Souper fails to turn multiplication by a power of two into a shift? about souper HOT 3 CLOSED

hi Nick, cost models have been a major problem for Souper since day 1.

here's the current cost function from Inst.cpp:

int Inst::getCost(Inst::Kind K) {
  switch (K) {
    case Var:
    case Const:
    case UntypedConst:
    case SAddO:
    case UAddO:
    case SSubO:
    case USubO:
    case SMulO:
    case UMulO:
      return 0;
    case BitReverse:
    case BSwap:
    case CtPop:
    case Cttz:
    case Ctlz:
    case SDiv:
    case UDiv:
    case SRem:
    case URem:
      return 5;
    case FShl:
    case FShr:
    case SAddSat:
    case UAddSat:
    case SSubSat:
    case USubSat:
    case Select:
      return 3;
    default:
      return 1;
  }
}

there are three things we can do. first, we can tweak this, for example by giving mul cost 2, which should cause the result you want to be produced. second, you can hack your local souper (and I suggest you try that now, and see if it does what I said). third, we could make it possible to provide souper with a cost model externally. I've tweaked this stupid cost model many times and there's no good answer even within LLVM, and of course you're sort of misusing souper for non-LLVM purposes, so...

from souper.

to give a bit more history, I've asked a number of smart and experienced LLVM people what we should do for a cost model over the years. everyone has one or more ideas, and none of them have ended up being satisfactory. so probably for now just do whatever you want in this function.

of course a better cost model would (optionally) look at multiple instructions instead of just one. but that way leads to madness.

from souper.

Thanks for the reply. Yeah, this definitely is a rabbit hole one could keep following.

Hacked up that method and I'm getting what I expect now. Thanks!

from souper.