Patch: GGC-ify type hashtable

Patch: GGC-ify type hashtable

[Dale Johannesen]

The hashtable for types in c_common_alias_set is not ggc'd, because I 
thought
type nodes would always be pointed to from somewhere else.  After our 
latest
merge it appears that this is no longer true (if it ever was), leading 
to an ICE
building SPEC with IMA.   I have no idea how to construct a small 
testcase.
The fix is obvious.  I am not planning to investigate how the type node 
became
otherwise unreferenced unless somebody thinks this really shouldn't 
happen.
Bootstrapped on Darwin.

2004-10-16  Dale Johannesen  <dalej@apple.com>

         * c-common.c (c_common_get_alias_set):  Use GGC for 
type_hash_table.

Index: c-common.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/c-common.c,v
retrieving revision 1.577
diff -u -d -b -w -p -c -3 -p -r1.577 c-common.c
cvs diff: conflicting specifications of output style
*** c-common.c  13 Oct 2004 23:31:19 -0000      1.577
--- c-common.c  16 Oct 2004 20:08:04 -0000
*************** c_type_hash (const void *p)
*** 2592,2597 ****
--- 2592,2599 ----
     return ((size << 24) | (i << shift));
   }

+ static GTY((param_is (union tree_node))) htab_t type_hash_table;
+
   /* Return the typed-based alias set for T, which may be an expression
      or a type.  Return -1 if we don't do anything special.  */

*************** c_common_get_alias_set (tree t)
*** 2600,2606 ****
   {
     tree u;
     PTR *slot;
-   static htab_t type_hash_table;

     /* Permit type-punning when accessing a union, provided the access
        is directly through the union.  For example, this code does not
--- 2602,2607 ----
*************** c_common_get_alias_set (tree t)
*** 2719,2725 ****
     /* Look up t in hash table.  Only one of the compatible types 
within each
        alias set is recorded in the table.  */
     if (!type_hash_table)
!     type_hash_table = htab_create (1021, c_type_hash,
             (htab_eq) lang_hooks.types_compatible_p,
             NULL);
     slot = htab_find_slot (type_hash_table, t, INSERT);
--- 2720,2726 ----
     /* Look up t in hash table.  Only one of the compatible types 
within each
        alias set is recorded in the table.  */
     if (!type_hash_table)
!     type_hash_table = htab_create_ggc (1021, c_type_hash,
             (htab_eq) lang_hooks.types_compatible_p,
             NULL);
     slot = htab_find_slot (type_hash_table, t, INSERT);

Re: Patch: GGC-ify type hashtable

[Andrew Pinski]

On Oct 16, 2004, at 4:14 PM, Dale Johannesen wrote:

> The hashtable for types in c_common_alias_set is not ggc'd, because I 
> thought
> type nodes would always be pointed to from somewhere else.  After our 
> latest
> merge it appears that this is no longer true (if it ever was), leading 
> to an ICE
> building SPEC with IMA.   I have no idea how to construct a small 
> testcase.
> The fix is obvious.  I am not planning to investigate how the type 
> node became
> otherwise unreferenced unless somebody thinks this really shouldn't 
> happen.
> Bootstrapped on Darwin.

I also see the same problem with --enable-intermodule and was trying to
figure out why, I found the patch which caused it, I think.

I also found a way to speed up comp_types for multiple files which I 
will
be submitting soon.

Thanks,
Andrew Pinski

Re: Patch: GGC-ify type hashtable

[Andrew Pinski]

On Oct 16, 2004, at 4:21 PM, Andrew Pinski wrote:

>
> On Oct 16, 2004, at 4:14 PM, Dale Johannesen wrote:
>
>> The hashtable for types in c_common_alias_set is not ggc'd, because I 
>> thought
>> type nodes would always be pointed to from somewhere else.  After our 
>> latest
>> merge it appears that this is no longer true (if it ever was), 
>> leading to an ICE
>> building SPEC with IMA.   I have no idea how to construct a small 
>> testcase.
>> The fix is obvious.  I am not planning to investigate how the type 
>> node became
>> otherwise unreferenced unless somebody thinks this really shouldn't 
>> happen.
>> Bootstrapped on Darwin.
>
> I also see the same problem with --enable-intermodule and was trying to
> figure out why, I found the patch which caused it, I think.
Forgot to say the patch would be:
http://gcc.gnu.org/ml/gcc-cvs/2004-09/msg00192.html

Which is connect to PR 17957 so maybe Jan can comment here.

Thanks,
Andrew Pinski

Re: Patch: GGC-ify type hashtable

[Jan Hubicka]

> 
> On Oct 16, 2004, at 4:21 PM, Andrew Pinski wrote:
> 
> >
> >On Oct 16, 2004, at 4:14 PM, Dale Johannesen wrote:
> >
> >>The hashtable for types in c_common_alias_set is not ggc'd, because I 
> >>thought
> >>type nodes would always be pointed to from somewhere else.  After our 
> >>latest
> >>merge it appears that this is no longer true (if it ever was), 
> >>leading to an ICE
> >>building SPEC with IMA.   I have no idea how to construct a small 
> >>testcase.
> >>The fix is obvious.  I am not planning to investigate how the type 
> >>node became
> >>otherwise unreferenced unless somebody thinks this really shouldn't 
> >>happen.
> >>Bootstrapped on Darwin.
> >
> >I also see the same problem with --enable-intermodule and was trying to
> >figure out why, I found the patch which caused it, I think.
> Forgot to say the patch would be:
> http://gcc.gnu.org/ml/gcc-cvs/2004-09/msg00192.html
> 
> Which is connect to PR 17957 so maybe Jan can comment here.

Well, there is not much to comment.  We didn't release the bodies for a
while by mistake and GTYing the pointer we keep around looks like proper
fix.

Honza
> 
> Thanks,
> Andrew Pinski

Re: Patch: GGC-ify type hashtable

[Richard Henderson]

On Sat, Oct 16, 2004 at 01:14:02PM -0700, Dale Johannesen wrote:
>         * c-common.c (c_common_get_alias_set):  Use GGC for 
> type_hash_table.

Ok.


r~