Merge commit 'mesa_7_5_rc4' into debian-experimental

Conflicts:
	progs/fp/point-position.c
	progs/fp/tri-depth.c
	progs/fp/tri-depth2.c
	progs/fp/tri-depthwrite.c
	progs/fp/tri-depthwrite2.c
	progs/fp/tri-inv.c
	progs/fp/tri-param.c
	progs/fp/tri-tex.c
	progs/tests/.gitignore
	progs/tests/Makefile
	progs/tests/zreaddraw.c

43 files changed, 654 insertions, 281 deletions

diff --git a/src/gallium/auxiliary/draw/draw_pt_elts.c b/src/gallium/auxiliary/draw/draw_pt_elts.c
index b7780fb5073..88f4d9f495a 100644
--- a/src/gallium/auxiliary/draw/draw_pt_elts.c
+++ b/src/gallium/auxiliary/draw/draw_pt_elts.c
@@ -54,7 +54,8 @@ static unsigned elt_ubyte( const void *elts, unsigned idx )
 
 static unsigned elt_vert( const void *elts, unsigned idx )
 {
-   return (const ubyte *)elts - (const ubyte *)NULL + idx;
+   /* unsigned index is packed in the pointer */
+   return (unsigned)(uintptr_t)elts + idx;
 }
 
 pt_elt_func draw_pt_elt_func( struct draw_context *draw )
diff --git a/src/gallium/auxiliary/indices/u_indices.c b/src/gallium/auxiliary/indices/u_indices.c
index 0cf7d88653c..03d7453f303 100644
--- a/src/gallium/auxiliary/indices/u_indices.c
+++ b/src/gallium/auxiliary/indices/u_indices.c
@@ -244,7 +244,7 @@ int u_index_generator( unsigned hw_mask,
 
       default:
          assert(0);
-         *out_generate = generate[out_idx][in_pv][out_pv][prim];
+         *out_generate = generate[out_idx][in_pv][out_pv][PIPE_PRIM_POINTS];
          *out_prim = PIPE_PRIM_POINTS;
          *out_nr = nr;
          return U_TRANSLATE_ERROR;
diff --git a/src/gallium/auxiliary/pipebuffer/pb_buffer.h b/src/gallium/auxiliary/pipebuffer/pb_buffer.h
index 92b6fd00564..2590546cb4a 100644
--- a/src/gallium/auxiliary/pipebuffer/pb_buffer.h
+++ b/src/gallium/auxiliary/pipebuffer/pb_buffer.h
@@ -72,6 +72,12 @@ struct pb_desc
 
 
 /**
+ * Size. Regular (32bit) unsigned for now.
+ */
+typedef unsigned pb_size;
+
+
+/**
  * Base class for all pb_* buffers.
  */
 struct pb_buffer 
@@ -126,7 +132,7 @@ struct pb_vtbl
     */
    void (*get_base_buffer)( struct pb_buffer *buf,
                             struct pb_buffer **base_buf,
-                            unsigned *offset );
+                            pb_size *offset );
    
 };
 
@@ -177,7 +183,7 @@ pb_unmap(struct pb_buffer *buf)
 static INLINE void
 pb_get_base_buffer( struct pb_buffer *buf,
 		    struct pb_buffer **base_buf,
-		    unsigned *offset )
+		    pb_size *offset )
 {
    assert(buf);
    if(!buf) {
@@ -241,7 +247,7 @@ pb_reference(struct pb_buffer **dst,
  * the requested or not.
  */
 static INLINE boolean
-pb_check_alignment(size_t requested, size_t provided)
+pb_check_alignment(pb_size requested, pb_size provided)
 {
    if(!requested)
       return TRUE;
@@ -269,7 +275,7 @@ pb_check_usage(unsigned requested, unsigned provided)
  * hardware.
  */
 struct pb_buffer *
-pb_malloc_buffer_create(size_t size, 
+pb_malloc_buffer_create(pb_size size, 
                         const struct pb_desc *desc);
 
 
diff --git a/src/gallium/auxiliary/pipebuffer/pb_buffer_fenced.c b/src/gallium/auxiliary/pipebuffer/pb_buffer_fenced.c
index 4698efa69cf..109ac7c9d63 100644
--- a/src/gallium/auxiliary/pipebuffer/pb_buffer_fenced.c
+++ b/src/gallium/auxiliary/pipebuffer/pb_buffer_fenced.c
@@ -65,11 +65,11 @@ struct fenced_buffer_list
    
    struct pb_fence_ops *ops;
    
-   size_t numDelayed;
+   pb_size numDelayed;
    struct list_head delayed;
    
 #ifdef DEBUG
-   size_t numUnfenced;
+   pb_size numUnfenced;
    struct list_head unfenced;
 #endif
 };
@@ -433,7 +433,7 @@ fenced_buffer_fence(struct pb_buffer *buf,
 static void
 fenced_buffer_get_base_buffer(struct pb_buffer *buf,
                               struct pb_buffer **base_buf,
-                              unsigned *offset)
+                              pb_size *offset)
 {
    struct fenced_buffer *fenced_buf = fenced_buffer(buf);
    pb_get_base_buffer(fenced_buf->buffer, base_buf, offset);
diff --git a/src/gallium/auxiliary/pipebuffer/pb_buffer_malloc.c b/src/gallium/auxiliary/pipebuffer/pb_buffer_malloc.c
index 689fd74771f..0760d607161 100644
--- a/src/gallium/auxiliary/pipebuffer/pb_buffer_malloc.c
+++ b/src/gallium/auxiliary/pipebuffer/pb_buffer_malloc.c
@@ -102,7 +102,7 @@ malloc_buffer_fence(struct pb_buffer *buf,
 static void
 malloc_buffer_get_base_buffer(struct pb_buffer *buf,
                               struct pb_buffer **base_buf,
-                              unsigned *offset)
+                              pb_size *offset)
 {
    *base_buf = buf;
    *offset = 0;
@@ -121,7 +121,7 @@ malloc_buffer_vtbl = {
 
 
 struct pb_buffer *
-pb_malloc_buffer_create(size_t size,
+pb_malloc_buffer_create(pb_size size,
                    	const struct pb_desc *desc) 
 {
    struct malloc_buffer *buf;
@@ -150,7 +150,7 @@ pb_malloc_buffer_create(size_t size,
 
 static struct pb_buffer *
 pb_malloc_bufmgr_create_buffer(struct pb_manager *mgr, 
-                               size_t size,
+                               pb_size size,
                                const struct pb_desc *desc) 
 {
    return pb_malloc_buffer_create(size, desc);
diff --git a/src/gallium/auxiliary/pipebuffer/pb_bufmgr.h b/src/gallium/auxiliary/pipebuffer/pb_bufmgr.h
index 74077f82774..39ab8e722c1 100644
--- a/src/gallium/auxiliary/pipebuffer/pb_bufmgr.h
+++ b/src/gallium/auxiliary/pipebuffer/pb_bufmgr.h
@@ -73,7 +73,7 @@ struct pb_manager
 
    struct pb_buffer *
    (*create_buffer)( struct pb_manager *mgr, 
-	             size_t size,
+	             pb_size size,
 	             const struct pb_desc *desc);
 
    /**
@@ -106,7 +106,7 @@ pb_malloc_bufmgr_create(void);
  */
 struct pb_manager *
 pool_bufmgr_create(struct pb_manager *provider, 
-                   size_t n, size_t size,
+                   pb_size n, pb_size size,
                    const struct pb_desc *desc);
 
 
@@ -119,7 +119,7 @@ pool_bufmgr_create(struct pb_manager *provider,
  */
 struct pb_manager *
 mm_bufmgr_create(struct pb_manager *provider, 
-                 size_t size, size_t align2);
+                 pb_size size, pb_size align2);
 
 /**
  * Same as mm_bufmgr_create.
@@ -128,7 +128,7 @@ mm_bufmgr_create(struct pb_manager *provider,
  */
 struct pb_manager *
 mm_bufmgr_create_from_buffer(struct pb_buffer *buffer, 
-                             size_t size, size_t align2);
+                             pb_size size, pb_size align2);
 
 
 /**
@@ -136,8 +136,8 @@ mm_bufmgr_create_from_buffer(struct pb_buffer *buffer,
  */
 struct pb_manager *
 pb_slab_manager_create(struct pb_manager *provider,
-                       size_t bufSize,
-                       size_t slabSize,
+                       pb_size bufSize,
+                       pb_size slabSize,
                        const struct pb_desc *desc);
 
 /**
@@ -146,9 +146,9 @@ pb_slab_manager_create(struct pb_manager *provider,
  */
 struct pb_manager *
 pb_slab_range_manager_create(struct pb_manager *provider,
-                             size_t minBufSize,
-                             size_t maxBufSize,
-                             size_t slabSize,
+                             pb_size minBufSize,
+                             pb_size maxBufSize,
+                             pb_size slabSize,
                              const struct pb_desc *desc);
 
 
@@ -204,7 +204,7 @@ pb_ondemand_manager_create(struct pb_manager *provider);
  */
 struct pb_manager *
 pb_debug_manager_create(struct pb_manager *provider,
-                        size_t underflow_size, size_t overflow_size); 
+                        pb_size underflow_size, pb_size overflow_size); 
 
 
 #ifdef __cplusplus
diff --git a/src/gallium/auxiliary/pipebuffer/pb_bufmgr_alt.c b/src/gallium/auxiliary/pipebuffer/pb_bufmgr_alt.c
index db67d46c561..f60c836f18b 100644
--- a/src/gallium/auxiliary/pipebuffer/pb_bufmgr_alt.c
+++ b/src/gallium/auxiliary/pipebuffer/pb_bufmgr_alt.c
@@ -60,7 +60,7 @@ pb_alt_manager(struct pb_manager *mgr)
 
 static struct pb_buffer *
 pb_alt_manager_create_buffer(struct pb_manager *_mgr, 
-                             size_t size,
+                             pb_size size,
                              const struct pb_desc *desc)
 {
    struct pb_alt_manager *mgr = pb_alt_manager(_mgr);
diff --git a/src/gallium/auxiliary/pipebuffer/pb_bufmgr_cache.c b/src/gallium/auxiliary/pipebuffer/pb_bufmgr_cache.c
index 35358430b43..57d1ede45a4 100644
--- a/src/gallium/auxiliary/pipebuffer/pb_bufmgr_cache.c
+++ b/src/gallium/auxiliary/pipebuffer/pb_bufmgr_cache.c
@@ -81,7 +81,7 @@ struct pb_cache_manager
    pipe_mutex mutex;
    
    struct list_head delayed;
-   size_t numDelayed;
+   pb_size numDelayed;
 };
 
 
@@ -204,7 +204,7 @@ pb_cache_buffer_fence(struct pb_buffer *_buf,
 static void
 pb_cache_buffer_get_base_buffer(struct pb_buffer *_buf,
                               struct pb_buffer **base_buf,
-                              unsigned *offset)
+                              pb_size *offset)
 {
    struct pb_cache_buffer *buf = pb_cache_buffer(_buf);
    pb_get_base_buffer(buf->buffer, base_buf, offset);
@@ -224,7 +224,7 @@ pb_cache_buffer_vtbl = {
 
 static INLINE boolean
 pb_cache_is_buffer_compat(struct pb_cache_buffer *buf,  
-                          size_t size,
+                          pb_size size,
                           const struct pb_desc *desc)
 {
    if(buf->base.base.size < size)
@@ -246,7 +246,7 @@ pb_cache_is_buffer_compat(struct pb_cache_buffer *buf,
 
 static struct pb_buffer *
 pb_cache_manager_create_buffer(struct pb_manager *_mgr, 
-                               size_t size,
+                               pb_size size,
                                const struct pb_desc *desc)
 {
    struct pb_cache_manager *mgr = pb_cache_manager(_mgr);
diff --git a/src/gallium/auxiliary/pipebuffer/pb_bufmgr_debug.c b/src/gallium/auxiliary/pipebuffer/pb_bufmgr_debug.c
index cedf745bdac..1b4df28c707 100644
--- a/src/gallium/auxiliary/pipebuffer/pb_bufmgr_debug.c
+++ b/src/gallium/auxiliary/pipebuffer/pb_bufmgr_debug.c
@@ -72,8 +72,8 @@ struct pb_debug_buffer
    struct pb_buffer *buffer;
    struct pb_debug_manager *mgr;
    
-   size_t underflow_size;
-   size_t overflow_size;
+   pb_size underflow_size;
+   pb_size overflow_size;
 
    struct debug_stack_frame create_backtrace[PB_DEBUG_CREATE_BACKTRACE];
 
@@ -91,8 +91,8 @@ struct pb_debug_manager
 
    struct pb_manager *provider;
 
-   size_t underflow_size;
-   size_t overflow_size;
+   pb_size underflow_size;
+   pb_size overflow_size;
    
    pipe_mutex mutex;
    struct list_head list;
@@ -124,9 +124,9 @@ static const uint8_t random_pattern[32] = {
 
 
 static INLINE void 
-fill_random_pattern(uint8_t *dst, size_t size)
+fill_random_pattern(uint8_t *dst, pb_size size)
 {
-   size_t i = 0;
+   pb_size i = 0;
    while(size--) {
       *dst++ = random_pattern[i++];
       i &= sizeof(random_pattern) - 1;
@@ -135,11 +135,11 @@ fill_random_pattern(uint8_t *dst, size_t size)
 
 
 static INLINE boolean 
-check_random_pattern(const uint8_t *dst, size_t size, 
-                     size_t *min_ofs, size_t *max_ofs) 
+check_random_pattern(const uint8_t *dst, pb_size size, 
+                     pb_size *min_ofs, pb_size *max_ofs) 
 {
    boolean result = TRUE;
-   size_t i;
+   pb_size i;
    *min_ofs = size;
    *max_ofs = 0;
    for(i = 0; i < size; ++i) {
@@ -183,7 +183,7 @@ pb_debug_buffer_check(struct pb_debug_buffer *buf)
    assert(map);
    if(map) {
       boolean underflow, overflow;
-      size_t min_ofs, max_ofs;
+      pb_size min_ofs, max_ofs;
       
       underflow = !check_random_pattern(map, buf->underflow_size, 
                                         &min_ofs, &max_ofs);
@@ -287,7 +287,7 @@ pb_debug_buffer_unmap(struct pb_buffer *_buf)
 static void
 pb_debug_buffer_get_base_buffer(struct pb_buffer *_buf,
                                 struct pb_buffer **base_buf,
-                                unsigned *offset)
+                                pb_size *offset)
 {
    struct pb_debug_buffer *buf = pb_debug_buffer(_buf);
    pb_get_base_buffer(buf->buffer, base_buf, offset);
@@ -363,13 +363,13 @@ pb_debug_manager_dump(struct pb_debug_manager *mgr)
 
 static struct pb_buffer *
 pb_debug_manager_create_buffer(struct pb_manager *_mgr, 
-                               size_t size,
+                               pb_size size,
                                const struct pb_desc *desc)
 {
    struct pb_debug_manager *mgr = pb_debug_manager(_mgr);
    struct pb_debug_buffer *buf;
    struct pb_desc real_desc;
-   size_t real_size;
+   pb_size real_size;
    
    buf = CALLOC_STRUCT(pb_debug_buffer);
    if(!buf)
@@ -455,7 +455,7 @@ pb_debug_manager_destroy(struct pb_manager *_mgr)
 
 struct pb_manager *
 pb_debug_manager_create(struct pb_manager *provider, 
-                        size_t underflow_size, size_t overflow_size) 
+                        pb_size underflow_size, pb_size overflow_size) 
 {
    struct pb_debug_manager *mgr;
 
@@ -485,7 +485,7 @@ pb_debug_manager_create(struct pb_manager *provider,
 
 struct pb_manager *
 pb_debug_manager_create(struct pb_manager *provider, 
-                        size_t underflow_size, size_t overflow_size) 
+                        pb_size underflow_size, pb_size overflow_size) 
 {
    return provider;
 }
diff --git a/src/gallium/auxiliary/pipebuffer/pb_bufmgr_fenced.c b/src/gallium/auxiliary/pipebuffer/pb_bufmgr_fenced.c
index 144db5669b6..97dd1427fda 100644
--- a/src/gallium/auxiliary/pipebuffer/pb_bufmgr_fenced.c
+++ b/src/gallium/auxiliary/pipebuffer/pb_bufmgr_fenced.c
@@ -62,7 +62,7 @@ fenced_pb_manager(struct pb_manager *mgr)
 
 static struct pb_buffer *
 fenced_bufmgr_create_buffer(struct pb_manager *mgr, 
-                            size_t size,
+                            pb_size size,
                             const struct pb_desc *desc)
 {
    struct fenced_pb_manager *fenced_mgr = fenced_pb_manager(mgr);
diff --git a/src/gallium/auxiliary/pipebuffer/pb_bufmgr_mm.c b/src/gallium/auxiliary/pipebuffer/pb_bufmgr_mm.c
index 5a342fbf3b1..6400fc5b0a3 100644
--- a/src/gallium/auxiliary/pipebuffer/pb_bufmgr_mm.c
+++ b/src/gallium/auxiliary/pipebuffer/pb_bufmgr_mm.c
@@ -55,10 +55,10 @@ struct mm_pb_manager
    
    pipe_mutex mutex;
    
-   size_t size;
+   pb_size size;
    struct mem_block *heap;
    
-   size_t align2;
+   pb_size align2;
    
    struct pb_buffer *buffer;
    void *map;
@@ -148,7 +148,7 @@ mm_buffer_fence(struct pb_buffer *buf,
 static void
 mm_buffer_get_base_buffer(struct pb_buffer *buf,
                           struct pb_buffer **base_buf,
-                          unsigned *offset)
+                          pb_size *offset)
 {
    struct mm_buffer *mm_buf = mm_buffer(buf);
    struct mm_pb_manager *mm = mm_buf->mgr;
@@ -170,15 +170,15 @@ mm_buffer_vtbl = {
 
 static struct pb_buffer *
 mm_bufmgr_create_buffer(struct pb_manager *mgr, 
-                        size_t size,
+                        pb_size size,
                         const struct pb_desc *desc)
 {
    struct mm_pb_manager *mm = mm_pb_manager(mgr);
    struct mm_buffer *mm_buf;
 
    /* We don't handle alignments larger then the one initially setup */
-   assert(pb_check_alignment(desc->alignment, 1 << mm->align2));
-   if(!pb_check_alignment(desc->alignment, 1 << mm->align2))
+   assert(pb_check_alignment(desc->alignment, (pb_size)1 << mm->align2));
+   if(!pb_check_alignment(desc->alignment, (pb_size)1 << mm->align2))
       return NULL;
    
    pipe_mutex_lock(mm->mutex);
@@ -198,10 +198,10 @@ mm_bufmgr_create_buffer(struct pb_manager *mgr,
    
    mm_buf->mgr = mm;
    
-   mm_buf->block = u_mmAllocMem(mm->heap, size, mm->align2, 0);
+   mm_buf->block = u_mmAllocMem(mm->heap, (int)size, (int)mm->align2, 0);
    if(!mm_buf->block) {
-      debug_printf("warning: heap full\n");
 #if 0
+      debug_printf("warning: heap full\n");
       mmDumpMemInfo(mm->heap);
 #endif
       FREE(mm_buf);
@@ -210,8 +210,8 @@ mm_bufmgr_create_buffer(struct pb_manager *mgr,
    }
    
    /* Some sanity checks */
-   assert(0 <= (unsigned)mm_buf->block->ofs && (unsigned)mm_buf->block->ofs < mm->size);
-   assert(size <= (unsigned)mm_buf->block->size && (unsigned)mm_buf->block->ofs + (unsigned)mm_buf->block->size <= mm->size);
+   assert(0 <= (pb_size)mm_buf->block->ofs && (pb_size)mm_buf->block->ofs < mm->size);
+   assert(size <= (pb_size)mm_buf->block->size && (pb_size)mm_buf->block->ofs + (pb_size)mm_buf->block->size <= mm->size);
    
    pipe_mutex_unlock(mm->mutex);
    return SUPER(mm_buf);
@@ -245,7 +245,7 @@ mm_bufmgr_destroy(struct pb_manager *mgr)
 
 struct pb_manager *
 mm_bufmgr_create_from_buffer(struct pb_buffer *buffer, 
-                             size_t size, size_t align2) 
+                             pb_size size, pb_size align2) 
 {
    struct mm_pb_manager *mm;
 
@@ -273,7 +273,7 @@ mm_bufmgr_create_from_buffer(struct pb_buffer *buffer,
    if(!mm->map)
       goto failure;
 
-   mm->heap = u_mmInit(0, size); 
+   mm->heap = u_mmInit(0, (int)size); 
    if (!mm->heap)
       goto failure;
 
@@ -292,7 +292,7 @@ if(mm->heap)
 
 struct pb_manager *
 mm_bufmgr_create(struct pb_manager *provider, 
-                 size_t size, size_t align2) 
+                 pb_size size, pb_size align2) 
 {
    struct pb_buffer *buffer;
    struct pb_manager *mgr;
diff --git a/src/gallium/auxiliary/pipebuffer/pb_bufmgr_ondemand.c b/src/gallium/auxiliary/pipebuffer/pb_bufmgr_ondemand.c
index 4f7e6b1c4df..bc3093f620f 100644
--- a/src/gallium/auxiliary/pipebuffer/pb_bufmgr_ondemand.c
+++ b/src/gallium/auxiliary/pipebuffer/pb_bufmgr_ondemand.c
@@ -55,7 +55,7 @@ struct pb_ondemand_buffer
    
    /** Real buffer */
    struct pb_buffer *buffer;
-   size_t size;
+   pb_size size;
    struct pb_desc desc;
 };
 
@@ -204,7 +204,7 @@ pb_ondemand_buffer_fence(struct pb_buffer *_buf,
 static void
 pb_ondemand_buffer_get_base_buffer(struct pb_buffer *_buf,
                                    struct pb_buffer **base_buf,
-                                   unsigned *offset)
+                                   pb_size *offset)
 {
    struct pb_ondemand_buffer *buf = pb_ondemand_buffer(_buf);
 
@@ -232,7 +232,7 @@ pb_ondemand_buffer_vtbl = {
 
 static struct pb_buffer *
 pb_ondemand_manager_create_buffer(struct pb_manager *_mgr, 
-                                  size_t size,
+                                  pb_size size,
                                   const struct pb_desc *desc) 
 {
    struct pb_ondemand_manager *mgr = pb_ondemand_manager(_mgr);
diff --git a/src/gallium/auxiliary/pipebuffer/pb_bufmgr_pool.c b/src/gallium/auxiliary/pipebuffer/pb_bufmgr_pool.c
index 07fd1a22d93..7fd65ed2261 100644
--- a/src/gallium/auxiliary/pipebuffer/pb_bufmgr_pool.c
+++ b/src/gallium/auxiliary/pipebuffer/pb_bufmgr_pool.c
@@ -58,11 +58,11 @@ struct pool_pb_manager
    
    pipe_mutex mutex;
    
-   size_t bufSize;
-   size_t bufAlign;
+   pb_size bufSize;
+   pb_size bufAlign;
    
-   size_t numFree;
-   size_t numTot;
+   pb_size numFree;
+   pb_size numTot;
    
    struct list_head free;
    
@@ -89,7 +89,7 @@ struct pool_buffer
    
    struct list_head head;
    
-   size_t start;
+   pb_size start;
 };
 
 
@@ -162,7 +162,7 @@ pool_buffer_fence(struct pb_buffer *buf,
 static void
 pool_buffer_get_base_buffer(struct pb_buffer *buf,
                             struct pb_buffer **base_buf,
-                            unsigned *offset)
+                            pb_size *offset)
 {
    struct pool_buffer *pool_buf = pool_buffer(buf);
    struct pool_pb_manager *pool = pool_buf->mgr;
@@ -184,7 +184,7 @@ pool_buffer_vtbl = {
 
 static struct pb_buffer *
 pool_bufmgr_create_buffer(struct pb_manager *mgr,
-                          size_t size,
+                          pb_size size,
                           const struct pb_desc *desc)
 {
    struct pool_pb_manager *pool = pool_pb_manager(mgr);
@@ -251,13 +251,13 @@ pool_bufmgr_destroy(struct pb_manager *mgr)
 
 struct pb_manager *
 pool_bufmgr_create(struct pb_manager *provider, 
-                   size_t numBufs, 
-                   size_t bufSize,
+                   pb_size numBufs, 
+                   pb_size bufSize,
                    const struct pb_desc *desc) 
 {
    struct pool_pb_manager *pool;
    struct pool_buffer *pool_buf;
-   size_t i;
+   pb_size i;
 
    if(!provider)
       return NULL;
diff --git a/src/gallium/auxiliary/pipebuffer/pb_bufmgr_slab.c b/src/gallium/auxiliary/pipebuffer/pb_bufmgr_slab.c
index 724aaadb436..e7352e90db9 100644
--- a/src/gallium/auxiliary/pipebuffer/pb_bufmgr_slab.c
+++ b/src/gallium/auxiliary/pipebuffer/pb_bufmgr_slab.c
@@ -68,7 +68,7 @@ struct pb_slab_buffer
    unsigned mapCount;
    
    /** Offset relative to the start of the slab buffer. */
-   size_t start;
+   pb_size start;
    
    /** Use when validating, to signal that all mappings are finished */
    /* TODO: Actually validation does not reach this stage yet */
@@ -83,8 +83,8 @@ struct pb_slab
 {
    struct list_head head;
    struct list_head freeBuffers;
-   size_t numBuffers;
-   size_t numFree;
+   pb_size numBuffers;
+   pb_size numFree;
    
    struct pb_slab_buffer *buffers;
    struct pb_slab_manager *mgr;
@@ -108,10 +108,10 @@ struct pb_slab_manager
    struct pb_manager *provider;
    
    /** Size of the buffers we hand on downstream */
-   size_t bufSize;
+   pb_size bufSize;
    
    /** Size of the buffers we request upstream */
-   size_t slabSize;
+   pb_size slabSize;
    
    /** 
     * Alignment, usage to be used to allocate the slab buffers.
@@ -150,14 +150,14 @@ struct pb_slab_range_manager
 
    struct pb_manager *provider;
    
-   size_t minBufSize;
-   size_t maxBufSize;
+   pb_size minBufSize;
+   pb_size maxBufSize;
    
    /** @sa pb_slab_manager::desc */ 
    struct pb_desc desc;
    
    unsigned numBuckets;
-   size_t *bucketSizes;
+   pb_size *bucketSizes;
    
    /** Array of pb_slab_manager, one for each bucket size */
    struct pb_manager **buckets;
@@ -270,7 +270,7 @@ pb_slab_buffer_fence(struct pb_buffer *_buf,
 static void
 pb_slab_buffer_get_base_buffer(struct pb_buffer *_buf,
                                struct pb_buffer **base_buf,
-                               unsigned *offset)
+                               pb_size *offset)
 {
    struct pb_slab_buffer *buf = pb_slab_buffer(_buf);
    pb_get_base_buffer(buf->slab->bo, base_buf, offset);
@@ -369,7 +369,7 @@ out_err0:
 
 static struct pb_buffer *
 pb_slab_manager_create_buffer(struct pb_manager *_mgr,
-                              size_t size,
+                              pb_size size,
                               const struct pb_desc *desc)
 {
    struct pb_slab_manager *mgr = pb_slab_manager(_mgr);
@@ -450,8 +450,8 @@ pb_slab_manager_destroy(struct pb_manager *_mgr)
 
 struct pb_manager *
 pb_slab_manager_create(struct pb_manager *provider,
-                       size_t bufSize,
-                       size_t slabSize,
+                       pb_size bufSize,
+                       pb_size slabSize,
                        const struct pb_desc *desc)
 {
    struct pb_slab_manager *mgr;
@@ -479,11 +479,11 @@ pb_slab_manager_create(struct pb_manager *provider,
 
 static struct pb_buffer *
 pb_slab_range_manager_create_buffer(struct pb_manager *_mgr,
-                                    size_t size,
+                                    pb_size size,
                                     const struct pb_desc *desc)
 {
    struct pb_slab_range_manager *mgr = pb_slab_range_manager(_mgr);
-   size_t bufSize;
+   pb_size bufSize;
    unsigned i;
 
    bufSize = mgr->minBufSize;
@@ -527,13 +527,13 @@ pb_slab_range_manager_destroy(struct pb_manager *_mgr)
 
 struct pb_manager *
 pb_slab_range_manager_create(struct pb_manager *provider,
-                             size_t minBufSize,
-                             size_t maxBufSize,
-                             size_t slabSize,
+                             pb_size minBufSize,
+                             pb_size maxBufSize,
+                             pb_size slabSize,
                              const struct pb_desc *desc)
 {
    struct pb_slab_range_manager *mgr;
-   size_t bufSize;
+   pb_size bufSize;
    unsigned i;
 
    if(!provider)
diff --git a/src/gallium/auxiliary/rtasm/rtasm_ppc.c b/src/gallium/auxiliary/rtasm/rtasm_ppc.c
index e3586482db4..ef4b306cb67 100644
--- a/src/gallium/auxiliary/rtasm/rtasm_ppc.c
+++ b/src/gallium/auxiliary/rtasm/rtasm_ppc.c
@@ -168,7 +168,7 @@ ppc_allocate_register(struct ppc_function *p)
 {
    unsigned i;
    for (i = 0; i < PPC_NUM_REGS; i++) {
-      const uint64_t mask = 1 << i;
+      const uint32_t mask = 1 << i;
       if ((p->reg_used & mask) == 0) {
          p->reg_used |= mask;
          return i;
@@ -200,7 +200,7 @@ ppc_allocate_fp_register(struct ppc_function *p)
 {
    unsigned i;
    for (i = 0; i < PPC_NUM_FP_REGS; i++) {
-      const uint64_t mask = 1 << i;
+      const uint32_t mask = 1 << i;
       if ((p->fp_used & mask) == 0) {
          p->fp_used |= mask;
          return i;
@@ -232,7 +232,7 @@ ppc_allocate_vec_register(struct ppc_function *p)
 {
    unsigned i;
    for (i = 0; i < PPC_NUM_VEC_REGS; i++) {
-      const uint64_t mask = 1 << i;
+      const uint32_t mask = 1 << i;
       if ((p->vec_used & mask) == 0) {
          p->vec_used |= mask;
          return i;
diff --git a/src/gallium/auxiliary/tgsi/tgsi_dump.c b/src/gallium/auxiliary/tgsi/tgsi_dump.c
index a784b7cc3c1..76a09af18ee 100644
--- a/src/gallium/auxiliary/tgsi/tgsi_dump.c
+++ b/src/gallium/auxiliary/tgsi/tgsi_dump.c
@@ -27,6 +27,7 @@
 
 #include "util/u_debug.h"
 #include "util/u_string.h"
+#include "util/u_math.h"
 #include "tgsi_dump.h"
 #include "tgsi_info.h"
 #include "tgsi_iterate.h"
@@ -516,7 +517,7 @@ struct str_dump_ctx
    struct dump_ctx base;
    char *str;
    char *ptr;
-   size_t left;
+   int left;
 };
 
 static void
@@ -525,13 +526,20 @@ str_dump_ctx_printf(struct dump_ctx *ctx, const char *format, ...)
    struct str_dump_ctx *sctx = (struct str_dump_ctx *)ctx;
    
    if(sctx->left > 1) {
-      size_t written;
+      int written;
       va_list ap;
       va_start(ap, format);
       written = util_vsnprintf(sctx->ptr, sctx->left, format, ap);
       va_end(ap);
-      sctx->ptr += written;
-      sctx->left -= written;
+
+      /* Some complicated logic needed to handle the return value of
+       * vsnprintf:
+       */
+      if (written > 0) {
+         written = MIN2(sctx->left, written);
+         sctx->ptr += written;
+         sctx->left -= written;
+      }
    }
 }
 
@@ -556,7 +564,7 @@ tgsi_dump_str(
    ctx.str = str;
    ctx.str[0] = 0;
    ctx.ptr = str;
-   ctx.left = size;
+   ctx.left = (int)size;
 
    tgsi_iterate_shader( tokens, &ctx.base.iter );
 }
diff --git a/src/gallium/auxiliary/util/u_gen_mipmap.c b/src/gallium/auxiliary/util/u_gen_mipmap.c
index 6fa13a8ce11..ca797486a0e 100644
--- a/src/gallium/auxiliary/util/u_gen_mipmap.c
+++ b/src/gallium/auxiliary/util/u_gen_mipmap.c
@@ -1385,7 +1385,7 @@ set_vertex_data(struct gen_mipmap_state *ctx,
           * Not +/-1 to avoid cube face selection ambiguity near the edges,
           * though that can still sometimes happen with this scale factor...
           */
-         const float scale = 0.9999;
+         const float scale = 0.9999f;
          const float sc = (2.0f * st[i][0] - 1.0f) * scale;
          const float tc = (2.0f * st[i][1] - 1.0f) * scale;
 
diff --git a/src/gallium/auxiliary/util/u_snprintf.c b/src/gallium/auxiliary/util/u_snprintf.c
index 0d54299b28a..65bdd0df2b7 100644
--- a/src/gallium/auxiliary/util/u_snprintf.c
+++ b/src/gallium/auxiliary/util/u_snprintf.c
@@ -829,7 +829,7 @@ util_vsnprintf(char *str, size_t size, const char *format, va_list args)
 					break;
 				default:
 					intptr = va_arg(args, int *);
-					*intptr = len;
+					*intptr = (int)len;
 					break;
 				}
 				break;
diff --git a/src/gallium/drivers/cell/ppu/cell_screen.c b/src/gallium/drivers/cell/ppu/cell_screen.c
index 512d85d3525..bd48ce70050 100644
--- a/src/gallium/drivers/cell/ppu/cell_screen.c
+++ b/src/gallium/drivers/cell/ppu/cell_screen.c
@@ -86,6 +86,8 @@ cell_get_param(struct pipe_screen *screen, int param)
       return 1; /* XXX not really true */
    case PIPE_CAP_TEXTURE_MIRROR_CLAMP:
       return 0; /* XXX to do */
+   case PIPE_CAP_TGSI_CONT_SUPPORTED:
+      return 1;
    default:
       return 0;
    }
diff --git a/src/gallium/drivers/softpipe/sp_context.c b/src/gallium/drivers/softpipe/sp_context.c
index 62e8d99cfd0..86df320ea8a 100644
--- a/src/gallium/drivers/softpipe/sp_context.c
+++ b/src/gallium/drivers/softpipe/sp_context.c
@@ -126,6 +126,22 @@ softpipe_is_texture_referenced( struct pipe_context *pipe,
 				struct pipe_texture *texture,
 				unsigned face, unsigned level)
 {
+   struct softpipe_context *softpipe = softpipe_context( pipe );
+   unsigned i;
+
+   if(softpipe->dirty_render_cache) {
+      for (i = 0; i < softpipe->framebuffer.nr_cbufs; i++) {
+         if(softpipe->framebuffer.cbufs[i] && 
+            softpipe->framebuffer.cbufs[i]->texture == texture)
+            return PIPE_REFERENCED_FOR_WRITE;
+      }
+      if(softpipe->framebuffer.zsbuf && 
+         softpipe->framebuffer.zsbuf->texture == texture)
+         return PIPE_REFERENCED_FOR_WRITE;
+   }
+   
+   /* FIXME: we also need to do the same for the texture cache */
+   
    return PIPE_UNREFERENCED;
 }
 
diff --git a/src/gallium/drivers/softpipe/sp_context.h b/src/gallium/drivers/softpipe/sp_context.h
index 59d6df8f2dd..dffc15a4f19 100644
--- a/src/gallium/drivers/softpipe/sp_context.h
+++ b/src/gallium/drivers/softpipe/sp_context.h
@@ -144,6 +144,8 @@ struct softpipe_context {
    struct draw_stage *vbuf;
    struct softpipe_vbuf_render *vbuf_render;
 
+   boolean dirty_render_cache;
+   
    struct softpipe_tile_cache *cbuf_cache[PIPE_MAX_COLOR_BUFS];
    struct softpipe_tile_cache *zsbuf_cache;
 
diff --git a/src/gallium/drivers/softpipe/sp_draw_arrays.c b/src/gallium/drivers/softpipe/sp_draw_arrays.c
index f117096bf73..ba2766ff139 100644
--- a/src/gallium/drivers/softpipe/sp_draw_arrays.c
+++ b/src/gallium/drivers/softpipe/sp_draw_arrays.c
@@ -182,6 +182,8 @@ softpipe_draw_range_elements(struct pipe_context *pipe,
    /* Note: leave drawing surfaces mapped */
    softpipe_unmap_constant_buffers(sp);
 
+   sp->dirty_render_cache = TRUE;
+   
    return TRUE;
 }
 
diff --git a/src/gallium/drivers/softpipe/sp_flush.c b/src/gallium/drivers/softpipe/sp_flush.c
index 035f4b963eb..4a14d49686e 100644
--- a/src/gallium/drivers/softpipe/sp_flush.c
+++ b/src/gallium/drivers/softpipe/sp_flush.c
@@ -71,6 +71,8 @@ softpipe_flush( struct pipe_context *pipe,
        * to unmap surfaces when flushing.
        */
       softpipe_unmap_transfers(softpipe);
+      
+      softpipe->dirty_render_cache = FALSE;
    }
 
    /* Enable to dump BMPs of the color/depth buffers each frame */
diff --git a/src/gallium/drivers/softpipe/sp_screen.c b/src/gallium/drivers/softpipe/sp_screen.c
index a32fd3a1ba0..cc39d33ede6 100644
--- a/src/gallium/drivers/softpipe/sp_screen.c
+++ b/src/gallium/drivers/softpipe/sp_screen.c
@@ -87,6 +87,8 @@ softpipe_get_param(struct pipe_screen *screen, int param)
       return 8;  /* max 128x128x128 */
    case PIPE_CAP_MAX_TEXTURE_CUBE_LEVELS:
       return 12; /* max 2Kx2K */
+   case PIPE_CAP_TGSI_CONT_SUPPORTED:
+      return 1;
    default:
       return 0;
    }
diff --git a/src/gallium/drivers/softpipe/sp_setup.c b/src/gallium/drivers/softpipe/sp_setup.c
index accc692b66f..e5be65242d6 100644
--- a/src/gallium/drivers/softpipe/sp_setup.c
+++ b/src/gallium/drivers/softpipe/sp_setup.c
@@ -444,7 +444,8 @@ static void flush_spans( struct setup_context *setup )
             mask |= MASK_TOP_RIGHT;
          if (x+1 >= xleft1 && x+1 < xright1)
             mask |= MASK_BOTTOM_RIGHT;
-         EMIT_QUAD( setup, x, setup->span.y, mask );
+         if (mask)
+            EMIT_QUAD( setup, x, setup->span.y, mask );
       }
       break;
 
@@ -458,7 +459,8 @@ static void flush_spans( struct setup_context *setup )
             mask |= MASK_TOP_LEFT;
          if (x+1 >= xleft0 && x+1 < xright0)
             mask |= MASK_TOP_RIGHT;
-         EMIT_QUAD( setup, x, setup->span.y, mask );
+         if (mask)
+            EMIT_QUAD( setup, x, setup->span.y, mask );
       }
       break;
 
@@ -472,7 +474,8 @@ static void flush_spans( struct setup_context *setup )
             mask |= MASK_BOTTOM_LEFT;
          if (x+1 >= xleft1 && x+1 < xright1)
             mask |= MASK_BOTTOM_RIGHT;
-         EMIT_QUAD( setup, x, setup->span.y, mask );
+         if (mask)
+            EMIT_QUAD( setup, x, setup->span.y, mask );
       }
       break;
 
diff --git a/src/gallium/drivers/softpipe/sp_tex_sample.c b/src/gallium/drivers/softpipe/sp_tex_sample.c
index adbd0cb7f0c..5de358dae93 100644
--- a/src/gallium/drivers/softpipe/sp_tex_sample.c
+++ b/src/gallium/drivers/softpipe/sp_tex_sample.c
@@ -744,7 +744,9 @@ shadow_compare(uint compare_func,
       break;
    }
 
+   /* XXX returning result for default GL_DEPTH_TEXTURE_MODE = GL_LUMINANCE */
    rgba[0][j] = rgba[1][j] = rgba[2][j] = (float) k;
+   rgba[3][j] = 1.0F;
 }
 
 
diff --git a/src/gallium/drivers/trace/tr_dump.c b/src/gallium/drivers/trace/tr_dump.c
index 3a1409e95a7..643587ab427 100644
--- a/src/gallium/drivers/trace/tr_dump.c
+++ b/src/gallium/drivers/trace/tr_dump.c
@@ -451,11 +451,11 @@ void trace_dump_float(double value)
 }
 
 void trace_dump_bytes(const void *data,
-                      long unsigned size)
+                      size_t size)
 {
    static const char hex_table[16] = "0123456789ABCDEF";
    const uint8_t *p = data;
-   long unsigned i;
+   size_t i;
 
    if (!dumping)
       return;
diff --git a/src/gallium/drivers/trace/tr_dump.h b/src/gallium/drivers/trace/tr_dump.h
index 31ac70802f0..32592bab12f 100644
--- a/src/gallium/drivers/trace/tr_dump.h
+++ b/src/gallium/drivers/trace/tr_dump.h
@@ -91,7 +91,7 @@ void trace_dump_bool(int value);
 void trace_dump_int(long long int value);
 void trace_dump_uint(long long unsigned value);
 void trace_dump_float(double value);
-void trace_dump_bytes(const void *data, long unsigned size);
+void trace_dump_bytes(const void *data, size_t size);
 void trace_dump_string(const char *str);
 void trace_dump_enum(const char *value);
 void trace_dump_array_begin(void);
@@ -130,7 +130,7 @@ void trace_dump_transfer_ptr(struct pipe_transfer *_transfer);
 
 #define trace_dump_array(_type, _obj, _size) \
    do { \
-      unsigned long idx; \
+      size_t idx; \
       trace_dump_array_begin(); \
       for(idx = 0; idx < (_size); ++idx) { \
          trace_dump_elem_begin(); \
@@ -142,7 +142,7 @@ void trace_dump_transfer_ptr(struct pipe_transfer *_transfer);
 
 #define trace_dump_struct_array(_type, _obj, _size) \
    do { \
-      unsigned long idx; \
+      size_t idx; \
       trace_dump_array_begin(); \
       for(idx = 0; idx < (_size); ++idx) { \
          trace_dump_elem_begin(); \
diff --git a/src/gallium/include/pipe/p_defines.h b/src/gallium/include/pipe/p_defines.h
index 82e23c413c8..ab57ed73c4c 100644
--- a/src/gallium/include/pipe/p_defines.h
+++ b/src/gallium/include/pipe/p_defines.h
@@ -310,6 +310,7 @@ enum pipe_transfer_usage {
 #define PIPE_CAP_TEXTURE_MIRROR_CLAMP    24
 #define PIPE_CAP_TEXTURE_MIRROR_REPEAT   25
 #define PIPE_CAP_MAX_VERTEX_TEXTURE_UNITS 26
+#define PIPE_CAP_TGSI_CONT_SUPPORTED     27
 
 
 /**
diff --git a/src/gallium/include/pipe/p_format.h b/src/gallium/include/pipe/p_format.h
index a279eefef9e..c4469d4a9e9 100644
--- a/src/gallium/include/pipe/p_format.h
+++ b/src/gallium/include/pipe/p_format.h
@@ -536,14 +536,40 @@ pf_get_nblocks(const struct pipe_format_block *block, unsigned width, unsigned h
    return pf_get_nblocksx(block, width)*pf_get_nblocksy(block, height);
 }
 
+static INLINE size_t
+pf_get_stride(const struct pipe_format_block *block, unsigned width)
+{
+   return pf_get_nblocksx(block, width)*block->size;
+}
+
+static INLINE size_t
+pf_get_2d_size(const struct pipe_format_block *block, size_t stride, unsigned height)
+{
+   return pf_get_nblocksy(block, height)*stride;
+}
+
 static INLINE boolean 
-pf_is_depth_stencil( enum pipe_format format )
+pf_is_depth_or_stencil( enum pipe_format format )
 {
    return (pf_get_component_bits( format, PIPE_FORMAT_COMP_Z ) +
            pf_get_component_bits( format, PIPE_FORMAT_COMP_S )) != 0;
 }
 
 static INLINE boolean 
+pf_is_depth_and_stencil( enum pipe_format format )
+{
+   return (pf_get_component_bits( format, PIPE_FORMAT_COMP_Z ) != 0 &&
+           pf_get_component_bits( format, PIPE_FORMAT_COMP_S ) != 0);
+}
+
+/** DEPRECATED: For backwards compatibility */
+static INLINE boolean
+pf_is_depth_stencil( enum pipe_format format )
+{
+   return pf_is_depth_or_stencil( format );
+}
+
+static INLINE boolean 
 pf_is_compressed( enum pipe_format format )
 {
    return pf_layout(format) == PIPE_FORMAT_LAYOUT_DXT ? TRUE : FALSE;
diff --git a/src/gallium/include/pipe/p_refcnt.h b/src/gallium/include/pipe/p_refcnt.h
index 1f89453e09a..1f9088b3e9c 100644
--- a/src/gallium/include/pipe/p_refcnt.h
+++ b/src/gallium/include/pipe/p_refcnt.h
@@ -62,29 +62,29 @@ pipe_is_referenced(struct pipe_reference *reference)
  * Set 'ptr' to point to 'reference' and update reference counting.
  * The old thing pointed to, if any, will be unreferenced first.
  * 'reference' may be NULL.
- *
- * XXX: thread safety issues!
  */
 static INLINE bool
 pipe_reference(struct pipe_reference **ptr, struct pipe_reference *reference)
 {
    bool destroy = FALSE;
 
-   /* bump the reference.count first */
-   if (reference) {
-      assert(pipe_is_referenced(reference));
-      p_atomic_inc(&reference->count);
-   }
-
-   if (*ptr) {
-      assert(pipe_is_referenced(*ptr));
-      if (p_atomic_dec_zero(&(*ptr)->count)) {
-         destroy = TRUE;
+   if(*ptr != reference) {
+      /* bump the reference.count first */
+      if (reference) {
+         assert(pipe_is_referenced(reference));
+         p_atomic_inc(&reference->count);
       }
+   
+      if (*ptr) {
+         assert(pipe_is_referenced(*ptr));
+         if (p_atomic_dec_zero(&(*ptr)->count)) {
+            destroy = TRUE;
+         }
+      }
+   
+      *ptr = reference;
    }
 
-   *ptr = reference;
-
    return destroy;
 }
 
diff --git a/src/gallium/include/pipe/p_thread.h b/src/gallium/include/pipe/p_thread.h
index de55e99ed49..ce8d79549ad 100644
--- a/src/gallium/include/pipe/p_thread.h
+++ b/src/gallium/include/pipe/p_thread.h
@@ -36,6 +36,7 @@
 
 
 #include "pipe/p_compiler.h"
+#include "util/u_debug.h" /* for assert */
 
 
 #if defined(PIPE_OS_LINUX) || defined(PIPE_OS_BSD) || defined(PIPE_OS_SOLARIS)
diff --git a/src/gallium/state_trackers/dri/dri_drawable.c b/src/gallium/state_trackers/dri/dri_drawable.c
index 15a2088df51..5846390d02c 100644
--- a/src/gallium/state_trackers/dri/dri_drawable.c
+++ b/src/gallium/state_trackers/dri/dri_drawable.c
@@ -231,7 +231,6 @@ dri_create_buffer(__DRIscreenPrivate * sPriv,
    if (visual->depthBits) {
       if (pscreen->is_format_supported(pscreen, PIPE_FORMAT_Z24S8_UNORM,
 				       PIPE_TEXTURE_2D,
-				       PIPE_TEXTURE_USAGE_RENDER_TARGET |
 				       PIPE_TEXTURE_USAGE_DEPTH_STENCIL, 0))
 	 depthFormat = PIPE_FORMAT_Z24S8_UNORM;
       else
@@ -242,7 +241,6 @@ dri_create_buffer(__DRIscreenPrivate * sPriv,
    if (visual->stencilBits) {
       if (pscreen->is_format_supported(pscreen, PIPE_FORMAT_Z24S8_UNORM,
 				       PIPE_TEXTURE_2D,
-				       PIPE_TEXTURE_USAGE_RENDER_TARGET |
 				       PIPE_TEXTURE_USAGE_DEPTH_STENCIL, 0))
 	 stencilFormat = PIPE_FORMAT_Z24S8_UNORM;
       else
diff --git a/src/gallium/state_trackers/glx/xlib/fakeglx.c b/src/gallium/state_trackers/glx/xlib/fakeglx.c
index 65e7048188e..6df4c7d6936 100644
--- a/src/gallium/state_trackers/glx/xlib/fakeglx.c
+++ b/src/gallium/state_trackers/glx/xlib/fakeglx.c
@@ -1817,7 +1817,7 @@ Fake_glXCreatePixmap( Display *dpy, GLXFBConfig config, Pixmap pixmap,
    if (!dpy || !config || !pixmap)
       return 0;
 
-   for (attr = attribList; *attr; attr++) {
+   for (attr = attribList; attr && *attr; attr++) {
       switch (*attr) {
       case GLX_TEXTURE_FORMAT_EXT:
          attr++;
diff --git a/src/gallium/state_trackers/python/retrace/interpreter.py b/src/gallium/state_trackers/python/retrace/interpreter.py
index 5d4d04498b2..5885e162c28 100755
--- a/src/gallium/state_trackers/python/retrace/interpreter.py
+++ b/src/gallium/state_trackers/python/retrace/interpreter.py
@@ -43,19 +43,27 @@ except ImportError:
         return struct.unpack(fmt, buf[offset:offset + size])
 
 
-def make_image(surface):
+def make_image(surface, x=None, y=None, w=None, h=None):
+    if x is None:
+        x = 0
+    if y is None:
+        y = 0
+    if w is None:
+        w = surface.width - x
+    if h is None:
+        h = surface.height - y
     data = surface.get_tile_rgba8(0, 0, surface.width, surface.height)
 
     import Image
     outimage = Image.fromstring('RGBA', (surface.width, surface.height), data, "raw", 'RGBA', 0, 1)
     return outimage
 
-def save_image(filename, surface):
-    outimage = make_image(surface)
+def save_image(filename, surface, x=None, y=None, w=None, h=None):
+    outimage = make_image(surface, x, y, w, h)
     outimage.save(filename, "PNG")
 
-def show_image(surface, title):
-    outimage = make_image(surface)
+def show_image(surface, title, x=None, y=None, w=None, h=None):
+    outimage = make_image(surface, x, y, w, h)
     
     import Tkinter as tk
     from PIL import Image, ImageTk
@@ -305,7 +313,11 @@ class Screen(Object):
     def get_tex_transfer(self, texture, face, level, zslice, usage, x, y, w, h):
         if texture is None:
             return None
-        return Transfer(texture.get_surface(face, level, zslice), x, y, w, h)
+        transfer = Transfer(texture.get_surface(face, level, zslice), x, y, w, h)
+        if transfer and usage != gallium.PIPE_TRANSFER_WRITE:
+            if self.interpreter.options.all:
+                self.interpreter.present(transfer.surface, 'transf_read', x, y, w, h)
+        return transfer
     
     def tex_transfer_destroy(self, transfer):
         self.interpreter.unregister_object(transfer)
@@ -314,6 +326,8 @@ class Screen(Object):
         if transfer is None:
             return
         transfer.surface.put_tile_raw(transfer.x, transfer.y, transfer.w, transfer.h, data, stride)
+        if self.interpreter.options.all:
+            self.interpreter.present(transfer.surface, 'transf_write', transfer.x, transfer.y, transfer.w, transfer.h)
 
     def user_buffer_create(self, data, size):
         # We don't really care to distinguish between user and regular buffers
@@ -577,6 +591,14 @@ class Context(Object):
         self.real.draw_range_elements(indexBuffer, indexSize, minIndex, maxIndex, mode, start, count)
         self._set_dirty()
         
+    def is_texture_referenced(self, texture, face, level):
+        #return self.real.is_texture_referenced(format, texture, face, level)
+        pass
+    
+    def is_buffer_referenced(self, buf):
+        #return self.real.is_buffer_referenced(format, buf)
+        pass
+    
     def _set_dirty(self):
         if self.interpreter.options.step:
             self._present()
@@ -602,6 +624,9 @@ class Context(Object):
     
         if self.cbufs and self.cbufs[0]:
             self.interpreter.present(self.cbufs[0], "cbuf")
+        if self.zsbuf:
+            if self.interpreter.options.all:
+                self.interpreter.present(self.zsbuf, "zsbuf")
     
 
 class Interpreter(parser.TraceDumper):
@@ -671,16 +696,16 @@ class Interpreter(parser.TraceDumper):
     def verbosity(self, level):
         return self.options.verbosity >= level
 
-    def present(self, surface, description):
+    def present(self, surface, description, x=None, y=None, w=None, h=None):
         if self.call_no < self.options.start:
             return
 
         if self.options.images:
-            filename = '%s_%04u.png' % (description, self.call_no)
-            save_image(filename, surface)
+            filename = '%04u_%s.png' % (self.call_no, description)
+            save_image(filename, surface, x, y, w, h)
         else:
             title = '%u. %s' % (self.call_no, description)
-            show_image(surface, title)
+            show_image(surface, title, x, y, w, h)
     
 
 class Main(parser.Main):
@@ -690,6 +715,7 @@ class Main(parser.Main):
         optparser.add_option("-q", "--quiet", action="store_const", const=0, dest="verbosity", help="no messages")
         optparser.add_option("-v", "--verbose", action="count", dest="verbosity", default=1, help="increase verbosity level")
         optparser.add_option("-i", "--images", action="store_true", dest="images", default=False, help="save images instead of showing them")
+        optparser.add_option("-a", "--all", action="store_true", dest="all", default=False, help="show depth, stencil, and transfers")
         optparser.add_option("-s", "--step", action="store_true", dest="step", default=False, help="step trhough every draw")
         optparser.add_option("-f", "--from", action="store", type="int", dest="start", default=0, help="from call no")
         optparser.add_option("-t", "--to", action="store", type="int", dest="stop", default=0, help="until call no")
diff --git a/src/gallium/state_trackers/python/tests/base.py b/src/gallium/state_trackers/python/tests/base.py
index 1fa7fe6f3b3..202ccfc3509 100755
--- a/src/gallium/state_trackers/python/tests/base.py
+++ b/src/gallium/state_trackers/python/tests/base.py
@@ -46,6 +46,14 @@ for name, value in globals().items():
     if name.startswith("PIPE_FORMAT_") and isinstance(value, int):
         formats[value] = name
 
+def is_depth_stencil_format(format):
+    # FIXME: make and use binding to pf_is_depth_stencil
+    return format in (
+        PIPE_FORMAT_Z32_UNORM,
+        PIPE_FORMAT_Z24S8_UNORM,
+        PIPE_FORMAT_Z24X8_UNORM,
+        PIPE_FORMAT_Z16_UNORM,
+    )
 
 def make_image(width, height, rgba):
     import Image
diff --git a/src/gallium/state_trackers/python/tests/texture_sample.py b/src/gallium/state_trackers/python/tests/texture_sample.py
index a382424667c..c7b78abbbec 100755
--- a/src/gallium/state_trackers/python/tests/texture_sample.py
+++ b/src/gallium/state_trackers/python/tests/texture_sample.py
@@ -99,7 +99,7 @@ def is_pot(n):
     return n & (n - 1) == 0
       
                 
-class TextureTest(TestCase):
+class TextureColorSampleTest(TestCase):
     
     tags = (
         'target',
@@ -286,6 +286,206 @@ class TextureTest(TestCase):
         self.assert_rgba(cbuf, x, y, w, h, expected_rgba, 4.0/256, 0.85)
         
 
+class TextureDepthSampleTest(TestCase):
+    
+    tags = (
+        'target',
+        'format',
+        'width',
+        'height',
+        'depth',
+        'last_level',
+        'face',
+        'level',
+        'zslice',
+    )
+
+    def test(self):
+        dev = self.dev
+        
+        target = self.target
+        format = self.format
+        width = self.width
+        height = self.height
+        depth = self.depth
+        last_level = self.last_level
+        face = self.face
+        level = self.level
+        zslice = self.zslice
+        
+        tex_usage = PIPE_TEXTURE_USAGE_SAMPLER
+        geom_flags = 0
+        if width != height:
+            geom_flags |= PIPE_TEXTURE_GEOM_NON_SQUARE
+        if not is_pot(width) or not is_pot(height) or not is_pot(depth):
+            geom_flags |= PIPE_TEXTURE_GEOM_NON_POWER_OF_TWO
+        
+        if not dev.is_format_supported(format, target, tex_usage, geom_flags):
+            raise TestSkip
+        
+        ctx = self.dev.context_create()
+    
+        # disabled blending/masking
+        blend = Blend()
+        blend.rgb_src_factor = PIPE_BLENDFACTOR_ONE
+        blend.alpha_src_factor = PIPE_BLENDFACTOR_ONE
+        blend.rgb_dst_factor = PIPE_BLENDFACTOR_ZERO
+        blend.alpha_dst_factor = PIPE_BLENDFACTOR_ZERO
+        blend.colormask = PIPE_MASK_RGBA
+        ctx.set_blend(blend)
+    
+        # depth/stencil/alpha
+        depth_stencil_alpha = DepthStencilAlpha()
+        depth_stencil_alpha.depth.enabled = 1
+        depth_stencil_alpha.depth.writemask = 1
+        depth_stencil_alpha.depth.func = PIPE_FUNC_LESS
+        ctx.set_depth_stencil_alpha(depth_stencil_alpha)
+    
+        # rasterizer
+        rasterizer = Rasterizer()
+        rasterizer.front_winding = PIPE_WINDING_CW
+        rasterizer.cull_mode = PIPE_WINDING_NONE
+        rasterizer.bypass_vs_clip_and_viewport = 1
+        ctx.set_rasterizer(rasterizer)
+    
+        # samplers
+        sampler = Sampler()
+        sampler.wrap_s = PIPE_TEX_WRAP_CLAMP_TO_EDGE
+        sampler.wrap_t = PIPE_TEX_WRAP_CLAMP_TO_EDGE
+        sampler.wrap_r = PIPE_TEX_WRAP_CLAMP_TO_EDGE
+        sampler.min_mip_filter = PIPE_TEX_MIPFILTER_NEAREST
+        sampler.min_img_filter = PIPE_TEX_MIPFILTER_NEAREST
+        sampler.mag_img_filter = PIPE_TEX_MIPFILTER_NEAREST
+        sampler.normalized_coords = 1
+        sampler.min_lod = 0
+        sampler.max_lod = PIPE_MAX_TEXTURE_LEVELS - 1
+        ctx.set_sampler(0, sampler)
+    
+        #  texture 
+        texture = dev.texture_create(
+            target = target,
+            format = format, 
+            width = width, 
+            height = height,
+            depth = depth, 
+            last_level = last_level,
+            tex_usage = tex_usage,
+        )
+        
+        expected_rgba = FloatArray(height*width*4) 
+        texture.get_surface(
+            face = face,
+            level = level,
+            zslice = zslice,
+        ).sample_rgba(expected_rgba)
+        
+        ctx.set_sampler_texture(0, texture)
+
+        #  framebuffer 
+        cbuf_tex = dev.texture_create(
+            PIPE_FORMAT_A8R8G8B8_UNORM, 
+            width, 
+            height,
+            tex_usage = PIPE_TEXTURE_USAGE_RENDER_TARGET,
+        )
+
+        zsbuf_tex = dev.texture_create(
+            PIPE_FORMAT_Z24X8_UNORM, 
+            width, 
+            height,
+            tex_usage = PIPE_TEXTURE_USAGE_RENDER_TARGET,
+        )
+
+        cbuf = cbuf_tex.get_surface()
+        zsbuf = zsbuf_tex.get_surface()
+        fb = Framebuffer()
+        fb.width = width
+        fb.height = height
+        fb.nr_cbufs = 1
+        fb.set_cbuf(0, cbuf)
+        fb.set_zsbuf(zsbuf)
+        ctx.set_framebuffer(fb)
+        rgba = FloatArray(4);
+        rgba[0] = 0.5
+        rgba[1] = 0.5
+        rgba[2] = 0.5
+        rgba[3] = 0.5
+        ctx.clear(PIPE_CLEAR_DEPTHSTENCIL, rgba, 1.0, 0)
+        del fb
+    
+        # vertex shader
+        vs = Shader('''
+            VERT1.1
+            DCL IN[0], POSITION, CONSTANT
+            DCL IN[1], GENERIC, CONSTANT
+            DCL OUT[0], POSITION, CONSTANT
+            DCL OUT[1], GENERIC, CONSTANT
+            0:MOV OUT[0], IN[0]
+            1:MOV OUT[1], IN[1]
+            2:END
+        ''')
+        #vs.dump()
+        ctx.set_vertex_shader(vs)
+    
+        # fragment shader
+        op = {
+            PIPE_TEXTURE_1D: "1D", 
+            PIPE_TEXTURE_2D: "2D", 
+            PIPE_TEXTURE_3D: "3D", 
+            PIPE_TEXTURE_CUBE: "CUBE",
+        }[target]
+        fs = Shader('''
+            FRAG1.1
+            DCL IN[0], GENERIC[0], LINEAR
+            DCL SAMP[0], CONSTANT
+            DCL OUT[0].z, POSITION
+            0:TEX OUT[0].z, IN[0], SAMP[0], %s
+            1:END
+        ''' % op)
+        #fs.dump()
+        ctx.set_fragment_shader(fs)
+
+        nverts = 4
+        nattrs = 2
+        verts = FloatArray(nverts * nattrs * 4)
+    
+        x = 0
+        y = 0
+        w, h = minify((width, height), level)
+    
+        pos = [
+            [x, y],
+            [x+w, y],
+            [x+w, y+h],
+            [x, y+h],
+        ]
+    
+        tex = tex_coords(texture, face, level, zslice)
+    
+        for i in range(0, 4):
+            j = 8*i
+            verts[j + 0] = pos[i][0] # x
+            verts[j + 1] = pos[i][1] # y
+            verts[j + 2] = 0.0 # z
+            verts[j + 3] = 1.0 # w
+            verts[j + 4] = tex[i][0] # s
+            verts[j + 5] = tex[i][1] # r
+            verts[j + 6] = tex[i][2] # q
+            verts[j + 7] = 1.0
+    
+        ctx.draw_vertices(PIPE_PRIM_TRIANGLE_FAN,
+                          nverts, 
+                          nattrs, 
+                          verts)
+    
+        ctx.flush()
+    
+        zsbuf = zsbuf_tex.get_surface()
+        
+        self.assert_rgba(zsbuf, x, y, w, h, expected_rgba, 4.0/256, 0.85)
+        
+
+
 
 def main():
     dev = Device()
@@ -297,18 +497,13 @@ def main():
         PIPE_TEXTURE_3D,
     ]
     
-    formats = [
+    color_formats = [
         PIPE_FORMAT_A8R8G8B8_UNORM,
         PIPE_FORMAT_X8R8G8B8_UNORM,
         #PIPE_FORMAT_A8R8G8B8_SRGB,
         PIPE_FORMAT_R5G6B5_UNORM,
         PIPE_FORMAT_A1R5G5B5_UNORM,
         PIPE_FORMAT_A4R4G4B4_UNORM,
-        #PIPE_FORMAT_Z32_UNORM,
-        #PIPE_FORMAT_Z24S8_UNORM,
-        #PIPE_FORMAT_Z24X8_UNORM,
-        #PIPE_FORMAT_Z16_UNORM,
-        #PIPE_FORMAT_S8_UNORM,
         PIPE_FORMAT_A8_UNORM,
         PIPE_FORMAT_L8_UNORM,
         PIPE_FORMAT_YCBCR,
@@ -318,6 +513,13 @@ def main():
         #PIPE_FORMAT_DXT5_RGBA,
     ]
     
+    depth_formats = [
+        PIPE_FORMAT_Z32_UNORM,
+        PIPE_FORMAT_Z24S8_UNORM,
+        PIPE_FORMAT_Z24X8_UNORM,
+        PIPE_FORMAT_Z16_UNORM,
+    ]
+    
     sizes = [64, 32, 16, 8, 4, 2, 1]
     #sizes = [1020, 508, 252, 62, 30, 14, 6, 3]
     #sizes = [64]
@@ -332,8 +534,8 @@ def main():
         PIPE_TEX_FACE_NEG_Z,
     ]
 
-    for target in targets:
-        for format in formats:
+    for format in color_formats:
+        for target in targets:
             for size in sizes:
                 if target == PIPE_TEXTURE_3D:
                     depth = size
@@ -347,7 +549,7 @@ def main():
                         for level in range(0, last_level + 1):
                             zslice = 0
                             while zslice < depth >> level:
-                                test = TextureTest(
+                                test = TextureColorSampleTest(
                                     dev = dev,
                                     target = target,
                                     format = format, 
@@ -361,6 +563,27 @@ def main():
                                 )
                                 suite.add_test(test)
                                 zslice = (zslice + 1)*2 - 1
+    for format in depth_formats:
+        target = PIPE_TEXTURE_2D
+        depth = 1
+        face = 0
+        last_level = 0
+        level = 0
+        zslice = 0
+        for size in sizes:
+            test = TextureDepthSampleTest(
+                dev = dev,
+                target = target,
+                format = format, 
+                width = size,
+                height = size,
+                depth = depth,
+                last_level = last_level,
+                face = face,
+                level = level,
+                zslice = zslice,
+            )
+            suite.add_test(test)
     suite.run()
 
 
diff --git a/src/gallium/state_trackers/wgl/shared/stw_context.c b/src/gallium/state_trackers/wgl/shared/stw_context.c
index 662b5fbcd2a..9df1ab76526 100644
--- a/src/gallium/state_trackers/wgl/shared/stw_context.c
+++ b/src/gallium/state_trackers/wgl/shared/stw_context.c
@@ -47,6 +47,23 @@
 #include "stw_context.h"
 #include "stw_tls.h"
 
+
+static INLINE struct stw_context *
+stw_context(GLcontext *glctx)
+{
+   if(!glctx)
+      return NULL;
+   assert(glctx->DriverCtx);
+   return (struct stw_context *)glctx->DriverCtx;
+}
+
+static INLINE struct stw_context *
+stw_current_context(void)
+{
+   GET_CURRENT_CONTEXT( glctx );
+   return stw_context(glctx);
+}
+
 BOOL
 stw_copy_context(
    UINT_PTR hglrcSrc,
@@ -99,6 +116,18 @@ stw_share_lists(
    return ret;
 }
 
+static void
+stw_viewport(GLcontext * glctx, GLint x, GLint y,
+             GLsizei width, GLsizei height)
+{
+   struct stw_context *ctx = (struct stw_context *)glctx->DriverCtx;
+   struct stw_framebuffer *fb;
+   
+   fb = stw_framebuffer_from_hdc( ctx->hdc );
+   if(fb)
+      stw_framebuffer_update(fb);
+}
+
 UINT_PTR
 stw_create_layer_context(
    HDC hdc,
@@ -158,6 +187,7 @@ stw_create_layer_context(
       goto no_st_ctx;
 
    ctx->st->ctx->DriverCtx = ctx;
+   ctx->st->ctx->Driver.Viewport = stw_viewport;
 
    pipe_mutex_lock( stw_dev->mutex );
    ctx->hglrc = handle_table_add(stw_dev->ctx_table, ctx);
@@ -194,11 +224,10 @@ stw_delete_context(
    pipe_mutex_unlock( stw_dev->mutex );
 
    if (ctx) {
-      GLcontext *glctx = ctx->st->ctx;
-      GET_CURRENT_CONTEXT( glcurctx );
-
+      struct stw_context *curctx = stw_current_context();
+      
       /* Unbind current if deleting current context. */
-      if (glcurctx == glctx)
+      if (curctx == ctx)
          st_make_current( NULL, NULL, NULL );
 
       st_destroy_context(ctx->st);
@@ -230,13 +259,8 @@ stw_release_context(
     * current for this thread.  We should check that and return False
     * if not the case.
     */
-   {
-      GLcontext *glctx = ctx->st->ctx;
-      GET_CURRENT_CONTEXT( glcurctx );
-
-      if (glcurctx != glctx)
-         return FALSE;
-   }
+   if (ctx != stw_current_context())
+      return FALSE;
 
    if (stw_make_current( NULL, 0 ) == FALSE)
       return FALSE;
@@ -248,14 +272,9 @@ stw_release_context(
 UINT_PTR
 stw_get_current_context( void )
 {
-   GET_CURRENT_CONTEXT( glcurctx );
    struct stw_context *ctx;
 
-   if(!glcurctx)
-      return 0;
-   
-   ctx = (struct stw_context *)glcurctx->DriverCtx;
-   assert(ctx);
+   ctx = stw_current_context();
    if(!ctx)
       return 0;
    
@@ -265,14 +284,9 @@ stw_get_current_context( void )
 HDC
 stw_get_current_dc( void )
 {
-   GET_CURRENT_CONTEXT( glcurctx );
    struct stw_context *ctx;
 
-   if(!glcurctx)
-      return NULL;
-   
-   ctx = (struct stw_context *)glcurctx->DriverCtx;
-   assert(ctx);
+   ctx = stw_current_context();
    if(!ctx)
       return NULL;
    
@@ -284,23 +298,24 @@ stw_make_current(
    HDC hdc,
    UINT_PTR hglrc )
 {
+   struct stw_context *curctx;
    struct stw_context *ctx;
-   GET_CURRENT_CONTEXT( glcurctx );
    struct stw_framebuffer *fb;
 
    if (!stw_dev)
       goto fail;
 
-   if (glcurctx != NULL) {
-      struct stw_context *curctx;
-      curctx = (struct stw_context *) glcurctx->DriverCtx;
-
+   curctx = stw_current_context();
+   if (curctx != NULL) {
       if (curctx->hglrc != hglrc)
-	 st_flush(glcurctx->st, PIPE_FLUSH_RENDER_CACHE, NULL);
+	 st_flush(curctx->st, PIPE_FLUSH_RENDER_CACHE, NULL);
       
       /* Return if already current. */
-      if (curctx->hglrc == hglrc && curctx->hdc == hdc)
-         return TRUE;
+      if (curctx->hglrc == hglrc && curctx->hdc == hdc) {
+         ctx = curctx;
+         fb = stw_framebuffer_from_hdc( hdc );
+         goto success;
+      }
    }
 
    if (hdc == NULL || hglrc == 0) {
@@ -344,7 +359,10 @@ stw_make_current(
    if(!st_make_current( ctx->st, fb->stfb, fb->stfb ))
       goto fail;
 
-   stw_framebuffer_resize(fb);
+success:
+   assert(fb);
+   if(fb)
+      stw_framebuffer_update(fb);
    
    return TRUE;
 
diff --git a/src/gallium/state_trackers/wgl/shared/stw_device.c b/src/gallium/state_trackers/wgl/shared/stw_device.c
index 1a6b29807d7..ce466241463 100644
--- a/src/gallium/state_trackers/wgl/shared/stw_device.c
+++ b/src/gallium/state_trackers/wgl/shared/stw_device.c
@@ -30,6 +30,7 @@
 #include "glapi/glthread.h"
 #include "util/u_debug.h"
 #include "pipe/p_screen.h"
+#include "state_tracker/st_public.h"
 
 #ifdef DEBUG
 #include "trace/tr_screen.h"
@@ -63,15 +64,39 @@ stw_flush_frontbuffer(struct pipe_screen *screen,
 {
    const struct stw_winsys *stw_winsys = stw_dev->stw_winsys;
    HDC hdc = (HDC)context_private;
+   struct stw_framebuffer *fb;
    
+   fb = stw_framebuffer_from_hdc( hdc );
+   /* fb can be NULL if window was destroyed already */
+   if (fb) {
+      pipe_mutex_lock( fb->mutex );
+
+#if DEBUG
+      {
+         struct pipe_surface *surface2;
+   
+         if(!st_get_framebuffer_surface( fb->stfb, ST_SURFACE_FRONT_LEFT, &surface2 ))
+            assert(0);
+         else
+            assert(surface2 == surface);
+      }
+#endif
+
 #ifdef DEBUG
-   if(stw_dev->trace_running) {
-      screen = trace_screen(screen)->screen;
-      surface = trace_surface(surface)->surface;
-   }
+      if(stw_dev->trace_running) {
+         screen = trace_screen(screen)->screen;
+         surface = trace_surface(surface)->surface;
+      }
 #endif
+   }
    
    stw_winsys->flush_frontbuffer(screen, surface, hdc);
+   
+   if(fb) {
+      stw_framebuffer_update(fb);
+      
+      pipe_mutex_unlock( fb->mutex );
+   }
 }
 
 
@@ -133,20 +158,13 @@ error1:
 boolean
 stw_init_thread(void)
 {
-   if (!stw_tls_init_thread())
-      return FALSE;
-
-   if (!stw_framebuffer_init_thread())
-      return FALSE;
-
-   return TRUE;
+   return stw_tls_init_thread();
 }
 
 
 void
 stw_cleanup_thread(void)
 {
-   stw_framebuffer_cleanup_thread();
    stw_tls_cleanup_thread();
 }
 
diff --git a/src/gallium/state_trackers/wgl/shared/stw_framebuffer.c b/src/gallium/state_trackers/wgl/shared/stw_framebuffer.c
index de37d135dfe..7d0e8f46482 100644
--- a/src/gallium/state_trackers/wgl/shared/stw_framebuffer.c
+++ b/src/gallium/state_trackers/wgl/shared/stw_framebuffer.c
@@ -45,7 +45,7 @@
 #include "stw_tls.h"
 
 
-struct stw_framebuffer *
+static INLINE struct stw_framebuffer *
 stw_framebuffer_from_hwnd_locked(
    HWND hwnd )
 {
@@ -84,7 +84,7 @@ stw_framebuffer_destroy_locked(
  * @sa http://msdn.microsoft.com/en-us/library/ms644975(VS.85).aspx
  * @sa http://msdn.microsoft.com/en-us/library/ms644960(VS.85).aspx
  */
-static LRESULT CALLBACK
+LRESULT CALLBACK
 stw_call_window_proc(
    int nCode,
    WPARAM wParam,
@@ -111,14 +111,10 @@ stw_call_window_proc(
          unsigned width = LOWORD( pParams->lParam );
          unsigned height = HIWORD( pParams->lParam );
          
-         /* FIXME: The mesa statetracker makes the assumptions that only
-          * one context is using the framebuffer, and that that context is the 
-          * current one. However neither holds true, as WGL allows more than
-          * one context to be bound to the same drawable, and this function can 
-          * be called from any thread.
-          */
          pipe_mutex_lock( fb->mutex );
-         st_resize_framebuffer( fb->stfb, width, height );
+         fb->must_resize = TRUE;
+         fb->width = width;
+         fb->height = height;
          pipe_mutex_unlock( fb->mutex );
       }
    }
@@ -139,6 +135,31 @@ stw_call_window_proc(
 }
 
 
+static void
+stw_framebuffer_get_size( struct stw_framebuffer *fb )
+{
+   unsigned width, height;
+   RECT rect;
+
+   assert(fb->hWnd);
+   
+   GetClientRect( fb->hWnd, &rect );
+   width = rect.right - rect.left;
+   height = rect.bottom - rect.top;
+
+   if(width < 1)
+      width = 1;
+   if(height < 1)
+      height = 1;
+
+   if(width != fb->width || height != fb->height) {
+      fb->must_resize = TRUE;
+      fb->width = width; 
+      fb->height = height; 
+   }
+}
+
+
 /**
  * Create a new framebuffer object which will correspond to the given HDC.
  */
@@ -168,6 +189,8 @@ stw_framebuffer_create_locked(
 
    stw_pixelformat_visual(&fb->visual, pfi);
    
+   stw_framebuffer_get_size(fb);
+
    pipe_mutex_init( fb->mutex );
 
    fb->next = stw_dev->fb_head;
@@ -177,32 +200,6 @@ stw_framebuffer_create_locked(
 }
 
 
-static void
-stw_framebuffer_get_size( struct stw_framebuffer *fb, GLuint *pwidth, GLuint *pheight )
-{
-   GLuint width, height;
-
-   if (fb->hWnd) {
-      RECT rect;
-      GetClientRect( fb->hWnd, &rect );
-      width = rect.right - rect.left;
-      height = rect.bottom - rect.top;
-   }
-   else {
-      width = GetDeviceCaps( fb->hDC, HORZRES );
-      height = GetDeviceCaps( fb->hDC, VERTRES );
-   }
-
-   if(width < 1)
-      width = 1;
-   if(height < 1)
-      height = 1;
-
-   *pwidth = width; 
-   *pheight = height; 
-}
-
-
 BOOL
 stw_framebuffer_allocate(
    struct stw_framebuffer *fb)
@@ -212,7 +209,6 @@ stw_framebuffer_allocate(
    if(!fb->stfb) {
       const struct stw_pixelformat_info *pfi = fb->pfi;
       enum pipe_format colorFormat, depthFormat, stencilFormat;
-      GLuint width, height; 
 
       colorFormat = pfi->color_format;
       
@@ -228,16 +224,21 @@ stw_framebuffer_allocate(
       else
          stencilFormat = PIPE_FORMAT_NONE;
    
-      stw_framebuffer_get_size(fb, &width, &height);
-      
+      assert(fb->must_resize);
+      assert(fb->width);
+      assert(fb->height);
+
       fb->stfb = st_create_framebuffer(
          &fb->visual,
          colorFormat,
          depthFormat,
          stencilFormat,
-         width,
-         height,
+         fb->width,
+         fb->height,
          (void *) fb );
+      
+      // to notify the context
+      fb->must_resize = TRUE;
    }
    
    pipe_mutex_unlock( fb->mutex );
@@ -246,14 +247,29 @@ stw_framebuffer_allocate(
 }
 
 
+/**
+ * Update the framebuffer's size if necessary.
+ */
 void
-stw_framebuffer_resize(
+stw_framebuffer_update(
    struct stw_framebuffer *fb)
 {
-   GLuint width, height; 
    assert(fb->stfb);
-   stw_framebuffer_get_size(fb, &width, &height);
-   st_resize_framebuffer(fb->stfb, width, height);
+   assert(fb->height);
+   assert(fb->width);
+   
+   /* XXX: It would be nice to avoid checking the size again -- in theory  
+    * stw_call_window_proc would have cought the resize and stored the right 
+    * size already, but unfortunately threads created before the DllMain is 
+    * called don't get a DLL_THREAD_ATTACH notification, and there is no way
+    * to know of their existing without using the not very portable PSAPI.
+    */
+   stw_framebuffer_get_size(fb);
+   
+   if(fb->must_resize) {
+      st_resize_framebuffer(fb->stfb, fb->width, fb->height);
+      fb->must_resize = FALSE;
+   }
 }                      
 
 
@@ -379,6 +395,9 @@ stw_swap_buffers(
    if (fb == NULL)
       return FALSE;
 
+   if (!(fb->pfi->pfd.dwFlags & PFD_DOUBLEBUFFER))
+      return TRUE;
+
    pipe_mutex_lock( fb->mutex );
 
    /* If we're swapping the buffer associated with the current context
@@ -403,6 +422,8 @@ stw_swap_buffers(
 
    stw_dev->stw_winsys->flush_frontbuffer( screen, surface, hdc );
    
+   stw_framebuffer_update(fb);
+   
    pipe_mutex_unlock( fb->mutex );
    
    return TRUE;
@@ -419,38 +440,3 @@ stw_swap_layer_buffers(
 
    return FALSE;
 }
-
-
-boolean
-stw_framebuffer_init_thread(void)
-{
-   struct stw_tls_data *tls_data;
-   
-   tls_data = stw_tls_get_data();
-   if(!tls_data)
-      return FALSE;
-   
-   tls_data->hCallWndProcHook = SetWindowsHookEx(WH_CALLWNDPROC,
-                                                 stw_call_window_proc,
-                                                 NULL,
-                                                 GetCurrentThreadId());
-   if(tls_data->hCallWndProcHook == NULL)
-      return FALSE;
-   
-   return TRUE;
-}
-
-void
-stw_framebuffer_cleanup_thread(void)
-{
-   struct stw_tls_data *tls_data;
-   
-   tls_data = stw_tls_get_data();
-   if(!tls_data)
-      return;
-   
-   if(tls_data->hCallWndProcHook) {
-      UnhookWindowsHookEx(tls_data->hCallWndProcHook);
-      tls_data->hCallWndProcHook = NULL;
-   }
-}
diff --git a/src/gallium/state_trackers/wgl/shared/stw_framebuffer.h b/src/gallium/state_trackers/wgl/shared/stw_framebuffer.h
index e7fa51c3a81..759e06b8914 100644
--- a/src/gallium/state_trackers/wgl/shared/stw_framebuffer.h
+++ b/src/gallium/state_trackers/wgl/shared/stw_framebuffer.h
@@ -51,6 +51,11 @@ struct stw_framebuffer
    pipe_mutex mutex;
    struct st_framebuffer *stfb;
    
+   /* FIXME: Make this work for multiple contexts bound to the same framebuffer */
+   boolean must_resize;
+   unsigned width;
+   unsigned height;
+   
    /** This is protected by stw_device::mutex, not the mutex above */
    struct stw_framebuffer *next;
 };
@@ -65,7 +70,7 @@ stw_framebuffer_allocate(
    struct stw_framebuffer *fb );
 
 void
-stw_framebuffer_resize(
+stw_framebuffer_update(
    struct stw_framebuffer *fb);
 
 void
@@ -79,10 +84,4 @@ struct stw_framebuffer *
 stw_framebuffer_from_hdc(
    HDC hdc );
 
-boolean
-stw_framebuffer_init_thread(void);
-
-void
-stw_framebuffer_cleanup_thread(void);
-
 #endif /* STW_FRAMEBUFFER_H */
diff --git a/src/gallium/state_trackers/wgl/shared/stw_tls.c b/src/gallium/state_trackers/wgl/shared/stw_tls.c
index 0c18a52352c..4bd6a9289c9 100644
--- a/src/gallium/state_trackers/wgl/shared/stw_tls.c
+++ b/src/gallium/state_trackers/wgl/shared/stw_tls.c
@@ -51,9 +51,23 @@ stw_tls_data_create()
 
    data = CALLOC_STRUCT(stw_tls_data);
    if (!data)
-      return NULL;
+      goto no_data;
+
+   data->hCallWndProcHook = SetWindowsHookEx(WH_CALLWNDPROC,
+                                             stw_call_window_proc,
+                                             NULL,
+                                             GetCurrentThreadId());
+   if(data->hCallWndProcHook == NULL)
+      goto no_hook;
+
+   TlsSetValue(tlsIndex, data);
 
    return data;
+
+no_hook:
+   FREE(data);
+no_data:
+   return NULL;
 }
 
 boolean
@@ -69,8 +83,6 @@ stw_tls_init_thread(void)
    if(!data)
       return FALSE;
 
-   TlsSetValue(tlsIndex, data);
-
    return TRUE;
 }
 
@@ -84,8 +96,16 @@ stw_tls_cleanup_thread(void)
    }
 
    data = (struct stw_tls_data *) TlsGetValue(tlsIndex);
-   TlsSetValue(tlsIndex, NULL);
-   FREE(data);
+   if(data) {
+      TlsSetValue(tlsIndex, NULL);
+   
+      if(data->hCallWndProcHook) {
+         UnhookWindowsHookEx(data->hCallWndProcHook);
+         data->hCallWndProcHook = NULL;
+      }
+   
+      FREE(data);
+   }
 }
 
 void
@@ -110,12 +130,9 @@ stw_tls_get_data(void)
    if(!data) {
       /* DllMain is called with DLL_THREAD_ATTACH only by threads created after 
        * the DLL is loaded by the process */
-      
       data = stw_tls_data_create();
       if(!data)
          return NULL;
-
-      TlsSetValue(tlsIndex, data);
    }
 
    return data;
diff --git a/src/gallium/state_trackers/wgl/shared/stw_tls.h b/src/gallium/state_trackers/wgl/shared/stw_tls.h
index 6af8be70c9f..fbf8b1cbee4 100644
--- a/src/gallium/state_trackers/wgl/shared/stw_tls.h
+++ b/src/gallium/state_trackers/wgl/shared/stw_tls.h
@@ -50,4 +50,10 @@ stw_tls_cleanup(void);
 struct stw_tls_data *
 stw_tls_get_data(void);
 
+LRESULT CALLBACK
+stw_call_window_proc(
+   int nCode,
+   WPARAM wParam,
+   LPARAM lParam );
+
 #endif /* STW_TLS_H */