llvm: small updatesllvm-backend

1 files changed, 104 insertions, 94 deletions

diff --git a/orc/orcprogram-llvm.c b/orc/orcprogram-llvm.c
index 3768068..a615523 100644
--- a/orc/orcprogram-llvm.c
+++ b/orc/orcprogram-llvm.c
@@ -29,7 +29,7 @@ orc_compiler_llvm_init (OrcCompiler *compiler)
   for(i=ORC_GP_REG_BASE;i<ORC_GP_REG_BASE+32;i++){
     compiler->valid_regs[i] = 1;
   }
-  compiler->loop_shift = 0;
+  compiler->loop_shift = 1;
 }
 
 const char *
@@ -102,6 +102,7 @@ output_prototype (OrcCompiler *compiler)
   OrcProgram *p = compiler->program;
   int i;
   int need_comma;
+  int mult = 1 << compiler->loop_shift;
 
   ORC_ASM_CODE(compiler,"define void @%s (", compiler->program->name);
   need_comma = FALSE;
@@ -110,7 +111,7 @@ output_prototype (OrcCompiler *compiler)
     var = &p->vars[ORC_VAR_D1 + i];
     if (var->size) {
       if (need_comma) ORC_ASM_CODE (compiler, ", ");
-      ORC_ASM_CODE (compiler, "<1 x i%d> * noalias %s", var->size*8,
+      ORC_ASM_CODE (compiler, "<%d x i%d> * noalias %s", mult, var->size*8,
           varnames[ORC_VAR_D1 + i]);
       if (p->is_2d) {
         ORC_ASM_CODE (compiler, ", i32 %s_stride", varnames[ORC_VAR_D1 + i]);
@@ -122,7 +123,7 @@ output_prototype (OrcCompiler *compiler)
     var = &p->vars[ORC_VAR_A1 + i];
     if (var->size) {
       if (need_comma) ORC_ASM_CODE (compiler, ", ");
-      ORC_ASM_CODE (compiler, "<1 x i%d> * noalias %s", var->size*8,
+      ORC_ASM_CODE (compiler, "<%d x i%d> * noalias %s", mult, var->size*8,
           varnames[ORC_VAR_A1 + i]);
       need_comma = TRUE;
     }
@@ -131,7 +132,7 @@ output_prototype (OrcCompiler *compiler)
     var = &p->vars[ORC_VAR_S1 + i];
     if (var->size) {
       if (need_comma) ORC_ASM_CODE (compiler, ", ");
-      ORC_ASM_CODE (compiler, "<1 x i%d> * noalias %s", var->size*8,
+      ORC_ASM_CODE (compiler, "<%d x i%d> * noalias %s", mult, var->size*8,
           varnames[ORC_VAR_S1 + i]);
       if (p->is_2d) {
         ORC_ASM_CODE (compiler, ", i32 %s_stride", varnames[ORC_VAR_S1 + i]);
@@ -194,6 +195,7 @@ orc_compiler_llvm_assemble (OrcCompiler *compiler)
   OrcStaticOpcode *opcode;
   OrcRule *rule;
   int prefix = 0;
+  int mult = 1 << compiler->loop_shift;
 
   if (!(compiler->target_flags & ORC_TARGET_C_BARE)) {
     output_prototype (compiler);
@@ -206,7 +208,7 @@ orc_compiler_llvm_assemble (OrcCompiler *compiler)
 
     ORC_ASM_CODE (compiler, "  store i32 0, i32* %%j\n\n");
 
-    ORC_ASM_CODE (compiler, "  %%index_m = load i32* %%j\n");
+    ORC_ASM_CODE (compiler, "  %%index_m = load i32, i32* %%j\n");
     ORC_ASM_CODE (compiler, "  %%cmp_m = icmp slt i32 %%index_m, %%m\n");
     ORC_ASM_CODE (compiler, "  br i1 %%cmp_m, label %%loop_m, label %%exit_m\n\n");
 
@@ -219,21 +221,13 @@ orc_compiler_llvm_assemble (OrcCompiler *compiler)
       if (var->name == NULL) continue;
       switch (var->vartype) {
         case ORC_VAR_TYPE_SRC:
+        case ORC_VAR_TYPE_DEST:
           {
             char s1[40], s2[40];
             get_varname(s1, compiler, i);
             get_varname_stride(s2, compiler, i);
-            ORC_ASM_CODE(compiler,"  %%ptr%d = getelementptr <1 x i%d>* %s\n",
-                i, var->size*8, varnames[i]);
-          }
-          break;
-        case ORC_VAR_TYPE_DEST:
-          {
-            char s1[40], s2[40];
-            get_varname(s1, compiler, i),
-            get_varname_stride(s2, compiler, i),
-            ORC_ASM_CODE(compiler,"  %%ptr%d = getelementptr <1 x i%d>* %s\n",
-                i, var->size*8, varnames[i]);
+            ORC_ASM_CODE(compiler,"  %%ptr%d = getelementptr <%d x i%d>, <%d x i%d>* %s\n",
+                i, mult, var->size*8, mult, var->size*8, varnames[i]);
           }
           break;
         default:
@@ -246,12 +240,9 @@ orc_compiler_llvm_assemble (OrcCompiler *compiler)
       if (var->name == NULL) continue;
       switch (var->vartype) {
         case ORC_VAR_TYPE_SRC:
-          ORC_ASM_CODE(compiler,"  %%ptr%d = getelementptr <1 x i%d>* %s\n",
-              i, var->size*8, varnames[i]);
-          break;
         case ORC_VAR_TYPE_DEST:
-          ORC_ASM_CODE(compiler,"  %%ptr%d = getelementptr <1 x i%d>* %s\n",
-              i, var->size*8, varnames[i]);
+          ORC_ASM_CODE(compiler,"  %%ptr%d = getelementptr <%d x i%d>, <%d x i%d>* %s\n",
+              i, mult, var->size*8, mult, var->size*8, varnames[i]);
           break;
         default:
           break;
@@ -282,7 +273,7 @@ orc_compiler_llvm_assemble (OrcCompiler *compiler)
   ORC_ASM_CODE (compiler, "%*s  br label %%head_n\n\n", prefix, "");
 
   ORC_ASM_CODE (compiler, "%*shead_n:\n", prefix, "");
-  ORC_ASM_CODE (compiler, "%*s  %%index_n = load i32* %%i\n", prefix, "");
+  ORC_ASM_CODE (compiler, "%*s  %%index_n = load i32, i32* %%i\n", prefix, "");
   ORC_ASM_CODE (compiler, "%*s  %%cmp_n = icmp slt i32 %%index_n, %%n\n", prefix, "");
   ORC_ASM_CODE (compiler, "%*s  br i1 %%cmp_n, label %%loop_n, label %%exit_n\n\n", prefix, "");
 
@@ -390,6 +381,20 @@ get_vector_constant (char *name, int len, int bits, orc_int64 val)
 }
 
 /* rules */
+static int
+get_shift (OrcCompiler *p, OrcInstruction *insn)
+{
+  int res;
+  if (insn && (insn->flags & ORC_INSTRUCTION_FLAG_X2)) {
+    res = 2;
+  } else if (insn && (insn->flags & ORC_INSTRUCTION_FLAG_X4)) {
+    res = 4;
+  } else {
+    res = 1;
+  }
+  return res << p->loop_shift;
+}
+
 
 static void
 llvm_get_name_int (char *name, OrcCompiler *p, OrcInstruction *insn, int var, int src)
@@ -422,26 +427,15 @@ llvm_get_name_int (char *name, OrcCompiler *p, OrcInstruction *insn, int var, in
     }
   } else if (p->vars[var].vartype == ORC_VAR_TYPE_CONST) {
     int bits = p->vars[var].size * 8;
-
-    if (insn && (insn->flags & (ORC_INSTRUCTION_FLAG_X2|ORC_INSTRUCTION_FLAG_X4))) {
-      if (insn->flags & ORC_INSTRUCTION_FLAG_X4)
-        get_vector_constant (name, 4, bits, p->vars[var].value.i);
-      else
-        get_vector_constant (name, 2, bits, p->vars[var].value.i);
-    } else
-      get_vector_constant (name, 1, bits, p->vars[var].value.i);
+    int sh = get_shift (p, insn);
+    get_vector_constant (name, sh, bits, p->vars[var].value.i);
   } else {
     if (insn && (insn->flags & (ORC_INSTRUCTION_FLAG_X2|ORC_INSTRUCTION_FLAG_X4))) {
       int bits = p->vars[var].size * 8;
-      int multiplier;
+      int multiplier = get_shift (p, insn);
 
       get_temp (name, p);
 
-      if (insn->flags & ORC_INSTRUCTION_FLAG_X4)
-        multiplier = 4;
-      else
-        multiplier = 2;
-
       if (src)
         ORC_ASM_CODE (p, "   %s = bitcast <1 x i%d> %%var%d to <%d x i%d>\n",
           name, bits, var, multiplier, bits / multiplier);
@@ -487,20 +481,6 @@ llvm_store_dst_int (char *name, OrcCompiler *p, OrcInstruction *insn, int var)
      var, multiplier, bits / multiplier, name, bits);
 }
 
-static int
-get_shift (OrcCompiler *p, OrcInstruction *insn)
-{
-  int res;;
-  if (insn && (insn->flags & ORC_INSTRUCTION_FLAG_X2)) {
-    res = 2;
-  } else if (insn && (insn->flags & ORC_INSTRUCTION_FLAG_X4)) {
-    res = 4;
-  } else {
-    res = 1;
-  }
-  return res;
-}
-
 static void
 llvm_rule_absb (OrcCompiler *p, void *user, OrcInstruction *insn)
 {
@@ -692,9 +672,9 @@ llvm_rule_loadb (OrcCompiler *p, void *user, OrcInstruction *insn)
 
   llvm_get_dst_int (dest, p, insn, insn->dest_args[0]);
 
-  ORC_ASM_CODE (p, "   %s = getelementptr <%d x i8>* %%ptr%d, i32 %%index_n\n",
-      t1, sh, insn->src_args[0]);
-  ORC_ASM_CODE (p, "   %s = load <%d x i8>* %s\n", dest, sh, t1);
+  ORC_ASM_CODE (p, "   %s = getelementptr <%d x i8>, <%d x i8>* %%ptr%d, i32 %%index_n\n",
+      t1, sh, sh, insn->src_args[0]);
+  ORC_ASM_CODE (p, "   %s = load <%d x i8>, <%d x i8>* %s\n", dest, sh, sh, t1);
 
   llvm_store_dst_int (dest, p, insn, insn->dest_args[0]);
 }
@@ -709,9 +689,9 @@ llvm_rule_loadoffb (OrcCompiler *p, void *user, OrcInstruction *insn)
   llvm_get_dst_int (dest, p, insn, insn->dest_args[0]);
 
   /* FIXME */
-  ORC_ASM_CODE (p, "   %s = getelementptr <%d x i8>* %%ptr%d, i32 %%index_n\n",
-      t1, sh, insn->src_args[0]);
-  ORC_ASM_CODE (p, "   %s = load <%d x i8>* %s\n", dest, sh, t1);
+  ORC_ASM_CODE (p, "   %s = getelementptr <%d x i8>, <%d x i8>* %%ptr%d, i32 %%index_n\n",
+      t1, sh, sh, insn->src_args[0]);
+  ORC_ASM_CODE (p, "   %s = load <%d x i8>, <%d x i8>* %s\n", dest, sh, sh, t1);
 
   llvm_store_dst_int (dest, p, insn, insn->dest_args[0]);
 }
@@ -726,9 +706,9 @@ llvm_rule_loadupdb (OrcCompiler *p, void *user, OrcInstruction *insn)
   llvm_get_dst_int (dest, p, insn, insn->dest_args[0]);
 
   /* FIXME */
-  ORC_ASM_CODE (p, "   %s = getelementptr <%d x i8>* %%ptr%d, i32 %%index_n\n",
-      t1, sh, insn->src_args[0]);
-  ORC_ASM_CODE (p, "   %s = load <%d x i8>* %s\n", dest, sh, t1);
+  ORC_ASM_CODE (p, "   %s = getelementptr <%d x i8>, <%d x i8>* %%ptr%d, i32 %%index_n\n",
+      t1, sh, sh, insn->src_args[0]);
+  ORC_ASM_CODE (p, "   %s = load <%d x i8>, <%d x i8>* %s\n", dest, sh, sh, t1);
 
   llvm_store_dst_int (dest, p, insn, insn->dest_args[0]);
 }
@@ -743,9 +723,9 @@ llvm_rule_loadupib (OrcCompiler *p, void *user, OrcInstruction *insn)
   llvm_get_dst_int (dest, p, insn, insn->dest_args[0]);
 
   /* FIXME */
-  ORC_ASM_CODE (p, "   %s = getelementptr <%d x i8>* %%ptr%d, i32 %%index_n\n",
-      t1, sh, insn->src_args[0]);
-  ORC_ASM_CODE (p, "   %s = load <%d x i8>* %s\n", dest, sh, t1);
+  ORC_ASM_CODE (p, "   %s = getelementptr <%d x i8>, <%d x i8>* %%ptr%d, i32 %%index_n\n",
+      t1, sh, sh, insn->src_args[0]);
+  ORC_ASM_CODE (p, "   %s = load <%d x i8>, <%d x i8>* %s\n", dest, sh, sh, t1);
 
   llvm_store_dst_int (dest, p, insn, insn->dest_args[0]);
 }
@@ -780,9 +760,9 @@ llvm_rule_ldresnearb (OrcCompiler *p, void *user, OrcInstruction *insn)
   llvm_get_dst_int (dest, p, insn, insn->dest_args[0]);
 
   /* FIXME */
-  ORC_ASM_CODE (p, "   %s = getelementptr <%d x i8>* %%ptr%d, i32 %%index_n\n",
-      t1, sh, insn->src_args[0]);
-  ORC_ASM_CODE (p, "   %s = load <%d x i8>* %s\n", dest, sh, t1);
+  ORC_ASM_CODE (p, "   %s = getelementptr <%d x i8>, <%d x i8>* %%ptr%d, i32 %%index_n\n",
+      t1, sh, sh, insn->src_args[0]);
+  ORC_ASM_CODE (p, "   %s = load <%d x i8>, <%d x i8>* %s\n", dest, sh, sh, t1);
 
   llvm_store_dst_int (dest, p, insn, insn->dest_args[0]);
 }
@@ -797,9 +777,9 @@ llvm_rule_ldresnearl (OrcCompiler *p, void *user, OrcInstruction *insn)
   llvm_get_dst_int (dest, p, insn, insn->dest_args[0]);
 
   /* FIXME */
-  ORC_ASM_CODE (p, "   %s = getelementptr <%d x i32>* %%ptr%d, i32 %%index_n\n",
-      t1, sh, insn->src_args[0]);
-  ORC_ASM_CODE (p, "   %s = load <%d x i32>* %s\n", dest, sh, t1);
+  ORC_ASM_CODE (p, "   %s = getelementptr <%d x i32>, <%d x i32>* %%ptr%d, i32 %%index_n\n",
+      t1, sh, sh, insn->src_args[0]);
+  ORC_ASM_CODE (p, "   %s = load <%d x i32>, <%d x i32>* %s\n", dest, sh, sh, t1);
 
   llvm_store_dst_int (dest, p, insn, insn->dest_args[0]);
 }
@@ -814,9 +794,9 @@ llvm_rule_ldreslinb (OrcCompiler *p, void *user, OrcInstruction *insn)
   llvm_get_dst_int (dest, p, insn, insn->dest_args[0]);
 
   /* FIXME */
-  ORC_ASM_CODE (p, "   %s = getelementptr <%d x i8>* %%ptr%d, i32 %%index_n\n",
-      t1, sh, insn->src_args[0]);
-  ORC_ASM_CODE (p, "   %s = load <%d x i8>* %s\n", dest, sh, t1);
+  ORC_ASM_CODE (p, "   %s = getelementptr <%d x i8>, <%d x i8>* %%ptr%d, i32 %%index_n\n",
+      t1, sh, sh, insn->src_args[0]);
+  ORC_ASM_CODE (p, "   %s = load <%d x i8>, <%d x i8>* %s\n", dest, sh, sh, t1);
 
   llvm_store_dst_int (dest, p, insn, insn->dest_args[0]);
 }
@@ -831,9 +811,9 @@ llvm_rule_ldreslinl (OrcCompiler *p, void *user, OrcInstruction *insn)
   llvm_get_dst_int (dest, p, insn, insn->dest_args[0]);
 
   /* FIXME */
-  ORC_ASM_CODE (p, "   %s = getelementptr <%d x i32>* %%ptr%d, i32 %%index_n\n",
-      t1, sh, insn->src_args[0]);
-  ORC_ASM_CODE (p, "   %s = load <%d x i32>* %s\n", dest, sh, t1);
+  ORC_ASM_CODE (p, "   %s = getelementptr <%d x i32>, <%d x i32>* %%ptr%d, i32 %%index_n\n",
+      t1, sh, sh, insn->src_args[0]);
+  ORC_ASM_CODE (p, "   %s = load <%d x i32>, <%d x i32>* %s\n", dest, sh, sh, t1);
 
   llvm_store_dst_int (dest, p, insn, insn->dest_args[0]);
 }
@@ -1012,8 +992,8 @@ llvm_rule_storeb (OrcCompiler *p, void *user, OrcInstruction *insn)
 
   llvm_get_dst_int (src, p, insn, insn->src_args[0]);
 
-  ORC_ASM_CODE (p, "   %s = getelementptr <%d x i8>* %%ptr%d, i32 %%index_n\n",
-      t1, sh, insn->dest_args[0]);
+  ORC_ASM_CODE (p, "   %s = getelementptr <%d x i8>, <%d x i8>* %%ptr%d, i32 %%index_n\n",
+      t1, sh, sh, insn->dest_args[0]);
   ORC_ASM_CODE (p, "   store <%d x i8> %s, <%d x i8>* %s\n", sh, src, sh, t1);
 }
 
@@ -1290,9 +1270,9 @@ llvm_rule_loadw (OrcCompiler *p, void *user, OrcInstruction *insn)
 
   llvm_get_dst_int (dest, p, insn, insn->dest_args[0]);
 
-  ORC_ASM_CODE (p, "   %s = getelementptr <%d x i16>* %%ptr%d, i32 %%index_n\n",
-      t1, sh, insn->src_args[0]);
-  ORC_ASM_CODE (p, "   %s = load <%d x i16>* %s\n", dest, sh, t1);
+  ORC_ASM_CODE (p, "   %s = getelementptr <%d x i16>, <%d x i16>* %%ptr%d, i32 %%index_n\n",
+      t1, sh, sh, insn->src_args[0]);
+  ORC_ASM_CODE (p, "   %s = load <%d x i16>, <%d x i16>* %s\n", dest, sh, sh, t1);
 
   llvm_store_dst_int (dest, p, insn, insn->dest_args[0]);
 }
@@ -1307,9 +1287,9 @@ llvm_rule_loadoffw (OrcCompiler *p, void *user, OrcInstruction *insn)
   llvm_get_dst_int (dest, p, insn, insn->dest_args[0]);
 
   /* FIXME */
-  ORC_ASM_CODE (p, "   %s = getelementptr <%d x i16>* %%ptr%d, i32 %%index_n\n",
-      t1, sh, insn->src_args[0]);
-  ORC_ASM_CODE (p, "   %s = load <%d x i16>* %s\n", dest, sh, t1);
+  ORC_ASM_CODE (p, "   %s = getelementptr <%d x i16>, <%d x i16>* %%ptr%d, i32 %%index_n\n",
+      t1, sh, sh, insn->src_args[0]);
+  ORC_ASM_CODE (p, "   %s = load <%d x i16>, <%d x i16>* %s\n", dest, sh, sh, t1);
 
   llvm_store_dst_int (dest, p, insn, insn->dest_args[0]);
 }
@@ -1522,8 +1502,8 @@ llvm_rule_storew (OrcCompiler *p, void *user, OrcInstruction *insn)
 
   llvm_get_dst_int (src, p, insn, insn->src_args[0]);
 
-  ORC_ASM_CODE (p, "   %s = getelementptr <%d x i16>* %%ptr%d, i32 %%index_n\n",
-      t1, sh, insn->dest_args[0]);
+  ORC_ASM_CODE (p, "   %s = getelementptr <%d x i16>, <%d x i16>* %%ptr%d, i32 %%index_n\n",
+      t1, sh, sh, insn->dest_args[0]);
   ORC_ASM_CODE (p, "   store <%d x i16> %s, <%d x i16>* %s\n", sh, src, sh, t1);
 }
 static void
@@ -1737,9 +1717,9 @@ llvm_rule_loadl (OrcCompiler *p, void *user, OrcInstruction *insn)
 
   llvm_get_dst_int (dest, p, insn, insn->dest_args[0]);
 
-  ORC_ASM_CODE (p, "   %s = getelementptr <%d x i32>* %%ptr%d, i32 %%index_n\n",
-      t1, sh, insn->src_args[0]);
-  ORC_ASM_CODE (p, "   %s = load <%d x i32>* %s\n", dest, sh, t1);
+  ORC_ASM_CODE (p, "   %s = getelementptr <%d x i32>, <%d x i32>* %%ptr%d, i32 %%index_n\n",
+      t1, sh, sh, insn->src_args[0]);
+  ORC_ASM_CODE (p, "   %s = load <%d x i32>, <%d x i32>* %s\n", dest, sh, sh, t1);
 
   llvm_store_dst_int (dest, p, insn, insn->dest_args[0]);
 }
@@ -1955,8 +1935,8 @@ llvm_rule_storel (OrcCompiler *p, void *user, OrcInstruction *insn)
 
   llvm_get_dst_int (src, p, insn, insn->src_args[0]);
 
-  ORC_ASM_CODE (p, "   %s = getelementptr <%d x i32>* %%ptr%d, i32 %%index_n\n",
-      t1, sh, insn->dest_args[0]);
+  ORC_ASM_CODE (p, "   %s = getelementptr <%d x i32>, <%d x i32>* %%ptr%d, i32 %%index_n\n",
+      t1, sh, sh, insn->dest_args[0]);
   ORC_ASM_CODE (p, "   store <%d x i32> %s, <%d x i32>* %s\n", sh, src, sh, t1);
 }
 static void
@@ -2026,9 +2006,9 @@ llvm_rule_loadq (OrcCompiler *p, void *user, OrcInstruction *insn)
 
   llvm_get_dst_int (dest, p, insn, insn->dest_args[0]);
 
-  ORC_ASM_CODE (p, "   %s = getelementptr <%d x i64>* %%ptr%d, i32 %%index_n\n",
-      t1, sh, insn->src_args[0]);
-  ORC_ASM_CODE (p, "   %s = load <%d x i64>* %s\n", dest, sh, t1);
+  ORC_ASM_CODE (p, "   %s = getelementptr <%d x i64>, <%d x i64>* %%ptr%d, i32 %%index_n\n",
+      t1, sh, sh, insn->src_args[0]);
+  ORC_ASM_CODE (p, "   %s = load <%d x i64>, <%d x i64>* %s\n", dest, sh, sh, t1);
 
   llvm_store_dst_int (dest, p, insn, insn->dest_args[0]);
 }
@@ -2072,8 +2052,8 @@ llvm_rule_storeq (OrcCompiler *p, void *user, OrcInstruction *insn)
 
   llvm_get_dst_int (src, p, insn, insn->src_args[0]);
 
-  ORC_ASM_CODE (p, "   %s = getelementptr <%d x i64>* %%ptr%d, i32 %%index_n\n",
-      t1, sh, insn->dest_args[0]);
+  ORC_ASM_CODE (p, "   %s = getelementptr <%d x i64>, <%d x i64>* %%ptr%d, i32 %%index_n\n",
+      t1, sh, sh, insn->dest_args[0]);
   ORC_ASM_CODE (p, "   store <%d x i64> %s, <%d x i64>* %s\n", sh, src, sh, t1);
 }
 static void
@@ -2545,9 +2525,39 @@ static void
 llvm_rule_convssslw (OrcCompiler *p, void *user, OrcInstruction *insn)
 {
   char dest[40], src[40];
+  int sh = get_shift (p, insn);
+  char t1[40], t2[40], t3[40], t4[40];
+  const char *mask1, *mask2;
+
+  get_temp (t1, p);
+  get_temp (t2, p);
+  get_temp (t3, p);
+  get_temp (t4, p);
 
   llvm_get_dst_int (dest, p, insn, insn->dest_args[0]);
   llvm_get_src_int (src, p, insn, insn->src_args[0]);
+
+  if (sh == 1) {
+    mask1 = "<i32 32767>";
+    mask2 = "<i32 -32768>";
+  } else if (sh == 2) {
+    mask1 = "<i32 32767, i16 32767>";
+    mask2 = "<i32 -32768, i16 -32768>";
+  } else if (sh == 4) {
+    mask1 = "<i32 32767, i16 32767, i16 32767, i16 32767>";
+    mask2 = "<i32 -32768, i16 -32768, i16 -32768, i16 -32768>";
+  } else {
+    mask1 = "ERROR";
+    mask2 = "ERROR";
+  }
+
+  ORC_ASM_CODE (p, "   %s = icmp sgt <%d x i32> %s, %s\n", t1, sh, src, mask1);
+  ORC_ASM_CODE (p, "   %s = select <%d x i1> %s, <%d x i32> %s, <%d x i32> %s\n", t2, sh, t1, sh, mask1, sh, src);
+  ORC_ASM_CODE (p, "   %s = icmp sle <%d x i32> %s, %s\n", t3, sh, src, mask2);
+  ORC_ASM_CODE (p, "   %s = select <%d x i1> %s, <%d x i32> %s, <%d x i32> %s\n", t4, sh, t3, sh, mask2, sh, t2);
+  ORC_ASM_CODE (p, "   %s = trunc <%d x i32> %s to <%d x i16>\n", dest, sh, t4, sh);
+
+  llvm_store_dst_int (dest, p, insn, insn->dest_args[0]);
 }
 static void
 llvm_rule_convsuslw (OrcCompiler *p, void *user, OrcInstruction *insn)