Fix larena aligned alloc

Makefile

@@ -1,39 +1,38 @@
 CC           =  gcc
 LD           =  $(CC)
-CFLAGS       =  -Wall -Wextra -Wno-unused-parameter -ggdb
+CFLAGS       =  -std=gnu17 -Wall -Wextra -Wno-unused-parameter -ggdb
 CPPFLAGS     += -Iinclude
 LDFLAGS      +=
 ifndef DEBUG
 CFLAGS       += -O2
 CPPFLAGS     += -DNDEBUG
 else
-CPPFLAGS     += -DDEBUG
+CPPFLAGS     += -DDEBUG -fsanitize=address,undefined
+LDFLAGS      += -fsanitize=address,undefined
 endif
 BUILDDIR	  =  build
 
 SOURCES = $(wildcard src/*.c)
-OBJECTS = $(addprefix $(BUILDDIR)/,$(SOURCES:.c=.o))
-DEPS	= $(addprefix $(BUILDDIR)/,$(SOURCES:.c=.d))
-LIB     = liblilalloc
 TEST_SOURCES = $(wildcard test/*.c)
+OBJECTS = $(addprefix $(BUILDDIR)/,$(SOURCES:.c=.o))
+DEPS	= $(addprefix $(BUILDDIR)/,$(SOURCES:.c=.d)) $(addprefix $(BUILDDIR)/,$(TEST_SOURCES:.c=.d))
+LIB     = liblilalloc.a
 TEST_BINS = $(patsubst test/%.c,$(BUILDDIR)/test/%,$(TEST_SOURCES))
 
 .PHONY: all
-all:  $(BUILDDIR)/$(LIB).a
+all:  $(BUILDDIR)/$(LIB)
 
 
-$(BUILDDIR)/%.o: %.c
+$(BUILDDIR)/%.o: %.c Makefile
 	mkdir -p $(@D)
 	$(CC) $(CPPFLAGS) $(CFLAGS) -MMD -MP -c -o $@ $<
 
-$(BUILDDIR)/$(LIB).a: $(OBJECTS)
+$(BUILDDIR)/$(LIB): $(OBJECTS)
 	ar crs $@ $^
 
-
-$(BUILDDIR)/test/%: test/%.c $(BUILDDIR)/$(LIB).a
+$(BUILDDIR)/test/%: test/%.c $(BUILDDIR)/$(LIB)
 	mkdir -p $(@D)
-	$(CC) $(CPPFLAGS) $(CFLAGS) -MMD -MP -o $@ $< $(BUILDDIR)/$(LIB).a
-
+	$(CC) $(CPPFLAGS) $(CFLAGS) $(LDFLAGS) -MMD -MP -o $@ $< $(BUILDDIR)/$(LIB)
 
 .PHONY: test
 test: $(TEST_BINS)
@@ -44,8 +43,6 @@ test: $(TEST_BINS)
 		echo ; \
 	done
 
-
-
 .PHONY: clean
 clean:
 	rm -rf $(BUILDDIR)

include/lalloc.h

@@ -34,7 +34,6 @@ typedef struct LAlloc LAlloc;
 struct LAlloc {
     uint8_t *base;             // backing store
     uint32_t size;             // total size
-    LBlockHeader *heap_start;  // first block
     LBlockHeader *next_alloc;  // start of free block search
 };
 

include/larena.h

@@ -94,9 +94,14 @@ struct LArenaTemp {
  * struct foo *foo = LARENA_ALLOC_TYPE(&my_allocator, struct foo);
  * // Array types can be used too:
  * struct Foo *str = LARENA_ALLOC_TYPE(&my_allocator, struct Foo[32]);
+ * // For arrays with runtime determined size, use:
+ * LARENA_ALLOC_ARRAY(&my_allocator, struct Foo, cnt_foos);
  * @endcode
  */
 #define LARENA_ALLOC_TYPE(allocator, type_) larena_alloc_aligned((allocator), sizeof(type_), _Alignof(type_))
+
+#define LARENA_ALLOC_ARRAY(allocator, type_, nr_elements_) larena_alloc_aligned((allocator), (nr_elements_) * sizeof(type_), _Alignof(type_))
+
 // Arena size in bytes
 static inline size_t larena_size(const LArena *arena)
 {
@@ -152,16 +157,17 @@ static inline void *larena_alloc(LArena *arena, size_t sz)
 
 // Allocate @sz bytes from the LArena, returned pointer is aligned to @align
 // Returns NULL if not enough space
-static inline void *larena_alloc_aligned(LArena *arena, size_t sz, unsigned int align) [[gnu::alloc_align(3)]]
+[[gnu::alloc_align(3)]]
+static inline void *larena_alloc_aligned(LArena *arena, size_t sz, unsigned int align)
 {
     assert(align != 0);
     assert((align & (align - 1)) == 0);  // power of 2
 
-    uintptr_t curr = (uintptr_t)arena->curr;
+    uintptr_t curr = (uintptr_t)arena->;
     uintptr_t aligned = (curr + (align - (uintptr_t)1)) & ~(align - (uintptr_t)1);
     uint8_t *start = (uint8_t *)aligned;
 
-    if (sz > (larena_available(arena))) {
+    if (sz > (size_t)(arena->end - start)) {
         return NULL;
     }
 

include/lpool.h

@@ -9,7 +9,7 @@ extern "C" {
 
 typedef union LPoolBlock LPoolBlock;
 union LPoolBlock {
-    LPoolBlock *next;       // next free block (when not allocated)
+    uint32_t next;          // offset from start to next free block (when not allocated)
     unsigned char data[0];  // data when allocated
 };
 
@@ -30,10 +30,10 @@ union LPoolBlock {
 typedef struct LPool LPool;
 struct LPool {
     LPoolBlock *free_list;  // free list
-    LPoolBlock *start;      // start of user buffer
-    LPoolBlock *end;        // one past end of user buffer
+    uint8_t *start;         // start of user buffer
     uint32_t block_size;    // aligned size of each block
-    uint32_t capacity;      // number of blocks
+    uint32_t capacity;      // total number of blocks
+    uint32_t used;          // used number of blocks
     uint32_t align;         // alignment
 };
 

src/lalloc.c

@@ -70,7 +70,6 @@ static void lmaybe_split_block(LAlloc *allocator, LBlockHeader *block, uint32_t
 
     lblock_set_size(block, size);
     lblock_set_used(block);
-
     new_block->size_and_flags = new_size | BLOCK_FREE;
     new_block->next = block->next;
     new_block->prev = lbh_offset(allocator, block);
@@ -116,7 +115,6 @@ void lalloc_init(LAlloc *allocator, void *backing_store, uint32_t size)
     first->next = NULL_BLOCK;
     first->prev = NULL_BLOCK;
 
-    allocator->heap_start = first;
     allocator->next_alloc = first;
 }
 
@@ -127,7 +125,7 @@ void *lalloc(LAlloc *allocator, uint32_t size)
     }
     size = ALIGN_UP(size);
 
-    LBlockHeader *start = allocator->next_alloc ? allocator->next_alloc : allocator->heap_start;
+    LBlockHeader *start = allocator->next_alloc ? allocator->next_alloc : (LBlockHeader *)allocator->base;
     LBlockHeader *curr = start;
 
     do {
@@ -135,11 +133,11 @@ void *lalloc(LAlloc *allocator, uint32_t size)
             lmaybe_split_block(allocator, curr, size);
             lblock_set_used(curr);
             LBlockHeader *next = lblock_from_offset(allocator, curr->next);
-            allocator->next_alloc = next ? next : allocator->heap_start;
+            allocator->next_alloc = next ? next : (LBlockHeader *)allocator->base;
             return (uint8_t *)curr + sizeof(LBlockHeader);
         }
         LBlockHeader *next = lblock_from_offset(allocator, curr->next);
-        curr = next ? next : allocator->heap_start;
+        curr = next ? next : (LBlockHeader *)allocator->base;
     } while (curr != start);
 
     return NULL;
@@ -153,7 +151,7 @@ void lfree(LAlloc *allocator, void *ptr)
 
     LBlockHeader *block = (LBlockHeader *)((uint8_t *)ptr - sizeof(LBlockHeader));
 
-    assert((uint8_t *)ptr - sizeof(LBlockHeader) >= allocator->base && (uint8_t *)ptr < allocator->base + allocator->size);
+    assert((uint8_t *)block >= allocator->base && (uint8_t *)ptr < allocator->base + allocator->size);
     // double free detection
     assert(!lblock_is_free(block));
 

src/lpool.c

@@ -3,13 +3,30 @@
 #include "lpool.h"
 
 #define ALIGN_UP(n, alignment) (((n) + (__typeof__(n))(alignment) - 1) & ~((__typeof__(n))(alignment) - 1))
+#define NULL_BLOCK (0xffffffff)
+
+static inline LPoolBlock *lpool_next(LPool *pool, LPoolBlock *block)
+{
+    if (block->next == NULL_BLOCK) {
+        return NULL;
+    }
+    return (LPoolBlock *)&pool->start[block->next];
+}
+
+static inline uint32_t lpool_offset(LPool *pool, LPoolBlock *block)
+{
+    if (block == NULL) {
+        return NULL_BLOCK;
+    }
+    return (uint8_t *)block - pool->start;
+}
 
 void lpool_init(LPool *pool, void *buffer, size_t buffer_size, uint32_t block_size, uint32_t align)
 {
     assert(align > 0);
     assert((align & (align - 1)) == 0);  // power of 2
 
-    // enforce minimum for storing free list pointers
+    // enforce minimum for storing free list offsets
 
     if (align < _Alignof(LPoolBlock)) {
         align = _Alignof(LPoolBlock);
@@ -19,91 +36,96 @@ void lpool_init(LPool *pool, void *buffer, size_t buffer_size, uint32_t block_si
         block_size = sizeof(LPoolBlock);
     }
     // ensure initial alignment
-    uint32_t stride = ALIGN_UP(block_size, align);
+    block_size = ALIGN_UP(block_size, align);
 
     // align first usable block
-    uintptr_t start_addr = (uintptr_t)buffer;
-    uintptr_t first_addr = ALIGN_UP(start_addr, align);
-    size_t padding = first_addr - start_addr;
+    uintptr_t buffer_addr = (uintptr_t)buffer;
+    uintptr_t first_addr = ALIGN_UP(buffer_addr, align);
+    size_t padding = first_addr - buffer_addr;
 
-    if (padding + stride > buffer_size) {
+    if (padding + block_size > buffer_size) {
         pool->capacity = 0;
         pool->free_list = NULL;
     } else {
         size_t usable = buffer_size - padding;
-        pool->capacity = usable / stride;
+        pool->capacity = usable / block_size;
 
         // build intrusive free list
         LPoolBlock *block = (LPoolBlock *)first_addr;
         pool->free_list = block;
-        for (uint32_t i = 0; i < pool->capacity - 1; i++) {
-            LPoolBlock *next = __builtin_assume_aligned(&block->data[0] + stride, _Alignof(LPoolBlock));
+        for (uint32_t i = 1; i < pool->capacity; i++) {
+            uint32_t next = i * block_size;
             block->next = next;
-            block = next;
+            block = (LPoolBlock *)(first_addr + next);
         }
-        block->next = NULL;  // last element
+        block->next = NULL_BLOCK;  // last element
     }
-    pool->block_size = stride;
+    pool->block_size = block_size;
     pool->align = align;
     pool->start = buffer;
-    pool->end = buffer + buffer_size;
+    pool->used = 0;
 }
 
 void *lpool_alloc(LPool *pool)
 {
-    if (!pool->free_list) {
+    if (pool->free_list == NULL) {
         return NULL;
     }
     LPoolBlock *block = pool->free_list;
-    pool->free_list = block->next;
+    LPoolBlock *next = lpool_next(pool, block);
+    pool->free_list = next;
 
 #ifdef DEBUG
     memset(block, 0xCB, pool->block_size);
 #endif
+    pool->used++;
     return block->data;
 }
 
-void lpool_free(LPool *pool, void *p)
+void lpool_free(LPool *pool, void *ptr)
 {
-    if (p == NULL) {
+    if (ptr == NULL) {
         return;
     }
-    LPoolBlock *block = p;
 #ifdef DEBUG
-    assert(block >= pool->start && block < pool->end);
-    assert(((uintptr_t)p - (uintptr_t)pool->start) % pool->block_size == 0);
+    assert(pool->used > 0);
+    uint8_t *end = pool->start + (pool->capacity * pool->block_size);
+    assert((uint8_t *)ptr >= pool->start && (uint8_t *)ptr < end);
+    assert((((uintptr_t)ptr - (uintptr_t)pool->start) & (pool->block_size - 1)) == 0);
 
     // double-free detection
-    for (LPoolBlock *f = pool->free_list; f; f = f->next) {
-        if (f == p) {
+    for (LPoolBlock *block = pool->free_list; block != NULL; block = lpool_next(pool, block)) {
+        if (block == ptr) {
             assert(0 && "Double free detected in pool!");
             return;
         }
     }
-
-    memset(p, 0xDD, pool->block_size);
+    memset(ptr, 0xDD, pool->block_size);
 #endif
+    LPoolBlock *block = ptr;
     // When not in use, store pointer to next free element inline in the block
-    block->next = pool->free_list;
-    pool->free_list = p;
+    block->next = lpool_offset(pool, pool->free_list);
+    pool->free_list = block;
+    pool->used--;
 }
 
 void lpool_reset(LPool *pool)
 {
-    if (pool->capacity == 0) return;
+    if (pool->capacity == 0) {
+        return;
+    }
 
-    uintptr_t start_addr = (uintptr_t)pool->start;
-    uintptr_t first_addr = ALIGN_UP(start_addr, pool->align);
+    uintptr_t first_addr = (uintptr_t)pool->start;
 
     // build intrusive free list
-    LPoolBlock *p = (LPoolBlock *)first_addr;
-    for (size_t i = 0; i < pool->capacity - 1; ++i) {
-        LPoolBlock *next = __builtin_assume_aligned(&p->data[0] + pool->block_size, _Alignof(LPoolBlock));
-        p->next = next;
-        p = next;
+    LPoolBlock *block = (LPoolBlock *)first_addr;
+    pool->free_list = block;
+    for (size_t i = 1; i < pool->capacity; ++i) {
+        uint32_t next = i * pool->block_size;
+        block->next = next;
+        block = (LPoolBlock *)(first_addr + next);
     }
 
-    p->next = NULL;  // last element
-
-    pool->free_list = (LPoolBlock *)first_addr;
+    block->next = NULL_BLOCK;  // last element
+    pool->used = 0;
 }

test/lalloc_test.c

@@ -7,7 +7,7 @@
 void debug_alloc(const LAlloc *allocator)
 {
     puts("--debug start--");
-    for (LBlockHeader *b = allocator->heap_start; b; b = lblock_from_offset(allocator, b->next)) {
+    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--");

test/lpool_test.c

@@ -1,4 +1,6 @@
 #include <stdio.h>
+#include <stdlib.h>
+
 #include "lpool.h"
 
 struct TestData {
@@ -11,8 +13,10 @@ struct TestData {
 int main(int argc, char *argv[])
 {
     _Alignas(16) uint8_t store[2048];
+    // uint8_t *store = malloc(1024 * 1024 - 17);
+
     LPool pool;
-    lpool_init(&pool, store, sizeof store, sizeof(struct TestData), _Alignof(struct TestData));
+    lpool_init(&pool, store, 2048, sizeof(struct TestData), _Alignof(struct TestData));
     printf("align=%u capacity=%u block_size=%u\n", pool.align, pool.capacity, pool.block_size);
     struct TestData *data = lpool_alloc(&pool);
     size_t allocs = 1;
@@ -23,10 +27,20 @@ int main(int argc, char *argv[])
         }
         allocs++;
     }
-    printf("Total allocs = %zu\n", allocs);
-    printf("alloc succeded = %d\n", lpool_alloc(&pool) != NULL);
+    printf("Total allocs = %zu, used = %u\n", allocs, pool.used);
+    printf("alloc succeded = %d used  = %u\n", lpool_alloc(&pool) != NULL, pool.used);
     lpool_free(&pool, data);
-    printf("alloc succeded = %d\n", lpool_alloc(&pool) != NULL);
+    printf("alloc succeded = %d used  = %u\n", lpool_alloc(&pool) != NULL, pool.used);
 
+    lpool_reset(&pool);
+    allocs = 0;
+    for (;;) {
+        struct TestData *t = lpool_alloc(&pool);
+        if (t == NULL) {
+            break;
+        }
+        allocs++;
+    }
+    printf("Total allocs = %zu, used = %u\n", allocs, pool.used);
     return 0;
 }