OLD | NEW |
---|---|
1 // Copyright 2010 the V8 project authors. All rights reserved. | 1 // Copyright 2010 the V8 project authors. All rights reserved. |
2 // Redistribution and use in source and binary forms, with or without | 2 // Redistribution and use in source and binary forms, with or without |
3 // modification, are permitted provided that the following conditions are | 3 // modification, are permitted provided that the following conditions are |
4 // met: | 4 // met: |
5 // | 5 // |
6 // * Redistributions of source code must retain the above copyright | 6 // * Redistributions of source code must retain the above copyright |
7 // notice, this list of conditions and the following disclaimer. | 7 // notice, this list of conditions and the following disclaimer. |
8 // * Redistributions in binary form must reproduce the above | 8 // * Redistributions in binary form must reproduce the above |
9 // copyright notice, this list of conditions and the following | 9 // copyright notice, this list of conditions and the following |
10 // disclaimer in the documentation and/or other materials provided | 10 // disclaimer in the documentation and/or other materials provided |
(...skipping 16 matching lines...) Expand all Loading... | |
27 | 27 |
28 #include "v8.h" | 28 #include "v8.h" |
29 | 29 |
30 #include "circular-queue-inl.h" | 30 #include "circular-queue-inl.h" |
31 | 31 |
32 namespace v8 { | 32 namespace v8 { |
33 namespace internal { | 33 namespace internal { |
34 | 34 |
35 | 35 |
36 SamplingCircularQueue::SamplingCircularQueue(size_t record_size_in_bytes, | 36 SamplingCircularQueue::SamplingCircularQueue(size_t record_size_in_bytes, |
37 size_t desired_chunk_size_in_bytes, | 37 unsigned buffer_size_in_records) |
38 unsigned buffer_size_in_chunks) | 38 : entry_size_(RoundUp(record_size_in_bytes + sizeof(Cell), |
39 : record_size_(record_size_in_bytes / sizeof(Cell)), | 39 kProcessorCacheLineSize) / sizeof(Cell)), |
40 chunk_size_in_bytes_(desired_chunk_size_in_bytes / record_size_in_bytes * | 40 buffer_size_(entry_size_ * buffer_size_in_records) { |
41 record_size_in_bytes + sizeof(Cell)), | 41 const size_t cache_line_size = kProcessorCacheLineSize / sizeof(Cell); |
42 chunk_size_(chunk_size_in_bytes_ / sizeof(Cell)), | 42 not_aligned_buffer_ = NewArray<Cell>(buffer_size_ + cache_line_size); |
43 buffer_size_(chunk_size_ * buffer_size_in_chunks), | 43 // Align on cache line boundaries. |
44 buffer_(NewArray<Cell>(buffer_size_)) { | 44 buffer_ = reinterpret_cast<Cell*>(RoundUp(not_aligned_buffer_, |
45 ASSERT(record_size_ * sizeof(Cell) == record_size_in_bytes); | 45 cache_line_size)); |
46 ASSERT(chunk_size_ * sizeof(Cell) == chunk_size_in_bytes_); | 46 // Mark all entries as empty. |
47 ASSERT(buffer_size_in_chunks > 2); | 47 for (size_t i = 0; i < buffer_size_; i += entry_size_) { |
48 // Mark all chunks as clear. | 48 buffer_[i] = kEmpty; |
49 for (size_t i = 0; i < buffer_size_; i += chunk_size_) { | |
50 buffer_[i] = kClear; | |
51 } | 49 } |
52 | 50 |
53 // Layout producer and consumer position pointers each on their own | 51 // Layout producer and consumer position pointers each on their own |
54 // cache lines to avoid cache lines thrashing due to simultaneous | 52 // cache lines to avoid cache lines thrashing due to simultaneous |
55 // updates of positions by different processor cores. | 53 // updates of positions by different processor cores. |
56 const int positions_size = | 54 const int positions_size = |
57 RoundUp(1, kProcessorCacheLineSize) + | 55 RoundUp(1, kProcessorCacheLineSize) + |
58 RoundUp(static_cast<int>(sizeof(ProducerPosition)), | 56 RoundUp(static_cast<int>(sizeof(ProducerPosition)), |
59 kProcessorCacheLineSize) + | 57 kProcessorCacheLineSize) + |
60 RoundUp(static_cast<int>(sizeof(ConsumerPosition)), | 58 RoundUp(static_cast<int>(sizeof(ConsumerPosition)), |
61 kProcessorCacheLineSize); | 59 kProcessorCacheLineSize); |
62 positions_ = NewArray<byte>(positions_size); | 60 positions_ = NewArray<byte>(positions_size); |
63 | 61 |
64 producer_pos_ = reinterpret_cast<ProducerPosition*>( | 62 producer_pos_ = reinterpret_cast<ProducerPosition*>( |
65 RoundUp(positions_, kProcessorCacheLineSize)); | 63 RoundUp(positions_, kProcessorCacheLineSize)); |
66 producer_pos_->next_chunk_pos = buffer_; | |
67 producer_pos_->enqueue_pos = buffer_; | 64 producer_pos_->enqueue_pos = buffer_; |
68 | 65 |
69 consumer_pos_ = reinterpret_cast<ConsumerPosition*>( | 66 consumer_pos_ = reinterpret_cast<ConsumerPosition*>( |
70 reinterpret_cast<byte*>(producer_pos_) + kProcessorCacheLineSize); | 67 reinterpret_cast<byte*>(producer_pos_) + kProcessorCacheLineSize); |
71 ASSERT(reinterpret_cast<byte*>(consumer_pos_ + 1) <= | 68 ASSERT(reinterpret_cast<byte*>(consumer_pos_ + 1) <= |
72 positions_ + positions_size); | 69 positions_ + positions_size); |
73 consumer_pos_->dequeue_chunk_pos = buffer_; | 70 consumer_pos_->dequeue_pos = buffer_; |
74 // The distance ensures that producer and consumer never step on | |
75 // each other's chunks and helps eviction of produced data from | |
76 // the CPU cache (having that chunk size is bigger than the cache.) | |
77 const size_t producer_consumer_distance = (2 * chunk_size_); | |
78 consumer_pos_->dequeue_chunk_poll_pos = buffer_ + producer_consumer_distance; | |
79 consumer_pos_->dequeue_pos = NULL; | |
80 } | 71 } |
81 | 72 |
82 | 73 |
83 SamplingCircularQueue::~SamplingCircularQueue() { | 74 SamplingCircularQueue::~SamplingCircularQueue() { |
84 DeleteArray(positions_); | 75 DeleteArray(positions_); |
85 DeleteArray(buffer_); | 76 DeleteArray(not_aligned_buffer_); |
86 } | 77 } |
87 | 78 |
88 | 79 |
89 void* SamplingCircularQueue::StartDequeue() { | 80 void* SamplingCircularQueue::StartDequeue() { |
90 if (consumer_pos_->dequeue_pos != NULL) { | 81 MemoryBarrier(); |
Benedikt Meurer
2013/08/13 09:31:32
Why do we need this memory barrier here?
yurys
2013/08/13 10:05:20
Because we'd like to see up-to-date value at consu
| |
91 return consumer_pos_->dequeue_pos; | 82 if (Acquire_Load(consumer_pos_->dequeue_pos) != kEmpty) { |
92 } else { | 83 // Skip marker. |
93 if (Acquire_Load(consumer_pos_->dequeue_chunk_poll_pos) != kClear) { | 84 return consumer_pos_->dequeue_pos + 1; |
94 // Skip marker. | |
95 consumer_pos_->dequeue_pos = consumer_pos_->dequeue_chunk_pos + 1; | |
96 consumer_pos_->dequeue_end_pos = | |
97 consumer_pos_->dequeue_chunk_pos + chunk_size_; | |
98 return consumer_pos_->dequeue_pos; | |
99 } else { | |
100 return NULL; | |
101 } | |
102 } | 85 } |
86 return NULL; | |
103 } | 87 } |
104 | 88 |
105 | 89 |
106 void SamplingCircularQueue::FinishDequeue() { | 90 void SamplingCircularQueue::FinishDequeue() { |
107 consumer_pos_->dequeue_pos += record_size_; | 91 Release_Store(consumer_pos_->dequeue_pos, kEmpty); |
108 if (consumer_pos_->dequeue_pos < consumer_pos_->dequeue_end_pos) return; | 92 consumer_pos_->dequeue_pos += entry_size_; |
109 // Move to next chunk. | 93 WrapPositionIfNeeded(&consumer_pos_->dequeue_pos); |
110 consumer_pos_->dequeue_pos = NULL; | |
111 *consumer_pos_->dequeue_chunk_pos = kClear; | |
112 consumer_pos_->dequeue_chunk_pos += chunk_size_; | |
113 WrapPositionIfNeeded(&consumer_pos_->dequeue_chunk_pos); | |
114 consumer_pos_->dequeue_chunk_poll_pos += chunk_size_; | |
115 WrapPositionIfNeeded(&consumer_pos_->dequeue_chunk_poll_pos); | |
116 } | 94 } |
Benedikt Meurer
2013/08/13 09:31:32
T* StartDequeue() {
if (Acquire_Load(&buffer_[de
yurys
2013/08/13 14:10:29
Changed that code.
| |
117 | 95 |
118 | |
119 void SamplingCircularQueue::FlushResidualRecords() { | |
120 // Eliminate producer / consumer distance. | |
121 consumer_pos_->dequeue_chunk_poll_pos = consumer_pos_->dequeue_chunk_pos; | |
122 } | |
123 | |
124 | |
125 } } // namespace v8::internal | 96 } } // namespace v8::internal |
OLD | NEW |