ok, I realized we need to realloc buffers allocated by FCEU_malloc (why didn't I guess that..) which makes the fact that they're aligned be horrible. so I added FCEU_amalloc and FCEU_afree instead to do aligned allocs and frees.

3 files changed, 61 insertions, 37 deletions

diff --git a/src/utils/memory.cpp b/src/utils/memory.cpp
index 93bc3af9..fa5d65b4 100644
--- a/src/utils/memory.cpp
+++ b/src/utils/memory.cpp
@@ -28,18 +28,34 @@
 #include "../fceu.h"
 #include "memory.h"
 
+void *FCEU_amalloc(size_t size, size_t alignment)

+{

+	size = (size + alignment - 1) & ~(alignment-1);

+

+	#ifdef _MSC_VER

+	void *ret = _aligned_malloc(size,alignment);

+	#else

+	void *ret = aligned_alloc(alignment,size);

+	#endif

+

+	if(!ret)

+		FCEU_abort("Error allocating memory!");

+

+	return ret;

+}
+
+void FCEU_afree(void* ptr)

+{

+	#ifdef _MSC_VER

+	_aligned_free(ptr);

+	#else

+	free(ptr);

+	#endif

+}
+
 static void *_FCEU_malloc(uint32 size)
 {
-	//do not add an aligned allocation function. if a larger alignment is needed, change this constant to use it for all allocations.
-	static const int alignment = 32;
-	
-	size = (size + alignment - 1) & ~(alignment-1);
-
-	#ifdef _MSC_VER
-	void *ret = _aligned_malloc(size,alignment);
-	#else
-	void *ret = aligned_alloc(alignment,size);
-	#endif
+	void* ret = malloc(size);
 
 	if(!ret)
 		FCEU_abort("Error allocating memory!");
@@ -49,22 +65,17 @@ static void *_FCEU_malloc(uint32 size)
 
 static void _FCEU_free(void* ptr)
 {
-	#ifdef _MSC_VER
-	_aligned_free(ptr);
-	#else
 	free(ptr);
-	#endif
 }
 
-///allocates the specified number of bytes. exits process if this fails
 void *FCEU_gmalloc(uint32 size)
 {
- void *ret = _FCEU_malloc(size);
+	void *ret = _FCEU_malloc(size);
  
- // initialize according to RAMInitOption, default zero
- FCEU_MemoryRand((uint8*)ret,size,true);
+	// initialize according to RAMInitOption, default zero
+	FCEU_MemoryRand((uint8*)ret,size,true);
 
- return ret;
+	return ret;
 }
 
 void *FCEU_malloc(uint32 size)
@@ -74,13 +85,11 @@ void *FCEU_malloc(uint32 size)
 	return ret;
 }
 
-//frees memory allocated with FCEU_gmalloc
 void FCEU_gfree(void *ptr)
 {
 	_FCEU_free(ptr);
 }
 
-//frees memory allocated with FCEU_malloc
 void FCEU_free(void *ptr)
 {
 	_FCEU_free(ptr);
@@ -96,8 +105,13 @@ void FCEU_dfree(void *ptr)
 	return FCEU_free(ptr);
 }
 
-void FCEU_abort(const char* message)

-{

+void* FCEU_realloc(void* ptr, size_t size)
+{
+	return realloc(ptr,size);
+}
+
+void FCEU_abort(const char* message)
+{
 	if(message) FCEU_PrintError(message);
 	abort();
-}

+}
diff --git a/src/utils/memory.h b/src/utils/memory.h
index b9e6f145..4fd97d9d 100644
--- a/src/utils/memory.h
+++ b/src/utils/memory.h
@@ -24,19 +24,29 @@
 
 #define FCEU_dwmemset(d,c,n) {int _x; for(_x=n-4;_x>=0;_x-=4) *(uint32 *)&(d)[_x]=c;}
 
-//returns a 32-aligned buffer, initialized to 0
-void *FCEU_malloc(uint32 size); 
+//returns a buffer initialized to 0
+void *FCEU_malloc(uint32 size);
 
-//returns a 32-aligned buffer, with jumbled initial contents
+//returns a buffer, with jumbled initial contents
 //used by boards for WRAM etc, initialized to 0 (default) or other via RAMInitOption
-void *FCEU_gmalloc(uint32 size); 
+void *FCEU_gmalloc(uint32 size);
 
 //free memory allocated with FCEU_gmalloc
 void FCEU_gfree(void *ptr);
 
-//free memory allocated with 
+//returns an aligned buffer, initialized to 0
+//the alignment will default to the largest thing you could ever sensibly want for massively aligned cache friendly buffers
+void *FCEU_amalloc(size_t size, size_t alignment = 256);
+
+//frees memory allocated with FCEU_amalloc
+void FCEU_afree(void* ptr);
+
+//free memory allocated with FCEU_malloc
 void FCEU_free(void *ptr);
 
+//reallocate memory allocated with FCEU_malloc
+void* FCEU_realloc(void* ptr, size_t size);
+
 //don't use these. change them if you find them.
 void *FCEU_dmalloc(uint32 size);
 
diff --git a/src/video.cpp b/src/video.cpp
index cbad58fc..20a48e76 100644
--- a/src/video.cpp
+++ b/src/video.cpp
@@ -91,19 +91,19 @@ void FCEU_KillVirtualVideo(void)
 {

 	if ( XBuf )

 	{

-		FCEU_free(XBuf); XBuf = NULL;

+		FCEU_afree(XBuf); XBuf = NULL;

 	}

 	if ( XBackBuf )

 	{

-		FCEU_free(XBackBuf); XBackBuf = NULL;

+		FCEU_afree(XBackBuf); XBackBuf = NULL;

 	}

 	if ( XDBuf )

 	{

-		FCEU_free(XDBuf); XDBuf = NULL;

+		FCEU_afree(XDBuf); XDBuf = NULL;

 	}

 	if ( XDBackBuf )

 	{

-		FCEU_free(XDBackBuf); XDBackBuf = NULL;

+		FCEU_afree(XDBackBuf); XDBackBuf = NULL;

 	}

 	//printf("Video Core Cleanup\n");

 }

@@ -120,10 +120,10 @@ int FCEU_InitVirtualVideo(void)
 	if(XBuf)

 		return 1;

 	

-	XBuf = (u8*)FCEU_malloc(256 * 256);

-	XBackBuf = (u8*)FCEU_malloc(256 * 256);

-	XDBuf = (u8*)FCEU_malloc(256 * 256);

-	XDBackBuf = (u8*)FCEU_malloc(256 * 256);

+	XBuf = (u8*)FCEU_amalloc(256 * 256);

+	XBackBuf = (u8*)FCEU_amalloc(256 * 256);

+	XDBuf = (u8*)FCEU_amalloc(256 * 256);

+	XDBackBuf = (u8*)FCEU_amalloc(256 * 256);

 

 

 	xbsave = XBuf;