[LLVMdev] ARMv4T Copy Lowering

Renato Golin renato.golin at linaro.org
Wed Aug 20 14:24:38 PDT 2014


On 20 August 2014 19:01, Jonathan Roelofs <jonathan at codesourcery.com> wrote:
> I've got the patch ready that implements (1). Does it make sense to commit
> this
> (or something close to it) in the near term to fix the bug, and let someone
> else worry about performance later? (Possibly me, but it will probably be a
> a while before I come back to this. Hopefully this thread serves as a nice
> bread-crumb for anyone who is interested)

Hi Jon,

Correctness before performance any day.

This patch is simple and obvious enough that you can commit after you
add some test showing that it'll select push/pop for v4/v5 on thumb1
and not for other combinations (up to v6 is enough). Also, please add
some short FIXME one-liner saying that we should try high registers
but beware of scheduling issues inside your current comment.

I'm sure it'll boil down as you working on this again, as I don't know
anyone else *that* interested in v4T. :)

Thanks!
--renato



More information about the llvm-dev mailing list