cmd/compile: reuse []Flow

Benchmarked using compilebench on a quiet
but rather old OS X laptop.

Benchmarks from others would be welcome,
since the numbers look too good to be true.

name      old time/op    new time/op    delta
Template     331ms ± 9%     303ms ± 4%   -8.25%  (p=0.000 n=24+24)
GoTypes      946ms ± 4%     888ms ± 3%   -6.17%  (p=0.000 n=24+25)
Compiler     3.20s ± 1%     3.10s ± 2%   -3.07%  (p=0.000 n=24+25)

name      old alloc/op   new alloc/op   delta
Template    72.5MB ± 0%    61.8MB ± 0%  -14.76%  (p=0.000 n=25+24)
GoTypes      224MB ± 0%     189MB ± 0%  -15.65%  (p=0.000 n=25+25)
Compiler     695MB ± 0%     561MB ± 0%  -19.26%  (p=0.000 n=25+25)

name      old allocs/op  new allocs/op  delta
Template      498k ± 0%      497k ± 0%   -0.21%  (p=0.000 n=25+23)
GoTypes      1.47M ± 0%     1.47M ± 0%   -0.25%  (p=0.000 n=25+25)
Compiler     4.09M ± 0%     4.08M ± 0%   -0.18%  (p=0.000 n=25+23)

Change-Id: I2394bc748128d721863453257fa5756c410f7898
Reviewed-on: https://go-review.googlesource.com/19771Reviewed-by: Russ Cox <rsc@golang.org>
Run-TryBot: Russ Cox <rsc@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>

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

@@ -241,6 +241,19 @@ var flowmark int
 // will not have flow graphs and consequently will not be optimized.
 const MaxFlowProg = 50000
 
+var ffcache []Flow // reusable []Flow, to reduce allocation
+
+func growffcache(n int) {
+	if n > cap(ffcache) {
+		n = (n * 5) / 4
+		if n > MaxFlowProg {
+			n = MaxFlowProg
+		}
+		ffcache = make([]Flow, n)
+	}
+	ffcache = ffcache[:n]
+}
+
 func Flowstart(firstp *obj.Prog, newData func() interface{}) *Graph {
 	// Count and mark instructions to annotate.
 	nf := 0
@@ -268,7 +281,9 @@ func Flowstart(firstp *obj.Prog, newData func() interface{}) *Graph {
 
 	// Allocate annotations and assign to instructions.
 	graph := new(Graph)
-	ff := make([]Flow, nf)
+
+	growffcache(nf)
+	ff := ffcache
 	start := &ff[0]
 	id := 0
 	var last *Flow
@@ -331,6 +346,10 @@ func Flowend(graph *Graph) {
 		f.Prog.Info.Flags = 0 // drop cached proginfo
 		f.Prog.Opt = nil
 	}
+	clear := ffcache[:graph.Num]
+	for i := range clear {
+		clear[i] = Flow{}
+	}
 }
 
 // find looping structure