[PATCH] D25020: [ARM] Fix 26% performance regression on Cortex-A9 caused by not using VMLA/VMLS

[Evgeny Astigeevich via llvm-commits]

eastig added a comment.

Hi Diana,

Based on the results of the LNT runs I think the code checking accumulator forwarding is not needed at all. If I am correct only SIMD VMLx instructions can have issues. So we should check only them.
My thoughts:

- We can have features: HasSIMDVMLxHazards and HasVFPVMLxHazards. HasVMLxHazards can be built on them.
- The current checks in MLxExpansion and in ARMDAGToDAGISel::canUseVMLxForwarding are used only for SIMD instructions.

What do you think?

Thanks,
Evgeny


https://reviews.llvm.org/D25020