math: improve sqrt for ppc64le,ppc64

The existing implementation uses code written in Go to
implement Sqrt; this adds the assembler to use the sqrt
instruction for Power and makes the necessary changes to
allow it to be inlined.

The following tests showed this relative improvement:

benchmark                 delta
BenchmarkSqrt             -97.91%
BenchmarkSqrtIndirect     -96.65%
BenchmarkSqrtGo           -35.93%
BenchmarkSqrtPrime        -96.94%

Fixes #14349

Change-Id: I8074f4dc63486e756587564ceb320aca300bf5fa
Reviewed-on: https://go-review.googlesource.com/19515Reviewed-by: Minux Ma <minux@golang.org>

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

@@ -677,7 +677,7 @@ opswitch:
 
 		if n.Left.Op == ONAME && n.Left.Sym.Name == "Sqrt" && n.Left.Sym.Pkg.Path == "math" {
 			switch Thearch.Thechar {
-			case '5', '6', '7':
+			case '5', '6', '7', '9':
 				n.Op = OSQRT
 				n.Left = n.List.First()
 				n.List.Set(nil)

src/cmd/compile/internal/ppc64/gsubr.go

@@ -706,6 +706,7 @@ func optoas(op gc.Op, t *gc.Type) obj.As {
 		OCMP_   = uint32(gc.OCMP) << 16
 		OAS_    = uint32(gc.OAS) << 16
 		OHMUL_  = uint32(gc.OHMUL) << 16
+		OSQRT_  = uint32(gc.OSQRT) << 16
 	)
 
 	a := obj.AXXX
@@ -1028,6 +1029,9 @@ func optoas(op gc.Op, t *gc.Type) obj.As {
 
 	case ODIV_ | gc.TFLOAT64:
 		a = ppc64.AFDIV
+
+	case OSQRT_ | gc.TFLOAT64:
+		a = ppc64.AFSQRT
 	}
 
 	return a

src/cmd/compile/internal/ppc64/peep.go

@@ -632,7 +632,8 @@ func copyu(p *obj.Prog, v *obj.Addr, s *obj.Addr) int {
 		ppc64.AFMOVD,
 		ppc64.AFRSP,
 		ppc64.AFNEG,
-		ppc64.AFNEGCC:
+		ppc64.AFNEGCC,
+		ppc64.AFSQRT:
 		if s != nil {
 			if copysub(&p.From, v, s, true) {
 				return 1

src/cmd/compile/internal/ppc64/prog.go

@@ -73,6 +73,7 @@ var progtable = [ppc64.ALAST & obj.AMask]obj.ProgInfo{
 	ppc64.AFCFID & obj.AMask:  {Flags: gc.SizeF | gc.LeftRead | gc.RegRead | gc.RightWrite},
 	ppc64.AFCMPU & obj.AMask:  {Flags: gc.SizeD | gc.LeftRead | gc.RightRead},
 	ppc64.AFRSP & obj.AMask:   {Flags: gc.SizeD | gc.LeftRead | gc.RightWrite | gc.Conv},
+	ppc64.AFSQRT & obj.AMask:  {Flags: gc.SizeD | gc.LeftRead | gc.RightWrite},
 
 	// Moves
 	ppc64.AMOVB & obj.AMask:  {Flags: gc.SizeB | gc.LeftRead | gc.RightWrite | gc.Move | gc.Conv},

src/math/sqrt_ppc64x.s

+// Copyright 2016 The Go Authors.  All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// +build ppc64 ppc64le
+
+#include "textflag.h"
+
+// func Sqrt(x float64) float64
+TEXT ·Sqrt(SB),NOSPLIT,$0
+	FMOVD	x+0(FP), F0
+	FSQRT	F0, F0
+	FMOVD	F0, ret+8(FP)
+	RET

src/math/stubs_ppc64x.s

@@ -84,8 +84,5 @@ TEXT ·Sin(SB),NOSPLIT,$0
 TEXT ·Cos(SB),NOSPLIT,$0
 	BR ·cos(SB)
 
-TEXT ·Sqrt(SB),NOSPLIT,$0
-	BR ·sqrt(SB)
-
 TEXT ·Tan(SB),NOSPLIT,$0
 	BR ·tan(SB)