Move dynarec code memory out of codeblock structure.

src/codegen.h

@@ -30,7 +30,10 @@
 */
 
 /*Hack until better memory management written*/
-#define BLOCK_DATA_SIZE 0x10000
+/*This is deliberately _not_ a power of two, to avoid cache aliasing problems.
+  Try changing this to a power of two, and watch the performance plummet :)
+  It's probably best for this to be a multiple of the cache line size though*/
+#define BLOCK_DATA_SIZE 0xff80
 
 typedef struct codeblock_t
 {
@@ -61,9 +64,11 @@ typedef struct codeblock_t
         uint32_t status;
         uint32_t flags;
 
-        uint8_t data[BLOCK_DATA_SIZE];
+        uint8_t *data;
 } codeblock_t;
 
+uint8_t *codeblock_data;
+
 /*Code block uses FPU*/
 #define CODEBLOCK_HAS_FPU 1
 /*Code block is always entered with the same FPU top-of-stack*/

src/codegen_backend.c

@@ -96,7 +96,10 @@ void codegen_reset()
         mem_reset_page_blocks();
         
         for (c = 0; c < BLOCK_SIZE; c++)
+        {
                 codeblock[c].pc = BLOCK_PC_INVALID;
+                codeblock[c].data = &codeblock_data[c * BLOCK_DATA_SIZE];
+        }
 }
 
 void dump_block()

src/codegen_backend_arm.c

@@ -294,23 +294,27 @@ void codegen_backend_init()
 #endif
 
 #if defined WIN32 || defined _WIN32 || defined _WIN32
-        codeblock = VirtualAlloc(NULL, (BLOCK_SIZE+1) * sizeof(codeblock_t), MEM_COMMIT, PAGE_EXECUTE_READWRITE);
+        codeblock_data = VirtualAlloc(NULL, (BLOCK_SIZE+1) * BLOCK_DATA_SIZE, MEM_COMMIT, PAGE_EXECUTE_READWRITE);
 #else
-        codeblock = malloc((BLOCK_SIZE+1) * sizeof(codeblock_t));
+        codeblock_data = malloc((BLOCK_SIZE+1) * BLOCK_DATA_SIZE);
 #endif
-	if (!codeblock)
+	if (!codeblock_data)
-		fatal("codeblock failed to alloc - %i\n", (BLOCK_SIZE+1) * sizeof(codeblock_t));
+		fatal("codeblock_data failed to alloc - %i\n", (BLOCK_SIZE+1) * BLOCK_DATA_SIZE);
+	codeblock = malloc((BLOCK_SIZE+1) * sizeof(codeblock_t));
         codeblock_hash = malloc(HASH_SIZE * sizeof(codeblock_t *));
 
         memset(codeblock, 0, (BLOCK_SIZE+1) * sizeof(codeblock_t));
         memset(codeblock_hash, 0, HASH_SIZE * sizeof(codeblock_t *));
 
-        for (c = 0; c < BLOCK_SIZE; c++)
+        for (c = 0; c < BLOCK_SIZE+1; c++)
+	{
                 codeblock[c].pc = BLOCK_PC_INVALID;
+		codeblock[c].data = &codeblock_data[c * BLOCK_DATA_SIZE];
+	}
 
 #if defined(__linux__) || defined(__APPLE__)
-	start = (void *)((long)codeblock & pagemask);
+	start = (void *)((long)codeblock_data & pagemask);
-	len = (((BLOCK_SIZE+1) * sizeof(codeblock_t)) + pagesize) & pagemask;
+	len = (((BLOCK_SIZE+1) * BLOCK_DATA_SIZE) + pagesize) & pagemask;
 	if (mprotect(start, len, PROT_READ | PROT_WRITE | PROT_EXEC) != 0)
 	{
 		perror("mprotect");

src/codegen_backend_arm64.c

@@ -289,23 +289,27 @@ void codegen_backend_init()
 #endif
 
 #if defined WIN32 || defined _WIN32 || defined _WIN32
-        codeblock = VirtualAlloc(NULL, (BLOCK_SIZE+1) * sizeof(codeblock_t), MEM_COMMIT, PAGE_EXECUTE_READWRITE);
+        codeblock_data = VirtualAlloc(NULL, (BLOCK_SIZE+1) * BLOCK_DATA_SIZE, MEM_COMMIT, PAGE_EXECUTE_READWRITE);
 #else
-        codeblock = malloc((BLOCK_SIZE+1) * sizeof(codeblock_t));
+        codeblock_data = malloc((BLOCK_SIZE+1) * BLOCK_DATA_SIZE);
 #endif
-	if (!codeblock)
+	if (!codeblock_data)
-		fatal("codeblock failed to alloc - %i\n", (BLOCK_SIZE+1) * sizeof(codeblock_t));
+		fatal("codeblock_data failed to alloc - %i\n", (BLOCK_SIZE+1) * BLOCK_DATA_SIZE);
+	codeblock = malloc((BLOCK_SIZE+1) * sizeof(codeblock_t));
         codeblock_hash = malloc(HASH_SIZE * sizeof(codeblock_t *));
 
         memset(codeblock, 0, (BLOCK_SIZE+1) * sizeof(codeblock_t));
         memset(codeblock_hash, 0, HASH_SIZE * sizeof(codeblock_t *));
 
-        for (c = 0; c < BLOCK_SIZE; c++)
+        for (c = 0; c < BLOCK_SIZE+1; c++)
+	{
                 codeblock[c].pc = BLOCK_PC_INVALID;
+		codeblock[c].data = &codeblock_data[c * BLOCK_SIZE];
+	}
 
 #if defined(__linux__) || defined(__APPLE__)
-	start = (void *)((long)codeblock & pagemask);
+	start = (void *)((long)codeblock_data & pagemask);
-	len = (((BLOCK_SIZE+1) * sizeof(codeblock_t)) + pagesize) & pagemask;
+	len = (((BLOCK_SIZE+1) * BLOCK_DATA_SIZE) + pagesize) & pagemask;
 	if (mprotect(start, len, PROT_READ | PROT_WRITE | PROT_EXEC) != 0)
 	{
 		perror("mprotect");

src/codegen_backend_x86-64.c

@@ -276,17 +276,21 @@ void codegen_backend_init()
 #endif
 
 #if defined WIN32 || defined _WIN32 || defined _WIN32
-        codeblock = VirtualAlloc(NULL, (BLOCK_SIZE+1) * sizeof(codeblock_t), MEM_COMMIT, PAGE_EXECUTE_READWRITE);
+        codeblock_data = VirtualAlloc(NULL, (BLOCK_SIZE+1) * BLOCK_DATA_SIZE, MEM_COMMIT, PAGE_EXECUTE_READWRITE);
 #else
-        codeblock = malloc((BLOCK_SIZE+1) * sizeof(codeblock_t));
+        codeblock_data = malloc((BLOCK_SIZE+1) * BLOCK_DATA_SIZE);
 #endif
+        codeblock = malloc((BLOCK_SIZE+1) * sizeof(codeblock_t));
         codeblock_hash = malloc(HASH_SIZE * sizeof(codeblock_t *));
 
         memset(codeblock, 0, (BLOCK_SIZE+1) * sizeof(codeblock_t));
         memset(codeblock_hash, 0, HASH_SIZE * sizeof(codeblock_t *));
 
-        for (c = 0; c < BLOCK_SIZE; c++)
+        for (c = 0; c < BLOCK_SIZE+1; c++)
+        {
+                codeblock[c].data = &codeblock_data[c * BLOCK_DATA_SIZE];
                 codeblock[c].pc = BLOCK_PC_INVALID;
+        }
 
 #if defined(__linux__) || defined(__APPLE__)
 	start = (void *)((long)codeblock & pagemask);

src/codegen_backend_x86-64.h

@@ -10,7 +10,7 @@
 #define HASH(l) ((l) & 0x1ffff)
 
 /*Hack until better memory management written*/
-#define BLOCK_EXIT_OFFSET 0xffe0
+#define BLOCK_EXIT_OFFSET (BLOCK_DATA_SIZE - 32)
 /*#define BLOCK_EXIT_OFFSET 0x7f0*/
 #define BLOCK_GPF_OFFSET (BLOCK_EXIT_OFFSET - 20)
 

src/codegen_backend_x86.c

@@ -261,22 +261,26 @@ void codegen_backend_init()
 	long pagemask = ~(pagesize - 1);
 #endif
 
-#if defined WIN32 || defined _WIN32 || defined _WIN32
-        codeblock = VirtualAlloc(NULL, (BLOCK_SIZE+1) * sizeof(codeblock_t), MEM_COMMIT, PAGE_EXECUTE_READWRITE);
-#else
         codeblock = malloc((BLOCK_SIZE+1) * sizeof(codeblock_t));
+#if defined WIN32 || defined _WIN32 || defined _WIN32
+        codeblock_data = VirtualAlloc(NULL, (BLOCK_SIZE+1) * BLOCK_DATA_SIZE, MEM_COMMIT, PAGE_EXECUTE_READWRITE);
+#else
+        codeblock_data = malloc((BLOCK_SIZE+1) * BLOCK_DATA_SIZE;
 #endif
         codeblock_hash = malloc(HASH_SIZE * sizeof(codeblock_t *));
 
         memset(codeblock, 0, (BLOCK_SIZE+1) * sizeof(codeblock_t));
         memset(codeblock_hash, 0, HASH_SIZE * sizeof(codeblock_t *));
 
-        for (c = 0; c < BLOCK_SIZE; c++)
+        for (c = 0; c < BLOCK_SIZE + 1; c++)
+        {
+                codeblock[c].data = &codeblock_data[c * BLOCK_DATA_SIZE];
                 codeblock[c].pc = BLOCK_PC_INVALID;
+        }
 
 #if defined(__linux__) || defined(__APPLE__)
-	start = (void *)((long)codeblock & pagemask);
+	start = (void *)((long)codeblock_data & pagemask);
-	len = (((BLOCK_SIZE+1) * sizeof(codeblock_t)) + pagesize) & pagemask;
+	len = (((BLOCK_SIZE+1) * BLOCK_DATA_SIZE) + pagesize) & pagemask;
 	if (mprotect(start, len, PROT_READ | PROT_WRITE | PROT_EXEC) != 0)
 	{
 		perror("mprotect");

src/codegen_backend_x86.h

@@ -10,7 +10,8 @@
 #define HASH(l) ((l) & 0x1ffff)
 
 /*Hack until better memory management written*/
-#define BLOCK_EXIT_OFFSET 0xfff0
+#define BLOCK_EXIT_OFFSET (BLOCK_DATA_SIZE - 16)
+//#define BLOCK_EXIT_OFFSET 0xfff0
 /*#define BLOCK_EXIT_OFFSET 0x7f0*/
 #define BLOCK_GPF_OFFSET (BLOCK_EXIT_OFFSET - 20)