llvm: updates

1 files changed, 136 insertions, 4 deletions

diff --git a/orc/orcprogram-llvm.c b/orc/orcprogram-llvm.c
index e30b493..ea8726a 100644
--- a/orc/orcprogram-llvm.c
+++ b/orc/orcprogram-llvm.c
@@ -541,9 +541,9 @@ 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, <%d x i64> %%i\n",
+  ORC_ASM_CODE (p, "   %s = getelementptr <%d x i8>* %%ptr%d, <%d x i64> %%i\n",
       t1, sh, insn->src_args[0], sh);
-  ORC_ASM_CODE (p, "   %s = load <%d x i8*> %s\n", dest, sh, t1);
+  ORC_ASM_CODE (p, "   %s = load <%d x i8>* %s\n", dest, sh, t1);
 
   llvm_store_dst_int (dest, p, insn, insn->dest_args[0]);
 }
@@ -745,9 +745,9 @@ 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, <%d x i64> %%i\n",
+  ORC_ASM_CODE (p, "   %s = getelementptr <%d x i8>* %%ptr%d, <%d x i64> %%i\n",
       t1, sh, insn->dest_args[0], sh);
-  ORC_ASM_CODE (p, "   store <%d x i8*> %s, %s\n", sh, src, t1);
+  ORC_ASM_CODE (p, "   store <%d x i8> %s, %s\n", sh, src, t1);
 }
 
 static void
@@ -968,6 +968,18 @@ llvm_rule_divluw (OrcCompiler *p, void *user, OrcInstruction *insn)
 static void
 llvm_rule_loadw (OrcCompiler *p, void *user, OrcInstruction *insn)
 {
+  char dest[40], t1[40];
+  int sh = get_shift (p, insn);
+
+  get_temp (t1, p);
+
+  llvm_get_dst_int (dest, p, insn, insn->dest_args[0]);
+
+  ORC_ASM_CODE (p, "   %s = getelementptr <%d x i16>* %%ptr%d, <%d x i64> %%i\n",
+      t1, sh, insn->src_args[0], sh);
+  ORC_ASM_CODE (p, "   %s = load <%d x i16>* %s\n", dest, sh, t1);
+
+  llvm_store_dst_int (dest, p, insn, insn->dest_args[0]);
 }
 static void
 llvm_rule_loadoffw (OrcCompiler *p, void *user, OrcInstruction *insn)
@@ -1142,6 +1154,16 @@ llvm_rule_signw (OrcCompiler *p, void *user, OrcInstruction *insn)
 static void
 llvm_rule_storew (OrcCompiler *p, void *user, OrcInstruction *insn)
 {
+  char src[40], t1[40];
+  int sh = get_shift (p, insn);
+
+  get_temp (t1, p);
+
+  llvm_get_dst_int (src, p, insn, insn->src_args[0]);
+
+  ORC_ASM_CODE (p, "   %s = getelementptr <%d x i16>* %%ptr%d, <%d x i64> %%i\n",
+      t1, sh, insn->dest_args[0], sh);
+  ORC_ASM_CODE (p, "   store <%d x i16> %s, %s\n", sh, src, t1);
 }
 static void
 llvm_rule_subw (OrcCompiler *p, void *user, OrcInstruction *insn)
@@ -1344,6 +1366,18 @@ llvm_rule_copyl (OrcCompiler *p, void *user, OrcInstruction *insn)
 static void
 llvm_rule_loadl (OrcCompiler *p, void *user, OrcInstruction *insn)
 {
+  char dest[40], t1[40];
+  int sh = get_shift (p, insn);
+
+  get_temp (t1, p);
+
+  llvm_get_dst_int (dest, p, insn, insn->dest_args[0]);
+
+  ORC_ASM_CODE (p, "   %s = getelementptr <%d x i32>* %%ptr%d, <%d x i64> %%i\n",
+      t1, sh, insn->src_args[0], sh);
+  ORC_ASM_CODE (p, "   %s = load <%d x i32>* %s\n", dest, sh, t1);
+
+  llvm_store_dst_int (dest, p, insn, insn->dest_args[0]);
 }
 static void
 llvm_rule_loadoffl (OrcCompiler *p, void *user, OrcInstruction *insn)
@@ -1518,6 +1552,16 @@ llvm_rule_signl (OrcCompiler *p, void *user, OrcInstruction *insn)
 static void
 llvm_rule_storel (OrcCompiler *p, void *user, OrcInstruction *insn)
 {
+  char src[40], t1[40];
+  int sh = get_shift (p, insn);
+
+  get_temp (t1, p);
+
+  llvm_get_dst_int (src, p, insn, insn->src_args[0]);
+
+  ORC_ASM_CODE (p, "   %s = getelementptr <%d x i32>* %%ptr%d, <%d x i64> %%i\n",
+      t1, sh, insn->dest_args[0], sh);
+  ORC_ASM_CODE (p, "   store <%d x i32> %s, %s\n", sh, src, t1);
 }
 static void
 llvm_rule_subl (OrcCompiler *p, void *user, OrcInstruction *insn)
@@ -1579,6 +1623,18 @@ llvm_rule_xorl (OrcCompiler *p, void *user, OrcInstruction *insn)
 static void
 llvm_rule_loadq (OrcCompiler *p, void *user, OrcInstruction *insn)
 {
+  char dest[40], t1[40];
+  int sh = get_shift (p, insn);
+
+  get_temp (t1, p);
+
+  llvm_get_dst_int (dest, p, insn, insn->dest_args[0]);
+
+  ORC_ASM_CODE (p, "   %s = getelementptr <%d x i64>* %%ptr%d, <%d x i64> %%i\n",
+      t1, sh, insn->src_args[0], sh);
+  ORC_ASM_CODE (p, "   %s = load <%d x i64>* %s\n", dest, sh, t1);
+
+  llvm_store_dst_int (dest, p, insn, insn->dest_args[0]);
 }
 static void
 llvm_rule_loadpq (OrcCompiler *p, void *user, OrcInstruction *insn)
@@ -1587,6 +1643,16 @@ llvm_rule_loadpq (OrcCompiler *p, void *user, OrcInstruction *insn)
 static void
 llvm_rule_storeq (OrcCompiler *p, void *user, OrcInstruction *insn)
 {
+  char src[40], t1[40];
+  int sh = get_shift (p, insn);
+
+  get_temp (t1, p);
+
+  llvm_get_dst_int (src, p, insn, insn->src_args[0]);
+
+  ORC_ASM_CODE (p, "   %s = getelementptr <%d x i64>* %%ptr%d, <%d x i64> %%i\n",
+      t1, sh, insn->dest_args[0], sh);
+  ORC_ASM_CODE (p, "   store <%d x i64> %s, %s\n", sh, src, t1);
 }
 static void
 llvm_rule_splatw3q (OrcCompiler *p, void *user, OrcInstruction *insn)
@@ -2280,14 +2346,48 @@ llvm_rule_accsadubl (OrcCompiler *p, void *user, OrcInstruction *insn)
   llvm_store_dst_int (dest, p, insn, insn->dest_args[0]);
 }
 
+static const char *
+get_swap_mask1 (OrcCompiler *p, int sh)
+{
+  const char *mask;
+
+  if (sh == 1)
+    mask = "<2 x i32> <i32 1, i32 0>";
+  else if (sh == 2)
+    mask = "<4 x i32> <i32 1, i32 0, i32 1, i32 0>";
+  else if (sh == 4)
+    mask = "<8 x i32> <i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0>";
+  else
+    mask = "ERROR";
+
+  return mask;
+}
+
 static void
 llvm_rule_swapw (OrcCompiler *p, void *user, OrcInstruction *insn)
 {
   char dest[40], src[40];
+  int sh = get_shift (p, insn);
+  char t1[40], t2[40];
+  const char *mask;
 
   llvm_get_dst_int (dest, p, insn, insn->dest_args[0]);
   llvm_get_src_int (src, p, insn, insn->src_args[0]);
+  get_temp (t1, p);
+  get_temp (t2, p);
+
+  mask = get_swap_mask1 (p, sh);
+
+  ORC_ASM_CODE (p, "   %s = bitcast <%d x i16> %s to <%d x i8>\n",
+      t1, sh, src, sh * 2);
+  ORC_ASM_CODE (p, "   %s = shufflevector <%d x i8> %s, <%d x i8> %s, %s\n",
+      t2, sh * 2, t1, sh * 2, t1, mask);
+  ORC_ASM_CODE (p, "   %s = bitcast <%d x i8> %s to <%d x i16>\n",
+      dest, sh * 2, t2, sh);
+
+  llvm_store_dst_int (dest, p, insn, insn->dest_args[0]);
 }
+
 static void
 llvm_rule_swapl (OrcCompiler *p, void *user, OrcInstruction *insn)
 {
@@ -2300,9 +2400,25 @@ static void
 llvm_rule_swapwl (OrcCompiler *p, void *user, OrcInstruction *insn)
 {
   char dest[40], src[40];
+  int sh = get_shift (p, insn);
+  char t1[40], t2[40];
+  const char *mask;
 
   llvm_get_dst_int (dest, p, insn, insn->dest_args[0]);
   llvm_get_src_int (src, p, insn, insn->src_args[0]);
+  get_temp (t1, p);
+  get_temp (t2, p);
+
+  mask = get_swap_mask1 (p, sh);
+
+  ORC_ASM_CODE (p, "   %s = bitcast <%d x i32> %s to <%d x i16>\n",
+      t1, sh, src, sh * 2);
+  ORC_ASM_CODE (p, "   %s = shufflevector <%d x i16> %s, <%d x i16> %s, %s\n",
+      dest, sh, t1, sh, t1, mask);
+  ORC_ASM_CODE (p, "   %s = bitcast <%d x i16> %s to <%d x i32>\n",
+      dest, sh * 2, t2, sh);
+
+  llvm_store_dst_int (dest, p, insn, insn->dest_args[0]);
 }
 static void
 llvm_rule_swapq (OrcCompiler *p, void *user, OrcInstruction *insn)
@@ -2316,9 +2432,25 @@ static void
 llvm_rule_swaplq (OrcCompiler *p, void *user, OrcInstruction *insn)
 {
   char dest[40], src[40];
+  int sh = get_shift (p, insn);
+  const char *mask;
+  char t1[40], t2[40];
 
   llvm_get_dst_int (dest, p, insn, insn->dest_args[0]);
   llvm_get_src_int (src, p, insn, insn->src_args[0]);
+  get_temp (t1, p);
+  get_temp (t2, p);
+
+  mask = get_swap_mask1 (p, sh);
+
+  ORC_ASM_CODE (p, "   %s = bitcast <%d x i64> %s to <%d x i32>\n",
+      t1, sh, src, sh * 2);
+  ORC_ASM_CODE (p, "   %s = shufflevector <%d x i32> %s, <%d x i32> %s, %s\n",
+      t2, sh, t1, sh, t1, mask);
+  ORC_ASM_CODE (p, "   %s = bitcast <%d x i32> %s to <%d x i64>\n",
+      dest, sh * 2, t2, sh);
+
+  llvm_store_dst_int (dest, p, insn, insn->dest_args[0]);
 }