Reapply "Create a single copy of stub templates

src/coreclr/clrdefinitions.cmake

@@ -218,6 +218,11 @@ if (FEATURE_STUBPRECODE_DYNAMIC_HELPERS)
   add_definitions(-DFEATURE_STUBPRECODE_DYNAMIC_HELPERS)
 endif()
 
+if (CLR_CMAKE_TARGET_APPLE)
+#  Re-enable when dbgshim containing https://github.com/dotnet/diagnostics/pull/5487 is generally available
+#  add_definitions(-DFEATURE_MAP_THUNKS_FROM_IMAGE)
+endif()
+
 # Use this function to enable building with a specific target OS and architecture set of defines
 # This is known to work for the set of defines used by the JIT and gcinfo, it is not likely correct for
 # other components of the runtime

src/coreclr/inc/executableallocator.h

@@ -182,6 +182,9 @@ class ExecutableAllocator
     // Return true if double mapping is enabled.
     static bool IsDoubleMappingEnabled();
 
+    // Release memory allocated via DoubleMapping for either templates or normal double mapped data
+    void ReleaseWorker(void* pRX, bool releaseTemplate);
+
     // Initialize the allocator instance
     bool Initialize();
 
@@ -262,6 +265,18 @@ class ExecutableAllocator
 
     // Unmap the RW mapping at the specified address
     void UnmapRW(void* pRW);
+
+    // Allocate thunks from a template. pTemplate is the return value from CreateTemplate
+    void* AllocateThunksFromTemplate(void *pTemplate, size_t templateSize);
+
+    // Free a set of thunks allocated from templates. pThunks must have been returned from AllocateThunksFromTemplate
+    void FreeThunksFromTemplate(void *pThunks, size_t templateSize);
+
+    // Create a template
+    // If templateInImage is not null, it will attempt to use it as the template, otherwise it will create an temporary in memory file to serve as the template
+    // Some OS/Architectures may/may not be able to work with this, so this api is permitted to return NULL, and callers should have an alternate approach using
+    // the codePageGenerator directly.
+    void* CreateTemplate(void* templateInImage, size_t templateSize, void (*codePageGenerator)(uint8_t* pageBase, uint8_t* pageBaseRX, size_t size));
 };
 
 #define ExecutableWriterHolder ExecutableWriterHolderNoLog

src/coreclr/inc/loaderheap.h

@@ -455,10 +455,19 @@ class UnlockedLoaderHeap : public UnlockedLoaderHeapBase
     static void WeGotAFaultNowWhat(UnlockedLoaderHeap *pHeap);
 };
 
+struct InterleavedLoaderHeapConfig
+{
+    uint32_t StubSize;
+    void* Template;
+    void (*CodePageGenerator)(uint8_t* pageBase, uint8_t* pageBaseRX, size_t size);
+};
+
+void InitializeLoaderHeapConfig(InterleavedLoaderHeapConfig *pConfig, size_t stubSize, void* templateInImage, void (*codePageGenerator)(uint8_t* pageBase, uint8_t* pageBaseRX, size_t size));
+
 //===============================================================================
 // This is the base class for InterleavedLoaderHeap It's used as a simple
 // allocator for stubs in a scheme where each stub is a small fixed size, and is paired
-// with memory which is GetOSStubPageSize() bytes away. In addition there is an
+// with memory which is GetStubCodePageSize() bytes away. In addition there is an
 // ability to free is via a "backout" mechanism that is not considered to have good performance.
 //
 //===============================================================================
@@ -492,16 +501,13 @@ class UnlockedInterleavedLoaderHeap : public UnlockedLoaderHeapBase
 
     InterleavedStubFreeListNode  *m_pFreeListHead;
 
-public:
-public:
-    void                (*m_codePageGenerator)(BYTE* pageBase, BYTE* pageBaseRX, SIZE_T size);
+    const InterleavedLoaderHeapConfig *m_pConfig;
 
 #ifndef DACCESS_COMPILE
 protected:
     UnlockedInterleavedLoaderHeap(
         RangeList *pRangeList,
-        void (*codePageGenerator)(BYTE* pageBase, BYTE* pageBaseRX, SIZE_T size),
-        DWORD dwGranularity);
+        const InterleavedLoaderHeapConfig *pConfig);
 
     virtual ~UnlockedInterleavedLoaderHeap();
 #endif
@@ -1039,13 +1045,11 @@ class InterleavedLoaderHeap : public UnlockedInterleavedLoaderHeap
 public:
     InterleavedLoaderHeap(RangeList *pRangeList,
                BOOL fUnlocked,
-               void (*codePageGenerator)(BYTE* pageBase, BYTE* pageBaseRX, SIZE_T size),
-               DWORD dwGranularity
+               const InterleavedLoaderHeapConfig *pConfig
                )
       : UnlockedInterleavedLoaderHeap(
                            pRangeList,
-                           codePageGenerator,
-                           dwGranularity),
+                           pConfig),
         m_CriticalSection(fUnlocked ? NULL : CreateLoaderHeapLock())
     {
         WRAPPER_NO_CONTRACT;