Uploaded image for project: 'JDK'
  1. JDK
  2. JDK-8216580

Fix generation of VNNI vector code by allowing adjacent LoadS nodes to be isomorphic

XMLWordPrintable

    • Icon: Enhancement Enhancement
    • Resolution: Fixed
    • Icon: P3 P3
    • 13
    • 12, 13
    • hotspot
    • b11
    • x86

      Fix generation of vector code by allowing adjacent LoadS nodes to be isomorphic when they have different control RangeCheck nodes for a[i] and a[i+1] accesses in same MulAddS2I node.
      Separating it from JDK-8216050 in a new bug.

          [JDK-8216580] Fix generation of VNNI vector code by allowing adjacent LoadS nodes to be isomorphic

          HG Updates added a comment -
          URL: http://hg.openjdk.java.net/jdk/jdk/rev/6ffb8d7fe1e4
          User: vdeshpande
          Date: 2019-03-05 22:08:21 +0000

          HG Updates added a comment - URL: http://hg.openjdk.java.net/jdk/jdk/rev/6ffb8d7fe1e4 User: vdeshpande Date: 2019-03-05 22:08:21 +0000

          Vivek Deshpande (Inactive) added a comment -
          Webrev with an added check against NULL.
          http://cr.openjdk.java.net/~vdeshpande/8216580/webrev.03/

          Vivek Deshpande (Inactive) added a comment - Webrev with an added check against NULL. http://cr.openjdk.java.net/~vdeshpande/8216580/webrev.03/

          Vladimir Kozlov added a comment -
          My tier1-tier3 testing passed with few unrelated failures. Note, we don't have VNNI machines so the code is not exercised.
          [~vdeshpande] is running testing on VNNI machine.

          Vladimir Kozlov added a comment - My tier1-tier3 testing passed with few unrelated failures. Note, we don't have VNNI machines so the code is not exercised. [~vdeshpande] is running testing on VNNI machine.

          Vladimir Kozlov added a comment -
          http://cr.openjdk.java.net/~vdeshpande/8216580/webrev.02/

          Vladimir Kozlov added a comment - http://cr.openjdk.java.net/~vdeshpande/8216580/webrev.02/

          Vivek Deshpande (Inactive) added a comment -
          Fixed parentheses at 1225,1226
          http://cr.openjdk.java.net/~vdeshpande/8216580/webrev.01/

          Vivek Deshpande (Inactive) added a comment - Fixed parentheses at 1225,1226 http://cr.openjdk.java.net/~vdeshpande/8216580/webrev.01/

          Vivek Deshpande (Inactive) added a comment -
          Hi Rahul,
          The jtreg test Vec_MulAddS2I.java is the failing case.
          It does not generate wrong code, but it does not get vectorized as it cannot pack the adjacent LoadS nodes.

          Vivek Deshpande (Inactive) added a comment - Hi Rahul, The jtreg test Vec_MulAddS2I.java is the failing case. It does not generate wrong code, but it does not get vectorized as it cannot pack the adjacent LoadS nodes.

          Tobias Hartmann added a comment -
          Shouldn't this be an enhancement?

          Tobias Hartmann added a comment - Shouldn't this be an enhancement?

          Rahul Raghavan added a comment -
          initial ILW = wrong vector code generation; not commonly reported!; disable vectorization = HLM = P3

          Hi [~vdeshpande], please add a sample failing test case if any. Thanks.

          Rahul Raghavan added a comment - initial ILW = wrong vector code generation; not commonly reported!; disable vectorization = HLM = P3 Hi [~vdeshpande], please add a sample failing test case if any. Thanks.

          Vivek Deshpande (Inactive) added a comment - - edited
          Fix generation of vector code by allowing adjacent LoadS nodes to be isomorphic when they have different control RangeCheck nodes for a[i] and a[i+1] accesses in same MulAddS2I node.
          Webrev:
          http://cr.openjdk.java.net/~vdeshpande/8216580/webrev.00/

          Vivek Deshpande (Inactive) added a comment - - edited Fix generation of vector code by allowing adjacent LoadS nodes to be isomorphic when they have different control RangeCheck nodes for a[i] and a[i+1] accesses in same MulAddS2I node. Webrev: http://cr.openjdk.java.net/~vdeshpande/8216580/webrev.00/

            vdeshpande Vivek Deshpande (Inactive)
            vdeshpande Vivek Deshpande (Inactive)
            Votes:
            0 Vote for this issue
            Watchers:
            4 Start watching this issue

              Created:
              Updated:
              Resolved: