[tree-ssa] Miscompilation (SIGSEGV) with -ffast-math (!?), huge compile time regression

[Diego Novillo]

On Mon, 2004-02-23 at 09:01, Richard Guenther wrote:

> g++ (GCC) 3.5-tree-ssa 20040217 (merged 20040211) endlessly loops,
> g++ (GCC) 3.5-tree-ssa 20040209 (merged 20040126) is fine.  Note that 
> between these two a huge compile time regression occoured, 20040209 
> takes 6m36s, 20040217 over 12m!
> 
Yeah, it certainly takes a while to build (roughly 4m30s on a 2.2Ghz P4
w/ 1Gb of RAM).  It also makes cc1plus to grow quite big (about 480Mb).

The biggest time consumers according to -ftime-report are

 dominator optimization:  47.83 (20%) usr   0.27 ( 4%) sys  48.29 (20%) wall
 expand                :  46.79 (20%) usr   0.08 ( 1%) sys  46.95 (19%) wall
 tree DSE              :  23.53 (10%) usr   0.01 ( 0%) sys  23.55 (10%) wall
 tree SSA rewrite      :  16.31 ( 7%) usr   0.00 ( 0%) sys  16.31 ( 7%) wall
 integration           :  13.86 ( 6%) usr   0.07 ( 1%) sys  13.93 ( 6%) wall


> So maybe its the merge from mainline that brought all these problems? 
> Though enough, I never checked mainline, but stayed with the 3.4 branch 
> which is fine in both regards.
> 
The compile time regression you experienced is probably not all because
of the merge.  We are spending a significant chunk of time in the tree
optimizers.


Diego.