Loop vectorizer may create redundant preheader and scalar loop

[duk-37]

Compiler Explorer

#include <cstdint>

uint32_t foo(uint32_t *A, int n) {
  uint32_t sum = 0;
  for (int i = 0; i < n * 32; i++)
        sum += A[i];
  return sum;
}

Output IR is attached. I believe what's happening here is that the smax generated by SCEV expansion during IndVarSimplify interacts badly with SCEV checks in the loop vectorizer. I'm not quite sure though.

output.ll.txt

[fhahn]

Basic support would be added with #111310. But there is extra work needed to preserve wrap flags on the mul/shl + using them from SCEV when computing the trip count expression

[fhahn]

(alive proof for redundant RT check with nsw flag on shl: https://alive2.llvm.org/ce/z/PpNtgW)