AMDGPU/GlobalISel: Add floating point med3 combine

[Petar.Avramovic]

Summary

Add floating point version of med3 combine.
Source is fminnum(fmaxnum(Val, K0), K1) or fmaxnum(fminnum(Val, K1), K0)
where K0 and K1 are constants and K0 <= K1.

[arsenm]

Missing tests for f16 and v2f16 cases

llvm/lib/Target/AMDGPU/AMDGPUPostLegalizerCombiner.cpp

174–175 ↗

(On Diff #300311)

Might as well handle this now

181–182 ↗

(On Diff #300311)

You can directly pass the APFloat to is the overload of isInlineConstant

[Petar.Avramovic]

llvm/lib/Target/AMDGPU/AMDGPUPostLegalizerCombiner.cpp

174–175 ↗

(On Diff #300311)

I am not really sure I know correct way to do it. Also this combine heavily depends on legalizer and primarily completing remaining of SNaN checks in isKnownNeverSNaN.

[foad]

llvm/lib/Target/AMDGPU/AMDGPUPostLegalizerCombiner.cpp

172 ↗

(On Diff #300311)

This will probably warn that Info is unused in a Release build.

[Petar.Avramovic]

Rebase/rework.

[Petar.Avramovic]

Ping.

[arsenm]

llvm/lib/Target/AMDGPU/AMDGPURegBankCombiner.cpp

178

Like in the other patches, why do you need to re-query getConstantFPVRegVal? Why doesn't the constant matcher just directly return the constant?

[Petar.Avramovic]

Use matcher that returns both VReg and APFloat from G_FCONSTANT.

[Petar.Avramovic]

Ping.

[Petar.Avramovic]

Ping.

[Petar.Avramovic]

Rebase and ping.

[arsenm]

llvm/lib/Target/AMDGPU/AMDGPURegBankCombiner.cpp

174

I don't see why to make these optional if they won't be set or used if the matcher return false

191

This should be cached in the helper?

[Petar.Avramovic]

Cache SIInstrInfo in the helper.

[Petar.Avramovic]

llvm/lib/Target/AMDGPU/AMDGPURegBankCombiner.cpp

174

Util helpers and matchers use Optional.

llvm/lib/Target/AMDGPU/AMDGPUCombine.td

def int_minmax_to_med3 : GICombineRule<
   (defs root:$min_or_max, med3_matchdata:$matchinfo),
   (match (wip_match_opcode G_SMAX,
                            G_SMIN,
                            G_UMAX,
                            G_UMIN):$min_or_max,
          [{ return RegBankHelper.matchIntMinMaxToMed3(*${min_or_max}, ${matchinfo}); }]),
   (apply [{ RegBankHelper.applyMed3(*${min_or_max}, ${matchinfo}); }])>;
 
+def fp_minmax_to_med3 : GICombineRule<
+  (defs root:$min_or_max, med3_matchdata:$matchinfo),
+  (match (wip_match_opcode G_FMAXNUM,
+                           G_FMINNUM,
+                           G_FMAXNUM_IEEE,
+                           G_FMINNUM_IEEE):$min_or_max,
+         [{ return RegBankHelper.matchFPMinMaxToMed3(*${min_or_max}, ${matchinfo}); }]),
+  (apply [{ RegBankHelper.applyMed3(*${min_or_max}, ${matchinfo}); }])>;
+
 def remove_fcanonicalize_matchinfo : GIDefMatchData<"Register">;
 
 def remove_fcanonicalize : GICombineRule<
   (defs root:$fcanonicalize, remove_fcanonicalize_matchinfo:$matchinfo),
   (match (wip_match_opcode G_FCANONICALIZE):$fcanonicalize,
          [{ return PostLegalizerHelper.matchRemoveFcanonicalize(*${fcanonicalize}, ${matchinfo}); }]),
   (apply [{ Helper.replaceSingleDefInstWithReg(*${fcanonicalize}, ${matchinfo}); }])>;
 
def AMDGPUPostLegalizerCombinerHelper: GICombinerHelper<
   [all_combines, gfx6gfx7_combines,
    uchar_to_float, cvt_f32_ubyteN, remove_fcanonicalize]> {
   let DisableRuleOption = "amdgpupostlegalizercombiner-disable-rule";
   let StateClass = "AMDGPUPostLegalizerCombinerHelperState";
   let AdditionalArguments = [];
 }
 
 def AMDGPURegBankCombinerHelper : GICombinerHelper<
-  "AMDGPUGenRegBankCombinerHelper", [zext_trunc_fold, int_minmax_to_med3]> {
+  "AMDGPUGenRegBankCombinerHelper",
+  [zext_trunc_fold, int_minmax_to_med3, fp_minmax_to_med3]> {
   let DisableRuleOption = "amdgpuregbankcombiner-disable-rule";
   let StateClass = "AMDGPURegBankCombinerHelperState";
   let AdditionalArguments = [];
 }

llvm/lib/Target/AMDGPU/AMDGPUGISel.td

 def : GINodeEquiv<G_AMDGPU_CVT_F32_UBYTE0, AMDGPUcvt_f32_ubyte0>;
 def : GINodeEquiv<G_AMDGPU_CVT_F32_UBYTE1, AMDGPUcvt_f32_ubyte1>;
 def : GINodeEquiv<G_AMDGPU_CVT_F32_UBYTE2, AMDGPUcvt_f32_ubyte2>;
 def : GINodeEquiv<G_AMDGPU_CVT_F32_UBYTE3, AMDGPUcvt_f32_ubyte3>;
 
 def : GINodeEquiv<G_AMDGPU_CVT_PK_I16_I32, AMDGPUpk_i16_i32_impl>;
 def : GINodeEquiv<G_AMDGPU_SMED3, AMDGPUsmed3>;
 def : GINodeEquiv<G_AMDGPU_UMED3, AMDGPUumed3>;
+def : GINodeEquiv<G_AMDGPU_FMED3, AMDGPUfmed3_impl>;
 
 def : GINodeEquiv<G_AMDGPU_ATOMIC_CMPXCHG, AMDGPUatomic_cmp_swap>;
 def : GINodeEquiv<G_AMDGPU_BUFFER_LOAD, SIbuffer_load>;
 def : GINodeEquiv<G_AMDGPU_BUFFER_LOAD_USHORT, SIbuffer_load_ushort>;
 def : GINodeEquiv<G_AMDGPU_BUFFER_LOAD_UBYTE, SIbuffer_load_ubyte>;
 def : GINodeEquiv<G_AMDGPU_BUFFER_LOAD_SSHORT, SIbuffer_load_short>;
 def : GINodeEquiv<G_AMDGPU_BUFFER_LOAD_SBYTE, SIbuffer_load_byte>;
 def : GINodeEquiv<G_AMDGPU_BUFFER_LOAD_FORMAT, SIbuffer_load_format>;

llvm/lib/Target/AMDGPU/AMDGPURegBankCombiner.cpp

 //
 //===----------------------------------------------------------------------===//
 
 #include "AMDGPU.h"
 #include "AMDGPULegalizerInfo.h"
 #include "AMDGPURegisterBankInfo.h"
 #include "GCNSubtarget.h"
 #include "MCTargetDesc/AMDGPUMCTargetDesc.h"
+#include "SIMachineFunctionInfo.h"
 #include "llvm/CodeGen/GlobalISel/Combiner.h"
 #include "llvm/CodeGen/GlobalISel/CombinerHelper.h"
 #include "llvm/CodeGen/GlobalISel/CombinerInfo.h"
 #include "llvm/CodeGen/GlobalISel/GISelKnownBits.h"
 #include "llvm/CodeGen/GlobalISel/MIPatternMatch.h"
 #include "llvm/CodeGen/MachineDominators.h"
 #include "llvm/CodeGen/TargetPassConfig.h"
 #include "llvm/Target/TargetMachine.h"
public:
 
   MinMaxMedOpc getMinMaxPair(unsigned Opc);
 
   template <class m_Cst, typename CstTy>
   bool matchMed(MachineInstr &MI, MachineRegisterInfo &MRI, MinMaxMedOpc MMMOpc,
                 Register &Val, CstTy &K0, CstTy &K1);
 
   bool matchIntMinMaxToMed3(MachineInstr &MI, Med3MatchInfo &MatchInfo);
+  bool matchFPMinMaxToMed3(MachineInstr &MI, Med3MatchInfo &MatchInfo);
   void applyMed3(MachineInstr &MI, Med3MatchInfo &MatchInfo);
+
+private:
+  AMDGPU::SIModeRegisterDefaults getMode();
+  bool getIEEE();
+  bool isFminnumIeee(const MachineInstr &MI);
 };
 
 bool AMDGPURegBankCombinerHelper::isVgprRegBank(Register Reg) {
   return RBI.getRegBank(Reg, MRI, TRI)->getID() == AMDGPU::VGPRRegBankID;
 }
 
 AMDGPURegBankCombinerHelper::MinMaxMedOpc
 AMDGPURegBankCombinerHelper::getMinMaxPair(unsigned Opc) {
   switch (Opc) {
   default:
     llvm_unreachable("Unsupported opcode");
   case AMDGPU::G_SMAX:
   case AMDGPU::G_SMIN:
     return {AMDGPU::G_SMIN, AMDGPU::G_SMAX, AMDGPU::G_AMDGPU_SMED3};
   case AMDGPU::G_UMAX:
   case AMDGPU::G_UMIN:
     return {AMDGPU::G_UMIN, AMDGPU::G_UMAX, AMDGPU::G_AMDGPU_UMED3};
+  case AMDGPU::G_FMAXNUM:
+  case AMDGPU::G_FMINNUM:
+    return {AMDGPU::G_FMINNUM, AMDGPU::G_FMAXNUM, AMDGPU::G_AMDGPU_FMED3};
+  case AMDGPU::G_FMAXNUM_IEEE:
+  case AMDGPU::G_FMINNUM_IEEE:
+    return {AMDGPU::G_FMINNUM_IEEE, AMDGPU::G_FMAXNUM_IEEE,
+            AMDGPU::G_AMDGPU_FMED3};
   }
 }
 
 template <class m_Cst, typename CstTy>
 bool AMDGPURegBankCombinerHelper::matchMed(MachineInstr &MI,
                                            MachineRegisterInfo &MRI,
                                            MinMaxMedOpc MMMOpc, Register &Val,
                                            CstTy &K0, CstTy &K1) {
if (OpcodeTriple.Med == AMDGPU::G_AMDGPU_SMED3 && K0->Value.sgt(K1->Value))
     return false;
   if (OpcodeTriple.Med == AMDGPU::G_AMDGPU_UMED3 && K0->Value.ugt(K1->Value))
     return false;
 
   MatchInfo = {OpcodeTriple.Med, Val, K0->VReg, K1->VReg};
   return true;
 }
 
+// fmed3(NaN, K0, K1) = min(min(NaN, K0), K1)
+// ieee = true  : min/max(SNaN, K) = QNaN, min/max(QNaN, K) = K
+// ieee = false : min/max(NaN, K) = K
+// clamp(NaN) = dx10_clamp ? 0.0 : NaN
+// Consider values of min(max(Val, K0), K1) and max(min(Val, K1), K0) as input.
+// Other operand commutes (see matchMed) give same result since min and max are
+// commutative.
+
+// Try to replace fp min(max(Val, K0), K1) or max(min(Val, K1), K0), KO<=K1
+// with fmed3(Val, K0, K1) or clamp(Val). Clamp requires K0 = 0.0 and K1 = 1.0.
+// Val = SNaN only for ieee = true
+// fmed3(SNaN, K0, K1) = min(min(SNaN, K0), K1) = min(QNaN, K1) = K1
+// min(max(SNaN, K0), K1) = min(QNaN, K1) = K1
+// max(min(SNaN, K1), K0) = max(K1, K0) = K1
+// Val = NaN,ieee = false or Val = QNaN,ieee = true
+// fmed3(NaN, K0, K1) = min(min(NaN, K0), K1) = min(K0, K1) = K0
+// min(max(NaN, K0), K1) = min(K0, K1) = K0 (can clamp when dx10_clamp = true)
+// max(min(NaN, K1), K0) = max(K1, K0) = K1 != K0
+bool AMDGPURegBankCombinerHelper::matchFPMinMaxToMed3(
+    MachineInstr &MI, Med3MatchInfo &MatchInfo) {
+  Register Dst = MI.getOperand(0).getReg();
+  LLT Ty = MRI.getType(Dst);
+  if (Ty != LLT::scalar(16) && Ty != LLT::scalar(32))
+    return false;
+
+  auto OpcodeTriple = getMinMaxPair(MI.getOpcode());
+
+  Register Val;
+  Optional<FPValueAndVReg> K0, K1;
+  // Match min(max(Val, K0), K1) or max(min(Val, K1), K0). Then see if K0 <= K1.
+  if (!matchMed<GFCstAndRegMatch>(MI, MRI, OpcodeTriple, Val, K0, K1))
+    return false;
+
+  if (K0->Value > K1->Value)
+    return false;
+
+  // For IEEE=false perform combine only when it's safe to assume that there are
+  // no NaN inputs. Most often MI is marked with nnan fast math flag.
+  // For IEEE=true consider NaN inputs. fmed3(NaN, K0, K1) is equivalent to
+  // min(min(NaN, K0), K1). Safe to fold for min(max(Val, K0), K1) since inner
+  // nodes(max/min) have same behavior when one input is NaN and other isn't.
+  // Don't consider max(min(SNaN, K1), K0) since there is no isKnownNeverQNaN,
+  // also post-legalizer inputs to min/max are fcanonicalized (never SNaN).
+  if ((getIEEE() && isFminnumIeee(MI)) || isKnownNeverNaN(Dst, MRI)) {
+    // Don't fold single use constant that can't be inlined.
+    const SIInstrInfo *TII = MF.getSubtarget<GCNSubtarget>().getInstrInfo();
+    if ((!MRI.hasOneNonDBGUse(K0->VReg) || TII->isInlineConstant(K0->Value)) &&
+        (!MRI.hasOneNonDBGUse(K1->VReg) || TII->isInlineConstant(K1->Value))) {
+      MatchInfo = {OpcodeTriple.Med, Val, K0->VReg, K1->VReg};
+      return true;
+    }
+  }
+
+  return false;
+}
 void AMDGPURegBankCombinerHelper::applyMed3(MachineInstr &MI,
                                             Med3MatchInfo &MatchInfo) {
   B.setInstrAndDebugLoc(MI);
   B.buildInstr(MatchInfo.Opc, {MI.getOperand(0)},
                {MatchInfo.Val0, MatchInfo.Val1, MatchInfo.Val2}, MI.getFlags());
   MI.eraseFromParent();
 }
 
+AMDGPU::SIModeRegisterDefaults AMDGPURegBankCombinerHelper::getMode() {
+  return MF.getInfo<SIMachineFunctionInfo>()->getMode();
+}
+
+bool AMDGPURegBankCombinerHelper::getIEEE() { return getMode().IEEE; }
+
+bool AMDGPURegBankCombinerHelper::isFminnumIeee(const MachineInstr &MI) {
+  return MI.getOpcode() == AMDGPU::G_FMINNUM_IEEE;
+}
+
 class AMDGPURegBankCombinerHelperState {
 protected:
   CombinerHelper &Helper;
   AMDGPURegBankCombinerHelper &RegBankHelper;
 
 public:
   AMDGPURegBankCombinerHelperState(CombinerHelper &Helper,
                                    AMDGPURegBankCombinerHelper &RegBankHelper)

llvm/lib/Target/AMDGPU/SIInstructions.td

 }
 
 def G_AMDGPU_UMED3 : AMDGPUGenericInstruction {
   let OutOperandList = (outs type0:$dst);
   let InOperandList = (ins type0:$src0, type0:$src1, type0:$src2);
   let hasSideEffects = 0;
 }
 
+def G_AMDGPU_FMED3 : AMDGPUGenericInstruction {
+  let OutOperandList = (outs type0:$dst);
+  let InOperandList = (ins type0:$src0, type0:$src1, type0:$src2);
+  let hasSideEffects = 0;
+}
+
 // Atomic cmpxchg. $cmpval ad $newval are packed in a single vector
 // operand Expects a MachineMemOperand in addition to explicit
 // operands.
 def G_AMDGPU_ATOMIC_CMPXCHG : AMDGPUGenericInstruction {
   let OutOperandList = (outs type0:$oldval);
   let InOperandList = (ins ptype1:$addr, type0:$cmpval_newval);
   let hasSideEffects = 0;
   let mayLoad = 1;

llvm/test/CodeGen/AMDGPU/GlobalISel/fmed3-min-max-const-combine.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -global-isel -mtriple=amdgcn-amd-mesa3d -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck -check-prefix=GFX10 %s
+
+define float @test_min_max_ValK0_K1_f32(float %a) #0 {
+; GFX10-LABEL: test_min_max_ValK0_K1_f32:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    v_med3_f32 v0, v0, 2.0, 4.0
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %maxnum = call nnan float @llvm.maxnum.f32(float %a, float 2.0)
+  %fmed = call nnan float @llvm.minnum.f32(float %maxnum, float 4.0)
+  ret float %fmed
+}
+
+define float @test_min_max_K0Val_K1_f32(float %a) #1 {
+; GFX10-LABEL: test_min_max_K0Val_K1_f32:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    v_med3_f32 v0, v0, 2.0, 4.0
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %maxnum = call nnan float @llvm.maxnum.f32(float 2.0, float %a)
+  %fmed = call nnan float @llvm.minnum.f32(float %maxnum, float 4.0)
+  ret float %fmed
+}
+
+; min-max patterns for ieee=true do not have to check for NaNs
+; 'v_max_f16_e32 v0, v0, v0' is from fcanonicalize of the input to fmin/fmax with ieee=true
+define half @test_min_K1max_ValK0_f16(half %a) #0 {
+; GFX10-LABEL: test_min_K1max_ValK0_f16:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    v_max_f16_e32 v0, v0, v0
+; GFX10-NEXT:    v_med3_f16 v0, v0, 2.0, 4.0
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %maxnum = call half @llvm.maxnum.f16(half %a, half 2.0)
+  %fmed = call half @llvm.minnum.f16(half 4.0, half %maxnum)
+  ret half %fmed
+}
+
+define half @test_min_K1max_K0Val_f16(half %a) #1 {
+; GFX10-LABEL: test_min_K1max_K0Val_f16:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    v_med3_f16 v0, v0, 2.0, 4.0
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %maxnum = call nnan half @llvm.maxnum.f16(half 2.0, half %a)
+  %fmed = call nnan half @llvm.minnum.f16(half 4.0, half %maxnum)
+  ret half %fmed
+}
+
+; max-mix patterns work only for non-NaN inputs
+define float @test_max_min_ValK1_K0_f32(float %a) #0 {
+; GFX10-LABEL: test_max_min_ValK1_K0_f32:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    v_med3_f32 v0, v0, 2.0, 4.0
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %minnum = call nnan float @llvm.minnum.f32(float %a, float 4.0)
+  %fmed = call nnan float @llvm.maxnum.f32(float %minnum, float 2.0)
+  ret float %fmed
+}
+
+define float @test_max_min_K1Val_K0_f32(float %a) #1 {
+; GFX10-LABEL: test_max_min_K1Val_K0_f32:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    v_med3_f32 v0, v0, 2.0, 4.0
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %minnum = call nnan float @llvm.minnum.f32(float 4.0, float %a)
+  %fmed = call nnan float @llvm.maxnum.f32(float %minnum, float 2.0)
+  ret float %fmed
+}
+
+define half @test_max_K0min_ValK1_f16(half %a) #0 {
+; GFX10-LABEL: test_max_K0min_ValK1_f16:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    v_med3_f16 v0, v0, 2.0, 4.0
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %minnum = call nnan half @llvm.minnum.f16(half %a, half 4.0)
+  %fmed = call nnan half @llvm.maxnum.f16(half 2.0, half %minnum)
+  ret half %fmed
+}
+
+define half @test_max_K0min_K1Val_f16(half %a) #1 {
+; GFX10-LABEL: test_max_K0min_K1Val_f16:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    v_med3_f16 v0, v0, 2.0, 4.0
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %minnum = call nnan half @llvm.minnum.f16(half 4.0, half %a)
+  %fmed = call nnan half @llvm.maxnum.f16(half 2.0, half %minnum)
+  ret half %fmed
+}
+
+; global nnan function attribute always forces fmed3 combine
+
+define float @test_min_max_global_nnan(float %a) #2 {
+; GFX10-LABEL: test_min_max_global_nnan:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    v_med3_f32 v0, v0, 2.0, 4.0
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %maxnum = call float @llvm.maxnum.f32(float %a, float 2.0)
+  %fmed = call float @llvm.minnum.f32(float %maxnum, float 4.0)
+  ret float %fmed
+}
+
+define float @test_max_min_global_nnan(float %a) #2 {
+; GFX10-LABEL: test_max_min_global_nnan:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    v_med3_f32 v0, v0, 2.0, 4.0
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %minnum = call float @llvm.minnum.f32(float %a, float 4.0)
+  %fmed = call float @llvm.maxnum.f32(float %minnum, float 2.0)
+  ret float %fmed
+}
+
+; ------------------------------------------------------------------------------
+; Negative patterns
+; ------------------------------------------------------------------------------
+
+; min(max(Val, K0), K1) K0 > K1, should be K0<=K1
+define float @test_min_max_K0_gt_K1(float %a) #0 {
+; GFX10-LABEL: test_min_max_K0_gt_K1:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    v_max_f32_e32 v0, 4.0, v0
+; GFX10-NEXT:    v_min_f32_e32 v0, 2.0, v0
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %maxnum = call nnan float @llvm.maxnum.f32(float %a, float 4.0)
+  %fmed = call nnan float @llvm.minnum.f32(float %maxnum, float 2.0)
+  ret float %fmed
+}
+
+; max(min(Val, K1), K0) K0 > K1, should be K0<=K1
+define float @test_max_min_K0_gt_K1(float %a) #0 {
+; GFX10-LABEL: test_max_min_K0_gt_K1:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    v_min_f32_e32 v0, 2.0, v0
+; GFX10-NEXT:    v_max_f32_e32 v0, 4.0, v0
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %minnum = call nnan float @llvm.minnum.f32(float %a, float 2.0)
+  %fmed = call nnan float @llvm.maxnum.f32(float %minnum, float 4.0)
+  ret float %fmed
+}
+
+; non-inline constant
+define float @test_min_max_non_inline_const(float %a) #0 {
+; GFX10-LABEL: test_min_max_non_inline_const:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    v_max_f32_e32 v0, 2.0, v0
+; GFX10-NEXT:    v_min_f32_e32 v0, 0x41000000, v0
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %maxnum = call nnan float @llvm.maxnum.f32(float %a, float 2.0)
+  %fmed = call nnan float @llvm.minnum.f32(float %maxnum, float 8.0)
+  ret float %fmed
+}
+
+; there is no fmed3 for f64 or v2f16 types
+
+define double @test_min_max_f64(double %a) #0 {
+; GFX10-LABEL: test_min_max_f64:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    v_max_f64 v[0:1], v[0:1], 2.0
+; GFX10-NEXT:    v_min_f64 v[0:1], v[0:1], 4.0
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %maxnum = call nnan double @llvm.maxnum.f64(double %a, double 2.0)
+  %fmed = call nnan double @llvm.minnum.f64(double %maxnum, double 4.0)
+  ret double %fmed
+}
+
+define <2 x half> @test_min_max_v2f16(<2 x half> %a) #0 {
+; GFX10-LABEL: test_min_max_v2f16:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    v_pk_max_f16 v0, v0, 2.0 op_sel_hi:[1,0]
+; GFX10-NEXT:    v_pk_min_f16 v0, v0, 4.0 op_sel_hi:[1,0]
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %maxnum = call nnan <2 x half> @llvm.maxnum.v2f16(<2 x half> %a, <2 x half> <half 2.0, half 2.0>)
+  %fmed = call nnan <2 x half> @llvm.minnum.v2f16(<2 x half> %maxnum, <2 x half> <half 4.0, half 4.0>)
+  ret <2 x half> %fmed
+}
+
+; input that can be NaN
+
+; min-max patterns for ieee=false require known non-NaN input
+define float @test_min_max_maybe_NaN_input_ieee_false(float %a) #1 {
+; GFX10-LABEL: test_min_max_maybe_NaN_input_ieee_false:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    v_max_f32_e32 v0, 2.0, v0
+; GFX10-NEXT:    v_min_f32_e32 v0, 4.0, v0
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %maxnum = call float @llvm.maxnum.f32(float %a, float 2.0)
+  %fmed = call float @llvm.minnum.f32(float %maxnum, float 4.0)
+  ret float %fmed
+}
+
+; max-min patterns always require known non-NaN input
+
+define float @test_max_min_maybe_NaN_input_ieee_false(float %a) #1 {
+; GFX10-LABEL: test_max_min_maybe_NaN_input_ieee_false:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    v_min_f32_e32 v0, 4.0, v0
+; GFX10-NEXT:    v_max_f32_e32 v0, 2.0, v0
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %minnum = call float @llvm.minnum.f32(float %a, float 4.0)
+  %fmed = call float @llvm.maxnum.f32(float %minnum, float 2.0)
+  ret float %fmed
+}
+
+; 'v_max_f32_e32 v0, v0, v0' is from fcanonicalize of the input to fmin/fmax with ieee=true
+define float @test_max_min_maybe_NaN_input_ieee_true(float %a) #0 {
+; GFX10-LABEL: test_max_min_maybe_NaN_input_ieee_true:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    v_max_f32_e32 v0, v0, v0
+; GFX10-NEXT:    v_min_f32_e32 v0, 4.0, v0
+; GFX10-NEXT:    v_max_f32_e32 v0, 2.0, v0
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %minnum = call float @llvm.minnum.f32(float %a, float 4.0)
+  %fmed = call float @llvm.maxnum.f32(float %minnum, float 2.0)
+  ret float %fmed
+}
+
+declare half @llvm.minnum.f16(half, half)
+declare half @llvm.maxnum.f16(half, half)
+declare float @llvm.minnum.f32(float, float)
+declare float @llvm.maxnum.f32(float, float)
+declare double @llvm.minnum.f64(double, double)
+declare double @llvm.maxnum.f64(double, double)
+declare <2 x half> @llvm.minnum.v2f16(<2 x half>, <2 x half>)
+declare <2 x half> @llvm.maxnum.v2f16(<2 x half>, <2 x half>)
+attributes #0 = {"amdgpu-ieee"="true"}
+attributes #1 = {"amdgpu-ieee"="false"}
+attributes #2 = {"no-nans-fp-math"="true"}

llvm/test/CodeGen/AMDGPU/GlobalISel/regbankcombiner-fmed3-minmax-const.mir

+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -mtriple=amdgcn-amd-mesa3d -mcpu=gfx1010 -run-pass=amdgpu-regbank-combiner -verify-machineinstrs %s -o - | FileCheck %s
+
+---
+name: test_min_max_ValK0_K1_f32
+legalized: true
+regBankSelected: true
+tracksRegLiveness: true
+machineFunctionInfo:
+  mode:
+    ieee: true
+body: |
+  bb.1:
+    liveins: $vgpr0
+
+    ; CHECK-LABEL: name: test_min_max_ValK0_K1_f32
+    ; CHECK: liveins: $vgpr0
+    ; CHECK: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
+    ; CHECK: [[C:%[0-9]+]]:sgpr(s32) = G_FCONSTANT float 2.000000e+00
+    ; CHECK: [[C1:%[0-9]+]]:sgpr(s32) = G_FCONSTANT float 4.000000e+00
+    ; CHECK: [[AMDGPU_FMED3_:%[0-9]+]]:vgpr(s32) = nnan G_AMDGPU_FMED3 [[COPY]], [[C]], [[C1]]
+    ; CHECK: $vgpr0 = COPY [[AMDGPU_FMED3_]](s32)
+    %0:vgpr(s32) = COPY $vgpr0
+    %2:sgpr(s32) = G_FCONSTANT float 2.000000e+00
+    %7:vgpr(s32) = COPY %2(s32)
+    %3:vgpr(s32) = nnan G_FMAXNUM_IEEE %0, %7
+    %4:sgpr(s32) = G_FCONSTANT float 4.000000e+00
+    %8:vgpr(s32) = COPY %4(s32)
+    %5:vgpr(s32) = nnan G_FMINNUM_IEEE %3, %8
+    $vgpr0 = COPY %5(s32)
+...
+
+---
+name: test_min_max_K0Val_K1_f32
+legalized: true
+regBankSelected: true
+tracksRegLiveness: true
+machineFunctionInfo:
+  mode:
+    ieee: false
+body: |
+  bb.1:
+    liveins: $vgpr0
+
+    ; CHECK-LABEL: name: test_min_max_K0Val_K1_f32
+    ; CHECK: liveins: $vgpr0
+    ; CHECK: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
+    ; CHECK: [[C:%[0-9]+]]:sgpr(s32) = G_FCONSTANT float 2.000000e+00
+    ; CHECK: [[C1:%[0-9]+]]:sgpr(s32) = G_FCONSTANT float 4.000000e+00
+    ; CHECK: [[AMDGPU_FMED3_:%[0-9]+]]:vgpr(s32) = nnan G_AMDGPU_FMED3 [[COPY]], [[C]], [[C1]]
+    ; CHECK: $vgpr0 = COPY [[AMDGPU_FMED3_]](s32)
+    %0:vgpr(s32) = COPY $vgpr0
+    %2:sgpr(s32) = G_FCONSTANT float 2.000000e+00
+    %7:vgpr(s32) = COPY %2(s32)
+    %3:vgpr(s32) = nnan G_FMAXNUM %7, %0
+    %4:sgpr(s32) = G_FCONSTANT float 4.000000e+00
+    %8:vgpr(s32) = COPY %4(s32)
+    %5:vgpr(s32) = nnan G_FMINNUM %3, %8
+    $vgpr0 = COPY %5(s32)
+...
+
+---
+name: test_min_K1max_ValK0_f16
+legalized: true
+regBankSelected: true
+tracksRegLiveness: true
+machineFunctionInfo:
+  mode:
+    ieee: true
+body: |
+  bb.1:
+    liveins: $vgpr0
+
+    ; CHECK-LABEL: name: test_min_K1max_ValK0_f16
+    ; CHECK: liveins: $vgpr0
+    ; CHECK: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
+    ; CHECK: [[TRUNC:%[0-9]+]]:vgpr(s16) = G_TRUNC [[COPY]](s32)
+    ; CHECK: [[C:%[0-9]+]]:sgpr(s16) = G_FCONSTANT half 0xH4000
+    ; CHECK: [[FCANONICALIZE:%[0-9]+]]:vgpr(s16) = G_FCANONICALIZE [[TRUNC]]
+    ; CHECK: [[C1:%[0-9]+]]:sgpr(s16) = G_FCONSTANT half 0xH4400
+    ; CHECK: [[AMDGPU_FMED3_:%[0-9]+]]:vgpr(s16) = G_AMDGPU_FMED3 [[FCANONICALIZE]], [[C]], [[C1]]
+    ; CHECK: [[ANYEXT:%[0-9]+]]:vgpr(s32) = G_ANYEXT [[AMDGPU_FMED3_]](s16)
+    ; CHECK: $vgpr0 = COPY [[ANYEXT]](s32)
+    %2:vgpr(s32) = COPY $vgpr0
+    %0:vgpr(s16) = G_TRUNC %2(s32)
+    %3:sgpr(s16) = G_FCONSTANT half 0xH4000
+    %9:vgpr(s16) = G_FCANONICALIZE %0
+    %10:vgpr(s16) = COPY %3(s16)
+    %4:vgpr(s16) = G_FMAXNUM_IEEE %9, %10
+    %5:sgpr(s16) = G_FCONSTANT half 0xH4400
+    %11:vgpr(s16) = COPY %5(s16)
+    %6:vgpr(s16) = G_FMINNUM_IEEE %11, %4
+    %8:vgpr(s32) = G_ANYEXT %6(s16)
+    $vgpr0 = COPY %8(s32)
+
+...
+
+---
+name: test_min_K1max_K0Val_f16
+legalized: true
+regBankSelected: true
+tracksRegLiveness: true
+machineFunctionInfo:
+  mode:
+    ieee: false
+body: |
+  bb.1:
+    liveins: $vgpr0, $sgpr30_sgpr31
+
+    ; CHECK-LABEL: name: test_min_K1max_K0Val_f16
+    ; CHECK: liveins: $vgpr0, $sgpr30_sgpr31
+    ; CHECK: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
+    ; CHECK: [[TRUNC:%[0-9]+]]:vgpr(s16) = G_TRUNC [[COPY]](s32)
+    ; CHECK: [[C:%[0-9]+]]:sgpr(s16) = G_FCONSTANT half 0xH4000
+    ; CHECK: [[C1:%[0-9]+]]:sgpr(s16) = G_FCONSTANT half 0xH4400
+    ; CHECK: [[AMDGPU_FMED3_:%[0-9]+]]:vgpr(s16) = nnan G_AMDGPU_FMED3 [[TRUNC]], [[C]], [[C1]]
+    ; CHECK: [[ANYEXT:%[0-9]+]]:vgpr(s32) = G_ANYEXT [[AMDGPU_FMED3_]](s16)
+    ; CHECK: $vgpr0 = COPY [[ANYEXT]](s32)
+    %2:vgpr(s32) = COPY $vgpr0
+    %0:vgpr(s16) = G_TRUNC %2(s32)
+    %3:sgpr(s16) = G_FCONSTANT half 0xH4000
+    %9:vgpr(s16) = COPY %3(s16)
+    %4:vgpr(s16) = nnan G_FMAXNUM %9, %0
+    %5:sgpr(s16) = G_FCONSTANT half 0xH4400
+    %10:vgpr(s16) = COPY %5(s16)
+    %6:vgpr(s16) = nnan G_FMINNUM %10, %4
+    %8:vgpr(s32) = G_ANYEXT %6(s16)
+    $vgpr0 = COPY %8(s32)
+...
+
+---
+name: test_max_min_ValK1_K0_f32
+legalized: true
+regBankSelected: true
+tracksRegLiveness: true
+machineFunctionInfo:
+  mode:
+    ieee: true
+body: |
+  bb.1:
+    liveins: $vgpr0
+
+    ; CHECK-LABEL: name: test_max_min_ValK1_K0_f32
+    ; CHECK: liveins: $vgpr0
+    ; CHECK: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
+    ; CHECK: [[C:%[0-9]+]]:sgpr(s32) = G_FCONSTANT float 4.000000e+00
+    ; CHECK: [[C1:%[0-9]+]]:sgpr(s32) = G_FCONSTANT float 2.000000e+00
+    ; CHECK: [[AMDGPU_FMED3_:%[0-9]+]]:vgpr(s32) = nnan G_AMDGPU_FMED3 [[COPY]], [[C1]], [[C]]
+    ; CHECK: $vgpr0 = COPY [[AMDGPU_FMED3_]](s32)
+    %0:vgpr(s32) = COPY $vgpr0
+    %2:sgpr(s32) = G_FCONSTANT float 4.000000e+00
+    %7:vgpr(s32) = COPY %2(s32)
+    %3:vgpr(s32) = nnan G_FMINNUM_IEEE %0, %7
+    %4:sgpr(s32) = G_FCONSTANT float 2.000000e+00
+    %8:vgpr(s32) = COPY %4(s32)
+    %5:vgpr(s32) = nnan G_FMAXNUM_IEEE %3, %8
+    $vgpr0 = COPY %5(s32)
+...
+
+---
+name: test_max_min_K1Val_K0_f32
+legalized: true
+regBankSelected: true
+tracksRegLiveness: true
+machineFunctionInfo:
+  mode:
+    ieee: false
+body: |
+  bb.1:
+    liveins: $vgpr0
+
+    ; CHECK-LABEL: name: test_max_min_K1Val_K0_f32
+    ; CHECK: liveins: $vgpr0
+    ; CHECK: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
+    ; CHECK: [[C:%[0-9]+]]:sgpr(s32) = G_FCONSTANT float 4.000000e+00
+    ; CHECK: [[C1:%[0-9]+]]:sgpr(s32) = G_FCONSTANT float 2.000000e+00
+    ; CHECK: [[AMDGPU_FMED3_:%[0-9]+]]:vgpr(s32) = nnan G_AMDGPU_FMED3 [[COPY]], [[C1]], [[C]]
+    ; CHECK: $vgpr0 = COPY [[AMDGPU_FMED3_]](s32)
+    %0:vgpr(s32) = COPY $vgpr0
+    %2:sgpr(s32) = G_FCONSTANT float 4.000000e+00
+    %7:vgpr(s32) = COPY %2(s32)
+    %3:vgpr(s32) = nnan G_FMINNUM %7, %0
+    %4:sgpr(s32) = G_FCONSTANT float 2.000000e+00
+    %8:vgpr(s32) = COPY %4(s32)
+    %5:vgpr(s32) = nnan G_FMAXNUM %3, %8
+    $vgpr0 = COPY %5(s32)
+...
+
+---
+name: test_max_K0min_ValK1_f16
+legalized: true
+regBankSelected: true
+tracksRegLiveness: true
+machineFunctionInfo:
+  mode:
+    ieee: true
+body: |
+  bb.1:
+    liveins: $vgpr0
+
+    ; CHECK-LABEL: name: test_max_K0min_ValK1_f16
+    ; CHECK: liveins: $vgpr0
+    ; CHECK: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
+    ; CHECK: [[TRUNC:%[0-9]+]]:vgpr(s16) = G_TRUNC [[COPY]](s32)
+    ; CHECK: [[C:%[0-9]+]]:sgpr(s16) = G_FCONSTANT half 0xH4400
+    ; CHECK: [[C1:%[0-9]+]]:sgpr(s16) = G_FCONSTANT half 0xH4000
+    ; CHECK: [[AMDGPU_FMED3_:%[0-9]+]]:vgpr(s16) = nnan G_AMDGPU_FMED3 [[TRUNC]], [[C1]], [[C]]
+    ; CHECK: [[ANYEXT:%[0-9]+]]:vgpr(s32) = G_ANYEXT [[AMDGPU_FMED3_]](s16)
+    ; CHECK: $vgpr0 = COPY [[ANYEXT]](s32)
+    %2:vgpr(s32) = COPY $vgpr0
+    %0:vgpr(s16) = G_TRUNC %2(s32)
+    %3:sgpr(s16) = G_FCONSTANT half 0xH4400
+    %9:vgpr(s16) = COPY %3(s16)
+    %4:vgpr(s16) = nnan G_FMINNUM_IEEE %0, %9
+    %5:sgpr(s16) = G_FCONSTANT half 0xH4000
+    %10:vgpr(s16) = COPY %5(s16)
+    %6:vgpr(s16) = nnan G_FMAXNUM_IEEE %10, %4
+    %8:vgpr(s32) = G_ANYEXT %6(s16)
+    $vgpr0 = COPY %8(s32)
+...
+
+---
+name: test_max_K0min_K1Val_f16
+legalized: true
+regBankSelected: true
+tracksRegLiveness: true
+machineFunctionInfo:
+  mode:
+    ieee: false
+body: |
+  bb.1:
+    liveins: $vgpr0, $sgpr30_sgpr31
+
+    ; CHECK-LABEL: name: test_max_K0min_K1Val_f16
+    ; CHECK: liveins: $vgpr0, $sgpr30_sgpr31
+    ; CHECK: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
+    ; CHECK: [[TRUNC:%[0-9]+]]:vgpr(s16) = G_TRUNC [[COPY]](s32)
+    ; CHECK: [[C:%[0-9]+]]:sgpr(s16) = G_FCONSTANT half 0xH4400
+    ; CHECK: [[C1:%[0-9]+]]:sgpr(s16) = G_FCONSTANT half 0xH4000
+    ; CHECK: [[AMDGPU_FMED3_:%[0-9]+]]:vgpr(s16) = nnan G_AMDGPU_FMED3 [[TRUNC]], [[C1]], [[C]]
+    ; CHECK: [[ANYEXT:%[0-9]+]]:vgpr(s32) = G_ANYEXT [[AMDGPU_FMED3_]](s16)
+    ; CHECK: $vgpr0 = COPY [[ANYEXT]](s32)
+    %2:vgpr(s32) = COPY $vgpr0
+    %0:vgpr(s16) = G_TRUNC %2(s32)
+    %3:sgpr(s16) = G_FCONSTANT half 0xH4400
+    %9:vgpr(s16) = COPY %3(s16)
+    %4:vgpr(s16) = nnan G_FMINNUM %9, %0
+    %5:sgpr(s16) = G_FCONSTANT half 0xH4000
+    %10:vgpr(s16) = COPY %5(s16)
+    %6:vgpr(s16) = nnan G_FMAXNUM %10, %4
+    %8:vgpr(s32) = G_ANYEXT %6(s16)
+    $vgpr0 = COPY %8(s32)
+...
+
+# FixMe: add tests with attributes #2 = {"no-nans-fp-math"="true"}
+
+---
+name: test_min_max_K0_gt_K1
+legalized: true
+regBankSelected: true
+tracksRegLiveness: true
+machineFunctionInfo:
+  mode:
+    ieee: true
+body: |
+  bb.1:
+    liveins: $vgpr0
+
+    ; CHECK-LABEL: name: test_min_max_K0_gt_K1
+    ; CHECK: liveins: $vgpr0
+    ; CHECK: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
+    ; CHECK: [[C:%[0-9]+]]:sgpr(s32) = G_FCONSTANT float 4.000000e+00
+    ; CHECK: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY [[C]](s32)
+    ; CHECK: [[FMAXNUM_IEEE:%[0-9]+]]:vgpr(s32) = nnan G_FMAXNUM_IEEE [[COPY]], [[COPY1]]
+    ; CHECK: [[C1:%[0-9]+]]:sgpr(s32) = G_FCONSTANT float 2.000000e+00
+    ; CHECK: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY [[C1]](s32)
+    ; CHECK: [[FMINNUM_IEEE:%[0-9]+]]:vgpr(s32) = nnan G_FMINNUM_IEEE [[FMAXNUM_IEEE]], [[COPY2]]
+    ; CHECK: $vgpr0 = COPY [[FMINNUM_IEEE]](s32)
+    %0:vgpr(s32) = COPY $vgpr0
+    %2:sgpr(s32) = G_FCONSTANT float 4.000000e+00
+    %7:vgpr(s32) = COPY %2(s32)
+    %3:vgpr(s32) = nnan G_FMAXNUM_IEEE %0, %7
+    %4:sgpr(s32) = G_FCONSTANT float 2.000000e+00
+    %8:vgpr(s32) = COPY %4(s32)
+    %5:vgpr(s32) = nnan G_FMINNUM_IEEE %3, %8
+    $vgpr0 = COPY %5(s32)
+...
+
+---
+name: test_max_min_K0_gt_K1
+legalized: true
+regBankSelected: true
+tracksRegLiveness: true
+machineFunctionInfo:
+  mode:
+    ieee: true
+body: |
+  bb.1:
+    liveins: $vgpr0
+
+    ; CHECK-LABEL: name: test_max_min_K0_gt_K1
+    ; CHECK: liveins: $vgpr0
+    ; CHECK: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
+    ; CHECK: [[C:%[0-9]+]]:sgpr(s32) = G_FCONSTANT float 2.000000e+00
+    ; CHECK: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY [[C]](s32)
+    ; CHECK: [[FMINNUM_IEEE:%[0-9]+]]:vgpr(s32) = nnan G_FMINNUM_IEEE [[COPY]], [[COPY1]]
+    ; CHECK: [[C1:%[0-9]+]]:sgpr(s32) = G_FCONSTANT float 4.000000e+00
+    ; CHECK: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY [[C1]](s32)
+    ; CHECK: [[FMAXNUM_IEEE:%[0-9]+]]:vgpr(s32) = nnan G_FMAXNUM_IEEE [[FMINNUM_IEEE]], [[COPY2]]
+    ; CHECK: $vgpr0 = COPY [[FMAXNUM_IEEE]](s32)
+    %0:vgpr(s32) = COPY $vgpr0
+    %2:sgpr(s32) = G_FCONSTANT float 2.000000e+00
+    %7:vgpr(s32) = COPY %2(s32)
+    %3:vgpr(s32) = nnan G_FMINNUM_IEEE %0, %7
+    %4:sgpr(s32) = G_FCONSTANT float 4.000000e+00
+    %8:vgpr(s32) = COPY %4(s32)
+    %5:vgpr(s32) = nnan G_FMAXNUM_IEEE %3, %8
+    $vgpr0 = COPY %5(s32)
+...
+
+---
+name: test_min_max_non_inline_const
+legalized: true
+regBankSelected: true
+tracksRegLiveness: true
+machineFunctionInfo:
+  mode:
+    ieee: true
+body: |
+  bb.1:
+    liveins: $vgpr0
+
+    ; CHECK-LABEL: name: test_min_max_non_inline_const
+    ; CHECK: liveins: $vgpr0
+    ; CHECK: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
+    ; CHECK: [[C:%[0-9]+]]:sgpr(s32) = G_FCONSTANT float 2.000000e+00
+    ; CHECK: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY [[C]](s32)
+    ; CHECK: [[FMAXNUM_IEEE:%[0-9]+]]:vgpr(s32) = nnan G_FMAXNUM_IEEE [[COPY]], [[COPY1]]
+    ; CHECK: [[C1:%[0-9]+]]:sgpr(s32) = G_FCONSTANT float 8.000000e+00
+    ; CHECK: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY [[C1]](s32)
+    ; CHECK: [[FMINNUM_IEEE:%[0-9]+]]:vgpr(s32) = nnan G_FMINNUM_IEEE [[FMAXNUM_IEEE]], [[COPY2]]
+    ; CHECK: $vgpr0 = COPY [[FMINNUM_IEEE]](s32)
+    %0:vgpr(s32) = COPY $vgpr0
+    %2:sgpr(s32) = G_FCONSTANT float 2.000000e+00
+    %7:vgpr(s32) = COPY %2(s32)
+    %3:vgpr(s32) = nnan G_FMAXNUM_IEEE %0, %7
+    %4:sgpr(s32) = G_FCONSTANT float 8.000000e+00
+    %8:vgpr(s32) = COPY %4(s32)
+    %5:vgpr(s32) = nnan G_FMINNUM_IEEE %3, %8
+    $vgpr0 = COPY %5(s32)
+...
+
+---
+name: test_min_max_f64
+legalized: true
+regBankSelected: true
+tracksRegLiveness: true
+machineFunctionInfo:
+  mode:
+    ieee: true
+body: |
+  bb.1:
+    liveins: $vgpr0_vgpr1
+
+    ; CHECK-LABEL: name: test_min_max_f64
+    ; CHECK: liveins: $vgpr0_vgpr1
+    ; CHECK: [[COPY:%[0-9]+]]:vgpr(s64) = COPY $vgpr0_vgpr1
+    ; CHECK: [[C:%[0-9]+]]:sgpr(s64) = G_FCONSTANT double 2.000000e+00
+    ; CHECK: [[COPY1:%[0-9]+]]:vgpr(s64) = COPY [[C]](s64)
+    ; CHECK: [[FMAXNUM_IEEE:%[0-9]+]]:vgpr(s64) = nnan G_FMAXNUM_IEEE [[COPY]], [[COPY1]]
+    ; CHECK: [[C1:%[0-9]+]]:sgpr(s64) = G_FCONSTANT double 4.000000e+00
+    ; CHECK: [[COPY2:%[0-9]+]]:vgpr(s64) = COPY [[C1]](s64)
+    ; CHECK: [[FMINNUM_IEEE:%[0-9]+]]:vgpr(s64) = nnan G_FMINNUM_IEEE [[FMAXNUM_IEEE]], [[COPY2]]
+    ; CHECK: $vgpr0_vgpr1 = COPY [[FMINNUM_IEEE]](s64)
+    %0:vgpr(s64) = COPY $vgpr0_vgpr1
+    %4:sgpr(s64) = G_FCONSTANT double 2.000000e+00
+    %11:vgpr(s64) = COPY %4(s64)
+    %5:vgpr(s64) = nnan G_FMAXNUM_IEEE %0, %11
+    %6:sgpr(s64) = G_FCONSTANT double 4.000000e+00
+    %12:vgpr(s64) = COPY %6(s64)
+    %7:vgpr(s64) = nnan G_FMINNUM_IEEE %5, %12
+    $vgpr0_vgpr1 = COPY %7(s64)
+...
+
+---
+name: test_min_max_v2f16
+legalized: true
+regBankSelected: true
+tracksRegLiveness: true
+machineFunctionInfo:
+  mode:
+    ieee: true
+body: |
+  bb.1:
+    liveins: $vgpr0
+
+    ; CHECK-LABEL: name: test_min_max_v2f16
+    ; CHECK: liveins: $vgpr0
+    ; CHECK: [[COPY:%[0-9]+]]:vgpr(<2 x s16>) = COPY $vgpr0
+    ; CHECK: [[C:%[0-9]+]]:sgpr(s16) = G_FCONSTANT half 0xH4000
+    ; CHECK: [[ANYEXT:%[0-9]+]]:sgpr(s32) = G_ANYEXT [[C]](s16)
+    ; CHECK: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:sgpr(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[ANYEXT]](s32), [[ANYEXT]](s32)
+    ; CHECK: [[C1:%[0-9]+]]:sgpr(s16) = G_FCONSTANT half 0xH4400
+    ; CHECK: [[ANYEXT1:%[0-9]+]]:sgpr(s32) = G_ANYEXT [[C1]](s16)
+    ; CHECK: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:sgpr(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[ANYEXT1]](s32), [[ANYEXT1]](s32)
+    ; CHECK: [[COPY1:%[0-9]+]]:vgpr(<2 x s16>) = COPY [[BUILD_VECTOR_TRUNC]](<2 x s16>)
+    ; CHECK: [[FMAXNUM_IEEE:%[0-9]+]]:vgpr(<2 x s16>) = nnan G_FMAXNUM_IEEE [[COPY]], [[COPY1]]
+    ; CHECK: [[COPY2:%[0-9]+]]:vgpr(<2 x s16>) = COPY [[BUILD_VECTOR_TRUNC1]](<2 x s16>)
+    ; CHECK: [[FMINNUM_IEEE:%[0-9]+]]:vgpr(<2 x s16>) = nnan G_FMINNUM_IEEE [[FMAXNUM_IEEE]], [[COPY2]]
+    ; CHECK: $vgpr0 = COPY [[FMINNUM_IEEE]](<2 x s16>)
+    %0:vgpr(<2 x s16>) = COPY $vgpr0
+    %3:sgpr(s16) = G_FCONSTANT half 0xH4000
+    %9:sgpr(s32) = G_ANYEXT %3(s16)
+    %2:sgpr(<2 x s16>) = G_BUILD_VECTOR_TRUNC %9(s32), %9(s32)
+    %6:sgpr(s16) = G_FCONSTANT half 0xH4400
+    %10:sgpr(s32) = G_ANYEXT %6(s16)
+    %5:sgpr(<2 x s16>) = G_BUILD_VECTOR_TRUNC %10(s32), %10(s32)
+    %11:vgpr(<2 x s16>) = COPY %2(<2 x s16>)
+    %4:vgpr(<2 x s16>) = nnan G_FMAXNUM_IEEE %0, %11
+    %12:vgpr(<2 x s16>) = COPY %5(<2 x s16>)
+    %7:vgpr(<2 x s16>) = nnan G_FMINNUM_IEEE %4, %12
+    $vgpr0 = COPY %7(<2 x s16>)
+...
+
+---
+name: test_min_max_maybe_NaN_input_ieee_false
+legalized: true
+regBankSelected: true
+tracksRegLiveness: true
+machineFunctionInfo:
+  mode:
+    ieee: false
+body: |
+  bb.1:
+    liveins: $vgpr0
+
+    ; CHECK-LABEL: name: test_min_max_maybe_NaN_input_ieee_false
+    ; CHECK: liveins: $vgpr0
+    ; CHECK: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
+    ; CHECK: [[C:%[0-9]+]]:sgpr(s32) = G_FCONSTANT float 2.000000e+00
+    ; CHECK: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY [[C]](s32)
+    ; CHECK: [[FMAXNUM:%[0-9]+]]:vgpr(s32) = G_FMAXNUM [[COPY]], [[COPY1]]
+    ; CHECK: [[C1:%[0-9]+]]:sgpr(s32) = G_FCONSTANT float 4.000000e+00
+    ; CHECK: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY [[C1]](s32)
+    ; CHECK: [[FMINNUM:%[0-9]+]]:vgpr(s32) = G_FMINNUM [[FMAXNUM]], [[COPY2]]
+    ; CHECK: $vgpr0 = COPY [[FMINNUM]](s32)
+    %0:vgpr(s32) = COPY $vgpr0
+    %2:sgpr(s32) = G_FCONSTANT float 2.000000e+00
+    %7:vgpr(s32) = COPY %2(s32)
+    %3:vgpr(s32) = G_FMAXNUM %0, %7
+    %4:sgpr(s32) = G_FCONSTANT float 4.000000e+00
+    %8:vgpr(s32) = COPY %4(s32)
+    %5:vgpr(s32) = G_FMINNUM %3, %8
+    $vgpr0 = COPY %5(s32)
+...
+
+---
+name: test_max_min_maybe_NaN_input_ieee_false
+legalized: true
+regBankSelected: true
+tracksRegLiveness: true
+machineFunctionInfo:
+  mode:
+    ieee: false
+body: |
+  bb.1:
+    liveins: $vgpr0
+
+    ; CHECK-LABEL: name: test_max_min_maybe_NaN_input_ieee_false
+    ; CHECK: liveins: $vgpr0
+    ; CHECK: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
+    ; CHECK: [[C:%[0-9]+]]:sgpr(s32) = G_FCONSTANT float 4.000000e+00
+    ; CHECK: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY [[C]](s32)
+    ; CHECK: [[FMINNUM:%[0-9]+]]:vgpr(s32) = G_FMINNUM [[COPY]], [[COPY1]]
+    ; CHECK: [[C1:%[0-9]+]]:sgpr(s32) = G_FCONSTANT float 2.000000e+00
+    ; CHECK: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY [[C1]](s32)
+    ; CHECK: [[FMAXNUM:%[0-9]+]]:vgpr(s32) = G_FMAXNUM [[FMINNUM]], [[COPY2]]
+    ; CHECK: $vgpr0 = COPY [[FMAXNUM]](s32)
+    %0:vgpr(s32) = COPY $vgpr0
+    %2:sgpr(s32) = G_FCONSTANT float 4.000000e+00
+    %7:vgpr(s32) = COPY %2(s32)
+    %3:vgpr(s32) = G_FMINNUM %0, %7
+    %4:sgpr(s32) = G_FCONSTANT float 2.000000e+00
+    %8:vgpr(s32) = COPY %4(s32)
+    %5:vgpr(s32) = G_FMAXNUM %3, %8
+    $vgpr0 = COPY %5(s32)
+...
+
+---
+name: test_max_min_maybe_NaN_input_ieee_true
+legalized: true
+regBankSelected: true
+tracksRegLiveness: true
+machineFunctionInfo:
+  mode:
+    ieee: true
+body: |
+  bb.1:
+    liveins: $vgpr0
+
+    ; CHECK-LABEL: name: test_max_min_maybe_NaN_input_ieee_true
+    ; CHECK: liveins: $vgpr0
+    ; CHECK: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
+    ; CHECK: [[C:%[0-9]+]]:sgpr(s32) = G_FCONSTANT float 4.000000e+00
+    ; CHECK: [[FCANONICALIZE:%[0-9]+]]:vgpr(s32) = G_FCANONICALIZE [[COPY]]
+    ; CHECK: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY [[C]](s32)
+    ; CHECK: [[FMINNUM_IEEE:%[0-9]+]]:vgpr(s32) = G_FMINNUM_IEEE [[FCANONICALIZE]], [[COPY1]]
+    ; CHECK: [[C1:%[0-9]+]]:sgpr(s32) = G_FCONSTANT float 2.000000e+00
+    ; CHECK: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY [[C1]](s32)
+    ; CHECK: [[FMAXNUM_IEEE:%[0-9]+]]:vgpr(s32) = G_FMAXNUM_IEEE [[FMINNUM_IEEE]], [[COPY2]]
+    ; CHECK: $vgpr0 = COPY [[FMAXNUM_IEEE]](s32)
+    %0:vgpr(s32) = COPY $vgpr0
+    %2:sgpr(s32) = G_FCONSTANT float 4.000000e+00
+    %7:vgpr(s32) = G_FCANONICALIZE %0
+    %8:vgpr(s32) = COPY %2(s32)
+    %3:vgpr(s32) = G_FMINNUM_IEEE %7, %8
+    %4:sgpr(s32) = G_FCONSTANT float 2.000000e+00
+    %9:vgpr(s32) = COPY %4(s32)
+    %5:vgpr(s32) = G_FMAXNUM_IEEE %3, %9
+    $vgpr0 = COPY %5(s32)
+...