gc.c 16 KB
Newer Older
Damien's avatar
Damien committed
1
2
3
4
5
#include <stdio.h>
#include <stdlib.h>
#include <stdint.h>
#include <string.h>

6
#include "mpconfig.h"
Damien's avatar
Damien committed
7
8
#include "gc.h"

9
10
#if MICROPY_ENABLE_GC

11
12
#if 0 // print debugging info
#define DEBUG_PRINT (1)
13
#define DEBUG_printf DEBUG_printf
14
#else // don't print debugging info
15
#define DEBUG_printf(...) (void)0
16
17
#endif

Damien's avatar
Damien committed
18
19
20
21
22
23
typedef unsigned char byte;

#define WORDS_PER_BLOCK (4)
#define BYTES_PER_BLOCK (WORDS_PER_BLOCK * BYTES_PER_WORD)
#define STACK_SIZE (64) // tunable; minimum is 1

24
25
26
27
STATIC byte *gc_alloc_table_start;
STATIC machine_uint_t gc_alloc_table_byte_len;
STATIC machine_uint_t *gc_pool_start;
STATIC machine_uint_t *gc_pool_end;
Damien's avatar
Damien committed
28

29
30
31
STATIC int gc_stack_overflow;
STATIC machine_uint_t gc_stack[STACK_SIZE];
STATIC machine_uint_t *gc_sp;
Damien's avatar
Damien committed
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

// ATB = allocation table byte
// 0b00 = FREE -- free block
// 0b01 = HEAD -- head of a chain of blocks
// 0b10 = TAIL -- in the tail of a chain of blocks
// 0b11 = MARK -- marked head block

#define AT_FREE (0)
#define AT_HEAD (1)
#define AT_TAIL (2)
#define AT_MARK (3)

#define BLOCKS_PER_ATB (4)
#define ATB_MASK_0 (0x03)
#define ATB_MASK_1 (0x0c)
#define ATB_MASK_2 (0x30)
#define ATB_MASK_3 (0xc0)

#define ATB_0_IS_FREE(a) (((a) & ATB_MASK_0) == 0)
#define ATB_1_IS_FREE(a) (((a) & ATB_MASK_1) == 0)
#define ATB_2_IS_FREE(a) (((a) & ATB_MASK_2) == 0)
#define ATB_3_IS_FREE(a) (((a) & ATB_MASK_3) == 0)

#define BLOCK_SHIFT(block) (2 * ((block) & (BLOCKS_PER_ATB - 1)))
#define ATB_GET_KIND(block) ((gc_alloc_table_start[(block) / BLOCKS_PER_ATB] >> BLOCK_SHIFT(block)) & 3)
#define ATB_ANY_TO_FREE(block) do { gc_alloc_table_start[(block) / BLOCKS_PER_ATB] &= (~(AT_MARK << BLOCK_SHIFT(block))); } while (0)
#define ATB_FREE_TO_HEAD(block) do { gc_alloc_table_start[(block) / BLOCKS_PER_ATB] |= (AT_HEAD << BLOCK_SHIFT(block)); } while (0)
#define ATB_FREE_TO_TAIL(block) do { gc_alloc_table_start[(block) / BLOCKS_PER_ATB] |= (AT_TAIL << BLOCK_SHIFT(block)); } while (0)
#define ATB_HEAD_TO_MARK(block) do { gc_alloc_table_start[(block) / BLOCKS_PER_ATB] |= (AT_MARK << BLOCK_SHIFT(block)); } while (0)
#define ATB_MARK_TO_HEAD(block) do { gc_alloc_table_start[(block) / BLOCKS_PER_ATB] &= (~(AT_TAIL << BLOCK_SHIFT(block))); } while (0)

#define BLOCK_FROM_PTR(ptr) (((ptr) - (machine_uint_t)gc_pool_start) / BYTES_PER_BLOCK)
#define PTR_FROM_BLOCK(block) (((block) * BYTES_PER_BLOCK + (machine_uint_t)gc_pool_start))
#define ATB_FROM_BLOCK(bl) ((bl) / BLOCKS_PER_ATB)

67
68
69
70
// TODO waste less memory; currently requires that all entries in alloc_table have a corresponding block in pool
void gc_init(void *start, void *end) {
    // align end pointer on block boundary
    end = (void*)((machine_uint_t)end & (~(BYTES_PER_BLOCK - 1)));
71
    DEBUG_printf("Initializing GC heap: %p-%p\n", start, end);
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91

    // calculate parameters for GC
    machine_uint_t total_word_len = (machine_uint_t*)end - (machine_uint_t*)start;
    gc_alloc_table_byte_len = total_word_len * BYTES_PER_WORD / (1 + BITS_PER_BYTE / 2 * BYTES_PER_BLOCK);
    gc_alloc_table_start = (byte*)start;
    machine_uint_t gc_pool_block_len = gc_alloc_table_byte_len * BITS_PER_BYTE / 2;
    machine_uint_t gc_pool_word_len = gc_pool_block_len * WORDS_PER_BLOCK;
    gc_pool_start = (machine_uint_t*)end - gc_pool_word_len;
    gc_pool_end = end;

    // clear ATBs
    memset(gc_alloc_table_start, 0, gc_alloc_table_byte_len);

    // allocate first block because gc_pool_start points there and it will never
    // be freed, so allocating 1 block with null pointers will minimise memory loss
    ATB_FREE_TO_HEAD(0);
    for (int i = 0; i < WORDS_PER_BLOCK; i++) {
        gc_pool_start[i] = 0;
    }

92
    DEBUG_printf("GC layout:\n");
93
94
    DEBUG_printf("  alloc table at %p, length " UINT_FMT " bytes\n", gc_alloc_table_start, gc_alloc_table_byte_len);
    DEBUG_printf("  pool at %p, length " UINT_FMT " blocks = " UINT_FMT " words = " UINT_FMT " bytes\n", gc_pool_start, gc_pool_block_len, gc_pool_word_len, gc_pool_word_len * BYTES_PER_WORD);
95
96
}

97
98
99
100
101
102
#define VERIFY_PTR(ptr) ( \
        (ptr & (BYTES_PER_BLOCK - 1)) == 0          /* must be aligned on a block */ \
        && ptr >= (machine_uint_t)gc_pool_start     /* must be above start of pool */ \
        && ptr < (machine_uint_t)gc_pool_end        /* must be below end of pool */ \
    )

Damien's avatar
Damien committed
103
104
#define VERIFY_MARK_AND_PUSH(ptr) \
    do { \
105
        if (VERIFY_PTR(ptr)) { \
Damien's avatar
Damien committed
106
107
108
109
110
111
112
113
114
115
116
117
118
            machine_uint_t _block = BLOCK_FROM_PTR(ptr); \
            if (ATB_GET_KIND(_block) == AT_HEAD) { \
                /* an unmarked head, mark it, and push it on gc stack */ \
                ATB_HEAD_TO_MARK(_block); \
                if (gc_sp < &gc_stack[STACK_SIZE]) { \
                    *gc_sp++ = _block; \
                } else { \
                    gc_stack_overflow = 1; \
                } \
            } \
        } \
    } while (0)

119
STATIC void gc_drain_stack(void) {
Damien's avatar
Damien committed
120
121
122
123
    while (gc_sp > gc_stack) {
        // pop the next block off the stack
        machine_uint_t block = *--gc_sp;

124
        // work out number of consecutive blocks in the chain starting with this one
Damien's avatar
Damien committed
125
126
127
128
129
130
131
132
133
134
135
136
137
138
        machine_uint_t n_blocks = 0;
        do {
            n_blocks += 1;
        } while (ATB_GET_KIND(block + n_blocks) == AT_TAIL);

        // check this block's children
        machine_uint_t *scan = (machine_uint_t*)PTR_FROM_BLOCK(block);
        for (machine_uint_t i = n_blocks * WORDS_PER_BLOCK; i > 0; i--, scan++) {
            machine_uint_t ptr2 = *scan;
            VERIFY_MARK_AND_PUSH(ptr2);
        }
    }
}

139
STATIC void gc_deal_with_stack_overflow(void) {
Damien's avatar
Damien committed
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
    while (gc_stack_overflow) {
        gc_stack_overflow = 0;
        gc_sp = gc_stack;

        // scan entire memory looking for blocks which have been marked but not their children
        for (machine_uint_t block = 0; block < gc_alloc_table_byte_len * BLOCKS_PER_ATB; block++) {
            // trace (again) if mark bit set
            if (ATB_GET_KIND(block) == AT_MARK) {
                *gc_sp++ = block;
                gc_drain_stack();
            }
        }
    }
}

155
STATIC void gc_sweep(void) {
Damien's avatar
Damien committed
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
    // free unmarked heads and their tails
    int free_tail = 0;
    for (machine_uint_t block = 0; block < gc_alloc_table_byte_len * BLOCKS_PER_ATB; block++) {
        switch (ATB_GET_KIND(block)) {
            case AT_HEAD:
                free_tail = 1;
                // fall through to free the head

            case AT_TAIL:
                if (free_tail) {
                    ATB_ANY_TO_FREE(block);
                }
                break;

            case AT_MARK:
                ATB_MARK_TO_HEAD(block);
                free_tail = 0;
                break;
        }
    }
}

178
void gc_collect_start(void) {
Damien's avatar
Damien committed
179
180
181
182
183
184
185
186
187
188
189
190
    gc_stack_overflow = 0;
    gc_sp = gc_stack;
}

void gc_collect_root(void **ptrs, machine_uint_t len) {
    for (machine_uint_t i = 0; i < len; i++) {
        machine_uint_t ptr = (machine_uint_t)ptrs[i];
        VERIFY_MARK_AND_PUSH(ptr);
        gc_drain_stack();
    }
}

191
void gc_collect_end(void) {
Damien's avatar
Damien committed
192
193
    gc_deal_with_stack_overflow();
    gc_sweep();
194
}
Damien's avatar
Damien committed
195

196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
void gc_info(gc_info_t *info) {
    info->total = (gc_pool_end - gc_pool_start) * sizeof(machine_uint_t);
    info->used = 0;
    info->free = 0;
    info->num_1block = 0;
    info->num_2block = 0;
    info->max_block = 0;
    for (machine_uint_t block = 0, len = 0; block < gc_alloc_table_byte_len * BLOCKS_PER_ATB; block++) {
        machine_uint_t kind = ATB_GET_KIND(block);
        if (kind == AT_FREE || kind == AT_HEAD) {
            if (len == 1) {
                info->num_1block += 1;
            } else if (len == 2) {
                info->num_2block += 1;
            }
            if (len > info->max_block) {
                info->max_block = len;
            }
        }
        switch (kind) {
Damien's avatar
Damien committed
216
            case AT_FREE:
217
218
                info->free += 1;
                len = 0;
Damien's avatar
Damien committed
219
220
221
                break;

            case AT_HEAD:
222
223
224
225
                info->used += 1;
                len = 1;
                break;

Damien's avatar
Damien committed
226
            case AT_TAIL:
227
228
                info->used += 1;
                len += 1;
Damien's avatar
Damien committed
229
230
231
                break;

            case AT_MARK:
232
                // shouldn't happen
Damien's avatar
Damien committed
233
234
235
236
                break;
        }
    }

237
238
    info->used *= BYTES_PER_BLOCK;
    info->free *= BYTES_PER_BLOCK;
Damien's avatar
Damien committed
239
240
241
242
}

void *gc_alloc(machine_uint_t n_bytes) {
    machine_uint_t n_blocks = ((n_bytes + BYTES_PER_BLOCK - 1) & (~(BYTES_PER_BLOCK - 1))) / BYTES_PER_BLOCK;
243
    DEBUG_printf("gc_alloc(" UINT_FMT " bytes -> " UINT_FMT " blocks)\n", n_bytes, n_blocks);
Damien's avatar
Damien committed
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269

    // check for 0 allocation
    if (n_blocks == 0) {
        return NULL;
    }

    machine_uint_t i;
    machine_uint_t end_block;
    machine_uint_t start_block;
    machine_uint_t n_free = 0;
    int collected = 0;
    for (;;) {

        // look for a run of n_blocks available blocks
        for (i = 0; i < gc_alloc_table_byte_len; i++) {
            byte a = gc_alloc_table_start[i];
            if (ATB_0_IS_FREE(a)) { if (++n_free >= n_blocks) { i = i * BLOCKS_PER_ATB + 0; goto found; } } else { n_free = 0; }
            if (ATB_1_IS_FREE(a)) { if (++n_free >= n_blocks) { i = i * BLOCKS_PER_ATB + 1; goto found; } } else { n_free = 0; }
            if (ATB_2_IS_FREE(a)) { if (++n_free >= n_blocks) { i = i * BLOCKS_PER_ATB + 2; goto found; } } else { n_free = 0; }
            if (ATB_3_IS_FREE(a)) { if (++n_free >= n_blocks) { i = i * BLOCKS_PER_ATB + 3; goto found; } } else { n_free = 0; }
        }

        // nothing found!
        if (collected) {
            return NULL;
        }
270
        DEBUG_printf("gc_alloc(" UINT_FMT "): no free mem, triggering GC\n", n_bytes);
Damien's avatar
Damien committed
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
        gc_collect();
        collected = 1;
    }

    // found, ending at block i inclusive
found:
    // get starting and end blocks, both inclusive
    end_block = i;
    start_block = i - n_free + 1;

    // mark first block as used head
    ATB_FREE_TO_HEAD(start_block);

    // mark rest of blocks as used tail
    // TODO for a run of many blocks can make this more efficient
    for (machine_uint_t bl = start_block + 1; bl <= end_block; bl++) {
        ATB_FREE_TO_TAIL(bl);
    }

    // return pointer to first block
    return (void*)(gc_pool_start + start_block * WORDS_PER_BLOCK);
}

294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
// force the freeing of a piece of memory
void gc_free(void *ptr_in) {
    machine_uint_t ptr = (machine_uint_t)ptr_in;

    if (VERIFY_PTR(ptr)) {
        machine_uint_t block = BLOCK_FROM_PTR(ptr);
        if (ATB_GET_KIND(block) == AT_HEAD) {
            // free head and all of its tail blocks
            do {
                ATB_ANY_TO_FREE(block);
                block += 1;
            } while (ATB_GET_KIND(block) == AT_TAIL);
        }
    }
}

Damien's avatar
Damien committed
310
311
312
machine_uint_t gc_nbytes(void *ptr_in) {
    machine_uint_t ptr = (machine_uint_t)ptr_in;

313
    if (VERIFY_PTR(ptr)) {
Damien's avatar
Damien committed
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
        machine_uint_t block = BLOCK_FROM_PTR(ptr);
        if (ATB_GET_KIND(block) == AT_HEAD) {
            // work out number of consecutive blocks in the chain starting with this on
            machine_uint_t n_blocks = 0;
            do {
                n_blocks += 1;
            } while (ATB_GET_KIND(block + n_blocks) == AT_TAIL);
            return n_blocks * BYTES_PER_BLOCK;
        }
    }

    // invalid pointer
    return 0;
}

329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
// use this realloc for now, one below is broken
void *gc_realloc(void *ptr, machine_uint_t n_bytes) {
    machine_uint_t n_existing = gc_nbytes(ptr);
    if (n_bytes <= n_existing) {
        return ptr;
    } else {
        // TODO check if we can grow inplace
        void *ptr2 = gc_alloc(n_bytes);
        if (ptr2 == NULL) {
            return ptr2;
        }
        memcpy(ptr2, ptr, n_existing);
        gc_free(ptr);
        return ptr2;
    }
}

#if 0
mux's avatar
mux committed
347
348
void *gc_realloc(void *ptr_in, machine_uint_t n_bytes) {
    void *ptr_out = NULL;
349
    machine_uint_t block = 0;
mux's avatar
mux committed
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
    machine_uint_t ptr = (machine_uint_t)ptr_in;

    if (ptr_in == NULL) {
        return gc_alloc(n_bytes);
    }

    if (VERIFY_PTR(ptr)                         /* verify pointer */
        && (block = BLOCK_FROM_PTR(ptr))        /* get first block */
        && ATB_GET_KIND(block) == AT_HEAD) {    /* make sure it's a HEAD block */

        byte block_type;
        machine_uint_t n_free   = 0;
        machine_uint_t n_blocks = 1; /* counting HEAD block */

        /* get the number of consecutive tail blocks and
           the number of free blocks after last tail block */
366
        // XXX make sure we stop if we get to end of heap
367
        do {
mux's avatar
mux committed
368
369
370
371
372
373
374
            block_type = ATB_GET_KIND(block + n_blocks + n_free);
            switch (block_type) {
                case AT_FREE: n_free++; break;
                case AT_TAIL: n_blocks++; break;
                default: break;
            }
        /* stop as soon as we find enough blocks for n_bytes */
375
        // XXX check for n_bytes is wrong since we don't include n_free
mux's avatar
mux committed
376
377
378
379
380
381
382
383
        } while (block_type != AT_HEAD && (n_bytes > (n_blocks * BYTES_PER_BLOCK)));

        /* number of allocated bytes */
        machine_uint_t n_existing = n_blocks * BYTES_PER_BLOCK;

        /* check if realloc'ing to a smaller size */
        if (n_bytes <= n_existing) {
            ptr_out = ptr_in;
384
385
386
            /* free unneeded tail blocks */
            for (machine_uint_t bl = block + n_blocks; ATB_GET_KIND(bl) == AT_TAIL; bl++) {
                ATB_ANY_TO_FREE(bl);
mux's avatar
mux committed
387
388
389
            }

        /* check if we can expand in place */
390
391
        // XXX disabled for now
        } else if (0 && n_bytes <= (n_existing + (n_free * BYTES_PER_BLOCK))) {
mux's avatar
mux committed
392
            /* number of blocks needed to expand +1 if there's a remainder */
393
            // XXX this has a bug, but don't know why; try: l=[i for i in range(1000)]; for i in l: print(i/3)
mux's avatar
mux committed
394
395
396
397
398
399
400
            machine_uint_t n_diff = ( n_bytes - n_existing)/BYTES_PER_BLOCK+
                                    ((n_bytes - n_existing)%BYTES_PER_BLOCK!=0);

            DEBUG_printf("gc_realloc: expanding " UINT_FMT " blocks (" UINT_FMT " bytes) to " UINT_FMT " blocks (" UINT_FMT " bytes)\n",
                    n_existing/BYTES_PER_BLOCK, n_existing, n_existing/BYTES_PER_BLOCK+n_diff, n_existing + n_diff*BYTES_PER_BLOCK);

            /* mark rest of blocks as used tail */
401
            for (machine_uint_t bl = block + n_blocks; bl < (block + n_blocks + n_diff); bl++) {
mux's avatar
mux committed
402
403
404
405
406
407
                ATB_FREE_TO_TAIL(bl);
            }
            ptr_out = ptr_in;

        /* try to find a new contiguous chain */
        } else if ((ptr_out = gc_alloc(n_bytes)) != NULL) {
408
            DEBUG_printf("gc_realloc: allocated new block \n");
mux's avatar
mux committed
409
410
            memcpy(ptr_out, ptr_in, n_existing);
            gc_free(ptr_in);
411
        }
Damien's avatar
Damien committed
412
    }
mux's avatar
mux committed
413
414

    return ptr_out;
Damien's avatar
Damien committed
415
}
416
#endif
Damien's avatar
Damien committed
417

418
419
420
421
422
423
424
425
void gc_dump_info() {
    gc_info_t info;
    gc_info(&info);
    printf("GC: total: " UINT_FMT ", used: " UINT_FMT ", free: " UINT_FMT "\n", info.total, info.used, info.free);
    printf(" No. of 1-blocks: " UINT_FMT ", 2-blocks: " UINT_FMT ", max blk sz: " UINT_FMT "\n",
           info.num_1block, info.num_2block, info.max_block);
}

426
427
void gc_dump_alloc_table(void) {
    printf("GC memory layout:");
428
    for (machine_uint_t bl = 0; bl < gc_alloc_table_byte_len * BLOCKS_PER_ATB; bl++) {
429
430
431
432
        if (bl % 64 == 0) {
            printf("\n%04x: ", (uint)bl);
        }
        int c = ' ';
433
        switch (ATB_GET_KIND(bl)) {
434
435
436
437
            case AT_FREE: c = '.'; break;
            case AT_HEAD: c = 'h'; break;
            case AT_TAIL: c = 't'; break;
            case AT_MARK: c = 'm'; break;
438
        }
439
        printf("%c", c);
440
    }
441
    printf("\n");
442
443
}

444
#if DEBUG_PRINT
445
446
void gc_test(void) {
    machine_uint_t len = 500;
Damien's avatar
Damien committed
447
448
449
450
    machine_uint_t *heap = malloc(len);
    gc_init(heap, heap + len / sizeof(machine_uint_t));
    void *ptrs[100];
    {
451
        machine_uint_t **p = gc_alloc(16);
Damien's avatar
Damien committed
452
453
454
455
        p[0] = gc_alloc(64);
        p[1] = gc_alloc(1);
        p[2] = gc_alloc(1);
        p[3] = gc_alloc(1);
456
        machine_uint_t ***p2 = gc_alloc(16);
Damien's avatar
Damien committed
457
458
459
460
        p2[0] = p;
        p2[1] = p;
        ptrs[0] = p2;
    }
461
    for (int i = 0; i < 25; i+=2) {
Damien's avatar
Damien committed
462
463
464
465
466
467
468
        machine_uint_t *p = gc_alloc(i);
        printf("p=%p\n", p);
        if (i & 3) {
            //ptrs[i] = p;
        }
    }

469
    printf("Before GC:\n");
470
    gc_dump_alloc_table();
471
472
473
474
475
    printf("Starting GC...\n");
    gc_collect_start();
    gc_collect_root(ptrs, sizeof(ptrs) / sizeof(void*));
    gc_collect_end();
    printf("After GC:\n");
476
    gc_dump_alloc_table();
Damien's avatar
Damien committed
477
}
478
#endif
479
480

#endif // MICROPY_ENABLE_GC