[PATCH] D19528: [ELF] - Implemented -z combrelocs/nocombreloc.

[Rafael Espíndola via llvm-commits]

On 6 May 2016 at 19:39, George Rimar <grimar at accesssoftek.com> wrote:
>>I was able to manually apply it.
>>
>>What I did was:
>>
>>* build clang with shared libraries.
>>* time with
>  >sudo schedtool -F -p 99 -a 0x4 -e perf stat -r 10 ./bin/clang
>>* build the patched version of lld
>>* delete bin/clang-3.9 and lib/*.so
>>* link again
>>* time the result
>>
>>What I got was
>>
>>* trunk: 0.059005660
>>* sort on RELATIVE, Symbol, offset, type: 0.040090757
>>* sort on Symbol, offset, type: 0.042367124
>>* sort on offset, type: 0.059119353
>>* sort on RELATIVE, Symbol, offset: 0.040131764
>>* sort on RELATIVE, Symbol: 0.040396584
>>* sort on RELATIVE: 0.058322444
>>* sort on Symbol: 0.039886032
>>
>>So it looks like the only thing that makes a difference is the symbol.
>>
>>BTW, one cannot sort .rel.plt, it would be nice to document why.
>>I think it is better to use stable_sort to be sure we have a
>>deterministic output without checking fields that don't make a
>>difference.
>>
>>So a change sorting only by symbol with stable sort LGTM. Please make
>>sure you are not computing the number of relative rolocations or
>>setting any DT_*.
>>
>>The partial apply that I used for benchmarking was.
>>
>
> I'll update the patch a bit later then for final review.
>
> At the same time I got 10% difference in startup cycles
> when compared with DT_* vs without: http://reviews.llvm.org/D19528#416734
> Do you think that was an computing error ?

Maybe. In any case, it would be better to commit first the simple
patch and then benchmark that independently.

Cheers,
Rafael