From: Lynn Boger <laboger@linux.vnet.ibm.com>
Date: Wed, 26 Sep 2018 14:43:15 +0000 (-0400)
Subject: cmd/internal/obj/ppc64: generate float 0 more efficiently on ppc64x
X-Git-Tag: go1.12beta1~832
X-Git-Url: http://www.git.cypherpunks.su/?a=commitdiff_plain;h=f776d51bf7fe90926f8b929aee65d802a9091d5e;p=gostls13.git

cmd/internal/obj/ppc64: generate float 0 more efficiently on ppc64x

This change makes use of a VSX instruction to generate the
float 0 value instead of generating a constant in memory and
loading it from there.

This uses 1 instruction instead of 2 and avoids a memory reference.
in the +0 case, uses 2 instructions in the -0 case but avoids
the memory reference.

Since this is done in the assembler for ppc64x, an update has
been made to the assembler test.

Change-Id: Ief7dddcb057bfb602f78215f6947664e8c841464
Reviewed-on: https://go-review.googlesource.com/c/139420
Reviewed-by: Michael Munday <mike.munday@ibm.com>
---

diff --git a/src/cmd/asm/internal/asm/testdata/ppc64enc.s b/src/cmd/asm/internal/asm/testdata/ppc64enc.s
index 7ab1a578f8..0133a85b98 100644
--- a/src/cmd/asm/internal/asm/testdata/ppc64enc.s
+++ b/src/cmd/asm/internal/asm/testdata/ppc64enc.s
@@ -98,4 +98,7 @@ TEXT asmtest(SB),DUPOK|NOSPLIT,$0
 	LDAR (R4),$0,R5                 // 7ca020a8
 	LDAR (R3),R5                    // 7ca018a8
 
+	// float constants
+	FMOVD $(0.0), F1                // f0210cd0
+	FMOVD $(-0.0), F1               // f0210cd0fc200850
 	RET
diff --git a/src/cmd/internal/obj/ppc64/asm9.go b/src/cmd/internal/obj/ppc64/asm9.go
index 756170bc55..66a77b308c 100644
--- a/src/cmd/internal/obj/ppc64/asm9.go
+++ b/src/cmd/internal/obj/ppc64/asm9.go
@@ -35,6 +35,7 @@ import (
 	"encoding/binary"
 	"fmt"
 	"log"
+	"math"
 	"sort"
 )
 
@@ -342,6 +343,8 @@ var optab = []Optab{
 	{AFMOVD, C_LEXT, C_NONE, C_NONE, C_FREG, 36, 8, REGSB},
 	{AFMOVD, C_LAUTO, C_NONE, C_NONE, C_FREG, 36, 8, REGSP},
 	{AFMOVD, C_LOREG, C_NONE, C_NONE, C_FREG, 36, 8, REGZERO},
+	{AFMOVD, C_ZCON, C_NONE, C_NONE, C_FREG, 24, 4, 0},
+	{AFMOVD, C_ADDCON, C_NONE, C_NONE, C_FREG, 24, 8, 0},
 	{AFMOVD, C_ADDR, C_NONE, C_NONE, C_FREG, 75, 8, 0},
 	{AFMOVD, C_FREG, C_NONE, C_NONE, C_SEXT, 7, 4, REGSB},
 	{AFMOVD, C_FREG, C_NONE, C_NONE, C_SAUTO, 7, 4, REGSP},
@@ -829,6 +832,18 @@ func (c *ctxt9) aclass(a *obj.Addr) int {
 	case obj.TYPE_TEXTSIZE:
 		return C_TEXTSIZE
 
+	case obj.TYPE_FCONST:
+		// The only cases where FCONST will occur are with float64 +/- 0.
+		// All other float constants are generated in memory.
+		f64 := a.Val.(float64)
+		if f64 == 0 {
+			if math.Signbit(f64) {
+				return C_ADDCON
+			}
+			return C_ZCON
+		}
+		log.Fatalf("Unexpected nonzero FCONST operand %v", a)
+
 	case obj.TYPE_CONST,
 		obj.TYPE_ADDR:
 		switch a.Name {
@@ -2763,6 +2778,13 @@ func (c *ctxt9) asmout(p *obj.Prog, o *Optab, out []uint32) {
 			c.ctxt.Diag("%v is not supported", p)
 		}
 
+	case 24: /* lfd fA,float64(0) -> xxlxor xsA,xsaA,xsaA + fneg for -0 */
+		o1 = AOP_XX3I(c.oprrr(AXXLXOR), uint32(p.To.Reg), uint32(p.To.Reg), uint32(p.To.Reg), uint32(0))
+		// This is needed for -0.
+		if o.size == 8 {
+			o2 = AOP_RRR(c.oprrr(AFNEG), uint32(p.To.Reg), 0, uint32(p.To.Reg))
+		}
+
 	case 25:
 		/* sld[.] $sh,rS,rA -> rldicr[.] $sh,rS,mask(0,63-sh),rA; srd[.] -> rldicl */
 		v := c.regoff(&p.From)
diff --git a/src/cmd/internal/obj/ppc64/obj9.go b/src/cmd/internal/obj/ppc64/obj9.go
index f42d675805..7a07b5058c 100644
--- a/src/cmd/internal/obj/ppc64/obj9.go
+++ b/src/cmd/internal/obj/ppc64/obj9.go
@@ -67,10 +67,13 @@ func progedit(ctxt *obj.Link, p *obj.Prog, newprog obj.ProgAlloc) {
 	case AFMOVD:
 		if p.From.Type == obj.TYPE_FCONST {
 			f64 := p.From.Val.(float64)
-			p.From.Type = obj.TYPE_MEM
-			p.From.Sym = ctxt.Float64Sym(f64)
-			p.From.Name = obj.NAME_EXTERN
-			p.From.Offset = 0
+			// Constant not needed in memory for float +/- 0
+			if f64 != 0 {
+				p.From.Type = obj.TYPE_MEM
+				p.From.Sym = ctxt.Float64Sym(f64)
+				p.From.Name = obj.NAME_EXTERN
+				p.From.Offset = 0
+			}
 		}
 
 		// Put >32-bit constants in memory and load them