[llvm] AArch64: Drop AllowFPOpFusion from the MachineCombiner FMA conditions (PR #221404)
via llvm-commits
llvm-commits at lists.llvm.org
Fri Sep 4 23:40:52 PDT 2026
llvmorg-github-actions[bot] wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-backend-aarch64
Author: Matt Arsenault (arsenm)
<details>
<summary>Changes</summary>
Only use the flags, and add missing tests for a missing flag on only
one of the participating instructions. This is a step toward removing
the AllowFPOpFusion field from TargetOptions.
Co-authored-by: Claude (Claude-Opus-4.8) <noreply@<!-- -->anthropic.com>
---
Full diff: https://github.com/llvm/llvm-project/pull/221404.diff
2 Files Affected:
- (modified) llvm/lib/Target/AArch64/AArch64InstrInfo.cpp (+3-5)
- (modified) llvm/test/CodeGen/AArch64/arm64-neon-2velem.ll (+56-6)
``````````diff
diff --git a/llvm/lib/Target/AArch64/AArch64InstrInfo.cpp b/llvm/lib/Target/AArch64/AArch64InstrInfo.cpp
index e69336ea04857..6bfb1f36295e1 100644
--- a/llvm/lib/Target/AArch64/AArch64InstrInfo.cpp
+++ b/llvm/lib/Target/AArch64/AArch64InstrInfo.cpp
@@ -7569,11 +7569,9 @@ static bool isCombineInstrCandidateFP(const MachineInstr &Inst) {
case AArch64::FSUBv2f32:
case AArch64::FSUBv2f64:
case AArch64::FSUBv4f32:
- TargetOptions Options = Inst.getParent()->getParent()->getTarget().Options;
- // We can fuse FADD/FSUB with FMUL, if fusion is either allowed globally by
- // the target options or if FADD/FSUB has the contract fast-math flag.
- return Options.AllowFPOpFusion == FPOpFusion::Fast ||
- Inst.getFlag(MachineInstr::FmContract);
+ // We can fuse FADD/FSUB with FMUL, if FADD/FSUB has the contract fast-math
+ // flag.
+ return Inst.getFlag(MachineInstr::FmContract);
}
return false;
}
diff --git a/llvm/test/CodeGen/AArch64/arm64-neon-2velem.ll b/llvm/test/CodeGen/AArch64/arm64-neon-2velem.ll
index 85d8b7c3e2866..f90333012c21f 100644
--- a/llvm/test/CodeGen/AArch64/arm64-neon-2velem.ll
+++ b/llvm/test/CodeGen/AArch64/arm64-neon-2velem.ll
@@ -1,6 +1,6 @@
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc < %s -verify-machineinstrs -mtriple=arm64-none-linux-gnu -fp-contract=fast | FileCheck %s --check-prefixes=CHECK,CHECK-SD
-; RUN: llc < %s -verify-machineinstrs -mtriple=arm64-none-linux-gnu -fp-contract=fast -global-isel -global-isel-abort=2 2>&1 | FileCheck %s --check-prefixes=CHECK,CHECK-GI
+; RUN: llc < %s -verify-machineinstrs -mtriple=arm64-none-linux-gnu | FileCheck %s --check-prefixes=CHECK,CHECK-SD
+; RUN: llc < %s -verify-machineinstrs -mtriple=arm64-none-linux-gnu -global-isel -global-isel-abort=2 2>&1 | FileCheck %s --check-prefixes=CHECK,CHECK-GI
; CHECK-GI: warning: Instruction selection used fallback path for test_vfms_lane_f32
; CHECK-GI-NEXT: warning: Instruction selection used fallback path for test_vfmsq_lane_f32
@@ -4231,8 +4231,8 @@ entry:
%lane1 = shufflevector <4 x float> %v, <4 x float> undef, <4 x i32> <i32 3, i32 3, i32 3, i32 3>
%0 = tail call <4 x float> @llvm.fma.v4f32(<4 x float> %lane1, <4 x float> %b, <4 x float> %a)
%lane2 = shufflevector <4 x float> %v, <4 x float> undef, <4 x i32> <i32 3, i32 3, i32 3, i32 3>
- %1 = fmul <4 x float> %lane2, %c
- %s = fsub <4 x float> %0, %1
+ %1 = fmul contract <4 x float> %lane2, %c
+ %s = fsub contract <4 x float> %0, %1
ret <4 x float> %s
}
@@ -4246,8 +4246,58 @@ entry:
%lane1 = shufflevector <4 x float> %v, <4 x float> undef, <4 x i32> <i32 3, i32 3, i32 3, i32 3>
%0 = tail call <4 x float> @llvm.fma.v4f32(<4 x float> %lane1, <4 x float> %b, <4 x float> %a)
%lane2 = shufflevector <4 x float> %v, <4 x float> undef, <4 x i32> <i32 1, i32 1, i32 1, i32 1>
- %1 = fmul <4 x float> %lane2, %c
- %s = fsub <4 x float> %0, %1
+ %1 = fmul contract <4 x float> %lane2, %c
+ %s = fsub contract <4 x float> %0, %1
+ ret <4 x float> %s
+}
+
+; The multiply-subtract is only fused when the fsub carries the contract flag;
+; without it the fmul stays separate.
+define <4 x float> @no_fuse_dup_fsub_no_contract(<4 x float> %a, <4 x float> %b, <4 x float> %c, <4 x float> %v) {
+; CHECK-SD-LABEL: no_fuse_dup_fsub_no_contract:
+; CHECK-SD: // %bb.0: // %entry
+; CHECK-SD-NEXT: fmul v2.4s, v2.4s, v3.s[3]
+; CHECK-SD-NEXT: fmla v0.4s, v1.4s, v3.s[3]
+; CHECK-SD-NEXT: fsub v0.4s, v0.4s, v2.4s
+; CHECK-SD-NEXT: ret
+;
+; CHECK-GI-LABEL: no_fuse_dup_fsub_no_contract:
+; CHECK-GI: // %bb.0: // %entry
+; CHECK-GI-NEXT: fmla v0.4s, v1.4s, v3.s[3]
+; CHECK-GI-NEXT: fmul v1.4s, v2.4s, v3.s[3]
+; CHECK-GI-NEXT: fsub v0.4s, v0.4s, v1.4s
+; CHECK-GI-NEXT: ret
+entry:
+ %lane1 = shufflevector <4 x float> %v, <4 x float> undef, <4 x i32> <i32 3, i32 3, i32 3, i32 3>
+ %0 = tail call <4 x float> @llvm.fma.v4f32(<4 x float> %lane1, <4 x float> %b, <4 x float> %a)
+ %lane2 = shufflevector <4 x float> %v, <4 x float> undef, <4 x i32> <i32 3, i32 3, i32 3, i32 3>
+ %1 = fmul contract <4 x float> %lane2, %c
+ %s = fsub <4 x float> %0, %1
+ ret <4 x float> %s
+}
+
+; Likewise, the multiply-add is only fused when the fadd carries the contract
+; flag.
+define <4 x float> @no_fuse_dup_fadd_no_contract(<4 x float> %a, <4 x float> %b, <4 x float> %c, <4 x float> %v) {
+; CHECK-SD-LABEL: no_fuse_dup_fadd_no_contract:
+; CHECK-SD: // %bb.0: // %entry
+; CHECK-SD-NEXT: fmul v2.4s, v2.4s, v3.s[3]
+; CHECK-SD-NEXT: fmla v0.4s, v1.4s, v3.s[3]
+; CHECK-SD-NEXT: fadd v0.4s, v0.4s, v2.4s
+; CHECK-SD-NEXT: ret
+;
+; CHECK-GI-LABEL: no_fuse_dup_fadd_no_contract:
+; CHECK-GI: // %bb.0: // %entry
+; CHECK-GI-NEXT: fmla v0.4s, v1.4s, v3.s[3]
+; CHECK-GI-NEXT: fmul v1.4s, v2.4s, v3.s[3]
+; CHECK-GI-NEXT: fadd v0.4s, v0.4s, v1.4s
+; CHECK-GI-NEXT: ret
+entry:
+ %lane1 = shufflevector <4 x float> %v, <4 x float> undef, <4 x i32> <i32 3, i32 3, i32 3, i32 3>
+ %0 = tail call <4 x float> @llvm.fma.v4f32(<4 x float> %lane1, <4 x float> %b, <4 x float> %a)
+ %lane2 = shufflevector <4 x float> %v, <4 x float> undef, <4 x i32> <i32 3, i32 3, i32 3, i32 3>
+ %1 = fmul contract <4 x float> %lane2, %c
+ %s = fadd <4 x float> %0, %1
ret <4 x float> %s
}
``````````
</details>
https://github.com/llvm/llvm-project/pull/221404
More information about the llvm-commits
mailing list