R600: Scheduling support for VLIW5 gen

[Vincent Lejeune]

Hi,

here is an updated version for the second patch ; I added your test case to it.
can you retest to see if there is some other regression ?

Thank,
Vincent




----- Mail original -----
> De : Tom Stellard <tom at stellard.net>
> À : Vincent Lejeune <vljn at ovi.com>
> Cc : "llvm-commits at cs.uiuc.edu" <llvm-commits at cs.uiuc.edu>
> Envoyé le : Mardi 30 juillet 2013 1h25
> Objet : Re: R600: Scheduling support for VLIW5 gen
> 
> On Mon, Jul 29, 2013 at 12:42:59PM -0700, Tom Stellard wrote:
>>  On Fri, Jul 26, 2013 at 07:00:27AM -0700, Vincent Lejeune wrote:
>>  > Hi,
>>  > 
>>  > the first 2 patches in this serie make our scheduler properly support 
> vliw5 hardware. Previously a vliw4 was always assumed;
>>  > it was compatible with vliw5 scheduling but provided not optimal 
> performance as the transALU slot was never used.
>>  > 
>>  > The last patch is a more general performance improvement that increase 
> performance in Lightmark 2008, and the other one
>>  > just make sure that there is never more than 4 litteral in a single ig 
> instead of crashing at R600ControlFlowFinalizer.
>>  > 
>>  > Vincent
>> 
>>  > From 13b24a905a9e61b354df4f24400fdc501d77fa51 Mon Sep 17 00:00:00 2001
>>  > From: Vincent Lejeune <vljn at ovi.com>
>>  > Date: Wed, 26 Jun 2013 18:32:52 +0200
>>  > Subject: [PATCH 2/4] R600: Non vector only instruction can be 
> scheduled on
>>  >  trans unit
>>  > 
>> 
>>  It would be nice to have a test for this.  You should be able to write a
>>  simple one with 5 fmul instructions.
>> 
>>  Other than that the series is:
>> 
>>  Reviewed-by: Tom Stellard <thomas.stellard at amd.com>
>> 
>>  I still would like to test this on r600g with compute, so I'll let you
>>  know when I've finished.
>> 
> 
> There are a few regressions with this branch.  It looks like some ALU
> instructions are being emitted after trans-only instructions like MULLO_INT.
> See the attached test case.  It would be great if you could reduce this
> to something smaller but either way the test should be added to the
> revised commit.
> 
> -Tom
> 
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 0001-R600-Non-vector-only-instruction-can-be-scheduled-on.patch
Type: text/x-patch
Size: 40783 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20130730/5a75467d/attachment.bin>