[llvm-commits] discouraging use of x86_64 R12 and R13

[Dan Gohman]

On Nov 17, 2008, at 4:08 PM, Stuart Hastings wrote:

> A trivial patch for a mild (and unproven) performance improvement.   
> (<rdar://problem/6351057
>> )
>
> The x86_64 R12 and R13 register encodings are related to the RBP and
> RSP encodings, and thus require slightly longer instructions than the
> other R8..R15 registers.  Specifically, just like [RBP], the X86_64
> cannot encode [R12] directly; this is expressed as [R12+0].  In like
> manner, [R13] resembles [RSP], and requires an SIB byte.
>
> Since R14 and R15 don't have these drawbacks, here is a patch for LLVM
> to prefer them over R12 & R13.  By "prefer," we really mean "allocate
> R14 and R15 first."

Along with Evan's comments, can you summarize the above two paragraphs
in a comment in the source?

Thanks,

Dan