LoopSimplifyPass and LoopAnalysis work incomplete

IR & Optimizations Loop Optimizations
,
1

I want to write a transformation pass that first needs to detect loops in the form

for (int i = 0; i < b && c; i++)

so a canonical loop where b is a constant and an additional condition.
I need the induction variable i for my transformation. Per function f (with the FunctionAnalysisManager fam) my pass tries to enforce a canonical form and gather loop information:

LoopSimplifyPass lsp;
lsp.run(F, fam);
ScalarEvolution &se = fam.getResult<ScalarEvolutionAnalysis>(F);
auto &loop_info = fam.getResult<LoopAnalysis>(F);

However, for loops like this

int main(int argc, char **argv) {
  int n = atoi(argv[1]);
  for (int i = 7; n != 3 && i < 13; i++) {
    n = atoi(argv[1]);
    printf("%d\n", i);
  }
}

the Loop Analysis neither detects an induction variable, nor is it in canonical form (only in loop-simplify-form). If i initialize i with 0 it works flawlessly.
Why doesn’t the loop simplify pass transform this loop to:

int main(int argc, char **argv) {
  int n = atoi(argv[1]);
  for (int i = 0; n != 3 && i < 6; i++) {
    n = atoi(argv[1]);
    printf("%d\n", i + 7);
  }
}

or at least report i as an induction variable?

My pass is registered with
registerPipelineParsingCallback for opt and with registerPipelineEarlySimplificationEPCallback for clang.

2

There are additional passes that normalize loops other than LoopSimplifyPass, including loop rotation, LCSSA, and IndVarSimplify.

You should not rely on the existence of a specific induction variable, use ScalarEvolution instead.

https://llvm.org/docs/LoopTerminology.html#more-canonical-loops

1 Like
3

IndVarSimplifyPass is what i am looking for.
It has a run method that takes a loop and additional parameters.
Is there a simple way to manually invoke it in a Function pass?

Thank you very much!

4

You got the wrong conclusion. I recommend against using Loop::getCanonicalInductionVariable(), Loop::getInductionVariable() or IndVarSimplify. DO use ScalarEvolution SCEV expressions instead. IndVarSimplify has been severely nerfed in Disable IV rewriting by default. See PR10916. · llvm/llvm-project@061d811 · GitHub anyway.

For some background, see e.g. Missed Loop peeling and thus vectorization opportunity - #4 by Meinersbur