[llvm-commits] [PATCH] Smarter unaligned stores of constant values.

[Lang Hames]

Postscript: Output with patch is exactly what you'd expect:

_foo:                                   @ @foo
@ BB#0:                                 @ %entry
        mvn     r1, #0
        strb    r1, [r0, #3]
        strb    r1, [r0, #2]
        strb    r1, [r0, #1]
        strb    r1, [r0]
        mov     pc, lr

Cheers,
Lang.

On Fri, Nov 4, 2011 at 4:30 PM, Lang Hames <lhames at gmail.com> wrote:

> Encountered the following when I tried outputting an unaligned store of -1
> on ARM:
>
> _foo:                                   @ @foo
> @ BB#0:                                 @ %entry
>         mov     r1, #255
>         mov     r2, #255
>         strb    r1, [r0, #3]
>         orr     r1, r2, #65280
>         strb    r1, [r0, #2]
>         mvn     r1, #-16777216
>         strb    r1, [r0, #1]
>         mvn     r1, #0
>         strb    r1, [r0]
>         mov     pc, lr
>
> The strange constants are coming out of the legalizer, which tries to
> legalize the unaligned store by repeatedly splitting it into two half-sized
> stores, and shifting the constant around. This results in several new and
> confusing constants being introduced.
>
> This patch checks for constant value stores, and if the two halves of the
> constant have the same bit-pattern it avoids creating new constants.
>
> Could someone who's more familiar with the legalizer take a look and let
> me know if this is Ok to commit?  (I'm also planning to enable unaligned
> stores on supported subtargets, but that's a different issue)
>
> Cheers,
> Lang.
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20111104/1f59633e/attachment.html>