Merge remote branch 'origin/gallium-0.1' into nouveau-gallium-0.1

9 files changed, 139 insertions, 158 deletions

diff --git a/src/gallium/drivers/cell/spu/spu_exec.c b/src/gallium/drivers/cell/spu/spu_exec.c
index 96393732ed8..42e5022f30b 100644
--- a/src/gallium/drivers/cell/spu/spu_exec.c
+++ b/src/gallium/drivers/cell/spu/spu_exec.c
@@ -603,8 +603,8 @@ store_dest(
  * Kill fragment if any of the four values is less than zero.
  */
 static void
-exec_kilp(struct spu_exec_machine *mach,
-          const struct tgsi_full_instruction *inst)
+exec_kil(struct spu_exec_machine *mach,
+         const struct tgsi_full_instruction *inst)
 {
    uint uniquemask;
    uint chan_index;
@@ -640,6 +640,20 @@ exec_kilp(struct spu_exec_machine *mach,
    mach->Temps[TEMP_KILMASK_I].xyzw[TEMP_KILMASK_C].u[0] |= kilmask;
 }
 
+/**
+ * Execute NVIDIA-style KIL which is predicated by a condition code.
+ * Kill fragment if the condition code is TRUE.
+ */
+static void
+exec_kilp(struct tgsi_exec_machine *mach,
+          const struct tgsi_full_instruction *inst)
+{
+   uint kilmask = 0; /* bit 0 = pixel 0, bit 1 = pixel 1, etc */
+
+   /* TODO: build kilmask from CC mask */
+
+   mach->Temps[TEMP_KILMASK_I].xyzw[TEMP_KILMASK_C].u[0] |= kilmask;
+}
 
 /*
  * Fetch a texel using STR texture coordinates.
@@ -1336,8 +1350,7 @@ exec_instruction(
       break;
 
    case TGSI_OPCODE_KIL:
-      /* for enabled ExecMask bits, set the killed bit */
-      mach->Temps[TEMP_KILMASK_I].xyzw[TEMP_KILMASK_C].u[0] |= mach->ExecMask;
+      exec_kil (mach, inst);
       break;
 
    case TGSI_OPCODE_PK2H:
diff --git a/src/gallium/drivers/i915simple/i915_fpc_translate.c b/src/gallium/drivers/i915simple/i915_fpc_translate.c
index 04507ab8adc..64432982c42 100644
--- a/src/gallium/drivers/i915simple/i915_fpc_translate.c
+++ b/src/gallium/drivers/i915simple/i915_fpc_translate.c
@@ -560,18 +560,6 @@ i915_translate_instruction(struct i915_fp_compile *p,
       break;
 
    case TGSI_OPCODE_KIL:
-      /* unconditional kill */
-      assert(0); /* not tested yet */
-#if 0
-      src0 = src_vector(p, &inst->FullSrcRegisters[0]);
-      tmp = i915_get_utemp(p);
-
-      i915_emit_texld(p, tmp, A0_DEST_CHANNEL_ALL,   /* use a dummy dest reg */
-                      0, src0, T0_TEXKILL);
-#endif
-      break;
-
-   case TGSI_OPCODE_KILP:
       /* kill if src[0].x < 0 || src[0].y < 0 ... */
       src0 = src_vector(p, &inst->FullSrcRegisters[0]);
       tmp = i915_get_utemp(p);
@@ -584,6 +572,10 @@ i915_translate_instruction(struct i915_fp_compile *p,
                       T0_TEXKILL);           /* opcode */
       break;
 
+   case TGSI_OPCODE_KILP:
+      assert(0); /* not tested yet */
+      break;
+
    case TGSI_OPCODE_LG2:
       src0 = src_vector(p, &inst->FullSrcRegisters[0]);
 
diff --git a/src/gallium/drivers/softpipe/sp_setup.c b/src/gallium/drivers/softpipe/sp_setup.c
index b48eec730b2..b7f2f16307e 100644
--- a/src/gallium/drivers/softpipe/sp_setup.c
+++ b/src/gallium/drivers/softpipe/sp_setup.c
@@ -39,6 +39,7 @@
 #include "sp_quad.h"
 #include "sp_state.h"
 #include "sp_prim_setup.h"
+#include "draw/draw_context.h"
 #include "draw/draw_private.h"
 #include "draw/draw_vertex.h"
 #include "pipe/p_util.h"
@@ -1272,10 +1273,7 @@ void setup_prepare( struct setup_context *setup )
    }
 
    /* Note: nr_attrs is only used for debugging (vertex printing) */
-   {
-      const struct sp_fragment_shader *fs = setup->softpipe->fs;
-      setup->quad.nr_attrs = fs->info.num_inputs + 1; /* +1 for vert pos */
-   }
+   setup->quad.nr_attrs = draw_num_vs_outputs(sp->draw);
 
    sp->quad.first->begin(sp->quad.first);
 
diff --git a/src/gallium/drivers/trace/tr_context.c b/src/gallium/drivers/trace/tr_context.c
index 47a217ec7c7..f82126370cf 100644
--- a/src/gallium/drivers/trace/tr_context.c
+++ b/src/gallium/drivers/trace/tr_context.c
@@ -40,8 +40,7 @@ trace_context_set_edgeflags(struct pipe_context *_pipe,
                             const unsigned *bitfield)
 {
    struct trace_context *tr_ctx = trace_context(_pipe);
-   struct trace_screen *tr_scr = trace_screen(_pipe->screen);
-   struct trace_stream *stream = tr_scr->stream;
+   struct trace_stream *stream = tr_ctx->stream;
    struct pipe_context *pipe = tr_ctx->pipe;
 
    trace_dump_call_begin(stream, "pipe_context", "set_edgeflags");
@@ -61,8 +60,7 @@ trace_context_draw_arrays(struct pipe_context *_pipe,
                           unsigned mode, unsigned start, unsigned count)
 {
    struct trace_context *tr_ctx = trace_context(_pipe);
-   struct trace_screen *tr_scr = trace_screen(_pipe->screen);
-   struct trace_stream *stream = tr_scr->stream;
+   struct trace_stream *stream = tr_ctx->stream;
    struct pipe_context *pipe = tr_ctx->pipe;
    boolean result;
 
@@ -90,8 +88,7 @@ trace_context_draw_elements(struct pipe_context *_pipe,
                           unsigned mode, unsigned start, unsigned count)
 {
    struct trace_context *tr_ctx = trace_context(_pipe);
-   struct trace_screen *tr_scr = trace_screen(_pipe->screen);
-   struct trace_stream *stream = tr_scr->stream;
+   struct trace_stream *stream = tr_ctx->stream;
    struct pipe_context *pipe = tr_ctx->pipe;
    boolean result;
 
@@ -125,8 +122,7 @@ trace_context_draw_range_elements(struct pipe_context *_pipe,
                                   unsigned count)
 {
    struct trace_context *tr_ctx = trace_context(_pipe);
-   struct trace_screen *tr_scr = trace_screen(_pipe->screen);
-   struct trace_stream *stream = tr_scr->stream;
+   struct trace_stream *stream = tr_ctx->stream;
    struct pipe_context *pipe = tr_ctx->pipe;
    boolean result;
 
@@ -159,8 +155,7 @@ trace_context_create_query(struct pipe_context *_pipe,
                            unsigned query_type)
 {
    struct trace_context *tr_ctx = trace_context(_pipe);
-   struct trace_screen *tr_scr = trace_screen(_pipe->screen);
-   struct trace_stream *stream = tr_scr->stream;
+   struct trace_stream *stream = tr_ctx->stream;
    struct pipe_context *pipe = tr_ctx->pipe;
    struct pipe_query *result;
 
@@ -184,8 +179,7 @@ trace_context_destroy_query(struct pipe_context *_pipe,
                             struct pipe_query *query)
 {
    struct trace_context *tr_ctx = trace_context(_pipe);
-   struct trace_screen *tr_scr = trace_screen(_pipe->screen);
-   struct trace_stream *stream = tr_scr->stream;
+   struct trace_stream *stream = tr_ctx->stream;
    struct pipe_context *pipe = tr_ctx->pipe;
 
    trace_dump_call_begin(stream, "pipe_context", "destroy_query");
@@ -204,8 +198,7 @@ trace_context_begin_query(struct pipe_context *_pipe,
                           struct pipe_query *query)
 {
    struct trace_context *tr_ctx = trace_context(_pipe);
-   struct trace_screen *tr_scr = trace_screen(_pipe->screen);
-   struct trace_stream *stream = tr_scr->stream;
+   struct trace_stream *stream = tr_ctx->stream;
    struct pipe_context *pipe = tr_ctx->pipe;
 
    trace_dump_call_begin(stream, "pipe_context", "begin_query");
@@ -224,8 +217,7 @@ trace_context_end_query(struct pipe_context *_pipe,
                         struct pipe_query *query)
 {
    struct trace_context *tr_ctx = trace_context(_pipe);
-   struct trace_screen *tr_scr = trace_screen(_pipe->screen);
-   struct trace_stream *stream = tr_scr->stream;
+   struct trace_stream *stream = tr_ctx->stream;
    struct pipe_context *pipe = tr_ctx->pipe;
 
    trace_dump_call_begin(stream, "pipe_context", "end_query");
@@ -246,8 +238,7 @@ trace_context_get_query_result(struct pipe_context *_pipe,
                                uint64 *presult)
 {
    struct trace_context *tr_ctx = trace_context(_pipe);
-   struct trace_screen *tr_scr = trace_screen(_pipe->screen);
-   struct trace_stream *stream = tr_scr->stream;
+   struct trace_stream *stream = tr_ctx->stream;
    struct pipe_context *pipe = tr_ctx->pipe;
    uint64 result;
    boolean _result;
@@ -273,8 +264,7 @@ trace_context_create_blend_state(struct pipe_context *_pipe,
                                  const struct pipe_blend_state *state)
 {
    struct trace_context *tr_ctx = trace_context(_pipe);
-   struct trace_screen *tr_scr = trace_screen(_pipe->screen);
-   struct trace_stream *stream = tr_scr->stream;
+   struct trace_stream *stream = tr_ctx->stream;
    struct pipe_context *pipe = tr_ctx->pipe;
    void * result;
 
@@ -285,6 +275,8 @@ trace_context_create_blend_state(struct pipe_context *_pipe,
 
    result = pipe->create_blend_state(pipe, state);;
 
+   trace_dump_ret(stream, ptr, result);
+
    trace_dump_call_end(stream);
    
    return result;
@@ -296,8 +288,7 @@ trace_context_bind_blend_state(struct pipe_context *_pipe,
                                void *state)
 {
    struct trace_context *tr_ctx = trace_context(_pipe);
-   struct trace_screen *tr_scr = trace_screen(_pipe->screen);
-   struct trace_stream *stream = tr_scr->stream;
+   struct trace_stream *stream = tr_ctx->stream;
    struct pipe_context *pipe = tr_ctx->pipe;
 
    trace_dump_call_begin(stream, "pipe_context", "bind_blend_state");
@@ -316,8 +307,7 @@ trace_context_delete_blend_state(struct pipe_context *_pipe,
                                  void *state)
 {
    struct trace_context *tr_ctx = trace_context(_pipe);
-   struct trace_screen *tr_scr = trace_screen(_pipe->screen);
-   struct trace_stream *stream = tr_scr->stream;
+   struct trace_stream *stream = tr_ctx->stream;
    struct pipe_context *pipe = tr_ctx->pipe;
 
    trace_dump_call_begin(stream, "pipe_context", "delete_blend_state");
@@ -336,8 +326,7 @@ trace_context_create_sampler_state(struct pipe_context *_pipe,
                                    const struct pipe_sampler_state *state)
 {
    struct trace_context *tr_ctx = trace_context(_pipe);
-   struct trace_screen *tr_scr = trace_screen(_pipe->screen);
-   struct trace_stream *stream = tr_scr->stream;
+   struct trace_stream *stream = tr_ctx->stream;
    struct pipe_context *pipe = tr_ctx->pipe;
    void * result;
 
@@ -361,8 +350,7 @@ trace_context_bind_sampler_states(struct pipe_context *_pipe,
                                   unsigned num_states, void **states)
 {
    struct trace_context *tr_ctx = trace_context(_pipe);
-   struct trace_screen *tr_scr = trace_screen(_pipe->screen);
-   struct trace_stream *stream = tr_scr->stream;
+   struct trace_stream *stream = tr_ctx->stream;
    struct pipe_context *pipe = tr_ctx->pipe;
 
    trace_dump_call_begin(stream, "pipe_context", "bind_sampler_states");
@@ -382,8 +370,7 @@ trace_context_delete_sampler_state(struct pipe_context *_pipe,
                                    void *state)
 {
    struct trace_context *tr_ctx = trace_context(_pipe);
-   struct trace_screen *tr_scr = trace_screen(_pipe->screen);
-   struct trace_stream *stream = tr_scr->stream;
+   struct trace_stream *stream = tr_ctx->stream;
    struct pipe_context *pipe = tr_ctx->pipe;
 
    trace_dump_call_begin(stream, "pipe_context", "delete_sampler_state");
@@ -402,8 +389,7 @@ trace_context_create_rasterizer_state(struct pipe_context *_pipe,
                                       const struct pipe_rasterizer_state *state)
 {
    struct trace_context *tr_ctx = trace_context(_pipe);
-   struct trace_screen *tr_scr = trace_screen(_pipe->screen);
-   struct trace_stream *stream = tr_scr->stream;
+   struct trace_stream *stream = tr_ctx->stream;
    struct pipe_context *pipe = tr_ctx->pipe;
    void * result;
 
@@ -427,8 +413,7 @@ trace_context_bind_rasterizer_state(struct pipe_context *_pipe,
                                     void *state)
 {
    struct trace_context *tr_ctx = trace_context(_pipe);
-   struct trace_screen *tr_scr = trace_screen(_pipe->screen);
-   struct trace_stream *stream = tr_scr->stream;
+   struct trace_stream *stream = tr_ctx->stream;
    struct pipe_context *pipe = tr_ctx->pipe;
 
    trace_dump_call_begin(stream, "pipe_context", "bind_rasterizer_state");
@@ -447,8 +432,7 @@ trace_context_delete_rasterizer_state(struct pipe_context *_pipe,
                                       void *state)
 {
    struct trace_context *tr_ctx = trace_context(_pipe);
-   struct trace_screen *tr_scr = trace_screen(_pipe->screen);
-   struct trace_stream *stream = tr_scr->stream;
+   struct trace_stream *stream = tr_ctx->stream;
    struct pipe_context *pipe = tr_ctx->pipe;
 
    trace_dump_call_begin(stream, "pipe_context", "delete_rasterizer_state");
@@ -467,8 +451,7 @@ trace_context_create_depth_stencil_alpha_state(struct pipe_context *_pipe,
                                                const struct pipe_depth_stencil_alpha_state *state)
 {
    struct trace_context *tr_ctx = trace_context(_pipe);
-   struct trace_screen *tr_scr = trace_screen(_pipe->screen);
-   struct trace_stream *stream = tr_scr->stream;
+   struct trace_stream *stream = tr_ctx->stream;
    struct pipe_context *pipe = tr_ctx->pipe;
    void * result;
 
@@ -476,8 +459,10 @@ trace_context_create_depth_stencil_alpha_state(struct pipe_context *_pipe,
 
    result = pipe->create_depth_stencil_alpha_state(pipe, state);;
 
-   trace_dump_ret(stream, ptr, result);
+   trace_dump_arg(stream, ptr, pipe);
    trace_dump_arg(stream, depth_stencil_alpha_state, state);
+   
+   trace_dump_ret(stream, ptr, result);
 
    trace_dump_call_end(stream);
    
@@ -490,8 +475,7 @@ trace_context_bind_depth_stencil_alpha_state(struct pipe_context *_pipe,
                                              void *state)
 {
    struct trace_context *tr_ctx = trace_context(_pipe);
-   struct trace_screen *tr_scr = trace_screen(_pipe->screen);
-   struct trace_stream *stream = tr_scr->stream;
+   struct trace_stream *stream = tr_ctx->stream;
    struct pipe_context *pipe = tr_ctx->pipe;
 
    trace_dump_call_begin(stream, "pipe_context", "bind_depth_stencil_alpha_state");
@@ -510,8 +494,7 @@ trace_context_delete_depth_stencil_alpha_state(struct pipe_context *_pipe,
                                                void *state)
 {
    struct trace_context *tr_ctx = trace_context(_pipe);
-   struct trace_screen *tr_scr = trace_screen(_pipe->screen);
-   struct trace_stream *stream = tr_scr->stream;
+   struct trace_stream *stream = tr_ctx->stream;
    struct pipe_context *pipe = tr_ctx->pipe;
 
    trace_dump_call_begin(stream, "pipe_context", "delete_depth_stencil_alpha_state");
@@ -530,8 +513,7 @@ trace_context_create_fs_state(struct pipe_context *_pipe,
                               const struct pipe_shader_state *state)
 {
    struct trace_context *tr_ctx = trace_context(_pipe);
-   struct trace_screen *tr_scr = trace_screen(_pipe->screen);
-   struct trace_stream *stream = tr_scr->stream;
+   struct trace_stream *stream = tr_ctx->stream;
    struct pipe_context *pipe = tr_ctx->pipe;
    void * result;
 
@@ -555,8 +537,7 @@ trace_context_bind_fs_state(struct pipe_context *_pipe,
                             void *state)
 {
    struct trace_context *tr_ctx = trace_context(_pipe);
-   struct trace_screen *tr_scr = trace_screen(_pipe->screen);
-   struct trace_stream *stream = tr_scr->stream;
+   struct trace_stream *stream = tr_ctx->stream;
    struct pipe_context *pipe = tr_ctx->pipe;
 
    trace_dump_call_begin(stream, "pipe_context", "bind_fs_state");
@@ -575,8 +556,7 @@ trace_context_delete_fs_state(struct pipe_context *_pipe,
                               void *state)
 {
    struct trace_context *tr_ctx = trace_context(_pipe);
-   struct trace_screen *tr_scr = trace_screen(_pipe->screen);
-   struct trace_stream *stream = tr_scr->stream;
+   struct trace_stream *stream = tr_ctx->stream;
    struct pipe_context *pipe = tr_ctx->pipe;
 
    trace_dump_call_begin(stream, "pipe_context", "delete_fs_state");
@@ -595,8 +575,7 @@ trace_context_create_vs_state(struct pipe_context *_pipe,
                               const struct pipe_shader_state *state)
 {
    struct trace_context *tr_ctx = trace_context(_pipe);
-   struct trace_screen *tr_scr = trace_screen(_pipe->screen);
-   struct trace_stream *stream = tr_scr->stream;
+   struct trace_stream *stream = tr_ctx->stream;
    struct pipe_context *pipe = tr_ctx->pipe;
    void * result;
 
@@ -620,8 +599,7 @@ trace_context_bind_vs_state(struct pipe_context *_pipe,
                             void *state)
 {
    struct trace_context *tr_ctx = trace_context(_pipe);
-   struct trace_screen *tr_scr = trace_screen(_pipe->screen);
-   struct trace_stream *stream = tr_scr->stream;
+   struct trace_stream *stream = tr_ctx->stream;
    struct pipe_context *pipe = tr_ctx->pipe;
 
    trace_dump_call_begin(stream, "pipe_context", "bind_vs_state");
@@ -640,8 +618,7 @@ trace_context_delete_vs_state(struct pipe_context *_pipe,
                               void *state)
 {
    struct trace_context *tr_ctx = trace_context(_pipe);
-   struct trace_screen *tr_scr = trace_screen(_pipe->screen);
-   struct trace_stream *stream = tr_scr->stream;
+   struct trace_stream *stream = tr_ctx->stream;
    struct pipe_context *pipe = tr_ctx->pipe;
 
    trace_dump_call_begin(stream, "pipe_context", "delete_vs_state");
@@ -660,8 +637,7 @@ trace_context_set_blend_color(struct pipe_context *_pipe,
                               const struct pipe_blend_color *state)
 {
    struct trace_context *tr_ctx = trace_context(_pipe);
-   struct trace_screen *tr_scr = trace_screen(_pipe->screen);
-   struct trace_stream *stream = tr_scr->stream;
+   struct trace_stream *stream = tr_ctx->stream;
    struct pipe_context *pipe = tr_ctx->pipe;
 
    trace_dump_call_begin(stream, "pipe_context", "set_blend_color");
@@ -680,8 +656,7 @@ trace_context_set_clip_state(struct pipe_context *_pipe,
                              const struct pipe_clip_state *state)
 {
    struct trace_context *tr_ctx = trace_context(_pipe);
-   struct trace_screen *tr_scr = trace_screen(_pipe->screen);
-   struct trace_stream *stream = tr_scr->stream;
+   struct trace_stream *stream = tr_ctx->stream;
    struct pipe_context *pipe = tr_ctx->pipe;
 
    trace_dump_call_begin(stream, "pipe_context", "set_clip_state");
@@ -701,8 +676,7 @@ trace_context_set_constant_buffer(struct pipe_context *_pipe,
                                   const struct pipe_constant_buffer *buffer)
 {
    struct trace_context *tr_ctx = trace_context(_pipe);
-   struct trace_screen *tr_scr = trace_screen(_pipe->screen);
-   struct trace_stream *stream = tr_scr->stream;
+   struct trace_stream *stream = tr_ctx->stream;
    struct pipe_context *pipe = tr_ctx->pipe;
 
    trace_dump_call_begin(stream, "pipe_context", "set_constant_buffer");
@@ -723,8 +697,7 @@ trace_context_set_framebuffer_state(struct pipe_context *_pipe,
                                     const struct pipe_framebuffer_state *state)
 {
    struct trace_context *tr_ctx = trace_context(_pipe);
-   struct trace_screen *tr_scr = trace_screen(_pipe->screen);
-   struct trace_stream *stream = tr_scr->stream;
+   struct trace_stream *stream = tr_ctx->stream;
    struct pipe_context *pipe = tr_ctx->pipe;
 
    trace_dump_call_begin(stream, "pipe_context", "set_framebuffer_state");
@@ -743,8 +716,7 @@ trace_context_set_polygon_stipple(struct pipe_context *_pipe,
                                   const struct pipe_poly_stipple *state)
 {
    struct trace_context *tr_ctx = trace_context(_pipe);
-   struct trace_screen *tr_scr = trace_screen(_pipe->screen);
-   struct trace_stream *stream = tr_scr->stream;
+   struct trace_stream *stream = tr_ctx->stream;
    struct pipe_context *pipe = tr_ctx->pipe;
 
    trace_dump_call_begin(stream, "pipe_context", "set_polygon_stipple");
@@ -763,8 +735,7 @@ trace_context_set_scissor_state(struct pipe_context *_pipe,
                                 const struct pipe_scissor_state *state)
 {
    struct trace_context *tr_ctx = trace_context(_pipe);
-   struct trace_screen *tr_scr = trace_screen(_pipe->screen);
-   struct trace_stream *stream = tr_scr->stream;
+   struct trace_stream *stream = tr_ctx->stream;
    struct pipe_context *pipe = tr_ctx->pipe;
 
    trace_dump_call_begin(stream, "pipe_context", "set_scissor_state");
@@ -783,8 +754,7 @@ trace_context_set_viewport_state(struct pipe_context *_pipe,
                                  const struct pipe_viewport_state *state)
 {
    struct trace_context *tr_ctx = trace_context(_pipe);
-   struct trace_screen *tr_scr = trace_screen(_pipe->screen);
-   struct trace_stream *stream = tr_scr->stream;
+   struct trace_stream *stream = tr_ctx->stream;
    struct pipe_context *pipe = tr_ctx->pipe;
 
    trace_dump_call_begin(stream, "pipe_context", "set_viewport_state");
@@ -804,8 +774,7 @@ trace_context_set_sampler_textures(struct pipe_context *_pipe,
                                    struct pipe_texture **textures)
 {
    struct trace_context *tr_ctx = trace_context(_pipe);
-   struct trace_screen *tr_scr = trace_screen(_pipe->screen);
-   struct trace_stream *stream = tr_scr->stream;
+   struct trace_stream *stream = tr_ctx->stream;
    struct pipe_context *pipe = tr_ctx->pipe;
 
    trace_dump_call_begin(stream, "pipe_context", "set_sampler_textures");
@@ -826,8 +795,7 @@ trace_context_set_vertex_buffers(struct pipe_context *_pipe,
                                  const struct pipe_vertex_buffer *buffers)
 {
    struct trace_context *tr_ctx = trace_context(_pipe);
-   struct trace_screen *tr_scr = trace_screen(_pipe->screen);
-   struct trace_stream *stream = tr_scr->stream;
+   struct trace_stream *stream = tr_ctx->stream;
    struct pipe_context *pipe = tr_ctx->pipe;
 
    trace_dump_call_begin(stream, "pipe_context", "set_vertex_buffers");
@@ -851,15 +819,13 @@ trace_context_set_vertex_elements(struct pipe_context *_pipe,
                                   const struct pipe_vertex_element *elements)
 {
    struct trace_context *tr_ctx = trace_context(_pipe);
-   struct trace_screen *tr_scr = trace_screen(_pipe->screen);
-   struct trace_stream *stream = tr_scr->stream;
+   struct trace_stream *stream = tr_ctx->stream;
    struct pipe_context *pipe = tr_ctx->pipe;
 
    trace_dump_call_begin(stream, "pipe_context", "set_vertex_elements");
 
    trace_dump_arg(stream, ptr, pipe);
    trace_dump_arg(stream, uint, num_elements);
-   trace_dump_arg(stream, ptr, elements);
 
    trace_dump_arg_begin(stream, "elements");
    trace_dump_struct_array(stream, vertex_element, elements, num_elements);
@@ -881,8 +847,7 @@ trace_context_surface_copy(struct pipe_context *_pipe,
                            unsigned width, unsigned height)
 {
    struct trace_context *tr_ctx = trace_context(_pipe);
-   struct trace_screen *tr_scr = trace_screen(_pipe->screen);
-   struct trace_stream *stream = tr_scr->stream;
+   struct trace_stream *stream = tr_ctx->stream;
    struct pipe_context *pipe = tr_ctx->pipe;
 
    trace_dump_call_begin(stream, "pipe_context", "surface_copy");
@@ -914,8 +879,7 @@ trace_context_surface_fill(struct pipe_context *_pipe,
                            unsigned value)
 {
    struct trace_context *tr_ctx = trace_context(_pipe);
-   struct trace_screen *tr_scr = trace_screen(_pipe->screen);
-   struct trace_stream *stream = tr_scr->stream;
+   struct trace_stream *stream = tr_ctx->stream;
    struct pipe_context *pipe = tr_ctx->pipe;
 
    trace_dump_call_begin(stream, "pipe_context", "surface_fill");
@@ -939,8 +903,7 @@ trace_context_clear(struct pipe_context *_pipe,
                     unsigned clearValue)
 {
    struct trace_context *tr_ctx = trace_context(_pipe);
-   struct trace_screen *tr_scr = trace_screen(_pipe->screen);
-   struct trace_stream *stream = tr_scr->stream;
+   struct trace_stream *stream = tr_ctx->stream;
    struct pipe_context *pipe = tr_ctx->pipe;
 
    trace_dump_call_begin(stream, "pipe_context", "clear");
@@ -961,8 +924,7 @@ trace_context_flush(struct pipe_context *_pipe,
                     struct pipe_fence_handle **fence)
 {
    struct trace_context *tr_ctx = trace_context(_pipe);
-   struct trace_screen *tr_scr = trace_screen(_pipe->screen);
-   struct trace_stream *stream = tr_scr->stream;
+   struct trace_stream *stream = tr_ctx->stream;
    struct pipe_context *pipe = tr_ctx->pipe;
 
    trace_dump_call_begin(stream, "pipe_context", "flush");
@@ -981,8 +943,7 @@ static INLINE void
 trace_context_destroy(struct pipe_context *_pipe)
 {
    struct trace_context *tr_ctx = trace_context(_pipe);
-   struct trace_screen *tr_scr = trace_screen(_pipe->screen);
-   struct trace_stream *stream = tr_scr->stream;
+   struct trace_stream *stream = tr_ctx->stream;
    struct pipe_context *pipe = tr_ctx->pipe;
 
    trace_dump_call_begin(stream, "pipe_context", "destroy");
@@ -1000,6 +961,7 @@ trace_context_destroy(struct pipe_context *_pipe)
 struct pipe_context *
 trace_context_create(struct pipe_context *pipe)
 {
+   struct trace_stream *stream;
    struct trace_context *tr_ctx;
    
    if(!debug_get_bool_option("GALLIUM_TRACE", FALSE))
@@ -1055,10 +1017,18 @@ trace_context_create(struct pipe_context *pipe)
    tr_ctx->base.flush = trace_context_flush;
 
    tr_ctx->pipe = pipe;
+   tr_ctx->stream = stream = trace_winsys(pipe->winsys)->stream;
    
    /* We don't want to trace the internal pipe calls */
    pipe->winsys = trace_winsys(pipe->winsys)->winsys;
    pipe->screen = trace_screen(pipe->screen)->screen;
    
+   trace_dump_call_begin(stream, "", "pipe_context_create");
+   trace_dump_arg_begin(stream, "screen");
+   trace_dump_ptr(stream, pipe->screen);
+   trace_dump_arg_end(stream);
+   trace_dump_ret(stream, ptr, pipe);
+   trace_dump_call_end(stream);
+
    return &tr_ctx->base;
 }
diff --git a/src/gallium/drivers/trace/tr_context.h b/src/gallium/drivers/trace/tr_context.h
index 80fb5980d6a..2c0b0c72e45 100644
--- a/src/gallium/drivers/trace/tr_context.h
+++ b/src/gallium/drivers/trace/tr_context.h
@@ -34,11 +34,16 @@
 #include "pipe/p_context.h"
 
 
+struct trace_stream;
+
+   
 struct trace_context
 {
    struct pipe_context base;
    
    struct pipe_context *pipe;
+   
+   struct trace_stream *stream;
 };
 
 
diff --git a/src/gallium/drivers/trace/tr_dump.c b/src/gallium/drivers/trace/tr_dump.c
index f545de30f4f..5269c4ddc84 100644
--- a/src/gallium/drivers/trace/tr_dump.c
+++ b/src/gallium/drivers/trace/tr_dump.c
@@ -276,17 +276,16 @@ void trace_dump_bytes(struct trace_stream *stream,
                       const void *data,
                       long unsigned size)
 {
-   static char hex_table[] = "0123456789ABCDE";
+   static const char hex_table[16] = "0123456789ABCDEF";
    const uint8_t *p = data;
    long unsigned i;
    trace_dump_write(stream, "<bytes>");
    for(i = 0; i < size; ++i) {
       uint8_t byte = *p++;
-      char str[3];
-      str[0] = hex_table[byte >> 4];
-      str[1] = hex_table[byte & 0xf];
-      str[2] = 0;
-      trace_dump_write(stream, str);
+      char hex[2];
+      hex[0] = hex_table[byte >> 4];
+      hex[1] = hex_table[byte & 0xf];
+      trace_stream_write(stream, hex, 2);
    }
    trace_dump_write(stream, "</bytes>");
 }
diff --git a/src/gallium/drivers/trace/tr_screen.c b/src/gallium/drivers/trace/tr_screen.c
index de885abae2c..27c218039ec 100644
--- a/src/gallium/drivers/trace/tr_screen.c
+++ b/src/gallium/drivers/trace/tr_screen.c
@@ -346,6 +346,7 @@ trace_screen_destroy(struct pipe_screen *_screen)
 struct pipe_screen *
 trace_screen_create(struct pipe_screen *screen)
 {
+   struct trace_stream *stream;
    struct trace_screen *tr_scr;
    
    if(!debug_get_bool_option("GALLIUM_TRACE", FALSE))
@@ -371,13 +372,17 @@ trace_screen_create(struct pipe_screen *screen)
    tr_scr->base.surface_unmap = trace_screen_surface_unmap;
    
    tr_scr->screen = screen;
-      
-   tr_scr->stream = trace_winsys(screen->winsys)->stream;
-   if(!tr_scr->stream)
-      return NULL;
+   tr_scr->stream = stream = trace_winsys(screen->winsys)->stream;
 
    /* We don't want to trace the internal pipe calls */
    screen->winsys = trace_winsys(screen->winsys)->winsys;
 
+   trace_dump_call_begin(stream, "", "pipe_screen_create");
+   trace_dump_arg_begin(stream, "winsys");
+   trace_dump_ptr(stream, screen->winsys);
+   trace_dump_arg_end(stream);
+   trace_dump_ret(stream, ptr, screen);
+   trace_dump_call_end(stream);
+
    return &tr_scr->base;
 }
diff --git a/src/gallium/drivers/trace/tr_state.c b/src/gallium/drivers/trace/tr_state.c
index 9ffe77146d9..e074ae7abc1 100644
--- a/src/gallium/drivers/trace/tr_state.c
+++ b/src/gallium/drivers/trace/tr_state.c
@@ -73,6 +73,10 @@ void trace_dump_template(struct trace_stream *stream,
    trace_dump_array(stream, uint, templat->height, 1);
    trace_dump_member_end(stream);
 
+   trace_dump_member_begin(stream, "depth");
+   trace_dump_array(stream, uint, templat->depth, 1);
+   trace_dump_member_end(stream);
+
    trace_dump_member_begin(stream, "block");
    trace_dump_block(stream, &templat->block);
    trace_dump_member_end(stream);
@@ -146,9 +150,10 @@ void trace_dump_poly_stipple(struct trace_stream *stream,
    trace_dump_struct_begin(stream, "pipe_poly_stipple");
 
    trace_dump_member_begin(stream, "stipple");
-   trace_dump_bytes(stream, 
+   trace_dump_array(stream,
+                    uint,
                     state->stipple, 
-                    sizeof(state->stipple));
+                    Elements(state->stipple));
    trace_dump_member_end(stream);
    
    trace_dump_struct_end(stream);
@@ -273,7 +278,7 @@ void trace_dump_depth_stencil_alpha_state(struct trace_stream *stream,
    trace_dump_struct_begin(stream, "pipe_depth_stencil_alpha_state");
 
    trace_dump_member_begin(stream, "depth");
-   trace_dump_struct_begin(stream, "");
+   trace_dump_struct_begin(stream, "pipe_depth_state");
    trace_dump_member(stream, bool, &state->depth, enabled);
    trace_dump_member(stream, bool, &state->depth, writemask);
    trace_dump_member(stream, uint, &state->depth, func);
@@ -285,7 +290,7 @@ void trace_dump_depth_stencil_alpha_state(struct trace_stream *stream,
    trace_dump_array_begin(stream);
    for(i = 0; i < Elements(state->stencil); ++i) {
       trace_dump_elem_begin(stream);
-      trace_dump_struct_begin(stream, "");
+      trace_dump_struct_begin(stream, "pipe_stencil_state");
       trace_dump_member(stream, bool, &state->stencil[i], enabled);
       trace_dump_member(stream, uint, &state->stencil[i], func);
       trace_dump_member(stream, uint, &state->stencil[i], fail_op);
@@ -301,7 +306,7 @@ void trace_dump_depth_stencil_alpha_state(struct trace_stream *stream,
    trace_dump_member_end(stream);
 
    trace_dump_member_begin(stream, "alpha");
-   trace_dump_struct_begin(stream, "");
+   trace_dump_struct_begin(stream, "pipe_alpha_state");
    trace_dump_member(stream, bool, &state->alpha, enabled);
    trace_dump_member(stream, uint, &state->alpha, func);
    trace_dump_member(stream, float, &state->alpha, ref);
diff --git a/src/gallium/drivers/trace/tr_winsys.c b/src/gallium/drivers/trace/tr_winsys.c
index 964da5677b2..eec84a19816 100644
--- a/src/gallium/drivers/trace/tr_winsys.c
+++ b/src/gallium/drivers/trace/tr_winsys.c
@@ -26,6 +26,7 @@
  **************************************************************************/
 
 #include "pipe/p_util.h"
+#include "pipe/p_state.h"
 #include "util/u_hash_table.h"
 
 #include "tr_stream.h"
@@ -178,7 +179,7 @@ trace_winsys_buffer_create(struct pipe_winsys *_winsys,
    struct trace_winsys *tr_ws = trace_winsys(_winsys);
    struct trace_stream *stream = tr_ws->stream;
    struct pipe_winsys *winsys = tr_ws->winsys;
-   struct pipe_buffer *result;
+   struct pipe_buffer *buffer;
    
    trace_dump_call_begin(stream, "pipe_winsys", "buffer_create");
    
@@ -187,13 +188,23 @@ trace_winsys_buffer_create(struct pipe_winsys *_winsys,
    trace_dump_arg(stream, uint, usage);
    trace_dump_arg(stream, uint, size);
 
-   result = winsys->buffer_create(winsys, alignment, usage, size);
+   buffer = winsys->buffer_create(winsys, alignment, usage, size);
    
-   trace_dump_ret(stream, ptr, result);
+   trace_dump_ret(stream, ptr, buffer);
    
    trace_dump_call_end(stream);
+
+   /* Zero the buffer to avoid dumping uninitialized memory */
+   if(buffer->usage & PIPE_BUFFER_USAGE_CPU_WRITE) {
+      void *map;
+      map = winsys->buffer_map(winsys, buffer, PIPE_BUFFER_USAGE_CPU_WRITE);
+      if(map) {
+         memset(map, 0, buffer->size);
+         winsys->buffer_unmap(winsys, buffer);
+      }
+   }
    
-   return result;
+   return buffer;
 }
 
 
@@ -229,30 +240,18 @@ trace_winsys_buffer_map(struct pipe_winsys *_winsys,
                         unsigned usage)
 {
    struct trace_winsys *tr_ws = trace_winsys(_winsys);
-   struct trace_stream *stream = tr_ws->stream;
    struct pipe_winsys *winsys = tr_ws->winsys;
-   void *result;
-   
-   trace_dump_call_begin(stream, "pipe_winsys", "buffer_map");
+   void *map;
    
-   trace_dump_arg(stream, ptr, winsys);
-   trace_dump_arg(stream, ptr, buffer);
-   trace_dump_arg(stream, uint, usage);
-
-   result = winsys->buffer_map(winsys, buffer, usage);
-   
-   trace_dump_ret(stream, ptr, result);
-   
-   trace_dump_call_end(stream);
-   
-   if(result) {
+   map = winsys->buffer_map(winsys, buffer, usage);
+   if(map) {
       if(usage & PIPE_BUFFER_USAGE_CPU_WRITE) {
          assert(!hash_table_get(tr_ws->buffer_maps, buffer));
-         hash_table_set(tr_ws->buffer_maps, buffer, result);
+         hash_table_set(tr_ws->buffer_maps, buffer, map);
       }
    }
    
-   return result;
+   return map;
 }
 
 
@@ -267,13 +266,13 @@ trace_winsys_buffer_unmap(struct pipe_winsys *_winsys,
    
    map = hash_table_get(tr_ws->buffer_maps, buffer);
    if(map) {
-      trace_dump_call_begin(stream, "", "memcpy");
+      trace_dump_call_begin(stream, "pipe_winsys", "buffer_write");
       
-      trace_dump_arg_begin(stream, "dst");
-      trace_dump_ptr(stream, map);
-      trace_dump_arg_end(stream);
+      trace_dump_arg(stream, ptr, winsys);
       
-      trace_dump_arg_begin(stream, "src");
+      trace_dump_arg(stream, ptr, buffer);
+      
+      trace_dump_arg_begin(stream, "data");
       trace_dump_bytes(stream, map, buffer->size);
       trace_dump_arg_end(stream);
 
@@ -283,19 +282,10 @@ trace_winsys_buffer_unmap(struct pipe_winsys *_winsys,
    
       trace_dump_call_end(stream);
 
-      winsys->buffer_unmap(winsys, buffer);
-
       hash_table_remove(tr_ws->buffer_maps, buffer);
    }
    
-   trace_dump_call_begin(stream, "pipe_winsys", "buffer_unmap");
-   
-   trace_dump_arg(stream, ptr, winsys);
-   trace_dump_arg(stream, ptr, buffer);
-
    winsys->buffer_unmap(winsys, buffer);
-   
-   trace_dump_call_end(stream);
 }
 
 
@@ -420,6 +410,7 @@ trace_winsys_destroy(struct pipe_winsys *_winsys)
 struct pipe_winsys *
 trace_winsys_create(struct pipe_winsys *winsys)
 {
+   struct trace_stream *stream;
    struct trace_winsys *tr_ws;
    
    if(!debug_get_bool_option("GALLIUM_TRACE", FALSE))
@@ -446,7 +437,7 @@ trace_winsys_create(struct pipe_winsys *winsys)
    
    tr_ws->winsys = winsys;
 
-   tr_ws->stream = trace_stream_create("gallium", "trace");
+   tr_ws->stream = stream = trace_stream_create("gallium", "trace");
    if(!tr_ws->stream)
       return NULL;
 
@@ -455,8 +446,11 @@ trace_winsys_create(struct pipe_winsys *winsys)
    if(!tr_ws->buffer_maps)
       return NULL;
    
-   
    trace_dump_trace_begin(tr_ws->stream, 0);
    
+   trace_dump_call_begin(stream, "", "pipe_winsys_create");
+   trace_dump_ret(stream, ptr, winsys);
+   trace_dump_call_end(stream);
+
    return &tr_ws->base;
 }