aboutsummaryrefslogtreecommitdiffhomepage
path: root/include/private/SkArenaAlloc.h
blob: 1915f8f9bed3a9d0a6f0d94a9c8caa18092a29e6 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
/*
 * Copyright 2016 Google Inc.
 *
 * Use of this source code is governed by a BSD-style license that can be
 * found in the LICENSE file.
 */

#ifndef SkArenaAlloc_DEFINED
#define SkArenaAlloc_DEFINED

#include <cassert>
#include <cstddef>
#include <cstdint>
#include <cstdlib>
#include <cstring>
#include <limits>
#include <new>
#include <type_traits>
#include <utility>
#include <vector>

// SkArenaAlloc allocates object and destroys the allocated objects when destroyed. It's designed
// to minimize the number of underlying block allocations. SkArenaAlloc allocates first out of an
// (optional) user-provided block of memory, and when that's exhausted it allocates on the heap,
// starting with an allocation of extraSize bytes.  If your data (plus a small overhead) fits in
// the user-provided block, SkArenaAlloc never uses the heap, and if it fits in extraSize bytes,
// it'll use the heap only once.  If you pass extraSize = 0, it allocates blocks for each call to
// make<T>.
//
// Examples:
//
//   char block[mostCasesSize];
//   SkArenaAlloc arena(block, mostCasesSize);
//
// If mostCasesSize is too large for the stack, you can use the following pattern.
//
//   std::unique_ptr<char[]> block{new char[mostCasesSize]};
//   SkArenaAlloc arena(block.get(), mostCasesSize, almostAllCasesSize);
//
// If the program only sometimes allocates memory, use the following pattern.
//
//   SkArenaAlloc arena(nullptr, 0, almostAllCasesSize);
//
// The storage does not necessarily need to be on the stack. Embedding the storage in a class also
// works.
//
//   class Foo {
//       char storage[mostCasesSize];
//       SkArenaAlloc arena (storage, mostCasesSize);
//   };
//
// In addition, the system is optimized to handle POD data including arrays of PODs (where
// POD is really data with no destructors). For POD data it has zero overhead per item, and a
// typical block overhead of 8 bytes. For non-POD objects there is a per item overhead of 4 bytes.
// For arrays of non-POD objects there is a per array overhead of typically 8 bytes. There is an
// addition overhead when switching from POD data to non-POD data of typically 8 bytes.
//
// If additional blocks are needed they are increased exponentially. This strategy bounds the
// recursion of the RunDtorsOnBlock to be limited to O(log size-of-memory). Block size grow using
// the Fibonacci sequence which means that for 2^32 memory there are 48 allocations, and for 2^48
// there are 71 allocations.
class SkArenaAlloc {
public:
    SkArenaAlloc(char* block, size_t blockSize, size_t extraSize);

    SkArenaAlloc(size_t extraSize)
        : SkArenaAlloc(nullptr, 0, extraSize)
    {}

    ~SkArenaAlloc();

    template <typename T, typename... Args>
    T* make(Args&&... args) {
        uint32_t size      = ToU32(sizeof(T));
        uint32_t alignment = ToU32(alignof(T));
        char* objStart;
        if (std::is_trivially_destructible<T>::value) {
            objStart = this->allocObject(size, alignment);
            fCursor = objStart + size;
        } else {
            objStart = this->allocObjectWithFooter(size + sizeof(Footer), alignment);
            // Can never be UB because max value is alignof(T).
            uint32_t padding = ToU32(objStart - fCursor);

            // Advance to end of object to install footer.
            fCursor = objStart + size;
            FooterAction* releaser = [](char* objEnd) {
                char* objStart = objEnd - (sizeof(T) + sizeof(Footer));
                ((T*)objStart)->~T();
                return objStart;
            };
            this->installFooter(releaser, padding);
        }

        // This must be last to make objects with nested use of this allocator work.
        return new(objStart) T(std::forward<Args>(args)...);
    }

    template <typename T>
    T* makeArrayDefault(size_t count) {
        uint32_t safeCount = ToU32(count);
        T* array = (T*)this->commonArrayAlloc<T>(safeCount);

        // If T is primitive then no initialization takes place.
        for (size_t i = 0; i < safeCount; i++) {
            new (&array[i]) T;
        }
        return array;
    }

    template <typename T>
    T* makeArray(size_t count) {
        uint32_t safeCount = ToU32(count);
        T* array = (T*)this->commonArrayAlloc<T>(safeCount);

        // If T is primitive then the memory is initialized. For example, an array of chars will
        // be zeroed.
        for (size_t i = 0; i < safeCount; i++) {
            new (&array[i]) T();
        }
        return array;
    }

    // Only use makeBytesAlignedTo if none of the typed variants are impractical to use.
    void* makeBytesAlignedTo(size_t size, size_t align) {
        auto objStart = this->allocObject(ToU32(size), ToU32(align));
        fCursor = objStart + size;
        return objStart;
    }

    // Destroy all allocated objects, free any heap allocations.
    void reset();

private:
    static void AssertRelease(bool cond) { if (!cond) { ::abort(); } }
    static uint32_t ToU32(size_t v) {
#if SIZE_MAX > 0xffffffff
        AssertRelease(v <= 0xffffffff);
#endif
        return (uint32_t)v;
    }

    using Footer = int64_t;
    using FooterAction = char* (char*);

    static char* SkipPod(char* footerEnd);
    static void RunDtorsOnBlock(char* footerEnd);
    static char* NextBlock(char* footerEnd);

    void installFooter(FooterAction* releaser, uint32_t padding);
    void installUint32Footer(FooterAction* action, uint32_t value, uint32_t padding);
    void installPtrFooter(FooterAction* action, char* ptr, uint32_t padding);

    void ensureSpace(uint32_t size, uint32_t alignment);

    char* allocObject(uint32_t size, uint32_t alignment) {
        uintptr_t mask = alignment - 1;
        uintptr_t alignedOffset = (~reinterpret_cast<uintptr_t>(fCursor) + 1) & mask;
        uintptr_t totalSize = size + alignedOffset;
        AssertRelease(totalSize >= size);
        if (totalSize > static_cast<uintptr_t>(fEnd - fCursor)) {
            this->ensureSpace(size, alignment);
            alignedOffset = (~reinterpret_cast<uintptr_t>(fCursor) + 1) & mask;
        }
        return fCursor + alignedOffset;
    }

    char* allocObjectWithFooter(uint32_t sizeIncludingFooter, uint32_t alignment);

    template <typename T>
    char* commonArrayAlloc(uint32_t count) {
        char* objStart;
        AssertRelease(count <= std::numeric_limits<uint32_t>::max() / sizeof(T));
        uint32_t arraySize = ToU32(count * sizeof(T));
        uint32_t alignment = ToU32(alignof(T));

        if (std::is_trivially_destructible<T>::value) {
            objStart = this->allocObject(arraySize, alignment);
            fCursor = objStart + arraySize;
        } else {
            constexpr uint32_t overhead = sizeof(Footer) + sizeof(uint32_t);
            AssertRelease(arraySize <= std::numeric_limits<uint32_t>::max() - overhead);
            uint32_t totalSize = arraySize + overhead;
            objStart = this->allocObjectWithFooter(totalSize, alignment);

            // Can never be UB because max value is alignof(T).
            uint32_t padding = ToU32(objStart - fCursor);

            // Advance to end of array to install footer.?
            fCursor = objStart + arraySize;
            this->installUint32Footer(
                [](char* footerEnd) {
                    char* objEnd = footerEnd - (sizeof(Footer) + sizeof(uint32_t));
                    uint32_t count;
                    memmove(&count, objEnd, sizeof(uint32_t));
                    char* objStart = objEnd - count * sizeof(T);
                    T* array = (T*) objStart;
                    for (uint32_t i = 0; i < count; i++) {
                        array[i].~T();
                    }
                    return objStart;
                },
                ToU32(count),
                padding);
        }

        return objStart;
    }

    char*          fDtorCursor;
    char*          fCursor;
    char*          fEnd;
    char* const    fFirstBlock;
    const uint32_t fFirstSize;
    const uint32_t fExtraSize;

    // Use the Fibonacci sequence as the growth factor for block size. The size of the block
    // allocated is fFib0 * fExtraSize. Using 2 ^ n * fExtraSize had too much slop for Android.
    uint32_t       fFib0 {1}, fFib1 {1};
};

// Helper for defining allocators with inline/reserved storage.
// For argument declarations, stick to the base type (SkArenaAlloc).
template <size_t InlineStorageSize>
class SkSTArenaAlloc : public SkArenaAlloc {
public:
    explicit SkSTArenaAlloc(size_t extraSize = InlineStorageSize)
        : INHERITED(fInlineStorage, InlineStorageSize, extraSize) {}

private:
    char fInlineStorage[InlineStorageSize];

    using INHERITED = SkArenaAlloc;
};

#endif  // SkArenaAlloc_DEFINED