From dfeaefab66b89b4084c8d793bafa04b589634f34 Mon Sep 17 00:00:00 2001
From: David Schleef <ds@schleef.org>
Date: Tue, 31 Mar 2009 22:11:46 -0700
Subject: [PATCH] Allow reading from dest

---
 orc/orccompiler.c    |  4 ++++
 orc/orcprogram-sse.c | 10 +++++++---
 orc/x86.c            | 23 ++++++++++++++++-------
 orc/x86.h            |  2 +-
 4 files changed, 28 insertions(+), 11 deletions(-)

diff --git a/orc/orccompiler.c b/orc/orccompiler.c
index 512d7d8..6db5493 100644
--- a/orc/orccompiler.c
+++ b/orc/orccompiler.c
@@ -149,9 +149,11 @@ orc_compiler_rewrite_vars (OrcCompiler *compiler)
       if (opcode->src_size[k] == 0) continue;
 
       var = insn->src_args[k];
+#if 0
       if (compiler->vars[var].vartype == ORC_VAR_TYPE_DEST) {
         ORC_PROGRAM_ERROR(compiler, "using dest var as source");
       }
+#endif
 
       actual_var = var;
       if (compiler->vars[var].replaced) {
@@ -194,9 +196,11 @@ orc_compiler_rewrite_vars (OrcCompiler *compiler)
         compiler->vars[actual_var].used = TRUE;
         compiler->vars[actual_var].first_use = j;
       } else {
+#if 0
         if (compiler->vars[var].vartype == ORC_VAR_TYPE_DEST) {
           ORC_PROGRAM_ERROR(compiler,"writing dest more than once");
         }
+#endif
         if (compiler->vars[var].vartype == ORC_VAR_TYPE_TEMP) {
           actual_var = orc_compiler_dup_temporary (compiler, var, j);
           compiler->vars[var].replaced = TRUE;
diff --git a/orc/orcprogram-sse.c b/orc/orcprogram-sse.c
index 9af1047..1db7892 100644
--- a/orc/orcprogram-sse.c
+++ b/orc/orcprogram-sse.c
@@ -164,13 +164,16 @@ sse_emit_load_src (OrcCompiler *compiler, OrcVariable *var)
       x86_emit_mov_reg_sse (compiler, X86_ECX, var->alloc);
       break;
     case 4:
-      x86_emit_mov_memoffset_sse (compiler, 4, 0, ptr_reg, var->alloc);
+      x86_emit_mov_memoffset_sse (compiler, 4, 0, ptr_reg, var->alloc,
+          var->is_aligned);
       break;
     case 8:
-      x86_emit_mov_memoffset_sse (compiler, 8, 0, ptr_reg, var->alloc);
+      x86_emit_mov_memoffset_sse (compiler, 8, 0, ptr_reg, var->alloc,
+          var->is_aligned);
       break;
     case 16:
-      x86_emit_mov_memoffset_sse (compiler, 16, 0, ptr_reg, var->alloc);
+      x86_emit_mov_memoffset_sse (compiler, 16, 0, ptr_reg, var->alloc,
+          var->is_aligned);
       break;
     default:
       ORC_PROGRAM_ERROR(compiler,"bad load size %d",
@@ -405,6 +408,7 @@ sse_emit_loop (OrcCompiler *compiler)
 
       switch (var->vartype) {
         case ORC_VAR_TYPE_SRC:
+        case ORC_VAR_TYPE_DEST:
           sse_emit_load_src (compiler, var);
           break;
         case ORC_VAR_TYPE_CONST:
diff --git a/orc/x86.c b/orc/x86.c
index 5f2855b..bda38d7 100644
--- a/orc/x86.c
+++ b/orc/x86.c
@@ -287,7 +287,7 @@ x86_emit_mov_memoffset_mmx (OrcCompiler *compiler, int size, int offset,
 
 void
 x86_emit_mov_memoffset_sse (OrcCompiler *compiler, int size, int offset,
-    int reg1, int reg2)
+    int reg1, int reg2, int is_aligned)
 {
   switch (size) {
     case 4:
@@ -306,12 +306,21 @@ x86_emit_mov_memoffset_sse (OrcCompiler *compiler, int size, int offset,
       *compiler->codeptr++ = 0x7e;
       break;
     case 16:
-      ORC_ASM_CODE(compiler,"  movdqu %d(%%%s), %%%s\n", offset, x86_get_regname_ptr(reg1),
-          x86_get_regname_sse(reg2));
-      *compiler->codeptr++ = 0xf3;
-      x86_emit_rex(compiler, 0, reg2, 0, reg1);
-      *compiler->codeptr++ = 0x0f;
-      *compiler->codeptr++ = 0x6f;
+      if (is_aligned) {
+        ORC_ASM_CODE(compiler,"  movdqa %d(%%%s), %%%s\n", offset, x86_get_regname_ptr(reg1),
+            x86_get_regname_sse(reg2));
+        *compiler->codeptr++ = 0x66;
+        x86_emit_rex(compiler, 0, reg2, 0, reg1);
+        *compiler->codeptr++ = 0x0f;
+        *compiler->codeptr++ = 0x6f;
+      } else {
+        ORC_ASM_CODE(compiler,"  movdqu %d(%%%s), %%%s\n", offset, x86_get_regname_ptr(reg1),
+            x86_get_regname_sse(reg2));
+        *compiler->codeptr++ = 0xf3;
+        x86_emit_rex(compiler, 0, reg2, 0, reg1);
+        *compiler->codeptr++ = 0x0f;
+        *compiler->codeptr++ = 0x6f;
+      }
       break;
     default:
       ORC_PROGRAM_ERROR(compiler, "bad size");
diff --git a/orc/x86.h b/orc/x86.h
index f67fbae..1685f18 100644
--- a/orc/x86.h
+++ b/orc/x86.h
@@ -14,7 +14,7 @@ void x86_emit_mov_memoffset_reg (OrcCompiler *compiler, int size, int offset, in
 void x86_emit_mov_memoffset_mmx (OrcCompiler *compiler, int size, int offset,
     int reg1, int reg2);
 void x86_emit_mov_memoffset_sse (OrcCompiler *compiler, int size, int offset,
-    int reg1, int reg2);
+    int reg1, int reg2, int is_aligned);
 void x86_emit_mov_reg_memoffset (OrcCompiler *compiler, int size, int reg1, int offset, int reg2);
 void x86_emit_mov_mmx_memoffset (OrcCompiler *compiler, int size, int reg1, int offset,
     int reg2);
-- 
2.7.4