[PATCH] D32093: [InstCombine] PR32078: convert scalar operations to vector.

[Sanjay Patel via Phabricator via llvm-commits]

spatel requested changes to this revision.
spatel added a reviewer: efriedma.
spatel added a comment.
This revision now requires changes to proceed.

This transform is not safe as written. This program compiled with -O1 will crash/cause exception after this patch, but it does not before:

  typedef int v4si __attribute__((__vector_size__(16)));
  
  v4si divs(v4si x, v4si y) {
    int eltX = x[1];
    int eltY = y[1];
    x[1] = eltX / eltY;
    return x;
  }
  
  int main() {
    v4si x = (v4si){0, 1, 2, 3};
    x = divs(x, x);
    return x[0];
  }

The division example shows that we're operating on elements that the original program does not, so any FP op would also be a concern. If some element in a vector is a denorm, that could cause a perf explosion that doesn't exist in the original program. Besides that, I'm not sure that we can actually do this transform for any op in a target-independent pass. We're replacing scalar ops with potentially more expensive vector ops.


https://reviews.llvm.org/D32093