[PATCH] D150851: [LoopVectorize] Vectorize select-cmp reduction pattern for increasing integer induction variable

[Ramkumar Ramachandra via Phabricator via llvm-commits]

artagnon added a comment.

Hi Mel,

I had the chance to try out your patch today, and it seems that it's really quite restricted in its scope. I tried it out on this simple C program:

  #include <stdio.h>
  
  int main() {
        int src[20000] = {4, 5, 2};
        int r = 331;
        for (int i = 0; i < 20000; i++) {
          if (src[i] > 3)
            r = i;
        }
        printf("%d\n", r);
        return 0;
  }

Unfortunately, your patch failed to vectorize this simple case, because when you match `NonPhi` against a `PHINode`, it is blocked on a `trunc`:

  for.body:                                         ; preds = %entry, %for.body
    %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
    %r.05 = phi i32 [ 331, %entry ], [ %spec.select, %for.body ]
    %arrayidx = getelementptr inbounds [20000 x i32], ptr %src, i64 0, i64 %indvars.iv
    %3 = load i32, ptr %arrayidx, align 4, !tbaa !6
    %cmp1 = icmp sgt i32 %3, 3
    %4 = trunc i64 %indvars.iv to i32 // blocked here
    %spec.select = select i1 %cmp1, i32 %4, i32 %r.05
    %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
    %exitcond.not = icmp eq i64 %indvars.iv.next, 20000
    br i1 %exitcond.not, label %for.cond.cleanup, label %for.body, !llvm.loop !10

May I suggest using the logic outlined in `isInductionMinMaxPattern` in D152693 <https://reviews.llvm.org/D152693>?

Thanks.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D150851/new/

https://reviews.llvm.org/D150851