Stream: t-compiler

Topic: what is perf.rlo measuring?


pnkfelix (Apr 05 2019 at 10:26, on Zulip):

Okay, I'm sorry to admit that I don't immediately remember the answer to this question ...

pnkfelix (Apr 05 2019 at 10:26, on Zulip):

(but I also don't feel that bad, since I think this information should be more apparent in the presentation itself)

pnkfelix (Apr 05 2019 at 10:26, on Zulip):

... but: Is perf.rlo measuring the performance of the generated object binary? Or is it measuring the compilation time?

pnkfelix (Apr 05 2019 at 10:27, on Zulip):

I'm looking in particular at stats like https://perf.rust-lang.org/compare.html?start=eab3eb38df8dca99110b6149b3a15deeb4ef0413&end=4cb14446465d8d3cabfb3706f424a2938628b9f6&stat=wall-time

Zoxc (Apr 05 2019 at 10:27, on Zulip):

It's all compilation time

pnkfelix (Apr 05 2019 at 10:27, on Zulip):

of a bootstrapped compiler?

Zoxc (Apr 05 2019 at 10:28, on Zulip):

Yes

pnkfelix (Apr 05 2019 at 10:28, on Zulip):

i.e.when we use perf to measure performance of a change like #59546

pnkfelix (Apr 05 2019 at 10:28, on Zulip):

It would not surprise me if compilation took longer

pnkfelix (Apr 05 2019 at 10:28, on Zulip):

since e.g. the compiler may be working harder to try to optimize the code in question

pnkfelix (Apr 05 2019 at 10:29, on Zulip):

the more important thing for me to know about a case like that is "how did the object code performance change?"

pnkfelix (Apr 05 2019 at 10:29, on Zulip):

(am I not looking at this problem the right way?)

Zoxc (Apr 05 2019 at 10:29, on Zulip):

Then https://github.com/anp/lolbench is for you

pnkfelix (Apr 05 2019 at 10:29, on Zulip):

okay, and we should be striving to integrate that with @rust-timer in some way?

Zoxc (Apr 05 2019 at 10:36, on Zulip):

Probably

simulacrum (Apr 05 2019 at 11:45, on Zulip):

(lolbench is pretty separate though so that might be long-term)

Last update: Nov 16 2019 at 02:40UTC