[llvm-branch-commits] [llvm] DAG: Look through fp_extend/fp_round when folding bitcast of fcopysign (PR #242587)
via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Fri Nov 13 21:50:43 PST 2026
llvmorg-github-actions[bot] wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-backend-amdgpu
Author: Matt Arsenault (arsenm)
<details>
<summary>Changes</summary>
visitFCOPYSIGN strips these from the sign operand, but the bitcast may be
visited first, leaving the conversion live in the integer form. Strip it
in the fold too, so the result does not depend on visitation order.
Co-authored-by: Claude Opus 5.5 <noreply@<!-- -->anthropic.com>
---
Patch is 30.22 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/242587.diff
5 Files Affected:
- (modified) llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp (+35-31)
- (modified) llvm/test/CodeGen/AMDGPU/fcopysign.bf16.ll (+14-131)
- (modified) llvm/test/CodeGen/AMDGPU/fcopysign.f16.ll (+19-223)
- (modified) llvm/test/CodeGen/AMDGPU/fcopysign.f32.ll (+7-14)
- (modified) llvm/test/CodeGen/AMDGPU/fcopysign.f64.ll (+7-12)
``````````diff
diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 99d34cae64392..2e09de55451cb 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -18445,6 +18445,34 @@ SDValue DAGCombiner::foldBitcastedFPLogic(SDNode *N, SelectionDAG &DAG,
return SDValue();
}
+/// copysign(x, fp_extend(y)) -> copysign(x, y)
+/// copysign(x, fp_round(y)) -> copysign(x, y)
+/// Operands to the functions are the type of X and Y respectively.
+static inline bool CanCombineFCOPYSIGN_EXTEND_ROUND(EVT XTy, EVT YTy) {
+ // Always fold no-op FP casts.
+ if (XTy == YTy)
+ return true;
+
+ // Do not optimize out type conversion of f128 type yet.
+ // For some targets like x86_64, configuration is changed to keep one f128
+ // value in one SSE register, but instruction selection cannot handle
+ // FCOPYSIGN on SSE registers yet.
+ if (YTy == MVT::f128)
+ return false;
+
+ // Avoid mismatched vector operand types, for better instruction selection.
+ return !YTy.isVector();
+}
+
+static inline bool CanCombineFCOPYSIGN_EXTEND_ROUND(SDNode *N) {
+ SDValue N1 = N->getOperand(1);
+ if (N1.getOpcode() != ISD::FP_EXTEND && N1.getOpcode() != ISD::FP_ROUND)
+ return false;
+ EVT N1VT = N1->getValueType(0);
+ EVT N1Op0VT = N1->getOperand(0).getValueType();
+ return CanCombineFCOPYSIGN_EXTEND_ROUND(N1VT, N1Op0VT);
+}
+
SDValue DAGCombiner::visitBITCAST(SDNode *N) {
SDValue N0 = N->getOperand(0);
EVT VT = N->getValueType(0);
@@ -18626,10 +18654,15 @@ SDValue DAGCombiner::visitBITCAST(SDNode *N) {
if (N0.getOpcode() == ISD::FCOPYSIGN && N0->hasOneUse() &&
isa<ConstantFPSDNode>(N0.getOperand(0)) && VT.isInteger() &&
!VT.isVector()) {
- unsigned OrigXWidth = N0.getOperand(1).getValueSizeInBits();
+ SDValue Sign = N0.getOperand(1);
+ if (N0.getValueType() != MVT::ppcf128 &&
+ CanCombineFCOPYSIGN_EXTEND_ROUND(N0.getNode()))
+ Sign = Sign.getOperand(0);
+
+ unsigned OrigXWidth = Sign.getValueSizeInBits();
EVT IntXVT = EVT::getIntegerVT(*DAG.getContext(), OrigXWidth);
if (isTypeLegal(IntXVT)) {
- SDValue X = DAG.getBitcast(IntXVT, N0.getOperand(1));
+ SDValue X = DAG.getBitcast(IntXVT, Sign);
AddToWorklist(X.getNode());
// If X has a different width than the result, extend or truncate it.
@@ -20571,35 +20604,6 @@ SDValue DAGCombiner::visitFSQRT(SDNode *N) {
return buildSqrtEstimate(N0, Flags);
}
-/// copysign(x, fp_extend(y)) -> copysign(x, y)
-/// copysign(x, fp_round(y)) -> copysign(x, y)
-/// Operands to the functions are the type of X and Y respectively.
-static inline bool CanCombineFCOPYSIGN_EXTEND_ROUND(EVT XTy, EVT YTy) {
- // Always fold no-op FP casts.
- if (XTy == YTy)
- return true;
-
- // Do not optimize out type conversion of f128 type yet.
- // For some targets like x86_64, configuration is changed to keep one f128
- // value in one SSE register, but instruction selection cannot handle
- // FCOPYSIGN on SSE registers yet.
- if (YTy == MVT::f128)
- return false;
-
- // Avoid mismatched vector operand types, for better instruction selection.
- return !YTy.isVector();
-}
-
-static inline bool CanCombineFCOPYSIGN_EXTEND_ROUND(SDNode *N) {
- SDValue N1 = N->getOperand(1);
- if (N1.getOpcode() != ISD::FP_EXTEND &&
- N1.getOpcode() != ISD::FP_ROUND)
- return false;
- EVT N1VT = N1->getValueType(0);
- EVT N1Op0VT = N1->getOperand(0).getValueType();
- return CanCombineFCOPYSIGN_EXTEND_ROUND(N1VT, N1Op0VT);
-}
-
SDValue DAGCombiner::visitFCOPYSIGN(SDNode *N) {
SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);
diff --git a/llvm/test/CodeGen/AMDGPU/fcopysign.bf16.ll b/llvm/test/CodeGen/AMDGPU/fcopysign.bf16.ll
index 5c7932f388b88..21a3868379ce4 100644
--- a/llvm/test/CodeGen/AMDGPU/fcopysign.bf16.ll
+++ b/llvm/test/CodeGen/AMDGPU/fcopysign.bf16.ll
@@ -7305,65 +7305,36 @@ define bfloat @v_copysign_bf16_0_bf16(bfloat %sign) {
define amdgpu_ps i32 @s_copysign_bf16_0_f32(float inreg %sign) {
; GCN-LABEL: s_copysign_bf16_0_f32:
; GCN: ; %bb.0:
-; GCN-NEXT: v_mul_f32_e64 v0, 1.0, s0
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_and_b32_e32 v0, 0x8000, v0
-; GCN-NEXT: v_readfirstlane_b32 s0, v0
+; GCN-NEXT: s_lshr_b32 s0, s0, 16
+; GCN-NEXT: s_and_b32 s0, s0, 0x8000
; GCN-NEXT: ; return to shader part epilog
;
; GFX7-LABEL: s_copysign_bf16_0_f32:
; GFX7: ; %bb.0:
-; GFX7-NEXT: v_mul_f32_e64 v0, 1.0, s0
-; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
-; GFX7-NEXT: v_and_b32_e32 v0, 0x8000, v0
-; GFX7-NEXT: v_readfirstlane_b32 s0, v0
+; GFX7-NEXT: s_lshr_b32 s0, s0, 16
+; GFX7-NEXT: s_and_b32 s0, s0, 0x8000
; GFX7-NEXT: ; return to shader part epilog
;
; GFX8-LABEL: s_copysign_bf16_0_f32:
; GFX8: ; %bb.0:
-; GFX8-NEXT: s_bfe_u32 s1, s0, 0x10010
-; GFX8-NEXT: s_add_i32 s1, s1, s0
-; GFX8-NEXT: s_addk_i32 s1, 0x7fff
-; GFX8-NEXT: v_cmp_u_f32_e64 s[2:3], s0, s0
-; GFX8-NEXT: s_cmp_lg_u64 s[2:3], 0
-; GFX8-NEXT: s_cselect_b32 s0, s0, s1
; GFX8-NEXT: s_lshr_b32 s0, s0, 16
; GFX8-NEXT: s_and_b32 s0, s0, 0x8000
; GFX8-NEXT: ; return to shader part epilog
;
; GFX9-LABEL: s_copysign_bf16_0_f32:
; GFX9: ; %bb.0:
-; GFX9-NEXT: s_bfe_u32 s1, s0, 0x10010
-; GFX9-NEXT: s_add_i32 s1, s1, s0
-; GFX9-NEXT: s_addk_i32 s1, 0x7fff
-; GFX9-NEXT: v_cmp_u_f32_e64 s[2:3], s0, s0
-; GFX9-NEXT: s_cmp_lg_u64 s[2:3], 0
-; GFX9-NEXT: s_cselect_b32 s0, s0, s1
; GFX9-NEXT: s_lshr_b32 s0, s0, 16
; GFX9-NEXT: s_and_b32 s0, s0, 0x8000
; GFX9-NEXT: ; return to shader part epilog
;
; GFX10-LABEL: s_copysign_bf16_0_f32:
; GFX10: ; %bb.0:
-; GFX10-NEXT: s_bfe_u32 s1, s0, 0x10010
-; GFX10-NEXT: v_cmp_u_f32_e64 s2, s0, s0
-; GFX10-NEXT: s_add_i32 s1, s1, s0
-; GFX10-NEXT: s_addk_i32 s1, 0x7fff
-; GFX10-NEXT: s_cmp_lg_u32 s2, 0
-; GFX10-NEXT: s_cselect_b32 s0, s0, s1
; GFX10-NEXT: s_lshr_b32 s0, s0, 16
; GFX10-NEXT: s_and_b32 s0, s0, 0x8000
; GFX10-NEXT: ; return to shader part epilog
;
; GFX11-LABEL: s_copysign_bf16_0_f32:
; GFX11: ; %bb.0:
-; GFX11-NEXT: s_bfe_u32 s1, s0, 0x10010
-; GFX11-NEXT: v_cmp_u_f32_e64 s2, s0, s0
-; GFX11-NEXT: s_add_i32 s1, s1, s0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-NEXT: s_addk_i32 s1, 0x7fff
-; GFX11-NEXT: s_cmp_lg_u32 s2, 0
-; GFX11-NEXT: s_cselect_b32 s0, s0, s1
; GFX11-NEXT: s_lshr_b32 s0, s0, 16
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: s_and_b32 s0, s0, 0x8000
@@ -7379,7 +7350,6 @@ define bfloat @v_copysign_bf16_0_f32(float %sign) {
; GCN-LABEL: v_copysign_bf16_0_f32:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GCN-NEXT: v_and_b32_e32 v0, 0x8000, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
@@ -7387,7 +7357,6 @@ define bfloat @v_copysign_bf16_0_f32(float %sign) {
; GFX7-LABEL: v_copysign_bf16_0_f32:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0
; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; GFX7-NEXT: v_and_b32_e32 v0, 0x8000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
@@ -7434,121 +7403,37 @@ define bfloat @v_copysign_bf16_0_f32(float %sign) {
define amdgpu_ps i32 @s_copysign_bf16_0_f64(double inreg %sign) {
; GCN-LABEL: s_copysign_bf16_0_f64:
; GCN: ; %bb.0:
-; GCN-NEXT: v_cvt_f32_f64_e32 v0, s[0:1]
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v0
-; GCN-NEXT: v_and_b32_e32 v0, 0x8000, v0
-; GCN-NEXT: v_readfirstlane_b32 s0, v0
+; GCN-NEXT: s_lshr_b32 s0, s1, 16
+; GCN-NEXT: s_and_b32 s0, s0, 0x8000
; GCN-NEXT: ; return to shader part epilog
;
; GFX7-LABEL: s_copysign_bf16_0_f64:
; GFX7: ; %bb.0:
-; GFX7-NEXT: v_cvt_f32_f64_e32 v0, s[0:1]
-; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
-; GFX7-NEXT: v_and_b32_e32 v0, 0x8000, v0
-; GFX7-NEXT: v_readfirstlane_b32 s0, v0
+; GFX7-NEXT: s_lshr_b32 s0, s1, 16
+; GFX7-NEXT: s_and_b32 s0, s0, 0x8000
; GFX7-NEXT: ; return to shader part epilog
;
; GFX8-LABEL: s_copysign_bf16_0_f64:
; GFX8: ; %bb.0:
-; GFX8-NEXT: v_cvt_f32_f64_e32 v2, s[0:1]
-; GFX8-NEXT: v_cvt_f64_f32_e32 v[0:1], v2
-; GFX8-NEXT: v_readfirstlane_b32 s6, v2
-; GFX8-NEXT: s_bitcmp1_b32 s6, 0
-; GFX8-NEXT: s_cselect_b64 s[4:5], -1, 0
-; GFX8-NEXT: v_cmp_nlg_f64_e32 vcc, s[0:1], v[0:1]
-; GFX8-NEXT: v_cmp_gt_f64_e64 s[2:3], |s[0:1]|, |v[0:1]|
-; GFX8-NEXT: v_cmp_u_f64_e64 s[0:1], s[0:1], s[0:1]
-; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX8-NEXT: s_cmp_lg_u64 s[2:3], 0
-; GFX8-NEXT: s_cselect_b32 s2, 1, -1
-; GFX8-NEXT: s_add_i32 s7, s6, s2
-; GFX8-NEXT: s_and_b64 s[2:3], s[4:5], exec
-; GFX8-NEXT: s_cselect_b32 s2, s6, s7
-; GFX8-NEXT: s_bfe_u32 s3, s2, 0x10010
-; GFX8-NEXT: s_add_i32 s3, s3, s2
-; GFX8-NEXT: s_addk_i32 s3, 0x7fff
-; GFX8-NEXT: s_cmp_lg_u64 s[0:1], 0
-; GFX8-NEXT: s_cselect_b32 s0, s2, s3
-; GFX8-NEXT: s_lshr_b32 s0, s0, 16
+; GFX8-NEXT: s_lshr_b32 s0, s1, 16
; GFX8-NEXT: s_and_b32 s0, s0, 0x8000
; GFX8-NEXT: ; return to shader part epilog
;
; GFX9-LABEL: s_copysign_bf16_0_f64:
; GFX9: ; %bb.0:
-; GFX9-NEXT: v_cvt_f32_f64_e32 v2, s[0:1]
-; GFX9-NEXT: v_cvt_f64_f32_e32 v[0:1], v2
-; GFX9-NEXT: v_readfirstlane_b32 s6, v2
-; GFX9-NEXT: s_bitcmp1_b32 s6, 0
-; GFX9-NEXT: s_cselect_b64 s[4:5], -1, 0
-; GFX9-NEXT: v_cmp_nlg_f64_e32 vcc, s[0:1], v[0:1]
-; GFX9-NEXT: v_cmp_gt_f64_e64 s[2:3], |s[0:1]|, |v[0:1]|
-; GFX9-NEXT: v_cmp_u_f64_e64 s[0:1], s[0:1], s[0:1]
-; GFX9-NEXT: s_or_b64 s[4:5], vcc, s[4:5]
-; GFX9-NEXT: s_cmp_lg_u64 s[2:3], 0
-; GFX9-NEXT: s_cselect_b32 s2, 1, -1
-; GFX9-NEXT: s_add_i32 s7, s6, s2
-; GFX9-NEXT: s_and_b64 s[2:3], s[4:5], exec
-; GFX9-NEXT: s_cselect_b32 s2, s6, s7
-; GFX9-NEXT: s_bfe_u32 s3, s2, 0x10010
-; GFX9-NEXT: s_add_i32 s3, s3, s2
-; GFX9-NEXT: s_addk_i32 s3, 0x7fff
-; GFX9-NEXT: s_cmp_lg_u64 s[0:1], 0
-; GFX9-NEXT: s_cselect_b32 s0, s2, s3
-; GFX9-NEXT: s_lshr_b32 s0, s0, 16
+; GFX9-NEXT: s_lshr_b32 s0, s1, 16
; GFX9-NEXT: s_and_b32 s0, s0, 0x8000
; GFX9-NEXT: ; return to shader part epilog
;
; GFX10-LABEL: s_copysign_bf16_0_f64:
; GFX10: ; %bb.0:
-; GFX10-NEXT: v_cvt_f32_f64_e32 v2, s[0:1]
-; GFX10-NEXT: v_cvt_f64_f32_e32 v[0:1], v2
-; GFX10-NEXT: v_readfirstlane_b32 s3, v2
-; GFX10-NEXT: s_bitcmp1_b32 s3, 0
-; GFX10-NEXT: v_cmp_nlg_f64_e32 vcc_lo, s[0:1], v[0:1]
-; GFX10-NEXT: v_cmp_gt_f64_e64 s2, |s[0:1]|, |v[0:1]|
-; GFX10-NEXT: v_cmp_u_f64_e64 s0, s[0:1], s[0:1]
-; GFX10-NEXT: s_cselect_b32 s1, -1, 0
-; GFX10-NEXT: s_or_b32 s1, vcc_lo, s1
-; GFX10-NEXT: s_cmp_lg_u32 s2, 0
-; GFX10-NEXT: s_cselect_b32 s2, 1, -1
-; GFX10-NEXT: s_add_i32 s2, s3, s2
-; GFX10-NEXT: s_and_b32 s1, s1, exec_lo
-; GFX10-NEXT: s_cselect_b32 s1, s3, s2
-; GFX10-NEXT: s_bfe_u32 s2, s1, 0x10010
-; GFX10-NEXT: s_add_i32 s2, s2, s1
-; GFX10-NEXT: s_addk_i32 s2, 0x7fff
-; GFX10-NEXT: s_cmp_lg_u32 s0, 0
-; GFX10-NEXT: s_cselect_b32 s0, s1, s2
-; GFX10-NEXT: s_lshr_b32 s0, s0, 16
+; GFX10-NEXT: s_lshr_b32 s0, s1, 16
; GFX10-NEXT: s_and_b32 s0, s0, 0x8000
; GFX10-NEXT: ; return to shader part epilog
;
; GFX11-LABEL: s_copysign_bf16_0_f64:
; GFX11: ; %bb.0:
-; GFX11-NEXT: v_cvt_f32_f64_e32 v2, s[0:1]
-; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT: v_cvt_f64_f32_e32 v[0:1], v2
-; GFX11-NEXT: v_cmp_nlg_f64_e32 vcc_lo, s[0:1], v[0:1]
-; GFX11-NEXT: v_cmp_gt_f64_e64 s2, |s[0:1]|, |v[0:1]|
-; GFX11-NEXT: v_cmp_u_f64_e64 s0, s[0:1], s[0:1]
-; GFX11-NEXT: v_readfirstlane_b32 s1, v2
-; GFX11-NEXT: s_bitcmp1_b32 s1, 0
-; GFX11-NEXT: s_cselect_b32 s3, -1, 0
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-NEXT: s_or_b32 s3, vcc_lo, s3
-; GFX11-NEXT: s_cmp_lg_u32 s2, 0
-; GFX11-NEXT: s_cselect_b32 s2, 1, -1
-; GFX11-NEXT: s_add_i32 s2, s1, s2
-; GFX11-NEXT: s_and_b32 s3, s3, exec_lo
-; GFX11-NEXT: s_cselect_b32 s1, s1, s2
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
-; GFX11-NEXT: s_bfe_u32 s2, s1, 0x10010
-; GFX11-NEXT: s_add_i32 s2, s2, s1
-; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_2) | instid1(SALU_CYCLE_1)
-; GFX11-NEXT: s_addk_i32 s2, 0x7fff
-; GFX11-NEXT: s_cmp_lg_u32 s0, 0
-; GFX11-NEXT: s_cselect_b32 s0, s1, s2
-; GFX11-NEXT: s_lshr_b32 s0, s0, 16
+; GFX11-NEXT: s_lshr_b32 s0, s1, 16
; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT: s_and_b32 s0, s0, 0x8000
; GFX11-NEXT: ; return to shader part epilog
@@ -7563,16 +7448,14 @@ define bfloat @v_copysign_bf16_0_f64(double %sign) {
; GCN-LABEL: v_copysign_bf16_0_f64:
; GCN: ; %bb.0:
; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GCN-NEXT: v_cvt_f32_f64_e32 v0, v[0:1]
-; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v1
; GCN-NEXT: v_and_b32_e32 v0, 0x8000, v0
; GCN-NEXT: s_setpc_b64 s[30:31]
;
; GFX7-LABEL: v_copysign_bf16_0_f64:
; GFX7: ; %bb.0:
; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7-NEXT: v_cvt_f32_f64_e32 v0, v[0:1]
-; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0
+; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v1
; GFX7-NEXT: v_and_b32_e32 v0, 0x8000, v0
; GFX7-NEXT: s_setpc_b64 s[30:31]
;
diff --git a/llvm/test/CodeGen/AMDGPU/fcopysign.f16.ll b/llvm/test/CodeGen/AMDGPU/fcopysign.f16.ll
index e28368fb85838..06824fb78912b 100644
--- a/llvm/test/CodeGen/AMDGPU/fcopysign.f16.ll
+++ b/llvm/test/CodeGen/AMDGPU/fcopysign.f16.ll
@@ -6650,40 +6650,28 @@ define half @v_copysign_f16_0_f16(half %sign) {
define amdgpu_ps i32 @s_copysign_f16_0_f32(float inreg %sign) {
; SI-LABEL: s_copysign_f16_0_f32:
; SI: ; %bb.0:
-; SI-NEXT: v_cvt_f16_f32_e32 v0, s0
-; SI-NEXT: v_and_b32_e32 v0, 0x8000, v0
-; SI-NEXT: v_readfirstlane_b32 s0, v0
+; SI-NEXT: s_lshr_b32 s0, s0, 16
+; SI-NEXT: s_and_b32 s0, s0, 0x8000
; SI-NEXT: ; return to shader part epilog
;
; VI-LABEL: s_copysign_f16_0_f32:
; VI: ; %bb.0:
-; VI-NEXT: v_cvt_f16_f32_e32 v0, s0
-; VI-NEXT: v_and_b32_e32 v0, 0x8000, v0
-; VI-NEXT: v_readfirstlane_b32 s0, v0
+; VI-NEXT: s_lshr_b32 s0, s0, 16
+; VI-NEXT: s_and_b32 s0, s0, 0x8000
; VI-NEXT: ; return to shader part epilog
;
; GFX9-LABEL: s_copysign_f16_0_f32:
; GFX9: ; %bb.0:
-; GFX9-NEXT: v_cvt_f16_f32_e32 v0, s0
-; GFX9-NEXT: v_and_b32_e32 v0, 0x8000, v0
-; GFX9-NEXT: v_readfirstlane_b32 s0, v0
+; GFX9-NEXT: s_lshr_b32 s0, s0, 16
+; GFX9-NEXT: s_and_b32 s0, s0, 0x8000
; GFX9-NEXT: ; return to shader part epilog
;
-; GFX11-TRUE16-LABEL: s_copysign_f16_0_f32:
-; GFX11-TRUE16: ; %bb.0:
-; GFX11-TRUE16-NEXT: v_cvt_f16_f32_e32 v0.l, s0
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0x8000, v0
-; GFX11-TRUE16-NEXT: v_readfirstlane_b32 s0, v0
-; GFX11-TRUE16-NEXT: ; return to shader part epilog
-;
-; GFX11-FAKE16-LABEL: s_copysign_f16_0_f32:
-; GFX11-FAKE16: ; %bb.0:
-; GFX11-FAKE16-NEXT: v_cvt_f16_f32_e32 v0, s0
-; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 0x8000, v0
-; GFX11-FAKE16-NEXT: v_readfirstlane_b32 s0, v0
-; GFX11-FAKE16-NEXT: ; return to shader part epilog
+; GFX11-LABEL: s_copysign_f16_0_f32:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: s_lshr_b32 s0, s0, 16
+; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-NEXT: s_and_b32 s0, s0, 0x8000
+; GFX11-NEXT: ; return to shader part epilog
%sign.trunc = fptrunc float %sign to half
%op = call half @llvm.copysign.f16(half 0.0, half %sign.trunc)
%cast = bitcast half %op to i16
@@ -6695,7 +6683,7 @@ define half @v_copysign_f16_0_f32(float %sign) {
; SI-LABEL: v_copysign_f16_0_f32:
; SI: ; %bb.0:
; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; SI-NEXT: v_cvt_f16_f32_e32 v0, v0
+; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v0
; SI-NEXT: v_and_b32_e32 v0, 0x8000, v0
; SI-NEXT: s_setpc_b64 s[30:31]
;
@@ -6734,181 +6722,25 @@ define half @v_copysign_f16_0_f32(float %sign) {
define amdgpu_ps i32 @s_copysign_f16_0_f64(double inreg %sign) {
; SI-LABEL: s_copysign_f16_0_f64:
; SI: ; %bb.0:
-; SI-NEXT: s_and_b32 s2, s1, 0x1ff
-; SI-NEXT: s_or_b32 s0, s2, s0
-; SI-NEXT: s_cselect_b32 s0, 1, 0
-; SI-NEXT: s_lshr_b32 s2, s1, 8
-; SI-NEXT: s_bfe_u32 s3, s1, 0xb0014
-; SI-NEXT: s_and_b32 s2, s2, 0xffe
-; SI-NEXT: s_sub_i32 s4, 0x3f1, s3
-; SI-NEXT: s_or_b32 s0, s2, s0
-; SI-NEXT: v_med3_i32 v0, s4, 0, 13
-; SI-NEXT: s_or_b32 s2, s0, 0x1000
-; SI-NEXT: v_readfirstlane_b32 s4, v0
-; SI-NEXT: s_lshr_b32 s5, s2, s4
-; SI-NEXT: s_lshl_b32 s4, s5, s4
-; SI-NEXT: s_cmp_lg_u32 s4, s2
-; SI-NEXT: s_cselect_b32 s2, 1, 0
-; SI-NEXT: s_add_i32 s4, s3, 0xfffffc10
-; SI-NEXT: s_lshl_b32 s3, s4, 12
-; SI-NEXT: s_or_b32 s2, s5, s2
-; SI-NEXT: s_or_b32 s0, s0, s3
-; SI-NEXT: s_cmp_lt_i32 s4, 1
-; SI-NEXT: s_cselect_b32 s0, s2, s0
-; SI-NEXT: s_and_b32 s2, s0, 7
-; SI-NEXT: s_cmp_gt_i32 s2, 5
-; SI-NEXT: s_cselect_b32 s3, 1, 0
-; SI-NEXT: s_cmp_eq_u32 s2, 3
-; SI-NEXT: s_cselect_b32 s2, 1, 0
-; SI-NEXT: s_or_b32 s2, s2, s3
-; SI-NEXT: s_lshr_b32 s0, s0, 2
-; SI-NEXT: s_add_i32 s0, s0, s2
-; SI-NEXT: s_cmp_lt_i32 s4, 31
-; SI-NEXT: s_cselect_b64 s[2:3], -1, 0
-; SI-NEXT: s_cmpk_lg_i32 s4, 0x40f
-; SI-NEXT: s_cselect_b64 s[4:5], -1, 0
-; SI-NEXT: s_and_b64 s[2:3], s[4:5], s[2:3]
-; SI-NEXT: s_and_b64 s[2:3], s[2:3], exec
-; SI-NEXT: s_cselect_b32 s0, s0, 0x7c00
-; SI-NEXT: s_lshr_b32 s1, s1, 16
-; SI-NEXT: s_or_b32 s0, s1, s0
+; SI-NEXT: s_lshr_b32 s0, s1, 16
; SI-NEXT: s_and_b32 s0, s0, 0x8000
; SI-NEXT: ; return to shader part epilog
;
; VI-LABEL: s_copysign_f16_0_f64:
; VI: ; %bb.0:
-; VI-NEXT: s_and_b32 s2, s1, 0x1ff
-; VI-NEXT: s_or_b32 s0, s2, s0
-; VI-NEXT: s_cselect_b32 s0, 1, 0
-; VI-NEXT: s_lshr_b32 s2, s1, 8
-; VI-NEXT: s_bfe_u32 s3, s1, 0xb0014
-; VI-NEXT: s_and_b32 s2, s2, 0xffe
-; VI-NEXT: s_sub_i32 s4, 0x3f1, s3
-; VI-NEXT: s_or_b32 s0, s2, s0
-; VI-NEXT: v_med3_i32 v0, s4, 0, 13
-; VI-NEXT: s_or_b32 s2, s0, 0x1000
-; VI-NEXT: v_readfirstlane_b32 s4, v0
-; VI-NEXT: s_lshr_b32 s5, s2, s4
-; VI-NEXT: s_lshl_b32 s4, s5, s4
-; VI-NEXT: s_cmp_lg_u32 s4, s2
-; VI-NEXT: s_cselect_b32 s2, 1, 0
-; VI-NEXT: s_add_i32 s4, s3, 0xfffffc10
-; VI-NEXT: s_lshl_b32 s3, s4, 12
-; VI-NEXT: s_or_b32 s2, s5, s2
-; VI-NEXT: s_or_b32 s0, s0, s3
-; VI-NEXT: s_cmp_lt_i32 s4, 1
-; VI-NEXT: s_cselect_b32 s0, s2, s0
-; VI-NEXT: s_and_b32 s2, s0, 7
-; VI-NEXT: s_cmp_gt_i32 s2, 5
-; VI-NEXT: s_cselect_b32 s3, 1, 0
-; VI-NEXT: s_cmp_eq_u32 s2, 3
-; VI-NEXT: s_cselect_b32 s2, 1, 0
-; VI-NEXT: s_or_b32 s2, s2, s3
-; VI-NEXT: s_lshr_b32 s0, s0, 2
-; VI-NEXT: s_add_i32 s0, s0, s2
-; VI-NEXT: s_cmp_lt_i32 s4, 31
-; VI-NEXT: s_cselect_b64 s[2:3], -1, 0
-; VI-NEXT: s_cmpk_lg_i32 s4, 0x40f
-; VI-NEXT: s_cselect_b64 s[4:5], -1, 0
-; VI-NEXT: s_and_b64 s[2:3], s[4:5], s[2:3]
-; VI-NEXT: s_and_b64 s[2:3], s[2:3], exec
-; VI-NEXT: s_cselect_b32 s0, s0, 0x7c00
-; VI-NEXT: s_lshr_b32 s1, s1, 16
-; VI-NEXT: s_or_b32 s0, s1, s0
+; VI-NEXT: s_lshr_b32 s0, s1, 16
; VI-NEXT: s_and_b32 s0, s0, 0x8000
; VI-NEXT: ; return to shader part epilog
;
; GFX9-LABEL: s_copysign_f16_0_f64:
; GFX9: ; %bb.0:
-; GFX9-NEXT: s_and_b32 s2, s1, 0x1ff
-; GFX9-NEXT: s_or_b32 s0, s2, s0
-; GFX9-NEXT: s_cselect_b32 s0, 1, 0
-; GFX9-NEXT: s_lshr_b32 s2, s1, 8
-; GFX9-NEXT: s_bfe_u32 s3, s1, 0xb0014
-; GFX9-NEXT: s_and_b32 s2, s2, 0xffe
-; GFX9-NEXT: s_sub_i32 s4, 0x3f1, s3
-; GFX9-NEXT: s_or_b32 s0, s2, s0
-; GFX9-NEXT: v_med3_i32 v0, s4, 0, 13
-; GFX...
[truncated]
``````````
</details>
https://github.com/llvm/llvm-project/pull/242587
More information about the llvm-branch-commits
mailing list