From 3df3bc3b1eef5fc073666c1627a152c6727d4f39 Mon Sep 17 00:00:00 2001
From: Iris Shi <0.0@owo.li>
Date: Fri, 9 May 2025 18:01:25 +0800
Subject: [PATCH 1/2] [SelectionDAG] Handle `fneg`/`fabs`/`fcopysign` in
 `SimplifyDemandedBits`

---
 llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp |  29 +---
 .../CodeGen/SelectionDAG/TargetLowering.cpp   |  71 +++++++++
 .../CodeGen/AArch64/extract-vector-elt.ll     |   5 +-
 llvm/test/CodeGen/AMDGPU/bf16-conversions.ll  |  38 ++---
 llvm/test/CodeGen/AMDGPU/bf16.ll              |  36 ++---
 llvm/test/CodeGen/AMDGPU/fnearbyint.ll        | 142 +++++++++---------
 llvm/test/CodeGen/AMDGPU/rcp-pattern.ll       |  14 +-
 llvm/test/CodeGen/AMDGPU/roundeven.ll         |  49 +++---
 .../AMDGPU/select-fabs-fneg-extract.v2f16.ll  |  20 +--
 9 files changed, 228 insertions(+), 176 deletions(-)

diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 564cc372f595c..706c30d08944d 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -18259,21 +18259,6 @@ SDValue DAGCombiner::visitFCOPYSIGN(SDNode *N) {
     }
   }
 
-  // copysign(fabs(x), y) -> copysign(x, y)
-  // copysign(fneg(x), y) -> copysign(x, y)
-  // copysign(copysign(x,z), y) -> copysign(x, y)
-  if (N0.getOpcode() == ISD::FABS || N0.getOpcode() == ISD::FNEG ||
-      N0.getOpcode() == ISD::FCOPYSIGN)
-    return DAG.getNode(ISD::FCOPYSIGN, DL, VT, N0.getOperand(0), N1);
-
-  // copysign(x, abs(y)) -> abs(x)
-  if (N1.getOpcode() == ISD::FABS)
-    return DAG.getNode(ISD::FABS, DL, VT, N0);
-
-  // copysign(x, copysign(y,z)) -> copysign(x, z)
-  if (N1.getOpcode() == ISD::FCOPYSIGN)
-    return DAG.getNode(ISD::FCOPYSIGN, DL, VT, N0, N1.getOperand(1));
-
   // copysign(x, fp_extend(y)) -> copysign(x, y)
   // copysign(x, fp_round(y)) -> copysign(x, y)
   if (CanCombineFCOPYSIGN_EXTEND_ROUND(N))
@@ -18814,6 +18799,9 @@ SDValue DAGCombiner::visitFNEG(SDNode *N) {
                        N0.getOperand(0));
   }
 
+  if (SimplifyDemandedBits(N0, APInt::getAllOnes(VT.getScalarSizeInBits())))
+    return SDValue(N, 0);
+
   if (SDValue Cast = foldSignChangeInBitcast(N))
     return Cast;
 
@@ -18887,14 +18875,9 @@ SDValue DAGCombiner::visitFABS(SDNode *N) {
   if (SDValue C = DAG.FoldConstantArithmetic(ISD::FABS, DL, VT, {N0}))
     return C;
 
-  // fold (fabs (fabs x)) -> (fabs x)
-  if (N0.getOpcode() == ISD::FABS)
-    return N->getOperand(0);
-
-  // fold (fabs (fneg x)) -> (fabs x)
-  // fold (fabs (fcopysign x, y)) -> (fabs x)
-  if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN)
-    return DAG.getNode(ISD::FABS, DL, VT, N0.getOperand(0));
+  if (SimplifyDemandedBits(N0,
+                           APInt::getSignedMaxValue(VT.getScalarSizeInBits())))
+    return SDValue(N, 0);
 
   if (SDValue Cast = foldSignChangeInBitcast(N))
     return Cast;
diff --git a/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp b/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
index ba34c72156228..ffd74d09a11ed 100644
--- a/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
@@ -2958,6 +2958,77 @@ bool TargetLowering::SimplifyDemandedBits(
     }
     break;
   }
+  case ISD::FABS: {
+    SDValue Op0 = Op.getOperand(0);
+    APInt SignMask = APInt::getSignMask(BitWidth);
+
+    if (!DemandedBits.intersects(SignMask))
+      return TLO.CombineTo(Op, Op0);
+
+    if (SimplifyDemandedBits(Op0, ~SignMask & DemandedBits, DemandedElts, Known,
+                             TLO, Depth + 1))
+      return true;
+
+    if (Known.isNonNegative())
+      return TLO.CombineTo(Op, Op0);
+    if (Known.isNegative())
+      return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::FNEG, dl, VT, Op0));
+
+    Known.Zero |= SignMask;
+    Known.One &= ~SignMask;
+
+    break;
+  }
+  case ISD::FCOPYSIGN: {
+    SDValue Op0 = Op.getOperand(0);
+    SDValue Op1 = Op.getOperand(1);
+    APInt SignMask = APInt::getSignMask(BitWidth);
+
+    if (!DemandedBits.intersects(SignMask))
+      return TLO.CombineTo(Op, Op0);
+
+    if (SimplifyDemandedBits(Op0, ~SignMask & DemandedBits, DemandedElts, Known,
+                             TLO, Depth + 1))
+      return true;
+    if (SimplifyDemandedBits(Op1, SignMask, DemandedElts, Known2, TLO,
+                             Depth + 1))
+      return true;
+
+    if (ShrinkDemandedOp(Op, BitWidth, DemandedBits, TLO))
+      return true;
+
+    if ((Known.isNonNegative() && Known2.isNonNegative()) ||
+        (Known.isNegative() && Known2.isNegative()))
+      return TLO.CombineTo(Op, Op0);
+
+    if (Known2.isNonNegative())
+      return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::FABS, dl, VT, Op0));
+
+    if (Known2.isNegative()) {
+      Known.One |= SignMask;
+      Known.Zero &= ~SignMask;
+    }
+
+    break;
+  }
+  case ISD::FNEG: {
+    SDValue Op0 = Op.getOperand(0);
+    APInt SignMask = APInt::getSignMask(BitWidth);
+
+    if (!DemandedBits.intersects(SignMask))
+      return TLO.CombineTo(Op, Op0);
+
+    if (SimplifyDemandedBits(Op0, DemandedBits, DemandedElts, Known, TLO,
+                             Depth + 1))
+      return true;
+
+    if (Known.isNonNegative() || Known.isNegative()) {
+      Known.Zero ^= SignMask;
+      Known.One ^= SignMask;
+    }
+
+    break;
+  }
   default:
     // We also ask the target about intrinsics (which could be specific to it).
     if (Op.getOpcode() >= ISD::BUILTIN_OP_END ||
diff --git a/llvm/test/CodeGen/AArch64/extract-vector-elt.ll b/llvm/test/CodeGen/AArch64/extract-vector-elt.ll
index 5e5fdd6d31705..554b2e3444fe4 100644
--- a/llvm/test/CodeGen/AArch64/extract-vector-elt.ll
+++ b/llvm/test/CodeGen/AArch64/extract-vector-elt.ll
@@ -425,10 +425,7 @@ entry:
 define float @extract_v4i32_copysign_build_vector_const(<4 x float> %a, <4 x float> %b, i32 %c) {
 ; CHECK-SD-LABEL: extract_v4i32_copysign_build_vector_const:
 ; CHECK-SD:       // %bb.0: // %entry
-; CHECK-SD-NEXT:    adrp x8, .LCPI17_0
-; CHECK-SD-NEXT:    mvni v1.4s, #128, lsl #24
-; CHECK-SD-NEXT:    ldr q2, [x8, :lo12:.LCPI17_0]
-; CHECK-SD-NEXT:    bif v0.16b, v2.16b, v1.16b
+; CHECK-SD-NEXT:    fabs v0.4s, v0.4s
 ; CHECK-SD-NEXT:    mov s0, v0.s[2]
 ; CHECK-SD-NEXT:    ret
 ;
diff --git a/llvm/test/CodeGen/AMDGPU/bf16-conversions.ll b/llvm/test/CodeGen/AMDGPU/bf16-conversions.ll
index a597faa028f22..f03958a967328 100644
--- a/llvm/test/CodeGen/AMDGPU/bf16-conversions.ll
+++ b/llvm/test/CodeGen/AMDGPU/bf16-conversions.ll
@@ -427,16 +427,18 @@ entry:
 define amdgpu_ps void @fptrunc_f64_to_bf16_abs(double %a, ptr %out) {
 ; GFX-942-LABEL: fptrunc_f64_to_bf16_abs:
 ; GFX-942:       ; %bb.0: ; %entry
-; GFX-942-NEXT:    v_cvt_f32_f64_e64 v6, |v[0:1]|
-; GFX-942-NEXT:    v_cvt_f64_f32_e32 v[4:5], v6
-; GFX-942-NEXT:    v_and_b32_e32 v7, 1, v6
-; GFX-942-NEXT:    v_cmp_gt_f64_e64 s[2:3], |v[0:1]|, |v[4:5]|
-; GFX-942-NEXT:    v_cmp_nlg_f64_e64 s[0:1], |v[0:1]|, v[4:5]
-; GFX-942-NEXT:    v_cmp_eq_u32_e32 vcc, 1, v7
+; GFX-942-NEXT:    v_cvt_f32_f64_e64 v8, |v[0:1]|
+; GFX-942-NEXT:    v_and_b32_e32 v5, 0x7fffffff, v1
+; GFX-942-NEXT:    v_mov_b32_e32 v4, v0
+; GFX-942-NEXT:    v_cvt_f64_f32_e32 v[6:7], v8
+; GFX-942-NEXT:    v_and_b32_e32 v9, 1, v8
+; GFX-942-NEXT:    v_cmp_gt_f64_e64 s[2:3], |v[4:5]|, |v[6:7]|
+; GFX-942-NEXT:    v_cmp_nlg_f64_e64 s[0:1], |v[0:1]|, v[6:7]
+; GFX-942-NEXT:    v_cmp_eq_u32_e32 vcc, 1, v9
 ; GFX-942-NEXT:    v_cndmask_b32_e64 v4, -1, 1, s[2:3]
-; GFX-942-NEXT:    v_add_u32_e32 v4, v6, v4
+; GFX-942-NEXT:    v_add_u32_e32 v4, v8, v4
 ; GFX-942-NEXT:    s_or_b64 vcc, s[0:1], vcc
-; GFX-942-NEXT:    v_cndmask_b32_e32 v4, v4, v6, vcc
+; GFX-942-NEXT:    v_cndmask_b32_e32 v4, v4, v8, vcc
 ; GFX-942-NEXT:    v_bfe_u32 v5, v4, 16, 1
 ; GFX-942-NEXT:    s_movk_i32 s0, 0x7fff
 ; GFX-942-NEXT:    v_add3_u32 v5, v5, v4, s0
@@ -449,16 +451,18 @@ define amdgpu_ps void @fptrunc_f64_to_bf16_abs(double %a, ptr %out) {
 ;
 ; GFX-950-LABEL: fptrunc_f64_to_bf16_abs:
 ; GFX-950:       ; %bb.0: ; %entry
-; GFX-950-NEXT:    v_cvt_f32_f64_e64 v6, |v[0:1]|
-; GFX-950-NEXT:    v_cvt_f64_f32_e32 v[4:5], v6
-; GFX-950-NEXT:    v_and_b32_e32 v7, 1, v6
-; GFX-950-NEXT:    v_cmp_gt_f64_e64 s[2:3], |v[0:1]|, |v[4:5]|
-; GFX-950-NEXT:    v_cmp_nlg_f64_e64 s[0:1], |v[0:1]|, v[4:5]
-; GFX-950-NEXT:    v_cmp_eq_u32_e32 vcc, 1, v7
-; GFX-950-NEXT:    v_cndmask_b32_e64 v0, -1, 1, s[2:3]
-; GFX-950-NEXT:    v_add_u32_e32 v0, v6, v0
+; GFX-950-NEXT:    v_cvt_f32_f64_e64 v8, |v[0:1]|
+; GFX-950-NEXT:    v_and_b32_e32 v5, 0x7fffffff, v1
+; GFX-950-NEXT:    v_mov_b32_e32 v4, v0
+; GFX-950-NEXT:    v_cvt_f64_f32_e32 v[6:7], v8
+; GFX-950-NEXT:    v_cmp_nlg_f64_e64 s[0:1], |v[0:1]|, v[6:7]
+; GFX-950-NEXT:    v_and_b32_e32 v0, 1, v8
+; GFX-950-NEXT:    v_cmp_gt_f64_e64 s[2:3], |v[4:5]|, |v[6:7]|
+; GFX-950-NEXT:    v_cmp_eq_u32_e32 vcc, 1, v0
 ; GFX-950-NEXT:    s_or_b64 vcc, s[0:1], vcc
-; GFX-950-NEXT:    v_cndmask_b32_e32 v0, v0, v6, vcc
+; GFX-950-NEXT:    v_cndmask_b32_e64 v0, -1, 1, s[2:3]
+; GFX-950-NEXT:    v_add_u32_e32 v0, v8, v0
+; GFX-950-NEXT:    v_cndmask_b32_e32 v0, v0, v8, vcc
 ; GFX-950-NEXT:    v_cvt_pk_bf16_f32 v0, v0, s0
 ; GFX-950-NEXT:    flat_store_short v[2:3], v0
 ; GFX-950-NEXT:    s_endpgm
diff --git a/llvm/test/CodeGen/AMDGPU/bf16.ll b/llvm/test/CodeGen/AMDGPU/bf16.ll
index c4957fd44e2be..0007cfb667b7d 100644
--- a/llvm/test/CodeGen/AMDGPU/bf16.ll
+++ b/llvm/test/CodeGen/AMDGPU/bf16.ll
@@ -18639,8 +18639,8 @@ define bfloat @v_fabs_bf16(bfloat %a) {
 ; 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_and_b32_e32 v0, 0xffff0000, v0
-; GCN-NEXT:    v_and_b32_e32 v0, 0x7fffffff, v0
+; GCN-NEXT:    v_and_b32_e32 v0, 0x7fff0000, v0
+; GCN-NEXT:    v_mul_f32_e64 v0, 1.0, |v0|
 ; GCN-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
 ; GCN-NEXT:    s_setpc_b64 s[30:31]
 ;
@@ -18648,8 +18648,8 @@ define bfloat @v_fabs_bf16(bfloat %a) {
 ; 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_and_b32_e32 v0, 0xffff0000, v0
-; GFX7-NEXT:    v_and_b32_e32 v0, 0x7fffffff, v0
+; GFX7-NEXT:    v_and_b32_e32 v0, 0x7fff0000, v0
+; GFX7-NEXT:    v_mul_f32_e64 v0, 1.0, |v0|
 ; GFX7-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
 ; GFX7-NEXT:    s_setpc_b64 s[30:31]
 ;
@@ -18832,8 +18832,8 @@ define bfloat @v_fneg_fabs_bf16(bfloat %a) {
 ; 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_and_b32_e32 v0, 0xffff0000, v0
-; GCN-NEXT:    v_and_b32_e32 v0, 0x7fffffff, v0
+; GCN-NEXT:    v_and_b32_e32 v0, 0x7fff0000, v0
+; GCN-NEXT:    v_mul_f32_e64 v0, 1.0, |v0|
 ; GCN-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
 ; GCN-NEXT:    v_xor_b32_e32 v0, 0x80000000, v0
 ; GCN-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
@@ -18843,8 +18843,8 @@ define bfloat @v_fneg_fabs_bf16(bfloat %a) {
 ; 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_and_b32_e32 v0, 0xffff0000, v0
-; GFX7-NEXT:    v_and_b32_e32 v0, 0x7fffffff, v0
+; GFX7-NEXT:    v_and_b32_e32 v0, 0x7fff0000, v0
+; GFX7-NEXT:    v_mul_f32_e64 v0, 1.0, |v0|
 ; GFX7-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
 ; GFX7-NEXT:    v_xor_b32_e32 v0, 0x80000000, v0
 ; GFX7-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
@@ -18889,23 +18889,23 @@ define amdgpu_ps i32 @s_fneg_fabs_bf16(bfloat inreg %a) {
 ; GCN-LABEL: s_fneg_fabs_bf16:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    v_mul_f32_e64 v0, 1.0, s0
+; GCN-NEXT:    v_and_b32_e32 v0, 0x7fff0000, v0
+; GCN-NEXT:    v_mul_f32_e64 v0, 1.0, |v0|
+; GCN-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
+; GCN-NEXT:    v_xor_b32_e32 v0, 0x80000000, v0
+; GCN-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
 ; GCN-NEXT:    v_readfirstlane_b32 s0, v0
-; GCN-NEXT:    s_and_b32 s0, s0, 0xffff0000
-; GCN-NEXT:    s_bitset0_b32 s0, 31
-; GCN-NEXT:    s_and_b32 s0, s0, 0xffff0000
-; GCN-NEXT:    s_xor_b32 s0, s0, 0x80000000
-; GCN-NEXT:    s_lshr_b32 s0, s0, 16
 ; GCN-NEXT:    ; return to shader part epilog
 ;
 ; GFX7-LABEL: s_fneg_fabs_bf16:
 ; GFX7:       ; %bb.0:
 ; GFX7-NEXT:    v_mul_f32_e64 v0, 1.0, s0
+; GFX7-NEXT:    v_and_b32_e32 v0, 0x7fff0000, v0
+; GFX7-NEXT:    v_mul_f32_e64 v0, 1.0, |v0|
+; GFX7-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
+; GFX7-NEXT:    v_xor_b32_e32 v0, 0x80000000, v0
+; GFX7-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
 ; GFX7-NEXT:    v_readfirstlane_b32 s0, v0
-; GFX7-NEXT:    s_and_b32 s0, s0, 0xffff0000
-; GFX7-NEXT:    s_bitset0_b32 s0, 31
-; GFX7-NEXT:    s_and_b32 s0, s0, 0xffff0000
-; GFX7-NEXT:    s_xor_b32 s0, s0, 0x80000000
-; GFX7-NEXT:    s_lshr_b32 s0, s0, 16
 ; GFX7-NEXT:    ; return to shader part epilog
 ;
 ; GFX8-LABEL: s_fneg_fabs_bf16:
diff --git a/llvm/test/CodeGen/AMDGPU/fnearbyint.ll b/llvm/test/CodeGen/AMDGPU/fnearbyint.ll
index e9fd6119d0c36..9f9b14d1c87a0 100644
--- a/llvm/test/CodeGen/AMDGPU/fnearbyint.ll
+++ b/llvm/test/CodeGen/AMDGPU/fnearbyint.ll
@@ -211,22 +211,22 @@ define amdgpu_kernel void @nearbyint_f64(ptr addrspace(1) %out, double %in) {
 ; SI-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x9
 ; SI-NEXT:    s_mov_b32 s7, 0xf000
 ; SI-NEXT:    s_mov_b32 s6, -1
-; SI-NEXT:    s_brev_b32 s8, -2
-; SI-NEXT:    v_mov_b32_e32 v1, 0x43300000
-; SI-NEXT:    v_mov_b32_e32 v0, 0
-; SI-NEXT:    v_mov_b32_e32 v2, -1
-; SI-NEXT:    v_mov_b32_e32 v3, 0x432fffff
+; SI-NEXT:    v_mov_b32_e32 v0, -1
+; SI-NEXT:    v_mov_b32_e32 v1, 0x432fffff
+; SI-NEXT:    v_mov_b32_e32 v2, 0
+; SI-NEXT:    v_mov_b32_e32 v3, 0x43300000
+; SI-NEXT:    s_mov_b32 s8, 0
+; SI-NEXT:    s_mov_b32 s9, 0xc3300000
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
 ; SI-NEXT:    s_mov_b32 s4, s0
 ; SI-NEXT:    s_mov_b32 s5, s1
-; SI-NEXT:    v_mov_b32_e32 v6, s3
-; SI-NEXT:    v_bfi_b32 v1, s8, v1, v6
-; SI-NEXT:    v_mov_b32_e32 v7, s2
-; SI-NEXT:    v_add_f64 v[4:5], s[2:3], v[0:1]
-; SI-NEXT:    v_add_f64 v[0:1], v[4:5], -v[0:1]
-; SI-NEXT:    v_cmp_gt_f64_e64 vcc, |s[2:3]|, v[2:3]
-; SI-NEXT:    v_cndmask_b32_e32 v1, v1, v6, vcc
-; SI-NEXT:    v_cndmask_b32_e32 v0, v0, v7, vcc
+; SI-NEXT:    v_add_f64 v[2:3], s[2:3], v[2:3]
+; SI-NEXT:    v_mov_b32_e32 v4, s3
+; SI-NEXT:    v_mov_b32_e32 v5, s2
+; SI-NEXT:    v_add_f64 v[2:3], v[2:3], s[8:9]
+; SI-NEXT:    v_cmp_gt_f64_e64 vcc, |s[2:3]|, v[0:1]
+; SI-NEXT:    v_cndmask_b32_e32 v1, v3, v4, vcc
+; SI-NEXT:    v_cndmask_b32_e32 v0, v2, v5, vcc
 ; SI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
 ; SI-NEXT:    s_endpgm
 ;
@@ -270,30 +270,31 @@ define amdgpu_kernel void @nearbyint_v2f64(ptr addrspace(1) %out, <2 x double> %
 ; SI-NEXT:    s_load_dwordx4 s[4:7], s[4:5], 0xd
 ; SI-NEXT:    s_mov_b32 s3, 0xf000
 ; SI-NEXT:    s_mov_b32 s2, -1
-; SI-NEXT:    s_brev_b32 s10, -2
-; SI-NEXT:    v_mov_b32_e32 v6, 0x43300000
 ; SI-NEXT:    s_mov_b32 s9, 0x432fffff
-; SI-NEXT:    v_mov_b32_e32 v0, 0
+; SI-NEXT:    s_mov_b32 s10, 0
+; SI-NEXT:    s_mov_b32 s11, 0x43300000
+; SI-NEXT:    s_mov_b32 s12, 0
+; SI-NEXT:    s_mov_b32 s13, 0xc3300000
 ; SI-NEXT:    s_mov_b32 s8, s2
-; SI-NEXT:    v_mov_b32_e32 v4, s8
-; SI-NEXT:    v_mov_b32_e32 v5, s9
+; SI-NEXT:    v_mov_b32_e32 v0, s8
+; SI-NEXT:    v_mov_b32_e32 v1, s9
+; SI-NEXT:    v_mov_b32_e32 v2, s10
+; SI-NEXT:    v_mov_b32_e32 v3, s11
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    v_mov_b32_e32 v7, s7
-; SI-NEXT:    v_bfi_b32 v1, s10, v6, v7
-; SI-NEXT:    v_mov_b32_e32 v8, s6
-; SI-NEXT:    v_mov_b32_e32 v9, s5
-; SI-NEXT:    v_mov_b32_e32 v10, s4
-; SI-NEXT:    v_add_f64 v[2:3], s[6:7], v[0:1]
-; SI-NEXT:    v_add_f64 v[2:3], v[2:3], -v[0:1]
-; SI-NEXT:    v_bfi_b32 v1, s10, v6, v9
-; SI-NEXT:    v_cmp_gt_f64_e64 vcc, |s[6:7]|, v[4:5]
-; SI-NEXT:    v_cndmask_b32_e32 v3, v3, v7, vcc
-; SI-NEXT:    v_cndmask_b32_e32 v2, v2, v8, vcc
-; SI-NEXT:    v_add_f64 v[6:7], s[4:5], v[0:1]
-; SI-NEXT:    v_add_f64 v[0:1], v[6:7], -v[0:1]
-; SI-NEXT:    v_cmp_gt_f64_e64 vcc, |s[4:5]|, v[4:5]
-; SI-NEXT:    v_cndmask_b32_e32 v1, v1, v9, vcc
-; SI-NEXT:    v_cndmask_b32_e32 v0, v0, v10, vcc
+; SI-NEXT:    v_add_f64 v[4:5], s[6:7], v[2:3]
+; SI-NEXT:    v_mov_b32_e32 v8, s7
+; SI-NEXT:    v_mov_b32_e32 v9, s6
+; SI-NEXT:    v_add_f64 v[2:3], s[4:5], v[2:3]
+; SI-NEXT:    v_mov_b32_e32 v10, s5
+; SI-NEXT:    v_mov_b32_e32 v11, s4
+; SI-NEXT:    v_add_f64 v[4:5], v[4:5], s[12:13]
+; SI-NEXT:    v_add_f64 v[6:7], v[2:3], s[12:13]
+; SI-NEXT:    v_cmp_gt_f64_e64 vcc, |s[6:7]|, v[0:1]
+; SI-NEXT:    v_cndmask_b32_e32 v3, v5, v8, vcc
+; SI-NEXT:    v_cndmask_b32_e32 v2, v4, v9, vcc
+; SI-NEXT:    v_cmp_gt_f64_e64 vcc, |s[4:5]|, v[0:1]
+; SI-NEXT:    v_cndmask_b32_e32 v1, v7, v10, vcc
+; SI-NEXT:    v_cndmask_b32_e32 v0, v6, v11, vcc
 ; SI-NEXT:    buffer_store_dwordx4 v[0:3], off, s[0:3], 0
 ; SI-NEXT:    s_endpgm
 ;
@@ -347,46 +348,45 @@ define amdgpu_kernel void @nearbyint_v4f64(ptr addrspace(1) %out, <4 x double> %
 ; SI-NEXT:    s_load_dwordx8 s[0:7], s[4:5], 0x11
 ; SI-NEXT:    s_mov_b32 s11, 0xf000
 ; SI-NEXT:    s_mov_b32 s10, -1
-; SI-NEXT:    s_brev_b32 s14, -2
-; SI-NEXT:    v_mov_b32_e32 v10, 0x43300000
 ; SI-NEXT:    s_mov_b32 s13, 0x432fffff
-; SI-NEXT:    v_mov_b32_e32 v4, 0
+; SI-NEXT:    s_mov_b32 s14, 0
+; SI-NEXT:    s_mov_b32 s15, 0x43300000
+; SI-NEXT:    s_mov_b32 s16, 0
+; SI-NEXT:    s_mov_b32 s17, 0xc3300000
 ; SI-NEXT:    s_mov_b32 s12, s10
-; SI-NEXT:    v_mov_b32_e32 v8, s12
-; SI-NEXT:    v_mov_b32_e32 v9, s13
+; SI-NEXT:    v_mov_b32_e32 v4, s12
+; SI-NEXT:    v_mov_b32_e32 v5, s13
+; SI-NEXT:    v_mov_b32_e32 v0, s14
+; SI-NEXT:    v_mov_b32_e32 v1, s15
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    v_mov_b32_e32 v2, s3
-; SI-NEXT:    v_bfi_b32 v5, s14, v10, v2
-; SI-NEXT:    v_mov_b32_e32 v6, s2
-; SI-NEXT:    v_mov_b32_e32 v7, s1
-; SI-NEXT:    v_mov_b32_e32 v11, s0
-; SI-NEXT:    v_mov_b32_e32 v12, s7
-; SI-NEXT:    v_mov_b32_e32 v13, s6
-; SI-NEXT:    v_mov_b32_e32 v14, s5
-; SI-NEXT:    v_mov_b32_e32 v15, s4
-; SI-NEXT:    v_add_f64 v[0:1], s[2:3], v[4:5]
-; SI-NEXT:    v_add_f64 v[0:1], v[0:1], -v[4:5]
-; SI-NEXT:    v_bfi_b32 v5, s14, v10, v7
-; SI-NEXT:    v_cmp_gt_f64_e64 vcc, |s[2:3]|, v[8:9]
-; SI-NEXT:    v_cndmask_b32_e32 v3, v1, v2, vcc
-; SI-NEXT:    v_cndmask_b32_e32 v2, v0, v6, vcc
-; SI-NEXT:    v_add_f64 v[0:1], s[0:1], v[4:5]
-; SI-NEXT:    v_add_f64 v[0:1], v[0:1], -v[4:5]
-; SI-NEXT:    v_bfi_b32 v5, s14, v10, v12
-; SI-NEXT:    v_cmp_gt_f64_e64 vcc, |s[0:1]|, v[8:9]
-; SI-NEXT:    v_cndmask_b32_e32 v1, v1, v7, vcc
-; SI-NEXT:    v_cndmask_b32_e32 v0, v0, v11, vcc
-; SI-NEXT:    v_add_f64 v[6:7], s[6:7], v[4:5]
-; SI-NEXT:    v_add_f64 v[6:7], v[6:7], -v[4:5]
-; SI-NEXT:    v_bfi_b32 v5, s14, v10, v14
-; SI-NEXT:    v_cmp_gt_f64_e64 vcc, |s[6:7]|, v[8:9]
-; SI-NEXT:    v_cndmask_b32_e32 v7, v7, v12, vcc
-; SI-NEXT:    v_cndmask_b32_e32 v6, v6, v13, vcc
-; SI-NEXT:    v_add_f64 v[10:11], s[4:5], v[4:5]
-; SI-NEXT:    v_add_f64 v[4:5], v[10:11], -v[4:5]
-; SI-NEXT:    v_cmp_gt_f64_e64 vcc, |s[4:5]|, v[8:9]
-; SI-NEXT:    v_cndmask_b32_e32 v5, v5, v14, vcc
-; SI-NEXT:    v_cndmask_b32_e32 v4, v4, v15, vcc
+; SI-NEXT:    v_add_f64 v[2:3], s[2:3], v[0:1]
+; SI-NEXT:    v_mov_b32_e32 v10, s3
+; SI-NEXT:    v_mov_b32_e32 v11, s2
+; SI-NEXT:    v_add_f64 v[6:7], s[0:1], v[0:1]
+; SI-NEXT:    v_mov_b32_e32 v12, s1
+; SI-NEXT:    v_mov_b32_e32 v13, s0
+; SI-NEXT:    v_add_f64 v[8:9], s[6:7], v[0:1]
+; SI-NEXT:    v_mov_b32_e32 v14, s7
+; SI-NEXT:    v_mov_b32_e32 v15, s6
+; SI-NEXT:    v_add_f64 v[0:1], s[4:5], v[0:1]
+; SI-NEXT:    v_mov_b32_e32 v16, s5
+; SI-NEXT:    v_add_f64 v[2:3], v[2:3], s[16:17]
+; SI-NEXT:    v_cmp_gt_f64_e64 vcc, |s[2:3]|, v[4:5]
+; SI-NEXT:    v_cndmask_b32_e32 v3, v3, v10, vcc
+; SI-NEXT:    v_cndmask_b32_e32 v2, v2, v11, vcc
+; SI-NEXT:    v_mov_b32_e32 v17, s4
+; SI-NEXT:    v_add_f64 v[6:7], v[6:7], s[16:17]
+; SI-NEXT:    v_add_f64 v[8:9], v[8:9], s[16:17]
+; SI-NEXT:    v_add_f64 v[10:11], v[0:1], s[16:17]
+; SI-NEXT:    v_cmp_gt_f64_e64 vcc, |s[0:1]|, v[4:5]
+; SI-NEXT:    v_cndmask_b32_e32 v1, v7, v12, vcc
+; SI-NEXT:    v_cndmask_b32_e32 v0, v6, v13, vcc
+; SI-NEXT:    v_cmp_gt_f64_e64 vcc, |s[6:7]|, v[4:5]
+; SI-NEXT:    v_cndmask_b32_e32 v7, v9, v14, vcc
+; SI-NEXT:    v_cndmask_b32_e32 v6, v8, v15, vcc
+; SI-NEXT:    v_cmp_gt_f64_e64 vcc, |s[4:5]|, v[4:5]
+; SI-NEXT:    v_cndmask_b32_e32 v5, v11, v16, vcc
+; SI-NEXT:    v_cndmask_b32_e32 v4, v10, v17, vcc
 ; SI-NEXT:    buffer_store_dwordx4 v[4:7], off, s[8:11], 0 offset:16
 ; SI-NEXT:    buffer_store_dwordx4 v[0:3], off, s[8:11], 0
 ; SI-NEXT:    s_endpgm
diff --git a/llvm/test/CodeGen/AMDGPU/rcp-pattern.ll b/llvm/test/CodeGen/AMDGPU/rcp-pattern.ll
index 228420ef0acb0..0b6baf4b5f504 100644
--- a/llvm/test/CodeGen/AMDGPU/rcp-pattern.ll
+++ b/llvm/test/CodeGen/AMDGPU/rcp-pattern.ll
@@ -559,10 +559,11 @@ define float @v_rcp_fabs_f32_ieee_ulp25(float %x) #3 {
 ; SI-LABEL: v_rcp_fabs_f32_ieee_ulp25:
 ; SI:       ; %bb.0:
 ; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v1, 0x7fffffff, v0
 ; SI-NEXT:    s_mov_b32 s4, 0x7f800000
-; SI-NEXT:    v_frexp_mant_f32_e64 v1, |v0|
-; SI-NEXT:    v_cmp_lt_f32_e64 s[4:5], |v0|, s4
-; SI-NEXT:    v_cndmask_b32_e64 v1, |v0|, v1, s[4:5]
+; SI-NEXT:    v_frexp_mant_f32_e64 v2, |v0|
+; SI-NEXT:    v_cmp_lt_f32_e64 s[4:5], |v1|, s4
+; SI-NEXT:    v_cndmask_b32_e64 v1, |v0|, v2, s[4:5]
 ; SI-NEXT:    v_rcp_f32_e32 v1, v1
 ; SI-NEXT:    v_frexp_exp_i32_f32_e32 v0, v0
 ; SI-NEXT:    v_sub_i32_e32 v0, vcc, 0, v0
@@ -700,10 +701,11 @@ define float @v_rcp_neg_fabs_f32_ieee_ulp25(float %x) #3 {
 ; SI-LABEL: v_rcp_neg_fabs_f32_ieee_ulp25:
 ; SI:       ; %bb.0:
 ; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v1, 0x7fffffff, v0
 ; SI-NEXT:    s_mov_b32 s4, 0x7f800000
-; SI-NEXT:    v_frexp_mant_f32_e64 v1, -|v0|
-; SI-NEXT:    v_cmp_lt_f32_e64 s[4:5], |v0|, s4
-; SI-NEXT:    v_cndmask_b32_e64 v1, -|v0|, v1, s[4:5]
+; SI-NEXT:    v_frexp_mant_f32_e64 v2, -|v0|
+; SI-NEXT:    v_cmp_lt_f32_e64 s[4:5], |v1|, s4
+; SI-NEXT:    v_cndmask_b32_e64 v1, -|v0|, v2, s[4:5]
 ; SI-NEXT:    v_rcp_f32_e32 v1, v1
 ; SI-NEXT:    v_frexp_exp_i32_f32_e32 v0, v0
 ; SI-NEXT:    v_sub_i32_e32 v0, vcc, 0, v0
diff --git a/llvm/test/CodeGen/AMDGPU/roundeven.ll b/llvm/test/CodeGen/AMDGPU/roundeven.ll
index 59a1fe041bf90..5ebc08cf8402c 100644
--- a/llvm/test/CodeGen/AMDGPU/roundeven.ll
+++ b/llvm/test/CodeGen/AMDGPU/roundeven.ll
@@ -1124,14 +1124,14 @@ define double @v_roundeven_f64(double %x) {
 ; SDAG_GFX6-LABEL: v_roundeven_f64:
 ; SDAG_GFX6:       ; %bb.0:
 ; SDAG_GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; SDAG_GFX6-NEXT:    s_brev_b32 s6, -2
-; SDAG_GFX6-NEXT:    v_mov_b32_e32 v2, 0x43300000
-; SDAG_GFX6-NEXT:    v_bfi_b32 v3, s6, v2, v1
-; SDAG_GFX6-NEXT:    v_mov_b32_e32 v2, 0
+; SDAG_GFX6-NEXT:    s_mov_b32 s6, 0
+; SDAG_GFX6-NEXT:    s_mov_b32 s7, 0x43300000
 ; SDAG_GFX6-NEXT:    s_mov_b32 s4, -1
-; SDAG_GFX6-NEXT:    v_add_f64 v[4:5], v[0:1], v[2:3]
+; SDAG_GFX6-NEXT:    v_add_f64 v[2:3], v[0:1], s[6:7]
+; SDAG_GFX6-NEXT:    s_mov_b32 s6, 0
 ; SDAG_GFX6-NEXT:    s_mov_b32 s5, 0x432fffff
-; SDAG_GFX6-NEXT:    v_add_f64 v[2:3], v[4:5], -v[2:3]
+; SDAG_GFX6-NEXT:    s_mov_b32 s7, 0xc3300000
+; SDAG_GFX6-NEXT:    v_add_f64 v[2:3], v[2:3], s[6:7]
 ; SDAG_GFX6-NEXT:    v_cmp_gt_f64_e64 vcc, |v[0:1]|, s[4:5]
 ; SDAG_GFX6-NEXT:    v_cndmask_b32_e32 v0, v2, v0, vcc
 ; SDAG_GFX6-NEXT:    v_cndmask_b32_e32 v1, v3, v1, vcc
@@ -1208,18 +1208,18 @@ define double @v_roundeven_f64_fneg(double %x) {
 ; SDAG_GFX6-LABEL: v_roundeven_f64_fneg:
 ; SDAG_GFX6:       ; %bb.0:
 ; SDAG_GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; SDAG_GFX6-NEXT:    v_xor_b32_e32 v6, 0x80000000, v1
-; SDAG_GFX6-NEXT:    s_brev_b32 s4, -2
-; SDAG_GFX6-NEXT:    v_mov_b32_e32 v2, 0x43300000
-; SDAG_GFX6-NEXT:    v_bfi_b32 v3, s4, v2, v6
-; SDAG_GFX6-NEXT:    v_mov_b32_e32 v2, 0
-; SDAG_GFX6-NEXT:    v_add_f64 v[4:5], -v[0:1], v[2:3]
+; SDAG_GFX6-NEXT:    s_mov_b32 s6, 0
+; SDAG_GFX6-NEXT:    s_mov_b32 s7, 0x43300000
 ; SDAG_GFX6-NEXT:    s_mov_b32 s4, -1
+; SDAG_GFX6-NEXT:    v_add_f64 v[2:3], -v[0:1], s[6:7]
+; SDAG_GFX6-NEXT:    s_mov_b32 s6, 0
 ; SDAG_GFX6-NEXT:    s_mov_b32 s5, 0x432fffff
-; SDAG_GFX6-NEXT:    v_add_f64 v[2:3], v[4:5], -v[2:3]
+; SDAG_GFX6-NEXT:    s_mov_b32 s7, 0xc3300000
+; SDAG_GFX6-NEXT:    v_add_f64 v[2:3], v[2:3], s[6:7]
 ; SDAG_GFX6-NEXT:    v_cmp_gt_f64_e64 vcc, |v[0:1]|, s[4:5]
+; SDAG_GFX6-NEXT:    v_xor_b32_e32 v4, 0x80000000, v1
 ; SDAG_GFX6-NEXT:    v_cndmask_b32_e32 v0, v2, v0, vcc
-; SDAG_GFX6-NEXT:    v_cndmask_b32_e32 v1, v3, v6, vcc
+; SDAG_GFX6-NEXT:    v_cndmask_b32_e32 v1, v3, v4, vcc
 ; SDAG_GFX6-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; SDAG_GFX7-LABEL: v_roundeven_f64_fneg:
@@ -1304,20 +1304,19 @@ define <2 x double> @v_roundeven_v2f64(<2 x double> %x) {
 ; SDAG_GFX6-LABEL: v_roundeven_v2f64:
 ; SDAG_GFX6:       ; %bb.0:
 ; SDAG_GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; SDAG_GFX6-NEXT:    s_brev_b32 s6, -2
-; SDAG_GFX6-NEXT:    v_mov_b32_e32 v8, 0x43300000
-; SDAG_GFX6-NEXT:    v_bfi_b32 v5, s6, v8, v1
-; SDAG_GFX6-NEXT:    v_mov_b32_e32 v4, 0
-; SDAG_GFX6-NEXT:    v_add_f64 v[6:7], v[0:1], v[4:5]
+; SDAG_GFX6-NEXT:    s_mov_b32 s6, 0
+; SDAG_GFX6-NEXT:    s_mov_b32 s7, 0x43300000
 ; SDAG_GFX6-NEXT:    s_mov_b32 s4, -1
+; SDAG_GFX6-NEXT:    v_add_f64 v[4:5], v[0:1], s[6:7]
+; SDAG_GFX6-NEXT:    s_mov_b32 s8, 0
 ; SDAG_GFX6-NEXT:    s_mov_b32 s5, 0x432fffff
-; SDAG_GFX6-NEXT:    v_add_f64 v[5:6], v[6:7], -v[4:5]
+; SDAG_GFX6-NEXT:    s_mov_b32 s9, 0xc3300000
+; SDAG_GFX6-NEXT:    v_add_f64 v[4:5], v[4:5], s[8:9]
 ; SDAG_GFX6-NEXT:    v_cmp_gt_f64_e64 vcc, |v[0:1]|, s[4:5]
-; SDAG_GFX6-NEXT:    v_cndmask_b32_e32 v0, v5, v0, vcc
-; SDAG_GFX6-NEXT:    v_bfi_b32 v5, s6, v8, v3
-; SDAG_GFX6-NEXT:    v_add_f64 v[7:8], v[2:3], v[4:5]
-; SDAG_GFX6-NEXT:    v_cndmask_b32_e32 v1, v6, v1, vcc
-; SDAG_GFX6-NEXT:    v_add_f64 v[4:5], v[7:8], -v[4:5]
+; SDAG_GFX6-NEXT:    v_add_f64 v[6:7], v[2:3], s[6:7]
+; SDAG_GFX6-NEXT:    v_cndmask_b32_e32 v0, v4, v0, vcc
+; SDAG_GFX6-NEXT:    v_cndmask_b32_e32 v1, v5, v1, vcc
+; SDAG_GFX6-NEXT:    v_add_f64 v[4:5], v[6:7], s[8:9]
 ; SDAG_GFX6-NEXT:    v_cmp_gt_f64_e64 vcc, |v[2:3]|, s[4:5]
 ; SDAG_GFX6-NEXT:    v_cndmask_b32_e32 v2, v4, v2, vcc
 ; SDAG_GFX6-NEXT:    v_cndmask_b32_e32 v3, v5, v3, vcc
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 7ed27f008083e..8691925797621 100644
--- a/llvm/test/CodeGen/AMDGPU/select-fabs-fneg-extract.v2f16.ll
+++ b/llvm/test/CodeGen/AMDGPU/select-fabs-fneg-extract.v2f16.ll
@@ -776,20 +776,16 @@ define <2 x half> @add_select_fabs_negk_negk_v2f16(<2 x i32> %c, <2 x half> %x)
 ; CI-LABEL: add_select_fabs_negk_negk_v2f16:
 ; CI:       ; %bb.0:
 ; CI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; CI-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v0
-; CI-NEXT:    v_cndmask_b32_e64 v0, -1.0, -2.0, vcc
-; CI-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v1
-; CI-NEXT:    v_cndmask_b32_e64 v1, -1.0, -2.0, vcc
 ; CI-NEXT:    v_cvt_f16_f32_e32 v3, v3
 ; CI-NEXT:    v_cvt_f16_f32_e32 v2, v2
-; CI-NEXT:    v_cvt_f16_f32_e32 v0, v0
-; CI-NEXT:    v_cvt_f16_f32_e32 v1, v1
+; CI-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v1
+; CI-NEXT:    v_cndmask_b32_e64 v1, -1.0, -2.0, vcc
 ; CI-NEXT:    v_cvt_f32_f16_e32 v3, v3
 ; CI-NEXT:    v_cvt_f32_f16_e32 v2, v2
-; CI-NEXT:    v_cvt_f32_f16_e64 v0, |v0|
-; CI-NEXT:    v_cvt_f32_f16_e64 v1, |v1|
-; CI-NEXT:    v_add_f32_e32 v0, v0, v2
-; CI-NEXT:    v_add_f32_e32 v1, v1, v3
+; CI-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v0
+; CI-NEXT:    v_cndmask_b32_e64 v0, -1.0, -2.0, vcc
+; CI-NEXT:    v_sub_f32_e32 v1, v3, v1
+; CI-NEXT:    v_sub_f32_e32 v0, v2, v0
 ; CI-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; VI-LABEL: add_select_fabs_negk_negk_v2f16:
@@ -801,8 +797,8 @@ define <2 x half> @add_select_fabs_negk_negk_v2f16(<2 x i32> %c, <2 x half> %x)
 ; VI-NEXT:    v_cndmask_b32_e32 v0, v3, v4, vcc
 ; VI-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v1
 ; VI-NEXT:    v_cndmask_b32_e32 v1, v3, v4, vcc
-; VI-NEXT:    v_add_f16_sdwa v1, |v1|, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
-; VI-NEXT:    v_add_f16_e64 v0, |v0|, v2
+; VI-NEXT:    v_sub_f16_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
+; VI-NEXT:    v_sub_f16_e32 v0, v2, v0
 ; VI-NEXT:    v_or_b32_e32 v0, v0, v1
 ; VI-NEXT:    s_setpc_b64 s[30:31]
 ;

From 61c772f9ecf31483ef9e4f9c9048a39b31d8d4f1 Mon Sep 17 00:00:00 2001
From: Iris Shi <0.0@owo.li>
Date: Sat, 10 May 2025 23:56:28 +0800
Subject: [PATCH 2/2] update

---
 llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp |  4 ++
 .../CodeGen/SelectionDAG/TargetLowering.cpp   | 14 +++----
 llvm/test/CodeGen/AMDGPU/bf16-conversions.ll  | 38 +++++++++----------
 llvm/test/CodeGen/AMDGPU/rcp-pattern.ll       | 14 +++----
 4 files changed, 33 insertions(+), 37 deletions(-)

diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 706c30d08944d..097e1bfe84f99 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -18875,6 +18875,10 @@ SDValue DAGCombiner::visitFABS(SDNode *N) {
   if (SDValue C = DAG.FoldConstantArithmetic(ISD::FABS, DL, VT, {N0}))
     return C;
 
+  // fold (fabs (fabs x)) -> (fabs x)
+  if (N0.getOpcode() == ISD::FABS)
+    return N->getOperand(0);
+
   if (SimplifyDemandedBits(N0,
                            APInt::getSignedMaxValue(VT.getScalarSizeInBits())))
     return SDValue(N, 0);
diff --git a/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp b/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
index ffd74d09a11ed..dd436b257b22d 100644
--- a/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
@@ -2972,7 +2972,8 @@ bool TargetLowering::SimplifyDemandedBits(
     if (Known.isNonNegative())
       return TLO.CombineTo(Op, Op0);
     if (Known.isNegative())
-      return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::FNEG, dl, VT, Op0));
+      return TLO.CombineTo(
+          Op, TLO.DAG.getNode(ISD::FNEG, dl, VT, Op0, Op->getFlags()));
 
     Known.Zero |= SignMask;
     Known.One &= ~SignMask;
@@ -2988,21 +2989,18 @@ bool TargetLowering::SimplifyDemandedBits(
       return TLO.CombineTo(Op, Op0);
 
     if (SimplifyDemandedBits(Op0, ~SignMask & DemandedBits, DemandedElts, Known,
-                             TLO, Depth + 1))
-      return true;
-    if (SimplifyDemandedBits(Op1, SignMask, DemandedElts, Known2, TLO,
+                             TLO, Depth + 1) ||
+        SimplifyDemandedBits(Op1, SignMask, DemandedElts, Known2, TLO,
                              Depth + 1))
       return true;
 
-    if (ShrinkDemandedOp(Op, BitWidth, DemandedBits, TLO))
-      return true;
-
     if ((Known.isNonNegative() && Known2.isNonNegative()) ||
         (Known.isNegative() && Known2.isNegative()))
       return TLO.CombineTo(Op, Op0);
 
     if (Known2.isNonNegative())
-      return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::FABS, dl, VT, Op0));
+      return TLO.CombineTo(
+          Op, TLO.DAG.getNode(ISD::FABS, dl, VT, Op0, Op->getFlags()));
 
     if (Known2.isNegative()) {
       Known.One |= SignMask;
diff --git a/llvm/test/CodeGen/AMDGPU/bf16-conversions.ll b/llvm/test/CodeGen/AMDGPU/bf16-conversions.ll
index f03958a967328..a597faa028f22 100644
--- a/llvm/test/CodeGen/AMDGPU/bf16-conversions.ll
+++ b/llvm/test/CodeGen/AMDGPU/bf16-conversions.ll
@@ -427,18 +427,16 @@ entry:
 define amdgpu_ps void @fptrunc_f64_to_bf16_abs(double %a, ptr %out) {
 ; GFX-942-LABEL: fptrunc_f64_to_bf16_abs:
 ; GFX-942:       ; %bb.0: ; %entry
-; GFX-942-NEXT:    v_cvt_f32_f64_e64 v8, |v[0:1]|
-; GFX-942-NEXT:    v_and_b32_e32 v5, 0x7fffffff, v1
-; GFX-942-NEXT:    v_mov_b32_e32 v4, v0
-; GFX-942-NEXT:    v_cvt_f64_f32_e32 v[6:7], v8
-; GFX-942-NEXT:    v_and_b32_e32 v9, 1, v8
-; GFX-942-NEXT:    v_cmp_gt_f64_e64 s[2:3], |v[4:5]|, |v[6:7]|
-; GFX-942-NEXT:    v_cmp_nlg_f64_e64 s[0:1], |v[0:1]|, v[6:7]
-; GFX-942-NEXT:    v_cmp_eq_u32_e32 vcc, 1, v9
+; GFX-942-NEXT:    v_cvt_f32_f64_e64 v6, |v[0:1]|
+; GFX-942-NEXT:    v_cvt_f64_f32_e32 v[4:5], v6
+; GFX-942-NEXT:    v_and_b32_e32 v7, 1, v6
+; GFX-942-NEXT:    v_cmp_gt_f64_e64 s[2:3], |v[0:1]|, |v[4:5]|
+; GFX-942-NEXT:    v_cmp_nlg_f64_e64 s[0:1], |v[0:1]|, v[4:5]
+; GFX-942-NEXT:    v_cmp_eq_u32_e32 vcc, 1, v7
 ; GFX-942-NEXT:    v_cndmask_b32_e64 v4, -1, 1, s[2:3]
-; GFX-942-NEXT:    v_add_u32_e32 v4, v8, v4
+; GFX-942-NEXT:    v_add_u32_e32 v4, v6, v4
 ; GFX-942-NEXT:    s_or_b64 vcc, s[0:1], vcc
-; GFX-942-NEXT:    v_cndmask_b32_e32 v4, v4, v8, vcc
+; GFX-942-NEXT:    v_cndmask_b32_e32 v4, v4, v6, vcc
 ; GFX-942-NEXT:    v_bfe_u32 v5, v4, 16, 1
 ; GFX-942-NEXT:    s_movk_i32 s0, 0x7fff
 ; GFX-942-NEXT:    v_add3_u32 v5, v5, v4, s0
@@ -451,18 +449,16 @@ define amdgpu_ps void @fptrunc_f64_to_bf16_abs(double %a, ptr %out) {
 ;
 ; GFX-950-LABEL: fptrunc_f64_to_bf16_abs:
 ; GFX-950:       ; %bb.0: ; %entry
-; GFX-950-NEXT:    v_cvt_f32_f64_e64 v8, |v[0:1]|
-; GFX-950-NEXT:    v_and_b32_e32 v5, 0x7fffffff, v1
-; GFX-950-NEXT:    v_mov_b32_e32 v4, v0
-; GFX-950-NEXT:    v_cvt_f64_f32_e32 v[6:7], v8
-; GFX-950-NEXT:    v_cmp_nlg_f64_e64 s[0:1], |v[0:1]|, v[6:7]
-; GFX-950-NEXT:    v_and_b32_e32 v0, 1, v8
-; GFX-950-NEXT:    v_cmp_gt_f64_e64 s[2:3], |v[4:5]|, |v[6:7]|
-; GFX-950-NEXT:    v_cmp_eq_u32_e32 vcc, 1, v0
-; GFX-950-NEXT:    s_or_b64 vcc, s[0:1], vcc
+; GFX-950-NEXT:    v_cvt_f32_f64_e64 v6, |v[0:1]|
+; GFX-950-NEXT:    v_cvt_f64_f32_e32 v[4:5], v6
+; GFX-950-NEXT:    v_and_b32_e32 v7, 1, v6
+; GFX-950-NEXT:    v_cmp_gt_f64_e64 s[2:3], |v[0:1]|, |v[4:5]|
+; GFX-950-NEXT:    v_cmp_nlg_f64_e64 s[0:1], |v[0:1]|, v[4:5]
+; GFX-950-NEXT:    v_cmp_eq_u32_e32 vcc, 1, v7
 ; GFX-950-NEXT:    v_cndmask_b32_e64 v0, -1, 1, s[2:3]
-; GFX-950-NEXT:    v_add_u32_e32 v0, v8, v0
-; GFX-950-NEXT:    v_cndmask_b32_e32 v0, v0, v8, vcc
+; GFX-950-NEXT:    v_add_u32_e32 v0, v6, v0
+; GFX-950-NEXT:    s_or_b64 vcc, s[0:1], vcc
+; GFX-950-NEXT:    v_cndmask_b32_e32 v0, v0, v6, vcc
 ; GFX-950-NEXT:    v_cvt_pk_bf16_f32 v0, v0, s0
 ; GFX-950-NEXT:    flat_store_short v[2:3], v0
 ; GFX-950-NEXT:    s_endpgm
diff --git a/llvm/test/CodeGen/AMDGPU/rcp-pattern.ll b/llvm/test/CodeGen/AMDGPU/rcp-pattern.ll
index 0b6baf4b5f504..228420ef0acb0 100644
--- a/llvm/test/CodeGen/AMDGPU/rcp-pattern.ll
+++ b/llvm/test/CodeGen/AMDGPU/rcp-pattern.ll
@@ -559,11 +559,10 @@ define float @v_rcp_fabs_f32_ieee_ulp25(float %x) #3 {
 ; SI-LABEL: v_rcp_fabs_f32_ieee_ulp25:
 ; SI:       ; %bb.0:
 ; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0x7fffffff, v0
 ; SI-NEXT:    s_mov_b32 s4, 0x7f800000
-; SI-NEXT:    v_frexp_mant_f32_e64 v2, |v0|
-; SI-NEXT:    v_cmp_lt_f32_e64 s[4:5], |v1|, s4
-; SI-NEXT:    v_cndmask_b32_e64 v1, |v0|, v2, s[4:5]
+; SI-NEXT:    v_frexp_mant_f32_e64 v1, |v0|
+; SI-NEXT:    v_cmp_lt_f32_e64 s[4:5], |v0|, s4
+; SI-NEXT:    v_cndmask_b32_e64 v1, |v0|, v1, s[4:5]
 ; SI-NEXT:    v_rcp_f32_e32 v1, v1
 ; SI-NEXT:    v_frexp_exp_i32_f32_e32 v0, v0
 ; SI-NEXT:    v_sub_i32_e32 v0, vcc, 0, v0
@@ -701,11 +700,10 @@ define float @v_rcp_neg_fabs_f32_ieee_ulp25(float %x) #3 {
 ; SI-LABEL: v_rcp_neg_fabs_f32_ieee_ulp25:
 ; SI:       ; %bb.0:
 ; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; SI-NEXT:    v_and_b32_e32 v1, 0x7fffffff, v0
 ; SI-NEXT:    s_mov_b32 s4, 0x7f800000
-; SI-NEXT:    v_frexp_mant_f32_e64 v2, -|v0|
-; SI-NEXT:    v_cmp_lt_f32_e64 s[4:5], |v1|, s4
-; SI-NEXT:    v_cndmask_b32_e64 v1, -|v0|, v2, s[4:5]
+; SI-NEXT:    v_frexp_mant_f32_e64 v1, -|v0|
+; SI-NEXT:    v_cmp_lt_f32_e64 s[4:5], |v0|, s4
+; SI-NEXT:    v_cndmask_b32_e64 v1, -|v0|, v1, s[4:5]
 ; SI-NEXT:    v_rcp_f32_e32 v1, v1
 ; SI-NEXT:    v_frexp_exp_i32_f32_e32 v0, v0
 ; SI-NEXT:    v_sub_i32_e32 v0, vcc, 0, v0