cmd/6g: avoid MOVSD between registers

MOVSD only copies the low half of the packed register pair,
while MOVAPD copies both halves.  I assume the internal
register renaming works better with the latter, since it makes
our code run 25% faster.

Before:
mandelbrot 16000
        gcc -O2 mandelbrot.c	28.44u 0.00s 28.45r
        gc mandelbrot	44.12u 0.00s 44.13r
        gc_B mandelbrot	44.17u 0.01s 44.19r

After:
mandelbrot 16000
        gcc -O2 mandelbrot.c	28.22u 0.00s 28.23r
        gc mandelbrot	32.81u 0.00s 32.82r
        gc_B mandelbrot	32.82u 0.00s 32.83r

R=ken2
CC=golang-dev
https://golang.org/cl/6248068

diff --git a/src/cmd/6g/peep.c b/src/cmd/6g/peep.c
index f55dfc8868fc1e7262e5c49b492c22245a31271c..51f4722149094e60331b823fd7e6029970165510 100644 (file)
--- a/src/cmd/6g/peep.c
+++ b/src/cmd/6g/peep.c
@@ -283,6 +283,12 @@ loop1:
        // copyprop.  Now that copyprop is done, remov MOVLQZX R1, R2
        // if it is dominated by an earlier ADDL/MOVL/etc into R1 that
        // will have already cleared the high bits.
+       //
+       // MOVSD removal.
+       // We never use packed registers, so a MOVSD between registers
+       // can be replaced by MOVAPD, which moves the pair of float64s
+       // instead of just the lower one.  We only use the lower one, but
+       // the processor can do better if we do moves using both.
        for(r=firstr; r!=R; r=r->link) {
                p = r->prog;
                if(p->as == AMOVLQZX)
@@ -290,6 +296,11 @@ loop1:
                if(p->from.type == p->to.type)
                if(prevl(r, p->from.type))
                        excise(r);
+               
+               if(p->as == AMOVSD)
+               if(regtyp(&p->from))
+               if(regtyp(&p->to))
+                       p->as = AMOVAPD;
        }
 
        // load pipelining