Skip to content

[X86] Check MinMax has NaN and replace with NewX for minimumnum/maximumnum #164546

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
Merged
merged 2 commits into from
Oct 22, 2025
Merged

[X86] Check MinMax has NaN and replace with NewX for minimumnum/maximumnum #164546

merged 2 commits into from
Oct 22, 2025

Conversation

@phoebewang
Copy link
Contributor

It is incorrect to just check for NewX and return its ordered elements.

@phoebewang
Copy link
Contributor Author

Thanks @wenju-he for the report!

@llvmbot
Copy link
Member

llvmbot commented Oct 22, 2025

@llvm/pr-subscribers-backend-x86

Author: Phoebe Wang (phoebewang)

Changes

It is incorrect to just check for NewX and return its ordered elements.

Patch is 104.96 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/164546.diff

2 Files Affected:

  • (modified) llvm/lib/Target/X86/X86ISelLowering.cpp (+10-3)
  • (modified) llvm/test/CodeGen/X86/fminimumnum-fmaximumnum.ll (+1029-751)
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 9580adebba712..6f25ce752325f 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -29502,16 +29502,23 @@ static SDValue LowerFMINIMUM_FMAXIMUM(SDValue Op, const X86Subtarget &Subtarget,
       DAG.isKnownNeverNaN(IsNum ? NewX : NewY))
     std::swap(NewX, NewY);
 
+  if (!IgnoreNaN && IsNum && !DAG.isKnownNeverNaN(NewY)) {
+    SDValue IsOrdered = DAG.getSetCC(DL, SetCCType, NewX, NewX, ISD::SETO);
+    SDValue X = DAG.getSelect(DL, VT, IsOrdered, NewY, NewX);
+    SDValue Y = DAG.getSelect(DL, VT, IsOrdered, NewX, NewY);
+    NewX = X;
+    NewY = Y;
+  }
+
   SDValue MinMax = DAG.getNode(MinMaxOp, DL, VT, NewX, NewY, Op->getFlags());
 
-  if (IgnoreNaN || DAG.isKnownNeverNaN(IsNum ? NewY : NewX))
+  if (IgnoreNaN || IsNum || DAG.isKnownNeverNaN(NewX))
     return MinMax;
 
   if (DAG.isKnownNeverNaN(NewX))
     NewX = NewY;
 
-  SDValue IsNaN =
-      DAG.getSetCC(DL, SetCCType, NewX, NewX, IsNum ? ISD::SETO : ISD::SETUO);
+  SDValue IsNaN = DAG.getSetCC(DL, SetCCType, NewX, NewX, ISD::SETUO);
 
   return DAG.getSelect(DL, VT, IsNaN, NewX, MinMax);
 }
diff --git a/llvm/test/CodeGen/X86/fminimumnum-fmaximumnum.ll b/llvm/test/CodeGen/X86/fminimumnum-fmaximumnum.ll
index 0fe107cbbeee3..72eb565842ad1 100644
--- a/llvm/test/CodeGen/X86/fminimumnum-fmaximumnum.ll
+++ b/llvm/test/CodeGen/X86/fminimumnum-fmaximumnum.ll
@@ -22,25 +22,28 @@ declare <4 x bfloat> @llvm.maximumnum.v4bf16(<4 x bfloat>, <4 x bfloat>)
 define float @test_fmaximumnum(float %x, float %y) nounwind {
 ; SSE2-LABEL: test_fmaximumnum:
 ; SSE2:       # %bb.0:
-; SSE2-NEXT:    movdqa %xmm0, %xmm2
 ; SSE2-NEXT:    movd %xmm0, %eax
 ; SSE2-NEXT:    testl %eax, %eax
-; SSE2-NEXT:    movdqa %xmm0, %xmm3
-; SSE2-NEXT:    js .LBB0_2
-; SSE2-NEXT:  # %bb.1:
+; SSE2-NEXT:    js .LBB0_1
+; SSE2-NEXT:  # %bb.2:
+; SSE2-NEXT:    movdqa %xmm0, %xmm2
+; SSE2-NEXT:    jmp .LBB0_3
+; SSE2-NEXT:  .LBB0_1:
+; SSE2-NEXT:    movdqa %xmm1, %xmm2
+; SSE2-NEXT:    movdqa %xmm0, %xmm1
+; SSE2-NEXT:  .LBB0_3:
+; SSE2-NEXT:    ucomiss %xmm1, %xmm1
 ; SSE2-NEXT:    movdqa %xmm1, %xmm3
-; SSE2-NEXT:  .LBB0_2:
-; SSE2-NEXT:    movdqa %xmm3, %xmm0
-; SSE2-NEXT:    cmpordss %xmm3, %xmm0
-; SSE2-NEXT:    movaps %xmm0, %xmm4
-; SSE2-NEXT:    andps %xmm3, %xmm4
-; SSE2-NEXT:    js .LBB0_4
-; SSE2-NEXT:  # %bb.3:
-; SSE2-NEXT:    movdqa %xmm2, %xmm1
-; SSE2-NEXT:  .LBB0_4:
-; SSE2-NEXT:    maxss %xmm1, %xmm3
-; SSE2-NEXT:    andnps %xmm3, %xmm0
-; SSE2-NEXT:    orps %xmm4, %xmm0
+; SSE2-NEXT:    jnp .LBB0_5
+; SSE2-NEXT:  # %bb.4:
+; SSE2-NEXT:    movdqa %xmm2, %xmm3
+; SSE2-NEXT:  .LBB0_5:
+; SSE2-NEXT:    movaps %xmm1, %xmm0
+; SSE2-NEXT:    cmpordss %xmm1, %xmm0
+; SSE2-NEXT:    pand %xmm0, %xmm2
+; SSE2-NEXT:    pandn %xmm1, %xmm0
+; SSE2-NEXT:    por %xmm2, %xmm0
+; SSE2-NEXT:    maxss %xmm3, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; AVX1-LABEL: test_fmaximumnum:
@@ -55,9 +58,15 @@ define float @test_fmaximumnum(float %x, float %y) nounwind {
 ; AVX1-NEXT:    vmovdqa %xmm1, %xmm2
 ; AVX1-NEXT:    vmovdqa %xmm0, %xmm1
 ; AVX1-NEXT:  .LBB0_3:
-; AVX1-NEXT:    vmaxss %xmm2, %xmm1, %xmm0
-; AVX1-NEXT:    vcmpordss %xmm1, %xmm1, %xmm2
-; AVX1-NEXT:    vblendvps %xmm2, %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    vucomiss %xmm1, %xmm1
+; AVX1-NEXT:    vmovdqa %xmm1, %xmm0
+; AVX1-NEXT:    jnp .LBB0_5
+; AVX1-NEXT:  # %bb.4:
+; AVX1-NEXT:    vmovdqa %xmm2, %xmm0
+; AVX1-NEXT:  .LBB0_5:
+; AVX1-NEXT:    vcmpordss %xmm1, %xmm1, %xmm3
+; AVX1-NEXT:    vblendvps %xmm3, %xmm2, %xmm1, %xmm1
+; AVX1-NEXT:    vmaxss %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX512-LABEL: test_fmaximumnum:
@@ -66,12 +75,16 @@ define float @test_fmaximumnum(float %x, float %y) nounwind {
 ; AVX512-NEXT:    testl %eax, %eax
 ; AVX512-NEXT:    sets %al
 ; AVX512-NEXT:    kmovw %eax, %k1
-; AVX512-NEXT:    vmovdqa %xmm0, %xmm2
-; AVX512-NEXT:    vmovss %xmm1, %xmm2, %xmm2 {%k1}
-; AVX512-NEXT:    vmovss %xmm0, %xmm1, %xmm1 {%k1}
-; AVX512-NEXT:    vmaxss %xmm2, %xmm1, %xmm0
-; AVX512-NEXT:    vcmpordss %xmm1, %xmm1, %k1
+; AVX512-NEXT:    vmovaps %xmm1, %xmm2
+; AVX512-NEXT:    vmovss %xmm0, %xmm2, %xmm2 {%k1}
 ; AVX512-NEXT:    vmovss %xmm1, %xmm0, %xmm0 {%k1}
+; AVX512-NEXT:    vucomiss %xmm2, %xmm2
+; AVX512-NEXT:    setnp %al
+; AVX512-NEXT:    kmovw %eax, %k1
+; AVX512-NEXT:    vmovaps %xmm0, %xmm1
+; AVX512-NEXT:    vmovss %xmm2, %xmm1, %xmm1 {%k1}
+; AVX512-NEXT:    vmovss %xmm0, %xmm2, %xmm2 {%k1}
+; AVX512-NEXT:    vmaxss %xmm1, %xmm2, %xmm0
 ; AVX512-NEXT:    retq
 ;
 ; AVX10_2-LABEL: test_fmaximumnum:
@@ -94,9 +107,15 @@ define float @test_fmaximumnum(float %x, float %y) nounwind {
 ; X86-NEXT:    vmovdqa %xmm0, %xmm1
 ; X86-NEXT:    vmovdqa %xmm2, %xmm0
 ; X86-NEXT:  .LBB0_3:
-; X86-NEXT:    vmaxss %xmm1, %xmm0, %xmm1
-; X86-NEXT:    vcmpordss %xmm0, %xmm0, %xmm2
-; X86-NEXT:    vblendvps %xmm2, %xmm0, %xmm1, %xmm0
+; X86-NEXT:    vucomiss %xmm0, %xmm0
+; X86-NEXT:    vmovdqa %xmm0, %xmm2
+; X86-NEXT:    jnp .LBB0_5
+; X86-NEXT:  # %bb.4:
+; X86-NEXT:    vmovdqa %xmm1, %xmm2
+; X86-NEXT:  .LBB0_5:
+; X86-NEXT:    vcmpordss %xmm0, %xmm0, %xmm3
+; X86-NEXT:    vblendvps %xmm3, %xmm1, %xmm0, %xmm0
+; X86-NEXT:    vmaxss %xmm2, %xmm0, %xmm0
 ; X86-NEXT:    vmovss %xmm0, (%esp)
 ; X86-NEXT:    flds (%esp)
 ; X86-NEXT:    popl %eax
@@ -370,29 +389,43 @@ define double @test_fmaximumnum_zero2(double %x, double %y) {
 define float @test_fmaximumnum_nsz(float %x, float %y) "no-signed-zeros-fp-math"="true" nounwind {
 ; SSE2-LABEL: test_fmaximumnum_nsz:
 ; SSE2:       # %bb.0:
+; SSE2-NEXT:    ucomiss %xmm0, %xmm0
+; SSE2-NEXT:    movaps %xmm0, %xmm3
+; SSE2-NEXT:    jnp .LBB8_2
+; SSE2-NEXT:  # %bb.1:
+; SSE2-NEXT:    movaps %xmm1, %xmm3
+; SSE2-NEXT:  .LBB8_2:
 ; SSE2-NEXT:    movaps %xmm0, %xmm2
 ; SSE2-NEXT:    cmpordss %xmm0, %xmm2
-; SSE2-NEXT:    movaps %xmm2, %xmm3
-; SSE2-NEXT:    andps %xmm0, %xmm3
-; SSE2-NEXT:    maxss %xmm1, %xmm0
+; SSE2-NEXT:    andps %xmm2, %xmm1
 ; SSE2-NEXT:    andnps %xmm0, %xmm2
-; SSE2-NEXT:    orps %xmm3, %xmm2
+; SSE2-NEXT:    orps %xmm1, %xmm2
+; SSE2-NEXT:    maxss %xmm3, %xmm2
 ; SSE2-NEXT:    movaps %xmm2, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; AVX1-LABEL: test_fmaximumnum_nsz:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vmaxss %xmm1, %xmm0, %xmm1
-; AVX1-NEXT:    vcmpordss %xmm0, %xmm0, %xmm2
-; AVX1-NEXT:    vblendvps %xmm2, %xmm0, %xmm1, %xmm0
+; AVX1-NEXT:    vucomiss %xmm0, %xmm0
+; AVX1-NEXT:    vmovaps %xmm0, %xmm2
+; AVX1-NEXT:    jnp .LBB8_2
+; AVX1-NEXT:  # %bb.1:
+; AVX1-NEXT:    vmovaps %xmm1, %xmm2
+; AVX1-NEXT:  .LBB8_2:
+; AVX1-NEXT:    vcmpordss %xmm0, %xmm0, %xmm3
+; AVX1-NEXT:    vblendvps %xmm3, %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    vmaxss %xmm2, %xmm0, %xmm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX512-LABEL: test_fmaximumnum_nsz:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    vmaxss %xmm1, %xmm0, %xmm1
-; AVX512-NEXT:    vcmpordss %xmm0, %xmm0, %k1
-; AVX512-NEXT:    vmovss %xmm0, %xmm1, %xmm1 {%k1}
-; AVX512-NEXT:    vmovaps %xmm1, %xmm0
+; AVX512-NEXT:    vucomiss %xmm0, %xmm0
+; AVX512-NEXT:    setnp %al
+; AVX512-NEXT:    kmovw %eax, %k1
+; AVX512-NEXT:    vmovaps %xmm1, %xmm2
+; AVX512-NEXT:    vmovss %xmm0, %xmm2, %xmm2 {%k1}
+; AVX512-NEXT:    vmovss %xmm1, %xmm0, %xmm0 {%k1}
+; AVX512-NEXT:    vmaxss %xmm2, %xmm0, %xmm0
 ; AVX512-NEXT:    retq
 ;
 ; AVX10_2-LABEL: test_fmaximumnum_nsz:
@@ -404,9 +437,16 @@ define float @test_fmaximumnum_nsz(float %x, float %y) "no-signed-zeros-fp-math"
 ; X86:       # %bb.0:
 ; X86-NEXT:    pushl %eax
 ; X86-NEXT:    vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero
-; X86-NEXT:    vcmpordss %xmm0, %xmm0, %xmm1
-; X86-NEXT:    vmaxss {{[0-9]+}}(%esp), %xmm0, %xmm2
-; X86-NEXT:    vblendvps %xmm1, %xmm0, %xmm2, %xmm0
+; X86-NEXT:    vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero
+; X86-NEXT:    vucomiss %xmm1, %xmm1
+; X86-NEXT:    vmovaps %xmm1, %xmm2
+; X86-NEXT:    jnp .LBB8_2
+; X86-NEXT:  # %bb.1:
+; X86-NEXT:    vmovaps %xmm0, %xmm2
+; X86-NEXT:  .LBB8_2:
+; X86-NEXT:    vcmpordss %xmm1, %xmm1, %xmm3
+; X86-NEXT:    vblendvps %xmm3, %xmm0, %xmm1, %xmm0
+; X86-NEXT:    vmaxss %xmm2, %xmm0, %xmm0
 ; X86-NEXT:    vmovss %xmm0, (%esp)
 ; X86-NEXT:    flds (%esp)
 ; X86-NEXT:    popl %eax
@@ -421,23 +461,26 @@ define float @test_fmaximumnum_combine_cmps(float %x, float %y) nounwind {
 ; SSE2-NEXT:    divss %xmm0, %xmm1
 ; SSE2-NEXT:    movd %xmm0, %eax
 ; SSE2-NEXT:    testl %eax, %eax
-; SSE2-NEXT:    movaps %xmm0, %xmm3
-; SSE2-NEXT:    js .LBB9_2
-; SSE2-NEXT:  # %bb.1:
-; SSE2-NEXT:    movaps %xmm1, %xmm3
-; SSE2-NEXT:  .LBB9_2:
-; SSE2-NEXT:    movaps %xmm3, %xmm2
-; SSE2-NEXT:    cmpordss %xmm3, %xmm2
-; SSE2-NEXT:    movaps %xmm2, %xmm4
-; SSE2-NEXT:    andps %xmm3, %xmm4
-; SSE2-NEXT:    js .LBB9_4
-; SSE2-NEXT:  # %bb.3:
+; SSE2-NEXT:    js .LBB9_1
+; SSE2-NEXT:  # %bb.2:
+; SSE2-NEXT:    movaps %xmm0, %xmm2
+; SSE2-NEXT:    jmp .LBB9_3
+; SSE2-NEXT:  .LBB9_1:
+; SSE2-NEXT:    movaps %xmm1, %xmm2
 ; SSE2-NEXT:    movaps %xmm0, %xmm1
-; SSE2-NEXT:  .LBB9_4:
-; SSE2-NEXT:    maxss %xmm1, %xmm3
-; SSE2-NEXT:    andnps %xmm3, %xmm2
-; SSE2-NEXT:    orps %xmm4, %xmm2
-; SSE2-NEXT:    movaps %xmm2, %xmm0
+; SSE2-NEXT:  .LBB9_3:
+; SSE2-NEXT:    ucomiss %xmm1, %xmm1
+; SSE2-NEXT:    movaps %xmm1, %xmm3
+; SSE2-NEXT:    jnp .LBB9_5
+; SSE2-NEXT:  # %bb.4:
+; SSE2-NEXT:    movaps %xmm2, %xmm3
+; SSE2-NEXT:  .LBB9_5:
+; SSE2-NEXT:    movaps %xmm1, %xmm0
+; SSE2-NEXT:    cmpordss %xmm1, %xmm0
+; SSE2-NEXT:    andps %xmm0, %xmm2
+; SSE2-NEXT:    andnps %xmm1, %xmm0
+; SSE2-NEXT:    orps %xmm2, %xmm0
+; SSE2-NEXT:    maxss %xmm3, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; AVX1-LABEL: test_fmaximumnum_combine_cmps:
@@ -453,9 +496,15 @@ define float @test_fmaximumnum_combine_cmps(float %x, float %y) nounwind {
 ; AVX1-NEXT:    vmovaps %xmm1, %xmm2
 ; AVX1-NEXT:    vmovaps %xmm0, %xmm1
 ; AVX1-NEXT:  .LBB9_3:
-; AVX1-NEXT:    vmaxss %xmm2, %xmm1, %xmm0
-; AVX1-NEXT:    vcmpordss %xmm1, %xmm1, %xmm2
-; AVX1-NEXT:    vblendvps %xmm2, %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    vucomiss %xmm1, %xmm1
+; AVX1-NEXT:    vmovaps %xmm1, %xmm0
+; AVX1-NEXT:    jnp .LBB9_5
+; AVX1-NEXT:  # %bb.4:
+; AVX1-NEXT:    vmovaps %xmm2, %xmm0
+; AVX1-NEXT:  .LBB9_5:
+; AVX1-NEXT:    vcmpordss %xmm1, %xmm1, %xmm3
+; AVX1-NEXT:    vblendvps %xmm3, %xmm2, %xmm1, %xmm1
+; AVX1-NEXT:    vmaxss %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX512F-LABEL: test_fmaximumnum_combine_cmps:
@@ -465,12 +514,16 @@ define float @test_fmaximumnum_combine_cmps(float %x, float %y) nounwind {
 ; AVX512F-NEXT:    testl %eax, %eax
 ; AVX512F-NEXT:    sets %al
 ; AVX512F-NEXT:    kmovw %eax, %k1
-; AVX512F-NEXT:    vmovaps %xmm0, %xmm2
-; AVX512F-NEXT:    vmovss %xmm1, %xmm2, %xmm2 {%k1}
-; AVX512F-NEXT:    vmovss %xmm0, %xmm1, %xmm1 {%k1}
-; AVX512F-NEXT:    vmaxss %xmm2, %xmm1, %xmm0
-; AVX512F-NEXT:    vcmpordss %xmm1, %xmm1, %k1
+; AVX512F-NEXT:    vmovaps %xmm1, %xmm2
+; AVX512F-NEXT:    vmovss %xmm0, %xmm2, %xmm2 {%k1}
 ; AVX512F-NEXT:    vmovss %xmm1, %xmm0, %xmm0 {%k1}
+; AVX512F-NEXT:    vucomiss %xmm2, %xmm2
+; AVX512F-NEXT:    setnp %al
+; AVX512F-NEXT:    kmovw %eax, %k1
+; AVX512F-NEXT:    vmovaps %xmm0, %xmm1
+; AVX512F-NEXT:    vmovss %xmm2, %xmm1, %xmm1 {%k1}
+; AVX512F-NEXT:    vmovss %xmm0, %xmm2, %xmm2 {%k1}
+; AVX512F-NEXT:    vmaxss %xmm1, %xmm2, %xmm0
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512DQ-LABEL: test_fmaximumnum_combine_cmps:
@@ -493,22 +546,28 @@ define float @test_fmaximumnum_combine_cmps(float %x, float %y) nounwind {
 ; X86-LABEL: test_fmaximumnum_combine_cmps:
 ; X86:       # %bb.0:
 ; X86-NEXT:    pushl %eax
-; X86-NEXT:    vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero
+; X86-NEXT:    vmovss {{.*#+}} xmm2 = mem[0],zero,zero,zero
 ; X86-NEXT:    vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero
-; X86-NEXT:    vdivss %xmm1, %xmm0, %xmm0
-; X86-NEXT:    vmovd %xmm1, %eax
+; X86-NEXT:    vdivss %xmm2, %xmm0, %xmm0
+; X86-NEXT:    vmovd %xmm2, %eax
 ; X86-NEXT:    testl %eax, %eax
 ; X86-NEXT:    js .LBB9_1
 ; X86-NEXT:  # %bb.2:
-; X86-NEXT:    vmovaps %xmm1, %xmm2
+; X86-NEXT:    vmovaps %xmm2, %xmm1
 ; X86-NEXT:    jmp .LBB9_3
 ; X86-NEXT:  .LBB9_1:
-; X86-NEXT:    vmovaps %xmm0, %xmm2
-; X86-NEXT:    vmovaps %xmm1, %xmm0
+; X86-NEXT:    vmovaps %xmm0, %xmm1
+; X86-NEXT:    vmovaps %xmm2, %xmm0
 ; X86-NEXT:  .LBB9_3:
-; X86-NEXT:    vmaxss %xmm2, %xmm0, %xmm1
-; X86-NEXT:    vcmpordss %xmm0, %xmm0, %xmm2
-; X86-NEXT:    vblendvps %xmm2, %xmm0, %xmm1, %xmm0
+; X86-NEXT:    vucomiss %xmm0, %xmm0
+; X86-NEXT:    vmovaps %xmm0, %xmm2
+; X86-NEXT:    jnp .LBB9_5
+; X86-NEXT:  # %bb.4:
+; X86-NEXT:    vmovaps %xmm1, %xmm2
+; X86-NEXT:  .LBB9_5:
+; X86-NEXT:    vcmpordss %xmm0, %xmm0, %xmm3
+; X86-NEXT:    vblendvps %xmm3, %xmm1, %xmm0, %xmm0
+; X86-NEXT:    vmaxss %xmm2, %xmm0, %xmm0
 ; X86-NEXT:    vmovss %xmm0, (%esp)
 ; X86-NEXT:    flds (%esp)
 ; X86-NEXT:    popl %eax
@@ -527,23 +586,27 @@ define float @test_fminimumnum(float %x, float %y) nounwind {
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movd %xmm0, %eax
 ; SSE2-NEXT:    testl %eax, %eax
-; SSE2-NEXT:    movdqa %xmm1, %xmm3
-; SSE2-NEXT:    js .LBB10_2
-; SSE2-NEXT:  # %bb.1:
-; SSE2-NEXT:    movdqa %xmm0, %xmm3
-; SSE2-NEXT:  .LBB10_2:
-; SSE2-NEXT:    movdqa %xmm3, %xmm2
-; SSE2-NEXT:    cmpordss %xmm3, %xmm2
-; SSE2-NEXT:    movaps %xmm2, %xmm4
-; SSE2-NEXT:    andps %xmm3, %xmm4
-; SSE2-NEXT:    js .LBB10_4
-; SSE2-NEXT:  # %bb.3:
+; SSE2-NEXT:    js .LBB10_1
+; SSE2-NEXT:  # %bb.2:
+; SSE2-NEXT:    movdqa %xmm1, %xmm2
+; SSE2-NEXT:    jmp .LBB10_3
+; SSE2-NEXT:  .LBB10_1:
+; SSE2-NEXT:    movdqa %xmm0, %xmm2
 ; SSE2-NEXT:    movdqa %xmm1, %xmm0
-; SSE2-NEXT:  .LBB10_4:
-; SSE2-NEXT:    minss %xmm0, %xmm3
-; SSE2-NEXT:    andnps %xmm3, %xmm2
-; SSE2-NEXT:    orps %xmm4, %xmm2
-; SSE2-NEXT:    movaps %xmm2, %xmm0
+; SSE2-NEXT:  .LBB10_3:
+; SSE2-NEXT:    ucomiss %xmm0, %xmm0
+; SSE2-NEXT:    movdqa %xmm0, %xmm3
+; SSE2-NEXT:    jnp .LBB10_5
+; SSE2-NEXT:  # %bb.4:
+; SSE2-NEXT:    movdqa %xmm2, %xmm3
+; SSE2-NEXT:  .LBB10_5:
+; SSE2-NEXT:    movaps %xmm0, %xmm1
+; SSE2-NEXT:    cmpordss %xmm0, %xmm1
+; SSE2-NEXT:    pand %xmm1, %xmm2
+; SSE2-NEXT:    pandn %xmm0, %xmm1
+; SSE2-NEXT:    por %xmm2, %xmm1
+; SSE2-NEXT:    minss %xmm3, %xmm1
+; SSE2-NEXT:    movaps %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; AVX1-LABEL: test_fminimumnum:
@@ -558,9 +621,15 @@ define float @test_fminimumnum(float %x, float %y) nounwind {
 ; AVX1-NEXT:    vmovdqa %xmm0, %xmm2
 ; AVX1-NEXT:    vmovdqa %xmm1, %xmm0
 ; AVX1-NEXT:  .LBB10_3:
-; AVX1-NEXT:    vminss %xmm2, %xmm0, %xmm1
-; AVX1-NEXT:    vcmpordss %xmm0, %xmm0, %xmm2
-; AVX1-NEXT:    vblendvps %xmm2, %xmm0, %xmm1, %xmm0
+; AVX1-NEXT:    vucomiss %xmm0, %xmm0
+; AVX1-NEXT:    vmovdqa %xmm0, %xmm1
+; AVX1-NEXT:    jnp .LBB10_5
+; AVX1-NEXT:  # %bb.4:
+; AVX1-NEXT:    vmovdqa %xmm2, %xmm1
+; AVX1-NEXT:  .LBB10_5:
+; AVX1-NEXT:    vcmpordss %xmm0, %xmm0, %xmm3
+; AVX1-NEXT:    vblendvps %xmm3, %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vminss %xmm1, %xmm0, %xmm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX512-LABEL: test_fminimumnum:
@@ -569,13 +638,16 @@ define float @test_fminimumnum(float %x, float %y) nounwind {
 ; AVX512-NEXT:    testl %eax, %eax
 ; AVX512-NEXT:    sets %al
 ; AVX512-NEXT:    kmovw %eax, %k1
-; AVX512-NEXT:    vmovaps %xmm1, %xmm2
-; AVX512-NEXT:    vmovss %xmm0, %xmm2, %xmm2 {%k1}
-; AVX512-NEXT:    vmovss %xmm1, %xmm0, %xmm0 {%k1}
-; AVX512-NEXT:    vminss %xmm2, %xmm0, %xmm1
-; AVX512-NEXT:    vcmpordss %xmm0, %xmm0, %k1
+; AVX512-NEXT:    vmovdqa %xmm0, %xmm2
+; AVX512-NEXT:    vmovss %xmm1, %xmm2, %xmm2 {%k1}
 ; AVX512-NEXT:    vmovss %xmm0, %xmm1, %xmm1 {%k1}
+; AVX512-NEXT:    vucomiss %xmm2, %xmm2
+; AVX512-NEXT:    setnp %al
+; AVX512-NEXT:    kmovw %eax, %k1
 ; AVX512-NEXT:    vmovaps %xmm1, %xmm0
+; AVX512-NEXT:    vmovss %xmm2, %xmm0, %xmm0 {%k1}
+; AVX512-NEXT:    vmovss %xmm1, %xmm2, %xmm2 {%k1}
+; AVX512-NEXT:    vminss %xmm0, %xmm2, %xmm0
 ; AVX512-NEXT:    retq
 ;
 ; AVX10_2-LABEL: test_fminimumnum:
@@ -586,21 +658,27 @@ define float @test_fminimumnum(float %x, float %y) nounwind {
 ; X86-LABEL: test_fminimumnum:
 ; X86:       # %bb.0:
 ; X86-NEXT:    pushl %eax
-; X86-NEXT:    vmovd {{.*#+}} xmm1 = mem[0],zero,zero,zero
+; X86-NEXT:    vmovd {{.*#+}} xmm2 = mem[0],zero,zero,zero
 ; X86-NEXT:    vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
 ; X86-NEXT:    vmovd %xmm0, %eax
 ; X86-NEXT:    testl %eax, %eax
 ; X86-NEXT:    js .LBB10_1
 ; X86-NEXT:  # %bb.2:
-; X86-NEXT:    vmovdqa %xmm1, %xmm2
+; X86-NEXT:    vmovdqa %xmm2, %xmm1
 ; X86-NEXT:    jmp .LBB10_3
 ; X86-NEXT:  .LBB10_1:
-; X86-NEXT:    vmovdqa %xmm0, %xmm2
-; X86-NEXT:    vmovdqa %xmm1, %xmm0
+; X86-NEXT:    vmovdqa %xmm0, %xmm1
+; X86-NEXT:    vmovdqa %xmm2, %xmm0
 ; X86-NEXT:  .LBB10_3:
-; X86-NEXT:    vminss %xmm2, %xmm0, %xmm1
-; X86-NEXT:    vcmpordss %xmm0, %xmm0, %xmm2
-; X86-NEXT:    vblendvps %xmm2, %xmm0, %xmm1, %xmm0
+; X86-NEXT:    vucomiss %xmm0, %xmm0
+; X86-NEXT:    vmovdqa %xmm0, %xmm2
+; X86-NEXT:    jnp .LBB10_5
+; X86-NEXT:  # %bb.4:
+; X86-NEXT:    vmovdqa %xmm1, %xmm2
+; X86-NEXT:  .LBB10_5:
+; X86-NEXT:    vcmpordss %xmm0, %xmm0, %xmm3
+; X86-NEXT:    vblendvps %xmm3, %xmm1, %xmm0, %xmm0
+; X86-NEXT:    vminss %xmm2, %xmm0, %xmm0
 ; X86-NEXT:    vmovss %xmm0, (%esp)
 ; X86-NEXT:    flds (%esp)
 ; X86-NEXT:    popl %eax
@@ -856,29 +934,43 @@ define double @test_fminimumnum_zero2(double %x, double %y) {
 define float @test_fminimumnum_nsz(float %x, float %y) nounwind {
 ; SSE2-LABEL: test_fminimumnum_nsz:
 ; SSE2:       # %bb.0:
+; SSE2-NEXT:    ucomiss %xmm0, %xmm0
+; SSE2-NEXT:    movaps %xmm0, %xmm3
+; SSE2-NEXT:    jnp .LBB18_2
+; SSE2-NEXT:  # %bb.1:
+; SSE2-NEXT:    movaps %xmm1, %xmm3
+; SSE2-NEXT:  .LBB18_2:
 ; SSE2-NEXT:    movaps %xmm0, %xmm2
 ; SSE2-NEXT:    cmpordss %xmm0, %xmm2
-; SSE2-NEXT:    movaps %xmm2, %xmm3
-; SSE2-NEXT:    andps %xmm0, %xmm3
-; SSE2-NEXT:    minss %xmm1, %xmm0
+; SSE2-NEXT:    andps %xmm2, %xmm1
 ; SSE2-NEXT:    andnps %xmm0, %xmm2
-; SSE2-NEXT:    orps %xmm3, %xmm2
+; SSE2-NEXT:    orps %xmm1, %xmm2
+; SSE2-NEXT:    minss %xmm3, %xmm2
 ; SSE2-NEXT:    movaps %xmm2, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; AVX1-LABEL: test_fminimumnum_nsz:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vminss %xmm1, %xmm0, %xmm1
-; AVX1-NEXT:    vcmpordss %xmm0, %xmm0, %xmm2
-; AVX1-NEXT:    vblendvps %xmm2, %xmm0, %xmm1, %xmm0
+; AVX1-NEXT:    vucomiss %xmm0, %xmm0
+; AVX1-NEXT:    vmovaps %xmm0, %xmm2
+; AVX1-NEXT:    jnp .LBB18_2
+; AVX1-NEXT:  # %bb.1:
+; AVX1-NEXT:    vmovaps %xmm1, %xmm2
+; AVX1-NEXT:  .LBB18_2:
+; AVX1-NEXT:    vcmpordss %xmm0, %xmm0, %xmm3
+; AVX1-NEXT:    vblendvps %xmm3, %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    vminss %xmm2, %xmm0, %xmm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX512-LABEL: test_fminimumnum_nsz:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    vminss %xmm1, %xmm0, %xmm1
-; AVX512-NEXT:    vcmpordss %xmm0, %xmm0, %k1
-; AVX512-NEXT:    vmovss %xmm0, %xmm1, %xmm1 {%k1}
-; AVX512-NEXT:    vmovaps %xmm1, %xmm0
+; AVX512-NEXT:    vucomiss %xmm0, %xmm0
+; AVX512-NEXT:    setnp %al
+; AVX512-NEXT:    kmovw %eax, %k1
+; AVX512-NEXT:    vmovaps %xmm1, %xmm2
+; AVX512-NEXT:    vmovss %xmm0, %xmm2, %xmm2 {%k1}
+; AVX512-NEXT:    vmovss %xmm1, %xmm0, %xmm0 {%k1}
+; AVX512-NEXT:    vminss %xmm2, %xmm0, %xmm0
 ; AVX512-NEXT:    retq
 ;
 ; AVX10_2-LABEL: test_fminimumnum_nsz:
@@ -890,9 +982,16 @@ define float @test_fminimumnum_nsz(float %x, float %y) nounwind {
 ; X86:       # %bb.0:
 ; X86-NEXT:    pushl %eax
 ; X86-NEXT:    vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero
-; X86-NEXT:    vcmpordss %xmm0, %xmm0, %xmm1
-; X86-NEXT:    vminss {{[0-9]+}}(%esp), %xmm0, %xmm2
-; X86-NEXT:    vblendvps %xmm1, %xmm0, %xmm2, %xmm0
+; X86-NEXT:    vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero
+; X86-NEXT:    vucomiss %xmm1, %xmm1
+; X86-NEXT:    vmovaps %xmm1, %xmm2
+; X86-NEXT:    jnp .LBB18_2
+; X86-NEXT:  # %bb.1:
+; X86-NEXT:    vmovaps %xmm0, %xmm2
+; X86-NEXT:  .LBB18_2:
+; X86-NEXT:    vcmpordss %xmm1, %xmm1, %xmm3
+; X86-NEXT:    vblendvps %xmm3, %xmm0, %xmm1, %xmm0
+; X86-NEXT:    vminss %xmm2, %xmm0, %xmm0
 ; X86-NEXT:    vmovss %xmm0, (%esp)
 ; X86-NEXT:    flds (%esp)
 ; X86-NEXT:    popl %eax
@@ -907,23 +1006,27 @@ define float @test_fminimumnum_combine_cmps(float %x, float %y) nounwind {
 ; SSE2-NEXT:    divss %xmm0, %xmm1
 ; SSE2-NEXT:    movd %xmm0, %eax
 ; SSE2-NEXT:    testl %eax, %eax
-; SSE2-NEXT:    movaps %xmm1, %xmm3
-; SSE2-NEXT:    js .LBB19_2
-; SSE2-NEXT:  # %bb.1:
+; SSE2-NEXT:    js .LBB19_1
+; ...
[truncated]

@wenju-he
Copy link
Contributor

Thanks @phoebewang, this PR can fix the failing tests in intel/llvm#20402

@wenju-he wenju-he mentioned this pull request Oct 22, 2025
Open
KanRobert
KanRobert reviewed Oct 22, 2025
View reviewed changes
llvm/lib/Target/X86/X86ISelLowering.cpp Outdated
return MinMax;

if (DAG.isKnownNeverNaN(NewX))
NewX = NewY;

SDValue IsNaN =
DAG.getSetCC(DL, SetCCType, NewX, NewX, IsNum ? ISD::SETO : ISD::SETUO);
SDValue IsNaN = DAG.getSetCC(DL, SetCCType, NewX, NewX, ISD::SETUO);
Copy link
Contributor

Choose a reason for hiding this comment

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

IsNum ? ISD::SETO : ISD::SETUO

->

ISD::SETUO

Why this change?

Copy link
Contributor Author

Choose a reason for hiding this comment

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

I optimized it with another approach. PTAL.

Copy link
Contributor Author

Choose a reason for hiding this comment

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

The prior code is to check if NewX has ordered elements. It's not correct for minimumnum/maximumnum. The new code check if the MinMax has unordered elements and repalce it with NewX.

@phoebewang
[X86] Swap NewX and NewY for ordered elements for minimumnum/maximumnum
e28163f 
It is incorrect to just check for NewX and return its ordered elements.
@phoebewang phoebewang changed the title [X86] Swap NewX and NewY for ordered elements for minimumnum/maximumnum [X86] Check MinMax has NaN and replace with NewX for minimumnum/maximumnum Oct 22, 2025
@phoebewang
Copy link
Contributor Author

Thanks @phoebewang, this PR can fix the failing tests in intel/llvm#20402

Thanks for the check @wenju-he! Sorry, I just have an idea to optimize the implementation. Could you help check it again? Thanks!

@wenju-he
Copy link
Contributor

Could you help check it again?

the newer version still works.

@phoebewang
Copy link
Contributor Author

Could you help check it again?

the newer version still works.

Thanks a lot!

KanRobert
KanRobert approved these changes Oct 22, 2025
View reviewed changes
Copy link
Contributor

@KanRobert KanRobert left a comment

Choose a reason for hiding this comment

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

LGTM. This logic is a bit hard to follow; hope it can be simplified using early returns to reduce complexity from line 29496 for IsNum=true in the future.

@phoebewang
Copy link
Contributor Author

LGTM. This logic is a bit hard to follow; hope it can be simplified using early returns to reduce complexity from line 29496 for IsNum=true in the future.

Simplified :)

@phoebewang phoebewang enabled auto-merge (squash) October 22, 2025 07:00
@phoebewang phoebewang merged commit 5ad2487 into llvm:main Oct 22, 2025
9 of 10 checks passed
@phoebewang phoebewang deleted the FMINIMUM_FMAXIMUM branch October 22, 2025 07:50
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@KanRobert KanRobert KanRobert approved these changes

@RKSimon RKSimon Awaiting requested review from RKSimon

@wenju-he wenju-he Awaiting requested review from wenju-he

@e-kud e-kud Awaiting requested review from e-kud

Assignees

No one assigned

Labels

backend:X86

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

4 participants

@phoebewang @llvmbot @wenju-he @KanRobert