[LLVMdev] Improving select_cc lowering for targets with conditional move

[Damien Vincent]

Let's consider the following piece of C code:  (incomplete and not
compilable ;) )
  result = initValue;
  for(i)
  {
    ...
    if(condition)
      result = updatedValue_i;
    ...
  }

For targets with conditional moves, the result is updated using the
following sequence of instructions:
regTmp = regFalse;
if(condition2) regTmp = regTrue;
regResult = regTmp;

Now, you have 2 cases:
1) either condition2 = condition
In this case: regFalse is a phi node between the initial value of result and
the current result. It's very likely that regFalse and regResult will be the
same hardware register. So the sequence of instructions reduce to just:
if(condition2) regResult = regTrue

2) either condition2 = !condition
In this case, regTrue is a phi node between the initial value of result and
the current result. In this case, the sequence of 3 instructions cannot be
reduced easily !

So, my question is (finally ;)):
Is there a way to introduce some intelligence in the select_cc lowering by
reversing the condition if it is likely to generate more efficient code ?
I am asking this question because the lowering works on basic blocks with a
set of input virtual registers and output virtual registers and all the
connections between these 2 set of registers seemed to be "lost" (I don't
think we know at the lowering stage that an input register is a phi node
between an output register and an initial value...)

  Thank you !

  Damien
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20110303/609f69ca/attachment.html>