[ reload-v2]

[ reload-v2]

[Jeff Law]

[-- Attachment #1: Type: text/plain, Size: 168 bytes --]


ggc_realloc is no longer supported, ggc_realloc_stat is the proper way 
to do things now.   This allows reload-v2 to bootstrap after monday's 
merge from the trunk.



[-- Attachment #2: P --]
[-- Type: text/plain, Size: 1502 bytes --]

Index: ChangeLog-ira-reload
===================================================================
*** ChangeLog-ira-reload	(revision 160799)
--- ChangeLog-ira-reload	(working copy)
***************
*** 1,3 ****
--- 1,7 ----
+ 2010-06-16  Jeff Law <law@redhat.com>
+ 
+ 	* ira-reload.c (localize_pseudos): Use ggc_realloc_stat.
+ 
  2010-06-09  Jeff Law <law@redhat.com>
  
  	* ira-build.c (initiate_cost_vectors): Initiate a cost vector for
Index: ira-reload.c
===================================================================
*** ira-reload.c	(revision 160799)
--- ira-reload.c	(working copy)
*************** localize_pseudos (basic_block bb, bitmap
*** 899,905 ****
        VEC_safe_grow_cleared (rtx, gc, reg_equiv_memory_loc_vec, max_regno);
        reg_equiv_memory_loc = VEC_address (rtx, reg_equiv_memory_loc_vec);
        reg_equiv_init
! 	= (rtx *) ggc_realloc (reg_equiv_init, max_regno * sizeof (rtx));
        memset (&reg_equiv_init[orig_max_reg_num], 0, nregs * sizeof (rtx));
        regstat_n_sets_and_refs
  	 = ((struct regstat_n_sets_and_refs_t *)
--- 899,905 ----
        VEC_safe_grow_cleared (rtx, gc, reg_equiv_memory_loc_vec, max_regno);
        reg_equiv_memory_loc = VEC_address (rtx, reg_equiv_memory_loc_vec);
        reg_equiv_init
! 	= (rtx *) ggc_realloc_stat (reg_equiv_init, max_regno * sizeof (rtx) PASS_MEM_STAT);
        memset (&reg_equiv_init[orig_max_reg_num], 0, nregs * sizeof (rtx));
        regstat_n_sets_and_refs
  	 = ((struct regstat_n_sets_and_refs_t *)

Re: [ reload-v2]

[Jan Hubicka]

>
> ggc_realloc is no longer supported, ggc_realloc_stat is the proper way  
> to do things now.   This allows reload-v2 to bootstrap after monday's  
> merge from the trunk.

Well, it was definitly not intention of my statistics code to require
users to add random PASS_MEM_STAT.
There should be always corresponding macro hidding that.

Also your patch will not compile with --enable-gather-detailed-stats
since you need MEM_STAT_INFO.
I would wote for returning ggc_realloc define, I guess it went away
only because it is currently unused (all callers of ggc_realloc are
_stat functions themselves).

Honza
>
>

> Index: ChangeLog-ira-reload
> ===================================================================
> *** ChangeLog-ira-reload	(revision 160799)
> --- ChangeLog-ira-reload	(working copy)
> ***************
> *** 1,3 ****
> --- 1,7 ----
> + 2010-06-16  Jeff Law <law@redhat.com>
> + 
> + 	* ira-reload.c (localize_pseudos): Use ggc_realloc_stat.
> + 
>   2010-06-09  Jeff Law <law@redhat.com>
>   
>   	* ira-build.c (initiate_cost_vectors): Initiate a cost vector for
> Index: ira-reload.c
> ===================================================================
> *** ira-reload.c	(revision 160799)
> --- ira-reload.c	(working copy)
> *************** localize_pseudos (basic_block bb, bitmap
> *** 899,905 ****
>         VEC_safe_grow_cleared (rtx, gc, reg_equiv_memory_loc_vec, max_regno);
>         reg_equiv_memory_loc = VEC_address (rtx, reg_equiv_memory_loc_vec);
>         reg_equiv_init
> ! 	= (rtx *) ggc_realloc (reg_equiv_init, max_regno * sizeof (rtx));
>         memset (&reg_equiv_init[orig_max_reg_num], 0, nregs * sizeof (rtx));
>         regstat_n_sets_and_refs
>   	 = ((struct regstat_n_sets_and_refs_t *)
> --- 899,905 ----
>         VEC_safe_grow_cleared (rtx, gc, reg_equiv_memory_loc_vec, max_regno);
>         reg_equiv_memory_loc = VEC_address (rtx, reg_equiv_memory_loc_vec);
>         reg_equiv_init
> ! 	= (rtx *) ggc_realloc_stat (reg_equiv_init, max_regno * sizeof (rtx) PASS_MEM_STAT);
>         memset (&reg_equiv_init[orig_max_reg_num], 0, nregs * sizeof (rtx));
>         regstat_n_sets_and_refs
>   	 = ((struct regstat_n_sets_and_refs_t *)

Re: [ reload-v2]

[Jeff Law]

>
>> ggc_realloc is no longer supported, ggc_realloc_stat is the proper way
>> to do things now.   This allows reload-v2 to bootstrap after monday's
>> merge from the trunk.
>>      
> Well, it was definitly not intention of my statistics code to require
> users to add random PASS_MEM_STAT.
> There should be always corresponding macro hidding that.
>    
OK.  I just found another ggc_realloc call and copied what it did.

> Also your patch will not compile with --enable-gather-detailed-stats
> since you need MEM_STAT_INFO.
>    
I'll take care of this.


Thanks,
Jeff

Re: [ reload-v2]

[Jan Hubicka]

>>
>>> ggc_realloc is no longer supported, ggc_realloc_stat is the proper way
>>> to do things now.   This allows reload-v2 to bootstrap after monday's
>>> merge from the trunk.
>>>      
>> Well, it was definitly not intention of my statistics code to require
>> users to add random PASS_MEM_STAT.
>> There should be always corresponding macro hidding that.
>>    
> OK.  I just found another ggc_realloc call and copied what it did.

Well, it is because it was inside _stat function.
those gets MEM_STAT_DECL that compiles to nothing unless you do detailed
mem stats and then it is source file and line, submitted by MEM_STAT_INFO.
However for some functions (like vec_reallocate) we don't want to record
the function itself and rather its callers, so then you create _stat
version of it getting MEM_STAT_DECL and use PASS_MEM_STAT to pass the info
further.

Honza
>
>> Also your patch will not compile with --enable-gather-detailed-stats
>> since you need MEM_STAT_INFO.
>>    
> I'll take care of this.
>
>
> Thanks,
> Jeff

Re: [ reload-v2]

[Laurynas Biveinis]

2010/6/16 Jeff Law <law@redhat.com>:
>
> ggc_realloc is no longer supported, ggc_realloc_stat is the proper way to do
> things now.   This allows reload-v2 to bootstrap after monday's merge from
> the trunk.

There are two ggc_realloc replacements: GGC_RESIZEVEC and
GGC_RESIZEVAR. Do they work for you?

-- 
Laurynas

Re: [ reload-v2]

[Jeff Law]

On 06/16/10 23:15, Laurynas Biveinis wrote:
> 2010/6/16 Jeff Law<law@redhat.com>:
>    
>> ggc_realloc is no longer supported, ggc_realloc_stat is the proper way to do
>> things now.   This allows reload-v2 to bootstrap after monday's merge from
>> the trunk.
>>      
> There are two ggc_realloc replacements: GGC_RESIZEVEC and
> GGC_RESIZEVAR. Do they work for you?
>    
I simply wasn't aware of them.  I'm testing with GGC_RESIZEVEC right now.

jeff