| Message ID | 20100726192415.GC20984@kam.mff.cuni.cz |
|---|---|
| State | New |
| Headers | show |
On 10-07-26 12:24 , Jan Hubicka wrote: > Hi, > Andi's changes introduced heap allocated object pointed from lto_file_data, so > we now ICE on everything big enough to trigger garbage collector at WPA stage. > Andi: this solves the cgraph hash table ICE you sent me > > Bootstrapping/regtesting x86_64-linux, OK? > Honza > > * lto-streamer.h (struct lto_file_decl_data): Mark resolutions with GTY((skip)) Would it make more sense to change the allocation on the resolutions field? Diego.
> On 10-07-26 12:24 , Jan Hubicka wrote: >> Hi, >> Andi's changes introduced heap allocated object pointed from lto_file_data, so >> we now ICE on everything big enough to trigger garbage collector at WPA stage. >> Andi: this solves the cgraph hash table ICE you sent me >> >> Bootstrapping/regtesting x86_64-linux, OK? >> Honza >> >> * lto-streamer.h (struct lto_file_decl_data): Mark resolutions with GTY((skip)) > > Would it make more sense to change the allocation on the resolutions field? Well, lto_file_decl_data has most of fields skipped anyway and I think it is good idea as I don't like moving random stuff into garbage collector. It is in GGC because it points to global_decl_state that points to decls, but the other streamer data are better to stay in explicitely allocated memory IMO. Honza > > > Diego.
On Mon, Jul 26, 2010 at 16:40, Jan Hubicka <hubicka@ucw.cz> wrote: > Well, lto_file_decl_data has most of fields skipped anyway and I think it is > good idea as I don't like moving random stuff into garbage collector. It is in > GGC because it points to global_decl_state that points to decls, but the > other streamer data are better to stay in explicitely allocated memory IMO. Ah, yes, you're right. I had forgotten that. Patch is OK. Diego.
On 7/26/2010 9:24 PM, Jan Hubicka wrote: > Hi, > Andi's changes introduced heap allocated object pointed from lto_file_data, so > we now ICE on everything big enough to trigger garbage collector at WPA stage. > Andi: this solves the cgraph hash table ICE you sent me Thanks for the fix! -Andi
On Mon, 26 Jul 2010, Jan Hubicka wrote: > Hi, > Andi's changes introduced heap allocated object pointed from lto_file_data, so > we now ICE on everything big enough to trigger garbage collector at WPA stage. > Andi: this solves the cgraph hash table ICE you sent me > > Bootstrapping/regtesting x86_64-linux, OK? Ok. Thanks, Richard. > Honza > > * lto-streamer.h (struct lto_file_decl_data): Mark resolutions with GTY((skip)) > Index: lto-streamer.h > =================================================================== > --- lto-streamer.h (revision 162539) > +++ lto-streamer.h (working copy) > @@ -602,13 +602,13 @@ > htab_t GTY((skip)) renaming_hash_table; > > /* Linked list used temporarily in reader */ > struct lto_file_decl_data *next; > > /* Sub ID for merged objects. */ > unsigned id; > > /* Symbol resolutions for this file */ > - VEC(ld_plugin_symbol_resolution_t,heap) *resolutions; > + VEC(ld_plugin_symbol_resolution_t,heap) * GTY((skip)) resolutions; > }; > > typedef struct lto_file_decl_data *lto_file_decl_data_ptr; > >
Index: lto-streamer.h =================================================================== --- lto-streamer.h (revision 162539) +++ lto-streamer.h (working copy) @@ -602,13 +602,13 @@ htab_t GTY((skip)) renaming_hash_table; /* Linked list used temporarily in reader */ struct lto_file_decl_data *next; /* Sub ID for merged objects. */ unsigned id; /* Symbol resolutions for this file */ - VEC(ld_plugin_symbol_resolution_t,heap) *resolutions; + VEC(ld_plugin_symbol_resolution_t,heap) * GTY((skip)) resolutions; }; typedef struct lto_file_decl_data *lto_file_decl_data_ptr;