[PATCH] D80811: [ARM] Balance register usage when sinking splats

[Dave Green via Phabricator via llvm-commits]

dmgreen created this revision.
dmgreen added reviewers: simon_tatham, SjoerdMeijer, efriedma, samparker, ostannard.
Herald added subscribers: danielkiss, hiraditya, kristof.beyls.
Herald added a project: LLVM.

If we sink a splat into a loop in order to make more use of MVE r reg instructions, we can sometimes get to a point where we go over the register limit, leading to more spills in the loop than we helped by not using the q reg. In order to alleviate that this patch attempts to come up with a rough estimate of the number of different registers used in a loop, and doesn't sink if there are free Q regs but no free R regs.

I had to add some LoopInfo update too, in order to prevent the LoopInfo containing stale information.


https://reviews.llvm.org/D80811

Files:
  llvm/include/llvm/CodeGen/TargetLowering.h
  llvm/lib/CodeGen/CodeGenPrepare.cpp
  llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
  llvm/lib/Target/AArch64/AArch64ISelLowering.h
  llvm/lib/Target/ARM/ARMISelLowering.cpp
  llvm/lib/Target/ARM/ARMISelLowering.h
  llvm/lib/Target/X86/X86ISelLowering.cpp
  llvm/lib/Target/X86/X86ISelLowering.h
  llvm/test/CodeGen/Thumb2/mve-gather-scatter-optimisation.ll
  llvm/test/CodeGen/Thumb2/mve-sinkbalance.ll

-------------- next part --------------
A non-text attachment was scrubbed...
Name: D80811.267242.patch
Type: text/x-patch
Size: 26433 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20200529/344835bd/attachment-0001.bin>