[FFmpeg-devel] [PATCH 1/2] ffbuild: compose linker response files in a loop

[Gyan Doshi]

On 2025-03-26 11:37 pm, Andreas Rheinhardt wrote:
> Gyan Doshi:
>>
>> On 2025-03-25 05:11 pm, Ramiro Polla wrote:
>>> Hi,
>>>
>>> On Sat, Mar 22, 2025 at 5:33 AM Gyan Doshi <ffmpeg at gyani.pro> wrote:
>>>> On 2025-03-22 03:29 am, Martin Storsjö wrote:
>>>>> On Fri, 21 Mar 2025, Gyan Doshi wrote:
>>>>>> ffbuild/library.mak | 7 +++++--
>>>>>> 1 file changed, 5 insertions(+), 2 deletions(-)
>>>>>>
>>>>>> diff --git a/ffbuild/library.mak b/ffbuild/library.mak
>>>>>> index 7e1871b74c..15302852ec 100644
>>>>>> --- a/ffbuild/library.mak
>>>>>> +++ b/ffbuild/library.mak
>>>>>> @@ -36,7 +36,8 @@ endif
>>>>>> $(SUBDIR)$(LIBNAME): $(OBJS) $(STLIBOBJS)
>>>>>>       $(RM) $@
>>>>>> ifeq ($(AR_OBJS),true)
>>>>>> -    $(Q)echo $^ > $@.objs
>>>>>> +    -$(RM) $@.objs
>>>>>> +    $(Q)$(foreach ARG,$^,echo -n "$(ARG) " >> $@.objs;)
>>>>> Does this instance even work, it looks broken, like it is missing
>>>>> something?
>>>> Tested both static and shared building. All working with no noticeable
>>>> speed difference although I didn't formally bench.
>>> It's generally a good idea to check for and report the speed
>>> differences. Even if just for the sake of curiosity.
>> So, with this configure:
>>
>>    --disable-doc --disable-w32threads --enable-pthreads --disable-
>> autodetect --cc="ccache gcc" --enable-static --enable-shared --extra-
>> ldflags="-static" --enable-gpl --enable-version3
>>
>> I ran one priming run + 5 repeat runs of    `time make -j4`.
>>
>> Before, real ranged from 2m10.188s to 2m12.154s
>> After, real ranged from 2m12.486s to 2m13.405s
>>
>> I also ran one priming run + 5 repeat runs of    `time make -j4
>> libavcodec/libavcodec.a libavcodec/avcodec.dll`.
>>
>> Before, real ranged from 1m34.232s to 1m42.721s
>> After, real ranged from 1m32.360s to 1m39.484s
>>
> I presume this command compiled the whole of FFmpeg/libavcodec, i.e. it
> did not just link, didn't it?

Its deps, are *.o in lavu, lswr, lavc, but after the priming run, all 
the cc steps would be cache hits,
so the bulk of the fresh executions are the linking steps.

Regards,
Gyan