[PATCH] D154075: [LoopVectorize] Add pre-commit tests for D152366

[David Sherwood via Phabricator via llvm-commits]

david-arm added a comment.

> - inner and outer inductions decreasing (constant step)

Hi @fhahn, perhaps I'm missing something but the runtime checks already look broken to me for decreasing IVs in the inner loop:

  #include <stdint.h>
  
  void decreasing_inner_iv(int32_t *dst, int32_t *src, int m, int n) {
    for (int i = 0; i < m; i++) {
      for (int j = n - 1; j >= 0; j--) {
        dst[(i * (n + 1)) + j] += src[(i * n) + j];
      }
    }
  }

These are the SCEVs for the checks:

  LAA: Adding RT check for range:
  Start: {%dst,+,(4 * (zext i32 (1 + %n) to i64))<nuw><nsw>}<%for.cond1.preheader.us> End: {((4 * (zext i32 %n to i64))<nuw><nsw> + %dst),+,(4 * (zext i32 (1 + %n) to i64))<nuw><nsw>}<%for.co
  nd1.preheader.us>
  LAA: Adding RT check for range:
  Start: {%src,+,(4 * (zext i32 %n to i64))<nuw><nsw>}<%for.cond1.preheader.us> End: {((4 * (zext i32 %n to i64))<nuw><nsw> + %src),+,(4 * (zext i32 %n to i64))<nuw><nsw>}<%for.cond1.preheade
  r.us>
  Calculating cost of runtime checks:
    1  for   %bound0 = icmp ult ptr %scevgep, %scevgep36
    1  for   %bound1 = icmp ult ptr %scevgep35, %scevgep34
    1  for   %found.conflict = and i1 %bound0, %bound1

I don't see how the Start and End values correspond to actual range of addresses accessed in the inner loop?


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D154075/new/

https://reviews.llvm.org/D154075