cmd/internal/gc: inline writeBarrierEnabled check before calling writebarrierptr

I believe the benchmarks that get slower are under register pressure,
and not making the call unconditionally makes the pressure worse,
and the register allocator doesn't do a great job. But part of the point
of this sequence is to get the write barriers out of the way so I can work
on the register allocator, so that's okay.

name                                       old                     new          delta
BenchmarkBinaryTree17              17.9s × (1.00,1.01)     18.0s × (0.99,1.01)  ~
BenchmarkFannkuch11                4.43s × (1.00,1.00)     4.43s × (1.00,1.00)  ~
BenchmarkFmtFprintfEmpty           110ns × (1.00,1.06)     114ns × (0.95,1.05)  ~
BenchmarkFmtFprintfString          487ns × (0.99,1.00)     468ns × (0.99,1.01)  -4.00%
BenchmarkFmtFprintfInt             450ns × (0.99,1.00)     433ns × (1.00,1.01)  -3.88%
BenchmarkFmtFprintfIntInt          762ns × (1.00,1.00)     748ns × (0.99,1.01)  -1.84%
BenchmarkFmtFprintfPrefixedInt     584ns × (0.99,1.01)     547ns × (0.99,1.01)  -6.26%
BenchmarkFmtFprintfFloat           738ns × (1.00,1.00)     756ns × (1.00,1.01)  +2.37%
BenchmarkFmtManyArgs              2.80µs × (1.00,1.01)    2.79µs × (1.00,1.01)  ~
BenchmarkGobDecode                39.0ms × (0.99,1.00)    39.6ms × (0.99,1.00)  +1.54%
BenchmarkGobEncode                37.8ms × (0.98,1.01)    37.6ms × (1.00,1.01)  ~
BenchmarkGzip                      661ms × (0.99,1.01)     663ms × (0.99,1.02)  ~
BenchmarkGunzip                    142ms × (1.00,1.00)     142ms × (1.00,1.00)  ~
BenchmarkHTTPClientServer          132µs × (0.99,1.01)     132µs × (0.99,1.01)  ~
BenchmarkJSONEncode               56.3ms × (0.99,1.01)    56.2ms × (0.99,1.01)  ~
BenchmarkJSONDecode                138ms × (0.99,1.01)     138ms × (1.00,1.00)  ~
BenchmarkMandelbrot200            6.01ms × (1.00,1.00)    6.03ms × (1.00,1.01)  +0.23%
BenchmarkGoParse                  10.2ms × (0.87,1.05)     9.8ms × (0.93,1.10)  ~
BenchmarkRegexpMatchEasy0_32       208ns × (1.00,1.00)     207ns × (1.00,1.00)  ~
BenchmarkRegexpMatchEasy0_1K       588ns × (1.00,1.00)     581ns × (1.00,1.01)  -1.27%
BenchmarkRegexpMatchEasy1_32       182ns × (0.99,1.01)     185ns × (0.99,1.01)  +1.65%
BenchmarkRegexpMatchEasy1_1K       986ns × (1.00,1.01)     975ns × (1.00,1.01)  -1.17%
BenchmarkRegexpMatchMedium_32      323ns × (1.00,1.01)     328ns × (0.99,1.00)  +1.55%
BenchmarkRegexpMatchMedium_1K     89.9µs × (1.00,1.00)    88.6µs × (1.00,1.01)  -1.38%
BenchmarkRegexpMatchHard_32       4.72µs × (0.95,1.01)    4.69µs × (0.95,1.03)  ~
BenchmarkRegexpMatchHard_1K        133µs × (1.00,1.01)     133µs × (1.00,1.01)  ~
BenchmarkRevcomp                   900ms × (1.00,1.05)     902ms × (0.99,1.05)  ~
BenchmarkTemplate                  168ms × (0.99,1.01)     174ms × (0.99,1.01)  +3.30%
BenchmarkTimeParse                 637ns × (1.00,1.00)     639ns × (1.00,1.00)  +0.31%
BenchmarkTimeFormat                738ns × (1.00,1.00)     736ns × (1.00,1.01)  ~

Change-Id: I03ce152852edec404538f6c20eb650fac82e2aa2
Reviewed-on: https://go-review.googlesource.com/9224Reviewed-by: Austin Clements <austin@google.com>

src/cmd/internal/gc/builtin.go

@@ -87,6 +87,7 @@ const runtimeimport = "" +
 	"func @\"\".chanrecv2 (@\"\".chanType·2 *byte, @\"\".hchan·3 <-chan any, @\"\".elem·4 *any) (? bool)\n" +
 	"func @\"\".chansend1 (@\"\".chanType·1 *byte, @\"\".hchan·2 chan<- any, @\"\".elem·3 *any)\n" +
 	"func @\"\".closechan (@\"\".hchan·1 any)\n" +
+	"var @\"\".writeBarrierEnabled bool\n" +
 	"func @\"\".writebarrierptr (@\"\".dst·1 *any, @\"\".src·2 any)\n" +
 	"func @\"\".writebarrierstring (@\"\".dst·1 *any, @\"\".src·2 any)\n" +
 	"func @\"\".writebarrierslice (@\"\".dst·1 *any, @\"\".src·2 any)\n" +

src/cmd/internal/gc/builtin/runtime.go

@@ -108,7 +108,8 @@ func chanrecv2(chanType *byte, hchan <-chan any, elem *any) bool
 func chansend1(chanType *byte, hchan chan<- any, elem *any)
 func closechan(hchan any)
 
-// *byte is really *runtime.Type
+var writeBarrierEnabled bool
+
 func writebarrierptr(dst *any, src any)
 func writebarrierstring(dst *any, src any)
 func writebarrierslice(dst *any, src any)
@@ -144,6 +145,7 @@ func writebarrierfat1101(dst *any, _ uintptr, src any)
 func writebarrierfat1110(dst *any, _ uintptr, src any)
 func writebarrierfat1111(dst *any, _ uintptr, src any)
 
+// *byte is really *runtime.Type
 func typedmemmove(typ *byte, dst *any, src *any)
 func typedslicecopy(typ *byte, dst any, src any) int
 

src/cmd/internal/gc/cgen.go

@@ -793,10 +793,24 @@ func cgen_wbptr(n, res *Node) {
 	if Debug_wb > 0 {
 		Warn("write barrier")
 	}
+
 	var dst, src Node
-	Agenr(res, &dst, nil)
-	Cgenr(n, &src, nil)
-	p := Thearch.Gins(Thearch.Optoas(OAS, Types[Tptr]), &dst, nil)
+	Igen(res, &dst, nil)
+	if n.Op == OREGISTER {
+		src = *n
+		Regrealloc(&src)
+	} else {
+		Cgenr(n, &src, nil)
+	}
+
+	Thearch.Gins(Thearch.Optoas(OCMP, Types[TUINT8]), syslook("writeBarrierEnabled", 0), Nodintconst(0))
+	pbr := Gbranch(Thearch.Optoas(ONE, Types[TUINT32]), nil, -1)
+	Thearch.Gins(Thearch.Optoas(OAS, Types[Tptr]), &src, &dst)
+	pjmp := Gbranch(obj.AJMP, nil, 0)
+	Patch(pbr, Pc)
+	var adst Node
+	Agenr(&dst, &adst, &dst)
+	p := Thearch.Gins(Thearch.Optoas(OAS, Types[Tptr]), &adst, nil)
 	a := &p.To
 	a.Type = obj.TYPE_MEM
 	a.Reg = int16(Thearch.REGSP)
@@ -807,12 +821,15 @@ func cgen_wbptr(n, res *Node) {
 	p2 := Thearch.Gins(Thearch.Optoas(OAS, Types[Tptr]), &src, nil)
 	p2.To = p.To
 	p2.To.Offset += int64(Widthptr)
-	Regfree(&dst)
-	Regfree(&src)
+	Regfree(&adst)
 	if sys_wbptr == nil {
 		sys_wbptr = writebarrierfn("writebarrierptr", Types[Tptr], Types[Tptr])
 	}
 	Ginscall(sys_wbptr, 0)
+	Patch(pjmp, Pc)
+
+	Regfree(&dst)
+	Regfree(&src)
 }
 
 func cgen_wbfat(n, res *Node) {