Skip to content

[DAGCombiner] Remove NoSignedZerosFPMath in visitFNEG #162052

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom
Open
wants to merge 4 commits into
base: main
Choose a base branch
from
Open

[DAGCombiner] Remove NoSignedZerosFPMath in visitFNEG #162052

wants to merge 4 commits into from
+1,405 −2,702

Conversation

paperchalice
Copy link
Contributor

@paperchalice paperchalice commented Oct 6, 2025 •
edited
Loading

Remove the NoSignedZerosFPMath use in visitFNEG. Now the only use of NoSignedZerosFPMath is in foldFPToIntToFP, but adding fast-math flags support for uitofp may introduce breaking changes.

@paperchalice paperchalice marked this pull request as ready for review October 7, 2025 05:32
@llvmbot llvmbot added backend:AMDGPU llvm:SelectionDAG SelectionDAGISel as well labels Oct 7, 2025
@llvmbot
Copy link
Member

llvmbot commented Oct 7, 2025

@llvm/pr-subscribers-backend-amdgpu

Author: None (paperchalice)

Changes

Remove the NoSignedZerosFPMath use in visitFNEG. Now the only use of NoSignedZerosFPMath is in foldFPToIntToFP, but adding fast-math flags support for uitofp may introduce breaking changes.

Full diff: https://github.com/llvm/llvm-project/pull/162052.diff

3 Files Affected:

  • (modified) llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp (+2-3)
  • (modified) llvm/test/CodeGen/AMDGPU/fsub.ll (+1-4)
  • (modified) llvm/test/CodeGen/AMDGPU/select-fabs-fneg-extract.v2f16.ll (+123-37)
diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 309f1bea8b77c..c5c38661f1d71 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -19319,9 +19319,8 @@ SDValue DAGCombiner::visitFNEG(SDNode *N) {
   // FIXME: This is duplicated in getNegatibleCost, but getNegatibleCost doesn't
   // know it was called from a context with a nsz flag if the input fsub does
   // not.
-  if (N0.getOpcode() == ISD::FSUB &&
-      (DAG.getTarget().Options.NoSignedZerosFPMath ||
-       N->getFlags().hasNoSignedZeros()) && N0.hasOneUse()) {
+  if (N0.getOpcode() == ISD::FSUB && N->getFlags().hasNoSignedZeros() &&
+      N0.hasOneUse()) {
     return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N0.getOperand(1),
                        N0.getOperand(0));
   }
diff --git a/llvm/test/CodeGen/AMDGPU/fsub.ll b/llvm/test/CodeGen/AMDGPU/fsub.ll
index 743431c7e0e67..984690fd0ea16 100644
--- a/llvm/test/CodeGen/AMDGPU/fsub.ll
+++ b/llvm/test/CodeGen/AMDGPU/fsub.ll
@@ -100,7 +100,7 @@ define amdgpu_kernel void @v_fneg_fsub_nsz_attribute_f32(ptr addrspace(1) %out,
   %a = load float, ptr addrspace(1) %in, align 4
   %b = load float, ptr addrspace(1) %b_ptr, align 4
   %result = fsub float %a, %b
-  %neg.result = fsub float -0.0, %result
+  %neg.result = fsub nsz float -0.0, %result
   store float %neg.result, ptr addrspace(1) %out, align 4
   ret void
 }
@@ -129,6 +129,3 @@ define amdgpu_kernel void @v_fsub_0_nsz_attribute_f32(ptr addrspace(1) %out, ptr
   store float %result, ptr addrspace(1) %out, align 4
   ret void
 }
-
-attributes #0 = { nounwind "no-signed-zeros-fp-math"="true" }
-attributes #1 = { nounwind "no-signed-zeros-fp-math"="false" }
diff --git a/llvm/test/CodeGen/AMDGPU/select-fabs-fneg-extract.v2f16.ll b/llvm/test/CodeGen/AMDGPU/select-fabs-fneg-extract.v2f16.ll
index bb22144b815a1..eb69a53a72c90 100644
--- a/llvm/test/CodeGen/AMDGPU/select-fabs-fneg-extract.v2f16.ll
+++ b/llvm/test/CodeGen/AMDGPU/select-fabs-fneg-extract.v2f16.ll
@@ -4387,28 +4387,28 @@ define <2 x half> @select_fneg_posk_src_add_v2f16_nsz(<2 x i32> %c, <2 x half> %
 }
 
 define <2 x half> @select_fneg_posk_src_sub_v2f16(<2 x i32> %c, <2 x half> %x) {
-; CI-SAFE-LABEL: select_fneg_posk_src_sub_v2f16:
-; CI-SAFE:       ; %bb.0:
-; CI-SAFE-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; CI-SAFE-NEXT:    v_cvt_f16_f32_e32 v3, v3
-; CI-SAFE-NEXT:    v_cvt_f16_f32_e32 v2, v2
-; CI-SAFE-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v0
-; CI-SAFE-NEXT:    v_cvt_f32_f16_e32 v3, v3
-; CI-SAFE-NEXT:    v_cvt_f32_f16_e32 v2, v2
-; CI-SAFE-NEXT:    v_add_f32_e32 v3, -4.0, v3
-; CI-SAFE-NEXT:    v_add_f32_e32 v2, -4.0, v2
-; CI-SAFE-NEXT:    v_cvt_f16_f32_e32 v3, v3
-; CI-SAFE-NEXT:    v_cvt_f16_f32_e32 v2, v2
-; CI-SAFE-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; CI-SAFE-NEXT:    v_or_b32_e32 v2, v2, v3
-; CI-SAFE-NEXT:    v_xor_b32_e32 v2, 0x80008000, v2
-; CI-SAFE-NEXT:    v_cvt_f32_f16_e32 v3, v2
-; CI-SAFE-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
-; CI-SAFE-NEXT:    v_cvt_f32_f16_e32 v2, v2
-; CI-SAFE-NEXT:    v_cndmask_b32_e32 v0, 2.0, v3, vcc
-; CI-SAFE-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v1
-; CI-SAFE-NEXT:    v_cndmask_b32_e32 v1, 2.0, v2, vcc
-; CI-SAFE-NEXT:    s_setpc_b64 s[30:31]
+; CI-LABEL: select_fneg_posk_src_sub_v2f16:
+; CI:       ; %bb.0:
+; CI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CI-NEXT:    v_cvt_f16_f32_e32 v3, v3
+; CI-NEXT:    v_cvt_f16_f32_e32 v2, v2
+; CI-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v0
+; CI-NEXT:    v_cvt_f32_f16_e32 v3, v3
+; CI-NEXT:    v_cvt_f32_f16_e32 v2, v2
+; CI-NEXT:    v_add_f32_e32 v3, -4.0, v3
+; CI-NEXT:    v_add_f32_e32 v2, -4.0, v2
+; CI-NEXT:    v_cvt_f16_f32_e32 v3, v3
+; CI-NEXT:    v_cvt_f16_f32_e32 v2, v2
+; CI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; CI-NEXT:    v_or_b32_e32 v2, v2, v3
+; CI-NEXT:    v_xor_b32_e32 v2, 0x80008000, v2
+; CI-NEXT:    v_cvt_f32_f16_e32 v3, v2
+; CI-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
+; CI-NEXT:    v_cvt_f32_f16_e32 v2, v2
+; CI-NEXT:    v_cndmask_b32_e32 v0, 2.0, v3, vcc
+; CI-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v1
+; CI-NEXT:    v_cndmask_b32_e32 v1, 2.0, v2, vcc
+; CI-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; VI-SAFE-LABEL: select_fneg_posk_src_sub_v2f16:
 ; VI-SAFE:       ; %bb.0:
@@ -4468,21 +4468,6 @@ define <2 x half> @select_fneg_posk_src_sub_v2f16(<2 x i32> %c, <2 x half> %x) {
 ; GFX11-SAFE-FAKE16-NEXT:    v_perm_b32 v0, v1, v0, 0x5040100
 ; GFX11-SAFE-FAKE16-NEXT:    s_setpc_b64 s[30:31]
 ;
-; CI-NSZ-LABEL: select_fneg_posk_src_sub_v2f16:
-; CI-NSZ:       ; %bb.0:
-; CI-NSZ-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; CI-NSZ-NEXT:    v_cvt_f16_f32_e32 v2, v2
-; CI-NSZ-NEXT:    v_cvt_f16_f32_e32 v3, v3
-; CI-NSZ-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v0
-; CI-NSZ-NEXT:    v_cvt_f32_f16_e32 v2, v2
-; CI-NSZ-NEXT:    v_cvt_f32_f16_e32 v3, v3
-; CI-NSZ-NEXT:    v_sub_f32_e32 v2, 4.0, v2
-; CI-NSZ-NEXT:    v_sub_f32_e32 v3, 4.0, v3
-; CI-NSZ-NEXT:    v_cndmask_b32_e32 v0, 2.0, v2, vcc
-; CI-NSZ-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v1
-; CI-NSZ-NEXT:    v_cndmask_b32_e32 v1, 2.0, v3, vcc
-; CI-NSZ-NEXT:    s_setpc_b64 s[30:31]
-;
 ; VI-NSZ-LABEL: select_fneg_posk_src_sub_v2f16:
 ; VI-NSZ:       ; %bb.0:
 ; VI-NSZ-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
@@ -4541,6 +4526,105 @@ define <2 x half> @select_fneg_posk_src_sub_v2f16(<2 x i32> %c, <2 x half> %x) {
   ret <2 x half> %select
 }
 
+define <2 x half> @select_fneg_posk_src_sub_v2f16_nsz(<2 x i32> %c, <2 x half> %x) {
+; CI-LABEL: select_fneg_posk_src_sub_v2f16_nsz:
+; CI:       ; %bb.0:
+; CI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CI-NEXT:    v_cvt_f16_f32_e32 v2, v2
+; CI-NEXT:    v_cvt_f16_f32_e32 v3, v3
+; CI-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v0
+; CI-NEXT:    v_cvt_f32_f16_e32 v2, v2
+; CI-NEXT:    v_cvt_f32_f16_e32 v3, v3
+; CI-NEXT:    v_sub_f32_e32 v2, 4.0, v2
+; CI-NEXT:    v_sub_f32_e32 v3, 4.0, v3
+; CI-NEXT:    v_cndmask_b32_e32 v0, 2.0, v2, vcc
+; CI-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v1
+; CI-NEXT:    v_cndmask_b32_e32 v1, 2.0, v3, vcc
+; CI-NEXT:    s_setpc_b64 s[30:31]
+;
+; VI-LABEL: select_fneg_posk_src_sub_v2f16_nsz:
+; VI:       ; %bb.0:
+; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v1
+; VI-NEXT:    v_mov_b32_e32 v1, 0x4400
+; VI-NEXT:    v_sub_f16_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; VI-NEXT:    v_sub_f16_e32 v2, 4.0, v2
+; VI-NEXT:    v_mov_b32_e32 v3, 0x4000
+; VI-NEXT:    v_cmp_eq_u32_e64 s[4:5], 0, v0
+; VI-NEXT:    v_cndmask_b32_e64 v0, v3, v2, s[4:5]
+; VI-NEXT:    v_cndmask_b32_sdwa v1, v3, v1, vcc dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: select_fneg_posk_src_sub_v2f16_nsz:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v1
+; GFX9-NEXT:    v_pk_add_f16 v1, v2, 4.0 op_sel_hi:[1,0] neg_lo:[1,0] neg_hi:[1,0]
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0x4000
+; GFX9-NEXT:    v_cmp_eq_u32_e64 s[4:5], 0, v0
+; GFX9-NEXT:    v_cndmask_b32_e64 v0, v2, v1, s[4:5]
+; GFX9-NEXT:    v_cndmask_b32_sdwa v1, v2, v1, vcc dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX9-NEXT:    s_mov_b32 s4, 0x5040100
+; GFX9-NEXT:    v_perm_b32 v0, v1, v0, s4
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX11-SAFE-TRUE16-LABEL: select_fneg_posk_src_sub_v2f16_nsz:
+; GFX11-SAFE-TRUE16:       ; %bb.0:
+; GFX11-SAFE-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-SAFE-TRUE16-NEXT:    v_cmp_eq_u32_e32 vcc_lo, 0, v0
+; GFX11-SAFE-TRUE16-NEXT:    v_pk_add_f16 v0, v2, 4.0 op_sel_hi:[1,0] neg_lo:[1,0] neg_hi:[1,0]
+; GFX11-SAFE-TRUE16-NEXT:    v_cmp_eq_u32_e64 s0, 0, v1
+; GFX11-SAFE-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-SAFE-TRUE16-NEXT:    v_cndmask_b16 v0.l, 0x4000, v0.l, vcc_lo
+; GFX11-SAFE-TRUE16-NEXT:    v_cndmask_b16 v0.h, 0x4000, v0.h, s0
+; GFX11-SAFE-TRUE16-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX11-SAFE-FAKE16-LABEL: select_fneg_posk_src_sub_v2f16_nsz:
+; GFX11-SAFE-FAKE16:       ; %bb.0:
+; GFX11-SAFE-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-SAFE-FAKE16-NEXT:    v_pk_add_f16 v2, v2, 4.0 op_sel_hi:[1,0] neg_lo:[1,0] neg_hi:[1,0]
+; GFX11-SAFE-FAKE16-NEXT:    v_cmp_eq_u32_e32 vcc_lo, 0, v0
+; GFX11-SAFE-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-SAFE-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v2
+; GFX11-SAFE-FAKE16-NEXT:    v_cndmask_b32_e32 v0, 0x4000, v2, vcc_lo
+; GFX11-SAFE-FAKE16-NEXT:    v_cmp_eq_u32_e32 vcc_lo, 0, v1
+; GFX11-SAFE-FAKE16-NEXT:    v_cndmask_b32_e32 v1, 0x4000, v3, vcc_lo
+; GFX11-SAFE-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-SAFE-FAKE16-NEXT:    v_perm_b32 v0, v1, v0, 0x5040100
+; GFX11-SAFE-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX11-NSZ-TRUE16-LABEL: select_fneg_posk_src_sub_v2f16_nsz:
+; GFX11-NSZ-TRUE16:       ; %bb.0:
+; GFX11-NSZ-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NSZ-TRUE16-NEXT:    v_cmp_eq_u32_e32 vcc_lo, 0, v0
+; GFX11-NSZ-TRUE16-NEXT:    v_pk_add_f16 v0, v2, 4.0 op_sel_hi:[1,0] neg_lo:[1,0] neg_hi:[1,0]
+; GFX11-NSZ-TRUE16-NEXT:    v_cmp_eq_u32_e64 s0, 0, v1
+; GFX11-NSZ-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NSZ-TRUE16-NEXT:    v_cndmask_b16 v0.l, 0x4000, v0.l, vcc_lo
+; GFX11-NSZ-TRUE16-NEXT:    v_cndmask_b16 v0.h, 0x4000, v0.h, s0
+; GFX11-NSZ-TRUE16-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX11-NSZ-FAKE16-LABEL: select_fneg_posk_src_sub_v2f16_nsz:
+; GFX11-NSZ-FAKE16:       ; %bb.0:
+; GFX11-NSZ-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NSZ-FAKE16-NEXT:    v_pk_add_f16 v2, v2, 4.0 op_sel_hi:[1,0] neg_lo:[1,0] neg_hi:[1,0]
+; GFX11-NSZ-FAKE16-NEXT:    v_cmp_eq_u32_e32 vcc_lo, 0, v0
+; GFX11-NSZ-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NSZ-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v2
+; GFX11-NSZ-FAKE16-NEXT:    v_cndmask_b32_e32 v0, 0x4000, v2, vcc_lo
+; GFX11-NSZ-FAKE16-NEXT:    v_cmp_eq_u32_e32 vcc_lo, 0, v1
+; GFX11-NSZ-FAKE16-NEXT:    v_cndmask_b32_e32 v1, 0x4000, v3, vcc_lo
+; GFX11-NSZ-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NSZ-FAKE16-NEXT:    v_perm_b32 v0, v1, v0, 0x5040100
+; GFX11-NSZ-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+  %cmp = icmp eq <2 x i32> %c, zeroinitializer
+  %add = fsub <2 x half> %x, <half 4.0, half 4.0>
+  %fneg = fneg nsz <2 x half> %add
+  %select = select <2 x i1> %cmp, <2 x half> %fneg, <2 x half> <half 2.0, half 2.0>
+  ret <2 x half> %select
+}
+
 define <2 x half> @select_fneg_posk_src_mul_v2f16(<2 x i32> %c, <2 x half> %x) {
 ; CI-LABEL: select_fneg_posk_src_mul_v2f16:
 ; CI:       ; %bb.0:
@@ -5048,6 +5132,8 @@ declare <2 x half> @llvm.fmuladd.v2f16(<2 x half>, <2 x half>, <2 x half>) #0
 
 attributes #0 = { nocallback nofree nosync nounwind speculatable willreturn memory(none) }
 ;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
+; CI-NSZ: {{.*}}
+; CI-SAFE: {{.*}}
 ; GFX11: {{.*}}
 ; GFX11-NSZ: {{.*}}
 ; GFX11-SAFE: {{.*}}

@llvmbot
Copy link
Member

llvmbot commented Oct 7, 2025

@llvm/pr-subscribers-llvm-selectiondag

Author: None (paperchalice)

Changes

Remove the NoSignedZerosFPMath use in visitFNEG. Now the only use of NoSignedZerosFPMath is in foldFPToIntToFP, but adding fast-math flags support for uitofp may introduce breaking changes.

Full diff: https://github.com/llvm/llvm-project/pull/162052.diff

3 Files Affected:

  • (modified) llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp (+2-3)
  • (modified) llvm/test/CodeGen/AMDGPU/fsub.ll (+1-4)
  • (modified) llvm/test/CodeGen/AMDGPU/select-fabs-fneg-extract.v2f16.ll (+123-37)
diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 309f1bea8b77c..c5c38661f1d71 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -19319,9 +19319,8 @@ SDValue DAGCombiner::visitFNEG(SDNode *N) {
   // FIXME: This is duplicated in getNegatibleCost, but getNegatibleCost doesn't
   // know it was called from a context with a nsz flag if the input fsub does
   // not.
-  if (N0.getOpcode() == ISD::FSUB &&
-      (DAG.getTarget().Options.NoSignedZerosFPMath ||
-       N->getFlags().hasNoSignedZeros()) && N0.hasOneUse()) {
+  if (N0.getOpcode() == ISD::FSUB && N->getFlags().hasNoSignedZeros() &&
+      N0.hasOneUse()) {
     return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N0.getOperand(1),
                        N0.getOperand(0));
   }
diff --git a/llvm/test/CodeGen/AMDGPU/fsub.ll b/llvm/test/CodeGen/AMDGPU/fsub.ll
index 743431c7e0e67..984690fd0ea16 100644
--- a/llvm/test/CodeGen/AMDGPU/fsub.ll
+++ b/llvm/test/CodeGen/AMDGPU/fsub.ll
@@ -100,7 +100,7 @@ define amdgpu_kernel void @v_fneg_fsub_nsz_attribute_f32(ptr addrspace(1) %out,
   %a = load float, ptr addrspace(1) %in, align 4
   %b = load float, ptr addrspace(1) %b_ptr, align 4
   %result = fsub float %a, %b
-  %neg.result = fsub float -0.0, %result
+  %neg.result = fsub nsz float -0.0, %result
   store float %neg.result, ptr addrspace(1) %out, align 4
   ret void
 }
@@ -129,6 +129,3 @@ define amdgpu_kernel void @v_fsub_0_nsz_attribute_f32(ptr addrspace(1) %out, ptr
   store float %result, ptr addrspace(1) %out, align 4
   ret void
 }
-
-attributes #0 = { nounwind "no-signed-zeros-fp-math"="true" }
-attributes #1 = { nounwind "no-signed-zeros-fp-math"="false" }
diff --git a/llvm/test/CodeGen/AMDGPU/select-fabs-fneg-extract.v2f16.ll b/llvm/test/CodeGen/AMDGPU/select-fabs-fneg-extract.v2f16.ll
index bb22144b815a1..eb69a53a72c90 100644
--- a/llvm/test/CodeGen/AMDGPU/select-fabs-fneg-extract.v2f16.ll
+++ b/llvm/test/CodeGen/AMDGPU/select-fabs-fneg-extract.v2f16.ll
@@ -4387,28 +4387,28 @@ define <2 x half> @select_fneg_posk_src_add_v2f16_nsz(<2 x i32> %c, <2 x half> %
 }
 
 define <2 x half> @select_fneg_posk_src_sub_v2f16(<2 x i32> %c, <2 x half> %x) {
-; CI-SAFE-LABEL: select_fneg_posk_src_sub_v2f16:
-; CI-SAFE:       ; %bb.0:
-; CI-SAFE-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; CI-SAFE-NEXT:    v_cvt_f16_f32_e32 v3, v3
-; CI-SAFE-NEXT:    v_cvt_f16_f32_e32 v2, v2
-; CI-SAFE-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v0
-; CI-SAFE-NEXT:    v_cvt_f32_f16_e32 v3, v3
-; CI-SAFE-NEXT:    v_cvt_f32_f16_e32 v2, v2
-; CI-SAFE-NEXT:    v_add_f32_e32 v3, -4.0, v3
-; CI-SAFE-NEXT:    v_add_f32_e32 v2, -4.0, v2
-; CI-SAFE-NEXT:    v_cvt_f16_f32_e32 v3, v3
-; CI-SAFE-NEXT:    v_cvt_f16_f32_e32 v2, v2
-; CI-SAFE-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; CI-SAFE-NEXT:    v_or_b32_e32 v2, v2, v3
-; CI-SAFE-NEXT:    v_xor_b32_e32 v2, 0x80008000, v2
-; CI-SAFE-NEXT:    v_cvt_f32_f16_e32 v3, v2
-; CI-SAFE-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
-; CI-SAFE-NEXT:    v_cvt_f32_f16_e32 v2, v2
-; CI-SAFE-NEXT:    v_cndmask_b32_e32 v0, 2.0, v3, vcc
-; CI-SAFE-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v1
-; CI-SAFE-NEXT:    v_cndmask_b32_e32 v1, 2.0, v2, vcc
-; CI-SAFE-NEXT:    s_setpc_b64 s[30:31]
+; CI-LABEL: select_fneg_posk_src_sub_v2f16:
+; CI:       ; %bb.0:
+; CI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CI-NEXT:    v_cvt_f16_f32_e32 v3, v3
+; CI-NEXT:    v_cvt_f16_f32_e32 v2, v2
+; CI-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v0
+; CI-NEXT:    v_cvt_f32_f16_e32 v3, v3
+; CI-NEXT:    v_cvt_f32_f16_e32 v2, v2
+; CI-NEXT:    v_add_f32_e32 v3, -4.0, v3
+; CI-NEXT:    v_add_f32_e32 v2, -4.0, v2
+; CI-NEXT:    v_cvt_f16_f32_e32 v3, v3
+; CI-NEXT:    v_cvt_f16_f32_e32 v2, v2
+; CI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
+; CI-NEXT:    v_or_b32_e32 v2, v2, v3
+; CI-NEXT:    v_xor_b32_e32 v2, 0x80008000, v2
+; CI-NEXT:    v_cvt_f32_f16_e32 v3, v2
+; CI-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
+; CI-NEXT:    v_cvt_f32_f16_e32 v2, v2
+; CI-NEXT:    v_cndmask_b32_e32 v0, 2.0, v3, vcc
+; CI-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v1
+; CI-NEXT:    v_cndmask_b32_e32 v1, 2.0, v2, vcc
+; CI-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; VI-SAFE-LABEL: select_fneg_posk_src_sub_v2f16:
 ; VI-SAFE:       ; %bb.0:
@@ -4468,21 +4468,6 @@ define <2 x half> @select_fneg_posk_src_sub_v2f16(<2 x i32> %c, <2 x half> %x) {
 ; GFX11-SAFE-FAKE16-NEXT:    v_perm_b32 v0, v1, v0, 0x5040100
 ; GFX11-SAFE-FAKE16-NEXT:    s_setpc_b64 s[30:31]
 ;
-; CI-NSZ-LABEL: select_fneg_posk_src_sub_v2f16:
-; CI-NSZ:       ; %bb.0:
-; CI-NSZ-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; CI-NSZ-NEXT:    v_cvt_f16_f32_e32 v2, v2
-; CI-NSZ-NEXT:    v_cvt_f16_f32_e32 v3, v3
-; CI-NSZ-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v0
-; CI-NSZ-NEXT:    v_cvt_f32_f16_e32 v2, v2
-; CI-NSZ-NEXT:    v_cvt_f32_f16_e32 v3, v3
-; CI-NSZ-NEXT:    v_sub_f32_e32 v2, 4.0, v2
-; CI-NSZ-NEXT:    v_sub_f32_e32 v3, 4.0, v3
-; CI-NSZ-NEXT:    v_cndmask_b32_e32 v0, 2.0, v2, vcc
-; CI-NSZ-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v1
-; CI-NSZ-NEXT:    v_cndmask_b32_e32 v1, 2.0, v3, vcc
-; CI-NSZ-NEXT:    s_setpc_b64 s[30:31]
-;
 ; VI-NSZ-LABEL: select_fneg_posk_src_sub_v2f16:
 ; VI-NSZ:       ; %bb.0:
 ; VI-NSZ-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
@@ -4541,6 +4526,105 @@ define <2 x half> @select_fneg_posk_src_sub_v2f16(<2 x i32> %c, <2 x half> %x) {
   ret <2 x half> %select
 }
 
+define <2 x half> @select_fneg_posk_src_sub_v2f16_nsz(<2 x i32> %c, <2 x half> %x) {
+; CI-LABEL: select_fneg_posk_src_sub_v2f16_nsz:
+; CI:       ; %bb.0:
+; CI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CI-NEXT:    v_cvt_f16_f32_e32 v2, v2
+; CI-NEXT:    v_cvt_f16_f32_e32 v3, v3
+; CI-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v0
+; CI-NEXT:    v_cvt_f32_f16_e32 v2, v2
+; CI-NEXT:    v_cvt_f32_f16_e32 v3, v3
+; CI-NEXT:    v_sub_f32_e32 v2, 4.0, v2
+; CI-NEXT:    v_sub_f32_e32 v3, 4.0, v3
+; CI-NEXT:    v_cndmask_b32_e32 v0, 2.0, v2, vcc
+; CI-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v1
+; CI-NEXT:    v_cndmask_b32_e32 v1, 2.0, v3, vcc
+; CI-NEXT:    s_setpc_b64 s[30:31]
+;
+; VI-LABEL: select_fneg_posk_src_sub_v2f16_nsz:
+; VI:       ; %bb.0:
+; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v1
+; VI-NEXT:    v_mov_b32_e32 v1, 0x4400
+; VI-NEXT:    v_sub_f16_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; VI-NEXT:    v_sub_f16_e32 v2, 4.0, v2
+; VI-NEXT:    v_mov_b32_e32 v3, 0x4000
+; VI-NEXT:    v_cmp_eq_u32_e64 s[4:5], 0, v0
+; VI-NEXT:    v_cndmask_b32_e64 v0, v3, v2, s[4:5]
+; VI-NEXT:    v_cndmask_b32_sdwa v1, v3, v1, vcc dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
+; VI-NEXT:    v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; VI-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: select_fneg_posk_src_sub_v2f16_nsz:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v1
+; GFX9-NEXT:    v_pk_add_f16 v1, v2, 4.0 op_sel_hi:[1,0] neg_lo:[1,0] neg_hi:[1,0]
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0x4000
+; GFX9-NEXT:    v_cmp_eq_u32_e64 s[4:5], 0, v0
+; GFX9-NEXT:    v_cndmask_b32_e64 v0, v2, v1, s[4:5]
+; GFX9-NEXT:    v_cndmask_b32_sdwa v1, v2, v1, vcc dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; GFX9-NEXT:    s_mov_b32 s4, 0x5040100
+; GFX9-NEXT:    v_perm_b32 v0, v1, v0, s4
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX11-SAFE-TRUE16-LABEL: select_fneg_posk_src_sub_v2f16_nsz:
+; GFX11-SAFE-TRUE16:       ; %bb.0:
+; GFX11-SAFE-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-SAFE-TRUE16-NEXT:    v_cmp_eq_u32_e32 vcc_lo, 0, v0
+; GFX11-SAFE-TRUE16-NEXT:    v_pk_add_f16 v0, v2, 4.0 op_sel_hi:[1,0] neg_lo:[1,0] neg_hi:[1,0]
+; GFX11-SAFE-TRUE16-NEXT:    v_cmp_eq_u32_e64 s0, 0, v1
+; GFX11-SAFE-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-SAFE-TRUE16-NEXT:    v_cndmask_b16 v0.l, 0x4000, v0.l, vcc_lo
+; GFX11-SAFE-TRUE16-NEXT:    v_cndmask_b16 v0.h, 0x4000, v0.h, s0
+; GFX11-SAFE-TRUE16-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX11-SAFE-FAKE16-LABEL: select_fneg_posk_src_sub_v2f16_nsz:
+; GFX11-SAFE-FAKE16:       ; %bb.0:
+; GFX11-SAFE-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-SAFE-FAKE16-NEXT:    v_pk_add_f16 v2, v2, 4.0 op_sel_hi:[1,0] neg_lo:[1,0] neg_hi:[1,0]
+; GFX11-SAFE-FAKE16-NEXT:    v_cmp_eq_u32_e32 vcc_lo, 0, v0
+; GFX11-SAFE-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-SAFE-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v2
+; GFX11-SAFE-FAKE16-NEXT:    v_cndmask_b32_e32 v0, 0x4000, v2, vcc_lo
+; GFX11-SAFE-FAKE16-NEXT:    v_cmp_eq_u32_e32 vcc_lo, 0, v1
+; GFX11-SAFE-FAKE16-NEXT:    v_cndmask_b32_e32 v1, 0x4000, v3, vcc_lo
+; GFX11-SAFE-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-SAFE-FAKE16-NEXT:    v_perm_b32 v0, v1, v0, 0x5040100
+; GFX11-SAFE-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX11-NSZ-TRUE16-LABEL: select_fneg_posk_src_sub_v2f16_nsz:
+; GFX11-NSZ-TRUE16:       ; %bb.0:
+; GFX11-NSZ-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NSZ-TRUE16-NEXT:    v_cmp_eq_u32_e32 vcc_lo, 0, v0
+; GFX11-NSZ-TRUE16-NEXT:    v_pk_add_f16 v0, v2, 4.0 op_sel_hi:[1,0] neg_lo:[1,0] neg_hi:[1,0]
+; GFX11-NSZ-TRUE16-NEXT:    v_cmp_eq_u32_e64 s0, 0, v1
+; GFX11-NSZ-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NSZ-TRUE16-NEXT:    v_cndmask_b16 v0.l, 0x4000, v0.l, vcc_lo
+; GFX11-NSZ-TRUE16-NEXT:    v_cndmask_b16 v0.h, 0x4000, v0.h, s0
+; GFX11-NSZ-TRUE16-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX11-NSZ-FAKE16-LABEL: select_fneg_posk_src_sub_v2f16_nsz:
+; GFX11-NSZ-FAKE16:       ; %bb.0:
+; GFX11-NSZ-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-NSZ-FAKE16-NEXT:    v_pk_add_f16 v2, v2, 4.0 op_sel_hi:[1,0] neg_lo:[1,0] neg_hi:[1,0]
+; GFX11-NSZ-FAKE16-NEXT:    v_cmp_eq_u32_e32 vcc_lo, 0, v0
+; GFX11-NSZ-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-NSZ-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v2
+; GFX11-NSZ-FAKE16-NEXT:    v_cndmask_b32_e32 v0, 0x4000, v2, vcc_lo
+; GFX11-NSZ-FAKE16-NEXT:    v_cmp_eq_u32_e32 vcc_lo, 0, v1
+; GFX11-NSZ-FAKE16-NEXT:    v_cndmask_b32_e32 v1, 0x4000, v3, vcc_lo
+; GFX11-NSZ-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NSZ-FAKE16-NEXT:    v_perm_b32 v0, v1, v0, 0x5040100
+; GFX11-NSZ-FAKE16-NEXT:    s_setpc_b64 s[30:31]
+  %cmp = icmp eq <2 x i32> %c, zeroinitializer
+  %add = fsub <2 x half> %x, <half 4.0, half 4.0>
+  %fneg = fneg nsz <2 x half> %add
+  %select = select <2 x i1> %cmp, <2 x half> %fneg, <2 x half> <half 2.0, half 2.0>
+  ret <2 x half> %select
+}
+
 define <2 x half> @select_fneg_posk_src_mul_v2f16(<2 x i32> %c, <2 x half> %x) {
 ; CI-LABEL: select_fneg_posk_src_mul_v2f16:
 ; CI:       ; %bb.0:
@@ -5048,6 +5132,8 @@ declare <2 x half> @llvm.fmuladd.v2f16(<2 x half>, <2 x half>, <2 x half>) #0
 
 attributes #0 = { nocallback nofree nosync nounwind speculatable willreturn memory(none) }
 ;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
+; CI-NSZ: {{.*}}
+; CI-SAFE: {{.*}}
 ; GFX11: {{.*}}
 ; GFX11-NSZ: {{.*}}
 ; GFX11-SAFE: {{.*}}

arsenm
arsenm reviewed Oct 7, 2025
View reviewed changes
llvm/test/CodeGen/AMDGPU/fsub.ll Outdated

; FUNC-LABEL: {{^}}v_fsub_0_nsz_attribute_f32:
; SI-NOT: v_sub
define amdgpu_kernel void @v_fsub_0_nsz_attribute_f32(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 {
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Rename function to flag? Or is the function redundant with a flag test already here?

arsenm
arsenm reviewed Oct 7, 2025
View reviewed changes
llvm/test/CodeGen/AMDGPU/select-fabs-fneg-extract.v2f16.ll Outdated
@@ -5048,6 +5132,8 @@ declare <2 x half> @llvm.fmuladd.v2f16(<2 x half>, <2 x half>, <2 x half>) #0

attributes #0 = { nocallback nofree nosync nounwind speculatable willreturn memory(none) }
;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
; CI-NSZ: {{.*}}
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Should this remove the flag from the run lines and merge them?

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@arsenm arsenm arsenm left review comments

Assignees
No one assigned
Labels
backend:AMDGPU llvm:SelectionDAG SelectionDAGISel as well
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
3 participants
@paperchalice @llvmbot @arsenm