[llvm-commits] [llvm] r122659 - in /llvm/trunk: lib/Transforms/Scalar/LoopIdiomRecognize.cpp test/Transforms/LoopIdiom/basic.ll

[Chris Lattner]

On Jan 1, 2011, at 4:54 PM, Frits van Bommel wrote:

> On Sat, Jan 1, 2011 at 8:39 PM, Chris Lattner <sabre at nondot.org> wrote:
>> +/// mayLoopModRefLocation - Return true if the specified loop might do a load or
>> +/// store to the same location that the specified store could store to, which is
>> +/// a loop-strided access.
>> +static bool mayLoopModRefLocation(StoreInst *SI, Loop *L, AliasAnalysis &AA) {
>> +  // Get the location that may be stored across the loop.  Since the access is
>> +  // strided positively through memory, we say that the modified location starts
>> +  // at the pointer and has infinite size.
>> +  // TODO: Could improve this for constant trip-count loops.
>> +  AliasAnalysis::Location StoreLoc =
>> +    AliasAnalysis::Location(SI->getPointerOperand());
>> +
>> +  for (Loop::block_iterator BI = L->block_begin(), E = L->block_end(); BI != E;
>> +       ++BI)
>> +    for (BasicBlock::iterator I = (*BI)->begin(), E = (*BI)->end(); I != E; ++I)
>> +      if (AA.getModRefInfo(I, StoreLoc) != AliasAnalysis::NoModRef)
>> +        return true;
> 
> Wouldn't it be cleaner to add an 'I != SI &&' to the condition here
> instead of temporarily removing SI from the block in before calling
> this?

Yes, I did this because I thought it would be useful for memcpy formation, but I changed my mind.  I switched to this approach.

> Of course, checking that extra condition for every instruction in the
> loop is probably less efficient, which I'm guessing is the reason you
> did it this way?

I did it this way because I wanted memcpy formation to take the load *and* store out, but I realize now that it would be a bad idea because I *want* an interference when the store aliases the load.  Anyway, please take a look at r122678.

-Chris