On Mon, 25 Jun 2018, Richard Biener wrote:
>> Sigh - I see GCC optimizes memmove as well as memcpy in this case, so
>> changing the offending memcpy calls to memmoves would be a bit cleaner. OK to
>> go with this instead?
>
> I think that's better. Or conditionalizing the offending ones on dest! = src?
That would work, but would require an extra comparison ('c->n == 3' test would
need to be kept) and also would introduce a poorly-predictable branch depending
on comparator outcomes - something the design of gcc_qsort strives to avoid in
the first place.
This is the patch I'm going to apply.
PR middle-end/86311
* sort.cc (REORDER_23): Avoid memcpy with same destination and source.
(REORDER_45): Likewise.
diff --git a/gcc/sort.cc b/gcc/sort.cc
index a48a477d4e8..293e2058f89 100644
--- a/gcc/sort.cc
+++ b/gcc/sort.cc
@@ -69,7 +69,7 @@ do { \
memcpy (&t1, e1 + OFFSET, sizeof (TYPE)); \
char *out = c->out + OFFSET; \
if (likely (c->n == 3)) \
- memcpy (out + 2*STRIDE, e2 + OFFSET, sizeof (TYPE)); \
+ memmove (out + 2*STRIDE, e2 + OFFSET, sizeof (TYPE));\
memcpy (out, &t0, sizeof (TYPE)); out += STRIDE; \
memcpy (out, &t1, sizeof (TYPE)); \
} while (0)
@@ -101,7 +101,7 @@ do { \
memcpy (&t3, e3 + OFFSET, sizeof (TYPE)); \
char *out = c->out + OFFSET; \
if (likely (c->n == 5)) \
- memcpy (out + 4*STRIDE, e4 + OFFSET, sizeof (TYPE)); \
+ memmove (out + 4*STRIDE, e4 + OFFSET, sizeof (TYPE));\
memcpy (out, &t0, sizeof (TYPE)); out += STRIDE; \
memcpy (out, &t1, sizeof (TYPE)); out += STRIDE; \
memcpy (out, &t2, sizeof (TYPE)); out += STRIDE; \
