Reordered some CPU variables and rewrote codegen emitted by the old recompiler's 64-bit version of codegen_accumulate_flush() to not assume 32-bit pointers, fixes 64-bit old recompiler, closes #1615.

src/codegen/codegen_accumulate_x86-64.c

@@ -48,11 +48,16 @@ void codegen_accumulate(int acc_reg, int delta)
 void codegen_accumulate_flush(void)
 {
 	if (acc_regs[0].count) {
-		addbyte(0x81); /*ADD $acc_regs[0].count,acc_regs[0].dest*/
-		addbyte(0x04);
-		addbyte(0x25);
-		addlong((uint32_t) acc_regs[0].dest_reg);
+		addbyte(0x55); /*push rbp*/
+		addbyte(0x48); /*mov rbp,val*/
+		addbyte(0xbd);
+		addlong((uint32_t) (acc_regs[0].dest_reg & 0xffffffffULL));
+		addlong((uint32_t) (acc_regs[0].dest_reg >> 32ULL));
+		addbyte(0x81); /* add d,[rbp][0],val */
+		addbyte(0x45);
+		addbyte(0x00);
 		addlong(acc_regs[0].count);
+		addbyte(0x5d); /*pop rbp*/
 	}
 
 	acc_regs[0].count = 0;

src/cpu/cpu.c

@@ -69,6 +69,9 @@ enum {
 #define CPUID_3DNOW	(1UL << 31UL)
 
 
+/* Make sure this is as low as possible. */
+cpu_state_t	cpu_state;
+
 #ifdef USE_DYNAREC
 const OpFn	*x86_dynarec_opcodes, *x86_dynarec_opcodes_0f,
 		*x86_dynarec_opcodes_d8_a16, *x86_dynarec_opcodes_d8_a32,
@@ -128,8 +131,6 @@ double		cpu_dmulti;
 
 msr_t		msr;
 
-cpu_state_t	cpu_state;
-
 cyrix_t		cyrix;
 
 cpu_family_t	*cpu_f;

src/cpu/cpu.h

@@ -461,6 +461,8 @@ COMPILE_TIME_ASSERT(sizeof(cpu_state_t) <= 128)
 
 
 /* Global variables. */
+extern cpu_state_t	cpu_state;
+
 extern const cpu_family_t cpu_families[];
 extern const cpu_legacy_machine_t cpu_legacy_table[];
 extern cpu_family_t *cpu_f;
@@ -505,7 +507,6 @@ extern uint32_t		cpu_cur_status;
 extern uint64_t		cpu_CR4_mask;
 extern uint64_t		tsc;
 extern msr_t		msr;
-extern cpu_state_t	cpu_state;
 extern uint8_t		opcode;
 extern int		cgate16;
 extern int		cpl_override;