From 8e19a0926f3fcc7e1d5bd1fec6e96de32274619f Mon Sep 17 00:00:00 2001
From: Hugh Sanderson <gamehaxe@gmail.com>
Date: Tue, 17 Sep 2024 16:01:43 +0800
Subject: [PATCH] Fix HXCPP_GC_GENERATIONAL in combination with
 HXCPP_ALIGN_ALLOC

---
 include/hx/GC.h              |  81 ++++++++++---------
 include/hxcpp.h              |   2 +-
 src/hx/gc/Immix.cpp          | 148 +++++++++++++++++++++--------------
 test/haxe/compile.hxml       |   2 +-
 toolchain/common-defines.xml |   2 +
 5 files changed, 140 insertions(+), 95 deletions(-)

diff --git a/include/hx/GC.h b/include/hx/GC.h
index 5370c142e..a9148c505 100644
--- a/include/hx/GC.h
+++ b/include/hx/GC.h
@@ -363,7 +363,12 @@ class ImmixAllocator
    {
       #ifdef HXCPP_GC_NURSERY
 
+         #ifdef HXCPP_ALIGN_ALLOC
+         // make sure buffer is 8-byte aligned
+         unsigned char *buffer = alloc->spaceFirst + ( (size_t)alloc->spaceFirst & 4 );
+         #else
          unsigned char *buffer = alloc->spaceFirst;
+         #endif
          unsigned char *end = buffer + (inSize + 4);
 
          if ( end > alloc->spaceOversize )
@@ -381,6 +386,10 @@ class ImmixAllocator
                ((unsigned int *)buffer)[-1] = inSize;
          }
 
+         #if defined(HXCPP_GC_CHECK_POINTER) && defined(HXCPP_GC_DEBUG_ALWAYS_MOVE)
+         hx::GCOnNewPointer(buffer);
+         #endif
+
          #ifdef HXCPP_TELEMETRY
          __hxt_gc_new((hx::StackContext *)alloc,buffer, inSize, inName);
          #endif
@@ -388,41 +397,43 @@ class ImmixAllocator
          return buffer;
 
       #else
-         #ifndef HXCPP_ALIGN_ALLOC
-            // Inline the fast-path if we can
-            // We know the object can hold a pointer (vtable) and that the size is int-aligned
-            int start = alloc->spaceStart;
-            int end = start + sizeof(int) + inSize;
-
-            if ( end <= alloc->spaceEnd )
-            {
-               alloc->spaceStart = end;
-
-               unsigned int *buffer = (unsigned int *)(alloc->allocBase + start);
-
-               int startRow = start>>IMMIX_LINE_BITS;
-
-               alloc->allocStartFlags[ startRow ] |= gImmixStartFlag[start&127];
-
-               if (inContainer)
-                  *buffer++ =  (( (end+(IMMIX_LINE_LEN-1))>>IMMIX_LINE_BITS) -startRow) |
-                               (inSize<<IMMIX_ALLOC_SIZE_SHIFT) |
-                               hx::gMarkIDWithContainer;
-               else
-                  *buffer++ =  (( (end+(IMMIX_LINE_LEN-1))>>IMMIX_LINE_BITS) -startRow) |
-                               (inSize<<IMMIX_ALLOC_SIZE_SHIFT) |
-                               hx::gMarkID;
-
-               #if defined(HXCPP_GC_CHECK_POINTER) && defined(HXCPP_GC_DEBUG_ALWAYS_MOVE)
-               hx::GCOnNewPointer(buffer);
-               #endif
-
-               #ifdef HXCPP_TELEMETRY
-               __hxt_gc_new((hx::StackContext *)alloc,buffer, inSize, inName);
-               #endif
-               return buffer;
-            }
-         #endif // HXCPP_ALIGN_ALLOC
+         // Inline the fast-path if we can
+         // We know the object can hold a pointer (vtable) and that the size is int-aligned
+         int start = alloc->spaceStart;
+         #ifdef HXCPP_ALIGN_ALLOC
+            // Ensure odd alignment in 8 bytes
+            start += 4 - (start & 4);
+         #endif
+         int end = start + sizeof(int) + inSize;
+
+         if ( end <= alloc->spaceEnd )
+         {
+            alloc->spaceStart = end;
+
+            unsigned int *buffer = (unsigned int *)(alloc->allocBase + start);
+
+            int startRow = start>>IMMIX_LINE_BITS;
+
+            alloc->allocStartFlags[ startRow ] |= gImmixStartFlag[start&127];
+
+            if (inContainer)
+               *buffer++ =  (( (end+(IMMIX_LINE_LEN-1))>>IMMIX_LINE_BITS) -startRow) |
+                            (inSize<<IMMIX_ALLOC_SIZE_SHIFT) |
+                            hx::gMarkIDWithContainer;
+            else
+               *buffer++ =  (( (end+(IMMIX_LINE_LEN-1))>>IMMIX_LINE_BITS) -startRow) |
+                            (inSize<<IMMIX_ALLOC_SIZE_SHIFT) |
+                            hx::gMarkID;
+
+            #if defined(HXCPP_GC_CHECK_POINTER) && defined(HXCPP_GC_DEBUG_ALWAYS_MOVE)
+            hx::GCOnNewPointer(buffer);
+            #endif
+
+            #ifdef HXCPP_TELEMETRY
+            __hxt_gc_new((hx::StackContext *)alloc,buffer, inSize, inName);
+            #endif
+            return buffer;
+         }
 
          // Fall back to external method
          void *result = alloc->CallAlloc(inSize, inContainer ? IMMIX_ALLOC_IS_CONTAINER : 0);
diff --git a/include/hxcpp.h b/include/hxcpp.h
index b2568d142..faa1b4db2 100755
--- a/include/hxcpp.h
+++ b/include/hxcpp.h
@@ -70,7 +70,7 @@
 
 // Must allign allocs to 8 bytes to match floating point requirement?
 // Ints must br read on 4-byte boundary
-#if defined(EMSCRIPTEN) || defined(GCW0)
+#if (!defined(HXCPP_ALIGN_FLOAT) && (defined(EMSCRIPTEN) || defined(GCW0)) )
    #define HXCPP_ALIGN_ALLOC
 #endif
 
diff --git a/src/hx/gc/Immix.cpp b/src/hx/gc/Immix.cpp
index 253199c53..05d0ac842 100644
--- a/src/hx/gc/Immix.cpp
+++ b/src/hx/gc/Immix.cpp
@@ -140,7 +140,7 @@ static size_t sgMaximumFreeSpace  = 1024*1024*1024;
 //#define HXCPP_GC_SUMMARY
 //#define PROFILE_COLLECT
 //#define PROFILE_THREAD_USAGE
-//#define HX_GC_VERIFY
+//#define HXCPP_GC_VERIFY
 //#define HX_GC_VERIFY_ALLOC_START
 //#define SHOW_MEM_EVENTS
 //#define SHOW_MEM_EVENTS_VERBOSE
@@ -158,7 +158,7 @@ static size_t sgMaximumFreeSpace  = 1024*1024*1024;
 //  every collect so you can see where it goes wrong (usually requires HX_GC_FIXED_BLOCKS)
 //#define HX_WATCH
 
-#if defined(HX_GC_VERIFY) && defined(HXCPP_GC_GENERATIONAL)
+#if defined(HXCPP_GC_VERIFY) && defined(HXCPP_GC_GENERATIONAL)
 #define HX_GC_VERIFY_GENERATIONAL
 #endif
 
@@ -192,7 +192,7 @@ static bool sGcVerifyGenerational = false;
 #endif
 
 
-#if HX_HAS_ATOMIC && (HXCPP_GC_DEBUG_LEVEL==0) && !defined(HX_GC_VERIFY) && !defined(EMSCRIPTEN)
+#if HX_HAS_ATOMIC && (HXCPP_GC_DEBUG_LEVEL==0) && !defined(HXCPP_GC_VERIFY) && !defined(EMSCRIPTEN)
   #if defined(HX_MACOS) || defined(HX_WINDOWS) || defined(HX_LINUX)
   enum { MAX_GC_THREADS = 4 };
   #else
@@ -728,6 +728,13 @@ static int gBlockInfoEmptySlots = 0;
 #define ZEROED_THREAD 1
 #define ZEROED_AUTO   2
 
+// Align padding based on block offset
+#ifdef HXCPP_ALIGN_ALLOC
+    #define ALIGN_PADDING(x) (4-(x&4))
+#else
+    #define ALIGN_PADDING(x) 0
+#endif
+
 struct BlockDataInfo
 {
    int             mId;
@@ -1202,6 +1209,13 @@ struct BlockDataInfo
          int last = scan + mRanges[h].length;
          if (inOffset<last)
          {
+            #ifdef HXCPP_ALIGN_ALLOC
+            // Make sure header scan is odd-int aligned so the following object will be even-int
+            // aligned
+            if (!(scan & 0x4))
+               scan += 4;
+            #endif
+
             // Found hole that the object was possibly allocated in
             while(scan<=inOffset)
             {
@@ -1237,6 +1251,12 @@ struct BlockDataInfo
                   return allocString;
                }
                scan = end;
+               #ifdef HXCPP_ALIGN_ALLOC
+               // Make sure scan is odd-int aligned so the following object will be even-int
+               // aligned
+               if (!(scan & 0x4))
+                  scan += 4;
+               #endif
             }
             break;
          }
@@ -1353,7 +1373,7 @@ struct BlockDataInfo
    }
    #endif
 
-   #ifdef HX_GC_VERIFY
+   #ifdef HXCPP_GC_VERIFY
    void verify(const char *inWhere)
    {
       for(int i=IMMIX_HEADER_LINES;i<IMMIX_LINES;i++)
@@ -1471,6 +1491,14 @@ void GCCheckPointer(void *inPtr)
 
 void GCOnNewPointer(void *inPtr)
 {
+   #ifdef HXCPP_ALIGN_ALLOC
+   if ( (size_t)inPtr & 0x7 )
+   {
+      GCLOG("Misaligned pointer %p\n", inPtr);
+      NullReference("Object", false);
+   }
+   #endif
+
    #ifdef HXCPP_GC_DEBUG_ALWAYS_MOVE
    hx::sgPointerMoved.erase(inPtr);
    _hx_atomic_add(&sgAllocsSinceLastSpam, 1);
@@ -3521,7 +3549,7 @@ class GlobalAllocator
       for(int i=0;i<BLOCK_OFSIZE_COUNT;i++)
          mNextFreeBlockOfSize[i] = newSize;
 
-      #ifdef HX_GC_VERIFY
+      #ifdef HXCPP_GC_VERIFY
       VerifyBlockOrder();
       #endif
 
@@ -3746,7 +3774,7 @@ class GlobalAllocator
                         int size = ((header & IMMIX_ALLOC_SIZE_MASK) >> IMMIX_ALLOC_SIZE_SHIFT);
                         int allocSize = size + sizeof(int);
 
-                        while(allocSize>destLen)
+                        while(allocSize + ALIGN_PADDING(destPos)>destLen)
                         {
                            hole++;
                            if (hole<holes)
@@ -3778,6 +3806,10 @@ class GlobalAllocator
                            }
                         }
 
+                        #ifdef HXCPP_ALIGN_ALLOC
+                        destPos += ALIGN_PADDING(destPos);
+                        #endif
+
                         int startRow = destPos>>IMMIX_LINE_BITS;
 
                         destStarts[ startRow ] |= hx::gImmixStartFlag[destPos&127];
@@ -3897,6 +3929,8 @@ class GlobalAllocator
       return 0;
    }
 
+ 
+
    int MoveSurvivors(hx::QuickVec<hx::Object *> *inRemembered)
    {
       int sourceScan = 0;
@@ -3951,7 +3985,7 @@ class GlobalAllocator
                            int allocSize = size + sizeof(int);
 
                            // Find dest reqion ...
-                           while(destHole==0 || destLen<allocSize)
+                           while(destHole==0 || destLen < ALIGN_PADDING(destPos) + allocSize)
                            {
                               if (destHole==0)
                               {
@@ -3964,7 +3998,7 @@ class GlobalAllocator
                                  destHole = -1;
                                  destLen = 0;
                               }
-                              if (destLen<allocSize)
+                              if (destLen + ALIGN_PADDING(0)<allocSize)
                               {
                                  destHole++;
                                  if (destHole>=destInfo->mHoles)
@@ -3981,6 +4015,11 @@ class GlobalAllocator
 
                            // TODO - not copy + paste
 
+                        printf("Move!\n");
+                           #ifdef HXCPP_ALIGN_ALLOC
+                           destPos += ALIGN_PADDING(destPos);
+                           #endif
+
                            int startRow = destPos>>IMMIX_LINE_BITS;
 
                            destStarts[ startRow ] |= hx::gImmixStartFlag[destPos&127];
@@ -4060,7 +4099,7 @@ class GlobalAllocator
              gAllocGroups[mAllBlocks[i]->mGroupId].isEmpty=false;
       }
 
-      #ifdef HX_GC_VERIFY
+      #ifdef HXCPP_GC_VERIFY
       typedef std::pair< void *, void * > ReleasedRange;
       std::vector<ReleasedRange> releasedRange;
       std::vector<int> releasedGids;
@@ -4076,7 +4115,7 @@ class GlobalAllocator
             #ifdef SHOW_MEM_EVENTS_VERBOSE
             GCLOG("Release group %d: %p -> %p\n", i, g.alloc, g.alloc+groupBytes);
             #endif
-            #ifdef HX_GC_VERIFY
+            #ifdef HXCPP_GC_VERIFY
             releasedRange.push_back( ReleasedRange(g.alloc, g.alloc+groupBytes) );
             releasedGids.push_back(i);
             #endif
@@ -4105,7 +4144,7 @@ class GlobalAllocator
          }
          else
          {
-            #ifdef HX_GC_VERIFY
+            #ifdef HXCPP_GC_VERIFY
             for(int g=0;g<releasedGids.size();g++)
                if (mAllBlocks[i]->mGroupId == releasedGids[g])
                {
@@ -4117,7 +4156,7 @@ class GlobalAllocator
          }
       }
 
-      #ifdef HX_GC_VERIFY
+      #ifdef HXCPP_GC_VERIFY
       for(int i=0;i<mAllBlocks.size();i++)
       {
          BlockDataInfo *info = mAllBlocks[i];
@@ -4456,7 +4495,7 @@ class GlobalAllocator
       {
          waitForThreadWake(inId);
 
-         #ifdef HX_GC_VERIFY
+         #ifdef HXCPP_GC_VERIFY
          if (! (sRunningThreads & (1<<inId)) )
             printf("Bad running threads!\n");
          #endif
@@ -4769,7 +4808,7 @@ class GlobalAllocator
 
       hx::RunFinalizers();
 
-      #ifdef HX_GC_VERIFY
+      #ifdef HXCPP_GC_VERIFY
       for(int i=0;i<mAllBlocks.size();i++)
          mAllBlocks[i]->verify("After mark");
       #endif
@@ -4793,7 +4832,7 @@ class GlobalAllocator
    }
    #endif
 
-   #ifdef HX_GC_VERIFY
+   #ifdef HXCPP_GC_VERIFY
    void VerifyBlockOrder()
    {
       for(int i=1;i<mAllBlocks.size();i++)
@@ -5205,7 +5244,7 @@ class GlobalAllocator
 
             std::stable_sort(&mAllBlocks[0], &mAllBlocks[0] + mAllBlocks.size(), SortByBlockPtr );
 
-            #ifdef HX_GC_VERIFY
+            #ifdef HXCPP_GC_VERIFY
             VerifyBlockOrder();
             #endif
          }
@@ -5331,7 +5370,7 @@ class GlobalAllocator
       #endif
 
 
-      #ifdef HX_GC_VERIFY
+      #ifdef HXCPP_GC_VERIFY
       VerifyBlockOrder();
       #endif
 
@@ -5642,7 +5681,13 @@ void MarkConservative(int *inBottom, int *inTop,hx::MarkContext *__inCtx)
       void *vptr = *(void **)ptr;
 
       MemType mem;
-      if (vptr && !((size_t)vptr & 0x03) && vptr!=prev && vptr!=lastPin)
+      #ifdef HXCPP_ALIGN_ALLOC
+      const size_t validObjectMask = 0x07;
+      #else
+      const size_t validObjectMask = 0x03;
+      #endif
+
+      if (vptr && !((size_t)vptr & validObjectMask) && vptr!=prev && vptr!=lastPin)
       {
 
          #ifdef PROFILE_COLLECT
@@ -6178,37 +6223,26 @@ class LocalAllocator : public hx::StackContext
 
 
 
-
-
-
-
-
-
-
-
-
-
    void ExpandAlloc(int &ioSize)
    {
-      #ifdef HXCPP_GC_NURSERY
-      int spaceStart = spaceFirst - allocBase - 4;
-      int spaceEnd = spaceOversize - allocBase - 4;
-      #endif
-
-
-      int size = ioSize + sizeof(int);
       #ifdef HXCPP_ALIGN_ALLOC
-      // If we start in even-int offset, we need to skip 8 bytes to get alloc on even-int
-      if (size+spaceStart>spaceEnd || !(spaceStart & 7))
-         size += 4;
+      // Do nothing here - aligning to the end of the row will bump the
+      // next allocation, so it's not clear if its a good idea.
+      #else
+         #ifdef HXCPP_GC_NURSERY
+         int spaceStart = spaceFirst - allocBase - 4;
+         int spaceEnd = spaceOversize - allocBase - 4;
+         #endif
+
+         int size = ioSize + sizeof(int);
+         int end = spaceStart + size;
+         if (end <= spaceEnd)
+         {
+            int linePad = IMMIX_LINE_LEN - (end & (IMMIX_LINE_LEN-1));
+            if (linePad>0 && linePad<=64)
+               ioSize += linePad;
+         }
       #endif
-      int end = spaceStart + size;
-      if (end <= spaceEnd)
-      {
-         int linePad = IMMIX_LINE_LEN - (end & (IMMIX_LINE_LEN-1));
-         if (linePad>0 && linePad<=64)
-            ioSize += linePad;
-      }
    }
 
 
@@ -6226,20 +6260,13 @@ class LocalAllocator : public hx::StackContext
       if (inSize==0)
          return hx::emptyAlloc;
 
-      #if defined(HXCPP_VISIT_ALLOCS) && defined(HXCPP_M64)
+      #if defined(HXCPP_VISIT_ALLOCS) && (defined(HXCPP_M64)||defined(HXCPP_ARM64))
       // Make sure we can fit a relocation pointer
       int allocSize = sizeof(int) + std::max(8,inSize);
       #else
       int allocSize = sizeof(int) + inSize;
       #endif
 
-      #ifdef HXCPP_ALIGN_ALLOC
-      // If we start in even-int offset, we need to skip 8 bytes to get alloc on even-int
-      int skip4 = allocSize+spaceStart>spaceEnd || !(spaceStart & 7) ? 4 : 0;
-      #else
-      enum { skip4 = 0 };
-      #endif
-
       #if HXCPP_GC_DEBUG_LEVEL>0
       if (inSize & 3) DebuggerTrap();
       #endif
@@ -6248,6 +6275,10 @@ class LocalAllocator : public hx::StackContext
       {
          #ifdef HXCPP_GC_NURSERY
             unsigned char *buffer = spaceFirst;
+            #ifdef HXCPP_ALIGN_ALLOC
+            if ((size_t)buffer & 0x4 )
+               buffer += 4;
+            #endif
             unsigned char *end = buffer + allocSize;
 
             if ( end <= spaceOversize )
@@ -6268,13 +6299,14 @@ class LocalAllocator : public hx::StackContext
             if (s>spaceFirst && mFraggedRows)
                *mFraggedRows += (s - spaceFirst)>>IMMIX_LINE_BITS;
          #else
-            int end = spaceStart + allocSize + skip4;
+            #ifdef HXCPP_ALIGN_ALLOC
+            if (!((size_t)spaceStart & 0x4 ))
+               spaceStart += 4;
+            #endif
+
+            int end = spaceStart + allocSize;
             if (end <= spaceEnd)
             {
-               #ifdef HXCPP_ALIGN_ALLOC
-               spaceStart += skip4;
-               #endif
-
                unsigned int *buffer = (unsigned int *)(allocBase + spaceStart);
 
                int startRow = spaceStart>>IMMIX_LINE_BITS;
diff --git a/test/haxe/compile.hxml b/test/haxe/compile.hxml
index 2a23d789e..eedd52a3d 100644
--- a/test/haxe/compile.hxml
+++ b/test/haxe/compile.hxml
@@ -2,4 +2,4 @@
 -r TestMain.hx
 -D HXCPP_GC_GENERATIONAL
 -L utest
---cpp bin
\ No newline at end of file
+--cpp bin
diff --git a/toolchain/common-defines.xml b/toolchain/common-defines.xml
index 517d9b401..3f8c111a9 100644
--- a/toolchain/common-defines.xml
+++ b/toolchain/common-defines.xml
@@ -24,6 +24,7 @@
  <flag value="-DHXCPP_GC_CHECK_POINTER" if="HXCPP_GC_CHECK_POINTER" tag="haxe" />
  <flag value="-DHXCPP_GC_DEBUG_ALWAYS_MOVE" if="HXCPP_GC_DEBUG_ALWAYS_MOVE" tag="haxe" />
  <flag value="-DHXCPP_GC_GENERATIONAL" if="HXCPP_GC_GENERATIONAL" tag="haxe" />
+ <flag value="-DHXCPP_ALIGN_ALLOC" if="HXCPP_ALIGN_ALLOC" tag="haxe" />
  <flag value="-DHXCPP_GC_NURSERY" if="HXCPP_GC_NURSERY" tag="haxe" />
  <flag value="-DHXCPP_DLL_IMPORT" if="dll_import" tag="haxe" />
  <flag value="-I${dll_import_include}" if="dll_import_include" tag="haxe" />
@@ -36,6 +37,7 @@
  <flag value='-DHXCPP_API_LEVEL=0' unless="hxcpp_api_level"  tag="haxe"/>
  <flag value='-DHXCPP_OBJC' if="objc"  tag="haxe"/>
  <flag value='-DHXCPP_GC_DEBUG_LEVEL=${HXCPP_GC_DEBUG_LEVEL}' if="HXCPP_GC_DEBUG_LEVEL" tag="gc" />
+ <flag value='-DHXCPP_GC_VERIFY' if="HXCPP_GC_VERIFY" tag="gc" />
  <flag value='-DHXCPP_WINXP_COMPAT' if="HXCPP_WINXP_COMPAT" tag="haxe"/>
  <mmflag value='-Wno-invalid-offsetof' />
  <mmflag value='-Wno-format-security' />