Use "andnot" instruction when vectorising

Issue #528 closed
Erik Schnetter created an issue

Use the "andnot" instruction to reduce the number of different bit masks that are required. Using fewer different bit masks may require fewer registers to hold them, or fewer load instructions to access them, thus potentially improving performance.

Do not scalarize ifpos when SSE 4.1 is not available; instead, use logical operations to create a bit mask.

Keyword:

Comments (4)

  1. Barry Wardell
    • removed comment

    This looks fine to me and the testsuites still pass with it applied. I think it can be committed.

  2. Log in to comment