simplified aligned allocation; improved codegen; fix mi_good_size with padding included; add MI_MAX_ALIGN_GUARANTEE

include/mimalloc/internal.h

@@ -196,7 +196,9 @@ mi_msecs_t  _mi_clock_end(mi_msecs_t start);
 mi_msecs_t  _mi_clock_start(void);
 
 // "alloc.c"
-void*       _mi_page_malloc(mi_heap_t* heap, mi_page_t* page, size_t size, bool zero) mi_attr_noexcept;  // called from `_mi_malloc_generic`
+void*       _mi_page_malloc_zero(mi_heap_t* heap, mi_page_t* page, size_t size, bool zero) mi_attr_noexcept;  // called from `_mi_malloc_generic`
+void*       _mi_page_malloc(mi_heap_t* heap, mi_page_t* page, size_t size) mi_attr_noexcept;                  // called from `_mi_heap_malloc_aligned`
+void*       _mi_page_malloc_zeroed(mi_heap_t* heap, mi_page_t* page, size_t size) mi_attr_noexcept;           // called from `_mi_heap_malloc_aligned`
 void*       _mi_heap_malloc_zero(mi_heap_t* heap, size_t size, bool zero) mi_attr_noexcept;
 void*       _mi_heap_malloc_zero_ex(mi_heap_t* heap, size_t size, bool zero, size_t huge_alignment) mi_attr_noexcept;     // called from `_mi_heap_malloc_aligned`
 void*       _mi_heap_realloc_zero(mi_heap_t* heap, void* p, size_t newsize, bool zero) mi_attr_noexcept;

include/mimalloc/types.h

@@ -200,6 +200,9 @@ typedef int32_t  mi_ssize_t;
 #error "mimalloc internal: define more bins"
 #endif
 
+// Maximum block size for which blocks are guarenteed to be block size aligned. (see `segment.c:_mi_segment_page_start`)
+#define MI_MAX_ALIGN_GUARANTEE   (MI_MEDIUM_OBJ_SIZE_MAX)
+
 // Alignments over MI_BLOCK_ALIGNMENT_MAX are allocated in dedicated huge page segments
 #define MI_BLOCK_ALIGNMENT_MAX   (MI_SEGMENT_SIZE >> 1)
 

src/alloc-aligned.c

@@ -15,15 +15,15 @@ terms of the MIT license. A copy of the license can be found in the file
 // Aligned Allocation
 // ------------------------------------------------------
 
-static inline bool mi_is_naturally_aligned( size_t size, size_t alignment ) {
-  // objects up to `MI_MEDIUM_OBJ_SIZE_MAX` are allocated aligned to their size (see `segment.c:_mi_segment_page_start`).
-  // note: the size may not be not an actual bin-size but it turns out the test below is still correct for our
-  // powers of two bin spacing (see test-api.c:test-aligned13).
+static bool mi_malloc_is_naturally_aligned( size_t size, size_t alignment ) {
+  // objects up to `MI_MAX_ALIGN_GUARANTEE` are allocated aligned to their size (see `segment.c:_mi_segment_page_start`).
   mi_assert_internal(_mi_is_power_of_two(alignment) && (alignment > 0));
-  return (size <= MI_MEDIUM_OBJ_SIZE_MAX && alignment <= size && ((size + MI_PADDING_SIZE) & (alignment-1)) == 0);
+  if (alignment > size) return false;
+  if (alignment <= MI_MAX_ALIGN_SIZE) return true;
+  const size_t bsize = mi_good_size(size);
+  return (bsize <= MI_MAX_ALIGN_GUARANTEE && (bsize & (alignment-1)) == 0);
 }
 
-
 // Fallback primitive aligned allocation -- split out for better codegen
 static mi_decl_noinline void* mi_heap_malloc_zero_aligned_at_fallback(mi_heap_t* const heap, const size_t size, const size_t alignment, const size_t offset, const bool zero) mi_attr_noexcept
 {
@@ -31,10 +31,18 @@ static mi_decl_noinline void* mi_heap_malloc_zero_aligned_at_fallback(mi_heap_t*
   mi_assert_internal(alignment != 0 && _mi_is_power_of_two(alignment));
 
   // use regular allocation if it is guaranteed to fit the alignment constraints.
-  if (offset == 0 && mi_is_naturally_aligned(size,alignment)) {
+  if (offset == 0 && mi_malloc_is_naturally_aligned(size,alignment)) {
     void* p = _mi_heap_malloc_zero(heap, size, zero);
     mi_assert_internal(p == NULL || ((uintptr_t)p % alignment) == 0);
-    return p;
+    const bool is_aligned_or_null = (((uintptr_t)p) & (alignment-1))==0;  
+    if mi_likely(is_aligned_or_null) {
+      return p;
+    }
+    else {
+      // this should never happen if the `mi_malloc_is_naturally_aligned` check is correct..
+      mi_assert(false);
+      mi_free(p); 
+    }
   }
 
   void* p;
@@ -106,33 +114,35 @@ static void* mi_heap_malloc_zero_aligned_at(mi_heap_t* const heap, const size_t
     #endif
     return NULL;
   }
+  
+  // try first if there happens to be a small block available with just the right alignment
+  if mi_likely(size <= MI_SMALL_SIZE_MAX && alignment <= size) {
+    const uintptr_t align_mask = alignment-1;       // for any x, `(x & align_mask) == (x % alignment)`
+    const size_t padsize = size + MI_PADDING_SIZE;  
+    mi_page_t* page = _mi_heap_get_free_small_page(heap, padsize);
+    if mi_likely(page->free != NULL) {
+      const bool is_aligned = (((uintptr_t)page->free + offset) & align_mask)==0;
+      if mi_likely(is_aligned)
+      {
+        #if MI_STAT>1
+        mi_heap_stat_increase(heap, malloc, size);
+        #endif
+        void* p = (zero ? _mi_page_malloc_zeroed(heap,page,padsize) : _mi_page_malloc(heap,page,padsize)); // call specific page malloc for better codegen
+        mi_assert_internal(p != NULL);
+        mi_assert_internal(((uintptr_t)p + offset) % alignment == 0);
+        mi_track_malloc(p,size,zero);
+        return p;
+      }
+    }
+  }
 
+  // fallback
   if mi_unlikely(size > (MI_MAX_ALLOC_SIZE - MI_PADDING_SIZE)) { // we don't allocate more than MI_MAX_ALLOC_SIZE (see <https://sourceware.org/ml/libc-announce/2019/msg00001.html>)
     #if MI_DEBUG > 0
     _mi_error_message(EOVERFLOW, "aligned allocation request is too large (size %zu, alignment %zu)\n", size, alignment);
     #endif
     return NULL;
   }
-  const uintptr_t align_mask = alignment-1;       // for any x, `(x & align_mask) == (x % alignment)`
-  const size_t padsize = size + MI_PADDING_SIZE;  // note: cannot overflow due to earlier size  check
-
-  // try first if there happens to be a small block available with just the right alignment
-  if mi_likely(padsize <= MI_SMALL_SIZE_MAX && alignment <= padsize) {
-    mi_page_t* page = _mi_heap_get_free_small_page(heap, padsize);
-    const bool is_aligned = (((uintptr_t)page->free+offset) & align_mask)==0;
-    if mi_likely(page->free != NULL && is_aligned)
-    {
-      #if MI_STAT>1
-      mi_heap_stat_increase(heap, malloc, size);
-      #endif
-      void* p = _mi_page_malloc(heap, page, padsize, zero); // TODO: inline _mi_page_malloc
-      mi_assert_internal(p != NULL);
-      mi_assert_internal(((uintptr_t)p + offset) % alignment == 0);
-      mi_track_malloc(p,size,zero);
-      return p;
-    }
-  }
-  // fallback
   return mi_heap_malloc_zero_aligned_at_fallback(heap, size, alignment, offset, zero);
 }
 
@@ -146,14 +156,7 @@ mi_decl_nodiscard mi_decl_restrict void* mi_heap_malloc_aligned_at(mi_heap_t* he
 }
 
 mi_decl_nodiscard mi_decl_restrict void* mi_heap_malloc_aligned(mi_heap_t* heap, size_t size, size_t alignment) mi_attr_noexcept {
-  if (alignment == 0 || !_mi_is_power_of_two(alignment)) return NULL;
-  if (size <= MI_SMALL_SIZE_MAX && mi_is_naturally_aligned(size,alignment)) {
-    // fast path for common alignment and size
-    return mi_heap_malloc_small(heap, size);
-  }
-  else {
-    return mi_heap_malloc_aligned_at(heap, size, alignment, 0);
-  }
+  return mi_heap_malloc_aligned_at(heap, size, alignment, 0);
 }
 
 // ------------------------------------------------------

src/alloc.c

@@ -28,7 +28,7 @@ terms of the MIT license. A copy of the license can be found in the file
 // Fast allocation in a page: just pop from the free list.
 // Fall back to generic allocation only if the list is empty.
 // Note: in release mode the (inlined) routine is about 7 instructions with a single test.
-extern inline void* _mi_page_malloc(mi_heap_t* heap, mi_page_t* page, size_t size, bool zero) mi_attr_noexcept 
+extern inline void* _mi_page_malloc_zero(mi_heap_t* heap, mi_page_t* page, size_t size, bool zero) mi_attr_noexcept 
 {
   mi_assert_internal(page->block_size == 0 /* empty heap */ || mi_page_block_size(page) >= size);
   mi_block_t* const block = page->free;
@@ -85,14 +85,14 @@ extern inline void* _mi_page_malloc(mi_heap_t* heap, mi_page_t* page, size_t siz
   #endif
 
   #if MI_PADDING // && !MI_TRACK_ENABLED
-  mi_padding_t* const padding = (mi_padding_t*)((uint8_t*)block + mi_page_usable_block_size(page));
-  ptrdiff_t delta = ((uint8_t*)padding - (uint8_t*)block - (size - MI_PADDING_SIZE));
+    mi_padding_t* const padding = (mi_padding_t*)((uint8_t*)block + mi_page_usable_block_size(page));
+    ptrdiff_t delta = ((uint8_t*)padding - (uint8_t*)block - (size - MI_PADDING_SIZE));
     #if (MI_DEBUG>=2)
     mi_assert_internal(delta >= 0 && mi_page_usable_block_size(page) >= (size - MI_PADDING_SIZE + delta));
     #endif
-  mi_track_mem_defined(padding,sizeof(mi_padding_t));  // note: re-enable since mi_page_usable_block_size may set noaccess
-  padding->canary = (uint32_t)(mi_ptr_encode(page,block,page->keys));
-  padding->delta  = (uint32_t)(delta);
+    mi_track_mem_defined(padding,sizeof(mi_padding_t));  // note: re-enable since mi_page_usable_block_size may set noaccess
+    padding->canary = (uint32_t)(mi_ptr_encode(page,block,page->keys));
+    padding->delta  = (uint32_t)(delta);
     #if MI_PADDING_CHECK
     if (!mi_page_is_huge(page)) {
       uint8_t* fill = (uint8_t*)padding - delta;
@@ -105,6 +105,14 @@ extern inline void* _mi_page_malloc(mi_heap_t* heap, mi_page_t* page, size_t siz
   return block;
 }
 
+// extra entries for improved efficiency in `alloc-aligned.c`.
+extern inline void* _mi_page_malloc(mi_heap_t* heap, mi_page_t* page, size_t size) mi_attr_noexcept {
+  return _mi_page_malloc_zero(heap,page,size,false);
+}
+extern inline void* _mi_page_malloc_zeroed(mi_heap_t* heap, mi_page_t* page, size_t size) mi_attr_noexcept {
+  return _mi_page_malloc_zero(heap,page,size,true);
+}
+
 static inline mi_decl_restrict void* mi_heap_malloc_small_zero(mi_heap_t* heap, size_t size, bool zero) mi_attr_noexcept {
   mi_assert(heap != NULL);
   #if MI_DEBUG
@@ -117,7 +125,7 @@ static inline mi_decl_restrict void* mi_heap_malloc_small_zero(mi_heap_t* heap,
   #endif
   
   mi_page_t* page = _mi_heap_get_free_small_page(heap, size + MI_PADDING_SIZE);
-  void* const p = _mi_page_malloc(heap, page, size + MI_PADDING_SIZE, zero);  
+  void* const p = _mi_page_malloc_zero(heap, page, size + MI_PADDING_SIZE, zero);  
   mi_track_malloc(p,size,zero);
 
   #if MI_STAT>1

src/page-queue.c

@@ -113,10 +113,10 @@ size_t _mi_bin_size(uint8_t bin) {
 // Good size for allocation
 size_t mi_good_size(size_t size) mi_attr_noexcept {
   if (size <= MI_LARGE_OBJ_SIZE_MAX) {
-    return _mi_bin_size(mi_bin(size));
+    return _mi_bin_size(mi_bin(size + MI_PADDING_SIZE));
   }
   else {
-    return _mi_align_up(size,_mi_os_page_size());
+    return _mi_align_up(size + MI_PADDING_SIZE,_mi_os_page_size());
   }
 }
 

src/page.c

@@ -914,12 +914,12 @@ void* _mi_malloc_generic(mi_heap_t* heap, size_t size, bool zero, size_t huge_al
   // and try again, this time succeeding! (i.e. this should never recurse through _mi_page_malloc)
   if mi_unlikely(zero && page->block_size == 0) {
     // note: we cannot call _mi_page_malloc with zeroing for huge blocks; we zero it afterwards in that case.
-    void* p = _mi_page_malloc(heap, page, size, false);
+    void* p = _mi_page_malloc(heap, page, size);
     mi_assert_internal(p != NULL);
     _mi_memzero_aligned(p, mi_page_usable_block_size(page));
     return p;
   }
   else {
-    return _mi_page_malloc(heap, page, size, zero);
+    return _mi_page_malloc_zero(heap, page, size, zero);
   }
 }

src/segment.c

@@ -426,15 +426,13 @@ uint8_t* _mi_segment_page_start(const mi_segment_t* segment, const mi_page_t* pa
   size_t   psize;
   uint8_t* p = mi_segment_raw_page_start(segment, page, &psize);
   const size_t block_size = mi_page_block_size(page);
-  if (page->segment_idx == 0 && block_size > 0 && segment->page_kind <= MI_PAGE_MEDIUM) {
+  if (/*page->segment_idx == 0 &&*/ block_size > 0 && block_size <= MI_MAX_ALIGN_GUARANTEE) {
     // for small and medium objects, ensure the page start is aligned with the block size (PR#66 by kickunderscore)
+    mi_assert_internal(segment->page_kind <= MI_PAGE_MEDIUM);
     size_t adjust = block_size - ((uintptr_t)p % block_size);
-    if (psize - adjust >= block_size) {
-      if (adjust < block_size) {
-        p += adjust;
-        psize -= adjust;
-        // if (pre_size != NULL) *pre_size = adjust;
-      }
+    if (adjust < block_size && psize >= block_size + adjust) {
+      p += adjust;
+      psize -= adjust;
       mi_assert_internal((uintptr_t)p % block_size == 0);
     }
   }