[llvm-dev] [ARM] Thumb code-gen for 8-bit imm arguments results in extra reg copies

[Tim Northover via llvm-dev]

On Thu, 18 Jun 2020 at 10:39, Prathamesh Kulkarni
<prathamesh.kulkarni at linaro.org> wrote:
> Sorry, I didn't entirely understand why it isn't safe to propagate
> while checking for single live def, could you please elaborate?

Imagine something like:

    %1:gpr = tMOVi8 123, def dead %cpsr
    [...]
    tCMPr %2:gpr, %3:gpr, def %cpsr
    %4:gpr = COPY %1:gpr
    tBcc %bb.1, 1, implicit %cpsr

The initial mov only has one live use, but reinserting it between the
cmp and the bne would make the branch go to the wrong place. You need
%cpsr to be dead at the place you're intending to insert the mov too.

> If reMaterializeTrivialDef is not the right place to handle the
> transform, could you please suggest where should I look for adding it?

It might be the right place, otherwise it might fit in with a separate
ARM pass of its own, or a known more expensive computation that hasn't
been written yet. I don't know, but we should at least consider
whether we want to complicate the interface for these functions.

Cheers.

Tim.