Re: [PATCH] Fix GC problem with complex types

[Andrew Pinski <pinskia at physics dot uc dot edu>]

On Nov 23, 2004, at 1:25 PM, Janis Johnson wrote:

> On Mon, Nov 22, 2004 at 09:28:54PM -0500, Andrew Pinski wrote:
> > We are GCing some complex types which we still have reference to.
> > The problem is the the variable which stores last vector type is not
> > marked with GTY, this fixes the problem.  The reason why I also marked
> > the int as GTY is because if we don't we could have a possible PCH
> > problem but I did not test that patch at all.
> >
> > OK? Bootstrapped and tested on ppc-darwin with no regressions.
>
> I tried this with my original testcase on powerpc-linux and it still
> segfaults, then tried it with the small testcase from the PR with ggc
> params to force frequent garbage collection and that fails.

Oh, know why this did not fix the problem, I had forgot to add it
to the GTFILES (why is needed again, why can't this be done
automatically). I also added the assert in build_word_mode_vector_type
like there is in the bug report.  And fixed where the GTY markers
should be, before the type.

OK? Bootstrapped and tested on powerpc-darwin with no regressions.

Thanks,
Andrew Pinski

Testcase:
/* { dg-do compile } */
/* { dg-options "--param ggc-min-expand=0 -param ggc-min-heapsize=0" } 
*/
__attribute__ ((vector_size (64))) unsigned char v1, v2, v3;
void
vadd (void)
{
  v1 = v2 + v3;
}
void
test_add (void)
{
  vadd ();
}
void
vsub (void)
{
  v1 = v2 - v3;
}

ChangeLog:

	* tree-complex.c (vector_inner_type): New variable moved from
	build_word_mode_vector_type.
	(vector_last_type): Likewise.
	(vector_last_nunits): Likewise.
	(build_word_mode_vector_type): Use the new variables.
	* Makefile.in (tree-complex.o): Add gt-tree-complex.h $(GGC_H).
	(GTFILES): Add tree-complex.c.
	(gt-tree-complex.h): New rule, add it to the rest of the gt-* rules.

Attachment: temp.diff.txt
Description: Text document