WHOPR partitioning algorithm

[Xinliang David Li]

On Mon, Sep 6, 2010 at 12:48 PM, Jack Howarth <howarth@bromo.med.uc.edu> wrote:
> On Mon, Sep 06, 2010 at 11:00:40AM -0700, Xinliang David Li wrote:
>> I have not followed the thread carefully, but what is the advantage of
>> doing code layout in IPO instead of doing this at link time? The
>> support of user directed code and data layout is already in gold  (and
>> mozilla folks have tried this functionality to layout functions
>> according to load time affinity data collected by instrumentation and
>> got large speed up).  The remaining part of the support is also almost
>> ready but not yet submitted upstream (needs more tuning on various
>> targets): 1) the compiler part of the support (e.g. writing down call
>> graph hot edges as note sections for linker to consume; 2) the code
>> layout algorithm in linker using the data from the compiler.
>>
>
>   Wouldn't that present issues for other targets like darwin and cygwin
> that can't depend on such support in the linker?

Yes, that is true. Both methods have usability limitations.

> Also, wouldn't such a change
> actually be be heading FSF gcc's LTO design towards llvm's design of using
> a libLTO linker module?

I don't see a connection here. It does not change how gcc's LTO works
and they can co-exist.

David

>               Jack
>
>> Thanks,
>>
>> David
>>
>