[llvm-commits] [llvm] r56286 - in /llvm/trunk: lib/Transforms/Utils/SimplifyCFG.cpp test/Transforms/SimplifyCFG/2008-09-17-SpeculativeHoist.ll

[Evan Cheng]

On Sep 18, 2008, at 10:12 AM, Devang Patel wrote:

>
> On Sep 18, 2008, at 9:57 AM, Evan Cheng wrote:
>
>>> -  // If we get here, we can hoist the instruction. Try to place it
>>> before the
>>> -  // icmp instruction preceeding the conditional branch.
>>> -  BasicBlock::iterator InsertPos = BI;
>>> -  if (InsertPos != BIParent->begin())
>>> -    --InsertPos;
>>> -  if (InsertPos == BrCond && !isa<PHINode>(BrCond))
>>> -    BIParent->getInstList().splice(InsertPos, BB1->getInstList(),
>>> I);
>>> -  else
>>> -    BIParent->getInstList().splice(BI, BB1->getInstList(), I);
>>> +  // If we get here, we can hoist the instruction.
>>> +  BIParent->getInstList().splice(BI, BB1->getInstList(), I);
>>
>> Rather than removing the logic, why not check if the instruction uses
>> the branch condition?
>
> What's the advantage of hoisting this instruction by one  additional
> instruction in some cases ?

As the comment said, it's trying to place it before the icmp  
instruction which produces the predicate. This is a micro-optimization  
for performance reason. Some optimization passes do not look very far,  
it may prefer the icmp and branch to be next to each other. Also fast- 
isel codegen will also prefer them to be together.

Evan

>
> -
> Devang
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits