MIPS64, part 1: Add MIPS64 support to interpreter.

Contributed by Djordje Kovacevic and Stefan Pejic from RT-RK.com. Sponsored by Cisco Systems, Inc.
2025-02-07 15:14:08 +00:00 · 2016-05-28 05:10:55 +02:00 · 2016-05-28 05:10:55 +02:00 · d9986fbadb
commit d9986fbadb
parent e3c4c9af0f
19 changed files with 5213 additions and 62 deletions
--- a/doc/ext_jit.html
+++ b/doc/ext_jit.html
@ -153,7 +153,7 @@ Contains the target OS name:
 <h3 id="jit_arch"><tt>jit.arch</tt></h3>
 <p>
 Contains the target architecture name:
-"x86", "x64", "arm", "ppc", or "mips".
+"x86", "x64", "arm", "arm64", "ppc", "mips" or "mips64".
 </p>

 <h2 id="jit_opt"><tt>jit.opt.*</tt> &mdash; JIT compiler optimization control</h2>
--- a/doc/install.html
+++ b/doc/install.html
@ -148,7 +148,7 @@ operating systems, CPUs and compilers:
 <td class="compatos">XEDK (<a href="#xbox360">Xbox 360</a>)</td>
 </tr>
 <tr class="even">
-<td class="compatcpu"><a href="#cross2">MIPS</a></td>
+<td class="compatcpu"><a href="#cross2">MIPS32<br>MIPS64</a></td>
 <td class="compatos">GCC 4.3+</td>
 <td class="compatos">GCC 4.3+</td>
 <td class="compatos compatno">&nbsp;</td>
@ -386,7 +386,7 @@ important to compile with the proper CPU or architecture settings:
 <li>The best way to get consistent results is to specify the correct settings when building the toolchain yourself.</li>
 <li>For a pre-built, generic toolchain add <tt>-mcpu=...</tt> or <tt>-march=...</tt> and other necessary flags to <tt>TARGET_CFLAGS</tt>.</li>
 <li>For ARM it's important to have the correct <tt>-mfloat-abi=...</tt> setting, too. Otherwise LuaJIT may not run at the full performance of your target CPU.</li>
-<li>For MIPS it's important to select a supported ABI (o32 on MIPS32) and consistently compile your project either with hard-float or soft-float compiler settings. Do not use <tt>-mips16</tt>.</li>
+<li>For MIPS it's important to select a supported ABI (o32 on MIPS32, n64 on MIPS64) and consistently compile your project either with hard-float or soft-float compiler settings. Do not use <tt>-mips16</tt>.</li>
 </ul>
 <p>
 Here are some examples for targets with a different CPU than the host:
@ -409,10 +409,15 @@ make CROSS=aarch64-linux-
 # PPC
 make HOST_CC="gcc -m32" CROSS=powerpc-linux-gnu-

-# MIPS big-endian
+# MIPS32 big-endian
 make HOST_CC="gcc -m32" CROSS=mips-linux-
-# MIPS little-endian
+# MIPS32 little-endian
 make HOST_CC="gcc -m32" CROSS=mipsel-linux-
+
+# MIPS64 big-endian
+make CROSS=mips-linux- TARGET_CFLAGS="-mips64r2 -mabi=64"
+# MIPS64 little-endian
+make CROSS=mipsel-linux- TARGET_CFLAGS="-mips64r2 -mabi=64"
 </pre>
 <p>
 You can cross-compile for <b id="android">Android</b> using the <a href="https://developer.android.com/ndk/index.html">Android NDK</a>.
--- a/doc/luajit.html
+++ b/doc/luajit.html
@ -169,10 +169,10 @@ LuaJIT is Copyright &copy; 2005-2016 Mike Pall, released under the
 <tr><td>PS3</td><td>PS4</td><td>PS Vita</td><td>Xbox 360</td><td>Xbox One</td></tr>
 </table>
 <table class="feature compiler">
-<tr><td>GCC</td><td>CLANG<br>LLVM</td><td>MSVC</td></tr>
+<tr><td>GCC</td><td>Clang<br>LLVM</td><td>MSVC</td></tr>
 </table>
 <table class="feature cpu">
-<tr><td>x86</td><td>x64</td><td>ARM</td><td>ARM64</td><td>PPC</td><td>MIPS</td></tr>
+<tr><td>x86<br>x64</td><td>ARM<br>ARM64</td><td>PPC</td><td>MIPS32<br>MIPS64</td></tr>
 </table>
 <table class="feature fcompat">
 <tr><td>Lua&nbsp;5.1<br>API+ABI</td><td>+&nbsp;JIT</td><td>+&nbsp;BitOp</td><td>+&nbsp;FFI</td><td>Drop-in<br>DLL/.so</td></tr>
--- a/src/Makefile
+++ b/src/Makefile
@ -257,7 +257,11 @@ ifneq (,$(findstring LJ_TARGET_MIPS ,$(TARGET_TESTARCH)))
  ifneq (,$(findstring MIPSEL ,$(TARGET_TESTARCH)))
    TARGET_ARCH= -D__MIPSEL__=1
  endif
+  ifneq (,$(findstring LJ_TARGET_MIPS64 ,$(TARGET_TESTARCH)))
+    TARGET_LJARCH= mips64
+  else
    TARGET_LJARCH= mips
+  endif
 else
  $(error Unsupported target architecture)
 endif
--- a/src/jit/dis_mips.lua
+++ b/src/jit/dis_mips.lua
@ -38,7 +38,7 @@ local map_special = {
  "multST",	"multuST",	"divST",	"divuST",
  false,	false,		false,		false,
  "addDST",	"addu|moveDST0", "subDST",	"subu|neguDS0T",
-  "andDST",	"orDST",	"xorDST",	"nor|notDST0",
+  "andDST",	"or|moveDST0",	"xorDST",	"nor|notDST0",
  false,	false,		"sltDST",	"sltuDST",
  false,	false,		false,		false,
  "tgeSTZ",	"tgeuSTZ",	"tltSTZ",	"tltuSTZ",
--- a/src/lib_jit.c
+++ b/src/lib_jit.c
@ -715,15 +715,15 @@ static uint32_t jit_cpudetect(lua_State *L)
 #if LJ_HASJIT
  /* Compile-time MIPS CPU detection. */
 #if LJ_ARCH_VERSION >= 20
-  flags |= JIT_F_MIPS32R2;
+  flags |= JIT_F_MIPSXXR2;
 #endif
  /* Runtime MIPS CPU detection. */
 #if defined(__GNUC__)
-  if (!(flags & JIT_F_MIPS32R2)) {
+  if (!(flags & JIT_F_MIPSXXR2)) {
    int x;
    /* On MIPS32R1 rotr is treated as srl. rotr r2,r2,1 -> srl r2,r2,1. */
    __asm__("li $2, 1\n\t.long 0x00221042\n\tmove %0, $2" : "=r"(x) : : "$2");
-    if (x) flags |= JIT_F_MIPS32R2;  /* Either 0x80000000 (R2) or 0 (R1). */
+    if (x) flags |= JIT_F_MIPSXXR2;  /* Either 0x80000000 (R2) or 0 (R1). */
  }
 #endif
 #endif
--- a/src/lj_arch.h
+++ b/src/lj_arch.h
@ -25,6 +25,10 @@
 #define LUAJIT_ARCH_ppc		5
 #define LUAJIT_ARCH_MIPS	6
 #define LUAJIT_ARCH_mips	6
+#define LUAJIT_ARCH_MIPS32	6
+#define LUAJIT_ARCH_mips32	6
+#define LUAJIT_ARCH_MIPS64	7
+#define LUAJIT_ARCH_mips64	7

 /* Target OS. */
 #define LUAJIT_OS_OTHER		0
@ -47,8 +51,10 @@
 #define LUAJIT_TARGET	LUAJIT_ARCH_ARM64
 #elif defined(__ppc__) || defined(__ppc) || defined(__PPC__) || defined(__PPC) || defined(__powerpc__) || defined(__powerpc) || defined(__POWERPC__) || defined(__POWERPC) || defined(_M_PPC)
 #define LUAJIT_TARGET	LUAJIT_ARCH_PPC
+#elif defined(__mips64__) || defined(__mips64) || defined(__MIPS64__) || defined(__MIPS64)
+#define LUAJIT_TARGET	LUAJIT_ARCH_MIPS64
 #elif defined(__mips__) || defined(__mips) || defined(__MIPS__) || defined(__MIPS)
-#define LUAJIT_TARGET	LUAJIT_ARCH_MIPS
+#define LUAJIT_TARGET	LUAJIT_ARCH_MIPS32
 #else
 #error "No support for this architecture (yet)"
 #endif
@ -289,13 +295,21 @@
 #define LJ_ARCH_XENON		1
 #endif

-#elif LUAJIT_TARGET == LUAJIT_ARCH_MIPS
+#elif LUAJIT_TARGET == LUAJIT_ARCH_MIPS32 || LUAJIT_TARGET == LUAJIT_ARCH_MIPS64

 #if defined(__MIPSEL__) || defined(__MIPSEL) || defined(_MIPSEL)
+#if LUAJIT_TARGET == LUAJIT_ARCH_MIPS32
 #define LJ_ARCH_NAME		"mipsel"
+#else
+#define LJ_ARCH_NAME		"mips64el"
+#endif
 #define LJ_ARCH_ENDIAN		LUAJIT_LE
 #else
+#if LUAJIT_TARGET == LUAJIT_ARCH_MIPS32
 #define LJ_ARCH_NAME		"mips"
+#else
+#define LJ_ARCH_NAME		"mips64"
+#endif
 #define LJ_ARCH_ENDIAN		LUAJIT_BE
 #endif

@ -307,11 +321,6 @@
 #endif
 #endif

-/* Temporarily disable features until the code has been merged. */
-#if !defined(LUAJIT_NO_UNWIND) && __GNU_COMPACT_EH__
-#define LUAJIT_NO_UNWIND	1
-#endif
-
 #if !defined(LJ_ABI_SOFTFP)
 #ifdef __mips_soft_float
 #define LJ_ABI_SOFTFP		1
@ -320,7 +329,15 @@
 #endif
 #endif

+#if LUAJIT_TARGET == LUAJIT_ARCH_MIPS32
 #define LJ_ARCH_BITS		32
+#define LJ_TARGET_MIPS32	1
+#else
+#define LJ_ARCH_BITS		64
+#define LJ_TARGET_MIPS64	1
+#define LJ_TARGET_GC64		1
+#define LJ_ARCH_NOJIT		1	/* NYI */
+#endif
 #define LJ_TARGET_MIPS		1
 #define LJ_TARGET_EHRETREG	4
 #define LJ_TARGET_JUMPRANGE	27	/* 2*2^27 = 256MB-aligned region */
@ -329,7 +346,7 @@
 #define LJ_TARGET_UNIFYROT	2	/* Want only IR_BROR. */
 #define LJ_ARCH_NUMMODE		LJ_NUMMODE_DUAL

-#if _MIPS_ARCH_MIPS32R2
+#if _MIPS_ARCH_MIPS32R2 || _MIPS_ARCH_MIPS64R2
 #define LJ_ARCH_VERSION		20
 #else
 #define LJ_ARCH_VERSION		10
@ -410,9 +427,13 @@
 #ifdef __NO_FPRS__
 #error "No support for PPC/e500 anymore (use LuaJIT 2.0)"
 #endif
-#elif LJ_TARGET_MIPS
-#if defined(_LP64)
-#error "No support for MIPS64"
+#elif LJ_TARGET_MIPS32
+#if _MIPS_SIM != _MIPS_SIM_ABI32
+#error "Only o32 ABI supported for MIPS32"
+#endif
+#elif LJ_TARGET_MIPS64
+#if _MIPS_SIM != _MIPS_SIM_ABI64
+#error "Only n64 ABI supported for MIPS64"
 #endif
 #endif
 #endif
@ -524,6 +545,11 @@
 #define LJ_NO_SYSTEM		1
 #endif

+#if !defined(LUAJIT_NO_UNWIND) && __GNU_COMPACT_EH__
+/* NYI: no support for compact unwind specification, yet. */
+#define LUAJIT_NO_UNWIND	1
+#endif
+
 #if defined(LUAJIT_NO_UNWIND) || defined(__symbian__) || LJ_TARGET_IOS || LJ_TARGET_PS3 || LJ_TARGET_PS4
 #define LJ_NO_UNWIND		1
 #endif
--- a/src/lj_asm_mips.h
+++ b/src/lj_asm_mips.h
@ -510,7 +510,7 @@ static void asm_conv(ASMState *as, IRIns *ir)
      Reg left = ra_alloc1(as, ir->op1, RSET_GPR);
      lua_assert(irt_isint(ir->t) || irt_isu32(ir->t));
      if ((ir->op2 & IRCONV_SEXT)) {
-	if ((as->flags & JIT_F_MIPS32R2)) {
+	if ((as->flags & JIT_F_MIPSXXR2)) {
 	  emit_dst(as, st == IRT_I8 ? MIPSI_SEB : MIPSI_SEH, dest, 0, left);
 	} else {
 	  uint32_t shift = st == IRT_I8 ? 24 : 16;
@ -739,7 +739,7 @@ static void asm_href(ASMState *as, IRIns *ir, IROp merge)
      emit_dst(as, MIPSI_SUBU, tmp2, tmp2, dest);
      if (LJ_SOFTFP ? (irkey[1].o == IR_HIOP) : irt_isnum(kt)) {
 	emit_dst(as, MIPSI_XOR, tmp2, tmp2, tmp1);
-	if ((as->flags & JIT_F_MIPS32R2)) {
+	if ((as->flags & JIT_F_MIPSXXR2)) {
 	  emit_dta(as, MIPSI_ROTR, dest, tmp1, (-HASH_ROT1)&31);
 	} else {
 	  emit_dst(as, MIPSI_OR, dest, dest, tmp1);
@ -1457,7 +1457,7 @@ static void asm_bswap(ASMState *as, IRIns *ir)
 {
  Reg dest = ra_dest(as, ir, RSET_GPR);
  Reg left = ra_alloc1(as, ir->op1, RSET_GPR);
-  if ((as->flags & JIT_F_MIPS32R2)) {
+  if ((as->flags & JIT_F_MIPSXXR2)) {
    emit_dta(as, MIPSI_ROTR, dest, RID_TMP, 16);
    emit_dst(as, MIPSI_WSBH, RID_TMP, 0, left);
  } else {
@ -1513,7 +1513,7 @@ static void asm_bitshift(ASMState *as, IRIns *ir, MIPSIns mi, MIPSIns mik)

 static void asm_bror(ASMState *as, IRIns *ir)
 {
-  if ((as->flags & JIT_F_MIPS32R2)) {
+  if ((as->flags & JIT_F_MIPSXXR2)) {
    asm_bitshift(as, ir, MIPSI_ROTRV, MIPSI_ROTR);
  } else {
    Reg dest = ra_dest(as, ir, RSET_GPR);
--- a/src/lj_ccall.c
+++ b/src/lj_ccall.c
@ -407,8 +407,8 @@
  if (ctype_isfp(ctr->info) && ctr->size == sizeof(float)) \
    ctr = ctype_get(cts, CTID_DOUBLE);  /* FPRs always hold doubles. */

-#elif LJ_TARGET_MIPS
-/* -- MIPS calling conventions -------------------------------------------- */
+#elif LJ_TARGET_MIPS32
+/* -- MIPS o32 calling conventions ---------------------------------------- */

 #define CCALL_HANDLE_STRUCTRET \
  cc->retref = 1;  /* Return all structs by reference. */ \
@ -483,6 +483,78 @@
    sp = (uint8_t *)&cc->fpr[0].f;
 #endif

+#elif LJ_TARGET_MIPS64
+/* -- MIPS n64 calling conventions ---------------------------------------- */
+
+#define CCALL_HANDLE_STRUCTRET \
+  cc->retref = !(sz <= 16); \
+  if (cc->retref) cc->gpr[ngpr++] = (GPRArg)dp;
+
+#define CCALL_HANDLE_STRUCTRET2 \
+  ccall_copy_struct(cc, ctr, dp, sp, ccall_classify_struct(cts, ctr, ct));
+
+#define CCALL_HANDLE_COMPLEXRET \
+  /* Complex values are returned in 1 or 2 FPRs. */ \
+  cc->retref = 0;
+
+#if LJ_ABI_SOFTFP	/* MIPS64 soft-float */
+
+#define CCALL_HANDLE_COMPLEXRET2 \
+  if (ctr->size == 2*sizeof(float)) {  /* Copy complex float from GPRs. */ \
+    ((intptr_t *)dp)[0] = cc->gpr[0]; \
+  } else {  /* Copy complex double from GPRs. */ \
+    ((intptr_t *)dp)[0] = cc->gpr[0]; \
+    ((intptr_t *)dp)[1] = cc->gpr[1]; \
+  }
+
+#define CCALL_HANDLE_COMPLEXARG \
+  /* Pass complex by value in 2 or 4 GPRs. */
+
+/* Position of soft-float 'float' return value depends on endianess.  */
+#define CCALL_HANDLE_RET \
+  if (ctype_isfp(ctr->info) && ctr->size == sizeof(float)) \
+    sp = (uint8_t *)cc->gpr + LJ_ENDIAN_SELECT(0, 4);
+
+#else			/* MIPS64 hard-float */
+
+#define CCALL_HANDLE_COMPLEXRET2 \
+  if (ctr->size == 2*sizeof(float)) {  /* Copy complex float from FPRs. */ \
+    ((float *)dp)[0] = cc->fpr[0].f; \
+    ((float *)dp)[1] = cc->fpr[1].f; \
+  } else {  /* Copy complex double from FPRs. */ \
+    ((double *)dp)[0] = cc->fpr[0].d; \
+    ((double *)dp)[1] = cc->fpr[1].d; \
+  }
+
+#define CCALL_HANDLE_COMPLEXARG \
+  if (sz == 2*sizeof(float)) { \
+    isfp = 2; \
+    if (ngpr < maxgpr) \
+      sz *= 2; \
+  }
+
+#define CCALL_HANDLE_RET \
+  if (ctype_isfp(ctr->info) && ctr->size == sizeof(float)) \
+    sp = (uint8_t *)&cc->fpr[0].f;
+
+#endif
+
+#define CCALL_HANDLE_STRUCTARG \
+  /* Pass all structs by value in registers and/or on the stack. */
+
+#define CCALL_HANDLE_REGARG \
+  if (ngpr < maxgpr) { \
+    dp = &cc->gpr[ngpr]; \
+    if (ngpr + n > maxgpr) { \
+      nsp += ngpr + n - maxgpr;  /* Assumes contiguous gpr/stack fields. */ \
+      if (nsp > CCALL_MAXSTACK) goto err_nyi;  /* Too many arguments. */ \
+      ngpr = maxgpr; \
+    } else { \
+      ngpr += n; \
+    } \
+    goto done; \
+  }
+
 #else
 #error "Missing calling convention definitions for this architecture"
 #endif
@ -722,6 +794,78 @@ noth:  /* Not a homogeneous float/double aggregate. */

 #endif

+/* -- MIPS64 ABI struct classification ---------------------------- */
+
+#if LJ_TARGET_MIPS64
+
+#define FTYPE_FLOAT	1
+#define FTYPE_DOUBLE	2
+
+/* Classify FP fields (max. 2) and their types. */
+static unsigned int ccall_classify_struct(CTState *cts, CType *ct, CType *ctf)
+{
+  int n = 0, ft = 0;
+  if ((ctf->info & CTF_VARARG) || (ct->info & CTF_UNION))
+    goto noth;
+  while (ct->sib) {
+    CType *sct;
+    ct = ctype_get(cts, ct->sib);
+    if (n == 2) {
+      goto noth;
+    } else if (ctype_isfield(ct->info)) {
+      sct = ctype_rawchild(cts, ct);
+      if (ctype_isfp(sct->info)) {
+	ft |= (sct->size == 4 ? FTYPE_FLOAT : FTYPE_DOUBLE) << 2*n;
+	n++;
+      } else {
+	goto noth;
+      }
+    } else if (ctype_isbitfield(ct->info) ||
+	       ctype_isxattrib(ct->info, CTA_SUBTYPE)) {
+      goto noth;
+    }
+  }
+  if (n <= 2)
+    return ft;
+noth:  /* Not a homogeneous float/double aggregate. */
+  return 0;  /* Struct is in GPRs. */
+}
+
+void ccall_copy_struct(CCallState *cc, CType *ctr, void *dp, void *sp, int ft)
+{
+  if (LJ_ABI_SOFTFP ? ft :
+      ((ft & 3) == FTYPE_FLOAT || (ft >> 2) == FTYPE_FLOAT)) {
+    int i, ofs = 0;
+    for (i = 0; ft != 0; i++, ft >>= 2) {
+      if ((ft & 3) == FTYPE_FLOAT) {
+#if LJ_ABI_SOFTFP
+	/* The 2nd FP struct result is in CARG1 (gpr[2]) and not CRET2. */
+	memcpy((uint8_t *)dp + ofs,
+	       (uint8_t *)&cc->gpr[2*i] + LJ_ENDIAN_SELECT(0, 4), 4);
+#else
+	*(float *)((uint8_t *)dp + ofs) = cc->fpr[i].f;
+#endif
+	ofs += 4;
+      } else {
+	ofs = (ofs + 7) & ~7;  /* 64 bit alignment. */
+#if LJ_ABI_SOFTFP
+	*(intptr_t *)((uint8_t *)dp + ofs) = cc->gpr[2*i];
+#else
+	*(double *)((uint8_t *)dp + ofs) = cc->fpr[i].d;
+#endif
+	ofs += 8;
+      }
+    }
+  } else {
+#if !LJ_ABI_SOFTFP
+    if (ft) sp = (uint8_t *)&cc->fpr[0];
+#endif
+    memcpy(dp, sp, ctr->size);
+  }
+}
+
+#endif
+
 /* -- Common C call handling ---------------------------------------------- */

 /* Infer the destination CTypeID for a vararg argument. */
@ -889,6 +1033,12 @@ static int ccall_set_args(lua_State *L, CTState *cts, CType *ct,
 	*(int32_t *)dp = d->size == 1 ? (int32_t)*(int8_t *)dp :
 					(int32_t)*(int16_t *)dp;
    }
+#if LJ_TARGET_MIPS64
+    if ((ctype_isinteger_or_bool(d->info) || ctype_isenum(d->info) ||
+	 (isfp && nsp == 0)) && d->size <= 4) {
+      *(int64_t *)dp = (int64_t)*(int32_t *)dp;  /* Sign-extend to 64 bit. */
+    }
+#endif
 #if LJ_TARGET_X64 && LJ_ABI_WIN
    if (isva) {  /* Windows/x64 mirrors varargs in both register sets. */
      if (nfpr == ngpr)
@ -904,7 +1054,7 @@ static int ccall_set_args(lua_State *L, CTState *cts, CType *ct,
      cc->fpr[nfpr-1].d[0] = cc->fpr[nfpr-2].d[1];  /* Split complex double. */
      cc->fpr[nfpr-2].d[1] = 0;
    }
-#elif LJ_TARGET_ARM64
+#elif LJ_TARGET_ARM64 || (LJ_TARGET_MIPS64 && !LJ_ABI_SOFTFP)
    if (isfp == 2 && (uint8_t *)dp < (uint8_t *)cc->stack) {
      /* Split float HFA or complex float into separate registers. */
      CTSize i = (sz >> 2) - 1;
@ -951,7 +1101,8 @@ static int ccall_get_results(lua_State *L, CTState *cts, CType *ct,
    CCALL_HANDLE_COMPLEXRET2
    return 1;  /* One GC step. */
  }
-  if (LJ_BE && ctype_isinteger_or_bool(ctr->info) && ctr->size < CTSIZE_PTR)
+  if (LJ_BE && ctr->size < CTSIZE_PTR &&
+      (ctype_isinteger_or_bool(ctr->info) || ctype_isenum(ctr->info)))
    sp += (CTSIZE_PTR - ctr->size);
 #if CCALL_NUM_FPR
  if (ctype_isfp(ctr->info) || ctype_isvector(ctr->info))
--- a/src/lj_ccall.h
+++ b/src/lj_ccall.h
@ -95,11 +95,11 @@ typedef union FPRArg {
 typedef intptr_t GPRArg;
 typedef double FPRArg;

-#elif LJ_TARGET_MIPS
+#elif LJ_TARGET_MIPS32

 #define CCALL_NARG_GPR		4
 #define CCALL_NARG_FPR		(LJ_ABI_SOFTFP ? 0 : 2)
-#define CCALL_NRET_GPR		2
+#define CCALL_NRET_GPR		(LJ_ABI_SOFTFP ? 4 : 2)
 #define CCALL_NRET_FPR		(LJ_ABI_SOFTFP ? 0 : 2)
 #define CCALL_SPS_EXTRA		7
 #define CCALL_SPS_FREE		1
@ -110,6 +110,22 @@ typedef union FPRArg {
  struct { LJ_ENDIAN_LOHI(float f; , float g;) };
 } FPRArg;

+#elif LJ_TARGET_MIPS64
+
+/* FP args are positional and overlay the GPR array. */
+#define CCALL_NARG_GPR		8
+#define CCALL_NARG_FPR		0
+#define CCALL_NRET_GPR		2
+#define CCALL_NRET_FPR		(LJ_ABI_SOFTFP ? 0 : 2)
+#define CCALL_SPS_EXTRA		3
+#define CCALL_SPS_FREE		1
+
+typedef intptr_t GPRArg;
+typedef union FPRArg {
+  double d;
+  struct { LJ_ENDIAN_LOHI(float f; , float g;) };
+} FPRArg;
+
 #else
 #error "Missing calling convention definitions for this architecture"
 #endif
--- a/src/lj_ccallback.c
+++ b/src/lj_ccallback.c
@ -63,9 +63,13 @@ static MSize CALLBACK_OFS2SLOT(MSize ofs)

 #define CALLBACK_MCODE_HEAD		24

-#elif LJ_TARGET_MIPS
+#elif LJ_TARGET_MIPS32

-#define CALLBACK_MCODE_HEAD		24
+#define CALLBACK_MCODE_HEAD		20
+
+#elif LJ_TARGET_MIPS64
+
+#define CALLBACK_MCODE_HEAD		52

 #else

@ -206,14 +210,27 @@ static void callback_mcode_init(global_State *g, uint32_t *page)
 static void callback_mcode_init(global_State *g, uint32_t *page)
 {
  uint32_t *p = page;
-  void *target = (void *)lj_vm_ffi_callback;
+  uintptr_t target = (uintptr_t)(void *)lj_vm_ffi_callback;
+  uintptr_t ug = (uintptr_t)(void *)g;
  MSize slot;
-  *p++ = MIPSI_SW | MIPSF_T(RID_R1)|MIPSF_S(RID_SP) | 0;
-  *p++ = MIPSI_LUI | MIPSF_T(RID_R3) | (u32ptr(target) >> 16);
-  *p++ = MIPSI_LUI | MIPSF_T(RID_R2) | (u32ptr(g) >> 16);
-  *p++ = MIPSI_ORI | MIPSF_T(RID_R3)|MIPSF_S(RID_R3) |(u32ptr(target)&0xffff);
+#if LJ_TARGET_MIPS32
+  *p++ = MIPSI_LUI | MIPSF_T(RID_R3) | (target >> 16);
+  *p++ = MIPSI_LUI | MIPSF_T(RID_R2) | (ug >> 16);
+#else
+  *p++ = MIPSI_LUI  | MIPSF_T(RID_R3) | (target >> 48);
+  *p++ = MIPSI_LUI  | MIPSF_T(RID_R2) | (ug >> 48);
+  *p++ = MIPSI_ORI  | MIPSF_T(RID_R3)|MIPSF_S(RID_R3) | ((target >> 32) & 0xffff);
+  *p++ = MIPSI_ORI  | MIPSF_T(RID_R2)|MIPSF_S(RID_R2) | ((ug >> 32) & 0xffff);
+  *p++ = MIPSI_DSLL | MIPSF_D(RID_R3)|MIPSF_T(RID_R3) | MIPSF_A(16);
+  *p++ = MIPSI_DSLL | MIPSF_D(RID_R2)|MIPSF_T(RID_R2) | MIPSF_A(16);
+  *p++ = MIPSI_ORI  | MIPSF_T(RID_R3)|MIPSF_S(RID_R3) | ((target >> 16) & 0xffff);
+  *p++ = MIPSI_ORI  | MIPSF_T(RID_R2)|MIPSF_S(RID_R2) | ((ug >> 16) & 0xffff);
+  *p++ = MIPSI_DSLL | MIPSF_D(RID_R3)|MIPSF_T(RID_R3) | MIPSF_A(16);
+  *p++ = MIPSI_DSLL | MIPSF_D(RID_R2)|MIPSF_T(RID_R2) | MIPSF_A(16);
+#endif
+  *p++ = MIPSI_ORI  | MIPSF_T(RID_R3)|MIPSF_S(RID_R3) | (target & 0xffff);
  *p++ = MIPSI_JR | MIPSF_S(RID_R3);
-  *p++ = MIPSI_ORI | MIPSF_T(RID_R2)|MIPSF_S(RID_R2) | (u32ptr(g)&0xffff);
+  *p++ = MIPSI_ORI | MIPSF_T(RID_R2)|MIPSF_S(RID_R2) | (ug & 0xffff);
  for (slot = 0; slot < CALLBACK_MAX_SLOT; slot++) {
    *p = MIPSI_B | ((page-p-1) & 0x0000ffffu);
    p++;
@ -425,7 +442,7 @@ void lj_ccallback_mcode_free(CTState *cts)
  if (ctype_isfp(ctr->info) && ctr->size == sizeof(float)) \
    *(double *)dp = *(float *)dp;  /* FPRs always hold doubles. */

-#elif LJ_TARGET_MIPS
+#elif LJ_TARGET_MIPS32

 #define CALLBACK_HANDLE_GPR \
  if (n > 1) ngpr = (ngpr + 1u) & ~1u;  /* Align to regpair. */ \
@ -451,6 +468,29 @@ void lj_ccallback_mcode_free(CTState *cts)
  UNUSED(isfp);
 #endif

+#define CALLBACK_HANDLE_RET \
+  if (ctype_isfp(ctr->info) && ctr->size == sizeof(float)) \
+    ((float *)dp)[1] = *(float *)dp;
+
+#elif LJ_TARGET_MIPS64
+
+#if !LJ_ABI_SOFTFP	/* MIPS64 hard-float */
+#define CALLBACK_HANDLE_REGARG \
+  if (ngpr + n <= maxgpr) { \
+    sp = isfp ? (void*) &cts->cb.fpr[ngpr] : (void*) &cts->cb.gpr[ngpr]; \
+    ngpr += n; \
+    goto done; \
+  }
+#else			/* MIPS64 soft-float */
+#define CALLBACK_HANDLE_REGARG \
+  if (ngpr + n <= maxgpr) { \
+    UNUSED(isfp); \
+    sp = (void*) &cts->cb.gpr[ngpr]; \
+    ngpr += n; \
+    goto done; \
+  }
+#endif
+
 #define CALLBACK_HANDLE_RET \
  if (ctype_isfp(ctr->info) && ctr->size == sizeof(float)) \
    ((float *)dp)[1] = *(float *)dp;
@ -542,7 +582,11 @@ static void callback_conv_args(CTState *cts, lua_State *L)
      nsp += n;

    done:
-      if (LJ_BE && cta->size < CTSIZE_PTR)
+      if (LJ_BE && cta->size < CTSIZE_PTR
+#if LJ_TARGET_MIPS64
+	  && !(isfp && nsp)
+#endif
+	 )
 	sp = (void *)((uint8_t *)sp + CTSIZE_PTR-cta->size);
      gcsteps += lj_cconv_tv_ct(cts, cta, 0, o++, sp);
    }
@ -593,6 +637,12 @@ static void callback_conv_result(CTState *cts, lua_State *L, TValue *o)
 	*(int32_t *)dp = ctr->size == 1 ? (int32_t)*(int8_t *)dp :
 					  (int32_t)*(int16_t *)dp;
    }
+#if LJ_TARGET_MIPS64
+    /* Always sign-extend results to 64 bits. Even a soft-fp 'float'. */
+    if (ctr->size <= 4 &&
+	(LJ_ABI_SOFTFP || ctype_isinteger_or_bool(ctr->info)))
+      *(int64_t *)dp = (int64_t)*(int32_t *)dp;
+#endif
 #if LJ_TARGET_X86
    if (ctype_isfp(ctr->info))
      cts->cb.gpr[2] = ctr->size == sizeof(float) ? 1 : 2;
--- a/src/lj_dispatch.c
+++ b/src/lj_dispatch.c
@ -75,7 +75,7 @@ void lj_dispatch_init(GG_State *GG)
  for (i = 0; i < GG_NUM_ASMFF; i++)
    GG->bcff[i] = BCINS_AD(BC__MAX+i, 0, 0);
 #if LJ_TARGET_MIPS
-  memcpy(GG->got, dispatch_got, LJ_GOT__MAX*4);
+  memcpy(GG->got, dispatch_got, LJ_GOT__MAX*sizeof(ASMFunction *));
 #endif
 }

--- a/src/lj_emit_mips.h
+++ b/src/lj_emit_mips.h
@ -35,7 +35,7 @@ static void emit_fgh(ASMState *as, MIPSIns mi, Reg rf, Reg rg, Reg rh)

 static void emit_rotr(ASMState *as, Reg dest, Reg src, Reg tmp, uint32_t shift)
 {
-  if ((as->flags & JIT_F_MIPS32R2)) {
+  if ((as->flags & JIT_F_MIPSXXR2)) {
    emit_dta(as, MIPSI_ROTR, dest, src, shift);
  } else {
    emit_dst(as, MIPSI_OR, dest, dest, tmp);
--- a/src/lj_frame.h
+++ b/src/lj_frame.h
@ -229,26 +229,41 @@ enum { LJ_CONT_TAILCALL, LJ_CONT_FFI_CALLBACK };  /* Special continuations. */
 #define CFRAME_SIZE		272
 #define CFRAME_SHIFT_MULTRES	3
 #endif
-#elif LJ_TARGET_MIPS
+#elif LJ_TARGET_MIPS32
 #if LJ_ARCH_HASFPU
 #define CFRAME_OFS_ERRF		124
 #define CFRAME_OFS_NRES		120
 #define CFRAME_OFS_PREV		116
 #define CFRAME_OFS_L		112
-#define CFRAME_OFS_PC		20
-#define CFRAME_OFS_MULTRES	16
 #define CFRAME_SIZE		112
-#define CFRAME_SHIFT_MULTRES	3
 #else
 #define CFRAME_OFS_ERRF		76
 #define CFRAME_OFS_NRES		72
 #define CFRAME_OFS_PREV		68
 #define CFRAME_OFS_L		64
+#define CFRAME_SIZE		64
+#endif
 #define CFRAME_OFS_PC		20
 #define CFRAME_OFS_MULTRES	16
-#define CFRAME_SIZE		64
 #define CFRAME_SHIFT_MULTRES	3
+#elif LJ_TARGET_MIPS64
+#if LJ_ARCH_HASFPU
+#define CFRAME_OFS_ERRF		188
+#define CFRAME_OFS_NRES		184
+#define CFRAME_OFS_PREV		176
+#define CFRAME_OFS_L		168
+#define CFRAME_OFS_PC		160
+#define CFRAME_SIZE		192
+#else
+#define CFRAME_OFS_ERRF		124
+#define CFRAME_OFS_NRES		120
+#define CFRAME_OFS_PREV		112
+#define CFRAME_OFS_L		104
+#define CFRAME_OFS_PC		96
+#define CFRAME_SIZE		128
 #endif
+#define CFRAME_OFS_MULTRES	0
+#define CFRAME_SHIFT_MULTRES	3
 #else
 #error "Missing CFRAME_* definitions for this architecture"
 #endif
--- a/src/lj_ircall.h
+++ b/src/lj_ircall.h
@ -78,13 +78,13 @@ typedef struct CCallInfo {
 #define IRCALLCOND_SOFTFP_FFI(x)	NULL
 #endif

-#if LJ_SOFTFP && LJ_TARGET_MIPS
+#if LJ_SOFTFP && LJ_TARGET_MIPS32
 #define IRCALLCOND_SOFTFP_MIPS(x)	x
 #else
 #define IRCALLCOND_SOFTFP_MIPS(x)	NULL
 #endif

-#define LJ_NEED_FP64	(LJ_TARGET_ARM || LJ_TARGET_PPC || LJ_TARGET_MIPS)
+#define LJ_NEED_FP64	(LJ_TARGET_ARM || LJ_TARGET_PPC || LJ_TARGET_MIPS32)

 #if LJ_HASFFI && (LJ_SOFTFP || LJ_NEED_FP64)
 #define IRCALLCOND_FP64_FFI(x)		x
--- a/src/lj_jit.h
+++ b/src/lj_jit.h
@ -46,12 +46,16 @@
 #define JIT_F_CPU_FIRST		JIT_F_SQRT
 #define JIT_F_CPUSTRING		"\4SQRT\5ROUND"
 #elif LJ_TARGET_MIPS
-#define JIT_F_MIPS32R2		0x00000010
+#define JIT_F_MIPSXXR2		0x00000010

 /* Names for the CPU-specific flags. Must match the order above. */
-#define JIT_F_CPU_FIRST		JIT_F_MIPS32R2
+#define JIT_F_CPU_FIRST		JIT_F_MIPSXXR2
+#if LJ_TARGET_MIPS32
 #define JIT_F_CPUSTRING		"\010MIPS32R2"
 #else
+#define JIT_F_CPUSTRING		"\010MIPS64R2"
+#endif
+#else
 #define JIT_F_CPU_FIRST		0
 #define JIT_F_CPUSTRING		""
 #endif
--- a/src/lj_target_mips.h
+++ b/src/lj_target_mips.h
@ -82,11 +82,15 @@ enum {
 #if LJ_SOFTFP
 #define RSET_FPR	0
 #else
+#if LJ_32
 #define RSET_FPR \
  (RID2RSET(RID_F0)|RID2RSET(RID_F2)|RID2RSET(RID_F4)|RID2RSET(RID_F6)|\
   RID2RSET(RID_F8)|RID2RSET(RID_F10)|RID2RSET(RID_F12)|RID2RSET(RID_F14)|\
   RID2RSET(RID_F16)|RID2RSET(RID_F18)|RID2RSET(RID_F20)|RID2RSET(RID_F22)|\
   RID2RSET(RID_F24)|RID2RSET(RID_F26)|RID2RSET(RID_F28)|RID2RSET(RID_F30))
+#else
+#define RSET_FPR	RSET_RANGE(RID_MIN_FPR, RID_MAX_FPR)
+#endif
 #endif
 #define RSET_ALL	(RSET_GPR|RSET_FPR)
 #define RSET_INIT	RSET_ALL
@ -97,23 +101,37 @@ enum {
 #if LJ_SOFTFP
 #define RSET_SCRATCH_FPR	0
 #else
+#if LJ_32
 #define RSET_SCRATCH_FPR \
  (RID2RSET(RID_F0)|RID2RSET(RID_F2)|RID2RSET(RID_F4)|RID2RSET(RID_F6)|\
   RID2RSET(RID_F8)|RID2RSET(RID_F10)|RID2RSET(RID_F12)|RID2RSET(RID_F14)|\
   RID2RSET(RID_F16)|RID2RSET(RID_F18))
+#else
+#define RSET_SCRATCH_FPR	RSET_RANGE(RID_F0, RID_F24)
+#endif
 #endif
 #define RSET_SCRATCH		(RSET_SCRATCH_GPR|RSET_SCRATCH_FPR)
 #define REGARG_FIRSTGPR		RID_R4
+#if LJ_32
 #define REGARG_LASTGPR		RID_R7
 #define REGARG_NUMGPR		4
+#else
+#define REGARG_LASTGPR		RID_R11
+#define REGARG_NUMGPR		8
+#endif
 #if LJ_ABI_SOFTFP
 #define REGARG_FIRSTFPR		0
 #define REGARG_LASTFPR		0
 #define REGARG_NUMFPR		0
 #else
 #define REGARG_FIRSTFPR		RID_F12
+#if LJ_32
 #define REGARG_LASTFPR		RID_F14
 #define REGARG_NUMFPR		2
+#else
+#define REGARG_LASTFPR		RID_F19
+#define REGARG_NUMFPR		8
+#endif
 #endif

 /* -- Spill slots --------------------------------------------------------- */
@ -125,7 +143,11 @@ enum {
 **
 ** SPS_FIRST: First spill slot for general use.
 */
+#if LJ_32
 #define SPS_FIXED	5
+#else
+#define SPS_FIXED	4
+#endif
 #define SPS_FIRST	4

 #define SPOFS_TMP	0
@ -140,7 +162,7 @@ typedef struct {
 #if !LJ_SOFTFP
  lua_Number fpr[RID_NUM_FPR];	/* Floating-point registers. */
 #endif
-  int32_t gpr[RID_NUM_GPR];	/* General-purpose registers. */
+  intptr_t gpr[RID_NUM_GPR];	/* General-purpose registers. */
  int32_t spill[256];		/* Spill slots. */
 } ExitState;

@ -172,7 +194,7 @@ static LJ_AINLINE uint32_t *exitstub_trace_addr_(uint32_t *p)

 typedef enum MIPSIns {
  /* Integer instructions. */
-  MIPSI_MOVE = 0x00000021,
+  MIPSI_MOVE = 0x00000025,
  MIPSI_NOP = 0x00000000,

  MIPSI_LI = 0x24000000,
@ -204,15 +226,15 @@ typedef enum MIPSIns {
  MIPSI_SLL = 0x00000000,
  MIPSI_SRL = 0x00000002,
  MIPSI_SRA = 0x00000003,
-  MIPSI_ROTR = 0x00200002,	/* MIPS32R2 */
+  MIPSI_ROTR = 0x00200002,	/* MIPSXXR2 */
  MIPSI_SLLV = 0x00000004,
  MIPSI_SRLV = 0x00000006,
  MIPSI_SRAV = 0x00000007,
-  MIPSI_ROTRV = 0x00000046,	/* MIPS32R2 */
+  MIPSI_ROTRV = 0x00000046,	/* MIPSXXR2 */

-  MIPSI_SEB = 0x7c000420,	/* MIPS32R2 */
-  MIPSI_SEH = 0x7c000620,	/* MIPS32R2 */
-  MIPSI_WSBH = 0x7c0000a0,	/* MIPS32R2 */
+  MIPSI_SEB = 0x7c000420,	/* MIPSXXR2 */
+  MIPSI_SEH = 0x7c000620,	/* MIPSXXR2 */
+  MIPSI_WSBH = 0x7c0000a0,	/* MIPSXXR2 */

  MIPSI_B = 0x10000000,
  MIPSI_J = 0x08000000,
@ -241,6 +263,15 @@ typedef enum MIPSIns {
  MIPSI_LDC1 = 0xd4000000,
  MIPSI_SDC1 = 0xf4000000,

+  /* MIPS64 instructions. */
+  MIPSI_DSLL = 0x00000038,
+  MIPSI_LD = 0xdc000000,
+  MIPSI_DADDIU = 0x64000000,
+  MIPSI_SD = 0xfc000000,
+  MIPSI_DMFC1 = 0x44200000,
+  MIPSI_DSRA32 = 0x0000003f,
+  MIPSI_MFHC1 = 0x44600000,
+
  /* FP instructions. */
  MIPSI_MOV_S = 0x46000006,
  MIPSI_MOV_D = 0x46200006,
--- a/src/vm_mips.dasc
+++ b/src/vm_mips.dasc
@ -57,7 +57,7 @@
 |.define TMP2,		r14
 |.define TMP3,		r15
 |
-|// Calling conventions.
+|// MIPS o32 calling convention.
 |.define CFUNCADDR,	r25
 |.define CARG1,		r4
 |.define CARG2,		r5
--- a/src/vm_mips64.dasc
+++ b/src/vm_mips64.dasc