[DAGCombiner] Relax nsz constraint with fp->int->fp optimizations

llvm/include/llvm/CodeGen/SelectionDAG.h

@@ -2322,6 +2322,10 @@ class SelectionDAG {
   /// +nan are considered positive, -0.0, -inf and -nan are not.
   LLVM_ABI bool cannotBeOrderedNegativeFP(SDValue Op) const;
 
+  /// Check if all uses of a floating-point value are insensitive to signed
+  /// zeros.
+  LLVM_ABI bool allUsesSignedZeroInsensitive(SDValue Op) const;
+
   /// Test whether two SDValues are known to compare equal. This
   /// is true if they are the same value, or if one is negative zero and the
   /// other positive zero.

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

@@ -18891,12 +18891,13 @@ static SDValue foldFPToIntToFP(SDNode *N, const SDLoc &DL, SelectionDAG &DAG,
   bool IsSigned = N->getOpcode() == ISD::SINT_TO_FP;
   assert(IsSigned || IsUnsigned);
 
-  bool IsSignedZeroSafe = DAG.getTarget().Options.NoSignedZerosFPMath;
+  bool IsSignedZeroSafe = DAG.getTarget().Options.NoSignedZerosFPMath ||
+                          DAG.allUsesSignedZeroInsensitive(SDValue(N, 0));
   // For signed conversions: The optimization changes signed zero behavior.
   if (IsSigned && !IsSignedZeroSafe)
     return SDValue();
   // For unsigned conversions, we need FABS to canonicalize -0.0 to +0.0
-  // (unless NoSignedZerosFPMath is set).
+  // (unless outputting a signed zero is OK).
   if (IsUnsigned && !IsSignedZeroSafe && !TLI.isFAbsFree(VT))
     return SDValue();
 
@@ -19375,10 +19376,17 @@ 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 && N->getFlags().hasNoSignedZeros() &&
-      N0.hasOneUse()) {
-    return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N0.getOperand(1),
-                       N0.getOperand(0));
+  if (N0.getOpcode() == ISD::FSUB && N0.hasOneUse()) {
+    SDValue X = N0.getOperand(0);
+    SDValue Y = N0.getOperand(1);
+
+    // Safe if NoSignedZeros, or if we can prove X != Y (avoiding the -0.0 vs
+    // +0.0 issue) For now, we use a conservative check: if either operand is
+    // known never zero, then X - Y can't produce a signed zero from X == Y.
+    if (N->getFlags().hasNoSignedZeros() || DAG.isKnownNeverZeroFloat(X) ||
+        DAG.isKnownNeverZeroFloat(Y)) {
+      return DAG.getNode(ISD::FSUB, SDLoc(N), VT, Y, X);
+    }
   }
 
   if (SimplifyDemandedBits(SDValue(N, 0)))

llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp

@@ -6075,6 +6075,35 @@ bool SelectionDAG::isKnownNeverZeroFloat(SDValue Op) const {
       Op, [](ConstantFPSDNode *C) { return !C->isZero(); });
 }
 
+bool SelectionDAG::allUsesSignedZeroInsensitive(SDValue Op) const {
+  assert(Op.getValueType().isFloatingPoint());
+  return all_of(Op->uses(), [&](SDUse &Use) {
+    SDNode *User = Use.getUser();
+    unsigned OperandNo = Use.getOperandNo();
+
+    // Check if this use is insensitive to the sign of zero
+    switch (User->getOpcode()) {
+    case ISD::SETCC:
+      // Comparisons: IEEE-754 specifies +0.0 == -0.0.
+    case ISD::FABS:
+      // fabs always produces +0.0.
+      return true;
+    case ISD::FCOPYSIGN:
+      // copysign overwrites the sign bit of the first operand.
+      return OperandNo == 0;
+    case ISD::FADD:
+    case ISD::FSUB: {
+      // Arithmetic with non-zero constants fixes the uncertainty around the
+      // sign bit.
+      SDValue Other = User->getOperand(1 - OperandNo);
+      return isKnownNeverZeroFloat(Other);
+    }
+    default:
+      return false;
+    }
+  });
+}
+
 bool SelectionDAG::isKnownNeverZero(SDValue Op, unsigned Depth) const {
   if (Depth >= MaxRecursionDepth)
     return false; // Limit search depth.

llvm/test/CodeGen/AArch64/fp-to-int-to-fp.ll

@@ -134,7 +134,111 @@ entry:
   ret float %f
 }
 
+define i1 @test_fcmp(float %x) {
+; CHECK-LABEL: test_fcmp:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    frintz s0, s0
+; CHECK-NEXT:    fcmp s0, #0.0
+; CHECK-NEXT:    cset w0, eq
+; CHECK-NEXT:    ret
+;
+; NO-SIGNED-ZEROS-LABEL: test_fcmp:
+; NO-SIGNED-ZEROS:       // %bb.0:
+; NO-SIGNED-ZEROS-NEXT:    frintz s0, s0
+; NO-SIGNED-ZEROS-NEXT:    fcmp s0, #0.0
+; NO-SIGNED-ZEROS-NEXT:    cset w0, eq
+; NO-SIGNED-ZEROS-NEXT:    ret
+  %conv1 = fptosi float %x to i32
+  %conv2 = sitofp i32 %conv1 to float
+  %cmp = fcmp oeq float %conv2, 0.0
+  ret i1 %cmp
+}
+
+define float @test_fabs(float %x) {
+; CHECK-LABEL: test_fabs:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    frintz s0, s0
+; CHECK-NEXT:    fabs s0, s0
+; CHECK-NEXT:    ret
+;
+; NO-SIGNED-ZEROS-LABEL: test_fabs:
+; NO-SIGNED-ZEROS:       // %bb.0:
+; NO-SIGNED-ZEROS-NEXT:    frintz s0, s0
+; NO-SIGNED-ZEROS-NEXT:    fabs s0, s0
+; NO-SIGNED-ZEROS-NEXT:    ret
+  %conv1 = fptosi float %x to i32
+  %conv2 = sitofp i32 %conv1 to float
+  %abs = call float @llvm.fabs.f32(float %conv2)
+  ret float %abs
+}
+
+define float @test_copysign(float %x, float %y) {
+; CHECK-LABEL: test_copysign:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    frintz s0, s0
+; CHECK-NEXT:    mvni v2.4s, #128, lsl #24
+; CHECK-NEXT:    // kill: def $s1 killed $s1 def $q1
+; CHECK-NEXT:    bif v0.16b, v1.16b, v2.16b
+; CHECK-NEXT:    // kill: def $s0 killed $s0 killed $q0
+; CHECK-NEXT:    ret
+;
+; NO-SIGNED-ZEROS-LABEL: test_copysign:
+; NO-SIGNED-ZEROS:       // %bb.0:
+; NO-SIGNED-ZEROS-NEXT:    frintz s0, s0
+; NO-SIGNED-ZEROS-NEXT:    mvni v2.4s, #128, lsl #24
+; NO-SIGNED-ZEROS-NEXT:    // kill: def $s1 killed $s1 def $q1
+; NO-SIGNED-ZEROS-NEXT:    bif v0.16b, v1.16b, v2.16b
+; NO-SIGNED-ZEROS-NEXT:    // kill: def $s0 killed $s0 killed $q0
+; NO-SIGNED-ZEROS-NEXT:    ret
+  %conv1 = fptosi float %x to i32
+  %conv2 = sitofp i32 %conv1 to float
+  %combine = call float @llvm.copysign.f32(float %conv2, float %y)
+  ret float %combine
+}
+
+define float @test_fadd(float %x) {
+; CHECK-LABEL: test_fadd:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    frintz s0, s0
+; CHECK-NEXT:    fmov s1, #1.00000000
+; CHECK-NEXT:    fadd s0, s0, s1
+; CHECK-NEXT:    ret
+;
+; NO-SIGNED-ZEROS-LABEL: test_fadd:
+; NO-SIGNED-ZEROS:       // %bb.0:
+; NO-SIGNED-ZEROS-NEXT:    frintz s0, s0
+; NO-SIGNED-ZEROS-NEXT:    fmov s1, #1.00000000
+; NO-SIGNED-ZEROS-NEXT:    fadd s0, s0, s1
+; NO-SIGNED-ZEROS-NEXT:    ret
+  %conv1 = fptosi float %x to i32
+  %conv2 = sitofp i32 %conv1 to float
+  %add = fadd float %conv2, 1.0
+  ret float %add
+}
+
+define float @test_fsub(float %x) {
+; CHECK-LABEL: test_fsub:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    frintz s0, s0
+; CHECK-NEXT:    fmov s1, #-1.00000000
+; CHECK-NEXT:    fadd s0, s0, s1
+; CHECK-NEXT:    ret
+;
+; NO-SIGNED-ZEROS-LABEL: test_fsub:
+; NO-SIGNED-ZEROS:       // %bb.0:
+; NO-SIGNED-ZEROS-NEXT:    frintz s0, s0
+; NO-SIGNED-ZEROS-NEXT:    fmov s1, #-1.00000000
+; NO-SIGNED-ZEROS-NEXT:    fadd s0, s0, s1
+; NO-SIGNED-ZEROS-NEXT:    ret
+  %conv1 = fptosi float %x to i32
+  %conv2 = sitofp i32 %conv1 to float
+  %sub = fsub float %conv2, 1.0
+  ret float %sub
+}
+
 declare i32 @llvm.smin.i32(i32, i32)
 declare i32 @llvm.smax.i32(i32, i32)
 declare i32 @llvm.umin.i32(i32, i32)
 declare i32 @llvm.umax.i32(i32, i32)
+declare float @llvm.fabs.f32(float)
+declare float @llvm.copysign.f32(float, float)

llvm/test/CodeGen/AMDGPU/select-fabs-fneg-extract.f16.ll

@@ -2615,65 +2615,43 @@ define half @select_fneg_posk_src_add_f16(i32 %c, half %x, half %y) {
 }
 
 define half @select_fneg_posk_src_sub_f16(i32 %c, half %x) {
-; CI-SAFE-LABEL: select_fneg_posk_src_sub_f16:
-; CI-SAFE:       ; %bb.0:
-; CI-SAFE-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; CI-SAFE-NEXT:    v_cvt_f16_f32_e32 v1, v1
-; CI-SAFE-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v0
-; CI-SAFE-NEXT:    v_cvt_f32_f16_e32 v1, v1
-; CI-SAFE-NEXT:    v_add_f32_e32 v1, -4.0, v1
-; CI-SAFE-NEXT:    v_cndmask_b32_e64 v0, 2.0, -v1, vcc
-; CI-SAFE-NEXT:    s_setpc_b64 s[30:31]
+; CI-LABEL: select_fneg_posk_src_sub_f16:
+; CI:       ; %bb.0:
+; CI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CI-NEXT:    v_cvt_f16_f32_e32 v1, v1
+; CI-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v0
+; CI-NEXT:    v_cvt_f32_f16_e32 v1, v1
+; CI-NEXT:    v_sub_f32_e32 v1, 4.0, v1
+; CI-NEXT:    v_cndmask_b32_e32 v0, 2.0, v1, vcc
+; CI-NEXT:    s_setpc_b64 s[30:31]
 ;
-; VI-SAFE-LABEL: select_fneg_posk_src_sub_f16:
-; VI-SAFE:       ; %bb.0:
-; VI-SAFE-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-SAFE-NEXT:    v_add_f16_e32 v1, -4.0, v1
-; VI-SAFE-NEXT:    v_xor_b32_e32 v1, 0x8000, v1
-; VI-SAFE-NEXT:    v_mov_b32_e32 v2, 0x4000
-; VI-SAFE-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v0
-; VI-SAFE-NEXT:    v_cndmask_b32_e32 v0, v2, v1, vcc
-; VI-SAFE-NEXT:    s_setpc_b64 s[30:31]
+; VI-LABEL: select_fneg_posk_src_sub_f16:
+; VI:       ; %bb.0:
+; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; VI-NEXT:    v_sub_f16_e32 v1, 4.0, v1
+; VI-NEXT:    v_mov_b32_e32 v2, 0x4000
+; VI-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v0
+; VI-NEXT:    v_cndmask_b32_e32 v0, v2, v1, vcc
+; VI-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX11-SAFE-TRUE16-LABEL: select_fneg_posk_src_sub_f16:
 ; GFX11-SAFE-TRUE16:       ; %bb.0:
 ; GFX11-SAFE-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-SAFE-TRUE16-NEXT:    v_add_f16_e32 v1.l, -4.0, v1.l
 ; GFX11-SAFE-TRUE16-NEXT:    v_cmp_eq_u32_e32 vcc_lo, 0, v0
-; GFX11-SAFE-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-SAFE-TRUE16-NEXT:    v_xor_b16 v0.l, 0x8000, v1.l
+; GFX11-SAFE-TRUE16-NEXT:    v_sub_f16_e32 v0.l, 4.0, v1.l
+; GFX11-SAFE-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 ; GFX11-SAFE-TRUE16-NEXT:    v_cndmask_b16 v0.l, 0x4000, v0.l, vcc_lo
 ; GFX11-SAFE-TRUE16-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX11-SAFE-FAKE16-LABEL: select_fneg_posk_src_sub_f16:
 ; GFX11-SAFE-FAKE16:       ; %bb.0:
 ; GFX11-SAFE-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-SAFE-FAKE16-NEXT:    v_add_f16_e32 v1, -4.0, v1
+; GFX11-SAFE-FAKE16-NEXT:    v_sub_f16_e32 v1, 4.0, v1
 ; 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(NEXT) | instid1(VALU_DEP_1)
-; GFX11-SAFE-FAKE16-NEXT:    v_xor_b32_e32 v1, 0x8000, v1
+; GFX11-SAFE-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
 ; GFX11-SAFE-FAKE16-NEXT:    v_cndmask_b32_e32 v0, 0x4000, v1, vcc_lo
 ; GFX11-SAFE-FAKE16-NEXT:    s_setpc_b64 s[30:31]
 ;
-; CI-NSZ-LABEL: select_fneg_posk_src_sub_f16:
-; CI-NSZ:       ; %bb.0:
-; CI-NSZ-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; CI-NSZ-NEXT:    v_cvt_f16_f32_e32 v1, v1
-; CI-NSZ-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v0
-; CI-NSZ-NEXT:    v_cvt_f32_f16_e32 v1, v1
-; CI-NSZ-NEXT:    v_sub_f32_e32 v1, 4.0, v1
-; CI-NSZ-NEXT:    v_cndmask_b32_e32 v0, 2.0, v1, vcc
-; CI-NSZ-NEXT:    s_setpc_b64 s[30:31]
-;
-; VI-NSZ-LABEL: select_fneg_posk_src_sub_f16:
-; VI-NSZ:       ; %bb.0:
-; VI-NSZ-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; VI-NSZ-NEXT:    v_sub_f16_e32 v1, 4.0, v1
-; VI-NSZ-NEXT:    v_mov_b32_e32 v2, 0x4000
-; VI-NSZ-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v0
-; VI-NSZ-NEXT:    v_cndmask_b32_e32 v0, v2, v1, vcc
-; VI-NSZ-NEXT:    s_setpc_b64 s[30:31]
-;
 ; GFX11-NSZ-TRUE16-LABEL: select_fneg_posk_src_sub_f16:
 ; GFX11-NSZ-TRUE16:       ; %bb.0:
 ; GFX11-NSZ-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)

llvm/test/CodeGen/AMDGPU/select-fabs-fneg-extract.v2f16.ll

@@ -3277,48 +3277,37 @@ define <2 x half> @select_fneg_posk_src_sub_v2f16(<2 x i32> %c, <2 x half> %x) {
 ; 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_cmp_eq_u32_e32 vcc, 0, v0
 ; CI-NEXT:    v_cvt_f32_f16_e32 v2, v2
-; CI-NEXT:    v_cndmask_b32_e32 v0, 2.0, v3, vcc
+; 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, v2, vcc
+; 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:
 ; 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, 0xc400
-; VI-NEXT:    v_add_f16_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; VI-NEXT:    v_add_f16_e32 v2, -4.0, v2
-; VI-NEXT:    v_or_b32_e32 v1, v2, v1
-; VI-NEXT:    v_xor_b32_e32 v1, 0x80008000, v1
-; VI-NEXT:    v_mov_b32_e32 v2, 0x4000
+; 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, v2, v1, s[4:5]
-; VI-NEXT:    v_cndmask_b32_sdwa v1, v2, v1, vcc dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
+; 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:
 ; 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]
-; GFX9-NEXT:    v_xor_b32_e32 v1, 0x80008000, 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]
@@ -3330,28 +3319,25 @@ define <2 x half> @select_fneg_posk_src_sub_v2f16(<2 x i32> %c, <2 x half> %x) {
 ; GFX11-TRUE16-LABEL: select_fneg_posk_src_sub_v2f16:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, v2, -4.0 op_sel_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_cmp_eq_u32_e32 vcc_lo, 0, v0
+; GFX11-TRUE16-NEXT:    v_pk_add_f16 v0, v2, 4.0 op_sel_hi:[1,0] neg_lo:[1,0] neg_hi:[1,0]
 ; GFX11-TRUE16-NEXT:    v_cmp_eq_u32_e64 s0, 0, v1
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT:    v_xor_b32_e32 v0, 0x80008000, v2
+; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.l, 0x4000, v0.l, vcc_lo
-; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
 ; GFX11-TRUE16-NEXT:    v_cndmask_b16 v0.h, 0x4000, v0.h, s0
 ; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX11-FAKE16-LABEL: select_fneg_posk_src_sub_v2f16:
 ; GFX11-FAKE16:       ; %bb.0:
 ; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-FAKE16-NEXT:    v_pk_add_f16 v2, v2, -4.0 op_sel_hi:[1,0]
+; GFX11-FAKE16-NEXT:    v_pk_add_f16 v2, v2, 4.0 op_sel_hi:[1,0] neg_lo:[1,0] neg_hi:[1,0]
 ; GFX11-FAKE16-NEXT:    v_cmp_eq_u32_e32 vcc_lo, 0, v0
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-FAKE16-NEXT:    v_xor_b32_e32 v2, 0x80008000, v2
+; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
 ; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v2
 ; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, 0x4000, v2, vcc_lo
 ; GFX11-FAKE16-NEXT:    v_cmp_eq_u32_e32 vcc_lo, 0, v1
-; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, 0x4000, v3, vcc_lo
+; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 ; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v1, v0, 0x5040100
 ; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
   %cmp = icmp eq <2 x i32> %c, zeroinitializer