Document how to build PGO-optimized GCC version

2022-12-27 Thread Alexander Zaitsev

Hello.

We are using GCC for our C++ projects. Our projects are huge, commit 
rate is quite huge, so our CI workers are always busy (so as any other 
CI workers, honestly). Since we want to increase build speed, one of the 
option is to optimize the compiler itself. Sounds like a good case for PGO.


Clang has the infrastructure for building the Clang itself with PGO: 
https://llvm.org/docs/HowToBuildWithPGO.html . I have tried to find 
something like that for GCC but with no success.


My proposal is:

 * add support for building PGO-optimized GCC into the GCC build
   infrastructure
 * add documentation to the GCC site, how to build GCC with PGO
   optimizations
 * (if GCC community provides prebuilt gcc binaries) use PGO for the
   prebuilt binaries. E.g. Clang and rustc already uses this approach.

Any feedback is appreciated.

Thanks in advance!

--
Best regards,
Alexander Zaitsev


Re: Document how to build PGO-optimized GCC version

2022-12-27 Thread Andrew Pinski via Gcc
On Tue, Dec 27, 2022 at 9:38 PM Alexander Zaitsev  wrote:
>
> Hello.
>
> We are using GCC for our C++ projects. Our projects are huge, commit
> rate is quite huge, so our CI workers are always busy (so as any other
> CI workers, honestly). Since we want to increase build speed, one of the
> option is to optimize the compiler itself. Sounds like a good case for PGO.
>
> Clang has the infrastructure for building the Clang itself with PGO:
> https://llvm.org/docs/HowToBuildWithPGO.html . I have tried to find
> something like that for GCC but with no success.
>
> My proposal is:
>
>   * add support for building PGO-optimized GCC into the GCC build
> infrastructure
>   * add documentation to the GCC site, how to build GCC with PGO
> optimizations

It is already there, see the last section of
https://gcc.gnu.org/install/build.html (Building with profile
feedback).  It has been included in GCC and documented since June
2003.
https://gcc.gnu.org/r0-50361-g8f231b5d874dcb .

>   * (if GCC community provides prebuilt gcc binaries) use PGO for the
> prebuilt binaries. E.g. Clang and rustc already uses this approach.

GCC community does not provide prebuilt gcc binaries for a few
different reasons.
But distros do provide more recent prebuilt binaries, you could ask
them to build using PGO (some do already I think).

Thanks,
Andrew Pinski

>
> Any feedback is appreciated.
>
> Thanks in advance!
>
> --
> Best regards,
> Alexander Zaitsev


Re: Document how to build PGO-optimized GCC version

2022-12-27 Thread NightStrike via Gcc
On Wed, Dec 28, 2022, 00:37 Alexander Zaitsev  wrote:

> Hello.
>
> We are using GCC for our C++ projects. Our projects are huge, commit
> rate is quite huge, so our CI workers are always busy (so as any other
> CI workers, honestly). Since we want to increase build speed, one of the
> option is to optimize the compiler itself. Sounds like a good case for PGO.
>
> Clang has the infrastructure for building the Clang itself with PGO:
> https://llvm.org/docs/HowToBuildWithPGO.html . I have tried to find
> something like that for GCC but with no success.
>
> My proposal is:
>
>   * add support for building PGO-optimized GCC into the GCC build
> infrastructure
>   * add documentation to the GCC site, how to build GCC with PGO
> optimizations
>   * (if GCC community provides prebuilt gcc binaries) use PGO for the
> prebuilt binaries. E.g. Clang and rustc already uses this approach.
>
> Any feedback is appreciated.
>
> Thanks in advance!
>
> --
> Best regards,
> Alexander Zaitsev
>

I would wager that you would get more bang for your buck out of 1) building
a more recent gcc yourself instead of using whatever comes packaged, and 2)
building with march=native for each processor type you run on. Not that PGO
won't help, of course it will since you are building the same software
repeatedly, but my personal experience doing the exact same thing is that I
saw a 50% performance improvement from just from that, and it's rather
trivial to do if your infrastructure is homogeneous (mine was a many node
compute cluster). The next biggest bottleneck for me was IO, because the
project this was for when compiling was very file intensive.

Anyway, just some alternative suggestions, since there's already a response
giving you what you asked for. Feel free to ignore me :)

>