[Bug ld/26766] Support standalone PGO build

[hjl.tools at gmail dot com]

https://sourceware.org/bugzilla/show_bug.cgi?id=26766

--- Comment #5 from H.J. Lu <hjl.tools at gmail dot com> ---
(In reply to Martin Liska from comment #4)
> > Please take a look at users/hjl/pgo-lto branch
> > 
> > https://gitlab.com/x86-binutils/binutils-gdb/-/commits/users/hjl/pgo-lto
> > 
> > If check-binutils check-gas check-ld training is good enough, I can drop
> > "make install" and "make check" steps.
> 
> That's very promising!
> I have a couple of comments:
> 
> 1) I would separate --enable-standalone-pgo-binutils=lto into 2
> configuration options: --enable-lto and --enable-pgo-bootstrap
>

The way I implement it works better with --enable-pgo-bootstrap=lto since
-flto is passed together with PGO options.  --enable-lto can't work without
--enable-pgo-bootstrap.

> 2) --enable-lto should add -flto=jobserver instead of -flto, but I noticed
> the following warning when doing that:
> 
> lto-wrapper: warning: jobserver is not available: '--jobserver-auth=' is not
> present in 'MAKEFLAGS'
> 
> That's likely cased by missing '+' as documented:
> You must prepend a ‘+’ to the command recipe in the parent Makefile for this
> to work.
> 
> Can you please check it?

I will give it a try.

> 3) I would use only -fprofile-generate without an additional folder
> location; or do you see a .gcda file name clashing?

I will give it a try.

> 4) I can confirm that gas is trained well; I'm testing that for a Firefox
> LTRANS file (~100MB in size).
> 
> Anyway, great progress you made.

Good to know.

-- 
You are receiving this mail because:
You are on the CC list for the bug.