nitsanw

Index: jctools-core/src/main/java/org/jctools/maps/NonBlockingHashMap.java
IDEA additional info:
Subsystem: com.intellij.openapi.diff.impl.patch.CharsetEP
<+>UTF-8
===================================================================
--- jctools-core/src/main/java/org/jctools/maps/NonBlockingHashMap.java	(revision a2191c1e0243c8cd4690990d6b250b26a3ca8bfa)
+++ jctools-core/src/main/java/org/jctools/maps/NonBlockingHashMap.java	(revision )
@@ -650,134 +650,130 @@
     int reprobe_cnt=0;
     Object K=null, V=null;

nitsanw

  0.18%    0.00%  │││   │  0x00007f485a84780d: lock addl $0x0,(%rsp)     ;*invokevirtual putObjectRelease {reexecute=0 rethrow=0 return_oop=0}
                  │││   │                                                ; - sun.misc.Unsafe::putOrderedObject@7 (line 1042)
                  │││   │                                                ; - org.jctools.util.UnsafeRefArrayAccess::soElement@6 (line 77)
                  │││   │                                                ; - org.jctools.queues.SpscArrayQueue::offer@64 (line 187)
                  │││   │                                                ; - org.jctools.jmh.throughput.QueueThroughputBackoffNone::offer@8 (line 92)
                  │││   │                                                ; - org.jctools.jmh.throughput.generated.QueueThroughputBackoffNone_tpt_jmhTest::offer_thrpt_jmhStub@17 (line 299)
 22.58%   31.59%  │││   │  0x00007f485a847812: inc    %rbp               ;*ladd {reexecute=0 rethrow=0 return_oop=0}
                  │││   │      

nitsanw

// CPP, OpenJDK C2 compiler style... this is from:
// http://hg.openjdk.java.net/jdk9/jdk9/hotspot/file/b756e7a2ec33/src/share/vm/opto/graphKit.cpp#l4285
void GraphKit::g1_write_barrier_post(Node* oop_store,
                                     Node* obj,
                                     Node* adr,
                                     uint alias_idx,
                                     Node* val,
                                     BasicType bt,
                                     bool use_precise) {
/// ... Loads of other crap here...

nitsanw

# c2, level 4, org.jctools.jmh.throughput.generated.QueueThroughputBackoffNone_tpt_jmhTest::offer_thrpt_jmhStub, version 643 (89 bytes)
 0,46%| 0xb254: mov    %r11,%rcx          ;*getfield buffer {reexecute=0 rethrow=0 return_oop=0}
      |                                   ; - org.jctools.queues.SpscArrayQueue::offer@14 (line 176)
      |                                   ; - org.jctools.jmh.throughput.QueueThroughputBackoffNone::offer@8 (line 92)
      |                                   ; - org.jctools.jmh.throughput.generated.QueueThroughputBackoffNone_tpt_jmhTest::offer_thrpt_jmhStub@17 (line 299)
      | 0xb257: lea    0x10(%r11,%r10,4),%r13  ;*invokevirtual getObjectVolatile {reexecute=0 rethrow=0 return_oop=0}
      |                                   ; - sun.misc.Unsafe::getObjectVolatile@5 (line 923)
      |                                   ; - org.jctools.util.UnsafeRefArrayAccess::lvElement@5 (line 103)
      |                                   ; - org.jctools.queues.SpscArrayQueue::offerSlowPath

nitsanw

protected final boolean offerColdPath(final E[] buffer, final long mask, final E e, final long index,
        final long offset) {
  final long lookAheadStep = this.lookAheadStep;
  // normal case, go around the buffer or resize if full (unless we hit max capacity)
  if (lookAheadStep > 0) {
    long lookAheadElementOffset = calcElementOffset(index + lookAheadStep, mask);
    // Try and look ahead a number of elements so we don't have to do this all the time
    if (null == lvElement(buffer, lookAheadElementOffset)) {
      producerBufferLimit = index + lookAheadStep - 1; // joy, there's plenty of room
      writeToQueue(buffer, e, index, offset);

nitsanw

@Override
protected final boolean offerColdPath(E[] buffer, long mask, E e, long pIndex, long offset) {
  // use a fixed lookahead step based on buffer capacity
  final long lookAheadStep = (mask + 1) / 4;
  long pBufferLimit = pIndex + lookAheadStep;
  long pQueueLimit = producerQueueLimit;

  // out of known space
  if (pIndex >= pQueueLimit) {
    // we tested against a potentially out of date queue limit, refresh it

nitsanw

private E newBufferPoll(E[] buffer, final long index) {
  final E[] newBuffer = lvNextArrayAndUnlink(buffer);
  final long newMask = newBuffer.length - 2; // because buffer last element is for linking
  
  // update consumer view on buffer
  consumerBuffer = nextBuffer;
  consumerMask = newMask;
  
  // poll element, must be visible in new buffer
  final long offsetInNew = calcElementOffset(index, newMask);

nitsanw

@Override
protected boolean offerColdPath(E[] buffer, long mask, E e, long pIndex, long offset) {
  // use a fixed lookahead step based on buffer capacity
  final long lookAheadStep = (mask + 1) / 4;
  long pBufferLimit = pIndex + lookAheadStep;

  // go around the buffer or add a new buffer
  if (null == lvElement(buffer, calcElementOffset(pBufferLimit, mask))) {
    // the element lookAheadStep away is empty, so we can use up to the element before it
    producerBufferLimit = pBufferLimit - 1;

nitsanw

@Override
public boolean offer(final E e) {
  Objects.requireNonNull(e);
  
  // local load of fields to avoid repeated loads after volatile reads
  final E[] buffer = producerBuffer;
  final long index = producerIndex;
  final long mask = producerMask;
  final long offset = calcElementOffset(index, mask);

nitsanw

TotalCount=241116
Period(ms)=59994
Throughput(ops/sec)=4019.00
Min=164
Mean=630.28
50.000ptile=425
90.000ptile=1116
99.000ptile=3080
99.900ptile=11517
99.990ptile=50757