Re: [PATCH RS6000] adding an unsigned SI and a DI together and subtracting an unsigned SI from an DI

[Andrew Pinski <pinskia at physics dot uc dot edu>]

Here is the patch as I do not have write access.
I have also added the strings for the `! WORDS_BIG_ENDIAN' case
for *addsidi3_noppc64 and *subsidi3_noppc64.

ChangeLog:
2003-05-14  Andrew Pinski  <pinskia@physics.uc.edu>
	* config/rs6000/rs6000.md (zero_extendsidi2): Always expand.
	(*zero_extendsidi2_noppc64): New insn and split for the non-PPC64
	case.
	(*zero_extendsidi2_ppc64): Add name for the PPC64 case of
	zero_extendsidi2.
	(*addsidi3_noppc64): New instruction for combine to use, which
	simplifies adddi3 with zero_extended SI.
	(*subsidi3_noppc64): Likewise but for subdi3.


Patch:

Attachment: temp.diff
Description: Binary data

Thanks,
Andrew Pinski


On Wednesday, May 14, 2003, at 14:04 US/Eastern, David Edelsohn wrote:

> 	* config/rs6000/rs6000.md (zero_extendsidi2): Always expand.
> 	(*zero_extendsidi2_noppc64): New insn and split for the non-PPC64
> 	case.
> 	(*zero_extendsidi2_ppc64): Add name for the PPC64 case of
> 	zero_extendsidi2.
> 	(*addsidi3_noppc64): New instruction for combine to use, which
> 	simplifies adddi3 with zero_extended SI.
> 	(*subsidi3_noppc64): Likewise but for subdi3.
>
> 1) Please add the "&& 1" in the define_insn_and_split as pointed out by
>    Richard.
> 2) Please remove the WORDS_BIG_ENDIAN variations from the splitter.
> 3) Please make sure you format the ChangeLog entry correctly (as above,
>    not word-wrapped as a block).
> 4) Please use "ppc64" and "noppc64" instead of "64" and "no64" in the
>    pattern names to be consistent with the other patterns.
> 5) Please use "*addsidi3_noppc64" and "subsidi3_noppc64" as the pattern
>    names instead of appending "sidi" to be consistent with the standard
>    mulsidi3 named pattern.
>
> With those changes, you can commit the patch.
>
>
> > The only problem I have with this patch is that it make
> > `zero_extendsidi2' done in an split which might cause worse code.
>
> 	The default code is the same as the code produced by the splitter,
> except it exposes the subregs in the initial RTL and surrounds it by a
> LIBCALL block.  Preserving the zero_extend and avoiding the LIBCALL 
> might
> actually be a win and probably is not any worse.
>
>
> Thanks, David