Add mergelq/select0ql/select1ql opcodes

diff --git a/orc/opcodes.h b/orc/opcodes.h
index 9581e1f..f411f00 100644 (file)
--- a/orc/opcodes.h
+++ b/orc/opcodes.h
@@ -126,12 +126,15 @@ BINARY_BW(mulubw, "(orc_uint8)%s * (orc_uint8)%s")
 BINARY_WL(mulswl, "%s * %s")
 BINARY_WL(muluwl, "(orc_uint16)%s * (orc_uint16)%s")
 
+BINARY_LQ(mergelq, "((orc_uint64)%s & ORC_UINT64_C(0xffffffff)) | ((orc_uint64)%s << 32)")
 BINARY_WL(mergewl, "((orc_uint16)%s) | ((orc_uint16)%s << 16)")
 BINARY_BW(mergebw, "((orc_uint8)%s) | ((orc_uint8)%s << 8)")
 UNARY_WB(select0wb, "(orc_uint16)%s & 0xff")
 UNARY_WB(select1wb, "((orc_uint16)%s >> 8)&0xff")
 UNARY_LW(select0lw, "(orc_uint32)%s & 0xffff")
 UNARY_LW(select1lw, "((orc_uint32)%s >> 16)&0xffff")
+UNARY_QL(select0ql, "(orc_uint64)%s & 0xffffffff")
+UNARY_QL(select1ql, "((orc_uint64)%s >> 32)&0xffffffff")
 UNARY_UW(swapw, "ORC_SWAP_W(%s)")
 UNARY_UL(swapl, "ORC_SWAP_L(%s)")
 UNARY_UQ(swapq, "ORC_SWAP_Q(%s)")

diff --git a/orc/orcemulateopcodes.c b/orc/orcemulateopcodes.c
index 8ad7474..3022304 100644 (file)
--- a/orc/orcemulateopcodes.c
+++ b/orc/orcemulateopcodes.c
@@ -4158,6 +4158,83 @@ emulate_select1lw (OrcOpcodeExecutor *ex, int offset, int n)
 }
 
 void
+emulate_select0ql (OrcOpcodeExecutor *ex, int offset, int n)
+{
+  int i;
+  orc_union32 * ORC_RESTRICT ptr0;
+  const orc_union64 * ORC_RESTRICT ptr4;
+  orc_union64 var32;
+  orc_union32 var33;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union64 *)ex->src_ptrs[0];
+
+
+  for (i = 0; i < n; i++) {
+    /* 0: loadq */
+    var32 = ptr4[i];
+    /* 1: select0ql */
+    var33.i = (orc_uint64)var32.i & 0xffffffff;
+    /* 2: storel */
+    ptr0[i] = var33;
+  }
+
+}
+
+void
+emulate_select1ql (OrcOpcodeExecutor *ex, int offset, int n)
+{
+  int i;
+  orc_union32 * ORC_RESTRICT ptr0;
+  const orc_union64 * ORC_RESTRICT ptr4;
+  orc_union64 var32;
+  orc_union32 var33;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union64 *)ex->src_ptrs[0];
+
+
+  for (i = 0; i < n; i++) {
+    /* 0: loadq */
+    var32 = ptr4[i];
+    /* 1: select1ql */
+    var33.i = ((orc_uint64)var32.i >> 32)&0xffffffff;
+    /* 2: storel */
+    ptr0[i] = var33;
+  }
+
+}
+
+void
+emulate_mergelq (OrcOpcodeExecutor *ex, int offset, int n)
+{
+  int i;
+  orc_union64 * ORC_RESTRICT ptr0;
+  const orc_union32 * ORC_RESTRICT ptr4;
+  const orc_union32 * ORC_RESTRICT ptr5;
+  orc_union32 var32;
+  orc_union32 var33;
+  orc_union64 var34;
+
+  ptr0 = (orc_union64 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+  ptr5 = (orc_union32 *)ex->src_ptrs[1];
+
+
+  for (i = 0; i < n; i++) {
+    /* 0: loadl */
+    var32 = ptr4[i];
+    /* 1: loadl */
+    var33 = ptr5[i];
+    /* 2: mergelq */
+    var34.i = ((orc_uint64)var32.i & ORC_UINT64_C(0xffffffff)) | ((orc_uint64)var33.i << 32);
+    /* 3: storeq */
+    ptr0[i] = var34;
+  }
+
+}
+
+void
 emulate_mergewl (OrcOpcodeExecutor *ex, int offset, int n)
 {
   int i;

diff --git a/orc/orcemulateopcodes.h b/orc/orcemulateopcodes.h
index e0d6fd1..df84191 100644 (file)
--- a/orc/orcemulateopcodes.h
+++ b/orc/orcemulateopcodes.h
@@ -162,6 +162,9 @@ void emulate_select0wb (OrcOpcodeExecutor *ex, int i, int n);
 void emulate_select1wb (OrcOpcodeExecutor *ex, int i, int n);
 void emulate_select0lw (OrcOpcodeExecutor *ex, int i, int n);
 void emulate_select1lw (OrcOpcodeExecutor *ex, int i, int n);
+void emulate_select0ql (OrcOpcodeExecutor *ex, int i, int n);
+void emulate_select1ql (OrcOpcodeExecutor *ex, int i, int n);
+void emulate_mergelq (OrcOpcodeExecutor *ex, int i, int n);
 void emulate_mergewl (OrcOpcodeExecutor *ex, int i, int n);
 void emulate_mergebw (OrcOpcodeExecutor *ex, int i, int n);
 void emulate_splitql (OrcOpcodeExecutor *ex, int i, int n);

diff --git a/orc/orcopcodes.c b/orc/orcopcodes.c
index 8201083..113d342 100644 (file)
--- a/orc/orcopcodes.c
+++ b/orc/orcopcodes.c
@@ -462,6 +462,9 @@ static OrcStaticOpcode opcodes[] = {
   { "select1wb", 0, { 1 }, { 2 }, emulate_select1wb },
   { "select0lw", 0, { 2 }, { 4 }, emulate_select0lw },
   { "select1lw", 0, { 2 }, { 4 }, emulate_select1lw },
+  { "select0ql", 0, { 4 }, { 8 }, emulate_select0ql },
+  { "select1ql", 0, { 4 }, { 8 }, emulate_select1ql },
+  { "mergelq", 0, { 8 }, { 4, 4 }, emulate_mergelq },
   { "mergewl", 0, { 4 }, { 2, 2 }, emulate_mergewl },
   { "mergebw", 0, { 2 }, { 1, 1 }, emulate_mergebw },
   { "splitql", 0, { 4, 4 }, { 8 }, emulate_splitql },

diff --git a/orc/orcprogram-c.c b/orc/orcprogram-c.c
index 7c49ab4..dfed3fe 100644 (file)
--- a/orc/orcprogram-c.c
+++ b/orc/orcprogram-c.c
@@ -685,6 +685,8 @@ c_rule_ ## name (OrcCompiler *p, void *user, OrcInstruction *insn) \
 #define UNARY_UQ(a,b) UNARY(a,b)
 #define BINARY_BW(a,b) BINARY(a,b)
 #define BINARY_WL(a,b) BINARY(a,b)
+#define BINARY_LQ(a,b) BINARY(a,b)
+#define BINARY_QL(a,b) BINARY(a,b)
 #define BINARY_LW(a,b) BINARY(a,b)
 #define BINARY_WB(a,b) BINARY(a,b)
 #define UNARY_BW(a,b) UNARY(a,b)
@@ -732,6 +734,8 @@ c_rule_ ## name (OrcCompiler *p, void *user, OrcInstruction *insn) \
 #undef UNARY_D
 #undef BINARY_BW
 #undef BINARY_WL
+#undef BINARY_LQ
+#undef BINARY_QL
 #undef BINARY_LW
 #undef BINARY_WB
 #undef UNARY_BW
@@ -1327,6 +1331,8 @@ orc_c_init (void)
 #define UNARY_D(a,b) orc_rule_register (rule_set, #a , c_rule_ ## a, NULL);
 #define BINARY_BW(a,b) orc_rule_register (rule_set, #a , c_rule_ ## a, NULL);
 #define BINARY_WL(a,b) orc_rule_register (rule_set, #a , c_rule_ ## a, NULL);
+#define BINARY_LQ(a,b) orc_rule_register (rule_set, #a , c_rule_ ## a, NULL);
+#define BINARY_QL(a,b) orc_rule_register (rule_set, #a , c_rule_ ## a, NULL);
 #define BINARY_LW(a,b) orc_rule_register (rule_set, #a , c_rule_ ## a, NULL);
 #define BINARY_WB(a,b) orc_rule_register (rule_set, #a , c_rule_ ## a, NULL);
 #define UNARY_BW(a,b) orc_rule_register (rule_set, #a , c_rule_ ## a, NULL);