Skip to content

expandFMINIMUMNUM_FMAXIMUMNUM: Improve compare between zeros #140193

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 2 commits into
base: main
Choose a base branch
from
Open

expandFMINIMUMNUM_FMAXIMUMNUM: Improve compare between zeros #140193

wants to merge 2 commits into from

Conversation

wzssyqa
Copy link
Contributor

@wzssyqa wzssyqa commented May 16, 2025

  1. On GPR32 platform, expandIS_FPCLASS may fail due to ISD::BITCAST
    double to int64 may fail. Let's FP_ROUND double to float first.
    Since we use it if MinMax is zero only, so the flushing won't
    break anything.

  2. Only one IS_FPCLASS is needed. MinMax will always be RHS if equal.
    So we can select between LHS and MinMax.
    It will even safe if FP_ROUND flush a small LHS, as if LHS is not zero
    then, MinMax won't be Zero, so we will always use MinMax.

@llvmbot
Copy link
Member

llvmbot commented May 16, 2025 •
edited
Loading

@llvm/pr-subscribers-backend-amdgpu

@llvm/pr-subscribers-backend-x86

Author: YunQiang Su (wzssyqa)

Changes

  1. On GPR32 platform, expandIS_FPCLASS may fail due to ISD::BITCAST
    double to int64 may fail. Let's FP_ROUND double to float first.
    Since we use it if MinMax is zero only, so the flushing won't
    break anything.

  2. Only one IS_FPCLASS is needed. MinMax will always be RHS if equal.
    So we can select between LHS and MinMax.
    It will even safe if FP_ROUND flush a small LHS, as if LHS is not zero
    then, MinMax won't be Zero, so we will always use MinMax.

Patch is 2.87 MiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/140193.diff

7 Files Affected:

  • (modified) llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp (+21-11)
  • (modified) llvm/test/CodeGen/AMDGPU/fmax3-maximumnum.ll (+193-493)
  • (modified) llvm/test/CodeGen/AMDGPU/fmin3-minimumnum.ll (+217-518)
  • (modified) llvm/test/CodeGen/AMDGPU/maximumnum.bf16.ll (+7105-12288)
  • (modified) llvm/test/CodeGen/AMDGPU/minimumnum.bf16.ll (+7214-12382)
  • (modified) llvm/test/CodeGen/Mips/fp-maximumnum-minimumnum.ll (+880)
  • (modified) llvm/test/CodeGen/X86/fminimumnum-fmaximumnum.ll (+116-162)
diff --git a/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp b/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
index f17d6a2787889..5166bc0ecd3b8 100644
--- a/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
@@ -8681,13 +8681,9 @@ SDValue TargetLowering::expandFMINIMUMNUM_FMAXIMUMNUM(SDNode *Node,
     RHS = DAG.getSelectCC(DL, RHS, RHS, LHS, RHS, ISD::SETUO);
   }
 
+  // Please always prefer RHS if equal.
   SDValue MinMax =
       DAG.getSelectCC(DL, LHS, RHS, LHS, RHS, IsMax ? ISD::SETGT : ISD::SETLT);
-  // If MinMax is NaN, let's quiet it.
-  if (!Flags.hasNoNaNs() && !DAG.isKnownNeverNaN(LHS) &&
-      !DAG.isKnownNeverNaN(RHS)) {
-    MinMax = DAG.getNode(ISD::FCANONICALIZE, DL, VT, MinMax, Flags);
-  }
 
   // Fixup signed zero behavior.
   if (Options.NoSignedZerosFPMath || Flags.hasNoSignedZeros() ||
@@ -8698,13 +8694,27 @@ SDValue TargetLowering::expandFMINIMUMNUM_FMAXIMUMNUM(SDNode *Node,
       DAG.getTargetConstant(IsMax ? fcPosZero : fcNegZero, DL, MVT::i32);
   SDValue IsZero = DAG.getSetCC(DL, CCVT, MinMax,
                                 DAG.getConstantFP(0.0, DL, VT), ISD::SETEQ);
-  SDValue LCmp = DAG.getSelect(
-      DL, VT, DAG.getNode(ISD::IS_FPCLASS, DL, CCVT, LHS, TestZero), LHS,
+  unsigned BitSize = VT.getScalarSizeInBits();
+  EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), BitSize);
+  EVT FloatVT = EVT::getFloatingPointVT(32);
+  if (VT.isVector()) {
+    IntVT =
+        EVT::getVectorVT(*DAG.getContext(), IntVT, VT.getVectorElementCount());
+    FloatVT = EVT::getVectorVT(*DAG.getContext(), FloatVT,
+                               VT.getVectorElementCount());
+  }
+  SDValue LHSTrunc = LHS;
+  if (!isOperationLegal(ISD::BITCAST, IntVT) &&
+      !isOperationLegal(ISD::IS_FPCLASS, VT)) {
+    LHSTrunc = DAG.getNode(ISD::FP_ROUND, DL, FloatVT, LHS,
+                           DAG.getIntPtrConstant(0, DL, /*isTarget=*/true));
+  }
+  // It's OK to select from LHS and MinMax, with only one ISD::IS_FPCLASS, as
+  // we preferred RHS when generate MinMax, if the operands are equal.
+  SDValue RetZero = DAG.getSelect(
+      DL, VT, DAG.getNode(ISD::IS_FPCLASS, DL, CCVT, LHSTrunc, TestZero), LHS,
       MinMax, Flags);
-  SDValue RCmp = DAG.getSelect(
-      DL, VT, DAG.getNode(ISD::IS_FPCLASS, DL, CCVT, RHS, TestZero), RHS, LCmp,
-      Flags);
-  return DAG.getSelect(DL, VT, IsZero, RCmp, MinMax, Flags);
+  return DAG.getSelect(DL, VT, IsZero, RetZero, MinMax, Flags);
 }
 
 /// Returns a true value if if this FPClassTest can be performed with an ordered
diff --git a/llvm/test/CodeGen/AMDGPU/fmax3-maximumnum.ll b/llvm/test/CodeGen/AMDGPU/fmax3-maximumnum.ll
index d458bb2492f23..02ea2cc2a1919 100644
--- a/llvm/test/CodeGen/AMDGPU/fmax3-maximumnum.ll
+++ b/llvm/test/CodeGen/AMDGPU/fmax3-maximumnum.ll
@@ -1696,23 +1696,13 @@ define bfloat @v_max3_bf16_maximumnum_maximumnum__v_v_v_0(bfloat %a, bfloat %b,
 ; GFX8-NEXT:    v_lshlrev_b32_e32 v4, 16, v0
 ; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v4, v3
 ; GFX8-NEXT:    v_cndmask_b32_e32 v3, v1, v0, vcc
-; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; GFX8-NEXT:    v_mul_f32_e32 v3, 1.0, v3
-; GFX8-NEXT:    v_bfe_u32 v4, v3, 16, 1
-; GFX8-NEXT:    v_add_u32_e32 v4, vcc, v4, v3
-; GFX8-NEXT:    s_movk_i32 s4, 0x7fff
-; GFX8-NEXT:    v_add_u32_e32 v4, vcc, s4, v4
-; GFX8-NEXT:    v_or_b32_e32 v5, 0x400000, v3
-; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
-; GFX8-NEXT:    v_cndmask_b32_e32 v3, v4, v5, vcc
-; GFX8-NEXT:    v_lshrrev_b32_e32 v4, 16, v3
 ; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
-; GFX8-NEXT:    v_cndmask_b32_e32 v0, v4, v0, vcc
+; GFX8-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
 ; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
 ; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc
-; GFX8-NEXT:    v_and_b32_e32 v1, 0xffff0000, v3
+; GFX8-NEXT:    v_lshlrev_b32_e32 v1, 16, v3
 ; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v1
-; GFX8-NEXT:    v_cndmask_b32_e32 v0, v4, v0, vcc
+; GFX8-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
 ; GFX8-NEXT:    v_lshlrev_b32_e32 v1, 16, v0
 ; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
 ; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
@@ -1722,22 +1712,14 @@ define bfloat @v_max3_bf16_maximumnum_maximumnum__v_v_v_0(bfloat %a, bfloat %b,
 ; GFX8-NEXT:    v_lshlrev_b32_e32 v1, 16, v0
 ; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
 ; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v1, v3
-; GFX8-NEXT:    v_cndmask_b32_sdwa v1, v2, v0, vcc dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX8-NEXT:    v_mul_f32_e32 v1, 1.0, v1
-; GFX8-NEXT:    v_bfe_u32 v3, v1, 16, 1
-; GFX8-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
-; GFX8-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
-; GFX8-NEXT:    v_or_b32_e32 v4, 0x400000, v1
-; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
-; GFX8-NEXT:    v_cndmask_b32_e32 v1, v3, v4, vcc
-; GFX8-NEXT:    v_lshrrev_b32_e32 v3, 16, v1
+; GFX8-NEXT:    v_cndmask_b32_e32 v1, v2, v0, vcc
 ; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
-; GFX8-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
+; GFX8-NEXT:    v_cndmask_b32_e32 v0, v1, v0, vcc
 ; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
-; GFX8-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
 ; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
-; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v1
-; GFX8-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
+; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
+; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v2
+; GFX8-NEXT:    v_cndmask_b32_e32 v0, v1, v0, vcc
 ; GFX8-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX900-SDAG-LABEL: v_max3_bf16_maximumnum_maximumnum__v_v_v_0:
@@ -1753,22 +1735,13 @@ define bfloat @v_max3_bf16_maximumnum_maximumnum__v_v_v_0(bfloat %a, bfloat %b,
 ; GFX900-SDAG-NEXT:    v_lshlrev_b32_e32 v4, 16, v1
 ; GFX900-SDAG-NEXT:    v_cmp_gt_f32_e32 vcc, v3, v4
 ; GFX900-SDAG-NEXT:    v_cndmask_b32_e32 v3, v1, v0, vcc
-; GFX900-SDAG-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; GFX900-SDAG-NEXT:    v_max_f32_e32 v3, v3, v3
-; GFX900-SDAG-NEXT:    v_bfe_u32 v4, v3, 16, 1
-; GFX900-SDAG-NEXT:    s_movk_i32 s4, 0x7fff
-; GFX900-SDAG-NEXT:    v_add3_u32 v4, v4, v3, s4
-; GFX900-SDAG-NEXT:    v_or_b32_e32 v5, 0x400000, v3
-; GFX900-SDAG-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
-; GFX900-SDAG-NEXT:    v_cndmask_b32_e32 v3, v4, v5, vcc
-; GFX900-SDAG-NEXT:    v_lshrrev_b32_e32 v4, 16, v3
 ; GFX900-SDAG-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
-; GFX900-SDAG-NEXT:    v_cndmask_b32_e32 v0, v4, v0, vcc
+; GFX900-SDAG-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
 ; GFX900-SDAG-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
 ; GFX900-SDAG-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc
-; GFX900-SDAG-NEXT:    v_and_b32_e32 v1, 0xffff0000, v3
+; GFX900-SDAG-NEXT:    v_lshlrev_b32_e32 v1, 16, v3
 ; GFX900-SDAG-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v1
-; GFX900-SDAG-NEXT:    v_cndmask_b32_e32 v0, v4, v0, vcc
+; GFX900-SDAG-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
 ; GFX900-SDAG-NEXT:    v_lshlrev_b32_e32 v1, 16, v0
 ; GFX900-SDAG-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
 ; GFX900-SDAG-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
@@ -1779,21 +1752,13 @@ define bfloat @v_max3_bf16_maximumnum_maximumnum__v_v_v_0(bfloat %a, bfloat %b,
 ; GFX900-SDAG-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
 ; GFX900-SDAG-NEXT:    v_cmp_gt_f32_e32 vcc, v1, v3
 ; GFX900-SDAG-NEXT:    v_cndmask_b32_e32 v1, v2, v0, vcc
-; GFX900-SDAG-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX900-SDAG-NEXT:    v_max_f32_e32 v1, v1, v1
-; GFX900-SDAG-NEXT:    v_bfe_u32 v3, v1, 16, 1
-; GFX900-SDAG-NEXT:    v_add3_u32 v3, v3, v1, s4
-; GFX900-SDAG-NEXT:    v_or_b32_e32 v4, 0x400000, v1
-; GFX900-SDAG-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
-; GFX900-SDAG-NEXT:    v_cndmask_b32_e32 v1, v3, v4, vcc
-; GFX900-SDAG-NEXT:    v_lshrrev_b32_e32 v3, 16, v1
 ; GFX900-SDAG-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
-; GFX900-SDAG-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
+; GFX900-SDAG-NEXT:    v_cndmask_b32_e32 v0, v1, v0, vcc
 ; GFX900-SDAG-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
-; GFX900-SDAG-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
 ; GFX900-SDAG-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
-; GFX900-SDAG-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v1
-; GFX900-SDAG-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
+; GFX900-SDAG-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
+; GFX900-SDAG-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v2
+; GFX900-SDAG-NEXT:    v_cndmask_b32_e32 v0, v1, v0, vcc
 ; GFX900-SDAG-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX950-SDAG-LABEL: v_max3_bf16_maximumnum_maximumnum__v_v_v_0:
@@ -1812,9 +1777,6 @@ define bfloat @v_max3_bf16_maximumnum_maximumnum__v_v_v_0(bfloat %a, bfloat %b,
 ; GFX950-SDAG-NEXT:    v_cmp_gt_f32_e32 vcc, v3, v4
 ; GFX950-SDAG-NEXT:    s_nop 1
 ; GFX950-SDAG-NEXT:    v_cndmask_b32_e32 v3, v1, v0, vcc
-; GFX950-SDAG-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; GFX950-SDAG-NEXT:    v_max_f32_e32 v3, v3, v3
-; GFX950-SDAG-NEXT:    v_cvt_pk_bf16_f32 v3, v3, s0
 ; GFX950-SDAG-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
 ; GFX950-SDAG-NEXT:    s_nop 1
 ; GFX950-SDAG-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
@@ -1838,9 +1800,6 @@ define bfloat @v_max3_bf16_maximumnum_maximumnum__v_v_v_0(bfloat %a, bfloat %b,
 ; GFX950-SDAG-NEXT:    v_cmp_gt_f32_e32 vcc, v1, v3
 ; GFX950-SDAG-NEXT:    s_nop 1
 ; GFX950-SDAG-NEXT:    v_cndmask_b32_e32 v1, v2, v0, vcc
-; GFX950-SDAG-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX950-SDAG-NEXT:    v_max_f32_e32 v1, v1, v1
-; GFX950-SDAG-NEXT:    v_cvt_pk_bf16_f32 v1, v1, s0
 ; GFX950-SDAG-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
 ; GFX950-SDAG-NEXT:    s_nop 1
 ; GFX950-SDAG-NEXT:    v_cndmask_b32_e32 v0, v1, v0, vcc
@@ -1866,22 +1825,14 @@ define bfloat @v_max3_bf16_maximumnum_maximumnum__v_v_v_0(bfloat %a, bfloat %b,
 ; GFX10-NEXT:    v_lshlrev_b32_e32 v4, 16, v1
 ; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v3, v4
 ; GFX10-NEXT:    v_cndmask_b32_e32 v3, v1, v0, vcc_lo
-; GFX10-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; GFX10-NEXT:    v_max_f32_e32 v3, v3, v3
-; GFX10-NEXT:    v_bfe_u32 v4, v3, 16, 1
-; GFX10-NEXT:    v_or_b32_e32 v5, 0x400000, v3
-; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX10-NEXT:    v_add3_u32 v4, v4, v3, 0x7fff
-; GFX10-NEXT:    v_cndmask_b32_e32 v3, v4, v5, vcc_lo
 ; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
-; GFX10-NEXT:    v_lshrrev_b32_e32 v4, 16, v3
-; GFX10-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
-; GFX10-NEXT:    v_cndmask_b32_e32 v0, v4, v0, vcc_lo
+; GFX10-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc_lo
 ; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
 ; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
-; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v3
+; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v4
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc_lo
 ; GFX10-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
-; GFX10-NEXT:    v_cndmask_b32_e32 v0, v4, v0, vcc_lo
 ; GFX10-NEXT:    v_lshlrev_b32_e32 v1, 16, v0
 ; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
 ; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
@@ -1891,21 +1842,13 @@ define bfloat @v_max3_bf16_maximumnum_maximumnum__v_v_v_0(bfloat %a, bfloat %b,
 ; GFX10-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
 ; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v2, v3
 ; GFX10-NEXT:    v_cndmask_b32_e32 v2, v1, v0, vcc_lo
-; GFX10-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX10-NEXT:    v_max_f32_e32 v2, v2, v2
-; GFX10-NEXT:    v_bfe_u32 v3, v2, 16, 1
-; GFX10-NEXT:    v_or_b32_e32 v4, 0x400000, v2
-; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX10-NEXT:    v_add3_u32 v3, v3, v2, 0x7fff
-; GFX10-NEXT:    v_cndmask_b32_e32 v2, v3, v4, vcc_lo
 ; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
-; GFX10-NEXT:    v_lshrrev_b32_e32 v3, 16, v2
-; GFX10-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
-; GFX10-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc_lo
+; GFX10-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v2, v0, vcc_lo
 ; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
 ; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
-; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v2
-; GFX10-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc_lo
+; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v3
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v2, v0, vcc_lo
 ; GFX10-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX11-LABEL: v_max3_bf16_maximumnum_maximumnum__v_v_v_0:
@@ -1921,61 +1864,38 @@ define bfloat @v_max3_bf16_maximumnum_maximumnum__v_v_v_0(bfloat %a, bfloat %b,
 ; GFX11-NEXT:    v_cndmask_b32_e32 v1, v1, v0, vcc_lo
 ; GFX11-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
 ; GFX11-NEXT:    v_lshlrev_b32_e32 v4, 16, v1
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v3, v4
 ; GFX11-NEXT:    v_cndmask_b32_e32 v3, v1, v0, vcc_lo
-; GFX11-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT:    v_max_f32_e32 v3, v3, v3
-; GFX11-NEXT:    v_bfe_u32 v4, v3, 16, 1
-; GFX11-NEXT:    v_or_b32_e32 v5, 0x400000, v3
-; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT:    v_add3_u32 v4, v4, v3, 0x7fff
-; GFX11-NEXT:    v_cndmask_b32_e32 v3, v4, v5, vcc_lo
 ; GFX11-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT:    v_lshrrev_b32_e32 v4, 16, v3
-; GFX11-NEXT:    v_dual_cndmask_b32 v0, v4, v0 :: v_dual_and_b32 v3, 0xffff0000, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
+; GFX11-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc_lo
 ; GFX11-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_4)
 ; GFX11-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
-; GFX11-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v3
+; GFX11-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v4
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT:    v_cndmask_b32_e32 v0, v4, v0, vcc_lo
+; GFX11-NEXT:    v_dual_cndmask_b32 v0, v3, v0 :: v_dual_lshlrev_b32 v3, 16, v2
 ; GFX11-NEXT:    v_lshlrev_b32_e32 v1, 16, v0
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_4)
 ; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
 ; GFX11-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
-; GFX11-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
 ; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
 ; GFX11-NEXT:    v_cndmask_b32_e32 v1, v2, v0, vcc_lo
 ; GFX11-NEXT:    v_lshlrev_b32_e32 v2, 16, v0
 ; GFX11-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v2, v3
 ; GFX11-NEXT:    v_cndmask_b32_e32 v2, v1, v0, vcc_lo
-; GFX11-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT:    v_max_f32_e32 v2, v2, v2
-; GFX11-NEXT:    v_bfe_u32 v3, v2, 16, 1
-; GFX11-NEXT:    v_or_b32_e32 v4, 0x400000, v2
-; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT:    v_add3_u32 v3, v3, v2, 0x7fff
-; GFX11-NEXT:    v_cndmask_b32_e32 v2, v3, v4, vcc_lo
 ; GFX11-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT:    v_lshrrev_b32_e32 v3, 16, v2
-; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
-; GFX11-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc_lo
+; GFX11-NEXT:    v_dual_cndmask_b32 v0, v2, v0 :: v_dual_lshlrev_b32 v3, 16, v2
 ; GFX11-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
 ; GFX11-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
-; GFX11-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v2
+; GFX11-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v3
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2)
-; GFX11-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc_lo
+; GFX11-NEXT:    v_cndmask_b32_e32 v0, v2, v0, vcc_lo
 ; GFX11-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX12-LABEL: v_max3_bf16_maximumnum_maximumnum__v_v_v_0:
@@ -1997,72 +1917,47 @@ define bfloat @v_max3_bf16_maximumnum_maximumnum__v_v_v_0(bfloat %a, bfloat %b,
 ; GFX12-NEXT:    v_cndmask_b32_e32 v1, v1, v0, vcc_lo
 ; GFX12-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
 ; GFX12-NEXT:    v_lshlrev_b32_e32 v4, 16, v1
-; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2)
 ; GFX12-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v3, v4
 ; GFX12-NEXT:    s_wait_alu 0xfffd
 ; GFX12-NEXT:    v_cndmask_b32_e32 v3, v1, v0, vcc_lo
-; GFX12-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-NEXT:    v_max_num_f32_e32 v3, v3, v3
-; GFX12-NEXT:    v_bfe_u32 v4, v3, 16, 1
-; GFX12-NEXT:    v_or_b32_e32 v5, 0x400000, v3
-; GFX12-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX12-NEXT:    v_add3_u32 v4, v4, v3, 0x7fff
-; GFX12-NEXT:    s_wait_alu 0xfffd
-; GFX12-NEXT:    v_cndmask_b32_e32 v3, v4, v5, vcc_lo
 ; GFX12-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
-; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX12-NEXT:    v_lshrrev_b32_e32 v4, 16, v3
+; GFX12-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
 ; GFX12-NEXT:    s_wait_alu 0xfffd
-; GFX12-NEXT:    v_dual_cndmask_b32 v0, v4, v0 :: v_dual_and_b32 v3, 0xffff0000, v3
+; GFX12-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc_lo
 ; GFX12-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
 ; GFX12-NEXT:    s_wait_alu 0xfffd
-; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX12-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
-; GFX12-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v3
+; GFX12-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v4
 ; GFX12-NEXT:    s_wait_alu 0xfffd
-; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-NEXT:    v_cndmask_b32_e32 v0, v4, v0, vcc_lo
+; GFX12-NEXT:    v_dual_cndmask_b32 v0, v3, v0 :: v_dual_lshlrev_b32 v3, 16, v2
+; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX12-NEXT:    v_lshlrev_b32_e32 v1, 16, v0
-; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
 ; GFX12-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
 ; GFX12-NEXT:    s_wait_alu 0xfffd
 ; GFX12-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
-; GFX12-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
 ; GFX12-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
 ; GFX12-NEXT:    s_wait_alu 0xfffd
-; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
 ; GFX12-NEXT:    v_cndmask_b32_e32 v1, v2, v0, vcc_lo
 ; GFX12-NEXT:    v_lshlrev_b32_e32 v2, 16, v0
 ; GFX12-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
-; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(VALU_DEP_2)
 ; GFX12-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v2, v3
 ; GFX12-NEXT:    s_wait_alu 0xfffd
 ; GFX12-NEXT:    v_cndmask_b32_e32 v2, v1, v0, vcc_lo
-; GFX12-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-NEXT:    v_max_num_f32_e32 v2, v2, v2
-; GFX12-NEXT:    v_bfe_u32 v3, v2, 16, 1
-; GFX12-NEXT:    v_or_b32_e32 v4, 0x400000, v2
-; GFX12-NEXT: ...
[truncated]

@llvmbot
Copy link
Member

llvmbot commented May 16, 2025

@llvm/pr-subscribers-llvm-selectiondag

Author: YunQiang Su (wzssyqa)

Changes

  1. On GPR32 platform, expandIS_FPCLASS may fail due to ISD::BITCAST
    double to int64 may fail. Let's FP_ROUND double to float first.
    Since we use it if MinMax is zero only, so the flushing won't
    break anything.

  2. Only one IS_FPCLASS is needed. MinMax will always be RHS if equal.
    So we can select between LHS and MinMax.
    It will even safe if FP_ROUND flush a small LHS, as if LHS is not zero
    then, MinMax won't be Zero, so we will always use MinMax.

Patch is 2.87 MiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/140193.diff

7 Files Affected:

  • (modified) llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp (+21-11)
  • (modified) llvm/test/CodeGen/AMDGPU/fmax3-maximumnum.ll (+193-493)
  • (modified) llvm/test/CodeGen/AMDGPU/fmin3-minimumnum.ll (+217-518)
  • (modified) llvm/test/CodeGen/AMDGPU/maximumnum.bf16.ll (+7105-12288)
  • (modified) llvm/test/CodeGen/AMDGPU/minimumnum.bf16.ll (+7214-12382)
  • (modified) llvm/test/CodeGen/Mips/fp-maximumnum-minimumnum.ll (+880)
  • (modified) llvm/test/CodeGen/X86/fminimumnum-fmaximumnum.ll (+116-162)
diff --git a/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp b/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
index f17d6a2787889..5166bc0ecd3b8 100644
--- a/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
@@ -8681,13 +8681,9 @@ SDValue TargetLowering::expandFMINIMUMNUM_FMAXIMUMNUM(SDNode *Node,
     RHS = DAG.getSelectCC(DL, RHS, RHS, LHS, RHS, ISD::SETUO);
   }
 
+  // Please always prefer RHS if equal.
   SDValue MinMax =
       DAG.getSelectCC(DL, LHS, RHS, LHS, RHS, IsMax ? ISD::SETGT : ISD::SETLT);
-  // If MinMax is NaN, let's quiet it.
-  if (!Flags.hasNoNaNs() && !DAG.isKnownNeverNaN(LHS) &&
-      !DAG.isKnownNeverNaN(RHS)) {
-    MinMax = DAG.getNode(ISD::FCANONICALIZE, DL, VT, MinMax, Flags);
-  }
 
   // Fixup signed zero behavior.
   if (Options.NoSignedZerosFPMath || Flags.hasNoSignedZeros() ||
@@ -8698,13 +8694,27 @@ SDValue TargetLowering::expandFMINIMUMNUM_FMAXIMUMNUM(SDNode *Node,
       DAG.getTargetConstant(IsMax ? fcPosZero : fcNegZero, DL, MVT::i32);
   SDValue IsZero = DAG.getSetCC(DL, CCVT, MinMax,
                                 DAG.getConstantFP(0.0, DL, VT), ISD::SETEQ);
-  SDValue LCmp = DAG.getSelect(
-      DL, VT, DAG.getNode(ISD::IS_FPCLASS, DL, CCVT, LHS, TestZero), LHS,
+  unsigned BitSize = VT.getScalarSizeInBits();
+  EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), BitSize);
+  EVT FloatVT = EVT::getFloatingPointVT(32);
+  if (VT.isVector()) {
+    IntVT =
+        EVT::getVectorVT(*DAG.getContext(), IntVT, VT.getVectorElementCount());
+    FloatVT = EVT::getVectorVT(*DAG.getContext(), FloatVT,
+                               VT.getVectorElementCount());
+  }
+  SDValue LHSTrunc = LHS;
+  if (!isOperationLegal(ISD::BITCAST, IntVT) &&
+      !isOperationLegal(ISD::IS_FPCLASS, VT)) {
+    LHSTrunc = DAG.getNode(ISD::FP_ROUND, DL, FloatVT, LHS,
+                           DAG.getIntPtrConstant(0, DL, /*isTarget=*/true));
+  }
+  // It's OK to select from LHS and MinMax, with only one ISD::IS_FPCLASS, as
+  // we preferred RHS when generate MinMax, if the operands are equal.
+  SDValue RetZero = DAG.getSelect(
+      DL, VT, DAG.getNode(ISD::IS_FPCLASS, DL, CCVT, LHSTrunc, TestZero), LHS,
       MinMax, Flags);
-  SDValue RCmp = DAG.getSelect(
-      DL, VT, DAG.getNode(ISD::IS_FPCLASS, DL, CCVT, RHS, TestZero), RHS, LCmp,
-      Flags);
-  return DAG.getSelect(DL, VT, IsZero, RCmp, MinMax, Flags);
+  return DAG.getSelect(DL, VT, IsZero, RetZero, MinMax, Flags);
 }
 
 /// Returns a true value if if this FPClassTest can be performed with an ordered
diff --git a/llvm/test/CodeGen/AMDGPU/fmax3-maximumnum.ll b/llvm/test/CodeGen/AMDGPU/fmax3-maximumnum.ll
index d458bb2492f23..02ea2cc2a1919 100644
--- a/llvm/test/CodeGen/AMDGPU/fmax3-maximumnum.ll
+++ b/llvm/test/CodeGen/AMDGPU/fmax3-maximumnum.ll
@@ -1696,23 +1696,13 @@ define bfloat @v_max3_bf16_maximumnum_maximumnum__v_v_v_0(bfloat %a, bfloat %b,
 ; GFX8-NEXT:    v_lshlrev_b32_e32 v4, 16, v0
 ; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v4, v3
 ; GFX8-NEXT:    v_cndmask_b32_e32 v3, v1, v0, vcc
-; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; GFX8-NEXT:    v_mul_f32_e32 v3, 1.0, v3
-; GFX8-NEXT:    v_bfe_u32 v4, v3, 16, 1
-; GFX8-NEXT:    v_add_u32_e32 v4, vcc, v4, v3
-; GFX8-NEXT:    s_movk_i32 s4, 0x7fff
-; GFX8-NEXT:    v_add_u32_e32 v4, vcc, s4, v4
-; GFX8-NEXT:    v_or_b32_e32 v5, 0x400000, v3
-; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
-; GFX8-NEXT:    v_cndmask_b32_e32 v3, v4, v5, vcc
-; GFX8-NEXT:    v_lshrrev_b32_e32 v4, 16, v3
 ; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
-; GFX8-NEXT:    v_cndmask_b32_e32 v0, v4, v0, vcc
+; GFX8-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
 ; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
 ; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc
-; GFX8-NEXT:    v_and_b32_e32 v1, 0xffff0000, v3
+; GFX8-NEXT:    v_lshlrev_b32_e32 v1, 16, v3
 ; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v1
-; GFX8-NEXT:    v_cndmask_b32_e32 v0, v4, v0, vcc
+; GFX8-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
 ; GFX8-NEXT:    v_lshlrev_b32_e32 v1, 16, v0
 ; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
 ; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
@@ -1722,22 +1712,14 @@ define bfloat @v_max3_bf16_maximumnum_maximumnum__v_v_v_0(bfloat %a, bfloat %b,
 ; GFX8-NEXT:    v_lshlrev_b32_e32 v1, 16, v0
 ; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
 ; GFX8-NEXT:    v_cmp_gt_f32_e32 vcc, v1, v3
-; GFX8-NEXT:    v_cndmask_b32_sdwa v1, v2, v0, vcc dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
-; GFX8-NEXT:    v_mul_f32_e32 v1, 1.0, v1
-; GFX8-NEXT:    v_bfe_u32 v3, v1, 16, 1
-; GFX8-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
-; GFX8-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
-; GFX8-NEXT:    v_or_b32_e32 v4, 0x400000, v1
-; GFX8-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
-; GFX8-NEXT:    v_cndmask_b32_e32 v1, v3, v4, vcc
-; GFX8-NEXT:    v_lshrrev_b32_e32 v3, 16, v1
+; GFX8-NEXT:    v_cndmask_b32_e32 v1, v2, v0, vcc
 ; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
-; GFX8-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
+; GFX8-NEXT:    v_cndmask_b32_e32 v0, v1, v0, vcc
 ; GFX8-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
-; GFX8-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
 ; GFX8-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
-; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v1
-; GFX8-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
+; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
+; GFX8-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v2
+; GFX8-NEXT:    v_cndmask_b32_e32 v0, v1, v0, vcc
 ; GFX8-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX900-SDAG-LABEL: v_max3_bf16_maximumnum_maximumnum__v_v_v_0:
@@ -1753,22 +1735,13 @@ define bfloat @v_max3_bf16_maximumnum_maximumnum__v_v_v_0(bfloat %a, bfloat %b,
 ; GFX900-SDAG-NEXT:    v_lshlrev_b32_e32 v4, 16, v1
 ; GFX900-SDAG-NEXT:    v_cmp_gt_f32_e32 vcc, v3, v4
 ; GFX900-SDAG-NEXT:    v_cndmask_b32_e32 v3, v1, v0, vcc
-; GFX900-SDAG-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; GFX900-SDAG-NEXT:    v_max_f32_e32 v3, v3, v3
-; GFX900-SDAG-NEXT:    v_bfe_u32 v4, v3, 16, 1
-; GFX900-SDAG-NEXT:    s_movk_i32 s4, 0x7fff
-; GFX900-SDAG-NEXT:    v_add3_u32 v4, v4, v3, s4
-; GFX900-SDAG-NEXT:    v_or_b32_e32 v5, 0x400000, v3
-; GFX900-SDAG-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
-; GFX900-SDAG-NEXT:    v_cndmask_b32_e32 v3, v4, v5, vcc
-; GFX900-SDAG-NEXT:    v_lshrrev_b32_e32 v4, 16, v3
 ; GFX900-SDAG-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
-; GFX900-SDAG-NEXT:    v_cndmask_b32_e32 v0, v4, v0, vcc
+; GFX900-SDAG-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
 ; GFX900-SDAG-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v1
 ; GFX900-SDAG-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc
-; GFX900-SDAG-NEXT:    v_and_b32_e32 v1, 0xffff0000, v3
+; GFX900-SDAG-NEXT:    v_lshlrev_b32_e32 v1, 16, v3
 ; GFX900-SDAG-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v1
-; GFX900-SDAG-NEXT:    v_cndmask_b32_e32 v0, v4, v0, vcc
+; GFX900-SDAG-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
 ; GFX900-SDAG-NEXT:    v_lshlrev_b32_e32 v1, 16, v0
 ; GFX900-SDAG-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
 ; GFX900-SDAG-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
@@ -1779,21 +1752,13 @@ define bfloat @v_max3_bf16_maximumnum_maximumnum__v_v_v_0(bfloat %a, bfloat %b,
 ; GFX900-SDAG-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
 ; GFX900-SDAG-NEXT:    v_cmp_gt_f32_e32 vcc, v1, v3
 ; GFX900-SDAG-NEXT:    v_cndmask_b32_e32 v1, v2, v0, vcc
-; GFX900-SDAG-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX900-SDAG-NEXT:    v_max_f32_e32 v1, v1, v1
-; GFX900-SDAG-NEXT:    v_bfe_u32 v3, v1, 16, 1
-; GFX900-SDAG-NEXT:    v_add3_u32 v3, v3, v1, s4
-; GFX900-SDAG-NEXT:    v_or_b32_e32 v4, 0x400000, v1
-; GFX900-SDAG-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
-; GFX900-SDAG-NEXT:    v_cndmask_b32_e32 v1, v3, v4, vcc
-; GFX900-SDAG-NEXT:    v_lshrrev_b32_e32 v3, 16, v1
 ; GFX900-SDAG-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
-; GFX900-SDAG-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
+; GFX900-SDAG-NEXT:    v_cndmask_b32_e32 v0, v1, v0, vcc
 ; GFX900-SDAG-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v2
-; GFX900-SDAG-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
 ; GFX900-SDAG-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc
-; GFX900-SDAG-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v1
-; GFX900-SDAG-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
+; GFX900-SDAG-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
+; GFX900-SDAG-NEXT:    v_cmp_eq_f32_e32 vcc, 0, v2
+; GFX900-SDAG-NEXT:    v_cndmask_b32_e32 v0, v1, v0, vcc
 ; GFX900-SDAG-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX950-SDAG-LABEL: v_max3_bf16_maximumnum_maximumnum__v_v_v_0:
@@ -1812,9 +1777,6 @@ define bfloat @v_max3_bf16_maximumnum_maximumnum__v_v_v_0(bfloat %a, bfloat %b,
 ; GFX950-SDAG-NEXT:    v_cmp_gt_f32_e32 vcc, v3, v4
 ; GFX950-SDAG-NEXT:    s_nop 1
 ; GFX950-SDAG-NEXT:    v_cndmask_b32_e32 v3, v1, v0, vcc
-; GFX950-SDAG-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; GFX950-SDAG-NEXT:    v_max_f32_e32 v3, v3, v3
-; GFX950-SDAG-NEXT:    v_cvt_pk_bf16_f32 v3, v3, s0
 ; GFX950-SDAG-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
 ; GFX950-SDAG-NEXT:    s_nop 1
 ; GFX950-SDAG-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc
@@ -1838,9 +1800,6 @@ define bfloat @v_max3_bf16_maximumnum_maximumnum__v_v_v_0(bfloat %a, bfloat %b,
 ; GFX950-SDAG-NEXT:    v_cmp_gt_f32_e32 vcc, v1, v3
 ; GFX950-SDAG-NEXT:    s_nop 1
 ; GFX950-SDAG-NEXT:    v_cndmask_b32_e32 v1, v2, v0, vcc
-; GFX950-SDAG-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX950-SDAG-NEXT:    v_max_f32_e32 v1, v1, v1
-; GFX950-SDAG-NEXT:    v_cvt_pk_bf16_f32 v1, v1, s0
 ; GFX950-SDAG-NEXT:    v_cmp_eq_u16_e32 vcc, 0, v0
 ; GFX950-SDAG-NEXT:    s_nop 1
 ; GFX950-SDAG-NEXT:    v_cndmask_b32_e32 v0, v1, v0, vcc
@@ -1866,22 +1825,14 @@ define bfloat @v_max3_bf16_maximumnum_maximumnum__v_v_v_0(bfloat %a, bfloat %b,
 ; GFX10-NEXT:    v_lshlrev_b32_e32 v4, 16, v1
 ; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v3, v4
 ; GFX10-NEXT:    v_cndmask_b32_e32 v3, v1, v0, vcc_lo
-; GFX10-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; GFX10-NEXT:    v_max_f32_e32 v3, v3, v3
-; GFX10-NEXT:    v_bfe_u32 v4, v3, 16, 1
-; GFX10-NEXT:    v_or_b32_e32 v5, 0x400000, v3
-; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX10-NEXT:    v_add3_u32 v4, v4, v3, 0x7fff
-; GFX10-NEXT:    v_cndmask_b32_e32 v3, v4, v5, vcc_lo
 ; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
-; GFX10-NEXT:    v_lshrrev_b32_e32 v4, 16, v3
-; GFX10-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
-; GFX10-NEXT:    v_cndmask_b32_e32 v0, v4, v0, vcc_lo
+; GFX10-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc_lo
 ; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
 ; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
-; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v3
+; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v4
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc_lo
 ; GFX10-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
-; GFX10-NEXT:    v_cndmask_b32_e32 v0, v4, v0, vcc_lo
 ; GFX10-NEXT:    v_lshlrev_b32_e32 v1, 16, v0
 ; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
 ; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
@@ -1891,21 +1842,13 @@ define bfloat @v_max3_bf16_maximumnum_maximumnum__v_v_v_0(bfloat %a, bfloat %b,
 ; GFX10-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
 ; GFX10-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v2, v3
 ; GFX10-NEXT:    v_cndmask_b32_e32 v2, v1, v0, vcc_lo
-; GFX10-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX10-NEXT:    v_max_f32_e32 v2, v2, v2
-; GFX10-NEXT:    v_bfe_u32 v3, v2, 16, 1
-; GFX10-NEXT:    v_or_b32_e32 v4, 0x400000, v2
-; GFX10-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX10-NEXT:    v_add3_u32 v3, v3, v2, 0x7fff
-; GFX10-NEXT:    v_cndmask_b32_e32 v2, v3, v4, vcc_lo
 ; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
-; GFX10-NEXT:    v_lshrrev_b32_e32 v3, 16, v2
-; GFX10-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
-; GFX10-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc_lo
+; GFX10-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v2, v0, vcc_lo
 ; GFX10-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
 ; GFX10-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
-; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v2
-; GFX10-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc_lo
+; GFX10-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v3
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v2, v0, vcc_lo
 ; GFX10-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX11-LABEL: v_max3_bf16_maximumnum_maximumnum__v_v_v_0:
@@ -1921,61 +1864,38 @@ define bfloat @v_max3_bf16_maximumnum_maximumnum__v_v_v_0(bfloat %a, bfloat %b,
 ; GFX11-NEXT:    v_cndmask_b32_e32 v1, v1, v0, vcc_lo
 ; GFX11-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
 ; GFX11-NEXT:    v_lshlrev_b32_e32 v4, 16, v1
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v3, v4
 ; GFX11-NEXT:    v_cndmask_b32_e32 v3, v1, v0, vcc_lo
-; GFX11-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT:    v_max_f32_e32 v3, v3, v3
-; GFX11-NEXT:    v_bfe_u32 v4, v3, 16, 1
-; GFX11-NEXT:    v_or_b32_e32 v5, 0x400000, v3
-; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT:    v_add3_u32 v4, v4, v3, 0x7fff
-; GFX11-NEXT:    v_cndmask_b32_e32 v3, v4, v5, vcc_lo
 ; GFX11-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT:    v_lshrrev_b32_e32 v4, 16, v3
-; GFX11-NEXT:    v_dual_cndmask_b32 v0, v4, v0 :: v_dual_and_b32 v3, 0xffff0000, v3
+; GFX11-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
+; GFX11-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc_lo
 ; GFX11-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_4)
 ; GFX11-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
-; GFX11-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v3
+; GFX11-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v4
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT:    v_cndmask_b32_e32 v0, v4, v0, vcc_lo
+; GFX11-NEXT:    v_dual_cndmask_b32 v0, v3, v0 :: v_dual_lshlrev_b32 v3, 16, v2
 ; GFX11-NEXT:    v_lshlrev_b32_e32 v1, 16, v0
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_4)
 ; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
 ; GFX11-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
-; GFX11-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
 ; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
 ; GFX11-NEXT:    v_cndmask_b32_e32 v1, v2, v0, vcc_lo
 ; GFX11-NEXT:    v_lshlrev_b32_e32 v2, 16, v0
 ; GFX11-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX11-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v2, v3
 ; GFX11-NEXT:    v_cndmask_b32_e32 v2, v1, v0, vcc_lo
-; GFX11-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT:    v_max_f32_e32 v2, v2, v2
-; GFX11-NEXT:    v_bfe_u32 v3, v2, 16, 1
-; GFX11-NEXT:    v_or_b32_e32 v4, 0x400000, v2
-; GFX11-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT:    v_add3_u32 v3, v3, v2, 0x7fff
-; GFX11-NEXT:    v_cndmask_b32_e32 v2, v3, v4, vcc_lo
 ; GFX11-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-NEXT:    v_lshrrev_b32_e32 v3, 16, v2
-; GFX11-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
-; GFX11-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc_lo
+; GFX11-NEXT:    v_dual_cndmask_b32 v0, v2, v0 :: v_dual_lshlrev_b32 v3, 16, v2
 ; GFX11-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
 ; GFX11-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
-; GFX11-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v2
+; GFX11-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v3
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2)
-; GFX11-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc_lo
+; GFX11-NEXT:    v_cndmask_b32_e32 v0, v2, v0, vcc_lo
 ; GFX11-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX12-LABEL: v_max3_bf16_maximumnum_maximumnum__v_v_v_0:
@@ -1997,72 +1917,47 @@ define bfloat @v_max3_bf16_maximumnum_maximumnum__v_v_v_0(bfloat %a, bfloat %b,
 ; GFX12-NEXT:    v_cndmask_b32_e32 v1, v1, v0, vcc_lo
 ; GFX12-NEXT:    v_lshlrev_b32_e32 v3, 16, v0
 ; GFX12-NEXT:    v_lshlrev_b32_e32 v4, 16, v1
-; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2)
 ; GFX12-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v3, v4
 ; GFX12-NEXT:    s_wait_alu 0xfffd
 ; GFX12-NEXT:    v_cndmask_b32_e32 v3, v1, v0, vcc_lo
-; GFX12-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
-; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-NEXT:    v_max_num_f32_e32 v3, v3, v3
-; GFX12-NEXT:    v_bfe_u32 v4, v3, 16, 1
-; GFX12-NEXT:    v_or_b32_e32 v5, 0x400000, v3
-; GFX12-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX12-NEXT:    v_add3_u32 v4, v4, v3, 0x7fff
-; GFX12-NEXT:    s_wait_alu 0xfffd
-; GFX12-NEXT:    v_cndmask_b32_e32 v3, v4, v5, vcc_lo
 ; GFX12-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v0
-; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX12-NEXT:    v_lshrrev_b32_e32 v4, 16, v3
+; GFX12-NEXT:    v_lshlrev_b32_e32 v4, 16, v3
 ; GFX12-NEXT:    s_wait_alu 0xfffd
-; GFX12-NEXT:    v_dual_cndmask_b32 v0, v4, v0 :: v_dual_and_b32 v3, 0xffff0000, v3
+; GFX12-NEXT:    v_cndmask_b32_e32 v0, v3, v0, vcc_lo
 ; GFX12-NEXT:    v_cmp_eq_u16_e32 vcc_lo, 0, v1
 ; GFX12-NEXT:    s_wait_alu 0xfffd
-; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
 ; GFX12-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
-; GFX12-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v3
+; GFX12-NEXT:    v_cmp_eq_f32_e32 vcc_lo, 0, v4
 ; GFX12-NEXT:    s_wait_alu 0xfffd
-; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-NEXT:    v_cndmask_b32_e32 v0, v4, v0, vcc_lo
+; GFX12-NEXT:    v_dual_cndmask_b32 v0, v3, v0 :: v_dual_lshlrev_b32 v3, 16, v2
+; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX12-NEXT:    v_lshlrev_b32_e32 v1, 16, v0
-; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
 ; GFX12-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
 ; GFX12-NEXT:    s_wait_alu 0xfffd
 ; GFX12-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
-; GFX12-NEXT:    v_lshlrev_b32_e32 v3, 16, v2
 ; GFX12-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
 ; GFX12-NEXT:    s_wait_alu 0xfffd
-; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
 ; GFX12-NEXT:    v_cndmask_b32_e32 v1, v2, v0, vcc_lo
 ; GFX12-NEXT:    v_lshlrev_b32_e32 v2, 16, v0
 ; GFX12-NEXT:    v_lshlrev_b32_e32 v3, 16, v1
-; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(VALU_DEP_2)
 ; GFX12-NEXT:    v_cmp_gt_f32_e32 vcc_lo, v2, v3
 ; GFX12-NEXT:    s_wait_alu 0xfffd
 ; GFX12-NEXT:    v_cndmask_b32_e32 v2, v1, v0, vcc_lo
-; GFX12-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
-; GFX12-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX12-NEXT:    v_max_num_f32_e32 v2, v2, v2
-; GFX12-NEXT:    v_bfe_u32 v3, v2, 16, 1
-; GFX12-NEXT:    v_or_b32_e32 v4, 0x400000, v2
-; GFX12-NEXT: ...
[truncated]

@wzssyqa wzssyqa requested a review from arsenm May 16, 2025 05:40
@wzssyqa wzssyqa mentioned this pull request Jun 4, 2025
Merged
@wzssyqa
expandFMINIMUMNUM_FMAXIMUMNUM: Improve compare between zeros
cd71163 
1. On GPR32 platform, expandIS_FPCLASS may fail due to ISD::BITCAST
double to int64 may fail. Let's FP_ROUND double to float first.
Since we use it if MinMax is zero only, so the flushing won't
break anything.

2. Only one IS_FPCLASS is needed. MinMax will always be RHS if equal.
So we can select between LHS and MinMax.
It will even safe if FP_ROUND flush a small LHS, as if LHS is not zero
then, MinMax won't be Zero, so we will always use MinMax.
@wzssyqa wzssyqa force-pushed the fminimumnum-32bit branch from b035f20 to cd71163 Compare June 4, 2025 01:19
@wzssyqa wzssyqa requested a review from nikic June 4, 2025 01:20
@nikic nikic mentioned this pull request Jun 4, 2025
Closed
@wzssyqa
Copy link
Contributor Author

wzssyqa commented Jun 9, 2025

@nikic any idea about how to sync the FMAXIMUMNUM and FMAXIMUM?
I prefer use FP_ROUND for FMAXIMUMNUM first and then sync to FMAXIMUM.

FP_ROUND seems performing better since it doesn't need stack operations.

nikic
nikic reviewed Jun 9, 2025
View reviewed changes
llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
EVT FloatVT = EVT::getFloatingPointVT(32);
if (VT.isVector()) {
IntVT =
EVT::getVectorVT(*DAG.getContext(), IntVT, VT.getVectorElementCount());
Copy link
Contributor

Choose a reason for hiding this comment

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

You can directly use VT.changeTypeToInteger() to get the corresponding integer type with correct vector handling.

llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
IntVT =
EVT::getVectorVT(*DAG.getContext(), IntVT, VT.getVectorElementCount());
FloatVT = EVT::getVectorVT(*DAG.getContext(), FloatVT,
VT.getVectorElementCount());
Copy link
Contributor

Choose a reason for hiding this comment

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

You can use VT.changeElementType(FloatVT) here.

llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
VT.getVectorElementCount());
}
SDValue LHSTrunc = LHS;
if (!isOperationLegal(ISD::BITCAST, IntVT) &&
Copy link
Contributor

Choose a reason for hiding this comment

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

Suggested change
if (!isOperationLegal(ISD::BITCAST, IntVT) &&
if (!isTypeLegal(IntVT) &&

Would make more sense here, I think?

llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
if (!isOperationLegal(ISD::BITCAST, IntVT) &&
!isOperationLegal(ISD::IS_FPCLASS, VT)) {
LHSTrunc = DAG.getNode(ISD::FP_ROUND, DL, FloatVT, LHS,
DAG.getIntPtrConstant(0, DL, /*isTarget=*/true));
Copy link
Contributor

Choose a reason for hiding this comment

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

So this assumes that for all cases where the FP type is legal and the int type is not, that float is a legal type and that it either has legal IS_FPCLASS or i32 is also legal. Do these assumptions hold in practice? I think it's reasonable to expect float to be a legal type if there are any legal FP types, but I'm less sure about the rest. I think we have targets where f32 is legal but i32 is not.

Could you please add tests for #139380, #139381 and #140445 (just for minimumnum instead of minimum)? I think all of those are going to be fine with this patch, but I'd like to be sure.

llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp
DAG.getIntPtrConstant(0, DL, /*isTarget=*/true));
}
// It's OK to select from LHS and MinMax, with only one ISD::IS_FPCLASS, as
// we preferred RHS when generate MinMax, if the operands are equal.
Copy link
Contributor

Choose a reason for hiding this comment

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

I think this is correct for how these opcodes are legalized, but wouldn't transfer to the maximum/minimum case with the current implementation. I guess the idea is that for the legalizations through other min/max opcodes we shouldn't need the signed zero handling at all in the future, because they should now be handling it themselves?

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

@nikic nikic nikic left review comments

@arsenm arsenm Awaiting requested review from arsenm

Assignees
No one assigned
Labels
backend:AMDGPU backend:X86 llvm:SelectionDAG SelectionDAGISel as well
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
3 participants
@wzssyqa @llvmbot @nikic