Re: [tree-ssa] cfg.texi needs reviewing by a native speaker (Was: Re: "Documentation by paper")

[Jan Hubicka <jh at suse dot cz>]

Hi,
I've run the SPEC on tree-ssa and 3.4 branch and get following overall
scores for tree-ssa

HEURISTICS                  BRANCHES  (REL)  HITRATE             COVERAGE  (REL)
DS theory                      31543  54.5%  74.89%/ 90.42%   9625583876  27.9%
combined                       57891 100.0%  79.12%/ 90.81%  34477566666 100.0%
opcode values nonequal          9358  16.2%  73.12%/ 88.57%   3505203791  10.2%
no prediction                  16383  28.3%  65.63%/ 86.73%  10242545226  29.7%
call                           17308  29.9%  71.64%/ 93.67%   3981422239  11.5%
first match                     9965  17.2%  91.38%/ 93.92%  14609437564  42.4%
loop branch                     4487   7.8%  91.33%/ 93.80%   6166574467  17.9%
opcode values positive          2285   3.9%  83.26%/ 85.90%   1465042435   4.2%
loop header                     4960   8.6%  63.76%/ 89.05%   1211747096   3.5%
loop exit                       4497   7.8%  91.16%/ 94.22%   7165216605  20.8%
loop iterations                 1037   1.8%  91.57%/ 91.57%   2226735613   6.5%
negative return                  245   0.4%  98.98%/ 99.97%     54555025   0.1%
pointer                         5173   8.9%  85.69%/ 95.23%   1443841057   4.2%
noreturn call                    657   1.1%  99.99%/ 99.99%    252431954   0.7%
const return                     211   0.4%  87.87%/ 97.13%     82434673   0.2%
early return                    1580   2.7%  27.22%/ 94.68%    186652617   0.5%
null return                      146   0.3%  93.82%/ 96.43%     46507848   0.1%

and for mainline

HEURISTICS                  BRANCHES  (REL)  HITRATE             COVERAGE  (REL)
DS theory                      31490  53.3%  71.79%/ 91.00%  10839022579  30.2%
combined                       59027 100.0%  76.61%/ 91.65%  35936313076 100.0%
opcode values nonequal          9122  15.5%  71.77%/ 89.50%   3896571147  10.8%
no prediction                  17112  29.0%  60.53%/ 88.63%   9999337183  27.8%
call                           17124  29.0%  73.00%/ 94.71%   4441597166  12.3%
first match                    10425  17.7%  90.72%/ 94.12%  15097953314  42.0%
loop branch                     5612   9.5%  89.66%/ 93.51%   7902327865  22.0%
opcode values positive          2466   4.2%  78.01%/ 83.39%   1364314266   3.8%
loop header                     5861   9.9%  63.88%/ 90.22%   1695466332   4.7%
continue                         635   1.1%  56.05%/ 86.42%   1107595846   3.1%
loop exit                       3838   6.5%  91.71%/ 95.23%   5934599476  16.5%
loop iterations                 1130   1.9%  91.58%/ 91.58%   2231770138   6.2%
negative return                  244   0.4%  98.99%/ 99.98%     54548576   0.1%
pointer                         3611   6.1%  86.64%/ 95.38%   1168152743   3.2%
noreturn call                    577   1.0%  99.99%/100.00%    206199365   0.6%
const return                     202   0.3%  97.15%/ 98.88%     83294074   0.2%
early return                    1512   2.6%  27.61%/ 94.73%    182473757   0.5%
null return                      150   0.3%  93.53%/ 96.86%     42332230   0.1%
goto                             334   0.6%  73.24%/ 90.10%     24026456   0.1%

This is quite interesting as the tree-ssa looks a lot better overall -
fewer branches and has better predictions.  Pointer heuristics has
slightly lower accurace, but better coverage so I would say it works
pretty well overall.  Still splitting it into two heuristics - one
predicting variables other arrays/structures may be interesting.

It is interesting how the loop branch heuirstics changed.
I am attacing the per-test analysis.

Honza

Attachment: cmp.tgz
Description: GNU Unix tar archive