), FPR64:$fj, sub_64), 0)),
+ sub_64))>;
def : Pat<(f32 (ftrunc FPR32:$fj)),
(f32 (EXTRACT_SUBREG (VFRINTRZ_S (VREPLVEI_W
- (SUBREG_TO_REG FPR32:$fj, sub_32), 0)), sub_32))>;
+ (INSERT_SUBREG (IMPLICIT_DEF), FPR32:$fj, sub_32), 0)),
+ sub_32))>;
def : Pat<(f64 (ftrunc FPR64:$fj)),
(f64 (EXTRACT_SUBREG (VFRINTRZ_D (VREPLVEI_D
- (SUBREG_TO_REG FPR64:$fj, sub_64), 0)), sub_64))>;
+ (INSERT_SUBREG (IMPLICIT_DEF), FPR64:$fj, sub_64), 0)),
+ sub_64))>;
def : Pat<(f32 (froundeven FPR32:$fj)),
(f32 (EXTRACT_SUBREG (VFRINTRNE_S (VREPLVEI_W
- (SUBREG_TO_REG FPR32:$fj, sub_32), 0)), sub_32))>;
+ (INSERT_SUBREG (IMPLICIT_DEF), FPR32:$fj, sub_32), 0)),
+ sub_32))>;
def : Pat<(f64 (froundeven FPR64:$fj)),
(f64 (EXTRACT_SUBREG (VFRINTRNE_D (VREPLVEI_D
- (SUBREG_TO_REG FPR64:$fj, sub_64), 0)), sub_64))>;
+ (INSERT_SUBREG (IMPLICIT_DEF), FPR64:$fj, sub_64), 0)),
+ sub_64))>;
defm : PatVrF<fceil, "VFRINTRP">;
defm : PatVrF<ffloor, "VFRINTRM">;
More information about the llvm-branch-commits
mailing list