cmd/5g: avoid temporary during constant binary op

This CL is an attempt to backport the abop code generation from 6g. This improves the generation of the range offset if the increment can be encoded directly via Operand2 shift encoding.

0023 (/home/dfc/src/range.go:7) BGE     ,29(APC)
0024 (/home/dfc/src/range.go:7) MOVW    0(R3),R5
0025 (/home/dfc/src/range.go:7) MOVW    $4,R1
0026 (/home/dfc/src/range.go:7) ADD     R1,R3,R3
0027 (/home/dfc/src/range.go:8) ADD     R5,R4,R4
0028 (/home/dfc/src/range.go:7) B       ,17(APC)

becomes

0023 (/home/dfc/src/range.go:7) BGE     ,28(APC)
0024 (/home/dfc/src/range.go:7) MOVW    0(R3),R0
0025 (/home/dfc/src/range.go:7) ADD     $4,R3,R3
0026 (/home/dfc/src/range.go:8) ADD     R0,R4,R4
0027 (/home/dfc/src/range.go:7) B       ,17(APC)

Benchmarks are unimpressive

dfc@qnap:~/go/test/bench/go1$ ~/go/misc/benchcmp {old,new}.txt
benchmark                 old ns/op    new ns/op    delta
BenchmarkBinaryTree17    2147483647   2147483647   +0.93%
BenchmarkFannkuch11      2147483647   2147483647   -2.52%
BenchmarkGobDecode        196135200    195842000   -0.15%
BenchmarkGobEncode         78581650     76734450   -2.35%
BenchmarkGzip            2147483647   2147483647   -0.47%
BenchmarkGunzip          1087243000   1070254000   -1.56%
BenchmarkJSONEncode      1107558000   1146077000   +3.48%
BenchmarkJSONDecode      2147483647   2147483647   -0.07%
BenchmarkMandelbrot200   2147483647   2147483647   -0.77%
BenchmarkParse             74328550     71653400   -3.60%
BenchmarkRevcomp          111123900    109325950   -1.62%
BenchmarkTemplate        2147483647   2147483647   -0.82%

benchmark                  old MB/s     new MB/s  speedup
BenchmarkGobDecode             3.91         3.92    1.00x
BenchmarkGobEncode             9.77        10.00    1.02x
BenchmarkGzip                  3.65         3.66    1.00x
BenchmarkGunzip               17.85        18.13    1.02x
BenchmarkJSONEncode            1.75         1.69    0.97x
BenchmarkJSONDecode            0.83         0.83    1.00x
BenchmarkParse                 0.78         0.81    1.04x
BenchmarkRevcomp              22.87        23.25    1.02x
BenchmarkTemplate              0.84         0.85    1.01x

R=remyoudompheng, minux.ma, rsc
CC=golang-dev
https://golang.org/cl/6564067

diff --git a/src/cmd/5g/cgen.c b/src/cmd/5g/cgen.c
index c0e9b4b6f89aa648de768c3b7e611fa61d024e78..79efab4947948f62f5093a31947e339ebeaa1c6c 100644 (file)
--- a/src/cmd/5g/cgen.c
+++ b/src/cmd/5g/cgen.c
@@ -439,18 +439,43 @@ abop:     // asymmetric binary
        if(nl->ullman >= nr->ullman) {
                regalloc(&n1, nl->type, res);
                cgen(nl, &n1);
-               regalloc(&n2, nr->type, N);
-               cgen(nr, &n2);
+               switch(n->op) {
+               case OADD:
+               case OSUB:
+               case OAND:
+               case OOR:
+               case OXOR:
+                       if(smallintconst(nr)) {
+                               n2 = *nr;
+                               break;
+                       }
+               default:
+                       regalloc(&n2, nr->type, N);
+                       cgen(nr, &n2);
+               }
        } else {
-               regalloc(&n2, nr->type, res);
-               cgen(nr, &n2);
+               switch(n->op) {
+               case OADD:
+               case OSUB:
+               case OAND:
+               case OOR:
+               case OXOR:
+                       if(smallintconst(nr)) {
+                               n2 = *nr;
+                               break;
+                       }
+               default:
+                       regalloc(&n2, nr->type, res);
+                       cgen(nr, &n2);
+               }
                regalloc(&n1, nl->type, N);
                cgen(nl, &n1);
        }
        gins(a, &n2, &n1);
        gmove(&n1, res);
        regfree(&n1);
-       regfree(&n2);
+       if(n2.op != OLITERAL)
+               regfree(&n2);
        goto ret;
 
 flt:   // floating-point.