Indent. Remove GL dependencies. Simplify. Add rtasm instructions.

3 files changed, 1400 insertions, 1351 deletions

diff --git a/src/mesa/pipe/tgsi/exec/tgsi_sse2.c b/src/mesa/pipe/tgsi/exec/tgsi_sse2.c
index 36d9b86f75..359775fdfb 100755
--- a/src/mesa/pipe/tgsi/exec/tgsi_sse2.c
+++ b/src/mesa/pipe/tgsi/exec/tgsi_sse2.c
@@ -1,7 +1,8 @@
 #include "tgsi_platform.h"
 #include "tgsi_core.h"
+#include "x86/rtasm/x86sse.h"
 
-#if defined USE_X86_ASM
+#if defined(USE_X86_ASM) || defined(SLANG_X86)
 
 #define FOR_EACH_CHANNEL( CHAN )\
    for( CHAN = 0; CHAN < 4; CHAN++ )
@@ -181,148 +182,257 @@ emit_tempf(
 }
 
 static void
-emit_temps (struct x86_function *func,
-            unsigned xmm,
-            unsigned vec,
-            unsigned chan)
+emit_temps (
+   struct x86_function *func,
+   unsigned xmm,
+   unsigned vec,
+   unsigned chan )
 {
-    sse_movaps (func,
-                get_temp (vec, chan),
-                make_xmm (xmm));
+   sse_movaps(
+      func,
+      get_temp( vec, chan ),
+      make_xmm( xmm ) );
 }
 
 static void
-emit_addrf( struct x86_function *func,
-	    unsigned xmm,
-	    unsigned vec,
-	    unsigned chan )
+emit_addrf(
+   struct x86_function *func,
+   unsigned xmm,
+   unsigned vec,
+   unsigned chan )
 {
-   emit_tempf( func, xmm, vec + TGSI_EXEC_NUM_TEMPS, chan );
+   emit_tempf(
+      func,
+      xmm,
+      vec + TGSI_EXEC_NUM_TEMPS,
+      chan );
 }
 
 static void
-emit_addrs( struct x86_function *func,
-	    unsigned xmm,
-	    unsigned vec,
-	    unsigned chan )
+emit_addrs(
+   struct x86_function *func,
+   unsigned xmm,
+   unsigned vec,
+   unsigned chan )
 {
-   emit_temps( func, xmm, vec + TGSI_EXEC_NUM_TEMPS, chan );
+   emit_temps(
+      func,
+      xmm,
+      vec + TGSI_EXEC_NUM_TEMPS,
+      chan );
 }
 
 static void
-emit_abs (struct x86_function *func,
-          unsigned xmm)
+emit_abs(
+   struct x86_function *func,
+   unsigned xmm )
 {
-    sse_andps (func,
-               make_xmm (xmm),
-               get_temp (TGSI_EXEC_TEMP_7FFFFFFF_I,
-                         TGSI_EXEC_TEMP_7FFFFFFF_C));
+   sse_andps(
+      func,
+      make_xmm( xmm ),
+      get_temp(
+         TGSI_EXEC_TEMP_7FFFFFFF_I,
+         TGSI_EXEC_TEMP_7FFFFFFF_C ) );
 }
 
 static void
-emit_neg (struct x86_function *func,
-          unsigned xmm)
+emit_neg(
+   struct x86_function *func,
+   unsigned xmm )
 {
-    sse_xorps (func,
-               make_xmm (xmm),
-               get_temp (TGSI_EXEC_TEMP_80000000_I,
-                         TGSI_EXEC_TEMP_80000000_C));
+   sse_xorps(
+      func,
+      make_xmm( xmm ),
+      get_temp(
+         TGSI_EXEC_TEMP_80000000_I,
+         TGSI_EXEC_TEMP_80000000_C ) );
 }
 
 static void
-emit_setsign (struct x86_function *func,
-              unsigned xmm)
+emit_setsign(
+   struct x86_function *func,
+   unsigned xmm )
 {
-    sse_orps (func,
-              make_xmm (xmm),
-              get_temp (TGSI_EXEC_TEMP_80000000_I,
-                        TGSI_EXEC_TEMP_80000000_C));
+   sse_orps(
+      func,
+      make_xmm( xmm ),
+      get_temp(
+         TGSI_EXEC_TEMP_80000000_I,
+         TGSI_EXEC_TEMP_80000000_C ) );
 }
 
 static void
-emit_add (struct x86_function *func,
-          unsigned xmm_dst,
-          unsigned xmm_src)
+emit_add(
+   struct x86_function *func,
+   unsigned xmm_dst,
+   unsigned xmm_src )
 {
-    sse_addps (func,
-               make_xmm (xmm_dst),
-               make_xmm (xmm_src));
+   sse_addps(
+      func,
+      make_xmm( xmm_dst ),
+      make_xmm( xmm_src ) );
 }
 
-static GLfloat g_cos_storage[4 + 3];
+static void
+emit_push_abcd(
+   struct x86_function *func )
+{
+   x86_push(
+      func,
+      x86_make_reg( file_REG32, reg_AX ) );
+   x86_push(
+      func,
+      x86_make_reg( file_REG32, reg_BX ) );
+   x86_push(
+      func,
+      x86_make_reg( file_REG32, reg_CX ) );
+   x86_push(
+      func,
+      x86_make_reg( file_REG32, reg_DX ) );
+}
 
 static void
-cos4f (void)
+emit_pop_dcba(
+   struct x86_function *func )
 {
-    GLfloat *store = (GLfloat *) ALIGN16((unsigned) g_cos_storage);
+   x86_pop(
+      func,
+      x86_make_reg( file_REG32, reg_DX ) );
+   x86_pop(
+      func,
+      x86_make_reg( file_REG32, reg_CX ) );
+   x86_pop(
+      func,
+      x86_make_reg( file_REG32, reg_BX ) );
+   x86_pop(
+      func,
+      x86_make_reg( file_REG32, reg_AX ) );
+}
+
+static void
+emit_func_call1(
+   struct x86_function *func,
+   unsigned xmm_dst,
+   unsigned storage,
+   unsigned char *code )
+{
+   x86_push(
+      func,
+      x86_make_reg( file_REG32, reg_AX ) );
+   x86_mov_reg_imm(
+      func,
+      x86_make_reg( file_REG32, reg_AX ),
+      storage );
+   sse_movaps(
+      func,
+      x86_deref( x86_make_reg( file_REG32, reg_AX ) ),
+      make_xmm( xmm_dst ) );
+   emit_push_abcd(
+      func );
+   x86_call(
+      func,
+      code );
+   emit_pop_dcba(
+      func );
+   sse_movaps(
+      func,
+      make_xmm( xmm_dst ),
+      x86_deref( x86_make_reg( file_REG32, reg_AX ) ) );
+   x86_pop(
+      func,
+      x86_make_reg( file_REG32, reg_AX ) );
+}
+
+static void
+emit_func_call2(
+   struct x86_function *func,
+   unsigned xmm_dst,
+   unsigned xmm_src,
+   unsigned storage,
+   unsigned char *code )
+{
+   x86_push(
+      func,
+      x86_make_reg( file_REG32, reg_AX ) );
+   x86_mov_reg_imm(
+      func,
+      x86_make_reg( file_REG32, reg_AX ),
+      storage );
+   sse_movaps(
+      func,
+      x86_deref( x86_make_reg( file_REG32, reg_AX ) ),
+      make_xmm( xmm_dst ) );
+   sse_movaps(
+      func,
+      x86_make_disp( x86_make_reg( file_REG32, reg_AX ), 16 ),
+      make_xmm( xmm_src ) );
+   emit_push_abcd(
+      func );
+   x86_call(
+      func,
+      code );
+   emit_pop_dcba(
+      func );
+   sse_movaps(
+      func,
+      make_xmm( xmm_dst ),
+      x86_deref( x86_make_reg( file_REG32, reg_AX ) ) );
+   x86_pop(
+      func,
+      x86_make_reg( file_REG32, reg_AX ) );
+}
+
+/* XXX: move into machine context */
+static float g_cos_storage[4 + 3];
+
+static void
+cos4f( void )
+{
+   float *store = (float *) ALIGN16( (unsigned) g_cos_storage );
 
 #ifdef WIN32
-    store[0] = (GLfloat) cos ((GLdouble) store[0]);
-    store[1] = (GLfloat) cos ((GLdouble) store[1]);
-    store[2] = (GLfloat) cos ((GLdouble) store[2]);
-    store[3] = (GLfloat) cos ((GLdouble) store[3]);
+   store[0] = (float) cos( (double) store[0] );
+   store[1] = (float) cos( (double) store[1] );
+   store[2] = (float) cos( (double) store[2] );
+   store[3] = (float) cos( (double) store[3] );
 #else
-    store[0] = cosf (store[0]);
-    store[1] = cosf (store[1]);
-    store[2] = cosf (store[2]);
-    store[3] = cosf (store[3]);
+   store[0] = cosf( store[0] );
+   store[1] = cosf( store[1] );
+   store[2] = cosf( store[2] );
+   store[3] = cosf( store[3] );
 #endif
 }
 
 static void
-emit_cos (struct x86_function *func,
-          unsigned xmm_dst)
+emit_cos(
+   struct x86_function *func,
+   unsigned xmm_dst )
 {
-    x86_push (func,
-              x86_make_reg (file_REG32, reg_AX));
-    x86_mov_reg_imm (func,
-                     x86_make_reg (file_REG32, reg_AX),
-                     ALIGN16((GLint) g_cos_storage));
-    sse_movaps (func,
-                x86_deref (x86_make_reg (file_REG32, reg_AX)),
-                make_xmm (xmm_dst));
-    x86_push (func,
-              x86_make_reg (file_REG32, reg_AX));
-    x86_push (func,
-              x86_make_reg (file_REG32, reg_BX));
-    x86_push (func,
-              x86_make_reg (file_REG32, reg_CX));
-    x86_push (func,
-              x86_make_reg (file_REG32, reg_DX));
-    x86_call (func,
-              (GLubyte *) cos4f);
-    x86_pop (func,
-             x86_make_reg (file_REG32, reg_DX));
-    x86_pop (func,
-             x86_make_reg (file_REG32, reg_CX));
-    x86_pop (func,
-             x86_make_reg (file_REG32, reg_BX));
-    x86_pop (func,
-             x86_make_reg (file_REG32, reg_AX));
-    sse_movaps (func,
-                make_xmm (xmm_dst),
-                x86_deref (x86_make_reg (file_REG32, reg_AX)));
-    x86_pop (func,
-             x86_make_reg (file_REG32, reg_AX));
+   emit_func_call1(
+      func,
+      xmm_dst,
+      ALIGN16( (unsigned) g_cos_storage ),
+      (unsigned char *) cos4f );
 }
 
-static GLfloat g_sin_storage[4 + 3];
+/* XXX: move into machine context */
+static float g_sin_storage[4 + 3];
 
 static void
-sin4f (void)
+sin4f( void )
 {
-    GLfloat *store = (GLfloat *) ALIGN16((unsigned) g_sin_storage);
+   float *store = (float *) ALIGN16( (unsigned) g_sin_storage );
 
 #ifdef WIN32
-    store[0] = (GLfloat) sin ((GLdouble) store[0]);
-    store[1] = (GLfloat) sin ((GLdouble) store[1]);
-    store[2] = (GLfloat) sin ((GLdouble) store[2]);
-    store[3] = (GLfloat) sin ((GLdouble) store[3]);
+   store[0] = (float) sin( (double) store[0] );
+   store[1] = (float) sin( (double) store[1] );
+   store[2] = (float) sin( (double) store[2] );
+   store[3] = (float) sin( (double) store[3] );
 #else
-    store[0] = sin (store[0]);
-    store[1] = sin (store[1]);
-    store[2] = sin (store[2]);
-    store[3] = sin (store[3]);
+   store[0] = sinf( store[0] );
+   store[1] = sinf( store[1] );
+   store[2] = sinf( store[2] );
+   store[3] = sinf( store[3] );
 #endif
 }
 
@@ -330,47 +440,23 @@ static void
 emit_sin (struct x86_function *func,
           unsigned xmm_dst)
 {
-    x86_push (func,
-              x86_make_reg (file_REG32, reg_AX));
-    x86_mov_reg_imm (func,
-                     x86_make_reg (file_REG32, reg_AX),
-                     ALIGN16((GLint) g_sin_storage));
-    sse_movaps (func,
-                x86_deref (x86_make_reg (file_REG32, reg_AX)),
-                make_xmm (xmm_dst));
-    x86_push (func,
-              x86_make_reg (file_REG32, reg_AX));
-    x86_push (func,
-              x86_make_reg (file_REG32, reg_BX));
-    x86_push (func,
-              x86_make_reg (file_REG32, reg_CX));
-    x86_push (func,
-              x86_make_reg (file_REG32, reg_DX));
-    x86_call (func,
-              (GLubyte *) sin4f);
-    x86_pop (func,
-             x86_make_reg (file_REG32, reg_DX));
-    x86_pop (func,
-             x86_make_reg (file_REG32, reg_CX));
-    x86_pop (func,
-             x86_make_reg (file_REG32, reg_BX));
-    x86_pop (func,
-             x86_make_reg (file_REG32, reg_AX));
-    sse_movaps (func,
-                make_xmm (xmm_dst),
-                x86_deref (x86_make_reg (file_REG32, reg_AX)));
-    x86_pop (func,
-             x86_make_reg (file_REG32, reg_AX));
+   emit_func_call1(
+      func,
+      xmm_dst,
+      ALIGN16( (unsigned) g_sin_storage ),
+      (unsigned char *) sin4f );
 }
 
 static void
-emit_mov (struct x86_function *func,
-          unsigned xmm_dst,
-          unsigned xmm_src)
+emit_mov(
+   struct x86_function *func,
+   unsigned xmm_dst,
+   unsigned xmm_src )
 {
-    sse_movups (func,
-                make_xmm (xmm_dst),
-                make_xmm (xmm_src));
+   sse_movups(
+      func,
+      make_xmm( xmm_dst ),
+      make_xmm( xmm_src ) );
 }
 
 static void
@@ -378,1353 +464,1255 @@ emit_mul (struct x86_function *func,
           unsigned xmm_dst,
           unsigned xmm_src)
 {
-    sse_mulps (func,
-               make_xmm (xmm_dst),
-               make_xmm (xmm_src));
+   sse_mulps(
+      func,
+      make_xmm( xmm_dst ),
+      make_xmm( xmm_src ) );
 }
 
-static GLfloat g_pow_storage[4 + 4 + 3];
+/* XXX: move into machine context */
+static float g_pow_storage[4 + 4 + 3];
 
 static void
-pow4f (void)
+pow4f( void )
 {
-    GLfloat *store = (GLfloat *) ALIGN16((unsigned) g_pow_storage);
+   float *store = (float *) ALIGN16( (unsigned) g_pow_storage );
 
 #ifdef WIN32
-    store[0] = (GLfloat) pow ((GLdouble) store[0], (GLdouble) store[4]);
-    store[1] = (GLfloat) pow ((GLdouble) store[1], (GLdouble) store[5]);
-    store[2] = (GLfloat) pow ((GLdouble) store[2], (GLdouble) store[6]);
-    store[3] = (GLfloat) pow ((GLdouble) store[3], (GLdouble) store[7]);
+   store[0] = (float) pow( (double) store[0], (double) store[4] );
+   store[1] = (float) pow( (double) store[1], (double) store[5] );
+   store[2] = (float) pow( (double) store[2], (double) store[6] );
+   store[3] = (float) pow( (double) store[3], (double) store[7] );
 #else
-    store[0] = powf (store[0], store[4]);
-    store[1] = powf (store[1], store[5]);
-    store[2] = powf (store[2], store[6]);
-    store[3] = powf (store[3], store[7]);
+   store[0] = powf( store[0], store[4] );
+   store[1] = powf( store[1], store[5] );
+   store[2] = powf( store[2], store[6] );
+   store[3] = powf( store[3], store[7] );
 #endif
 }
 
 static void
-emit_pow (struct x86_function *func,
-          unsigned xmm_dst,
-          unsigned xmm_src)
+emit_pow(
+   struct x86_function *func,
+   unsigned xmm_dst,
+   unsigned xmm_src )
 {
-    x86_push (func,
-              x86_make_reg (file_REG32, reg_AX));
-    x86_mov_reg_imm (func,
-                     x86_make_reg (file_REG32, reg_AX),
-                     ALIGN16((GLint) g_pow_storage));
-    sse_movaps (func,
-                x86_make_disp (x86_make_reg (file_REG32, reg_AX), 0),
-                make_xmm (xmm_dst));
-    sse_movaps (func,
-                x86_make_disp (x86_make_reg (file_REG32, reg_AX), 16),
-                make_xmm (xmm_src));
-    x86_push (func,
-              x86_make_reg (file_REG32, reg_AX));
-    x86_push (func,
-              x86_make_reg (file_REG32, reg_BX));
-    x86_push (func,
-              x86_make_reg (file_REG32, reg_CX));
-    x86_push (func,
-              x86_make_reg (file_REG32, reg_DX));
-    x86_call (func,
-              (GLubyte *) pow4f);
-    x86_pop (func,
-             x86_make_reg (file_REG32, reg_DX));
-    x86_pop (func,
-             x86_make_reg (file_REG32, reg_CX));
-    x86_pop (func,
-             x86_make_reg (file_REG32, reg_BX));
-    x86_pop (func,
-             x86_make_reg (file_REG32, reg_AX));
-    sse_movaps (func,
-                make_xmm (xmm_dst),
-                x86_deref (x86_make_reg (file_REG32, reg_AX)));
-    x86_pop (func,
-             x86_make_reg (file_REG32, reg_AX));
+   emit_func_call2(
+      func,
+      xmm_dst,
+      xmm_src,
+      ALIGN16( (unsigned) g_pow_storage ),
+      (unsigned char *) pow4f );
 }
 
-static GLfloat g_ex2_storage[4 + 3];
+/* XXX: move into machine context */
+static float g_ex2_storage[4 + 3];
 
 static void
-ex24f (void)
+ex24f( void )
 {
-    GLfloat *store = (GLfloat *) ALIGN16((unsigned) g_ex2_storage);
+   float *store = (float *) ALIGN16( (unsigned) g_ex2_storage );
 
-    store[0] = (GLfloat) pow (2.0, (GLdouble) store[0]);
-    store[1] = (GLfloat) pow (2.0, (GLdouble) store[1]);
-    store[2] = (GLfloat) pow (2.0, (GLdouble) store[2]);
-    store[3] = (GLfloat) pow (2.0, (GLdouble) store[3]);
+#ifdef WIN32
+   store[0] = (float) pow( 2.0, (double) store[0] );
+   store[1] = (float) pow( 2.0, (double) store[1] );
+   store[2] = (float) pow( 2.0, (double) store[2] );
+   store[3] = (float) pow( 2.0, (double) store[3] );
+#else
+   store[0] = powf( 2.0f, store[0] );
+   store[1] = powf( 2.0f, store[1] );
+   store[2] = powf( 2.0f, store[2] );
+   store[3] = powf( 2.0f, store[3] );
+#endif
 }
 
 static void
-emit_ex2 (struct x86_function *func,
-          unsigned xmm_dst)
+emit_ex2(
+   struct x86_function *func,
+   unsigned xmm_dst )
 {
-    x86_push (func,
-              x86_make_reg (file_REG32, reg_AX));
-    x86_mov_reg_imm (func,
-                     x86_make_reg (file_REG32, reg_AX),
-                     ALIGN16((GLint) g_ex2_storage));
-    sse_movaps (func,
-                x86_deref (x86_make_reg (file_REG32, reg_AX)),
-                make_xmm (xmm_dst));
-    x86_push (func,
-              x86_make_reg (file_REG32, reg_AX));
-    x86_push (func,
-              x86_make_reg (file_REG32, reg_BX));
-    x86_push (func,
-              x86_make_reg (file_REG32, reg_CX));
-    x86_push (func,
-              x86_make_reg (file_REG32, reg_DX));
-    x86_call (func,
-              (GLubyte *) ex24f);
-    x86_pop (func,
-             x86_make_reg (file_REG32, reg_DX));
-    x86_pop (func,
-             x86_make_reg (file_REG32, reg_CX));
-    x86_pop (func,
-             x86_make_reg (file_REG32, reg_BX));
-    x86_pop (func,
-             x86_make_reg (file_REG32, reg_AX));
-    sse_movaps (func,
-                make_xmm (xmm_dst),
-                x86_deref (x86_make_reg (file_REG32, reg_AX)));
-    x86_pop (func,
-             x86_make_reg (file_REG32, reg_AX));
+   emit_func_call1(
+      func,
+      xmm_dst,
+      ALIGN16( (unsigned) g_ex2_storage ),
+      (unsigned char *) ex24f );
 }
 
-static GLfloat g_lg2_storage[4 + 3];
+/* XXX: move into machine context */
+static float g_lg2_storage[4 + 3];
 
 static void
-lg24f (void)
+lg24f( void )
 {
-    GLfloat *store = (GLfloat *) ALIGN16((unsigned) g_lg2_storage);
+   float *store = (float *) ALIGN16( (unsigned) g_lg2_storage );
 
-    store[0] = LOG2 (store[0]);
-    store[1] = LOG2 (store[1]);
-    store[2] = LOG2 (store[2]);
-    store[3] = LOG2 (store[3]);
+   store[0] = LOG2( store[0] );
+   store[1] = LOG2( store[1] );
+   store[2] = LOG2( store[2] );
+   store[3] = LOG2( store[3] );
 }
 
 static void
-emit_lg2 (struct x86_function *func,
-          unsigned xmm_dst)
+emit_lg2(
+   struct x86_function *func,
+   unsigned xmm_dst )
 {
-    x86_push (func,
-              x86_make_reg (file_REG32, reg_AX));
-    x86_mov_reg_imm (func,
-                     x86_make_reg (file_REG32, reg_AX),
-                     ALIGN16((GLint) g_lg2_storage));
-    sse_movaps (func,
-                x86_deref (x86_make_reg (file_REG32, reg_AX)),
-                make_xmm (xmm_dst));
-    x86_push (func,
-              x86_make_reg (file_REG32, reg_AX));
-    x86_push (func,
-              x86_make_reg (file_REG32, reg_BX));
-    x86_push (func,
-              x86_make_reg (file_REG32, reg_CX));
-    x86_push (func,
-              x86_make_reg (file_REG32, reg_DX));
-    x86_call (func,
-              (GLubyte *) lg24f);
-    x86_pop (func,
-             x86_make_reg (file_REG32, reg_DX));
-    x86_pop (func,
-             x86_make_reg (file_REG32, reg_CX));
-    x86_pop (func,
-             x86_make_reg (file_REG32, reg_BX));
-    x86_pop (func,
-             x86_make_reg (file_REG32, reg_AX));
-    sse_movaps (func,
-                make_xmm (xmm_dst),
-                x86_deref (x86_make_reg (file_REG32, reg_AX)));
-    x86_pop (func,
-             x86_make_reg (file_REG32, reg_AX));
+   emit_func_call1(
+      func,
+      xmm_dst,
+      ALIGN16( (unsigned) g_lg2_storage ),
+      (unsigned char *) lg24f );
 }
 
-static GLfloat g_flr_storage[4 + 3];
+/* XXX: move into machine context */
+static float g_flr_storage[4 + 3];
 
 static void
-flr4f (void)
+flr4f( void )
 {
-    GLfloat *store = (GLfloat *) ALIGN16((unsigned) g_flr_storage);
+   float *store = (float *) ALIGN16( (unsigned) g_flr_storage );
 
-    store[0] = (GLfloat) floor ((GLdouble) store[0]);
-    store[1] = (GLfloat) floor ((GLdouble) store[1]);
-    store[2] = (GLfloat) floor ((GLdouble) store[2]);
-    store[3] = (GLfloat) floor ((GLdouble) store[3]);
+   store[0] = (float) floor( (double) store[0] );
+   store[1] = (float) floor( (double) store[1] );
+   store[2] = (float) floor( (double) store[2] );
+   store[3] = (float) floor( (double) store[3] );
 }
 
 static void
-emit_flr (struct x86_function *func,
-          unsigned xmm_dst)
+emit_flr(
+   struct x86_function *func,
+   unsigned xmm_dst )
 {
-    x86_push (func,
-              x86_make_reg (file_REG32, reg_AX));
-    x86_mov_reg_imm (func,
-                     x86_make_reg (file_REG32, reg_AX),
-                     ALIGN16((GLint) g_flr_storage));
-    sse_movaps (func,
-                x86_deref (x86_make_reg (file_REG32, reg_AX)),
-                make_xmm (xmm_dst));
-    x86_push (func,
-              x86_make_reg (file_REG32, reg_AX));
-    x86_push (func,
-              x86_make_reg (file_REG32, reg_BX));
-    x86_push (func,
-              x86_make_reg (file_REG32, reg_CX));
-    x86_push (func,
-              x86_make_reg (file_REG32, reg_DX));
-    x86_call (func,
-              (GLubyte *) flr4f);
-    x86_pop (func,
-             x86_make_reg (file_REG32, reg_DX));
-    x86_pop (func,
-             x86_make_reg (file_REG32, reg_CX));
-    x86_pop (func,
-             x86_make_reg (file_REG32, reg_BX));
-    x86_pop (func,
-             x86_make_reg (file_REG32, reg_AX));
-    sse_movaps (func,
-                make_xmm (xmm_dst),
-                x86_deref (x86_make_reg (file_REG32, reg_AX)));
-    x86_pop (func,
-             x86_make_reg (file_REG32, reg_AX));
+   emit_func_call1(
+      func,
+      xmm_dst,
+      ALIGN16( (unsigned) g_flr_storage ),
+      (unsigned char *) flr4f );
 }
 
-static GLfloat g_frc_storage[4 + 3];
+/* XXX: move into machine context */
+static float g_frc_storage[4 + 3];
 
 static void
-frc4f (void)
+frc4f( void )
 {
-    GLfloat *store = (GLfloat *) ALIGN16((unsigned) g_frc_storage);
+   float *store = (float *) ALIGN16( (unsigned) g_frc_storage );
 
-    store[0] -= (GLfloat) floor ((GLdouble) store[0]);
-    store[1] -= (GLfloat) floor ((GLdouble) store[1]);
-    store[2] -= (GLfloat) floor ((GLdouble) store[2]);
-    store[3] -= (GLfloat) floor ((GLdouble) store[3]);
+   store[0] -= (float) floor( (double) store[0] );
+   store[1] -= (float) floor( (double) store[1] );
+   store[2] -= (float) floor( (double) store[2] );
+   store[3] -= (float) floor( (double) store[3] );
 }
 
 static void
-emit_frc (struct x86_function *func,
-          unsigned xmm_dst)
+emit_frc(
+   struct x86_function *func,
+   unsigned xmm_dst )
 {
-    x86_push (func,
-              x86_make_reg (file_REG32, reg_AX));
-    x86_mov_reg_imm (func,
-                     x86_make_reg (file_REG32, reg_AX),
-                     ALIGN16((GLint) g_frc_storage));
-    sse_movaps (func,
-                x86_deref (x86_make_reg (file_REG32, reg_AX)),
-                make_xmm (xmm_dst));
-    x86_push (func,
-              x86_make_reg (file_REG32, reg_AX));
-    x86_push (func,
-              x86_make_reg (file_REG32, reg_BX));
-    x86_push (func,
-              x86_make_reg (file_REG32, reg_CX));
-    x86_push (func,
-              x86_make_reg (file_REG32, reg_DX));
-    x86_call (func,
-              (GLubyte *) frc4f);
-    x86_pop (func,
-             x86_make_reg (file_REG32, reg_DX));
-    x86_pop (func,
-             x86_make_reg (file_REG32, reg_CX));
-    x86_pop (func,
-             x86_make_reg (file_REG32, reg_BX));
-    x86_pop (func,
-             x86_make_reg (file_REG32, reg_AX));
-    sse_movaps (func,
-                make_xmm (xmm_dst),
-                x86_deref (x86_make_reg (file_REG32, reg_AX)));
-    x86_pop (func,
-             x86_make_reg (file_REG32, reg_AX));
+   emit_func_call1(
+      func,
+      xmm_dst,
+      ALIGN16( (unsigned) g_frc_storage ),
+      (unsigned char *) frc4f );
 }
 
 static void
-emit_rcp (struct x86_function *func,
-          unsigned xmm_dst,
-          unsigned xmm_src)
+emit_rcp (
+   struct x86_function *func,
+   unsigned xmm_dst,
+   unsigned xmm_src )
 {
-    sse2_rcpps (func,
-                make_xmm (xmm_dst),
-                make_xmm (xmm_src));
+   sse2_rcpps(
+      func,
+      make_xmm( xmm_dst ),
+      make_xmm( xmm_src ) );
 }
 
 static void
-emit_rsqrt (struct x86_function *func,
-            unsigned xmm_dst,
-            unsigned xmm_src)
+emit_rsqrt(
+   struct x86_function *func,
+   unsigned xmm_dst,
+   unsigned xmm_src )
 {
-    sse_rsqrtps (func,
-                 make_xmm (xmm_dst),
-                 make_xmm (xmm_src));
+   sse_rsqrtps(
+      func,
+      make_xmm( xmm_dst ),
+      make_xmm( xmm_src ) );
 }
 
 static void
-emit_sub (struct x86_function *func,
-          unsigned xmm_dst,
-          unsigned xmm_src)
+emit_sub(
+   struct x86_function *func,
+   unsigned xmm_dst,
+   unsigned xmm_src )
 {
-    sse_subps (func,
-               make_xmm (xmm_dst),
-               make_xmm (xmm_src));
+   sse_subps(
+      func,
+      make_xmm( xmm_dst ),
+      make_xmm( xmm_src ) );
 }
 
 static void
-emit_fetch (struct x86_function *func,
-            unsigned xmm,
-            const struct tgsi_full_src_register *reg,
-            const unsigned chan_index)
+emit_fetch(
+   struct x86_function *func,
+   unsigned xmm,
+   const struct tgsi_full_src_register *reg,
+   const unsigned chan_index )
 {
-    unsigned swizzle = tgsi_util_get_full_src_register_extswizzle (reg, chan_index);
-
-    switch (swizzle)
-    {
-    case TGSI_EXTSWIZZLE_X:
-    case TGSI_EXTSWIZZLE_Y:
-    case TGSI_EXTSWIZZLE_Z:
-    case TGSI_EXTSWIZZLE_W:
-        switch (reg->SrcRegister.File)
-        {
-        case TGSI_FILE_CONSTANT:
-            emit_const (func, xmm, reg->SrcRegister.Index, swizzle);
-            break;
-
-        case TGSI_FILE_INPUT:
-            emit_input (func, xmm, reg->SrcRegister.Index, swizzle);
-            break;
-
-        case TGSI_FILE_TEMPORARY:
-            emit_tempf (func, xmm, reg->SrcRegister.Index, swizzle);
-            break;
-
-        default:
-            assert (0);
-        }
-        break;
-
-    case TGSI_EXTSWIZZLE_ZERO:
-        emit_tempf (func,
-                    xmm,
-                    TGSI_EXEC_TEMP_00000000_I,
-                    TGSI_EXEC_TEMP_00000000_C);
-        break;
-
-    case TGSI_EXTSWIZZLE_ONE:
-        emit_tempf (func,
-                    xmm,
-                    TGSI_EXEC_TEMP_ONE_I,
-                    TGSI_EXEC_TEMP_ONE_C);
-        break;
-
-    default:
-        assert (0);
-    }
-
-    switch (tgsi_util_get_full_src_register_sign_mode (reg, chan_index))
-    {
-    case TGSI_UTIL_SIGN_CLEAR:
-        emit_abs (func, xmm);
-        break;
-
-    case TGSI_UTIL_SIGN_SET:
-        emit_setsign (func, xmm);
-        break;
-
-    case TGSI_UTIL_SIGN_TOGGLE:
-        emit_neg (func, xmm);
-        break;
-
-    case TGSI_UTIL_SIGN_KEEP:
-        break;
-    }
+   unsigned swizzle = tgsi_util_get_full_src_register_extswizzle( reg, chan_index );
+
+   switch( swizzle ) {
+   case TGSI_EXTSWIZZLE_X:
+   case TGSI_EXTSWIZZLE_Y:
+   case TGSI_EXTSWIZZLE_Z:
+   case TGSI_EXTSWIZZLE_W:
+      switch( reg->SrcRegister.File ) {
+      case TGSI_FILE_CONSTANT:
+         emit_const(
+            func,
+            xmm,
+            reg->SrcRegister.Index,
+            swizzle );
+         break;
+
+      case TGSI_FILE_INPUT:
+         emit_input(
+            func,
+            xmm,
+            reg->SrcRegister.Index,
+            swizzle );
+         break;
+
+      case TGSI_FILE_TEMPORARY:
+         emit_tempf(
+            func,
+            xmm,
+            reg->SrcRegister.Index,
+            swizzle );
+         break;
+
+      default:
+         assert( 0 );
+      }
+      break;
+
+   case TGSI_EXTSWIZZLE_ZERO:
+      emit_tempf(
+         func,
+         xmm,
+         TGSI_EXEC_TEMP_00000000_I,
+         TGSI_EXEC_TEMP_00000000_C );
+      break;
+
+   case TGSI_EXTSWIZZLE_ONE:
+      emit_tempf(
+         func,
+         xmm,
+         TGSI_EXEC_TEMP_ONE_I,
+         TGSI_EXEC_TEMP_ONE_C );
+      break;
+
+   default:
+      assert( 0 );
+   }
+
+   switch( tgsi_util_get_full_src_register_sign_mode( reg, chan_index ) ) {
+   case TGSI_UTIL_SIGN_CLEAR:
+      emit_abs( func, xmm );
+      break;
+
+   case TGSI_UTIL_SIGN_SET:
+      emit_setsign( func, xmm );
+      break;
+
+   case TGSI_UTIL_SIGN_TOGGLE:
+      emit_neg( func, xmm );
+      break;
+
+   case TGSI_UTIL_SIGN_KEEP:
+      break;
+   }
 }
 
 static void
-emit_store (struct x86_function *func,
-            unsigned xmm,
-            const struct tgsi_full_dst_register *reg,
-            const struct tgsi_full_instruction *inst,
-            unsigned chan_index)
+emit_store(
+   struct x86_function *func,
+   unsigned xmm,
+   const struct tgsi_full_dst_register *reg,
+   const struct tgsi_full_instruction *inst,
+   unsigned chan_index )
 {
-    switch (reg->DstRegister.File)
-    {
-    case TGSI_FILE_OUTPUT:
-        emit_output (func, xmm, reg->DstRegister.Index, chan_index);
-        break;
+   switch( reg->DstRegister.File ) {
+   case TGSI_FILE_OUTPUT:
+      emit_output(
+         func,
+         xmm,
+         reg->DstRegister.Index,
+         chan_index );
+      break;
 
-    case TGSI_FILE_TEMPORARY:
-        emit_temps (func, xmm, reg->DstRegister.Index, chan_index);
-        break;
+   case TGSI_FILE_TEMPORARY:
+      emit_temps(
+         func,
+         xmm,
+         reg->DstRegister.Index,
+         chan_index );
+      break;
 
    case TGSI_FILE_ADDRESS:
-      emit_addrs( func, xmm, reg->DstRegister.Index, chan_index );
+      emit_addrs(
+         func,
+         xmm,
+         reg->DstRegister.Index,
+         chan_index );
       break;
 
-    default:
-        assert (0);
-    }
+   default:
+      assert( 0 );
+   }
 
-    switch (inst->Instruction.Saturate)
-    {
-    case TGSI_SAT_NONE:
-        break;
+   switch( inst->Instruction.Saturate ) {
+   case TGSI_SAT_NONE:
+      break;
 
-    case TGSI_SAT_ZERO_ONE:
-//        assert (0);
-        break;
+   case TGSI_SAT_ZERO_ONE:
+//      assert( 0 );
+      break;
 
-    case TGSI_SAT_MINUS_PLUS_ONE:
-        assert (0);
-        break;
-    }
+   case TGSI_SAT_MINUS_PLUS_ONE:
+      assert( 0 );
+      break;
+   }
 }
 
-#define FETCH(XMM,INDEX,CHAN)\
-    emit_fetch (func, XMM, &inst->FullSrcRegisters[INDEX], CHAN)
-
-#define STORE(XMM,INDEX,CHAN)\
-    emit_store (func, XMM, &inst->FullDstRegisters[INDEX], inst, CHAN)
-
 static void
-emit_kil (struct x86_function *func,
-          const struct tgsi_full_src_register *reg)
+emit_kil(
+   struct x86_function *func,
+   const struct tgsi_full_src_register *reg )
 {
-    unsigned uniquemask;
-    unsigned registers[4];
-    unsigned nextregister = 0;
-    unsigned firstchan = ~0;
-    unsigned chan_index;
-
-    /* This mask stores component bits that were already tested. Note that
-     * we test if the value is less than zero, so 1.0 and 0.0 need not to be
-     * tested. */
-    uniquemask = (1 << TGSI_EXTSWIZZLE_ZERO) | (1 << TGSI_EXTSWIZZLE_ONE);
-
-    for (chan_index = 0; chan_index < 4; chan_index++)
-    {
-        unsigned swizzle;
-
-        /* unswizzle channel */
-        swizzle = tgsi_util_get_full_src_register_extswizzle (reg, chan_index);
-
-        /* check if the component has not been already tested */
-        if (!(uniquemask & (1 << swizzle)))
-        {
-            uniquemask |= 1 << swizzle;
-
-            /* allocate register */
-            registers[chan_index] = nextregister;
-            emit_fetch (func, nextregister, reg, chan_index);
-            nextregister++;
-
-            /* mark the first channel used */
-            if (firstchan == ~0)
-                firstchan = chan_index;
-        }
-    }
-
-    x86_push (func,
-              x86_make_reg (file_REG32, reg_AX));
-    x86_push (func,
-              x86_make_reg (file_REG32, reg_DX));
-
-    for (chan_index = 0; chan_index < 4; chan_index++)
-    {
-        if (uniquemask & (1 << chan_index))
-        {
-            sse_cmpps (func,
-                       make_xmm (registers[chan_index]),
-                       get_temp (TGSI_EXEC_TEMP_00000000_I,
-                                 TGSI_EXEC_TEMP_00000000_C),
-                       cc_LessThan);
-
-            if (chan_index == firstchan)
-            {
-                sse_pmovmskb (func,
-                              x86_make_reg (file_REG32, reg_AX),
-                              make_xmm (registers[chan_index]));
-            }
-            else
-            {
-                sse_pmovmskb (func,
-                              x86_make_reg (file_REG32, reg_DX),
-                              make_xmm (registers[chan_index]));
-                x86_or (func,
-                        x86_make_reg (file_REG32, reg_AX),
-                        x86_make_reg (file_REG32, reg_DX));
-            }
-        }
-    }
-
-    x86_or (func,
-            get_temp(TGSI_EXEC_TEMP_KILMASK_I,
-                     TGSI_EXEC_TEMP_KILMASK_C),
-            x86_make_reg (file_REG32, reg_AX));
-
-    x86_pop (func,
-             x86_make_reg (file_REG32, reg_DX));
-    x86_pop (func,
-             x86_make_reg (file_REG32, reg_AX));
-}
+   unsigned uniquemask;
+   unsigned registers[4];
+   unsigned nextregister = 0;
+   unsigned firstchan = ~0;
+   unsigned chan_index;
+
+   /* This mask stores component bits that were already tested. Note that
+    * we test if the value is less than zero, so 1.0 and 0.0 need not to be
+    * tested. */
+   uniquemask = (1 << TGSI_EXTSWIZZLE_ZERO) | (1 << TGSI_EXTSWIZZLE_ONE);
+
+   FOR_EACH_CHANNEL( chan_index ) {
+      unsigned swizzle;
+
+      /* unswizzle channel */
+      swizzle = tgsi_util_get_full_src_register_extswizzle(
+         reg,
+         chan_index );
+
+      /* check if the component has not been already tested */
+      if( !(uniquemask & (1 << swizzle)) ) {
+         uniquemask |= 1 << swizzle;
+
+         /* allocate register */
+         registers[chan_index] = nextregister;
+         emit_fetch(
+            func,
+            nextregister,
+            reg,
+            chan_index );
+         nextregister++;
+
+         /* mark the first channel used */
+         if( firstchan == ~0 ) {
+            firstchan = chan_index;
+         }
+      }
+   }
 
-static void
-emit_setcc (struct x86_function *func,
-            struct tgsi_full_instruction *inst,
-            enum sse_cc cc)
-{
-    unsigned chan_index;
+   x86_push(
+      func,
+      x86_make_reg( file_REG32, reg_AX ) );
+   x86_push(
+      func,
+      x86_make_reg( file_REG32, reg_DX ) );
+
+   FOR_EACH_CHANNEL( chan_index ) {
+      if( uniquemask & (1 << chan_index) ) {
+         sse_cmpps(
+            func,
+            make_xmm( registers[chan_index] ),
+            get_temp(
+               TGSI_EXEC_TEMP_00000000_I,
+               TGSI_EXEC_TEMP_00000000_C ),
+            cc_LessThan );
+
+         if( chan_index == firstchan ) {
+            sse_pmovmskb(
+               func,
+               x86_make_reg( file_REG32, reg_AX ),
+               make_xmm( registers[chan_index] ) );
+         }
+         else {
+            sse_pmovmskb(
+               func,
+               x86_make_reg( file_REG32, reg_DX ),
+               make_xmm( registers[chan_index] ) );
+            x86_or(
+               func,
+               x86_make_reg( file_REG32, reg_AX ),
+               x86_make_reg( file_REG32, reg_DX ) );
+         }
+      }
+   }
+
+   x86_or(
+      func,
+      get_temp(
+         TGSI_EXEC_TEMP_KILMASK_I,
+         TGSI_EXEC_TEMP_KILMASK_C ),
+      x86_make_reg( file_REG32, reg_AX ) );
 
-    FOR_EACH_ENABLED_CHANNEL
-    {
-        emit_fetch (func, 0, &inst->FullSrcRegisters[0], chan_index);
-        emit_fetch (func, 1, &inst->FullSrcRegisters[1], chan_index);
+   x86_pop(
+      func,
+      x86_make_reg( file_REG32, reg_DX ) );
+   x86_pop(
+      func,
+      x86_make_reg( file_REG32, reg_AX ) );
+}
 
-        sse_cmpps (func,
-                   make_xmm (0),
-                   make_xmm (1),
-                   cc);
+#define FETCH( FUNC, INST, XMM, INDEX, CHAN )\
+   emit_fetch( FUNC, XMM, &(INST).FullSrcRegisters[INDEX], CHAN )
 
-        sse_andps (func,
-                   make_xmm (0),
-                   get_temp (TGSI_EXEC_TEMP_ONE_I,
-                             TGSI_EXEC_TEMP_ONE_C));
+#define STORE( FUNC, INST, XMM, INDEX, CHAN )\
+   emit_store( FUNC, XMM, &(INST).FullDstRegisters[INDEX], &(INST), CHAN )
 
-        emit_store (func, 0, &inst->FullDstRegisters[0], inst, chan_index);
-    }
+static void
+emit_setcc(
+   struct x86_function *func,
+   struct tgsi_full_instruction *inst,
+   enum sse_cc cc )
+{
+   unsigned chan_index;
+
+   FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
+      FETCH( func, *inst, 0, 0, chan_index );
+      FETCH( func, *inst, 1, 1, chan_index );
+      sse_cmpps(
+         func,
+         make_xmm( 0 ),
+         make_xmm( 1 ),
+         cc );
+      sse_andps(
+         func,
+         make_xmm( 0 ),
+         get_temp(
+            TGSI_EXEC_TEMP_ONE_I,
+            TGSI_EXEC_TEMP_ONE_C ) );
+      STORE( func, *inst, 0, 0, chan_index );
+   }
 }
 
 static void
-emit_cmp (struct x86_function *func,
-          struct tgsi_full_instruction *inst)
+emit_cmp(
+   struct x86_function *func,
+   struct tgsi_full_instruction *inst )
 {
-    unsigned chan_index;
-
-    FOR_EACH_ENABLED_CHANNEL
-    {
-        emit_fetch (func, 0, &inst->FullSrcRegisters[0], chan_index);
-        emit_fetch (func, 1, &inst->FullSrcRegisters[1], chan_index);
-        emit_fetch (func, 2, &inst->FullSrcRegisters[2], chan_index);
-
-        sse_cmpps (func,
-                   make_xmm (0),
-                   get_temp (TGSI_EXEC_TEMP_00000000_I,
-                             TGSI_EXEC_TEMP_00000000_C),
-                   cc_LessThan);
-
-        sse_andps (func,
-                   make_xmm (1),
-                   make_xmm (0));
-        sse_andnps (func,
-                    make_xmm (0),
-                    make_xmm (2));
-        sse_orps (func,
-                  make_xmm (0),
-                  make_xmm (1));
-
-        emit_store (func, 0, &inst->FullDstRegisters[0], inst, chan_index);
-    }
+   unsigned chan_index;
+
+   FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
+      FETCH( func, *inst, 0, 0, chan_index );
+      FETCH( func, *inst, 1, 1, chan_index );
+      FETCH( func, *inst, 2, 2, chan_index );
+      sse_cmpps(
+         func,
+         make_xmm( 0 ),
+         get_temp(
+            TGSI_EXEC_TEMP_00000000_I,
+            TGSI_EXEC_TEMP_00000000_C ),
+         cc_LessThan );
+      sse_andps(
+         func,
+         make_xmm( 1 ),
+         make_xmm( 0 ) );
+      sse_andnps(
+         func,
+         make_xmm( 0 ),
+         make_xmm( 2 ) );
+      sse_orps(
+         func,
+         make_xmm( 0 ),
+         make_xmm( 1 ) );
+      STORE( func, *inst, 0, 0, chan_index );
+   }
 }
 
 static void
-emit_f2it( struct x86_function *func,
-	   unsigned xmm )
+emit_f2it(
+   struct x86_function *func,
+   unsigned xmm )
 {
-   sse2_cvttps2dq( func, make_xmm( xmm ), make_xmm( xmm ) );
+   sse2_cvttps2dq(
+      func,
+      make_xmm( xmm ),
+      make_xmm( xmm ) );
 }
 
 static void
-emit_instruction (struct x86_function *func,
-                  struct tgsi_full_instruction *inst)
+emit_instruction(
+   struct x86_function *func,
+   struct tgsi_full_instruction *inst )
 {
-    unsigned chan_index;
+   unsigned chan_index;
 
-   switch (inst->Instruction.Opcode) {
+   switch( inst->Instruction.Opcode ) {
    case TGSI_OPCODE_ARL:
-      FOR_EACH_ENABLED_CHANNEL {
-	 FETCH( 0, 0, chan_index );
-	 emit_f2it( func, 0 );
-	 STORE( 0, 0, chan_index );
+      FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( func, *inst, 0, 0, chan_index );
+         emit_f2it( func, 0 );
+         STORE( func, *inst, 0, 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_MOV:
+   /* TGSI_OPCODE_SWZ */
+      FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( func, *inst, 0, 0, chan_index );
+         STORE( func, *inst, 0, 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_LIT:
+      if( IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) ||
+          IS_DST0_CHANNEL_ENABLED( *inst, CHAN_W ) ) {
+         emit_tempf(
+            func,
+            0,
+            TGSI_EXEC_TEMP_ONE_I,
+            TGSI_EXEC_TEMP_ONE_C);
+         if( IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) ) {
+            STORE( func, *inst, 0, 0, CHAN_X );
+         }
+         if( IS_DST0_CHANNEL_ENABLED( *inst, CHAN_W ) ) {
+            STORE( func, *inst, 0, 0, CHAN_W );
+         }
+      }
+      if( IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y ) ||
+          IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Z ) ) {
+         if( IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y ) ) {
+            FETCH( func, *inst, 0, 0, CHAN_X );
+            sse_maxps(
+               func,
+               make_xmm( 0 ),
+               get_temp(
+                  TGSI_EXEC_TEMP_00000000_I,
+                  TGSI_EXEC_TEMP_00000000_C ) );
+            STORE( func, *inst, 0, 0, CHAN_Y );
+         }
+         if( IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Z ) ) {
+            FETCH( func, *inst, 1, 0, CHAN_Y );
+            sse_maxps(
+               func,
+               make_xmm( 1 ),
+               get_temp(
+                  TGSI_EXEC_TEMP_00000000_I,
+                  TGSI_EXEC_TEMP_00000000_C ) );
+            FETCH( func, *inst, 2, 0, CHAN_W );
+            sse_minps(
+               func,
+               make_xmm( 2 ),
+               get_temp(
+                  TGSI_EXEC_TEMP_128_I,
+                  TGSI_EXEC_TEMP_128_C ) );
+            sse_maxps(
+               func,
+               make_xmm( 2 ),
+               get_temp(
+                  TGSI_EXEC_TEMP_MINUS_128_I,
+                  TGSI_EXEC_TEMP_MINUS_128_C ) );
+            emit_pow( func, 1, 2 );
+            FETCH( func, *inst, 0, 0, CHAN_X );
+            sse_xorps(
+               func,
+               make_xmm( 2 ),
+               make_xmm( 2 ) );
+            sse_cmpps(
+               func,
+               make_xmm( 2 ),
+               make_xmm( 0 ),
+               cc_LessThanEqual );
+            sse_andps(
+               func,
+               make_xmm( 2 ),
+               make_xmm( 1 ) );
+            STORE( func, *inst, 2, 0, CHAN_Z );
+         }
+      }
+      break;
+
+   case TGSI_OPCODE_RCP:
+   /* TGSI_OPCODE_RECIP */
+      FETCH( func, *inst, 0, 0, CHAN_X );
+      emit_rcp( func, 0, 0 );
+      FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
+         STORE( func, *inst, 0, 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_RSQ:
+   /* TGSI_OPCODE_RECIPSQRT */
+      FETCH( func, *inst, 0, 0, CHAN_X );
+      emit_rsqrt( func, 0, 0 );
+      FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
+         STORE( func, *inst, 0, 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_EXP:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_LOG:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_MUL:
+      FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( func, *inst, 0, 0, chan_index );
+         FETCH( func, *inst, 1, 1, chan_index );
+         emit_mul( func, 0, 1 );
+         STORE( func, *inst, 0, 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_ADD:
+      FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( func, *inst, 0, 0, chan_index );
+         FETCH( func, *inst, 1, 1, chan_index );
+         emit_add( func, 0, 1 );
+         STORE( func, *inst, 0, 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_DP3:
+   /* TGSI_OPCODE_DOT3 */
+      FETCH( func, *inst, 0, 0, CHAN_X );
+      FETCH( func, *inst, 1, 1, CHAN_X );
+      emit_mul( func, 0, 1 );
+      FETCH( func, *inst, 1, 0, CHAN_Y );
+      FETCH( func, *inst, 2, 1, CHAN_Y );
+      emit_mul( func, 1, 2 );
+      emit_add( func, 0, 1 );
+      FETCH( func, *inst, 1, 0, CHAN_Z );
+      FETCH( func, *inst, 2, 1, CHAN_Z );
+      emit_mul( func, 1, 2 );
+      emit_add( func, 0, 1 );
+      FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
+         STORE( func, *inst, 0, 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_DP4:
+   /* TGSI_OPCODE_DOT4 */
+      FETCH( func, *inst, 0, 0, CHAN_X );
+      FETCH( func, *inst, 1, 1, CHAN_X );
+      emit_mul( func, 0, 1 );
+      FETCH( func, *inst, 1, 0, CHAN_Y );
+      FETCH( func, *inst, 2, 1, CHAN_Y );
+      emit_mul( func, 1, 2 );
+      emit_add( func, 0, 1 );
+      FETCH( func, *inst, 1, 0, CHAN_Z );
+      FETCH( func, *inst, 2, 1, CHAN_Z );
+      emit_mul(func, 1, 2 );
+      emit_add(func, 0, 1 );
+      FETCH( func, *inst, 1, 0, CHAN_W );
+      FETCH( func, *inst, 2, 1, CHAN_W );
+      emit_mul( func, 1, 2 );
+      emit_add( func, 0, 1 );
+      FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
+         STORE( func, *inst, 0, 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_DST:
+      IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) {
+         emit_tempf(
+            func,
+            0,
+            TGSI_EXEC_TEMP_ONE_I,
+            TGSI_EXEC_TEMP_ONE_C );
+         STORE( func, *inst, 0, 0, CHAN_X );
+      }
+      IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y ) {
+         FETCH( func, *inst, 0, 0, CHAN_Y );
+         FETCH( func, *inst, 1, 1, CHAN_Y );
+         emit_mul( func, 0, 1 );
+         STORE( func, *inst, 0, 0, CHAN_Y );
       }
+      IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Z ) {
+         FETCH( func, *inst, 0, 0, CHAN_Z );
+         STORE( func, *inst, 0, 0, CHAN_Z );
+      }
+      IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_W ) {
+         FETCH( func, *inst, 0, 1, CHAN_W );
+         STORE( func, *inst, 0, 0, CHAN_W );
+      }
+      break;
+
+   case TGSI_OPCODE_MIN:
+      FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( func, *inst, 0, 0, chan_index );
+         FETCH( func, *inst, 1, 1, chan_index );
+         sse_minps(
+            func,
+            make_xmm( 0 ),
+            make_xmm( 1 ) );
+         STORE( func, *inst, 0, 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_MAX:
+      FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( func, *inst, 0, 0, chan_index );
+         FETCH( func, *inst, 1, 1, chan_index );
+         sse_maxps(
+            func,
+            make_xmm( 0 ),
+            make_xmm( 1 ) );
+         STORE( func, *inst, 0, 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_SLT:
+   /* TGSI_OPCODE_SETLT */
+      emit_setcc( func, inst, cc_LessThan );
+      break;
+
+   case TGSI_OPCODE_SGE:
+   /* TGSI_OPCODE_SETGE */
+      emit_setcc( func, inst, cc_NotLessThan );
+      break;
+
+   case TGSI_OPCODE_MAD:
+   /* TGSI_OPCODE_MADD */
+      FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( func, *inst, 0, 0, chan_index );
+         FETCH( func, *inst, 1, 1, chan_index );
+         FETCH( func, *inst, 2, 2, chan_index );
+         emit_mul( func, 0, 1 );
+         emit_add( func, 0, 2 );
+         STORE( func, *inst, 0, 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_SUB:
+      FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( func, *inst, 0, 0, chan_index );
+         FETCH( func, *inst, 1, 1, chan_index );
+         emit_sub( func, 0, 1 );
+         STORE( func, *inst, 0, 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_LERP:
+   /* TGSI_OPCODE_LRP */
+      FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( func, *inst, 0, 0, chan_index );
+         FETCH( func, *inst, 1, 1, chan_index );
+         FETCH( func, *inst, 2, 2, chan_index );
+         emit_sub( func, 1, 2 );
+         emit_mul( func, 0, 1 );
+         emit_add( func, 0, 2 );
+         STORE( func, *inst, 0, 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_CND:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_CND0:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_DOT2ADD:
+   /* TGSI_OPCODE_DP2A */
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_INDEX:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_NEGATE:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_FRAC:
+   /* TGSI_OPCODE_FRC */
+      FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( func, *inst, 0, 0, chan_index );
+         emit_frc( func, 0 );
+         STORE( func, *inst, 0, 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_CLAMP:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_FLOOR:
+   /* TGSI_OPCODE_FLR */
+      FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( func, *inst, 0, 0, chan_index );
+         emit_flr( func, 0 );
+         STORE( func, *inst, 0, 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_ROUND:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_EXPBASE2:
+   /* TGSI_OPCODE_EX2 */
+      FETCH( func, *inst, 0, 0, CHAN_X );
+      emit_ex2( func, 0 );
+      FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
+         STORE( func, *inst, 0, 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_LOGBASE2:
+   /* TGSI_OPCODE_LG2 */
+      FETCH( func, *inst, 0, 0, CHAN_X );
+      emit_lg2( func, 0 );
+      FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
+         STORE( func, *inst, 0, 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_POWER:
+   /* TGSI_OPCODE_POW */
+      FETCH( func, *inst, 0, 0, CHAN_X );
+      FETCH( func, *inst, 1, 1, CHAN_X );
+      emit_pow( func, 0, 1 );
+      FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
+         STORE( func, *inst, 0, 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_CROSSPRODUCT:
+   /* TGSI_OPCODE_XPD */
+      if( IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) ||
+          IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y ) ) {
+         FETCH( func, *inst, 1, 1, CHAN_Z );
+         FETCH( func, *inst, 3, 0, CHAN_Z );
+      }
+      if( IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) ||
+          IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Z ) ) {
+         FETCH( func, *inst, 0, 0, CHAN_Y );
+         FETCH( func, *inst, 4, 1, CHAN_Y );
+      }
+      IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) {
+         emit_mov( func, 2, 0 );
+         emit_mul( func, 2, 1 );
+         emit_mov( func, 5, 3 );
+         emit_mul( func, 5, 4 );
+         emit_sub( func, 2, 5 );
+         STORE( func, *inst, 2, 0, CHAN_X );
+      }
+      if( IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y ) ||
+          IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Z ) ) {
+         FETCH( func, *inst, 2, 1, CHAN_X );
+         FETCH( func, *inst, 5, 0, CHAN_X );
+      }
+      IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y ) {
+         emit_mul( func, 3, 2 );
+         emit_mul( func, 1, 5 );
+         emit_sub( func, 3, 1 );
+         STORE( func, *inst, 3, 0, CHAN_Y );
+      }
+      IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Z ) {
+         emit_mul( func, 5, 4 );
+         emit_mul( func, 0, 2 );
+         emit_sub( func, 5, 0 );
+         STORE( func, *inst, 5, 0, CHAN_Z );
+      }
+      IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_W ) {
+         FETCH( func, *inst, 0, TGSI_EXEC_TEMP_ONE_I, TGSI_EXEC_TEMP_ONE_C );
+         STORE( func, *inst, 0, 0, CHAN_W );
+      }
+      break;
+
+   case TGSI_OPCODE_MULTIPLYMATRIX:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_ABS:
+      FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
+         FETCH( func, *inst, 0, 0, chan_index );
+         emit_abs( func, 0) ;
+
+         STORE( func, *inst, 0, 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_RCC:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_DPH:
+      FETCH( func, *inst, 0, 0, CHAN_X );
+      FETCH( func, *inst, 1, 1, CHAN_X );
+      emit_mul( func, 0, 1 );
+      FETCH( func, *inst, 1, 0, CHAN_Y );
+      FETCH( func, *inst, 2, 1, CHAN_Y );
+      emit_mul( func, 1, 2 );
+      emit_add( func, 0, 1 );
+      FETCH( func, *inst, 1, 0, CHAN_Z );
+      FETCH( func, *inst, 2, 1, CHAN_Z );
+      emit_mul( func, 1, 2 );
+      emit_add( func, 0, 1 );
+      FETCH( func, *inst, 1, 1, CHAN_W );
+      emit_add( func, 0, 1 );
+      FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
+         STORE( func, *inst, 0, 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_COS:
+      FETCH( func, *inst, 0, 0, CHAN_X );
+      emit_cos( func, 0 );
+      FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
+         STORE( func, *inst, 0, 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_DDX:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_DDY:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_KIL:
+      emit_kil( func, &inst->FullSrcRegisters[0] );
+      break;
+
+   case TGSI_OPCODE_PK2H:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_PK2US:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_PK4B:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_PK4UB:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_RFL:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_SEQ:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_SFL:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_SGT:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_SIN:
+      FETCH( func, *inst, 0, 0, CHAN_X );
+      emit_sin( func, 0 );
+      FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
+         STORE( func, *inst, 0, 0, chan_index );
+      }
+      break;
+
+   case TGSI_OPCODE_SLE:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_SNE:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_STR:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_TEX:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_TXD:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_UP2H:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_UP2US:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_UP4B:
+      assert( 0 );
       break;
 
-    case TGSI_OPCODE_MOV:
-    /* TGSI_OPCODE_SWZ */
-        FOR_EACH_ENABLED_CHANNEL
-        {
-            FETCH(0, 0, chan_index);
-            STORE(0, 0, chan_index);
-        }
-        break;
-
-    case TGSI_OPCODE_LIT:
-        if (IS_CHANNEL_ENABLED(CHAN_X) || IS_CHANNEL_ENABLED(CHAN_W))
-        {
-            emit_tempf (func, 0, TGSI_EXEC_TEMP_ONE_I, TGSI_EXEC_TEMP_ONE_C);
-            if (IS_CHANNEL_ENABLED(CHAN_X))
-                STORE(0, 0, CHAN_X);
-            if (IS_CHANNEL_ENABLED(CHAN_W))
-                STORE(0, 0, CHAN_W);
-        }
-        if (IS_CHANNEL_ENABLED(CHAN_Y) || IS_CHANNEL_ENABLED(CHAN_Z))
-        {
-            if (IS_CHANNEL_ENABLED(CHAN_Y))
-            {
-                FETCH(0, 0, CHAN_X);
-                sse_maxps (func,
-                           make_xmm (0),
-                           get_temp (TGSI_EXEC_TEMP_00000000_I,
-                                     TGSI_EXEC_TEMP_00000000_C));
-                STORE(0, 0, CHAN_Y);
-            }
-            if (IS_CHANNEL_ENABLED(CHAN_Z))
-            {
-                FETCH(1, 0, CHAN_Y);
-                sse_maxps (func,
-                           make_xmm (1),
-                           get_temp (TGSI_EXEC_TEMP_00000000_I,
-                                     TGSI_EXEC_TEMP_00000000_C));
-
-                FETCH(2, 0, CHAN_W);
-                sse_minps (func,
-                           make_xmm (2),
-                           get_temp (TGSI_EXEC_TEMP_128_I,
-                                     TGSI_EXEC_TEMP_128_C));
-                sse_maxps (func,
-                           make_xmm (2),
-                           get_temp (TGSI_EXEC_TEMP_MINUS_128_I,
-                                     TGSI_EXEC_TEMP_MINUS_128_C));
-
-                emit_pow (func, 1, 2);
-
-                FETCH(0, 0, CHAN_X);
-                sse_xorps (func, make_xmm (2), make_xmm (2));
-                sse_cmpps (func,
-                           make_xmm (2),
-                           make_xmm (0),
-                           cc_LessThanEqual);
-                sse_andps (func,
-                           make_xmm (2),
-                           make_xmm (1));
-
-                emit_store (func, 2, &inst->FullDstRegisters[0], inst, CHAN_Z);
-            }
-        }
-        break;
-
-    case TGSI_OPCODE_RCP:
-    /* TGSI_OPCODE_RECIP */
-        FETCH(0, 0, CHAN_X);
-        emit_rcp (func, 0, 0);
-
-        FOR_EACH_ENABLED_CHANNEL
-        {
-            STORE(0, 0, chan_index);
-        }
-        break;
-
-    case TGSI_OPCODE_RSQ:
-    /* TGSI_OPCODE_RECIPSQRT */
-        FETCH(0, 0, CHAN_X);
-        emit_rsqrt (func, 0, 0);
-
-        FOR_EACH_ENABLED_CHANNEL
-        {
-            STORE(0, 0, chan_index);
-        }
-        break;
-
-    case TGSI_OPCODE_EXP:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_LOG:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_MUL:
-        FOR_EACH_ENABLED_CHANNEL
-        {
-            FETCH(0, 0, chan_index);
-            FETCH(1, 1, chan_index);
-            emit_mul (func, 0, 1);
-
-            STORE(0, 0, chan_index);
-        }
-        break;
-
-    case TGSI_OPCODE_ADD:
-        FOR_EACH_ENABLED_CHANNEL
-        {
-            FETCH(0, 0, chan_index);
-            FETCH(1, 1, chan_index);
-            emit_add (func, 0, 1);
-
-            STORE(0, 0, chan_index);
-        }
-        break;
-
-    case TGSI_OPCODE_DP3:
-    /* TGSI_OPCODE_DOT3 */
-        FETCH(0, 0, CHAN_X);
-        FETCH(1, 1, CHAN_X);
-        emit_mul (func, 0, 1);
-
-        FETCH(1, 0, CHAN_Y);
-        FETCH(2, 1, CHAN_Y);
-        emit_mul (func, 1, 2);
-        emit_add (func, 0, 1);
-
-        FETCH(1, 0, CHAN_Z);
-        FETCH(2, 1, CHAN_Z);
-        emit_mul (func, 1, 2);
-        emit_add (func, 0, 1);
-
-        FOR_EACH_ENABLED_CHANNEL
-        {
-            STORE(0, 0, chan_index);
-        }
-        break;
-
-    case TGSI_OPCODE_DP4:
-    /* TGSI_OPCODE_DOT4 */
-        FETCH(0, 0, CHAN_X);
-        FETCH(1, 1, CHAN_X);
-        emit_mul (func, 0, 1);
-
-        FETCH(1, 0, CHAN_Y);
-        FETCH(2, 1, CHAN_Y);
-        emit_mul (func, 1, 2);
-        emit_add (func, 0, 1);
-
-        FETCH(1, 0, CHAN_Z);
-        FETCH(2, 1, CHAN_Z);
-        emit_mul (func, 1, 2);
-        emit_add (func, 0, 1);
-
-        FETCH(1, 0, CHAN_W);
-        FETCH(2, 1, CHAN_W);
-        emit_mul (func, 1, 2);
-        emit_add (func, 0, 1);
-
-        FOR_EACH_ENABLED_CHANNEL
-        {
-            STORE(0, 0, chan_index);
-        }
-        break;
-
-    case TGSI_OPCODE_DST:
-        IF_IS_CHANNEL_ENABLED(CHAN_X)
-        {
-            emit_tempf (func,
-                        0,
-                        TGSI_EXEC_TEMP_ONE_I,
-                        TGSI_EXEC_TEMP_ONE_C);
-            STORE(0, 0, CHAN_X);
-        }
-        IF_IS_CHANNEL_ENABLED(CHAN_Y)
-        {
-            FETCH(0, 0, CHAN_Y);
-            FETCH(1, 1, CHAN_Y);
-            emit_mul (func, 0, 1);
-            STORE(0, 0, CHAN_Y);
-        }
-        IF_IS_CHANNEL_ENABLED(CHAN_Z)
-        {
-            FETCH(0, 0, CHAN_Z);
-            STORE(0, 0, CHAN_Z);
-        }
-        IF_IS_CHANNEL_ENABLED(CHAN_W)
-        {
-            FETCH(0, 1, CHAN_W);
-            STORE(0, 0, CHAN_W);
-        }
-        break;
-
-    case TGSI_OPCODE_MIN:
-        FOR_EACH_ENABLED_CHANNEL
-        {
-            FETCH(0, 0, chan_index);
-            FETCH(1, 1, chan_index);
-            sse_minps (func,
-                       make_xmm (0),
-                       make_xmm (1));
-            STORE(0, 0, chan_index);
-        }
-        break;
-
-    case TGSI_OPCODE_MAX:
-        FOR_EACH_ENABLED_CHANNEL
-        {
-            FETCH(0, 0, chan_index);
-            FETCH(1, 1, chan_index);
-            sse_maxps (func,
-                       make_xmm (0),
-                       make_xmm (1));
-            STORE(0, 0, chan_index);
-        }
-        break;
-
-    case TGSI_OPCODE_SLT:
-    /* TGSI_OPCODE_SETLT */
-        emit_setcc (func, inst, cc_LessThan);
-        break;
-
-    case TGSI_OPCODE_SGE:
-    /* TGSI_OPCODE_SETGE */
-        emit_setcc (func, inst, cc_NotLessThan);
-        break;
-
-    case TGSI_OPCODE_MAD:
-    /* TGSI_OPCODE_MADD */
-        FOR_EACH_ENABLED_CHANNEL
-        {
-            FETCH(0, 0, chan_index);
-            FETCH(1, 1, chan_index);
-            FETCH(2, 2, chan_index);
-            emit_mul (func, 0, 1);
-            emit_add (func, 0, 2);
-
-            STORE(0, 0, chan_index);
-        }
-        break;
-
-    case TGSI_OPCODE_SUB:
-        FOR_EACH_ENABLED_CHANNEL
-        {
-            FETCH(0, 0, chan_index);
-            FETCH(1, 1, chan_index);
-            emit_sub (func, 0, 1);
-
-            STORE(0, 0, chan_index);
-        }
-        break;
-
-    case TGSI_OPCODE_LERP:
-    /* TGSI_OPCODE_LRP */
-        FOR_EACH_ENABLED_CHANNEL
-        {
-            FETCH(0, 0, chan_index);
-            FETCH(1, 1, chan_index);
-            FETCH(2, 2, chan_index);
-            emit_sub (func, 1, 2);
-            emit_mul (func, 0, 1);
-            emit_add (func, 0, 2);
-
-            STORE(0, 0, chan_index);
-        }
-        break;
-
-    case TGSI_OPCODE_CND:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_CND0:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_DOT2ADD:
-    /* TGSI_OPCODE_DP2A */
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_INDEX:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_NEGATE:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_FRAC:
-    /* TGSI_OPCODE_FRC */
-        FOR_EACH_ENABLED_CHANNEL
-        {
-            FETCH(0, 0, chan_index);
-            emit_frc (func, 0);
-            STORE(0, 0, chan_index);
-        }
-        break;
-
-    case TGSI_OPCODE_CLAMP:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_FLOOR:
-    /* TGSI_OPCODE_FLR */
-        FOR_EACH_ENABLED_CHANNEL
-        {
-            FETCH(0, 0, chan_index);
-            emit_flr (func, 0);
-            STORE(0, 0, chan_index);
-        }
-        break;
-
-    case TGSI_OPCODE_ROUND:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_EXPBASE2:
-    /* TGSI_OPCODE_EX2 */
-        FETCH(0, 0, CHAN_X);
-        emit_ex2 (func, 0);
-
-        FOR_EACH_ENABLED_CHANNEL
-        {
-            STORE(0, 0, chan_index);
-        }
-        break;
-
-    case TGSI_OPCODE_LOGBASE2:
-    /* TGSI_OPCODE_LG2 */
-        FETCH(0, 0, CHAN_X);
-        emit_lg2 (func, 0);
-
-        FOR_EACH_ENABLED_CHANNEL
-        {
-            STORE(0, 0, chan_index);
-        }
-        break;
-
-    case TGSI_OPCODE_POWER:
-    /* TGSI_OPCODE_POW */
-        FETCH(0, 0, CHAN_X);
-        FETCH(1, 1, CHAN_X);
-        emit_pow (func, 0, 1);
-
-        FOR_EACH_ENABLED_CHANNEL
-        {
-            STORE(0, 0, chan_index);
-        }
-        break;
-
-    case TGSI_OPCODE_CROSSPRODUCT:
-    /* TGSI_OPCODE_XPD */
-        if (IS_CHANNEL_ENABLED(CHAN_X) || IS_CHANNEL_ENABLED(CHAN_Y))
-        {
-            FETCH(1, 1, CHAN_Z);
-            FETCH(3, 0, CHAN_Z);
-        }
-        if (IS_CHANNEL_ENABLED(CHAN_X) || IS_CHANNEL_ENABLED(CHAN_Z))
-        {
-            FETCH(0, 0, CHAN_Y);
-            FETCH(4, 1, CHAN_Y);
-        }
-        IF_IS_CHANNEL_ENABLED(CHAN_X)
-        {
-            emit_mov (func, 2, 0);
-            emit_mul (func, 2, 1);
-            emit_mov (func, 5, 3);
-            emit_mul (func, 5, 4);
-            emit_sub (func, 2, 5);
-            STORE(2, 0, CHAN_X);
-        }
-
-        if (IS_CHANNEL_ENABLED(CHAN_Y) || IS_CHANNEL_ENABLED(CHAN_Z))
-        {
-            FETCH(2, 1, CHAN_X);
-            FETCH(5, 0, CHAN_X);
-        }
-        IF_IS_CHANNEL_ENABLED(CHAN_Y)
-        {
-            emit_mul (func, 3, 2);
-            emit_mul (func, 1, 5);
-            emit_sub (func, 3, 1);
-            STORE(3, 0, CHAN_Y);
-        }
-
-        IF_IS_CHANNEL_ENABLED(CHAN_Z)
-        {
-            emit_mul (func, 5, 4);
-            emit_mul (func, 0, 2);
-            emit_sub (func, 5, 0);
-            STORE(5, 0, CHAN_Z);
-        }
-
-        IF_IS_CHANNEL_ENABLED(CHAN_W)
-        {
-            FETCH(0, TGSI_EXEC_TEMP_ONE_I, TGSI_EXEC_TEMP_ONE_C);
-            STORE(0, 0, CHAN_W);
-        }
-        break;
-
-    case TGSI_OPCODE_MULTIPLYMATRIX:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_ABS:
-        FOR_EACH_ENABLED_CHANNEL
-        {
-            FETCH(0, 0, chan_index);
-            emit_abs (func, 0);
-
-            STORE(0, 0, chan_index);
-        }
-        break;
-
-    case TGSI_OPCODE_RCC:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_DPH:
-        FETCH(0, 0, CHAN_X);
-        FETCH(1, 1, CHAN_X);
-        emit_mul (func, 0, 1);
-
-        FETCH(1, 0, CHAN_Y);
-        FETCH(2, 1, CHAN_Y);
-        emit_mul (func, 1, 2);
-        emit_add (func, 0, 1);
-
-        FETCH(1, 0, CHAN_Z);
-        FETCH(2, 1, CHAN_Z);
-        emit_mul (func, 1, 2);
-        emit_add (func, 0, 1);
-
-        FETCH(1, 1, CHAN_W);
-        emit_add (func, 0, 1);
-
-        FOR_EACH_ENABLED_CHANNEL
-        {
-            STORE(0, 0, chan_index);
-        }
-        break;
-
-    case TGSI_OPCODE_COS:
-        FETCH(0, 0, CHAN_X);
-        emit_cos (func, 0);
-
-        FOR_EACH_ENABLED_CHANNEL
-        {
-            STORE(0, 0, chan_index);
-        }
-        break;
-
-    case TGSI_OPCODE_DDX:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_DDY:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_KIL:
-        emit_kil (func, &inst->FullSrcRegisters[0]);
-        break;
-
-    case TGSI_OPCODE_PK2H:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_PK2US:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_PK4B:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_PK4UB:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_RFL:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_SEQ:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_SFL:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_SGT:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_SIN:
-        FETCH(0, 0, CHAN_X);
-        emit_sin (func, 0);
-
-        FOR_EACH_ENABLED_CHANNEL
-        {
-            STORE(0, 0, chan_index);
-        }
-        break;
-
-    case TGSI_OPCODE_SLE:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_SNE:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_STR:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_TEX:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_TXD:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_UP2H:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_UP2US:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_UP4B:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_UP4UB:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_X2D:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_ARA:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_ARR:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_BRA:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_CAL:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_RET:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_SSG:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_CMP:
-        emit_cmp (func, inst);
-        break;
-
-    case TGSI_OPCODE_SCS:
-        IF_IS_CHANNEL_ENABLED(CHAN_X)
-        {
-            FETCH(0, 0, CHAN_X);
-            emit_cos (func, 0);
-            STORE(0, 0, CHAN_X);
-        }
-
-        IF_IS_CHANNEL_ENABLED(CHAN_Y)
-        {
-            FETCH(0, 0, CHAN_Y);
-            emit_sin (func, 0);
-            STORE(0, 0, CHAN_Y);
-        }
-
-        IF_IS_CHANNEL_ENABLED(CHAN_Z)
-        {
-            FETCH(0, TGSI_EXEC_TEMP_00000000_I, TGSI_EXEC_TEMP_00000000_C);
-            STORE(0, 0, CHAN_Z);
-        }
-
-        IF_IS_CHANNEL_ENABLED(CHAN_W)
-        {
-            FETCH(0, TGSI_EXEC_TEMP_ONE_I, TGSI_EXEC_TEMP_ONE_C);
-            STORE(0, 0, CHAN_W);
-        }
-        break;
-
-    case TGSI_OPCODE_TXB:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_NRM:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_DIV:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_DP2:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_TXL:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_BRK:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_IF:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_LOOP:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_REP:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_ELSE:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_ENDIF:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_ENDLOOP:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_ENDREP:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_PUSHA:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_POPA:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_CEIL:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_I2F:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_NOT:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_TRUNC:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_SHL:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_SHR:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_AND:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_OR:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_MOD:
-        assert (0);
-        break;
-
-    case TGSI_OPCODE_XOR:
-        assert (0);
-        break;
+   case TGSI_OPCODE_UP4UB:
+      assert( 0 );
+      break;
 
-    case TGSI_OPCODE_SAD:
-        assert (0);
-        break;
+   case TGSI_OPCODE_X2D:
+      assert( 0 );
+      break;
 
-    case TGSI_OPCODE_TXF:
-        assert (0);
-        break;
+   case TGSI_OPCODE_ARA:
+      assert( 0 );
+      break;
 
-    case TGSI_OPCODE_TXQ:
-        assert (0);
-        break;
+   case TGSI_OPCODE_ARR:
+      assert( 0 );
+      break;
 
-    case TGSI_OPCODE_CONT:
-        assert (0);
-        break;
+   case TGSI_OPCODE_BRA:
+      assert( 0 );
+      break;
 
-    case TGSI_OPCODE_EMIT:
-        assert (0);
-        break;
+   case TGSI_OPCODE_CAL:
+      assert( 0 );
+      break;
 
-    case TGSI_OPCODE_ENDPRIM:
-        assert (0);
-        break;
+   case TGSI_OPCODE_RET:
+      assert( 0 );
+      break;
 
-    default:
-        assert (0);
-    }
+   case TGSI_OPCODE_SSG:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_CMP:
+      emit_cmp (func, inst);
+      break;
+
+   case TGSI_OPCODE_SCS:
+      IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_X ) {
+         FETCH( func, *inst, 0, 0, CHAN_X );
+         emit_cos( func, 0 );
+         STORE( func, *inst, 0, 0, CHAN_X );
+      }
+      IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Y ) {
+         FETCH( func, *inst, 0, 0, CHAN_Y );
+         emit_sin( func, 0 );
+         STORE( func, *inst, 0, 0, CHAN_Y );
+      }
+      IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_Z ) {
+         FETCH( func, *inst, 0, TGSI_EXEC_TEMP_00000000_I, TGSI_EXEC_TEMP_00000000_C );
+         STORE( func, *inst, 0, 0, CHAN_Z );
+      }
+      IF_IS_DST0_CHANNEL_ENABLED( *inst, CHAN_W ) {
+         FETCH( func, *inst, 0, TGSI_EXEC_TEMP_ONE_I, TGSI_EXEC_TEMP_ONE_C );
+         STORE( func, *inst, 0, 0, CHAN_W );
+      }
+      break;
+
+   case TGSI_OPCODE_TXB:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_NRM:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_DIV:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_DP2:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_TXL:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_BRK:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_IF:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_LOOP:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_REP:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_ELSE:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_ENDIF:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_ENDLOOP:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_ENDREP:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_PUSHA:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_POPA:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_CEIL:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_I2F:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_NOT:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_TRUNC:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_SHL:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_SHR:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_AND:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_OR:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_MOD:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_XOR:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_SAD:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_TXF:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_TXQ:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_CONT:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_EMIT:
+      assert( 0 );
+      break;
+
+   case TGSI_OPCODE_ENDPRIM:
+      assert( 0 );
+      break;
+
+   default:
+      assert( 0 );
+   }
 }
 
-GLboolean
-tgsi_emit_sse (struct tgsi_token *tokens,
-               struct x86_function *function)
+unsigned
+tgsi_emit_sse(
+   struct tgsi_token *tokens,
+   struct x86_function *func )
 {
-    struct tgsi_parse_context parse;
+   struct tgsi_parse_context parse;
 
-    x86_init_func (function);
+   x86_init_func( func );
 
-    x86_mov (function, get_input_base (), get_argument (0));
-    x86_mov (function, get_output_base (), get_argument (1));
-    x86_mov (function, get_const_base (), get_argument (2));
-    x86_mov (function, get_temp_base (), get_argument (3));
+   x86_mov(
+      func,
+      get_input_base(),
+      get_argument( 0 ) );
+   x86_mov(
+      func,
+      get_output_base(),
+      get_argument( 1 ) );
+   x86_mov(
+      func,
+      get_const_base(),
+      get_argument( 2 ) );
+   x86_mov(
+      func,
+      get_temp_base(),
+      get_argument( 3 ) );
 
-    tgsi_parse_init (&parse, tokens);
+   tgsi_parse_init( &parse, tokens );
 
-    while (!tgsi_parse_end_of_tokens (&parse))
-    {
-        tgsi_parse_token (&parse);
+   while( !tgsi_parse_end_of_tokens( &parse ) ) {
+      tgsi_parse_token( &parse );
 
-        switch (parse.FullToken.Token.Type)
-        {
-        case TGSI_TOKEN_TYPE_DECLARATION:
-            break;
+      switch( parse.FullToken.Token.Type ) {
+      case TGSI_TOKEN_TYPE_DECLARATION:
+         break;
 
-        case TGSI_TOKEN_TYPE_INSTRUCTION:
-            emit_instruction (function, &parse.FullToken.FullInstruction);
-            break;
+      case TGSI_TOKEN_TYPE_INSTRUCTION:
+         emit_instruction(
+            func,
+            &parse.FullToken.FullInstruction );
+         break;
 
-        default:
-            assert (0);
-        }
-    }
+      default:
+         assert( 0 );
+      }
+   }
 
-    tgsi_parse_free (&parse);
+   tgsi_parse_free( &parse );
 
 #ifdef WIN32
-    x86_retw (function, 16);
+   x86_retw( func, 16 );
 #else
-    x86_ret (function);
+   x86_ret( func );
 #endif
 
-    return GL_FALSE;
+   return 1;
 }
 
 #endif
diff --git a/src/mesa/x86/rtasm/x86sse.c b/src/mesa/x86/rtasm/x86sse.c
index 3ea37bb5e7..f5b0ccdb9b 100644
--- a/src/mesa/x86/rtasm/x86sse.c
+++ b/src/mesa/x86/rtasm/x86sse.c
@@ -502,6 +502,14 @@ void sse_addss( struct x86_function *p,
    emit_modrm( p, dst, src );
 }
 
+void sse_andnps( struct x86_function *p,
+                 struct x86_reg dst,
+                 struct x86_reg src )
+{
+   emit_2ub(p, X86_TWOB, 0x55);
+   emit_modrm( p, dst, src );
+}
+
 void sse_andps( struct x86_function *p,
 		struct x86_reg dst,
 		struct x86_reg src )
@@ -510,6 +518,13 @@ void sse_andps( struct x86_function *p,
    emit_modrm( p, dst, src );
 }
 
+void sse_rsqrtps( struct x86_function *p,
+                  struct x86_reg dst,
+                  struct x86_reg src )
+{
+   emit_2ub(p, X86_TWOB, 0x52);
+   emit_modrm( p, dst, src );
+}
 
 void sse_rsqrtss( struct x86_function *p,
 		  struct x86_reg dst,
@@ -538,6 +553,21 @@ void sse_movlhps( struct x86_function *p,
    emit_modrm( p, dst, src );
 }
 
+void sse_orps( struct x86_function *p,
+               struct x86_reg dst,
+               struct x86_reg src )
+{
+   emit_2ub(p, X86_TWOB, 0x56);
+   emit_modrm( p, dst, src );
+}
+
+void sse_xorps( struct x86_function *p,
+                struct x86_reg dst,
+                struct x86_reg src )
+{
+   emit_2ub(p, X86_TWOB, 0x57);
+   emit_modrm( p, dst, src );
+}
 
 void sse_cvtps2pi( struct x86_function *p,
 		   struct x86_reg dst,
@@ -576,6 +606,14 @@ void sse_cmpps( struct x86_function *p,
    emit_1ub(p, cc); 
 }
 
+void sse_pmovmskb( struct x86_function *p,
+                   struct x86_reg dest,
+                   struct x86_reg src)
+{
+    emit_3ub(p, 0x66, X86_TWOB, 0xD7);
+    emit_modrm(p, dest, src);
+}
+
 /***********************************************************************
  * SSE2 instructions
  */
@@ -593,6 +631,14 @@ void sse2_pshufd( struct x86_function *p,
    emit_1ub(p, shuf); 
 }
 
+void sse2_cvttps2dq( struct x86_function *p,
+                     struct x86_reg dst,
+                     struct x86_reg src )
+{
+   emit_3ub( p, 0xF3, X86_TWOB, 0x5B );
+   emit_modrm( p, dst, src );
+}
+
 void sse2_cvtps2dq( struct x86_function *p,
 		    struct x86_reg dst,
 		    struct x86_reg src )
@@ -625,6 +671,14 @@ void sse2_packuswb( struct x86_function *p,
    emit_modrm( p, dst, src );
 }
 
+void sse2_rcpps( struct x86_function *p,
+                 struct x86_reg dst,
+                 struct x86_reg src )
+{
+   emit_2ub(p, X86_TWOB, 0x53);
+   emit_modrm( p, dst, src );
+}
+
 void sse2_rcpss( struct x86_function *p,
 		struct x86_reg dst,
 		struct x86_reg src )
diff --git a/src/mesa/x86/rtasm/x86sse.h b/src/mesa/x86/rtasm/x86sse.h
index 66fb852ac9..c6236395b2 100644
--- a/src/mesa/x86/rtasm/x86sse.h
+++ b/src/mesa/x86/rtasm/x86sse.h
@@ -142,17 +142,20 @@ void mmx_packssdw( struct x86_function *p, struct x86_reg dst, struct x86_reg sr
 void mmx_packuswb( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
 
 void sse2_cvtps2dq( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
+void sse2_cvttps2dq( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
 void sse2_movd( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
 void sse2_packssdw( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
 void sse2_packsswb( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
 void sse2_packuswb( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
 void sse2_pshufd( struct x86_function *p, struct x86_reg dest, struct x86_reg arg0, GLubyte shuf );
+void sse2_rcpps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
 void sse2_rcpss( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
 
 void sse_addps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
 void sse_addss( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
 void sse_cvtps2pi( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
 void sse_divss( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
+void sse_andnps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
 void sse_andps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
 void sse_cmpps( struct x86_function *p, struct x86_reg dst, struct x86_reg src, GLubyte cc );
 void sse_maxps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
@@ -167,9 +170,13 @@ void sse_movss( struct x86_function *p, struct x86_reg dst, struct x86_reg src )
 void sse_movups( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
 void sse_mulps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
 void sse_mulss( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
+void sse_orps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
+void sse_xorps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
 void sse_subps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
+void sse_rsqrtps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
 void sse_rsqrtss( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
 void sse_shufps( struct x86_function *p, struct x86_reg dest, struct x86_reg arg0, GLubyte shuf );
+void sse_pmovmskb( struct x86_function *p, struct x86_reg dest, struct x86_reg src );
 
 void x86_add( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
 void x86_and( struct x86_function *p, struct x86_reg dst, struct x86_reg src );