[llvm] r300878 - ARM: handle post-indexed NEON ops where the offset isn't the access width.

[Friedman, Eli via llvm-commits]

On 4/20/2017 1:54 PM, Tim Northover wrote:
>> We might want to consider keeping this check; even if the transform is legal, it's likely not profitable given we have to materialize the immediate.
> I was a bit ambivalent about that too. The main reason it's there is that otherwise the whole thing is virtually untestable -- the kind of IR that only becomes obviously constant on the 3rd run of combine isn't the kind of IR you want in your regression tests (see the attachment to https://bugs.llvm.org/show_bug.cgi?id=32658).
>
> I'm still willing to try and come up with something if we really want (I might be able to tweak that example to cover the other kinds of load/store, though it'll always be fragile).

If it's just a matter of testing, we can add a command-line option to 
control the heuristic.

Granted, I'm only guess about the effect on performance; I'll try 
measuring it.

-Eli

-- 
Employee of Qualcomm Innovation Center, Inc.
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project