[PATCH] D102568: [Driver] Delete -mimplicit-it=

Wed May 19 12:13:15 PDT 2021

dblaikie added a comment.

In D102568#2769341 <https://reviews.llvm.org/D102568#2769341>, @MaskRay wrote:

> I think libvpx's ASFLAGS usage is about GNU as, not the driver option.
>
> In D102568#2769296 <https://reviews.llvm.org/D102568#2769296>, @dblaikie wrote:
>
>>> I think "waiting for a few releases" is too much and doesn't improve things (they will notice issues until you remove the option). I can accept "waiting for one major release".
>>
>> Having fairly broad windows of not breaking backwards compatibility is a fairly reasonable request. Take LLVM itself, for instance - which supports building with Clang back as far as 3.5. (not suggesting we need to have a window that large for every feature/piece of surface area - but that there is scope for fairly wide windows)
>
> I understand the broad Windows and I also strive for this portability.
>
> However, reverting this patch is now a trade-off. It would make x264/openh264's mingw build happy but break musl's arm build (2018-09 ~ ).
> There can be other projects doing detection on both -mimplicit-it and -Wa,-mimplicit-it and will be broken by reverting this patch.

Not sure I follow - presumably musl's arm build wasn't working with clang before this patch? Or something changed there after this patch landed?

It's not OK to say "this patch fixed a bug for someone, so now reverting it is on equal footing with the regressions the patch caused" - if a patch broke an existing use case that's different from fixing a use case that wasn't working to begin with & reverting to the known/existing failures would be where things tend to lean (not absolute by any means - there are cases where we tradeoff adding a new bug to fix an existing bug) if all other things are equal.

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D102568/new/

https://reviews.llvm.org/D102568