cmd/compile,math: improve code generation for math.Abs

Implement int reg <-> fp reg moves on amd64.
If we see a load to int reg followed by an int->fp move, then we can just
load to the fp reg instead.  Same for stores.

math.Abs is now:

MOVQ	"".x+8(SP), AX
SHLQ	$1, AX
SHRQ	$1, AX
MOVQ	AX, "".~r1+16(SP)

math.Copysign is now:

MOVQ	"".x+8(SP), AX
SHLQ	$1, AX
SHRQ	$1, AX
MOVQ	"".y+16(SP), CX
SHRQ	$63, CX
SHLQ	$63, CX
ORQ	CX, AX
MOVQ	AX, "".~r2+24(SP)

math.Float64bits is now:

MOVSD	"".x+8(SP), X0
MOVSD	X0, "".~r1+16(SP)
(it would be nicer to use a non-SSE reg for this, nothing is perfect)

And due to the fix for #21440, the inlined version of these improve as well.

name      old time/op  new time/op  delta
Abs       1.38ns ± 5%  0.89ns ±10%  -35.54%  (p=0.000 n=10+10)
Copysign  1.56ns ± 7%  1.35ns ± 6%  -13.77%  (p=0.000 n=9+10)

Fixes #13095

Change-Id: Ibd7f2792412a6668608780b0688a77062e1f1499
Reviewed-on: https://go-review.googlesource.com/58732
Run-TryBot: Keith Randall <khr@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Cherry Zhang <cherryyz@google.com>
Reviewed-by: Ilya Tocar <ilya.tocar@intel.com>

src/cmd/compile/internal/amd64/ssa.go

@@ -678,6 +678,18 @@ func ssaGenValue(s *gc.SSAGenState, v *ssa.Value) {
 		// Break false dependency on destination register.
 		opregreg(s, x86.AXORPS, r, r)
 		opregreg(s, v.Op.Asm(), r, v.Args[0].Reg())
+	case ssa.OpAMD64MOVQi2f, ssa.OpAMD64MOVQf2i:
+		p := s.Prog(x86.AMOVQ)
+		p.From.Type = obj.TYPE_REG
+		p.From.Reg = v.Args[0].Reg()
+		p.To.Type = obj.TYPE_REG
+		p.To.Reg = v.Reg()
+	case ssa.OpAMD64MOVLi2f, ssa.OpAMD64MOVLf2i:
+		p := s.Prog(x86.AMOVL)
+		p.From.Type = obj.TYPE_REG
+		p.From.Reg = v.Args[0].Reg()
+		p.To.Type = obj.TYPE_REG
+		p.To.Reg = v.Reg()
 	case ssa.OpAMD64ADDQmem, ssa.OpAMD64ADDLmem, ssa.OpAMD64SUBQmem, ssa.OpAMD64SUBLmem,
 		ssa.OpAMD64ANDQmem, ssa.OpAMD64ANDLmem, ssa.OpAMD64ORQmem, ssa.OpAMD64ORLmem,
 		ssa.OpAMD64XORQmem, ssa.OpAMD64XORLmem, ssa.OpAMD64ADDSDmem, ssa.OpAMD64ADDSSmem,

src/cmd/compile/internal/gc/asm_test.go

@@ -214,7 +214,7 @@ var allAsmTests = []*asmTests{
 	{
 		arch:    "amd64",
 		os:      "linux",
-		imports: []string{"encoding/binary", "math/bits", "unsafe"},
+		imports: []string{"encoding/binary", "math", "math/bits", "unsafe"},
 		tests:   linuxAMD64Tests,
 	},
 	{
@@ -990,6 +990,57 @@ var linuxAMD64Tests = []*asmTest{
 		`,
 		[]string{"TEXT\t.*, [$]0-8"},
 	},
+	// math.Abs using integer registers
+	{
+		`
+		func $(x float64) float64 {
+			return math.Abs(x)
+		}
+		`,
+		[]string{"\tSHLQ\t[$]1,", "\tSHRQ\t[$]1,"},
+	},
+	// math.Copysign using integer registers
+	{
+		`
+		func $(x, y float64) float64 {
+			return math.Copysign(x, y)
+		}
+		`,
+		[]string{"\tSHLQ\t[$]1,", "\tSHRQ\t[$]1,", "\tSHRQ\t[$]63,", "\tSHLQ\t[$]63,", "\tORQ\t"},
+	},
+	// int <-> fp moves
+	{
+		`
+		func $(x float64) uint64 {
+			return math.Float64bits(x+1) + 1
+		}
+		`,
+		[]string{"\tMOVQ\tX.*, [^X].*"},
+	},
+	{
+		`
+		func $(x float32) uint32 {
+			return math.Float32bits(x+1) + 1
+		}
+		`,
+		[]string{"\tMOVL\tX.*, [^X].*"},
+	},
+	{
+		`
+		func $(x uint64) float64 {
+			return math.Float64frombits(x+1) + 1
+		}
+		`,
+		[]string{"\tMOVQ\t[^X].*, X.*"},
+	},
+	{
+		`
+		func $(x uint32) float32 {
+			return math.Float32frombits(x+1) + 1
+		}
+		`,
+		[]string{"\tMOVL\t[^X].*, X.*"},
+	},
 }
 
 var linux386Tests = []*asmTest{

src/cmd/compile/internal/ssa/gen/AMD64.rules

@@ -2456,3 +2456,47 @@
 (MOVLstore [off] {sym} ptr a:(ADDLconst [c] l:(MOVLload [off] {sym} ptr2 mem)) mem)
 	&& isSamePtr(ptr, ptr2) && a.Uses == 1 && l.Uses == 1 && validValAndOff(c,off) ->
 	(ADDLconstmem {sym} [makeValAndOff(c,off)] ptr mem)
+
+// float <-> int register moves, with no conversion.
+// These come up when compiling math.{Float{32,64}bits,Float{32,64}frombits}.
+(MOVQload  [off] {sym} ptr (MOVSDstore [off] {sym} ptr val _)) -> (MOVQf2i val)
+(MOVLload  [off] {sym} ptr (MOVSSstore [off] {sym} ptr val _)) -> (MOVLf2i val)
+(MOVSDload [off] {sym} ptr (MOVQstore  [off] {sym} ptr val _)) -> (MOVQi2f val)
+(MOVSSload [off] {sym} ptr (MOVLstore  [off] {sym} ptr val _)) -> (MOVLi2f val)
+
+// Other load-like ops.
+(ADDQmem x [off] {sym} ptr (MOVSDstore [off] {sym} ptr y _)) -> (ADDQ x (MOVQf2i y))
+(ADDLmem x [off] {sym} ptr (MOVSSstore [off] {sym} ptr y _)) -> (ADDL x (MOVLf2i y))
+(SUBQmem x [off] {sym} ptr (MOVSDstore [off] {sym} ptr y _)) -> (SUBQ x (MOVQf2i y))
+(SUBLmem x [off] {sym} ptr (MOVSSstore [off] {sym} ptr y _)) -> (SUBL x (MOVLf2i y))
+(ANDQmem x [off] {sym} ptr (MOVSDstore [off] {sym} ptr y _)) -> (ANDQ x (MOVQf2i y))
+(ANDLmem x [off] {sym} ptr (MOVSSstore [off] {sym} ptr y _)) -> (ANDL x (MOVLf2i y))
+( ORQmem x [off] {sym} ptr (MOVSDstore [off] {sym} ptr y _)) -> ( ORQ x (MOVQf2i y))
+( ORLmem x [off] {sym} ptr (MOVSSstore [off] {sym} ptr y _)) -> ( ORL x (MOVLf2i y))
+(XORQmem x [off] {sym} ptr (MOVSDstore [off] {sym} ptr y _)) -> (XORQ x (MOVQf2i y))
+(XORLmem x [off] {sym} ptr (MOVSSstore [off] {sym} ptr y _)) -> (XORL x (MOVLf2i y))
+
+(ADDQconstmem [valOff] {sym} ptr (MOVSDstore [ValAndOff(valOff).Off()] {sym} ptr x _)) ->
+  (ADDQconst [ValAndOff(valOff).Val()] (MOVQf2i x))
+(ADDLconstmem [valOff] {sym} ptr (MOVSSstore [ValAndOff(valOff).Off()] {sym} ptr x _)) ->
+  (ADDLconst [ValAndOff(valOff).Val()] (MOVLf2i x))
+
+(ADDSDmem x [off] {sym} ptr (MOVQstore [off] {sym} ptr y _)) -> (ADDSD x (MOVQi2f y))
+(ADDSSmem x [off] {sym} ptr (MOVLstore [off] {sym} ptr y _)) -> (ADDSS x (MOVLi2f y))
+(SUBSDmem x [off] {sym} ptr (MOVQstore [off] {sym} ptr y _)) -> (SUBSD x (MOVQi2f y))
+(SUBSSmem x [off] {sym} ptr (MOVLstore [off] {sym} ptr y _)) -> (SUBSS x (MOVLi2f y))
+(MULSDmem x [off] {sym} ptr (MOVQstore [off] {sym} ptr y _)) -> (MULSD x (MOVQi2f y))
+(MULSSmem x [off] {sym} ptr (MOVLstore [off] {sym} ptr y _)) -> (MULSS x (MOVLi2f y))
+
+// Redirect stores to use the other register set.
+(MOVQstore  [off] {sym} ptr (MOVQf2i val) mem) -> (MOVSDstore [off] {sym} ptr val mem)
+(MOVLstore  [off] {sym} ptr (MOVLf2i val) mem) -> (MOVSSstore [off] {sym} ptr val mem)
+(MOVSDstore [off] {sym} ptr (MOVQi2f val) mem) -> (MOVQstore  [off] {sym} ptr val mem)
+(MOVSSstore [off] {sym} ptr (MOVLi2f val) mem) -> (MOVLstore  [off] {sym} ptr val mem)
+
+// Load args directly into the register class where it will be used.
+// We do this by just modifying the type of the Arg.
+(MOVQf2i <t> (Arg [off] {sym})) -> @b.Func.Entry (Arg <t> [off] {sym})
+(MOVLf2i <t> (Arg [off] {sym})) -> @b.Func.Entry (Arg <t> [off] {sym})
+(MOVQi2f <t> (Arg [off] {sym})) -> @b.Func.Entry (Arg <t> [off] {sym})
+(MOVLi2f <t> (Arg [off] {sym})) -> @b.Func.Entry (Arg <t> [off] {sym})

src/cmd/compile/internal/ssa/gen/AMD64Ops.go

@@ -386,6 +386,13 @@ func init() {
 		{name: "CVTSD2SS", argLength: 1, reg: fp11, asm: "CVTSD2SS"},   // convert float64 to float32
 		{name: "CVTSS2SD", argLength: 1, reg: fp11, asm: "CVTSS2SD"},   // convert float32 to float64
 
+		// Move values between int and float registers, with no conversion.
+		// TODO: should we have generic versions of these?
+		{name: "MOVQi2f", argLength: 1, reg: gpfp, typ: "Float64"}, // move 64 bits from int to float reg
+		{name: "MOVQf2i", argLength: 1, reg: fpgp, typ: "UInt64"},  // move 64 bits from float to int reg
+		{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
+
 		{name: "PXOR", argLength: 2, reg: fp21, asm: "PXOR", commutative: true, resultInArg0: true}, // exclusive or, applied to X regs for float negation.
 
 		{name: "LEAQ", argLength: 1, reg: gp11sb, asm: "LEAQ", aux: "SymOff", rematerializeable: true, symEffect: "Addr"}, // arg0 + auxint + offset encoded in aux

src/cmd/compile/internal/ssa/opGen.go

@@ -591,6 +591,10 @@ const (
 	OpAMD64CVTSQ2SD
 	OpAMD64CVTSD2SS
 	OpAMD64CVTSS2SD
+	OpAMD64MOVQi2f
+	OpAMD64MOVQf2i
+	OpAMD64MOVLi2f
+	OpAMD64MOVLf2i
 	OpAMD64PXOR
 	OpAMD64LEAQ
 	OpAMD64LEAQ1
@@ -7017,6 +7021,54 @@ var opcodeTable = [...]opInfo{
 			},
 		},
 	},
+	{
+		name:   "MOVQi2f",
+		argLen: 1,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{0, 65519}, // AX CX DX BX BP SI DI R8 R9 R10 R11 R12 R13 R14 R15
+			},
+			outputs: []outputInfo{
+				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+			},
+		},
+	},
+	{
+		name:   "MOVQf2i",
+		argLen: 1,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+			},
+			outputs: []outputInfo{
+				{0, 65519}, // AX CX DX BX BP SI DI R8 R9 R10 R11 R12 R13 R14 R15
+			},
+		},
+	},
+	{
+		name:   "MOVLi2f",
+		argLen: 1,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{0, 65519}, // AX CX DX BX BP SI DI R8 R9 R10 R11 R12 R13 R14 R15
+			},
+			outputs: []outputInfo{
+				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+			},
+		},
+	},
+	{
+		name:   "MOVLf2i",
+		argLen: 1,
+		reg: regInfo{
+			inputs: []inputInfo{
+				{0, 4294901760}, // X0 X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 X12 X13 X14 X15
+			},
+			outputs: []outputInfo{
+				{0, 65519}, // AX CX DX BX BP SI DI R8 R9 R10 R11 R12 R13 R14 R15
+			},
+		},
+	},
 	{
 		name:         "PXOR",
 		argLen:       2,

src/math/abs.go

@@ -10,14 +10,5 @@ package math
 //	Abs(±Inf) = +Inf
 //	Abs(NaN) = NaN
 func Abs(x float64) float64 {
-	// TODO: once golang.org/issue/13095 is fixed, change this to:
+	return Float64frombits(Float64bits(x) &^ (1 << 63))
-	// return Float64frombits(Float64bits(x) &^ (1 << 63))
-	// But for now, this generates better code and can also be inlined:
-	if x < 0 {
-		return -x
-	}
-	if x == 0 {
-		return 0 // return correctly abs(-0)
-	}
-	return x
 }