[SDAG] Check fminnum for non zero operand.

[samparker]

Summary

Following on from D143106...

Currently, in TargetLowering, if the target does not support fminnum, we lower to fminimum if neither operand could be a NaN. But this isn't quite correct because fminnum and fminimum treat +/-0 differently; so, we need to prove that one of the operands isn't a zero.

[spatel]

llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp

7995

Could we also check hasNoSignedZeros() and allow this?

[dmgreen]

It would be good if you could copy the Arm tests so there are versions that can transform and versions that can't too.

[samparker]

• Added check for no signed zeros.
• Fixed up vector support.
• Added arm tests.

[spatel]

llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp

7980

Is this different than

auto *C = isConstOrConstSplatFP(N));
return C && C->isZero();

[spatel]

llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp

7980

Oops - invert that isZero() result (we're looking for non-zero constant).

[samparker]

llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp

7980

For splats, I think your suggestion would work, and that's what I originally tried to do... but it doesn't work for the BUILD_VECTOR case, where we need to check if any element is a zero.

[spatel]

llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp

7980

Ah, right. It looks like we do have the basic call already:
SelectionDAG::isKnownNeverZeroFloat().
...and it has a TODO comment about handling a vector constant, so just append that part there?

[samparker]

llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp

7980

Nice, will do.

[samparker]

Modified, and now using, isKnownNeverZeroFloat.

[spatel]

LGTM - would be best to pre-commit the new baseline tests, so we have that coverage in place even if the patch is reverted for some reason.

llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp

bool SelectionDAG::isKnownNeverNaN(SDValue Op, bool SNaN, unsigned Depth) const {
   }
 }
 
 bool SelectionDAG::isKnownNeverZeroFloat(SDValue Op) const {
   assert(Op.getValueType().isFloatingPoint() &&
          "Floating point type expected");
 
   // If the value is a constant, we can obviously see if it is a zero or not.
-  // TODO: Add BuildVector support.
   if (const ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Op))
     return !C->isZero();
+
+  // Return false if we find any zero in a vector.
+  if (Op->getOpcode() == ISD::BUILD_VECTOR ||
+      Op->getOpcode() == ISD::SPLAT_VECTOR) {
+    for (const SDValue &OpVal : Op->op_values()) {
+      if (OpVal.isUndef())
+        return false;
+      if (auto *C = dyn_cast<ConstantFPSDNode>(OpVal))
+        if (C->isZero())
+          return false;
+    }
+    return true;
+  }
   return false;
 }
 
 bool SelectionDAG::isKnownNeverZero(SDValue Op) const {
   assert(!Op.getValueType().isFloatingPoint() &&
          "Floating point types unsupported - use isKnownNeverZeroFloat");
 
   // If the value is a constant, we can obviously see if it is a zero or not.

llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp

if (!Node->getFlags().hasNoNaNs()) {
       // Insert canonicalizes if it's possible we need to quiet to get correct
       // sNaN behavior.
       if (!DAG.isKnownNeverSNaN(Quiet0)) {
         Quiet0 = DAG.getNode(ISD::FCANONICALIZE, dl, VT, Quiet0,
                              Node->getFlags());
       }
       if (!DAG.isKnownNeverSNaN(Quiet1)) {
         Quiet1 = DAG.getNode(ISD::FCANONICALIZE, dl, VT, Quiet1,
                              Node->getFlags());
       }
     }
 
     return DAG.getNode(NewOp, dl, VT, Quiet0, Quiet1, Node->getFlags());
   }
 
   // If the target has FMINIMUM/FMAXIMUM but not FMINNUM/FMAXNUM use that
-  // instead if there are no NaNs.
-  if (Node->getFlags().hasNoNaNs()) {
+  // instead if there are no NaNs and there can't be an incompatiable zero
+  // compare: at least one operand isn't +/-0, or there are no signed-zeros.
+  if (Node->getFlags().hasNoNaNs() &&
+      (Node->getFlags().hasNoSignedZeros() ||
+       DAG.isKnownNeverZeroFloat(Node->getOperand(0)) ||
+       DAG.isKnownNeverZeroFloat(Node->getOperand(1)))) {
     unsigned IEEE2018Op =
         Node->getOpcode() == ISD::FMINNUM ? ISD::FMINIMUM : ISD::FMAXIMUM;
-    if (isOperationLegalOrCustom(IEEE2018Op, VT)) {
+    if (isOperationLegalOrCustom(IEEE2018Op, VT))
       return DAG.getNode(IEEE2018Op, dl, VT, Node->getOperand(0),
                          Node->getOperand(1), Node->getFlags());
-    }
-  }
 
   if (SDValue SelCC = createSelectForFMINNUM_FMAXNUM(Node, DAG))
     return SelCC;
 
   return SDValue();
 }
 
 SDValue TargetLowering::expandIS_FPCLASS(EVT ResultVT, SDValue Op,

llvm/test/CodeGen/ARM/lower-vmax.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc -mtriple=arm-eabihf -mattr=+neon < %s | FileCheck -check-prefixes=CHECK-NO_NEON %s
 ; RUN: llc -mtriple=arm-eabihf -mattr=+neon,+neonfp < %s | FileCheck -check-prefixes=CHECK-NEON %s
 
 define float @max_f32(float, float) {
-;CHECK-NEON: vmax.f32
-;CHECK-NO_NEON: vcmp.f32
-;CHECK-NO_NEON: vmrs
-;CHECK-NO_NEON: vmovgt.f32
+; CHECK-NO_NEON-LABEL: max_f32:
+; CHECK-NO_NEON:       @ %bb.0:
+; CHECK-NO_NEON-NEXT:    vcmp.f32 s1, s0
+; CHECK-NO_NEON-NEXT:    vmrs APSR_nzcv, fpscr
+; CHECK-NO_NEON-NEXT:    vmovgt.f32 s0, s1
+; CHECK-NO_NEON-NEXT:    mov pc, lr
+;
+; CHECK-NEON-LABEL: max_f32:
+; CHECK-NEON:       @ %bb.0:
+; CHECK-NEON-NEXT:    vcmp.f32 s1, s0
+; CHECK-NEON-NEXT:    vmrs APSR_nzcv, fpscr
+; CHECK-NEON-NEXT:    vmovgt.f32 s0, s1
+; CHECK-NEON-NEXT:    mov pc, lr
   %3 = call nnan float @llvm.maxnum.f32(float %1, float %0)
   ret float %3
 }
 
 declare float @llvm.maxnum.f32(float, float) #1
 
 define float @min_f32(float, float) {
-;CHECK-NEON: vmin.f32
-;CHECK-NO_NEON: vcmp.f32
-;CHECK-NO_NEON: vmrs
-;CHECK-NO_NEON: vmovlt.f32
+; CHECK-NO_NEON-LABEL: min_f32:
+; CHECK-NO_NEON:       @ %bb.0:
+; CHECK-NO_NEON-NEXT:    vcmp.f32 s1, s0
+; CHECK-NO_NEON-NEXT:    vmrs APSR_nzcv, fpscr
+; CHECK-NO_NEON-NEXT:    vmovlt.f32 s0, s1
+; CHECK-NO_NEON-NEXT:    mov pc, lr
+;
+; CHECK-NEON-LABEL: min_f32:
+; CHECK-NEON:       @ %bb.0:
+; CHECK-NEON-NEXT:    vcmp.f32 s1, s0
+; CHECK-NEON-NEXT:    vmrs APSR_nzcv, fpscr
+; CHECK-NEON-NEXT:    vmovlt.f32 s0, s1
+; CHECK-NEON-NEXT:    mov pc, lr
   %3 = call nnan float @llvm.minnum.f32(float %1, float %0)
   ret float %3
 }
 
 declare float @llvm.minnum.f32(float, float) #1
 

llvm/test/CodeGen/ARM/minnum-maxnum-intrinsics.ll

 ; ARMV8M-NEXT:    bx lr
   %a = call nnan double @llvm.maxnum.f64(double %x, double 1.0)
   ret double %a
 }
 
 define <4 x float> @fminnumv432_intrinsic(<4 x float> %x, <4 x float> %y) {
 ; ARMV7-LABEL: fminnumv432_intrinsic:
 ; ARMV7:       @ %bb.0:
-; ARMV7-NEXT:    vmov d17, r2, r3
-; ARMV7-NEXT:    vmov d16, r0, r1
-; ARMV7-NEXT:    mov r0, sp
-; ARMV7-NEXT:    vld1.64 {d18, d19}, [r0]
-; ARMV7-NEXT:    vmin.f32 q8, q8, q9
-; ARMV7-NEXT:    vmov r0, r1, d16
-; ARMV7-NEXT:    vmov r2, r3, d17
+; ARMV7-NEXT:    mov r12, sp
+; ARMV7-NEXT:    vld1.64 {d0, d1}, [r12]
+; ARMV7-NEXT:    vmov d3, r2, r3
+; ARMV7-NEXT:    vmov d2, r0, r1
+; ARMV7-NEXT:    vcmp.f32 s7, s3
+; ARMV7-NEXT:    vmrs APSR_nzcv, fpscr
+; ARMV7-NEXT:    vcmp.f32 s6, s2
+; ARMV7-NEXT:    vmovlt.f32 s3, s7
+; ARMV7-NEXT:    vmrs APSR_nzcv, fpscr
+; ARMV7-NEXT:    vcmp.f32 s5, s1
+; ARMV7-NEXT:    vmovlt.f32 s2, s6
+; ARMV7-NEXT:    vmrs APSR_nzcv, fpscr
+; ARMV7-NEXT:    vcmp.f32 s4, s0
+; ARMV7-NEXT:    vmovlt.f32 s1, s5
+; ARMV7-NEXT:    vmrs APSR_nzcv, fpscr
+; ARMV7-NEXT:    vmovlt.f32 s0, s4
+; ARMV7-NEXT:    vmov r2, r3, d1
+; ARMV7-NEXT:    vmov r0, r1, d0
 ; ARMV7-NEXT:    bx lr
 ;
 ; ARMV8-LABEL: fminnumv432_intrinsic:
 ; ARMV8:       @ %bb.0:
 ; ARMV8-NEXT:    vldr s0, [sp, #4]
 ; ARMV8-NEXT:    vmov s12, r1
 ; ARMV8-NEXT:    vldr s2, [sp, #8]
 ; ARMV8-NEXT:    vmov s10, r2

 ; ARMV8M-NEXT:    bx lr
   %a = call nnan <4 x float> @llvm.minnum.v4f32(<4 x float> %x, <4 x float><float -1.0, float -1.0, float -1.0, float -1.0>)
   ret <4 x float> %a
 }
 
 define <4 x float> @fminnumv432_one_zero_intrinsic(<4 x float> %x) {
 ; ARMV7-LABEL: fminnumv432_one_zero_intrinsic:
 ; ARMV7:       @ %bb.0:
-; ARMV7-NEXT:    vmov d17, r2, r3
-; ARMV7-NEXT:    vmov d16, r0, r1
-; ARMV7-NEXT:    adr r0, .LCPI18_0
-; ARMV7-NEXT:    vld1.64 {d18, d19}, [r0:128]
-; ARMV7-NEXT:    vmin.f32 q8, q8, q9
-; ARMV7-NEXT:    vmov r0, r1, d16
-; ARMV7-NEXT:    vmov r2, r3, d17
+; ARMV7-NEXT:    vmov d3, r2, r3
+; ARMV7-NEXT:    vmov d2, r0, r1
+; ARMV7-NEXT:    vmov.f32 s0, #-1.000000e+00
+; ARMV7-NEXT:    vcmp.f32 s5, #0
+; ARMV7-NEXT:    vldr s1, .LCPI18_0
+; ARMV7-NEXT:    vmrs APSR_nzcv, fpscr
+; ARMV7-NEXT:    vcmp.f32 s7, s0
+; ARMV7-NEXT:    vmovlt.f32 s1, s5
+; ARMV7-NEXT:    vmrs APSR_nzcv, fpscr
+; ARMV7-NEXT:    vmov.f32 s3, s0
+; ARMV7-NEXT:    vcmp.f32 s6, s0
+; ARMV7-NEXT:    vmovlt.f32 s3, s7
+; ARMV7-NEXT:    vmrs APSR_nzcv, fpscr
+; ARMV7-NEXT:    vmov.f32 s2, s0
+; ARMV7-NEXT:    vcmp.f32 s4, s0
+; ARMV7-NEXT:    vmovlt.f32 s2, s6
+; ARMV7-NEXT:    vmrs APSR_nzcv, fpscr
+; ARMV7-NEXT:    vmovlt.f32 s0, s4
+; ARMV7-NEXT:    vmov r2, r3, d1
+; ARMV7-NEXT:    vmov r0, r1, d0
 ; ARMV7-NEXT:    bx lr
-; ARMV7-NEXT:    .p2align 4
+; ARMV7-NEXT:    .p2align 2
 ; ARMV7-NEXT:  @ %bb.1:
 ; ARMV7-NEXT:  .LCPI18_0:
-; ARMV7-NEXT:    .long 0xbf800000 @ float -1
 ; ARMV7-NEXT:    .long 0x00000000 @ float 0
-; ARMV7-NEXT:    .long 0xbf800000 @ float -1
-; ARMV7-NEXT:    .long 0xbf800000 @ float -1
 ;
 ; ARMV8-LABEL: fminnumv432_one_zero_intrinsic:
 ; ARMV8:       @ %bb.0:
 ; ARMV8-NEXT:    vldr s0, .LCPI18_0
 ; ARMV8-NEXT:    vmov s8, r1
 ; ARMV8-NEXT:    vmov.f32 s2, #-1.000000e+00
 ; ARMV8-NEXT:    vminnm.f32 s0, s8, s0
 ; ARMV8-NEXT:    vmov s6, r2

 ; ARMV8M-NEXT:    .long 0xbf800000 @ float -1
   %a = call nnan <4 x float> @llvm.minnum.v4f32(<4 x float> %x, <4 x float><float -1.0, float 0.0, float -1.0, float -1.0>)
   ret <4 x float> %a
 }
 
 define <4 x float> @fmaxnumv432_intrinsic(<4 x float> %x, <4 x float> %y) {
 ; ARMV7-LABEL: fmaxnumv432_intrinsic:
 ; ARMV7:       @ %bb.0:
-; ARMV7-NEXT:    vmov d17, r2, r3
-; ARMV7-NEXT:    vmov d16, r0, r1
-; ARMV7-NEXT:    mov r0, sp
-; ARMV7-NEXT:    vld1.64 {d18, d19}, [r0]
-; ARMV7-NEXT:    vmax.f32 q8, q8, q9
-; ARMV7-NEXT:    vmov r0, r1, d16
-; ARMV7-NEXT:    vmov r2, r3, d17
+; ARMV7-NEXT:    mov r12, sp
+; ARMV7-NEXT:    vld1.64 {d0, d1}, [r12]
+; ARMV7-NEXT:    vmov d3, r2, r3
+; ARMV7-NEXT:    vmov d2, r0, r1
+; ARMV7-NEXT:    vcmp.f32 s7, s3
+; ARMV7-NEXT:    vmrs APSR_nzcv, fpscr
+; ARMV7-NEXT:    vcmp.f32 s6, s2
+; ARMV7-NEXT:    vmovgt.f32 s3, s7
+; ARMV7-NEXT:    vmrs APSR_nzcv, fpscr
+; ARMV7-NEXT:    vcmp.f32 s5, s1
+; ARMV7-NEXT:    vmovgt.f32 s2, s6
+; ARMV7-NEXT:    vmrs APSR_nzcv, fpscr
+; ARMV7-NEXT:    vcmp.f32 s4, s0
+; ARMV7-NEXT:    vmovgt.f32 s1, s5
+; ARMV7-NEXT:    vmrs APSR_nzcv, fpscr
+; ARMV7-NEXT:    vmovgt.f32 s0, s4
+; ARMV7-NEXT:    vmov r2, r3, d1
+; ARMV7-NEXT:    vmov r0, r1, d0
 ; ARMV7-NEXT:    bx lr
 ;
 ; ARMV8-LABEL: fmaxnumv432_intrinsic:
 ; ARMV8:       @ %bb.0:
 ; ARMV8-NEXT:    vldr s0, [sp, #4]
 ; ARMV8-NEXT:    vmov s12, r1
 ; ARMV8-NEXT:    vldr s2, [sp, #8]
 ; ARMV8-NEXT:    vmov s10, r2

 ; ARMV8M-NEXT:    bx lr
   %a = call nnan nsz <4 x float> @llvm.maxnum.v4f32(<4 x float> %x, <4 x float> %y)
   ret <4 x float> %a
 }
 
 define <4 x float> @fmaxnumv432_zero_intrinsic(<4 x float> %x) {
 ; ARMV7-LABEL: fmaxnumv432_zero_intrinsic:
 ; ARMV7:       @ %bb.0:
-; ARMV7-NEXT:    vmov d19, r2, r3
-; ARMV7-NEXT:    vmov.i32 q8, #0x0
-; ARMV7-NEXT:    vmov d18, r0, r1
-; ARMV7-NEXT:    vmax.f32 q8, q9, q8
-; ARMV7-NEXT:    vmov r0, r1, d16
-; ARMV7-NEXT:    vmov r2, r3, d17
+; ARMV7-NEXT:    vmov d3, r2, r3
+; ARMV7-NEXT:    vldr s0, .LCPI21_0
+; ARMV7-NEXT:    vmov d2, r0, r1
+; ARMV7-NEXT:    vcmp.f32 s7, #0
+; ARMV7-NEXT:    vmrs APSR_nzcv, fpscr
+; ARMV7-NEXT:    vmov.f32 s3, s0
+; ARMV7-NEXT:    vcmp.f32 s6, #0
+; ARMV7-NEXT:    vmovgt.f32 s3, s7
+; ARMV7-NEXT:    vmrs APSR_nzcv, fpscr
+; ARMV7-NEXT:    vmov.f32 s2, s0
+; ARMV7-NEXT:    vcmp.f32 s5, #0
+; ARMV7-NEXT:    vmovgt.f32 s2, s6
+; ARMV7-NEXT:    vmrs APSR_nzcv, fpscr
+; ARMV7-NEXT:    vmov.f32 s1, s0
+; ARMV7-NEXT:    vcmp.f32 s4, #0
+; ARMV7-NEXT:    vmovgt.f32 s1, s5
+; ARMV7-NEXT:    vmrs APSR_nzcv, fpscr
+; ARMV7-NEXT:    vmovgt.f32 s0, s4
+; ARMV7-NEXT:    vmov r2, r3, d1
+; ARMV7-NEXT:    vmov r0, r1, d0
 ; ARMV7-NEXT:    bx lr
+; ARMV7-NEXT:    .p2align 2
+; ARMV7-NEXT:  @ %bb.1:
+; ARMV7-NEXT:  .LCPI21_0:
+; ARMV7-NEXT:    .long 0x00000000 @ float 0
 ;
 ; ARMV8-LABEL: fmaxnumv432_zero_intrinsic:
 ; ARMV8:       @ %bb.0:
 ; ARMV8-NEXT:    vldr s0, .LCPI21_0
 ; ARMV8-NEXT:    vmov s4, r2
 ; ARMV8-NEXT:    vmov s6, r1
 ; ARMV8-NEXT:    vmov s8, r0
 ; ARMV8-NEXT:    vmaxnm.f32 s6, s6, s0

 ; ARMV8M-NEXT:    bx lr
   %a = call nnan <4 x float> @llvm.maxnum.v4f32(<4 x float> %x, <4 x float><float 0.0, float 0.0, float 0.0, float 0.0>)
   ret <4 x float> %a
 }
 
 define <4 x float> @fmaxnumv432_minus_zero_intrinsic(<4 x float> %x) {
 ; ARMV7-LABEL: fmaxnumv432_minus_zero_intrinsic:
 ; ARMV7:       @ %bb.0:
-; ARMV7-NEXT:    vmov d19, r2, r3
-; ARMV7-NEXT:    vmov.i32 q8, #0x80000000
-; ARMV7-NEXT:    vmov d18, r0, r1
-; ARMV7-NEXT:    vmax.f32 q8, q9, q8
-; ARMV7-NEXT:    vmov r0, r1, d16
-; ARMV7-NEXT:    vmov r2, r3, d17
+; ARMV7-NEXT:    vldr s0, .LCPI22_0
+; ARMV7-NEXT:    vmov d3, r2, r3
+; ARMV7-NEXT:    vmov d2, r0, r1
+; ARMV7-NEXT:    vcmp.f32 s7, s0
+; ARMV7-NEXT:    vmrs APSR_nzcv, fpscr
+; ARMV7-NEXT:    vmov.f32 s3, s0
+; ARMV7-NEXT:    vcmp.f32 s6, s0
+; ARMV7-NEXT:    vmovgt.f32 s3, s7
+; ARMV7-NEXT:    vmrs APSR_nzcv, fpscr
+; ARMV7-NEXT:    vmov.f32 s2, s0
+; ARMV7-NEXT:    vcmp.f32 s5, s0
+; ARMV7-NEXT:    vmovgt.f32 s2, s6
+; ARMV7-NEXT:    vmrs APSR_nzcv, fpscr
+; ARMV7-NEXT:    vmov.f32 s1, s0
+; ARMV7-NEXT:    vcmp.f32 s4, s0
+; ARMV7-NEXT:    vmovgt.f32 s1, s5
+; ARMV7-NEXT:    vmrs APSR_nzcv, fpscr
+; ARMV7-NEXT:    vmovgt.f32 s0, s4
+; ARMV7-NEXT:    vmov r2, r3, d1
+; ARMV7-NEXT:    vmov r0, r1, d0
 ; ARMV7-NEXT:    bx lr
+; ARMV7-NEXT:    .p2align 2
+; ARMV7-NEXT:  @ %bb.1:
+; ARMV7-NEXT:  .LCPI22_0:
+; ARMV7-NEXT:    .long 0x80000000 @ float -0
 ;
 ; ARMV8-LABEL: fmaxnumv432_minus_zero_intrinsic:
 ; ARMV8:       @ %bb.0:
 ; ARMV8-NEXT:    vldr s0, .LCPI22_0
 ; ARMV8-NEXT:    vmov s4, r2
 ; ARMV8-NEXT:    vmov s6, r1
 ; ARMV8-NEXT:    vmov s8, r0
 ; ARMV8-NEXT:    vmaxnm.f32 s6, s6, s0

llvm/test/CodeGen/WebAssembly/f32.ll

; CHECK-NEXT:    return $pop0
   ret float %a
 }
 
 declare float @llvm.minnum.f32(float, float)
 define float @fminnum32_intrinsic(float %x, float %y) {
 ; CHECK-LABEL: fminnum32_intrinsic:
 ; CHECK:         .functype fminnum32_intrinsic (f32, f32) -> (f32)
 ; CHECK-NEXT:  # %bb.0:
-; CHECK-NEXT:    local.get $push2=, 0
-; CHECK-NEXT:    local.get $push1=, 1
-; CHECK-NEXT:    f32.min $push0=, $pop2, $pop1
-; CHECK-NEXT:    return $pop0
+; CHECK-NEXT:    local.get $push5=, 0
+; CHECK-NEXT:    local.get $push4=, 1
+; CHECK-NEXT:    local.get $push3=, 0
+; CHECK-NEXT:    local.get $push2=, 1
+; CHECK-NEXT:    f32.lt $push0=, $pop3, $pop2
+; CHECK-NEXT:    f32.select $push1=, $pop5, $pop4, $pop0
+; CHECK-NEXT:    return $pop1
   %a = call nnan float @llvm.minnum.f32(float %x, float %y)
   ret float %a
 }
 
 define float @fminnum32_non_zero_intrinsic(float %x) {
 ; CHECK-LABEL: fminnum32_non_zero_intrinsic:
 ; CHECK:         .functype fminnum32_non_zero_intrinsic (f32) -> (f32)
 ; CHECK-NEXT:  # %bb.0:
; CHECK-NEXT:    return $pop0
   ret float %a
 }
 
 declare float @llvm.maxnum.f32(float, float)
 define float @fmaxnum32_intrinsic(float %x, float %y) {
 ; CHECK-LABEL: fmaxnum32_intrinsic:
 ; CHECK:         .functype fmaxnum32_intrinsic (f32, f32) -> (f32)
 ; CHECK-NEXT:  # %bb.0:
-; CHECK-NEXT:    local.get $push2=, 0
-; CHECK-NEXT:    local.get $push1=, 1
-; CHECK-NEXT:    f32.max $push0=, $pop2, $pop1
-; CHECK-NEXT:    return $pop0
+; CHECK-NEXT:    local.get $push5=, 0
+; CHECK-NEXT:    local.get $push4=, 1
+; CHECK-NEXT:    local.get $push3=, 0
+; CHECK-NEXT:    local.get $push2=, 1
+; CHECK-NEXT:    f32.gt $push0=, $pop3, $pop2
+; CHECK-NEXT:    f32.select $push1=, $pop5, $pop4, $pop0
+; CHECK-NEXT:    return $pop1
   %a = call nnan float @llvm.maxnum.f32(float %x, float %y)
   ret float %a
 }
 
 define float @fmaxnum32_nsz_intrinsic(float %x, float %y) {
 ; CHECK-LABEL: fmaxnum32_nsz_intrinsic:
 ; CHECK:         .functype fmaxnum32_nsz_intrinsic (f32, f32) -> (f32)
 ; CHECK-NEXT:  # %bb.0:
 ; CHECK-NEXT:    local.get $push2=, 0
 ; CHECK-NEXT:    local.get $push1=, 1
 ; CHECK-NEXT:    f32.max $push0=, $pop2, $pop1
 ; CHECK-NEXT:    return $pop0
   %a = call nnan nsz float @llvm.maxnum.f32(float %x, float %y)
   ret float %a
 }
 
 define float @fmaxnum32_zero_intrinsic(float %x) {
 ; CHECK-LABEL: fmaxnum32_zero_intrinsic:
 ; CHECK:         .functype fmaxnum32_zero_intrinsic (f32) -> (f32)
 ; CHECK-NEXT:  # %bb.0:
-; CHECK-NEXT:    local.get $push2=, 0
+; CHECK-NEXT:    local.get $push5=, 0
 ; CHECK-NEXT:    f32.const $push0=, 0x0p0
-; CHECK-NEXT:    f32.max $push1=, $pop2, $pop0
-; CHECK-NEXT:    return $pop1
+; CHECK-NEXT:    local.get $push4=, 0
+; CHECK-NEXT:    f32.const $push3=, 0x0p0
+; CHECK-NEXT:    f32.gt $push1=, $pop4, $pop3
+; CHECK-NEXT:    f32.select $push2=, $pop5, $pop0, $pop1
+; CHECK-NEXT:    return $pop2
   %a = call nnan float @llvm.maxnum.f32(float %x, float 0.0)
   ret float %a
 }
 
 define float @fmaxnum32_non_zero_intrinsic(float %x) {
 ; CHECK-LABEL: fmaxnum32_non_zero_intrinsic:
 ; CHECK:         .functype fmaxnum32_non_zero_intrinsic (f32) -> (f32)
 ; CHECK-NEXT:  # %bb.0:

llvm/test/CodeGen/WebAssembly/f64.ll

; CHECK-NEXT:    return $pop0
   ret double %a
 }
 
 declare double @llvm.minnum.f64(double, double)
 define double @fminnum64_intrinsic(double %x, double %y) {
 ; CHECK-LABEL: fminnum64_intrinsic:
 ; CHECK:         .functype fminnum64_intrinsic (f64, f64) -> (f64)
 ; CHECK-NEXT:  # %bb.0:
-; CHECK-NEXT:    local.get $push2=, 0
-; CHECK-NEXT:    local.get $push1=, 1
-; CHECK-NEXT:    f64.min $push0=, $pop2, $pop1
-; CHECK-NEXT:    return $pop0
+; CHECK-NEXT:    local.get $push5=, 0
+; CHECK-NEXT:    local.get $push4=, 1
+; CHECK-NEXT:    local.get $push3=, 0
+; CHECK-NEXT:    local.get $push2=, 1
+; CHECK-NEXT:    f64.lt $push0=, $pop3, $pop2
+; CHECK-NEXT:    f64.select $push1=, $pop5, $pop4, $pop0
+; CHECK-NEXT:    return $pop1
   %a = call nnan double @llvm.minnum.f64(double %x, double %y)
   ret double %a
 }
 
 define double @fminnum64_nsz_intrinsic(double %x, double %y) {
 ; CHECK-LABEL: fminnum64_nsz_intrinsic:
 ; CHECK:         .functype fminnum64_nsz_intrinsic (f64, f64) -> (f64)
 ; CHECK-NEXT:  # %bb.0:
 ; CHECK-NEXT:    local.get $push2=, 0
 ; CHECK-NEXT:    local.get $push1=, 1
 ; CHECK-NEXT:    f64.min $push0=, $pop2, $pop1
 ; CHECK-NEXT:    return $pop0
   %a = call nnan nsz double @llvm.minnum.f64(double %x, double %y)
   ret double %a
 }
 
 define double @fminnum64_zero_intrinsic(double %x) {
 ; CHECK-LABEL: fminnum64_zero_intrinsic:
 ; CHECK:         .functype fminnum64_zero_intrinsic (f64) -> (f64)
 ; CHECK-NEXT:  # %bb.0:
-; CHECK-NEXT:    local.get $push2=, 0
+; CHECK-NEXT:    local.get $push5=, 0
 ; CHECK-NEXT:    f64.const $push0=, -0x0p0
-; CHECK-NEXT:    f64.min $push1=, $pop2, $pop0
-; CHECK-NEXT:    return $pop1
+; CHECK-NEXT:    local.get $push4=, 0
+; CHECK-NEXT:    f64.const $push3=, -0x0p0
+; CHECK-NEXT:    f64.lt $push1=, $pop4, $pop3
+; CHECK-NEXT:    f64.select $push2=, $pop5, $pop0, $pop1
+; CHECK-NEXT:    return $pop2
   %a = call nnan double @llvm.minnum.f64(double %x, double -0.0)
   ret double %a
 }
 
 define double @fminnum64_non_zero_intrinsic(double %x) {
 ; CHECK-LABEL: fminnum64_non_zero_intrinsic:
 ; CHECK:         .functype fminnum64_non_zero_intrinsic (f64) -> (f64)
 ; CHECK-NEXT:  # %bb.0:
; CHECK-NEXT:    return $pop0
   ret double %a
 }
 
 declare double @llvm.maxnum.f64(double, double)
 define double@fmaxnum64_intrinsic(double %x, double %y) {
 ; CHECK-LABEL: fmaxnum64_intrinsic:
 ; CHECK:         .functype fmaxnum64_intrinsic (f64, f64) -> (f64)
 ; CHECK-NEXT:  # %bb.0:
-; CHECK-NEXT:    local.get $push2=, 0
-; CHECK-NEXT:    local.get $push1=, 1
-; CHECK-NEXT:    f64.max $push0=, $pop2, $pop1
-; CHECK-NEXT:    return $pop0
+; CHECK-NEXT:    local.get $push5=, 0
+; CHECK-NEXT:    local.get $push4=, 1
+; CHECK-NEXT:    local.get $push3=, 0
+; CHECK-NEXT:    local.get $push2=, 1
+; CHECK-NEXT:    f64.gt $push0=, $pop3, $pop2
+; CHECK-NEXT:    f64.select $push1=, $pop5, $pop4, $pop0
+; CHECK-NEXT:    return $pop1
   %a = call nnan double @llvm.maxnum.f64(double %x, double %y)
   ret double %a
 }
 
 define double@fmaxnum64_nsz_intrinsic(double %x, double %y) {
 ; CHECK-LABEL: fmaxnum64_nsz_intrinsic:
 ; CHECK:         .functype fmaxnum64_nsz_intrinsic (f64, f64) -> (f64)
 ; CHECK-NEXT:  # %bb.0:
 ; CHECK-NEXT:    local.get $push2=, 0
 ; CHECK-NEXT:    local.get $push1=, 1
 ; CHECK-NEXT:    f64.max $push0=, $pop2, $pop1
 ; CHECK-NEXT:    return $pop0
   %a = call nnan nsz double @llvm.maxnum.f64(double %x, double %y)
   ret double %a
 }
 
 define double @fmaxnum64_zero_intrinsic(double %x) {
 ; CHECK-LABEL: fmaxnum64_zero_intrinsic:
 ; CHECK:         .functype fmaxnum64_zero_intrinsic (f64) -> (f64)
 ; CHECK-NEXT:  # %bb.0:
-; CHECK-NEXT:    local.get $push2=, 0
+; CHECK-NEXT:    local.get $push5=, 0
 ; CHECK-NEXT:    f64.const $push0=, 0x0p0
-; CHECK-NEXT:    f64.max $push1=, $pop2, $pop0
-; CHECK-NEXT:    return $pop1
+; CHECK-NEXT:    local.get $push4=, 0
+; CHECK-NEXT:    f64.const $push3=, 0x0p0
+; CHECK-NEXT:    f64.gt $push1=, $pop4, $pop3
+; CHECK-NEXT:    f64.select $push2=, $pop5, $pop0, $pop1
+; CHECK-NEXT:    return $pop2
   %a = call nnan double @llvm.maxnum.f64(double %x, double 0.0)
   ret double %a
 }
 
 define double @fmaxnum64_non_zero_intrinsic(double %x) {
 ; CHECK-LABEL: fmaxnum64_non_zero_intrinsic:
 ; CHECK:         .functype fmaxnum64_non_zero_intrinsic (f64) -> (f64)
 ; CHECK-NEXT:  # %bb.0:

llvm/test/CodeGen/WebAssembly/simd-arith.ll

; NO-SIMD128-FAST-NEXT:    return
   ret <4 x float> %a
 }
 
 declare <4 x float> @llvm.minnum.v4f32(<4 x float>, <4 x float>)
 define <4 x float> @minnum_intrinsic_v4f32(<4 x float> %x, <4 x float> %y) {
 ; SIMD128-LABEL: minnum_intrinsic_v4f32:
 ; SIMD128:         .functype minnum_intrinsic_v4f32 (v128, v128) -> (v128)
 ; SIMD128-NEXT:  # %bb.0:
-; SIMD128-NEXT:    f32x4.min $push0=, $0, $1
-; SIMD128-NEXT:    return $pop0
+; SIMD128-NEXT:    f32x4.extract_lane $push27=, $0, 0
+; SIMD128-NEXT:    local.tee $push26=, $3=, $pop27
+; SIMD128-NEXT:    f32x4.extract_lane $push25=, $1, 0
+; SIMD128-NEXT:    local.tee $push24=, $2=, $pop25
+; SIMD128-NEXT:    f32.lt $push2=, $3, $2
+; SIMD128-NEXT:    f32.select $push3=, $pop26, $pop24, $pop2
+; SIMD128-NEXT:    f32x4.splat $push4=, $pop3
+; SIMD128-NEXT:    f32x4.extract_lane $push23=, $0, 1
+; SIMD128-NEXT:    local.tee $push22=, $3=, $pop23
+; SIMD128-NEXT:    f32x4.extract_lane $push21=, $1, 1
+; SIMD128-NEXT:    local.tee $push20=, $2=, $pop21
+; SIMD128-NEXT:    f32.lt $push0=, $3, $2
+; SIMD128-NEXT:    f32.select $push1=, $pop22, $pop20, $pop0
+; SIMD128-NEXT:    f32x4.replace_lane $push5=, $pop4, 1, $pop1
+; SIMD128-NEXT:    f32x4.extract_lane $push19=, $0, 2
+; SIMD128-NEXT:    local.tee $push18=, $3=, $pop19
+; SIMD128-NEXT:    f32x4.extract_lane $push17=, $1, 2
+; SIMD128-NEXT:    local.tee $push16=, $2=, $pop17
+; SIMD128-NEXT:    f32.lt $push6=, $3, $2
+; SIMD128-NEXT:    f32.select $push7=, $pop18, $pop16, $pop6
+; SIMD128-NEXT:    f32x4.replace_lane $push8=, $pop5, 2, $pop7
+; SIMD128-NEXT:    f32x4.extract_lane $push15=, $0, 3
+; SIMD128-NEXT:    local.tee $push14=, $3=, $pop15
+; SIMD128-NEXT:    f32x4.extract_lane $push13=, $1, 3
+; SIMD128-NEXT:    local.tee $push12=, $2=, $pop13
+; SIMD128-NEXT:    f32.lt $push9=, $3, $2
+; SIMD128-NEXT:    f32.select $push10=, $pop14, $pop12, $pop9
+; SIMD128-NEXT:    f32x4.replace_lane $push11=, $pop8, 3, $pop10
+; SIMD128-NEXT:    return $pop11
 ;
 ; SIMD128-FAST-LABEL: minnum_intrinsic_v4f32:
 ; SIMD128-FAST:         .functype minnum_intrinsic_v4f32 (v128, v128) -> (v128)
 ; SIMD128-FAST-NEXT:  # %bb.0:
-; SIMD128-FAST-NEXT:    f32x4.min $push0=, $0, $1
+; SIMD128-FAST-NEXT:    f32x4.extract_lane $push27=, $0, 0
+; SIMD128-FAST-NEXT:    local.tee $push26=, $3=, $pop27
+; SIMD128-FAST-NEXT:    f32x4.extract_lane $push25=, $1, 0
+; SIMD128-FAST-NEXT:    local.tee $push24=, $2=, $pop25
+; SIMD128-FAST-NEXT:    f32.lt $push3=, $3, $2
+; SIMD128-FAST-NEXT:    f32.select $push4=, $pop26, $pop24, $pop3
+; SIMD128-FAST-NEXT:    f32x4.splat $push5=, $pop4
+; SIMD128-FAST-NEXT:    f32x4.extract_lane $push23=, $0, 1
+; SIMD128-FAST-NEXT:    local.tee $push22=, $3=, $pop23
+; SIMD128-FAST-NEXT:    f32x4.extract_lane $push21=, $1, 1
+; SIMD128-FAST-NEXT:    local.tee $push20=, $2=, $pop21
+; SIMD128-FAST-NEXT:    f32.lt $push1=, $3, $2
+; SIMD128-FAST-NEXT:    f32.select $push2=, $pop22, $pop20, $pop1
+; SIMD128-FAST-NEXT:    f32x4.replace_lane $push6=, $pop5, 1, $pop2
+; SIMD128-FAST-NEXT:    f32x4.extract_lane $push19=, $0, 2
+; SIMD128-FAST-NEXT:    local.tee $push18=, $3=, $pop19
+; SIMD128-FAST-NEXT:    f32x4.extract_lane $push17=, $1, 2
+; SIMD128-FAST-NEXT:    local.tee $push16=, $2=, $pop17
+; SIMD128-FAST-NEXT:    f32.lt $push7=, $3, $2
+; SIMD128-FAST-NEXT:    f32.select $push8=, $pop18, $pop16, $pop7
+; SIMD128-FAST-NEXT:    f32x4.replace_lane $push9=, $pop6, 2, $pop8
+; SIMD128-FAST-NEXT:    f32x4.extract_lane $push15=, $0, 3
+; SIMD128-FAST-NEXT:    local.tee $push14=, $3=, $pop15
+; SIMD128-FAST-NEXT:    f32x4.extract_lane $push13=, $1, 3
+; SIMD128-FAST-NEXT:    local.tee $push12=, $2=, $pop13
+; SIMD128-FAST-NEXT:    f32.lt $push10=, $3, $2
+; SIMD128-FAST-NEXT:    f32.select $push11=, $pop14, $pop12, $pop10
+; SIMD128-FAST-NEXT:    f32x4.replace_lane $push0=, $pop9, 3, $pop11
 ; SIMD128-FAST-NEXT:    return $pop0
 ;
 ; NO-SIMD128-LABEL: minnum_intrinsic_v4f32:
 ; NO-SIMD128:         .functype minnum_intrinsic_v4f32 (i32, f32, f32, f32, f32, f32, f32, f32, f32) -> ()
 ; NO-SIMD128-NEXT:  # %bb.0:
 ; NO-SIMD128-NEXT:    call $push0=, fminf, $3, $7
 ; NO-SIMD128-NEXT:    f32.store 8($0), $pop0
 ; NO-SIMD128-NEXT:    call $push1=, fminf, $2, $6
; NO-SIMD128-FAST-NEXT:    return
   %a = call nnan <4 x float> @llvm.minnum.v4f32(<4 x float> %x, <4 x float><float -1.0, float -1.0, float -1.0, float -1.0>)
   ret <4 x float> %a
 }
 
 define <4 x float> @fminnumv432_one_zero_intrinsic(<4 x float> %x) {
 ; SIMD128-LABEL: fminnumv432_one_zero_intrinsic:
 ; SIMD128:         .functype fminnumv432_one_zero_intrinsic (v128) -> (v128)
 ; SIMD128-NEXT:  # %bb.0:
-; SIMD128-NEXT:    v128.const $push0=, -0x1p0, 0x0p0, -0x1p0, -0x1p0
-; SIMD128-NEXT:    f32x4.min $push1=, $0, $pop0
-; SIMD128-NEXT:    return $pop1
+; SIMD128-NEXT:    f32x4.extract_lane $push27=, $0, 0
+; SIMD128-NEXT:    local.tee $push26=, $1=, $pop27
+; SIMD128-NEXT:    f32.const $push3=, -0x1p0
+; SIMD128-NEXT:    f32.const $push25=, -0x1p0
+; SIMD128-NEXT:    f32.lt $push4=, $1, $pop25
+; SIMD128-NEXT:    f32.select $push5=, $pop26, $pop3, $pop4
+; SIMD128-NEXT:    f32x4.splat $push6=, $pop5
+; SIMD128-NEXT:    f32x4.extract_lane $push24=, $0, 1
+; SIMD128-NEXT:    local.tee $push23=, $1=, $pop24
+; SIMD128-NEXT:    f32.const $push0=, 0x0p0
+; SIMD128-NEXT:    f32.const $push22=, 0x0p0
+; SIMD128-NEXT:    f32.lt $push1=, $1, $pop22
+; SIMD128-NEXT:    f32.select $push2=, $pop23, $pop0, $pop1
+; SIMD128-NEXT:    f32x4.replace_lane $push7=, $pop6, 1, $pop2
+; SIMD128-NEXT:    f32x4.extract_lane $push21=, $0, 2
+; SIMD128-NEXT:    local.tee $push20=, $1=, $pop21
+; SIMD128-NEXT:    f32.const $push19=, -0x1p0
+; SIMD128-NEXT:    f32.const $push18=, -0x1p0
+; SIMD128-NEXT:    f32.lt $push8=, $1, $pop18
+; SIMD128-NEXT:    f32.select $push9=, $pop20, $pop19, $pop8
+; SIMD128-NEXT:    f32x4.replace_lane $push10=, $pop7, 2, $pop9
+; SIMD128-NEXT:    f32x4.extract_lane $push17=, $0, 3
+; SIMD128-NEXT:    local.tee $push16=, $1=, $pop17
+; SIMD128-NEXT:    f32.const $push15=, -0x1p0
+; SIMD128-NEXT:    f32.const $push14=, -0x1p0
+; SIMD128-NEXT:    f32.lt $push11=, $1, $pop14
+; SIMD128-NEXT:    f32.select $push12=, $pop16, $pop15, $pop11
+; SIMD128-NEXT:    f32x4.replace_lane $push13=, $pop10, 3, $pop12
+; SIMD128-NEXT:    return $pop13
 ;
 ; SIMD128-FAST-LABEL: fminnumv432_one_zero_intrinsic:
 ; SIMD128-FAST:         .functype fminnumv432_one_zero_intrinsic (v128) -> (v128)
 ; SIMD128-FAST-NEXT:  # %bb.0:
-; SIMD128-FAST-NEXT:    v128.const $push1=, -0x1p0, 0x0p0, -0x1p0, -0x1p0
-; SIMD128-FAST-NEXT:    f32x4.min $push0=, $0, $pop1
+; SIMD128-FAST-NEXT:    f32x4.extract_lane $push27=, $0, 0
+; SIMD128-FAST-NEXT:    local.tee $push26=, $1=, $pop27
+; SIMD128-FAST-NEXT:    f32.const $push4=, -0x1p0
+; SIMD128-FAST-NEXT:    f32.const $push25=, -0x1p0
+; SIMD128-FAST-NEXT:    f32.lt $push5=, $1, $pop25
+; SIMD128-FAST-NEXT:    f32.select $push6=, $pop26, $pop4, $pop5
+; SIMD128-FAST-NEXT:    f32x4.splat $push7=, $pop6
+; SIMD128-FAST-NEXT:    f32x4.extract_lane $push24=, $0, 1
+; SIMD128-FAST-NEXT:    local.tee $push23=, $1=, $pop24
+; SIMD128-FAST-NEXT:    f32.const $push1=, 0x0p0
+; SIMD128-FAST-NEXT:    f32.const $push22=, 0x0p0
+; SIMD128-FAST-NEXT:    f32.lt $push2=, $1, $pop22
+; SIMD128-FAST-NEXT:    f32.select $push3=, $pop23, $pop1, $pop2
+; SIMD128-FAST-NEXT:    f32x4.replace_lane $push8=, $pop7, 1, $pop3
+; SIMD128-FAST-NEXT:    f32x4.extract_lane $push21=, $0, 2
+; SIMD128-FAST-NEXT:    local.tee $push20=, $1=, $pop21
+; SIMD128-FAST-NEXT:    f32.const $push19=, -0x1p0
+; SIMD128-FAST-NEXT:    f32.const $push18=, -0x1p0
+; SIMD128-FAST-NEXT:    f32.lt $push9=, $1, $pop18
+; SIMD128-FAST-NEXT:    f32.select $push10=, $pop20, $pop19, $pop9
+; SIMD128-FAST-NEXT:    f32x4.replace_lane $push11=, $pop8, 2, $pop10
+; SIMD128-FAST-NEXT:    f32x4.extract_lane $push17=, $0, 3
+; SIMD128-FAST-NEXT:    local.tee $push16=, $1=, $pop17
+; SIMD128-FAST-NEXT:    f32.const $push15=, -0x1p0
+; SIMD128-FAST-NEXT:    f32.const $push14=, -0x1p0
+; SIMD128-FAST-NEXT:    f32.lt $push12=, $1, $pop14
+; SIMD128-FAST-NEXT:    f32.select $push13=, $pop16, $pop15, $pop12
+; SIMD128-FAST-NEXT:    f32x4.replace_lane $push0=, $pop11, 3, $pop13
 ; SIMD128-FAST-NEXT:    return $pop0
 ;
 ; NO-SIMD128-LABEL: fminnumv432_one_zero_intrinsic:
 ; NO-SIMD128:         .functype fminnumv432_one_zero_intrinsic (i32, f32, f32, f32, f32) -> ()
 ; NO-SIMD128-NEXT:  # %bb.0:
 ; NO-SIMD128-NEXT:    f32.const $push0=, -0x1p0
 ; NO-SIMD128-NEXT:    call $push1=, fminf, $3, $pop0
 ; NO-SIMD128-NEXT:    f32.store 8($0), $pop1
; NO-SIMD128-FAST-NEXT:    return
   ret <4 x float> %a
 }
 
 declare <4 x float> @llvm.maxnum.v4f32(<4 x float>, <4 x float>)
 define <4 x float> @maxnum_intrinsic_v4f32(<4 x float> %x, <4 x float> %y) {
 ; SIMD128-LABEL: maxnum_intrinsic_v4f32:
 ; SIMD128:         .functype maxnum_intrinsic_v4f32 (v128, v128) -> (v128)
 ; SIMD128-NEXT:  # %bb.0:
-; SIMD128-NEXT:    f32x4.max $push0=, $0, $1
-; SIMD128-NEXT:    return $pop0
+; SIMD128-NEXT:    f32x4.extract_lane $push27=, $0, 0
+; SIMD128-NEXT:    local.tee $push26=, $3=, $pop27
+; SIMD128-NEXT:    f32x4.extract_lane $push25=, $1, 0
+; SIMD128-NEXT:    local.tee $push24=, $2=, $pop25
+; SIMD128-NEXT:    f32.gt $push2=, $3, $2
+; SIMD128-NEXT:    f32.select $push3=, $pop26, $pop24, $pop2
+; SIMD128-NEXT:    f32x4.splat $push4=, $pop3
+; SIMD128-NEXT:    f32x4.extract_lane $push23=, $0, 1
+; SIMD128-NEXT:    local.tee $push22=, $3=, $pop23
+; SIMD128-NEXT:    f32x4.extract_lane $push21=, $1, 1
+; SIMD128-NEXT:    local.tee $push20=, $2=, $pop21
+; SIMD128-NEXT:    f32.gt $push0=, $3, $2
+; SIMD128-NEXT:    f32.select $push1=, $pop22, $pop20, $pop0
+; SIMD128-NEXT:    f32x4.replace_lane $push5=, $pop4, 1, $pop1
+; SIMD128-NEXT:    f32x4.extract_lane $push19=, $0, 2
+; SIMD128-NEXT:    local.tee $push18=, $3=, $pop19
+; SIMD128-NEXT:    f32x4.extract_lane $push17=, $1, 2
+; SIMD128-NEXT:    local.tee $push16=, $2=, $pop17
+; SIMD128-NEXT:    f32.gt $push6=, $3, $2
+; SIMD128-NEXT:    f32.select $push7=, $pop18, $pop16, $pop6
+; SIMD128-NEXT:    f32x4.replace_lane $push8=, $pop5, 2, $pop7
+; SIMD128-NEXT:    f32x4.extract_lane $push15=, $0, 3
+; SIMD128-NEXT:    local.tee $push14=, $3=, $pop15
+; SIMD128-NEXT:    f32x4.extract_lane $push13=, $1, 3
+; SIMD128-NEXT:    local.tee $push12=, $2=, $pop13
+; SIMD128-NEXT:    f32.gt $push9=, $3, $2
+; SIMD128-NEXT:    f32.select $push10=, $pop14, $pop12, $pop9
+; SIMD128-NEXT:    f32x4.replace_lane $push11=, $pop8, 3, $pop10
+; SIMD128-NEXT:    return $pop11
 ;
 ; SIMD128-FAST-LABEL: maxnum_intrinsic_v4f32:
 ; SIMD128-FAST:         .functype maxnum_intrinsic_v4f32 (v128, v128) -> (v128)
 ; SIMD128-FAST-NEXT:  # %bb.0:
-; SIMD128-FAST-NEXT:    f32x4.max $push0=, $0, $1
+; SIMD128-FAST-NEXT:    f32x4.extract_lane $push27=, $0, 0
+; SIMD128-FAST-NEXT:    local.tee $push26=, $3=, $pop27
+; SIMD128-FAST-NEXT:    f32x4.extract_lane $push25=, $1, 0
+; SIMD128-FAST-NEXT:    local.tee $push24=, $2=, $pop25
+; SIMD128-FAST-NEXT:    f32.gt $push3=, $3, $2
+; SIMD128-FAST-NEXT:    f32.select $push4=, $pop26, $pop24, $pop3
+; SIMD128-FAST-NEXT:    f32x4.splat $push5=, $pop4
+; SIMD128-FAST-NEXT:    f32x4.extract_lane $push23=, $0, 1
+; SIMD128-FAST-NEXT:    local.tee $push22=, $3=, $pop23
+; SIMD128-FAST-NEXT:    f32x4.extract_lane $push21=, $1, 1
+; SIMD128-FAST-NEXT:    local.tee $push20=, $2=, $pop21
+; SIMD128-FAST-NEXT:    f32.gt $push1=, $3, $2
+; SIMD128-FAST-NEXT:    f32.select $push2=, $pop22, $pop20, $pop1
+; SIMD128-FAST-NEXT:    f32x4.replace_lane $push6=, $pop5, 1, $pop2
+; SIMD128-FAST-NEXT:    f32x4.extract_lane $push19=, $0, 2
+; SIMD128-FAST-NEXT:    local.tee $push18=, $3=, $pop19
+; SIMD128-FAST-NEXT:    f32x4.extract_lane $push17=, $1, 2
+; SIMD128-FAST-NEXT:    local.tee $push16=, $2=, $pop17
+; SIMD128-FAST-NEXT:    f32.gt $push7=, $3, $2
+; SIMD128-FAST-NEXT:    f32.select $push8=, $pop18, $pop16, $pop7
+; SIMD128-FAST-NEXT:    f32x4.replace_lane $push9=, $pop6, 2, $pop8
+; SIMD128-FAST-NEXT:    f32x4.extract_lane $push15=, $0, 3
+; SIMD128-FAST-NEXT:    local.tee $push14=, $3=, $pop15
+; SIMD128-FAST-NEXT:    f32x4.extract_lane $push13=, $1, 3
+; SIMD128-FAST-NEXT:    local.tee $push12=, $2=, $pop13
+; SIMD128-FAST-NEXT:    f32.gt $push10=, $3, $2
+; SIMD128-FAST-NEXT:    f32.select $push11=, $pop14, $pop12, $pop10
+; SIMD128-FAST-NEXT:    f32x4.replace_lane $push0=, $pop9, 3, $pop11
 ; SIMD128-FAST-NEXT:    return $pop0
 ;
 ; NO-SIMD128-LABEL: maxnum_intrinsic_v4f32:
 ; NO-SIMD128:         .functype maxnum_intrinsic_v4f32 (i32, f32, f32, f32, f32, f32, f32, f32, f32) -> ()
 ; NO-SIMD128-NEXT:  # %bb.0:
 ; NO-SIMD128-NEXT:    call $push0=, fmaxf, $3, $7
 ; NO-SIMD128-NEXT:    f32.store 8($0), $pop0
 ; NO-SIMD128-NEXT:    call $push1=, fmaxf, $2, $6
; NO-SIMD128-FAST-NEXT:    return
   %a = call nnan nsz <4 x float> @llvm.maxnum.v4f32(<4 x float> %x, <4 x float> %y)
   ret <4 x float> %a
 }
 
 define <4 x float> @maxnum_one_zero_intrinsic_v4f32(<4 x float> %x, <4 x float> %y) {
 ; SIMD128-LABEL: maxnum_one_zero_intrinsic_v4f32:
 ; SIMD128:         .functype maxnum_one_zero_intrinsic_v4f32 (v128, v128) -> (v128)
 ; SIMD128-NEXT:  # %bb.0:
-; SIMD128-NEXT:    v128.const $push0=, -0x1p0, 0x0p0, -0x1p0, -0x1p0
-; SIMD128-NEXT:    f32x4.max $push1=, $0, $pop0
-; SIMD128-NEXT:    return $pop1
+; SIMD128-NEXT:    f32x4.extract_lane $push27=, $0, 0
+; SIMD128-NEXT:    local.tee $push26=, $2=, $pop27
+; SIMD128-NEXT:    f32.const $push3=, -0x1p0
+; SIMD128-NEXT:    f32.const $push25=, -0x1p0
+; SIMD128-NEXT:    f32.gt $push4=, $2, $pop25
+; SIMD128-NEXT:    f32.select $push5=, $pop26, $pop3, $pop4
+; SIMD128-NEXT:    f32x4.splat $push6=, $pop5
+; SIMD128-NEXT:    f32x4.extract_lane $push24=, $0, 1
+; SIMD128-NEXT:    local.tee $push23=, $2=, $pop24
+; SIMD128-NEXT:    f32.const $push0=, 0x0p0
+; SIMD128-NEXT:    f32.const $push22=, 0x0p0
+; SIMD128-NEXT:    f32.gt $push1=, $2, $pop22
+; SIMD128-NEXT:    f32.select $push2=, $pop23, $pop0, $pop1
+; SIMD128-NEXT:    f32x4.replace_lane $push7=, $pop6, 1, $pop2
+; SIMD128-NEXT:    f32x4.extract_lane $push21=, $0, 2
+; SIMD128-NEXT:    local.tee $push20=, $2=, $pop21
+; SIMD128-NEXT:    f32.const $push19=, -0x1p0
+; SIMD128-NEXT:    f32.const $push18=, -0x1p0
+; SIMD128-NEXT:    f32.gt $push8=, $2, $pop18
+; SIMD128-NEXT:    f32.select $push9=, $pop20, $pop19, $pop8
+; SIMD128-NEXT:    f32x4.replace_lane $push10=, $pop7, 2, $pop9
+; SIMD128-NEXT:    f32x4.extract_lane $push17=, $0, 3
+; SIMD128-NEXT:    local.tee $push16=, $2=, $pop17
+; SIMD128-NEXT:    f32.const $push15=, -0x1p0
+; SIMD128-NEXT:    f32.const $push14=, -0x1p0
+; SIMD128-NEXT:    f32.gt $push11=, $2, $pop14
+; SIMD128-NEXT:    f32.select $push12=, $pop16, $pop15, $pop11
+; SIMD128-NEXT:    f32x4.replace_lane $push13=, $pop10, 3, $pop12
+; SIMD128-NEXT:    return $pop13
 ;
 ; SIMD128-FAST-LABEL: maxnum_one_zero_intrinsic_v4f32:
 ; SIMD128-FAST:         .functype maxnum_one_zero_intrinsic_v4f32 (v128, v128) -> (v128)
 ; SIMD128-FAST-NEXT:  # %bb.0:
-; SIMD128-FAST-NEXT:    v128.const $push1=, -0x1p0, 0x0p0, -0x1p0, -0x1p0
-; SIMD128-FAST-NEXT:    f32x4.max $push0=, $0, $pop1
+; SIMD128-FAST-NEXT:    f32x4.extract_lane $push27=, $0, 0
+; SIMD128-FAST-NEXT:    local.tee $push26=, $2=, $pop27
+; SIMD128-FAST-NEXT:    f32.const $push4=, -0x1p0
+; SIMD128-FAST-NEXT:    f32.const $push25=, -0x1p0
+; SIMD128-FAST-NEXT:    f32.gt $push5=, $2, $pop25
+; SIMD128-FAST-NEXT:    f32.select $push6=, $pop26, $pop4, $pop5
+; SIMD128-FAST-NEXT:    f32x4.splat $push7=, $pop6
+; SIMD128-FAST-NEXT:    f32x4.extract_lane $push24=, $0, 1
+; SIMD128-FAST-NEXT:    local.tee $push23=, $2=, $pop24
+; SIMD128-FAST-NEXT:    f32.const $push1=, 0x0p0
+; SIMD128-FAST-NEXT:    f32.const $push22=, 0x0p0
+; SIMD128-FAST-NEXT:    f32.gt $push2=, $2, $pop22
+; SIMD128-FAST-NEXT:    f32.select $push3=, $pop23, $pop1, $pop2
+; SIMD128-FAST-NEXT:    f32x4.replace_lane $push8=, $pop7, 1, $pop3
+; SIMD128-FAST-NEXT:    f32x4.extract_lane $push21=, $0, 2
+; SIMD128-FAST-NEXT:    local.tee $push20=, $2=, $pop21
+; SIMD128-FAST-NEXT:    f32.const $push19=, -0x1p0
+; SIMD128-FAST-NEXT:    f32.const $push18=, -0x1p0
+; SIMD128-FAST-NEXT:    f32.gt $push9=, $2, $pop18
+; SIMD128-FAST-NEXT:    f32.select $push10=, $pop20, $pop19, $pop9
+; SIMD128-FAST-NEXT:    f32x4.replace_lane $push11=, $pop8, 2, $pop10
+; SIMD128-FAST-NEXT:    f32x4.extract_lane $push17=, $0, 3
+; SIMD128-FAST-NEXT:    local.tee $push16=, $2=, $pop17
+; SIMD128-FAST-NEXT:    f32.const $push15=, -0x1p0
+; SIMD128-FAST-NEXT:    f32.const $push14=, -0x1p0
+; SIMD128-FAST-NEXT:    f32.gt $push12=, $2, $pop14
+; SIMD128-FAST-NEXT:    f32.select $push13=, $pop16, $pop15, $pop12
+; SIMD128-FAST-NEXT:    f32x4.replace_lane $push0=, $pop11, 3, $pop13
 ; SIMD128-FAST-NEXT:    return $pop0
 ;
 ; NO-SIMD128-LABEL: maxnum_one_zero_intrinsic_v4f32:
 ; NO-SIMD128:         .functype maxnum_one_zero_intrinsic_v4f32 (i32, f32, f32, f32, f32, f32, f32, f32, f32) -> ()
 ; NO-SIMD128-NEXT:  # %bb.0:
 ; NO-SIMD128-NEXT:    f32.const $push0=, -0x1p0
 ; NO-SIMD128-NEXT:    call $push1=, fmaxf, $3, $pop0
 ; NO-SIMD128-NEXT:    f32.store 8($0), $pop1