From 9f5602ad073b62d269c034fb78fe08abb52def16 Mon Sep 17 00:00:00 2001
From: David Schleef <ds@schleef.org>
Date: Tue, 27 Jul 2010 17:08:02 -0700
Subject: [PATCH] Reimplement emulation

---
 orc/Makefile.am          |   10 +-
 orc/generate-emulation.c |  302 +++++
 orc/opcodes.h            |   20 +-
 orc/orcemulateopcodes.c  | 3302 ++++++++++++++++++++++++++++++++++++++++++++++
 orc/orcemulateopcodes.h  |  248 ++++
 orc/orcexecutor.c        |  199 +--
 orc/orcopcodes.c         |  267 ++--
 orc/orcprogram-c.c       |   66 +-
 orc/orcprogram.h         |    9 +
 9 files changed, 4178 insertions(+), 245 deletions(-)
 create mode 100644 orc/generate-emulation.c
 create mode 100644 orc/orcemulateopcodes.c
 create mode 100644 orc/orcemulateopcodes.h

diff --git a/orc/Makefile.am b/orc/Makefile.am
index 1ad75bc..6d9178a 100644
--- a/orc/Makefile.am
+++ b/orc/Makefile.am
@@ -11,6 +11,7 @@ liborc_@ORC_MAJORMINOR@_la_CFLAGS = $(ORC_CFLAGS) \
 
 liborc_@ORC_MAJORMINOR@_la_SOURCES = \
 	orc.c \
+	orcemulateopcodes.c \
 	orcexecutor.c \
 	orcfunctions.c \
 	orcutils.c \
@@ -88,8 +89,13 @@ nodist_pkginclude_HEADERS = orc-stdint.h
 
 DISTCLEANFILES = orc-stdint.h
 
+noinst_PROGRAMS = generate-emulation
 
-orcfunctions:
+generate_emulation_LDADD = $(ORC_LIBS)
+generate_emulation_CFLAGS = $(ORC_CFLAGS)
+
+update:
 	$(top_builddir)/tools/orcc$(EXEEXT) --implementation -o orcfunctions.c orcfunctions.orc
 	$(top_builddir)/tools/orcc$(EXEEXT) --header -o orcfunctions.h orcfunctions.orc
-
+	./generate-emulation$(EXEEXT) -o orcemulateopcodes.c
+	
diff --git a/orc/generate-emulation.c b/orc/generate-emulation.c
new file mode 100644
index 0000000..b26e2aa
--- /dev/null
+++ b/orc/generate-emulation.c
@@ -0,0 +1,302 @@
+
+#include "config.h"
+
+#include <orc/orc.h>
+#include <orc/orcparse.h>
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <ctype.h>
+
+void output_code_emulate (OrcProgram *p, FILE *output);
+
+int verbose = 0;
+int error = 0;
+int compat;
+
+char *target = "sse";
+
+#define ORC_VERSION(a,b,c,d) ((a)*1000000 + (b)*10000 + (c)*100 + (d))
+#define REQUIRE(a,b,c,d) do { \
+  if (ORC_VERSION((a),(b),(c),(d)) > compat) { \
+    fprintf(stderr, "Feature used that is incompatible with --compat\n"); \
+    exit (1); \
+  } \
+} while (0)
+
+void help (void)
+{
+  printf("Usage:\n");
+  printf("  generate-emulation [OPTION...]\n");
+  printf("\n");
+  printf("Help Options:\n");
+  printf("  -h, --help              Show help options\n");
+  printf("\n");
+  printf("Application Options:\n");
+  printf("  -o, --output FILE       Write output to FILE\n");
+  printf("\n");
+
+  exit (0);
+}
+
+int
+main (int argc, char *argv[])
+{
+  char *output_file = NULL;
+  char *input_file = NULL;
+  char *include_file = NULL;
+  FILE *output;
+  int i;
+  OrcOpcodeSet *opcode_set;
+
+  orc_init ();
+
+  for(i=1;i<argc;i++) {
+    if (strcmp (argv[i], "--output") == 0 ||
+        strcmp(argv[i], "-o") == 0) {
+      if (i+1 < argc) {
+        output_file = argv[i+1];
+        i++;
+      } else {
+        help();
+      }
+    } else if (strncmp(argv[i], "-", 1) == 0) {
+      printf("Unknown option: %s\n", argv[i]);
+      exit (1);
+    } else {
+      if (input_file == NULL) {
+        input_file = argv[i];
+      } else {
+        printf("More than one input file specified: %s\n", argv[i]);
+        exit (1);
+      }
+    }
+  }
+
+  if (output_file == NULL) {
+    output_file = "out.c";
+  }
+
+  output = fopen (output_file, "w");
+  if (!output) {
+    printf("Could not write output file: %s\n", output_file);
+    exit(1);
+  }
+
+  fprintf(output, "\n");
+  fprintf(output, "/* autogenerated by generate-emulation */\n");
+  fprintf(output, "\n");
+
+  fprintf(output, "#ifdef HAVE_CONFIG_H\n");
+  fprintf(output, "#include \"config.h\"\n");
+  fprintf(output, "#endif\n");
+  fprintf(output, "#include <math.h>\n");
+  fprintf(output, "#include <orc/orc.h>\n");
+  if (include_file) {
+    fprintf(output, "#include <%s>\n", include_file);
+  }
+  fprintf(output, "\n");
+  fprintf(output, "%s", orc_target_get_asm_preamble ("c"));
+  fprintf(output, "\n");
+
+  opcode_set = orc_opcode_set_get ("sys");
+
+  for(i=0;i<opcode_set->n_opcodes;i++){
+    char s[40];
+    OrcProgram *program;
+    OrcStaticOpcode *opcode = opcode_set->opcodes + i;
+
+    program = orc_program_new ();
+
+    sprintf(s, "emulate_%s", opcode->name);
+    orc_program_set_name (program, s);
+
+    if (opcode->flags & ORC_STATIC_OPCODE_ACCUMULATOR) {
+      orc_program_add_accumulator (program, opcode->dest_size[0], "d1");
+    } else {
+      orc_program_add_destination (program, opcode->dest_size[0], "d1");
+    }
+    if (opcode->dest_size[1]) {
+      orc_program_add_destination (program, opcode->dest_size[1], "d2");
+    }
+    orc_program_add_source (program, opcode->src_size[0], "s1");
+    if (opcode->src_size[1]) {
+      if (opcode->flags & ORC_STATIC_OPCODE_SCALAR) {
+        orc_program_add_parameter (program, opcode->src_size[1], "s2");
+      } else {
+        orc_program_add_source (program, opcode->src_size[1], "s2");
+      }
+    }
+    if (opcode->src_size[1]) {
+      orc_program_append_str (program, opcode->name, "d1", "s1", "s2");
+    } else {
+      if (opcode->dest_size[1]) {
+        orc_program_append_dds_str (program, opcode->name, "d1", "d2", "s1");
+      } else {
+        orc_program_append_ds_str (program, opcode->name, "d1", "s1");
+      }
+    }
+
+    output_code_emulate (program, output);
+  }
+
+  fclose (output);
+
+  if (error) exit(1);
+
+  return 0;
+}
+
+
+const char *varnames[] = {
+  "d1", "d2", "d3", "d4",
+  "s1", "s2", "s3", "s4",
+  "s5", "s6", "s7", "s8",
+  "a1", "a2", "a3", "d4",
+  "c1", "c2", "c3", "c4",
+  "c5", "c6", "c7", "c8",
+  "p1", "p2", "p3", "p4",
+  "p5", "p6", "p7", "p8",
+  "t1", "t2", "t3", "t4",
+  "t5", "t6", "t7", "t8",
+  "t9", "t10", "t11", "t12",
+  "t13", "t14", "t15", "t16"
+};
+
+const char *enumnames[] = {
+  "ORC_VAR_D1", "ORC_VAR_D2", "ORC_VAR_D3", "ORC_VAR_D4",
+  "ORC_VAR_S1", "ORC_VAR_S2", "ORC_VAR_S3", "ORC_VAR_S4",
+  "ORC_VAR_S5", "ORC_VAR_S6", "ORC_VAR_S7", "ORC_VAR_S8",
+  "ORC_VAR_A1", "ORC_VAR_A2", "ORC_VAR_A3", "ORC_VAR_A4",
+  "ORC_VAR_C1", "ORC_VAR_C2", "ORC_VAR_C3", "ORC_VAR_C4",
+  "ORC_VAR_C5", "ORC_VAR_C6", "ORC_VAR_C7", "ORC_VAR_C8",
+  "ORC_VAR_P1", "ORC_VAR_P2", "ORC_VAR_P3", "ORC_VAR_P4",
+  "ORC_VAR_P5", "ORC_VAR_P6", "ORC_VAR_P7", "ORC_VAR_P8",
+  "ORC_VAR_T1", "ORC_VAR_T2", "ORC_VAR_T3", "ORC_VAR_T4",
+  "ORC_VAR_T5", "ORC_VAR_T6", "ORC_VAR_T7", "ORC_VAR_T8",
+  "ORC_VAR_T9", "ORC_VAR_T10", "ORC_VAR_T11", "ORC_VAR_T12",
+  "ORC_VAR_T13", "ORC_VAR_T14", "ORC_VAR_T15", "ORC_VAR_T16"
+};
+
+#if 0
+void
+output_prototype (OrcProgram *p, FILE *output)
+{
+  OrcVariable *var;
+  int i;
+  int need_comma;
+
+  fprintf(output, "%s (", p->name);
+  need_comma = FALSE;
+  for(i=0;i<4;i++){
+    var = &p->vars[ORC_VAR_D1 + i];
+    if (var->size) {
+      if (need_comma) fprintf(output, ", ");
+      if (var->type_name) {
+        fprintf(output, "%s * %s", var->type_name,
+            varnames[ORC_VAR_D1 + i]);
+      } else {
+        fprintf(output, "orc_uint%d * %s", var->size*8,
+            varnames[ORC_VAR_D1 + i]);
+      }
+      if (p->is_2d) {
+        fprintf(output, ", int %s_stride", varnames[ORC_VAR_D1 + i]);
+      }
+      need_comma = TRUE;
+    }
+  }
+  for(i=0;i<4;i++){
+    var = &p->vars[ORC_VAR_A1 + i];
+    if (var->size) {
+      if (need_comma) fprintf(output, ", ");
+      if (var->type_name) {
+        fprintf(output, "%s * %s", var->type_name,
+            varnames[ORC_VAR_A1 + i]);
+      } else {
+        fprintf(output, "orc_uint%d * %s", var->size*8,
+            varnames[ORC_VAR_A1 + i]);
+      }
+      need_comma = TRUE;
+    }
+  }
+  for(i=0;i<8;i++){
+    var = &p->vars[ORC_VAR_S1 + i];
+    if (var->size) {
+      if (need_comma) fprintf(output, ", ");
+      if (var->type_name) {
+        fprintf(output, "const %s * %s", var->type_name,
+            varnames[ORC_VAR_S1 + i]);
+      } else {
+        fprintf(output, "const orc_uint%d * %s", var->size*8,
+            varnames[ORC_VAR_S1 + i]);
+      }
+      if (p->is_2d) {
+        fprintf(output, ", int %s_stride", varnames[ORC_VAR_S1 + i]);
+      }
+      need_comma = TRUE;
+    }
+  }
+  for(i=0;i<8;i++){
+    var = &p->vars[ORC_VAR_P1 + i];
+    if (var->size) {
+      if (need_comma) fprintf(output, ", ");
+      if (var->is_float_param) {
+        REQUIRE(0,4,5,1);
+      }
+      fprintf(output, "%s %s",
+          var->is_float_param ? "float" : "int",
+          varnames[ORC_VAR_P1 + i]);
+      need_comma = TRUE;
+    }
+  }
+  if (p->constant_n == 0) {
+    if (need_comma) fprintf(output, ", ");
+    fprintf(output, "int n");
+    need_comma = TRUE;
+  }
+  if (p->is_2d && p->constant_m == 0) {
+    if (need_comma) fprintf(output, ", ");
+    fprintf(output, "int m");
+  }
+  fprintf(output, ")");
+}
+
+void
+output_code_header (OrcProgram *p, FILE *output)
+{
+  fprintf(output, "void ");
+  output_prototype (p, output);
+  fprintf(output, ";\n");
+}
+#endif
+
+void
+output_code_emulate (OrcProgram *p, FILE *output)
+{
+
+  fprintf(output, "void\n");
+  if (p->constant_n) {
+    fprintf(output, "%s (OrcOpcodeExecutor *ex)\n", p->name);
+  } else {
+    fprintf(output, "%s (OrcOpcodeExecutor *ex, int n)\n", p->name);
+  }
+  fprintf(output, "{\n");
+  {
+    OrcCompileResult result;
+
+    result = orc_program_compile_full (p, orc_target_get_by_name("c"),
+        ORC_TARGET_C_BARE | ORC_TARGET_C_OPCODE);
+    if (ORC_COMPILE_RESULT_IS_SUCCESSFUL(result)) {
+      fprintf(output, "%s\n", orc_program_get_asm_code (p));
+    } else {
+      printf("Failed to compile %s\n", p->name);
+      error = TRUE;
+    }
+  }
+  fprintf(output, "}\n");
+  fprintf(output, "\n");
+
+}
+
+
diff --git a/orc/opcodes.h b/orc/opcodes.h
index 3b13344..15883ef 100644
--- a/orc/opcodes.h
+++ b/orc/opcodes.h
@@ -111,16 +111,16 @@ UNARY_LW(select1lw, "((orc_uint32)%s >> 16)&0xffff")
 UNARY_UW(swapw, "ORC_SWAP_W(%s)")
 UNARY_UL(swapl, "ORC_SWAP_L(%s)")
 
-BINARY_F(addf, "%s + %s")
-BINARY_F(subf, "%s - %s")
-BINARY_F(mulf, "%s * %s")
-BINARY_F(divf, "%s / %s")
-UNARY_F(sqrtf, "sqrt(%s)")
-BINARY_F(maxf, "ORC_MAX(%s,%s)")
-BINARY_F(minf, "ORC_MIN(%s,%s)")
-BINARY_FL(cmpeqf, "(%s == %s) ? (~0) : 0")
-BINARY_FL(cmpltf, "(%s < %s) ? (~0) : 0")
-BINARY_FL(cmplef, "(%s <= %s) ? (~0) : 0")
+BINARY_F(addf, "ORC_DENORMAL(ORC_DENORMAL(%s) + ORC_DENORMAL(%s))")
+BINARY_F(subf, "ORC_DENORMAL(ORC_DENORMAL(%s) - ORC_DENORMAL(%s))")
+BINARY_F(mulf, "ORC_DENORMAL(ORC_DENORMAL(%s) * ORC_DENORMAL(%s))")
+BINARY_F(divf, "ORC_DENORMAL(ORC_DENORMAL(%s) / ORC_DENORMAL(%s))")
+UNARY_F(sqrtf, "sqrt(ORC_DENORMAL(%s))")
+BINARY_F(maxf, "ORC_MAXF(ORC_DENORMAL(%s),ORC_DENORMAL(%s))")
+BINARY_F(minf, "ORC_MINF(ORC_DENORMAL(%s),ORC_DENORMAL(%s))")
+BINARY_FL(cmpeqf, "(ORC_DENORMAL(%s) == ORC_DENORMAL(%s)) ? (~0) : 0")
+BINARY_FL(cmpltf, "(ORC_DENORMAL(%s) < ORC_DENORMAL(%s)) ? (~0) : 0")
+BINARY_FL(cmplef, "(ORC_DENORMAL(%s) <= ORC_DENORMAL(%s)) ? (~0) : 0")
 UNARY_FL(convfl, "(int)%s")
 UNARY_LF(convlf, "%s")
 
diff --git a/orc/orcemulateopcodes.c b/orc/orcemulateopcodes.c
new file mode 100644
index 0000000..7f4d9da
--- /dev/null
+++ b/orc/orcemulateopcodes.c
@@ -0,0 +1,3302 @@
+
+/* autogenerated by generate-emulation */
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+#include <math.h>
+#include <orc/orc.h>
+
+
+/* begin Orc C target preamble */
+#define ORC_CLAMP(x,a,b) ((x)<(a) ? (a) : ((x)>(b) ? (b) : (x)))
+#define ORC_ABS(a) ((a)<0 ? -(a) : (a))
+#define ORC_MIN(a,b) ((a)<(b) ? (a) : (b))
+#define ORC_MAX(a,b) ((a)>(b) ? (a) : (b))
+#define ORC_SB_MAX 127
+#define ORC_SB_MIN (-1-ORC_SB_MAX)
+#define ORC_UB_MAX 255
+#define ORC_UB_MIN 0
+#define ORC_SW_MAX 32767
+#define ORC_SW_MIN (-1-ORC_SW_MAX)
+#define ORC_UW_MAX 65535
+#define ORC_UW_MIN 0
+#define ORC_SL_MAX 2147483647
+#define ORC_SL_MIN (-1-ORC_SL_MAX)
+#define ORC_UL_MAX 4294967295U
+#define ORC_UL_MIN 0
+#define ORC_CLAMP_SB(x) ORC_CLAMP(x,ORC_SB_MIN,ORC_SB_MAX)
+#define ORC_CLAMP_UB(x) ORC_CLAMP(x,ORC_UB_MIN,ORC_UB_MAX)
+#define ORC_CLAMP_SW(x) ORC_CLAMP(x,ORC_SW_MIN,ORC_SW_MAX)
+#define ORC_CLAMP_UW(x) ORC_CLAMP(x,ORC_UW_MIN,ORC_UW_MAX)
+#define ORC_CLAMP_SL(x) ORC_CLAMP(x,ORC_SL_MIN,ORC_SL_MAX)
+#define ORC_CLAMP_UL(x) ORC_CLAMP(x,ORC_UL_MIN,ORC_UL_MAX)
+#define ORC_SWAP_W(x) ((((x)&0xff)<<8) | (((x)&0xff00)>>8))
+#define ORC_SWAP_L(x) ((((x)&0xff)<<24) | (((x)&0xff00)<<8) | (((x)&0xff0000)>>8) | (((x)&0xff000000)>>24))
+#define ORC_PTR_OFFSET(ptr,offset) ((void *)(((unsigned char *)(ptr)) + (offset)))
+#define ORC_MIN_NORMAL (1.1754944909521339405e-38)
+#define ORC_DENORMAL(x) (((x) > -ORC_MIN_NORMAL && (x) < ORC_MIN_NORMAL) ? ((x)<0 ? (-0.0f) : (0.0f)) : (x))
+#define ORC_MINF(a,b) (isnan(a) ? a : isnan(b) ? b : ((a)<(b)) ? (a) : (b))
+#define ORC_MAXF(a,b) (isnan(a) ? a : isnan(b) ? b : ((a)>(b)) ? (a) : (b))
+/* end Orc C target preamble */
+
+
+void
+emulate_absb (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int8 var0;
+  orc_int8 * ptr0;
+  orc_int8 var4;
+  const orc_int8 * ptr4;
+
+  ptr0 = (orc_int8 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int8 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: absb */
+    var0 = ORC_ABS(var4);
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_addb (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int8 var0;
+  orc_int8 * ptr0;
+  orc_int8 var4;
+  const orc_int8 * ptr4;
+  orc_int8 var5;
+  const orc_int8 * ptr5;
+
+  ptr0 = (orc_int8 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int8 *)ex->src_ptrs[0];
+  ptr5 = (orc_int8 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: addb */
+    var0 = var4 + var5;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_addssb (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int8 var0;
+  orc_int8 * ptr0;
+  orc_int8 var4;
+  const orc_int8 * ptr4;
+  orc_int8 var5;
+  const orc_int8 * ptr5;
+
+  ptr0 = (orc_int8 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int8 *)ex->src_ptrs[0];
+  ptr5 = (orc_int8 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: addssb */
+    var0 = ORC_CLAMP_SB(var4 + var5);
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_addusb (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int8 var0;
+  orc_int8 * ptr0;
+  orc_int8 var4;
+  const orc_int8 * ptr4;
+  orc_int8 var5;
+  const orc_int8 * ptr5;
+
+  ptr0 = (orc_int8 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int8 *)ex->src_ptrs[0];
+  ptr5 = (orc_int8 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: addusb */
+    var0 = ORC_CLAMP_UB((orc_uint8)var4 + (orc_uint8)var5);
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_andb (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int8 var0;
+  orc_int8 * ptr0;
+  orc_int8 var4;
+  const orc_int8 * ptr4;
+  orc_int8 var5;
+  const orc_int8 * ptr5;
+
+  ptr0 = (orc_int8 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int8 *)ex->src_ptrs[0];
+  ptr5 = (orc_int8 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: andb */
+    var0 = var4 & var5;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_andnb (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int8 var0;
+  orc_int8 * ptr0;
+  orc_int8 var4;
+  const orc_int8 * ptr4;
+  orc_int8 var5;
+  const orc_int8 * ptr5;
+
+  ptr0 = (orc_int8 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int8 *)ex->src_ptrs[0];
+  ptr5 = (orc_int8 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: andnb */
+    var0 = (~var4) & var5;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_avgsb (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int8 var0;
+  orc_int8 * ptr0;
+  orc_int8 var4;
+  const orc_int8 * ptr4;
+  orc_int8 var5;
+  const orc_int8 * ptr5;
+
+  ptr0 = (orc_int8 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int8 *)ex->src_ptrs[0];
+  ptr5 = (orc_int8 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: avgsb */
+    var0 = (var4 + var5 + 1)>>1;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_avgub (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int8 var0;
+  orc_int8 * ptr0;
+  orc_int8 var4;
+  const orc_int8 * ptr4;
+  orc_int8 var5;
+  const orc_int8 * ptr5;
+
+  ptr0 = (orc_int8 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int8 *)ex->src_ptrs[0];
+  ptr5 = (orc_int8 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: avgub */
+    var0 = ((orc_uint8)var4 + (orc_uint8)var5 + 1)>>1;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_cmpeqb (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int8 var0;
+  orc_int8 * ptr0;
+  orc_int8 var4;
+  const orc_int8 * ptr4;
+  orc_int8 var5;
+  const orc_int8 * ptr5;
+
+  ptr0 = (orc_int8 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int8 *)ex->src_ptrs[0];
+  ptr5 = (orc_int8 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: cmpeqb */
+    var0 = (var4 == var5) ? (~0) : 0;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_cmpgtsb (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int8 var0;
+  orc_int8 * ptr0;
+  orc_int8 var4;
+  const orc_int8 * ptr4;
+  orc_int8 var5;
+  const orc_int8 * ptr5;
+
+  ptr0 = (orc_int8 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int8 *)ex->src_ptrs[0];
+  ptr5 = (orc_int8 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: cmpgtsb */
+    var0 = (var4 > var5) ? (~0) : 0;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_copyb (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int8 var0;
+  orc_int8 * ptr0;
+  orc_int8 var4;
+  const orc_int8 * ptr4;
+
+  ptr0 = (orc_int8 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int8 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: copyb */
+    var0 = var4;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_maxsb (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int8 var0;
+  orc_int8 * ptr0;
+  orc_int8 var4;
+  const orc_int8 * ptr4;
+  orc_int8 var5;
+  const orc_int8 * ptr5;
+
+  ptr0 = (orc_int8 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int8 *)ex->src_ptrs[0];
+  ptr5 = (orc_int8 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: maxsb */
+    var0 = ORC_MAX(var4, var5);
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_maxub (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int8 var0;
+  orc_int8 * ptr0;
+  orc_int8 var4;
+  const orc_int8 * ptr4;
+  orc_int8 var5;
+  const orc_int8 * ptr5;
+
+  ptr0 = (orc_int8 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int8 *)ex->src_ptrs[0];
+  ptr5 = (orc_int8 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: maxub */
+    var0 = ORC_MAX((orc_uint8)var4, (orc_uint8)var5);
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_minsb (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int8 var0;
+  orc_int8 * ptr0;
+  orc_int8 var4;
+  const orc_int8 * ptr4;
+  orc_int8 var5;
+  const orc_int8 * ptr5;
+
+  ptr0 = (orc_int8 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int8 *)ex->src_ptrs[0];
+  ptr5 = (orc_int8 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: minsb */
+    var0 = ORC_MIN(var4, var5);
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_minub (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int8 var0;
+  orc_int8 * ptr0;
+  orc_int8 var4;
+  const orc_int8 * ptr4;
+  orc_int8 var5;
+  const orc_int8 * ptr5;
+
+  ptr0 = (orc_int8 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int8 *)ex->src_ptrs[0];
+  ptr5 = (orc_int8 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: minub */
+    var0 = ORC_MIN((orc_uint8)var4, (orc_uint8)var5);
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_mullb (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int8 var0;
+  orc_int8 * ptr0;
+  orc_int8 var4;
+  const orc_int8 * ptr4;
+  orc_int8 var5;
+  const orc_int8 * ptr5;
+
+  ptr0 = (orc_int8 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int8 *)ex->src_ptrs[0];
+  ptr5 = (orc_int8 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: mullb */
+    var0 = (var4 * var5) & 0xff;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_mulhsb (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int8 var0;
+  orc_int8 * ptr0;
+  orc_int8 var4;
+  const orc_int8 * ptr4;
+  orc_int8 var5;
+  const orc_int8 * ptr5;
+
+  ptr0 = (orc_int8 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int8 *)ex->src_ptrs[0];
+  ptr5 = (orc_int8 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: mulhsb */
+    var0 = (var4 * var5) >> 8;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_mulhub (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int8 var0;
+  orc_int8 * ptr0;
+  orc_int8 var4;
+  const orc_int8 * ptr4;
+  orc_int8 var5;
+  const orc_int8 * ptr5;
+
+  ptr0 = (orc_int8 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int8 *)ex->src_ptrs[0];
+  ptr5 = (orc_int8 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: mulhub */
+    var0 = ((orc_uint32)(orc_uint8)var4 * (orc_uint32)(orc_uint8)var5) >> 8;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_orb (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int8 var0;
+  orc_int8 * ptr0;
+  orc_int8 var4;
+  const orc_int8 * ptr4;
+  orc_int8 var5;
+  const orc_int8 * ptr5;
+
+  ptr0 = (orc_int8 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int8 *)ex->src_ptrs[0];
+  ptr5 = (orc_int8 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: orb */
+    var0 = var4 | var5;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_shlb (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int8 var0;
+  orc_int8 * ptr0;
+  orc_int8 var4;
+  const orc_int8 * ptr4;
+  const int var24 = ((orc_union32 *)(ex->src_ptrs[1]))->i;
+
+  ptr0 = (orc_int8 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int8 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: shlb */
+    var0 = var4 << var24;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_shrsb (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int8 var0;
+  orc_int8 * ptr0;
+  orc_int8 var4;
+  const orc_int8 * ptr4;
+  const int var24 = ((orc_union32 *)(ex->src_ptrs[1]))->i;
+
+  ptr0 = (orc_int8 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int8 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: shrsb */
+    var0 = var4 >> var24;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_shrub (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int8 var0;
+  orc_int8 * ptr0;
+  orc_int8 var4;
+  const orc_int8 * ptr4;
+  const int var24 = ((orc_union32 *)(ex->src_ptrs[1]))->i;
+
+  ptr0 = (orc_int8 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int8 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: shrub */
+    var0 = ((orc_uint8)var4) >> var24;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_signb (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int8 var0;
+  orc_int8 * ptr0;
+  orc_int8 var4;
+  const orc_int8 * ptr4;
+
+  ptr0 = (orc_int8 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int8 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: signb */
+    var0 = ORC_CLAMP(var4,-1,1);
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_subb (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int8 var0;
+  orc_int8 * ptr0;
+  orc_int8 var4;
+  const orc_int8 * ptr4;
+  orc_int8 var5;
+  const orc_int8 * ptr5;
+
+  ptr0 = (orc_int8 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int8 *)ex->src_ptrs[0];
+  ptr5 = (orc_int8 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: subb */
+    var0 = var4 - var5;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_subssb (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int8 var0;
+  orc_int8 * ptr0;
+  orc_int8 var4;
+  const orc_int8 * ptr4;
+  orc_int8 var5;
+  const orc_int8 * ptr5;
+
+  ptr0 = (orc_int8 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int8 *)ex->src_ptrs[0];
+  ptr5 = (orc_int8 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: subssb */
+    var0 = ORC_CLAMP_SB(var4 - var5);
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_subusb (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int8 var0;
+  orc_int8 * ptr0;
+  orc_int8 var4;
+  const orc_int8 * ptr4;
+  orc_int8 var5;
+  const orc_int8 * ptr5;
+
+  ptr0 = (orc_int8 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int8 *)ex->src_ptrs[0];
+  ptr5 = (orc_int8 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: subusb */
+    var0 = ORC_CLAMP_UB((orc_uint8)var4 - (orc_uint8)var5);
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_xorb (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int8 var0;
+  orc_int8 * ptr0;
+  orc_int8 var4;
+  const orc_int8 * ptr4;
+  orc_int8 var5;
+  const orc_int8 * ptr5;
+
+  ptr0 = (orc_int8 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int8 *)ex->src_ptrs[0];
+  ptr5 = (orc_int8 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: xorb */
+    var0 = var4 ^ var5;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_absw (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int16 var0;
+  orc_int16 * ptr0;
+  orc_int16 var4;
+  const orc_int16 * ptr4;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: absw */
+    var0 = ORC_ABS(var4);
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_addw (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int16 var0;
+  orc_int16 * ptr0;
+  orc_int16 var4;
+  const orc_int16 * ptr4;
+  orc_int16 var5;
+  const orc_int16 * ptr5;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+  ptr5 = (orc_int16 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: addw */
+    var0 = var4 + var5;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_addssw (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int16 var0;
+  orc_int16 * ptr0;
+  orc_int16 var4;
+  const orc_int16 * ptr4;
+  orc_int16 var5;
+  const orc_int16 * ptr5;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+  ptr5 = (orc_int16 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: addssw */
+    var0 = ORC_CLAMP_SW(var4 + var5);
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_addusw (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int16 var0;
+  orc_int16 * ptr0;
+  orc_int16 var4;
+  const orc_int16 * ptr4;
+  orc_int16 var5;
+  const orc_int16 * ptr5;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+  ptr5 = (orc_int16 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: addusw */
+    var0 = ORC_CLAMP_UW((orc_uint16)var4 + (orc_uint16)var5);
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_andw (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int16 var0;
+  orc_int16 * ptr0;
+  orc_int16 var4;
+  const orc_int16 * ptr4;
+  orc_int16 var5;
+  const orc_int16 * ptr5;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+  ptr5 = (orc_int16 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: andw */
+    var0 = var4 & var5;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_andnw (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int16 var0;
+  orc_int16 * ptr0;
+  orc_int16 var4;
+  const orc_int16 * ptr4;
+  orc_int16 var5;
+  const orc_int16 * ptr5;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+  ptr5 = (orc_int16 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: andnw */
+    var0 = (~var4) & var5;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_avgsw (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int16 var0;
+  orc_int16 * ptr0;
+  orc_int16 var4;
+  const orc_int16 * ptr4;
+  orc_int16 var5;
+  const orc_int16 * ptr5;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+  ptr5 = (orc_int16 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: avgsw */
+    var0 = (var4 + var5 + 1)>>1;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_avguw (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int16 var0;
+  orc_int16 * ptr0;
+  orc_int16 var4;
+  const orc_int16 * ptr4;
+  orc_int16 var5;
+  const orc_int16 * ptr5;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+  ptr5 = (orc_int16 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: avguw */
+    var0 = ((orc_uint16)var4 + (orc_uint16)var5 + 1)>>1;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_cmpeqw (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int16 var0;
+  orc_int16 * ptr0;
+  orc_int16 var4;
+  const orc_int16 * ptr4;
+  orc_int16 var5;
+  const orc_int16 * ptr5;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+  ptr5 = (orc_int16 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: cmpeqw */
+    var0 = (var4 == var5) ? (~0) : 0;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_cmpgtsw (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int16 var0;
+  orc_int16 * ptr0;
+  orc_int16 var4;
+  const orc_int16 * ptr4;
+  orc_int16 var5;
+  const orc_int16 * ptr5;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+  ptr5 = (orc_int16 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: cmpgtsw */
+    var0 = (var4 > var5) ? (~0) : 0;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_copyw (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int16 var0;
+  orc_int16 * ptr0;
+  orc_int16 var4;
+  const orc_int16 * ptr4;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: copyw */
+    var0 = var4;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_maxsw (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int16 var0;
+  orc_int16 * ptr0;
+  orc_int16 var4;
+  const orc_int16 * ptr4;
+  orc_int16 var5;
+  const orc_int16 * ptr5;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+  ptr5 = (orc_int16 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: maxsw */
+    var0 = ORC_MAX(var4, var5);
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_maxuw (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int16 var0;
+  orc_int16 * ptr0;
+  orc_int16 var4;
+  const orc_int16 * ptr4;
+  orc_int16 var5;
+  const orc_int16 * ptr5;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+  ptr5 = (orc_int16 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: maxuw */
+    var0 = ORC_MAX((orc_uint16)var4, (orc_uint16)var5);
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_minsw (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int16 var0;
+  orc_int16 * ptr0;
+  orc_int16 var4;
+  const orc_int16 * ptr4;
+  orc_int16 var5;
+  const orc_int16 * ptr5;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+  ptr5 = (orc_int16 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: minsw */
+    var0 = ORC_MIN(var4, var5);
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_minuw (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int16 var0;
+  orc_int16 * ptr0;
+  orc_int16 var4;
+  const orc_int16 * ptr4;
+  orc_int16 var5;
+  const orc_int16 * ptr5;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+  ptr5 = (orc_int16 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: minuw */
+    var0 = ORC_MIN((orc_uint16)var4, (orc_uint16)var5);
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_mullw (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int16 var0;
+  orc_int16 * ptr0;
+  orc_int16 var4;
+  const orc_int16 * ptr4;
+  orc_int16 var5;
+  const orc_int16 * ptr5;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+  ptr5 = (orc_int16 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: mullw */
+    var0 = (var4 * var5) & 0xffff;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_mulhsw (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int16 var0;
+  orc_int16 * ptr0;
+  orc_int16 var4;
+  const orc_int16 * ptr4;
+  orc_int16 var5;
+  const orc_int16 * ptr5;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+  ptr5 = (orc_int16 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: mulhsw */
+    var0 = (var4 * var5) >> 16;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_mulhuw (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int16 var0;
+  orc_int16 * ptr0;
+  orc_int16 var4;
+  const orc_int16 * ptr4;
+  orc_int16 var5;
+  const orc_int16 * ptr5;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+  ptr5 = (orc_int16 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: mulhuw */
+    var0 = ((orc_uint32)((orc_uint16)var4) * (orc_uint32)((orc_uint16)var5)) >> 16;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_orw (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int16 var0;
+  orc_int16 * ptr0;
+  orc_int16 var4;
+  const orc_int16 * ptr4;
+  orc_int16 var5;
+  const orc_int16 * ptr5;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+  ptr5 = (orc_int16 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: orw */
+    var0 = var4 | var5;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_shlw (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int16 var0;
+  orc_int16 * ptr0;
+  orc_int16 var4;
+  const orc_int16 * ptr4;
+  const int var24 = ((orc_union32 *)(ex->src_ptrs[1]))->i;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: shlw */
+    var0 = var4 << var24;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_shrsw (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int16 var0;
+  orc_int16 * ptr0;
+  orc_int16 var4;
+  const orc_int16 * ptr4;
+  const int var24 = ((orc_union32 *)(ex->src_ptrs[1]))->i;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: shrsw */
+    var0 = var4 >> var24;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_shruw (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int16 var0;
+  orc_int16 * ptr0;
+  orc_int16 var4;
+  const orc_int16 * ptr4;
+  const int var24 = ((orc_union32 *)(ex->src_ptrs[1]))->i;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: shruw */
+    var0 = ((orc_uint16)var4) >> var24;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_signw (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int16 var0;
+  orc_int16 * ptr0;
+  orc_int16 var4;
+  const orc_int16 * ptr4;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: signw */
+    var0 = ORC_CLAMP(var4,-1,1);
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_subw (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int16 var0;
+  orc_int16 * ptr0;
+  orc_int16 var4;
+  const orc_int16 * ptr4;
+  orc_int16 var5;
+  const orc_int16 * ptr5;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+  ptr5 = (orc_int16 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: subw */
+    var0 = var4 - var5;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_subssw (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int16 var0;
+  orc_int16 * ptr0;
+  orc_int16 var4;
+  const orc_int16 * ptr4;
+  orc_int16 var5;
+  const orc_int16 * ptr5;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+  ptr5 = (orc_int16 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: subssw */
+    var0 = ORC_CLAMP_SW(var4 - var5);
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_subusw (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int16 var0;
+  orc_int16 * ptr0;
+  orc_int16 var4;
+  const orc_int16 * ptr4;
+  orc_int16 var5;
+  const orc_int16 * ptr5;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+  ptr5 = (orc_int16 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: subusw */
+    var0 = ORC_CLAMP_UW((orc_uint16)var4 - (orc_uint16)var5);
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_xorw (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int16 var0;
+  orc_int16 * ptr0;
+  orc_int16 var4;
+  const orc_int16 * ptr4;
+  orc_int16 var5;
+  const orc_int16 * ptr5;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+  ptr5 = (orc_int16 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: xorw */
+    var0 = var4 ^ var5;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_absl (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: absl */
+    var0.i = ORC_ABS(var4.i);
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_addl (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+  orc_union32 var5;
+  const orc_union32 * ptr5;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+  ptr5 = (orc_union32 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: addl */
+    var0.i = var4.i + var5.i;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_addssl (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+  orc_union32 var5;
+  const orc_union32 * ptr5;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+  ptr5 = (orc_union32 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: addssl */
+    var0.i = ORC_CLAMP_SL((orc_int64)var4.i + (orc_int64)var5.i);
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_addusl (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+  orc_union32 var5;
+  const orc_union32 * ptr5;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+  ptr5 = (orc_union32 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: addusl */
+    var0.i = ORC_CLAMP_UL((orc_int64)(orc_uint32)var4.i + (orc_int64)(orc_uint32)var5.i);
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_andl (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+  orc_union32 var5;
+  const orc_union32 * ptr5;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+  ptr5 = (orc_union32 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: andl */
+    var0.i = var4.i & var5.i;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_andnl (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+  orc_union32 var5;
+  const orc_union32 * ptr5;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+  ptr5 = (orc_union32 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: andnl */
+    var0.i = (~var4.i) & var5.i;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_avgsl (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+  orc_union32 var5;
+  const orc_union32 * ptr5;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+  ptr5 = (orc_union32 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: avgsl */
+    var0.i = ((orc_int64)var4.i + (orc_int64)var5.i + 1)>>1;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_avgul (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+  orc_union32 var5;
+  const orc_union32 * ptr5;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+  ptr5 = (orc_union32 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: avgul */
+    var0.i = ((orc_uint64)(orc_uint32)var4.i + (orc_uint64)(orc_uint32)var5.i + 1)>>1;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_cmpeql (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+  orc_union32 var5;
+  const orc_union32 * ptr5;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+  ptr5 = (orc_union32 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: cmpeql */
+    var0.i = (var4.i == var5.i) ? (~0) : 0;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_cmpgtsl (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+  orc_union32 var5;
+  const orc_union32 * ptr5;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+  ptr5 = (orc_union32 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: cmpgtsl */
+    var0.i = (var4.i > var5.i) ? (~0) : 0;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_copyl (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: copyl */
+    var0.i = var4.i;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_maxsl (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+  orc_union32 var5;
+  const orc_union32 * ptr5;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+  ptr5 = (orc_union32 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: maxsl */
+    var0.i = ORC_MAX(var4.i, var5.i);
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_maxul (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+  orc_union32 var5;
+  const orc_union32 * ptr5;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+  ptr5 = (orc_union32 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: maxul */
+    var0.i = ORC_MAX((orc_uint32)var4.i, (orc_uint32)var5.i);
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_minsl (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+  orc_union32 var5;
+  const orc_union32 * ptr5;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+  ptr5 = (orc_union32 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: minsl */
+    var0.i = ORC_MIN(var4.i, var5.i);
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_minul (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+  orc_union32 var5;
+  const orc_union32 * ptr5;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+  ptr5 = (orc_union32 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: minul */
+    var0.i = ORC_MIN((orc_uint32)var4.i, (orc_uint32)var5.i);
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_mulll (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+  orc_union32 var5;
+  const orc_union32 * ptr5;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+  ptr5 = (orc_union32 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: mulll */
+    var0.i = (var4.i * var5.i) & 0xffffffff;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_mulhsl (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+  orc_union32 var5;
+  const orc_union32 * ptr5;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+  ptr5 = (orc_union32 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: mulhsl */
+    var0.i = ((orc_int64)var4.i * (orc_int64)var5.i) >> 32;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_mulhul (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+  orc_union32 var5;
+  const orc_union32 * ptr5;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+  ptr5 = (orc_union32 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: mulhul */
+    var0.i = ((orc_uint64)var4.i * (orc_uint64)var5.i) >> 32;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_orl (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+  orc_union32 var5;
+  const orc_union32 * ptr5;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+  ptr5 = (orc_union32 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: orl */
+    var0.i = var4.i | var5.i;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_shll (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+  const int var24 = ((orc_union32 *)(ex->src_ptrs[1]))->i;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: shll */
+    var0.i = var4.i << var24;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_shrsl (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+  const int var24 = ((orc_union32 *)(ex->src_ptrs[1]))->i;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: shrsl */
+    var0.i = var4.i >> var24;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_shrul (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+  const int var24 = ((orc_union32 *)(ex->src_ptrs[1]))->i;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: shrul */
+    var0.i = ((orc_uint32)var4.i) >> var24;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_signl (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: signl */
+    var0.i = ORC_CLAMP(var4.i,-1,1);
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_subl (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+  orc_union32 var5;
+  const orc_union32 * ptr5;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+  ptr5 = (orc_union32 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: subl */
+    var0.i = var4.i - var5.i;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_subssl (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+  orc_union32 var5;
+  const orc_union32 * ptr5;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+  ptr5 = (orc_union32 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: subssl */
+    var0.i = ORC_CLAMP_SL((orc_int64)var4.i - (orc_int64)var5.i);
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_subusl (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+  orc_union32 var5;
+  const orc_union32 * ptr5;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+  ptr5 = (orc_union32 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: subusl */
+    var0.i = ORC_CLAMP_UL((orc_int64)(orc_uint32)var4.i - (orc_int64)(orc_uint32)var5.i);
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_xorl (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+  orc_union32 var5;
+  const orc_union32 * ptr5;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+  ptr5 = (orc_union32 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: xorl */
+    var0.i = var4.i ^ var5.i;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_convsbw (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int16 var0;
+  orc_int16 * ptr0;
+  orc_int8 var4;
+  const orc_int8 * ptr4;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int8 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: convsbw */
+    var0 = var4;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_convubw (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int16 var0;
+  orc_int16 * ptr0;
+  orc_int8 var4;
+  const orc_int8 * ptr4;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int8 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: convubw */
+    var0 = (orc_uint8)var4;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_convswl (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_int16 var4;
+  const orc_int16 * ptr4;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: convswl */
+    var0.i = var4;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_convuwl (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_int16 var4;
+  const orc_int16 * ptr4;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: convuwl */
+    var0.i = (orc_uint16)var4;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_convwb (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int8 var0;
+  orc_int8 * ptr0;
+  orc_int16 var4;
+  const orc_int16 * ptr4;
+
+  ptr0 = (orc_int8 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: convwb */
+    var0 = var4;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_convssswb (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int8 var0;
+  orc_int8 * ptr0;
+  orc_int16 var4;
+  const orc_int16 * ptr4;
+
+  ptr0 = (orc_int8 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: convssswb */
+    var0 = ORC_CLAMP_SB(var4);
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_convsuswb (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int8 var0;
+  orc_int8 * ptr0;
+  orc_int16 var4;
+  const orc_int16 * ptr4;
+
+  ptr0 = (orc_int8 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: convsuswb */
+    var0 = ORC_CLAMP_UB(var4);
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_convusswb (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int8 var0;
+  orc_int8 * ptr0;
+  orc_int16 var4;
+  const orc_int16 * ptr4;
+
+  ptr0 = (orc_int8 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: convusswb */
+    var0 = ORC_CLAMP_SB((orc_uint16)var4);
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_convuuswb (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int8 var0;
+  orc_int8 * ptr0;
+  orc_int16 var4;
+  const orc_int16 * ptr4;
+
+  ptr0 = (orc_int8 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: convuuswb */
+    var0 = ORC_CLAMP_UB((orc_uint16)var4);
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_convlw (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int16 var0;
+  orc_int16 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: convlw */
+    var0 = var4.i;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_convssslw (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int16 var0;
+  orc_int16 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: convssslw */
+    var0 = ORC_CLAMP_SW(var4.i);
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_convsuslw (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int16 var0;
+  orc_int16 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: convsuslw */
+    var0 = ORC_CLAMP_UW(var4.i);
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_convusslw (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int16 var0;
+  orc_int16 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: convusslw */
+    var0 = ORC_CLAMP_SW((orc_uint32)var4.i);
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_convuuslw (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int16 var0;
+  orc_int16 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: convuuslw */
+    var0 = ORC_CLAMP_UW((orc_uint32)var4.i);
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_mulsbw (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int16 var0;
+  orc_int16 * ptr0;
+  orc_int8 var4;
+  const orc_int8 * ptr4;
+  orc_int8 var5;
+  const orc_int8 * ptr5;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int8 *)ex->src_ptrs[0];
+  ptr5 = (orc_int8 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: mulsbw */
+    var0 = var4 * var5;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_mulubw (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int16 var0;
+  orc_int16 * ptr0;
+  orc_int8 var4;
+  const orc_int8 * ptr4;
+  orc_int8 var5;
+  const orc_int8 * ptr5;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int8 *)ex->src_ptrs[0];
+  ptr5 = (orc_int8 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: mulubw */
+    var0 = (orc_uint8)var4 * (orc_uint8)var5;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_mulswl (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_int16 var4;
+  const orc_int16 * ptr4;
+  orc_int16 var5;
+  const orc_int16 * ptr5;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+  ptr5 = (orc_int16 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: mulswl */
+    var0.i = var4 * var5;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_muluwl (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_int16 var4;
+  const orc_int16 * ptr4;
+  orc_int16 var5;
+  const orc_int16 * ptr5;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+  ptr5 = (orc_int16 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: muluwl */
+    var0.i = (orc_uint16)var4 * (orc_uint16)var5;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_accw (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int16 var4;
+  const orc_int16 * ptr4;
+  orc_int16 var12 = 0;
+
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: accw */
+    var12 = var12 + var4;
+  }
+  ((orc_union32 *)ex->dest_ptrs[0])->i = (var12 + ((orc_union32 *)ex->dest_ptrs[0])->i) & 0xffff;
+
+}
+
+void
+emulate_accl (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+  orc_union32 var12 =  { 0 };
+
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: accl */
+    var12.i = var12.i + var4.i;
+  }
+  ((orc_union32 *)ex->dest_ptrs[0])->i += var12.i;
+
+}
+
+void
+emulate_accsadubl (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int8 var4;
+  const orc_int8 * ptr4;
+  orc_int8 var5;
+  const orc_int8 * ptr5;
+  orc_union32 var12 =  { 0 };
+
+  ptr4 = (orc_int8 *)ex->src_ptrs[0];
+  ptr5 = (orc_int8 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: accsadubl */
+    var12.i = var12.i + ORC_ABS((orc_int32)(orc_uint8)var4 - (orc_int32)(orc_uint8)var5);
+  }
+  ((orc_union32 *)ex->dest_ptrs[0])->i += var12.i;
+
+}
+
+void
+emulate_swapw (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int16 var0;
+  orc_int16 * ptr0;
+  orc_int16 var4;
+  const orc_int16 * ptr4;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: swapw */
+    var0 = ORC_SWAP_W(var4);
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_swapl (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: swapl */
+    var0.i = ORC_SWAP_L(var4.i);
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_select0wb (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int8 var0;
+  orc_int8 * ptr0;
+  orc_int16 var4;
+  const orc_int16 * ptr4;
+
+  ptr0 = (orc_int8 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: select0wb */
+    var0 = (orc_uint16)var4 & 0xff;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_select1wb (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int8 var0;
+  orc_int8 * ptr0;
+  orc_int16 var4;
+  const orc_int16 * ptr4;
+
+  ptr0 = (orc_int8 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: select1wb */
+    var0 = ((orc_uint16)var4 >> 8)&0xff;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_select0lw (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int16 var0;
+  orc_int16 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: select0lw */
+    var0 = (orc_uint32)var4.i & 0xffff;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_select1lw (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int16 var0;
+  orc_int16 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: select1lw */
+    var0 = ((orc_uint32)var4.i >> 16)&0xffff;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_mergewl (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_int16 var4;
+  const orc_int16 * ptr4;
+  orc_int16 var5;
+  const orc_int16 * ptr5;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+  ptr5 = (orc_int16 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: mergewl */
+    var0.i = ((orc_uint16)var4) | ((orc_uint16)var5 << 16);
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_mergebw (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int16 var0;
+  orc_int16 * ptr0;
+  orc_int8 var4;
+  const orc_int8 * ptr4;
+  orc_int8 var5;
+  const orc_int8 * ptr5;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int8 *)ex->src_ptrs[0];
+  ptr5 = (orc_int8 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: mergebw */
+    var0 = ((orc_uint8)var4) | ((orc_uint8)var5 << 8);
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_splitlw (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int16 var0;
+  orc_int16 * ptr0;
+  orc_int16 var1;
+  orc_int16 * ptr1;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr1 = (orc_int16 *)ex->dest_ptrs[1];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: splitlw */
+    var0 = (var4.i >> 16) & 0xffff;
+    var1 = var4.i & 0xffff;
+    *ptr0 = var0;
+    ptr0++;
+    *ptr1 = var1;
+    ptr1++;
+  }
+
+}
+
+void
+emulate_splitwb (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_int8 var0;
+  orc_int8 * ptr0;
+  orc_int8 var1;
+  orc_int8 * ptr1;
+  orc_int16 var4;
+  const orc_int16 * ptr4;
+
+  ptr0 = (orc_int8 *)ex->dest_ptrs[0];
+  ptr1 = (orc_int8 *)ex->dest_ptrs[1];
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: splitwb */
+    var0 = (var4 >> 8) & 0xff;
+    var1 = var4 & 0xff;
+    *ptr0 = var0;
+    ptr0++;
+    *ptr1 = var1;
+    ptr1++;
+  }
+
+}
+
+void
+emulate_addf (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+  orc_union32 var5;
+  const orc_union32 * ptr5;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+  ptr5 = (orc_union32 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: addf */
+    var0.f = ORC_DENORMAL(ORC_DENORMAL(var4.f) + ORC_DENORMAL(var5.f));
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_subf (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+  orc_union32 var5;
+  const orc_union32 * ptr5;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+  ptr5 = (orc_union32 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: subf */
+    var0.f = ORC_DENORMAL(ORC_DENORMAL(var4.f) - ORC_DENORMAL(var5.f));
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_mulf (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+  orc_union32 var5;
+  const orc_union32 * ptr5;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+  ptr5 = (orc_union32 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: mulf */
+    var0.f = ORC_DENORMAL(ORC_DENORMAL(var4.f) * ORC_DENORMAL(var5.f));
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_divf (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+  orc_union32 var5;
+  const orc_union32 * ptr5;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+  ptr5 = (orc_union32 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: divf */
+    var0.f = ORC_DENORMAL(ORC_DENORMAL(var4.f) / ORC_DENORMAL(var5.f));
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_sqrtf (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: sqrtf */
+    var0.f = sqrt(ORC_DENORMAL(var4.f));
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_maxf (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+  orc_union32 var5;
+  const orc_union32 * ptr5;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+  ptr5 = (orc_union32 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: maxf */
+    var0.f = ORC_MAXF(ORC_DENORMAL(var4.f),ORC_DENORMAL(var5.f));
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_minf (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+  orc_union32 var5;
+  const orc_union32 * ptr5;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+  ptr5 = (orc_union32 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: minf */
+    var0.f = ORC_MINF(ORC_DENORMAL(var4.f),ORC_DENORMAL(var5.f));
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_cmpeqf (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+  orc_union32 var5;
+  const orc_union32 * ptr5;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+  ptr5 = (orc_union32 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: cmpeqf */
+    var0.i = (ORC_DENORMAL(var4.f) == ORC_DENORMAL(var5.f)) ? (~0) : 0;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_cmpltf (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+  orc_union32 var5;
+  const orc_union32 * ptr5;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+  ptr5 = (orc_union32 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: cmpltf */
+    var0.i = (ORC_DENORMAL(var4.f) < ORC_DENORMAL(var5.f)) ? (~0) : 0;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_cmplef (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+  orc_union32 var5;
+  const orc_union32 * ptr5;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+  ptr5 = (orc_union32 *)ex->src_ptrs[1];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    var5 = *ptr5;
+    ptr5++;
+    /* 0: cmplef */
+    var0.i = (ORC_DENORMAL(var4.f) <= ORC_DENORMAL(var5.f)) ? (~0) : 0;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_convfl (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: convfl */
+    var0.i = (int)var4.f;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
+void
+emulate_convlf (OrcOpcodeExecutor *ex, int n)
+{
+  int i;
+  orc_union32 var0;
+  orc_union32 * ptr0;
+  orc_union32 var4;
+  const orc_union32 * ptr4;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    var4 = *ptr4;
+    ptr4++;
+    /* 0: convlf */
+    var0.f = var4.i;
+    *ptr0 = var0;
+    ptr0++;
+  }
+
+}
+
diff --git a/orc/orcemulateopcodes.h b/orc/orcemulateopcodes.h
new file mode 100644
index 0000000..4d4a827
--- /dev/null
+++ b/orc/orcemulateopcodes.h
@@ -0,0 +1,248 @@
+void emulate_absb (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_absb (OrcOpcodeExecutor *ex);
+void emulate_addb (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_addb (OrcOpcodeExecutor *ex);
+void emulate_addssb (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_addssb (OrcOpcodeExecutor *ex);
+void emulate_addusb (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_addusb (OrcOpcodeExecutor *ex);
+void emulate_andb (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_andb (OrcOpcodeExecutor *ex);
+void emulate_andnb (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_andnb (OrcOpcodeExecutor *ex);
+void emulate_avgsb (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_avgsb (OrcOpcodeExecutor *ex);
+void emulate_avgub (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_avgub (OrcOpcodeExecutor *ex);
+void emulate_cmpeqb (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_cmpeqb (OrcOpcodeExecutor *ex);
+void emulate_cmpgtsb (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_cmpgtsb (OrcOpcodeExecutor *ex);
+void emulate_copyb (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_copyb (OrcOpcodeExecutor *ex);
+void emulate_maxsb (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_maxsb (OrcOpcodeExecutor *ex);
+void emulate_maxub (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_maxub (OrcOpcodeExecutor *ex);
+void emulate_minsb (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_minsb (OrcOpcodeExecutor *ex);
+void emulate_minub (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_minub (OrcOpcodeExecutor *ex);
+void emulate_mullb (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_mullb (OrcOpcodeExecutor *ex);
+void emulate_mulhsb (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_mulhsb (OrcOpcodeExecutor *ex);
+void emulate_mulhub (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_mulhub (OrcOpcodeExecutor *ex);
+void emulate_orb (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_orb (OrcOpcodeExecutor *ex);
+void emulate_shlb (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_shlb (OrcOpcodeExecutor *ex);
+void emulate_shrsb (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_shrsb (OrcOpcodeExecutor *ex);
+void emulate_shrub (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_shrub (OrcOpcodeExecutor *ex);
+void emulate_signb (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_signb (OrcOpcodeExecutor *ex);
+void emulate_subb (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_subb (OrcOpcodeExecutor *ex);
+void emulate_subssb (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_subssb (OrcOpcodeExecutor *ex);
+void emulate_subusb (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_subusb (OrcOpcodeExecutor *ex);
+void emulate_xorb (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_xorb (OrcOpcodeExecutor *ex);
+void emulate_absw (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_absw (OrcOpcodeExecutor *ex);
+void emulate_addw (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_addw (OrcOpcodeExecutor *ex);
+void emulate_addssw (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_addssw (OrcOpcodeExecutor *ex);
+void emulate_addusw (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_addusw (OrcOpcodeExecutor *ex);
+void emulate_andw (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_andw (OrcOpcodeExecutor *ex);
+void emulate_andnw (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_andnw (OrcOpcodeExecutor *ex);
+void emulate_avgsw (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_avgsw (OrcOpcodeExecutor *ex);
+void emulate_avguw (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_avguw (OrcOpcodeExecutor *ex);
+void emulate_cmpeqw (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_cmpeqw (OrcOpcodeExecutor *ex);
+void emulate_cmpgtsw (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_cmpgtsw (OrcOpcodeExecutor *ex);
+void emulate_copyw (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_copyw (OrcOpcodeExecutor *ex);
+void emulate_maxsw (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_maxsw (OrcOpcodeExecutor *ex);
+void emulate_maxuw (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_maxuw (OrcOpcodeExecutor *ex);
+void emulate_minsw (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_minsw (OrcOpcodeExecutor *ex);
+void emulate_minuw (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_minuw (OrcOpcodeExecutor *ex);
+void emulate_mullw (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_mullw (OrcOpcodeExecutor *ex);
+void emulate_mulhsw (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_mulhsw (OrcOpcodeExecutor *ex);
+void emulate_mulhuw (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_mulhuw (OrcOpcodeExecutor *ex);
+void emulate_orw (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_orw (OrcOpcodeExecutor *ex);
+void emulate_shlw (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_shlw (OrcOpcodeExecutor *ex);
+void emulate_shrsw (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_shrsw (OrcOpcodeExecutor *ex);
+void emulate_shruw (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_shruw (OrcOpcodeExecutor *ex);
+void emulate_signw (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_signw (OrcOpcodeExecutor *ex);
+void emulate_subw (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_subw (OrcOpcodeExecutor *ex);
+void emulate_subssw (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_subssw (OrcOpcodeExecutor *ex);
+void emulate_subusw (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_subusw (OrcOpcodeExecutor *ex);
+void emulate_xorw (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_xorw (OrcOpcodeExecutor *ex);
+void emulate_absl (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_absl (OrcOpcodeExecutor *ex);
+void emulate_addl (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_addl (OrcOpcodeExecutor *ex);
+void emulate_addssl (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_addssl (OrcOpcodeExecutor *ex);
+void emulate_addusl (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_addusl (OrcOpcodeExecutor *ex);
+void emulate_andl (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_andl (OrcOpcodeExecutor *ex);
+void emulate_andnl (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_andnl (OrcOpcodeExecutor *ex);
+void emulate_avgsl (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_avgsl (OrcOpcodeExecutor *ex);
+void emulate_avgul (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_avgul (OrcOpcodeExecutor *ex);
+void emulate_cmpeql (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_cmpeql (OrcOpcodeExecutor *ex);
+void emulate_cmpgtsl (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_cmpgtsl (OrcOpcodeExecutor *ex);
+void emulate_copyl (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_copyl (OrcOpcodeExecutor *ex);
+void emulate_maxsl (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_maxsl (OrcOpcodeExecutor *ex);
+void emulate_maxul (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_maxul (OrcOpcodeExecutor *ex);
+void emulate_minsl (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_minsl (OrcOpcodeExecutor *ex);
+void emulate_minul (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_minul (OrcOpcodeExecutor *ex);
+void emulate_mulll (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_mulll (OrcOpcodeExecutor *ex);
+void emulate_mulhsl (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_mulhsl (OrcOpcodeExecutor *ex);
+void emulate_mulhul (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_mulhul (OrcOpcodeExecutor *ex);
+void emulate_orl (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_orl (OrcOpcodeExecutor *ex);
+void emulate_shll (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_shll (OrcOpcodeExecutor *ex);
+void emulate_shrsl (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_shrsl (OrcOpcodeExecutor *ex);
+void emulate_shrul (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_shrul (OrcOpcodeExecutor *ex);
+void emulate_signl (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_signl (OrcOpcodeExecutor *ex);
+void emulate_subl (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_subl (OrcOpcodeExecutor *ex);
+void emulate_subssl (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_subssl (OrcOpcodeExecutor *ex);
+void emulate_subusl (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_subusl (OrcOpcodeExecutor *ex);
+void emulate_xorl (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_xorl (OrcOpcodeExecutor *ex);
+void emulate_convsbw (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_convsbw (OrcOpcodeExecutor *ex);
+void emulate_convubw (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_convubw (OrcOpcodeExecutor *ex);
+void emulate_convswl (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_convswl (OrcOpcodeExecutor *ex);
+void emulate_convuwl (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_convuwl (OrcOpcodeExecutor *ex);
+void emulate_convwb (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_convwb (OrcOpcodeExecutor *ex);
+void emulate_convssswb (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_convssswb (OrcOpcodeExecutor *ex);
+void emulate_convsuswb (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_convsuswb (OrcOpcodeExecutor *ex);
+void emulate_convusswb (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_convusswb (OrcOpcodeExecutor *ex);
+void emulate_convuuswb (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_convuuswb (OrcOpcodeExecutor *ex);
+void emulate_convlw (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_convlw (OrcOpcodeExecutor *ex);
+void emulate_convssslw (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_convssslw (OrcOpcodeExecutor *ex);
+void emulate_convsuslw (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_convsuslw (OrcOpcodeExecutor *ex);
+void emulate_convusslw (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_convusslw (OrcOpcodeExecutor *ex);
+void emulate_convuuslw (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_convuuslw (OrcOpcodeExecutor *ex);
+void emulate_mulsbw (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_mulsbw (OrcOpcodeExecutor *ex);
+void emulate_mulubw (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_mulubw (OrcOpcodeExecutor *ex);
+void emulate_mulswl (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_mulswl (OrcOpcodeExecutor *ex);
+void emulate_muluwl (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_muluwl (OrcOpcodeExecutor *ex);
+void emulate_accw (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_accw (OrcOpcodeExecutor *ex);
+void emulate_accl (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_accl (OrcOpcodeExecutor *ex);
+void emulate_accsadubl (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_accsadubl (OrcOpcodeExecutor *ex);
+void emulate_swapw (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_swapw (OrcOpcodeExecutor *ex);
+void emulate_swapl (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_swapl (OrcOpcodeExecutor *ex);
+void emulate_select0wb (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_select0wb (OrcOpcodeExecutor *ex);
+void emulate_select1wb (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_select1wb (OrcOpcodeExecutor *ex);
+void emulate_select0lw (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_select0lw (OrcOpcodeExecutor *ex);
+void emulate_select1lw (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_select1lw (OrcOpcodeExecutor *ex);
+void emulate_mergewl (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_mergewl (OrcOpcodeExecutor *ex);
+void emulate_mergebw (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_mergebw (OrcOpcodeExecutor *ex);
+void emulate_splitlw (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_splitlw (OrcOpcodeExecutor *ex);
+void emulate_splitwb (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_splitwb (OrcOpcodeExecutor *ex);
+void emulate_addf (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_addf (OrcOpcodeExecutor *ex);
+void emulate_subf (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_subf (OrcOpcodeExecutor *ex);
+void emulate_mulf (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_mulf (OrcOpcodeExecutor *ex);
+void emulate_divf (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_divf (OrcOpcodeExecutor *ex);
+void emulate_sqrtf (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_sqrtf (OrcOpcodeExecutor *ex);
+void emulate_maxf (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_maxf (OrcOpcodeExecutor *ex);
+void emulate_minf (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_minf (OrcOpcodeExecutor *ex);
+void emulate_cmpeqf (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_cmpeqf (OrcOpcodeExecutor *ex);
+void emulate_cmpltf (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_cmpltf (OrcOpcodeExecutor *ex);
+void emulate_cmplef (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_cmplef (OrcOpcodeExecutor *ex);
+void emulate_convfl (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_convfl (OrcOpcodeExecutor *ex);
+void emulate_convlf (OrcOpcodeExecutor *ex, int n);
+void emulate_n16_convlf (OrcOpcodeExecutor *ex);
diff --git a/orc/orcexecutor.c b/orc/orcexecutor.c
index d37c76c..98c1e14 100644
--- a/orc/orcexecutor.c
+++ b/orc/orcexecutor.c
@@ -119,6 +119,43 @@ orc_executor_set_m (OrcExecutor *ex, int m)
   ORC_EXECUTOR_M(ex) = m;
 }
 
+static void
+load_constant (void *data, int size, int value)
+{
+  switch (size) {
+    case 1:
+      {
+        int l;
+        orc_int8 *d = data;
+        for(l=0;l<16;l++) {
+          d[l] = value;
+        }
+      }
+      break;
+    case 2:
+      {
+        int l;
+        orc_int16 *d = data;
+        for(l=0;l<16;l++) {
+          d[l] = value;
+        }
+      }
+      break;
+    case 4:
+      {
+        int l;
+        orc_int32 *d = data;
+        for(l=0;l<16;l++) {
+          d[l] = value;
+        }
+      }
+      break;
+    default:
+      ORC_ASSERT(0);
+  }
+
+}
+
 void
 orc_executor_emulate (OrcExecutor *ex)
 {
@@ -129,7 +166,8 @@ orc_executor_emulate (OrcExecutor *ex)
   OrcProgram *program = ex->program;
   OrcInstruction *insn;
   OrcStaticOpcode *opcode;
-  OrcOpcodeExecutor opcode_ex;
+  OrcOpcodeExecutor *opcode_ex;
+  void *tmpspace[ORC_VAR_T15+1] = { 0 };
 
   ex->accumulators[0] = 0;
   ex->accumulators[1] = 0;
@@ -145,104 +183,105 @@ orc_executor_emulate (OrcExecutor *ex)
   } else {
     m = 1;
   }
+
+  for(i=0;i<ORC_VAR_T15+1;i++){
+    OrcVariable *var = program->vars + i;
+
+    if (var->size) {
+      tmpspace[i] = malloc(4 * 16);
+    }
+  }
+
+  opcode_ex = malloc(sizeof(OrcOpcodeExecutor)*program->n_insns);
+
+  for(j=0;j<program->n_insns;j++){
+    insn = program->insns + j;
+    opcode = insn->opcode;
+
+    opcode_ex[j].emulateN = opcode->emulateN;
+    for(k=0;k<ORC_STATIC_OPCODE_N_SRC;k++) {
+      OrcVariable *var = program->vars + insn->src_args[k];
+
+      if (var->vartype == ORC_VAR_TYPE_CONST) {
+        opcode_ex[j].src_ptrs[k] = tmpspace[insn->src_args[k]];
+        load_constant (tmpspace[insn->src_args[k]], opcode->src_size[k],
+            var->value);
+      } else if (var->vartype == ORC_VAR_TYPE_PARAM) {
+        opcode_ex[j].src_ptrs[k] = tmpspace[insn->src_args[k]];
+        load_constant (tmpspace[insn->src_args[k]], opcode->src_size[k],
+            ex->params[insn->src_args[k]]);
+      } else if (var->vartype == ORC_VAR_TYPE_TEMP) {
+        opcode_ex[j].src_ptrs[k] = tmpspace[insn->src_args[k]];
+      } else if (var->vartype == ORC_VAR_TYPE_SRC) {
+        opcode_ex[j].src_ptrs[k] = ex->arrays[insn->src_args[k]];
+      } else if (var->vartype == ORC_VAR_TYPE_DEST) {
+        opcode_ex[j].src_ptrs[k] = ex->arrays[insn->src_args[k]];
+      }
+    }
+    for(k=0;k<ORC_STATIC_OPCODE_N_DEST;k++) {
+      OrcVariable *var = program->vars + insn->dest_args[k];
+
+      if (var->vartype == ORC_VAR_TYPE_TEMP) {
+        opcode_ex[j].dest_ptrs[k] = tmpspace[insn->dest_args[k]];
+      } else if (var->vartype == ORC_VAR_TYPE_ACCUMULATOR) {
+        opcode_ex[j].dest_ptrs[k] =
+          &ex->accumulators[insn->dest_args[k] - ORC_VAR_A1];
+      } else if (var->vartype == ORC_VAR_TYPE_DEST) {
+        opcode_ex[j].dest_ptrs[k] = ex->arrays[insn->dest_args[k]];
+      }
+    }
+  }
+  
   ORC_DEBUG("src ptr %p stride %d", ex->arrays[ORC_VAR_S1], ex->params[ORC_VAR_S1]);
   for(m_index=0;m_index<m;m_index++){
     ORC_DEBUG("m_index %d m %d", m_index, m);
-    for(i=0;i<ex->n;i++){
+
+    for(i=0;i<ex->n;i+=16){
       for(j=0;j<program->n_insns;j++){
         insn = program->insns + j;
         opcode = insn->opcode;
 
-        /* set up args */
         for(k=0;k<ORC_STATIC_OPCODE_N_SRC;k++) {
           OrcVariable *var = program->vars + insn->src_args[k];
 
-          if (opcode->src_size[k] == 0) continue;
-
-          if (var->vartype == ORC_VAR_TYPE_CONST) {
-            opcode_ex.src_values[k] = var->value;
-          } else if (var->vartype == ORC_VAR_TYPE_PARAM) {
-            opcode_ex.src_values[k] = ex->params[insn->src_args[k]];
-          } else if (var->vartype == ORC_VAR_TYPE_TEMP) {
-            /* FIXME shouldn't store executor stuff in program */
-            opcode_ex.src_values[k] = var->value;
-          } else if (var->vartype == ORC_VAR_TYPE_SRC ||
-              var->vartype == ORC_VAR_TYPE_DEST) {
-            void *ptr = ORC_PTR_OFFSET(ex->arrays[insn->src_args[k]],
-                var->size*i + ex->params[insn->src_args[k]]*m_index);
-
-            switch (var->size) {
-              case 1:
-                opcode_ex.src_values[k] = *(orc_int8 *)ptr;
-                break;
-              case 2:
-                opcode_ex.src_values[k] = *(orc_int16 *)ptr;
-                break;
-              case 4:
-                opcode_ex.src_values[k] = *(orc_int32 *)ptr;
-                break;
-              case 8:
-                opcode_ex.src_values[k] = *(orc_int64 *)ptr;
-                break;
-              default:
-                ORC_ERROR("unhandled size %d", program->vars[insn->src_args[k]].size);
-            }
-          } else {
-            ORC_ERROR("shouldn't be reached (%d)", var->vartype);
+          if (var->vartype == ORC_VAR_TYPE_SRC) {
+            opcode_ex[j].src_ptrs[k] =
+              ORC_PTR_OFFSET(ex->arrays[insn->src_args[k]],
+                  var->size*i + ex->params[insn->src_args[k]]*m_index);
+          } else if (var->vartype == ORC_VAR_TYPE_DEST) {
+            opcode_ex[j].src_ptrs[k] =
+              ORC_PTR_OFFSET(ex->arrays[insn->src_args[k]],
+                  var->size*i + ex->params[insn->src_args[k]]*m_index);
           }
         }
-
-        opcode->emulate (&opcode_ex, opcode->emulate_user);
-
-        for(k=0;k<ORC_STATIC_OPCODE_N_DEST;k++){
+        for(k=0;k<ORC_STATIC_OPCODE_N_DEST;k++) {
           OrcVariable *var = program->vars + insn->dest_args[k];
 
-          if (opcode->dest_size[k] == 0) continue;
-
-          if (var->vartype == ORC_VAR_TYPE_TEMP) {
-            /* FIXME shouldn't store executor stuff in program */
-            var->value = opcode_ex.dest_values[k];
-          } else if (var->vartype == ORC_VAR_TYPE_DEST) {
-            void *ptr = ORC_PTR_OFFSET(ex->arrays[insn->dest_args[k]],
-                var->size*i + ex->params[insn->dest_args[k]]*m_index);
-
-            switch (var->size) {
-              case 1:
-                *(orc_int8 *)ptr = opcode_ex.dest_values[k];
-                break;
-              case 2:
-                *(orc_int16 *)ptr = opcode_ex.dest_values[k];
-                break;
-              case 4:
-                *(orc_int32 *)ptr = opcode_ex.dest_values[k];
-                break;
-              case 8:
-                *(orc_int64 *)ptr = opcode_ex.dest_values[k];
-                break;
-              default:
-                ORC_ERROR("unhandled size %d", program->vars[insn->dest_args[k]].size);
-            }
-          } else if (var->vartype == ORC_VAR_TYPE_ACCUMULATOR) {
-            switch (var->size) {
-              case 2:
-                ex->accumulators[insn->dest_args[k] - ORC_VAR_A1] +=
-                  opcode_ex.dest_values[k];
-                ex->accumulators[insn->dest_args[k] - ORC_VAR_A1] &= 0xffff;
-                break;
-              case 4:
-                ex->accumulators[insn->dest_args[k] - ORC_VAR_A1] +=
-                  opcode_ex.dest_values[k];
-                break;
-              default:
-                ORC_ERROR("unhandled size %d", program->vars[insn->dest_args[k]].size);
-            }
-          } else {
-            ORC_ERROR("shouldn't be reached (%d)", var->vartype);
+          if (var->vartype == ORC_VAR_TYPE_DEST) {
+            opcode_ex[j].dest_ptrs[k] =
+              ORC_PTR_OFFSET(ex->arrays[insn->dest_args[k]],
+                  var->size*i + ex->params[insn->dest_args[k]]*m_index);
           }
         }
       }
+
+      for(j=0;j<program->n_insns;j++){
+        insn = program->insns + j;
+        opcode = insn->opcode;
+
+        if (ex->n - i >= 16) {
+          opcode_ex[j].emulateN (opcode_ex + j, 16);
+        } else {
+          opcode_ex[j].emulateN (opcode_ex + j, ex->n - i);
+        }
+      }
     }
   }
+
+  free (opcode_ex);
+  for(i=0;i<ORC_VAR_T15+1;i++){
+    if (tmpspace[i]) free (tmpspace[i]);
+  }
 }
 
 
diff --git a/orc/orcopcodes.c b/orc/orcopcodes.c
index ab751ad..ada3d02 100644
--- a/orc/orcopcodes.c
+++ b/orc/orcopcodes.c
@@ -696,161 +696,162 @@ convlf (OrcOpcodeExecutor *ex, void *user)
   ORC_FLOAT_WRITE(&ex->dest_values[0], ex->src_values[0]);
 }
 
+#include "orc/orcemulateopcodes.h"
 
 static OrcStaticOpcode opcodes[] = {
 
   /* byte ops */
-  { "absb", absb, NULL, 0, { 1 }, { 1 } },
-  { "addb", addb, NULL, 0, { 1 }, { 1, 1 } },
-  { "addssb", addssb, NULL, 0, { 1 }, { 1, 1 } },
-  { "addusb", addusb, NULL, 0, { 1 }, { 1, 1 } },
-  { "andb", andb, NULL, 0, { 1 }, { 1, 1 } },
-  { "andnb", andnb, NULL, 0, { 1 }, { 1, 1 } },
-  { "avgsb", avgsb, NULL, 0, { 1 }, { 1, 1 } },
-  { "avgub", avgub, NULL, 0, { 1 }, { 1, 1 } },
-  { "cmpeqb", cmpeqb, NULL, 0, { 1 }, { 1, 1 } },
-  { "cmpgtsb", cmpgtsb, NULL, 0, { 1 }, { 1, 1 } },
-  { "copyb", copyb, NULL, 0, { 1 }, { 1 } },
-  { "maxsb", maxsb, NULL, 0, { 1 }, { 1, 1 } },
-  { "maxub", maxub, NULL, 0, { 1 }, { 1, 1 } },
-  { "minsb", minsb, NULL, 0, { 1 }, { 1, 1 } },
-  { "minub", minub, NULL, 0, { 1 }, { 1, 1 } },
-  { "mullb", mullb, NULL, 0, { 1 }, { 1, 1 } },
-  { "mulhsb", mulhsb, NULL, 0, { 1 }, { 1, 1 } },
-  { "mulhub", mulhub, NULL, 0, { 1 }, { 1, 1 } },
-  { "orb", orb, NULL, 0, { 1 }, { 1, 1 } },
-  { "shlb", shlb, NULL, ORC_STATIC_OPCODE_SCALAR, { 1 }, { 1, 1 } },
-  { "shrsb", shrsb, NULL, ORC_STATIC_OPCODE_SCALAR, { 1 }, { 1, 1 } },
-  { "shrub", shrub, NULL, ORC_STATIC_OPCODE_SCALAR, { 1 }, { 1, 1 } },
-  { "signb", signb, NULL, 0, { 1 }, { 1 } },
-  { "subb", subb, NULL, 0, { 1 }, { 1, 1 } },
-  { "subssb", subssb, NULL, 0, { 1 }, { 1, 1 } },
-  { "subusb", subusb, NULL, 0, { 1 }, { 1, 1 } },
-  { "xorb", xorb, NULL, 0, { 1 }, { 1, 1 } },
+  { "absb", absb, NULL, 0, { 1 }, { 1 }, emulate_absb },
+  { "addb", addb, NULL, 0, { 1 }, { 1, 1 }, emulate_addb },
+  { "addssb", addssb, NULL, 0, { 1 }, { 1, 1 }, emulate_addssb },
+  { "addusb", addusb, NULL, 0, { 1 }, { 1, 1 }, emulate_addusb },
+  { "andb", andb, NULL, 0, { 1 }, { 1, 1 }, emulate_andb },
+  { "andnb", andnb, NULL, 0, { 1 }, { 1, 1 }, emulate_andnb },
+  { "avgsb", avgsb, NULL, 0, { 1 }, { 1, 1 }, emulate_avgsb },
+  { "avgub", avgub, NULL, 0, { 1 }, { 1, 1 }, emulate_avgub },
+  { "cmpeqb", cmpeqb, NULL, 0, { 1 }, { 1, 1 }, emulate_cmpeqb },
+  { "cmpgtsb", cmpgtsb, NULL, 0, { 1 }, { 1, 1 }, emulate_cmpgtsb },
+  { "copyb", copyb, NULL, 0, { 1 }, { 1 }, emulate_copyb },
+  { "maxsb", maxsb, NULL, 0, { 1 }, { 1, 1 }, emulate_maxsb },
+  { "maxub", maxub, NULL, 0, { 1 }, { 1, 1 }, emulate_maxub },
+  { "minsb", minsb, NULL, 0, { 1 }, { 1, 1 }, emulate_minsb },
+  { "minub", minub, NULL, 0, { 1 }, { 1, 1 }, emulate_minub },
+  { "mullb", mullb, NULL, 0, { 1 }, { 1, 1 }, emulate_mullb },
+  { "mulhsb", mulhsb, NULL, 0, { 1 }, { 1, 1 }, emulate_mulhsb },
+  { "mulhub", mulhub, NULL, 0, { 1 }, { 1, 1 }, emulate_mulhub },
+  { "orb", orb, NULL, 0, { 1 }, { 1, 1 }, emulate_orb },
+  { "shlb", shlb, NULL, ORC_STATIC_OPCODE_SCALAR, { 1 }, { 1, 1 }, emulate_shlb },
+  { "shrsb", shrsb, NULL, ORC_STATIC_OPCODE_SCALAR, { 1 }, { 1, 1 }, emulate_shrsb },
+  { "shrub", shrub, NULL, ORC_STATIC_OPCODE_SCALAR, { 1 }, { 1, 1 }, emulate_shrub },
+  { "signb", signb, NULL, 0, { 1 }, { 1 }, emulate_signb },
+  { "subb", subb, NULL, 0, { 1 }, { 1, 1 }, emulate_subb },
+  { "subssb", subssb, NULL, 0, { 1 }, { 1, 1 }, emulate_subssb },
+  { "subusb", subusb, NULL, 0, { 1 }, { 1, 1 }, emulate_subusb },
+  { "xorb", xorb, NULL, 0, { 1 }, { 1, 1 }, emulate_xorb },
 
   /* word ops */
-  { "absw", absw, NULL, 0, { 2 }, { 2 } },
-  { "addw", addw, NULL, 0, { 2 }, { 2, 2 } },
-  { "addssw", addssw, NULL, 0, { 2 }, { 2, 2 } },
-  { "addusw", addusw, NULL, 0, { 2 }, { 2, 2 } },
-  { "andw", andw, NULL, 0, { 2 }, { 2, 2 } },
-  { "andnw", andnw, NULL, 0, { 2 }, { 2, 2 } },
-  { "avgsw", avgsw, NULL, 0, { 2 }, { 2, 2 } },
-  { "avguw", avguw, NULL, 0, { 2 }, { 2, 2 } },
-  { "cmpeqw", cmpeqw, NULL, 0, { 2 }, { 2, 2 } },
-  { "cmpgtsw", cmpgtsw, NULL, 0, { 2 }, { 2, 2 } },
-  { "copyw", copyw, NULL, 0, { 2 }, { 2 } },
-  { "maxsw", maxsw, NULL, 0, { 2 }, { 2, 2 } },
-  { "maxuw", maxuw, NULL, 0, { 2 }, { 2, 2 } },
-  { "minsw", minsw, NULL, 0, { 2 }, { 2, 2 } },
-  { "minuw", minuw, NULL, 0, { 2 }, { 2, 2 } },
-  { "mullw", mullw, NULL, 0, { 2 }, { 2, 2 } },
-  { "mulhsw", mulhsw, NULL, 0, { 2 }, { 2, 2 } },
-  { "mulhuw", mulhuw, NULL, 0, { 2 }, { 2, 2 } },
-  { "orw", orw, NULL, 0, { 2 }, { 2, 2 } },
-  { "shlw", shlw, NULL, ORC_STATIC_OPCODE_SCALAR, { 2 }, { 2, 2 } },
-  { "shrsw", shrsw, NULL, ORC_STATIC_OPCODE_SCALAR, { 2 }, { 2, 2 } },
-  { "shruw", shruw, NULL, ORC_STATIC_OPCODE_SCALAR, { 2 }, { 2, 2 } },
-  { "signw", signw, NULL, 0, { 2 }, { 2 } },
-  { "subw", subw, NULL, 0, { 2 }, { 2, 2 } },
-  { "subssw", subssw, NULL, 0, { 2 }, { 2, 2 } },
-  { "subusw", subusw, NULL, 0, { 2 }, { 2, 2 } },
-  { "xorw", xorw, NULL, 0, { 2 }, { 2, 2 } },
+  { "absw", absw, NULL, 0, { 2 }, { 2 }, emulate_absw },
+  { "addw", addw, NULL, 0, { 2 }, { 2, 2 }, emulate_addw },
+  { "addssw", addssw, NULL, 0, { 2 }, { 2, 2 }, emulate_addssw },
+  { "addusw", addusw, NULL, 0, { 2 }, { 2, 2 }, emulate_addusw },
+  { "andw", andw, NULL, 0, { 2 }, { 2, 2 }, emulate_andw },
+  { "andnw", andnw, NULL, 0, { 2 }, { 2, 2 }, emulate_andnw },
+  { "avgsw", avgsw, NULL, 0, { 2 }, { 2, 2 }, emulate_avgsw },
+  { "avguw", avguw, NULL, 0, { 2 }, { 2, 2 }, emulate_avguw },
+  { "cmpeqw", cmpeqw, NULL, 0, { 2 }, { 2, 2 }, emulate_cmpeqw },
+  { "cmpgtsw", cmpgtsw, NULL, 0, { 2 }, { 2, 2 }, emulate_cmpgtsw },
+  { "copyw", copyw, NULL, 0, { 2 }, { 2 }, emulate_copyw },
+  { "maxsw", maxsw, NULL, 0, { 2 }, { 2, 2 }, emulate_maxsw },
+  { "maxuw", maxuw, NULL, 0, { 2 }, { 2, 2 }, emulate_maxuw },
+  { "minsw", minsw, NULL, 0, { 2 }, { 2, 2 }, emulate_minsw },
+  { "minuw", minuw, NULL, 0, { 2 }, { 2, 2 }, emulate_minuw },
+  { "mullw", mullw, NULL, 0, { 2 }, { 2, 2 }, emulate_mullw },
+  { "mulhsw", mulhsw, NULL, 0, { 2 }, { 2, 2 }, emulate_mulhsw },
+  { "mulhuw", mulhuw, NULL, 0, { 2 }, { 2, 2 }, emulate_mulhuw },
+  { "orw", orw, NULL, 0, { 2 }, { 2, 2 }, emulate_orw },
+  { "shlw", shlw, NULL, ORC_STATIC_OPCODE_SCALAR, { 2 }, { 2, 2 }, emulate_shlw },
+  { "shrsw", shrsw, NULL, ORC_STATIC_OPCODE_SCALAR, { 2 }, { 2, 2 }, emulate_shrsw },
+  { "shruw", shruw, NULL, ORC_STATIC_OPCODE_SCALAR, { 2 }, { 2, 2 }, emulate_shruw },
+  { "signw", signw, NULL, 0, { 2 }, { 2 }, emulate_signw },
+  { "subw", subw, NULL, 0, { 2 }, { 2, 2 }, emulate_subw },
+  { "subssw", subssw, NULL, 0, { 2 }, { 2, 2 }, emulate_subssw },
+  { "subusw", subusw, NULL, 0, { 2 }, { 2, 2 }, emulate_subusw },
+  { "xorw", xorw, NULL, 0, { 2 }, { 2, 2 }, emulate_xorw },
 
   /* long ops */
-  { "absl", absl, NULL, 0, { 4 }, { 4 } },
-  { "addl", addl, NULL, 0, { 4 }, { 4, 4 } },
-  { "addssl", addssl, NULL, 0, { 4 }, { 4, 4 } },
-  { "addusl", addusl, NULL, 0, { 4 }, { 4, 4 } },
-  { "andl", andl, NULL, 0, { 4 }, { 4, 4 } },
-  { "andnl", andnl, NULL, 0, { 4 }, { 4, 4 } },
-  { "avgsl", avgsl, NULL, 0, { 4 }, { 4, 4 } },
-  { "avgul", avgul, NULL, 0, { 4 }, { 4, 4 } },
-  { "cmpeql", cmpeql, NULL, 0, { 4 }, { 4, 4 } },
-  { "cmpgtsl", cmpgtsl, NULL, 0, { 4 }, { 4, 4 } },
-  { "copyl", copyl, NULL, 0, { 4 }, { 4 } },
-  { "maxsl", maxsl, NULL, 0, { 4 }, { 4, 4 } },
-  { "maxul", maxul, NULL, 0, { 4 }, { 4, 4 } },
-  { "minsl", minsl, NULL, 0, { 4 }, { 4, 4 } },
-  { "minul", minul, NULL, 0, { 4 }, { 4, 4 } },
-  { "mulll", mulll, NULL, 0, { 4 }, { 4, 4 } },
-  { "mulhsl", mulhsl, NULL, 0, { 4 }, { 4, 4 } },
-  { "mulhul", mulhul, NULL, 0, { 4 }, { 4, 4 } },
-  { "orl", orl, NULL, 0, { 4 }, { 4, 4 } },
-  { "shll", shll, NULL, ORC_STATIC_OPCODE_SCALAR, { 4 }, { 4, 4 } },
-  { "shrsl", shrsl, NULL, ORC_STATIC_OPCODE_SCALAR, { 4 }, { 4, 4 } },
-  { "shrul", shrul, NULL, ORC_STATIC_OPCODE_SCALAR, { 4 }, { 4, 4 } },
-  { "signl", signl, NULL, 0, { 4 }, { 4 } },
-  { "subl", subl, NULL, 0, { 4 }, { 4, 4 } },
-  { "subssl", subssl, NULL, 0, { 4 }, { 4, 4 } },
-  { "subusl", subusl, NULL, 0, { 4 }, { 4, 4 } },
-  { "xorl", xorl, NULL, 0, { 4 }, { 4, 4 } },
-
-  { "convsbw", convsbw, NULL, 0, { 2 }, { 1 } },
-  { "convubw", convubw, NULL, 0, { 2 }, { 1 } },
-  { "convswl", convswl, NULL, 0, { 4 }, { 2 } },
-  { "convuwl", convuwl, NULL, 0, { 4 }, { 2 } },
+  { "absl", absl, NULL, 0, { 4 }, { 4 }, emulate_absl },
+  { "addl", addl, NULL, 0, { 4 }, { 4, 4 }, emulate_addl },
+  { "addssl", addssl, NULL, 0, { 4 }, { 4, 4 }, emulate_addssl },
+  { "addusl", addusl, NULL, 0, { 4 }, { 4, 4 }, emulate_addusl },
+  { "andl", andl, NULL, 0, { 4 }, { 4, 4 }, emulate_andl },
+  { "andnl", andnl, NULL, 0, { 4 }, { 4, 4 }, emulate_andnl },
+  { "avgsl", avgsl, NULL, 0, { 4 }, { 4, 4 }, emulate_avgsl },
+  { "avgul", avgul, NULL, 0, { 4 }, { 4, 4 }, emulate_avgul },
+  { "cmpeql", cmpeql, NULL, 0, { 4 }, { 4, 4 }, emulate_cmpeql },
+  { "cmpgtsl", cmpgtsl, NULL, 0, { 4 }, { 4, 4 }, emulate_cmpgtsl },
+  { "copyl", copyl, NULL, 0, { 4 }, { 4 }, emulate_copyl },
+  { "maxsl", maxsl, NULL, 0, { 4 }, { 4, 4 }, emulate_maxsl },
+  { "maxul", maxul, NULL, 0, { 4 }, { 4, 4 }, emulate_maxul },
+  { "minsl", minsl, NULL, 0, { 4 }, { 4, 4 }, emulate_minsl },
+  { "minul", minul, NULL, 0, { 4 }, { 4, 4 }, emulate_minul },
+  { "mulll", mulll, NULL, 0, { 4 }, { 4, 4 }, emulate_mulll },
+  { "mulhsl", mulhsl, NULL, 0, { 4 }, { 4, 4 }, emulate_mulhsl },
+  { "mulhul", mulhul, NULL, 0, { 4 }, { 4, 4 }, emulate_mulhul },
+  { "orl", orl, NULL, 0, { 4 }, { 4, 4 }, emulate_orl },
+  { "shll", shll, NULL, ORC_STATIC_OPCODE_SCALAR, { 4 }, { 4, 4 }, emulate_shll },
+  { "shrsl", shrsl, NULL, ORC_STATIC_OPCODE_SCALAR, { 4 }, { 4, 4 }, emulate_shrsl },
+  { "shrul", shrul, NULL, ORC_STATIC_OPCODE_SCALAR, { 4 }, { 4, 4 }, emulate_shrul },
+  { "signl", signl, NULL, 0, { 4 }, { 4 }, emulate_signl },
+  { "subl", subl, NULL, 0, { 4 }, { 4, 4 }, emulate_subl },
+  { "subssl", subssl, NULL, 0, { 4 }, { 4, 4 }, emulate_subssl },
+  { "subusl", subusl, NULL, 0, { 4 }, { 4, 4 }, emulate_subusl },
+  { "xorl", xorl, NULL, 0, { 4 }, { 4, 4 }, emulate_xorl },
+
+  { "convsbw", convsbw, NULL, 0, { 2 }, { 1 }, emulate_convsbw },
+  { "convubw", convubw, NULL, 0, { 2 }, { 1 }, emulate_convubw },
+  { "convswl", convswl, NULL, 0, { 4 }, { 2 }, emulate_convswl },
+  { "convuwl", convuwl, NULL, 0, { 4 }, { 2 }, emulate_convuwl },
 #ifdef ENABLE_64BIT
-  { "convslq", convslq, NULL, 0, { 8 }, { 4 } },
-  { "convulq", convulq, NULL, 0, { 8 }, { 4 } },
+  { "convslq", convslq, NULL, 0, { 8 }, { 4 }, emulate_convslq },
+  { "convulq", convulq, NULL, 0, { 8 }, { 4 }, emulate_convulq },
 #endif
 
-  { "convwb", convwb, NULL, 0, { 1 }, { 2 } },
-  { "convssswb", convssswb, NULL, 0, { 1 }, { 2 } },
-  { "convsuswb", convsuswb, NULL, 0, { 1 }, { 2 } },
-  { "convusswb", convusswb, NULL, 0, { 1 }, { 2 } },
-  { "convuuswb", convuuswb, NULL, 0, { 1 }, { 2 } },
+  { "convwb", convwb, NULL, 0, { 1 }, { 2 }, emulate_convwb },
+  { "convssswb", convssswb, NULL, 0, { 1 }, { 2 }, emulate_convssswb },
+  { "convsuswb", convsuswb, NULL, 0, { 1 }, { 2 }, emulate_convsuswb },
+  { "convusswb", convusswb, NULL, 0, { 1 }, { 2 }, emulate_convusswb },
+  { "convuuswb", convuuswb, NULL, 0, { 1 }, { 2 }, emulate_convuuswb },
 
-  { "convlw", convlw, NULL, 0, { 2 }, { 4 } },
-  { "convssslw", convssslw, NULL, 0, { 2 }, { 4 } },
-  { "convsuslw", convsuslw, NULL, 0, { 2 }, { 4 } },
-  { "convusslw", convusslw, NULL, 0, { 2 }, { 4 } },
-  { "convuuslw", convuuslw, NULL, 0, { 2 }, { 4 } },
+  { "convlw", convlw, NULL, 0, { 2 }, { 4 }, emulate_convlw },
+  { "convssslw", convssslw, NULL, 0, { 2 }, { 4 }, emulate_convssslw },
+  { "convsuslw", convsuslw, NULL, 0, { 2 }, { 4 }, emulate_convsuslw },
+  { "convusslw", convusslw, NULL, 0, { 2 }, { 4 }, emulate_convusslw },
+  { "convuuslw", convuuslw, NULL, 0, { 2 }, { 4 }, emulate_convuuslw },
 
 #ifdef ENABLE_64BIT
-  { "convql", convql, NULL, 0, { 4 }, { 8 } },
-  { "convssql", convssql, NULL, 0, { 4 }, { 8 } },
-  { "convusql", convusql, NULL, 0, { 4 }, { 8 } },
+  { "convql", convql, NULL, 0, { 4 }, { 8 }, emulate_convql },
+  { "convssql", convssql, NULL, 0, { 4 }, { 8 }, emulate_convssql },
+  { "convusql", convusql, NULL, 0, { 4 }, { 8 }, emulate_convusql },
 #endif
 
-  { "mulsbw", mulsbw, NULL, 0, { 2 }, { 1, 1 } },
-  { "mulubw", mulubw, NULL, 0, { 2 }, { 1, 1 } },
-  { "mulswl", mulswl, NULL, 0, { 4 }, { 2, 2 } },
-  { "muluwl", muluwl, NULL, 0, { 4 }, { 2, 2 } },
+  { "mulsbw", mulsbw, NULL, 0, { 2 }, { 1, 1 }, emulate_mulsbw },
+  { "mulubw", mulubw, NULL, 0, { 2 }, { 1, 1 }, emulate_mulubw },
+  { "mulswl", mulswl, NULL, 0, { 4 }, { 2, 2 }, emulate_mulswl },
+  { "muluwl", muluwl, NULL, 0, { 4 }, { 2, 2 }, emulate_muluwl },
 #ifdef ENABLE_64BIT
-  { "mulslq", mulslq, NULL, 0, { 8 }, { 4, 4 } },
-  { "mululq", mululq, NULL, 0, { 8 }, { 4, 4 } },
+  { "mulslq", mulslq, NULL, 0, { 8 }, { 4, 4 }, emulate_mulslq },
+  { "mululq", mululq, NULL, 0, { 8 }, { 4, 4 }, emulate_mululq },
 #endif
 
   /* accumulators */
-  { "accw", accw, NULL, ORC_STATIC_OPCODE_ACCUMULATOR, { 2 }, { 2 } },
-  { "accl", accl, NULL, ORC_STATIC_OPCODE_ACCUMULATOR, { 4 }, { 4 } },
-  { "accsadubl", accsadubl, NULL, ORC_STATIC_OPCODE_ACCUMULATOR, { 4 }, { 1, 1 } },
-
-  { "swapw", swapw, NULL, 0, { 2 }, { 2 } },
-  { "swapl", swapl, NULL, 0, { 4 }, { 4 } },
-  { "select0wb", select0wb, NULL, 0, { 1 }, { 2 } },
-  { "select1wb", select1wb, NULL, 0, { 1 }, { 2 } },
-  { "select0lw", select0lw, NULL, 0, { 2 }, { 4 } },
-  { "select1lw", select1lw, NULL, 0, { 2 }, { 4 } },
-  { "mergewl", mergewl, NULL, 0, { 4 }, { 2, 2 } },
-  { "mergebw", mergebw, NULL, 0, { 2 }, { 1, 1 } },
-  { "splitlw", splitlw, NULL, 0, { 2, 2 }, { 4 } },
-  { "splitwb", splitwb, NULL, 0, { 1, 1 }, { 2 } },
+  { "accw", accw, NULL, ORC_STATIC_OPCODE_ACCUMULATOR, { 2 }, { 2 }, emulate_accw },
+  { "accl", accl, NULL, ORC_STATIC_OPCODE_ACCUMULATOR, { 4 }, { 4 }, emulate_accl },
+  { "accsadubl", accsadubl, NULL, ORC_STATIC_OPCODE_ACCUMULATOR, { 4 }, { 1, 1 }, emulate_accsadubl },
+
+  { "swapw", swapw, NULL, 0, { 2 }, { 2 }, emulate_swapw },
+  { "swapl", swapl, NULL, 0, { 4 }, { 4 }, emulate_swapl },
+  { "select0wb", select0wb, NULL, 0, { 1 }, { 2 }, emulate_select0wb },
+  { "select1wb", select1wb, NULL, 0, { 1 }, { 2 }, emulate_select1wb },
+  { "select0lw", select0lw, NULL, 0, { 2 }, { 4 }, emulate_select0lw },
+  { "select1lw", select1lw, NULL, 0, { 2 }, { 4 }, emulate_select1lw },
+  { "mergewl", mergewl, NULL, 0, { 4 }, { 2, 2 }, emulate_mergewl },
+  { "mergebw", mergebw, NULL, 0, { 2 }, { 1, 1 }, emulate_mergebw },
+  { "splitlw", splitlw, NULL, 0, { 2, 2 }, { 4 }, emulate_splitlw },
+  { "splitwb", splitwb, NULL, 0, { 1, 1 }, { 2 }, emulate_splitwb },
 
   /* float ops */
-  { "addf", addf, NULL, ORC_STATIC_OPCODE_FLOAT, { 4 }, { 4, 4 } },
-  { "subf", subf, NULL, ORC_STATIC_OPCODE_FLOAT, { 4 }, { 4, 4 } },
-  { "mulf", mulf, NULL, ORC_STATIC_OPCODE_FLOAT, { 4 }, { 4, 4 } },
-  { "divf", divf, NULL, ORC_STATIC_OPCODE_FLOAT, { 4 }, { 4, 4 } },
-  { "sqrtf", orc_sqrtf, NULL, ORC_STATIC_OPCODE_FLOAT, { 4 }, { 4 } },
-  { "maxf", maxf, NULL, ORC_STATIC_OPCODE_FLOAT, { 4 }, { 4, 4 } },
-  { "minf", minf, NULL, ORC_STATIC_OPCODE_FLOAT, { 4 }, { 4, 4 } },
-  { "cmpeqf", cmpeqf, NULL, ORC_STATIC_OPCODE_FLOAT_SRC, { 4 }, { 4, 4 } },
-  { "cmpltf", cmpltf, NULL, ORC_STATIC_OPCODE_FLOAT_SRC, { 4 }, { 4, 4 } },
-  { "cmplef", cmplef, NULL, ORC_STATIC_OPCODE_FLOAT_SRC, { 4 }, { 4, 4 } },
-  { "convfl", convfl, NULL, ORC_STATIC_OPCODE_FLOAT_SRC, { 4 }, { 4 } },
-  { "convlf", convlf, NULL, ORC_STATIC_OPCODE_FLOAT_DEST, { 4 }, { 4 } },
+  { "addf", addf, NULL, ORC_STATIC_OPCODE_FLOAT, { 4 }, { 4, 4 }, emulate_addf },
+  { "subf", subf, NULL, ORC_STATIC_OPCODE_FLOAT, { 4 }, { 4, 4 }, emulate_subf },
+  { "mulf", mulf, NULL, ORC_STATIC_OPCODE_FLOAT, { 4 }, { 4, 4 }, emulate_mulf },
+  { "divf", divf, NULL, ORC_STATIC_OPCODE_FLOAT, { 4 }, { 4, 4 }, emulate_divf },
+  { "sqrtf", orc_sqrtf, NULL, ORC_STATIC_OPCODE_FLOAT, { 4 }, { 4 }, emulate_sqrtf },
+  { "maxf", maxf, NULL, ORC_STATIC_OPCODE_FLOAT, { 4 }, { 4, 4 }, emulate_maxf },
+  { "minf", minf, NULL, ORC_STATIC_OPCODE_FLOAT, { 4 }, { 4, 4 }, emulate_minf },
+  { "cmpeqf", cmpeqf, NULL, ORC_STATIC_OPCODE_FLOAT_SRC, { 4 }, { 4, 4 }, emulate_cmpeqf },
+  { "cmpltf", cmpltf, NULL, ORC_STATIC_OPCODE_FLOAT_SRC, { 4 }, { 4, 4 }, emulate_cmpltf },
+  { "cmplef", cmplef, NULL, ORC_STATIC_OPCODE_FLOAT_SRC, { 4 }, { 4, 4 }, emulate_cmplef },
+  { "convfl", convfl, NULL, ORC_STATIC_OPCODE_FLOAT_SRC, { 4 }, { 4 }, emulate_convfl },
+  { "convlf", convlf, NULL, ORC_STATIC_OPCODE_FLOAT_DEST, { 4 }, { 4 }, emulate_convlf },
 
   { "" }
 };
diff --git a/orc/orcprogram-c.c b/orc/orcprogram-c.c
index 844a8b5..d085f7b 100644
--- a/orc/orcprogram-c.c
+++ b/orc/orcprogram-c.c
@@ -102,6 +102,10 @@ orc_target_c_get_asm_preamble (void)
     "#define ORC_SWAP_W(x) ((((x)&0xff)<<8) | (((x)&0xff00)>>8))\n"
     "#define ORC_SWAP_L(x) ((((x)&0xff)<<24) | (((x)&0xff00)<<8) | (((x)&0xff0000)>>8) | (((x)&0xff000000)>>24))\n"
     "#define ORC_PTR_OFFSET(ptr,offset) ((void *)(((unsigned char *)(ptr)) + (offset)))\n"
+    "#define ORC_MIN_NORMAL (1.1754944909521339405e-38)\n"
+    "#define ORC_DENORMAL(x) (((x) > -ORC_MIN_NORMAL && (x) < ORC_MIN_NORMAL) ? ((x)<0 ? (-0.0f) : (0.0f)) : (x))\n"
+    "#define ORC_MINF(a,b) (isnan(a) ? a : isnan(b) ? b : ((a)<(b)) ? (a) : (b))\n"
+    "#define ORC_MAXF(a,b) (isnan(a) ? a : isnan(b) ? b : ((a)>(b)) ? (a) : (b))\n"
     "/* end Orc C target preamble */\n\n";
 }
 
@@ -129,24 +133,30 @@ static const char *varnames[] = {
 static void
 get_varname (char *s, OrcCompiler *compiler, int var)
 {
-  if (!(compiler->target_flags & ORC_TARGET_C_NOEXEC)) {
-    sprintf(s, "ex->arrays[%d]", var);
-  } else {
+  if (compiler->target_flags & ORC_TARGET_C_NOEXEC) {
     if (var < 48) {
       strcpy (s, varnames[var]);
     } else {
       sprintf(s, "t%d", var-32);
     }
+  } else if (compiler->target_flags & ORC_TARGET_C_OPCODE) {
+    if (var < ORC_VAR_S1) {
+      sprintf(s, "ex->dest_ptrs[%d]", var-ORC_VAR_D1);
+    } else {
+      sprintf(s, "ex->src_ptrs[%d]", var-ORC_VAR_S1);
+    }
+  } else {
+    sprintf(s, "ex->arrays[%d]", var);
   }
 }
 
 static void
 get_varname_stride (char *s, OrcCompiler *compiler, int var)
 {
-  if (!(compiler->target_flags & ORC_TARGET_C_NOEXEC)) {
-    sprintf(s, "ex->params[%d]", var);
-  } else {
+  if (compiler->target_flags & ORC_TARGET_C_NOEXEC) {
     sprintf(s, "%s_stride", varnames[var]);
+  } else {
+    sprintf(s, "ex->params[%d]", var);
   }
 }
 
@@ -171,7 +181,8 @@ orc_compiler_c_assemble (OrcCompiler *compiler)
     ORC_ASM_CODE(compiler,"  int j;\n");
   }
   if (compiler->program->constant_n == 0) {
-    if (!(compiler->target_flags & ORC_TARGET_C_NOEXEC)) {
+    if (!(compiler->target_flags & ORC_TARGET_C_NOEXEC) &&
+        !(compiler->target_flags & ORC_TARGET_C_OPCODE)) {
       ORC_ASM_CODE(compiler,"  int n = ex->n;\n");
     }
   } else {
@@ -241,7 +252,19 @@ orc_compiler_c_assemble (OrcCompiler *compiler)
         break;
       case ORC_VAR_TYPE_PARAM:
         c_get_name (varname, compiler, i);
-        if (!(compiler->target_flags & ORC_TARGET_C_NOEXEC)) {
+        if (compiler->target_flags & ORC_TARGET_C_NOEXEC) {
+          ORC_ASM_CODE(compiler,"  const %s %s = %s;\n",
+              var->is_float_param ? "float" : "int",
+              varname, varnames[i]);
+        } else if (compiler->target_flags & ORC_TARGET_C_OPCODE) {
+          if (var->is_float_param) {
+            ORC_ASM_CODE(compiler,"  const float %s = ((orc_union32 *)(ex->src_ptrs[%d]))->f;\n",
+                varname, i - ORC_VAR_P1 + compiler->program->n_src_vars);
+          } else {
+            ORC_ASM_CODE(compiler,"  const int %s = ((orc_union32 *)(ex->src_ptrs[%d]))->i;\n",
+                varname, i - ORC_VAR_P1 + compiler->program->n_src_vars);
+          }
+        } else {
           if (var->is_float_param) {
             ORC_ASM_CODE(compiler,"  const float %s = ((orc_union32 *)(ex->params+%d))->f;\n",
                 varname, i);
@@ -249,10 +272,6 @@ orc_compiler_c_assemble (OrcCompiler *compiler)
             ORC_ASM_CODE(compiler,"  const int %s = ex->params[%d];\n",
                 varname, i);
           }
-        } else {
-          ORC_ASM_CODE(compiler,"  const %s %s = %s;\n",
-              var->is_float_param ? "float" : "int",
-              varname, varnames[i]);
         }
         break;
       default:
@@ -374,20 +393,27 @@ orc_compiler_c_assemble (OrcCompiler *compiler)
     switch (var->vartype) {
       case ORC_VAR_TYPE_ACCUMULATOR:
         if (var->size == 2) {
-          if (!(compiler->target_flags & ORC_TARGET_C_NOEXEC)) {
-            ORC_ASM_CODE(compiler,"  ex->accumulators[%d] = (%s & 0xffff);\n",
-                i - ORC_VAR_A1, varname);
-          } else {
+          if (compiler->target_flags & ORC_TARGET_C_NOEXEC) {
             ORC_ASM_CODE(compiler,"  *%s = (%s & 0xffff);\n",
                 varnames[i], varname);
+          } else if (compiler->target_flags & ORC_TARGET_C_OPCODE) {
+            ORC_ASM_CODE(compiler,"  ((orc_union32 *)ex->dest_ptrs[%d])->i = "
+                "(%s + ((orc_union32 *)ex->dest_ptrs[%d])->i) & 0xffff;\n",
+                i - ORC_VAR_A1, varname, i - ORC_VAR_A1);
+          } else {
+            ORC_ASM_CODE(compiler,"  ex->accumulators[%d] = (%s & 0xffff);\n",
+                i - ORC_VAR_A1, varname);
           }
         } else {
-          if (!(compiler->target_flags & ORC_TARGET_C_NOEXEC)) {
-            ORC_ASM_CODE(compiler,"  ex->accumulators[%d] = %s;\n",
-                i - ORC_VAR_A1, varname);
-          } else {
+          if (compiler->target_flags & ORC_TARGET_C_NOEXEC) {
             ORC_ASM_CODE(compiler,"  *%s = %s;\n",
                 varnames[i], varname);
+          } else if (compiler->target_flags & ORC_TARGET_C_OPCODE) {
+            ORC_ASM_CODE(compiler,"  ((orc_union32 *)ex->dest_ptrs[%d])->i += %s;\n",
+                i - ORC_VAR_A1, varname);
+          } else {
+            ORC_ASM_CODE(compiler,"  ex->accumulators[%d] = %s;\n",
+                i - ORC_VAR_A1, varname);
           }
         }
         break;
diff --git a/orc/orcprogram.h b/orc/orcprogram.h
index daa03ed..a4b1a66 100644
--- a/orc/orcprogram.h
+++ b/orc/orcprogram.h
@@ -20,6 +20,8 @@ typedef struct _OrcFixup OrcFixup;
 typedef struct _OrcTarget OrcTarget;
 
 typedef void (*OrcOpcodeEmulateFunc)(OrcOpcodeExecutor *ex, void *user);
+typedef void (*OrcOpcodeEmulateNFunc)(OrcOpcodeExecutor *ex, int n);
+typedef void (*OrcOpcodeEmulate16Func)(OrcOpcodeExecutor *ex);
 typedef void (*OrcRuleEmitFunc)(OrcCompiler *p, void *user, OrcInstruction *insn);
 typedef void (*OrcExecutorFunc)(OrcExecutor *ex);
 
@@ -75,6 +77,7 @@ enum {
   ORC_TARGET_C_C99 = (1<<0),
   ORC_TARGET_C_BARE = (1<<1),
   ORC_TARGET_C_NOEXEC = (1<<2),
+  ORC_TARGET_C_OPCODE = (1<<3),
   ORC_TARGET_FAST_NAN = (1<<30),
   ORC_TARGET_FAST_DENORMAL = (1<<31)
 };
@@ -269,6 +272,7 @@ struct _OrcStaticOpcode {
   unsigned int flags;
   int dest_size[ORC_STATIC_OPCODE_N_DEST];
   int src_size[ORC_STATIC_OPCODE_N_SRC];
+  OrcOpcodeEmulateNFunc emulateN;
 };
 
 /**
@@ -416,6 +420,11 @@ struct _OrcOpcodeExecutor {
   /*< private >*/
   int src_values[ORC_STATIC_OPCODE_N_SRC];
   int dest_values[ORC_STATIC_OPCODE_N_DEST];
+
+  OrcOpcodeEmulateNFunc emulateN;
+
+  void *src_ptrs[ORC_STATIC_OPCODE_N_SRC];
+  void *dest_ptrs[ORC_STATIC_OPCODE_N_DEST];
 };
 
 /**
-- 
2.7.4