cmd/internal/gc: use MOV R0, R1 instead of LEA 0(R0), R1 in Agen
Minor code generation optimization I've been meaning to do
for a while and noticed while working on the emitted write
barrier code. Using MOV lets the compiler and maybe the
processor do copy propagation.
name old new delta
BenchmarkBinaryTree17 17.9s × (0.99,1.01) 18.0s × (0.99,1.01) ~
BenchmarkFannkuch11 4.42s × (1.00,1.00) 4.36s × (1.00,1.00) -1.39%
BenchmarkFmtFprintfEmpty 118ns × (0.96,1.02) 120ns × (0.99,1.06) ~
BenchmarkFmtFprintfString 486ns × (0.99,1.01) 480ns × (0.99,1.01) -1.34%
BenchmarkFmtFprintfInt 457ns × (0.99,1.01) 451ns × (0.99,1.01) -1.31%
BenchmarkFmtFprintfIntInt 768ns × (1.00,1.01) 766ns × (0.99,1.01) ~
BenchmarkFmtFprintfPrefixedInt 584ns × (0.99,1.03) 569ns × (0.99,1.01) -2.57%
BenchmarkFmtFprintfFloat 739ns × (0.99,1.00) 728ns × (1.00,1.01) -1.49%
BenchmarkFmtManyArgs 2.77µs × (1.00,1.00) 2.81µs × (1.00,1.01) +1.53%
BenchmarkGobDecode 39.3ms × (0.99,1.01) 39.4ms × (0.99,1.01) ~
BenchmarkGobEncode 39.4ms × (0.99,1.00) 39.4ms × (0.99,1.00) ~
BenchmarkGzip 661ms × (0.99,1.01) 660ms × (1.00,1.01) ~
BenchmarkGunzip 142ms × (1.00,1.00) 143ms × (1.00,1.00) +0.20%
BenchmarkHTTPClientServer 133µs × (0.98,1.01) 132µs × (0.99,1.01) ~
BenchmarkJSONEncode 56.5ms × (0.99,1.01) 57.1ms × (0.99,1.01) +0.94%
BenchmarkJSONDecode 143ms × (1.00,1.00) 138ms × (1.00,1.01) -3.22%
BenchmarkMandelbrot200 6.01ms × (1.00,1.00) 6.02ms × (1.00,1.00) ~
BenchmarkGoParse 9.63ms × (0.94,1.07) 9.79ms × (0.92,1.07) ~
BenchmarkRegexpMatchEasy0_32 210ns × (1.00,1.00) 210ns × (1.00,1.01) ~
BenchmarkRegexpMatchEasy0_1K 596ns × (0.99,1.01) 593ns × (0.99,1.01) ~
BenchmarkRegexpMatchEasy1_32 184ns × (0.99,1.01) 182ns × (0.99,1.01) ~
BenchmarkRegexpMatchEasy1_1K 1.01µs × (0.99,1.01) 1.01µs × (1.00,1.01) ~
BenchmarkRegexpMatchMedium_32 327ns × (1.00,1.01) 331ns × (1.00,1.00) +1.22%
BenchmarkRegexpMatchMedium_1K 93.0µs × (1.00,1.02) 92.6µs × (1.00,1.01) ~
BenchmarkRegexpMatchHard_32 4.76µs × (0.95,1.01) 4.58µs × (0.99,1.05) ~
BenchmarkRegexpMatchHard_1K 136µs × (1.00,1.01) 136µs × (1.00,1.01) ~
BenchmarkRevcomp 892ms × (1.00,1.01) 900ms × (0.99,1.06) ~
BenchmarkTemplate 175ms × (0.99,1.00) 171ms × (1.00,1.01) -2.36%
BenchmarkTimeParse 638ns × (1.00,1.00) 637ns × (1.00,1.00) ~
BenchmarkTimeFormat 772ns × (1.00,1.00) 742ns × (1.00,1.00) -3.95%
Change-Id: I6504e310cb9cf48a73d539c478b4dbcacde208b2
Reviewed-on: https://go-review.googlesource.com/9308
Reviewed-by: Austin Clements <austin@google.com>
diff --git a/src/cmd/internal/gc/cgen.go b/src/cmd/internal/gc/cgen.go
index c996be7..427c671 100644
--- a/src/cmd/internal/gc/cgen.go
+++ b/src/cmd/internal/gc/cgen.go
@@ -1512,6 +1512,17 @@
return
}
+ if n.Op == OINDREG && n.Xoffset == 0 {
+ // Generate MOVW R0, R1 instead of MOVW $0(R0), R1.
+ // This allows better move propagation in the back ends
+ // (and maybe it helps the processor).
+ n1 := *n
+ n1.Op = OREGISTER
+ n1.Type = res.Type
+ Thearch.Gmove(&n1, res)
+ return
+ }
+
if n.Addable {
if n.Op == OREGISTER {
Fatal("agen OREGISTER")
