1 files changed, 28 insertions, 13 deletions

diff --git a/src/lj_target_x86.h b/src/lj_target_x86.h
index 69aec37c..69cb8ca5 100644
--- a/src/lj_target_x86.h
+++ b/src/lj_target_x86.h
@@ -22,7 +22,7 @@
   _(XMM0) _(XMM1) _(XMM2) _(XMM3) _(XMM4) _(XMM5) _(XMM6) _(XMM7)
 #endif
 #define VRIDDEF(_) \
-  _(MRM)
+  _(MRM) _(RIP)
 
 #define RIDENUM(name)   RID_##name,
 
@@ -31,15 +31,16 @@ enum {
   FPRDEF(RIDENUM)               /* Floating-point registers (FPRs). */
   RID_MAX,
   RID_MRM = RID_MAX,            /* Pseudo-id for ModRM operand. */
+  RID_RIP = RID_MAX+5,          /* Pseudo-id for RIP (x64 only), rm bits = 5. */
 
   /* Calling conventions. */
+  RID_SP = RID_ESP,
   RID_RET = RID_EAX,
 #if LJ_64
   RID_FPRET = RID_XMM0,
-#else
+#endif
   RID_RETLO = RID_EAX,
   RID_RETHI = RID_EDX,
-#endif
 
   /* These definitions must match with the *.dasc file(s): */
   RID_BASE = RID_EDX,           /* Interpreter BASE. */
@@ -62,8 +63,10 @@ enum {
 
 /* -- Register sets ------------------------------------------------------- */
 
-/* Make use of all registers, except the stack pointer. */
-#define RSET_GPR        (RSET_RANGE(RID_MIN_GPR, RID_MAX_GPR)-RID2RSET(RID_ESP))
+/* Make use of all registers, except the stack pointer (and maybe DISPATCH). */
+#define RSET_GPR        (RSET_RANGE(RID_MIN_GPR, RID_MAX_GPR) \
+                         - RID2RSET(RID_ESP) \
+                         - LJ_GC64*RID2RSET(RID_DISPATCH))
 #define RSET_FPR        (RSET_RANGE(RID_MIN_FPR, RID_MAX_FPR))
 #define RSET_ALL        (RSET_GPR|RSET_FPR)
 #define RSET_INIT       RSET_ALL
@@ -131,7 +134,11 @@ 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
@@ -157,6 +164,8 @@ typedef struct {
 #define EXITSTUB_SPACING        (2+2)
 #define EXITSTUBS_PER_GROUP     32
 
+#define EXITTRACE_VMSTATE       1       /* g->vmstate has traceno on exit. */
+
 /* -- x86 ModRM operand encoding ------------------------------------------ */
 
 typedef enum {
@@ -184,12 +193,18 @@ typedef struct {
 #define XO_f20f(o)      ((uint32_t)(0x0ff2fc + (0x##o<<24)))
 #define XO_f30f(o)      ((uint32_t)(0x0ff3fc + (0x##o<<24)))
 
+#define XV_660f38(o)    ((uint32_t)(0x79e2c4 + (0x##o<<24)))
+#define XV_f20f38(o)    ((uint32_t)(0x7be2c4 + (0x##o<<24)))
+#define XV_f20f3a(o)    ((uint32_t)(0x7be3c4 + (0x##o<<24)))
+#define XV_f30f38(o)    ((uint32_t)(0x7ae2c4 + (0x##o<<24)))
+
 /* This list of x86 opcodes is not intended to be complete. Opcodes are only
 ** included when needed. Take a look at DynASM or jit.dis_x86 to see the
 ** whole mess.
 */
 typedef enum {
   /* Fixed length opcodes. XI_* prefix. */
+  XI_O16 =      0x66,
   XI_NOP =      0x90,
   XI_XCHGa =    0x90,
   XI_CALL =     0xe8,
@@ -207,26 +222,28 @@ typedef enum {
   XI_PUSHi8 =   0x6a,
   XI_TESTb =    0x84,
   XI_TEST =     0x85,
+  XI_INT3 =     0xcc,
   XI_MOVmi =    0xc7,
   XI_GROUP5 =   0xff,
 
   /* Note: little-endian byte-order! */
   XI_FLDZ =     0xeed9,
   XI_FLD1 =     0xe8d9,
-  XI_FLDLG2 =   0xecd9,
-  XI_FLDLN2 =   0xedd9,
   XI_FDUP =     0xc0d9,  /* Really fld st0. */
   XI_FPOP =     0xd8dd,  /* Really fstp st0. */
   XI_FPOP1 =    0xd9dd,  /* Really fstp st1. */
   XI_FRNDINT =  0xfcd9,
-  XI_FSIN =     0xfed9,
-  XI_FCOS =     0xffd9,
-  XI_FPTAN =    0xf2d9,
-  XI_FPATAN =   0xf3d9,
   XI_FSCALE =   0xfdd9,
   XI_FYL2X =    0xf1d9,
 
+  /* VEX-encoded instructions. XV_* prefix. */
+  XV_RORX =     XV_f20f3a(f0),
+  XV_SARX =     XV_f30f38(f7),
+  XV_SHLX =     XV_660f38(f7),
+  XV_SHRX =     XV_f20f38(f7),
+
   /* Variable-length opcodes. XO_* prefix. */
+  XO_OR =       XO_(0b),
   XO_MOV =      XO_(8b),
   XO_MOVto =    XO_(89),
   XO_MOVtow =   XO_66(89),
@@ -277,10 +294,8 @@ typedef enum {
   XO_ROUNDSD =  0x0b3a0ffc,  /* Really 66 0f 3a 0b. See asm_fpmath. */
   XO_UCOMISD =  XO_660f(2e),
   XO_CVTSI2SD = XO_f20f(2a),
-  XO_CVTSD2SI = XO_f20f(2d),
   XO_CVTTSD2SI= XO_f20f(2c),
   XO_CVTSI2SS = XO_f30f(2a),
-  XO_CVTSS2SI = XO_f30f(2d),
   XO_CVTTSS2SI= XO_f30f(2c),
   XO_CVTSS2SD = XO_f30f(5a),
   XO_CVTSD2SS = XO_f20f(5a),