[llvm-commits] [llvm] r108109 - /llvm/trunk/lib/CodeGen/MachineLICM.cpp

[Evan Cheng]

On Jul 12, 2010, at 8:58 PM, Chris Lattner wrote:

> 
> On Jul 12, 2010, at 8:55 PM, Evan Cheng wrote:
> 
>>>> RA, or rather ProcessImplicitDef pass, turn them into trivial short live ranges when it's possible. But it does miss some cases. It's relatively rare but it does happen.
>>> 
>>> I don't think that LICM should be complicated by a deficiency handling implicit defs.
>> 
>> I agree. But it's already neutered by register allocator deficiency. We are only hoisting some of the loop invariants for fear of increasing register pressure. We should remove all such restrictions only after live range splitting work lands.
> 
> I don't see what that has to do with hoisting implicit defs.

It's the same issue. It can lead to increased register pressure. You guys should look at how implicit defs are being handled before jumping to conclusions. There are legitimate cases where implicit defs are not easily removed.

Evan

> 
> -Chris