[LLVMbugs] Reassociation Transform

[Casey Carter]

Chris Lattner wrote:

>>For each of these bytecodes, I collected stats for -reassociate, -licm,
>>and -reassociate -licm (attached). Reassociation was always followed by
>>-constprop -instcombine (this reduces the inflation of LICM statistics
>>due to the negation instructions from reassociation). There were no
>>fatalaties in either the original bytecodes or any of the transformed
>>bytecodes.
>>    
>>
>
>Sounds great!  I'll go ahead and install the patch.  Note that gcse is
>also helped by reassociation: -reassociate -instcombine -licm -gcse is a
>good combination (not that I expect you to regen the results!  :)
>
Its only a one-line change in the Makefile that generates the results, 
so I will redo them. (attached)

>>The results read like an advertisement for reassociation.  For
>>example, 254.gap.bc has no LICM without reassociation, and 486 hoisted
>>instructions with reassociation.
>>    
>>
>
>Sounds great!  You should forward your results to the llvmbugs list so
>they can be archived.  They are really good, especially considering these
>bytecode files are already optimized.  I'll go ahead and reenable
>reassociation in gccas also.
>  
>
254.gap.bc looks good for gcse as well: reassociation exposes 163 gcs's 
to e.

-- 
Casey Carter
Casey at Carter.net
ccarter at uiuc.edu
AIM: cartec69


-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: results
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20021031/57d3a67d/attachment.ksh>