[PATCH] D13427: RFC: faster isa<IntrinsicInst> (bugged tests?)

[escha via llvm-commits]

escha added a comment.

I stuck an rdtsc timer around code that traversed the function like this:

if (isa<callinst>(I)) {

  START_TIMER;
  if (isa<intrinsicinst>(I) and intrinsic ID is assume) {
     do a thing that will never happen
   }
  STOP_TIMER;

}

It was ~10 cycles per timer hit when using the ID-based approach, ~50 when using the current approach. The timer already subtracts out the cost of the timer itself and the effects of context switches (approximately).

I don't know why it was so slow, but I still don't think we should be constantly doing string comparisons either way, at least if we can avoid it.


Repository:
  rL LLVM

http://reviews.llvm.org/D13427