[tree-ssa] POOMA compile time / memory requirement comparison

[Richard Guenther]

Mark Mitchell wrote:
>>> I see two options here: (a) unfreeze the branch and do whatever work is
>>> necessary to bring POOMA in line with mainline, (b) continue with the
>>> merge and address these problems in mainline.
>>>
>>> IMO option (a) is too costly and may only serve to delay tree-ssa even
>>> more.  I suspect that much of the necessary work to reduce this gap
>>> involves removing RTL passes and/or doing major surgery on trees, memory
>>> allocation (the system times are pretty bad).
>>
>>
> 
> A compromise position would be to attempt to prove (by measuring) that 
> eliminating the RTL passes would help.  Can you quantify the time spent 
> in the RTL passes that will hopefully soon be eliminated?

Another way is to make tree-ssa optimize gcc itself as good as mainline.
A hint that this is presently not the case (at least for ia32 target) is
-fsyntax-only times of f.i. POOMA tramp3d-v3 testcase:

tree-ssa:
Execution times (seconds)
  garbage collection    :   1.70 ( 6%) usr   0.00 ( 0%) sys   1.72 ( 6%) 
wall
  preprocessing         :   0.66 ( 2%) usr   0.19 ( 7%) sys   0.89 ( 3%) 
wall
  parser                :  15.50 (56%) usr   1.32 (45%) sys  17.02 (55%) 
wall
  name lookup           :   5.61 (20%) usr   1.29 (44%) sys   6.95 (22%) 
wall
  tree gimplify         :   3.68 (13%) usr   0.10 ( 3%) sys   3.80 (12%) 
wall
  varconst              :   0.68 ( 2%) usr   0.00 ( 0%) sys   0.68 ( 2%) 
wall
  symout                :   0.01 ( 0%) usr   0.00 ( 0%) sys   0.01 ( 0%) 
wall
  TOTAL                 :  27.86             2.91            31.10

mainline:
Execution times (seconds)
  preprocessing         :   0.72 ( 4%) usr   0.17 ( 6%) sys   0.90 ( 4%) 
wall
  parser                :  13.29 (68%) usr   1.32 (50%) sys  14.76 (66%) 
wall
  name lookup           :   5.06 (26%) usr   1.09 (42%) sys   6.23 (28%) 
wall
  varconst              :   0.57 ( 3%) usr   0.00 ( 0%) sys   0.57 ( 3%) 
wall
  TOTAL                 :  19.65             2.62            22.51

That would be a 30% regression for -fsyntax-only.  Ugh.  Can anyone 
confirm this?  Also, why do we gimplify and symout for -fsyntax-only?

Richard.