[llvm] [ARM] Replace ABS and tABS machine nodes with custom lowering (PR #156717)

[David Green via llvm-commits]

davemgreen wrote:

> We don't have to worry about T1 since the behavior for that is unchanged.

Yeah everything I said was for the most part referring to Thumb2.

> Anyway, the costs are fine and more accurate now. if we want to worry about cycles, that should be in another PR because I do not know how to do this without having it apply to all intrinsics that involve shifted operands. @davemgreen.

It would be ARMTTIImpl::getIntrinsicInstrCost that needs to be adjusted, to add a better cost for scalar instructions. It seems a bit off to me still, in terms of number of instructions and what I consider to be high-cost, but it's probably not far off.

Can you rebase the patch?

https://github.com/llvm/llvm-project/pull/156717