On Wed, Feb 15, 2023 at 20:15:37 -1000, Richard Henderson wrote:
> On 2/10/23 02:35, Emilio Cota wrote:
> > I ran yesterday linux-user SPEC06 benchmarks from your tcg-life branch.
> > I do see perf regressions for two workloads (sjeng and xalancbmk).
> > With perf(1) I see liveness_pass* are at 0.00%, so I wonder: is it
> > possible that the emitted code isn't quite the same?
>
> Everything that I checked by hand was the same, but it's possible.
> It's a tedious process. You'd definitely want to turn off ASR.
I've checked with -jitdump and perf whether there was any difference
in the generated code before vs. after for the most common TBs.
They were identical.
Benchmarking without ASR didn't make a difference, unfortunately.
> My current branch has __attribute__((noreturn)) added to all of the liveness
> passes, so that they don't get folded into tcg_gen_code. But I still would
> expect 0%.
I'll bisect the series in the next few days see exactly where
the perf regression begins so that at least we know where to look.
Thanks,
Emilio
