Re: [PATCH][RFC] Add new ipa-reorder pass

[Fangrui Song <i at maskray dot me>]

On 2019-10-03, Andrew Pinski wrote:
> On Wed, Oct 2, 2019@9:52 PM Fangrui Song <i@maskray.me> wrote:
> > On 2019-09-24, Martin Liška wrote:
> > >On 9/19/19 10:33 AM, Martin Liška wrote:
> > >> - One needs modified binutils and I that would probably require a configure detection. The only way
> > >>   which I see is based on ld --version. I'm planning to make the binutils submission soon.
> > >
> > >The patch submission link:
> > >https://sourceware.org/ml/binutils/2019-09/msg00219.html
> >
> > Hi Martin,
> >
> > I have a question about why .text.sorted.* are needed.
> >
> > The Sony presentation (your [2] link) embedded a new section
> > .llvm.call-graph-profile[3] to represent edges in the object files. The
> > linker (lld) collects all .llvm.call-graph-profile sections and does a
> > C3 layout. There is no need for new section type .text.sorted.*
> >
> > [3]: https://github.com/llvm/llvm-project/blob/master/lld/test/ELF/cgprofile-obj.s
> >
> > (Please CC me. I am not subscribed.)
>
> The idea of GCC's version is to that the modification needed to the
> linker is very little.  And even without patching the linker
> script/linker, you don't need to much special and you don't need LD to
> do much work@all.

I am afraid this can be a large limitation. Then .text.sorted.* can
only a) reorder functions within a translation unit, b) or reorder all
functions when LTO is enabled.

b) is possible only if all .text.sorted.* sections can be numbered.

For the LLVM case, call graph profile can be used without LTO. When both
ThinLTO+PGO are enabled, however, the additional performance improvement
offered by the global call graph profile reordering is insignificant,
smaller than 1%.