[llvm-commits] [llvm] r91378 - in /llvm/trunk: lib/CodeGen/SelectionDAG/DAGCombiner.cpp test/CodeGen/X86/setcc.ll test/CodeGen/X86/zext-shl.ll

[Evan Cheng]

On Dec 14, 2009, at 5:47 PM, Chris Lattner wrote:

> 
> On Dec 14, 2009, at 5:45 PM, Evan Cheng wrote:
> 
>> 
>> On Dec 14, 2009, at 5:00 PM, Chris Lattner wrote:
>> 
>>> 
>>> On Dec 14, 2009, at 4:41 PM, Evan Cheng wrote:
>>> 
>>>> Author: evancheng
>>>> Date: Mon Dec 14 18:41:36 2009
>>>> New Revision: 91378
>>>> 
>>>> URL: http://llvm.org/viewvc/llvm-project?rev=91378&view=rev
>>>> Log:
>>>> Propagate zest through logical shift.
>>> 
>>> Is this really profitable in the general case?  It seems that this should only be done if at least one of the inputs to the shift will fold away.  Adding a nice "WouldFoldAZExt" predicate that would return true for zext/trunc would make sense, no?
>> 
>> Are there cases where (zext (zext)) would not fold into a single zext?
> 
> No, I'm worried that you're turning something like:
> 
> zext(shr(call, call)) -> shr(zext(call), zext(call))
> 
> which introduces two zexts and deletes one.

No. The transformation only happens when one of the shift's operand is a zext.

Evan

> 
> -Chris