A recent patch increased GCC's memory consumption in some cases!

[Jan Hubicka]

> The heuristics of the old  SRA decide not to scalarize D.17614 whereas
> the sub-access  propagation of the  new one leads to  scalarization of
> this aggregate.  There is no  fwprop run afterwards and so this bigger

This seems like sane decision.

> 					Old -O2		New -O2
>
> maxmem2 Overall allocated memory:	1194833 kB	1312909 kB
> mem-report Total Allocated:		396955168	409442360
> Decls:	   	 			129778		118316
> SSA_NAME nodes allocated:		217116		220153
> SSA_NAME nodes reused:			54964		49655

Hmm, so we don't really have that many of new trees created but still  
have so much of mem use growth?  It seems almost like we are  
exercising some new kind of memory leak in one of hand allocated  
sturctures?

Honza
> refs:	       				38447		10057
> exprs:					349178		371744
> assignments:   				212847		210677
> refs_may_alias_p disambiguations:	693		0
> refs_may_alias_p queries:		1421		0
> ref_maybe_used_by_call_p disambig.:	0		0
> ref_maybe_used_by_call_p queries:	1420		0
> pt_solutions_intersect disambig.:	0		0
> pt_solutions_intersect queries:		44		44
>
> SRA Alloc pools:
>
> SRA accesses              128         36   14972032(    116969)     
> 14949504(    116793)          0(         0)
> SRA links                  40         36     935760(     23394)       
> 935280(     23382)          0(         0)
>
>
> SRA bitmaps:
>
> tree-sra.c:463 (sra_initialize)                 36           46880    
>          44520               0     465834
> tree-sra.c:464 (sra_initialize)                 36            2320    
>             80               0         44
>
>
> SRA heap vectors:
>
> tree-sra.c:607 (create_access)                            0: 0.0%     
>    11920              27: 0.0%
> tree-sra.c:605 (create_access)                            0: 0.0%     
> 17301240           34354: 6.2%
>