On 04/30/2010 03:13 AM, Andrew Pinski wrote:
> On Thu, Apr 29, 2010 at 5:32 PM, Bernd Schmidt<bernds@codesourcery.com>  wrote:
>> The idea is to represent the sign-extension in the RTL call patterns,
>> choosing extending variants in calls.c when we notice a suitable
>> promotion.  Then we let combine do its work; this needed another little
>> tweak to make sure it recorded the extension of the return register.
>
> On PowerPC (and MIPS even with 4.3.3), this is already optimized
> without a sign extend.
>
>   (call_insn 5 4 6 3 t.c:4 (parallel [
>              (set (reg:SI 3 3)
>                  (call (mem:SI (symbol_ref:SI ("shortv2") [flags 0x41]
> <function_decl 0x7fc0bfb81a00 shortv2>) [0 S4 A8])
>                      (const_int 0 [0x0])))
>              (use (const_int 0 [0x0]))
>              (clobber (reg:SI 65 lr))
>          ]) -1 (nil)
>      (nil))
>
> (insn 6 5 7 3 t.c:4 (set (reg:SI 121)
>          (reg:SI 3 3)) -1 (nil))
>
> (insn 7 6 8 3 t.c:4 (set (reg:SI 119 [ D.2012+-2 ])
>          (reg:SI 121)) -1 (nil))
>
> (insn 8 7 9 3 t.c:4 (set (reg:SI 120 [<retval>+-2 ])
>          (reg:SI 119 [ D.2012+-2 ])) -1 (nil))
>
> --- CUT ---
>
> Maybe I am missing something but the arm RTL should look something
> similar without a sign extension.

I suspect that's due to a different definition of promote_prototypes and 
other hooks.

However, I agree with Andrew that all the information should already be 
there (through the promote_function_mode hook), it's just that combine 
isn't trying hard enough.  Maybe something based on the attached patch?

Paolo