diff --git a/src/matrix.h b/src/matrix.h index 7ec7df6a..c5cf5cd8 100644 --- a/src/matrix.h +++ b/src/matrix.h @@ -45,6 +45,26 @@ \ } \ \ + static name##_t *name##_new_aligned(size_t m, size_t n, size_t alignment) { \ + name##_t *matrix = malloc(sizeof(name##_t)); \ + \ + if (matrix == NULL) { \ + return NULL; \ + } \ + \ + matrix->m = m; \ + matrix->n = n; \ + \ + matrix->values = _aligned_malloc(sizeof(type) * m * n, alignment); \ + if (matrix->values == NULL) { \ + free(matrix); \ + return NULL; \ + } \ + \ + return matrix; \ + \ + } \ + \ static void name##_destroy(name##_t *self) { \ if (self == NULL) return; \ \ @@ -55,6 +75,16 @@ free(self); \ } \ \ + static void name##_destroy_aligned(name##_t *self) { \ + if (self == NULL) return; \ + \ + if (self->values != NULL) { \ + _aligned_free(self->values); \ + } \ + \ + free(self); \ + } \ + \ static inline void name##_zero(name##_t *self) { \ memset(self->values, 0, self->m * self->n * sizeof(type)); \ } \ @@ -64,10 +94,30 @@ if (self == NULL) return false; \ \ if (m * n > (self->m * self->n)) { \ - self->values = realloc(self->values, sizeof(type) * m * n); \ - if (self->values == NULL) { \ + type *ptr = realloc(self->values, sizeof(type) * m * n); \ + if (ptr == NULL) { \ return false; \ } \ + self->values = ptr; \ + } \ + \ + self->m = m; \ + self->n = n; \ + \ + name##_zero(self); \ + \ + return true; \ + } \ + \ + static inline bool name##_resize_aligned(name##_t *self, size_t m, size_t n, size_t alignment) { \ + if (self == NULL) return false; \ + \ + if (m * n > (self->m * self->n)) { \ + type *ptr = _aligned_realloc(self->values, sizeof(type) * m * n, alignment); \ + if (ptr == NULL) { \ + return false; \ + } \ + self->values = ptr; \ } \ \ self->m = m; \ @@ -90,7 +140,7 @@ if (self->m != other->m || self->n != other->n) { \ return false; \ } \ - size_t num_values = self->n * self->n; \ + size_t num_values = self->m * self->n; \ \ memcpy(other->values, self->values, num_values * sizeof(type)); \ return true; \ diff --git a/src/vector.h b/src/vector.h index ad3a75fa..c72644a1 100644 --- a/src/vector.h +++ b/src/vector.h @@ -13,7 +13,35 @@ static inline void *_aligned_malloc(size_t size, size_t alignment) { void *p; int ret = posix_memalign(&p, alignment, size); - return (ret == 0) ? p : 0; + return (ret == 0) ? p : NULL; +} +static inline void *_aligned_realloc(void *p, size_t size, size_t alignment) +{ + if ((alignment == 0) || ((alignment & (alignment - 1)) != 0) || (alignment < sizeof(void *))) { + return NULL; + } + + if (size == 0) { + return NULL; + } + + void *rp = realloc(p, size); + + /* If realloc result is not already at an aligned boundary, + _aligned_malloc a new block and copy the contents of the realloc'd + pointer to the aligned block, free the realloc'd pointer and return + the aligned pointer. + */ + if ( ((size_t)rp & (alignment - 1)) != 0) { + void *p1 = _aligned_malloc(size, alignment); + if (p1 != NULL) { + memcpy(p1, rp, size); + } + free(rp); + rp = p1; + } + + return rp; } static inline void _aligned_free(void *p) { @@ -28,76 +56,84 @@ static inline void _aligned_free(void *p) #endif // Based kvec.h, dynamic vectors of any type -#define __VECTOR_BASE(name, type) typedef struct { size_t n, m; type *a; } name; \ - static inline name *name##_new_size(size_t size) { \ - name *array = malloc(sizeof(name)); \ - if (array == NULL) return NULL; \ - array->n = array->m = 0; \ - array->a = malloc((size > 0 ? size : 1) * sizeof(type)); \ - if (array->a == NULL) return NULL; \ - array->m = size; \ - return array; \ - } \ - static inline name *name##_new(void) { \ - return name##_new_size(DEFAULT_VECTOR_SIZE); \ - } \ - static inline name *name##_new_aligned(size_t size, size_t alignment) { \ - name *array = malloc(sizeof(name)); \ - if (array == NULL) return NULL; \ - array->n = array->m = 0; \ - array->a = _aligned_malloc(size * sizeof(type), alignment); \ - if (array->a == NULL) return NULL; \ - array->m = size; \ - return array; \ - } \ - static inline bool name##_resize(name *array, size_t size) { \ - if (size <= array->m) return true; \ - type *ptr = realloc(array->a, sizeof(type) * size); \ - if (ptr == NULL) return false; \ - array->a = ptr; \ - array->m = size; \ - return true; \ - } \ - static inline void name##_push(name *array, type value) { \ - if (array->n == array->m) { \ - size_t size = array->m ? array->m << 1 : 2; \ - type *ptr = realloc(array->a, sizeof(type) * size); \ - if (ptr == NULL) { \ - fprintf(stderr, "realloc failed during " #name "_push\n"); \ - exit(EXIT_FAILURE); \ - } \ - array->a = ptr; \ - array->m = size; \ - } \ - array->a[array->n++] = value; \ - } \ - static inline bool name##_extend(name *array, name *other) { \ - bool ret = false; \ - size_t new_size = array->n + other->n; \ - if (new_size > array->m) ret = name##_resize(array, new_size); \ - if (!ret) return false; \ - memcpy(array->a + array->n, other->a, other->n * sizeof(type)); \ - array->n = new_size; \ - return ret; \ - } \ - static inline void name##_pop(name *array) { \ - if (array->n > 0) array->n--; \ - } \ - static inline void name##_clear(name *array) { \ - array->n = 0; \ - } \ - static inline bool name##_copy(name *dst, name *src, size_t n) { \ - bool ret = true; \ - if (dst->m < n) ret = name##_resize(dst, n); \ - if (!ret) return false; \ - memcpy(dst->a, src->a, n * sizeof(type)); \ - dst->n = n; \ - return ret; \ - } \ - static inline name *name##_new_copy(name *vector, size_t n) { \ - name *cpy = name##_new_size(n); \ - if (!name##_copy(cpy, vector, n)) return NULL; \ - return cpy; \ +#define __VECTOR_BASE(name, type) typedef struct { size_t n, m; type *a; } name; \ + static inline name *name##_new_size(size_t size) { \ + name *array = malloc(sizeof(name)); \ + if (array == NULL) return NULL; \ + array->n = array->m = 0; \ + array->a = malloc((size > 0 ? size : 1) * sizeof(type)); \ + if (array->a == NULL) return NULL; \ + array->m = size; \ + return array; \ + } \ + static inline name *name##_new(void) { \ + return name##_new_size(DEFAULT_VECTOR_SIZE); \ + } \ + static inline name *name##_new_aligned(size_t size, size_t alignment) { \ + name *array = malloc(sizeof(name)); \ + if (array == NULL) return NULL; \ + array->n = array->m = 0; \ + array->a = _aligned_malloc(size * sizeof(type), alignment); \ + if (array->a == NULL) return NULL; \ + array->m = size; \ + return array; \ + } \ + static inline bool name##_resize(name *array, size_t size) { \ + if (size <= array->m) return true; \ + type *ptr = realloc(array->a, sizeof(type) * size); \ + if (ptr == NULL) return false; \ + array->a = ptr; \ + array->m = size; \ + return true; \ + } \ + static inline bool name##_resize_aligned(name *array, size_t size, size_t alignment) { \ + if (size <= array->m) return true; \ + type *ptr = _aligned_realloc(array->a, sizeof(type) * size, alignment); \ + if (ptr == NULL) return false; \ + array->a = ptr; \ + array->m = size; \ + return true; \ + } \ + static inline void name##_push(name *array, type value) { \ + if (array->n == array->m) { \ + size_t size = array->m ? array->m << 1 : 2; \ + type *ptr = realloc(array->a, sizeof(type) * size); \ + if (ptr == NULL) { \ + fprintf(stderr, "realloc failed during " #name "_push\n"); \ + exit(EXIT_FAILURE); \ + } \ + array->a = ptr; \ + array->m = size; \ + } \ + array->a[array->n++] = value; \ + } \ + static inline bool name##_extend(name *array, name *other) { \ + bool ret = false; \ + size_t new_size = array->n + other->n; \ + if (new_size > array->m) ret = name##_resize(array, new_size); \ + if (!ret) return false; \ + memcpy(array->a + array->n, other->a, other->n * sizeof(type)); \ + array->n = new_size; \ + return ret; \ + } \ + static inline void name##_pop(name *array) { \ + if (array->n > 0) array->n--; \ + } \ + static inline void name##_clear(name *array) { \ + array->n = 0; \ + } \ + static inline bool name##_copy(name *dst, name *src, size_t n) { \ + bool ret = true; \ + if (dst->m < n) ret = name##_resize(dst, n); \ + if (!ret) return false; \ + memcpy(dst->a, src->a, n * sizeof(type)); \ + dst->n = n; \ + return ret; \ + } \ + static inline name *name##_new_copy(name *vector, size_t n) { \ + name *cpy = name##_new_size(n); \ + if (!name##_copy(cpy, vector, n)) return NULL; \ + return cpy; \ } #define __VECTOR_DESTROY(name, type) \ @@ -105,9 +141,13 @@ static inline void _aligned_free(void *p) if (array == NULL) return; \ if (array->a != NULL) free(array->a); \ free(array); \ + } \ + static inline void name##_destroy_aligned(name *array) { \ + if (array == NULL) return; \ + if (array->a != NULL) _aligned_free(array->a); \ + free(array); \ } - #define __VECTOR_DESTROY_FREE_DATA(name, type, free_func) \ static inline void name##_destroy(name *array) { \ if (array == NULL) return; \ @@ -118,6 +158,16 @@ static inline void _aligned_free(void *p) } \ free(array->a); \ free(array); \ + } \ + static inline void name##_destroy_aligned(name *array) { \ + if (array == NULL) return; \ + if (array->a != NULL) { \ + for (size_t i = 0; i < array->n; i++) { \ + free_func(array->a[i]); \ + } \ + } \ + _aligned_free(array->a); \ + free(array); \ } #define VECTOR_INIT(name, type) \