Template intrinsics user machine code

src/lib_ffi.c

@@ -492,6 +492,16 @@ LJLIB_CF(ffi_cdef)
   return 0;
 }
 
+LJLIB_CF(ffi_intrinsic)
+{
+#if LJ_HASINTRINSICS
+  lj_intrinsic_create(L);
+  return 1;
+#else
+  lj_err_callermsg(L, "Intrinsics disabled");
+#endif
+}
+
 LJLIB_CF(ffi_new)	LJLIB_REC(.)
 {
   CTState *cts = ctype_cts(L);

src/lj_asm.c

@@ -2569,6 +2569,7 @@ static void wrap_intrins(jit_State *J, CIntrinsic *intrins, IntrinWrapState *sta
   IntrinBuildState info;
   AsmHeader *hdr;
   MCode *asmofs = NULL, *origtop;
+  void* target = state->target;
   int spadj = 0;
 
   lj_asm_setup_intrins(J, as);
@@ -2587,6 +2588,12 @@ static void wrap_intrins(jit_State *J, CIntrinsic *intrins, IntrinWrapState *sta
     ra_modified(as, info.outcontext);
   }
 
+  /* Embed the mcode after the wrapper */
+  if ((intrins->flags & INTRINSFLAG_CALLED) && state->targetsz) {
+    *--as->mcp = XI_RET;
+    target = asmofs = asm_mcode(as, target, state->targetsz);
+  }
+
 restart:
   if (info.contexspill || rset_test(info.outset, info.outcontext)) {
     /* add some extra space for context spill and temp spill */
@@ -2630,8 +2637,20 @@ restart:
     emit_storeofsirt(as, IRT_INTP, info.outcontext, RID_SP, TEMPSPILL);
   }
 
-  /* Append the user supplied machine code */
+  if (intrins->flags & INTRINSFLAG_CALLED) {
-  asmofs = asm_mcode(as, state->target, state->targetsz);
+    Reg rin = 0;
+#if LJ_64
+    /* Pick a scratch register in case the relative distance for the call is
+    ** larger than a signed 32bit value
+    */
+    rin = intrinsic_scratch(as, RSET_GPR);
+#endif
+    /* emit a call to the target which may be collocated after us */
+    emit_intrins(as, intrins, rin, (uintptr_t)target);
+  } else {
+    /* Append the user supplied machine code */
+    asmofs = asm_mcode(as, state->target, state->targetsz);
+  }
   
   /* Move values out the context into there respective input registers */
   intrins_loadregs(as, intrins, &info);
@@ -2657,9 +2676,14 @@ restart:
   memset(hdr, 0, sizeof(AsmHeader));
   hdr->totalzs = (uint32_t)(origtop-as->mcp);
 
-  lua_assert((asmofs-as->mcp) < 0xffff);
+  /* Embed info before the code to support multiple versions of a user intrinsic intrinsic */
-  hdr->asmofs = (uint16_t)(asmofs-as->mcp);
+  if ((intrins->flags & INTRINSFLAG_CALLEDIND) == INTRINSFLAG_CALLEDIND) {
-  hdr->asmsz = state->targetsz;
+    hdr->target = (uintptr_t)target;
+  } else if (asmofs){
+    lua_assert((asmofs-as->mcp) < 0xffff);
+    hdr->asmofs = (uint16_t)(asmofs-as->mcp);
+    hdr->asmsz = state->targetsz;
+  }
 
   as->mcp = (MCode*)hdr;
 

src/lj_emit_x86.h

@@ -614,6 +614,24 @@ static void emit_addptr(ASMState *as, Reg r, int32_t ofs)
 #define TEMPSPILL (1*sizeof(intptr_t))
 #define CONTEXTSPILL (0)
 
+
+static MCode* emit_intrins(ASMState *as, CIntrinsic *intrins, Reg r1, 
+                           uintptr_t r2)
+{
+  if (intrins->flags & INTRINSFLAG_CALLED) {
+    lua_assert(r2);
+    emit_call_(as, (MCode*)r2, r1);
+    return NULL;
+  } else {
+    AsmHeader *hdr = ((AsmHeader*)r2)-1;
+    lua_assert((hdr->asmofs != 0 || hdr->asmofs < hdr->totalzs));
+
+    /* Directly copy the unmodified machine code of the intrinsic in */
+    asm_mcode(as, ((char*)r2)+hdr->asmofs, hdr->asmsz);
+  }
+  return as->mcp;
+}
+
 static int lj_popcnt(uint32_t i)
 {
   i = i - ((i >> 1) & 0x55555555);

src/lj_intrinsic.c

@@ -83,6 +83,16 @@ static CTypeID register_intrinsic(lua_State *L, CIntrinsic* src, CType *func)
   return id;
 }
 
+static void lj_intrinsic_new(lua_State *L, CTypeID id, void* wrapmc)
+{
+  CTState *cts = ctype_cts(L);
+  GCcdata *cd;
+  lua_assert(ctype_isintrinsic(ctype_get(cts, id)->info));
+  cd = lj_cdata_new(cts, id, CTSIZE_PTR);
+  *(void **)cdataptr(cd) = wrapmc;
+  setcdataV(L, L->top++, cd);
+}
+
 static int parse_fprreg(const char *name, uint32_t len)
 {
   uint32_t rid = 0, kind = REGKIND_FPR64;
@@ -282,25 +292,31 @@ static void setopcode(lua_State *L, CIntrinsic *intrins, uint32_t opcode)
   intrins->opcode = opcode;
 }
 
-static int parse_opstr(lua_State *L, GCstr *opstr)
+static int parse_opstr(lua_State *L, GCstr *opstr, CIntrinsic *intrins, int* buildflags)
 {
   const char *op = strdata(opstr);
   uint32_t opcode = 0;
   uint32_t i;
 
-  /* Find the end of the opcode number */
+  /* Parse the opcode number if this is not a template */
-  for (i = 0; i < opstr->len && lj_char_isxdigit((uint8_t)op[i]); i++) {
+  if (op[0] != '?') {
-  }
+    for (i = 0; i < opstr->len && lj_char_isxdigit((uint8_t)op[i]); i++) {
+    }
 
-  if (i == 0 || i > 8) {
+    if (i == 0 || i > 8) {
-    /* invalid or no hex number */
+      /* invalid or no hex number */
-    lj_err_callerv(L, LJ_ERR_FFI_BADOPSTR, op, "invalid opcode number");
+      lj_err_callerv(L, LJ_ERR_FFI_BADOPSTR, op, "invalid opcode number");
-  }
+    }
 
-  /* Scan hex digits. */
+    /* Scan hex digits. */
-  for (; i; i--, op++) {
+    for (; i; i--, op++) {
-    uint32_t d = *op; if (d > '9') d += 9;
+      uint32_t d = *op; if (d > '9') d += 9;
-    opcode = (opcode << 4) + (d & 15);
+      opcode = (opcode << 4) + (d & 15);
+    }
+
+  } else {
+    *buildflags |= INTRINSFLAG_TEMPLATE;
+    op++;
   }
   
   return opcode;
@@ -337,6 +353,54 @@ static IntrinsicWrapper lj_intrinsic_buildwrap(lua_State *L, CIntrinsic *intrins
   return (IntrinsicWrapper)state.wrapper;
 }
 
+CTypeID lj_intrinsic_template(lua_State *L, int narg)
+{
+  CTState *cts = ctype_cts(L);
+  CType *ct;
+  CTypeID id;
+  CIntrinsic* intrins;
+  GCstr *name = lj_lib_checkstr(L, narg);
+
+  id = lj_ctype_getname(cts, &ct, name, 1u << CT_FUNC);
+
+  if (!id) {
+    lj_err_argv(L, narg, LJ_ERR_FFI_NODECL, name);
+  } else if (!ctype_isintrinsic(ct->info)) {
+    lj_err_arg(L, narg, LJ_ERR_FFI_INVTYPE);
+  }
+
+  intrins = lj_intrinsic_get(cts, ct->size);
+
+  /* Can't be a template if it an opcode */
+  if ((intrins->opcode && intrins->outsz <= 4) || intrins->wrapped)
+    lj_err_arg(L, narg, LJ_ERR_FFI_INVTYPE);
+
+  return id;
+}
+
+int lj_intrinsic_create(lua_State *L)
+{
+  CTState *cts = ctype_cts(L);  
+  CTypeID id = lj_intrinsic_template(L, 1);
+  void *intrinsmc;
+  MSize asmsz;
+  CIntrinsic* intrins = lj_intrinsic_get(cts, ctype_get(cts, id)->size);
+  
+  lj_cconv_ct_tv(cts, ctype_get(cts, CTID_P_CVOID), (uint8_t *)&intrinsmc,
+                 L->base+1, CCF_ARG(2));
+
+  asmsz = lj_lib_checkint(L, 3);
+  if (asmsz <= 0 || asmsz > 0xffff ||
+      asmsz > (MSize)(L2J(L)->param[JIT_P_sizemcode] << 10)) {
+    lj_err_callermsg(L, "bad code size");
+  }
+
+  intrinsmc = lj_intrinsic_buildwrap(L, intrins, intrinsmc, asmsz, 
+                                     intrin_getmodrset(cts, intrins));
+  lj_intrinsic_new(L, id, intrinsmc);
+  return 1;
+}
+
 GCcdata *lj_intrinsic_createffi(CTState *cts, CType *func)
 {
   GCcdata *cd;
@@ -346,6 +410,10 @@ GCcdata *lj_intrinsic_createffi(CTState *cts, CType *func)
   uint32_t op = intrins->opcode;
   void* mcode = ((char*)&op) + (4-intrin_oplen(intrins));
   
+  if (intrins->opcode == 0) {
+    lj_err_callermsg(cts->L, "expected non template intrinsic");
+  }
+
   intrins->wrapped = lj_intrinsic_buildwrap(cts->L, intrins, mcode,
                                             intrin_oplen(intrins), mod);
 
@@ -360,13 +428,14 @@ int lj_intrinsic_fromcdef(lua_State *L, CTypeID fid, GCstr *opstr, uint32_t imm)
   CType *func = ctype_get(cts, fid);
   CTypeID sib = func->sib, retid = ctype_cid(func->info);
   uint32_t opcode;
+  int buildflags = 0;
   CIntrinsic _intrins;
   CIntrinsic* intrins = &_intrins;
   memset(intrins, 0, sizeof(CIntrinsic));
   
-  opcode = parse_opstr(L, opstr);
+  opcode = parse_opstr(L, opstr, intrins, &buildflags);
 
-  if (!opcode) {
+  if (!opcode && !(buildflags & INTRINSFLAG_TEMPLATE)) {
     return 0;
   }
 
@@ -387,7 +456,10 @@ int lj_intrinsic_fromcdef(lua_State *L, CTypeID fid, GCstr *opstr, uint32_t imm)
     sib = retid;
   }
 
-  setopcode(L, intrins, opcode);
+  /* If were a template theres no opcode to set */
+  if (opcode) {
+    setopcode(L, intrins, opcode);
+  } 
   register_intrinsic(L, intrins, ctype_get(cts, fid));
 
   lua_assert(sib > 0 && sib < cts->top);

src/lj_intrinsic.h

@@ -26,6 +26,10 @@ typedef enum INTRINSFLAGS {
   INTRINSFLAG_CALLED = 0x20,
   /* MODRM should always be set as indirect mode */
   INTRINSFLAG_INDIRECT = 0x40,
+  /* Intrinsic is a template with no machine code set until instantiate at runtime with
+  ** user supplied code.
+  */
+  INTRINSFLAG_TEMPLATE    = 0x40000,
 
   INTRINSFLAG_CALLEDIND = INTRINSFLAG_CALLED | INTRINSFLAG_INDIRECT
 } INTRINSFLAGS;
@@ -97,6 +101,7 @@ CTypeID1 regkind_ct[16];
 #define rk_ctype(rid, kind) ((rid) < RID_MAX_GPR ? rk_ctypegpr(kind) : rk_ctypefpr(kind))
 
 LJ_FUNC void lj_intrinsic_init(lua_State *L);
+LJ_FUNC int lj_intrinsic_create(lua_State *L);
 LJ_FUNC GCcdata *lj_intrinsic_createffi(CTState *cts, CType *func);
 LJ_FUNC int lj_intrinsic_fromcdef(lua_State *L, CTypeID fid, GCstr *opcode, uint32_t imm);
 LJ_FUNC int lj_intrinsic_call(CTState *cts, CType *ct);

tests/intrinsic_spec.lua

@@ -30,7 +30,7 @@ describe("intrinsic tests", function()
 context("nop inout", function()
 
   it("fpr", function()
-    assert_cdef([[void fpr_nop1(double xmm0) __mcode("90") __reglist(out, double xmm0)]], "fpr_nop1")
+    assert_cdef([[void fpr_nop1(double xmm0) __mcode("90_E") __reglist(out, double xmm0)]], "fpr_nop1")
     local fpr1 = ffi.C.fpr_nop1
 
     assert_error(function() fpr1() end)
@@ -39,10 +39,34 @@ context("nop inout", function()
   
     assert_jit(123.075, function(num) return (fpr1(num)) end, 123.075)
     assert_noexit(-123567.075, function(num) return (fpr1(num)) end, -123567.075)
+    
+    assert_cdef([[void fpr_all(double xmm0, double xmm1, double xmm2, double xmm3, double xmm4, double xmm5, double xmm6, double xmm7) __mcode("?E") 
+                                __reglist(out,double xmm0, double xmm1, double xmm2, double xmm3, double xmm4, double xmm5, double xmm6, double xmm7)]])
+    local fpr_all = ffi.intrinsic("fpr_all", "\x90", 1)
+                                       
+    local function testfpr_all(i, r1, r2, r3, r4, r5, r6, r7, r8)
+      local spilled = r1*2*i
+      local ro1, ro2, ro3, ro4, ro5, ro6, ro7, ro8 = fpr_all(r1, r2, r3, r4, r5, r6, r7, r8)
+      return ro1+i, ro2+i, ro3+i, ro4+i, ro5+i, ro6+i, ro7+i, (ro8*ro3)+i, ro2+spilled
+    end
+
+    local function checker(i, ro1, ro2, ro3, ro4, ro5, ro6, ro7, ro8, spilled)
+      assert(ro1 == 1.5+i)
+      assert(ro2 == 2.5+i)
+      assert(ro3 == 3.5+i)
+      assert(ro4 == 4.5+i)
+      assert(ro5 == 5.5+i)
+      assert(ro6 == 60000.525+i)
+      assert(ro7 == i+-7.5)
+      assert(ro8 == (-100*3.5)+i)
+      assert(spilled == 2.5+(1.5*2*i))
+    end
+    
+    assert_jitchecker(checker, testfpr_all, 1.5, 2.5, 3.5, 4.5, 5.5, 60000.525, -7.5, -100)
   end)
   
   it("gpr", function()
-    assert_cdef([[void gpr_nop1(int32_t eax) __mcode("90") __reglist(out, int32_t eax)]], "gpr_nop1")
+    assert_cdef([[void gpr_nop1(int32_t eax) __mcode("90_E") __reglist(out, int32_t eax)]], "gpr_nop1")
 
     local function testgpr1(num) 
       return (ffi.C.gpr_nop1(num)) 
@@ -67,11 +91,35 @@ context("nop inout", function()
     end
 
     assert_jitchecker(checker, testgpr_scratch, 0, 1, 30000)  
+
+    assert_cdef([[void gpr_all(int32_t ebp, int32_t esi, int32_t edi, int32_t eax, int32_t ebx, int32_t ecx, int32_t edx) __mcode("?E") 
+                    __reglist(out, int32_t ebp, int32_t esi, int32_t edi, int32_t eax, int32_t ebx, int32_t ecx, int32_t edx)]])
+                    
+    local gpr_all = ffi.intrinsic("gpr_all", "\x90", 1)
+    
+    local function testgpr_all(i, r1, r2, r3, r4, r5, r6, r7)
+      local spilled = r1+(10000*i)
+      local ro1, ro2, ro3, ro4, ro5, ro6, ro7 = gpr_all(r1, r2, r3, r4, 100, r6, r7)
+      return spilled+(ro1+ro2+ro3+ro4+ro5+ro6+ro7), ro3+i, ro2+i, ro1+i, ro4+i, ro5, ro6+i, ro7+i
+    end
+
+    local function checker(i, spilled, ro3, ro2, ro1, ro4, ro5, ro6, ro7)
+      assert(ro1 == 1+i)
+      assert(ro2 == 2+i)
+      assert(ro3 == 3+i)
+      assert(ro4 == 4+i)
+      assert(ro5 == 100)
+      assert(ro6 == 6+i)
+      assert(ro7 == 7+i)
+      assert(spilled == 124+(10000*i))
+    end
+
+    assert_jitchecker(checker, testgpr_all, 1, 2, 3, 4, 5, 6, 7)
   end)
 
 if ffi.arch == "x64" then
   it("gpr64", function()
-    assert_cdef([[void gpr64_1(int64_t rdx) __mcode("90") __reglist(out, int64_t rdx)]], "gpr64_1")
+    assert_cdef([[void gpr64_1(int64_t rdx) __mcode("90_E") __reglist(out, int64_t rdx)]], "gpr64_1")
 
     local function testgpr1(num) 
       return (ffi.C.gpr64_1(num)) 
@@ -79,10 +127,35 @@ if ffi.arch == "x64" then
 
     assert_jit(1235678ull, testgpr1, 1235678)
     assert_noexit(-1LL, testgpr1, -1)
+
+    assert_cdef([[void gpr64(int64_t rbp, int64_t rsi, int64_t rdi, int64_t rax, int64_t rbx, int64_t rcx, int64_t rdx) __mcode("?E") 
+                    __reglist(out, int64_t rbp, int64_t rsi, int64_t rdi, int64_t rax, int64_t rbx, int64_t rcx, int64_t rdx)]])
+                    
+    local gpr7 = ffi.intrinsic("gpr64", "\x90", 1)
+        
+    local function testgpr_all(i, r1, r2, r3, r4, r5, r6, r7)
+      local spilled = r1+(10000*i)
+      local ro1, ro2, ro3, ro4, ro5, ro6, ro7 = gpr7(r1, r2, r3, 68719476735ll, r5, r6, r7)
+      return spilled, ro3+i, ro2+i, ro1+i, ro4+i, ro5+i, ro6+i, ro7+i
+    end
+    
+    local function checker(i, spilled, ro3, ro2, ro1, ro4, ro5, ro6, ro7)
+      local sp = (10000*i)+1
+      assert(ro1 == 1+i)
+      assert(ro2 == 2+i)
+      assert(ro3 == 3+i)
+      assert(type(ro4) == "cdata" and ro4 == 68719476735ll+i)
+      assert(type(ro5) == "cdata" and ro5 == 5ll+i)
+      assert(type(ro6) == "cdata" and ro6 == 68719476721ll+i)
+      assert(type(ro7) == "cdata" and ro7 == (-7ll)+i)
+      assert(spilled == sp)
+    end
+    
+    assert_jitchecker(checker, testgpr_all, 1, 2, 3, 4, 5ll, 68719476721ll, -7ll)
   end)
   
   it("rex fpr", function()
-    assert_cdef([[void fpr_reg(double xmm9, double xmm0) __mcode("90") __reglist(out, double xmm0, double xmm9)]], "fpr_reg")
+    assert_cdef([[void fpr_reg(double xmm9, double xmm0) __mcode("90_E") __reglist(out, double xmm0, double xmm9)]], "fpr_reg")
     local fpr = ffi.C.fpr_reg
   
     local function testrex(n1, n2)
@@ -95,15 +168,38 @@ if ffi.arch == "x64" then
 end
   
   it("fpr_vec", function()
-    assert_cdef([[void fpr_vec(void* xmm7v) __mcode("90") __reglist(out, float4 xmm7v)]], "fpr_vec")
+    assert_cdef([[void fpr_vec(void* xmm7v) __mcode("90_E") __reglist(out, float4 xmm7v)]], "fpr_vec")
   
     local v1 = ffi.new("float[4]", 1, 2, 3, 4)
     local xmmout = ffi.C.fpr_vec(v1)  
     assert_v4eq(xmmout, 1, 2, 3, 4)    
   end)
   
+  it("check extra register spill", function()
+    local array = ffi.new("float4", 1, 2, 3, 4)
+    
+    -- Use up all gpr scatch registers before loading the vec causing wrapper builder to restart with an extra spill
+    assert_cdef([[void spillrestart(float4 xmm0v, int32_t eax, int32_t ecx, int32_t edx, int32_t esi, int32_t edi, int32_t ebx) __mcode("?E") 
+                                  __reglist(out, int32_t eax, int32_t ecx, int32_t edx, int32_t esi, int32_t edi, int32_t ebx, float4 xmm0v) 
+                 ]])
+    
+    local xmmtest = ffi.intrinsic("spillrestart", "\x90", 1)
+    local eax, ecx, edx, esi, edi, ebx, xmmout = xmmtest(array, 1, 2, 3, 4, 5, 6)
+    
+    assert_equal(eax, 1)
+    assert_equal(ecx, 2)
+    assert_equal(edx, 3)
+    assert_equal(esi, 4)
+    assert_equal(edi, 5)
+    assert_equal(ebx, 6)
+
+    for i=1,4 do
+      assert_equal(xmmout[i-1], i)
+    end
+  end) 
+
   it("idiv", function()
-    assert_cdef([[void idiv(int32_t eax, int32_t ecx) __mcode("99F7F9") __reglist(out, int32_t eax, int32_t edx)]], "idiv")
+    assert_cdef([[void idiv(int32_t eax, int32_t ecx) __mcode("99F7F9_E") __reglist(out, int32_t eax, int32_t edx)]], "idiv")
 
     local function checker(i, result, remainder)
       local rem = i%3
@@ -169,22 +265,22 @@ context("__mcode", function()
   end)
   
   it("invalid registers", function()
-    assert_cdef([[void validreg_gpr(int eax) __mcode("90");]], "validreg_gpr")
+    assert_cdef([[void validreg_gpr(int eax) __mcode("90_E");]], "validreg_gpr")
     
-    assert_cdeferr([[void badreg_1(int e) __mcode("90");]], "invalid")
+    assert_cdeferr([[void badreg_1(int e) __mcode("90_E");]], "invalid")
-    assert_cdeferr([[void badreg_1(int r20d) __mcode("90");]], "invalid") 
+    assert_cdeferr([[void badreg_1(int r20d) __mcode("90_E");]], "invalid") 
-    assert_cdeferr([[void badreg_gpr1() __mcode("90") __reglist(out, int e);]], "invalid")
+    assert_cdeferr([[void badreg_gpr1() __mcode("90_E") __reglist(out, int e);]], "invalid")
-    assert_cdeferr([[void badreg_gpr2() __mcode("90") __reglist(mod, e);]], "invalid")
+    assert_cdeferr([[void badreg_gpr2() __mcode("90_E") __reglist(mod, e);]], "invalid")
     
-    assert_cdef([[void validreg_fpr(float xmm0) __mcode("90");]], "validreg_fpr")
+    assert_cdef([[void validreg_fpr(float xmm0) __mcode("90_E");]], "validreg_fpr")
     
-    assert_cdeferr([[void badreg_fpr1(float x) __mcode("90");]], "invalid")
+    assert_cdeferr([[void badreg_fpr1(float x) __mcode("90_E");]], "invalid")
-    assert_cdeferr([[void badreg_fpr1(float xm) __mcode("90");]], "invalid")
+    assert_cdeferr([[void badreg_fpr1(float xm) __mcode("90_E");]], "invalid")
-    assert_cdeferr([[void badreg_fpr1(float xm0) __mcode("90");]], "invalid")
+    assert_cdeferr([[void badreg_fpr1(float xm0) __mcode("90_E");]], "invalid")
-    assert_cdeferr([[void badreg_fpr1(float xmmm0) __mcode("90");]], "invalid")
+    assert_cdeferr([[void badreg_fpr1(float xmmm0) __mcode("90_E");]], "invalid")
-    assert_cdeferr([[void badreg_fpr2(float xmm0vf) __mcode("90");]], "invalid")
+    assert_cdeferr([[void badreg_fpr2(float xmm0vf) __mcode("90_E");]], "invalid")
     --xmm register number too large
-    assert_cdeferr([[void badreg_fpr1(float xmm20) __mcode("90");]], "invalid")
+    assert_cdeferr([[void badreg_fpr1(float xmm20) __mcode("90_E");]], "invalid")
   end)
   
   it("multidef rollback", function()
@@ -199,7 +295,7 @@ context("__mcode", function()
     assert_error(function() ffi.C.multi2() end)
     
     assert_not_error(function() ffi.cdef[[
-      void multi1(int32_t eax) __mcode("90") __reglist(out, int32_t eax);
+      void multi1(int32_t eax) __mcode("90_E") __reglist(out, int32_t eax);
     ]] end)
 
     assert_equal(ffi.C.multi1(1.1), 1)
@@ -217,7 +313,7 @@ context("__mcode", function()
   end)
 
   it("bad args", function()
-    assert_cdef([[void idiv2(int32_t eax, int32_t ecx) __mcode("99F7F9") __reglist(out, int32_t eax, int32_t edx)]], "idiv2")
+    assert_cdef([[void idiv2(int32_t eax, int32_t ecx) __mcode("99F7F9_E") __reglist(out, int32_t eax, int32_t edx)]], "idiv2")
     
     local idiv = ffi.C.idiv2
     
@@ -234,8 +330,42 @@ context("__mcode", function()
     assert_error(function() idiv(1, 2, 3, 4) end)
   end) 
   
+  it("idiv(template)", function()
+    assert_cdef([[void idivT(int32_t eax, int32_t ecx) __mcode("?E") __reglist(out, int32_t eax, int32_t edx)]])
+    --trying to create template intrinsic through C library should always fail
+    assert_error(function() return ffi.C.idivT end)
+    
+    local idiv = ffi.intrinsic("idivT", "\x99\xF7\xF9", 3)
+
+    local function checker(i, result, remainder)
+      local rem = i%2
+    
+      if rem ~= remainder then
+        return rem, remainder
+      end
+    
+      local expected = (i-rem)/2
+      
+      if expected ~= result then
+        return expected, result
+      end
+    end
+    
+    local function test_idiv(value, divisor)
+      local result, remainder = idiv(value, divisor)
+      return result, remainder
+    end
+  
+    assert_jitchecker(checker, test_idiv, 2)
+    
+    -- create a second instance and check guard for wrapper pointer fails
+    idiv = ffi.intrinsic("idivT", "\x90", 1)
+    
+    assert_exit(10, test_idiv, 10, 5)
+  end)
+  
   it("cpuid_brand", function()
-    assert_cdef([[void cpuid(int32_t eax, int32_t ecx) __mcode("0FA2") __reglist(out, int32_t eax, int32_t ebx, int32_t ecx, int32_t edx);]], "cpuid")
+    assert_cdef([[void cpuid(int32_t eax, int32_t ecx) __mcode("0FA2_E") __reglist(out, int32_t eax, int32_t ebx, int32_t ecx, int32_t edx);]], "cpuid")
 
     local cpuid = ffi.C.cpuid
     
@@ -304,20 +434,20 @@ context("__reglist", function()
   
   it("stack pointer blacklist", function()
   
-    assert_cdeferr([[void blacklist_in(int esp) __mcode("90")]], "blacklist")
+    assert_cdeferr([[void blacklist_in(int esp) __mcode("90_E")]], "blacklist")
-    assert_cdeferr([[void blacklist_out(int eax) __mcode("90") __reglist(out, int esp)]], "blacklist")
+    assert_cdeferr([[void blacklist_out(int eax) __mcode("90_E") __reglist(out, int esp)]], "blacklist")
     --FIXME
-    --assert_cdeferr([[void blacklist_mod(int eax) __mcode("90") __reglist(mod, esp)]], "blacklist")
+    --assert_cdeferr([[void blacklist_mod(int eax) __mcode("90_E") __reglist(mod, esp)]], "blacklist")
   
     if ffi.arch == "x64" then
-      assert_cdeferr([[void blacklist_64(int rsp) __mcode("90")]], "blacklist")
+      assert_cdeferr([[void blacklist_64(int rsp) __mcode("90_E")]], "blacklist")
     end
   end)
   
   it("duplicate regs", function()
-    assert_cdeferr([[void duplicate_in(int eax, int eax) __mcode("90")]], "duplicate")
+    assert_cdeferr([[void duplicate_in(int eax, int eax) __mcode("90_E")]], "duplicate")
-    assert_cdeferr([[void duplicate_inxmm(float4 xmm0, float4 xmm0) __mcode("90")]], "duplicate")
+    assert_cdeferr([[void duplicate_inxmm(float4 xmm0, float4 xmm0) __mcode("90_E")]], "duplicate")
-    assert_cdeferr([[void duplicate_out(int eax) __mcode("90") __reglist(out, int eax, int eax)]], "duplicate")
+    assert_cdeferr([[void duplicate_out(int eax) __mcode("90_E") __reglist(out, int eax, int eax)]], "duplicate")
     --FIXME assert_cdeferr([[void duplicate_mod(int eax) __mcode("90_E") __reglist(mod, eax, eax)]], "duplicate")
   end)