x64: Add LJ_GC64 mode interpreter.

Enable this mode with: make XCFLAGS=-DLUAJIT_ENABLE_GC64

src/lib_base.c

@@ -428,7 +428,7 @@ LJLIB_CF(dofile)
 
 LJLIB_CF(gcinfo)
 {
-  setintV(L->top++, (G(L)->gc.total >> 10));
+  setintV(L->top++, (int32_t)(G(L)->gc.total >> 10));
   return 1;
 }
 

src/lj_arch.h

@@ -158,6 +158,9 @@
 #define LJ_TARGET_MASKROT	1
 #define LJ_TARGET_UNALIGNED	1
 #define LJ_ARCH_NUMMODE		LJ_NUMMODE_SINGLE_DUAL
+#ifdef LUAJIT_ENABLE_GC64
+#define LJ_TARGET_GC64		1
+#endif
 
 #elif LUAJIT_TARGET == LUAJIT_ARCH_ARM
 

src/lj_ccallback.c

@@ -35,7 +35,7 @@
 #elif LJ_TARGET_X86ORX64
 
 #define CALLBACK_MCODE_HEAD	(LJ_64 ? 8 : 0)
-#define CALLBACK_MCODE_GROUP	(-2+1+2+5+(LJ_64 ? 6 : 5))
+#define CALLBACK_MCODE_GROUP	(-2+1+2+(LJ_GC64 ? 10 : 5)+(LJ_64 ? 6 : 5))
 
 #define CALLBACK_SLOT2OFS(slot) \
   (CALLBACK_MCODE_HEAD + CALLBACK_MCODE_GROUP*((slot)/32) + 4*(slot))
@@ -120,8 +120,13 @@ static void callback_mcode_init(global_State *g, uint8_t *page)
       /* push ebp/rbp; mov ah, slot>>8; mov ebp, &g. */
       *p++ = XI_PUSH + RID_EBP;
       *p++ = XI_MOVrib | (RID_EAX+4); *p++ = (uint8_t)(slot >> 8);
+#if LJ_GC64
+      *p++ = 0x48; *p++ = XI_MOVri | RID_EBP;
+      *(uint64_t *)p = (uint64_t)(g); p += 8;
+#else
       *p++ = XI_MOVri | RID_EBP;
       *(int32_t *)p = i32ptr(g); p += 4;
+#endif
 #if LJ_64
       /* jmp [rip-pageofs] where lj_vm_ffi_callback is stored. */
       *p++ = XI_GROUP5; *p++ = XM_OFS0 + (XOg_JMP<<3) + RID_EBP;

src/lj_cdata.c

@@ -121,7 +121,12 @@ collect_attrib:
     idx = (ptrdiff_t)intV(key);
     goto integer_key;
   } else if (tvisnum(key)) {  /* Numeric key. */
-    idx = LJ_64 ? (ptrdiff_t)numV(key) : (ptrdiff_t)lj_num2int(numV(key));
+#ifdef _MSC_VER
+    /* Workaround for MSVC bug. */
+    volatile
+#endif
+    lua_Number n = numV(key);
+    idx = LJ_64 ? (ptrdiff_t)n : (ptrdiff_t)lj_num2int(n);
   integer_key:
     if (ctype_ispointer(ct->info)) {
       CTSize sz = lj_ctype_size(cts, ctype_cid(ct->info));  /* Element size. */

src/lj_frame.h

@@ -127,21 +127,37 @@ enum { LJ_CONT_TAILCALL, LJ_CONT_FFI_CALLBACK };  /* Special continuations. */
 #elif LJ_TARGET_X64
 #if LJ_ABI_WIN
 #define CFRAME_OFS_PREV		(13*8)
+#if LJ_GC64
+#define CFRAME_OFS_PC		(12*8)
+#define CFRAME_OFS_L		(11*8)
+#define CFRAME_OFS_ERRF		(21*4)
+#define CFRAME_OFS_NRES		(20*4)
+#define CFRAME_OFS_MULTRES	(8*4)
+#else
 #define CFRAME_OFS_PC		(25*4)
 #define CFRAME_OFS_L		(24*4)
 #define CFRAME_OFS_ERRF		(23*4)
 #define CFRAME_OFS_NRES		(22*4)
 #define CFRAME_OFS_MULTRES	(21*4)
+#endif
 #define CFRAME_SIZE		(10*8)
 #define CFRAME_SIZE_JIT		(CFRAME_SIZE + 9*16 + 4*8)
 #define CFRAME_SHIFT_MULTRES	0
 #else
 #define CFRAME_OFS_PREV		(4*8)
+#if LJ_GC64
+#define CFRAME_OFS_PC		(3*8)
+#define CFRAME_OFS_L		(2*8)
+#define CFRAME_OFS_ERRF		(3*4)
+#define CFRAME_OFS_NRES		(2*4)
+#define CFRAME_OFS_MULTRES	(0*4)
+#else
 #define CFRAME_OFS_PC		(7*4)
 #define CFRAME_OFS_L		(6*4)
 #define CFRAME_OFS_ERRF		(5*4)
 #define CFRAME_OFS_NRES		(4*4)
 #define CFRAME_OFS_MULTRES	(1*4)
+#endif
 #define CFRAME_SIZE		(10*8)
 #define CFRAME_SIZE_JIT		(CFRAME_SIZE + 16)
 #define CFRAME_SHIFT_MULTRES	0

src/lj_gdbjit.c

@@ -556,8 +556,8 @@ static void LJ_FASTCALL gdbjit_ehframe(GDBJITctx *ctx)
     DB(DW_CFA_offset|DW_REG_15); DUV(4);
     DB(DW_CFA_offset|DW_REG_14); DUV(5);
     /* Extra registers saved for JIT-compiled code. */
-    DB(DW_CFA_offset|DW_REG_13); DUV(9);
-    DB(DW_CFA_offset|DW_REG_12); DUV(10);
+    DB(DW_CFA_offset|DW_REG_13); DUV(LJ_GC64 ? 10 : 9);
+    DB(DW_CFA_offset|DW_REG_12); DUV(LJ_GC64 ? 11 : 10);
 #elif LJ_TARGET_ARM
     {
       int i;

src/lj_target_x86.h

@@ -132,7 +132,11 @@ enum {
 #define SPS_FIXED	(4*2)
 #define SPS_FIRST	(4*2)	/* Don't use callee register save area. */
 #else
+#if LJ_GC64
+#define SPS_FIXED	2
+#else
 #define SPS_FIXED	4
+#endif
 #define SPS_FIRST	2
 #endif
 #else