| Submitter | Jakub Jelinek |
|---|---|
| Date | Jan. 21, 2013, 2:36 p.m. |
| Message ID | <20130121143623.GN7269@tucnak.redhat.com> |
| Download | mbox | patch |
| Permalink | /patch/214185/ |
| State | New |
| Headers | show |
Comments
On 01/21/2013 07:36 AM, Jakub Jelinek wrote: > Hi! > > While working on the multiversioning patch I've just posted, > I've noticed that on say: > void foo (); > void foo () __attribute__((target ("avx"))); > void foo () __attribute__((target ("default"))); > __attribute__((target ("default"))) void foo () > { > } > __attribute__((target ("avx"))) void foo () > { > } > void (*fn) () = foo; > we leak memory, because init_reload is called several times (due to target > attribute) and it never destroys reload_obstack already created before > initializing a new one. Fixed by initializing it just once. > > Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk? > > 2013-01-21 Jakub Jelinek <jakub@redhat.com> > > * reload1.c (init_reload): Only initialize reload_obstack > during the first call. OK. jeff
Patch
--- gcc/reload1.c.jj 2013-01-11 09:03:17.000000000 +0100 +++ gcc/reload1.c 2013-01-21 12:47:01.091040655 +0100 @@ -468,8 +468,11 @@ init_reload (void) } /* Initialize obstack for our rtl allocation. */ - gcc_obstack_init (&reload_obstack); - reload_startobj = XOBNEWVAR (&reload_obstack, char, 0); + if (reload_startobj == NULL) + { + gcc_obstack_init (&reload_obstack); + reload_startobj = XOBNEWVAR (&reload_obstack, char, 0); + } INIT_REG_SET (&spilled_pseudos); INIT_REG_SET (&changed_allocation_pseudos);
Hi! While working on the multiversioning patch I've just posted, I've noticed that on say: void foo (); void foo () __attribute__((target ("avx"))); void foo () __attribute__((target ("default"))); __attribute__((target ("default"))) void foo () { } __attribute__((target ("avx"))) void foo () { } void (*fn) () = foo; we leak memory, because init_reload is called several times (due to target attribute) and it never destroys reload_obstack already created before initializing a new one. Fixed by initializing it just once. Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk? 2013-01-21 Jakub Jelinek <jakub@redhat.com> * reload1.c (init_reload): Only initialize reload_obstack during the first call. Jakub