Add orc_compiler_try_get_constant_long()

For loading long constants if it happens to be fast.  Use case is
for when there's a backup rule that is almost as fast.  (The
assumption is that loading long constants is slow.)

3 files changed, 79 insertions, 20 deletions

diff --git a/orc/orccompiler.c b/orc/orccompiler.c
index 16b4d76..7d60eb9 100644
--- a/orc/orccompiler.c
+++ b/orc/orccompiler.c
@@ -1081,6 +1081,41 @@ orc_compiler_get_constant_long (OrcCompiler *compiler,
 }
 
 int
+orc_compiler_try_get_constant_long (OrcCompiler *compiler,
+    orc_uint32 a, orc_uint32 b, orc_uint32 c, orc_uint32 d)
+{
+  int i;
+
+  for(i=0;i<compiler->n_constants;i++){
+    if (compiler->constants[i].is_long == TRUE &&
+        compiler->constants[i].full_value[0] == a &&
+        compiler->constants[i].full_value[1] == b &&
+        compiler->constants[i].full_value[2] == c &&
+        compiler->constants[i].full_value[3] == d) {
+      break;
+    }
+  }
+  if (i == compiler->n_constants) {
+    compiler->n_constants++;
+    compiler->constants[i].full_value[0] = a;
+    compiler->constants[i].full_value[1] = b;
+    compiler->constants[i].full_value[2] = c;
+    compiler->constants[i].full_value[3] = d;
+    compiler->constants[i].is_long = TRUE;
+    compiler->constants[i].alloc_reg = 0;
+    compiler->constants[i].use_count = 0;
+  }
+
+  compiler->constants[i].use_count++;
+
+  if (compiler->constants[i].alloc_reg != 0) {;
+    return compiler->constants[i].alloc_reg;
+  }
+  return ORC_REG_INVALID;
+}
+
+
+int
 orc_compiler_get_constant_reg (OrcCompiler *compiler)
 {
   int j;
diff --git a/orc/orcprogram.h b/orc/orcprogram.h
index e74de23..c88ccf2 100644
--- a/orc/orcprogram.h
+++ b/orc/orcprogram.h
@@ -722,6 +722,8 @@ int orc_compiler_label_new (OrcCompiler *compiler);
 int orc_compiler_get_constant (OrcCompiler *compiler, int size, int value);
 int orc_compiler_get_constant_long (OrcCompiler *compiler, orc_uint32 a,
     orc_uint32 b, orc_uint32 c, orc_uint32 d);
+int orc_compiler_try_get_constant_long (OrcCompiler *compiler, orc_uint32 a,
+    orc_uint32 b, orc_uint32 c, orc_uint32 d);
 int orc_compiler_get_temp_constant (OrcCompiler *compiler, int size, int value);
 int orc_compiler_get_temp_reg (OrcCompiler *compiler);
 int orc_compiler_get_constant_reg (OrcCompiler *compiler);
diff --git a/orc/orcrules-sse.c b/orc/orcrules-sse.c
index f880423..841fa28 100644
--- a/orc/orcrules-sse.c
+++ b/orc/orcrules-sse.c
@@ -1550,9 +1550,13 @@ sse_rule_swapw_ssse3 (OrcCompiler *p, void *user, OrcInstruction *insn)
   int dest = p->vars[insn->dest_args[0]].alloc;
   int tmp;
 
-  tmp = orc_compiler_get_constant_long (p,
+  tmp = orc_compiler_try_get_constant_long (p,
       0x02030001, 0x06070405, 0x0a0b0809, 0x0e0f0c0d);
-  orc_sse_emit_pshufb (p, tmp, dest);
+  if (tmp != ORC_REG_INVALID) {
+    orc_sse_emit_pshufb (p, tmp, dest);
+  } else {
+    sse_rule_swapw (p, user, insn);
+  }
 }
 
 static void
@@ -1561,10 +1565,13 @@ sse_rule_swapl_ssse3 (OrcCompiler *p, void *user, OrcInstruction *insn)
   int dest = p->vars[insn->dest_args[0]].alloc;
   int tmp;
 
-  tmp = orc_compiler_get_constant_long (p,
+  tmp = orc_compiler_try_get_constant_long (p,
       0x00010203, 0x04050607, 0x08090a0b, 0x0c0d0e0f);
-
-  orc_sse_emit_pshufb (p, tmp, dest);
+  if (tmp != ORC_REG_INVALID) {
+    orc_sse_emit_pshufb (p, tmp, dest);
+  } else {
+    sse_rule_swapl (p, user, insn);
+  }
 }
 
 static void
@@ -1573,10 +1580,13 @@ sse_rule_swapq_ssse3 (OrcCompiler *p, void *user, OrcInstruction *insn)
   int dest = p->vars[insn->dest_args[0]].alloc;
   int tmp;
 
-  tmp = orc_compiler_get_constant_long (p,
+  tmp = orc_compiler_try_get_constant_long (p,
       0x04050607, 0x00010203, 0x0c0d0e0f, 0x08090a0b);
-
-  orc_sse_emit_pshufb (p, tmp, dest);
+  if (tmp != ORC_REG_INVALID) {
+    orc_sse_emit_pshufb (p, tmp, dest);
+  } else {
+    sse_rule_swapq (p, user, insn);
+  }
 }
 
 static void
@@ -1585,10 +1595,13 @@ sse_rule_select0lw_ssse3 (OrcCompiler *p, void *user, OrcInstruction *insn)
   int dest = p->vars[insn->dest_args[0]].alloc;
   int tmp;
 
-  tmp = orc_compiler_get_constant_long (p,
+  tmp = orc_compiler_try_get_constant_long (p,
       0x05040100, 0x0d0c0908, 0x05040100, 0x0d0c0908);
-
-  orc_sse_emit_pshufb (p, tmp, dest);
+  if (tmp != ORC_REG_INVALID) {
+    orc_sse_emit_pshufb (p, tmp, dest);
+  } else {
+    sse_rule_select0lw (p, user, insn);
+  }
 }
 
 static void
@@ -1597,10 +1610,13 @@ sse_rule_select1lw_ssse3 (OrcCompiler *p, void *user, OrcInstruction *insn)
   int dest = p->vars[insn->dest_args[0]].alloc;
   int tmp;
 
-  tmp = orc_compiler_get_constant_long (p,
+  tmp = orc_compiler_try_get_constant_long (p,
       0x07060302, 0x0f0e0b0a, 0x07060302, 0x0f0e0b0a);
-
-  orc_sse_emit_pshufb (p, tmp, dest);
+  if (tmp != ORC_REG_INVALID) {
+    orc_sse_emit_pshufb (p, tmp, dest);
+  } else {
+    sse_rule_select1lw (p, user, insn);
+  }
 }
 
 static void
@@ -1609,10 +1625,13 @@ sse_rule_select0wb_ssse3 (OrcCompiler *p, void *user, OrcInstruction *insn)
   int dest = p->vars[insn->dest_args[0]].alloc;
   int tmp;
 
-  tmp = orc_compiler_get_constant_long (p,
+  tmp = orc_compiler_try_get_constant_long (p,
       0x06040200, 0x0e0c0a08, 0x06040200, 0x0e0c0a08);
-
-  orc_sse_emit_pshufb (p, tmp, dest);
+  if (tmp != ORC_REG_INVALID) {
+    orc_sse_emit_pshufb (p, tmp, dest);
+  } else {
+    sse_rule_select0wb (p, user, insn);
+  }
 }
 
 static void
@@ -1621,10 +1640,13 @@ sse_rule_select1wb_ssse3 (OrcCompiler *p, void *user, OrcInstruction *insn)
   int dest = p->vars[insn->dest_args[0]].alloc;
   int tmp;
 
-  tmp = orc_compiler_get_constant_long (p,
+  tmp = orc_compiler_try_get_constant_long (p,
       0x07050301, 0x0f0d0b09, 0x07050301, 0x0f0d0b09);
-
-  orc_sse_emit_pshufb (p, tmp, dest);
+  if (tmp != ORC_REG_INVALID) {
+    orc_sse_emit_pshufb (p, tmp, dest);
+  } else {
+    sse_rule_select1wb (p, user, insn);
+  }
 }
 #endif