[llvm] r317753 - AMDGPU: Merge BUFFER_LOAD_DWORD_OFFSET into x2, x4

[Marek Olšák via llvm-commits]

On Fri, Nov 10, 2017 at 11:37 AM, Michel Dänzer via llvm-commits
<llvm-commits at lists.llvm.org> wrote:
>
> Hi Marek,
>
>
>> I had looked at the issue and it seems variable indexing has always been
>> somewhat broken in LLVM and this commit uncovered the bug.
>
> That's unfortunate, but doesn't change the situation that tests which
> were previously passing are now failing. If we let these tests keep
> failing, we might miss further regressions affecting them.
>
> If there's no solution on the horizon, it would be better to revert this
> change until it doesn't break these tests.

It's an issue with scheduling. LLVM moves M0 assignments next to each
other, so movrel instructions only receive the last M0 assignment,
which is incorrect for most of them. It has nothing to do with this
commit, and thus I don't see a reason for reverting.

Marek