[llvm-branch-commits] [llvm] [DAGCombiner] Mark uint_to_fp widened through trunc nuw as nneg (PR #242916)
Dmitry Sidorov via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Mon Nov 16 01:57:29 PST 2026
https://github.com/MrSidims updated https://github.com/llvm/llvm-project/pull/242916
>From 99dafcb99cc68afdcbd6c337cb70adb965226a97 Mon Sep 17 00:00:00 2001
From: Dmitry Sidorov <Dmitry.Sidorov at amd.com>
Date: Sun, 15 Nov 2026 15:09:14 +0100
Subject: [PATCH 1/2] [DAGCombiner] Mark uint_to_fp widened through trunc nuw
as nneg
The source of a trunc nuw fits in a strictly narrower type, so it is
non-negative. Without nneg x86 lowers the widened conversion through the
unsigned i64 expansion.
Assisted-by: Claude Code Opus 5.5
---
llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp | 9 +-
.../CodeGen/X86/avx512fp16-int-to-fp-trunc.ll | 2 +-
llvm/test/CodeGen/X86/combine-uitofp.ll | 86 +++----------------
3 files changed, 19 insertions(+), 78 deletions(-)
diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 9b598e92f9eab..17443cad6321b 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -21009,10 +21009,13 @@ SDValue DAGCombiner::visitUINT_TO_FP(SDNode *N) {
return DAG.getNode(ISD::UINT_TO_FP, DL, VT, Trunc);
}
- // fold (uint_to_fp (trunc nuw x)) -> (uint_to_fp x)
+ // fold (uint_to_fp (trunc nuw x)) -> (uint_to_fp nneg x)
if (N0.getOpcode() == ISD::TRUNCATE && N0->getFlags().hasNoUnsignedWrap() &&
- TLI.isTypeDesirableForOp(N, N0.getOperand(0).getValueType()))
- return DAG.getNode(ISD::UINT_TO_FP, DL, VT, N0.getOperand(0));
+ TLI.isTypeDesirableForOp(N, N0.getOperand(0).getValueType())) {
+ SDNodeFlags Flags = N->getFlags();
+ Flags.setNonNeg(true);
+ return DAG.getNode(ISD::UINT_TO_FP, DL, VT, N0.getOperand(0), Flags);
+ }
return SDValue();
}
diff --git a/llvm/test/CodeGen/X86/avx512fp16-int-to-fp-trunc.ll b/llvm/test/CodeGen/X86/avx512fp16-int-to-fp-trunc.ll
index cc918b3e94de4..a1273541680f8 100644
--- a/llvm/test/CodeGen/X86/avx512fp16-int-to-fp-trunc.ll
+++ b/llvm/test/CodeGen/X86/avx512fp16-int-to-fp-trunc.ll
@@ -244,7 +244,7 @@ define <4 x float> @uitofp_trunc_nuw_v4i64_v4f32(<4 x i64> %a0) {
;
; VLDQ-LABEL: uitofp_trunc_nuw_v4i64_v4f32:
; VLDQ: # %bb.0:
-; VLDQ-NEXT: vcvtuqq2ps %ymm0, %xmm0
+; VLDQ-NEXT: vcvtqq2ps %ymm0, %xmm0
; VLDQ-NEXT: vzeroupper
; VLDQ-NEXT: retq
%t = trunc nuw <4 x i64> %a0 to <4 x i32>
diff --git a/llvm/test/CodeGen/X86/combine-uitofp.ll b/llvm/test/CodeGen/X86/combine-uitofp.ll
index 34c6176a90fb3..d6f615137cc0e 100644
--- a/llvm/test/CodeGen/X86/combine-uitofp.ll
+++ b/llvm/test/CodeGen/X86/combine-uitofp.ll
@@ -766,50 +766,32 @@ define <16 x float> @concat_uitofp_v16f32_v8i32(<8 x i32> %a0, <8 x i32> %a1, <8
define double @uitofp_trunc_nuw_i64_i8(i64 %a0) {
; SSE2-LABEL: uitofp_trunc_nuw_i64_i8:
; SSE2: # %bb.0:
-; SSE2-NEXT: movq %rdi, %xmm1
-; SSE2-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]
-; SSE2-NEXT: subpd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE2-NEXT: movapd %xmm1, %xmm0
-; SSE2-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
-; SSE2-NEXT: addsd %xmm1, %xmm0
+; SSE2-NEXT: cvtsi2sd %rdi, %xmm0
; SSE2-NEXT: retq
;
; SSE4-LABEL: uitofp_trunc_nuw_i64_i8:
; SSE4: # %bb.0:
-; SSE4-NEXT: movq %rdi, %xmm1
-; SSE4-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]
-; SSE4-NEXT: subpd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
-; SSE4-NEXT: movapd %xmm1, %xmm0
-; SSE4-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1],xmm1[1]
-; SSE4-NEXT: addsd %xmm1, %xmm0
+; SSE4-NEXT: cvtsi2sd %rdi, %xmm0
; SSE4-NEXT: retq
;
; AVX1-LABEL: uitofp_trunc_nuw_i64_i8:
; AVX1: # %bb.0:
-; AVX1-NEXT: vmovq %rdi, %xmm0
-; AVX1-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
-; AVX1-NEXT: vsubpd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
-; AVX1-NEXT: vshufpd {{.*#+}} xmm1 = xmm0[1,0]
-; AVX1-NEXT: vaddsd %xmm0, %xmm1, %xmm0
+; AVX1-NEXT: vcvtsi2sd %rdi, %xmm15, %xmm0
; AVX1-NEXT: retq
;
; AVX2-LABEL: uitofp_trunc_nuw_i64_i8:
; AVX2: # %bb.0:
-; AVX2-NEXT: vmovq %rdi, %xmm0
-; AVX2-NEXT: vpunpckldq {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
-; AVX2-NEXT: vsubpd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
-; AVX2-NEXT: vshufpd {{.*#+}} xmm1 = xmm0[1,0]
-; AVX2-NEXT: vaddsd %xmm0, %xmm1, %xmm0
+; AVX2-NEXT: vcvtsi2sd %rdi, %xmm15, %xmm0
; AVX2-NEXT: retq
;
; AVX512-LABEL: uitofp_trunc_nuw_i64_i8:
; AVX512: # %bb.0:
-; AVX512-NEXT: vcvtusi2sd %rdi, %xmm15, %xmm0
+; AVX512-NEXT: vcvtsi2sd %rdi, %xmm15, %xmm0
; AVX512-NEXT: retq
;
; AVX512F-LABEL: uitofp_trunc_nuw_i64_i8:
; AVX512F: # %bb.0:
-; AVX512F-NEXT: vcvtusi2sd %rdi, %xmm15, %xmm0
+; AVX512F-NEXT: vcvtsi2sd %rdi, %xmm15, %xmm0
; AVX512F-NEXT: retq
%t = trunc nuw i64 %a0 to i8
%r = uitofp i8 %t to double
@@ -819,76 +801,32 @@ define double @uitofp_trunc_nuw_i64_i8(i64 %a0) {
define float @uitofp_trunc_nuw_i64_i32(i64 %a0) {
; SSE2-LABEL: uitofp_trunc_nuw_i64_i32:
; SSE2: # %bb.0:
-; SSE2-NEXT: testq %rdi, %rdi
-; SSE2-NEXT: js .LBB7_1
-; SSE2-NEXT: # %bb.2:
; SSE2-NEXT: cvtsi2ss %rdi, %xmm0
; SSE2-NEXT: retq
-; SSE2-NEXT: .LBB7_1:
-; SSE2-NEXT: movq %rdi, %rax
-; SSE2-NEXT: shrq %rax
-; SSE2-NEXT: andl $1, %edi
-; SSE2-NEXT: orq %rax, %rdi
-; SSE2-NEXT: cvtsi2ss %rdi, %xmm0
-; SSE2-NEXT: addss %xmm0, %xmm0
-; SSE2-NEXT: retq
;
; SSE4-LABEL: uitofp_trunc_nuw_i64_i32:
; SSE4: # %bb.0:
-; SSE4-NEXT: testq %rdi, %rdi
-; SSE4-NEXT: js .LBB7_1
-; SSE4-NEXT: # %bb.2:
-; SSE4-NEXT: cvtsi2ss %rdi, %xmm0
-; SSE4-NEXT: retq
-; SSE4-NEXT: .LBB7_1:
-; SSE4-NEXT: movq %rdi, %rax
-; SSE4-NEXT: shrq %rax
-; SSE4-NEXT: andl $1, %edi
-; SSE4-NEXT: orq %rax, %rdi
; SSE4-NEXT: cvtsi2ss %rdi, %xmm0
-; SSE4-NEXT: addss %xmm0, %xmm0
; SSE4-NEXT: retq
;
; AVX1-LABEL: uitofp_trunc_nuw_i64_i32:
; AVX1: # %bb.0:
-; AVX1-NEXT: testq %rdi, %rdi
-; AVX1-NEXT: js .LBB7_1
-; AVX1-NEXT: # %bb.2:
; AVX1-NEXT: vcvtsi2ss %rdi, %xmm15, %xmm0
; AVX1-NEXT: retq
-; AVX1-NEXT: .LBB7_1:
-; AVX1-NEXT: movq %rdi, %rax
-; AVX1-NEXT: shrq %rax
-; AVX1-NEXT: andl $1, %edi
-; AVX1-NEXT: orq %rax, %rdi
-; AVX1-NEXT: vcvtsi2ss %rdi, %xmm15, %xmm0
-; AVX1-NEXT: vaddss %xmm0, %xmm0, %xmm0
-; AVX1-NEXT: retq
;
; AVX2-LABEL: uitofp_trunc_nuw_i64_i32:
; AVX2: # %bb.0:
-; AVX2-NEXT: testq %rdi, %rdi
-; AVX2-NEXT: js .LBB7_1
-; AVX2-NEXT: # %bb.2:
-; AVX2-NEXT: vcvtsi2ss %rdi, %xmm15, %xmm0
-; AVX2-NEXT: retq
-; AVX2-NEXT: .LBB7_1:
-; AVX2-NEXT: movq %rdi, %rax
-; AVX2-NEXT: shrq %rax
-; AVX2-NEXT: andl $1, %edi
-; AVX2-NEXT: orq %rax, %rdi
; AVX2-NEXT: vcvtsi2ss %rdi, %xmm15, %xmm0
-; AVX2-NEXT: vaddss %xmm0, %xmm0, %xmm0
; AVX2-NEXT: retq
;
; AVX512-LABEL: uitofp_trunc_nuw_i64_i32:
; AVX512: # %bb.0:
-; AVX512-NEXT: vcvtusi2ss %rdi, %xmm15, %xmm0
+; AVX512-NEXT: vcvtsi2ss %rdi, %xmm15, %xmm0
; AVX512-NEXT: retq
;
; AVX512F-LABEL: uitofp_trunc_nuw_i64_i32:
; AVX512F: # %bb.0:
-; AVX512F-NEXT: vcvtusi2ss %rdi, %xmm15, %xmm0
+; AVX512F-NEXT: vcvtsi2ss %rdi, %xmm15, %xmm0
; AVX512F-NEXT: retq
%t = trunc nuw i64 %a0 to i32
%r = uitofp i32 %t to float
@@ -974,7 +912,7 @@ define <2 x double> @uitofp_trunc_nuw_v2i64_v2f64(<2 x i64> %a0) {
;
; AVX512-LABEL: uitofp_trunc_nuw_v2i64_v2f64:
; AVX512: # %bb.0:
-; AVX512-NEXT: vcvtuqq2pd %xmm0, %xmm0
+; AVX512-NEXT: vcvtqq2pd %xmm0, %xmm0
; AVX512-NEXT: retq
;
; AVX512F-LABEL: uitofp_trunc_nuw_v2i64_v2f64:
@@ -1036,7 +974,7 @@ define <4 x double> @uitofp_trunc_nuw_v4i64_v4f64(<4 x i64> %a0) {
;
; AVX512-LABEL: uitofp_trunc_nuw_v4i64_v4f64:
; AVX512: # %bb.0:
-; AVX512-NEXT: vcvtuqq2pd %ymm0, %ymm0
+; AVX512-NEXT: vcvtqq2pd %ymm0, %ymm0
; AVX512-NEXT: retq
;
; AVX512F-LABEL: uitofp_trunc_nuw_v4i64_v4f64:
@@ -1105,7 +1043,7 @@ define <4 x float> @uitofp_trunc_nuw_v4i64_v4f32(<4 x i64> %a0) {
;
; AVX512-LABEL: uitofp_trunc_nuw_v4i64_v4f32:
; AVX512: # %bb.0:
-; AVX512-NEXT: vcvtuqq2ps %ymm0, %xmm0
+; AVX512-NEXT: vcvtqq2ps %ymm0, %xmm0
; AVX512-NEXT: vzeroupper
; AVX512-NEXT: retq
;
@@ -1187,7 +1125,7 @@ define <8 x double> @uitofp_trunc_nuw_v8i64_v8f64(<8 x i64> %a0) {
;
; AVX512-LABEL: uitofp_trunc_nuw_v8i64_v8f64:
; AVX512: # %bb.0:
-; AVX512-NEXT: vcvtuqq2pd %zmm0, %zmm0
+; AVX512-NEXT: vcvtqq2pd %zmm0, %zmm0
; AVX512-NEXT: retq
;
; AVX512F-LABEL: uitofp_trunc_nuw_v8i64_v8f64:
>From f11c747d3b433802f6ed38e557dc91ccbd26bb3a Mon Sep 17 00:00:00 2001
From: Dmitry Sidorov <Dmitry.Sidorov at amd.com>
Date: Mon, 16 Nov 2026 10:23:36 +0100
Subject: [PATCH 2/2] Inline nneg flag
---
llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp | 8 +++-----
1 file changed, 3 insertions(+), 5 deletions(-)
diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 17443cad6321b..4feba7112a6c9 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -21011,11 +21011,9 @@ SDValue DAGCombiner::visitUINT_TO_FP(SDNode *N) {
// fold (uint_to_fp (trunc nuw x)) -> (uint_to_fp nneg x)
if (N0.getOpcode() == ISD::TRUNCATE && N0->getFlags().hasNoUnsignedWrap() &&
- TLI.isTypeDesirableForOp(N, N0.getOperand(0).getValueType())) {
- SDNodeFlags Flags = N->getFlags();
- Flags.setNonNeg(true);
- return DAG.getNode(ISD::UINT_TO_FP, DL, VT, N0.getOperand(0), Flags);
- }
+ TLI.isTypeDesirableForOp(N, N0.getOperand(0).getValueType()))
+ return DAG.getNode(ISD::UINT_TO_FP, DL, VT, N0.getOperand(0),
+ N->getFlags() | SDNodeFlags::NonNeg);
return SDValue();
}
More information about the llvm-branch-commits
mailing list