6 files changed, 64 insertions, 10 deletions
diff --git a/src/mesa/pipe/p_state.h b/src/mesa/pipe/p_state.h
index fd5e7ad3af..e3f62a80ad 100644
--- a/src/mesa/pipe/p_state.h
+++ b/src/mesa/pipe/p_state.h
@@ -121,8 +121,11 @@ struct pipe_clip_state {
    GLuint nr;
 };
 
+
 struct pipe_fs_state {
-   struct gl_fragment_program *fp;
+   GLuint inputs_read;		/* FRAG_ATTRIB_* */
+   const struct tgsi_token *tokens;
+   
 };
 
 struct pipe_constant_buffer {
diff --git a/src/mesa/pipe/softpipe/sp_quad_fs.c b/src/mesa/pipe/softpipe/sp_quad_fs.c
index 7b1c90cba6..536365f219 100644
--- a/src/mesa/pipe/softpipe/sp_quad_fs.c
+++ b/src/mesa/pipe/softpipe/sp_quad_fs.c
@@ -37,6 +37,7 @@
 #include "sp_context.h"
 #include "sp_headers.h"
 #include "sp_quad.h"
+#include "core/tgsi_core.h"
 
 struct exec_machine {
    const struct setup_coefficient *coef; /**< will point to quad->coef */
@@ -180,8 +181,47 @@ shade_quad( struct quad_stage *qs, struct quad_header *quad )
 #endif
    }
 
-#if 0
-   softpipe->run_fs( tri->fp, quad, &tri->outputs );
+#if 1
+   /*softpipe->run_fs( tri->fp, quad, &tri->outputs );*/
+
+   {
+      struct tgsi_exec_machine machine;
+      struct tgsi_exec_vector inputs[FRAG_ATTRIB_MAX + 1];
+      struct tgsi_exec_vector outputs[FRAG_ATTRIB_MAX + 1];
+      struct tgsi_exec_vector *ainputs;
+      struct tgsi_exec_vector *aoutputs;
+      GLuint i, total;
+
+      ainputs = (struct tgsi_exec_vector *) tgsi_align_128bit( inputs );
+      aoutputs = (struct tgsi_exec_vector *) tgsi_align_128bit( outputs );
+
+      for( i = total = 0; i < PIPE_ATTRIB_MAX; i++ ) {
+         GLuint attr;
+
+         attr = softpipe->fp_attr_to_slot[i];
+         if( attr ) {
+            assert( total < FRAG_ATTRIB_MAX );
+            assert( attr < FRAG_ATTRIB_MAX );
+            assert( sizeof( ainputs[0] ) == sizeof( exec.attr[0] ) );
+
+            memcpy(
+               &ainputs[total],
+               exec.attr[attr],
+               sizeof( ainputs[0] ) );
+            total++;
+         }
+      }
+
+      tgsi_exec_machine_init(
+         &machine,
+         softpipe->fs.tokens );
+
+      machine.Inputs = ainputs;
+      machine.Outputs = aoutputs;
+
+      tgsi_exec_machine_run(
+         &machine );
+   }
 #else
    {
       GLuint attr = softpipe->fp_attr_to_slot[FRAG_ATTRIB_COL0];
diff --git a/src/mesa/pipe/softpipe/sp_state_derived.c b/src/mesa/pipe/softpipe/sp_state_derived.c
index 18dfb50e38..fcdedb54a9 100644
--- a/src/mesa/pipe/softpipe/sp_state_derived.c
+++ b/src/mesa/pipe/softpipe/sp_state_derived.c
@@ -28,7 +28,6 @@
 #include "main/glheader.h"
 #include "main/macros.h"
 #include "main/enums.h"
-#include "shader/program.h"
 
 #include "vf/vf.h"
 #include "pipe/draw/draw_context.h"
@@ -68,8 +67,7 @@ static const GLuint frag_to_vf[FRAG_ATTRIB_MAX] =
  */
 static void calculate_vertex_layout( struct softpipe_context *softpipe )
 {
-   struct gl_fragment_program *fp = softpipe->fs.fp;
-   const GLuint inputsRead = fp->Base.InputsRead;
+   const GLuint inputsRead = softpipe->fs.inputs_read;
    GLuint slot_to_vf_attr[VF_ATTRIB_MAX];
    GLbitfield attr_mask = 0x0;
    GLuint i;
diff --git a/src/mesa/pipe/tgsi/core/tgsi_dump.c b/src/mesa/pipe/tgsi/core/tgsi_dump.c
index fecb246ab1..0345fd93f7 100644
--- a/src/mesa/pipe/tgsi/core/tgsi_dump.c
+++ b/src/mesa/pipe/tgsi/core/tgsi_dump.c
@@ -400,12 +400,16 @@ tgsi_dump(
    GLuint deflt = !(flags & TGSI_DUMP_NO_DEFAULT);
 
    {
+#if 0
       static GLuint counter = 0;
       char buffer[64];
 
       sprintf( buffer, "sbir-dump-%.4u.txt", counter++ );
       dump.file = fopen( buffer, "wt" );
+#else
+      dump.file = stderr;
       dump.tabs = 0;
+#endif
    }
 
    tgsi_parse_init( &parse, tokens );
diff --git a/src/mesa/pipe/tgsi/core/tgsi_util.c b/src/mesa/pipe/tgsi/core/tgsi_util.c
index 2331affdfd..38d6d6e6bc 100644
--- a/src/mesa/pipe/tgsi/core/tgsi_util.c
+++ b/src/mesa/pipe/tgsi/core/tgsi_util.c
@@ -1,15 +1,22 @@
 #include "tgsi_platform.h"
 #include "tgsi_core.h"
 
+union pointer_hack
+{
+   void *pointer;
+   unsigned long long uint64;
+};
+
 void *
 tgsi_align_128bit(
    void *unaligned )
 {
-   GLuint *ptr, addr;
+   union pointer_hack ph;
 
-   ptr = (GLuint *) unaligned;
-   addr = (*(GLuint *) &ptr + 15) & ~15;
-   return *(void **) &addr;
+   ph.uint64 = 0;
+   ph.pointer = unaligned;
+   ph.uint64 = (ph.uint64 + 15) & ~15;
+   return ph.pointer;
 }
 
 GLuint
diff --git a/src/mesa/pipe/tgsi/mesa/mesa_to_tgsi.h b/src/mesa/pipe/tgsi/mesa/mesa_to_tgsi.h
index 4c1141e579..9256318997 100644
--- a/src/mesa/pipe/tgsi/mesa/mesa_to_tgsi.h
+++ b/src/mesa/pipe/tgsi/mesa/mesa_to_tgsi.h
@@ -5,6 +5,8 @@
 extern "C" {
 #endif // defined __cplusplus
 
+struct tgsi_token;
+
 GLboolean
 tgsi_mesa_compile_fp_program(
    const struct gl_fragment_program *program,