merge from dev (bitmap split)

CMakeLists.txt

@@ -29,6 +29,7 @@ set(mi_sources
     src/stats.c
     src/random.c
     src/os.c
+    src/bitmap.c
     src/arena.c
     src/segment.c
     src/page.c

ide/vs2017/mimalloc-override.vcxproj

@@ -215,6 +215,7 @@
     <ClInclude Include="..\..\include\mimalloc-new-delete.h" />
     <ClInclude Include="..\..\include\mimalloc-override.h" />
     <ClInclude Include="..\..\include\mimalloc-types.h" />
+    <ClInclude Include="..\..\src\bitmap.h" />
   </ItemGroup>
   <ItemGroup>
     <ClCompile Include="..\..\src\alloc-aligned.c">
@@ -232,6 +233,7 @@
     <ClCompile Include="..\..\src\alloc-posix.c" />
     <ClCompile Include="..\..\src\alloc.c" />
     <ClCompile Include="..\..\src\arena.c" />
+    <ClCompile Include="..\..\src\bitmap.c" />
     <ClCompile Include="..\..\src\heap.c" />
     <ClCompile Include="..\..\src\init.c" />
     <ClCompile Include="..\..\src\options.c" />

ide/vs2017/mimalloc-override.vcxproj.filters

@@ -29,6 +29,9 @@
     <ClInclude Include="..\..\include\mimalloc-new-delete.h">
       <Filter>Header Files</Filter>
     </ClInclude>
+    <ClInclude Include="..\..\src\bitmap.h">
+      <Filter>Header Files</Filter>
+    </ClInclude>
   </ItemGroup>
   <ItemGroup>
     <ClCompile Include="..\..\src\alloc.c">
@@ -73,5 +76,8 @@
     <ClCompile Include="..\..\src\random.c">
       <Filter>Source Files</Filter>
     </ClCompile>
+    <ClCompile Include="..\..\src\bitmap.c">
+      <Filter>Source Files</Filter>
+    </ClCompile>
   </ItemGroup>
 </Project>

ide/vs2017/mimalloc.vcxproj

@@ -230,6 +230,7 @@
     <ClCompile Include="..\..\src\alloc-posix.c" />
     <ClCompile Include="..\..\src\alloc.c" />
     <ClCompile Include="..\..\src\arena.c" />
+    <ClCompile Include="..\..\src\bitmap.c" />
     <ClCompile Include="..\..\src\heap.c" />
     <ClCompile Include="..\..\src\init.c" />
     <ClCompile Include="..\..\src\options.c" />
@@ -252,6 +253,7 @@
     <ClInclude Include="$(ProjectDir)..\..\include\mimalloc-override.h" />
     <ClInclude Include="$(ProjectDir)..\..\include\mimalloc-types.h" />
     <ClInclude Include="..\..\include\mimalloc-new-delete.h" />
+    <ClInclude Include="..\..\src\bitmap.h" />
   </ItemGroup>
   <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
   <ImportGroup Label="ExtensionTargets">

ide/vs2017/mimalloc.vcxproj.filters

@@ -59,6 +59,9 @@
     <ClCompile Include="..\..\src\random.c">
       <Filter>Source Files</Filter>
     </ClCompile>
+    <ClCompile Include="..\..\src\bitmap.c">
+      <Filter>Source Files</Filter>
+    </ClCompile>
   </ItemGroup>
   <ItemGroup>
     <ClInclude Include="$(ProjectDir)..\..\include\mimalloc.h">
@@ -79,5 +82,8 @@
     <ClInclude Include="..\..\include\mimalloc-new-delete.h">
       <Filter>Header Files</Filter>
     </ClInclude>
+    <ClInclude Include="..\..\src\bitmap.h">
+      <Filter>Header Files</Filter>
+    </ClInclude>
   </ItemGroup>
 </Project>

ide/vs2019/mimalloc-override.vcxproj

@@ -215,6 +215,7 @@
     <ClInclude Include="..\..\include\mimalloc-new-delete.h" />
     <ClInclude Include="..\..\include\mimalloc-override.h" />
     <ClInclude Include="..\..\include\mimalloc-types.h" />
+    <ClInclude Include="..\..\src\bitmap.h" />
   </ItemGroup>
   <ItemGroup>
     <ClCompile Include="..\..\src\alloc-aligned.c">
@@ -232,9 +233,7 @@
     <ClCompile Include="..\..\src\alloc-posix.c" />
     <ClCompile Include="..\..\src\alloc.c" />
     <ClCompile Include="..\..\src\arena.c" />
-    <ClCompile Include="..\..\src\bitmap.inc.c">
+    <ClCompile Include="..\..\src\bitmap.c" />
-      <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">true</ExcludedFromBuild>
-    </ClCompile>
     <ClCompile Include="..\..\src\heap.c" />
     <ClCompile Include="..\..\src\init.c" />
     <ClCompile Include="..\..\src\options.c" />

ide/vs2019/mimalloc-override.vcxproj.filters

@@ -37,15 +37,15 @@
     <ClCompile Include="..\..\src\arena.c">
       <Filter>Source Files</Filter>
     </ClCompile>
-    <ClCompile Include="..\..\src\bitmap.inc.c">
-      <Filter>Source Files</Filter>
-    </ClCompile>
     <ClCompile Include="..\..\src\random.c">
       <Filter>Source Files</Filter>
     </ClCompile>
     <ClCompile Include="..\..\src\options.c">
       <Filter>Source Files</Filter>
     </ClCompile>
+    <ClCompile Include="..\..\src\bitmap.c">
+      <Filter>Source Files</Filter>
+    </ClCompile>
   </ItemGroup>
   <ItemGroup>
     <ClInclude Include="$(ProjectDir)..\..\include\mimalloc.h">
@@ -66,6 +66,9 @@
     <ClInclude Include="..\..\include\mimalloc-types.h">
       <Filter>Header Files</Filter>
     </ClInclude>
+    <ClInclude Include="..\..\src\bitmap.h">
+      <Filter>Source Files</Filter>
+    </ClInclude>
   </ItemGroup>
   <ItemGroup>
     <Filter Include="Header Files">

ide/vs2019/mimalloc.vcxproj

@@ -220,8 +220,8 @@
     <ClCompile Include="..\..\src\alloc-posix.c" />
     <ClCompile Include="..\..\src\alloc.c" />
     <ClCompile Include="..\..\src\arena.c" />
-    <ClCompile Include="..\..\src\bitmap.inc.c">
+    <ClCompile Include="..\..\src\bitmap.c">
-      <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">true</ExcludedFromBuild>
+      <ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">false</ExcludedFromBuild>
     </ClCompile>
     <ClCompile Include="..\..\src\heap.c" />
     <ClCompile Include="..\..\src\init.c" />
@@ -245,6 +245,7 @@
     <ClInclude Include="$(ProjectDir)..\..\include\mimalloc-override.h" />
     <ClInclude Include="$(ProjectDir)..\..\include\mimalloc-types.h" />
     <ClInclude Include="..\..\include\mimalloc-new-delete.h" />
+    <ClInclude Include="..\..\src\bitmap.h" />
   </ItemGroup>
   <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
   <ImportGroup Label="ExtensionTargets">

ide/vs2019/mimalloc.vcxproj.filters

@@ -43,10 +43,10 @@
     <ClCompile Include="..\..\src\arena.c">
       <Filter>Source Files</Filter>
     </ClCompile>
-    <ClCompile Include="..\..\src\bitmap.inc.c">
+    <ClCompile Include="..\..\src\random.c">
       <Filter>Source Files</Filter>
     </ClCompile>
-    <ClCompile Include="..\..\src\random.c">
+    <ClCompile Include="..\..\src\bitmap.c">
       <Filter>Source Files</Filter>
     </ClCompile>
   </ItemGroup>
@@ -69,6 +69,9 @@
     <ClInclude Include="$(ProjectDir)..\..\include\mimalloc-types.h">
       <Filter>Header Files</Filter>
     </ClInclude>
+    <ClInclude Include="..\..\src\bitmap.h">
+      <Filter>Source Files</Filter>
+    </ClInclude>
   </ItemGroup>
   <ItemGroup>
     <Filter Include="Header Files">

include/mimalloc-internal.h

@@ -109,7 +109,6 @@ void       _mi_page_reclaim(mi_heap_t* heap, mi_page_t* page);   // callback fro
 
 size_t     _mi_bin_size(uint8_t bin);           // for stats
 uint8_t    _mi_bin(size_t size);                // for stats
-uint8_t    _mi_bsr(uintptr_t x);                // bit-scan-right, used on BSD in "os.c"
 
 // "heap.c"
 void       _mi_heap_destroy_pages(mi_heap_t* heap);
@@ -859,5 +858,108 @@ static inline uintptr_t _mi_thread_id(void) mi_attr_noexcept {
 }
 #endif
 
+// -----------------------------------------------------------------------
+// Count bits: trailing or leading zeros (with MI_INTPTR_BITS on all zero)
+// -----------------------------------------------------------------------
+
+#if defined(__GNUC__)
+
+#include <limits.h>       // LONG_MAX
+#define MI_HAVE_FAST_BITSCAN
+static inline size_t mi_clz(uintptr_t x) {
+  if (x==0) return MI_INTPTR_BITS;
+#if (INTPTR_MAX == LONG_MAX)
+  return __builtin_clzl(x);
+#else
+  return __builtin_clzll(x);
+#endif
+}
+static inline size_t mi_ctz(uintptr_t x) {
+  if (x==0) return MI_INTPTR_BITS;
+#if (INTPTR_MAX == LONG_MAX)
+  return __builtin_ctzl(x);
+#else
+  return __builtin_ctzll(x);
+#endif
+}
+
+#elif defined(_MSC_VER) 
+
+#include <limits.h>       // LONG_MAX
+#define MI_HAVE_FAST_BITSCAN
+static inline size_t mi_clz(uintptr_t x) {
+  if (x==0) return MI_INTPTR_BITS;
+  unsigned long idx;
+#if (INTPTR_MAX == LONG_MAX)
+  _BitScanReverse(&idx, x);
+#else
+  _BitScanReverse64(&idx, x);
+#endif  
+  return ((MI_INTPTR_BITS - 1) - idx);
+}
+static inline size_t mi_ctz(uintptr_t x) {
+  if (x==0) return MI_INTPTR_BITS;
+  unsigned long idx;
+#if (INTPTR_MAX == LONG_MAX)
+  _BitScanForward(&idx, x);
+#else
+  _BitScanForward64(&idx, x);
+#endif  
+  return idx;
+}
+
+#else
+static inline size_t mi_ctz32(uint32_t x) {
+  // de Bruijn multiplication, see <http://supertech.csail.mit.edu/papers/debruijn.pdf>
+  static const unsigned char debruijn[32] = {
+    0, 1, 28, 2, 29, 14, 24, 3, 30, 22, 20, 15, 25, 17, 4, 8,
+    31, 27, 13, 23, 21, 19, 16, 7, 26, 12, 18, 6, 11, 5, 10, 9
+  };
+  if (x==0) return 32;
+  return debruijn[((x & -(int32_t)x) * 0x077CB531UL) >> 27];
+}
+static inline size_t mi_clz32(uint32_t x) {
+  // de Bruijn multiplication, see <http://supertech.csail.mit.edu/papers/debruijn.pdf>
+  static const uint8_t debruijn[32] = {
+    31, 22, 30, 21, 18, 10, 29, 2, 20, 17, 15, 13, 9, 6, 28, 1,
+    23, 19, 11, 3, 16, 14, 7, 24, 12, 4, 8, 25, 5, 26, 27, 0
+  };
+  if (x==0) return 32;
+  x |= x >> 1;
+  x |= x >> 2;
+  x |= x >> 4;
+  x |= x >> 8;
+  x |= x >> 16;
+  return debruijn[(uint32_t)(x * 0x07C4ACDDUL) >> 27];
+}
+
+static inline size_t mi_clz(uintptr_t x) {
+  if (x==0) return MI_INTPTR_BITS;  
+#if (MI_INTPTR_BITS <= 32)
+  return mi_clz32((uint32_t)x);
+#else
+  size_t count = mi_clz32((uint32_t)(x >> 32));
+  if (count < 32) return count;
+  return (32 + mi_clz32((uint32_t)x));
+#endif
+}
+static inline size_t mi_ctz(uintptr_t x) {
+  if (x==0) return MI_INTPTR_BITS;
+#if (MI_INTPTR_BITS <= 32)
+  return mi_ctz32((uint32_t)x);
+#else
+  size_t count = mi_ctz32((uint32_t)x);
+  if (count < 32) return count;
+  return (32 + mi_ctz32((uint32_t)(x>>32)));
+#endif
+}
+
+#endif
+
+// "bit scan reverse": Return index of the highest bit (or MI_INTPTR_BITS if `x` is zero)
+static inline size_t mi_bsr(uintptr_t x) {
+  return (x==0 ? MI_INTPTR_BITS : MI_INTPTR_BITS - 1 - mi_clz(x));
+}
+
 
 #endif

src/arena.c

@@ -28,7 +28,7 @@ The arena allocation needs to be thread safe and we use an atomic bitmap to allo
 #include <string.h>  // memset
 #include <errno.h> // ENOMEM
 
-#include "bitmap.inc.c"  // atomic bitmap
+#include "bitmap.h"  // atomic bitmap
 
 
 // os.c
@@ -105,7 +105,7 @@ static size_t mi_block_count_of_size(size_t size) {
 static bool mi_arena_alloc(mi_arena_t* arena, size_t blocks, mi_bitmap_index_t* bitmap_idx)
 {
   size_t idx = mi_atomic_load_acquire(&arena->search_idx);  // start from last search
-  if (mi_bitmap_try_find_from_claim_across(arena->blocks_inuse, arena->field_count, idx, blocks, bitmap_idx)) {
+  if (_mi_bitmap_try_find_from_claim_across(arena->blocks_inuse, arena->field_count, idx, blocks, bitmap_idx)) {
     mi_atomic_store_release(&arena->search_idx, idx);  // start search from here next time
     return true;
   };
@@ -154,11 +154,11 @@ static void* mi_cache_pop(int numa_node, size_t size, size_t alignment, bool com
   mi_bitmap_index_t bitidx = 0;
   bool claimed = false;
   if (*large) {  // large allowed?
-    claimed = mi_bitmap_try_find_from_claim(cache_available_large, MI_CACHE_FIELDS, start_field, 1, &bitidx);
+    claimed = _mi_bitmap_try_find_from_claim(cache_available_large, MI_CACHE_FIELDS, start_field, 1, &bitidx);
     if (claimed) *large = true;
   }
   if (!claimed) {
-    claimed = mi_bitmap_try_find_from_claim(cache_available, MI_CACHE_FIELDS, start_field, 1, &bitidx);
+    claimed = _mi_bitmap_try_find_from_claim(cache_available, MI_CACHE_FIELDS, start_field, 1, &bitidx);
     if (claimed) *large = false;
   }
 
@@ -189,8 +189,8 @@ static void* mi_cache_pop(int numa_node, size_t size, size_t alignment, bool com
   *commit_mask = cmask;
   
   // mark the slot as free again
-  mi_assert_internal(mi_bitmap_is_claimed(cache_inuse, MI_CACHE_FIELDS, 1, bitidx));
+  mi_assert_internal(_mi_bitmap_is_claimed(cache_inuse, MI_CACHE_FIELDS, 1, bitidx));
-  mi_bitmap_unclaim(cache_inuse, MI_CACHE_FIELDS, 1, bitidx);
+  _mi_bitmap_unclaim(cache_inuse, MI_CACHE_FIELDS, 1, bitidx);
   return p;
 }
 
@@ -231,19 +231,19 @@ static void mi_cache_purge(mi_os_tld_t* tld) {
       // seems expired, first claim it from available
       purged++;
       mi_bitmap_index_t bitidx = mi_bitmap_index_create_from_bit(idx);
-      if (mi_bitmap_claim(cache_available, MI_CACHE_FIELDS, 1, bitidx, NULL)) {
+      if (_mi_bitmap_claim(cache_available, MI_CACHE_FIELDS, 1, bitidx, NULL)) {
         // was available, we claimed it
         expire = mi_atomic_loadi64_acquire(&slot->expire);
         if (expire != 0 && now >= expire) {  // safe read
           // still expired, decommit it
           mi_atomic_storei64_relaxed(&slot->expire,(mi_msecs_t)0);
-          mi_assert_internal(!mi_commit_mask_is_empty(slot->commit_mask) && mi_bitmap_is_claimed(cache_available_large, MI_CACHE_FIELDS, 1, bitidx));
+          mi_assert_internal(!mi_commit_mask_is_empty(slot->commit_mask) && _mi_bitmap_is_claimed(cache_available_large, MI_CACHE_FIELDS, 1, bitidx));
           _mi_abandoned_await_readers();  // wait until safe to decommit
           // decommit committed parts
           mi_commit_mask_decommit(&slot->commit_mask, slot->p, MI_SEGMENT_SIZE, tld->stats);
           //_mi_os_decommit(slot->p, MI_SEGMENT_SIZE, tld->stats);
         }
-        mi_bitmap_unclaim(cache_available, MI_CACHE_FIELDS, 1, bitidx); // make it available again for a pop
+        _mi_bitmap_unclaim(cache_available, MI_CACHE_FIELDS, 1, bitidx); // make it available again for a pop
       }
       if (purged > 4) break;  // bound to no more than 4 purge tries per push
     }
@@ -268,11 +268,11 @@ static bool mi_cache_push(void* start, size_t size, size_t memid, mi_commit_mask
 
   // find an available slot
   mi_bitmap_index_t bitidx;
-  bool claimed = mi_bitmap_try_find_from_claim(cache_inuse, MI_CACHE_FIELDS, start_field, 1, &bitidx);
+  bool claimed = _mi_bitmap_try_find_from_claim(cache_inuse, MI_CACHE_FIELDS, start_field, 1, &bitidx);
   if (!claimed) return false;
 
-  mi_assert_internal(mi_bitmap_is_claimed(cache_available, MI_CACHE_FIELDS, 1, bitidx));
+  mi_assert_internal(_mi_bitmap_is_claimed(cache_available, MI_CACHE_FIELDS, 1, bitidx));
-  mi_assert_internal(mi_bitmap_is_claimed(cache_available_large, MI_CACHE_FIELDS, 1, bitidx));
+  mi_assert_internal(_mi_bitmap_is_claimed(cache_available_large, MI_CACHE_FIELDS, 1, bitidx));
 
   // set the slot
   mi_cache_slot_t* slot = &cache[mi_bitmap_index_bit(bitidx)];
@@ -292,7 +292,7 @@ static bool mi_cache_push(void* start, size_t size, size_t memid, mi_commit_mask
   }
   
   // make it available
-  mi_bitmap_unclaim((is_large ? cache_available_large : cache_available), MI_CACHE_FIELDS, 1, bitidx);
+  _mi_bitmap_unclaim((is_large ? cache_available_large : cache_available), MI_CACHE_FIELDS, 1, bitidx);
   return true;
 }
 
@@ -310,7 +310,7 @@ static void* mi_arena_alloc_from(mi_arena_t* arena, size_t arena_index, size_t n
   // claimed it! set the dirty bits (todo: no need for an atomic op here?)
   void* p  = arena->start + (mi_bitmap_index_bit(bitmap_index)*MI_ARENA_BLOCK_SIZE);
   *memid   = mi_arena_id_create(arena_index, bitmap_index);
-  *is_zero = mi_bitmap_claim_across(arena->blocks_dirty, arena->field_count, needed_bcount, bitmap_index, NULL);
+  *is_zero = _mi_bitmap_claim_across(arena->blocks_dirty, arena->field_count, needed_bcount, bitmap_index, NULL);
   *large   = arena->is_large;
   if (arena->is_committed) {
     // always committed
@@ -319,7 +319,7 @@ static void* mi_arena_alloc_from(mi_arena_t* arena, size_t arena_index, size_t n
   else if (*commit) {
     // arena not committed as a whole, but commit requested: ensure commit now
     bool any_uncommitted;
-    mi_bitmap_claim_across(arena->blocks_committed, arena->field_count, needed_bcount, bitmap_index, &any_uncommitted);
+    _mi_bitmap_claim_across(arena->blocks_committed, arena->field_count, needed_bcount, bitmap_index, &any_uncommitted);
     if (any_uncommitted) {
       bool commit_zero;
       _mi_os_commit(p, needed_bcount * MI_ARENA_BLOCK_SIZE, &commit_zero, tld->stats);
@@ -328,7 +328,7 @@ static void* mi_arena_alloc_from(mi_arena_t* arena, size_t arena_index, size_t n
   }
   else {
     // no need to commit, but check if already fully committed
-    *commit = mi_bitmap_is_claimed_across(arena->blocks_committed, arena->field_count, needed_bcount, bitmap_index);
+    *commit = _mi_bitmap_is_claimed_across(arena->blocks_committed, arena->field_count, needed_bcount, bitmap_index);
   }
   return p;
 }
@@ -443,7 +443,7 @@ void _mi_arena_free(void* p, size_t size, size_t memid, mi_commit_mask_t commit_
       return;
     }
     const size_t blocks = mi_block_count_of_size(size);
-    bool ones = mi_bitmap_unclaim_across(arena->blocks_inuse, arena->field_count, blocks, bitmap_idx);
+    bool ones = _mi_bitmap_unclaim_across(arena->blocks_inuse, arena->field_count, blocks, bitmap_idx);
     if (!ones) {
       _mi_error_message(EAGAIN, "trying to free an already freed block: %p, size %zu\n", p, size);
       return;
@@ -495,7 +495,7 @@ bool mi_manage_os_memory(void* start, size_t size, bool is_committed, bool is_la
   if (post > 0) {
     // don't use leftover bits at the end
     mi_bitmap_index_t postidx = mi_bitmap_index_create(fields - 1, MI_BITMAP_FIELD_BITS - post);
-    mi_bitmap_claim(arena->blocks_inuse, fields, post, postidx, NULL);
+    _mi_bitmap_claim(arena->blocks_inuse, fields, post, postidx, NULL);
   }
 
   mi_arena_add(arena);

src/bitmap.c

@@ -1,147 +1,46 @@
 /* ----------------------------------------------------------------------------
-Copyright (c) 2019, Microsoft Research, Daan Leijen
+Copyright (c) 2019,2020 Microsoft Research, Daan Leijen
 This is free software; you can redistribute it and/or modify it under the
 terms of the MIT license. A copy of the license can be found in the file
 "LICENSE" at the root of this distribution.
 -----------------------------------------------------------------------------*/
 
 /* ----------------------------------------------------------------------------
-This file is meant to be included in other files for efficiency.
+Concurrent bitmap that can set/reset sequences of bits atomically,
-It implements a bitmap that can set/reset sequences of bits atomically
+represeted as an array of fields where each field is a machine word (`uintptr_t`)
-and is used to concurrently claim memory ranges.
 
-A bitmap is an array of fields where each field is a machine word (`uintptr_t`)
+There are two api's; the standard one cannot have sequences that cross
+between the bitmap fields (and a sequence must be <= MI_BITMAP_FIELD_BITS).
+(this is used in region allocation)
 
-A current limitation is that the bit sequences cannot cross fields
+The `_across` postfixed functions do allow sequences that can cross over
-and that the sequence must be smaller or equal to the bits in a field.
+between the fields. (This is used in arena allocation)
 ---------------------------------------------------------------------------- */
-#pragma once
-#ifndef MI_BITMAP_C
-#define MI_BITMAP_C
 
 #include "mimalloc.h"
 #include "mimalloc-internal.h"
+#include "bitmap.h"
 
 /* -----------------------------------------------------------
   Bitmap definition
 ----------------------------------------------------------- */
 
-#define MI_BITMAP_FIELD_BITS   (8*MI_INTPTR_SIZE)
-#define MI_BITMAP_FIELD_FULL   (~((uintptr_t)0))   // all bits set
-
-// An atomic bitmap of `uintptr_t` fields
-typedef _Atomic(uintptr_t)  mi_bitmap_field_t;
-typedef mi_bitmap_field_t*           mi_bitmap_t;
-
-// A bitmap index is the index of the bit in a bitmap.
-typedef size_t mi_bitmap_index_t;
-
-// Create a bit index.
-static inline mi_bitmap_index_t mi_bitmap_index_create(size_t idx, size_t bitidx) {
-  mi_assert_internal(bitidx < MI_BITMAP_FIELD_BITS);
-  return (idx*MI_BITMAP_FIELD_BITS) + bitidx;
-}
-
-// Create a bit index.
-static inline mi_bitmap_index_t mi_bitmap_index_create_from_bit(size_t full_bitidx) {  
-  return mi_bitmap_index_create(full_bitidx / MI_BITMAP_FIELD_BITS, full_bitidx % MI_BITMAP_FIELD_BITS);
-}
-
-// Get the field index from a bit index.
-static inline size_t mi_bitmap_index_field(mi_bitmap_index_t bitmap_idx) {
-  return (bitmap_idx / MI_BITMAP_FIELD_BITS);
-}
-
-// Get the bit index in a bitmap field
-static inline size_t mi_bitmap_index_bit_in_field(mi_bitmap_index_t bitmap_idx) {
-  return (bitmap_idx % MI_BITMAP_FIELD_BITS);
-}
-
-// Get the full bit index
-static inline size_t mi_bitmap_index_bit(mi_bitmap_index_t bitmap_idx) {
-  return bitmap_idx;
-}
-
-
 // The bit mask for a given number of blocks at a specified bit index.
 static inline uintptr_t mi_bitmap_mask_(size_t count, size_t bitidx) {
   mi_assert_internal(count + bitidx <= MI_BITMAP_FIELD_BITS);
   mi_assert_internal(count > 0);
-  //if (count >= MI_BITMAP_FIELD_BITS) return MI_BITMAP_FIELD_FULL;
+  if (count >= MI_BITMAP_FIELD_BITS) return MI_BITMAP_FIELD_FULL;
-  //if (count == 0) return 0;
   return ((((uintptr_t)1 << count) - 1) << bitidx);
 }
 
 
-/* -----------------------------------------------------------
-  Use bit scan forward/reverse to quickly find the first zero bit if it is available
------------------------------------------------------------ */
-#if defined(_MSC_VER)
-#define MI_HAVE_BITSCAN
-#include <intrin.h>
-#ifndef MI_64
-#if MI_INTPTR_SIZE==8
-#define MI_64(f) f##64
-#else
-#define MI_64(f) f
-#endif
-#endif
-
-static inline size_t mi_bsf(uintptr_t x) {
-  if (x==0) return 8*MI_INTPTR_SIZE;
-  DWORD idx;
-  MI_64(_BitScanForward)(&idx, x);
-  return idx;
-}
-static inline size_t mi_bsr(uintptr_t x) {
-  if (x==0) return 8*MI_INTPTR_SIZE;
-  DWORD idx;
-  MI_64(_BitScanReverse)(&idx, x);
-  return idx;
-}
-#elif defined(__GNUC__) || defined(__clang__)
-#include <limits.h> // LONG_MAX
-#define MI_HAVE_BITSCAN
-#if (INTPTR_MAX == LONG_MAX)
-# define MI_L(x)  x##l
-#else
-# define MI_L(x)  x##ll
-#endif
-static inline size_t mi_bsf(uintptr_t x) {
-  return (x==0 ? 8*MI_INTPTR_SIZE : MI_L(__builtin_ctz)(x));
-}
-static inline size_t mi_bsr(uintptr_t x) {
-  return (x==0 ? 8*MI_INTPTR_SIZE : (8*MI_INTPTR_SIZE - 1) - MI_L(__builtin_clz)(x));
-}
-#endif
-
 /* -----------------------------------------------------------
   Claim a bit sequence atomically
 ----------------------------------------------------------- */
 
-// Try to atomically claim a sequence of `count` bits at in `idx`
-// in the bitmap field. Returns `true` on success.
-static inline bool mi_bitmap_try_claim_field(mi_bitmap_t bitmap, size_t bitmap_fields, const size_t count, mi_bitmap_index_t bitmap_idx) {
-  const size_t idx = mi_bitmap_index_field(bitmap_idx);
-  const size_t bitidx = mi_bitmap_index_bit_in_field(bitmap_idx);
-  const uintptr_t mask = mi_bitmap_mask_(count, bitidx);
-  mi_assert_internal(bitmap_fields > idx); UNUSED(bitmap_fields);
-  mi_assert_internal(bitidx + count <= MI_BITMAP_FIELD_BITS);
-
-  uintptr_t field = mi_atomic_load_relaxed(&bitmap[idx]);
-  if ((field & mask) == 0) { // free?
-    if (mi_atomic_cas_strong_acq_rel(&bitmap[idx], &field, (field|mask))) {
-      // claimed!
-      return true;
-    }
-  }
-  return false;
-}
-
-
 // Try to atomically claim a sequence of `count` bits in a single
 // field at `idx` in `bitmap`. Returns `true` on success.
-static inline bool mi_bitmap_try_find_claim_field(mi_bitmap_t bitmap, size_t idx, const size_t count, mi_bitmap_index_t* bitmap_idx)
+bool _mi_bitmap_try_find_claim_field(mi_bitmap_t bitmap, size_t idx, const size_t count, mi_bitmap_index_t* bitmap_idx)
 {
   mi_assert_internal(bitmap_idx != NULL);
   mi_assert_internal(count <= MI_BITMAP_FIELD_BITS);
@@ -154,8 +53,8 @@ static inline bool mi_bitmap_try_find_claim_field(mi_bitmap_t bitmap, size_t idx
   const uintptr_t mask = mi_bitmap_mask_(count, 0);
   const size_t    bitidx_max = MI_BITMAP_FIELD_BITS - count;
 
-#ifdef MI_HAVE_BITSCAN
+#ifdef MI_HAVE_FAST_BITSCAN
-  size_t bitidx = mi_bsf(~map);    // quickly find the first zero bit if possible
+  size_t bitidx = mi_ctz(~map);    // quickly find the first zero bit if possible
 #else
   size_t bitidx = 0;               // otherwise start at 0
 #endif
@@ -163,7 +62,8 @@ static inline bool mi_bitmap_try_find_claim_field(mi_bitmap_t bitmap, size_t idx
 
   // scan linearly for a free range of zero bits
   while (bitidx <= bitidx_max) {
-    if ((map & m) == 0) {  // are the mask bits free at bitidx?
+    const uintptr_t mapm = map & m;
+    if (mapm == 0) {  // are the mask bits free at bitidx?
       mi_assert_internal((m >> bitidx) == mask); // no overflow?
       const uintptr_t newmap = map | m;
       mi_assert_internal((newmap^map) >> bitidx == mask);
@@ -179,8 +79,8 @@ static inline bool mi_bitmap_try_find_claim_field(mi_bitmap_t bitmap, size_t idx
     }
     else {
       // on to the next bit range
-#ifdef MI_HAVE_BITSCAN
+#ifdef MI_HAVE_FAST_BITSCAN
-      const size_t shift = (count == 1 ? 1 : mi_bsr(map & m) - bitidx + 1);
+      const size_t shift = (count == 1 ? 1 : mi_bsr(mapm) - bitidx + 1);
       mi_assert_internal(shift > 0 && shift <= count);
 #else
       const size_t shift = 1;
@@ -196,28 +96,28 @@ static inline bool mi_bitmap_try_find_claim_field(mi_bitmap_t bitmap, size_t idx
 // Find `count` bits of 0 and set them to 1 atomically; returns `true` on success.
 // Starts at idx, and wraps around to search in all `bitmap_fields` fields.
 // `count` can be at most MI_BITMAP_FIELD_BITS and will never cross fields.
-static inline bool mi_bitmap_try_find_from_claim(mi_bitmap_t bitmap, const size_t bitmap_fields, const size_t start_field_idx, const size_t count, mi_bitmap_index_t* bitmap_idx) {
+bool _mi_bitmap_try_find_from_claim(mi_bitmap_t bitmap, const size_t bitmap_fields, const size_t start_field_idx, const size_t count, mi_bitmap_index_t* bitmap_idx) {
   size_t idx = start_field_idx;
   for (size_t visited = 0; visited < bitmap_fields; visited++, idx++) {
     if (idx >= bitmap_fields) idx = 0; // wrap
-    if (mi_bitmap_try_find_claim_field(bitmap, idx, count, bitmap_idx)) {
+    if (_mi_bitmap_try_find_claim_field(bitmap, idx, count, bitmap_idx)) {
       return true;
     }
   }
   return false;
 }
 
-
+/*
 // Find `count` bits of 0 and set them to 1 atomically; returns `true` on success.
 // For now, `count` can be at most MI_BITMAP_FIELD_BITS and will never span fields.
-static inline bool mi_bitmap_try_find_claim(mi_bitmap_t bitmap, const size_t bitmap_fields, const size_t count, mi_bitmap_index_t* bitmap_idx) {
+bool _mi_bitmap_try_find_claim(mi_bitmap_t bitmap, const size_t bitmap_fields, const size_t count, mi_bitmap_index_t* bitmap_idx) {
-  return mi_bitmap_try_find_from_claim(bitmap, bitmap_fields, 0, count, bitmap_idx);
+  return _mi_bitmap_try_find_from_claim(bitmap, bitmap_fields, 0, count, bitmap_idx);
 }
-
+*/
 
 // Set `count` bits at `bitmap_idx` to 0 atomically
 // Returns `true` if all `count` bits were 1 previously.
-static inline bool mi_bitmap_unclaim(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx) {
+bool _mi_bitmap_unclaim(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx) {
   const size_t idx = mi_bitmap_index_field(bitmap_idx);
   const size_t bitidx = mi_bitmap_index_bit_in_field(bitmap_idx);
   const uintptr_t mask = mi_bitmap_mask_(count, bitidx);
@@ -230,7 +130,7 @@ static inline bool mi_bitmap_unclaim(mi_bitmap_t bitmap, size_t bitmap_fields, s
 
 // Set `count` bits at `bitmap_idx` to 1 atomically
 // Returns `true` if all `count` bits were 0 previously. `any_zero` is `true` if there was at least one zero bit.
-static inline bool mi_bitmap_claim(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx, bool* any_zero) {
+bool _mi_bitmap_claim(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx, bool* any_zero) {
   const size_t idx = mi_bitmap_index_field(bitmap_idx);
   const size_t bitidx = mi_bitmap_index_bit_in_field(bitmap_idx);
   const uintptr_t mask = mi_bitmap_mask_(count, bitidx);
@@ -242,7 +142,7 @@ static inline bool mi_bitmap_claim(mi_bitmap_t bitmap, size_t bitmap_fields, siz
 }
 
 // Returns `true` if all `count` bits were 1. `any_ones` is `true` if there was at least one bit set to one.
-static inline bool mi_bitmap_is_claimedx(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx, bool* any_ones) {
+static bool mi_bitmap_is_claimedx(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx, bool* any_ones) {
   const size_t idx = mi_bitmap_index_field(bitmap_idx);
   const size_t bitidx = mi_bitmap_index_bit_in_field(bitmap_idx);
   const uintptr_t mask = mi_bitmap_mask_(count, bitidx);
@@ -252,11 +152,11 @@ static inline bool mi_bitmap_is_claimedx(mi_bitmap_t bitmap, size_t bitmap_field
   return ((field & mask) == mask);
 }
 
-static inline bool mi_bitmap_is_claimed(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx) {
+bool _mi_bitmap_is_claimed(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx) {
   return mi_bitmap_is_claimedx(bitmap, bitmap_fields, count, bitmap_idx, NULL);
 }
 
-static inline bool mi_bitmap_is_any_claimed(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx) {
+bool _mi_bitmap_is_any_claimed(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx) {
   bool any_ones;
   mi_bitmap_is_claimedx(bitmap, bitmap_fields, count, bitmap_idx, &any_ones);
   return any_ones;
@@ -270,17 +170,17 @@ static inline bool mi_bitmap_is_any_claimed(mi_bitmap_t bitmap, size_t bitmap_fi
 
 // Try to atomically claim a sequence of `count` bits starting from the field 
 // at `idx` in `bitmap` and crossing into subsequent fields. Returns `true` on success.
-static inline bool mi_bitmap_try_find_claim_field_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t idx, const size_t count, const size_t retries, mi_bitmap_index_t* bitmap_idx)
+static bool mi_bitmap_try_find_claim_field_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t idx, const size_t count, const size_t retries, mi_bitmap_index_t* bitmap_idx)
 {
   mi_assert_internal(bitmap_idx != NULL);
   
   // check initial trailing zeros
   _Atomic(uintptr_t)* field = &bitmap[idx];
   uintptr_t map = mi_atomic_load_relaxed(field);  
-  const uintptr_t bitidx = (map==0 ? 0 : mi_bsr(map) + 1);
+  const size_t initial = mi_clz(map);  // count of initial zeros starting at idx
-  const size_t initial = MI_BITMAP_FIELD_BITS - bitidx;    // count of initial zeros starting at idx
+  mi_assert_internal(initial >= 0 && initial <= MI_BITMAP_FIELD_BITS);
   if (initial == 0)     return false;
-  if (initial >= count) return mi_bitmap_try_find_claim_field(bitmap, idx, count, bitmap_idx);     // no need to cross fields
+  if (initial >= count) return _mi_bitmap_try_find_claim_field(bitmap, idx, count, bitmap_idx);     // no need to cross fields
   if (_mi_divide_up(count - initial, MI_BITMAP_FIELD_BITS) >= (bitmap_fields - idx)) return false; // not enough entries
 
   // scan ahead
@@ -314,7 +214,7 @@ static inline bool mi_bitmap_try_find_claim_field_across(mi_bitmap_t bitmap, siz
   
   // intermediate fields
   while (++field < final_field) {
-    newmap = mi_bitmap_mask_(MI_BITMAP_FIELD_BITS, 0);
+    newmap = MI_BITMAP_FIELD_FULL;
     map = 0;
     if (!mi_atomic_cas_strong_acq_rel(field, &map, newmap)) { goto rollback; }
   }
@@ -328,14 +228,14 @@ static inline bool mi_bitmap_try_find_claim_field_across(mi_bitmap_t bitmap, siz
   } while (!mi_atomic_cas_strong_acq_rel(field, &map, newmap));
 
   // claimed!
-  *bitmap_idx = mi_bitmap_index_create(idx, bitidx);
+  *bitmap_idx = mi_bitmap_index_create(idx, MI_BITMAP_FIELD_BITS - initial);
   return true;
 
 rollback: 
   // roll back intermediate fields
   while (--field > initial_field) {
     newmap = 0;
-    map = mi_bitmap_mask_(MI_BITMAP_FIELD_BITS, 0);
+    map = MI_BITMAP_FIELD_FULL;
     mi_assert_internal(mi_atomic_load_relaxed(field) == map);
     mi_atomic_store_release(field, newmap);
   }
@@ -358,15 +258,15 @@ rollback:
 
 // Find `count` bits of zeros and set them to 1 atomically; returns `true` on success.
 // Starts at idx, and wraps around to search in all `bitmap_fields` fields.
-static inline bool mi_bitmap_try_find_from_claim_across(mi_bitmap_t bitmap, const size_t bitmap_fields, const size_t start_field_idx, const size_t count, mi_bitmap_index_t* bitmap_idx) {
+bool _mi_bitmap_try_find_from_claim_across(mi_bitmap_t bitmap, const size_t bitmap_fields, const size_t start_field_idx, const size_t count, mi_bitmap_index_t* bitmap_idx) {
   mi_assert_internal(count > 0);
-  if (count==1) return mi_bitmap_try_find_from_claim(bitmap, bitmap_fields, start_field_idx, count, bitmap_idx);
+  if (count==1) return _mi_bitmap_try_find_from_claim(bitmap, bitmap_fields, start_field_idx, count, bitmap_idx);
   size_t idx = start_field_idx;
   for (size_t visited = 0; visited < bitmap_fields; visited++, idx++) {
     if (idx >= bitmap_fields) idx = 0; // wrap
     // try to claim inside the field
     if (count <= MI_BITMAP_FIELD_BITS) {
-      if (mi_bitmap_try_find_claim_field(bitmap, idx, count, bitmap_idx)) {
+      if (_mi_bitmap_try_find_claim_field(bitmap, idx, count, bitmap_idx)) {
         return true;
       }
     }
@@ -379,7 +279,7 @@ static inline bool mi_bitmap_try_find_from_claim_across(mi_bitmap_t bitmap, cons
 }
 
 // Helper for masks across fields; returns the mid count, post_mask may be 0
-static inline size_t mi_bitmap_mask_across(mi_bitmap_index_t bitmap_idx, size_t bitmap_fields, size_t count, uintptr_t* pre_mask, uintptr_t* mid_mask, uintptr_t* post_mask) {
+static size_t mi_bitmap_mask_across(mi_bitmap_index_t bitmap_idx, size_t bitmap_fields, size_t count, uintptr_t* pre_mask, uintptr_t* mid_mask, uintptr_t* post_mask) {
   UNUSED_RELEASE(bitmap_fields);
   const size_t bitidx = mi_bitmap_index_bit_in_field(bitmap_idx);
   if (mi_likely(bitidx + count <= MI_BITMAP_FIELD_BITS)) {
@@ -405,7 +305,7 @@ static inline size_t mi_bitmap_mask_across(mi_bitmap_index_t bitmap_idx, size_t
 
 // Set `count` bits at `bitmap_idx` to 0 atomically
 // Returns `true` if all `count` bits were 1 previously.
-static inline bool mi_bitmap_unclaim_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx) {
+bool _mi_bitmap_unclaim_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx) {
   size_t idx = mi_bitmap_index_field(bitmap_idx);
   uintptr_t pre_mask;
   uintptr_t mid_mask;
@@ -428,7 +328,7 @@ static inline bool mi_bitmap_unclaim_across(mi_bitmap_t bitmap, size_t bitmap_fi
 
 // Set `count` bits at `bitmap_idx` to 1 atomically
 // Returns `true` if all `count` bits were 0 previously. `any_zero` is `true` if there was at least one zero bit.
-static inline bool mi_bitmap_claim_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx, bool* pany_zero) {
+bool _mi_bitmap_claim_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx, bool* pany_zero) {
   size_t idx = mi_bitmap_index_field(bitmap_idx);
   uintptr_t pre_mask;
   uintptr_t mid_mask;
@@ -457,7 +357,7 @@ static inline bool mi_bitmap_claim_across(mi_bitmap_t bitmap, size_t bitmap_fiel
 
 // Returns `true` if all `count` bits were 1. 
 // `any_ones` is `true` if there was at least one bit set to one.
-static inline bool mi_bitmap_is_claimedx_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx, bool* pany_ones) {
+static bool mi_bitmap_is_claimedx_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx, bool* pany_ones) {
   size_t idx = mi_bitmap_index_field(bitmap_idx);
   uintptr_t pre_mask;
   uintptr_t mid_mask;
@@ -483,14 +383,14 @@ static inline bool mi_bitmap_is_claimedx_across(mi_bitmap_t bitmap, size_t bitma
   return all_ones;
 }
 
-static inline bool mi_bitmap_is_claimed_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx) {
+bool _mi_bitmap_is_claimed_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx) {
   return mi_bitmap_is_claimedx_across(bitmap, bitmap_fields, count, bitmap_idx, NULL);
 }
 
-static inline bool mi_bitmap_is_any_claimed_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx) {
+/*
+bool _mi_bitmap_is_any_claimed_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx) {
   bool any_ones;
   mi_bitmap_is_claimedx_across(bitmap, bitmap_fields, count, bitmap_idx, &any_ones);
   return any_ones;
 }
-
+*/
-#endif

src/bitmap.h

@@ -0,0 +1,106 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2019,2020 Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+
+/* ----------------------------------------------------------------------------
+Concurrent bitmap that can set/reset sequences of bits atomically,
+represeted as an array of fields where each field is a machine word (`uintptr_t`)
+
+There are two api's; the standard one cannot have sequences that cross
+between the bitmap fields (and a sequence must be <= MI_BITMAP_FIELD_BITS).
+(this is used in region allocation)
+
+The `_across` postfixed functions do allow sequences that can cross over
+between the fields. (This is used in arena allocation)
+---------------------------------------------------------------------------- */
+#pragma once
+#ifndef MI_BITMAP_H
+#define MI_BITMAP_H
+
+/* -----------------------------------------------------------
+  Bitmap definition
+----------------------------------------------------------- */
+
+#define MI_BITMAP_FIELD_BITS   (8*MI_INTPTR_SIZE)
+#define MI_BITMAP_FIELD_FULL   (~((uintptr_t)0))   // all bits set
+
+// An atomic bitmap of `uintptr_t` fields
+typedef _Atomic(uintptr_t)  mi_bitmap_field_t;
+typedef mi_bitmap_field_t*  mi_bitmap_t;
+
+// A bitmap index is the index of the bit in a bitmap.
+typedef size_t mi_bitmap_index_t;
+
+// Create a bit index.
+static inline mi_bitmap_index_t mi_bitmap_index_create(size_t idx, size_t bitidx) {
+  mi_assert_internal(bitidx < MI_BITMAP_FIELD_BITS);
+  return (idx*MI_BITMAP_FIELD_BITS) + bitidx;
+}
+
+// Create a bit index.
+static inline mi_bitmap_index_t mi_bitmap_index_create_from_bit(size_t full_bitidx) {  
+  return mi_bitmap_index_create(full_bitidx / MI_BITMAP_FIELD_BITS, full_bitidx % MI_BITMAP_FIELD_BITS);
+}
+
+// Get the field index from a bit index.
+static inline size_t mi_bitmap_index_field(mi_bitmap_index_t bitmap_idx) {
+  return (bitmap_idx / MI_BITMAP_FIELD_BITS);
+}
+
+// Get the bit index in a bitmap field
+static inline size_t mi_bitmap_index_bit_in_field(mi_bitmap_index_t bitmap_idx) {
+  return (bitmap_idx % MI_BITMAP_FIELD_BITS);
+}
+
+// Get the full bit index
+static inline size_t mi_bitmap_index_bit(mi_bitmap_index_t bitmap_idx) {
+  return bitmap_idx;
+}
+
+/* -----------------------------------------------------------
+  Claim a bit sequence atomically
+----------------------------------------------------------- */
+
+// Try to atomically claim a sequence of `count` bits in a single
+// field at `idx` in `bitmap`. Returns `true` on success.
+bool _mi_bitmap_try_find_claim_field(mi_bitmap_t bitmap, size_t idx, const size_t count, mi_bitmap_index_t* bitmap_idx);
+
+// Starts at idx, and wraps around to search in all `bitmap_fields` fields.
+// For now, `count` can be at most MI_BITMAP_FIELD_BITS and will never cross fields.
+bool _mi_bitmap_try_find_from_claim(mi_bitmap_t bitmap, const size_t bitmap_fields, const size_t start_field_idx, const size_t count, mi_bitmap_index_t* bitmap_idx);
+
+// Set `count` bits at `bitmap_idx` to 0 atomically
+// Returns `true` if all `count` bits were 1 previously.
+bool _mi_bitmap_unclaim(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx);
+
+// Set `count` bits at `bitmap_idx` to 1 atomically
+// Returns `true` if all `count` bits were 0 previously. `any_zero` is `true` if there was at least one zero bit.
+bool _mi_bitmap_claim(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx, bool* any_zero);
+
+bool _mi_bitmap_is_claimed(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx);
+bool _mi_bitmap_is_any_claimed(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx);
+
+
+//--------------------------------------------------------------------------
+// the `_across` functions work on bitmaps where sequences can cross over
+// between the fields. This is used in arena allocation
+//--------------------------------------------------------------------------
+
+// Find `count` bits of zeros and set them to 1 atomically; returns `true` on success.
+// Starts at idx, and wraps around to search in all `bitmap_fields` fields.
+bool _mi_bitmap_try_find_from_claim_across(mi_bitmap_t bitmap, const size_t bitmap_fields, const size_t start_field_idx, const size_t count, mi_bitmap_index_t* bitmap_idx);
+
+// Set `count` bits at `bitmap_idx` to 0 atomically
+// Returns `true` if all `count` bits were 1 previously.
+bool _mi_bitmap_unclaim_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx);
+
+// Set `count` bits at `bitmap_idx` to 1 atomically
+// Returns `true` if all `count` bits were 0 previously. `any_zero` is `true` if there was at least one zero bit.
+bool _mi_bitmap_claim_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx, bool* pany_zero);
+
+bool _mi_bitmap_is_claimed_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx);
+
+#endif

src/os.c

@@ -368,7 +368,7 @@ static void* mi_unix_mmap(void* addr, size_t size, size_t try_alignment, int pro
   const int fd = mi_unix_mmap_fd();
   #if defined(MAP_ALIGNED)  // BSD
   if (try_alignment > 0) {
-    size_t n = _mi_bsr(try_alignment);
+    size_t n = mi_bsr(try_alignment);
     if (((size_t)1 << n) == try_alignment && n >= 12 && n <= 30) {  // alignment is a power of 2 and 4096 <= alignment <= 1GiB
       flags |= MAP_ALIGNED(n);
     }

src/page-queue.c

@@ -49,50 +49,6 @@ static inline bool mi_page_queue_is_special(const mi_page_queue_t* pq) {
   Bins
 ----------------------------------------------------------- */
 
-// Bit scan reverse: return the index of the highest bit.
-static inline uint8_t mi_bsr32(uint32_t x);
-
-#if defined(_MSC_VER)
-#include <intrin.h>
-static inline uint8_t mi_bsr32(uint32_t x) {
-  uint32_t idx;
-  _BitScanReverse((DWORD*)&idx, x);
-  return (uint8_t)idx;
-}
-#elif defined(__GNUC__) || defined(__clang__)
-static inline uint8_t mi_bsr32(uint32_t x) {
-  return (31 - __builtin_clz(x));
-}
-#else
-static inline uint8_t mi_bsr32(uint32_t x) {
-  // de Bruijn multiplication, see <http://supertech.csail.mit.edu/papers/debruijn.pdf>
-  static const uint8_t debruijn[32] = {
-     31,  0, 22,  1, 28, 23, 18,  2, 29, 26, 24, 10, 19,  7,  3, 12,
-     30, 21, 27, 17, 25,  9,  6, 11, 20, 16,  8,  5, 15,  4, 14, 13,
-  };
-  x |= x >> 1;
-  x |= x >> 2;
-  x |= x >> 4;
-  x |= x >> 8;
-  x |= x >> 16;
-  x++;
-  return debruijn[(x*0x076be629) >> 27];
-}
-#endif
-
-// Bit scan reverse: return the index of the highest bit.
-uint8_t _mi_bsr(uintptr_t x) {
-  if (x == 0) return 0;
-#if MI_INTPTR_SIZE==8
-  uint32_t hi = (x >> 32);
-  return (hi == 0 ? mi_bsr32((uint32_t)x) : 32 + mi_bsr32(hi));
-#elif MI_INTPTR_SIZE==4
-  return mi_bsr32(x);
-#else
-# error "define bsr for non-32 or 64-bit platforms"
-#endif
-}
-
 // Return the bin for a given field size.
 // Returns MI_BIN_HUGE if the size is too large.
 // We use `wsize` for the size in "machine word sizes",
@@ -125,7 +81,7 @@ extern inline uint8_t _mi_bin(size_t size) {
     #endif
     wsize--;
     // find the highest bit
-    uint8_t b = mi_bsr32((uint32_t)wsize);
+    uint8_t b = (uint8_t)mi_bsr(wsize);  // note: wsize != 0
     // and use the top 3 bits to determine the bin (~12.5% worst internal fragmentation).
     // - adjust with 3 because we use do not round the first 8 sizes
     //   which each get an exact bin

src/region.c

@@ -37,7 +37,7 @@ Possible issues:
 
 #include <string.h>  // memset
 
-#include "bitmap.inc.c"
+#include "bitmap.h"
 
 // Internal raw OS interface
 size_t  _mi_os_large_page_size();
@@ -200,7 +200,7 @@ static bool mi_region_try_alloc_os(size_t blocks, bool commit, bool allow_large,
   mi_atomic_store_release(&r->commit, (region_commit ? MI_BITMAP_FIELD_FULL : 0));
   mi_atomic_store_release(&r->reset, (uintptr_t)0);
   *bit_idx = 0;
-  mi_bitmap_claim(&r->in_use, 1, blocks, *bit_idx, NULL);
+  _mi_bitmap_claim(&r->in_use, 1, blocks, *bit_idx, NULL);
   mi_atomic_store_ptr_release(void,&r->start, start);
 
   // and share it 
@@ -248,7 +248,7 @@ static bool mi_region_try_claim(int numa_node, size_t blocks, bool allow_large,
     // if this region suits our demand (numa node matches, large OS page matches)
     if (mi_region_is_suitable(r, numa_node, allow_large)) {
       // then try to atomically claim a segment(s) in this region
-      if (mi_bitmap_try_find_claim_field(&r->in_use, 0, blocks, bit_idx)) {
+      if (_mi_bitmap_try_find_claim_field(&r->in_use, 0, blocks, bit_idx)) {
         tld->region_idx = idx;    // remember the last found position
         *region = r;
         return true;
@@ -277,7 +277,7 @@ static void* mi_region_try_alloc(size_t blocks, bool* commit, bool* is_large, bo
   // ------------------------------------------------
   // found a region and claimed `blocks` at `bit_idx`, initialize them now
   mi_assert_internal(region != NULL);
-  mi_assert_internal(mi_bitmap_is_claimed(&region->in_use, 1, blocks, bit_idx));
+  mi_assert_internal(_mi_bitmap_is_claimed(&region->in_use, 1, blocks, bit_idx));
 
   mi_region_info_t info;
   info.value = mi_atomic_load_acquire(&region->info);
@@ -285,7 +285,7 @@ static void* mi_region_try_alloc(size_t blocks, bool* commit, bool* is_large, bo
   mi_assert_internal(!(info.x.is_large && !*is_large));
   mi_assert_internal(start != NULL);
 
-  *is_zero = mi_bitmap_claim(&region->dirty, 1, blocks, bit_idx, NULL);  
+  *is_zero = _mi_bitmap_claim(&region->dirty, 1, blocks, bit_idx, NULL);  
   *is_large = info.x.is_large;
   *memid = mi_memid_create(region, bit_idx);
   void* p = start + (mi_bitmap_index_bit_in_field(bit_idx) * MI_SEGMENT_SIZE);
@@ -294,7 +294,7 @@ static void* mi_region_try_alloc(size_t blocks, bool* commit, bool* is_large, bo
   if (*commit) {
     // ensure commit
     bool any_uncommitted;
-    mi_bitmap_claim(&region->commit, 1, blocks, bit_idx, &any_uncommitted);
+    _mi_bitmap_claim(&region->commit, 1, blocks, bit_idx, &any_uncommitted);
     if (any_uncommitted) {
       mi_assert_internal(!info.x.is_large);
       bool commit_zero;
@@ -304,12 +304,12 @@ static void* mi_region_try_alloc(size_t blocks, bool* commit, bool* is_large, bo
   }
   else {
     // no need to commit, but check if already fully committed
-    *commit = mi_bitmap_is_claimed(&region->commit, 1, blocks, bit_idx);
+    *commit = _mi_bitmap_is_claimed(&region->commit, 1, blocks, bit_idx);
   }  
-  mi_assert_internal(!*commit || mi_bitmap_is_claimed(&region->commit, 1, blocks, bit_idx));
+  mi_assert_internal(!*commit || _mi_bitmap_is_claimed(&region->commit, 1, blocks, bit_idx));
 
   // unreset reset blocks
-  if (mi_bitmap_is_any_claimed(&region->reset, 1, blocks, bit_idx)) {
+  if (_mi_bitmap_is_any_claimed(&region->reset, 1, blocks, bit_idx)) {
     // some blocks are still reset
     mi_assert_internal(!info.x.is_large);
     mi_assert_internal(!mi_option_is_enabled(mi_option_eager_commit) || *commit || mi_option_get(mi_option_eager_commit_delay) > 0); 
@@ -320,7 +320,7 @@ static void* mi_region_try_alloc(size_t blocks, bool* commit, bool* is_large, bo
       if (reset_zero) *is_zero = true;
     }
   }
-  mi_assert_internal(!mi_bitmap_is_any_claimed(&region->reset, 1, blocks, bit_idx));
+  mi_assert_internal(!_mi_bitmap_is_any_claimed(&region->reset, 1, blocks, bit_idx));
   
   #if (MI_DEBUG>=2)
   if (*commit) { ((uint8_t*)p)[0] = 0; }
@@ -409,12 +409,12 @@ void _mi_mem_free(void* p, size_t size, size_t id, bool full_commit, bool any_re
 
     // committed?
     if (full_commit && (size % MI_SEGMENT_SIZE) == 0) {
-      mi_bitmap_claim(&region->commit, 1, blocks, bit_idx, NULL);
+      _mi_bitmap_claim(&region->commit, 1, blocks, bit_idx, NULL);
     }
 
     if (any_reset) {
       // set the is_reset bits if any pages were reset
-      mi_bitmap_claim(&region->reset, 1, blocks, bit_idx, NULL);
+      _mi_bitmap_claim(&region->reset, 1, blocks, bit_idx, NULL);
     }
 
     // reset the blocks to reduce the working set.
@@ -423,7 +423,7 @@ void _mi_mem_free(void* p, size_t size, size_t id, bool full_commit, bool any_re
            mi_option_is_enabled(mi_option_reset_decommits))) // cannot reset halfway committed segments, use only `option_page_reset` instead            
     {
       bool any_unreset;
-      mi_bitmap_claim(&region->reset, 1, blocks, bit_idx, &any_unreset);
+      _mi_bitmap_claim(&region->reset, 1, blocks, bit_idx, &any_unreset);
       if (any_unreset) {
         _mi_abandoned_await_readers(); // ensure no more pending write (in case reset = decommit)
         _mi_mem_reset(p, blocks * MI_SEGMENT_SIZE, tld);

src/segment.c

@@ -7,7 +7,6 @@ terms of the MIT license. A copy of the license can be found in the file
 #include "mimalloc.h"
 #include "mimalloc-internal.h"
 #include "mimalloc-atomic.h"
-#include "bitmap.inc.c"  // mi_bsr
 
 #include <string.h>  // memset
 #include <stdio.h>
@@ -49,29 +48,26 @@ static uint8_t* mi_slice_start(const mi_slice_t* slice) {
    Bins
 ----------------------------------------------------------- */
 // Use bit scan forward to quickly find the first zero bit if it is available
-#if !defined(MI_HAVE_BITSCAN)
-#error "define bsr for your platform"
-#endif
 
-static size_t mi_slice_bin8(size_t slice_count) {
+static inline size_t mi_slice_bin8(size_t slice_count) {
   if (slice_count<=1) return slice_count;
   mi_assert_internal(slice_count <= MI_SLICES_PER_SEGMENT);
   slice_count--;
-  size_t s = mi_bsr(slice_count);
+  size_t s = mi_bsr(slice_count);  // slice_count > 1
   if (s <= 2) return slice_count + 1;
   size_t bin = ((s << 2) | ((slice_count >> (s - 2))&0x03)) - 4;
   return bin;
 }
 
-static size_t mi_slice_bin(size_t slice_count) {
+static inline size_t mi_slice_bin(size_t slice_count) {
   mi_assert_internal(slice_count*MI_SEGMENT_SLICE_SIZE <= MI_SEGMENT_SIZE);
   mi_assert_internal(mi_slice_bin8(MI_SLICES_PER_SEGMENT) <= MI_SEGMENT_BIN_MAX);
-  size_t bin = (slice_count==0 ? 0 : mi_slice_bin8(slice_count));
+  size_t bin = mi_slice_bin8(slice_count);
   mi_assert_internal(bin <= MI_SEGMENT_BIN_MAX);
   return bin;
 }
 
-static size_t mi_slice_index(const mi_slice_t* slice) {
+static inline size_t mi_slice_index(const mi_slice_t* slice) {
   mi_segment_t* segment = _mi_ptr_segment(slice);
   ptrdiff_t index = slice - segment->slices;
   mi_assert_internal(index >= 0 && index < (ptrdiff_t)segment->slice_entries);
@@ -1436,7 +1432,7 @@ static mi_segment_t* _mi_segment_of(const void* p) {
   }
   if (index==0) return NULL;
   // search downwards for the first segment in case it is an interior pointer
-  // could be slow but searches in MI_INTPTR_SIZE * MI_SEGMENT_SIZE (4GiB) steps trough 
+  // could be slow but searches in MI_INTPTR_SIZE * MI_SEGMENT_SIZE (512MiB) steps trough 
   // valid huge objects
   // note: we could maintain a lowest index to speed up the path for invalid pointers?
   size_t lobitidx;
@@ -1444,14 +1440,14 @@ static mi_segment_t* _mi_segment_of(const void* p) {
   uintptr_t lobits = mask & (((uintptr_t)1 << bitidx) - 1);
   if (lobits != 0) {
     loindex = index;
-    lobitidx = _mi_bsr(lobits);
+    lobitidx = mi_bsr(lobits);    // lobits != 0
   }
   else {
     uintptr_t lomask = mask;
     loindex = index - 1;
     while (loindex > 0 && (lomask = mi_atomic_load_relaxed(&mi_segment_map[loindex])) == 0) loindex--;
     if (loindex==0) return NULL;
-    lobitidx = _mi_bsr(lomask);
+    lobitidx = mi_bsr(lomask);    // lomask != 0
   }
   // take difference as the addresses could be larger than the MAX_ADDRESS space.
   size_t diff = (((index - loindex) * (8*MI_INTPTR_SIZE)) + bitidx - lobitidx) * MI_SEGMENT_SIZE;

src/static.c

@@ -23,7 +23,7 @@ terms of the MIT license. A copy of the license can be found in the file
 #include "stats.c"
 #include "random.c"
 #include "os.c"
-//#include "memory.c"
+#include "bitmap.c"
 #include "arena.c"
 #include "segment.c"
 #include "page.c"