[llvm] r324557 - gold-plugin: Do not set codegen opt level based on LTO opt level.

[Peter Collingbourne via llvm-commits]

Unfortunately reverting would regress the user that requested this patch.
They use -plugin-opt=O0 and expect performance consistent with setting cg
opt level to default.

Instead of reverting can we make a change that just sets the cg opt level
to aggressive if -plugin-opt=O3 is passed?

Peter


On Fri, Apr 6, 2018, 09:18 Chad Rosier <mcrosier at codeaurora.org> wrote:

> Thanks, Rafael.  Sound reasonable to you, Peter?
>
>
> On 4/6/2018 12:12 PM, Rafael Avila de Espindola wrote:
> > I am OK with reverting it until we have a better framework in place.
> >
> > Cheers,
> > Rafael
> >
> > Chad Rosier via llvm-commits <llvm-commits at lists.llvm.org> writes:
> >
> >> Peter/Rafael,
> >>
> >> Over the past few days I've been investigating several regressions with
> >> our weekly LTO configuration.  As a result, I identified that the
> >> following regressions were due to this commit:
> >>
> >> SPEC2006/h264ref: -2.68%
> >> SPEC2006/libquantum: -4.36%
> >> SPEC2006/perlbench: -2.67%
> >> SPEC2017/mcf: -4.47%
> >> SPEC2017/perlbench: -3.82%
> >>
> >> I did not test SPEC2000, SPEC2006fp, or SPEC2017fp, but I'm seeing
> >> similar small regressions in our weekly runs there as well (i.e., the
> >> above is just what I was able to test last night, but isn't likely the
> >> only regressions).  I'd like to kindly request that this change be
> >> reverted until something is in place to properly propagate the codegen
> >> optimization level.
> >>
> >>    Chad
> >>
> >>
> >> On 2/7/2018 9:41 PM, Peter Collingbourne via llvm-commits wrote:
> >>> Author: pcc
> >>> Date: Wed Feb  7 18:41:22 2018
> >>> New Revision: 324557
> >>>
> >>> URL: http://llvm.org/viewvc/llvm-project?rev=324557&view=rev
> >>> Log:
> >>> gold-plugin: Do not set codegen opt level based on LTO opt level.
> >>>
> >>> The LTO opt level should not affect the codegen opt level, and indeed
> >>> it does not affect it in lld. Ideally the codegen opt level should
> >>> be controlled by an IR-level attribute based on the compile-time opt
> >>> level, but that hasn't been implemented yet.
> >>>
> >>> Differential Revision: https://reviews.llvm.org/D43040
> >>>
> >>> Modified:
> >>>       llvm/trunk/tools/gold/gold-plugin.cpp
> >>>
> >>> Modified: llvm/trunk/tools/gold/gold-plugin.cpp
> >>> URL:
> http://llvm.org/viewvc/llvm-project/llvm/trunk/tools/gold/gold-plugin.cpp?rev=324557&r1=324556&r2=324557&view=diff
> >>>
> ==============================================================================
> >>> --- llvm/trunk/tools/gold/gold-plugin.cpp (original)
> >>> +++ llvm/trunk/tools/gold/gold-plugin.cpp Wed Feb  7 18:41:22 2018
> >>> @@ -727,20 +727,6 @@ static int getOutputFileName(StringRef I
> >>>      return FD;
> >>>    }
> >>>
> >>> -static CodeGenOpt::Level getCGOptLevel() {
> >>> -  switch (options::OptLevel) {
> >>> -  case 0:
> >>> -    return CodeGenOpt::None;
> >>> -  case 1:
> >>> -    return CodeGenOpt::Less;
> >>> -  case 2:
> >>> -    return CodeGenOpt::Default;
> >>> -  case 3:
> >>> -    return CodeGenOpt::Aggressive;
> >>> -  }
> >>> -  llvm_unreachable("Invalid optimization level");
> >>> -}
> >>> -
> >>>    /// Parse the thinlto_prefix_replace option into the \p OldPrefix
> and
> >>>    /// \p NewPrefix strings, if it was specified.
> >>>    static void getThinLTOOldAndNewPrefix(std::string &OldPrefix,
> >>> @@ -767,7 +753,6 @@ static std::unique_ptr<LTO> createLTO(In
> >>>
> >>>      Conf.MAttrs = MAttrs;
> >>>      Conf.RelocModel = RelocationModel;
> >>> -  Conf.CGOptLevel = getCGOptLevel();
> >>>      Conf.DisableVerify = options::DisableVerify;
> >>>      Conf.OptLevel = options::OptLevel;
> >>>      if (options::Parallelism)
> >>>
> >>>
> >>> _______________________________________________
> >>> llvm-commits mailing list
> >>> llvm-commits at lists.llvm.org
> >>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits
> >> _______________________________________________
> >> llvm-commits mailing list
> >> llvm-commits at lists.llvm.org
> >> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20180406/54ea28bd/attachment.html>