cmd/internal/obj/x86: faster Assemble for non-NaCl hosts

Make span6 function (used as LinkArch.Assemble) faster
by avoiding redundant re-assemble rounds on hosts
that are not NaCl.

NaCl is excluded because it needs Prog.Isize to fix alignment.

For make.bash, there are around 50% of functions that can
be encoded in a single trip. With this change, those function
will be assembled with 1 round instead of 2.

compilebench results:

    name        old time/op       new time/op       delta
    Template          305ms ± 2%        299ms ± 2%  -1.99%  (p=0.001 n=10+10)
    Unicode           139ms ± 3%        138ms ± 4%    ~     (p=0.222 n=9+9)
    GoTypes           1.05s ± 1%        1.04s ± 1%  -1.34%  (p=0.000 n=10+9)
    Compiler          4.78s ± 1%        4.71s ± 1%  -1.45%  (p=0.000 n=9+9)
    SSA               12.2s ± 1%        12.0s ± 1%  -1.90%  (p=0.000 n=9+10)
    Flate             204ms ± 3%        202ms ± 3%    ~     (p=0.052 n=10+10)
    GoParser          248ms ± 1%        244ms ± 2%  -1.79%  (p=0.000 n=10+9)
    Reflect           671ms ± 1%        664ms ± 1%  -0.96%  (p=0.001 n=9+9)
    Tar               287ms ± 2%        285ms ± 3%    ~     (p=0.393 n=10+10)
    XML               362ms ± 1%        353ms ± 2%  -2.60%  (p=0.000 n=10+9)
    StdCmd            29.2s ± 1%        29.0s ± 1%  -0.63%  (p=0.021 n=10+8)
    [Geo mean]        888ms             875ms       -1.40%

    name        old user-time/op  new user-time/op  delta
    Template          393ms ± 5%        373ms ± 8%  -5.12%  (p=0.013 n=9+10)
    Unicode           185ms ± 6%        184ms ± 5%    ~     (p=0.825 n=10+10)
    GoTypes           1.33s ± 1%        1.31s ± 3%  -1.60%  (p=0.004 n=10+10)
    Compiler          5.98s ± 3%        5.92s ± 1%    ~     (p=0.050 n=10+10)
    SSA               15.5s ± 2%        15.3s ± 0%    ~     (p=0.156 n=10+9)
    Flate             255ms ± 5%        252ms ± 5%    ~     (p=0.362 n=10+10)
    GoParser          309ms ± 1%        304ms ± 3%  -1.79%  (p=0.021 n=7+10)
    Reflect           839ms ± 2%        833ms ± 1%    ~     (p=0.160 n=10+9)
    Tar               363ms ± 3%        358ms ± 4%    ~     (p=0.194 n=8+10)
    XML               446ms ± 3%        442ms ± 3%    ~     (p=0.503 n=10+10)
    [Geo mean]        791ms             779ms       -1.55%

Passes toolstash-check.

Change-Id: Ibcdb09f2c28907932581b7566f46d34be292594b
Reviewed-on: https://go-review.googlesource.com/108895
Run-TryBot: Iskander Sharipov <iskander.sharipov@intel.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Cherry Zhang <cherryyz@google.com>

src/cmd/internal/obj/x86/asm6.go

@@ -2168,7 +2168,9 @@ func span6(ctxt *obj.Link, s *obj.LSym, newprog obj.ProgAlloc) {
 	var c int32
 	errors := ctxt.Errors
 	for {
-		loop := int32(0)
+		// This loop continues while there are reasons to re-assemble
+		// whole block, like the presence of long forward jumps.
+		reAssemble := false
 		for i := range s.R {
 			s.R[i] = obj.Reloc{}
 		}
@@ -2228,7 +2230,7 @@ func span6(ctxt *obj.Link, s *obj.LSym, newprog obj.ProgAlloc) {
 				v := int32(p.Pc - (q.Pc + int64(q.Isize)))
 				if q.Back&branchShort != 0 {
 					if v > 127 {
-						loop++
+						reAssemble = true
 						q.Back ^= branchShort
 					}
 
@@ -2249,7 +2251,11 @@ func span6(ctxt *obj.Link, s *obj.LSym, newprog obj.ProgAlloc) {
 			m := ab.Len()
 			if int(p.Isize) != m {
 				p.Isize = uint8(m)
-				loop++
+				// When building for NaCl, we currently need
+				// at least 2 rounds to ensure proper 32-byte alignment.
+				if ctxt.Headtype == objabi.Hnacl {
+					reAssemble = true
+				}
 			}
 
 			s.Grow(p.Pc + int64(m))
@@ -2262,7 +2268,7 @@ func span6(ctxt *obj.Link, s *obj.LSym, newprog obj.ProgAlloc) {
 			ctxt.Diag("span must be looping")
 			log.Fatalf("loop")
 		}
-		if loop == 0 {
+		if !reAssemble {
 			break
 		}
 		if ctxt.Errors > errors {