[PATCH] D13269: Improved X86-FMA3 mem-folding & coalescing

[Vyacheslav Klochkov via llvm-commits]

v_klochkov added a comment.

Hi Quentin,

I do not have permissions for committing changes to LLVM trunc.
Andy (akaylor) who committed the changes for me previously reasonable refused to commit 
this change-set as it is now because of Ahmed Bougacha’s comment:

  “IMHO” Implement FMA*_Int instructions before landing this change-set.”

That is probably good move as the current fix potentially can produce regressions on tests where the operand 1 and 3 of scalar FMA intrinsics should not be commuted but can be commuted by the new changes. 
Also, my senior colleges recommended to implement FMA*_Int instructions first too.

I am ready to submit a new code-review tracker for FMA*_Int instructions.
The corresponding changes are already implemented and tested locally.

Adding FMA*_Int instructions will cause conflicts in X86InstrFMA.td file.

Do you agree with the following plan?

1. To add FMA*_Int instructions first;
2. To update this change-set (http://reviews.llvm.org/D13269), i.e. to resolve conflicts in X86InstrFMA.td, and ask for additional review for changes in that file.

3. In order to minimize the patch for (1), the commute optimization for FMA*_Int should be implemented as a separate change-set.

I am asking you as Ahmed stopped answering in this code-review tracker and also did not answer
to e-mails sent to him separately.

Thanks you,
Slava


http://reviews.llvm.org/D13269