• mark.lam@apple.com's avatar
    Poor man's fast breakpoints for a 2.3x debugger speedup. · 945b1399
    mark.lam@apple.com authored
    Reviewed by Geoffrey Garen.
    Previously we gained back some performance (run at baseline JIT speeds)
    when the WebInspector is opened provided no breakpoints are set. This
    was achieved by simply skipping all op_debug callbacks to the debugger
    if no breakpoints are set. If any breakpoints are set, the debugger will
    set a m_needsOpDebugCallbacks flag which causes the callbacks to be
    called, and we don't get the baseline JIT speeds anymore.
    With this patch, we will now track the number of breakpoints set in the
    CodeBlock that they are set in. The LLINT and baseline JIT code will
    check CodeBlock::m_numBreakpoints to determine if the op_debug callbacks
    need to be called. With this, we will only enable op_debug callbacks for
    CodeBlocks that need it i.e. those with breakpoints set in them.
    Debugger::m_needsOpDebugCallbacks is now obsoleted. The LLINT and baseline
    JIT code still needs to check Debugger::m_shouldPause to determine if the
    debugger is in stepping mode and hence, needs op_debug callbacks enabled
    for everything until the debugger "continues" the run and exit stepping
    Also in this patch, I fixed a regression in DOM breakpoints which relies
    Debugger::breakProgram() to pause the debugger.
    * bytecode/CodeBlock.cpp:
    - Missed accounting for op_debug's new hasBreakpointFlag operand here when
      it was added.
    - This is needed in Debugger::toggleBreakpoint() to determine if a
      breakpoint falls within a CodeBlock or not. Simply checking the bounds
      of the CodeBlock is insufficient. For example, let's say we have the
      following JS code:
          // begin global scope
          function f1() {
              function f2() {
                 ... // set breakpoint here.
          // end global scope
      Using the CodeBlock bounds alone, the breakpoint above will to appear
      to be in the global program CodeBlock, and the CodeBlocks for function
      f1() and f2(). With CodeBlock::hasOpDebugForLineAndColumn() we can
      rule out the global program CodeBlock and f1(), and only apply the
      breakpoint to f2(0 where it belongs.
      CodeBlock::hasOpDebugForLineAndColumn() works by iterating over all
      the opcodes in the CodeBlock to look for op_debug's. For each op_debug,
      it calls CodeBlock::expressionRangeForBytecodeOffset() to do a binary
      seach to get the line and column info for that op_debug. This is a
      N * log(N) algorithm. However, a quick hands on test using the
      WebInspector (with this patch applied) to exercise setting, breaking
      on, and clearing breakpoints, as well as stepping through some code
      shows no noticeable degradation of the user experience compared to the
      baseline without this patch.
    * bytecode/CodeBlock.h:
    * debugger/Breakpoint.h:
    - defined Breakpoint::unspecifiedColumn so that we can explicitly indicate
      when the WebInspector was setting a line breakpoint and did not provide
      a column value. CodeBlock::hasOpDebugForLineAndColumn() needs this
      information in order to loosen its matching criteria for op_debug
      bytecodes for the specified breakpoint line and column values provided
      by the debugger.
      Previously, we just hijack a 0 value column as an unspecified column.
      However, the WebInspector operates on 0-based ints for column values.
      Hence, 0 should be a valid column value and should not be hijacked to
      mean an unspecified column.
    * debugger/Debugger.cpp:
    - added tracking of the VM that the debugger is used with. This is
      needed by Debugger::breakProgram().
      The VM pointer is attained from the first JSGlobalObject that the debugger
      attaches to. When the debugger detaches from the last JSGlobalObject, it
      will nullify its VM pointer to allow a new one to be set on the next
      We were always only using each debugger instance with one VM. This change
      makes it explicit with an assert to ensure that all globalObjects that
      the debugger attaches to beongs to the same VM.
    - registerCodeBlock() is responsible for applying pre-existing breakpoints
      to new CodeBlocks being installed. Similarly, unregisterCodeBlock()
      clears the breakpoints.
    - This is the workhorse function that checks if a breakpoint falls within
      a CodeBlock or not. If it does, then it can either enable or disable
      said breakpoint in the CodeBlock. In the current implementation,
      enabling/disabling the breakpoint simply means incrementing/decrementing
      the CodeBlock's m_numBreakpoints.
    - Iterates all relevant CodeBlocks and apply the specified breakpoint
      if appropriate. This is called when a new breakpoint is being defined
      by the WebInspector and needs to be applied to an already installed
    - Fixed a regression that broke DOM breakpoints. The issue is that with
      the skipping of op_debug callbacks, we don't always have an updated
      m_currentCallFrame. Normally, m_currentCallFrame is provided as arg
      in the op_debug callback. In this case, we can get the CallFrame* from
    * debugger/Debugger.h:
    * heap/CodeBlockSet.h:
    * heap/Heap.h:
    - Added utility to iterate all CodeBlocks in the heap / VM.
    * interpreter/Interpreter.cpp:
    * jit/JITOpcodes.cpp:
    * jit/JITOpcodes32_64.cpp:
    * llint/LowLevelInterpreter.asm:
    - These now checks CodeBlock::m_numBreakpoints and Debugger::m_shouldPause
      instead of Debugger::m_needsOpDebugCallbacks.
    * runtime/Executable.cpp:
    git-svn-id: http://svn.webkit.org/repository/webkit/trunk@162598 268f45cc-cd09-0410-ab3c-d52691b4dbfc
CodeBlock.h 47.3 KB