On Fri, May 06, 2016 at 04:41:41PM +0200, Martin Liška wrote:
> On 05/06/2016 03:25 PM, Jakub Jelinek wrote:
> > Well, we already have the gimple poisoning/unpoisoning code on RTL (emitted
> > after the prologue and before the epilogue), so it shouldn't be that hard.
> > I'd only do the most common/easy cases inline though, like 1/2/4/8/16/32
> > bytes long variables.
> >
> > Jakub
>
> You are right, I didn't realize it earlier.
> As I've collected statistics for tramp3d, poisoning code has following
> distribution:
>
> 4:1.62%
> 8:3.53%
> 12:94.76%
>
> which is quite interesting that 12B are such a common size :)
> Probably due to a lot of time spent in ::evaluate (MultiArgEvaluator and
> MultiArgEvaluator).
> Considering just variables which needs_to_live_in_memory, tramp3d is still
> ~15x slower.
Please look at other testcases, not just tramp3d - we in the end don't want
to tune it to just tramp3d. Pick up some 3-4 C/C++ benchmarks, tramp3d can
be one of them ;)
Jakub
