[llvm-dev] Loop invariant values and scalarization overhead

[Jonas Paulsson via llvm-dev]

Hi

Recently we have had some discussion about the scalarization overhead 
and loop invariant values. Before I propose a patch for this I would 
like to know if this is needed and wanted right now.

Basically, getOperandsScalarizationOverhead() could be improved to check 
for a loop invariant value, just like it checks for a constant.

In order to do this, I imagine that the Loop* pointer should be passed 
to that function, so that

bool IsLoopInvariant = (L != nullptr && !L->contains(A));

Just like a constant, this value would never need to be extracted from a 
vector operand.

This would involve passing the Loop* pointer to many of the TTI cost 
functions, it seems.

As a good side effect of this is that some cost functions could become 
loop aware. For instance, on SystemZ a vector permute can be done with a 
constant mask, which cost is 0 if inside a loop, since it would be 
hoisted out of the loop. If not in a loop, that instruction should 
probably also be included in the cost.

Any other related issues?

I don't know if this would mean any improvements right now on 
benchmarks, but if there is a general agreement that this should be 
done, I would be happy to do so.

/Jonas