[PATCH] D32721: Accept archive files with no symbol table instad of warning on them.

[Mehdi AMINI via llvm-commits]

On Wed, May 3, 2017 at 5:26 PM Rui Ueyama <ruiu at google.com> wrote:

> On Wed, May 3, 2017 at 5:13 PM, Mehdi AMINI <joker.eph at gmail.com> wrote:
>
>> Clang is incrementally linking in a matter of a few seconds, so 0.5s to
>> read the symbols is a double digit percentage of that.
>> And there are over 50 binaries in LLVM, not just one.
>>
>
> We do not support incremental linking,
>

I'm talking about ThinLTO incremental linking, which we support.

but even if we support it, we don't need to read archives that haven't
> changed since the last build, so the overhead in that hypothetical case
> would be much smaller than 0.5s.
>

So yes we need to read all the archives.


> And you still don't address the "principle of least surprise": the
>> configuration is *not* what is expected from the user.
>>
>
> As a naive user of LTO, I was surprised that LTO needs llvm-ar, which is
> certainly I didn't expect (due to lack of knowledge).
>

That is why the warning is deserved.

-- 
Mehdi



>
>> --
>> Mehdi
>>
>>
>> 2017-05-03 16:51 GMT-07:00 Rui Ueyama <ruiu at google.com>:
>>
>>> The cost of reading symbols from object files in archive files is
>>> probably much cheaper than you might be thinking. If I strip all symbols
>>> from archives from a clang debug build, LLD takes 8.16 seconds to link,
>>> while it can usually link it in 7.65 seconds. So the difference is only 0.5
>>> seconds, and clang is a fairly large program as a test. That test case uses
>>> ELF, but with Peter's patch I believe reading symbols from bitcode files is
>>> fast too.
>>>
>>> To me 0.5 seconds is too small that I want the tool to "just work"
>>> instead of annoy me every time I run make/ninja until I change the build
>>> configuration to shave off 0.5 seconds from a LTO build.
>>>
>>> On Wed, May 3, 2017 at 4:23 PM, Mehdi AMINI via Phabricator <
>>> reviews at reviews.llvm.org> wrote:
>>>
>>>> mehdi_amini added a comment.
>>>>
>>>> I personally  think *not* warn is a terrible thing to do when there is
>>>> a configuration issue. Erroring is annoying, but warning should be intended
>>>> in such cases!
>>>>
>>>> > True, but on the other hand, it's pretty much the exact same work
>>>> that the archiver would need to do,
>>>>
>>>> The archiver do it once for potentially a lot of linker invocations.
>>>>
>>>> > and asking the user to change their archiver and rebuild would
>>>> probably consume even more time.
>>>>
>>>> This is a one time thing, and the user can live with the warning (or
>>>> pass a flag to disable the warning maybe) if they choose to.
>>>>
>>>>
>>>> Repository:
>>>>   rL LLVM
>>>>
>>>> https://reviews.llvm.org/D32721
>>>>
>>>>
>>>>
>>>>
>>>
>>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20170504/dac69237/attachment.html>