cmd/compile: implement float min/max in hardware for amd64 and arm64

Update #59488

Change-Id: I89f5ea494cbcc887f6fae8560e57bcbd8749be86
Reviewed-on: https://go-review.googlesource.com/c/go/+/514596
Reviewed-by: Keith Randall <khr@google.com>
TryBot-Result: Gopher Robot <gobot@golang.org>
Run-TryBot: Keith Randall <khr@golang.org>
Reviewed-by: Cherry Mui <cherryyz@google.com>

11 files changed, 298 insertions, 3 deletions

diff --git a/src/cmd/compile/internal/amd64/ssa.go b/src/cmd/compile/internal/amd64/ssa.go
index 9bcca32eb4..d32ea7ec16 100644
--- a/src/cmd/compile/internal/amd64/ssa.go
+++ b/src/cmd/compile/internal/amd64/ssa.go
@@ -252,7 +252,8 @@ func ssaGenValue(s *ssagen.State, v *ssa.Value) {
 		ssa.OpAMD64RORQ, ssa.OpAMD64RORL, ssa.OpAMD64RORW, ssa.OpAMD64RORB,
 		ssa.OpAMD64ADDSS, ssa.OpAMD64ADDSD, ssa.OpAMD64SUBSS, ssa.OpAMD64SUBSD,
 		ssa.OpAMD64MULSS, ssa.OpAMD64MULSD, ssa.OpAMD64DIVSS, ssa.OpAMD64DIVSD,
-		ssa.OpAMD64PXOR,
+		ssa.OpAMD64MINSS, ssa.OpAMD64MINSD,
+		ssa.OpAMD64POR, ssa.OpAMD64PXOR,
 		ssa.OpAMD64BTSL, ssa.OpAMD64BTSQ,
 		ssa.OpAMD64BTCL, ssa.OpAMD64BTCQ,
 		ssa.OpAMD64BTRL, ssa.OpAMD64BTRQ:
diff --git a/src/cmd/compile/internal/arm64/ssa.go b/src/cmd/compile/internal/arm64/ssa.go
index a0b432bd97..27b4e881c0 100644
--- a/src/cmd/compile/internal/arm64/ssa.go
+++ b/src/cmd/compile/internal/arm64/ssa.go
@@ -215,6 +215,10 @@ func ssaGenValue(s *ssagen.State, v *ssa.Value) {
 		ssa.OpARM64FNMULD,
 		ssa.OpARM64FDIVS,
 		ssa.OpARM64FDIVD,
+		ssa.OpARM64FMINS,
+		ssa.OpARM64FMIND,
+		ssa.OpARM64FMAXS,
+		ssa.OpARM64FMAXD,
 		ssa.OpARM64ROR,
 		ssa.OpARM64RORW:
 		r := v.Reg()
diff --git a/src/cmd/compile/internal/ssa/_gen/AMD64.rules b/src/cmd/compile/internal/ssa/_gen/AMD64.rules
index 5db5deb4bb..b6937de800 100644
--- a/src/cmd/compile/internal/ssa/_gen/AMD64.rules
+++ b/src/cmd/compile/internal/ssa/_gen/AMD64.rules
@@ -172,6 +172,20 @@
 
 (Round(32|64)F ...) => (Copy ...)
 
+// Floating-point min is tricky, as the hardware op isn't right for various special
+// cases (-0 and NaN). We use two hardware ops organized just right to make the
+// result come out how we want it. See https://github.com/golang/go/issues/59488#issuecomment-1553493207
+// (although that comment isn't exactly right, as the value overwritten is not simulated correctly).
+//    t1 = MINSD x, y   => incorrect if x==NaN or x==-0,y==+0
+//    t2 = MINSD t1, x  => fixes x==NaN case
+//   res = POR t1, t2   => fixes x==-0,y==+0 case
+// Note that this trick depends on the special property that (NaN OR x) produces a NaN (although
+// it might not produce the same NaN as the input).
+(Min(64|32)F <t> x y) => (POR (MINS(D|S) <t> (MINS(D|S) <t> x y) x) (MINS(D|S) <t> x y))
+// Floating-point max is even trickier. Punt to using min instead.
+// max(x,y) == -min(-x,-y)
+(Max(64|32)F <t> x y) => (Neg(64|32)F <t> (Min(64|32)F <t> (Neg(64|32)F <t> x) (Neg(64|32)F <t> y)))
+
 (CvtBoolToUint8 ...) => (Copy ...)
 
 // Lowering shifts
diff --git a/src/cmd/compile/internal/ssa/_gen/AMD64Ops.go b/src/cmd/compile/internal/ssa/_gen/AMD64Ops.go
index 1a5bc44ca6..e9205d56c6 100644
--- a/src/cmd/compile/internal/ssa/_gen/AMD64Ops.go
+++ b/src/cmd/compile/internal/ssa/_gen/AMD64Ops.go
@@ -681,6 +681,12 @@ func init() {
 		// Any use must be preceded by a successful check of runtime.support_fma.
 		{name: "VFMADD231SD", argLength: 3, reg: fp31, resultInArg0: true, asm: "VFMADD231SD"},
 
+		// Note that these operations don't exactly match the semantics of Go's
+		// builtin min. In particular, these aren't commutative, because on various
+		// special cases the 2nd argument is preferred.
+		{name: "MINSD", argLength: 2, reg: fp21, resultInArg0: true, asm: "MINSD"}, // min(arg0,arg1)
+		{name: "MINSS", argLength: 2, reg: fp21, resultInArg0: true, asm: "MINSS"}, // min(arg0,arg1)
+
 		{name: "SBBQcarrymask", argLength: 1, reg: flagsgp, asm: "SBBQ"}, // (int64)(-1) if carry is set, 0 if carry is clear.
 		{name: "SBBLcarrymask", argLength: 1, reg: flagsgp, asm: "SBBL"}, // (int32)(-1) if carry is set, 0 if carry is clear.
 		// Note: SBBW and SBBB are subsumed by SBBL
@@ -757,7 +763,8 @@ func init() {
 		{name: "MOVLi2f", argLength: 1, reg: gpfp, typ: "Float32"}, // move 32 bits from int to float reg
 		{name: "MOVLf2i", argLength: 1, reg: fpgp, typ: "UInt32"},  // move 32 bits from float to int reg, zero extend
 
-		{name: "PXOR", argLength: 2, reg: fp21, asm: "PXOR", commutative: true, resultInArg0: true}, // exclusive or, applied to X regs for float negation.
+		{name: "PXOR", argLength: 2, reg: fp21, asm: "PXOR", commutative: true, resultInArg0: true}, // exclusive or, applied to X regs (for float negation).
+		{name: "POR", argLength: 2, reg: fp21, asm: "POR", commutative: true, resultInArg0: true},   // inclusive or, applied to X regs (for float min/max).
 
 		{name: "LEAQ", argLength: 1, reg: gp11sb, asm: "LEAQ", aux: "SymOff", rematerializeable: true, symEffect: "Addr"}, // arg0 + auxint + offset encoded in aux
 		{name: "LEAL", argLength: 1, reg: gp11sb, asm: "LEAL", aux: "SymOff", rematerializeable: true, symEffect: "Addr"}, // arg0 + auxint + offset encoded in aux
diff --git a/src/cmd/compile/internal/ssa/_gen/ARM64.rules b/src/cmd/compile/internal/ssa/_gen/ARM64.rules
index bb9ad1006d..76fc9ed256 100644
--- a/src/cmd/compile/internal/ssa/_gen/ARM64.rules
+++ b/src/cmd/compile/internal/ssa/_gen/ARM64.rules
@@ -61,6 +61,9 @@
 
 (Sqrt32 ...) => (FSQRTS ...)
 
+(Min(64|32)F ...) => (FMIN(D|S) ...)
+(Max(64|32)F ...) => (FMAX(D|S) ...)
+
 // lowering rotates
 // we do rotate detection in generic rules, if the following rules need to be changed, check generic rules first.
 (RotateLeft8  <t> x (MOVDconst [c])) => (Or8 (Lsh8x64 <t> x (MOVDconst [c&7])) (Rsh8Ux64 <t> x (MOVDconst [-c&7])))
diff --git a/src/cmd/compile/internal/ssa/_gen/ARM64Ops.go b/src/cmd/compile/internal/ssa/_gen/ARM64Ops.go
index ca8e52e210..3ded6e7a55 100644
--- a/src/cmd/compile/internal/ssa/_gen/ARM64Ops.go
+++ b/src/cmd/compile/internal/ssa/_gen/ARM64Ops.go
@@ -235,6 +235,10 @@ func init() {
 		{name: "FNEGD", argLength: 1, reg: fp11, asm: "FNEGD"},                                // -arg0, float64
 		{name: "FSQRTD", argLength: 1, reg: fp11, asm: "FSQRTD"},                              // sqrt(arg0), float64
 		{name: "FSQRTS", argLength: 1, reg: fp11, asm: "FSQRTS"},                              // sqrt(arg0), float32
+		{name: "FMIND", argLength: 2, reg: fp21, asm: "FMIND"},                                // min(arg0, arg1)
+		{name: "FMINS", argLength: 2, reg: fp21, asm: "FMINS"},                                // min(arg0, arg1)
+		{name: "FMAXD", argLength: 2, reg: fp21, asm: "FMAXD"},                                // max(arg0, arg1)
+		{name: "FMAXS", argLength: 2, reg: fp21, asm: "FMAXS"},                                // max(arg0, arg1)
 		{name: "REV", argLength: 1, reg: gp11, asm: "REV"},                                    // byte reverse, 64-bit
 		{name: "REVW", argLength: 1, reg: gp11, asm: "REVW"},                                  // byte reverse, 32-bit
 		{name: "REV16", argLength: 1, reg: gp11, asm: "REV16"},                                // byte reverse in each 16-bit halfword, 64-bit
diff --git a/src/cmd/compile/internal/ssa/_gen/genericOps.go b/src/cmd/compile/internal/ssa/_gen/genericOps.go
index 53ff57f6b1..fb18319263 100644
--- a/src/cmd/compile/internal/ssa/_gen/genericOps.go
+++ b/src/cmd/compile/internal/ssa/_gen/genericOps.go
@@ -285,6 +285,12 @@ var genericOps = []opData{
 	{name: "Abs", argLength: 1},      // absolute value arg0
 	{name: "Copysign", argLength: 2}, // copy sign from arg0 to arg1
 
+	// Float min/max implementation, if hardware is available.
+	{name: "Min64F", argLength: 2}, // min(arg0,arg1)
+	{name: "Min32F", argLength: 2}, // min(arg0,arg1)
+	{name: "Max64F", argLength: 2}, // max(arg0,arg1)
+	{name: "Max32F", argLength: 2}, // max(arg0,arg1)
+
 	// 3-input opcode.
 	// Fused-multiply-add, float64 only.
 	// When a*b+c is exactly zero (before rounding), then the result is +0 or -0.
diff --git a/src/cmd/compile/internal/ssa/opGen.go b/src/cmd/compile/internal/ssa/opGen.go
index a495a04752..64aea38afe 100644
--- a/src/cmd/compile/internal/ssa/opGen.go
+++ b/src/cmd/compile/internal/ssa/opGen.go
@@ -912,6 +912,8 @@ const (
 	OpAMD64SQRTSS
 	OpAMD64ROUNDSD
 	OpAMD64VFMADD231SD
+	OpAMD64MINSD
+	OpAMD64MINSS
 	OpAMD64SBBQcarrymask
 	OpAMD64SBBLcarrymask
 	OpAMD64SETEQ
@@ -974,6 +976,7 @@ const (
 	OpAMD64MOVLi2f
 	OpAMD64MOVLf2i
 	OpAMD64PXOR
+	OpAMD64POR
 	OpAMD64LEAQ
 	OpAMD64LEAL
 	OpAMD64LEAW
@@ -1451,6 +1454,10 @@ const (
 	OpARM64FNEGD
 	OpARM64FSQRTD
 	OpARM64FSQRTS
+	OpARM64FMIND
+	OpARM64FMINS
+	OpARM64FMAXD
+	OpARM64FMAXS
 	OpARM64REV
 	OpARM64REVW
 	OpARM64REV16
@@ -3016,6 +3023,10 @@ const (
 	OpRoundToEven
 	OpAbs
 	OpCopysign
+	OpMin64F
+	OpMin32F
+	OpMax64F
+	OpMax32F
 	OpFMA
 	OpPhi
 	OpCopy
@@ -11901,6 +11912,36 @@ var opcodeTable = [...]opInfo{
 		},
 	},
 	{
+		name:         "MINSD",
+		argLen:       2,
+		resultInArg0: true,
+		asm:          x86.AMINSD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
+				{1, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
+			},
+			outputs: []outputInfo{
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
+			},
+		},
+	},
+	{
+		name:         "MINSS",
+		argLen:       2,
+		resultInArg0: true,
+		asm:          x86.AMINSS,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
+				{1, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
+			},
+			outputs: []outputInfo{
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
+			},
+		},
+	},
+	{
 		name:   "SBBQcarrymask",
 		argLen: 1,
 		asm:    x86.ASBBQ,
@@ -12671,6 +12712,22 @@ var opcodeTable = [...]opInfo{
 		},
 	},
 	{
+		name:         "POR",
+		argLen:       2,
+		commutative:  true,
+		resultInArg0: true,
+		asm:          x86.APOR,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
+				{1, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
+			},
+			outputs: []outputInfo{
+				{0, 2147418112}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14
+			},
+		},
+	},
+	{
 		name:              "LEAQ",
 		auxType:           auxSymOff,
 		argLen:            1,
@@ -19438,6 +19495,62 @@ var opcodeTable = [...]opInfo{
 		},
 	},
 	{
+		name:   "FMIND",
+		argLen: 2,
+		asm:    arm64.AFMIND,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{0, 9223372034707292160}, // F0 F1 F2 F3 F4 F5 F6 F7 F8 F9 F10 F11 F12 F13 F14 F15 F16 F17 F18 F19 F20 F21 F22 F23 F24 F25 F26 F27 F28 F29 F30 F31
+				{1, 9223372034707292160}, // F0 F1 F2 F3 F4 F5 F6 F7 F8 F9 F10 F11 F12 F13 F14 F15 F16 F17 F18 F19 F20 F21 F22 F23 F24 F25 F26 F27 F28 F29 F30 F31
+			},
+			outputs: []outputInfo{
+				{0, 9223372034707292160}, // F0 F1 F2 F3 F4 F5 F6 F7 F8 F9 F10 F11 F12 F13 F14 F15 F16 F17 F18 F19 F20 F21 F22 F23 F24 F25 F26 F27 F28 F29 F30 F31
+			},
+		},
+	},
+	{
+		name:   "FMINS",
+		argLen: 2,
+		asm:    arm64.AFMINS,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{0, 9223372034707292160}, // F0 F1 F2 F3 F4 F5 F6 F7 F8 F9 F10 F11 F12 F13 F14 F15 F16 F17 F18 F19 F20 F21 F22 F23 F24 F25 F26 F27 F28 F29 F30 F31
+				{1, 9223372034707292160}, // F0 F1 F2 F3 F4 F5 F6 F7 F8 F9 F10 F11 F12 F13 F14 F15 F16 F17 F18 F19 F20 F21 F22 F23 F24 F25 F26 F27 F28 F29 F30 F31
+			},
+			outputs: []outputInfo{
+				{0, 9223372034707292160}, // F0 F1 F2 F3 F4 F5 F6 F7 F8 F9 F10 F11 F12 F13 F14 F15 F16 F17 F18 F19 F20 F21 F22 F23 F24 F25 F26 F27 F28 F29 F30 F31
+			},
+		},
+	},
+	{
+		name:   "FMAXD",
+		argLen: 2,
+		asm:    arm64.AFMAXD,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{0, 9223372034707292160}, // F0 F1 F2 F3 F4 F5 F6 F7 F8 F9 F10 F11 F12 F13 F14 F15 F16 F17 F18 F19 F20 F21 F22 F23 F24 F25 F26 F27 F28 F29 F30 F31
+				{1, 9223372034707292160}, // F0 F1 F2 F3 F4 F5 F6 F7 F8 F9 F10 F11 F12 F13 F14 F15 F16 F17 F18 F19 F20 F21 F22 F23 F24 F25 F26 F27 F28 F29 F30 F31
+			},
+			outputs: []outputInfo{
+				{0, 9223372034707292160}, // F0 F1 F2 F3 F4 F5 F6 F7 F8 F9 F10 F11 F12 F13 F14 F15 F16 F17 F18 F19 F20 F21 F22 F23 F24 F25 F26 F27 F28 F29 F30 F31
+			},
+		},
+	},
+	{
+		name:   "FMAXS",
+		argLen: 2,
+		asm:    arm64.AFMAXS,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{0, 9223372034707292160}, // F0 F1 F2 F3 F4 F5 F6 F7 F8 F9 F10 F11 F12 F13 F14 F15 F16 F17 F18 F19 F20 F21 F22 F23 F24 F25 F26 F27 F28 F29 F30 F31
+				{1, 9223372034707292160}, // F0 F1 F2 F3 F4 F5 F6 F7 F8 F9 F10 F11 F12 F13 F14 F15 F16 F17 F18 F19 F20 F21 F22 F23 F24 F25 F26 F27 F28 F29 F30 F31
+			},
+			outputs: []outputInfo{
+				{0, 9223372034707292160}, // F0 F1 F2 F3 F4 F5 F6 F7 F8 F9 F10 F11 F12 F13 F14 F15 F16 F17 F18 F19 F20 F21 F22 F23 F24 F25 F26 F27 F28 F29 F30 F31
+			},
+		},
+	},
+	{
 		name:   "REV",
 		argLen: 1,
 		asm:    arm64.AREV,
@@ -39083,6 +39196,26 @@ var opcodeTable = [...]opInfo{
 		generic: true,
 	},
 	{
+		name:    "Min64F",
+		argLen:  2,
+		generic: true,
+	},
+	{
+		name:    "Min32F",
+		argLen:  2,
+		generic: true,
+	},
+	{
+		name:    "Max64F",
+		argLen:  2,
+		generic: true,
+	},
+	{
+		name:    "Max32F",
+		argLen:  2,
+		generic: true,
+	},
+	{
 		name:    "FMA",
 		argLen:  3,
 		generic: true,
diff --git a/src/cmd/compile/internal/ssa/rewriteAMD64.go b/src/cmd/compile/internal/ssa/rewriteAMD64.go
index 2d4a886ea5..afe9ed257a 100644
--- a/src/cmd/compile/internal/ssa/rewriteAMD64.go
+++ b/src/cmd/compile/internal/ssa/rewriteAMD64.go
@@ -871,6 +871,14 @@ func rewriteValueAMD64(v *Value) bool {
 		return rewriteValueAMD64_OpLsh8x64(v)
 	case OpLsh8x8:
 		return rewriteValueAMD64_OpLsh8x8(v)
+	case OpMax32F:
+		return rewriteValueAMD64_OpMax32F(v)
+	case OpMax64F:
+		return rewriteValueAMD64_OpMax64F(v)
+	case OpMin32F:
+		return rewriteValueAMD64_OpMin32F(v)
+	case OpMin64F:
+		return rewriteValueAMD64_OpMin64F(v)
 	case OpMod16:
 		return rewriteValueAMD64_OpMod16(v)
 	case OpMod16u:
@@ -27352,6 +27360,88 @@ func rewriteValueAMD64_OpLsh8x8(v *Value) bool {
 	}
 	return false
 }
+func rewriteValueAMD64_OpMax32F(v *Value) bool {
+	v_1 := v.Args[1]
+	v_0 := v.Args[0]
+	b := v.Block
+	// match: (Max32F <t> x y)
+	// result: (Neg32F <t> (Min32F <t> (Neg32F <t> x) (Neg32F <t> y)))
+	for {
+		t := v.Type
+		x := v_0
+		y := v_1
+		v.reset(OpNeg32F)
+		v.Type = t
+		v0 := b.NewValue0(v.Pos, OpMin32F, t)
+		v1 := b.NewValue0(v.Pos, OpNeg32F, t)
+		v1.AddArg(x)
+		v2 := b.NewValue0(v.Pos, OpNeg32F, t)
+		v2.AddArg(y)
+		v0.AddArg2(v1, v2)
+		v.AddArg(v0)
+		return true
+	}
+}
+func rewriteValueAMD64_OpMax64F(v *Value) bool {
+	v_1 := v.Args[1]
+	v_0 := v.Args[0]
+	b := v.Block
+	// match: (Max64F <t> x y)
+	// result: (Neg64F <t> (Min64F <t> (Neg64F <t> x) (Neg64F <t> y)))
+	for {
+		t := v.Type
+		x := v_0
+		y := v_1
+		v.reset(OpNeg64F)
+		v.Type = t
+		v0 := b.NewValue0(v.Pos, OpMin64F, t)
+		v1 := b.NewValue0(v.Pos, OpNeg64F, t)
+		v1.AddArg(x)
+		v2 := b.NewValue0(v.Pos, OpNeg64F, t)
+		v2.AddArg(y)
+		v0.AddArg2(v1, v2)
+		v.AddArg(v0)
+		return true
+	}
+}
+func rewriteValueAMD64_OpMin32F(v *Value) bool {
+	v_1 := v.Args[1]
+	v_0 := v.Args[0]
+	b := v.Block
+	// match: (Min32F <t> x y)
+	// result: (POR (MINSS <t> (MINSS <t> x y) x) (MINSS <t> x y))
+	for {
+		t := v.Type
+		x := v_0
+		y := v_1
+		v.reset(OpAMD64POR)
+		v0 := b.NewValue0(v.Pos, OpAMD64MINSS, t)
+		v1 := b.NewValue0(v.Pos, OpAMD64MINSS, t)
+		v1.AddArg2(x, y)
+		v0.AddArg2(v1, x)
+		v.AddArg2(v0, v1)
+		return true
+	}
+}
+func rewriteValueAMD64_OpMin64F(v *Value) bool {
+	v_1 := v.Args[1]
+	v_0 := v.Args[0]
+	b := v.Block
+	// match: (Min64F <t> x y)
+	// result: (POR (MINSD <t> (MINSD <t> x y) x) (MINSD <t> x y))
+	for {
+		t := v.Type
+		x := v_0
+		y := v_1
+		v.reset(OpAMD64POR)
+		v0 := b.NewValue0(v.Pos, OpAMD64MINSD, t)
+		v1 := b.NewValue0(v.Pos, OpAMD64MINSD, t)
+		v1.AddArg2(x, y)
+		v0.AddArg2(v1, x)
+		v.AddArg2(v0, v1)
+		return true
+	}
+}
 func rewriteValueAMD64_OpMod16(v *Value) bool {
 	v_1 := v.Args[1]
 	v_0 := v.Args[0]
diff --git a/src/cmd/compile/internal/ssa/rewriteARM64.go b/src/cmd/compile/internal/ssa/rewriteARM64.go
index b655c62720..e9b4749fbc 100644
--- a/src/cmd/compile/internal/ssa/rewriteARM64.go
+++ b/src/cmd/compile/internal/ssa/rewriteARM64.go
@@ -816,6 +816,18 @@ func rewriteValueARM64(v *Value) bool {
 		return rewriteValueARM64_OpLsh8x64(v)
 	case OpLsh8x8:
 		return rewriteValueARM64_OpLsh8x8(v)
+	case OpMax32F:
+		v.Op = OpARM64FMAXS
+		return true
+	case OpMax64F:
+		v.Op = OpARM64FMAXD
+		return true
+	case OpMin32F:
+		v.Op = OpARM64FMINS
+		return true
+	case OpMin64F:
+		v.Op = OpARM64FMIND
+		return true
 	case OpMod16:
 		return rewriteValueARM64_OpMod16(v)
 	case OpMod16u:
diff --git a/src/cmd/compile/internal/ssagen/ssa.go b/src/cmd/compile/internal/ssagen/ssa.go
index 03f9958098..9796978f4a 100644
--- a/src/cmd/compile/internal/ssagen/ssa.go
+++ b/src/cmd/compile/internal/ssagen/ssa.go
@@ -3567,11 +3567,32 @@ func (s *state) minMax(n *ir.CallExpr) *ssa.Value {
 
 	if typ.IsFloat() || typ.IsString() {
 		// min/max semantics for floats are tricky because of NaNs and
-		// negative zero, so we let the runtime handle this instead.
+		// negative zero. Some architectures have instructions which
+		// we can use to generate the right result. For others we must
+		// call into the runtime instead.
 		//
 		// Strings are conceptually simpler, but we currently desugar
 		// string comparisons during walk, not ssagen.
 
+		if typ.IsFloat() {
+			switch Arch.LinkArch.Family {
+			case sys.AMD64, sys.ARM64:
+				var op ssa.Op
+				switch {
+				case typ.Kind() == types.TFLOAT64 && n.Op() == ir.OMIN:
+					op = ssa.OpMin64F
+				case typ.Kind() == types.TFLOAT64 && n.Op() == ir.OMAX:
+					op = ssa.OpMax64F
+				case typ.Kind() == types.TFLOAT32 && n.Op() == ir.OMIN:
+					op = ssa.OpMin32F
+				case typ.Kind() == types.TFLOAT32 && n.Op() == ir.OMAX:
+					op = ssa.OpMax32F
+				}
+				return fold(func(x, a *ssa.Value) *ssa.Value {
+					return s.newValue2(op, typ, x, a)
+				})
+			}
+		}
 		var name string
 		switch typ.Kind() {
 		case types.TFLOAT32: