Uploaded image for project: 'JDK'
  1. JDK
  2. JDK-8274197

abstract assembly is imprecise wrt code comments

    XMLWordPrintable

    Details

      Description

      The nice abstract assembly dumping added by JDK-8213084 is not so precise with respect to the position of code comments. On x86, this means post processing MachCode sections with a disassembler can get confused.

      This is best shown with an example. With Graal, a ScopeDesc can be associated with the instruction immediately following a call (which is often a nop).
      So we expect to see something like this:

      ----------------------------------------------------------------------------------------------------------------
        0x00000001101bf5f4: ; {runtime_call DeoptimizationBlob}
        0x00000001101bf5f4: e881 6cfd | ff

        0x00000001101bf5f9: ; ImmutableOopMap {}
                            ;*iload_1 {reexecute=1 rethrow=0 return_oop=0}
                            ; - (reexecute) java.util.stream.StreamOpFlag::combineOpFlags@0 (line 708)
                            ; {runtime_call Stub<ExceptionHandlerStub.exceptionHandler>}
        0x00000001101bf5f9: 90 e801 | cdff ff90
      ----------------------------------------------------------------------------------------------------------------

      However, the output produced for this case is:

      ----------------------------------------------------------------------------------------------------------------
        0x00000001101bf5f4: ; {runtime_call DeoptimizationBlob}
        0x00000001101bf5f4: e881 6cfd

        0x00000001101bf5f8: ; ImmutableOopMap {}
                            ;*iload_1 {reexecute=1 rethrow=0 return_oop=0}
                            ; - (reexecute) java.util.stream.StreamOpFlag::combineOpFlags@0 (line 708)
                            ; {runtime_call Stub<ExceptionHandlerStub.exceptionHandler>}
        0x00000001101bf5f8: ff90 e801 | cdff ff90
      ----------------------------------------------------------------------------------------------------------------

      When post-processing with a disassembler, this means we get:

      ----------------------------------------------------------------------------------------------------------------
        0x00000001101bf5f4: <unknown: e8 81 6c fd> ; {runtime_call DeoptimizationBlob}
        0x00000001101bf5f8: call qword ptr [rax - 0x32fe18] ; ImmutableOopMap {}
                                                                   ;*iload_1 {reexecute=1 rethrow=0 return_oop=0}
                                                                   ; - (reexecute) java.util.stream.StreamOpFlag::combineOpFlags@0 (line 708)
                                                                   ; {runtime_call Stub<ExceptionHandlerStub.exceptionHandler>}
      [Deopt Handler Code]
        0x00000001101bf600: call 0x110196020 ; {runtime_call DeoptimizationBlob}
      ----------------------------------------------------------------------------------------------------------------

      instead of:

      ----------------------------------------------------------------------------------------------------------------
        0x00000001101bf5f4: call 0x11019627a ; {runtime_call DeoptimizationBlob}
        0x00000001101bf5f9: nop ; ImmutableOopMap {}
                                                                   ;*iload_1 {reexecute=1 rethrow=0 return_oop=0}
                                                                   ; - (reexecute) java.util.stream.StreamOpFlag::combineOpFlags@0 (line 708)
                                                                   ; {runtime_call Stub<ExceptionHandlerStub.exceptionHandler>}
        0x00000001101bf5fa: call 0x1101bc300
      ----------------------------------------------------------------------------------------------------------------

      I think the main change needed is for a code comment to be emitted at exactly the start address of the instruction it is associated with. That will prevent the instruction stream in a MachNode section from being interrupted in the middle of an instruction.

        Attachments

          Issue Links

            Activity

              People

              Assignee:
              Unassigned Unassigned
              Reporter:
              dnsimon Douglas Simon
              Votes:
              0 Vote for this issue
              Watchers:
              2 Start watching this issue

                Dates

                Created:
                Updated: