Add a buddy allocator

src/base/lalloc.c

@@ -0,0 +1,174 @@
+#include "lalloc.h"
+
+#define BLOCK_ALIGNMENT sizeof(LBlockHeader)
+#define BLOCK_SPLIT_MIN_SIZE BLOCK_ALIGNMENT
+#define NULL_BLOCK (0xffffffff)
+#define BLOCK_FREE ((U32)1)
+
+static inline void lblock_set_free(LBlockHeader *block)
+{
+    block->size_and_flags |= BLOCK_FREE;
+}
+
+static inline void lblock_set_used(LBlockHeader *block)
+{
+    block->size_and_flags &= ~BLOCK_FREE;
+}
+
+static inline void lblock_set_size(LBlockHeader *block, U32 size)
+{
+    U32 flags = block->size_and_flags & BLOCK_FREE;
+    Assert((size & BLOCK_FREE) == 0);
+    block->size_and_flags = size | flags;
+}
+
+U32 lblock_size(const LBlockHeader *block)
+{
+    return block->size_and_flags & ~BLOCK_FREE;
+}
+
+int lblock_is_free(const LBlockHeader *block)
+{
+    return (block->size_and_flags & BLOCK_FREE) != 0;
+}
+
+LBlockHeader *lblock_from_offset(const LAlloc *allocator, U32 offset)
+{
+    if (offset == NULL_BLOCK) {
+        return NULL;
+    }
+
+    return (LBlockHeader *)((UMM)allocator->base + offset);
+}
+
+static inline U32 lbh_offset(LAlloc *allocator, LBlockHeader *block)
+{
+    if (block == NULL) {
+        return NULL_BLOCK;
+    }
+
+    return (U32)((uintptr_t)block - (uintptr_t)allocator->base);
+}
+
+static void lmaybe_split_block(LAlloc *allocator, LBlockHeader *block, U32 size)
+{
+    // this function assumes size is already aligned
+
+    U32 bsize = lblock_size(block);
+    if (size > U32_Max - (sizeof(LBlockHeader) + BLOCK_SPLIT_MIN_SIZE)) {
+        return; // size is so large that size + header + min would overflow U32 anyway
+    }
+    if (bsize < size + sizeof(LBlockHeader) + BLOCK_SPLIT_MIN_SIZE) {
+        return; // no gain from splitting this block
+    }
+
+    U8 *base = (U8 *)block;
+    LBlockHeader *new_block = (LBlockHeader *)(base + size + sizeof(LBlockHeader) );
+
+    U32 new_size = bsize - size - sizeof(LBlockHeader);
+
+    lblock_set_size(block, size);
+    lblock_set_used(block);
+    lblock_set_size(new_block, new_size);
+    lblock_set_free(new_block);
+    new_block->next = block->next;
+    new_block->prev = lbh_offset(allocator, block);
+
+    LBlockHeader *next = lblock_from_offset(allocator, new_block->next);
+    if (next) {
+        next->prev = lbh_offset(allocator, new_block);
+    }
+
+    block->next = lbh_offset(allocator, new_block);
+}
+
+static void lmerge_with_next(LAlloc *allocator, LBlockHeader *block)
+{
+    LBlockHeader *next = lblock_from_offset(allocator, block->next);
+    if (next == NULL || !lblock_is_free(next)) {
+        return;
+    }
+
+    // block and next are both existing blocks within the same backing store, whose
+    // total size is asserted in lalloc_init to be < 0xfffffff0, so this can't overflow
+    // unless allocator state is already corrupted.
+    Assert(lblock_size(block) <= U32_Max - sizeof(LBlockHeader) - lblock_size(next));
+    U32 new_size = lblock_size(block) + sizeof(LBlockHeader) + lblock_size(next);
+
+    lblock_set_size(block, new_size);
+    block->next = next->next;
+    LBlockHeader *nextnext = lblock_from_offset(allocator, next->next);
+    if (nextnext) nextnext->prev = lbh_offset(allocator, block);
+    // note - caller have to adjust next_alloc on return
+}
+
+void lalloc_init(LAlloc *allocator, void *backing_store, U32 size)
+{
+    // backing store must be aligned, otherwise  all alignment guarantees are lies
+
+    Assert(IsAlignedTo((UMM)backing_store, BLOCK_ALIGNMENT));
+    Assert(size >= sizeof(LBlockHeader));
+    Assert(size < 0xfffffff0);
+
+    allocator->base = backing_store;
+    allocator->size = size;
+
+    LBlockHeader *first = backing_store;
+    // We need to keep size aligned to at least 2 as we reuse the LSB for the FREE flag
+    U32 usable = AlignDown(size, 2) - sizeof(LBlockHeader);
+    lblock_set_free(first);
+    lblock_set_size(first, usable);
+    first->next = NULL_BLOCK;
+    first->prev = NULL_BLOCK;
+
+    allocator->next_alloc = first;
+}
+
+void *lalloc(LAlloc *allocator, U32 size)
+{
+    if (size > U32_Max - (BLOCK_ALIGNMENT - 1)) { // -1 so AlignUp below doesn't overflow
+        return NULL;
+    }
+    size = AlignUp(size, BLOCK_ALIGNMENT);
+
+    LBlockHeader *start = allocator->next_alloc ? allocator->next_alloc : (LBlockHeader *)allocator->base;
+    LBlockHeader *curr = start;
+
+    do {
+        if (lblock_is_free(curr) && lblock_size(curr) >= size) {
+            lmaybe_split_block(allocator, curr, size);
+            lblock_set_used(curr);
+            LBlockHeader *next = lblock_from_offset(allocator, curr->next);
+            allocator->next_alloc = next ? next : (LBlockHeader *)allocator->base;
+            return (U8 *)curr + sizeof(LBlockHeader);
+        }
+        LBlockHeader *next = lblock_from_offset(allocator, curr->next);
+        curr = next ? next : (LBlockHeader *)allocator->base;
+    } while (curr != start);
+
+    return NULL;
+}
+
+void lfree(LAlloc *allocator, void *ptr)
+{
+    if (ptr == NULL) {
+        return;
+    }
+
+    LBlockHeader *block = (LBlockHeader *)((U8 *)ptr - sizeof(LBlockHeader));
+
+    Assert((U8 *)block >= allocator->base && (U8 *)ptr < allocator->base + allocator->size);
+    // double free detection
+    Assert(!lblock_is_free(block));
+
+    lblock_set_free(block);
+    lmerge_with_next(allocator, block);
+
+    LBlockHeader *prev = lblock_from_offset(allocator, block->prev);
+    if (prev != NULL && lblock_is_free(prev)) {
+        lmerge_with_next(allocator, prev);
+        allocator->next_alloc = prev;
+    } else {
+        allocator->next_alloc = block;
+    }
+}

src/base/lalloc.h

@@ -0,0 +1,41 @@
+#pragma once
+#include "core.h"
+
+// Internal header added in front of each allocation
+typedef struct LBlockHeader LBlockHeader;
+struct LBlockHeader {
+    U32 next;            // offset from base
+    U32 prev;            // offset from base
+    U32 size_and_flags;  // size | free_bit
+    // Padding to ensure we always return 16 byte aligned pointers.
+    uint8_t pad[4];
+};
+_Static_assert(sizeof(LBlockHeader) == 16, "sizeof(BlockHeader) is not 16");
+
+/**
+ * General purpose allocator from a pre-defined backing store
+ * - Allocated pointers are aligned to 16 bytes.
+ * - Backing buffer must be aligned to 16 bytes
+ * - 4Gb - 16 is max for the allocator backing buffer.
+ * - Allocation of > 0 bytes returns a size aligned up to 16 bytes
+ * - Allocation of 0 bytes returns a valid pointer with 0 usable space
+ * The allocator maintains a doubly linked list of blocks,
+ * when blocks are free'd, adjacent blocks are merged to keep.
+ * This is also known as a buddy allocator
+ * fragmentation to a minimum
+ */
+typedef struct LAlloc LAlloc;
+struct LAlloc {
+    uint8_t *base;              // backing store
+    U32 size;                   // total size
+    LBlockHeader *next_alloc;   // start of free block search
+};
+
+void lalloc_init(LAlloc *allocator, void *backing_store, U32 size);
+
+void *lalloc(LAlloc *allocator, U32 size);
+void lfree(LAlloc *allocator, void *ptr);
+
+LBlockHeader *lblock_from_offset(const LAlloc *allocator, U32 offset);
+U32 lblock_size(const LBlockHeader *block);
+int lblock_is_free(const LBlockHeader *block);

src/base/pool.h

@@ -32,7 +32,7 @@ struct APoolCreateOptions {
         apool_create_ex(arena, &(APoolCreateOptions){sizeof(type), alignof(type), n_reserved, 0})
 APool *apool_create_ex(Allocator *allocator, const APoolCreateOptions *options);
 void *apool_alloc(APool *pool);
-INLINE void apool_allocz(APool *pool)
+INLINE void *apool_allocz(APool *pool)
 {
     void *ptr = apool_alloc(pool);
     if (ptr != NULL) {

src/main.c

@@ -4,6 +4,7 @@
 #include "base/platform.h"
 #include "base/pool.h"
 #include "base/string.h"
+#include "base/lalloc.h"
 #include <stdio.h>
 #include <stdlib.h>
 
@@ -689,6 +690,49 @@ void string_test_dlist(void)
 
 }
 
+void debug_lalloc(const LAlloc *allocator)
+{
+    puts("--debug start--");
+    for (LBlockHeader *b = (LBlockHeader *)allocator->base; b; b = lblock_from_offset(allocator, b->next)) {
+        printf("Block size %u free %d prev %u next %u \n", lblock_size(b), lblock_is_free(b), b->prev, b->next);
+    }
+    puts("--debug end--");
+}
+
+static void lalloc_test_basic(void)
+{
+    printf("bh size %zu\n", sizeof(LBlockHeader));
+#define SZ 64
+    LAlloc a;
+    char *ptrs[SZ] = {};
+    alignas(16) char buff[1023];
+    lalloc_init(&a, buff, sizeof(buff));
+    debug_lalloc(&a);
+    ptrs[0] = lalloc(&a, 1023 - 16 * 2 - 144);
+    ptrs[1] = lalloc(&a, 80);
+    ptrs[2] = lalloc(&a, 32);
+    debug_lalloc(&a);
+    lfree(&a, ptrs[0]);
+    ptrs[3] = lalloc(&a, 12);
+    ptrs[4] = lalloc(&a, 12);
+    debug_lalloc(&a);
+
+    lfree(&a, ptrs[1]);
+    lfree(&a, ptrs[2]);
+    lfree(&a, ptrs[3]);
+    lfree(&a, ptrs[4]);
+
+    for (int i = 0; i < SZ; i++) {
+        ptrs[i] = lalloc(&a, rand() % 256 + 1);
+    }
+
+    for (int i = 0; i < SZ; i++) {
+        // if (rand() % 2 == 0) {
+        lfree(&a, ptrs[i]);
+        // }
+    }
+    debug_lalloc(&a);
+}
 int main(int argc, char *argv[])
 {
     arena_test_1();
@@ -709,6 +753,7 @@ int main(int argc, char *argv[])
     string_test_hex();
     string_test_slist();
     string_test_dlist();
+    lalloc_test_basic();
 
 }