diff --git a/ide/vs2022/mimalloc-override.vcxproj b/ide/vs2022/mimalloc-override.vcxproj
index f22d54a4..13722051 100644
--- a/ide/vs2022/mimalloc-override.vcxproj
+++ b/ide/vs2022/mimalloc-override.vcxproj
@@ -123,7 +123,7 @@
       <SDLCheck>true</SDLCheck>
       <ConformanceMode>true</ConformanceMode>
       <AdditionalIncludeDirectories>../../include</AdditionalIncludeDirectories>
-      <PreprocessorDefinitions>MI_DEBUG=3;MI_SHARED_LIB;MI_SHARED_LIB_EXPORT;MI_MALLOC_OVERRIDE;%(PreprocessorDefinitions);</PreprocessorDefinitions>
+      <PreprocessorDefinitions>MI_DEBUG=4;MI_SHARED_LIB;MI_SHARED_LIB_EXPORT;MI_MALLOC_OVERRIDE;%(PreprocessorDefinitions);</PreprocessorDefinitions>
       <RuntimeLibrary>MultiThreadedDebugDLL</RuntimeLibrary>
       <SupportJustMyCode>false</SupportJustMyCode>
       <CompileAs>Default</CompileAs>
diff --git a/ide/vs2022/mimalloc.vcxproj b/ide/vs2022/mimalloc.vcxproj
index 00aaaffa..616c226c 100644
--- a/ide/vs2022/mimalloc.vcxproj
+++ b/ide/vs2022/mimalloc.vcxproj
@@ -116,7 +116,7 @@
       <SDLCheck>true</SDLCheck>
       <ConformanceMode>true</ConformanceMode>
       <AdditionalIncludeDirectories>../../include</AdditionalIncludeDirectories>
-      <PreprocessorDefinitions>MI_DEBUG=3;%(PreprocessorDefinitions);</PreprocessorDefinitions>
+      <PreprocessorDefinitions>MI_DEBUG=4;%(PreprocessorDefinitions);</PreprocessorDefinitions>
       <CompileAs>CompileAsCpp</CompileAs>
       <SupportJustMyCode>false</SupportJustMyCode>
       <LanguageStandard>stdcpp20</LanguageStandard>
diff --git a/include/mimalloc-internal.h b/include/mimalloc-internal.h
index 550b6543..40bb1349 100644
--- a/include/mimalloc-internal.h
+++ b/include/mimalloc-internal.h
@@ -88,10 +88,13 @@ bool       _mi_os_reset(void* p, size_t size, mi_stats_t* stats);
 size_t     _mi_os_good_alloc_size(size_t size);
 bool       _mi_os_has_overcommit(void);
 
+void*      _mi_os_alloc_aligned_offset(size_t size, size_t alignment, size_t align_offset, bool commit, bool* large, mi_stats_t* tld_stats);
+void       _mi_os_free_aligned(void* p, size_t size, size_t alignment, size_t align_offset, bool was_committed, mi_stats_t* tld_stats);
+
 // arena.c
-void*      _mi_arena_alloc_aligned(size_t size, size_t alignment, bool* commit, bool* large, bool* is_pinned, bool* is_zero, mi_arena_id_t req_arena_id, size_t* memid, mi_os_tld_t* tld);
+void*      _mi_arena_alloc_aligned(size_t size, size_t alignment, size_t align_offset, bool* commit, bool* large, bool* is_pinned, bool* is_zero, mi_arena_id_t req_arena_id, size_t* memid, mi_os_tld_t* tld);
 void*      _mi_arena_alloc(size_t size, bool* commit, bool* large, bool* is_pinned, bool* is_zero, mi_arena_id_t req_arena_id, size_t* memid, mi_os_tld_t* tld);
-void       _mi_arena_free(void* p, size_t size, size_t memid, bool is_committed, mi_os_tld_t* tld);
+void       _mi_arena_free(void* p, size_t size, size_t alignment, size_t align_offset, size_t memid, bool all_committed, mi_stats_t* stats);
 mi_arena_id_t _mi_arena_id_none(void);
 bool       _mi_arena_memid_is_suitable(size_t memid, mi_arena_id_t req_arena_id);
 
@@ -103,7 +106,7 @@ void       _mi_segment_map_allocated_at(const mi_segment_t* segment);
 void       _mi_segment_map_freed_at(const mi_segment_t* segment);
 
 // "segment.c"
-mi_page_t* _mi_segment_page_alloc(mi_heap_t* heap, size_t block_wsize, mi_segments_tld_t* tld, mi_os_tld_t* os_tld);
+mi_page_t* _mi_segment_page_alloc(mi_heap_t* heap, size_t block_size, size_t page_alignment, mi_segments_tld_t* tld, mi_os_tld_t* os_tld);
 void       _mi_segment_page_free(mi_page_t* page, bool force, mi_segments_tld_t* tld);
 void       _mi_segment_page_abandon(mi_page_t* page, mi_segments_tld_t* tld);
 bool       _mi_segment_try_reclaim_abandoned( mi_heap_t* heap, bool try_all, mi_segments_tld_t* tld);
@@ -118,7 +121,7 @@ void       _mi_abandoned_collect(mi_heap_t* heap, bool force, mi_segments_tld_t*
 
 
 // "page.c"
-void*      _mi_malloc_generic(mi_heap_t* heap, size_t size, bool zero)  mi_attr_noexcept mi_attr_malloc;
+void*      _mi_malloc_generic(mi_heap_t* heap, size_t size, bool zero, size_t huge_alignment)  mi_attr_noexcept mi_attr_malloc;
 
 void       _mi_page_retire(mi_page_t* page) mi_attr_noexcept;                  // free the page if there are no other pages with many free blocks
 void       _mi_page_unfull(mi_page_t* page);
@@ -155,6 +158,7 @@ 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_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;
 mi_block_t* _mi_page_ptr_unalign(const mi_segment_t* segment, const mi_page_t* page, const void* p);
 bool        _mi_free_delayed_block(mi_block_t* block);
@@ -446,8 +450,8 @@ static inline mi_page_t* _mi_get_free_small_page(size_t size) {
 
 // Segment that contains the pointer
 static inline mi_segment_t* _mi_ptr_segment(const void* p) {
-  // mi_assert_internal(p != NULL);
-  return (mi_segment_t*)((uintptr_t)p & ~MI_SEGMENT_MASK);
+  mi_assert_internal(p != NULL);
+  return (mi_segment_t*)(((uintptr_t)p - 1) & ~MI_SEGMENT_MASK);
 }
 
 static inline mi_page_t* mi_slice_to_page(mi_slice_t* s) {
@@ -478,7 +482,8 @@ static inline mi_slice_t* mi_slice_first(const mi_slice_t* slice) {
 // Get the page containing the pointer
 static inline mi_page_t* _mi_segment_page_of(const mi_segment_t* segment, const void* p) {
   ptrdiff_t diff = (uint8_t*)p - (uint8_t*)segment;
-  mi_assert_internal(diff >= 0 && diff < (ptrdiff_t)MI_SEGMENT_SIZE);
+  mi_assert_internal(diff >= 0 && diff <= (ptrdiff_t)MI_SEGMENT_SIZE /* can be equal for large alignment */);
+  if (diff == MI_SEGMENT_SIZE) diff--;
   size_t idx = (size_t)diff >> MI_SEGMENT_SLICE_SHIFT;
   mi_assert_internal(idx < segment->slice_entries);
   mi_slice_t* slice0 = (mi_slice_t*)&segment->slices[idx];
diff --git a/include/mimalloc-types.h b/include/mimalloc-types.h
index 800d9413..b5931789 100644
--- a/include/mimalloc-types.h
+++ b/include/mimalloc-types.h
@@ -174,7 +174,8 @@ typedef int32_t  mi_ssize_t;
 #endif
 
 // Maximum slice offset (15)
-#define MI_MAX_SLICE_OFFSET               ((MI_ALIGNMENT_MAX / MI_SEGMENT_SLICE_SIZE) - 1)
+// #define MI_MAX_SLICE_OFFSET               ((MI_ALIGNMENT_MAX / MI_SEGMENT_SLICE_SIZE) - 1)
+#define MI_MAX_SLICE_OFFSET               ((MI_SEGMENT_SIZE / MI_SEGMENT_SLICE_SIZE))
 
 // Used as a special value to encode block sizes in 32 bits.
 #define MI_HUGE_BLOCK_SIZE                ((uint32_t)(2*MI_GiB))
@@ -355,6 +356,8 @@ typedef struct mi_segment_s {
   bool              mem_is_pinned;      // `true` if we cannot decommit/reset/protect in this memory (i.e. when allocated using large OS pages)    
   bool              mem_is_large;       // in large/huge os pages?
   bool              mem_is_committed;   // `true` if the whole segment is eagerly committed
+  size_t            mem_alignment;      // page alignment for huge pages (only used for alignment > MI_ALIGNMENT_MAX)
+  size_t            mem_align_offset;   // offset for huge page alignment (only used for alignment > MI_ALIGNMENT_MAX)
 
   bool              allow_decommit;     
   mi_msecs_t        decommit_expire;
diff --git a/include/mimalloc.h b/include/mimalloc.h
index 32eab19e..2a875522 100644
--- a/include/mimalloc.h
+++ b/include/mimalloc.h
@@ -167,10 +167,12 @@ mi_decl_export void mi_process_info(size_t* elapsed_msecs, size_t* user_msecs, s
 // Note that `alignment` always follows `size` for consistency with unaligned
 // allocation, but unfortunately this differs from `posix_memalign` and `aligned_alloc`.
 // -------------------------------------------------------------------------------------
+
+// The MI_ALIGNMENT_MAX is deprecated; any alignment is supported but alignments up to MI_ALIGNMENT_MAX may be cheaper.
 #if (INTPTR_MAX > INT32_MAX)
-#define MI_ALIGNMENT_MAX   (16*1024*1024UL)  // maximum supported alignment is 16MiB
+#define MI_ALIGNMENT_MAX   (32*1024*1024UL)  
 #else
-#define MI_ALIGNMENT_MAX   (1024*1024UL)     // maximum supported alignment for 32-bit systems is 1MiB
+#define MI_ALIGNMENT_MAX   (2*1024*1024UL)   
 #endif
 
 mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_malloc_aligned(size_t size, size_t alignment) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(1) mi_attr_alloc_align(2);
diff --git a/src/alloc-aligned.c b/src/alloc-aligned.c
index 9614aa09..76ed0ed7 100644
--- a/src/alloc-aligned.c
+++ b/src/alloc-aligned.c
@@ -18,9 +18,9 @@ terms of the MIT license. A copy of the license can be found in the file
 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
 {
   mi_assert_internal(size <= PTRDIFF_MAX);
-  mi_assert_internal(alignment!=0 && _mi_is_power_of_two(alignment) && alignment <= MI_ALIGNMENT_MAX);
+  mi_assert_internal(alignment != 0 && _mi_is_power_of_two(alignment));
 
-  const uintptr_t align_mask = alignment-1;  // for any x, `(x & align_mask) == (x % alignment)`
+  const uintptr_t align_mask = alignment - 1;  // for any x, `(x & align_mask) == (x % alignment)`
   const size_t padsize = size + MI_PADDING_SIZE;
 
   // use regular allocation if it is guaranteed to fit the alignment constraints
@@ -30,29 +30,52 @@ static mi_decl_noinline void* mi_heap_malloc_zero_aligned_at_fallback(mi_heap_t*
     return p;
   }
 
-  // otherwise over-allocate
-  const size_t oversize = size + alignment - 1;
-  void* p = _mi_heap_malloc_zero(heap, oversize, zero);
-  if (p == NULL) return NULL;
+  void* p;
+  size_t oversize;
+  if mi_unlikely(alignment > MI_ALIGNMENT_MAX) {
+    // use OS allocation for very large alignment and allocate inside a huge page (dedicated segment with 1 page)
+    if mi_unlikely(offset != 0) {
+      // todo: cannot support offset alignment for very large alignments yet
+#if MI_DEBUG > 0
+      _mi_error_message(EOVERFLOW, "aligned allocation with a very large alignment cannot be used with an alignment offset (size %zu, alignment %zu, offset %zu)\n", size, alignment, offset);
+#endif
+      return NULL;
+    }
+    oversize = (size <= MI_SMALL_SIZE_MAX ? MI_SMALL_SIZE_MAX + 1 /* ensure we use generic malloc path */ : size);
+    p = _mi_heap_malloc_zero_ex(heap, oversize, zero, alignment); // the page block size should be large enough to align in the single huge page block
+    if (p == NULL) return NULL;
+    const uintptr_t adjustx = alignment - (((uintptr_t)p + offset) & align_mask);
+    const mi_page_t* page = _mi_ptr_page(p);
+    const size_t bsize = mi_page_usable_block_size(page);
+    mi_assert_internal(bsize >= adjustx + size);
+    mi_assert_internal(true);
+  }
+  else {
+    // otherwise over-allocate
+    oversize = size + alignment - 1;
+    p = _mi_heap_malloc_zero(heap, oversize, zero);
+    if (p == NULL) return NULL;
+  }
 
   // .. and align within the allocation
   uintptr_t adjust = alignment - (((uintptr_t)p + offset) & align_mask);
   mi_assert_internal(adjust <= alignment);
   void* aligned_p = (adjust == alignment ? p : (void*)((uintptr_t)p + adjust));
-  if (aligned_p != p) mi_page_set_has_aligned(_mi_ptr_page(p), true);
-  mi_assert_internal(((uintptr_t)aligned_p + offset) % alignment == 0);
+  if (aligned_p != p) { mi_page_set_has_aligned(_mi_ptr_page(p), true);  }
+    
   mi_assert_internal(p == _mi_page_ptr_unalign(_mi_ptr_segment(aligned_p), _mi_ptr_page(aligned_p), aligned_p));
-
+  mi_assert_internal(((uintptr_t)aligned_p + offset) % alignment == 0);
+  
   #if MI_TRACK_ENABLED
   if (p != aligned_p) {
     mi_track_free(p);
-    mi_track_malloc(aligned_p,size,zero);
+    mi_track_malloc(aligned_p, size, zero);
   }
   else {
-    mi_track_resize(aligned_p,oversize,size);
+    mi_track_resize(aligned_p, oversize, size);
   }
   #endif
-  return aligned_p;
+  return aligned_p;  
 }
 
 // Primitive aligned allocation
@@ -60,18 +83,20 @@ static void* mi_heap_malloc_zero_aligned_at(mi_heap_t* const heap, const size_t
 {
   // note: we don't require `size > offset`, we just guarantee that the address at offset is aligned regardless of the allocated size.
   mi_assert(alignment > 0);
-  if mi_unlikely(alignment==0 || !_mi_is_power_of_two(alignment)) { // require power-of-two (see <https://en.cppreference.com/w/c/memory/aligned_alloc>)
+  if mi_unlikely(alignment == 0 || !_mi_is_power_of_two(alignment)) { // require power-of-two (see <https://en.cppreference.com/w/c/memory/aligned_alloc>)
     #if MI_DEBUG > 0
     _mi_error_message(EOVERFLOW, "aligned allocation requires the alignment to be a power-of-two (size %zu, alignment %zu)\n", size, alignment);
     #endif
     return NULL;
   }
+  /*
   if mi_unlikely(alignment > MI_ALIGNMENT_MAX) {  // we cannot align at a boundary larger than this (or otherwise we cannot find segment headers)
     #if MI_DEBUG > 0
     _mi_error_message(EOVERFLOW, "aligned allocation has a maximum alignment of %zu (size %zu, alignment %zu)\n", MI_ALIGNMENT_MAX, size, alignment);
     #endif
     return NULL;
   }
+  */
   if mi_unlikely(size > PTRDIFF_MAX) {          // we don't allocate more than PTRDIFF_MAX (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);
@@ -82,7 +107,7 @@ static void* mi_heap_malloc_zero_aligned_at(mi_heap_t* const heap, const size_t
   const size_t padsize = size + MI_PADDING_SIZE;  // note: cannot overflow due to earlier size > PTRDIFF_MAX 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) {
+  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)
diff --git a/src/alloc.c b/src/alloc.c
index 02d009e2..1117b5fc 100644
--- a/src/alloc.c
+++ b/src/alloc.c
@@ -30,7 +30,7 @@ extern inline void* _mi_page_malloc(mi_heap_t* heap, mi_page_t* page, size_t siz
   mi_assert_internal(page->xblock_size==0||mi_page_block_size(page) >= size);
   mi_block_t* const block = page->free;
   if mi_unlikely(block == NULL) {
-    return _mi_malloc_generic(heap, size, zero); 
+    return _mi_malloc_generic(heap, size, zero, 0); 
   }
   mi_assert_internal(block != NULL && _mi_ptr_page(block) == page);
   // pop from the free list
@@ -117,14 +117,15 @@ mi_decl_nodiscard extern inline mi_decl_restrict void* mi_malloc_small(size_t si
 }
 
 // The main allocation function
-extern inline void* _mi_heap_malloc_zero(mi_heap_t* heap, size_t size, bool zero) mi_attr_noexcept {
+inline void* _mi_heap_malloc_zero_ex(mi_heap_t* heap, size_t size, bool zero, size_t huge_alignment) mi_attr_noexcept {
   if mi_likely(size <= MI_SMALL_SIZE_MAX) {
+    mi_assert_internal(huge_alignment == 0);
     return mi_heap_malloc_small_zero(heap, size, zero);
   }
   else {
     mi_assert(heap!=NULL);
     mi_assert(heap->thread_id == 0 || heap->thread_id == _mi_thread_id());   // heaps are thread local
-    void* const p = _mi_malloc_generic(heap, size + MI_PADDING_SIZE, zero);  // note: size can overflow but it is detected in malloc_generic
+    void* const p = _mi_malloc_generic(heap, size + MI_PADDING_SIZE, zero, huge_alignment);  // note: size can overflow but it is detected in malloc_generic
     mi_assert_internal(p == NULL || mi_usable_size(p) >= size);
     #if MI_STAT>1
     if (p != NULL) {
@@ -137,6 +138,10 @@ extern inline void* _mi_heap_malloc_zero(mi_heap_t* heap, size_t size, bool zero
   }
 }
 
+extern inline void* _mi_heap_malloc_zero(mi_heap_t* heap, size_t size, bool zero) mi_attr_noexcept {
+  return _mi_heap_malloc_zero_ex(heap, size, zero, 0);
+}
+
 mi_decl_nodiscard extern inline mi_decl_restrict void* mi_heap_malloc(mi_heap_t* heap, size_t size) mi_attr_noexcept {
   return _mi_heap_malloc_zero(heap, size, false);
 }
@@ -471,8 +476,8 @@ static inline mi_segment_t* mi_checked_ptr_segment(const void* p, const char* ms
   }
 #endif
 
+  if mi_unlikely(p == NULL) return NULL;
   mi_segment_t* const segment = _mi_ptr_segment(p);
-  if mi_unlikely(segment == NULL) return NULL;  // checks also for (p==NULL)
 
 #if (MI_DEBUG>0)
   if mi_unlikely(!mi_is_in_heap_region(p)) {
diff --git a/src/arena.c b/src/arena.c
index 239c5667..f952559e 100644
--- a/src/arena.c
+++ b/src/arena.c
@@ -245,7 +245,7 @@ static mi_decl_noinline void* mi_arena_allocate(int numa_node, size_t size, size
 }
 
 
-void* _mi_arena_alloc_aligned(size_t size, size_t alignment, bool* commit, bool* large, bool* is_pinned, bool* is_zero,
+void* _mi_arena_alloc_aligned(size_t size, size_t alignment, size_t align_offset, bool* commit, bool* large, bool* is_pinned, bool* is_zero,
                               mi_arena_id_t req_arena_id, size_t* memid, mi_os_tld_t* tld)
 {
   mi_assert_internal(commit != NULL && is_pinned != NULL && is_zero != NULL && memid != NULL && tld != NULL);
@@ -259,7 +259,7 @@ void* _mi_arena_alloc_aligned(size_t size, size_t alignment, bool* commit, bool*
   const int numa_node = _mi_os_numa_node(tld); // current numa node
 
   // try to allocate in an arena if the alignment is small enough and the object is not too small (as for heap meta data)
-  if (size >= MI_ARENA_MIN_OBJ_SIZE && alignment <= MI_SEGMENT_ALIGN) {
+  if (size >= MI_ARENA_MIN_OBJ_SIZE && alignment <= MI_SEGMENT_ALIGN && align_offset == 0) {
     void* p = mi_arena_allocate(numa_node, size, alignment, commit, large, is_pinned, is_zero, req_arena_id, memid, tld);
     if (p != NULL) return p;
   }
@@ -271,14 +271,14 @@ void* _mi_arena_alloc_aligned(size_t size, size_t alignment, bool* commit, bool*
   }
   *is_zero = true;
   *memid   = MI_MEMID_OS;  
-  void* p = _mi_os_alloc_aligned(size, alignment, *commit, large, tld->stats);
+  void* p = _mi_os_alloc_aligned_offset(size, alignment, align_offset, *commit, large, tld->stats);
   if (p != NULL) *is_pinned = *large;
   return p;
 }
 
 void* _mi_arena_alloc(size_t size, bool* commit, bool* large, bool* is_pinned, bool* is_zero, mi_arena_id_t req_arena_id, size_t* memid, mi_os_tld_t* tld)
 {
-  return _mi_arena_alloc_aligned(size, MI_ARENA_BLOCK_SIZE, commit, large, is_pinned, is_zero, req_arena_id, memid, tld);
+  return _mi_arena_alloc_aligned(size, MI_ARENA_BLOCK_SIZE, 0, commit, large, is_pinned, is_zero, req_arena_id, memid, tld);
 }
 
 void* mi_arena_area(mi_arena_id_t arena_id, size_t* size) {
@@ -295,17 +295,18 @@ void* mi_arena_area(mi_arena_id_t arena_id, size_t* size) {
   Arena free
 ----------------------------------------------------------- */
 
-void _mi_arena_free(void* p, size_t size, size_t memid, bool all_committed, mi_os_tld_t* tld) {
-  mi_assert_internal(size > 0 && tld->stats != NULL);
+void _mi_arena_free(void* p, size_t size, size_t alignment, size_t align_offset, size_t memid, bool all_committed, mi_stats_t* stats) {
+  mi_assert_internal(size > 0 && stats != NULL);
   if (p==NULL) return;
   if (size==0) return;
 
   if (memid == MI_MEMID_OS) {
     // was a direct OS allocation, pass through
-    _mi_os_free_ex(p, size, all_committed, tld->stats);
+    _mi_os_free_aligned(p, size, alignment, align_offset, all_committed, stats);
   }
   else {
     // allocated in an arena
+    mi_assert_internal(align_offset == 0);
     size_t arena_idx;
     size_t bitmap_idx;
     mi_arena_memid_indices(memid, &arena_idx, &bitmap_idx);
@@ -329,7 +330,7 @@ void _mi_arena_free(void* p, size_t size, size_t memid, bool all_committed, mi_o
     }
     else {
       mi_assert_internal(arena->blocks_committed != NULL);
-      _mi_os_decommit(p, blocks * MI_ARENA_BLOCK_SIZE, tld->stats); // ok if this fails
+      _mi_os_decommit(p, blocks * MI_ARENA_BLOCK_SIZE, stats); // ok if this fails
       _mi_bitmap_unclaim_across(arena->blocks_committed, arena->field_count, blocks, bitmap_idx);
     }
     // and make it available to others again 
diff --git a/src/os.c b/src/os.c
index 6d724987..0f4ab3e1 100644
--- a/src/os.c
+++ b/src/os.c
@@ -837,8 +837,45 @@ void* _mi_os_alloc_aligned(size_t size, size_t alignment, bool commit, bool* lar
   return mi_os_mem_alloc_aligned(size, alignment, commit, allow_large, (large!=NULL?large:&allow_large), &_mi_stats_main /*tld->stats*/ );
 }
 
+/* -----------------------------------------------------------
+  OS aligned allocation with an offset. This is used
+  for large alignments > MI_SEGMENT_SIZE so we can align
+  the first page at an offset from the start of the segment.
+  As we may need to overallocate, we need to free such pointers
+  using `mi_free_aligned` to use the actual start of the 
+  memory region.
+----------------------------------------------------------- */
 
 
+void* _mi_os_alloc_aligned_offset(size_t size, size_t alignment, size_t offset, bool commit, bool* large, mi_stats_t* tld_stats) {
+  mi_assert(offset <= MI_SEGMENT_SIZE);
+  mi_assert(offset <= size);
+  mi_assert((alignment % _mi_os_page_size()) == 0);
+  if (offset > MI_SEGMENT_SIZE) return NULL;
+  if (offset == 0) {
+    return _mi_os_alloc_aligned(size, alignment, commit, large, tld_stats);
+  }
+  else {
+    const size_t extra = _mi_align_up(offset, alignment) - offset;
+    const size_t oversize = size + extra;
+    void* start = _mi_os_alloc_aligned(oversize, alignment, commit, large, tld_stats);
+    if (start == NULL) return NULL;
+    void* p = (uint8_t*)start + extra;
+    mi_assert(_mi_is_aligned((uint8_t*)p + offset, alignment));
+    if (commit && extra > _mi_os_page_size()) {
+      _mi_os_decommit(start, extra, tld_stats);
+    }
+    return p;
+  }
+}
+
+void _mi_os_free_aligned(void* p, size_t size, size_t alignment, size_t align_offset, bool was_committed, mi_stats_t* tld_stats) {
+  mi_assert(align_offset <= MI_SEGMENT_SIZE);
+  const size_t extra = _mi_align_up(align_offset, alignment) - align_offset;
+  void* start = (uint8_t*)p - extra;
+  _mi_os_free_ex(start, size + extra, was_committed, tld_stats);  
+}
+
 /* -----------------------------------------------------------
   OS memory API: reset, commit, decommit, protect, unprotect.
 ----------------------------------------------------------- */
diff --git a/src/page.c b/src/page.c
index 4b321156..ceee5f60 100644
--- a/src/page.c
+++ b/src/page.c
@@ -254,15 +254,21 @@ void _mi_page_reclaim(mi_heap_t* heap, mi_page_t* page) {
 }
 
 // allocate a fresh page from a segment
-static mi_page_t* mi_page_fresh_alloc(mi_heap_t* heap, mi_page_queue_t* pq, size_t block_size) {
+static mi_page_t* mi_page_fresh_alloc(mi_heap_t* heap, mi_page_queue_t* pq, size_t block_size, size_t page_alignment) {
   mi_assert_internal(pq==NULL||mi_heap_contains_queue(heap, pq));
-  mi_page_t* page = _mi_segment_page_alloc(heap, block_size, &heap->tld->segments, &heap->tld->os);
+  mi_assert_internal(pq==NULL || block_size > MI_MEDIUM_OBJ_SIZE_MAX || block_size == pq->block_size);
+  mi_page_t* page = _mi_segment_page_alloc(heap, block_size, page_alignment, &heap->tld->segments, &heap->tld->os);
   if (page == NULL) {
     // this may be out-of-memory, or an abandoned page was reclaimed (and in our queue)
     return NULL;
   }
   mi_assert_internal(pq==NULL || _mi_page_segment(page)->kind != MI_SEGMENT_HUGE);
-  mi_page_init(heap, page, block_size, heap->tld);
+  mi_assert_internal(pq!=NULL || page->xblock_size != 0);
+  mi_assert_internal(pq!=NULL || mi_page_block_size(page) >= block_size);
+  // a fresh page was found, initialize it
+  const size_t full_block_size = (pq == NULL ? mi_page_block_size(page) : block_size); // see also: mi_segment_huge_page_alloc
+  mi_assert_internal(full_block_size >= block_size);
+  mi_page_init(heap, page, full_block_size, heap->tld);
   mi_heap_stat_increase(heap, pages, 1);
   if (pq!=NULL) mi_page_queue_push(heap, pq, page); // huge pages use pq==NULL
   mi_assert_expensive(_mi_page_is_valid(page));
@@ -272,7 +278,7 @@ static mi_page_t* mi_page_fresh_alloc(mi_heap_t* heap, mi_page_queue_t* pq, size
 // Get a fresh page to use
 static mi_page_t* mi_page_fresh(mi_heap_t* heap, mi_page_queue_t* pq) {
   mi_assert_internal(mi_heap_contains_queue(heap, pq));
-  mi_page_t* page = mi_page_fresh_alloc(heap, pq, pq->block_size);
+  mi_page_t* page = mi_page_fresh_alloc(heap, pq, pq->block_size, 0);
   if (page==NULL) return NULL;
   mi_assert_internal(pq->block_size==mi_page_block_size(page));
   mi_assert_internal(pq==mi_page_queue(heap, mi_page_block_size(page)));
@@ -648,6 +654,7 @@ static void mi_page_init(mi_heap_t* heap, mi_page_t* page, size_t block_size, mi
   mi_assert_internal(page_size <= page->slice_count*MI_SEGMENT_SLICE_SIZE);
   mi_assert_internal(page_size / block_size < (1L<<16));
   page->reserved = (uint16_t)(page_size / block_size);
+  mi_assert_internal(page->reserved > 0);
   #ifdef MI_ENCODE_FREELIST
   page->keys[0] = _mi_heap_random_next(heap);
   page->keys[1] = _mi_heap_random_next(heap);
@@ -797,12 +804,12 @@ void mi_register_deferred_free(mi_deferred_free_fun* fn, void* arg) mi_attr_noex
 // Because huge pages contain just one block, and the segment contains
 // just that page, we always treat them as abandoned and any thread
 // that frees the block can free the whole page and segment directly.
-static mi_page_t* mi_large_huge_page_alloc(mi_heap_t* heap, size_t size) {
+static mi_page_t* mi_large_huge_page_alloc(mi_heap_t* heap, size_t size, size_t page_alignment) {
   size_t block_size = _mi_os_good_alloc_size(size);
-  mi_assert_internal(mi_bin(block_size) == MI_BIN_HUGE);
-  bool is_huge = (block_size > MI_LARGE_OBJ_SIZE_MAX);
+  mi_assert_internal(mi_bin(block_size) == MI_BIN_HUGE || page_alignment > 0);
+  bool is_huge = (block_size > MI_LARGE_OBJ_SIZE_MAX || page_alignment > 0);
   mi_page_queue_t* pq = (is_huge ? NULL : mi_page_queue(heap, block_size));
-  mi_page_t* page = mi_page_fresh_alloc(heap, pq, block_size);
+  mi_page_t* page = mi_page_fresh_alloc(heap, pq, block_size, page_alignment);
   if (page != NULL) {
     mi_assert_internal(mi_page_immediate_available(page));
     
@@ -833,16 +840,16 @@ static mi_page_t* mi_large_huge_page_alloc(mi_heap_t* heap, size_t size) {
 
 // Allocate a page
 // Note: in debug mode the size includes MI_PADDING_SIZE and might have overflowed.
-static mi_page_t* mi_find_page(mi_heap_t* heap, size_t size) mi_attr_noexcept {
+static mi_page_t* mi_find_page(mi_heap_t* heap, size_t size, size_t huge_alignment) mi_attr_noexcept {
   // huge allocation?
   const size_t req_size = size - MI_PADDING_SIZE;  // correct for padding_size in case of an overflow on `size`  
-  if mi_unlikely(req_size > (MI_MEDIUM_OBJ_SIZE_MAX - MI_PADDING_SIZE)) {
+  if mi_unlikely(req_size > (MI_MEDIUM_OBJ_SIZE_MAX - MI_PADDING_SIZE) || huge_alignment > 0) {
     if mi_unlikely(req_size > PTRDIFF_MAX) {  // we don't allocate more than PTRDIFF_MAX (see <https://sourceware.org/ml/libc-announce/2019/msg00001.html>)
       _mi_error_message(EOVERFLOW, "allocation request is too large (%zu bytes)\n", req_size);
       return NULL;
     }
     else {
-      return mi_large_huge_page_alloc(heap,size);
+      return mi_large_huge_page_alloc(heap,size,huge_alignment);
     }
   }
   else {
@@ -854,7 +861,7 @@ static mi_page_t* mi_find_page(mi_heap_t* heap, size_t size) mi_attr_noexcept {
 
 // Generic allocation routine if the fast path (`alloc.c:mi_page_malloc`) does not succeed.
 // Note: in debug mode the size includes MI_PADDING_SIZE and might have overflowed.
-void* _mi_malloc_generic(mi_heap_t* heap, size_t size, bool zero) mi_attr_noexcept
+void* _mi_malloc_generic(mi_heap_t* heap, size_t size, bool zero, size_t huge_alignment) mi_attr_noexcept
 {
   mi_assert_internal(heap != NULL);
 
@@ -873,10 +880,10 @@ void* _mi_malloc_generic(mi_heap_t* heap, size_t size, bool zero) mi_attr_noexce
   _mi_heap_delayed_free_partial(heap);
 
   // find (or allocate) a page of the right size
-  mi_page_t* page = mi_find_page(heap, size);
+  mi_page_t* page = mi_find_page(heap, size, huge_alignment);
   if mi_unlikely(page == NULL) { // first time out of memory, try to collect and retry the allocation once more
     mi_heap_collect(heap, true /* force */);
-    page = mi_find_page(heap, size);
+    page = mi_find_page(heap, size, huge_alignment);
   }
 
   if mi_unlikely(page == NULL) { // out of memory
diff --git a/src/region.c b/src/region.c
index 8b04387d..ea376aa4 100644
--- a/src/region.c
+++ b/src/region.c
@@ -50,9 +50,9 @@ bool    _mi_os_unreset(void* p, size_t size, bool* is_zero, mi_stats_t* stats);
 
 // arena.c
 mi_arena_id_t _mi_arena_id_none(void);
-void    _mi_arena_free(void* p, size_t size, size_t memid, bool all_committed, mi_stats_t* stats);
+void    _mi_arena_free(void* p, size_t size, size_t alignment, size_t align_offset, size_t memid, bool all_committed, mi_stats_t* stats);
 void*   _mi_arena_alloc(size_t size, bool* commit, bool* large, bool* is_pinned, bool* is_zero, mi_arena_id_t req_arena_id, size_t* memid, mi_os_tld_t* tld);
-void*   _mi_arena_alloc_aligned(size_t size, size_t alignment, bool* commit, bool* large, bool* is_pinned, bool* is_zero, mi_arena_id_t req_arena_id, size_t* memid, mi_os_tld_t* tld);
+void*   _mi_arena_alloc_aligned(size_t size, size_t alignment, size_t align_offset, bool* commit, bool* large, bool* is_pinned, bool* is_zero, mi_arena_id_t req_arena_id, size_t* memid, mi_os_tld_t* tld);
 
 
 
@@ -181,7 +181,7 @@ static bool mi_region_try_alloc_os(size_t blocks, bool commit, bool allow_large,
   bool is_zero = false;
   bool is_pinned = false;
   size_t arena_memid = 0;
-  void* const start = _mi_arena_alloc_aligned(MI_REGION_SIZE, MI_SEGMENT_ALIGN, &region_commit, &region_large, &is_pinned, &is_zero, _mi_arena_id_none(),  & arena_memid, tld);
+  void* const start = _mi_arena_alloc_aligned(MI_REGION_SIZE, MI_SEGMENT_ALIGN, 0, &region_commit, &region_large, &is_pinned, &is_zero, _mi_arena_id_none(),  & arena_memid, tld);
   if (start == NULL) return false;
   mi_assert_internal(!(region_large && !allow_large));
   mi_assert_internal(!region_large || region_commit);
@@ -190,7 +190,7 @@ static bool mi_region_try_alloc_os(size_t blocks, bool commit, bool allow_large,
   const size_t idx = mi_atomic_increment_acq_rel(&regions_count);
   if (idx >= MI_REGION_MAX) {
     mi_atomic_decrement_acq_rel(&regions_count);
-    _mi_arena_free(start, MI_REGION_SIZE, arena_memid, region_commit, tld->stats);
+    _mi_arena_free(start, MI_REGION_SIZE, MI_SEGMENT_ALIGN, 0, arena_memid, region_commit, tld->stats);
     _mi_warning_message("maximum regions used: %zu GiB (perhaps recompile with a larger setting for MI_HEAP_REGION_MAX_SIZE)", _mi_divide_up(MI_HEAP_REGION_MAX_SIZE, MI_GiB));
     return false;
   }
@@ -347,7 +347,7 @@ static void* mi_region_try_alloc(size_t blocks, bool* commit, bool* large, bool*
 
 // Allocate `size` memory aligned at `alignment`. Return non NULL on success, with a given memory `id`.
 // (`id` is abstract, but `id = idx*MI_REGION_MAP_BITS + bitidx`)
-void* _mi_mem_alloc_aligned(size_t size, size_t alignment, bool* commit, bool* large, bool* is_pinned, bool* is_zero, size_t* memid, mi_os_tld_t* tld)
+void* _mi_mem_alloc_aligned(size_t size, size_t alignment, size_t align_offset, bool* commit, bool* large, bool* is_pinned, bool* is_zero, size_t* memid, mi_os_tld_t* tld)
 {
   mi_assert_internal(memid != NULL && tld != NULL);
   mi_assert_internal(size > 0);
@@ -363,7 +363,7 @@ void* _mi_mem_alloc_aligned(size_t size, size_t alignment, bool* commit, bool* l
   void* p = NULL;
   size_t arena_memid;
   const size_t blocks = mi_region_block_count(size);
-  if (blocks <= MI_REGION_MAX_OBJ_BLOCKS && alignment <= MI_SEGMENT_ALIGN) {
+  if (blocks <= MI_REGION_MAX_OBJ_BLOCKS && alignment <= MI_SEGMENT_ALIGN && align_offset == 0) {
     p = mi_region_try_alloc(blocks, commit, large, is_pinned, is_zero, memid, tld);    
     if (p == NULL) {
       _mi_warning_message("unable to allocate from region: size %zu\n", size);
@@ -371,12 +371,12 @@ void* _mi_mem_alloc_aligned(size_t size, size_t alignment, bool* commit, bool* l
   }
   if (p == NULL) {
     // and otherwise fall back to the OS
-    p = _mi_arena_alloc_aligned(size, alignment, commit, large, is_pinned, is_zero, _mi_arena_id_none(),  & arena_memid, tld);
+    p = _mi_arena_alloc_aligned(size, alignment, align_offset, commit, large, is_pinned, is_zero, _mi_arena_id_none(),  & arena_memid, tld);
     *memid = mi_memid_create_from_arena(arena_memid);
   }
 
   if (p != NULL) {
-    mi_assert_internal((uintptr_t)p % alignment == 0);
+    mi_assert_internal(((uintptr_t)p + align_offset) % alignment == 0);
     #if (MI_DEBUG>=2) && !MI_TRACK_ENABLED
     if (*commit) { ((uint8_t*)p)[0] = 0; } // ensure the memory is committed
     #endif
@@ -391,7 +391,7 @@ Free
 -----------------------------------------------------------------------------*/
 
 // Free previously allocated memory with a given id.
-void _mi_mem_free(void* p, size_t size, size_t id, bool full_commit, bool any_reset, mi_os_tld_t* tld) {
+void _mi_mem_free(void* p, size_t size, size_t alignment, size_t align_offset, size_t id, bool full_commit, bool any_reset, mi_os_tld_t* tld) {
   mi_assert_internal(size > 0 && tld != NULL);
   if (p==NULL) return;
   if (size==0) return;
@@ -402,10 +402,11 @@ void _mi_mem_free(void* p, size_t size, size_t id, bool full_commit, bool any_re
   mem_region_t* region;
   if (mi_memid_is_arena(id,&region,&bit_idx,&arena_memid)) {
    // was a direct arena allocation, pass through
-    _mi_arena_free(p, size, arena_memid, full_commit, tld->stats);
+    _mi_arena_free(p, size, alignment, align_offset, arena_memid, full_commit, tld->stats);
   }
   else {
     // allocated in a region
+    mi_assert_internal(align_offset == 0);
     mi_assert_internal(size <= MI_REGION_MAX_OBJ_SIZE); if (size > MI_REGION_MAX_OBJ_SIZE) return;
     const size_t blocks = mi_region_block_count(size);
     mi_assert_internal(blocks + bit_idx <= MI_BITMAP_FIELD_BITS);
@@ -469,7 +470,7 @@ void _mi_mem_collect(mi_os_tld_t* tld) {
         mi_atomic_store_release(&region->info, (size_t)0);
         if (start != NULL) { // && !_mi_os_is_huge_reserved(start)) {         
           _mi_abandoned_await_readers(); // ensure no pending reads
-          _mi_arena_free(start, MI_REGION_SIZE, arena_memid, (~commit == 0), tld->stats);
+          _mi_arena_free(start, MI_REGION_SIZE, MI_SEGMENT_ALIGN, 0, arena_memid, (~commit == 0), tld->stats);
         }
       }
     }
diff --git a/src/segment-cache.c b/src/segment-cache.c
index da726716..34c8b029 100644
--- a/src/segment-cache.c
+++ b/src/segment-cache.c
@@ -245,7 +245,7 @@ mi_decl_noinline bool _mi_segment_cache_push(void* start, size_t size, size_t me
 static _Atomic(uintptr_t) mi_segment_map[MI_SEGMENT_MAP_WSIZE + 1];  // 2KiB per TB with 64MiB segments
 
 static size_t mi_segment_map_index_of(const mi_segment_t* segment, size_t* bitidx) {
-  mi_assert_internal(_mi_ptr_segment(segment) == segment); // is it aligned on MI_SEGMENT_SIZE?
+  mi_assert_internal(_mi_ptr_segment(segment + 1) == segment); // is it aligned on MI_SEGMENT_SIZE?
   if ((uintptr_t)segment >= MI_MAX_ADDRESS) {
     *bitidx = 0;
     return MI_SEGMENT_MAP_WSIZE;
diff --git a/src/segment.c b/src/segment.c
index 2ae591fd..c743f02f 100644
--- a/src/segment.c
+++ b/src/segment.c
@@ -386,11 +386,13 @@ static void mi_segment_os_free(mi_segment_t* segment, mi_segments_tld_t* tld) {
   
   // _mi_os_free(segment, mi_segment_size(segment), /*segment->memid,*/ tld->stats);
   const size_t size = mi_segment_size(segment);
-  if (size != MI_SEGMENT_SIZE || !_mi_segment_cache_push(segment, size, segment->memid, &segment->commit_mask, &segment->decommit_mask, segment->mem_is_large, segment->mem_is_pinned, tld->os)) {
+  if (size != MI_SEGMENT_SIZE || segment->mem_align_offset != 0 || // only push regular segments on the cache
+       !_mi_segment_cache_push(segment, size, segment->memid, &segment->commit_mask, &segment->decommit_mask, segment->mem_is_large, segment->mem_is_pinned, tld->os)) 
+  {
     const size_t csize = _mi_commit_mask_committed_size(&segment->commit_mask, size);
     if (csize > 0 && !segment->mem_is_pinned) _mi_stat_decrease(&_mi_stats_main.committed, csize);
     _mi_abandoned_await_readers();  // wait until safe to free
-    _mi_arena_free(segment, mi_segment_size(segment), segment->memid, segment->mem_is_pinned /* pretend not committed to not double count decommits */, tld->os);
+    _mi_arena_free(segment, mi_segment_size(segment), segment->mem_alignment, segment->mem_align_offset, segment->memid, segment->mem_is_pinned /* pretend not committed to not double count decommits */, tld->stats);
   }
 }
 
@@ -406,7 +408,7 @@ void _mi_segment_thread_collect(mi_segments_tld_t* tld) {
 ----------------------------------------------------------- */
 
 static void mi_segment_commit_mask(mi_segment_t* segment, bool conservative, uint8_t* p, size_t size, uint8_t** start_p, size_t* full_size, mi_commit_mask_t* cm) {
-  mi_assert_internal(_mi_ptr_segment(p) == segment);
+  mi_assert_internal(_mi_ptr_segment(p + 1) == segment);
   mi_assert_internal(segment->kind != MI_SEGMENT_HUGE);
   mi_commit_mask_create_empty(cm);
   if (size == 0 || size > MI_SEGMENT_SIZE || segment->kind == MI_SEGMENT_HUGE) return;
@@ -761,17 +763,15 @@ static mi_page_t* mi_segments_page_find_and_allocate(size_t slice_count, mi_aren
 ----------------------------------------------------------- */
 
 // Allocate a segment from the OS aligned to `MI_SEGMENT_SIZE` .
-static mi_segment_t* mi_segment_init(mi_segment_t* segment, size_t required, mi_arena_id_t req_arena_id, mi_segments_tld_t* tld, mi_os_tld_t* os_tld, mi_page_t** huge_page)
+static mi_segment_t* mi_segment_init(mi_segment_t* segment, size_t required, size_t page_alignment, mi_arena_id_t req_arena_id, mi_segments_tld_t* tld, mi_os_tld_t* os_tld, mi_page_t** huge_page)
 {
   mi_assert_internal((required==0 && huge_page==NULL) || (required>0 && huge_page != NULL));
   mi_assert_internal((segment==NULL) || (segment!=NULL && required==0));
   // calculate needed sizes first
   size_t info_slices;
   size_t pre_size;
-  const size_t segment_slices = mi_segment_calculate_slices(required, &pre_size, &info_slices);
-  const size_t slice_entries = (segment_slices > MI_SLICES_PER_SEGMENT ? MI_SLICES_PER_SEGMENT : segment_slices);
-  const size_t segment_size = segment_slices * MI_SEGMENT_SLICE_SIZE;
-
+  size_t segment_slices = mi_segment_calculate_slices(required, &pre_size, &info_slices);
+  
   // Commit eagerly only if not the first N lazy segments (to reduce impact of many threads that allocate just a little)
   const bool eager_delay = (// !_mi_os_has_overcommit() &&             // never delay on overcommit systems
                             _mi_current_thread_count() > 1 &&       // do not delay for the first N threads
@@ -797,9 +797,25 @@ static mi_segment_t* mi_segment_init(mi_segment_t* segment, size_t required, mi_
     bool mem_large = (!eager_delay && (MI_SECURE==0)); // only allow large OS pages once we are no longer lazy    
     bool is_pinned = false;
     size_t memid = 0;
-    segment = (mi_segment_t*)_mi_segment_cache_pop(segment_size, &commit_mask, &decommit_mask, &mem_large, &is_pinned, &is_zero, req_arena_id, &memid, os_tld);
+    size_t align_offset = 0;
+    size_t alignment = MI_SEGMENT_SIZE;
+    size_t segment_size = segment_slices * MI_SEGMENT_SLICE_SIZE;
+
+    if (page_alignment > 0) {
+      mi_assert_internal(huge_page != NULL);
+      mi_assert_internal(page_alignment >= MI_SEGMENT_ALIGN);
+      alignment = page_alignment;
+      const size_t info_size = info_slices * MI_SEGMENT_SLICE_SIZE;
+      align_offset = _mi_align_up( info_size, MI_SEGMENT_ALIGN );
+      segment_size += _mi_align_up(align_offset - info_size, MI_SEGMENT_SLICE_SIZE);
+      segment_slices = segment_size / MI_SEGMENT_SLICE_SIZE;
+    }
+    else {
+      segment = (mi_segment_t*)_mi_segment_cache_pop(segment_size, &commit_mask, &decommit_mask, &mem_large, &is_pinned, &is_zero, req_arena_id, &memid, os_tld);
+    }
+    
     if (segment==NULL) {
-      segment = (mi_segment_t*)_mi_arena_alloc_aligned(segment_size, MI_SEGMENT_SIZE, &commit, &mem_large, &is_pinned, &is_zero, req_arena_id, &memid, os_tld);
+      segment = (mi_segment_t*)_mi_arena_alloc_aligned(segment_size, alignment, align_offset, &commit, &mem_large, &is_pinned, &is_zero, req_arena_id, &memid, os_tld);
       if (segment == NULL) return NULL;  // failed to allocate
       if (commit) {
         mi_commit_mask_create_full(&commit_mask);
@@ -818,7 +834,7 @@ static mi_segment_t* mi_segment_init(mi_segment_t* segment, size_t required, mi_
       // at least commit the info slices
       mi_assert_internal(commit_needed*MI_COMMIT_SIZE >= info_slices*MI_SEGMENT_SLICE_SIZE);
       bool ok = _mi_os_commit(segment, commit_needed*MI_COMMIT_SIZE, &is_zero, tld->stats);
-      if (!ok) return NULL; // failed to commit   
+      if (!ok) return NULL; // failed to commit 
       mi_commit_mask_set(&commit_mask, &commit_needed_mask); 
     }
     mi_track_mem_undefined(segment,commit_needed);
@@ -826,6 +842,8 @@ static mi_segment_t* mi_segment_init(mi_segment_t* segment, size_t required, mi_
     segment->mem_is_pinned = is_pinned;
     segment->mem_is_large = mem_large;
     segment->mem_is_committed = mi_commit_mask_is_full(&commit_mask);
+    segment->mem_alignment = alignment;
+    segment->mem_align_offset = align_offset;
     mi_segments_track_size((long)(segment_size), tld);
     _mi_segment_map_allocated_at(segment);
   }
@@ -862,6 +880,7 @@ static mi_segment_t* mi_segment_init(mi_segment_t* segment, size_t required, mi_
   
 
   // initialize segment info
+  const size_t slice_entries = (segment_slices > MI_SLICES_PER_SEGMENT ? MI_SLICES_PER_SEGMENT : segment_slices);
   segment->segment_slices = segment_slices;
   segment->segment_info_slices = info_slices;
   segment->thread_id = _mi_thread_id();
@@ -912,8 +931,8 @@ static mi_segment_t* mi_segment_init(mi_segment_t* segment, size_t required, mi_
 
 
 // Allocate a segment from the OS aligned to `MI_SEGMENT_SIZE` .
-static mi_segment_t* mi_segment_alloc(size_t required, mi_arena_id_t req_arena_id, mi_segments_tld_t* tld, mi_os_tld_t* os_tld, mi_page_t** huge_page) {
-  return mi_segment_init(NULL, required, req_arena_id, tld, os_tld, huge_page);
+static mi_segment_t* mi_segment_alloc(size_t required, size_t page_alignment, mi_arena_id_t req_arena_id, mi_segments_tld_t* tld, mi_os_tld_t* os_tld, mi_page_t** huge_page) {
+  return mi_segment_init(NULL, required, page_alignment, req_arena_id, tld, os_tld, huge_page);
 }
 
 
@@ -1450,7 +1469,7 @@ static mi_segment_t* mi_segment_reclaim_or_alloc(mi_heap_t* heap, size_t needed_
     return segment;
   }
   // 2. otherwise allocate a fresh segment
-  return mi_segment_alloc(0, heap->arena_id, tld, os_tld, NULL);  
+  return mi_segment_alloc(0, 0, heap->arena_id, tld, os_tld, NULL);  
 }
 
 
@@ -1490,14 +1509,28 @@ static mi_page_t* mi_segments_page_alloc(mi_heap_t* heap, mi_page_kind_t page_ki
    Huge page allocation
 ----------------------------------------------------------- */
 
-static mi_page_t* mi_segment_huge_page_alloc(size_t size, mi_arena_id_t req_arena_id, mi_segments_tld_t* tld, mi_os_tld_t* os_tld)
+static mi_page_t* mi_segment_huge_page_alloc(size_t size, size_t page_alignment, mi_arena_id_t req_arena_id, mi_segments_tld_t* tld, mi_os_tld_t* os_tld)
 {
   mi_page_t* page = NULL;
-  mi_segment_t* segment = mi_segment_alloc(size,req_arena_id,tld,os_tld,&page);
+  mi_segment_t* segment = mi_segment_alloc(size,page_alignment,req_arena_id,tld,os_tld,&page);
   if (segment == NULL || page==NULL) return NULL;
   mi_assert_internal(segment->used==1);
   mi_assert_internal(mi_page_block_size(page) >= size);  
   segment->thread_id = 0; // huge segments are immediately abandoned
+  #if MI_DEBUG > 3
+  if (page_alignment > 0) {
+    size_t psize;
+    void* p = _mi_segment_page_start(segment, page, &psize);
+    void* aligned_p = (void*)_mi_align_up((uintptr_t)p, page_alignment);
+    mi_assert_internal(page_alignment == 0 || _mi_is_aligned(aligned_p, page_alignment));
+    mi_assert_internal(page_alignment == 0 || psize - ((uint8_t*)aligned_p - (uint8_t*)p) >= size);
+  }
+  #endif
+  // for huge pages we initialize the xblock_size as we may
+  // overallocate to accommodate large alignments.
+  size_t psize;
+  _mi_segment_page_start(segment, page, &psize);
+  page->xblock_size = (psize > MI_HUGE_BLOCK_SIZE ? MI_HUGE_BLOCK_SIZE : (uint32_t)psize);
   return page;
 }
 
@@ -1531,9 +1564,17 @@ void _mi_segment_huge_page_free(mi_segment_t* segment, mi_page_t* page, mi_block
 /* -----------------------------------------------------------
    Page allocation and free
 ----------------------------------------------------------- */
-mi_page_t* _mi_segment_page_alloc(mi_heap_t* heap, size_t block_size, mi_segments_tld_t* tld, mi_os_tld_t* os_tld) {
+mi_page_t* _mi_segment_page_alloc(mi_heap_t* heap, size_t block_size, size_t page_alignment, mi_segments_tld_t* tld, mi_os_tld_t* os_tld) {
   mi_page_t* page;
-  if (block_size <= MI_SMALL_OBJ_SIZE_MAX) {
+  if mi_unlikely(page_alignment > MI_ALIGNMENT_MAX) {
+    mi_assert_internal(_mi_is_power_of_two(page_alignment));
+    mi_assert_internal(page_alignment >= MI_SEGMENT_SIZE);
+    if (page_alignment < MI_SEGMENT_SIZE) {
+      page_alignment = MI_SEGMENT_SIZE;      
+    }
+    page = mi_segment_huge_page_alloc(block_size,page_alignment,heap->arena_id,tld,os_tld);
+  }
+  else if (block_size <= MI_SMALL_OBJ_SIZE_MAX) {
     page = mi_segments_page_alloc(heap,MI_PAGE_SMALL,block_size,block_size,tld,os_tld);
   }
   else if (block_size <= MI_MEDIUM_OBJ_SIZE_MAX) {
@@ -1543,7 +1584,7 @@ mi_page_t* _mi_segment_page_alloc(mi_heap_t* heap, size_t block_size, mi_segment
     page = mi_segments_page_alloc(heap,MI_PAGE_LARGE,block_size,block_size,tld, os_tld);
   }
   else {
-    page = mi_segment_huge_page_alloc(block_size,heap->arena_id,tld,os_tld);    
+    page = mi_segment_huge_page_alloc(block_size,page_alignment,heap->arena_id,tld,os_tld);    
   }
   mi_assert_internal(page == NULL || _mi_heap_memid_is_suitable(heap, _mi_page_segment(page)->memid));
   mi_assert_expensive(page == NULL || mi_segment_is_valid(_mi_page_segment(page),tld));
diff --git a/test/test-api.c b/test/test-api.c
index 3c2ef7e4..312b3f1b 100644
--- a/test/test-api.c
+++ b/test/test-api.c
@@ -161,10 +161,29 @@ int main(void) {
     result = ok;
   };
   CHECK_BODY("malloc-aligned7") {
-    void* p = mi_malloc_aligned(1024,MI_ALIGNMENT_MAX); mi_free(p);
-    };
+    void* p = mi_malloc_aligned(1024,MI_ALIGNMENT_MAX);
+    mi_free(p);
+    result = ((uintptr_t)p % MI_ALIGNMENT_MAX) == 0;
+  };
   CHECK_BODY("malloc-aligned8") {
-    void* p = mi_malloc_aligned(1024,2*MI_ALIGNMENT_MAX); mi_free(p);
+    bool ok = true;
+    for (int i = 0; i < 5 && ok; i++) {
+      int n = (1 << i);
+      void* p = mi_malloc_aligned(1024, n * MI_ALIGNMENT_MAX);
+      ok = ((uintptr_t)p % (n*MI_ALIGNMENT_MAX)) == 0;
+      mi_free(p);
+    }
+    result = ok;
+  };
+  CHECK_BODY("malloc-aligned9") {
+    bool ok = true;
+    for (int i = 0; i < 5 && ok; i++) {
+      int n = (1 << i);
+      void* p = mi_malloc_aligned( 2*n*MI_ALIGNMENT_MAX, n*MI_ALIGNMENT_MAX);
+      ok = ((uintptr_t)p % (n*MI_ALIGNMENT_MAX)) == 0;
+      mi_free(p);
+    }
+    result = ok;
   };
   CHECK_BODY("malloc-aligned-at1") {
     void* p = mi_malloc_aligned_at(48,32,0); result = (p != NULL && ((uintptr_t)(p) + 0) % 32 == 0); mi_free(p);