[llvm-branch-commits] [llvm] DAG: Allow narrowing uint_to_fp source to the zext source type (PR #242668)
Matt Arsenault via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Sat Nov 14 09:22:12 PST 2026
https://github.com/arsenm created https://github.com/llvm/llvm-project/pull/242668
The narrowing fold skipped uint_to_fp (zext x) when the narrow type
matched x, to avoid undoing a target combine that widened the source.
AMDGPU was the target doing this, and stopped in 50d1a75d978c. This
prevented the fold in the most direct case, so uitofp (zext i32 to i64)
was expanded as a 64-bit conversion. This also hit with topological
DAG combine ordering, where zext i16 -> i64 is first rewritten to a
zext from i32.
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 737368a97ee3fcf675c47115986bdf46b3acc292 Mon Sep 17 00:00:00 2001
From: Matt Arsenault <Matthew.Arsenault at amd.com>
Date: Thu, 29 Oct 2026 19:33:01 -0700
Subject: [PATCH] DAG: Allow narrowing uint_to_fp source to the zext source
type
The narrowing fold skipped uint_to_fp (zext x) when the narrow type
matched x, to avoid undoing a target combine that widened the source.
AMDGPU was the target doing this, and stopped in 50d1a75d978c. This
prevented the fold in the most direct case, so uitofp (zext i32 to i64)
was expanded as a 64-bit conversion. This also hit with topological
DAG combine ordering, where zext i16 -> i64 is first rewritten to a
zext from i32.
Co-authored-by: Claude Opus 5.5 <noreply at anthropic.com>
---
llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp | 5 -
.../CodeGen/AMDGPU/int_to_fp_narrow_i64.ll | 255 ++----------------
2 files changed, 17 insertions(+), 243 deletions(-)
diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index a503881aa918f..3d5f193900284 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -21001,11 +21001,6 @@ SDValue DAGCombiner::visitUINT_TO_FP(SDNode *N) {
if ((LegalTypes || OpVT.isVector()) && !TLI.isTypeLegal(NarrowVT))
continue;
- // Avoid undoing a target combine that widened the source operand.
- if (N0.getOpcode() == ISD::ZERO_EXTEND &&
- NarrowVT == N0.getOperand(0).getValueType())
- continue;
-
SDValue Trunc = DAG.getNode(ISD::TRUNCATE, DL, NarrowVT, N0);
return DAG.getNode(ISD::UINT_TO_FP, DL, VT, Trunc);
}
diff --git a/llvm/test/CodeGen/AMDGPU/int_to_fp_narrow_i64.ll b/llvm/test/CodeGen/AMDGPU/int_to_fp_narrow_i64.ll
index 73f15fe4bef6b..5fb0b0c4b1ec6 100644
--- a/llvm/test/CodeGen/AMDGPU/int_to_fp_narrow_i64.ll
+++ b/llvm/test/CodeGen/AMDGPU/int_to_fp_narrow_i64.ll
@@ -2903,48 +2903,21 @@ define half @v_uint_to_fp_zext_i32_to_f16(i32 %in) #0 {
; GFX6-LABEL: v_uint_to_fp_zext_i32_to_f16:
; GFX6: ; %bb.0:
; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX6-NEXT: v_ffbh_u32_e32 v2, 0
-; GFX6-NEXT: v_mov_b32_e32 v1, 0
-; GFX6-NEXT: v_min_u32_e32 v2, 32, v2
-; GFX6-NEXT: v_lshl_b64 v[0:1], v[0:1], v2
-; GFX6-NEXT: v_min_u32_e32 v0, 1, v0
-; GFX6-NEXT: v_or_b32_e32 v0, v1, v0
; GFX6-NEXT: v_cvt_f32_u32_e32 v0, v0
-; GFX6-NEXT: v_sub_i32_e32 v1, vcc, 32, v2
-; GFX6-NEXT: v_ldexp_f32_e32 v0, v0, v1
; GFX6-NEXT: v_cvt_f16_f32_e32 v0, v0
; GFX6-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_uint_to_fp_zext_i32_to_f16:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: v_ffbh_u32_e32 v2, 0
-; GFX8-NEXT: v_mov_b32_e32 v1, 0
-; GFX8-NEXT: v_min_u32_e32 v2, 32, v2
-; GFX8-NEXT: v_lshlrev_b64 v[0:1], v2, v[0:1]
-; GFX8-NEXT: v_min_u32_e32 v0, 1, v0
-; GFX8-NEXT: v_or_b32_e32 v0, v1, v0
; GFX8-NEXT: v_cvt_f32_u32_e32 v0, v0
-; GFX8-NEXT: v_sub_u32_e32 v1, vcc, 32, v2
-; GFX8-NEXT: v_ldexp_f32 v0, v0, v1
; GFX8-NEXT: v_cvt_f16_f32_e32 v0, v0
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: v_uint_to_fp_zext_i32_to_f16:
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT: v_clz_i32_u32_e32 v2, 0
-; GFX11-TRUE16-NEXT: v_mov_b32_e32 v1, 0
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_min_u32_e32 v2, 32, v2
-; GFX11-TRUE16-NEXT: v_lshlrev_b64 v[0:1], v2, v[0:1]
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_min_u32_e32 v0, 1, v0
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v1, v0
-; GFX11-TRUE16-NEXT: v_sub_nc_u32_e32 v1, 32, v2
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_cvt_f32_u32_e32 v0, v0
-; GFX11-TRUE16-NEXT: v_ldexp_f32 v0, v0, v1
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_cvt_f16_f32_e32 v0.l, v0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
@@ -2952,18 +2925,7 @@ define half @v_uint_to_fp_zext_i32_to_f16(i32 %in) #0 {
; GFX11-FAKE16-LABEL: v_uint_to_fp_zext_i32_to_f16:
; GFX11-FAKE16: ; %bb.0:
; GFX11-FAKE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT: v_clz_i32_u32_e32 v2, 0
-; GFX11-FAKE16-NEXT: v_mov_b32_e32 v1, 0
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT: v_min_u32_e32 v2, 32, v2
-; GFX11-FAKE16-NEXT: v_lshlrev_b64 v[0:1], v2, v[0:1]
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT: v_min_u32_e32 v0, 1, v0
-; GFX11-FAKE16-NEXT: v_or_b32_e32 v0, v1, v0
-; GFX11-FAKE16-NEXT: v_sub_nc_u32_e32 v1, 32, v2
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT: v_cvt_f32_u32_e32 v0, v0
-; GFX11-FAKE16-NEXT: v_ldexp_f32 v0, v0, v1
; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
; GFX11-FAKE16-NEXT: v_cvt_f16_f32_e32 v0, v0
; GFX11-FAKE16-NEXT: s_setpc_b64 s[30:31]
@@ -2971,45 +2933,21 @@ define half @v_uint_to_fp_zext_i32_to_f16(i32 %in) #0 {
; GFX701-LABEL: v_uint_to_fp_zext_i32_to_f16:
; GFX701: ; %bb.0:
; GFX701-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX701-NEXT: v_ffbh_u32_e32 v2, 0
-; GFX701-NEXT: v_mov_b32_e32 v1, 0
-; GFX701-NEXT: v_min_u32_e32 v2, 32, v2
-; GFX701-NEXT: v_lshl_b64 v[0:1], v[0:1], v2
-; GFX701-NEXT: v_min_u32_e32 v0, 1, v0
-; GFX701-NEXT: v_or_b32_e32 v0, v1, v0
; GFX701-NEXT: v_cvt_f32_u32_e32 v0, v0
-; GFX701-NEXT: v_sub_i32_e32 v1, vcc, 32, v2
-; GFX701-NEXT: v_ldexp_f32_e32 v0, v0, v1
; GFX701-NEXT: v_cvt_f16_f32_e32 v0, v0
; GFX701-NEXT: s_setpc_b64 s[30:31]
;
; GFX803-LABEL: v_uint_to_fp_zext_i32_to_f16:
; GFX803: ; %bb.0:
; GFX803-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX803-NEXT: v_ffbh_u32_e32 v2, 0
-; GFX803-NEXT: v_mov_b32_e32 v1, 0
-; GFX803-NEXT: v_min_u32_e32 v2, 32, v2
-; GFX803-NEXT: v_lshlrev_b64 v[0:1], v2, v[0:1]
-; GFX803-NEXT: v_min_u32_e32 v0, 1, v0
-; GFX803-NEXT: v_or_b32_e32 v0, v1, v0
; GFX803-NEXT: v_cvt_f32_u32_e32 v0, v0
-; GFX803-NEXT: v_sub_u32_e32 v1, vcc, 32, v2
-; GFX803-NEXT: v_ldexp_f32 v0, v0, v1
; GFX803-NEXT: v_cvt_f16_f32_e32 v0, v0
; GFX803-NEXT: s_setpc_b64 s[30:31]
;
; GFX942-LABEL: v_uint_to_fp_zext_i32_to_f16:
; GFX942: ; %bb.0:
; GFX942-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX942-NEXT: v_ffbh_u32_e32 v2, 0
-; GFX942-NEXT: v_mov_b32_e32 v1, 0
-; GFX942-NEXT: v_min_u32_e32 v2, 32, v2
-; GFX942-NEXT: v_lshlrev_b64 v[0:1], v2, v[0:1]
-; GFX942-NEXT: v_min_u32_e32 v0, 1, v0
-; GFX942-NEXT: v_or_b32_e32 v0, v1, v0
; GFX942-NEXT: v_cvt_f32_u32_e32 v0, v0
-; GFX942-NEXT: v_sub_u32_e32 v1, 32, v2
-; GFX942-NEXT: v_ldexp_f32 v0, v0, v1
; GFX942-NEXT: v_cvt_f16_f32_e32 v0, v0
; GFX942-NEXT: s_setpc_b64 s[30:31]
%wide = zext i32 %in to i64
@@ -3021,88 +2959,37 @@ define float @v_uint_to_fp_zext_i32_to_f32(i32 %in) #0 {
; GFX6-LABEL: v_uint_to_fp_zext_i32_to_f32:
; GFX6: ; %bb.0:
; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX6-NEXT: v_ffbh_u32_e32 v2, 0
-; GFX6-NEXT: v_mov_b32_e32 v1, 0
-; GFX6-NEXT: v_min_u32_e32 v2, 32, v2
-; GFX6-NEXT: v_lshl_b64 v[0:1], v[0:1], v2
-; GFX6-NEXT: v_min_u32_e32 v0, 1, v0
-; GFX6-NEXT: v_or_b32_e32 v0, v1, v0
; GFX6-NEXT: v_cvt_f32_u32_e32 v0, v0
-; GFX6-NEXT: v_sub_i32_e32 v1, vcc, 32, v2
-; GFX6-NEXT: v_ldexp_f32_e32 v0, v0, v1
; GFX6-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_uint_to_fp_zext_i32_to_f32:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: v_ffbh_u32_e32 v2, 0
-; GFX8-NEXT: v_mov_b32_e32 v1, 0
-; GFX8-NEXT: v_min_u32_e32 v2, 32, v2
-; GFX8-NEXT: v_lshlrev_b64 v[0:1], v2, v[0:1]
-; GFX8-NEXT: v_min_u32_e32 v0, 1, v0
-; GFX8-NEXT: v_or_b32_e32 v0, v1, v0
; GFX8-NEXT: v_cvt_f32_u32_e32 v0, v0
-; GFX8-NEXT: v_sub_u32_e32 v1, vcc, 32, v2
-; GFX8-NEXT: v_ldexp_f32 v0, v0, v1
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_uint_to_fp_zext_i32_to_f32:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_clz_i32_u32_e32 v2, 0
-; GFX11-NEXT: v_mov_b32_e32 v1, 0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_min_u32_e32 v2, 32, v2
-; GFX11-NEXT: v_lshlrev_b64 v[0:1], v2, v[0:1]
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_min_u32_e32 v0, 1, v0
-; GFX11-NEXT: v_or_b32_e32 v0, v1, v0
-; GFX11-NEXT: v_sub_nc_u32_e32 v1, 32, v2
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-NEXT: v_cvt_f32_u32_e32 v0, v0
-; GFX11-NEXT: v_ldexp_f32 v0, v0, v1
; GFX11-NEXT: s_setpc_b64 s[30:31]
;
; GFX701-LABEL: v_uint_to_fp_zext_i32_to_f32:
; GFX701: ; %bb.0:
; GFX701-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX701-NEXT: v_ffbh_u32_e32 v2, 0
-; GFX701-NEXT: v_mov_b32_e32 v1, 0
-; GFX701-NEXT: v_min_u32_e32 v2, 32, v2
-; GFX701-NEXT: v_lshl_b64 v[0:1], v[0:1], v2
-; GFX701-NEXT: v_min_u32_e32 v0, 1, v0
-; GFX701-NEXT: v_or_b32_e32 v0, v1, v0
; GFX701-NEXT: v_cvt_f32_u32_e32 v0, v0
-; GFX701-NEXT: v_sub_i32_e32 v1, vcc, 32, v2
-; GFX701-NEXT: v_ldexp_f32_e32 v0, v0, v1
; GFX701-NEXT: s_setpc_b64 s[30:31]
;
; GFX803-LABEL: v_uint_to_fp_zext_i32_to_f32:
; GFX803: ; %bb.0:
; GFX803-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX803-NEXT: v_ffbh_u32_e32 v2, 0
-; GFX803-NEXT: v_mov_b32_e32 v1, 0
-; GFX803-NEXT: v_min_u32_e32 v2, 32, v2
-; GFX803-NEXT: v_lshlrev_b64 v[0:1], v2, v[0:1]
-; GFX803-NEXT: v_min_u32_e32 v0, 1, v0
-; GFX803-NEXT: v_or_b32_e32 v0, v1, v0
; GFX803-NEXT: v_cvt_f32_u32_e32 v0, v0
-; GFX803-NEXT: v_sub_u32_e32 v1, vcc, 32, v2
-; GFX803-NEXT: v_ldexp_f32 v0, v0, v1
; GFX803-NEXT: s_setpc_b64 s[30:31]
;
; GFX942-LABEL: v_uint_to_fp_zext_i32_to_f32:
; GFX942: ; %bb.0:
; GFX942-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX942-NEXT: v_ffbh_u32_e32 v2, 0
-; GFX942-NEXT: v_mov_b32_e32 v1, 0
-; GFX942-NEXT: v_min_u32_e32 v2, 32, v2
-; GFX942-NEXT: v_lshlrev_b64 v[0:1], v2, v[0:1]
-; GFX942-NEXT: v_min_u32_e32 v0, 1, v0
-; GFX942-NEXT: v_or_b32_e32 v0, v1, v0
; GFX942-NEXT: v_cvt_f32_u32_e32 v0, v0
-; GFX942-NEXT: v_sub_u32_e32 v1, 32, v2
-; GFX942-NEXT: v_ldexp_f32 v0, v0, v1
; GFX942-NEXT: s_setpc_b64 s[30:31]
%wide = zext i32 %in to i64
%result = uitofp i64 %wide to float
@@ -3113,56 +3000,37 @@ define double @v_uint_to_fp_zext_i32_to_f64(i32 %in) #0 {
; GFX6-LABEL: v_uint_to_fp_zext_i32_to_f64:
; GFX6: ; %bb.0:
; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX6-NEXT: v_cvt_f64_u32_e32 v[1:2], 0
-; GFX6-NEXT: v_cvt_f64_u32_e32 v[3:4], v0
-; GFX6-NEXT: v_ldexp_f64 v[0:1], v[1:2], 32
-; GFX6-NEXT: v_add_f64 v[0:1], v[0:1], v[3:4]
+; GFX6-NEXT: v_cvt_f64_u32_e32 v[0:1], v0
; GFX6-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_uint_to_fp_zext_i32_to_f64:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: v_cvt_f64_u32_e32 v[1:2], 0
-; GFX8-NEXT: v_cvt_f64_u32_e32 v[3:4], v0
-; GFX8-NEXT: v_ldexp_f64 v[1:2], v[1:2], 32
-; GFX8-NEXT: v_add_f64 v[0:1], v[1:2], v[3:4]
+; GFX8-NEXT: v_cvt_f64_u32_e32 v[0:1], v0
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_uint_to_fp_zext_i32_to_f64:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_cvt_f64_u32_e32 v[1:2], 0
-; GFX11-NEXT: v_cvt_f64_u32_e32 v[3:4], v0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_ldexp_f64 v[0:1], v[1:2], 32
-; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], v[3:4]
+; GFX11-NEXT: v_cvt_f64_u32_e32 v[0:1], v0
; GFX11-NEXT: s_setpc_b64 s[30:31]
;
; GFX701-LABEL: v_uint_to_fp_zext_i32_to_f64:
; GFX701: ; %bb.0:
; GFX701-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX701-NEXT: v_cvt_f64_u32_e32 v[1:2], 0
-; GFX701-NEXT: v_cvt_f64_u32_e32 v[3:4], v0
-; GFX701-NEXT: v_ldexp_f64 v[0:1], v[1:2], 32
-; GFX701-NEXT: v_add_f64 v[0:1], v[0:1], v[3:4]
+; GFX701-NEXT: v_cvt_f64_u32_e32 v[0:1], v0
; GFX701-NEXT: s_setpc_b64 s[30:31]
;
; GFX803-LABEL: v_uint_to_fp_zext_i32_to_f64:
; GFX803: ; %bb.0:
; GFX803-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX803-NEXT: v_cvt_f64_u32_e32 v[1:2], 0
-; GFX803-NEXT: v_cvt_f64_u32_e32 v[3:4], v0
-; GFX803-NEXT: v_ldexp_f64 v[1:2], v[1:2], 32
-; GFX803-NEXT: v_add_f64 v[0:1], v[1:2], v[3:4]
+; GFX803-NEXT: v_cvt_f64_u32_e32 v[0:1], v0
; GFX803-NEXT: s_setpc_b64 s[30:31]
;
; GFX942-LABEL: v_uint_to_fp_zext_i32_to_f64:
; GFX942: ; %bb.0:
; GFX942-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX942-NEXT: v_cvt_f64_u32_e32 v[2:3], 0
-; GFX942-NEXT: v_ldexp_f64 v[2:3], v[2:3], 32
; GFX942-NEXT: v_cvt_f64_u32_e32 v[0:1], v0
-; GFX942-NEXT: v_add_f64 v[0:1], v[2:3], v[0:1]
; GFX942-NEXT: s_setpc_b64 s[30:31]
%wide = zext i32 %in to i64
%result = uitofp i64 %wide to double
@@ -3173,132 +3041,43 @@ define <2 x float> @v_uint_to_fp_zext_v2i32_to_v2f32(<2 x i32> %in) #0 {
; GFX6-LABEL: v_uint_to_fp_zext_v2i32_to_v2f32:
; GFX6: ; %bb.0:
; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX6-NEXT: v_mov_b32_e32 v3, v0
-; GFX6-NEXT: v_mov_b32_e32 v2, 0
-; GFX6-NEXT: v_ffbh_u32_e32 v0, 0
-; GFX6-NEXT: v_mov_b32_e32 v4, v2
-; GFX6-NEXT: v_min_u32_e32 v5, 32, v0
-; GFX6-NEXT: v_lshl_b64 v[3:4], v[3:4], v5
-; GFX6-NEXT: v_min_u32_e32 v0, 1, v3
-; GFX6-NEXT: v_or_b32_e32 v3, v4, v0
-; GFX6-NEXT: v_lshl_b64 v[0:1], v[1:2], v5
-; GFX6-NEXT: v_cvt_f32_u32_e32 v2, v3
-; GFX6-NEXT: v_min_u32_e32 v0, 1, v0
-; GFX6-NEXT: v_or_b32_e32 v0, v1, v0
-; GFX6-NEXT: v_cvt_f32_u32_e32 v1, v0
-; GFX6-NEXT: v_sub_i32_e32 v3, vcc, 32, v5
-; GFX6-NEXT: v_ldexp_f32_e32 v0, v2, v3
-; GFX6-NEXT: v_ldexp_f32_e32 v1, v1, v3
+; GFX6-NEXT: v_cvt_f32_u32_e32 v0, v0
+; GFX6-NEXT: v_cvt_f32_u32_e32 v1, v1
; GFX6-NEXT: s_setpc_b64 s[30:31]
;
; GFX8-LABEL: v_uint_to_fp_zext_v2i32_to_v2f32:
; GFX8: ; %bb.0:
; GFX8-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8-NEXT: v_mov_b32_e32 v3, v0
-; GFX8-NEXT: v_mov_b32_e32 v2, 0
-; GFX8-NEXT: v_ffbh_u32_e32 v0, 0
-; GFX8-NEXT: v_mov_b32_e32 v4, v2
-; GFX8-NEXT: v_min_u32_e32 v5, 32, v0
-; GFX8-NEXT: v_lshlrev_b64 v[3:4], v5, v[3:4]
-; GFX8-NEXT: v_min_u32_e32 v0, 1, v3
-; GFX8-NEXT: v_or_b32_e32 v3, v4, v0
-; GFX8-NEXT: v_lshlrev_b64 v[0:1], v5, v[1:2]
-; GFX8-NEXT: v_cvt_f32_u32_e32 v2, v3
-; GFX8-NEXT: v_min_u32_e32 v0, 1, v0
-; GFX8-NEXT: v_or_b32_e32 v0, v1, v0
-; GFX8-NEXT: v_cvt_f32_u32_e32 v1, v0
-; GFX8-NEXT: v_sub_u32_e32 v3, vcc, 32, v5
-; GFX8-NEXT: v_ldexp_f32 v0, v2, v3
-; GFX8-NEXT: v_ldexp_f32 v1, v1, v3
+; GFX8-NEXT: v_cvt_f32_u32_e32 v0, v0
+; GFX8-NEXT: v_cvt_f32_u32_e32 v1, v1
; GFX8-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-LABEL: v_uint_to_fp_zext_v2i32_to_v2f32:
; GFX11: ; %bb.0:
; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-NEXT: v_dual_mov_b32 v3, v0 :: v_dual_mov_b32 v2, 0
-; GFX11-NEXT: v_clz_i32_u32_e32 v0, 0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_mov_b32_e32 v4, v2
-; GFX11-NEXT: v_min_u32_e32 v5, 32, v0
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_lshlrev_b64 v[3:4], v5, v[3:4]
-; GFX11-NEXT: v_lshlrev_b64 v[0:1], v5, v[1:2]
-; GFX11-NEXT: v_min_u32_e32 v2, 1, v3
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_min_u32_e32 v0, 1, v0
-; GFX11-NEXT: v_or_b32_e32 v2, v4, v2
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-NEXT: v_or_b32_e32 v0, v1, v0
-; GFX11-NEXT: v_sub_nc_u32_e32 v1, 32, v5
-; GFX11-NEXT: v_cvt_f32_u32_e32 v2, v2
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT: v_cvt_f32_u32_e32 v3, v0
-; GFX11-NEXT: v_ldexp_f32 v0, v2, v1
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX11-NEXT: v_ldexp_f32 v1, v3, v1
+; GFX11-NEXT: v_cvt_f32_u32_e32 v0, v0
+; GFX11-NEXT: v_cvt_f32_u32_e32 v1, v1
; GFX11-NEXT: s_setpc_b64 s[30:31]
;
; GFX701-LABEL: v_uint_to_fp_zext_v2i32_to_v2f32:
; GFX701: ; %bb.0:
; GFX701-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX701-NEXT: v_mov_b32_e32 v3, v0
-; GFX701-NEXT: v_mov_b32_e32 v2, 0
-; GFX701-NEXT: v_ffbh_u32_e32 v0, 0
-; GFX701-NEXT: v_mov_b32_e32 v4, v2
-; GFX701-NEXT: v_min_u32_e32 v5, 32, v0
-; GFX701-NEXT: v_lshl_b64 v[3:4], v[3:4], v5
-; GFX701-NEXT: v_min_u32_e32 v0, 1, v3
-; GFX701-NEXT: v_or_b32_e32 v3, v4, v0
-; GFX701-NEXT: v_lshl_b64 v[0:1], v[1:2], v5
-; GFX701-NEXT: v_cvt_f32_u32_e32 v2, v3
-; GFX701-NEXT: v_min_u32_e32 v0, 1, v0
-; GFX701-NEXT: v_or_b32_e32 v0, v1, v0
-; GFX701-NEXT: v_cvt_f32_u32_e32 v1, v0
-; GFX701-NEXT: v_sub_i32_e32 v3, vcc, 32, v5
-; GFX701-NEXT: v_ldexp_f32_e32 v0, v2, v3
-; GFX701-NEXT: v_ldexp_f32_e32 v1, v1, v3
+; GFX701-NEXT: v_cvt_f32_u32_e32 v0, v0
+; GFX701-NEXT: v_cvt_f32_u32_e32 v1, v1
; GFX701-NEXT: s_setpc_b64 s[30:31]
;
; GFX803-LABEL: v_uint_to_fp_zext_v2i32_to_v2f32:
; GFX803: ; %bb.0:
; GFX803-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX803-NEXT: v_mov_b32_e32 v3, v0
-; GFX803-NEXT: v_mov_b32_e32 v2, 0
-; GFX803-NEXT: v_ffbh_u32_e32 v0, 0
-; GFX803-NEXT: v_mov_b32_e32 v4, v2
-; GFX803-NEXT: v_min_u32_e32 v5, 32, v0
-; GFX803-NEXT: v_lshlrev_b64 v[3:4], v5, v[3:4]
-; GFX803-NEXT: v_min_u32_e32 v0, 1, v3
-; GFX803-NEXT: v_or_b32_e32 v3, v4, v0
-; GFX803-NEXT: v_lshlrev_b64 v[0:1], v5, v[1:2]
-; GFX803-NEXT: v_cvt_f32_u32_e32 v2, v3
-; GFX803-NEXT: v_min_u32_e32 v0, 1, v0
-; GFX803-NEXT: v_or_b32_e32 v0, v1, v0
-; GFX803-NEXT: v_cvt_f32_u32_e32 v1, v0
-; GFX803-NEXT: v_sub_u32_e32 v3, vcc, 32, v5
-; GFX803-NEXT: v_ldexp_f32 v0, v2, v3
-; GFX803-NEXT: v_ldexp_f32 v1, v1, v3
+; GFX803-NEXT: v_cvt_f32_u32_e32 v0, v0
+; GFX803-NEXT: v_cvt_f32_u32_e32 v1, v1
; GFX803-NEXT: s_setpc_b64 s[30:31]
;
; GFX942-LABEL: v_uint_to_fp_zext_v2i32_to_v2f32:
; GFX942: ; %bb.0:
; GFX942-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX942-NEXT: v_mov_b32_e32 v3, 0
-; GFX942-NEXT: v_ffbh_u32_e32 v4, 0
-; GFX942-NEXT: v_mov_b32_e32 v2, v1
-; GFX942-NEXT: v_mov_b32_e32 v1, v3
-; GFX942-NEXT: v_min_u32_e32 v4, 32, v4
-; GFX942-NEXT: v_lshlrev_b64 v[0:1], v4, v[0:1]
-; GFX942-NEXT: v_min_u32_e32 v0, 1, v0
-; GFX942-NEXT: v_or_b32_e32 v0, v1, v0
-; GFX942-NEXT: v_cvt_f32_u32_e32 v5, v0
-; GFX942-NEXT: v_lshlrev_b64 v[0:1], v4, v[2:3]
-; GFX942-NEXT: v_min_u32_e32 v0, 1, v0
-; GFX942-NEXT: v_or_b32_e32 v0, v1, v0
-; GFX942-NEXT: v_cvt_f32_u32_e32 v1, v0
-; GFX942-NEXT: v_sub_u32_e32 v2, 32, v4
-; GFX942-NEXT: v_ldexp_f32 v0, v5, v2
-; GFX942-NEXT: v_ldexp_f32 v1, v1, v2
+; GFX942-NEXT: v_cvt_f32_u32_e32 v0, v0
+; GFX942-NEXT: v_cvt_f32_u32_e32 v1, v1
; GFX942-NEXT: s_setpc_b64 s[30:31]
%wide = zext <2 x i32> %in to <2 x i64>
%result = uitofp <2 x i64> %wide to <2 x float>
More information about the llvm-branch-commits
mailing list