Class BytecodeOSRMetadata

java.lang.Object
com.oracle.truffle.runtime.BytecodeOSRMetadata
public final class BytecodeOSRMetadata extends Object
Represents the metadata required to perform OSR compilation on Graal. An instance of this class is stored in the metadata field of a BytecodeOSRNode.

Performance note: We do not require the metadata field to be volatile. As long as the field is initialized under double-checked locking (as is done in Accessor.RuntimeSupport.pollBytecodeOSRBackEdge(BytecodeOSRNode)), all threads will observe the same metadata instance. The JMM guarantees that the instance's final fields will be safely initialized before it is published; the non-final + non-volatile fields (e.g., the back edge counter) may not be, but we tolerate this inaccuracy in order to avoid volatile accesses in the hot path.

  • Field Details

  • Method Details

    • incrementAndPoll

      public boolean incrementAndPoll()
      Increment back edge count and return whether compilation should be polled.

      When the OSR threshold is reached, this method will return true after every OSR_POLL_INTERVAL back-edges. This method is thread-safe, but could under-count.

    • incrementAndPoll

      public boolean incrementAndPoll(int loopCountIncrement)

    • isDisabled

      public boolean isDisabled()
      Returns whether OSR compilation is disabled for the current unit.

    • forceDisable

      public void forceDisable()
      Force disabling of OSR compilation for this method. Used for testing purposes.

    • transferFrame

      public void transferFrame(com.oracle.truffle.api.impl.FrameWithoutBoxing source, com.oracle.truffle.api.impl.FrameWithoutBoxing target, Object targetMetadata)
      Transfer state from source to target. Can be used to transfer state into an OSR frame.

    • restoreFrame

      public void restoreFrame(com.oracle.truffle.api.impl.FrameWithoutBoxing source, com.oracle.truffle.api.impl.FrameWithoutBoxing target)
      Transfer state from source to target. Can be used to transfer state from an OSR frame to a parent frame. Overall less efficient than its counterpart, mainly due to not being able to speculate on the source tags: While entering bytecode OSR is done through specific entry points (likely back edges), returning could be done from anywhere within a method body (through regular returns, or exception thrown). While we could theoretically have the same mechanism as on entries (caching encountered return state), we could not efficiently be able to retrieve from the cache (as we do not get the equivalent of the int osrBytecodeTarget for returns), even ignoring the potential memory cost of such a cache. Therefore, we are doing a best-effort of copying over source to target, reading from the actual frames rather than a cache: If the tag cannot be constant-folded at this point, we get a switch in compiled code. Since we are at a boundary back to interpreted code, this cost should not be too high.

    • getOSRCompilations

      public Map<Long, OptimizedCallTarget> getOSRCompilations()

    • getBackEdgeCount

      public int getBackEdgeCount()