This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [PATCH] doloop_optimize miscompiles openssl
On Mon, 25 Sep 2000, Geoff Keating wrote:
> Franz Sirl <Franz.Sirl-kernel@lauterbach.com> writes:
> > * doloop.c (doloop_modify): Prevent delete_insn() from
> > deleting too much. Prefer loop->top over loop->start as
> > target for the new JUMP insn.
> >
> > Index: gcc/doloop.c
> > ===================================================================
> > RCS file: /cvs/gcc/egcs/gcc/doloop.c,v
> > retrieving revision 1.2
> > diff -u -p -r1.2 doloop.c
> > --- doloop.c 2000/08/14 20:55:11 1.2
> > +++ doloop.c 2000/09/24 22:00:04
> > @@ -416,11 +416,14 @@ doloop_modify (loop, iterations, iterati
> > }
> >
> > /* Discard original jump to continue loop. The original compare
> > - result may still be live, so it cannot be discarded explicitly. */
> > + result may still be live, so it cannot be discarded explicitly.
> > + Don't delete too much, increment the use count of the target label.
> > */ + LABEL_NUSES (JUMP_LABEL (jump_insn))++;
> > delete_insn (jump_insn);
> > + LABEL_NUSES (JUMP_LABEL (jump_insn))--;
> >
> > - /* Emit the label that will delimit the start of the loop. */
> > - emit_label_after (start_label, loop->start);
> > + /* Emit the label that will delimit the top of the loop. */
> > + emit_label_after (start_label, loop->top ? loop->top : loop->start);
> > LABEL_NUSES (start_label)++;
>
> Franz,
>
> would it work to simply emit the start_label before calling
> delete_insn? After all, the label that jump_insn branches to is
> really not used after this transformation, right?
Yes, it works. In the meantime I found another bug with the appended
testcase, it seems testing for loop_info->has_multiple_exit_targets is not
enough for ruling out unsuitable loops. For the testcase
loop_info->has_multiple_exit_targets is 0, but loop->exit_counmt is 3.
Franz.
* doloop.c (doloop_modify): Prevent delete_insn() from
deleting too much. Prefer loop->top over loop->start as
target for the new JUMP insn.
(doloop_valid_p): Ignore loop with exit_count > 1.
Index: gcc/doloop.c
===================================================================
RCS file: /cvs/gcc/egcs/gcc/doloop.c,v
retrieving revision 1.2
diff -u -p -r1.2 doloop.c
--- gcc/doloop.c 2000/08/14 20:55:11 1.2
+++ gcc/doloop.c 2000/09/25 21:43:34
@@ -281,7 +281,7 @@ doloop_valid_p (loop, jump_insn)
statement within a loop will generate multiple loop exits.
Another example of a loop that currently generates multiple exit
targets is for (i = 0; i < (foo ? 8 : 4); i++) { }. */
- if (loop_info->has_multiple_exit_targets)
+ if (loop_info->has_multiple_exit_targets || loop->exit_count > 1)
{
if (loop_dump_stream)
fprintf (loop_dump_stream,
@@ -415,13 +415,13 @@ doloop_modify (loop, iterations, iterati
fputs (" iterations).", loop_dump_stream);
}
+ /* Emit the label that will delimit the top of the loop. */
+ emit_label_after (start_label, loop->top ? loop->top : loop->start);
+ LABEL_NUSES (start_label)++;
+
/* Discard original jump to continue loop. The original compare
result may still be live, so it cannot be discarded explicitly. */
delete_insn (jump_insn);
-
- /* Emit the label that will delimit the start of the loop. */
- emit_label_after (start_label, loop->start);
- LABEL_NUSES (start_label)++;
counter_reg = XEXP (condition, 0);
if (GET_CODE (counter_reg) == PLUS)
unsigned long bn_add_words(unsigned long *r, unsigned long *a, unsigned long *b, int n)
{
unsigned long long ll=0;
if (n <= 0) return((unsigned long)0);
for (;;)
{
ll+=(unsigned long long)a[0]+b[0];
r[0]=(unsigned long)ll&(0xffffffffL);
ll>>=32;
if (--n <= 0) break;
ll+=(unsigned long long)a[1]+b[1];
r[1]=(unsigned long)ll&(0xffffffffL);
ll>>=32;
if (--n <= 0) break;
ll+=(unsigned long long)a[2]+b[2];
r[2]=(unsigned long)ll&(0xffffffffL);
ll>>=32;
if (--n <= 0) break;
ll+=(unsigned long long)a[3]+b[3];
r[3]=(unsigned long)ll&(0xffffffffL);
ll>>=32;
if (--n <= 0) break;
a+=4;
b+=4;
r+=4;
}
return((unsigned long)ll);
}