Uploaded image for project: 'MariaDB Server'
  1. MariaDB Server
  2. MDEV-15843

mysys: remove optimized memcpy from 18 years ago

Details

    Description

      While this code has remained dormant for 18 years, libc implementers have used assembly features to gain improvements using architecture features optimized and by the buffer length like:

      https://svnweb.freebsd.org/base/head/lib/libc/amd64/string/memcmp.S
      https://sourceware.org/git/?p=glibc.git;a=blob;f=sysdeps/aarch64/memcmp.S
      https://sourceware.org/git/?p=glibc.git;a=blob;f=sysdeps/powerpc/powerpc64/memcpy.S
      While the perf results in commit message don't show a great improvement the varying length and codebase shows optimizations based on length.

      Pull request: https://github.com/MariaDB/server/pull/698

      Attachments

        Activity

          Ascending order - Click to sort in descending order
          monty Michael Widenius added a comment -

          Thanks a lot for the testing and results! Very interesting!

          I don't want yet to remove the ptr_compare code as the results depends a lot on:

          • Size of strings and if strings are aligned
          • If the cpu allows unaligned access

          There is some benefit in knowing the exact length and alignment in advance instead of doing the check for each call. However your tests shows that for our most common platform the libc memcmp is better and I agree we should use that.

          I prefer to keep the old code around, as it's still possible to do a faster memcmp based on the principle of ptr_cmp. I will take your excellent patch comment and disable the ptr_cmp code for now in preference of the standard memcmp.
          @montywi

          monty Michael Widenius added a comment - Thanks a lot for the testing and results! Very interesting! I don't want yet to remove the ptr_compare code as the results depends a lot on: Size of strings and if strings are aligned If the cpu allows unaligned access There is some benefit in knowing the exact length and alignment in advance instead of doing the check for each call. However your tests shows that for our most common platform the libc memcmp is better and I agree we should use that. I prefer to keep the old code around, as it's still possible to do a faster memcmp based on the principle of ptr_cmp. I will take your excellent patch comment and disable the ptr_cmp code for now in preference of the standard memcmp. @montywi
          monty Michael Widenius added a comment -

          Pushed

          monty Michael Widenius added a comment - Pushed

          People

            monty Michael Widenius
            svoj Sergey Vojtovich
            Votes:
            0 Vote for this issue
            Watchers:
            2 Start watching this issue

            Dates

              Created:
              Updated:
              Resolved:

              Git Integration

                Error rendering 'com.xiplink.jira.git.jira_git_plugin:git-issue-webpanel'. Please contact your Jira administrators.