cmd/compile: duplicate some scalar stores in wb blocks

DO NOT SUBMIT

[mailing for initial feedback, needs benchmark numbers]

Duplicate some scalar stores on both halves of a
writeBarrier.enabled block, to reduce branching.

There are other memory ops besides OpVarXXX and OpStore
that we could consider here: OpMove, OpZero, Op*Call,
atomic stores, OpSelect1, OpPhi.

However, OpVarXXX and OpStore are by far the most common,
and they correspond to a fixed number of instructions.

The others are all rare and are either tricky
(OpSelect1, OpPhi, atomic stores) or
correspond to potentially lots of code.

Fixes golang#19838

Change-Id: I0d5a68c2dfb3b6a6916b92151b3e64df533d4581

Author: josharian

src/cmd/compile/internal/ssa/writebarrier.go

@@ -9,6 +9,10 @@ import (
 	"cmd/internal/src"
 )
 
+// nonWBBudget is the number of non-WB stores we can duplicate
+// on either side of a check of runtime.writeBarrier.enabled.
+const nonWBBudget = 2
+
 // needwb returns whether we need write barrier for store op v.
 // v must be Store/Move/Zero.
 func needwb(v *Value) bool {
@@ -116,6 +120,7 @@ func writebarrier(f *Func) {
 		var last *Value
 		var start, end int
 		values := b.Values
+		budget := nonWBBudget
 	FindSeq:
 		for i := len(values) - 1; i >= 0; i-- {
 			w := values[i]
@@ -126,6 +131,15 @@ func writebarrier(f *Func) {
 					last = w
 					end = i + 1
 				}
+			case OpStore:
+				if last == nil {
+					continue
+				}
+				if budget > 0 {
+					budget--
+				} else {
+					break FindSeq
+				}
 			case OpVarDef, OpVarLive, OpVarKill:
 				continue
 			default:
@@ -194,6 +208,8 @@ func writebarrier(f *Func) {
 			case OpZeroWB:
 				fn = typedmemclr
 				typ = &ExternSymbol{Sym: w.Aux.(Type).Symbol()}
+			case OpStore:
+				val = w.Args[1]
 			case OpVarDef, OpVarLive, OpVarKill:
 			}
 
@@ -202,13 +218,15 @@ func writebarrier(f *Func) {
 			case OpStoreWB, OpMoveWB, OpZeroWB:
 				volatile := w.Op == OpMoveWB && isVolatile(val)
 				memThen = wbcall(pos, bThen, fn, typ, ptr, val, memThen, sp, sb, volatile)
+			case OpStore:
+				memThen = bThen.NewValue3A(pos, OpStore, TypeMem, w.Aux, ptr, val, memThen)
 			case OpVarDef, OpVarLive, OpVarKill:
 				memThen = bThen.NewValue1A(pos, w.Op, TypeMem, w.Aux, memThen)
 			}
 
 			// else block: normal store
 			switch w.Op {
-			case OpStoreWB:
+			case OpStoreWB, OpStore:
 				memElse = bElse.NewValue3A(pos, OpStore, TypeMem, w.Aux, ptr, val, memElse)
 			case OpMoveWB:
 				memElse = bElse.NewValue3I(pos, OpMove, TypeMem, w.AuxInt, ptr, val, memElse)