[Bug tree-optimization/80155] [7 regression] Performance regression with code hoisting enabled

[rguenther at suse dot de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80155

--- Comment #6 from rguenther at suse dot de <rguenther at suse dot de> ---
On Thu, 23 Mar 2017, thopre01 at gcc dot gnu.org wrote:

> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80155
> 
> --- Comment #5 from Thomas Preud'homme <thopre01 at gcc dot gnu.org> ---
> (In reply to Richard Biener from comment #4)
> > Can you benchmark the 2nd candidate (most appropriate at this stage I think,
> > would also fix 77498).
> > 
> > It will cause pessimizations for code that benefits from predictive
> > commoning transforms that PRE can do at -O2 because predictive commoning
> > doesn't run at -O2.
> > The fix for that would be to run predictive commoning at -O2 but in a mode
> > that doesn't perform unrolling for example.  [even pcom can blow through
> > register pressure limits though]
> > 
> > Any other approach to fixing this particular bug cannot be done for GCC 7
> > (even this one is quite late ...).  So in the end I'd vote for WONTFIX for
> > GCC 7.
> 
> Sadly it doesn't. Well not quite true, there is a very slight improvement but
> we're far from the uplift -fno-code-hoisting provides.
> 
> There's maybe more than one bug at play and the reduced testcase only show one.

The testcase shows hardly profitable PRE which the patches should
disable (I didn't verify the patches fix the testcase!)

code-hoisting (as well as CSE) can increase register pressure and neither
transform is register-pressure aware (and the patch doesn't try to
mitigate that).