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

[Vyacheslav Klochkov via llvm-commits]

v_klochkov added a comment.

Hi,

I created the FMA*_Int opcodes in the patch for ( http://reviews.llvm.org/D13710 ) and it has been committed to LLVM trunc.
That patch conflicted with the changes I did here in X86InstrFMA.td.
Thus, I had to update my local workspace and upload the new RE-BASED patch this time.

Please review the updated changes in X86InstrFMA.td.
I would like to comment some additional changes I did to resolve the conflicts and to simplify the opcode definitions:

- The parameters 'IsRVariantCommutable' and 'IsMVariantCommutable' were just removed because otherwise, they would be always set to 1.
- Moved some comments from fma3{p,s}_forms multiclasses to fma3{p,s}_rm multiclasses. (the multiclasses fma3{p,s}_forms stopped mentioning the commute features, so having those comments there seemed not quite appropriate).

Also, reviewers asked me to use update_llc_test_checks.py for the new test fma-commute-x86.ll.
For some unknown reasons that tool did not work for me, it printed error for any input test.  
So, I just added the better checks to the test manually.

This patch does not implement commute transformations for FMA*_Int opcodes.
That can/should be done in a separate patch.

Thank you,
Slava


http://reviews.llvm.org/D13269