[PATCH] R600/SI: Use source modifiers for f64 fabs / fneg

Matt Arsenault arsenm2 at gmail.com
Mon Aug 11 12:03:41 PDT 2014


On Aug 6, 2014, at 6:56 PM, Matt Arsenault <arsenm2 at gmail.com> wrote:

> 
> On Aug 6, 2014, at 10:34 AM, Matt Arsenault <arsenm2 at gmail.com> wrote:
> 
>> 
>> On Aug 6, 2014, at 12:25 AM, Michel Dänzer <michel at daenzer.net> wrote:
>> 
>>> 
>>>> Here is an updated set that implements that for f64. These also
>>>> switch to preferring using the literal constant as the first operand,
>>>> rather than loading the immediate into a register.
>>> 
>>> Sounds good, but I have to defer to Tom or others for review.
>>> 
>> 
>> I think I’m going to go back on this. It’s mostly a cosmetic change in the scalar case, trading a move of an immediate for one from an SGPR to VGPR. In the vector case, this would require N SGPR copies, but the other way the SGPRs could stay and only 1 copy of the immediate would need to be loaded.
> 
> 
> 
> These switch back to preferring the old order of literal loaded into a VGPR, and remove the FABS_SI / FNEG_SI pseudoinstructions. I don’t see any need for them anymore, and the pattern can directly match the instructions
> 
> 
> <0001-R600-SI-Refactor-fneg-fabs-patterns.patch>
> <0002-R600-SI-Use-source-modifier-for-f64-fabs.patch>
> <0003-R600-SI-Use-source-modifiers-for-f64-fneg.patch>
> <0004-R600-SI-Move-all-fabs-fneg-handling-to-patterns.patch>


Ping
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20140811/ea28c654/attachment.html>


More information about the llvm-commits mailing list