Merge commit 'origin/gallium-0.1' into gallium-0.2

Conflicts:

	Makefile
	progs/demos/Makefile
	progs/glsl/Makefile
	progs/redbook/Makefile
	progs/samples/Makefile
	progs/tests/Makefile
	progs/trivial/Makefile
	progs/xdemos/Makefile
	src/gallium/Makefile
	src/mesa/main/attrib.c
	src/mesa/main/bufferobj.c
	src/mesa/vbo/vbo_exec_draw.c

4 files changed, 105 insertions, 69 deletions

diff --git a/src/gallium/auxiliary/draw/draw_pt_vcache.c b/src/gallium/auxiliary/draw/draw_pt_vcache.c
index e8467b2ae3..80d7200ca6 100644
--- a/src/gallium/auxiliary/draw/draw_pt_vcache.c
+++ b/src/gallium/auxiliary/draw/draw_pt_vcache.c
@@ -63,7 +63,8 @@ struct vcache_frontend {
    unsigned opt;
 };
 
-static void vcache_flush( struct vcache_frontend *vcache )
+static INLINE void 
+vcache_flush( struct vcache_frontend *vcache )
 {
    if (vcache->middle_prim != vcache->output_prim) {
       vcache->middle_prim = vcache->output_prim;
@@ -86,7 +87,8 @@ static void vcache_flush( struct vcache_frontend *vcache )
    vcache->draw_count = 0;
 }
 
-static void vcache_check_flush( struct vcache_frontend *vcache )
+static INLINE void 
+vcache_check_flush( struct vcache_frontend *vcache )
 {
    if ( vcache->draw_count + 6 >= DRAW_MAX ||
         vcache->fetch_count + 4 >= FETCH_MAX )
@@ -96,9 +98,10 @@ static void vcache_check_flush( struct vcache_frontend *vcache )
 }
 
 
-static INLINE void vcache_elt( struct vcache_frontend *vcache,
-                               unsigned felt,
-                               ushort flags )
+static INLINE void 
+vcache_elt( struct vcache_frontend *vcache,
+            unsigned felt,
+            ushort flags )
 {
    unsigned idx = felt % CACHE_MAX;
 
@@ -115,10 +118,11 @@ static INLINE void vcache_elt( struct vcache_frontend *vcache,
 
 
                    
-static void vcache_triangle( struct vcache_frontend *vcache,
-                             unsigned i0,
-                             unsigned i1,
-                             unsigned i2 )
+static INLINE void 
+vcache_triangle( struct vcache_frontend *vcache,
+                 unsigned i0,
+                 unsigned i1,
+                 unsigned i2 )
 {
    vcache_elt(vcache, i0, 0);
    vcache_elt(vcache, i1, 0);
@@ -127,11 +131,12 @@ static void vcache_triangle( struct vcache_frontend *vcache,
 }
 
 			  
-static void vcache_triangle_flags( struct vcache_frontend *vcache,
-                                   ushort flags,
-                                   unsigned i0,
-                                   unsigned i1,
-                                   unsigned i2 )
+static INLINE void 
+vcache_triangle_flags( struct vcache_frontend *vcache,
+                       ushort flags,
+                       unsigned i0,
+                       unsigned i1,
+                       unsigned i2 )
 {
    vcache_elt(vcache, i0, flags);
    vcache_elt(vcache, i1, 0);
@@ -139,9 +144,10 @@ static void vcache_triangle_flags( struct vcache_frontend *vcache,
    vcache_check_flush(vcache);
 }
 
-static void vcache_line( struct vcache_frontend *vcache,
-                         unsigned i0,
-                         unsigned i1 )
+static INLINE void 
+vcache_line( struct vcache_frontend *vcache,
+             unsigned i0,
+             unsigned i1 )
 {
    vcache_elt(vcache, i0, 0);
    vcache_elt(vcache, i1, 0);
@@ -149,10 +155,11 @@ static void vcache_line( struct vcache_frontend *vcache,
 }
 
 
-static void vcache_line_flags( struct vcache_frontend *vcache,
-                               ushort flags,
-                               unsigned i0,
-                               unsigned i1 )
+static INLINE void 
+vcache_line_flags( struct vcache_frontend *vcache,
+                   ushort flags,
+                   unsigned i0,
+                   unsigned i1 )
 {
    vcache_elt(vcache, i0, flags);
    vcache_elt(vcache, i1, 0);
@@ -160,28 +167,31 @@ static void vcache_line_flags( struct vcache_frontend *vcache,
 }
 
 
-static void vcache_point( struct vcache_frontend *vcache,
-                          unsigned i0 )
+static INLINE void 
+vcache_point( struct vcache_frontend *vcache,
+              unsigned i0 )
 {
    vcache_elt(vcache, i0, 0);
    vcache_check_flush(vcache);
 }
 
-static void vcache_quad( struct vcache_frontend *vcache,
-                         unsigned i0,
-                         unsigned i1,
-                         unsigned i2,
-                         unsigned i3 )
+static INLINE void 
+vcache_quad( struct vcache_frontend *vcache,
+             unsigned i0,
+             unsigned i1,
+             unsigned i2,
+             unsigned i3 )
 {
    vcache_triangle( vcache, i0, i1, i3 );
    vcache_triangle( vcache, i1, i2, i3 );
 }
 
-static void vcache_ef_quad( struct vcache_frontend *vcache,
-                            unsigned i0,
-                            unsigned i1,
-                            unsigned i2,
-                            unsigned i3 )
+static INLINE void 
+vcache_ef_quad( struct vcache_frontend *vcache,
+                unsigned i0,
+                unsigned i1,
+                unsigned i2,
+                unsigned i3 )
 {
    vcache_triangle_flags( vcache,
                           ( DRAW_PIPE_RESET_STIPPLE |
@@ -213,10 +223,11 @@ static void vcache_ef_quad( struct vcache_frontend *vcache,
 #define FUNC vcache_run
 #include "draw_pt_vcache_tmp.h"
 
-static void rebase_uint_elts( const unsigned *src,
-                              unsigned count,
-                              int delta,
-                              ushort *dest )
+static INLINE void 
+rebase_uint_elts( const unsigned *src,
+                  unsigned count,
+                  int delta,
+                  ushort *dest )
 {
    unsigned i;
 
@@ -224,9 +235,10 @@ static void rebase_uint_elts( const unsigned *src,
       dest[i] = (ushort)(src[i] + delta);
 }
 
-static void rebase_ushort_elts( const ushort *src,
-                                unsigned count,
-                                int delta,
+static INLINE void 
+rebase_ushort_elts( const ushort *src,
+                    unsigned count,
+                    int delta,
                                 ushort *dest )
 {
    unsigned i;
@@ -235,10 +247,11 @@ static void rebase_ushort_elts( const ushort *src,
       dest[i] = (ushort)(src[i] + delta);
 }
 
-static void rebase_ubyte_elts( const ubyte *src,
-                               unsigned count,
-                               int delta,
-                               ushort *dest )
+static INLINE void 
+rebase_ubyte_elts( const ubyte *src,
+                   unsigned count,
+                   int delta,
+                   ushort *dest )
 {
    unsigned i;
 
@@ -248,9 +261,10 @@ static void rebase_ubyte_elts( const ubyte *src,
 
 
 
-static void translate_uint_elts( const unsigned *src,
-                                 unsigned count,
-                                 ushort *dest )
+static INLINE void 
+translate_uint_elts( const unsigned *src,
+                     unsigned count,
+                     ushort *dest )
 {
    unsigned i;
 
@@ -258,9 +272,10 @@ static void translate_uint_elts( const unsigned *src,
       dest[i] = (ushort)(src[i]);
 }
 
-static void translate_ushort_elts( const ushort *src,
-                                   unsigned count,
-                                   ushort *dest )
+static INLINE void 
+translate_ushort_elts( const ushort *src,
+                       unsigned count,
+                       ushort *dest )
 {
    unsigned i;
 
@@ -268,9 +283,10 @@ static void translate_ushort_elts( const ushort *src,
       dest[i] = (ushort)(src[i]);
 }
 
-static void translate_ubyte_elts( const ubyte *src,
-                                  unsigned count,
-                                  ushort *dest )
+static INLINE void 
+translate_ubyte_elts( const ubyte *src,
+                      unsigned count,
+                      ushort *dest )
 {
    unsigned i;
 
@@ -282,7 +298,8 @@ static void translate_ubyte_elts( const ubyte *src,
 
 
 #if 0
-static enum pipe_format format_from_get_elt( pt_elt_func get_elt )
+static INLINE enum pipe_format 
+format_from_get_elt( pt_elt_func get_elt )
 {
    switch (draw->pt.user.eltSize) {
    case 1: return PIPE_FORMAT_R8_UNORM;
@@ -293,10 +310,11 @@ static enum pipe_format format_from_get_elt( pt_elt_func get_elt )
 }
 #endif
 
-static void vcache_check_run( struct draw_pt_front_end *frontend, 
-                              pt_elt_func get_elt,
-                              const void *elts,
-                              unsigned draw_count )
+static INLINE void 
+vcache_check_run( struct draw_pt_front_end *frontend, 
+                  pt_elt_func get_elt,
+                  const void *elts,
+                  unsigned draw_count )
 {
    struct vcache_frontend *vcache = (struct vcache_frontend *)frontend; 
    struct draw_context *draw = vcache->draw;
@@ -416,10 +434,11 @@ static void vcache_check_run( struct draw_pt_front_end *frontend,
 
 
 
-static void vcache_prepare( struct draw_pt_front_end *frontend,
-                            unsigned prim,
-                            struct draw_pt_middle_end *middle,
-			    unsigned opt )
+static void 
+vcache_prepare( struct draw_pt_front_end *frontend,
+                unsigned prim,
+                struct draw_pt_middle_end *middle,
+                unsigned opt )
 {
    struct vcache_frontend *vcache = (struct vcache_frontend *)frontend;
 
@@ -448,14 +467,16 @@ static void vcache_prepare( struct draw_pt_front_end *frontend,
 
 
 
-static void vcache_finish( struct draw_pt_front_end *frontend )
+static void 
+vcache_finish( struct draw_pt_front_end *frontend )
 {
    struct vcache_frontend *vcache = (struct vcache_frontend *)frontend;
    vcache->middle->finish( vcache->middle );
    vcache->middle = NULL;
 }
 
-static void vcache_destroy( struct draw_pt_front_end *frontend )
+static void 
+vcache_destroy( struct draw_pt_front_end *frontend )
 {
    FREE(frontend);
 }
diff --git a/src/gallium/auxiliary/draw/draw_vs_aos.c b/src/gallium/auxiliary/draw/draw_vs_aos.c
index 760fcb389f..a556477a76 100644
--- a/src/gallium/auxiliary/draw/draw_vs_aos.c
+++ b/src/gallium/auxiliary/draw/draw_vs_aos.c
@@ -32,6 +32,7 @@
 #include "util/u_memory.h"
 #include "util/u_math.h"
 #include "pipe/p_shader_tokens.h"
+#include "pipe/p_debug.h"
 #include "tgsi/tgsi_parse.h"
 #include "tgsi/tgsi_util.h"
 #include "tgsi/tgsi_exec.h"
@@ -119,21 +120,26 @@ static struct x86_reg get_reg_ptr(struct aos_compilation *cp,
 
    switch (file) {
    case TGSI_FILE_INPUT:
+      assert(idx < MAX_INPUTS);
       return x86_make_disp(ptr, Offset(struct aos_machine, input[idx]));
 
    case TGSI_FILE_OUTPUT:
       return x86_make_disp(ptr, Offset(struct aos_machine, output[idx]));
 
    case TGSI_FILE_TEMPORARY:
+      assert(idx < MAX_TEMPS);
       return x86_make_disp(ptr, Offset(struct aos_machine, temp[idx]));
 
    case AOS_FILE_INTERNAL:
+      assert(idx < MAX_INTERNALS);
       return x86_make_disp(ptr, Offset(struct aos_machine, internal[idx]));
 
    case TGSI_FILE_IMMEDIATE: 
+      assert(idx < MAX_IMMEDIATES);  /* just a sanity check */
       return x86_make_disp(aos_get_x86(cp, 0, X86_IMMEDIATES), idx * 4 * sizeof(float));
 
    case TGSI_FILE_CONSTANT: 
+      assert(idx < MAX_CONSTANTS);  /* just a sanity check */
       return x86_make_disp(aos_get_x86(cp, 1, X86_CONSTANTS), idx * 4 * sizeof(float));
 
    default:
@@ -2108,6 +2114,11 @@ static void PIPE_CDECL vaos_run_linear( struct draw_vs_varient *varient,
                          start,
                          count,
                          output_buffer );
+
+   /* Sanity spot checks to make sure we didn't trash our constants */
+   assert(machine->internal[IMM_ONES][0] == 1.0f);
+   assert(machine->internal[IMM_IDENTITY][0] == 0.0f);
+   assert(machine->internal[IMM_NEGS][0] == -1.0f);
 }
 
 
diff --git a/src/gallium/auxiliary/draw/draw_vs_aos.h b/src/gallium/auxiliary/draw/draw_vs_aos.h
index 64e021ff6b..7fe6f79db0 100644
--- a/src/gallium/auxiliary/draw/draw_vs_aos.h
+++ b/src/gallium/auxiliary/draw/draw_vs_aos.h
@@ -51,11 +51,11 @@ struct x86_function;
 #define W    3
 
 #define MAX_INPUTS     PIPE_MAX_ATTRIBS
-#define MAX_OUTPUTS    PIPE_MAX_ATTRIBS
-#define MAX_TEMPS      PIPE_MAX_ATTRIBS /* say */
-#define MAX_CONSTANTS  PIPE_MAX_ATTRIBS /* say */
-#define MAX_IMMEDIATES PIPE_MAX_ATTRIBS /* say */
-#define MAX_INTERNALS  8
+#define MAX_OUTPUTS    PIPE_MAX_SHADER_OUTPUTS
+#define MAX_TEMPS      TGSI_EXEC_NUM_TEMPS
+#define MAX_CONSTANTS  1024  /** only used for sanity checking */
+#define MAX_IMMEDIATES 1024  /** only used for sanity checking */
+#define MAX_INTERNALS  8     /** see IMM_x values below */
 
 #define AOS_FILE_INTERNAL TGSI_FILE_COUNT
 
diff --git a/src/gallium/auxiliary/draw/draw_vs_aos_io.c b/src/gallium/auxiliary/draw/draw_vs_aos_io.c
index ab3c5b94a5..26297c74f8 100644
--- a/src/gallium/auxiliary/draw/draw_vs_aos_io.c
+++ b/src/gallium/auxiliary/draw/draw_vs_aos_io.c
@@ -55,9 +55,13 @@ static void emit_load_R32G32B32( struct aos_compilation *cp,
 				 struct x86_reg src_ptr )
 {
    sse_movss(cp->func, data, x86_make_disp(src_ptr, 8));
+   /* data = z ? ? ? */
    sse_shufps(cp->func, data, aos_get_internal_xmm( cp, IMM_IDENTITY ), SHUF(X,Y,Z,W) );
+   /* data = z ? 0 1 */
    sse_shufps(cp->func, data, data, SHUF(Y,Z,X,W) );
+   /* data = ? 0 z 1 */
    sse_movlps(cp->func, data, src_ptr);
+   /* data = x y z 1 */
 }
 
 static void emit_load_R32G32( struct aos_compilation *cp,