AMDGPU/GlobalISel: Add clamp combine

[Petar.Avramovic]

Summary

Add clamp combine. Source is fminnum(fmaxnum(Val, K0), K1) or
fmaxnum(fminnum(Val, K1), K0) with K0=0.0 and K1=1.0 or fmed3
intrinsic with 0.0 and 1.0 as two out of three operands.

[arsenm]

Missing tests for f16 cases

llvm/lib/Target/AMDGPU/AMDGPUPostLegalizerCombiner.cpp

187 ↗

(On Diff #300312)

Check this first around the whole section?

207–216 ↗

(On Diff #300312)

These look like two distinct apply actions to me

224–228 ↗

(On Diff #300312)

getConstantFPVRegVal

233 ↗

(On Diff #300312)

This should be implied by the matcher opcode

234 ↗

(On Diff #300312)

Early return on intrinsic check

235–246 ↗

(On Diff #300312)

I think the flow of the DAG version is easier to handle, swapping the values and then checking isClampZeroToOne at the end

247 ↗

(On Diff #300312)

Typo remainig

254 ↗

(On Diff #300312)

Should get this from the MachineIRBuilder

256–260 ↗

(On Diff #300312)

I think something is confused here. IEEE mode only changes signaling nan behavior. I also think trying to handle IEEE=0 correctly with snans is a fools errand since *no* operations will behave correctly. The DAG combine for this does not consider IEEE mode

275 ↗

(On Diff #300312)

Should not construct a new MachineIRBuilder, there should be one existing already in the combiner

[Petar.Avramovic]

llvm/lib/Target/AMDGPU/AMDGPUPostLegalizerCombiner.cpp

207–216 ↗

(On Diff #300312)

They share same match action. If we want to split them to matchMinMaxToClamp and matchMinMaxToMed3 we might end up matching min/max pattern twice.

235–246 ↗

(On Diff #300312)

That swaps operands and might produce different result with IEEE=1

256–260 ↗

(On Diff #300312)

IEEE=0 does not treat SNaN differently and needs no SNaN checks. With IEEE=0 DX10Clamp=1 (shaders want all possible x2, x4, div2, and clamp modifiers) there is no need for any nan check. Other operation don't really mention that backend needs to check for special values. They do have different behavior when input is SNaN but correctness in not mentioned. What do we have to do force correct behavior for IEEE=0?
The considering of SNaN only comes from compute kernel functions being in IEEE=1 mode and trying to lower fminnum to fminnum_ieee (If higher language used fminnum_ieee with IEEE=1 there would be no need for SNaN checks at all).

[foad]

Is it possible to separate this into two different combines:

1. min(max(...)) or max(min(...)) -> med(x, K0, K1)
2. med(x, 0.0, 1.0) -> clamp(x)

?

[arsenm]

llvm/lib/Target/AMDGPU/AMDGPUPostLegalizerCombiner.cpp

235–246 ↗

(On Diff #300312)

Are you saying the DAG version is broken?

256–260 ↗

(On Diff #300312)

With IEEE=0, no operations quiet signalling nans. All operations behave incorrectly and pass through the snan instead of quieting it. The lack of quieting the snan is a correctness issue. It would be a massive amount of work to actually implement the correct behavior and manually quiet snan inputs to implement the correct generic operation semantics, which we will probably never do.

[Petar.Avramovic]

llvm/lib/Target/AMDGPU/AMDGPUPostLegalizerCombiner.cpp

235–246 ↗

(On Diff #300312)

DAG version should also check for SNaN (This is for IEEE=true) alongside DX10Clamp.
result of fmed3 depends on operand order when one is SNaN
swapping the operands and deciding to give up since input might be SNaN might create different result:
med3_f32 (0, 1, SNaN) = min(min(0,1), SNaN) = QNaN
med3_f32 (0, 1, SNaN) -(swap)-> med3_f32 (SNaN, 0, 1) = min(min(SNaN, 0), 1) = 1

swapping is fine for QNaN.

256–260 ↗

(On Diff #300312)

All operations behave incorrectly and pass through the snan instead of quieting it. The lack of quieting the snan is a correctness issue.

Is this for llvm-ir? Would it then be acceptable to only perform combine if nnan flag (check using isKnownNeverNaN) is present on the instruction with IEEE=false?
Also when IEEE=false we should never use isKnownNeverSNaN.
For IEEE=true we could then check for NaN input or DX10Clamp.

[arsenm]

llvm/lib/Target/AMDGPU/AMDGPUPostLegalizerCombiner.cpp

256–260 ↗

(On Diff #300312)

Yes. Officially the IR description pretends snans do not exist, but really they do and need to be quieted by FP canonicalizing FP operations as per IEEE754. I'm not sure it even makes sense to try to get correct snan behavior with IEEE=0 since it's never going to be implemented. The comment here absolutely needs to make clear the distinction between snan and qnan handling

Getting IEEE=true correct is far more interesting/important.

[arsenm]

reverse ping

[Petar.Avramovic]

Rebase/rework.

[Petar.Avramovic]

Add one test for splat padded with undef with ieee=true and dx10_clamp=true.

[Petar.Avramovic]

Ping.

[arsenm]

llvm/lib/Target/AMDGPU/AMDGPURegBankCombiner.cpp

218

I would expect the helper above to return the two constants directly rather than having to requery the registers

227–229

Should swap these conditions so the cheap checks are first

330–336

There's already a helper for this, getDefIgnoringCopies/getDefSrcRegIgnoringCopies

[Petar.Avramovic]

Use matcher that returns APFloat.

[foad]

llvm/lib/Target/AMDGPU/AMDGPUCombine.td

68

Can you use the standard register_matchdata for this?

[Petar.Avramovic]

Use register_matchinfo.

llvm/lib/Target/AMDGPU/AMDGPUCombine.td

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 clamp_matchdata : GIDefMatchData<"Register">;
+
+def fp_minmax_to_clamp : GICombineRule<
+  (defs root:$min_or_max, clamp_matchdata:$matchinfo),
+  (match (wip_match_opcode G_FMAXNUM,
+                           G_FMINNUM,
+                           G_FMAXNUM_IEEE,
+                           G_FMINNUM_IEEE):$min_or_max,
+         [{ return RegBankHelper.matchFPMinMaxToClamp(*${min_or_max}, ${matchinfo}); }]),
+  (apply [{ RegBankHelper.applyClamp(*${min_or_max}, ${matchinfo}); }])>;
+
+def fmed3_intrinsic_to_clamp : GICombineRule<
+  (defs root:$fmed3, clamp_matchdata:$matchinfo),
+  (match (wip_match_opcode G_INTRINSIC):$fmed3,
+         [{ return RegBankHelper.matchFPMed3ToClamp(*${fmed3}, ${matchinfo}); }]),
+  (apply [{ RegBankHelper.applyClamp(*${fmed3}, ${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<
    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, fp_minmax_to_med3]> {
+  [zext_trunc_fold, int_minmax_to_med3, fp_minmax_to_clamp,
+   fp_minmax_to_med3, fmed3_intrinsic_to_clamp]> {
   let DisableRuleOption = "amdgpuregbankcombiner-disable-rule";
   let StateClass = "AMDGPURegBankCombinerHelperState";
   let AdditionalArguments = [];
 }

llvm/lib/Target/AMDGPU/AMDGPUGISel.td

 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_CLAMP, AMDGPUclamp>;
 
 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 "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/IR/IntrinsicsAMDGPU.h"
 #include "llvm/Target/TargetMachine.h"
 #define DEBUG_TYPE "amdgpu-regbank-combiner"
 
 using namespace llvm;
 using namespace MIPatternMatch;
 
 class AMDGPURegBankCombinerHelper {
 protected:
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);
+  bool matchFPMinMaxToClamp(MachineInstr &MI, Register &Reg);
+  bool matchFPMed3ToClamp(MachineInstr &MI, Register &Reg);
   void applyMed3(MachineInstr &MI, Med3MatchInfo &MatchInfo);
+  void applyClamp(MachineInstr &MI, Register &Reg);
 
 private:
   AMDGPU::SIModeRegisterDefaults getMode();
   bool getIEEE();
+  bool getDX10Clamp();
   bool isFminnumIeee(const MachineInstr &MI);
+  bool isFCst(MachineInstr *MI);
+  bool isClampZeroToOne(MachineInstr *K0, MachineInstr *K1);
 };
 
 bool AMDGPURegBankCombinerHelper::isVgprRegBank(Register Reg) {
   return RBI.getRegBank(Reg, MRI, TRI)->getID() == AMDGPU::VGPRRegBankID;
 }
 
 AMDGPURegBankCombinerHelper::MinMaxMedOpc
 AMDGPURegBankCombinerHelper::getMinMaxPair(unsigned Opc) {
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;
 }
+
+bool AMDGPURegBankCombinerHelper::matchFPMinMaxToClamp(MachineInstr &MI,
+                                                       Register &Reg) {
+  // Clamp is available on all types after regbankselect (f16, f32, f64, v2f16).
+  auto OpcodeTriple = getMinMaxPair(MI.getOpcode());
+  Register Val;
+  Optional<FPValueAndVReg> K0, K1;
+  // Match min(max(Val, K0), K1) or max(min(Val, K1), K0).
+  if (!matchMed<GFCstOrSplatGFCstMatch>(MI, MRI, OpcodeTriple, Val, K0, K1))
+    return false;
+
+  if (!K0->Value.isExactlyValue(0.0) || !K1->Value.isExactlyValue(1.0))
+    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. Only min(max(QNaN, 0.0), 1.0) evaluates
+  // to 0.0 requires dx10_clamp = true.
+  if ((getIEEE() && getDX10Clamp() && isFminnumIeee(MI) &&
+       isKnownNeverSNaN(Val, MRI)) ||
+      isKnownNeverNaN(MI.getOperand(0).getReg(), MRI)) {
+    Reg = Val;
+    return true;
+  }
+
+  return false;
+}
+
+// Replacing fmed3(NaN, 0.0, 1.0) with clamp. Requires dx10_clamp = true.
+// Val = SNaN only for ieee = true. It is important which operand is NaN.
+// min(min(SNaN, 0.0), 1.0) = min(QNaN, 1.0) = 1.0
+// min(min(SNaN, 1.0), 0.0) = min(QNaN, 0.0) = 0.0
+// min(min(0.0, 1.0), SNaN) = min(0.0, SNaN) = QNaN
+// Val = NaN,ieee = false or Val = QNaN,ieee = true
+// min(min(NaN, 0.0), 1.0) = min(0.0, 1.0) = 0.0
+// min(min(NaN, 1.0), 0.0) = min(1.0, 0.0) = 0.0
+// min(min(0.0, 1.0), NaN) = min(0.0, NaN) = 0.0
+bool AMDGPURegBankCombinerHelper::matchFPMed3ToClamp(MachineInstr &MI,
+                                                     Register &Reg) {
+  if (MI.getIntrinsicID() != Intrinsic::amdgcn_fmed3)
+    return false;
+
+  // In llvm-ir, clamp is often represented as an intrinsic call to
+  // @llvm.amdgcn.fmed3.f32(%Val, 0.0, 1.0). Check for other operand orders.
+  MachineInstr *Src0 = getDefIgnoringCopies(MI.getOperand(2).getReg(), MRI);
+  MachineInstr *Src1 = getDefIgnoringCopies(MI.getOperand(3).getReg(), MRI);
+  MachineInstr *Src2 = getDefIgnoringCopies(MI.getOperand(4).getReg(), MRI);
+
+  if (isFCst(Src0) && !isFCst(Src1))
+    std::swap(Src0, Src1);
+  if (isFCst(Src1) && !isFCst(Src2))
+    std::swap(Src1, Src2);
+  if (isFCst(Src0) && !isFCst(Src1))
+    std::swap(Src0, Src1);
+  if (!isClampZeroToOne(Src1, Src2))
+    return false;
+
+  Register Val = Src0->getOperand(0).getReg();
+
+  auto isOp3Zero = [&]() {
+    MachineInstr *Op3 = getDefIgnoringCopies(MI.getOperand(4).getReg(), MRI);
+    if (Op3->getOpcode() == TargetOpcode::G_FCONSTANT)
+      return Op3->getOperand(1).getFPImm()->isExactlyValue(0.0);
+    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. Requires dx10_clamp = true. Safe to fold
+  // when Val could be QNaN. If Val can also be SNaN third input should be 0.0.
+  if (isKnownNeverNaN(MI.getOperand(0).getReg(), MRI) ||
+      (getIEEE() && getDX10Clamp() &&
+       (isKnownNeverSNaN(Val, MRI) || isOp3Zero()))) {
+    Reg = Val;
+    return true;
+  }
+
+  return false;
+}
+
+void AMDGPURegBankCombinerHelper::applyClamp(MachineInstr &MI, Register &Reg) {
+  B.setInstrAndDebugLoc(MI);
+  B.buildInstr(AMDGPU::G_AMDGPU_CLAMP, {MI.getOperand(0)}, {Reg},
+               MI.getFlags());
+  MI.eraseFromParent();
+}
+
 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::getDX10Clamp() { return getMode().DX10Clamp; }
+
 bool AMDGPURegBankCombinerHelper::isFminnumIeee(const MachineInstr &MI) {
   return MI.getOpcode() == AMDGPU::G_FMINNUM_IEEE;
 }
 
+bool AMDGPURegBankCombinerHelper::isFCst(MachineInstr *MI) {
+  return MI->getOpcode() == AMDGPU::G_FCONSTANT;
+}
+
+bool AMDGPURegBankCombinerHelper::isClampZeroToOne(MachineInstr *K0,
+                                                   MachineInstr *K1) {
+  if (isFCst(K0) && isFCst(K1)) {
+    const ConstantFP *KO_FPImm = K0->getOperand(1).getFPImm();
+    const ConstantFP *K1_FPImm = K1->getOperand(1).getFPImm();
+    return (KO_FPImm->isExactlyValue(0.0) && K1_FPImm->isExactlyValue(1.0)) ||
+           (KO_FPImm->isExactlyValue(1.0) && K1_FPImm->isExactlyValue(0.0));
+  }
+  return false;
+}
+
 class AMDGPURegBankCombinerHelperState {
 protected:
   CombinerHelper &Helper;
   AMDGPURegBankCombinerHelper &RegBankHelper;
 
 public:
   AMDGPURegBankCombinerHelperState(CombinerHelper &Helper,
                                    AMDGPURegBankCombinerHelper &RegBankHelper)
       : Helper(Helper), RegBankHelper(RegBankHelper) {}
 };
 
 #define AMDGPUREGBANKCOMBINERHELPER_GENCOMBINERHELPER_DEPS
 #include "AMDGPUGenRegBankGICombiner.inc"
 #undef AMDGPUREGBANKCOMBINERHELPER_GENCOMBINERHELPER_DEPS
 

llvm/lib/Target/AMDGPU/SIInstructions.td

 }
 
 def G_AMDGPU_FMED3 : AMDGPUGenericInstruction {
   let OutOperandList = (outs type0:$dst);
   let InOperandList = (ins type0:$src0, type0:$src1, type0:$src2);
   let hasSideEffects = 0;
 }
 
+def G_AMDGPU_CLAMP : AMDGPUGenericInstruction {
+  let OutOperandList = (outs type0:$dst);
+  let InOperandList = (ins type0:$src);
+  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/clamp-fmed3-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_fmed3_f32_known_nnan_ieee_true(float %a) #0 {
+; GFX10-LABEL: test_fmed3_f32_known_nnan_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_mul_f32_e64 v0, v0, 2.0 clamp
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %fmul = fmul float %a, 2.0
+  %fmed = call nnan float @llvm.amdgcn.fmed3.f32(float %fmul, float 0.0, float 1.0)
+  ret float %fmed
+}
+
+define half @test_fmed3_f16_known_nnan_ieee_false(half %a) #1 {
+; GFX10-LABEL: test_fmed3_f16_known_nnan_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_mul_f16_e64 v0, v0, 2.0 clamp
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %fmul = fmul half %a, 2.0
+  %fmed = call nnan half @llvm.amdgcn.fmed3.f16(half %fmul, half 0.0, half 1.0)
+  ret half %fmed
+}
+
+; %fmin is known non-SNaN because fmin inputs are fcanonicalized
+define float @test_fmed3_non_SNaN_input_ieee_true_dx10clamp_true(float %a) #2 {
+; GFX10-LABEL: test_fmed3_non_SNaN_input_ieee_true_dx10clamp_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_e64 v0, 0x41200000, v0 clamp
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %fmin = call float @llvm.minnum.f32(float %a, float 10.0)
+  %fmed = call float @llvm.amdgcn.fmed3.f32(float %fmin, float 0.0, float 1.0)
+  ret float %fmed
+}
+
+; input may be SNaN. It's safe to clamp since third operand in fmed3 is 0.0
+define float @test_fmed3_maybe_SNaN_input_zero_third_operand_ieee_true_dx10clamp_true(float %a) #2 {
+; GFX10-LABEL: test_fmed3_maybe_SNaN_input_zero_third_operand_ieee_true_dx10clamp_true:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    v_mul_f32_e64 v0, v0, 2.0 clamp
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %fmul = fmul float %a, 2.0
+  %fmed = call float @llvm.amdgcn.fmed3.f32(float %fmul, float 1.0, float 0.0)
+  ret float %fmed
+}
+
+; global nnan function attribute always forces clamp combine
+
+define float @test_fmed3_global_nnan(float %a) #3 {
+; GFX10-LABEL: test_fmed3_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_mul_f32_e64 v0, v0, 2.0 clamp
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %fmul = fmul float %a, 2.0
+  %fmed = call float @llvm.amdgcn.fmed3.f32(float %fmul, float 0.0, float 1.0)
+  ret float %fmed
+}
+
+; ------------------------------------------------------------------------------
+; Negative patterns
+; ------------------------------------------------------------------------------
+
+; ieee=false requires known never NaN input
+define float @test_fmed3_f32_maybe_NaN_ieee_false(float %a) #1 {
+; GFX10-LABEL: test_fmed3_f32_maybe_NaN_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_mul_f32_e32 v0, 2.0, v0
+; GFX10-NEXT:    v_med3_f32 v0, v0, 1.0, 0
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %fmul = fmul float %a, 2.0
+  %fmed = call float @llvm.amdgcn.fmed3.f32(float %fmul, float 1.0, float 0.0)
+  ret float %fmed
+}
+
+; ieee=true input is known non-SNaN but dx10_clamp=false
+define float @test_fmed3_non_SNaN_input_ieee_true_dx10clamp_false(float %a) #4 {
+; GFX10-LABEL: test_fmed3_non_SNaN_input_ieee_true_dx10clamp_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, v0, v0
+; GFX10-NEXT:    v_min_f32_e32 v0, 0x41200000, v0
+; GFX10-NEXT:    v_med3_f32 v0, v0, 0, 1.0
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %fmin = call float @llvm.minnum.f32(float %a, float 10.0)
+  %fmed = call float @llvm.amdgcn.fmed3.f32(float %fmin, float 0.0, float 1.0)
+  ret float %fmed
+}
+
+; ieee=true dx10_clamp=true but input may be SNaN, clamp requires third operand in fmed3 to be 0.0
+define float @test_fmed3_maybe_SNaN_input_ieee_true_dx10clamp_true(float %a) #2 {
+; GFX10-LABEL: test_fmed3_maybe_SNaN_input_ieee_true_dx10clamp_true:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    v_mul_f32_e32 v0, 2.0, v0
+; GFX10-NEXT:    v_med3_f32 v0, v0, 0, 1.0
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %fmul = fmul float %a, 2.0
+  %fmed = call float @llvm.amdgcn.fmed3.f32(float %fmul, float 0.0, float 1.0)
+  ret float %fmed
+}
+
+declare half @llvm.amdgcn.fmed3.f16(half, half, half)
+declare float @llvm.amdgcn.fmed3.f32(float, float, float)
+declare float @llvm.minnum.f32(float, float)
+
+attributes #0 = {"amdgpu-ieee"="true"}
+attributes #1 = {"amdgpu-ieee"="false"}
+attributes #2 = {"amdgpu-ieee"="true" "amdgpu-dx10-clamp"="true"}
+attributes #3 = {"no-nans-fp-math"="true"}
+attributes #4 = {"amdgpu-ieee"="true" "amdgpu-dx10-clamp"="false"}

llvm/test/CodeGen/AMDGPU/GlobalISel/clamp-minmax-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_mul_f32_e64 v0, v0, 2.0 clamp
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %fmul = fmul float %a, 2.0
+  %maxnum = call nnan float @llvm.maxnum.f32(float %fmul, float 0.0)
+  %fmed = call nnan float @llvm.minnum.f32(float %maxnum, float 1.0)
+  ret float %fmed
+}
+
+define double @test_min_max_K0Val_K1_f64(double %a) #1 {
+; GFX10-LABEL: test_min_max_K0Val_K1_f64:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    v_mul_f64 v[0:1], v[0:1], 2.0 clamp
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %fmul = fmul double %a, 2.0
+  %maxnum = call nnan double @llvm.maxnum.f64(double 0.0, double %fmul)
+  %fmed = call nnan double @llvm.minnum.f64(double %maxnum, double 1.0)
+  ret double %fmed
+}
+
+; min-max patterns for ieee=true, dx10_clamp=true don't have to check for NaNs
+define half @test_min_K1max_ValK0_f16(half %a) #2 {
+; 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_mul_f16_e64 v0, v0, 2.0 clamp
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %fmul = fmul half %a, 2.0
+  %maxnum = call half @llvm.maxnum.f16(half %fmul, half 0.0)
+  %fmed = call half @llvm.minnum.f16(half 1.0, half %maxnum)
+  ret half %fmed
+}
+
+define <2 x half> @test_min_K1max_K0Val_f16(<2 x 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_pk_mul_f16 v0, v0, 2.0 op_sel_hi:[1,0] clamp
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %fmul = fmul <2 x half> %a, <half 2.0, half 2.0>
+  %maxnum = call nnan <2 x half> @llvm.maxnum.v2f16(<2 x half> <half 0.0, half 0.0>, <2 x half> %fmul)
+  %fmed = call nnan <2 x half> @llvm.minnum.v2f16(<2 x half> <half 1.0, half 1.0>, <2 x half> %maxnum)
+  ret <2 x half> %fmed
+}
+
+define <2 x half> @test_min_max_splat_padded_with_undef(<2 x half> %a) #2 {
+; GFX10-LABEL: test_min_max_splat_padded_with_undef:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    v_pk_mul_f16 v0, v0, 2.0 op_sel_hi:[1,0] clamp
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %fmul = fmul <2 x half> %a, <half 2.0, half 2.0>
+  %maxnum = call <2 x half> @llvm.maxnum.v2f16(<2 x half> <half 0.0, half undef>, <2 x half> %fmul)
+  %fmed = call <2 x half> @llvm.minnum.v2f16(<2 x half> <half 1.0, half undef>, <2 x half> %maxnum)
+  ret <2 x half> %fmed
+}
+
+; max-mix patterns work only for known 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_mul_f32_e64 v0, v0, 2.0 clamp
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %fmul = fmul float %a, 2.0
+  %minnum = call nnan float @llvm.minnum.f32(float %fmul, float 1.0)
+  %fmed = call nnan float @llvm.maxnum.f32(float %minnum, float 0.0)
+  ret float %fmed
+}
+
+define double @test_max_min_K1Val_K0_f64(double %a) #1 {
+; GFX10-LABEL: test_max_min_K1Val_K0_f64:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    v_mul_f64 v[0:1], v[0:1], 2.0 clamp
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %fmul = fmul double %a, 2.0
+  %minnum = call nnan double @llvm.minnum.f64(double 1.0, double %fmul)
+  %fmed = call nnan double @llvm.maxnum.f64(double %minnum, double 0.0)
+  ret double %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_mul_f16_e64 v0, v0, 2.0 clamp
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %fmul = fmul half %a, 2.0
+  %minnum = call nnan half @llvm.minnum.f16(half %fmul, half 1.0)
+  %fmed = call nnan half @llvm.maxnum.f16(half 0.0, half %minnum)
+  ret half %fmed
+}
+
+; treat undef as value that will result in a constant splat
+define <2 x half> @test_max_K0min_K1Val_v2f16(<2 x half> %a) #1 {
+; GFX10-LABEL: test_max_K0min_K1Val_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_mul_f16 v0, v0, 2.0 op_sel_hi:[1,0] clamp
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %fmul = fmul <2 x half> %a, <half 2.0, half 2.0>
+  %minnum = call nnan <2 x half> @llvm.minnum.v2f16(<2 x half> <half 1.0, half undef>, <2 x half> %fmul)
+  %fmed = call nnan <2 x half> @llvm.maxnum.v2f16(<2 x half> <half undef, half 0.0>, <2 x half> %minnum)
+  ret <2 x half> %fmed
+}
+
+; global nnan function attribute always forces clamp combine
+
+define float @test_min_max_global_nnan(float %a) #3 {
+; 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_max_f32_e64 v0, v0, v0 clamp
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %maxnum = call float @llvm.maxnum.f32(float %a, float 0.0)
+  %fmed = call float @llvm.minnum.f32(float %maxnum, float 1.0)
+  ret float %fmed
+}
+
+define float @test_max_min_global_nnan(float %a) #3 {
+; 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_max_f32_e64 v0, v0, v0 clamp
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %minnum = call float @llvm.minnum.f32(float %a, float 1.0)
+  %fmed = call float @llvm.maxnum.f32(float %minnum, float 0.0)
+  ret float %fmed
+}
+
+; ------------------------------------------------------------------------------
+; Negative patterns
+; ------------------------------------------------------------------------------
+
+; min(max(Val, 1.0), 0.0), should be min(max(Val, 0.0), 1.0)
+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, 1.0, v0
+; GFX10-NEXT:    v_min_f32_e32 v0, 0, v0
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %maxnum = call nnan float @llvm.maxnum.f32(float %a, float 1.0)
+  %fmed = call nnan float @llvm.minnum.f32(float %maxnum, float 0.0)
+  ret float %fmed
+}
+
+; max(min(Val, 0.0), 1.0), should be max(min(Val, 1.0), 0.0)
+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, 0, v0
+; GFX10-NEXT:    v_max_f32_e32 v0, 1.0, v0
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %minnum = call nnan float @llvm.minnum.f32(float %a, float 0.0)
+  %fmed = call nnan float @llvm.maxnum.f32(float %minnum, float 1.0)
+  ret float %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_mul_f32_e32 v0, 2.0, v0
+; GFX10-NEXT:    v_max_f32_e32 v0, 0, v0
+; GFX10-NEXT:    v_min_f32_e32 v0, 1.0, v0
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %fmul = fmul float %a, 2.0
+  %maxnum = call float @llvm.maxnum.f32(float %fmul, float 0.0)
+  %fmed = call float @llvm.minnum.f32(float %maxnum, float 1.0)
+  ret float %fmed
+}
+
+; clamp fails here since input can be NaN and dx10_clamp=false; fmed3 succeds
+define float @test_min_max_maybe_NaN_input_ieee_true_dx10clamp_false(float %a) #4 {
+; GFX10-LABEL: test_min_max_maybe_NaN_input_ieee_true_dx10clamp_false:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    v_mul_f32_e32 v0, 2.0, v0
+; GFX10-NEXT:    v_max_f32_e32 v0, v0, v0
+; GFX10-NEXT:    v_med3_f32 v0, v0, 0, 1.0
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %fmul = fmul float %a, 2.0
+  %maxnum = call float @llvm.maxnum.f32(float %fmul, float 0.0)
+  %fmed = call float @llvm.minnum.f32(float %maxnum, float 1.0)
+  ret float %fmed
+}
+
+; max-min patterns always require known non-NaN input
+
+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_mul_f32_e32 v0, 2.0, v0
+; GFX10-NEXT:    v_max_f32_e32 v0, v0, v0
+; GFX10-NEXT:    v_min_f32_e32 v0, 1.0, v0
+; GFX10-NEXT:    v_max_f32_e32 v0, 0, v0
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %fmul = fmul float %a, 2.0
+  %minnum = call float @llvm.minnum.f32(float %fmul, float 1.0)
+  %fmed = call float @llvm.maxnum.f32(float %minnum, float 0.0)
+  ret float %fmed
+}
+
+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_mul_f32_e32 v0, 2.0, v0
+; GFX10-NEXT:    v_min_f32_e32 v0, 1.0, v0
+; GFX10-NEXT:    v_max_f32_e32 v0, 0, v0
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+  %fmul = fmul float %a, 2.0
+  %minnum = call float @llvm.minnum.f32(float %fmul, float 1.0)
+  %fmed = call float @llvm.maxnum.f32(float %minnum, float 0.0)
+  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 = {"amdgpu-ieee"="true" "amdgpu-dx10-clamp"="true"}
+attributes #3 = {"no-nans-fp-math"="true"}
+attributes #4 = {"amdgpu-ieee"="true" "amdgpu-dx10-clamp"="false"}

llvm/test/CodeGen/AMDGPU/GlobalISel/regbankcombiner-clamp-fmed3-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_fmed3_f32_known_nnan_ieee_true
+legalized: true
+regBankSelected: true
+tracksRegLiveness: true
+machineFunctionInfo:
+  mode:
+    ieee: true
+    dx10-clamp: true
+body: |
+  bb.1 :
+    liveins: $vgpr0
+
+    ; CHECK-LABEL: name: test_fmed3_f32_known_nnan_ieee_true
+    ; 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: [[FMUL:%[0-9]+]]:vgpr(s32) = G_FMUL [[COPY]], [[COPY1]]
+    ; CHECK: [[AMDGPU_CLAMP:%[0-9]+]]:vgpr(s32) = nnan G_AMDGPU_CLAMP [[FMUL]]
+    ; CHECK: $vgpr0 = COPY [[AMDGPU_CLAMP]](s32)
+    %0:vgpr(s32) = COPY $vgpr0
+    %2:sgpr(s32) = G_FCONSTANT float 2.000000e+00
+    %8:vgpr(s32) = COPY %2(s32)
+    %3:vgpr(s32) = G_FMUL %0, %8
+    %6:sgpr(s32) = G_FCONSTANT float 1.000000e+00
+    %5:sgpr(s32) = G_FCONSTANT float 0.000000e+00
+    %9:vgpr(s32) = COPY %5(s32)
+    %10:vgpr(s32) = COPY %6(s32)
+    %4:vgpr(s32) = nnan G_INTRINSIC intrinsic(@llvm.amdgcn.fmed3), %3(s32), %9(s32), %10(s32)
+    $vgpr0 = COPY %4(s32)
+...
+
+---
+name: test_fmed3_f16_known_nnan_ieee_false
+legalized: true
+regBankSelected: true
+tracksRegLiveness: true
+machineFunctionInfo:
+  mode:
+    ieee: false
+    dx10-clamp: true
+body: |
+  bb.1 :
+    liveins: $vgpr0
+
+    ; CHECK-LABEL: name: test_fmed3_f16_known_nnan_ieee_false
+    ; 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: [[COPY1:%[0-9]+]]:vgpr(s16) = COPY [[C]](s16)
+    ; CHECK: [[FMUL:%[0-9]+]]:vgpr(s16) = G_FMUL [[TRUNC]], [[COPY1]]
+    ; CHECK: [[AMDGPU_CLAMP:%[0-9]+]]:vgpr(s16) = nnan G_AMDGPU_CLAMP [[FMUL]]
+    ; CHECK: [[ANYEXT:%[0-9]+]]:vgpr(s32) = G_ANYEXT [[AMDGPU_CLAMP]](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
+    %10:vgpr(s16) = COPY %3(s16)
+    %4:vgpr(s16) = G_FMUL %0, %10
+    %7:sgpr(s16) = G_FCONSTANT half 0xH3C00
+    %6:sgpr(s16) = G_FCONSTANT half 0xH0000
+    %11:vgpr(s16) = COPY %6(s16)
+    %12:vgpr(s16) = COPY %7(s16)
+    %5:vgpr(s16) = nnan G_INTRINSIC intrinsic(@llvm.amdgcn.fmed3), %4(s16), %11(s16), %12(s16)
+    %9:vgpr(s32) = G_ANYEXT %5(s16)
+    $vgpr0 = COPY %9(s32)
+...
+
+---
+name: test_fmed3_non_SNaN_input_ieee_true_dx10clamp_true
+legalized: true
+regBankSelected: true
+tracksRegLiveness: true
+machineFunctionInfo:
+  mode:
+    ieee:            true
+    dx10-clamp:      true
+body: |
+  bb.1 :
+    liveins: $vgpr0
+
+    ; CHECK-LABEL: name: test_fmed3_non_SNaN_input_ieee_true_dx10clamp_true
+    ; CHECK: liveins: $vgpr0
+    ; CHECK: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
+    ; CHECK: [[C:%[0-9]+]]:sgpr(s32) = G_FCONSTANT float 1.000000e+01
+    ; 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: [[AMDGPU_CLAMP:%[0-9]+]]:vgpr(s32) = G_AMDGPU_CLAMP [[FMINNUM_IEEE]]
+    ; CHECK: $vgpr0 = COPY [[AMDGPU_CLAMP]](s32)
+    %0:vgpr(s32) = COPY $vgpr0
+    %2:sgpr(s32) = G_FCONSTANT float 1.000000e+01
+    %8:vgpr(s32) = G_FCANONICALIZE %0
+    %9:vgpr(s32) = COPY %2(s32)
+    %3:vgpr(s32) = G_FMINNUM_IEEE %8, %9
+    %6:sgpr(s32) = G_FCONSTANT float 1.000000e+00
+    %5:sgpr(s32) = G_FCONSTANT float 0.000000e+00
+    %10:vgpr(s32) = COPY %5(s32)
+    %11:vgpr(s32) = COPY %6(s32)
+    %4:vgpr(s32) = G_INTRINSIC intrinsic(@llvm.amdgcn.fmed3), %3(s32), %10(s32), %11(s32)
+    $vgpr0 = COPY %4(s32)
+...
+
+---
+name: test_fmed3_maybe_SNaN_input_zero_third_operand_ieee_true_dx10clamp_true
+legalized: true
+regBankSelected: true
+tracksRegLiveness: true
+machineFunctionInfo:
+  mode:
+    ieee: true
+    dx10-clamp: true
+body: |
+  bb.1 :
+    liveins: $vgpr0
+
+    ; CHECK-LABEL: name: test_fmed3_maybe_SNaN_input_zero_third_operand_ieee_true_dx10clamp_true
+    ; 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: [[FMUL:%[0-9]+]]:vgpr(s32) = G_FMUL [[COPY]], [[COPY1]]
+    ; CHECK: [[AMDGPU_CLAMP:%[0-9]+]]:vgpr(s32) = G_AMDGPU_CLAMP [[FMUL]]
+    ; CHECK: $vgpr0 = COPY [[AMDGPU_CLAMP]](s32)
+    %0:vgpr(s32) = COPY $vgpr0
+    %2:sgpr(s32) = G_FCONSTANT float 2.000000e+00
+    %8:vgpr(s32) = COPY %2(s32)
+    %3:vgpr(s32) = G_FMUL %0, %8
+    %6:sgpr(s32) = G_FCONSTANT float 0.000000e+00
+    %5:sgpr(s32) = G_FCONSTANT float 1.000000e+00
+    %9:vgpr(s32) = COPY %5(s32)
+    %10:vgpr(s32) = COPY %6(s32)
+    %4:vgpr(s32) = G_INTRINSIC intrinsic(@llvm.amdgcn.fmed3), %3(s32), %9(s32), %10(s32)
+    $vgpr0 = COPY %4(s32)
+...
+
+# FixMe: add tests with attributes #3 = {"no-nans-fp-math"="true"}
+
+---
+name: test_fmed3_f32_maybe_NaN_ieee_false
+legalized: true
+regBankSelected: true
+tracksRegLiveness: true
+machineFunctionInfo:
+  mode:
+    ieee: false
+    dx10-clamp: true
+body: |
+  bb.1 :
+    liveins: $vgpr0
+
+    ; CHECK-LABEL: name: test_fmed3_f32_maybe_NaN_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: [[FMUL:%[0-9]+]]:vgpr(s32) = G_FMUL [[COPY]], [[COPY1]]
+    ; CHECK: [[C1:%[0-9]+]]:sgpr(s32) = G_FCONSTANT float 0.000000e+00
+    ; CHECK: [[C2:%[0-9]+]]:sgpr(s32) = G_FCONSTANT float 1.000000e+00
+    ; CHECK: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY [[C2]](s32)
+    ; CHECK: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY [[C1]](s32)
+    ; CHECK: [[INT:%[0-9]+]]:vgpr(s32) = G_INTRINSIC intrinsic(@llvm.amdgcn.fmed3), [[FMUL]](s32), [[COPY2]](s32), [[COPY3]](s32)
+    ; CHECK: $vgpr0 = COPY [[INT]](s32)
+    %0:vgpr(s32) = COPY $vgpr0
+    %2:sgpr(s32) = G_FCONSTANT float 2.000000e+00
+    %8:vgpr(s32) = COPY %2(s32)
+    %3:vgpr(s32) = G_FMUL %0, %8
+    %6:sgpr(s32) = G_FCONSTANT float 0.000000e+00
+    %5:sgpr(s32) = G_FCONSTANT float 1.000000e+00
+    %9:vgpr(s32) = COPY %5(s32)
+    %10:vgpr(s32) = COPY %6(s32)
+    %4:vgpr(s32) = G_INTRINSIC intrinsic(@llvm.amdgcn.fmed3), %3(s32), %9(s32), %10(s32)
+    $vgpr0 = COPY %4(s32)
+...
+
+---
+name: test_fmed3_non_SNaN_input_ieee_true_dx10clamp_false
+legalized: true
+regBankSelected: true
+tracksRegLiveness: true
+machineFunctionInfo:
+  mode:
+    ieee: true
+    dx10-clamp: false
+body: |
+  bb.1 :
+    liveins: $vgpr0
+
+    ; CHECK-LABEL: name: test_fmed3_non_SNaN_input_ieee_true_dx10clamp_false
+    ; CHECK: liveins: $vgpr0
+    ; CHECK: [[COPY:%[0-9]+]]:vgpr(s32) = COPY $vgpr0
+    ; CHECK: [[C:%[0-9]+]]:sgpr(s32) = G_FCONSTANT float 1.000000e+01
+    ; 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 1.000000e+00
+    ; CHECK: [[C2:%[0-9]+]]:sgpr(s32) = G_FCONSTANT float 0.000000e+00
+    ; CHECK: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY [[C2]](s32)
+    ; CHECK: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY [[C1]](s32)
+    ; CHECK: [[INT:%[0-9]+]]:vgpr(s32) = G_INTRINSIC intrinsic(@llvm.amdgcn.fmed3), [[FMINNUM_IEEE]](s32), [[COPY2]](s32), [[COPY3]](s32)
+    ; CHECK: $vgpr0 = COPY [[INT]](s32)
+    %0:vgpr(s32) = COPY $vgpr0
+    %2:sgpr(s32) = G_FCONSTANT float 1.000000e+01
+    %8:vgpr(s32) = G_FCANONICALIZE %0
+    %9:vgpr(s32) = COPY %2(s32)
+    %3:vgpr(s32) = G_FMINNUM_IEEE %8, %9
+    %6:sgpr(s32) = G_FCONSTANT float 1.000000e+00
+    %5:sgpr(s32) = G_FCONSTANT float 0.000000e+00
+    %10:vgpr(s32) = COPY %5(s32)
+    %11:vgpr(s32) = COPY %6(s32)
+    %4:vgpr(s32) = G_INTRINSIC intrinsic(@llvm.amdgcn.fmed3), %3(s32), %10(s32), %11(s32)
+    $vgpr0 = COPY %4(s32)
+...
+
+---
+name: test_fmed3_maybe_SNaN_input_ieee_true_dx10clamp_true
+legalized: true
+regBankSelected: true
+tracksRegLiveness: true
+machineFunctionInfo:
+  mode:
+    ieee: true
+    dx10-clamp: true
+body: |
+  bb.1 :
+    liveins: $vgpr0
+
+    ; CHECK-LABEL: name: test_fmed3_maybe_SNaN_input_ieee_true_dx10clamp_true
+    ; 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: [[FMUL:%[0-9]+]]:vgpr(s32) = G_FMUL [[COPY]], [[COPY1]]
+    ; CHECK: [[C1:%[0-9]+]]:sgpr(s32) = G_FCONSTANT float 1.000000e+00
+    ; CHECK: [[C2:%[0-9]+]]:sgpr(s32) = G_FCONSTANT float 0.000000e+00
+    ; CHECK: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY [[C2]](s32)
+    ; CHECK: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY [[C1]](s32)
+    ; CHECK: [[INT:%[0-9]+]]:vgpr(s32) = G_INTRINSIC intrinsic(@llvm.amdgcn.fmed3), [[FMUL]](s32), [[COPY2]](s32), [[COPY3]](s32)
+    ; CHECK: $vgpr0 = COPY [[INT]](s32)
+    %0:vgpr(s32) = COPY $vgpr0
+    %2:sgpr(s32) = G_FCONSTANT float 2.000000e+00
+    %8:vgpr(s32) = COPY %2(s32)
+    %3:vgpr(s32) = G_FMUL %0, %8
+    %6:sgpr(s32) = G_FCONSTANT float 1.000000e+00
+    %5:sgpr(s32) = G_FCONSTANT float 0.000000e+00
+    %9:vgpr(s32) = COPY %5(s32)
+    %10:vgpr(s32) = COPY %6(s32)
+    %4:vgpr(s32) = G_INTRINSIC intrinsic(@llvm.amdgcn.fmed3), %3(s32), %9(s32), %10(s32)
+    $vgpr0 = COPY %4(s32)
+...

llvm/test/CodeGen/AMDGPU/GlobalISel/regbankcombiner-clamp-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
+    dx10-clamp: 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: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY [[C]](s32)
+    ; CHECK: [[FMUL:%[0-9]+]]:vgpr(s32) = G_FMUL [[COPY]], [[COPY1]]
+    ; CHECK: [[AMDGPU_CLAMP:%[0-9]+]]:vgpr(s32) = nnan G_AMDGPU_CLAMP [[FMUL]]
+    ; CHECK: $vgpr0 = COPY [[AMDGPU_CLAMP]](s32)
+    %0:vgpr(s32) = COPY $vgpr0
+    %2:sgpr(s32) = G_FCONSTANT float 2.000000e+00
+    %9:vgpr(s32) = COPY %2(s32)
+    %3:vgpr(s32) = G_FMUL %0, %9
+    %4:sgpr(s32) = G_FCONSTANT float 0.000000e+00
+    %10:vgpr(s32) = COPY %4(s32)
+    %5:vgpr(s32) = nnan G_FMAXNUM_IEEE %3, %10
+    %6:sgpr(s32) = G_FCONSTANT float 1.000000e+00
+    %11:vgpr(s32) = COPY %6(s32)
+    %7:vgpr(s32) = nnan G_FMINNUM_IEEE %5, %11
+    $vgpr0 = COPY %7(s32)
+...
+
+---
+name: test_min_max_K0Val_K1_f64
+legalized: true
+regBankSelected: true
+tracksRegLiveness: true
+machineFunctionInfo:
+  mode:
+    ieee: false
+    dx10-clamp: true
+body: |
+  bb.1 :
+    liveins: $vgpr0_vgpr1
+
+    ; CHECK-LABEL: name: test_min_max_K0Val_K1_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: [[FMUL:%[0-9]+]]:vgpr(s64) = G_FMUL [[COPY]], [[COPY1]]
+    ; CHECK: [[AMDGPU_CLAMP:%[0-9]+]]:vgpr(s64) = nnan G_AMDGPU_CLAMP [[FMUL]]
+    ; CHECK: $vgpr0_vgpr1 = COPY [[AMDGPU_CLAMP]](s64)
+    %0:vgpr(s64) = COPY $vgpr0_vgpr1
+    %4:sgpr(s64) = G_FCONSTANT double 2.000000e+00
+    %13:vgpr(s64) = COPY %4(s64)
+    %5:vgpr(s64) = G_FMUL %0, %13
+    %6:sgpr(s64) = G_FCONSTANT double 0.000000e+00
+    %14:vgpr(s64) = COPY %6(s64)
+    %7:vgpr(s64) = nnan G_FMAXNUM %14, %5
+    %8:sgpr(s64) = G_FCONSTANT double 1.000000e+00
+    %15:vgpr(s64) = COPY %8(s64)
+    %9:vgpr(s64) = nnan G_FMINNUM %7, %15
+    $vgpr0_vgpr1 = COPY %9(s64)
+...
+
+---
+name: test_min_K1max_ValK0_f16
+legalized: true
+regBankSelected: true
+tracksRegLiveness: true
+machineFunctionInfo:
+  mode:
+    ieee: true
+    dx10-clamp: 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: [[COPY1:%[0-9]+]]:vgpr(s16) = COPY [[C]](s16)
+    ; CHECK: [[FMUL:%[0-9]+]]:vgpr(s16) = G_FMUL [[TRUNC]], [[COPY1]]
+    ; CHECK: [[FCANONICALIZE:%[0-9]+]]:vgpr(s16) = G_FCANONICALIZE [[FMUL]]
+    ; CHECK: [[AMDGPU_CLAMP:%[0-9]+]]:vgpr(s16) = G_AMDGPU_CLAMP [[FCANONICALIZE]]
+    ; CHECK: [[ANYEXT:%[0-9]+]]:vgpr(s32) = G_ANYEXT [[AMDGPU_CLAMP]](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
+    %12:vgpr(s16) = COPY %3(s16)
+    %4:vgpr(s16) = G_FMUL %0, %12
+    %5:sgpr(s16) = G_FCONSTANT half 0xH0000
+    %11:vgpr(s16) = G_FCANONICALIZE %4
+    %13:vgpr(s16) = COPY %5(s16)
+    %6:vgpr(s16) = G_FMAXNUM_IEEE %11, %13
+    %7:sgpr(s16) = G_FCONSTANT half 0xH3C00
+    %14:vgpr(s16) = COPY %7(s16)
+    %8:vgpr(s16) = G_FMINNUM_IEEE %14, %6
+    %10:vgpr(s32) = G_ANYEXT %8(s16)
+    $vgpr0 = COPY %10(s32)
+...
+
+---
+name: test_min_K1max_K0Val_f16
+legalized: true
+regBankSelected: true
+tracksRegLiveness: true
+machineFunctionInfo:
+  mode:
+    ieee: false
+    dx10-clamp: true
+body: |
+  bb.1 :
+    liveins: $vgpr0
+
+    ; CHECK-LABEL: name: test_min_K1max_K0Val_f16
+    ; 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: [[COPY1:%[0-9]+]]:vgpr(<2 x s16>) = COPY [[BUILD_VECTOR_TRUNC]](<2 x s16>)
+    ; CHECK: [[FMUL:%[0-9]+]]:vgpr(<2 x s16>) = G_FMUL [[COPY]], [[COPY1]]
+    ; CHECK: [[AMDGPU_CLAMP:%[0-9]+]]:vgpr(<2 x s16>) = nnan G_AMDGPU_CLAMP [[FMUL]]
+    ; CHECK: $vgpr0 = COPY [[AMDGPU_CLAMP]](<2 x s16>)
+    %0:vgpr(<2 x s16>) = COPY $vgpr0
+    %3:sgpr(s16) = G_FCONSTANT half 0xH4000
+    %12:sgpr(s32) = G_ANYEXT %3(s16)
+    %2:sgpr(<2 x s16>) = G_BUILD_VECTOR_TRUNC %12(s32), %12(s32)
+    %6:sgpr(s16) = G_FCONSTANT half 0xH0000
+    %13:sgpr(s32) = G_ANYEXT %6(s16)
+    %5:sgpr(<2 x s16>) = G_BUILD_VECTOR_TRUNC %13(s32), %13(s32)
+    %9:sgpr(s16) = G_FCONSTANT half 0xH3C00
+    %14:sgpr(s32) = G_ANYEXT %9(s16)
+    %8:sgpr(<2 x s16>) = G_BUILD_VECTOR_TRUNC %14(s32), %14(s32)
+    %15:vgpr(<2 x s16>) = COPY %2(<2 x s16>)
+    %4:vgpr(<2 x s16>) = G_FMUL %0, %15
+    %16:vgpr(<2 x s16>) = COPY %5(<2 x s16>)
+    %7:vgpr(<2 x s16>) = nnan G_FMAXNUM %16, %4
+    %17:vgpr(<2 x s16>) = COPY %8(<2 x s16>)
+    %10:vgpr(<2 x s16>) = nnan G_FMINNUM %17, %7
+    $vgpr0 = COPY %10(<2 x s16>)
+...
+
+---
+name: test_min_max_splat_padded_with_undef
+legalized: true
+regBankSelected: true
+tracksRegLiveness: true
+machineFunctionInfo:
+  mode:
+    ieee: true
+    dx10-clamp: true
+body: |
+  bb.1 :
+    liveins: $vgpr0
+
+    ; CHECK-LABEL: name: test_min_max_splat_padded_with_undef
+    ; 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: [[COPY1:%[0-9]+]]:vgpr(<2 x s16>) = COPY [[BUILD_VECTOR_TRUNC]](<2 x s16>)
+    ; CHECK: [[FMUL:%[0-9]+]]:vgpr(<2 x s16>) = G_FMUL [[COPY]], [[COPY1]]
+    ; CHECK: [[FCANONICALIZE:%[0-9]+]]:vgpr(<2 x s16>) = G_FCANONICALIZE [[FMUL]]
+    ; CHECK: [[AMDGPU_CLAMP:%[0-9]+]]:vgpr(<2 x s16>) = G_AMDGPU_CLAMP [[FCANONICALIZE]]
+    ; CHECK: $vgpr0 = COPY [[AMDGPU_CLAMP]](<2 x s16>)
+    %0:vgpr(<2 x s16>) = COPY $vgpr0
+    %3:sgpr(s16) = G_FCONSTANT half 0xH4000
+    %17:sgpr(s32) = G_ANYEXT %3(s16)
+    %2:sgpr(<2 x s16>) = G_BUILD_VECTOR_TRUNC %17(s32), %17(s32)
+    %6:sgpr(s16) = G_FCONSTANT half 0xH0000
+    %18:sgpr(s32) = G_ANYEXT %6(s16)
+    %19:sgpr(s32) = G_IMPLICIT_DEF
+    %5:sgpr(<2 x s16>) = G_BUILD_VECTOR_TRUNC %18(s32), %19(s32)
+    %10:sgpr(s16) = G_FCONSTANT half 0xH3C00
+    %20:sgpr(s32) = G_ANYEXT %10(s16)
+    %9:sgpr(<2 x s16>) = G_BUILD_VECTOR_TRUNC %20(s32), %19(s32)
+    %21:vgpr(<2 x s16>) = COPY %2(<2 x s16>)
+    %4:vgpr(<2 x s16>) = G_FMUL %0, %21
+    %16:vgpr(<2 x s16>) = G_FCANONICALIZE %4
+    %22:vgpr(<2 x s16>) = COPY %5(<2 x s16>)
+    %8:vgpr(<2 x s16>) = G_FMAXNUM_IEEE %22, %16
+    %23:vgpr(<2 x s16>) = COPY %9(<2 x s16>)
+    %11:vgpr(<2 x s16>) = G_FMINNUM_IEEE %23, %8
+    $vgpr0 = COPY %11(<2 x s16>)
+...
+
+---
+name: test_max_min_ValK1_K0_f32
+machineFunctionInfo:
+  mode:
+    ieee: true
+    dx10-clamp: true
+body: |
+  bb.1 :
+    liveins: $vgpr0
+
+    ; CHECK-LABEL: name: test_max_min_ValK1_K0_f32
+    ; 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: [[FMUL:%[0-9]+]]:vgpr(s32) = G_FMUL [[COPY]], [[COPY1]]
+    ; CHECK: [[AMDGPU_CLAMP:%[0-9]+]]:vgpr(s32) = nnan G_AMDGPU_CLAMP [[FMUL]]
+    ; CHECK: $vgpr0 = COPY [[AMDGPU_CLAMP]](s32)
+    %0:vgpr(s32) = COPY $vgpr0
+    %2:sgpr(s32) = G_FCONSTANT float 2.000000e+00
+    %9:vgpr(s32) = COPY %2(s32)
+    %3:vgpr(s32) = G_FMUL %0, %9
+    %4:sgpr(s32) = G_FCONSTANT float 1.000000e+00
+    %10:vgpr(s32) = COPY %4(s32)
+    %5:vgpr(s32) = nnan G_FMINNUM_IEEE %3, %10
+    %6:sgpr(s32) = G_FCONSTANT float 0.000000e+00
+    %11:vgpr(s32) = COPY %6(s32)
+    %7:vgpr(s32) = nnan G_FMAXNUM_IEEE %5, %11
+    $vgpr0 = COPY %7(s32)
+...
+
+---
+name: test_max_min_K1Val_K0_f64
+legalized: true
+regBankSelected: true
+tracksRegLiveness: true
+machineFunctionInfo:
+  mode:
+    ieee: false
+    dx10-clamp: true
+body: |
+  bb.1 :
+    liveins: $vgpr0_vgpr1
+
+    ; CHECK-LABEL: name: test_max_min_K1Val_K0_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: [[FMUL:%[0-9]+]]:vgpr(s64) = G_FMUL [[COPY]], [[COPY1]]
+    ; CHECK: [[AMDGPU_CLAMP:%[0-9]+]]:vgpr(s64) = nnan G_AMDGPU_CLAMP [[FMUL]]
+    ; CHECK: $vgpr0_vgpr1 = COPY [[AMDGPU_CLAMP]](s64)
+    %0:vgpr(s64) = COPY $vgpr0_vgpr1
+    %4:sgpr(s64) = G_FCONSTANT double 2.000000e+00
+    %13:vgpr(s64) = COPY %4(s64)
+    %5:vgpr(s64) = G_FMUL %0, %13
+    %6:sgpr(s64) = G_FCONSTANT double 1.000000e+00
+    %14:vgpr(s64) = COPY %6(s64)
+    %7:vgpr(s64) = nnan G_FMINNUM %14, %5
+    %8:sgpr(s64) = G_FCONSTANT double 0.000000e+00
+    %15:vgpr(s64) = COPY %8(s64)
+    %9:vgpr(s64) = nnan G_FMAXNUM %7, %15
+    $vgpr0_vgpr1 = COPY %9(s64)
+...
+
+---
+name: test_max_K0min_ValK1_f16
+legalized: true
+regBankSelected: true
+tracksRegLiveness: true
+machineFunctionInfo:
+  mode:
+    ieee: true
+    dx10-clamp: 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 0xH4000
+    ; CHECK: [[COPY1:%[0-9]+]]:vgpr(s16) = COPY [[C]](s16)
+    ; CHECK: [[FMUL:%[0-9]+]]:vgpr(s16) = G_FMUL [[TRUNC]], [[COPY1]]
+    ; CHECK: [[AMDGPU_CLAMP:%[0-9]+]]:vgpr(s16) = nnan G_AMDGPU_CLAMP [[FMUL]]
+    ; CHECK: [[ANYEXT:%[0-9]+]]:vgpr(s32) = G_ANYEXT [[AMDGPU_CLAMP]](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
+    %11:vgpr(s16) = COPY %3(s16)
+    %4:vgpr(s16) = G_FMUL %0, %11
+    %5:sgpr(s16) = G_FCONSTANT half 0xH3C00
+    %12:vgpr(s16) = COPY %5(s16)
+    %6:vgpr(s16) = nnan G_FMINNUM_IEEE %4, %12
+    %7:sgpr(s16) = G_FCONSTANT half 0xH0000
+    %13:vgpr(s16) = COPY %7(s16)
+    %8:vgpr(s16) = nnan G_FMAXNUM_IEEE %13, %6
+    %10:vgpr(s32) = G_ANYEXT %8(s16)
+    $vgpr0 = COPY %10(s32)
+...
+
+---
+name: test_max_K0min_K1Val_v2f16
+legalized: true
+regBankSelected: true
+tracksRegLiveness: true
+machineFunctionInfo:
+  mode:
+    ieee: false
+    dx10-clamp: true
+body: |
+  bb.1 :
+    liveins: $vgpr0
+
+    ; CHECK-LABEL: name: test_max_K0min_K1Val_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: [[COPY1:%[0-9]+]]:vgpr(<2 x s16>) = COPY [[BUILD_VECTOR_TRUNC]](<2 x s16>)
+    ; CHECK: [[FMUL:%[0-9]+]]:vgpr(<2 x s16>) = G_FMUL [[COPY]], [[COPY1]]
+    ; CHECK: [[AMDGPU_CLAMP:%[0-9]+]]:vgpr(<2 x s16>) = nnan G_AMDGPU_CLAMP [[FMUL]]
+    ; CHECK: $vgpr0 = COPY [[AMDGPU_CLAMP]](<2 x s16>)
+    %0:vgpr(<2 x s16>) = COPY $vgpr0
+    %3:sgpr(s16) = G_FCONSTANT half 0xH4000
+    %13:sgpr(s32) = G_ANYEXT %3(s16)
+    %2:sgpr(<2 x s16>) = G_BUILD_VECTOR_TRUNC %13(s32), %13(s32)
+    %6:sgpr(s16) = G_FCONSTANT half 0xH3C00
+    %14:sgpr(s32) = G_ANYEXT %6(s16)
+    %15:sgpr(s32) = G_IMPLICIT_DEF
+    %5:sgpr(<2 x s16>) = G_BUILD_VECTOR_TRUNC %14(s32), %15(s32)
+    %10:sgpr(s16) = G_FCONSTANT half 0xH0000
+    %16:sgpr(s32) = G_ANYEXT %10(s16)
+    %9:sgpr(<2 x s16>) = G_BUILD_VECTOR_TRUNC %15(s32), %16(s32)
+    %17:vgpr(<2 x s16>) = COPY %2(<2 x s16>)
+    %4:vgpr(<2 x s16>) = G_FMUL %0, %17
+    %18:vgpr(<2 x s16>) = COPY %5(<2 x s16>)
+    %8:vgpr(<2 x s16>) = nnan G_FMINNUM %18, %4
+    %19:vgpr(<2 x s16>) = COPY %9(<2 x s16>)
+    %11:vgpr(<2 x s16>) = nnan G_FMAXNUM %19, %8
+    $vgpr0 = COPY %11(<2 x s16>)
+...
+
+# FixMe: add tests with attributes #3 = {"no-nans-fp-math"="true"}
+
+---
+name: test_min_max_K0_gt_K1
+legalized: true
+regBankSelected: true
+tracksRegLiveness: true
+machineFunctionInfo:
+  mode:
+    ieee: true
+    dx10-clamp: 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 1.000000e+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 0.000000e+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 1.000000e+00
+    %7:vgpr(s32) = COPY %2(s32)
+    %3:vgpr(s32) = nnan G_FMAXNUM_IEEE %0, %7
+    %4:sgpr(s32) = G_FCONSTANT float 0.000000e+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
+    dx10-clamp: 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 0.000000e+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 1.000000e+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 0.000000e+00
+    %7:vgpr(s32) = COPY %2(s32)
+    %3:vgpr(s32) = nnan G_FMINNUM_IEEE %0, %7
+    %4:sgpr(s32) = G_FCONSTANT float 1.000000e+00
+    %8:vgpr(s32) = COPY %4(s32)
+    %5:vgpr(s32) = nnan G_FMAXNUM_IEEE %3, %8
+    $vgpr0 = COPY %5(s32)
+...
+
+---
+name: test_min_max_maybe_NaN_input_ieee_false
+legalized: true
+regBankSelected: true
+tracksRegLiveness: true
+machineFunctionInfo:
+  mode:
+    ieee: false
+    dx10-clamp: true
+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: [[FMUL:%[0-9]+]]:vgpr(s32) = G_FMUL [[COPY]], [[COPY1]]
+    ; CHECK: [[C1:%[0-9]+]]:sgpr(s32) = G_FCONSTANT float 0.000000e+00
+    ; CHECK: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY [[C1]](s32)
+    ; CHECK: [[FMAXNUM:%[0-9]+]]:vgpr(s32) = G_FMAXNUM [[FMUL]], [[COPY2]]
+    ; CHECK: [[C2:%[0-9]+]]:sgpr(s32) = G_FCONSTANT float 1.000000e+00
+    ; CHECK: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY [[C2]](s32)
+    ; CHECK: [[FMINNUM:%[0-9]+]]:vgpr(s32) = G_FMINNUM [[FMAXNUM]], [[COPY3]]
+    ; CHECK: $vgpr0 = COPY [[FMINNUM]](s32)
+    %0:vgpr(s32) = COPY $vgpr0
+    %2:sgpr(s32) = G_FCONSTANT float 2.000000e+00
+    %9:vgpr(s32) = COPY %2(s32)
+    %3:vgpr(s32) = G_FMUL %0, %9
+    %4:sgpr(s32) = G_FCONSTANT float 0.000000e+00
+    %10:vgpr(s32) = COPY %4(s32)
+    %5:vgpr(s32) = G_FMAXNUM %3, %10
+    %6:sgpr(s32) = G_FCONSTANT float 1.000000e+00
+    %11:vgpr(s32) = COPY %6(s32)
+    %7:vgpr(s32) = G_FMINNUM %5, %11
+    $vgpr0 = COPY %7(s32)
+...
+
+---
+name: test_min_max_maybe_NaN_input_ieee_true_dx10clamp_false
+legalized: true
+regBankSelected: true
+tracksRegLiveness: true
+machineFunctionInfo:
+  mode:
+    ieee: true
+    dx10-clamp: false
+body: |
+  bb.1 :
+    liveins: $vgpr0
+
+    ; CHECK-LABEL: name: test_min_max_maybe_NaN_input_ieee_true_dx10clamp_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: [[FMUL:%[0-9]+]]:vgpr(s32) = G_FMUL [[COPY]], [[COPY1]]
+    ; CHECK: [[C1:%[0-9]+]]:sgpr(s32) = G_FCONSTANT float 0.000000e+00
+    ; CHECK: [[FCANONICALIZE:%[0-9]+]]:vgpr(s32) = G_FCANONICALIZE [[FMUL]]
+    ; CHECK: [[C2:%[0-9]+]]:sgpr(s32) = G_FCONSTANT float 1.000000e+00
+    ; CHECK: [[AMDGPU_FMED3_:%[0-9]+]]:vgpr(s32) = G_AMDGPU_FMED3 [[FCANONICALIZE]], [[C1]], [[C2]]
+    ; CHECK: $vgpr0 = COPY [[AMDGPU_FMED3_]](s32)
+    %0:vgpr(s32) = COPY $vgpr0
+    %2:sgpr(s32) = G_FCONSTANT float 2.000000e+00
+    %10:vgpr(s32) = COPY %2(s32)
+    %3:vgpr(s32) = G_FMUL %0, %10
+    %4:sgpr(s32) = G_FCONSTANT float 0.000000e+00
+    %9:vgpr(s32) = G_FCANONICALIZE %3
+    %11:vgpr(s32) = COPY %4(s32)
+    %5:vgpr(s32) = G_FMAXNUM_IEEE %9, %11
+    %6:sgpr(s32) = G_FCONSTANT float 1.000000e+00
+    %12:vgpr(s32) = COPY %6(s32)
+    %7:vgpr(s32) = G_FMINNUM_IEEE %5, %12
+    $vgpr0 = COPY %7(s32)
+...
+
+---
+name: test_max_min_maybe_NaN_input_ieee_true
+legalized: true
+regBankSelected: true
+tracksRegLiveness: true
+machineFunctionInfo:
+  mode:
+    ieee: true
+    dx10-clamp: 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 2.000000e+00
+    ; CHECK: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY [[C]](s32)
+    ; CHECK: [[FMUL:%[0-9]+]]:vgpr(s32) = G_FMUL [[COPY]], [[COPY1]]
+    ; CHECK: [[C1:%[0-9]+]]:sgpr(s32) = G_FCONSTANT float 1.000000e+00
+    ; CHECK: [[FCANONICALIZE:%[0-9]+]]:vgpr(s32) = G_FCANONICALIZE [[FMUL]]
+    ; CHECK: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY [[C1]](s32)
+    ; CHECK: [[FMINNUM_IEEE:%[0-9]+]]:vgpr(s32) = G_FMINNUM_IEEE [[FCANONICALIZE]], [[COPY2]]
+    ; CHECK: [[C2:%[0-9]+]]:sgpr(s32) = G_FCONSTANT float 0.000000e+00
+    ; CHECK: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY [[C2]](s32)
+    ; CHECK: [[FMAXNUM_IEEE:%[0-9]+]]:vgpr(s32) = G_FMAXNUM_IEEE [[FMINNUM_IEEE]], [[COPY3]]
+    ; CHECK: $vgpr0 = COPY [[FMAXNUM_IEEE]](s32)
+    %0:vgpr(s32) = COPY $vgpr0
+    %2:sgpr(s32) = G_FCONSTANT float 2.000000e+00
+    %10:vgpr(s32) = COPY %2(s32)
+    %3:vgpr(s32) = G_FMUL %0, %10
+    %4:sgpr(s32) = G_FCONSTANT float 1.000000e+00
+    %9:vgpr(s32) = G_FCANONICALIZE %3
+    %11:vgpr(s32) = COPY %4(s32)
+    %5:vgpr(s32) = G_FMINNUM_IEEE %9, %11
+    %6:sgpr(s32) = G_FCONSTANT float 0.000000e+00
+    %12:vgpr(s32) = COPY %6(s32)
+    %7:vgpr(s32) = G_FMAXNUM_IEEE %5, %12
+    $vgpr0 = COPY %7(s32)
+...
+
+---
+name: test_max_min_maybe_NaN_input_ieee_false
+legalized: true
+regBankSelected: true
+tracksRegLiveness: true
+machineFunctionInfo:
+  mode:
+    ieee: false
+    dx10-clamp: true
+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 2.000000e+00
+    ; CHECK: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY [[C]](s32)
+    ; CHECK: [[FMUL:%[0-9]+]]:vgpr(s32) = G_FMUL [[COPY]], [[COPY1]]
+    ; CHECK: [[C1:%[0-9]+]]:sgpr(s32) = G_FCONSTANT float 1.000000e+00
+    ; CHECK: [[COPY2:%[0-9]+]]:vgpr(s32) = COPY [[C1]](s32)
+    ; CHECK: [[FMINNUM:%[0-9]+]]:vgpr(s32) = G_FMINNUM [[FMUL]], [[COPY2]]
+    ; CHECK: [[C2:%[0-9]+]]:sgpr(s32) = G_FCONSTANT float 0.000000e+00
+    ; CHECK: [[COPY3:%[0-9]+]]:vgpr(s32) = COPY [[C2]](s32)
+    ; CHECK: [[FMAXNUM:%[0-9]+]]:vgpr(s32) = G_FMAXNUM [[FMINNUM]], [[COPY3]]
+    ; CHECK: $vgpr0 = COPY [[FMAXNUM]](s32)
+    %0:vgpr(s32) = COPY $vgpr0
+    %2:sgpr(s32) = G_FCONSTANT float 2.000000e+00
+    %9:vgpr(s32) = COPY %2(s32)
+    %3:vgpr(s32) = G_FMUL %0, %9
+    %4:sgpr(s32) = G_FCONSTANT float 1.000000e+00
+    %10:vgpr(s32) = COPY %4(s32)
+    %5:vgpr(s32) = G_FMINNUM %3, %10
+    %6:sgpr(s32) = G_FCONSTANT float 0.000000e+00
+    %11:vgpr(s32) = COPY %6(s32)
+    %7:vgpr(s32) = G_FMAXNUM %5, %11
+    $vgpr0 = COPY %7(s32)
+...