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

28 files changed, 365 insertions, 237 deletions

diff --git a/src/gallium/auxiliary/cso_cache/cso_context.c b/src/gallium/auxiliary/cso_cache/cso_context.c
index eef898f4865..a1a3a9efaf3 100644
--- a/src/gallium/auxiliary/cso_cache/cso_context.c
+++ b/src/gallium/auxiliary/cso_cache/cso_context.c
@@ -84,7 +84,7 @@ static boolean delete_blend_state(struct cso_context *ctx, void *state)
 {
    struct cso_blend *cso = (struct cso_blend *)state;
 
-   if (ctx->blend == state)
+   if (ctx->blend == cso->data)
       return FALSE;
 
    if (cso->delete_state)
diff --git a/src/gallium/auxiliary/draw/draw_pipe_vbuf.c b/src/gallium/auxiliary/draw/draw_pipe_vbuf.c
index 2a19e6916af..86a7d1c7303 100644
--- a/src/gallium/auxiliary/draw/draw_pipe_vbuf.c
+++ b/src/gallium/auxiliary/draw/draw_pipe_vbuf.c
@@ -42,6 +42,7 @@
 #include "draw_vertex.h"
 #include "draw_pipe.h"
 #include "translate/translate.h"
+#include "translate/translate_cache.h"
 
 
 /**
@@ -75,6 +76,8 @@ struct vbuf_stage {
    /* Cache point size somewhere it's address won't change:
     */
    float point_size;
+
+   struct translate_cache *cache;
 };
 
 
@@ -220,7 +223,6 @@ vbuf_set_prim( struct vbuf_stage *vbuf, uint prim )
    /* Translate from pipeline vertices to hw vertices.
     */
    dst_offset = 0;
-   memset(&hw_key, 0, sizeof(hw_key));
 
    for (i = 0; i < vbuf->vinfo->num_attribs; i++) {
       unsigned emit_sz = 0;
@@ -277,12 +279,10 @@ vbuf_set_prim( struct vbuf_stage *vbuf, uint prim )
    /* Don't bother with caching at this stage:
     */
    if (!vbuf->translate ||
-       memcmp(&vbuf->translate->key, &hw_key, sizeof(hw_key)) != 0) 
+       translate_key_compare(&vbuf->translate->key, &hw_key) != 0) 
    {
-      if (vbuf->translate)
-	 vbuf->translate->release(vbuf->translate);
-
-      vbuf->translate = translate_create( &hw_key );
+      translate_key_sanitize(&hw_key);
+      vbuf->translate = translate_cache_find(vbuf->cache, &hw_key);
 
       vbuf->translate->set_buffer(vbuf->translate, 1, &vbuf->point_size, 0);
    }
@@ -433,6 +433,9 @@ static void vbuf_destroy( struct draw_stage *stage )
    if (vbuf->render)
       vbuf->render->destroy( vbuf->render );
 
+   if (vbuf->cache)
+      translate_cache_destroy(vbuf->cache);
+
    FREE( stage );
 }
 
@@ -463,6 +466,11 @@ struct draw_stage *draw_vbuf_stage( struct draw_context *draw,
 					    16 );
    if (!vbuf->indices)
       goto fail;
+
+   vbuf->cache = translate_cache_create();
+   if (!vbuf->cache) 
+      goto fail;
+      
    
    vbuf->vertices = NULL;
    vbuf->vertex_ptr = vbuf->vertices;
diff --git a/src/gallium/auxiliary/draw/draw_pt_emit.c b/src/gallium/auxiliary/draw/draw_pt_emit.c
index f9ac16786ef..ce3a153f647 100644
--- a/src/gallium/auxiliary/draw/draw_pt_emit.c
+++ b/src/gallium/auxiliary/draw/draw_pt_emit.c
@@ -58,8 +58,6 @@ void draw_pt_emit_prepare( struct pt_emit *emit,
       return;
    }
 
-   memset(&hw_key, 0, sizeof(hw_key));
-
    /* Must do this after set_primitive() above:
     */
    vinfo = draw->render->get_vertex_info(draw->render);
@@ -122,8 +120,9 @@ void draw_pt_emit_prepare( struct pt_emit *emit,
    hw_key.output_stride = vinfo->size * 4;
 
    if (!emit->translate ||
-       memcmp(&emit->translate->key, &hw_key, sizeof(hw_key)) != 0)
+       translate_key_compare(&emit->translate->key, &hw_key) != 0)
    {
+      translate_key_sanitize(&hw_key);
       emit->translate = translate_cache_find(emit->cache, &hw_key);
    }
 }
@@ -147,7 +146,7 @@ void draw_pt_emit( struct pt_emit *emit,
 
    hw_verts = render->allocate_vertices(render,
 					(ushort)translate->key.output_stride,
-					(ushort)count);
+					(ushort)vertex_count);
    if (!hw_verts) {
       assert(0);
       return;
@@ -197,7 +196,8 @@ struct pt_emit *draw_pt_emit_create( struct draw_context *draw )
 
 void draw_pt_emit_destroy( struct pt_emit *emit )
 {
-   translate_cache_destroy(emit->cache);
+   if (emit->cache)
+      translate_cache_destroy(emit->cache);
 
    FREE(emit);
 }
diff --git a/src/gallium/auxiliary/draw/draw_pt_fetch.c b/src/gallium/auxiliary/draw/draw_pt_fetch.c
index 1f765b73adf..100117a9aef 100644
--- a/src/gallium/auxiliary/draw/draw_pt_fetch.c
+++ b/src/gallium/auxiliary/draw/draw_pt_fetch.c
@@ -65,8 +65,6 @@ void draw_pt_fetch_prepare( struct pt_fetch *fetch,
 
    fetch->vertex_size = vertex_size;
 
-   memset(&key, 0, sizeof(key));
-
    /* Always emit/leave space for a vertex header.
     *
     * It's worth considering whether the vertex headers should contain
@@ -110,8 +108,9 @@ void draw_pt_fetch_prepare( struct pt_fetch *fetch,
 
 
    if (!fetch->translate ||
-       memcmp(&fetch->translate->key, &key, sizeof(key)) != 0)
+       translate_key_compare(&fetch->translate->key, &key) != 0)
    {
+      translate_key_sanitize(&key);
       fetch->translate = translate_cache_find(fetch->cache, &key);
 
       {
diff --git a/src/gallium/auxiliary/draw/draw_pt_fetch_emit.c b/src/gallium/auxiliary/draw/draw_pt_fetch_emit.c
index a4de341df8e..b7b970a297f 100644
--- a/src/gallium/auxiliary/draw/draw_pt_fetch_emit.c
+++ b/src/gallium/auxiliary/draw/draw_pt_fetch_emit.c
@@ -174,8 +174,9 @@ static void fetch_emit_prepare( struct draw_pt_middle_end *middle,
    /* Don't bother with caching at this stage:
     */
    if (!feme->translate ||
-       memcmp(&feme->translate->key, &key, sizeof(key)) != 0) 
+       translate_key_compare(&feme->translate->key, &key) != 0) 
    {
+      translate_key_sanitize(&key);
       feme->translate = translate_cache_find(feme->cache,
                                              &key);
 
diff --git a/src/gallium/auxiliary/gallivm/instructions.cpp b/src/gallium/auxiliary/gallivm/instructions.cpp
index 95a670edafc..1a98491b82a 100644
--- a/src/gallium/auxiliary/gallivm/instructions.cpp
+++ b/src/gallium/auxiliary/gallivm/instructions.cpp
@@ -166,10 +166,9 @@ llvm::Value * Instructions::rsq(llvm::Value *in1)
    Value *abs  = callFAbs(x);
    Value *sqrt = callFSqrt(abs);
 
-   Value *rsqrt = m_builder.CreateFDiv(ConstantFP::get(Type::FloatTy,
-                                                                APFloat(1.f)),
-                                                sqrt,
-                                                name("rsqrt"));
+   Value *rsqrt = m_builder.CreateFDiv(ConstantFP::get(APFloat(1.f)),
+                                       sqrt,
+                                       name("rsqrt"));
    return vectorFromVals(rsqrt, rsqrt, rsqrt, rsqrt);
 }
 
@@ -278,9 +277,8 @@ llvm::Value * Instructions::rcp(llvm::Value *in1)
    Value *x1 = m_builder.CreateExtractElement(in1,
                                               m_storage->constantInt(0),
                                               name("x1"));
-   Value *res = m_builder.CreateFDiv(ConstantFP::get(Type::FloatTy,
-                                                              APFloat(1.f)),
-                                              x1, name("rcp"));
+   Value *res = m_builder.CreateFDiv(ConstantFP::get(APFloat(1.f)),
+                                     x1, name("rcp"));
    return vectorFromVals(res, res, res, res);
 }
 
@@ -319,13 +317,13 @@ llvm::Value * Instructions::dst(llvm::Value *in1, llvm::Value *in2)
                                              m_storage->constantInt(3),
                                              name("w"));
    Value *ry = m_builder.CreateMul(y1, y2, name("tyuy"));
-   return vectorFromVals(ConstantFP::get(Type::FloatTy, APFloat(1.f)),
+   return vectorFromVals(ConstantFP::get(APFloat(1.f)),
                          ry, z, w);
 }
 
 llvm::Value * Instructions::ex2(llvm::Value *in)
 {
-   llvm::Value *val = callPow(ConstantFP::get(Type::FloatTy, APFloat(2.f)),
+   llvm::Value *val = callPow(ConstantFP::get(APFloat(2.f)),
                               m_builder.CreateExtractElement(
                                  in, m_storage->constantInt(0),
                                  name("x1")));
@@ -526,7 +524,7 @@ llvm::Function * Instructions::declarePrintf()
 
 llvm::Value * Instructions::sgt(llvm::Value *in1, llvm::Value *in2)
 {
-   Constant *const1f = ConstantFP::get(Type::FloatTy, APFloat(1.000000e+00f));
+   Constant *const1f = ConstantFP::get(APFloat(1.000000e+00f));
    Constant *const0f = Constant::getNullValue(Type::FloatTy);
 
    std::vector<llvm::Value*> vec1 = extractVector(in1);
@@ -547,7 +545,7 @@ llvm::Value * Instructions::sgt(llvm::Value *in1, llvm::Value *in2)
 }
 llvm::Value * Instructions::sge(llvm::Value *in1, llvm::Value *in2)
 {
-   Constant *const1f = ConstantFP::get(Type::FloatTy, APFloat(1.000000e+00f));
+   Constant *const1f = ConstantFP::get(APFloat(1.000000e+00f));
    Constant *const0f = Constant::getNullValue(Type::FloatTy);
 
    std::vector<llvm::Value*> vec1 = extractVector(in1);
@@ -571,7 +569,7 @@ llvm::Value * Instructions::sge(llvm::Value *in1, llvm::Value *in2)
 
 llvm::Value * Instructions::slt(llvm::Value *in1, llvm::Value *in2)
 {
-   Constant *const1f = ConstantFP::get(Type::FloatTy, APFloat(1.000000e+00f));
+   Constant *const1f = ConstantFP::get(APFloat(1.000000e+00f));
    Constant *const0f = Constant::getNullValue(Type::FloatTy);
 
    std::vector<llvm::Value*> vec1 = extractVector(in1);
@@ -814,10 +812,10 @@ llvm::Function * Instructions::findFunction(int label)
 llvm::Value * Instructions::constVector(float x, float y, float z, float w)
 {
    std::vector<Constant*> vec(4);
-   vec[0] = ConstantFP::get(Type::FloatTy, APFloat(x));
-   vec[1] = ConstantFP::get(Type::FloatTy, APFloat(y));
-   vec[2] = ConstantFP::get(Type::FloatTy, APFloat(z));
-   vec[3] = ConstantFP::get(Type::FloatTy, APFloat(w));
+   vec[0] = ConstantFP::get(APFloat(x));
+   vec[1] = ConstantFP::get(APFloat(y));
+   vec[2] = ConstantFP::get(APFloat(z));
+   vec[3] = ConstantFP::get(APFloat(w));
    return ConstantVector::get(m_floatVecType, vec);
 }
 
diff --git a/src/gallium/auxiliary/gallivm/storage.cpp b/src/gallium/auxiliary/gallivm/storage.cpp
index 9d9fd123604..6f373f6dd5e 100644
--- a/src/gallium/auxiliary/gallivm/storage.cpp
+++ b/src/gallium/auxiliary/gallivm/storage.cpp
@@ -69,10 +69,10 @@ llvm::Constant *Storage::shuffleMask(int vec)
 {
    if (!m_extSwizzleVec) {
       std::vector<Constant*> elems;
-      elems.push_back(ConstantFP::get(Type::FloatTy, APFloat(0.f)));
-      elems.push_back(ConstantFP::get(Type::FloatTy, APFloat(1.f)));
-      elems.push_back(ConstantFP::get(Type::FloatTy, APFloat(0.f)));
-      elems.push_back(ConstantFP::get(Type::FloatTy, APFloat(1.f)));
+      elems.push_back(ConstantFP::get(APFloat(0.f)));
+      elems.push_back(ConstantFP::get(APFloat(1.f)));
+      elems.push_back(ConstantFP::get(APFloat(0.f)));
+      elems.push_back(ConstantFP::get(APFloat(1.f)));
       m_extSwizzleVec = ConstantVector::get(m_floatVecType, elems);
    }
 
@@ -295,10 +295,10 @@ llvm::Value * Storage::immediateElement(int idx)
 void Storage::addImmediate(float *val)
 {
    std::vector<Constant*> vec(4);
-   vec[0] = ConstantFP::get(Type::FloatTy, APFloat(val[0]));
-   vec[1] = ConstantFP::get(Type::FloatTy, APFloat(val[1]));
-   vec[2] = ConstantFP::get(Type::FloatTy, APFloat(val[2]));
-   vec[3] = ConstantFP::get(Type::FloatTy, APFloat(val[3]));
+   vec[0] = ConstantFP::get(APFloat(val[0]));
+   vec[1] = ConstantFP::get(APFloat(val[1]));
+   vec[2] = ConstantFP::get(APFloat(val[2]));
+   vec[3] = ConstantFP::get(APFloat(val[3]));
    m_immediates.push_back(ConstantVector::get(m_floatVecType, vec));
 }
 
diff --git a/src/gallium/auxiliary/gallivm/storagesoa.cpp b/src/gallium/auxiliary/gallivm/storagesoa.cpp
index 0e6e68c9d70..78d754371f0 100644
--- a/src/gallium/auxiliary/gallivm/storagesoa.cpp
+++ b/src/gallium/auxiliary/gallivm/storagesoa.cpp
@@ -264,10 +264,10 @@ llvm::Constant * StorageSoa::createConstGlobalVector(const std::vector<float> &v
 {
    VectorType *vectorType = VectorType::get(Type::FloatTy, 4);
    std::vector<Constant*> immValues;
-   ConstantFP *constx = ConstantFP::get(Type::FloatTy, APFloat(vec[0]));
-   ConstantFP *consty = ConstantFP::get(Type::FloatTy, APFloat(vec[1]));
-   ConstantFP *constz = ConstantFP::get(Type::FloatTy, APFloat(vec[2]));
-   ConstantFP *constw = ConstantFP::get(Type::FloatTy, APFloat(vec[3]));
+   ConstantFP *constx = ConstantFP::get(APFloat(vec[0]));
+   ConstantFP *consty = ConstantFP::get(APFloat(vec[1]));
+   ConstantFP *constz = ConstantFP::get(APFloat(vec[2]));
+   ConstantFP *constw = ConstantFP::get(APFloat(vec[3]));
    immValues.push_back(constx);
    immValues.push_back(consty);
    immValues.push_back(constz);
diff --git a/src/gallium/auxiliary/translate/translate.h b/src/gallium/auxiliary/translate/translate.h
index 6c15d7e4dc7..b8210af50cf 100644
--- a/src/gallium/auxiliary/translate/translate.h
+++ b/src/gallium/auxiliary/translate/translate.h
@@ -47,10 +47,9 @@
 struct translate_element 
 {
    enum pipe_format input_format;
-   unsigned input_buffer;
-   unsigned input_offset;
-
    enum pipe_format output_format;
+   unsigned input_buffer;
+   unsigned input_offset;       /* can't really reduce the size of these */
    unsigned output_offset;
 };
 
@@ -96,6 +95,27 @@ struct translate *translate_lookup_or_create( struct translate_context *tctx,
 
 struct translate *translate_create( const struct translate_key *key );
 
+static INLINE int translate_keysize( const struct translate_key *key )
+{
+   return 2 * sizeof(int) + key->nr_elements * sizeof(struct translate_element);
+}
+
+static INLINE int translate_key_compare( const struct translate_key *a,
+                                         const struct translate_key *b )
+{
+   int keysize = translate_keysize(a);
+   return memcmp(a, b, keysize);
+}
+
+
+static INLINE void translate_key_sanitize( struct translate_key *a )
+{
+   int keysize = translate_keysize(a);
+   char *ptr = (char *)a;
+   memset(ptr + keysize, 0, sizeof(*a) - keysize);
+}
+
+
 /*******************************************************************************
  *  Private:
  */
diff --git a/src/gallium/drivers/softpipe/sp_quad_fs.c b/src/gallium/drivers/softpipe/sp_quad_fs.c
index 625d0f9b489..8c88c192f8f 100644
--- a/src/gallium/drivers/softpipe/sp_quad_fs.c
+++ b/src/gallium/drivers/softpipe/sp_quad_fs.c
@@ -53,7 +53,6 @@ struct quad_shade_stage
    struct tgsi_sampler samplers[PIPE_MAX_SAMPLERS];
    struct tgsi_exec_machine machine;
    struct tgsi_exec_vector *inputs, *outputs;
-   int colorOutSlot, depthOutSlot;
 };
 
 
@@ -156,20 +155,6 @@ static void shade_begin(struct quad_stage *qs)
       qss->samplers[i].texture = softpipe->texture[i];
    }
 
-   /* find output slots for depth, color */
-   qss->colorOutSlot = -1;
-   qss->depthOutSlot = -1;
-   for (i = 0; i < qss->stage.softpipe->fs->info.num_outputs; i++) {
-      switch (qss->stage.softpipe->fs->info.output_semantic_name[i]) {
-      case TGSI_SEMANTIC_POSITION:
-         qss->depthOutSlot = i;
-         break;
-      case TGSI_SEMANTIC_COLOR:
-         qss->colorOutSlot = i;
-         break;
-      }
-   }
-   
    softpipe->fs->prepare( softpipe->fs, 
 			  &qss->machine,
 			  qss->samplers );
diff --git a/src/gallium/drivers/softpipe/sp_setup.c b/src/gallium/drivers/softpipe/sp_setup.c
index df7be01fcd5..543d86a5cb9 100644
--- a/src/gallium/drivers/softpipe/sp_setup.c
+++ b/src/gallium/drivers/softpipe/sp_setup.c
@@ -209,77 +209,76 @@ static INLINE int block( int x )
 
 
 /**
- * Compute mask which indicates which pixels in the 2x2 quad are actually inside
- * the triangle's bounds.
- *
- * this is pretty nasty...  may need to rework flush_spans again to
- * fix it, if possible.
- */
-static unsigned calculate_mask( struct setup_context *setup, int x )
-{
-   unsigned mask = 0x0;
-
-   if (x >= setup->span.left[0] && x < setup->span.right[0])
-      mask |= MASK_TOP_LEFT;
-
-   if (x >= setup->span.left[1] && x < setup->span.right[1])
-      mask |= MASK_BOTTOM_LEFT;
-
-   if (x+1 >= setup->span.left[0] && x+1 < setup->span.right[0])
-      mask |= MASK_TOP_RIGHT;
-
-   if (x+1 >= setup->span.left[1] && x+1 < setup->span.right[1])
-      mask |= MASK_BOTTOM_RIGHT;
-
-   return mask;
-}
-
-
-/**
  * Render a horizontal span of quads
  */
 static void flush_spans( struct setup_context *setup )
 {
+   const int xleft0 = setup->span.left[0];
+   const int xleft1 = setup->span.left[1];
+   const int xright0 = setup->span.right[0];
+   const int xright1 = setup->span.right[1];
    int minleft, maxright;
    int x;
 
    switch (setup->span.y_flags) {
    case 0x3:
       /* both odd and even lines written (both quad rows) */
-      minleft = MIN2(setup->span.left[0], setup->span.left[1]);
-      maxright = MAX2(setup->span.right[0], setup->span.right[1]);
+      minleft = block(MIN2(xleft0, xleft1));
+      maxright = block(MAX2(xright0, xright1));
+      for (x = minleft; x <= maxright; x += 2) {
+         /* determine which of the four pixels is inside the span bounds */
+         uint mask = 0x0;
+         if (x >= xleft0 && x < xright0)
+            mask |= MASK_TOP_LEFT;
+         if (x >= xleft1 && x < xright1)
+            mask |= MASK_BOTTOM_LEFT;
+         if (x+1 >= xleft0 && x+1 < xright0)
+            mask |= MASK_TOP_RIGHT;
+         if (x+1 >= xleft1 && x+1 < xright1)
+            mask |= MASK_BOTTOM_RIGHT;
+         emit_quad( setup, x, setup->span.y, mask );
+      }
       break;
 
    case 0x1:
       /* only even line written (quad top row) */
-      minleft = setup->span.left[0];
-      maxright = setup->span.right[0];
+      minleft = block(xleft0);
+      maxright = block(xright0);
+      for (x = minleft; x <= maxright; x += 2) {
+         uint mask = 0x0;
+         if (x >= xleft0 && x < xright0)
+            mask |= MASK_TOP_LEFT;
+         if (x+1 >= xleft0 && x+1 < xright0)
+            mask |= MASK_TOP_RIGHT;
+         emit_quad( setup, x, setup->span.y, mask );
+      }
       break;
 
    case 0x2:
       /* only odd line written (quad bottom row) */
-      minleft = setup->span.left[1];
-      maxright = setup->span.right[1];
+      minleft = block(xleft1);
+      maxright = block(xright1);
+      for (x = minleft; x <= maxright; x += 2) {
+         uint mask = 0x0;
+         if (x >= xleft1 && x < xright1)
+            mask |= MASK_BOTTOM_LEFT;
+         if (x+1 >= xleft1 && x+1 < xright1)
+            mask |= MASK_BOTTOM_RIGHT;
+         emit_quad( setup, x, setup->span.y, mask );
+      }
       break;
 
    default:
       return;
    }
 
-   /* XXX this loop could be moved into the above switch cases and
-    * calculate_mask() could be simplified a bit...
-    */
-   for (x = block(minleft); x <= block(maxright); x += 2) {
-      emit_quad( setup, x, setup->span.y,
-                 calculate_mask( setup, x ) );
-   }
-
    setup->span.y = 0;
    setup->span.y_flags = 0;
    setup->span.right[0] = 0;
    setup->span.right[1] = 0;
 }
 
+
 #if DEBUG_VERTS
 static void print_vertex(const struct setup_context *setup,
                          const float (*v)[4])
@@ -717,7 +716,7 @@ void setup_tri( struct setup_context *setup,
                 const float (*v1)[4],
                 const float (*v2)[4] )
 {
-   float det = calc_det(v0, v1, v2);
+   float det;
 
 #if DEBUG_VERTS
    debug_printf("Setup triangle:\n");
@@ -728,7 +727,8 @@ void setup_tri( struct setup_context *setup,
 
    if (setup->softpipe->no_rast)
       return;
-
+   
+   det = calc_det(v0, v1, v2);
    /*
    debug_printf("%s\n", __FUNCTION__ );
    */
diff --git a/src/gallium/winsys/xlib/fakeglx.c b/src/gallium/winsys/xlib/fakeglx.c
index ec77e81fedb..2c0075e9345 100644
--- a/src/gallium/winsys/xlib/fakeglx.c
+++ b/src/gallium/winsys/xlib/fakeglx.c
@@ -1689,6 +1689,15 @@ static void
 Fake_glXSwapBuffers( Display *dpy, GLXDrawable drawable )
 {
    XMesaBuffer buffer = XMesaFindBuffer( dpy, drawable );
+   static boolean firsttime = 1, no_rast = 0;
+
+   if (firsttime) {
+      no_rast = getenv("SP_NO_RAST") != NULL;
+      firsttime = 0;
+   }
+
+   if (no_rast)
+      return;
 
    if (buffer) {
       XMesaSwapBuffers(buffer);
diff --git a/src/gallium/winsys/xlib/xm_api.c b/src/gallium/winsys/xlib/xm_api.c
index 0c248344b15..26b722f3439 100644
--- a/src/gallium/winsys/xlib/xm_api.c
+++ b/src/gallium/winsys/xlib/xm_api.c
@@ -110,6 +110,9 @@ int xmesa_check_for_xshm( XMesaDisplay *display )
    int major, minor, ignore;
    Bool pixmaps;
 
+   if (getenv("SP_NO_RAST")) 
+      return 0;
+
    if (getenv("MESA_NOSHM")) {
       return 0;
    }
diff --git a/src/mesa/main/context.c b/src/mesa/main/context.c
index 7b8d9341701..893c79f28ca 100755
--- a/src/mesa/main/context.c
+++ b/src/mesa/main/context.c
@@ -150,8 +150,6 @@ int MESA_DEBUG_FLAGS = 0;
 /* ubyte -> float conversion */
 GLfloat _mesa_ubyte_to_float_color_tab[256];
 
-static void
-free_shared_state( GLcontext *ctx, struct gl_shared_state *ss );
 
 
 /**
@@ -423,12 +421,14 @@ alloc_shared_state( GLcontext *ctx )
 #endif
 
 #if FEATURE_ARB_vertex_program
-   ss->DefaultVertexProgram = ctx->Driver.NewProgram(ctx, GL_VERTEX_PROGRAM_ARB, 0);
+   ss->DefaultVertexProgram = (struct gl_vertex_program *)
+      ctx->Driver.NewProgram(ctx, GL_VERTEX_PROGRAM_ARB, 0);
    if (!ss->DefaultVertexProgram)
       goto cleanup;
 #endif
 #if FEATURE_ARB_fragment_program
-   ss->DefaultFragmentProgram = ctx->Driver.NewProgram(ctx, GL_FRAGMENT_PROGRAM_ARB, 0);
+   ss->DefaultFragmentProgram = (struct gl_fragment_program *)
+      ctx->Driver.NewProgram(ctx, GL_FRAGMENT_PROGRAM_ARB, 0);
    if (!ss->DefaultFragmentProgram)
       goto cleanup;
 #endif
@@ -513,12 +513,10 @@ alloc_shared_state( GLcontext *ctx )
       _mesa_DeleteHashTable(ss->Programs);
 #endif
 #if FEATURE_ARB_vertex_program
-   if (ss->DefaultVertexProgram)
-      ctx->Driver.DeleteProgram(ctx, ss->DefaultVertexProgram);
+   _mesa_reference_vertprog(ctx, &ss->DefaultVertexProgram, NULL);
 #endif
 #if FEATURE_ARB_fragment_program
-   if (ss->DefaultFragmentProgram)
-      ctx->Driver.DeleteProgram(ctx, ss->DefaultFragmentProgram);
+   _mesa_reference_fragprog(ctx, &ss->DefaultFragmentProgram, NULL);
 #endif
 #if FEATURE_ATI_fragment_shader
    if (ss->DefaultFragmentShader)
@@ -633,6 +631,21 @@ delete_arrayobj_cb(GLuint id, void *data, void *userData)
 }
 
 /**
+ * Callback for freeing shader program data. Call it before delete_shader_cb
+ * to avoid memory access error.
+ */
+static void
+free_shader_program_data_cb(GLuint id, void *data, void *userData)
+{
+   GLcontext *ctx = (GLcontext *) userData;
+   struct gl_shader_program *shProg = (struct gl_shader_program *) data;
+
+   if (shProg->Type == GL_SHADER_PROGRAM_MESA) {
+       _mesa_free_shader_program_data(ctx, shProg);
+   }
+}
+
+/**
  * Callback for deleting shader and shader programs objects.
  * Called by _mesa_HashDeleteAll().
  */
@@ -695,10 +708,10 @@ free_shared_state( GLcontext *ctx, struct gl_shared_state *ss )
    _mesa_DeleteHashTable(ss->Programs);
 #endif
 #if FEATURE_ARB_vertex_program
-   ctx->Driver.DeleteProgram(ctx, ss->DefaultVertexProgram);
+   _mesa_reference_vertprog(ctx, &ss->DefaultVertexProgram, NULL);
 #endif
 #if FEATURE_ARB_fragment_program
-   ctx->Driver.DeleteProgram(ctx, ss->DefaultFragmentProgram);
+   _mesa_reference_fragprog(ctx, &ss->DefaultFragmentProgram, NULL);
 #endif
 
 #if FEATURE_ATI_fragment_shader
@@ -716,6 +729,7 @@ free_shared_state( GLcontext *ctx, struct gl_shared_state *ss )
    _mesa_DeleteHashTable(ss->ArrayObjects);
 
 #if FEATURE_ARB_shader_objects
+   _mesa_HashWalk(ss->ShaderObjects, free_shader_program_data_cb, ctx);
    _mesa_HashDeleteAll(ss->ShaderObjects, delete_shader_cb, ctx);
    _mesa_DeleteHashTable(ss->ShaderObjects);
 #endif
@@ -1190,6 +1204,14 @@ _mesa_free_context_data( GLcontext *ctx )
       _mesa_unreference_framebuffer(&ctx->ReadBuffer);
    }
 
+   _mesa_reference_vertprog(ctx, &ctx->VertexProgram.Current, NULL);
+   _mesa_reference_vertprog(ctx, &ctx->VertexProgram._Current, NULL);
+   _mesa_reference_vertprog(ctx, &ctx->VertexProgram._TnlProgram, NULL);
+
+   _mesa_reference_fragprog(ctx, &ctx->FragmentProgram.Current, NULL);
+   _mesa_reference_fragprog(ctx, &ctx->FragmentProgram._Current, NULL);
+   _mesa_reference_fragprog(ctx, &ctx->FragmentProgram._TexEnvProgram, NULL);
+
    _mesa_free_lighting_data( ctx );
    _mesa_free_eval_data( ctx );
    _mesa_free_texture_data( ctx );
diff --git a/src/mesa/main/ffvertex_prog.c b/src/mesa/main/ffvertex_prog.c
index f648d081c0e..810af9e33e6 100644
--- a/src/mesa/main/ffvertex_prog.c
+++ b/src/mesa/main/ffvertex_prog.c
@@ -946,35 +946,33 @@ static void build_lighting( struct tnl_program *p )
 	 _bfc1 = _bfc0;
    }
 
-
    /* If no lights, still need to emit the scenecolor.
     */
-      {
-	 struct ureg res0 = register_output( p, VERT_RESULT_COL0 );
-	 emit_op1(p, OPCODE_MOV, res0, 0, _col0);
-      }
+   {
+      struct ureg res0 = register_output( p, VERT_RESULT_COL0 );
+      emit_op1(p, OPCODE_MOV, res0, 0, _col0);
+   }
 
-      if (separate) {
-	 struct ureg res1 = register_output( p, VERT_RESULT_COL1 );
-	 emit_op1(p, OPCODE_MOV, res1, 0, _col1);
-      }
+   if (separate) {
+      struct ureg res1 = register_output( p, VERT_RESULT_COL1 );
+      emit_op1(p, OPCODE_MOV, res1, 0, _col1);
+   }
 
-      if (twoside) {
-	 struct ureg res0 = register_output( p, VERT_RESULT_BFC0 );
-	 emit_op1(p, OPCODE_MOV, res0, 0, _bfc0);
-      }
+   if (twoside) {
+      struct ureg res0 = register_output( p, VERT_RESULT_BFC0 );
+      emit_op1(p, OPCODE_MOV, res0, 0, _bfc0);
+   }
       
-      if (twoside && separate) {
-	 struct ureg res1 = register_output( p, VERT_RESULT_BFC1 );
-	 emit_op1(p, OPCODE_MOV, res1, 0, _bfc1);
-      }
+   if (twoside && separate) {
+      struct ureg res1 = register_output( p, VERT_RESULT_BFC1 );
+      emit_op1(p, OPCODE_MOV, res1, 0, _bfc1);
+   }
       
    if (nr_lights == 0) {
       release_temps(p);
       return;
    }
 
-
    for (i = 0; i < MAX_LIGHTS; i++) {
       if (p->state->unit[i].light_enabled) {
 	 struct ureg half = undef;
@@ -1006,7 +1004,7 @@ static void build_lighting( struct tnl_program *p )
 	    VPpli = get_temp(p); 
 	    half = get_temp(p);
  
-	    /* Calulate VPpli vector
+	    /* Calculate VPpli vector
 	     */
 	    emit_op2(p, OPCODE_SUB, VPpli, 0, Ppli, V); 
 
@@ -1017,15 +1015,13 @@ static void build_lighting( struct tnl_program *p )
 	    emit_op1(p, OPCODE_RSQ, dist, 0, dist);
 	    emit_op2(p, OPCODE_MUL, VPpli, 0, VPpli, dist);
 
-
-	    /* Calculate  attenuation:
+	    /* Calculate attenuation:
 	     */ 
 	    if (!p->state->unit[i].light_spotcutoff_is_180 ||
 		p->state->unit[i].light_attenuated) {
 	       att = calculate_light_attenuation(p, i, VPpli, dist);
 	    }
-	 
-      
+
 	    /* Calculate viewer direction, or use infinite viewer:
 	     */
 	    if (p->state->light_local_viewer) {
@@ -1047,7 +1043,6 @@ static void build_lighting( struct tnl_program *p )
 	 emit_op2(p, OPCODE_DP3, dots, WRITEMASK_X, normal, VPpli);
 	 emit_op2(p, OPCODE_DP3, dots, WRITEMASK_Y, normal, half);
 
-	
 	 /* Front face lighting:
 	  */
 	 {
diff --git a/src/mesa/main/imports.c b/src/mesa/main/imports.c
index d8d35af15e2..d798f80e253 100644
--- a/src/mesa/main/imports.c
+++ b/src/mesa/main/imports.c
@@ -542,26 +542,24 @@ int
 _mesa_ffs(int i)
 {
 #if (defined(_WIN32) && !defined(__MINGW32__) ) || defined(__IBMC__) || defined(__IBMCPP__)
-   register int bit = 0;
-   if (i != 0) {
-      if ((i & 0xffff) == 0) {
-         bit += 16;
-         i >>= 16;
-      }
-      if ((i & 0xff) == 0) {
-         bit += 8;
-         i >>= 8;
-      }
-      if ((i & 0xf) == 0) {
-         bit += 4;
-         i >>= 4;
-      }
-      while ((i & 1) == 0) {
-         bit++;
-         i >>= 1;
-      }
+   register int bit = 1;
+   if ((i & 0xffff) == 0) {
+      bit += 16;
+      i >>= 16;
+   }
+   if ((i & 0xff) == 0) {
+      bit += 8;
+      i >>= 8;
+   }
+   if ((i & 0xf) == 0) {
+      bit += 4;
+      i >>= 4;
+   }
+   if ((i & 0x3) == 0) {
+      bit += 2;
+      i >>= 2;
    }
-   return bit;
+   return (i) ? (bit + ((i + 1) & 0x01)) : 0;
 #else
    return ffs(i);
 #endif
diff --git a/src/mesa/main/mtypes.h b/src/mesa/main/mtypes.h
index 50b22d25bf3..463142fe395 100644
--- a/src/mesa/main/mtypes.h
+++ b/src/mesa/main/mtypes.h
@@ -2203,10 +2203,10 @@ struct gl_shared_state
    /*@{*/
    struct _mesa_HashTable *Programs; /**< All vertex/fragment programs */
 #if FEATURE_ARB_vertex_program
-   struct gl_program *DefaultVertexProgram;
+   struct gl_vertex_program *DefaultVertexProgram;
 #endif
 #if FEATURE_ARB_fragment_program
-   struct gl_program *DefaultFragmentProgram;
+   struct gl_fragment_program *DefaultFragmentProgram;
 #endif
    /*@}*/
 
diff --git a/src/mesa/main/state.c b/src/mesa/main/state.c
index 0b1c56fdd5f..90379a17726 100644
--- a/src/mesa/main/state.c
+++ b/src/mesa/main/state.c
@@ -982,16 +982,20 @@ update_program(GLcontext *ctx)
 #endif
    if (shProg && shProg->LinkStatus && shProg->FragmentProgram) {
       /* user-defined fragment shader */
-      ctx->FragmentProgram._Current = shProg->FragmentProgram;
+      _mesa_reference_fragprog(ctx, &ctx->FragmentProgram._Current,
+                               shProg->FragmentProgram);
    }
    else if (ctx->FragmentProgram._Enabled) {
       /* use user-defined fragment program */
-      ctx->FragmentProgram._Current = ctx->FragmentProgram.Current;
+      _mesa_reference_fragprog(ctx, &ctx->FragmentProgram._Current,
+                               ctx->FragmentProgram.Current);
    }
    else if (ctx->FragmentProgram._MaintainTexEnvProgram) {
       /* fragment program generated from fixed-function state */
-      ctx->FragmentProgram._Current = _mesa_get_fixed_func_fragment_program(ctx);
-      ctx->FragmentProgram._TexEnvProgram = ctx->FragmentProgram._Current;
+      _mesa_reference_fragprog(ctx, &ctx->FragmentProgram._Current,
+                               _mesa_get_fixed_func_fragment_program(ctx));
+      _mesa_reference_fragprog(ctx, &ctx->FragmentProgram._TexEnvProgram,
+                               ctx->FragmentProgram._Current);
 
       /* XXX get rid of this confusing stuff someday? */
       ctx->FragmentProgram._Active = ctx->FragmentProgram._Enabled;
@@ -1000,7 +1004,7 @@ update_program(GLcontext *ctx)
    }
    else {
       /* no fragment program */
-      ctx->FragmentProgram._Current = NULL;
+      _mesa_reference_fragprog(ctx, &ctx->FragmentProgram._Current, NULL);
    }
 
    if (ctx->FragmentProgram._Current != prevFP && ctx->Driver.BindProgram) {
@@ -1013,29 +1017,33 @@ update_program(GLcontext *ctx)
     **/
 #if 1
    /* XXX get rid of this someday? */
-   ctx->VertexProgram._TnlProgram = NULL;
+   _mesa_reference_vertprog(ctx, &ctx->VertexProgram._TnlProgram, NULL);
 #endif
    if (shProg && shProg->LinkStatus && shProg->VertexProgram) {
       /* user-defined vertex shader */
-      ctx->VertexProgram._Current = shProg->VertexProgram;
+      _mesa_reference_vertprog(ctx, &ctx->VertexProgram._Current,
+                               shProg->VertexProgram);
    }
    else if (ctx->VertexProgram._Enabled) {
       /* use user-defined vertex program */
-      ctx->VertexProgram._Current = ctx->VertexProgram.Current;
+      _mesa_reference_vertprog(ctx, &ctx->VertexProgram._Current,
+                               ctx->VertexProgram.Current);
    }
    else if (ctx->VertexProgram._MaintainTnlProgram) {
       /* vertex program generated from fixed-function state */
-      ctx->VertexProgram._Current = _mesa_get_fixed_func_vertex_program(ctx);
-      ctx->VertexProgram._TnlProgram = ctx->VertexProgram._Current;
+      _mesa_reference_vertprog(ctx, &ctx->VertexProgram._Current,
+                               _mesa_get_fixed_func_vertex_program(ctx));
+      _mesa_reference_vertprog(ctx, &ctx->VertexProgram._TnlProgram,
+                               ctx->VertexProgram._Current);
    }
    else {
       /* no vertex program / used fixed-function code */
-      ctx->VertexProgram._Current = NULL;
+      _mesa_reference_vertprog(ctx, &ctx->VertexProgram._Current, NULL);
    }
 
    if (ctx->VertexProgram._Current != prevVP && ctx->Driver.BindProgram) {
       ctx->Driver.BindProgram(ctx, GL_VERTEX_PROGRAM_ARB,
-                              (struct gl_program *) ctx->VertexProgram._Current);
+                           (struct gl_program *) ctx->VertexProgram._Current);
    }
 }
 
diff --git a/src/mesa/shader/prog_cache.c b/src/mesa/shader/prog_cache.c
index dd0241ef24e..36a25377c55 100644
--- a/src/mesa/shader/prog_cache.c
+++ b/src/mesa/shader/prog_cache.c
@@ -30,6 +30,7 @@
 #include "main/mtypes.h"
 #include "main/imports.h"
 #include "shader/prog_cache.h"
+#include "shader/program.h"
 
 
 struct cache_item
@@ -109,7 +110,7 @@ clear_cache(GLcontext *ctx, struct gl_program_cache *cache)
       for (c = cache->items[i]; c; c = next) {
 	 next = c->next;
 	 _mesa_free(c->key);
-	 ctx->Driver.DeleteProgram(ctx, c->program);
+         _mesa_reference_program(ctx, &c->program, NULL);
 	 _mesa_free(c);
       }
       cache->items[i] = NULL;
@@ -177,7 +178,7 @@ _mesa_program_cache_insert(GLcontext *ctx,
    c->key = _mesa_malloc(keysize);
    memcpy(c->key, key, keysize);
 
-   c->program = program;
+   c->program = program;  /* no refcount change */
 
    if (cache->n_items > cache->size * 1.5) {
       if (cache->size < 1000)
diff --git a/src/mesa/shader/program.c b/src/mesa/shader/program.c
index 0ed7f833d20..9a23c5d7d31 100644
--- a/src/mesa/shader/program.c
+++ b/src/mesa/shader/program.c
@@ -60,9 +60,9 @@ _mesa_init_program(GLcontext *ctx)
    ctx->VertexProgram.Enabled = GL_FALSE;
    ctx->VertexProgram.PointSizeEnabled = GL_FALSE;
    ctx->VertexProgram.TwoSideEnabled = GL_FALSE;
-   ctx->VertexProgram.Current = (struct gl_vertex_program *) ctx->Shared->DefaultVertexProgram;
+   _mesa_reference_vertprog(ctx, &ctx->VertexProgram.Current,
+                            ctx->Shared->DefaultVertexProgram);
    assert(ctx->VertexProgram.Current);
-   ctx->VertexProgram.Current->Base.RefCount++;
    for (i = 0; i < MAX_NV_VERTEX_PROGRAM_PARAMS / 4; i++) {
       ctx->VertexProgram.TrackMatrix[i] = GL_NONE;
       ctx->VertexProgram.TrackMatrixTransform[i] = GL_IDENTITY_NV;
@@ -72,9 +72,9 @@ _mesa_init_program(GLcontext *ctx)
 
 #if FEATURE_NV_fragment_program || FEATURE_ARB_fragment_program
    ctx->FragmentProgram.Enabled = GL_FALSE;
-   ctx->FragmentProgram.Current = (struct gl_fragment_program *) ctx->Shared->DefaultFragmentProgram;
+   _mesa_reference_fragprog(ctx, &ctx->FragmentProgram.Current,
+                            ctx->Shared->DefaultFragmentProgram);
    assert(ctx->FragmentProgram.Current);
-   ctx->FragmentProgram.Current->Base.RefCount++;
    ctx->FragmentProgram.Cache = _mesa_new_program_cache();
 #endif
 
@@ -96,19 +96,11 @@ void
 _mesa_free_program_data(GLcontext *ctx)
 {
 #if FEATURE_NV_vertex_program || FEATURE_ARB_vertex_program
-   if (ctx->VertexProgram.Current) {
-      ctx->VertexProgram.Current->Base.RefCount--;
-      if (ctx->VertexProgram.Current->Base.RefCount <= 0)
-         ctx->Driver.DeleteProgram(ctx, &(ctx->VertexProgram.Current->Base));
-   }
+   _mesa_reference_vertprog(ctx, &ctx->VertexProgram.Current, NULL);
    _mesa_delete_program_cache(ctx, ctx->VertexProgram.Cache);
 #endif
 #if FEATURE_NV_fragment_program || FEATURE_ARB_fragment_program
-   if (ctx->FragmentProgram.Current) {
-      ctx->FragmentProgram.Current->Base.RefCount--;
-      if (ctx->FragmentProgram.Current->Base.RefCount <= 0)
-         ctx->Driver.DeleteProgram(ctx, &(ctx->FragmentProgram.Current->Base));
-   }
+   _mesa_reference_fragprog(ctx, &ctx->FragmentProgram.Current, NULL);
    _mesa_delete_program_cache(ctx, ctx->FragmentProgram.Cache);
 #endif
    /* XXX probably move this stuff */
@@ -326,6 +318,59 @@ _mesa_lookup_program(GLcontext *ctx, GLuint id)
 
 
 /**
+ * Reference counting for vertex/fragment programs
+ */
+void
+_mesa_reference_program(GLcontext *ctx,
+                        struct gl_program **ptr,
+                        struct gl_program *prog)
+{
+   assert(ptr);
+   if (*ptr && prog) {
+      /* sanity check */
+      ASSERT((*ptr)->Target == prog->Target);
+   }
+   if (*ptr == prog) {
+      return;  /* no change */
+   }
+   if (*ptr) {
+      GLboolean deleteFlag;
+
+      /*_glthread_LOCK_MUTEX((*ptr)->Mutex);*/
+#if 0
+      printf("Program %p %u 0x%x  Refcount-- to %d\n",
+             *ptr, (*ptr)->Id, (*ptr)->Target, (*ptr)->RefCount - 1);
+#endif
+      ASSERT((*ptr)->RefCount > 0);
+      (*ptr)->RefCount--;
+
+      deleteFlag = ((*ptr)->RefCount == 0);
+      /*_glthread_UNLOCK_MUTEX((*ptr)->Mutex);*/
+      
+      if (deleteFlag) {
+         ASSERT(ctx);
+         ctx->Driver.DeleteProgram(ctx, *ptr);
+      }
+
+      *ptr = NULL;
+   }
+
+   assert(!*ptr);
+   if (prog) {
+      /*_glthread_LOCK_MUTEX(prog->Mutex);*/
+      prog->RefCount++;
+#if 0
+      printf("Program %p %u 0x%x  Refcount++ to %d\n",
+             prog, prog->Id, prog->Target, prog->RefCount);
+#endif
+      /*_glthread_UNLOCK_MUTEX(prog->Mutex);*/
+   }
+
+   *ptr = prog;
+}
+
+
+/**
  * Return a copy of a program.
  * XXX Problem here if the program object is actually OO-derivation
  * made by a device driver.
@@ -340,8 +385,9 @@ _mesa_clone_program(GLcontext *ctx, const struct gl_program *prog)
       return NULL;
 
    assert(clone->Target == prog->Target);
+   assert(clone->RefCount == 1);
+
    clone->String = (GLubyte *) _mesa_strdup((char *) prog->String);
-   clone->RefCount = 1;
    clone->Format = prog->Format;
    clone->Instructions = _mesa_alloc_instructions(prog->NumInstructions);
    if (!clone->Instructions) {
@@ -704,9 +750,9 @@ _mesa_BindProgram(GLenum target, GLuint id)
       /* Bind a default program */
       newProg = NULL;
       if (target == GL_VERTEX_PROGRAM_ARB) /* == GL_VERTEX_PROGRAM_NV */
-         newProg = ctx->Shared->DefaultVertexProgram;
+         newProg = &ctx->Shared->DefaultVertexProgram->Base;
       else
-         newProg = ctx->Shared->DefaultFragmentProgram;
+         newProg = &ctx->Shared->DefaultFragmentProgram->Base;
    }
    else {
       /* Bind a user program */
@@ -734,26 +780,16 @@ _mesa_BindProgram(GLenum target, GLuint id)
       return;
    }
 
-   /* unbind/delete oldProg */
-   if (curProg->Id != 0) {
-      /* decrement refcount on previously bound fragment program */
-      curProg->RefCount--;
-      /* and delete if refcount goes below one */
-      if (curProg->RefCount <= 0) {
-         /* the program ID was already removed from the hash table */
-         ctx->Driver.DeleteProgram(ctx, curProg);
-      }
-   }
-
    /* bind newProg */
    if (target == GL_VERTEX_PROGRAM_ARB) { /* == GL_VERTEX_PROGRAM_NV */
-      ctx->VertexProgram.Current = (struct gl_vertex_program *) newProg;
+      _mesa_reference_vertprog(ctx, &ctx->VertexProgram.Current,
+                               (struct gl_vertex_program *) newProg);
    }
    else if (target == GL_FRAGMENT_PROGRAM_NV ||
             target == GL_FRAGMENT_PROGRAM_ARB) {
-      ctx->FragmentProgram.Current = (struct gl_fragment_program *) newProg;
+      _mesa_reference_fragprog(ctx, &ctx->FragmentProgram.Current,
+                               (struct gl_fragment_program *) newProg);
    }
-   newProg->RefCount++;
 
    /* Never null pointers */
    ASSERT(ctx->VertexProgram.Current);
@@ -811,10 +847,7 @@ _mesa_DeletePrograms(GLsizei n, const GLuint *ids)
             }
             /* The ID is immediately available for re-use now */
             _mesa_HashRemove(ctx->Shared->Programs, ids[i]);
-            prog->RefCount--;
-            if (prog->RefCount <= 0) {
-               ctx->Driver.DeleteProgram(ctx, prog);
-            }
+            _mesa_reference_program(ctx, &prog, NULL);
          }
       }
    }
diff --git a/src/mesa/shader/program.h b/src/mesa/shader/program.h
index 414a57d39cc..08fe576afc1 100644
--- a/src/mesa/shader/program.h
+++ b/src/mesa/shader/program.h
@@ -83,6 +83,28 @@ _mesa_delete_program(GLcontext *ctx, struct gl_program *prog);
 extern struct gl_program *
 _mesa_lookup_program(GLcontext *ctx, GLuint id);
 
+extern void
+_mesa_reference_program(GLcontext *ctx,
+                        struct gl_program **ptr,
+                        struct gl_program *prog);
+
+static INLINE void
+_mesa_reference_vertprog(GLcontext *ctx,
+                         struct gl_vertex_program **ptr,
+                         struct gl_vertex_program *prog)
+{
+   _mesa_reference_program(ctx, (struct gl_program **) ptr,
+                           (struct gl_program *) prog);
+}
+
+static INLINE void
+_mesa_reference_fragprog(GLcontext *ctx,
+                         struct gl_fragment_program **ptr,
+                         struct gl_fragment_program *prog)
+{
+   _mesa_reference_program(ctx, (struct gl_program **) ptr,
+                           (struct gl_program *) prog);
+}
 
 extern struct gl_program *
 _mesa_clone_program(GLcontext *ctx, const struct gl_program *prog);
diff --git a/src/mesa/shader/shader_api.c b/src/mesa/shader/shader_api.c
index 9c419c9903f..f12fa28d97b 100644
--- a/src/mesa/shader/shader_api.c
+++ b/src/mesa/shader/shader_api.c
@@ -80,8 +80,7 @@ _mesa_clear_shader_program_data(GLcontext *ctx,
        * original/unlinked program.
        */
       shProg->VertexProgram->Base.Parameters = NULL;
-      ctx->Driver.DeleteProgram(ctx, &shProg->VertexProgram->Base);
-      shProg->VertexProgram = NULL;
+      _mesa_reference_vertprog(ctx, &shProg->VertexProgram, NULL);
    }
 
    if (shProg->FragmentProgram) {
@@ -89,8 +88,7 @@ _mesa_clear_shader_program_data(GLcontext *ctx,
        * original/unlinked program.
        */
       shProg->FragmentProgram->Base.Parameters = NULL;
-      ctx->Driver.DeleteProgram(ctx, &shProg->FragmentProgram->Base);
-      shProg->FragmentProgram = NULL;
+      _mesa_reference_fragprog(ctx, &shProg->FragmentProgram, NULL);
    }
 
    if (shProg->Uniforms) {
diff --git a/src/mesa/shader/slang/slang_link.c b/src/mesa/shader/slang/slang_link.c
index addff204218..ae581553dc7 100644
--- a/src/mesa/shader/slang/slang_link.c
+++ b/src/mesa/shader/slang/slang_link.c
@@ -410,19 +410,19 @@ _slang_link(GLcontext *ctx,
     * changing src/dst registers after merging the uniforms and varying vars.
     */
    if (vertProg) {
-      shProg->VertexProgram
-         = vertex_program(_mesa_clone_program(ctx, &vertProg->Base));
+      _mesa_reference_vertprog(ctx, &shProg->VertexProgram,
+                               vertex_program(_mesa_clone_program(ctx, &vertProg->Base)));
    }
    else {
-      shProg->VertexProgram = NULL;
+      _mesa_reference_vertprog(ctx, &shProg->VertexProgram, NULL);
    }
 
    if (fragProg) {
-      shProg->FragmentProgram
-         = fragment_program(_mesa_clone_program(ctx, &fragProg->Base));
+      _mesa_reference_fragprog(ctx, &shProg->FragmentProgram,
+                               fragment_program(_mesa_clone_program(ctx, &fragProg->Base)));
    }
    else {
-      shProg->FragmentProgram = NULL;
+      _mesa_reference_fragprog(ctx, &shProg->FragmentProgram, NULL);
    }
 
    /* link varying vars */
diff --git a/src/mesa/state_tracker/st_atom.h b/src/mesa/state_tracker/st_atom.h
index 3a63e2dec03..c6c6eba8121 100644
--- a/src/mesa/state_tracker/st_atom.h
+++ b/src/mesa/state_tracker/st_atom.h
@@ -44,20 +44,20 @@ void st_destroy_atoms( struct st_context *st );
 void st_validate_state( struct st_context *st );
 
 
-const struct st_tracked_state st_update_framebuffer;
-const struct st_tracked_state st_update_clip;
-const struct st_tracked_state st_update_depth_stencil_alpha;
-const struct st_tracked_state st_update_shader;
-const struct st_tracked_state st_update_rasterizer;
-const struct st_tracked_state st_update_polygon_stipple;
-const struct st_tracked_state st_update_viewport;
-const struct st_tracked_state st_update_scissor;
-const struct st_tracked_state st_update_blend;
-const struct st_tracked_state st_update_sampler;
-const struct st_tracked_state st_update_texture;
-const struct st_tracked_state st_update_fs_constants;
-const struct st_tracked_state st_update_vs_constants;
-const struct st_tracked_state st_update_pixel_transfer;
+extern const struct st_tracked_state st_update_framebuffer;
+extern const struct st_tracked_state st_update_clip;
+extern const struct st_tracked_state st_update_depth_stencil_alpha;
+extern const struct st_tracked_state st_update_shader;
+extern const struct st_tracked_state st_update_rasterizer;
+extern const struct st_tracked_state st_update_polygon_stipple;
+extern const struct st_tracked_state st_update_viewport;
+extern const struct st_tracked_state st_update_scissor;
+extern const struct st_tracked_state st_update_blend;
+extern const struct st_tracked_state st_update_sampler;
+extern const struct st_tracked_state st_update_texture;
+extern const struct st_tracked_state st_update_fs_constants;
+extern const struct st_tracked_state st_update_vs_constants;
+extern const struct st_tracked_state st_update_pixel_transfer;
 
 
 uint st_compare_func_to_pipe(GLenum func);
diff --git a/src/mesa/state_tracker/st_atom_shader.c b/src/mesa/state_tracker/st_atom_shader.c
index 652500f52a2..7745591afba 100644
--- a/src/mesa/state_tracker/st_atom_shader.c
+++ b/src/mesa/state_tracker/st_atom_shader.c
@@ -39,6 +39,7 @@
 
 #include "main/imports.h"
 #include "main/mtypes.h"
+#include "shader/program.h"
 
 #include "pipe/p_context.h"
 #include "pipe/p_shader_tokens.h"
@@ -264,14 +265,16 @@ update_linkage( struct st_context *st )
     */
    assert(st->ctx->VertexProgram._Current);
    stvp = st_vertex_program(st->ctx->VertexProgram._Current);
+   assert(stvp->Base.Base.Target == GL_VERTEX_PROGRAM_ARB);
 
    assert(st->ctx->FragmentProgram._Current);
    stfp = st_fragment_program(st->ctx->FragmentProgram._Current);
+   assert(stfp->Base.Base.Target == GL_FRAGMENT_PROGRAM_ARB);
 
    xvp = find_translated_vp(st, stvp, stfp);
 
-   st->vp = stvp;
-   st->fp = stfp;
+   st_reference_vertprog(st, &st->vp, stvp);
+   st_reference_fragprog(st, &st->fp, stfp);
 
    cso_set_vertex_shader_handle(st->cso_context, stvp->driver_shader);
    cso_set_fragment_shader_handle(st->cso_context, stfp->driver_shader);
diff --git a/src/mesa/state_tracker/st_cb_texture.c b/src/mesa/state_tracker/st_cb_texture.c
index 02ef961e4d3..866683c23e9 100644
--- a/src/mesa/state_tracker/st_cb_texture.c
+++ b/src/mesa/state_tracker/st_cb_texture.c
@@ -1366,7 +1366,7 @@ calculate_first_last_level(struct st_texture_object *stObj)
       }
       else {
          firstLevel = 0;
-         lastLevel = MIN2(tObj->MaxLevel, tObj->Image[0][0]->WidthLog2);
+         lastLevel = MIN2(tObj->MaxLevel, tObj->Image[0][tObj->BaseLevel]->WidthLog2);
       }
       break;
    case GL_TEXTURE_RECTANGLE_NV:
diff --git a/src/mesa/state_tracker/st_context.c b/src/mesa/state_tracker/st_context.c
index c900064f2b8..8db55a179f8 100644
--- a/src/mesa/state_tracker/st_context.c
+++ b/src/mesa/state_tracker/st_context.c
@@ -158,6 +158,9 @@ static void st_destroy_context_priv( struct st_context *st )
 {
    uint i;
 
+   st_reference_fragprog(st, &st->fp, NULL);
+   st_reference_vertprog(st, &st->vp, NULL);
+
    draw_destroy(st->draw);
    st_destroy_atoms( st );
    st_destroy_draw( st );
diff --git a/src/mesa/state_tracker/st_program.h b/src/mesa/state_tracker/st_program.h
index d8f26da2eeb..bf07a50789e 100644
--- a/src/mesa/state_tracker/st_program.h
+++ b/src/mesa/state_tracker/st_program.h
@@ -34,7 +34,8 @@
 #ifndef ST_PROGRAM_H
 #define ST_PROGRAM_H
 
-#include "mtypes.h"
+#include "main/mtypes.h"
+#include "shader/program.h"
 #include "pipe/p_shader_tokens.h"
 
 
@@ -115,6 +116,27 @@ st_vertex_program( struct gl_vertex_program *vp )
 }
 
 
+static INLINE void
+st_reference_vertprog(struct st_context *st,
+                      struct st_vertex_program **ptr,
+                      struct st_vertex_program *prog)
+{
+   _mesa_reference_program(st->ctx,
+                           (struct gl_program **) ptr,
+                           (struct gl_program *) prog);
+}
+
+static INLINE void
+st_reference_fragprog(struct st_context *st,
+                      struct st_fragment_program **ptr,
+                      struct st_fragment_program *prog)
+{
+   _mesa_reference_program(st->ctx,
+                           (struct gl_program **) ptr,
+                           (struct gl_program *) prog);
+}
+
+
 extern void
 st_translate_fragment_program(struct st_context *st,
                               struct st_fragment_program *fp,