Automatic PGO - Initial implementation (1/N)
echristo at gmail.com
Wed Sep 25 08:04:59 PDT 2013
On Wed, Sep 25, 2013 at 6:20 AM, Diego Novillo <dnovillo at google.com> wrote:
> On Wed, Sep 25, 2013 at 1:25 AM, Kostya Serebryany <kcc at google.com> wrote:
>> On Wed, Sep 25, 2013 at 12:38 AM, Renato Golin <renato.golin at linaro.org>
>>> On 24 September 2013 19:38, Eric Christopher <echristo at gmail.com> wrote:
>>>> -gline-tables-only (or -gmlt in gcc parlance). This will give you
>>>> accurate call stack information for each sample. Or at least as
>>>> accurate as I can get it.
>>> FWIW, I'm using perf with "-O3 -g" and LLVM has retained the line
>>> information quite well throughout all the inlining and optimizations that
>>> I assume it's easier to keep line information than variable location
>>> ranges and types. The biggest problem is to separate the two emissions, I
>>> guess, not the quality of the resulting information.
>> We've been using -gline-tables-only for a long time and are satisfied with
>> the quality.
>> It even gives use inlined frames (when we ask).
> OK, I think that just having line number information should be enough.
> I take it that it would not be a problem if I forced
> -gline-tables-only in the front end when -mauto-profile is used?
Nope, but -mauto-profile? Why not -fauto-profile?
More information about the llvm-commits