[RISCV] Add lowering for scalar fmaximum/fminimum.

[craig.topper]

Summary

Unlike fmaxnum and fminnum, these operations propagate nan and
consider -0.0 to be less than +0.0.

Without Zfa, we don't have a single instruction for this. The
lowering I've used forces the other input to nan if one input
is a nan. If both inputs are nan, they get swapped. Then use
the fmax or fmin instruction.

New ISD nodes are needed because fmaxnum/fminnum to not define
the order of -0.0 and +0.0.

This loweing ensures the snans are quieted though that is probably not
required in default environment). Also ensures non-canonical nans
are canonicalized, though I'm also not sure that's needed.

Another option could be to use fmax/fmin and then overwrite the
result based on the inputs being nan, but I'm not sure we can do
that with any less code.

Future work will handle nonans FMF, and handling the case where
we can prove the input isn't nan.

This does fix the crash in #64022, but we need to do more work
to avoid scalarization.

[fakepaper56]

LGTM.

llvm/lib/Target/RISCV/RISCVISelLowering.h

enum NodeType : unsigned {
   // Rounds an FP value to its corresponding integer in the same FP format.
   // First operand is the value to round, the second operand is the largest
   // integer that can be represented exactly in the FP format. This will be
   // expanded into multiple instructions and basic blocks with a custom
   // inserter.
   FROUND,
 
   FPCLASS,
+
+  // Floating point fmax and fmin matching the RISC-V instruction semantics.
+  FMAX, FMIN,
+
   // READ_CYCLE_WIDE - A read of the 64-bit cycle CSR on a 32-bit target
   // (returns (Lo, Hi)). It takes a chain operand.
   READ_CYCLE_WIDE,
   // brev8, orc.b, zip, and unzip from Zbb and Zbkb. All operands are i32 or
   // XLenVT.
   BREV8,
   ORC_B,
   ZIP,

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

setOperationAction({ISD::STRICT_FCEIL, ISD::STRICT_FFLOOR,
                         ISD::STRICT_FNEARBYINT, ISD::STRICT_FRINT,
                         ISD::STRICT_FROUND, ISD::STRICT_FROUNDEVEN,
                         ISD::STRICT_FTRUNC},
                        MVT::f16, Promote);
 
     // We need to custom promote this.
     if (Subtarget.is64Bit())
       setOperationAction(ISD::FPOWI, MVT::i32, Custom);
+
+    if (!Subtarget.hasStdExtZfa())
+      setOperationAction({ISD::FMAXIMUM, ISD::FMINIMUM}, MVT::f16, Custom);
   }
 
   if (Subtarget.hasStdExtFOrZfinx()) {
     setOperationAction(FPLegalNodeTypes, MVT::f32, Legal);
     setOperationAction(FPRndMode, MVT::f32,
                        Subtarget.hasStdExtZfa() ? Legal : Custom);
     setCondCodeAction(FPCCToExpand, MVT::f32, Expand);
     setOperationAction(ISD::SELECT_CC, MVT::f32, Expand);
     setOperationAction(ISD::SELECT, MVT::f32, Custom);
     setOperationAction(ISD::BR_CC, MVT::f32, Expand);
     setOperationAction(FPOpToExpand, MVT::f32, Expand);
     setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
     setTruncStoreAction(MVT::f32, MVT::f16, Expand);
     setOperationAction(ISD::IS_FPCLASS, MVT::f32, Custom);
     setOperationAction(ISD::BF16_TO_FP, MVT::f32, Custom);
     setOperationAction(ISD::FP_TO_BF16, MVT::f32,
                        Subtarget.isSoftFPABI() ? LibCall : Custom);
     setOperationAction(ISD::FP_TO_FP16, MVT::f32, Custom);
     setOperationAction(ISD::FP16_TO_FP, MVT::f32, Custom);
 
     if (Subtarget.hasStdExtZfa())
       setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
+    else
+      setOperationAction({ISD::FMAXIMUM, ISD::FMINIMUM}, MVT::f32, Custom);
   }
 
   if (Subtarget.hasStdExtFOrZfinx() && Subtarget.is64Bit())
     setOperationAction(ISD::BITCAST, MVT::i32, Custom);
 
   if (Subtarget.hasStdExtDOrZdinx()) {
     setOperationAction(FPLegalNodeTypes, MVT::f64, Legal);
 
     if (Subtarget.hasStdExtZfa()) {
       setOperationAction(FPRndMode, MVT::f64, Legal);
       setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
       setOperationAction(ISD::BITCAST, MVT::i64, Custom);
       setOperationAction(ISD::BITCAST, MVT::f64, Custom);
     } else {
       if (Subtarget.is64Bit())
         setOperationAction(FPRndMode, MVT::f64, Custom);
+
+      setOperationAction({ISD::FMAXIMUM, ISD::FMINIMUM}, MVT::f64, Custom);
     }
 
     setOperationAction(ISD::STRICT_FP_ROUND, MVT::f32, Legal);
     setOperationAction(ISD::STRICT_FP_EXTEND, MVT::f64, Legal);
     setCondCodeAction(FPCCToExpand, MVT::f64, Expand);
     setOperationAction(ISD::SELECT_CC, MVT::f64, Expand);
     setOperationAction(ISD::SELECT, MVT::f64, Custom);
     setOperationAction(ISD::BR_CC, MVT::f64, Expand);
SDValue RISCVTargetLowering::LowerIS_FPCLASS(SDValue Op,
   }
 
   SDValue FPCLASS = DAG.getNode(RISCVISD::FPCLASS, DL, VT, Op.getOperand(0));
   SDValue AND = DAG.getNode(ISD::AND, DL, VT, FPCLASS, TDCMaskV);
   return DAG.getSetCC(DL, VT, AND, DAG.getConstant(0, DL, XLenVT),
                       ISD::CondCode::SETNE);
 }
 
+// Lower fmaximum and fminimum. Unlike our fmax and fmin instructions, these
+// operations propagate nans.
+static SDValue lowerFMAXIMUM_FMINIMUM(SDValue Op, SelectionDAG &DAG,
+                                      const RISCVSubtarget &Subtarget) {
+  SDLoc DL(Op);
+  EVT VT = Op.getValueType();
+
+  SDValue X = Op.getOperand(0);
+  SDValue Y = Op.getOperand(1);
+
+  MVT XLenVT = Subtarget.getXLenVT();
+
+  // If X is a nan, replace Y with X. If Y is a nan, replace X with Y. This
+  // ensures that when one input is a nan, the other will also be a nan allowing
+  // the nan to propagate. If both inputs are nan, this will swap the inputs
+  // which is harmless.
+  // FIXME: Handle nonans FMF and use isKnownNeverNaN.
+  SDValue XIsNonNan = DAG.getSetCC(DL, XLenVT, X, X, ISD::SETOEQ);
+  SDValue NewY = DAG.getSelect(DL, VT, XIsNonNan, Y, X);
+
+  SDValue YIsNonNan = DAG.getSetCC(DL, XLenVT, Y, Y, ISD::SETOEQ);
+  SDValue NewX = DAG.getSelect(DL, VT, YIsNonNan, X, Y);
+
+  unsigned Opc =
+      Op.getOpcode() == ISD::FMAXIMUM ? RISCVISD::FMAX : RISCVISD::FMIN;
+  return DAG.getNode(Opc, DL, VT, NewX, NewY);
+}
+
 /// Get a RISCV target specified VL op for a given SDNode.
 static unsigned getRISCVVLOp(SDValue Op) {
 #define OP_CASE(NODE)                                                          \
   case ISD::NODE:                                                              \
     return RISCVISD::NODE##_VL;
   switch (Op.getOpcode()) {
   default:
     llvm_unreachable("don't have RISC-V specified VL op for this SDNode");
if (Op.getValueType() == MVT::f16 && Subtarget.is64Bit() &&
       SDValue Op0 = DAG.getNode(ISD::FP_EXTEND, DL, MVT::f32, Op.getOperand(0));
       SDValue Powi =
           DAG.getNode(ISD::FPOWI, DL, MVT::f32, Op0, Op.getOperand(1));
       return DAG.getNode(ISD::FP_ROUND, DL, MVT::f16, Powi,
                          DAG.getIntPtrConstant(0, DL, /*isTarget=*/true));
     }
     return SDValue();
   }
+  case ISD::FMAXIMUM:
+  case ISD::FMINIMUM:
+    return lowerFMAXIMUM_FMINIMUM(Op, DAG, Subtarget);
   case ISD::FP_EXTEND: {
     SDLoc DL(Op);
     EVT VT = Op.getValueType();
     SDValue Op0 = Op.getOperand(0);
     EVT Op0VT = Op0.getValueType();
     if (VT == MVT::f32 && Op0VT == MVT::bf16 && Subtarget.hasStdExtZfbfmin())
       return DAG.getNode(RISCVISD::FP_EXTEND_BF16, DL, MVT::f32, Op0);
     if (VT == MVT::f64 && Op0VT == MVT::bf16 && Subtarget.hasStdExtZfbfmin()) {
#define NODE_NAME_CASE(NODE)                                                   \
   NODE_NAME_CASE(FCVT_W_RV64)
   NODE_NAME_CASE(FCVT_WU_RV64)
   NODE_NAME_CASE(STRICT_FCVT_W_RV64)
   NODE_NAME_CASE(STRICT_FCVT_WU_RV64)
   NODE_NAME_CASE(FP_ROUND_BF16)
   NODE_NAME_CASE(FP_EXTEND_BF16)
   NODE_NAME_CASE(FROUND)
   NODE_NAME_CASE(FPCLASS)
+  NODE_NAME_CASE(FMAX)
+  NODE_NAME_CASE(FMIN)
   NODE_NAME_CASE(READ_CYCLE_WIDE)
   NODE_NAME_CASE(BREV8)
   NODE_NAME_CASE(ORC_B)
   NODE_NAME_CASE(ZIP)
   NODE_NAME_CASE(UNZIP)
   NODE_NAME_CASE(CLMUL)
   NODE_NAME_CASE(CLMULH)
   NODE_NAME_CASE(CLMULR)

llvm/lib/Target/RISCV/RISCVInstrInfoD.td

 } // Predicates = [HasStdExtZdinx, IsRV32]
 
 // The ratified 20191213 ISA spec defines fmin and fmax in a way that matches
 // LLVM's fminnum and fmaxnum.
 // <https://github.com/riscv/riscv-isa-manual/commit/cd20cee7efd9bac7c5aa127ec3b451749d2b3cce>.
 foreach Ext = DExts in {
   defm : PatFprFpr_m<fminnum, FMIN_D, Ext>;
   defm : PatFprFpr_m<fmaxnum, FMAX_D, Ext>;
+  defm : PatFprFpr_m<riscv_fmin, FMIN_D, Ext>;
+  defm : PatFprFpr_m<riscv_fmax, FMAX_D, Ext>;
 }
 
 /// Setcc
 // FIXME: SETEQ/SETLT/SETLE imply nonans, can we pick better instructions for
 // strict versions of those.
 
 // Match non-signaling FEQ_D
 foreach Ext = DExts in {

llvm/lib/Target/RISCV/RISCVInstrInfoF.td

def riscv_fcvt_w_rv64
     : SDNode<"RISCVISD::FCVT_W_RV64", SDT_RISCVFCVT_W_RV64>;
 def riscv_fcvt_wu_rv64
     : SDNode<"RISCVISD::FCVT_WU_RV64", SDT_RISCVFCVT_W_RV64>;
 def riscv_fcvt_x
     : SDNode<"RISCVISD::FCVT_X", SDT_RISCVFCVT_X>;
 def riscv_fcvt_xu
     : SDNode<"RISCVISD::FCVT_XU", SDT_RISCVFCVT_X>;
 
+def riscv_fmin : SDNode<"RISCVISD::FMIN", SDTFPBinOp>;
+def riscv_fmax : SDNode<"RISCVISD::FMAX", SDTFPBinOp>;
+
 def riscv_strict_fcvt_w_rv64
     : SDNode<"RISCVISD::STRICT_FCVT_W_RV64", SDT_RISCVFCVT_W_RV64,
              [SDNPHasChain]>;
 def riscv_strict_fcvt_wu_rv64
     : SDNode<"RISCVISD::STRICT_FCVT_WU_RV64", SDT_RISCVFCVT_W_RV64,
              [SDNPHasChain]>;
 
 def riscv_any_fcvt_w_rv64 : PatFrags<(ops node:$src, node:$frm),

 } // Predicates = [HasStdExtZfinx]
 
 // The ratified 20191213 ISA spec defines fmin and fmax in a way that matches
 // LLVM's fminnum and fmaxnum
 // <https://github.com/riscv/riscv-isa-manual/commit/cd20cee7efd9bac7c5aa127ec3b451749d2b3cce>.
 foreach Ext = FExts in {
   defm : PatFprFpr_m<fminnum, FMIN_S, Ext>;
   defm : PatFprFpr_m<fmaxnum, FMAX_S, Ext>;
+  defm : PatFprFpr_m<riscv_fmin, FMIN_S, Ext>;
+  defm : PatFprFpr_m<riscv_fmax, FMAX_S, Ext>;
 }
 
 /// Setcc
 // FIXME: SETEQ/SETLT/SETLE imply nonans, can we pick better instructions for
 // strict versions of those.
 
 // Match non-signaling FEQ_S
 foreach Ext = FExts in {

llvm/lib/Target/RISCV/RISCVInstrInfoZfh.td

 } // Predicates = [HasStdExtZhinx]
 
 // The ratified 20191213 ISA spec defines fmin and fmax in a way that matches
 // LLVM's fminnum and fmaxnum
 // <https://github.com/riscv/riscv-isa-manual/commit/cd20cee7efd9bac7c5aa127ec3b451749d2b3cce>.
 foreach Ext = ZfhExts in {
   defm : PatFprFpr_m<fminnum, FMIN_H, Ext>;
   defm : PatFprFpr_m<fmaxnum, FMAX_H, Ext>;
+  defm : PatFprFpr_m<riscv_fmin, FMIN_H, Ext>;
+  defm : PatFprFpr_m<riscv_fmax, FMAX_H, Ext>;
 }
 
 /// Setcc
 // FIXME: SETEQ/SETLT/SETLE imply nonans, can we pick better instructions for
 // strict versions of those.
 
 // Match non-signaling FEQ_D
 foreach Ext = ZfhExts in {

llvm/test/Analysis/CostModel/RISCV/fp-min-max-abs.ll

;
   call <vscale x 2 x double> @llvm.maxnum.nvx2f64(<vscale x 2 x double> undef, <vscale x 2 x double> undef)
   call <vscale x 4 x double> @llvm.maxnum.nvx4f64(<vscale x 4 x double> undef, <vscale x 4 x double> undef)
   call <vscale x 8 x double> @llvm.maxnum.nvx8f64(<vscale x 8 x double> undef, <vscale x 8 x double> undef)
   ret void
 }
 
 define void @minimum() {
 ; CHECK-LABEL: 'minimum'
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %1 = call float @llvm.minimum.f32(float undef, float undef)
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 23 for instruction: %2 = call <2 x float> @llvm.minimum.v2f32(<2 x float> undef, <2 x float> undef)
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 47 for instruction: %3 = call <4 x float> @llvm.minimum.v4f32(<4 x float> undef, <4 x float> undef)
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 95 for instruction: %4 = call <8 x float> @llvm.minimum.v8f32(<8 x float> undef, <8 x float> undef)
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 191 for instruction: %5 = call <16 x float> @llvm.minimum.v16f32(<16 x float> undef, <16 x float> undef)
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %1 = call float @llvm.minimum.f32(float undef, float undef)
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 7 for instruction: %2 = call <2 x float> @llvm.minimum.v2f32(<2 x float> undef, <2 x float> undef)
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 15 for instruction: %3 = call <4 x float> @llvm.minimum.v4f32(<4 x float> undef, <4 x float> undef)
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 31 for instruction: %4 = call <8 x float> @llvm.minimum.v8f32(<8 x float> undef, <8 x float> undef)
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 63 for instruction: %5 = call <16 x float> @llvm.minimum.v16f32(<16 x float> undef, <16 x float> undef)
 ; CHECK-NEXT:  Cost Model: Invalid cost for instruction: %6 = call <vscale x 1 x float> @llvm.minimum.nxv1f32(<vscale x 1 x float> undef, <vscale x 1 x float> undef)
 ; CHECK-NEXT:  Cost Model: Invalid cost for instruction: %7 = call <vscale x 2 x float> @llvm.minimum.nxv2f32(<vscale x 2 x float> undef, <vscale x 2 x float> undef)
 ; CHECK-NEXT:  Cost Model: Invalid cost for instruction: %8 = call <vscale x 4 x float> @llvm.minimum.nxv4f32(<vscale x 4 x float> undef, <vscale x 4 x float> undef)
 ; CHECK-NEXT:  Cost Model: Invalid cost for instruction: %9 = call <vscale x 8 x float> @llvm.minimum.nxv8f32(<vscale x 8 x float> undef, <vscale x 8 x float> undef)
 ; CHECK-NEXT:  Cost Model: Invalid cost for instruction: %10 = call <vscale x 16 x float> @llvm.minimum.nxv16f32(<vscale x 16 x float> undef, <vscale x 16 x float> undef)
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %11 = call double @llvm.minimum.f64(double undef, double undef)
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 23 for instruction: %12 = call <2 x double> @llvm.minimum.v2f64(<2 x double> undef, <2 x double> undef)
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 47 for instruction: %13 = call <4 x double> @llvm.minimum.v4f64(<4 x double> undef, <4 x double> undef)
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 95 for instruction: %14 = call <8 x double> @llvm.minimum.v8f64(<8 x double> undef, <8 x double> undef)
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 191 for instruction: %15 = call <16 x double> @llvm.minimum.v16f64(<16 x double> undef, <16 x double> undef)
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %11 = call double @llvm.minimum.f64(double undef, double undef)
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 7 for instruction: %12 = call <2 x double> @llvm.minimum.v2f64(<2 x double> undef, <2 x double> undef)
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 15 for instruction: %13 = call <4 x double> @llvm.minimum.v4f64(<4 x double> undef, <4 x double> undef)
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 31 for instruction: %14 = call <8 x double> @llvm.minimum.v8f64(<8 x double> undef, <8 x double> undef)
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 63 for instruction: %15 = call <16 x double> @llvm.minimum.v16f64(<16 x double> undef, <16 x double> undef)
 ; CHECK-NEXT:  Cost Model: Invalid cost for instruction: %16 = call <vscale x 1 x double> @llvm.minimum.nxv1f64(<vscale x 1 x double> undef, <vscale x 1 x double> undef)
 ; CHECK-NEXT:  Cost Model: Invalid cost for instruction: %17 = call <vscale x 2 x double> @llvm.minimum.nxv2f64(<vscale x 2 x double> undef, <vscale x 2 x double> undef)
 ; CHECK-NEXT:  Cost Model: Invalid cost for instruction: %18 = call <vscale x 4 x double> @llvm.minimum.nxv4f64(<vscale x 4 x double> undef, <vscale x 4 x double> undef)
 ; CHECK-NEXT:  Cost Model: Invalid cost for instruction: %19 = call <vscale x 8 x double> @llvm.minimum.nxv8f64(<vscale x 8 x double> undef, <vscale x 8 x double> undef)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret void
 ;
   call float @llvm.minimum.f32(float undef, float undef)
   call <2 x float> @llvm.minimum.v2f32(<2 x float> undef, <2 x float> undef)
;
   call <vscale x 2 x double> @llvm.minimum.nvx2f64(<vscale x 2 x double> undef, <vscale x 2 x double> undef)
   call <vscale x 4 x double> @llvm.minimum.nvx4f64(<vscale x 4 x double> undef, <vscale x 4 x double> undef)
   call <vscale x 8 x double> @llvm.minimum.nvx8f64(<vscale x 8 x double> undef, <vscale x 8 x double> undef)
   ret void
 }
 
 define void @maximum() {
 ; CHECK-LABEL: 'maximum'
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %1 = call float @llvm.maximum.f32(float undef, float undef)
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 23 for instruction: %2 = call <2 x float> @llvm.maximum.v2f32(<2 x float> undef, <2 x float> undef)
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 47 for instruction: %3 = call <4 x float> @llvm.maximum.v4f32(<4 x float> undef, <4 x float> undef)
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 95 for instruction: %4 = call <8 x float> @llvm.maximum.v8f32(<8 x float> undef, <8 x float> undef)
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 191 for instruction: %5 = call <16 x float> @llvm.maximum.v16f32(<16 x float> undef, <16 x float> undef)
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %1 = call float @llvm.maximum.f32(float undef, float undef)
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 7 for instruction: %2 = call <2 x float> @llvm.maximum.v2f32(<2 x float> undef, <2 x float> undef)
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 15 for instruction: %3 = call <4 x float> @llvm.maximum.v4f32(<4 x float> undef, <4 x float> undef)
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 31 for instruction: %4 = call <8 x float> @llvm.maximum.v8f32(<8 x float> undef, <8 x float> undef)
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 63 for instruction: %5 = call <16 x float> @llvm.maximum.v16f32(<16 x float> undef, <16 x float> undef)
 ; CHECK-NEXT:  Cost Model: Invalid cost for instruction: %6 = call <vscale x 1 x float> @llvm.maximum.nxv1f32(<vscale x 1 x float> undef, <vscale x 1 x float> undef)
 ; CHECK-NEXT:  Cost Model: Invalid cost for instruction: %7 = call <vscale x 2 x float> @llvm.maximum.nxv2f32(<vscale x 2 x float> undef, <vscale x 2 x float> undef)
 ; CHECK-NEXT:  Cost Model: Invalid cost for instruction: %8 = call <vscale x 4 x float> @llvm.maximum.nxv4f32(<vscale x 4 x float> undef, <vscale x 4 x float> undef)
 ; CHECK-NEXT:  Cost Model: Invalid cost for instruction: %9 = call <vscale x 8 x float> @llvm.maximum.nxv8f32(<vscale x 8 x float> undef, <vscale x 8 x float> undef)
 ; CHECK-NEXT:  Cost Model: Invalid cost for instruction: %10 = call <vscale x 16 x float> @llvm.maximum.nxv16f32(<vscale x 16 x float> undef, <vscale x 16 x float> undef)
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 10 for instruction: %11 = call double @llvm.maximum.f64(double undef, double undef)
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 23 for instruction: %12 = call <2 x double> @llvm.maximum.v2f64(<2 x double> undef, <2 x double> undef)
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 47 for instruction: %13 = call <4 x double> @llvm.maximum.v4f64(<4 x double> undef, <4 x double> undef)
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 95 for instruction: %14 = call <8 x double> @llvm.maximum.v8f64(<8 x double> undef, <8 x double> undef)
-; CHECK-NEXT:  Cost Model: Found an estimated cost of 191 for instruction: %15 = call <16 x double> @llvm.maximum.v16f64(<16 x double> undef, <16 x double> undef)
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 2 for instruction: %11 = call double @llvm.maximum.f64(double undef, double undef)
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 7 for instruction: %12 = call <2 x double> @llvm.maximum.v2f64(<2 x double> undef, <2 x double> undef)
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 15 for instruction: %13 = call <4 x double> @llvm.maximum.v4f64(<4 x double> undef, <4 x double> undef)
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 31 for instruction: %14 = call <8 x double> @llvm.maximum.v8f64(<8 x double> undef, <8 x double> undef)
+; CHECK-NEXT:  Cost Model: Found an estimated cost of 63 for instruction: %15 = call <16 x double> @llvm.maximum.v16f64(<16 x double> undef, <16 x double> undef)
 ; CHECK-NEXT:  Cost Model: Invalid cost for instruction: %16 = call <vscale x 1 x double> @llvm.maximum.nxv1f64(<vscale x 1 x double> undef, <vscale x 1 x double> undef)
 ; CHECK-NEXT:  Cost Model: Invalid cost for instruction: %17 = call <vscale x 2 x double> @llvm.maximum.nxv2f64(<vscale x 2 x double> undef, <vscale x 2 x double> undef)
 ; CHECK-NEXT:  Cost Model: Invalid cost for instruction: %18 = call <vscale x 4 x double> @llvm.maximum.nxv4f64(<vscale x 4 x double> undef, <vscale x 4 x double> undef)
 ; CHECK-NEXT:  Cost Model: Invalid cost for instruction: %19 = call <vscale x 8 x double> @llvm.maximum.nxv8f64(<vscale x 8 x double> undef, <vscale x 8 x double> undef)
 ; CHECK-NEXT:  Cost Model: Found an estimated cost of 1 for instruction: ret void
 ;
   call float @llvm.maximum.f32(float undef, float undef)
   call <2 x float> @llvm.maximum.v2f32(<2 x float> undef, <2 x float> undef)

llvm/test/CodeGen/RISCV/double-maximum-minimum.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: sed 's/iXLen/i32/g' %s | llc -mtriple=riscv32 -mattr=+d \
+; RUN:   -verify-machineinstrs -target-abi=ilp32d \
+; RUN:   | FileCheck -check-prefixes=CHECKIFD,RV32IFD %s
+; RUN: sed 's/iXLen/i64/g' %s | llc -mtriple=riscv64 -mattr=+d \
+; RUN:   -verify-machineinstrs -target-abi=lp64d \
+; RUN:   | FileCheck -check-prefixes=CHECKIFD,RV64IFD %s
+; RUN: sed 's/iXLen/i32/g' %s | llc -mtriple=riscv32 -mattr=+zdinx \
+; RUN:   -verify-machineinstrs -target-abi=ilp32 \
+; RUN:   | FileCheck -check-prefix=RV32IZFINXZDINX %s
+; RUN: sed 's/iXLen/i64/g' %s | llc -mtriple=riscv64 -mattr=+zdinx \
+; RUN:   -verify-machineinstrs -target-abi=lp64 \
+; RUN:   | FileCheck -check-prefix=RV64IZFINXZDINX %s
+
+declare double @llvm.minimum.f64(double, double)
+
+define double @fminimum_f64(double %a, double %b) nounwind {
+; CHECKIFD-LABEL: fminimum_f64:
+; CHECKIFD:       # %bb.0:
+; CHECKIFD-NEXT:    feq.d a0, fa0, fa0
+; CHECKIFD-NEXT:    fmv.d fa5, fa1
+; CHECKIFD-NEXT:    beqz a0, .LBB0_3
+; CHECKIFD-NEXT:  # %bb.1:
+; CHECKIFD-NEXT:    feq.d a0, fa1, fa1
+; CHECKIFD-NEXT:    beqz a0, .LBB0_4
+; CHECKIFD-NEXT:  .LBB0_2:
+; CHECKIFD-NEXT:    fmin.d fa0, fa0, fa5
+; CHECKIFD-NEXT:    ret
+; CHECKIFD-NEXT:  .LBB0_3:
+; CHECKIFD-NEXT:    fmv.d fa5, fa0
+; CHECKIFD-NEXT:    feq.d a0, fa1, fa1
+; CHECKIFD-NEXT:    bnez a0, .LBB0_2
+; CHECKIFD-NEXT:  .LBB0_4:
+; CHECKIFD-NEXT:    fmin.d fa0, fa1, fa5
+; CHECKIFD-NEXT:    ret
+;
+; RV32IZFINXZDINX-LABEL: fminimum_f64:
+; RV32IZFINXZDINX:       # %bb.0:
+; RV32IZFINXZDINX-NEXT:    addi sp, sp, -16
+; RV32IZFINXZDINX-NEXT:    sw a2, 8(sp)
+; RV32IZFINXZDINX-NEXT:    sw a3, 12(sp)
+; RV32IZFINXZDINX-NEXT:    lw a2, 8(sp)
+; RV32IZFINXZDINX-NEXT:    lw a3, 12(sp)
+; RV32IZFINXZDINX-NEXT:    sw a0, 8(sp)
+; RV32IZFINXZDINX-NEXT:    sw a1, 12(sp)
+; RV32IZFINXZDINX-NEXT:    lw a0, 8(sp)
+; RV32IZFINXZDINX-NEXT:    lw a1, 12(sp)
+; RV32IZFINXZDINX-NEXT:    feq.d a6, a0, a0
+; RV32IZFINXZDINX-NEXT:    mv a4, a2
+; RV32IZFINXZDINX-NEXT:    bnez a6, .LBB0_2
+; RV32IZFINXZDINX-NEXT:  # %bb.1:
+; RV32IZFINXZDINX-NEXT:    mv a4, a0
+; RV32IZFINXZDINX-NEXT:  .LBB0_2:
+; RV32IZFINXZDINX-NEXT:    feq.d a6, a2, a2
+; RV32IZFINXZDINX-NEXT:    bnez a6, .LBB0_4
+; RV32IZFINXZDINX-NEXT:  # %bb.3:
+; RV32IZFINXZDINX-NEXT:    mv a0, a2
+; RV32IZFINXZDINX-NEXT:  .LBB0_4:
+; RV32IZFINXZDINX-NEXT:    fmin.d a0, a0, a4
+; RV32IZFINXZDINX-NEXT:    sw a0, 8(sp)
+; RV32IZFINXZDINX-NEXT:    sw a1, 12(sp)
+; RV32IZFINXZDINX-NEXT:    lw a0, 8(sp)
+; RV32IZFINXZDINX-NEXT:    lw a1, 12(sp)
+; RV32IZFINXZDINX-NEXT:    addi sp, sp, 16
+; RV32IZFINXZDINX-NEXT:    ret
+;
+; RV64IZFINXZDINX-LABEL: fminimum_f64:
+; RV64IZFINXZDINX:       # %bb.0:
+; RV64IZFINXZDINX-NEXT:    feq.d a3, a0, a0
+; RV64IZFINXZDINX-NEXT:    mv a2, a1
+; RV64IZFINXZDINX-NEXT:    beqz a3, .LBB0_3
+; RV64IZFINXZDINX-NEXT:  # %bb.1:
+; RV64IZFINXZDINX-NEXT:    feq.d a3, a1, a1
+; RV64IZFINXZDINX-NEXT:    beqz a3, .LBB0_4
+; RV64IZFINXZDINX-NEXT:  .LBB0_2:
+; RV64IZFINXZDINX-NEXT:    fmin.d a0, a0, a2
+; RV64IZFINXZDINX-NEXT:    ret
+; RV64IZFINXZDINX-NEXT:  .LBB0_3:
+; RV64IZFINXZDINX-NEXT:    mv a2, a0
+; RV64IZFINXZDINX-NEXT:    feq.d a3, a1, a1
+; RV64IZFINXZDINX-NEXT:    bnez a3, .LBB0_2
+; RV64IZFINXZDINX-NEXT:  .LBB0_4:
+; RV64IZFINXZDINX-NEXT:    fmin.d a0, a1, a2
+; RV64IZFINXZDINX-NEXT:    ret
+  %1 = call double @llvm.minimum.f64(double %a, double %b)
+  ret double %1
+}
+
+declare double @llvm.maximum.f64(double, double)
+
+define double @fmaximum_f64(double %a, double %b) nounwind {
+; CHECKIFD-LABEL: fmaximum_f64:
+; CHECKIFD:       # %bb.0:
+; CHECKIFD-NEXT:    feq.d a0, fa0, fa0
+; CHECKIFD-NEXT:    fmv.d fa5, fa1
+; CHECKIFD-NEXT:    beqz a0, .LBB1_3
+; CHECKIFD-NEXT:  # %bb.1:
+; CHECKIFD-NEXT:    feq.d a0, fa1, fa1
+; CHECKIFD-NEXT:    beqz a0, .LBB1_4
+; CHECKIFD-NEXT:  .LBB1_2:
+; CHECKIFD-NEXT:    fmax.d fa0, fa0, fa5
+; CHECKIFD-NEXT:    ret
+; CHECKIFD-NEXT:  .LBB1_3:
+; CHECKIFD-NEXT:    fmv.d fa5, fa0
+; CHECKIFD-NEXT:    feq.d a0, fa1, fa1
+; CHECKIFD-NEXT:    bnez a0, .LBB1_2
+; CHECKIFD-NEXT:  .LBB1_4:
+; CHECKIFD-NEXT:    fmax.d fa0, fa1, fa5
+; CHECKIFD-NEXT:    ret
+;
+; RV32IZFINXZDINX-LABEL: fmaximum_f64:
+; RV32IZFINXZDINX:       # %bb.0:
+; RV32IZFINXZDINX-NEXT:    addi sp, sp, -16
+; RV32IZFINXZDINX-NEXT:    sw a2, 8(sp)
+; RV32IZFINXZDINX-NEXT:    sw a3, 12(sp)
+; RV32IZFINXZDINX-NEXT:    lw a2, 8(sp)
+; RV32IZFINXZDINX-NEXT:    lw a3, 12(sp)
+; RV32IZFINXZDINX-NEXT:    sw a0, 8(sp)
+; RV32IZFINXZDINX-NEXT:    sw a1, 12(sp)
+; RV32IZFINXZDINX-NEXT:    lw a0, 8(sp)
+; RV32IZFINXZDINX-NEXT:    lw a1, 12(sp)
+; RV32IZFINXZDINX-NEXT:    feq.d a6, a0, a0
+; RV32IZFINXZDINX-NEXT:    mv a4, a2
+; RV32IZFINXZDINX-NEXT:    bnez a6, .LBB1_2
+; RV32IZFINXZDINX-NEXT:  # %bb.1:
+; RV32IZFINXZDINX-NEXT:    mv a4, a0
+; RV32IZFINXZDINX-NEXT:  .LBB1_2:
+; RV32IZFINXZDINX-NEXT:    feq.d a6, a2, a2
+; RV32IZFINXZDINX-NEXT:    bnez a6, .LBB1_4
+; RV32IZFINXZDINX-NEXT:  # %bb.3:
+; RV32IZFINXZDINX-NEXT:    mv a0, a2
+; RV32IZFINXZDINX-NEXT:  .LBB1_4:
+; RV32IZFINXZDINX-NEXT:    fmax.d a0, a0, a4
+; RV32IZFINXZDINX-NEXT:    sw a0, 8(sp)
+; RV32IZFINXZDINX-NEXT:    sw a1, 12(sp)
+; RV32IZFINXZDINX-NEXT:    lw a0, 8(sp)
+; RV32IZFINXZDINX-NEXT:    lw a1, 12(sp)
+; RV32IZFINXZDINX-NEXT:    addi sp, sp, 16
+; RV32IZFINXZDINX-NEXT:    ret
+;
+; RV64IZFINXZDINX-LABEL: fmaximum_f64:
+; RV64IZFINXZDINX:       # %bb.0:
+; RV64IZFINXZDINX-NEXT:    feq.d a3, a0, a0
+; RV64IZFINXZDINX-NEXT:    mv a2, a1
+; RV64IZFINXZDINX-NEXT:    beqz a3, .LBB1_3
+; RV64IZFINXZDINX-NEXT:  # %bb.1:
+; RV64IZFINXZDINX-NEXT:    feq.d a3, a1, a1
+; RV64IZFINXZDINX-NEXT:    beqz a3, .LBB1_4
+; RV64IZFINXZDINX-NEXT:  .LBB1_2:
+; RV64IZFINXZDINX-NEXT:    fmax.d a0, a0, a2
+; RV64IZFINXZDINX-NEXT:    ret
+; RV64IZFINXZDINX-NEXT:  .LBB1_3:
+; RV64IZFINXZDINX-NEXT:    mv a2, a0
+; RV64IZFINXZDINX-NEXT:    feq.d a3, a1, a1
+; RV64IZFINXZDINX-NEXT:    bnez a3, .LBB1_2
+; RV64IZFINXZDINX-NEXT:  .LBB1_4:
+; RV64IZFINXZDINX-NEXT:    fmax.d a0, a1, a2
+; RV64IZFINXZDINX-NEXT:    ret
+  %1 = call double @llvm.maximum.f64(double %a, double %b)
+  ret double %1
+}
+;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
+; RV32IFD: {{.*}}
+; RV64IFD: {{.*}}

llvm/test/CodeGen/RISCV/float-maximum-minimum.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: sed 's/iXLen/i32/g' %s | llc -mtriple=riscv32 -mattr=+f \
+; RUN:   -verify-machineinstrs -target-abi=ilp32f \
+; RUN:   | FileCheck -check-prefix=RV32IF %s
+; RUN: sed 's/iXLen/i32/g' %s | llc -mtriple=riscv32 -mattr=+zfinx \
+; RUN:   -verify-machineinstrs -target-abi=ilp32 \
+; RUN:   | FileCheck -check-prefix=RV32IZFINX %s
+; RUN: sed 's/iXLen/i32/g' %s | llc -mtriple=riscv32 -mattr=+d \
+; RUN:   -verify-machineinstrs -target-abi=ilp32f \
+; RUN:   | FileCheck -check-prefix=RV32IF %s
+; RUN: sed 's/iXLen/i64/g' %s | llc -mtriple=riscv64 -mattr=+f \
+; RUN:   -verify-machineinstrs -target-abi=lp64f \
+; RUN:   | FileCheck -check-prefix=RV64IF %s
+; RUN: sed 's/iXLen/i64/g' %s | llc -mtriple=riscv64 -mattr=+zfinx \
+; RUN:   -verify-machineinstrs -target-abi=lp64 \
+; RUN:   | FileCheck -check-prefix=RV64IZFINX %s
+; RUN: sed 's/iXLen/i64/g' %s | llc -mtriple=riscv64 -mattr=+d \
+; RUN:   -verify-machineinstrs -target-abi=lp64d \
+; RUN:   | FileCheck -check-prefix=RV64IF %s
+
+declare float @llvm.minimum.f32(float, float)
+
+define float @fminimum_f32(float %a, float %b) nounwind {
+; RV32IF-LABEL: fminimum_f32:
+; RV32IF:       # %bb.0:
+; RV32IF-NEXT:    feq.s a0, fa0, fa0
+; RV32IF-NEXT:    fmv.s fa5, fa1
+; RV32IF-NEXT:    beqz a0, .LBB0_3
+; RV32IF-NEXT:  # %bb.1:
+; RV32IF-NEXT:    feq.s a0, fa1, fa1
+; RV32IF-NEXT:    beqz a0, .LBB0_4
+; RV32IF-NEXT:  .LBB0_2:
+; RV32IF-NEXT:    fmin.s fa0, fa0, fa5
+; RV32IF-NEXT:    ret
+; RV32IF-NEXT:  .LBB0_3:
+; RV32IF-NEXT:    fmv.s fa5, fa0
+; RV32IF-NEXT:    feq.s a0, fa1, fa1
+; RV32IF-NEXT:    bnez a0, .LBB0_2
+; RV32IF-NEXT:  .LBB0_4:
+; RV32IF-NEXT:    fmin.s fa0, fa1, fa5
+; RV32IF-NEXT:    ret
+;
+; RV32IZFINX-LABEL: fminimum_f32:
+; RV32IZFINX:       # %bb.0:
+; RV32IZFINX-NEXT:    feq.s a3, a0, a0
+; RV32IZFINX-NEXT:    mv a2, a1
+; RV32IZFINX-NEXT:    beqz a3, .LBB0_3
+; RV32IZFINX-NEXT:  # %bb.1:
+; RV32IZFINX-NEXT:    feq.s a3, a1, a1
+; RV32IZFINX-NEXT:    beqz a3, .LBB0_4
+; RV32IZFINX-NEXT:  .LBB0_2:
+; RV32IZFINX-NEXT:    fmin.s a0, a0, a2
+; RV32IZFINX-NEXT:    ret
+; RV32IZFINX-NEXT:  .LBB0_3:
+; RV32IZFINX-NEXT:    mv a2, a0
+; RV32IZFINX-NEXT:    feq.s a3, a1, a1
+; RV32IZFINX-NEXT:    bnez a3, .LBB0_2
+; RV32IZFINX-NEXT:  .LBB0_4:
+; RV32IZFINX-NEXT:    fmin.s a0, a1, a2
+; RV32IZFINX-NEXT:    ret
+;
+; RV64IF-LABEL: fminimum_f32:
+; RV64IF:       # %bb.0:
+; RV64IF-NEXT:    feq.s a0, fa0, fa0
+; RV64IF-NEXT:    fmv.s fa5, fa1
+; RV64IF-NEXT:    beqz a0, .LBB0_3
+; RV64IF-NEXT:  # %bb.1:
+; RV64IF-NEXT:    feq.s a0, fa1, fa1
+; RV64IF-NEXT:    beqz a0, .LBB0_4
+; RV64IF-NEXT:  .LBB0_2:
+; RV64IF-NEXT:    fmin.s fa0, fa0, fa5
+; RV64IF-NEXT:    ret
+; RV64IF-NEXT:  .LBB0_3:
+; RV64IF-NEXT:    fmv.s fa5, fa0
+; RV64IF-NEXT:    feq.s a0, fa1, fa1
+; RV64IF-NEXT:    bnez a0, .LBB0_2
+; RV64IF-NEXT:  .LBB0_4:
+; RV64IF-NEXT:    fmin.s fa0, fa1, fa5
+; RV64IF-NEXT:    ret
+;
+; RV64IZFINX-LABEL: fminimum_f32:
+; RV64IZFINX:       # %bb.0:
+; RV64IZFINX-NEXT:    feq.s a3, a0, a0
+; RV64IZFINX-NEXT:    mv a2, a1
+; RV64IZFINX-NEXT:    beqz a3, .LBB0_3
+; RV64IZFINX-NEXT:  # %bb.1:
+; RV64IZFINX-NEXT:    feq.s a3, a1, a1
+; RV64IZFINX-NEXT:    beqz a3, .LBB0_4
+; RV64IZFINX-NEXT:  .LBB0_2:
+; RV64IZFINX-NEXT:    fmin.s a0, a0, a2
+; RV64IZFINX-NEXT:    ret
+; RV64IZFINX-NEXT:  .LBB0_3:
+; RV64IZFINX-NEXT:    mv a2, a0
+; RV64IZFINX-NEXT:    feq.s a3, a1, a1
+; RV64IZFINX-NEXT:    bnez a3, .LBB0_2
+; RV64IZFINX-NEXT:  .LBB0_4:
+; RV64IZFINX-NEXT:    fmin.s a0, a1, a2
+; RV64IZFINX-NEXT:    ret
+  %1 = call float @llvm.minimum.f32(float %a, float %b)
+  ret float %1
+}
+
+declare float @llvm.maximum.f32(float, float)
+
+define float @fmaximum_f32(float %a, float %b) nounwind {
+; RV32IF-LABEL: fmaximum_f32:
+; RV32IF:       # %bb.0:
+; RV32IF-NEXT:    feq.s a0, fa0, fa0
+; RV32IF-NEXT:    fmv.s fa5, fa1
+; RV32IF-NEXT:    beqz a0, .LBB1_3
+; RV32IF-NEXT:  # %bb.1:
+; RV32IF-NEXT:    feq.s a0, fa1, fa1
+; RV32IF-NEXT:    beqz a0, .LBB1_4
+; RV32IF-NEXT:  .LBB1_2:
+; RV32IF-NEXT:    fmax.s fa0, fa0, fa5
+; RV32IF-NEXT:    ret
+; RV32IF-NEXT:  .LBB1_3:
+; RV32IF-NEXT:    fmv.s fa5, fa0
+; RV32IF-NEXT:    feq.s a0, fa1, fa1
+; RV32IF-NEXT:    bnez a0, .LBB1_2
+; RV32IF-NEXT:  .LBB1_4:
+; RV32IF-NEXT:    fmax.s fa0, fa1, fa5
+; RV32IF-NEXT:    ret
+;
+; RV32IZFINX-LABEL: fmaximum_f32:
+; RV32IZFINX:       # %bb.0:
+; RV32IZFINX-NEXT:    feq.s a3, a0, a0
+; RV32IZFINX-NEXT:    mv a2, a1
+; RV32IZFINX-NEXT:    beqz a3, .LBB1_3
+; RV32IZFINX-NEXT:  # %bb.1:
+; RV32IZFINX-NEXT:    feq.s a3, a1, a1
+; RV32IZFINX-NEXT:    beqz a3, .LBB1_4
+; RV32IZFINX-NEXT:  .LBB1_2:
+; RV32IZFINX-NEXT:    fmax.s a0, a0, a2
+; RV32IZFINX-NEXT:    ret
+; RV32IZFINX-NEXT:  .LBB1_3:
+; RV32IZFINX-NEXT:    mv a2, a0
+; RV32IZFINX-NEXT:    feq.s a3, a1, a1
+; RV32IZFINX-NEXT:    bnez a3, .LBB1_2
+; RV32IZFINX-NEXT:  .LBB1_4:
+; RV32IZFINX-NEXT:    fmax.s a0, a1, a2
+; RV32IZFINX-NEXT:    ret
+;
+; RV64IF-LABEL: fmaximum_f32:
+; RV64IF:       # %bb.0:
+; RV64IF-NEXT:    feq.s a0, fa0, fa0
+; RV64IF-NEXT:    fmv.s fa5, fa1
+; RV64IF-NEXT:    beqz a0, .LBB1_3
+; RV64IF-NEXT:  # %bb.1:
+; RV64IF-NEXT:    feq.s a0, fa1, fa1
+; RV64IF-NEXT:    beqz a0, .LBB1_4
+; RV64IF-NEXT:  .LBB1_2:
+; RV64IF-NEXT:    fmax.s fa0, fa0, fa5
+; RV64IF-NEXT:    ret
+; RV64IF-NEXT:  .LBB1_3:
+; RV64IF-NEXT:    fmv.s fa5, fa0
+; RV64IF-NEXT:    feq.s a0, fa1, fa1
+; RV64IF-NEXT:    bnez a0, .LBB1_2
+; RV64IF-NEXT:  .LBB1_4:
+; RV64IF-NEXT:    fmax.s fa0, fa1, fa5
+; RV64IF-NEXT:    ret
+;
+; RV64IZFINX-LABEL: fmaximum_f32:
+; RV64IZFINX:       # %bb.0:
+; RV64IZFINX-NEXT:    feq.s a3, a0, a0
+; RV64IZFINX-NEXT:    mv a2, a1
+; RV64IZFINX-NEXT:    beqz a3, .LBB1_3
+; RV64IZFINX-NEXT:  # %bb.1:
+; RV64IZFINX-NEXT:    feq.s a3, a1, a1
+; RV64IZFINX-NEXT:    beqz a3, .LBB1_4
+; RV64IZFINX-NEXT:  .LBB1_2:
+; RV64IZFINX-NEXT:    fmax.s a0, a0, a2
+; RV64IZFINX-NEXT:    ret
+; RV64IZFINX-NEXT:  .LBB1_3:
+; RV64IZFINX-NEXT:    mv a2, a0
+; RV64IZFINX-NEXT:    feq.s a3, a1, a1
+; RV64IZFINX-NEXT:    bnez a3, .LBB1_2
+; RV64IZFINX-NEXT:  .LBB1_4:
+; RV64IZFINX-NEXT:    fmax.s a0, a1, a2
+; RV64IZFINX-NEXT:    ret
+  %1 = call float @llvm.maximum.f32(float %a, float %b)
+  ret float %1
+}

llvm/test/CodeGen/RISCV/half-maximum-minimum.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: sed 's/iXLen/i32/g' %s | llc -mtriple=riscv32 -mattr=+zfh \
+; RUN:   -verify-machineinstrs -target-abi ilp32f | \
+; RUN:   FileCheck -check-prefixes=CHECKIZFH %s
+; RUN: sed 's/iXLen/i64/g' %s | llc -mtriple=riscv64 -mattr=+zfh \
+; RUN:   -verify-machineinstrs -target-abi lp64f | \
+; RUN:   FileCheck -check-prefixes=CHECKIZFH %s
+; RUN: sed 's/iXLen/i32/g' %s | llc -mtriple=riscv32 -mattr=+zhinx \
+; RUN:   -verify-machineinstrs -target-abi ilp32 | \
+; RUN:   FileCheck -check-prefixes=CHECKIZHINX %s
+; RUN: sed 's/iXLen/i64/g' %s | llc -mtriple=riscv64 -mattr=+zhinx \
+; RUN:   -verify-machineinstrs -target-abi lp64 | \
+; RUN:   FileCheck -check-prefixes=CHECKIZHINX %s
+
+declare half @llvm.minimum.f16(half, half)
+
+define half @fminimum_f16(half %a, half %b) nounwind {
+; CHECKIZFH-LABEL: fminimum_f16:
+; CHECKIZFH:       # %bb.0:
+; CHECKIZFH-NEXT:    feq.h a0, fa0, fa0
+; CHECKIZFH-NEXT:    fmv.h fa5, fa1
+; CHECKIZFH-NEXT:    beqz a0, .LBB0_3
+; CHECKIZFH-NEXT:  # %bb.1:
+; CHECKIZFH-NEXT:    feq.h a0, fa1, fa1
+; CHECKIZFH-NEXT:    beqz a0, .LBB0_4
+; CHECKIZFH-NEXT:  .LBB0_2:
+; CHECKIZFH-NEXT:    fmin.h fa0, fa0, fa5
+; CHECKIZFH-NEXT:    ret
+; CHECKIZFH-NEXT:  .LBB0_3:
+; CHECKIZFH-NEXT:    fmv.h fa5, fa0
+; CHECKIZFH-NEXT:    feq.h a0, fa1, fa1
+; CHECKIZFH-NEXT:    bnez a0, .LBB0_2
+; CHECKIZFH-NEXT:  .LBB0_4:
+; CHECKIZFH-NEXT:    fmin.h fa0, fa1, fa5
+; CHECKIZFH-NEXT:    ret
+;
+; CHECKIZHINX-LABEL: fminimum_f16:
+; CHECKIZHINX:       # %bb.0:
+; CHECKIZHINX-NEXT:    feq.h a3, a0, a0
+; CHECKIZHINX-NEXT:    mv a2, a1
+; CHECKIZHINX-NEXT:    beqz a3, .LBB0_3
+; CHECKIZHINX-NEXT:  # %bb.1:
+; CHECKIZHINX-NEXT:    feq.h a3, a1, a1
+; CHECKIZHINX-NEXT:    beqz a3, .LBB0_4
+; CHECKIZHINX-NEXT:  .LBB0_2:
+; CHECKIZHINX-NEXT:    fmin.h a0, a0, a2
+; CHECKIZHINX-NEXT:    ret
+; CHECKIZHINX-NEXT:  .LBB0_3:
+; CHECKIZHINX-NEXT:    mv a2, a0
+; CHECKIZHINX-NEXT:    feq.h a3, a1, a1
+; CHECKIZHINX-NEXT:    bnez a3, .LBB0_2
+; CHECKIZHINX-NEXT:  .LBB0_4:
+; CHECKIZHINX-NEXT:    fmin.h a0, a1, a2
+; CHECKIZHINX-NEXT:    ret
+  %1 = call half @llvm.minimum.f16(half %a, half %b)
+  ret half %1
+}
+
+declare half @llvm.maximum.f16(half, half)
+
+define half @fmaximum_f16(half %a, half %b) nounwind {
+; CHECKIZFH-LABEL: fmaximum_f16:
+; CHECKIZFH:       # %bb.0:
+; CHECKIZFH-NEXT:    feq.h a0, fa0, fa0
+; CHECKIZFH-NEXT:    fmv.h fa5, fa1
+; CHECKIZFH-NEXT:    beqz a0, .LBB1_3
+; CHECKIZFH-NEXT:  # %bb.1:
+; CHECKIZFH-NEXT:    feq.h a0, fa1, fa1
+; CHECKIZFH-NEXT:    beqz a0, .LBB1_4
+; CHECKIZFH-NEXT:  .LBB1_2:
+; CHECKIZFH-NEXT:    fmax.h fa0, fa0, fa5
+; CHECKIZFH-NEXT:    ret
+; CHECKIZFH-NEXT:  .LBB1_3:
+; CHECKIZFH-NEXT:    fmv.h fa5, fa0
+; CHECKIZFH-NEXT:    feq.h a0, fa1, fa1
+; CHECKIZFH-NEXT:    bnez a0, .LBB1_2
+; CHECKIZFH-NEXT:  .LBB1_4:
+; CHECKIZFH-NEXT:    fmax.h fa0, fa1, fa5
+; CHECKIZFH-NEXT:    ret
+;
+; CHECKIZHINX-LABEL: fmaximum_f16:
+; CHECKIZHINX:       # %bb.0:
+; CHECKIZHINX-NEXT:    feq.h a3, a0, a0
+; CHECKIZHINX-NEXT:    mv a2, a1
+; CHECKIZHINX-NEXT:    beqz a3, .LBB1_3
+; CHECKIZHINX-NEXT:  # %bb.1:
+; CHECKIZHINX-NEXT:    feq.h a3, a1, a1
+; CHECKIZHINX-NEXT:    beqz a3, .LBB1_4
+; CHECKIZHINX-NEXT:  .LBB1_2:
+; CHECKIZHINX-NEXT:    fmax.h a0, a0, a2
+; CHECKIZHINX-NEXT:    ret
+; CHECKIZHINX-NEXT:  .LBB1_3:
+; CHECKIZHINX-NEXT:    mv a2, a0
+; CHECKIZHINX-NEXT:    feq.h a3, a1, a1
+; CHECKIZHINX-NEXT:    bnez a3, .LBB1_2
+; CHECKIZHINX-NEXT:  .LBB1_4:
+; CHECKIZHINX-NEXT:    fmax.h a0, a1, a2
+; CHECKIZHINX-NEXT:    ret
+  %1 = call half @llvm.maximum.f16(half %a, half %b)
+  ret half %1
+}