AMDGPU/GlobalISel: Calculate isKnownNeverNaN for fminnum and fmaxnum

[Petar.Avramovic]

Summary

Implements same logis as in SelectionDAG.
G_FMINNUM_IEEE and G_FMAXNUM_IEEE are never SNaN by definition and
never NaN when one operand is known non-NaN and other known non-SNaN.
G_FMINNUM_IEEE and G_FMAXNUM_IEEE are never NaN/SNaN when one of the
operands is known non-NaN/SNaN.

[arsenm]

llvm/lib/CodeGen/GlobalISel/Utils.cpp

514–520

Generic opcodes should not have target dependent semantics like this, especially since these already have semantics inherited from the IR/C fmin/fmax definition. They return the non-nan argument regardless of snan or not

[Petar.Avramovic]

llvm/lib/CodeGen/GlobalISel/Utils.cpp

518–519

Can this then be return true?

[arsenm]

llvm/lib/CodeGen/GlobalISel/Utils.cpp

518–519

No, it depends on the inputs. You can refer to the existing SelectionDAG::isKnownNeverNaN:

// Only one needs to be known not-nan, since it will be returned if the
// other ends up being one.
return isKnownNeverNaN(Op.getOperand(0), SNaN, Depth + 1) ||
       isKnownNeverNaN(Op.getOperand(1), SNaN, Depth + 1);

[Petar.Avramovic]

llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-fminnum.mir

823

return isKnownNeverNaN(Op.getOperand(0), SNaN, Depth + 1) ||
       isKnownNeverNaN(Op.getOperand(1), SNaN, Depth + 1);
}

This is legalized first. %2 is isKnownNeverSNaN with IEEE = true, but since inputs for %2:_(s32) = G_FMAXNUM %0, %1 are not yet canonicalized isKnownNeverSNaN will fail and canonicalize is inserted. However since we know what will happen with '%2:_(s32) = G_FMAXNUM %0, %1 we declare it isKnownNeverSNaN to not depend on order of legalization.

[arsenm]

llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-fminnum.mir

823

The legalization order doesn't matter. These operations have their own independent semantics. isKnownNeverNaN needs to understand both pairs

[Petar.Avramovic]

llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-fminnum.mir

818–820

Using only that formula this will not change and we are left with unnecessary G_FCANONICALIZE(G_FMAXNUM_IEEE) which stops us from performing further combines both from td file and combiner. Do you have other suggestion how to avoid making G_FCANONICALIZE?

823

How can this be done?

The legalization order doesn't matter.

Formula above gives different result depending if input was legalized already or not. It would work if we would legalize uses first.

[arsenm]

llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-fminnum.mir

823

The same way SelectionDAG handles this:

case ISD::FMINNUM_IEEE:
case ISD::FMAXNUM_IEEE: {
  if (SNaN)
    return true;
  // This can return a NaN if either operand is an sNaN, or if both operands
  // are NaN.
  return (isKnownNeverNaN(Op.getOperand(0), false, Depth + 1) &&
          isKnownNeverSNaN(Op.getOperand(1), Depth + 1)) ||
         (isKnownNeverNaN(Op.getOperand(1), false, Depth + 1) &&
          isKnownNeverSNaN(Op.getOperand(0), Depth + 1));
}

[Petar.Avramovic]

llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-fminnum.mir

823

I meant for FMINNUM, writing same code as in SDAG doesn't deal with unnecessary canonicalize. FMINNUM_IEEE is the same (there is only if (SNaN) part atm).

[arsenm]

llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-fminnum.mir

823

The lowering for fminnum to fminnum_ieee tries to avoid the canonicalize depending on the instruction/inputs. This is how SelectionDAG also does it. For this patch, it just needs to fully handle the correct semanics for the two pairs

[Petar.Avramovic]

llvm/lib/CodeGen/GlobalISel/Utils.cpp

514–520

Generic opcodes should not have target dependent semantics like this, especially since these already have semantics inherited from the IR/C fmin/fmax definition.

I would say that only "is Known Never any NaN" is given. This patch does not change that (isKnownNeverNaN(SNaN = false)).

They return the non-nan argument regardless of snan or not

still stands. We are just aware of the way we lower instruction and use it in isKnownNeverSNaN only. This is only used for efficiency to avoid having to quiet "known non SNaN" into instruction that is about to be lowered.

518–519

I don't think

return isKnownNeverNaN(DefMI->getOperand(1).getReg(), MRI, SNaN) ||
       isKnownNeverNaN(DefMI->getOperand(2).getReg(), MRI, SNaN);

is enough for globalisel since it does not look at the whole picture. The formula above works when target naturally has FMINNUM instruction and knowing the inputs gives conclusion about the output (one input is not nan result is not nan). This would be generic formula (FIXME, use this formula in isKnownNeverNaNForMI instead of return false).
But subtarget with IEEE=true does not have FMINNUM and has to lower it to FMINNUM_IEEE. This gives additional critical piece of information FMINNUM is never SNaN with IEEE=true.
Are test changes correct?

llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-fminnum.mir

823

About examining the input: We are likely examining inputs in the middle of the legalization. With IEEE=true fminnum will never be the input but fminnum_ieee. Recursively asking for fminnum's inputs is not correct in this context since (maybe quieted)inputs will go to fminnum_ieee (its result is never SNaN) not fminnum and this is target dependent.
I don't see how this breaks correct semantics.

[arsenm]

llvm/lib/CodeGen/GlobalISel/Utils.cpp

518–519

The single instruction is the whole picture and does not depend on the target. These have defined, universal NAN handling semantics regardless of however they are lowered. This patch should only aim to faithfully implement these semantics. Ensuring we don't have redundant quieting canonicalizes is an optimization beyond this patch.

As an optimization, the expansion to the IEEE version tries to avoid inserting a quiet, which mostly works well enough. We are missing some optimization hints. I have long wished that we had separate no-qnan and no-snan fast math flags (as well as renaming the IR intrinsics to match the real fmin/fmax semantics, and introduce a new pair for the IEEE-758 2008 semantics)

[arsenm]

reverse ping

[Petar.Avramovic]

Removing target dependent semantics.

[Petar.Avramovic]

llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-fmaxnum.mir

721

Thoughts on adding

%0:_(s32) = G_FMINNUM_IEEE/G_FMAXNUM_IEEE ...
%1:_(s32) = G_FCANONICALIZE %0
->
%0:_(s32) = G_FMINNUM_IEEE/G_FMAXNUM_IEEE ...
%1:_(s32) = COPY %0

combine post legalizer?

[arsenm]

llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-fmaxnum.mir

721

Yes, redundant canonicalization elimination belongs post legalizer

[Petar.Avramovic]

Ping.

llvm/lib/CodeGen/GlobalISel/Utils.cpp

bool llvm::isKnownNeverNaN(Register Val, const MachineRegisterInfo &MRI,
 
   if (DefMI->getOpcode() == TargetOpcode::G_BUILD_VECTOR) {
     for (auto Op : DefMI->uses())
       if (!isKnownNeverNaN(Op.getReg(), MRI, SNaN))
         return false;
     return true;
   }
 
+  switch (DefMI->getOpcode()) {
+  default:
+    break;
+  case TargetOpcode::G_FMINNUM_IEEE:
+  case TargetOpcode::G_FMAXNUM_IEEE: {
+    if (SNaN)
+      return true;
+    // This can return a NaN if either operand is an sNaN, or if both operands
+    // are NaN.
+    return (isKnownNeverNaN(DefMI->getOperand(1).getReg(), MRI) &&
+            isKnownNeverSNaN(DefMI->getOperand(2).getReg(), MRI)) ||
+           (isKnownNeverSNaN(DefMI->getOperand(1).getReg(), MRI) &&
+            isKnownNeverNaN(DefMI->getOperand(2).getReg(), MRI));
+  }
+  case TargetOpcode::G_FMINNUM:
+  case TargetOpcode::G_FMAXNUM: {
+    // Only one needs to be known not-nan, since it will be returned if the
+    // other ends up being one.
+    return isKnownNeverNaN(DefMI->getOperand(1).getReg(), MRI, SNaN) ||
+           isKnownNeverNaN(DefMI->getOperand(2).getReg(), MRI, SNaN);
+  }
+  }
+
   if (SNaN) {
     // FP operations quiet. For now, just handle the ones inserted during
     // legalization.
     switch (DefMI->getOpcode()) {
     case TargetOpcode::G_FPEXT:
     case TargetOpcode::G_FPTRUNC:
     case TargetOpcode::G_FCANONICALIZE:
       return true;
     default:
       return false;
     }
   }
 
   return false;
 }
 
 Align llvm::inferAlignFromPtrInfo(MachineFunction &MF,
                                   const MachinePointerInfo &MPO) {
   auto PSV = MPO.V.dyn_cast<const PseudoSourceValue *>();
   if (auto FSPV = dyn_cast_or_null<FixedStackPseudoSourceValue>(PSV)) {
     MachineFrameInfo &MFI = MF.getFrameInfo();
     return commonAlignment(MFI.getObjectAlign(FSPV->getFrameIndex()),
                            MPO.Offset);

llvm/test/CodeGen/AMDGPU/GlobalISel/fmed3.ll

 ; SI-NEXT:    s_mov_b64 s[8:9], s[6:7]
 ; SI-NEXT:    buffer_load_dword v4, v[0:1], s[8:11], 0 addr64
 ; SI-NEXT:    s_waitcnt vmcnt(2)
 ; SI-NEXT:    v_add_f32_e32 v2, 1.0, v2
 ; SI-NEXT:    s_waitcnt vmcnt(1)
 ; SI-NEXT:    v_add_f32_e32 v3, 2.0, v3
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_add_f32_e32 v4, 4.0, v4
-; SI-NEXT:    v_min_f32_e32 v5, v2, v3
-; SI-NEXT:    v_max_f32_e32 v2, v2, v3
-; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
-; SI-NEXT:    v_min_f32_e32 v2, v2, v4
-; SI-NEXT:    v_mul_f32_e32 v3, 1.0, v5
-; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
-; SI-NEXT:    v_max_f32_e32 v2, v3, v2
+; SI-NEXT:    v_med3_f32 v2, v2, v3, v4
 ; SI-NEXT:    s_mov_b64 s[2:3], s[10:11]
 ; SI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: v_nnan_inputs_med3_f32_pat0:
 ; VI:       ; %bb.0:
 ; VI-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
 ; VI-NEXT:    v_lshlrev_b32_e32 v6, 2, v0

 ; VI-NEXT:    v_mov_b32_e32 v0, s0
 ; VI-NEXT:    v_mov_b32_e32 v1, s1
 ; VI-NEXT:    v_add_u32_e32 v0, vcc, v0, v6
 ; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
 ; VI-NEXT:    s_waitcnt vmcnt(2)
 ; VI-NEXT:    v_add_f32_e32 v4, 1.0, v7
 ; VI-NEXT:    s_waitcnt vmcnt(1)
 ; VI-NEXT:    v_add_f32_e32 v2, 2.0, v2
-; VI-NEXT:    v_min_f32_e32 v5, v4, v2
-; VI-NEXT:    v_max_f32_e32 v2, v4, v2
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_add_f32_e32 v3, 4.0, v3
-; VI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
-; VI-NEXT:    v_min_f32_e32 v2, v2, v3
-; VI-NEXT:    v_mul_f32_e32 v3, 1.0, v5
-; VI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
-; VI-NEXT:    v_max_f32_e32 v2, v3, v2
+; VI-NEXT:    v_med3_f32 v2, v4, v2, v3
 ; VI-NEXT:    flat_store_dword v[0:1], v2
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: v_nnan_inputs_med3_f32_pat0:
 ; GFX9:       ; %bb.0:
 ; GFX9-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
 ; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_load_dword v1, v0, s[2:3]
 ; GFX9-NEXT:    global_load_dword v2, v0, s[4:5]
 ; GFX9-NEXT:    global_load_dword v3, v0, s[6:7]
 ; GFX9-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-NEXT:    v_add_f32_e32 v1, 1.0, v1
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-NEXT:    v_add_f32_e32 v2, 2.0, v2
-; GFX9-NEXT:    v_min_f32_e32 v4, v1, v2
-; GFX9-NEXT:    v_max_f32_e32 v1, v1, v2
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    v_add_f32_e32 v3, 4.0, v3
-; GFX9-NEXT:    v_max_f32_e32 v1, v1, v1
-; GFX9-NEXT:    v_min_f32_e32 v1, v1, v3
-; GFX9-NEXT:    v_max_f32_e32 v2, v4, v4
-; GFX9-NEXT:    v_max_f32_e32 v1, v1, v1
-; GFX9-NEXT:    v_max_f32_e32 v1, v2, v1
+; GFX9-NEXT:    v_med3_f32 v1, v1, v2, v3
 ; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
 ; GFX9-NEXT:    s_endpgm
   %tid = call i32 @llvm.amdgcn.workitem.id.x()
   %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
   %gep1 = getelementptr float, float addrspace(1)* %bptr, i32 %tid
   %gep2 = getelementptr float, float addrspace(1)* %cptr, i32 %tid
   %outgep = getelementptr float, float addrspace(1)* %out, i32 %tid
   %a = load volatile float, float addrspace(1)* %gep0

llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-fmaxnum.mir

bb.0:
     ; VI: $vgpr0 = COPY [[FMAXNUM_IEEE1]](s32)
     ; GFX9-LABEL: name: test_fmaxnum_with_fmaxnum_argument_s32_ieee_mode_on
     ; GFX9: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
     ; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
     ; GFX9: [[FCANONICALIZE:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[COPY]]
     ; GFX9: [[FCANONICALIZE1:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[COPY1]]
     ; GFX9: [[FMAXNUM_IEEE:%[0-9]+]]:_(s32) = G_FMAXNUM_IEEE [[FCANONICALIZE]], [[FCANONICALIZE1]]
     ; GFX9: [[COPY2:%[0-9]+]]:_(s32) = COPY $vgpr2
     ; GFX9: [[FCANONICALIZE2:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[FMAXNUM_IEEE]]
     ; GFX9: [[FCANONICALIZE3:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[COPY2]]
     ; GFX9: [[FMAXNUM_IEEE1:%[0-9]+]]:_(s32) = G_FMAXNUM_IEEE [[FCANONICALIZE2]], [[FCANONICALIZE3]]
     ; GFX9: $vgpr0 = COPY [[FMAXNUM_IEEE1]](s32)
     %0:_(s32) = COPY $vgpr0
     %1:_(s32) = COPY $vgpr1
     %2:_(s32) = G_FMAXNUM %0, %1
     %3:_(s32) = COPY $vgpr2
     %4:_(s32) = G_FMAXNUM %2, %3
bb.0:
     liveins: $vgpr0, $vgpr1
 
     ; SI-LABEL: name: test_fmaxnum_with_nonNaN_fmaxnum_argument_s32_ieee_mode_on
     ; SI: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
     ; SI: [[C:%[0-9]+]]:_(s32) = G_FCONSTANT float 0.000000e+00
     ; SI: [[FCANONICALIZE:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[COPY]]
     ; SI: [[FMAXNUM_IEEE:%[0-9]+]]:_(s32) = G_FMAXNUM_IEEE [[FCANONICALIZE]], [[C]]
     ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
-    ; SI: [[FCANONICALIZE1:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[FMAXNUM_IEEE]]
-    ; SI: [[FCANONICALIZE2:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[COPY1]]
-    ; SI: [[FMAXNUM_IEEE1:%[0-9]+]]:_(s32) = G_FMAXNUM_IEEE [[FCANONICALIZE1]], [[FCANONICALIZE2]]
+    ; SI: [[FCANONICALIZE1:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[COPY1]]
+    ; SI: [[FMAXNUM_IEEE1:%[0-9]+]]:_(s32) = G_FMAXNUM_IEEE [[FMAXNUM_IEEE]], [[FCANONICALIZE1]]
     ; SI: $vgpr0 = COPY [[FMAXNUM_IEEE1]](s32)
     ; VI-LABEL: name: test_fmaxnum_with_nonNaN_fmaxnum_argument_s32_ieee_mode_on
     ; VI: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
     ; VI: [[C:%[0-9]+]]:_(s32) = G_FCONSTANT float 0.000000e+00
     ; VI: [[FCANONICALIZE:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[COPY]]
     ; VI: [[FMAXNUM_IEEE:%[0-9]+]]:_(s32) = G_FMAXNUM_IEEE [[FCANONICALIZE]], [[C]]
     ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
-    ; VI: [[FCANONICALIZE1:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[FMAXNUM_IEEE]]
-    ; VI: [[FCANONICALIZE2:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[COPY1]]
-    ; VI: [[FMAXNUM_IEEE1:%[0-9]+]]:_(s32) = G_FMAXNUM_IEEE [[FCANONICALIZE1]], [[FCANONICALIZE2]]
+    ; VI: [[FCANONICALIZE1:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[COPY1]]
+    ; VI: [[FMAXNUM_IEEE1:%[0-9]+]]:_(s32) = G_FMAXNUM_IEEE [[FMAXNUM_IEEE]], [[FCANONICALIZE1]]
     ; VI: $vgpr0 = COPY [[FMAXNUM_IEEE1]](s32)
     ; GFX9-LABEL: name: test_fmaxnum_with_nonNaN_fmaxnum_argument_s32_ieee_mode_on
     ; GFX9: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
     ; GFX9: [[C:%[0-9]+]]:_(s32) = G_FCONSTANT float 0.000000e+00
     ; GFX9: [[FCANONICALIZE:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[COPY]]
     ; GFX9: [[FMAXNUM_IEEE:%[0-9]+]]:_(s32) = G_FMAXNUM_IEEE [[FCANONICALIZE]], [[C]]
     ; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
-    ; GFX9: [[FCANONICALIZE1:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[FMAXNUM_IEEE]]
-    ; GFX9: [[FCANONICALIZE2:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[COPY1]]
-    ; GFX9: [[FMAXNUM_IEEE1:%[0-9]+]]:_(s32) = G_FMAXNUM_IEEE [[FCANONICALIZE1]], [[FCANONICALIZE2]]
+    ; GFX9: [[FCANONICALIZE1:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[COPY1]]
+    ; GFX9: [[FMAXNUM_IEEE1:%[0-9]+]]:_(s32) = G_FMAXNUM_IEEE [[FMAXNUM_IEEE]], [[FCANONICALIZE1]]
     ; GFX9: $vgpr0 = COPY [[FMAXNUM_IEEE1]](s32)
     %0:_(s32) = COPY $vgpr0
     %1:_(s32) = G_FCONSTANT float 0.000000e+00
     %2:_(s32) = G_FMAXNUM %0, %1
     %3:_(s32) = COPY $vgpr1
     %4:_(s32) = G_FMAXNUM %2, %3
     $vgpr0 = COPY %4
 ...
bb.0:
     liveins: $vgpr0, $vgpr1
 
     ; SI-LABEL: name: test_fmaxnum_with_nonNaN_fminnum_argument_s32_ieee_mode_on
     ; SI: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
     ; SI: [[C:%[0-9]+]]:_(s32) = G_FCONSTANT float 0.000000e+00
     ; SI: [[FCANONICALIZE:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[COPY]]
     ; SI: [[FMINNUM_IEEE:%[0-9]+]]:_(s32) = G_FMINNUM_IEEE [[FCANONICALIZE]], [[C]]
     ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
-    ; SI: [[FCANONICALIZE1:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[FMINNUM_IEEE]]
-    ; SI: [[FCANONICALIZE2:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[COPY1]]
-    ; SI: [[FMAXNUM_IEEE:%[0-9]+]]:_(s32) = G_FMAXNUM_IEEE [[FCANONICALIZE1]], [[FCANONICALIZE2]]
+    ; SI: [[FCANONICALIZE1:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[COPY1]]
+    ; SI: [[FMAXNUM_IEEE:%[0-9]+]]:_(s32) = G_FMAXNUM_IEEE [[FMINNUM_IEEE]], [[FCANONICALIZE1]]
     ; SI: $vgpr0 = COPY [[FMAXNUM_IEEE]](s32)
     ; VI-LABEL: name: test_fmaxnum_with_nonNaN_fminnum_argument_s32_ieee_mode_on
     ; VI: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
     ; VI: [[C:%[0-9]+]]:_(s32) = G_FCONSTANT float 0.000000e+00
     ; VI: [[FCANONICALIZE:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[COPY]]
     ; VI: [[FMINNUM_IEEE:%[0-9]+]]:_(s32) = G_FMINNUM_IEEE [[FCANONICALIZE]], [[C]]
     ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
-    ; VI: [[FCANONICALIZE1:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[FMINNUM_IEEE]]
-    ; VI: [[FCANONICALIZE2:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[COPY1]]
-    ; VI: [[FMAXNUM_IEEE:%[0-9]+]]:_(s32) = G_FMAXNUM_IEEE [[FCANONICALIZE1]], [[FCANONICALIZE2]]
+    ; VI: [[FCANONICALIZE1:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[COPY1]]
+    ; VI: [[FMAXNUM_IEEE:%[0-9]+]]:_(s32) = G_FMAXNUM_IEEE [[FMINNUM_IEEE]], [[FCANONICALIZE1]]
     ; VI: $vgpr0 = COPY [[FMAXNUM_IEEE]](s32)
     ; GFX9-LABEL: name: test_fmaxnum_with_nonNaN_fminnum_argument_s32_ieee_mode_on
     ; GFX9: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
     ; GFX9: [[C:%[0-9]+]]:_(s32) = G_FCONSTANT float 0.000000e+00
     ; GFX9: [[FCANONICALIZE:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[COPY]]
     ; GFX9: [[FMINNUM_IEEE:%[0-9]+]]:_(s32) = G_FMINNUM_IEEE [[FCANONICALIZE]], [[C]]
     ; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
-    ; GFX9: [[FCANONICALIZE1:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[FMINNUM_IEEE]]
-    ; GFX9: [[FCANONICALIZE2:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[COPY1]]
-    ; GFX9: [[FMAXNUM_IEEE:%[0-9]+]]:_(s32) = G_FMAXNUM_IEEE [[FCANONICALIZE1]], [[FCANONICALIZE2]]
+    ; GFX9: [[FCANONICALIZE1:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[COPY1]]
+    ; GFX9: [[FMAXNUM_IEEE:%[0-9]+]]:_(s32) = G_FMAXNUM_IEEE [[FMINNUM_IEEE]], [[FCANONICALIZE1]]
     ; GFX9: $vgpr0 = COPY [[FMAXNUM_IEEE]](s32)
     %0:_(s32) = COPY $vgpr0
     %1:_(s32) = G_FCONSTANT float 0.000000e+00
     %2:_(s32) = G_FMINNUM %0, %1
     %3:_(s32) = COPY $vgpr1
     %4:_(s32) = G_FMAXNUM %2, %3
     $vgpr0 = COPY %4
 ...

llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-fminnum.mir

bb.0:
     liveins: $vgpr0, $vgpr1
 
     ; SI-LABEL: name: test_fminnum_with_nonNaN_fminnum_argument_s32_ieee_mode_on
     ; SI: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
     ; SI: [[C:%[0-9]+]]:_(s32) = G_FCONSTANT float 0.000000e+00
     ; SI: [[FCANONICALIZE:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[COPY]]
     ; SI: [[FMINNUM_IEEE:%[0-9]+]]:_(s32) = G_FMINNUM_IEEE [[FCANONICALIZE]], [[C]]
     ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
-    ; SI: [[FCANONICALIZE1:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[FMINNUM_IEEE]]
-    ; SI: [[FCANONICALIZE2:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[COPY1]]
-    ; SI: [[FMINNUM_IEEE1:%[0-9]+]]:_(s32) = G_FMINNUM_IEEE [[FCANONICALIZE1]], [[FCANONICALIZE2]]
+    ; SI: [[FCANONICALIZE1:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[COPY1]]
+    ; SI: [[FMINNUM_IEEE1:%[0-9]+]]:_(s32) = G_FMINNUM_IEEE [[FMINNUM_IEEE]], [[FCANONICALIZE1]]
     ; SI: $vgpr0 = COPY [[FMINNUM_IEEE1]](s32)
     ; VI-LABEL: name: test_fminnum_with_nonNaN_fminnum_argument_s32_ieee_mode_on
     ; VI: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
     ; VI: [[C:%[0-9]+]]:_(s32) = G_FCONSTANT float 0.000000e+00
     ; VI: [[FCANONICALIZE:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[COPY]]
     ; VI: [[FMINNUM_IEEE:%[0-9]+]]:_(s32) = G_FMINNUM_IEEE [[FCANONICALIZE]], [[C]]
     ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
-    ; VI: [[FCANONICALIZE1:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[FMINNUM_IEEE]]
-    ; VI: [[FCANONICALIZE2:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[COPY1]]
-    ; VI: [[FMINNUM_IEEE1:%[0-9]+]]:_(s32) = G_FMINNUM_IEEE [[FCANONICALIZE1]], [[FCANONICALIZE2]]
+    ; VI: [[FCANONICALIZE1:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[COPY1]]
+    ; VI: [[FMINNUM_IEEE1:%[0-9]+]]:_(s32) = G_FMINNUM_IEEE [[FMINNUM_IEEE]], [[FCANONICALIZE1]]
     ; VI: $vgpr0 = COPY [[FMINNUM_IEEE1]](s32)
     ; GFX9-LABEL: name: test_fminnum_with_nonNaN_fminnum_argument_s32_ieee_mode_on
     ; GFX9: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
     ; GFX9: [[C:%[0-9]+]]:_(s32) = G_FCONSTANT float 0.000000e+00
     ; GFX9: [[FCANONICALIZE:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[COPY]]
     ; GFX9: [[FMINNUM_IEEE:%[0-9]+]]:_(s32) = G_FMINNUM_IEEE [[FCANONICALIZE]], [[C]]
     ; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
-    ; GFX9: [[FCANONICALIZE1:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[FMINNUM_IEEE]]
-    ; GFX9: [[FCANONICALIZE2:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[COPY1]]
-    ; GFX9: [[FMINNUM_IEEE1:%[0-9]+]]:_(s32) = G_FMINNUM_IEEE [[FCANONICALIZE1]], [[FCANONICALIZE2]]
+    ; GFX9: [[FCANONICALIZE1:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[COPY1]]
+    ; GFX9: [[FMINNUM_IEEE1:%[0-9]+]]:_(s32) = G_FMINNUM_IEEE [[FMINNUM_IEEE]], [[FCANONICALIZE1]]
     ; GFX9: $vgpr0 = COPY [[FMINNUM_IEEE1]](s32)
     %0:_(s32) = COPY $vgpr0
     %1:_(s32) = G_FCONSTANT float 0.000000e+00
     %2:_(s32) = G_FMINNUM %0, %1
     %3:_(s32) = COPY $vgpr1
     %4:_(s32) = G_FMINNUM %2, %3
     $vgpr0 = COPY %4
 ...
bb.0:
     ; VI: $vgpr0 = COPY [[FMINNUM_IEEE]](s32)
     ; GFX9-LABEL: name: test_fminnum_with_fmaxnum_argument_s32_ieee_mode_on
     ; GFX9: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
     ; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
     ; GFX9: [[FCANONICALIZE:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[COPY]]
     ; GFX9: [[FCANONICALIZE1:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[COPY1]]
     ; GFX9: [[FMAXNUM_IEEE:%[0-9]+]]:_(s32) = G_FMAXNUM_IEEE [[FCANONICALIZE]], [[FCANONICALIZE1]]
     ; GFX9: [[COPY2:%[0-9]+]]:_(s32) = COPY $vgpr2
     ; GFX9: [[FCANONICALIZE2:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[FMAXNUM_IEEE]]
     ; GFX9: [[FCANONICALIZE3:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[COPY2]]
     ; GFX9: [[FMINNUM_IEEE:%[0-9]+]]:_(s32) = G_FMINNUM_IEEE [[FCANONICALIZE2]], [[FCANONICALIZE3]]
     ; GFX9: $vgpr0 = COPY [[FMINNUM_IEEE]](s32)
     %0:_(s32) = COPY $vgpr0
     %1:_(s32) = COPY $vgpr1
     %2:_(s32) = G_FMAXNUM %0, %1
     %3:_(s32) = COPY $vgpr2
     %4:_(s32) = G_FMINNUM %2, %3
     $vgpr0 = COPY %4
 ...
 
 ---
 name: test_fminnum_with_nonNaN_fmaxnum_argument_s32_ieee_mode_on
 machineFunctionInfo:
   mode:
     ieee: true
 body: |
   bb.0:
     liveins: $vgpr0, $vgpr1
 
     ; SI-LABEL: name: test_fminnum_with_nonNaN_fmaxnum_argument_s32_ieee_mode_on
     ; SI: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
     ; SI: [[C:%[0-9]+]]:_(s32) = G_FCONSTANT float 0.000000e+00
     ; SI: [[FCANONICALIZE:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[COPY]]
     ; SI: [[FMAXNUM_IEEE:%[0-9]+]]:_(s32) = G_FMAXNUM_IEEE [[FCANONICALIZE]], [[C]]
     ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
-    ; SI: [[FCANONICALIZE1:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[FMAXNUM_IEEE]]
-    ; SI: [[FCANONICALIZE2:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[COPY1]]
-    ; SI: [[FMINNUM_IEEE:%[0-9]+]]:_(s32) = G_FMINNUM_IEEE [[FCANONICALIZE1]], [[FCANONICALIZE2]]
+    ; SI: [[FCANONICALIZE1:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[COPY1]]
+    ; SI: [[FMINNUM_IEEE:%[0-9]+]]:_(s32) = G_FMINNUM_IEEE [[FMAXNUM_IEEE]], [[FCANONICALIZE1]]
     ; SI: $vgpr0 = COPY [[FMINNUM_IEEE]](s32)
     ; VI-LABEL: name: test_fminnum_with_nonNaN_fmaxnum_argument_s32_ieee_mode_on
     ; VI: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
     ; VI: [[C:%[0-9]+]]:_(s32) = G_FCONSTANT float 0.000000e+00
     ; VI: [[FCANONICALIZE:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[COPY]]
     ; VI: [[FMAXNUM_IEEE:%[0-9]+]]:_(s32) = G_FMAXNUM_IEEE [[FCANONICALIZE]], [[C]]
     ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
-    ; VI: [[FCANONICALIZE1:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[FMAXNUM_IEEE]]
-    ; VI: [[FCANONICALIZE2:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[COPY1]]
-    ; VI: [[FMINNUM_IEEE:%[0-9]+]]:_(s32) = G_FMINNUM_IEEE [[FCANONICALIZE1]], [[FCANONICALIZE2]]
+    ; VI: [[FCANONICALIZE1:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[COPY1]]
+    ; VI: [[FMINNUM_IEEE:%[0-9]+]]:_(s32) = G_FMINNUM_IEEE [[FMAXNUM_IEEE]], [[FCANONICALIZE1]]
     ; VI: $vgpr0 = COPY [[FMINNUM_IEEE]](s32)
     ; GFX9-LABEL: name: test_fminnum_with_nonNaN_fmaxnum_argument_s32_ieee_mode_on
     ; GFX9: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
     ; GFX9: [[C:%[0-9]+]]:_(s32) = G_FCONSTANT float 0.000000e+00
     ; GFX9: [[FCANONICALIZE:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[COPY]]
     ; GFX9: [[FMAXNUM_IEEE:%[0-9]+]]:_(s32) = G_FMAXNUM_IEEE [[FCANONICALIZE]], [[C]]
     ; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
-    ; GFX9: [[FCANONICALIZE1:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[FMAXNUM_IEEE]]
-    ; GFX9: [[FCANONICALIZE2:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[COPY1]]
-    ; GFX9: [[FMINNUM_IEEE:%[0-9]+]]:_(s32) = G_FMINNUM_IEEE [[FCANONICALIZE1]], [[FCANONICALIZE2]]
+    ; GFX9: [[FCANONICALIZE1:%[0-9]+]]:_(s32) = G_FCANONICALIZE [[COPY1]]
+    ; GFX9: [[FMINNUM_IEEE:%[0-9]+]]:_(s32) = G_FMINNUM_IEEE [[FMAXNUM_IEEE]], [[FCANONICALIZE1]]
     ; GFX9: $vgpr0 = COPY [[FMINNUM_IEEE]](s32)
     %0:_(s32) = COPY $vgpr0
     %1:_(s32) = G_FCONSTANT float 0.000000e+00
     %2:_(s32) = G_FMAXNUM %0, %1
     %3:_(s32) = COPY $vgpr1
     %4:_(s32) = G_FMINNUM %2, %3
     $vgpr0 = COPY %4
 ...