[llvm] [LAA] Support different strides & non constant dep distances using SCEV. (PR #88039)

[Pete Steinfeld via llvm-commits]

psteinfeld wrote:

@fhahn, after this change, we started failing one of our tests -- the Fortran Polyhedron benchmark "rnflow".  When I build flang-new and compile the benchmark with the "flang-new -Ofast -march=native rnflow.f90", then run the benchmark, and then run the Polyhedron utility "pbvalid", you'll see failures.  The failures only happen when both compilation options are used.

Do you have access to the Polyhedron "rnflow" benchmark?

https://github.com/llvm/llvm-project/pull/88039