Merge commit 'origin/gallium-0.2' into gallium-master-merge

46 files changed, 876 insertions, 534 deletions

diff --git a/progs/SConscript b/progs/SConscript
index 631923a38f..6484c761fc 100644
--- a/progs/SConscript
+++ b/progs/SConscript
@@ -1,6 +1,8 @@
 SConscript([
     'util/SConscript',
     'demos/SConscript',
+    'redbook/SConscript',
+    'samples/SConscript',
     'trivial/SConscript',
     'vp/SConscript',
 ])
diff --git a/progs/redbook/SConscript b/progs/redbook/SConscript
new file mode 100644
index 0000000000..242cb6647f
--- /dev/null
+++ b/progs/redbook/SConscript
@@ -0,0 +1,91 @@
+Import('*')
+
+if not env['GLUT']:
+    Return()
+
+env = env.Clone()
+
+env.Prepend(CPPPATH = [
+	'../util',
+])
+
+env.Prepend(LIBS = [
+	util,
+	'$GLUT_LIB'
+])
+
+if env['platform'] == 'windows':
+    env.Append(CPPDEFINES = ['NOMINMAX'])
+    env.Prepend(LIBS = ['winmm'])
+
+progs = [
+    'aaindex',
+    'aapoly',
+    'aargb',
+    'accanti',
+    'accpersp',
+    'alpha3D',
+    'alpha',
+    'anti',
+    'bezcurve',
+    'bezmesh',
+    'checker',
+    'clip',
+    'colormat',
+    'cube',
+    'depthcue',
+    'dof',
+    'double',
+    'drawf',
+    'feedback',
+    'fog',
+    'fogindex',
+    'font',
+    'hello',
+    'image',
+    'light',
+    'lines',
+    'list',
+    'material',
+    'mipmap',
+    'model',
+    'movelight',
+    'nurbs',
+    'pickdepth',
+    'picksquare',
+    'plane',
+    'planet',
+    'polyoff',
+    'polys',
+    'quadric',
+    'robot',
+    'sccolorlight',
+    'scenebamb',
+    'scene',
+    'sceneflat',
+    'select',
+    'smooth',
+    'stencil',
+    'stroke',
+    'surface',
+    'teaambient',
+    'teapots',
+    'tess',
+    'tesswind',
+    'texbind',
+    'texgen',
+    'texprox',
+    'texsub',
+    'texturesurf',
+    'torus',
+    'trim',
+    'unproject',
+    'varray',
+    'wrap',
+]
+
+for prog in progs:
+    env.Program(
+        target = prog,
+        source = prog + '.c',
+    )
diff --git a/progs/samples/Makefile b/progs/samples/Makefile
index 63dca07d2e..b300e38b9c 100644
--- a/progs/samples/Makefile
+++ b/progs/samples/Makefile
@@ -7,7 +7,7 @@ INCDIR = $(TOP)/include
 
 LIB_DEP = $(TOP)/$(LIB_DIR)/$(GL_LIB_NAME) $(TOP)/$(LIB_DIR)/$(GLU_LIB_NAME) $(TOP)/$(LIB_DIR)/$(GLUT_LIB_NAME)
 
-LIBS = -L$(TOP)/$(LIB_DIR) -l$(GLUT_LIB) -l$(GLU_LIB) -l$(GL_LIB) $(APP_LIB_DEPS)
+LIBS = -L$(TOP)/$(LIB_DIR) -l$(GLUT_LIB) -l$(GLEW_LIB) -l$(GLU_LIB) -l$(GL_LIB) $(APP_LIB_DEPS)
 
 PROGS = accum bitmap1 bitmap2 blendeq blendxor copy cursor depth eval fog \
 	font line logo nurb olympic overlay point prim quad select \
diff --git a/progs/samples/SConscript b/progs/samples/SConscript
new file mode 100644
index 0000000000..7a8a0d62d8
--- /dev/null
+++ b/progs/samples/SConscript
@@ -0,0 +1,58 @@
+Import('*')
+
+if not env['GLUT']:
+    Return()
+
+env = env.Clone()
+
+env.Prepend(CPPPATH = [
+	'../util',
+])
+
+env.Prepend(LIBS = [
+	util,
+	'$GLUT_LIB'
+])
+
+if env['platform'] == 'windows':
+    env.Append(CPPDEFINES = ['NOMINMAX'])
+    env.Prepend(LIBS = ['winmm'])
+
+progs = [
+    'accum',
+    'bitmap1',
+    'bitmap2',
+    'blendeq',
+    'blendxor',
+    'copy',
+    'cursor',
+    'depth',
+    'eval',
+    'fog',
+    'font',
+    'line',
+    'logo',
+    'nurb',
+    #'oglinfo',
+    'olympic',
+    'overlay',
+    'point',
+    'prim',
+    'quad',
+    'rgbtoppm',
+    'select',
+    'shape',
+    'sphere',
+    'star',
+    'stencil',
+    'stretch',
+    'texture',
+    'tri',
+    'wave',
+]
+
+for prog in progs:
+    env.Program(
+        target = prog,
+        source = prog + '.c',
+    )
diff --git a/progs/samples/blendeq.c b/progs/samples/blendeq.c
index d5143ecdf5..cd0474c5ce 100644
--- a/progs/samples/blendeq.c
+++ b/progs/samples/blendeq.c
@@ -11,10 +11,7 @@
 #include <string.h>
 #include <stdlib.h>
 #include <stdio.h>
-#ifdef _WIN32
-#include <windows.h>
-#endif
-#define GL_GLEXT_PROTOTYPES
+#include <GL/glew.h>
 #include <GL/glut.h>
 
 GLenum doubleBuffer;
@@ -274,6 +271,8 @@ int main(int argc, char **argv)
 	exit(1);
     }
 
+    glewInit();
+
     /* Make sure blend_logic_op extension is there. */
     s = (char *) glGetString(GL_EXTENSIONS);
     version = (char*) glGetString(GL_VERSION);
diff --git a/progs/samples/blendxor.c b/progs/samples/blendxor.c
index 5bc4aa9bea..057143651f 100644
--- a/progs/samples/blendxor.c
+++ b/progs/samples/blendxor.c
@@ -10,13 +10,8 @@
 #include <unistd.h>
 #endif
 #include <stdlib.h>
-#ifdef _WIN32
-#include <windows.h>
-#endif
-#define GL_GLEXT_LEGACY
-#define GL_GLEXT_PROTOTYPES
+#include <GL/glew.h>
 #include <GL/glut.h>
-#include <GL/glext.h>
 
 
 GLenum doubleBuffer;
@@ -176,6 +171,8 @@ int main(int argc, char **argv)
 	exit(1);
     }
 
+    glewInit();
+
     /* Make sure blend_logic_op extension is there. */
     s = (char *) glGetString(GL_EXTENSIONS);
     version = (char*) glGetString(GL_VERSION);
diff --git a/scons/gallium.py b/scons/gallium.py
index 065c53c54b..fc1ed08e47 100644
--- a/scons/gallium.py
+++ b/scons/gallium.py
@@ -163,6 +163,25 @@ def createInstallMethods(env):
     env.AddMethod(install_shared_library, 'InstallSharedLibrary')
 
 
+def num_jobs():
+    try:
+        return int(os.environ['NUMBER_OF_PROCESSORS'])
+    except (ValueError, KeyError):
+        pass
+
+    try:
+        return os.sysconf('SC_NPROCESSORS_ONLN')
+    except (ValueError, OSError, AttributeError):
+        pass
+
+    try:
+        return int(os.popen2("sysctl -n hw.ncpu")[1].read())
+    except ValueError:
+        pass
+
+    return 1
+
+
 def generate(env):
     """Common environment generation code"""
 
@@ -207,6 +226,10 @@ def generate(env):
     env.SConsignFile(os.path.join(build_dir, '.sconsign'))
     env.CacheDir('build/cache')
 
+    # Parallel build
+    if env.GetOption('num_jobs') <= 1:
+        env.SetOption('num_jobs', num_jobs())
+
     # C preprocessor options
     cppdefines = []
     if debug:
diff --git a/scons/generic.py b/scons/generic.py
index 05f7356b76..2323196673 100644
--- a/scons/generic.py
+++ b/scons/generic.py
@@ -206,6 +206,25 @@ _bool_map = {
 }
 
 
+def num_jobs():
+    try:
+        return int(os.environ['NUMBER_OF_PROCESSORS'])
+    except (ValueError, KeyError):
+        pass
+
+    try:
+        return os.sysconf('SC_NPROCESSORS_ONLN')
+    except (ValueError, OSError, AttributeError):
+        pass
+
+    try:
+        return int(os.popen2("sysctl -n hw.ncpu")[1].read())
+    except ValueError:
+        pass
+
+    return 1
+
+
 def generate(env):
     """Common environment generation code"""
 
@@ -266,6 +285,10 @@ def generate(env):
     # different scons versions building the same source file
     env.SConsignFile(os.path.join(env['build'], '.sconsign'))
 
+    # Parallel build
+    if env.GetOption('num_jobs') <= 1:
+        env.SetOption('num_jobs', num_jobs())
+
     # Summary
     print
     print '  platform=%s' % env['platform']
@@ -274,6 +297,7 @@ def generate(env):
     print '  debug=%s' % ['no', 'yes'][env['debug']]
     print '  profile=%s' % ['no', 'yes'][env['profile']]
     print '  build=%s' % env['build']
+    print '  %s jobs' % env.GetOption('num_jobs')
     print
 
     # Load tool chain
diff --git a/src/gallium/auxiliary/draw/draw_pipe_pstipple.c b/src/gallium/auxiliary/draw/draw_pipe_pstipple.c
index b764d9c518..a0f9716dac 100644
--- a/src/gallium/auxiliary/draw/draw_pipe_pstipple.c
+++ b/src/gallium/auxiliary/draw/draw_pipe_pstipple.c
@@ -256,7 +256,7 @@ pstip_transform_inst(struct tgsi_transform_context *ctx,
          struct tgsi_full_immediate immed;
          uint size = 4;
          immed = tgsi_default_full_immediate();
-         immed.Immediate.Size = 1 + size; /* one for the token itself */
+         immed.Immediate.NrTokens = 1 + size; /* one for the token itself */
          immed.u.Pointer = (void *) value;
          ctx->emit_immediate(ctx, &immed);
       }
diff --git a/src/gallium/auxiliary/draw/draw_vs_aos.c b/src/gallium/auxiliary/draw/draw_vs_aos.c
index 6817f29c2a..0c693a4a65 100644
--- a/src/gallium/auxiliary/draw/draw_vs_aos.c
+++ b/src/gallium/auxiliary/draw/draw_vs_aos.c
@@ -1870,7 +1870,7 @@ static boolean note_immediate( struct aos_compilation *cp,
    unsigned pos = cp->num_immediates++;
    unsigned j;
 
-   for (j = 0; j < imm->Immediate.Size; j++) {
+   for (j = 0; j < imm->Immediate.NrTokens - 1; j++) {
       cp->vaos->machine->immediate[pos][j] = imm->u.ImmediateFloat32[j].Float;
    }
 
diff --git a/src/gallium/auxiliary/gallivm/tgsitollvm.cpp b/src/gallium/auxiliary/gallivm/tgsitollvm.cpp
index 6b18a68fe6..5b08200d14 100644
--- a/src/gallium/auxiliary/gallivm/tgsitollvm.cpp
+++ b/src/gallium/auxiliary/gallivm/tgsitollvm.cpp
@@ -160,7 +160,7 @@ translate_immediate(Storage *storage,
 {
    float vec[4];
    int i;
-   for (i = 0; i < imm->Immediate.Size - 1; ++i) {
+   for (i = 0; i < imm->Immediate.NrTokens - 1; ++i) {
       switch (imm->Immediate.DataType) {
       case TGSI_IMM_FLOAT32:
          vec[i] = imm->u.ImmediateFloat32[i].Float;
@@ -179,7 +179,7 @@ translate_immediateir(StorageSoa *storage,
 {
    float vec[4];
    int i;
-   for (i = 0; i < imm->Immediate.Size - 1; ++i) {
+   for (i = 0; i < imm->Immediate.NrTokens - 1; ++i) {
       switch (imm->Immediate.DataType) {
       case TGSI_IMM_FLOAT32:
          vec[i] = imm->u.ImmediateFloat32[i].Float;
diff --git a/src/gallium/auxiliary/indices/u_indices_gen.c b/src/gallium/auxiliary/indices/u_indices_gen.c
index 01953c3244..4c05b3eedb 100644
--- a/src/gallium/auxiliary/indices/u_indices_gen.c
+++ b/src/gallium/auxiliary/indices/u_indices_gen.c
@@ -1133,7 +1133,7 @@ static void translate_points_ubyte2ushort_first2first(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -1146,7 +1146,7 @@ static void translate_lines_ubyte2ushort_first2first(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -1160,7 +1160,7 @@ static void translate_linestrip_ubyte2ushort_first2first(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -1174,7 +1174,7 @@ static void translate_lineloop_ubyte2ushort_first2first(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -1190,7 +1190,7 @@ static void translate_tris_ubyte2ushort_first2first(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -1205,7 +1205,7 @@ static void translate_tristrip_ubyte2ushort_first2first(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -1220,7 +1220,7 @@ static void translate_trifan_ubyte2ushort_first2first(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -1235,7 +1235,7 @@ static void translate_quads_ubyte2ushort_first2first(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -1253,7 +1253,7 @@ static void translate_quadstrip_ubyte2ushort_first2first(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -1271,7 +1271,7 @@ static void translate_polygon_ubyte2ushort_first2first(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -1286,7 +1286,7 @@ static void translate_points_ubyte2ushort_first2last(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -1299,7 +1299,7 @@ static void translate_lines_ubyte2ushort_first2last(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -1313,7 +1313,7 @@ static void translate_linestrip_ubyte2ushort_first2last(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -1327,7 +1327,7 @@ static void translate_lineloop_ubyte2ushort_first2last(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -1343,7 +1343,7 @@ static void translate_tris_ubyte2ushort_first2last(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -1358,7 +1358,7 @@ static void translate_tristrip_ubyte2ushort_first2last(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -1373,7 +1373,7 @@ static void translate_trifan_ubyte2ushort_first2last(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -1388,7 +1388,7 @@ static void translate_quads_ubyte2ushort_first2last(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -1406,7 +1406,7 @@ static void translate_quadstrip_ubyte2ushort_first2last(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -1424,7 +1424,7 @@ static void translate_polygon_ubyte2ushort_first2last(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -1439,7 +1439,7 @@ static void translate_points_ubyte2ushort_last2first(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -1452,7 +1452,7 @@ static void translate_lines_ubyte2ushort_last2first(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -1466,7 +1466,7 @@ static void translate_linestrip_ubyte2ushort_last2first(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -1480,7 +1480,7 @@ static void translate_lineloop_ubyte2ushort_last2first(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -1496,7 +1496,7 @@ static void translate_tris_ubyte2ushort_last2first(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -1511,7 +1511,7 @@ static void translate_tristrip_ubyte2ushort_last2first(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -1526,7 +1526,7 @@ static void translate_trifan_ubyte2ushort_last2first(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -1541,7 +1541,7 @@ static void translate_quads_ubyte2ushort_last2first(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -1559,7 +1559,7 @@ static void translate_quadstrip_ubyte2ushort_last2first(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -1577,7 +1577,7 @@ static void translate_polygon_ubyte2ushort_last2first(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -1592,7 +1592,7 @@ static void translate_points_ubyte2ushort_last2last(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -1605,7 +1605,7 @@ static void translate_lines_ubyte2ushort_last2last(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -1619,7 +1619,7 @@ static void translate_linestrip_ubyte2ushort_last2last(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -1633,7 +1633,7 @@ static void translate_lineloop_ubyte2ushort_last2last(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -1649,7 +1649,7 @@ static void translate_tris_ubyte2ushort_last2last(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -1664,7 +1664,7 @@ static void translate_tristrip_ubyte2ushort_last2last(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -1679,7 +1679,7 @@ static void translate_trifan_ubyte2ushort_last2last(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -1694,7 +1694,7 @@ static void translate_quads_ubyte2ushort_last2last(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -1712,7 +1712,7 @@ static void translate_quadstrip_ubyte2ushort_last2last(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -1730,7 +1730,7 @@ static void translate_polygon_ubyte2ushort_last2last(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -1745,7 +1745,7 @@ static void translate_points_ubyte2uint_first2first(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -1758,7 +1758,7 @@ static void translate_lines_ubyte2uint_first2first(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -1772,7 +1772,7 @@ static void translate_linestrip_ubyte2uint_first2first(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -1786,7 +1786,7 @@ static void translate_lineloop_ubyte2uint_first2first(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -1802,7 +1802,7 @@ static void translate_tris_ubyte2uint_first2first(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -1817,7 +1817,7 @@ static void translate_tristrip_ubyte2uint_first2first(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -1832,7 +1832,7 @@ static void translate_trifan_ubyte2uint_first2first(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -1847,7 +1847,7 @@ static void translate_quads_ubyte2uint_first2first(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -1865,7 +1865,7 @@ static void translate_quadstrip_ubyte2uint_first2first(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -1883,7 +1883,7 @@ static void translate_polygon_ubyte2uint_first2first(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -1898,7 +1898,7 @@ static void translate_points_ubyte2uint_first2last(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -1911,7 +1911,7 @@ static void translate_lines_ubyte2uint_first2last(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -1925,7 +1925,7 @@ static void translate_linestrip_ubyte2uint_first2last(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -1939,7 +1939,7 @@ static void translate_lineloop_ubyte2uint_first2last(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -1955,7 +1955,7 @@ static void translate_tris_ubyte2uint_first2last(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -1970,7 +1970,7 @@ static void translate_tristrip_ubyte2uint_first2last(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -1985,7 +1985,7 @@ static void translate_trifan_ubyte2uint_first2last(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -2000,7 +2000,7 @@ static void translate_quads_ubyte2uint_first2last(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -2018,7 +2018,7 @@ static void translate_quadstrip_ubyte2uint_first2last(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -2036,7 +2036,7 @@ static void translate_polygon_ubyte2uint_first2last(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -2051,7 +2051,7 @@ static void translate_points_ubyte2uint_last2first(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -2064,7 +2064,7 @@ static void translate_lines_ubyte2uint_last2first(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -2078,7 +2078,7 @@ static void translate_linestrip_ubyte2uint_last2first(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -2092,7 +2092,7 @@ static void translate_lineloop_ubyte2uint_last2first(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -2108,7 +2108,7 @@ static void translate_tris_ubyte2uint_last2first(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -2123,7 +2123,7 @@ static void translate_tristrip_ubyte2uint_last2first(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -2138,7 +2138,7 @@ static void translate_trifan_ubyte2uint_last2first(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -2153,7 +2153,7 @@ static void translate_quads_ubyte2uint_last2first(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -2171,7 +2171,7 @@ static void translate_quadstrip_ubyte2uint_last2first(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -2189,7 +2189,7 @@ static void translate_polygon_ubyte2uint_last2first(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -2204,7 +2204,7 @@ static void translate_points_ubyte2uint_last2last(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -2217,7 +2217,7 @@ static void translate_lines_ubyte2uint_last2last(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -2231,7 +2231,7 @@ static void translate_linestrip_ubyte2uint_last2last(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -2245,7 +2245,7 @@ static void translate_lineloop_ubyte2uint_last2last(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -2261,7 +2261,7 @@ static void translate_tris_ubyte2uint_last2last(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -2276,7 +2276,7 @@ static void translate_tristrip_ubyte2uint_last2last(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -2291,7 +2291,7 @@ static void translate_trifan_ubyte2uint_last2last(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -2306,7 +2306,7 @@ static void translate_quads_ubyte2uint_last2last(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -2324,7 +2324,7 @@ static void translate_quadstrip_ubyte2uint_last2last(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -2342,7 +2342,7 @@ static void translate_polygon_ubyte2uint_last2last(
     unsigned nr,
     void *_out )
 {
-  const ubyte*in = (const ubyte*)in;
+  const ubyte*in = (const ubyte*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -2357,7 +2357,7 @@ static void translate_points_ushort2ushort_first2first(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -2370,7 +2370,7 @@ static void translate_lines_ushort2ushort_first2first(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -2384,7 +2384,7 @@ static void translate_linestrip_ushort2ushort_first2first(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -2398,7 +2398,7 @@ static void translate_lineloop_ushort2ushort_first2first(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -2414,7 +2414,7 @@ static void translate_tris_ushort2ushort_first2first(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -2429,7 +2429,7 @@ static void translate_tristrip_ushort2ushort_first2first(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -2444,7 +2444,7 @@ static void translate_trifan_ushort2ushort_first2first(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -2459,7 +2459,7 @@ static void translate_quads_ushort2ushort_first2first(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -2477,7 +2477,7 @@ static void translate_quadstrip_ushort2ushort_first2first(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -2495,7 +2495,7 @@ static void translate_polygon_ushort2ushort_first2first(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -2510,7 +2510,7 @@ static void translate_points_ushort2ushort_first2last(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -2523,7 +2523,7 @@ static void translate_lines_ushort2ushort_first2last(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -2537,7 +2537,7 @@ static void translate_linestrip_ushort2ushort_first2last(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -2551,7 +2551,7 @@ static void translate_lineloop_ushort2ushort_first2last(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -2567,7 +2567,7 @@ static void translate_tris_ushort2ushort_first2last(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -2582,7 +2582,7 @@ static void translate_tristrip_ushort2ushort_first2last(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -2597,7 +2597,7 @@ static void translate_trifan_ushort2ushort_first2last(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -2612,7 +2612,7 @@ static void translate_quads_ushort2ushort_first2last(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -2630,7 +2630,7 @@ static void translate_quadstrip_ushort2ushort_first2last(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -2648,7 +2648,7 @@ static void translate_polygon_ushort2ushort_first2last(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -2663,7 +2663,7 @@ static void translate_points_ushort2ushort_last2first(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -2676,7 +2676,7 @@ static void translate_lines_ushort2ushort_last2first(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -2690,7 +2690,7 @@ static void translate_linestrip_ushort2ushort_last2first(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -2704,7 +2704,7 @@ static void translate_lineloop_ushort2ushort_last2first(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -2720,7 +2720,7 @@ static void translate_tris_ushort2ushort_last2first(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -2735,7 +2735,7 @@ static void translate_tristrip_ushort2ushort_last2first(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -2750,7 +2750,7 @@ static void translate_trifan_ushort2ushort_last2first(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -2765,7 +2765,7 @@ static void translate_quads_ushort2ushort_last2first(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -2783,7 +2783,7 @@ static void translate_quadstrip_ushort2ushort_last2first(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -2801,7 +2801,7 @@ static void translate_polygon_ushort2ushort_last2first(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -2816,7 +2816,7 @@ static void translate_points_ushort2ushort_last2last(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -2829,7 +2829,7 @@ static void translate_lines_ushort2ushort_last2last(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -2843,7 +2843,7 @@ static void translate_linestrip_ushort2ushort_last2last(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -2857,7 +2857,7 @@ static void translate_lineloop_ushort2ushort_last2last(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -2873,7 +2873,7 @@ static void translate_tris_ushort2ushort_last2last(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -2888,7 +2888,7 @@ static void translate_tristrip_ushort2ushort_last2last(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -2903,7 +2903,7 @@ static void translate_trifan_ushort2ushort_last2last(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -2918,7 +2918,7 @@ static void translate_quads_ushort2ushort_last2last(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -2936,7 +2936,7 @@ static void translate_quadstrip_ushort2ushort_last2last(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -2954,7 +2954,7 @@ static void translate_polygon_ushort2ushort_last2last(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -2969,7 +2969,7 @@ static void translate_points_ushort2uint_first2first(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -2982,7 +2982,7 @@ static void translate_lines_ushort2uint_first2first(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -2996,7 +2996,7 @@ static void translate_linestrip_ushort2uint_first2first(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -3010,7 +3010,7 @@ static void translate_lineloop_ushort2uint_first2first(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -3026,7 +3026,7 @@ static void translate_tris_ushort2uint_first2first(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -3041,7 +3041,7 @@ static void translate_tristrip_ushort2uint_first2first(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -3056,7 +3056,7 @@ static void translate_trifan_ushort2uint_first2first(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -3071,7 +3071,7 @@ static void translate_quads_ushort2uint_first2first(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -3089,7 +3089,7 @@ static void translate_quadstrip_ushort2uint_first2first(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -3107,7 +3107,7 @@ static void translate_polygon_ushort2uint_first2first(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -3122,7 +3122,7 @@ static void translate_points_ushort2uint_first2last(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -3135,7 +3135,7 @@ static void translate_lines_ushort2uint_first2last(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -3149,7 +3149,7 @@ static void translate_linestrip_ushort2uint_first2last(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -3163,7 +3163,7 @@ static void translate_lineloop_ushort2uint_first2last(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -3179,7 +3179,7 @@ static void translate_tris_ushort2uint_first2last(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -3194,7 +3194,7 @@ static void translate_tristrip_ushort2uint_first2last(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -3209,7 +3209,7 @@ static void translate_trifan_ushort2uint_first2last(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -3224,7 +3224,7 @@ static void translate_quads_ushort2uint_first2last(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -3242,7 +3242,7 @@ static void translate_quadstrip_ushort2uint_first2last(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -3260,7 +3260,7 @@ static void translate_polygon_ushort2uint_first2last(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -3275,7 +3275,7 @@ static void translate_points_ushort2uint_last2first(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -3288,7 +3288,7 @@ static void translate_lines_ushort2uint_last2first(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -3302,7 +3302,7 @@ static void translate_linestrip_ushort2uint_last2first(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -3316,7 +3316,7 @@ static void translate_lineloop_ushort2uint_last2first(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -3332,7 +3332,7 @@ static void translate_tris_ushort2uint_last2first(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -3347,7 +3347,7 @@ static void translate_tristrip_ushort2uint_last2first(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -3362,7 +3362,7 @@ static void translate_trifan_ushort2uint_last2first(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -3377,7 +3377,7 @@ static void translate_quads_ushort2uint_last2first(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -3395,7 +3395,7 @@ static void translate_quadstrip_ushort2uint_last2first(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -3413,7 +3413,7 @@ static void translate_polygon_ushort2uint_last2first(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -3428,7 +3428,7 @@ static void translate_points_ushort2uint_last2last(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -3441,7 +3441,7 @@ static void translate_lines_ushort2uint_last2last(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -3455,7 +3455,7 @@ static void translate_linestrip_ushort2uint_last2last(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -3469,7 +3469,7 @@ static void translate_lineloop_ushort2uint_last2last(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -3485,7 +3485,7 @@ static void translate_tris_ushort2uint_last2last(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -3500,7 +3500,7 @@ static void translate_tristrip_ushort2uint_last2last(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -3515,7 +3515,7 @@ static void translate_trifan_ushort2uint_last2last(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -3530,7 +3530,7 @@ static void translate_quads_ushort2uint_last2last(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -3548,7 +3548,7 @@ static void translate_quadstrip_ushort2uint_last2last(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -3566,7 +3566,7 @@ static void translate_polygon_ushort2uint_last2last(
     unsigned nr,
     void *_out )
 {
-  const ushort*in = (const ushort*)in;
+  const ushort*in = (const ushort*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -3581,7 +3581,7 @@ static void translate_points_uint2ushort_first2first(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -3594,7 +3594,7 @@ static void translate_lines_uint2ushort_first2first(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -3608,7 +3608,7 @@ static void translate_linestrip_uint2ushort_first2first(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -3622,7 +3622,7 @@ static void translate_lineloop_uint2ushort_first2first(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -3638,7 +3638,7 @@ static void translate_tris_uint2ushort_first2first(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -3653,7 +3653,7 @@ static void translate_tristrip_uint2ushort_first2first(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -3668,7 +3668,7 @@ static void translate_trifan_uint2ushort_first2first(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -3683,7 +3683,7 @@ static void translate_quads_uint2ushort_first2first(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -3701,7 +3701,7 @@ static void translate_quadstrip_uint2ushort_first2first(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -3719,7 +3719,7 @@ static void translate_polygon_uint2ushort_first2first(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -3734,7 +3734,7 @@ static void translate_points_uint2ushort_first2last(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -3747,7 +3747,7 @@ static void translate_lines_uint2ushort_first2last(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -3761,7 +3761,7 @@ static void translate_linestrip_uint2ushort_first2last(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -3775,7 +3775,7 @@ static void translate_lineloop_uint2ushort_first2last(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -3791,7 +3791,7 @@ static void translate_tris_uint2ushort_first2last(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -3806,7 +3806,7 @@ static void translate_tristrip_uint2ushort_first2last(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -3821,7 +3821,7 @@ static void translate_trifan_uint2ushort_first2last(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -3836,7 +3836,7 @@ static void translate_quads_uint2ushort_first2last(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -3854,7 +3854,7 @@ static void translate_quadstrip_uint2ushort_first2last(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -3872,7 +3872,7 @@ static void translate_polygon_uint2ushort_first2last(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -3887,7 +3887,7 @@ static void translate_points_uint2ushort_last2first(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -3900,7 +3900,7 @@ static void translate_lines_uint2ushort_last2first(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -3914,7 +3914,7 @@ static void translate_linestrip_uint2ushort_last2first(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -3928,7 +3928,7 @@ static void translate_lineloop_uint2ushort_last2first(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -3944,7 +3944,7 @@ static void translate_tris_uint2ushort_last2first(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -3959,7 +3959,7 @@ static void translate_tristrip_uint2ushort_last2first(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -3974,7 +3974,7 @@ static void translate_trifan_uint2ushort_last2first(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -3989,7 +3989,7 @@ static void translate_quads_uint2ushort_last2first(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -4007,7 +4007,7 @@ static void translate_quadstrip_uint2ushort_last2first(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -4025,7 +4025,7 @@ static void translate_polygon_uint2ushort_last2first(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -4040,7 +4040,7 @@ static void translate_points_uint2ushort_last2last(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -4053,7 +4053,7 @@ static void translate_lines_uint2ushort_last2last(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -4067,7 +4067,7 @@ static void translate_linestrip_uint2ushort_last2last(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -4081,7 +4081,7 @@ static void translate_lineloop_uint2ushort_last2last(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -4097,7 +4097,7 @@ static void translate_tris_uint2ushort_last2last(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -4112,7 +4112,7 @@ static void translate_tristrip_uint2ushort_last2last(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -4127,7 +4127,7 @@ static void translate_trifan_uint2ushort_last2last(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -4142,7 +4142,7 @@ static void translate_quads_uint2ushort_last2last(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -4160,7 +4160,7 @@ static void translate_quadstrip_uint2ushort_last2last(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -4178,7 +4178,7 @@ static void translate_polygon_uint2ushort_last2last(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
@@ -4193,7 +4193,7 @@ static void translate_points_uint2uint_first2first(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -4206,7 +4206,7 @@ static void translate_lines_uint2uint_first2first(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -4220,7 +4220,7 @@ static void translate_linestrip_uint2uint_first2first(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -4234,7 +4234,7 @@ static void translate_lineloop_uint2uint_first2first(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -4250,7 +4250,7 @@ static void translate_tris_uint2uint_first2first(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -4265,7 +4265,7 @@ static void translate_tristrip_uint2uint_first2first(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -4280,7 +4280,7 @@ static void translate_trifan_uint2uint_first2first(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -4295,7 +4295,7 @@ static void translate_quads_uint2uint_first2first(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -4313,7 +4313,7 @@ static void translate_quadstrip_uint2uint_first2first(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -4331,7 +4331,7 @@ static void translate_polygon_uint2uint_first2first(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -4346,7 +4346,7 @@ static void translate_points_uint2uint_first2last(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -4359,7 +4359,7 @@ static void translate_lines_uint2uint_first2last(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -4373,7 +4373,7 @@ static void translate_linestrip_uint2uint_first2last(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -4387,7 +4387,7 @@ static void translate_lineloop_uint2uint_first2last(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -4403,7 +4403,7 @@ static void translate_tris_uint2uint_first2last(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -4418,7 +4418,7 @@ static void translate_tristrip_uint2uint_first2last(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -4433,7 +4433,7 @@ static void translate_trifan_uint2uint_first2last(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -4448,7 +4448,7 @@ static void translate_quads_uint2uint_first2last(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -4466,7 +4466,7 @@ static void translate_quadstrip_uint2uint_first2last(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -4484,7 +4484,7 @@ static void translate_polygon_uint2uint_first2last(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -4499,7 +4499,7 @@ static void translate_points_uint2uint_last2first(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -4512,7 +4512,7 @@ static void translate_lines_uint2uint_last2first(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -4526,7 +4526,7 @@ static void translate_linestrip_uint2uint_last2first(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -4540,7 +4540,7 @@ static void translate_lineloop_uint2uint_last2first(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -4556,7 +4556,7 @@ static void translate_tris_uint2uint_last2first(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -4571,7 +4571,7 @@ static void translate_tristrip_uint2uint_last2first(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -4586,7 +4586,7 @@ static void translate_trifan_uint2uint_last2first(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -4601,7 +4601,7 @@ static void translate_quads_uint2uint_last2first(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -4619,7 +4619,7 @@ static void translate_quadstrip_uint2uint_last2first(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -4637,7 +4637,7 @@ static void translate_polygon_uint2uint_last2first(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -4652,7 +4652,7 @@ static void translate_points_uint2uint_last2last(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -4665,7 +4665,7 @@ static void translate_lines_uint2uint_last2last(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -4679,7 +4679,7 @@ static void translate_linestrip_uint2uint_last2last(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -4693,7 +4693,7 @@ static void translate_lineloop_uint2uint_last2last(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -4709,7 +4709,7 @@ static void translate_tris_uint2uint_last2last(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -4724,7 +4724,7 @@ static void translate_tristrip_uint2uint_last2last(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -4739,7 +4739,7 @@ static void translate_trifan_uint2uint_last2last(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -4754,7 +4754,7 @@ static void translate_quads_uint2uint_last2last(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -4772,7 +4772,7 @@ static void translate_quadstrip_uint2uint_last2last(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
@@ -4790,7 +4790,7 @@ static void translate_polygon_uint2uint_last2last(
     unsigned nr,
     void *_out )
 {
-  const uint*in = (const uint*)in;
+  const uint*in = (const uint*)_in;
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
diff --git a/src/gallium/auxiliary/indices/u_indices_gen.py b/src/gallium/auxiliary/indices/u_indices_gen.py
index 40b047bb58..0dc58d0cd0 100644
--- a/src/gallium/auxiliary/indices/u_indices_gen.py
+++ b/src/gallium/auxiliary/indices/u_indices_gen.py
@@ -157,7 +157,7 @@ def preamble(intype, outtype, inpv, outpv, prim):
     print '    void *_out )'
     print '{'
     if intype != GENERATE:
-        print '  const ' + intype + '*in = (const ' + intype + '*)in;'
+        print '  const ' + intype + '*in = (const ' + intype + '*)_in;'
     print '  ' + outtype + ' *out = (' + outtype + '*)_out;'
     print '  unsigned i, j;'
     print '  (void)j;'
diff --git a/src/gallium/auxiliary/pipebuffer/pb_buffer_fenced.c b/src/gallium/auxiliary/pipebuffer/pb_buffer_fenced.c
index 61afdfe82a..f9e6226436 100644
--- a/src/gallium/auxiliary/pipebuffer/pb_buffer_fenced.c
+++ b/src/gallium/auxiliary/pipebuffer/pb_buffer_fenced.c
@@ -44,7 +44,6 @@
 #include "pipe/p_compiler.h"
 #include "pipe/p_error.h"
 #include "pipe/p_debug.h"
-#include "pipe/internal/p_winsys_screen.h"
 #include "pipe/p_thread.h"
 #include "util/u_memory.h"
 #include "util/u_double_list.h"
@@ -64,7 +63,7 @@ struct fenced_buffer_list
 {
    pipe_mutex mutex;
    
-   struct pipe_winsys *winsys;
+   struct pb_fence_ops *ops;
    
    size_t numDelayed;
    
@@ -140,12 +139,12 @@ static INLINE void
 _fenced_buffer_remove(struct fenced_buffer_list *fenced_list,
                       struct fenced_buffer *fenced_buf)
 {
-   struct pipe_winsys *winsys = fenced_list->winsys;
+   struct pb_fence_ops *ops = fenced_list->ops;
 
    assert(fenced_buf->fence);
    assert(fenced_buf->list == fenced_list);
    
-   winsys->fence_reference(winsys, &fenced_buf->fence, NULL);
+   ops->fence_reference(ops, &fenced_buf->fence, NULL);
    fenced_buf->flags &= ~PIPE_BUFFER_USAGE_GPU_READ_WRITE;
    
    assert(fenced_buf->head.prev);
@@ -168,7 +167,7 @@ static INLINE enum pipe_error
 _fenced_buffer_finish(struct fenced_buffer *fenced_buf)
 {
    struct fenced_buffer_list *fenced_list = fenced_buf->list;
-   struct pipe_winsys *winsys = fenced_list->winsys;
+   struct pb_fence_ops *ops = fenced_list->ops;
 
 #if 0
    debug_warning("waiting for GPU");
@@ -176,7 +175,7 @@ _fenced_buffer_finish(struct fenced_buffer *fenced_buf)
 
    assert(fenced_buf->fence);
    if(fenced_buf->fence) {
-      if(winsys->fence_finish(winsys, fenced_buf->fence, 0) != 0) {
+      if(ops->fence_finish(ops, fenced_buf->fence, 0) != 0) {
 	 return PIPE_ERROR;
       }
       /* Remove from the fenced list */
@@ -196,7 +195,7 @@ static void
 _fenced_buffer_list_check_free(struct fenced_buffer_list *fenced_list, 
                                int wait)
 {
-   struct pipe_winsys *winsys = fenced_list->winsys;
+   struct pb_fence_ops *ops = fenced_list->ops;
    struct list_head *curr, *next;
    struct fenced_buffer *fenced_buf;
    struct pipe_fence_handle *prev_fence = NULL;
@@ -209,15 +208,15 @@ _fenced_buffer_list_check_free(struct fenced_buffer_list *fenced_list,
       if(fenced_buf->fence != prev_fence) {
 	 int signaled;
 	 if (wait)
-	    signaled = winsys->fence_finish(winsys, fenced_buf->fence, 0);
+	    signaled = ops->fence_finish(ops, fenced_buf->fence, 0);
 	 else
-	    signaled = winsys->fence_signalled(winsys, fenced_buf->fence, 0);
+	    signaled = ops->fence_signalled(ops, fenced_buf->fence, 0);
 	 if (signaled != 0)
 	    break;
 	 prev_fence = fenced_buf->fence;
       }
       else {
-	 assert(winsys->fence_signalled(winsys, fenced_buf->fence, 0) == 0);
+	 assert(ops->fence_signalled(ops, fenced_buf->fence, 0) == 0);
       }
 
       _fenced_buffer_remove(fenced_list, fenced_buf);
@@ -237,14 +236,14 @@ fenced_buffer_destroy(struct pb_buffer *buf)
    pipe_mutex_lock(fenced_list->mutex);
    assert(fenced_buf->base.base.refcount == 0);
    if (fenced_buf->fence) {
-      struct pipe_winsys *winsys = fenced_list->winsys;
-      if(winsys->fence_signalled(winsys, fenced_buf->fence, 0) == 0) {
+      struct pb_fence_ops *ops = fenced_list->ops;
+      if(ops->fence_signalled(ops, fenced_buf->fence, 0) == 0) {
 	 struct list_head *curr, *prev;
 	 curr = &fenced_buf->head;
 	 prev = curr->prev;
 	 do {
 	    fenced_buf = LIST_ENTRY(struct fenced_buffer, curr, head);
-	    assert(winsys->fence_signalled(winsys, fenced_buf->fence, 0) == 0);
+	    assert(ops->fence_signalled(ops, fenced_buf->fence, 0) == 0);
 	    _fenced_buffer_remove(fenced_list, fenced_buf);
 	    curr = prev;
 	    prev = curr->prev;
@@ -366,11 +365,11 @@ fenced_buffer_fence(struct pb_buffer *buf,
 {
    struct fenced_buffer *fenced_buf;
    struct fenced_buffer_list *fenced_list;
-   struct pipe_winsys *winsys;
+   struct pb_fence_ops *ops;
 
    fenced_buf = fenced_buffer(buf);
    fenced_list = fenced_buf->list;
-   winsys = fenced_list->winsys;
+   ops = fenced_list->ops;
    
    if(fence == fenced_buf->fence) {
       /* Nothing to do */
@@ -384,7 +383,7 @@ fenced_buffer_fence(struct pb_buffer *buf,
    if (fenced_buf->fence)
       _fenced_buffer_remove(fenced_list, fenced_buf);
    if (fence) {
-      winsys->fence_reference(winsys, &fenced_buf->fence, fence);
+      ops->fence_reference(ops, &fenced_buf->fence, fence);
       fenced_buf->flags |= fenced_buf->validation_flags;
       _fenced_buffer_add(fenced_buf);
    }
@@ -447,7 +446,7 @@ fenced_buffer_create(struct fenced_buffer_list *fenced_list,
 
 
 struct fenced_buffer_list *
-fenced_buffer_list_create(struct pipe_winsys *winsys) 
+fenced_buffer_list_create(struct pb_fence_ops *ops) 
 {
    struct fenced_buffer_list *fenced_list;
 
@@ -455,7 +454,7 @@ fenced_buffer_list_create(struct pipe_winsys *winsys)
    if (!fenced_list)
       return NULL;
 
-   fenced_list->winsys = winsys;
+   fenced_list->ops = ops;
 
    LIST_INITHEAD(&fenced_list->delayed);
 
@@ -494,6 +493,8 @@ fenced_buffer_list_destroy(struct fenced_buffer_list *fenced_list)
 
    pipe_mutex_unlock(fenced_list->mutex);
    
+   fenced_list->ops->destroy(fenced_list->ops);
+   
    FREE(fenced_list);
 }
 
diff --git a/src/gallium/auxiliary/pipebuffer/pb_buffer_fenced.h b/src/gallium/auxiliary/pipebuffer/pb_buffer_fenced.h
index b15c676194..d1c9d4c17d 100644
--- a/src/gallium/auxiliary/pipebuffer/pb_buffer_fenced.h
+++ b/src/gallium/auxiliary/pipebuffer/pb_buffer_fenced.h
@@ -59,7 +59,6 @@ extern "C" {
 #endif
 
 
-struct pipe_winsys;
 struct pipe_buffer;
 struct pipe_fence_handle;
 
@@ -70,13 +69,42 @@ struct pipe_fence_handle;
 struct fenced_buffer_list;
 
 
+struct pb_fence_ops
+{
+   void (*destroy)( struct pb_fence_ops *ops );
+
+   /** Set ptr = fence, with reference counting */
+   void (*fence_reference)( struct pb_fence_ops *ops,
+                            struct pipe_fence_handle **ptr,
+                            struct pipe_fence_handle *fence );
+
+   /**
+    * Checks whether the fence has been signalled.
+    * \param flags  driver-specific meaning
+    * \return zero on success.
+    */
+   int (*fence_signalled)( struct pb_fence_ops *ops,
+                           struct pipe_fence_handle *fence,
+                           unsigned flag );
+
+   /**
+    * Wait for the fence to finish.
+    * \param flags  driver-specific meaning
+    * \return zero on success.
+    */
+   int (*fence_finish)( struct pb_fence_ops *ops,
+                        struct pipe_fence_handle *fence,
+                        unsigned flag );
+};
+
+
 /**
  * Create a fenced buffer list.
  * 
  * See also fenced_bufmgr_create for a more convenient way to use this.
  */
 struct fenced_buffer_list *
-fenced_buffer_list_create(struct pipe_winsys *winsys);
+fenced_buffer_list_create(struct pb_fence_ops *ops);
 
 
 /**
diff --git a/src/gallium/auxiliary/pipebuffer/pb_bufmgr.h b/src/gallium/auxiliary/pipebuffer/pb_bufmgr.h
index 0a8264a924..fec8db91c7 100644
--- a/src/gallium/auxiliary/pipebuffer/pb_bufmgr.h
+++ b/src/gallium/auxiliary/pipebuffer/pb_bufmgr.h
@@ -61,7 +61,6 @@ extern "C" {
 
 struct pb_desc;
 struct pipe_buffer;
-struct pipe_winsys;
 
 
 /** 
@@ -163,6 +162,8 @@ pb_cache_manager_create(struct pb_manager *provider,
                      	unsigned usecs); 
 
 
+struct pb_fence_ops;
+
 /** 
  * Fenced buffer manager.
  *
@@ -174,7 +175,7 @@ pb_cache_manager_create(struct pb_manager *provider,
  */
 struct pb_manager *
 fenced_bufmgr_create(struct pb_manager *provider,
-                     struct pipe_winsys *winsys);
+                     struct pb_fence_ops *ops);
 
 
 struct pb_manager *
diff --git a/src/gallium/auxiliary/pipebuffer/pb_bufmgr_fenced.c b/src/gallium/auxiliary/pipebuffer/pb_bufmgr_fenced.c
index 513ed28ca6..47e9fee533 100644
--- a/src/gallium/auxiliary/pipebuffer/pb_bufmgr_fenced.c
+++ b/src/gallium/auxiliary/pipebuffer/pb_bufmgr_fenced.c
@@ -122,7 +122,7 @@ fenced_bufmgr_destroy(struct pb_manager *mgr)
 
 struct pb_manager *
 fenced_bufmgr_create(struct pb_manager *provider, 
-                     struct pipe_winsys *winsys) 
+                     struct pb_fence_ops *ops) 
 {
    struct fenced_pb_manager *fenced_mgr;
 
@@ -138,7 +138,7 @@ fenced_bufmgr_create(struct pb_manager *provider,
    fenced_mgr->base.flush = fenced_bufmgr_flush;
 
    fenced_mgr->provider = provider;
-   fenced_mgr->fenced_list = fenced_buffer_list_create(winsys);
+   fenced_mgr->fenced_list = fenced_buffer_list_create(ops);
    if(!fenced_mgr->fenced_list) {
       FREE(fenced_mgr);
       return NULL;
diff --git a/src/gallium/auxiliary/tgsi/tgsi_build.c b/src/gallium/auxiliary/tgsi/tgsi_build.c
index fd02c2c87c..17886540cf 100644
--- a/src/gallium/auxiliary/tgsi/tgsi_build.c
+++ b/src/gallium/auxiliary/tgsi/tgsi_build.c
@@ -114,7 +114,7 @@ tgsi_default_declaration( void )
    struct tgsi_declaration declaration;
 
    declaration.Type = TGSI_TOKEN_TYPE_DECLARATION;
-   declaration.Size = 1;
+   declaration.NrTokens = 1;
    declaration.File = TGSI_FILE_NULL;
    declaration.UsageMask = TGSI_WRITEMASK_XYZW;
    declaration.Interpolate = TGSI_INTERPOLATE_CONSTANT;
@@ -160,9 +160,9 @@ declaration_grow(
    struct tgsi_declaration *declaration,
    struct tgsi_header *header )
 {
-   assert( declaration->Size < 0xFF );
+   assert( declaration->NrTokens < 0xFF );
 
-   declaration->Size++;
+   declaration->NrTokens++;
 
    header_bodysize_grow( header );
 }
@@ -308,7 +308,7 @@ tgsi_default_immediate( void )
    struct tgsi_immediate immediate;
 
    immediate.Type = TGSI_TOKEN_TYPE_IMMEDIATE;
-   immediate.Size = 1;
+   immediate.NrTokens = 1;
    immediate.DataType = TGSI_IMM_FLOAT32;
    immediate.Padding = 0;
    immediate.Extended = 0;
@@ -345,9 +345,9 @@ immediate_grow(
    struct tgsi_immediate *immediate,
    struct tgsi_header *header )
 {
-   assert( immediate->Size < 0xFF );
+   assert( immediate->NrTokens < 0xFF );
 
-   immediate->Size++;
+   immediate->NrTokens++;
 
    header_bodysize_grow( header );
 }
@@ -384,7 +384,7 @@ tgsi_build_full_immediate(
 
    *immediate = tgsi_build_immediate( header );
 
-   for( i = 0; i < full_imm->Immediate.Size - 1; i++ ) {
+   for( i = 0; i < full_imm->Immediate.NrTokens - 1; i++ ) {
       struct tgsi_immediate_float32 *if32;
 
       if( maxsize <= size )
@@ -411,7 +411,7 @@ tgsi_default_instruction( void )
    struct tgsi_instruction instruction;
 
    instruction.Type = TGSI_TOKEN_TYPE_INSTRUCTION;
-   instruction.Size = 1;
+   instruction.NrTokens = 1;
    instruction.Opcode = TGSI_OPCODE_MOV;
    instruction.Saturate = TGSI_SAT_NONE;
    instruction.NumDstRegs = 1;
@@ -453,9 +453,9 @@ instruction_grow(
    struct tgsi_instruction *instruction,
    struct tgsi_header *header )
 {
-   assert (instruction->Size <   0xFF);
+   assert (instruction->NrTokens <   0xFF);
 
-   instruction->Size++;
+   instruction->NrTokens++;
 
    header_bodysize_grow( header );
 }
diff --git a/src/gallium/auxiliary/tgsi/tgsi_dump.c b/src/gallium/auxiliary/tgsi/tgsi_dump.c
index 2ed8c2bf07..ab2b1f2c58 100644
--- a/src/gallium/auxiliary/tgsi/tgsi_dump.c
+++ b/src/gallium/auxiliary/tgsi/tgsi_dump.c
@@ -285,7 +285,7 @@ iter_immediate(
    ENM( imm->Immediate.DataType, immediate_type_names );
 
    TXT( " { " );
-   for (i = 0; i < imm->Immediate.Size - 1; i++) {
+   for (i = 0; i < imm->Immediate.NrTokens - 1; i++) {
       switch (imm->Immediate.DataType) {
       case TGSI_IMM_FLOAT32:
          FLT( imm->u.ImmediateFloat32[i].Float );
@@ -294,7 +294,7 @@ iter_immediate(
          assert( 0 );
       }
 
-      if (i < imm->Immediate.Size - 2)
+      if (i < imm->Immediate.NrTokens - 2)
          TXT( ", " );
    }
    TXT( " }" );
diff --git a/src/gallium/auxiliary/tgsi/tgsi_dump_c.c b/src/gallium/auxiliary/tgsi/tgsi_dump_c.c
index c575b6c3e1..2ecf1e2f14 100644
--- a/src/gallium/auxiliary/tgsi/tgsi_dump_c.c
+++ b/src/gallium/auxiliary/tgsi/tgsi_dump_c.c
@@ -283,7 +283,7 @@ dump_immediate_verbose(
       UIX( imm->Immediate.Padding );
    }
 
-   for( i = 0; i < imm->Immediate.Size - 1; i++ ) {
+   for( i = 0; i < imm->Immediate.NrTokens - 1; i++ ) {
       EOL();
       switch( imm->Immediate.DataType ) {
       case TGSI_IMM_FLOAT32:
@@ -675,7 +675,7 @@ tgsi_dump_c(
       ENM( parse.FullToken.Token.Type, TGSI_TOKEN_TYPES );
       if( ignored ) {
          TXT( "\nSize       : " );
-         UID( parse.FullToken.Token.Size );
+         UID( parse.FullToken.Token.NrTokens );
          if( deflt || parse.FullToken.Token.Extended ) {
             TXT( "\nExtended   : " );
             UID( parse.FullToken.Token.Extended );
diff --git a/src/gallium/auxiliary/tgsi/tgsi_exec.c b/src/gallium/auxiliary/tgsi/tgsi_exec.c
index a182e679da..ab641efb60 100644
--- a/src/gallium/auxiliary/tgsi/tgsi_exec.c
+++ b/src/gallium/auxiliary/tgsi/tgsi_exec.c
@@ -202,7 +202,7 @@ tgsi_exec_machine_bind_shader(
 
       case TGSI_TOKEN_TYPE_IMMEDIATE:
          {
-            uint size = parse.FullToken.FullImmediate.Immediate.Size - 1;
+            uint size = parse.FullToken.FullImmediate.Immediate.NrTokens - 1;
             assert( size % 4 == 0 );
             assert( mach->ImmLimit + size / 4 <= TGSI_EXEC_NUM_IMMEDIATES );
 
diff --git a/src/gallium/auxiliary/tgsi/tgsi_parse.c b/src/gallium/auxiliary/tgsi/tgsi_parse.c
index 2cd56e413a..d374b16f9a 100644
--- a/src/gallium/auxiliary/tgsi/tgsi_parse.c
+++ b/src/gallium/auxiliary/tgsi/tgsi_parse.c
@@ -155,8 +155,8 @@ tgsi_parse_token(
       switch (imm->Immediate.DataType) {
       case TGSI_IMM_FLOAT32:
          imm->u.Pointer = MALLOC(
-            sizeof( struct tgsi_immediate_float32 ) * (imm->Immediate.Size - 1) );
-         for( i = 0; i < imm->Immediate.Size - 1; i++ ) {
+            sizeof( struct tgsi_immediate_float32 ) * (imm->Immediate.NrTokens - 1) );
+         for( i = 0; i < imm->Immediate.NrTokens - 1; i++ ) {
             next_token( ctx, (struct tgsi_immediate_float32 *) &imm->u.ImmediateFloat32[i] );
          }
          break;
diff --git a/src/gallium/auxiliary/tgsi/tgsi_ppc.c b/src/gallium/auxiliary/tgsi/tgsi_ppc.c
index 1a4db47501..f365030e52 100644
--- a/src/gallium/auxiliary/tgsi/tgsi_ppc.c
+++ b/src/gallium/auxiliary/tgsi/tgsi_ppc.c
@@ -1327,7 +1327,7 @@ tgsi_emit_ppc(const struct tgsi_token *tokens,
       case TGSI_TOKEN_TYPE_IMMEDIATE:
          /* splat each immediate component into a float[4] vector for SoA */
          {
-            const uint size = parse.FullToken.FullImmediate.Immediate.Size - 1;
+            const uint size = parse.FullToken.FullImmediate.Immediate.NrTokens - 1;
             uint i;
             assert(size <= 4);
             assert(num_immediates < TGSI_EXEC_NUM_IMMEDIATES);
diff --git a/src/gallium/auxiliary/tgsi/tgsi_sse2.c b/src/gallium/auxiliary/tgsi/tgsi_sse2.c
index cac44af7f4..481ba89c5e 100644
--- a/src/gallium/auxiliary/tgsi/tgsi_sse2.c
+++ b/src/gallium/auxiliary/tgsi/tgsi_sse2.c
@@ -2671,7 +2671,7 @@ tgsi_emit_sse2(
       case TGSI_TOKEN_TYPE_IMMEDIATE:
          /* simply copy the immediate values into the next immediates[] slot */
          {
-            const uint size = parse.FullToken.FullImmediate.Immediate.Size - 1;
+            const uint size = parse.FullToken.FullImmediate.Immediate.NrTokens - 1;
             uint i;
             assert(size <= 4);
             assert(num_immediates < TGSI_EXEC_NUM_IMMEDIATES);
diff --git a/src/gallium/auxiliary/tgsi/tgsi_text.c b/src/gallium/auxiliary/tgsi/tgsi_text.c
index 9454563361..1e822fbbea 100644
--- a/src/gallium/auxiliary/tgsi/tgsi_text.c
+++ b/src/gallium/auxiliary/tgsi/tgsi_text.c
@@ -1023,7 +1023,7 @@ static boolean parse_immediate( struct translate_ctx *ctx )
    ctx->cur++;
 
    imm = tgsi_default_full_immediate();
-   imm.Immediate.Size += 4;
+   imm.Immediate.NrTokens += 4;
    imm.Immediate.DataType = TGSI_IMM_FLOAT32;
    imm.u.Pointer = values;
 
diff --git a/src/gallium/drivers/i915simple/i915_fpc_translate.c b/src/gallium/drivers/i915simple/i915_fpc_translate.c
index 43d62c5176..d92bdc1bc6 100644
--- a/src/gallium/drivers/i915simple/i915_fpc_translate.c
+++ b/src/gallium/drivers/i915simple/i915_fpc_translate.c
@@ -964,7 +964,7 @@ i915_translate_instructions(struct i915_fp_compile *p,
                = &parse.FullToken.FullImmediate;
             const uint pos = p->num_immediates++;
             uint j;
-            for (j = 0; j < imm->Immediate.Size; j++) {
+            for (j = 0; j < imm->Immediate.NrTokens - 1; j++) {
                p->immediates[pos][j] = imm->u.ImmediateFloat32[j].Float;
             }
          }
diff --git a/src/gallium/drivers/i965simple/brw_vs_emit.c b/src/gallium/drivers/i965simple/brw_vs_emit.c
index 34dbc0624d..e03d653482 100644
--- a/src/gallium/drivers/i965simple/brw_vs_emit.c
+++ b/src/gallium/drivers/i965simple/brw_vs_emit.c
@@ -1293,7 +1293,7 @@ void brw_vs_emit(struct brw_vs_compile *c)
          break;
       case TGSI_TOKEN_TYPE_IMMEDIATE: {
          struct tgsi_full_immediate *imm = &parse.FullToken.FullImmediate;
-         /*assert(imm->Immediate.Size == 4);*/
+         assert(imm->Immediate.NrTokens == 4 + 1);
          c->prog_data.imm_buf[c->prog_data.num_imm][0] = imm->u.ImmediateFloat32[0].Float;
          c->prog_data.imm_buf[c->prog_data.num_imm][1] = imm->u.ImmediateFloat32[1].Float;
          c->prog_data.imm_buf[c->prog_data.num_imm][2] = imm->u.ImmediateFloat32[2].Float;
diff --git a/src/gallium/drivers/nv04/nv04_surface_2d.c b/src/gallium/drivers/nv04/nv04_surface_2d.c
index 1b5d980468..230cfd17dd 100644
--- a/src/gallium/drivers/nv04/nv04_surface_2d.c
+++ b/src/gallium/drivers/nv04/nv04_surface_2d.c
@@ -1,7 +1,6 @@
 #include "pipe/p_context.h"
 #include "pipe/p_format.h"
 #include "util/u_memory.h"
-#include "util/u_math.h"
 
 #include "nouveau/nouveau_winsys.h"
 #include "nouveau/nouveau_util.h"
@@ -106,9 +105,8 @@ nv04_surface_copy_swizzle(struct nv04_surface_2d *ctx,
 	const unsigned max_h = 1024;
 	const unsigned sub_w = w > max_w ? max_w : w;
 	const unsigned sub_h = h > max_h ? max_h : h;
-	unsigned cx = 0;
-	unsigned cy = 0;
-	int i, offset = 0;
+	unsigned cx;
+	unsigned cy;
 
 	/* POT or GTFO */
 	assert(!(w & (w - 1)) && !(h & (h - 1)));
@@ -116,54 +114,46 @@ nv04_surface_copy_swizzle(struct nv04_surface_2d *ctx,
 	BEGIN_RING(chan, swzsurf, NV04_SWIZZLED_SURFACE_DMA_IMAGE, 1);
 	OUT_RELOCo(chan, dst_bo,
 	                 NOUVEAU_BO_GART | NOUVEAU_BO_VRAM | NOUVEAU_BO_WR);
+
 	BEGIN_RING(chan, swzsurf, NV04_SWIZZLED_SURFACE_FORMAT, 1);
 	OUT_RING  (chan, nv04_surface_format(dst->format) |
 	                 log2i(w) << NV04_SWIZZLED_SURFACE_FORMAT_BASE_SIZE_U_SHIFT |
 	                 log2i(h) << NV04_SWIZZLED_SURFACE_FORMAT_BASE_SIZE_V_SHIFT);
-
+ 
 	BEGIN_RING(chan, sifm, NV04_SCALED_IMAGE_FROM_MEMORY_DMA_IMAGE, 1);
 	OUT_RELOCo(chan, src_bo,
 	                 NOUVEAU_BO_GART | NOUVEAU_BO_VRAM | NOUVEAU_BO_RD);
 	BEGIN_RING(chan, sifm, NV04_SCALED_IMAGE_FROM_MEMORY_SURFACE, 1);
 	OUT_RING  (chan, swzsurf->handle);
 
-	/* Upload, then swizzle each mipmap level in turn */
-	for (i=0; i<src->texture->last_level; i++) {
-
-		for (cy = 0; cy < h; cy += sub_h) {
-		  for (cx = 0; cx < w; cx += sub_w) {
-		    BEGIN_RING(chan, swzsurf, NV04_SWIZZLED_SURFACE_OFFSET, 1);
-		    OUT_RELOCl(chan, dst_bo, dst->offset + nv04_swizzle_bits(cx, cy) *
-				     dst->block.size + offset, NOUVEAU_BO_GART |
-				     NOUVEAU_BO_VRAM | NOUVEAU_BO_WR);
-
-		    BEGIN_RING(chan, sifm, NV04_SCALED_IMAGE_FROM_MEMORY_COLOR_CONVERSION, 9);
-		    OUT_RING  (chan, NV04_SCALED_IMAGE_FROM_MEMORY_COLOR_CONVERSION_TRUNCATE);
-		    OUT_RING  (chan, nv04_scaled_image_format(src->format));
-		    OUT_RING  (chan, NV04_SCALED_IMAGE_FROM_MEMORY_OPERATION_SRCCOPY);
-		    OUT_RING  (chan, 0);
-		    OUT_RING  (chan, sub_h << 16 | sub_w);
-		    OUT_RING  (chan, 0);
-		    OUT_RING  (chan, sub_h << 16 | sub_w);
-		    OUT_RING  (chan, 1 << 20);
-		    OUT_RING  (chan, 1 << 20);
-
-		    BEGIN_RING(chan, sifm, NV04_SCALED_IMAGE_FROM_MEMORY_SIZE, 4);
-		    OUT_RING  (chan, sub_h << 16 | sub_w);
-		    OUT_RING  (chan, src->stride |
-		                     NV04_SCALED_IMAGE_FROM_MEMORY_FORMAT_ORIGIN_CENTER |
-		                     NV04_SCALED_IMAGE_FROM_MEMORY_FORMAT_FILTER_POINT_SAMPLE);
-		    OUT_RELOCl(chan, src_bo, src->offset + cy * src->stride +
-				     cx * src->block.size + offset, NOUVEAU_BO_GART |
-				     NOUVEAU_BO_VRAM | NOUVEAU_BO_RD);
-		    OUT_RING  (chan, 0);
-		  }
-		}
-
-		/* FIXME: need to know how many bytes per pixel */
-		offset += align(w * h * 2 /*src->block.size*/, 64);
-		w >>= 1;
-		h >>= 1;
+	for (cy = 0; cy < h; cy += sub_h) {
+	  for (cx = 0; cx < w; cx += sub_w) {
+	    BEGIN_RING(chan, swzsurf, NV04_SWIZZLED_SURFACE_OFFSET, 1);
+	    OUT_RELOCl(chan, dst_bo, dst->offset + nv04_swizzle_bits(cx, cy) *
+			     dst->block.size, NOUVEAU_BO_GART |
+			     NOUVEAU_BO_VRAM | NOUVEAU_BO_WR);
+
+	    BEGIN_RING(chan, sifm, NV04_SCALED_IMAGE_FROM_MEMORY_COLOR_CONVERSION, 9);
+	    OUT_RING  (chan, NV04_SCALED_IMAGE_FROM_MEMORY_COLOR_CONVERSION_TRUNCATE);
+	    OUT_RING  (chan, nv04_scaled_image_format(src->format));
+	    OUT_RING  (chan, NV04_SCALED_IMAGE_FROM_MEMORY_OPERATION_SRCCOPY);
+	    OUT_RING  (chan, 0);
+	    OUT_RING  (chan, sub_h << 16 | sub_w);
+	    OUT_RING  (chan, 0);
+	    OUT_RING  (chan, sub_h << 16 | sub_w);
+	    OUT_RING  (chan, 1 << 20);
+	    OUT_RING  (chan, 1 << 20);
+
+	    BEGIN_RING(chan, sifm, NV04_SCALED_IMAGE_FROM_MEMORY_SIZE, 4);
+	    OUT_RING  (chan, sub_h << 16 | sub_w);
+	    OUT_RING  (chan, src->stride |
+			     NV04_SCALED_IMAGE_FROM_MEMORY_FORMAT_ORIGIN_CENTER |
+			     NV04_SCALED_IMAGE_FROM_MEMORY_FORMAT_FILTER_POINT_SAMPLE);
+	    OUT_RELOCl(chan, src_bo, src->offset + cy * src->stride +
+			     cx * src->block.size, NOUVEAU_BO_GART |
+			     NOUVEAU_BO_VRAM | NOUVEAU_BO_RD);
+	    OUT_RING  (chan, 0);
+	  }
 	}
 
 	return 0;
@@ -258,8 +248,7 @@ nv04_surface_copy(struct nv04_surface_2d *ctx, struct pipe_surface *dst,
 	assert(src->format == dst->format);
 
 	/* Setup transfer to swizzle the texture to vram if needed */
-	/* FIXME/TODO: check proper limits of this operation */
-	if (src_linear ^ dst_linear) {
+	if (src_linear && !dst_linear && w > 1 && h > 1) {
 		nv04_surface_copy_swizzle(ctx, dst, dx, dy, src, sx, sy, w, h);
 		return;
 	}
@@ -267,7 +256,8 @@ nv04_surface_copy(struct nv04_surface_2d *ctx, struct pipe_surface *dst,
 	/* NV_CONTEXT_SURFACES_2D has buffer alignment restrictions, fallback
 	 * to NV_MEMORY_TO_MEMORY_FORMAT in this case.
 	 */
-	if ((src->offset & 63) || (dst->offset & 63)) {
+	if ((src->offset & 63) || (dst->offset & 63) ||
+	    (src->stride & 63) || (dst->stride & 63)) {
 		nv04_surface_copy_m2mf(ctx, dst, dx, dy, src, sx, sy, w, h);
 		return;
 	}
@@ -456,5 +446,3 @@ nv04_surface_2d_init(struct nouveau_winsys *nvws)
 	ctx->fill = nv04_surface_fill;
 	return ctx;
 }
-
-
diff --git a/src/gallium/drivers/nv20/nv20_vertprog.c b/src/gallium/drivers/nv20/nv20_vertprog.c
index a885fcd7a5..5db0e807ff 100644
--- a/src/gallium/drivers/nv20/nv20_vertprog.c
+++ b/src/gallium/drivers/nv20/nv20_vertprog.c
@@ -613,7 +613,7 @@ nv20_vertprog_translate(struct nv20_context *nv20,
 
 			imm = &parse.FullToken.FullImmediate;
 			assert(imm->Immediate.DataType == TGSI_IMM_FLOAT32);
-//			assert(imm->Immediate.Size == 4);
+			assert(imm->Immediate.NrTokens == 4 + 1);
 			vpc->imm[vpc->nr_imm++] =
 				constant(vpc, -1,
 					 imm->u.ImmediateFloat32[0].Float,
diff --git a/src/gallium/drivers/nv30/nv30_miptree.c b/src/gallium/drivers/nv30/nv30_miptree.c
index 23f8829321..fe13f50ebb 100644
--- a/src/gallium/drivers/nv30/nv30_miptree.c
+++ b/src/gallium/drivers/nv30/nv30_miptree.c
@@ -8,10 +8,14 @@ static void
 nv30_miptree_layout(struct nv30_miptree *nv30mt)
 {
 	struct pipe_texture *pt = &nv30mt->base;
-	boolean swizzled = FALSE;
 	uint width = pt->width[0], height = pt->height[0], depth = pt->depth[0];
 	uint offset = 0;
-	int nr_faces, l, f, pitch;
+	int nr_faces, l, f;
+	uint wide_pitch = pt->tex_usage & (PIPE_TEXTURE_USAGE_SAMPLER |
+		                           PIPE_TEXTURE_USAGE_DEPTH_STENCIL |
+		                           PIPE_TEXTURE_USAGE_RENDER_TARGET |
+		                           PIPE_TEXTURE_USAGE_DISPLAY_TARGET |
+		                           PIPE_TEXTURE_USAGE_PRIMARY);
 
 	if (pt->target == PIPE_TEXTURE_CUBE) {
 		nr_faces = 6;
@@ -22,7 +26,6 @@ nv30_miptree_layout(struct nv30_miptree *nv30mt)
 		nr_faces = 1;
 	}
 
-	pitch = pt->width[0];
 	for (l = 0; l <= pt->last_level; l++) {
 		pt->width[l] = width;
 		pt->height[l] = height;
@@ -30,11 +33,11 @@ nv30_miptree_layout(struct nv30_miptree *nv30mt)
 		pt->nblocksx[l] = pf_get_nblocksx(&pt->block, width);
 		pt->nblocksy[l] = pf_get_nblocksy(&pt->block, height);
 
-		if (swizzled)
-			pitch = pt->nblocksx[l];
-		pitch = align(pitch, 64);
+		if (wide_pitch && (pt->tex_usage & NOUVEAU_TEXTURE_USAGE_LINEAR))
+			nv30mt->level[l].pitch = align(pt->width[0] * pt->block.size, 64);
+		else
+			nv30mt->level[l].pitch = pt->width[l] * pt->block.size;
 
-		nv30mt->level[l].pitch = pitch * pt->block.size;
 		nv30mt->level[l].image_offset =
 			CALLOC(nr_faces, sizeof(unsigned));
 
@@ -44,10 +47,17 @@ nv30_miptree_layout(struct nv30_miptree *nv30mt)
 	}
 
 	for (f = 0; f < nr_faces; f++) {
-		for (l = 0; l <= pt->last_level; l++) {
+		for (l = 0; l < pt->last_level; l++) {
 			nv30mt->level[l].image_offset[f] = offset;
-			offset += nv30mt->level[l].pitch * pt->height[l];
+
+			if (!(pt->tex_usage & NOUVEAU_TEXTURE_USAGE_LINEAR))
+				offset += align(nv30mt->level[l].pitch * pt->height[l], 64);
+			else
+				offset += nv30mt->level[l].pitch * pt->height[l];
 		}
+
+		nv30mt->level[l].image_offset[f] = offset;
+		offset += nv30mt->level[l].pitch * pt->height[l];
 	}
 
 	nv30mt->total_size = offset;
@@ -74,7 +84,8 @@ nv30_miptree_create(struct pipe_screen *pscreen, const struct pipe_texture *pt)
 		mt->base.tex_usage |= NOUVEAU_TEXTURE_USAGE_LINEAR;
 	else
 	if (pt->tex_usage & (PIPE_TEXTURE_USAGE_PRIMARY |
-	                     PIPE_TEXTURE_USAGE_DISPLAY_TARGET))
++	                     PIPE_TEXTURE_USAGE_DISPLAY_TARGET |
++	                     PIPE_TEXTURE_USAGE_DEPTH_STENCIL))
 		mt->base.tex_usage |= NOUVEAU_TEXTURE_USAGE_LINEAR;
 	else
 	if (pt->tex_usage & PIPE_TEXTURE_USAGE_DYNAMIC)
@@ -85,7 +96,11 @@ nv30_miptree_create(struct pipe_screen *pscreen, const struct pipe_texture *pt)
 		case PIPE_FORMAT_A8R8G8B8_UNORM:
 		case PIPE_FORMAT_X8R8G8B8_UNORM:
 		case PIPE_FORMAT_R16_SNORM:
-			break;
+		{
+			if (debug_get_bool_option("NOUVEAU_NO_SWIZZLE", FALSE))
+				mt->base.tex_usage |= NOUVEAU_TEXTURE_USAGE_LINEAR;
+ 			break;
+		}
 		default:
 			mt->base.tex_usage |= NOUVEAU_TEXTURE_USAGE_LINEAR;
 		}
@@ -148,8 +163,8 @@ nv30_miptree_release(struct pipe_screen *pscreen, struct pipe_texture **ppt)
 	}
 
 	if (mt->shadow_tex) {
-		assert(mt->shadow_surface);
-		pscreen->tex_surface_release(pscreen, &mt->shadow_surface);
+		if (mt->shadow_surface)
+			pscreen->tex_surface_release(pscreen, &mt->shadow_surface);
 		nv30_miptree_release(pscreen, &mt->shadow_tex);
 	}
 
diff --git a/src/gallium/drivers/nv30/nv30_screen.c b/src/gallium/drivers/nv30/nv30_screen.c
index e3c9b42044..c97a73f0b1 100644
--- a/src/gallium/drivers/nv30/nv30_screen.c
+++ b/src/gallium/drivers/nv30/nv30_screen.c
@@ -206,6 +206,8 @@ nv30_surface_unmap(struct pipe_screen *screen, struct pipe_surface *surface)
 		nvscreen->eng2d->copy(nvscreen->eng2d, surface, 0, 0,
 		                      surface_to_unmap, 0, 0,
 		                      surface->width, surface->height);
+
+		screen->tex_surface_release(screen, &surface_to_unmap);
 	}
 }
 
diff --git a/src/gallium/drivers/nv30/nv30_vertprog.c b/src/gallium/drivers/nv30/nv30_vertprog.c
index 72824559e8..d262725057 100644
--- a/src/gallium/drivers/nv30/nv30_vertprog.c
+++ b/src/gallium/drivers/nv30/nv30_vertprog.c
@@ -613,7 +613,7 @@ nv30_vertprog_translate(struct nv30_context *nv30,
 
 			imm = &parse.FullToken.FullImmediate;
 			assert(imm->Immediate.DataType == TGSI_IMM_FLOAT32);
-//			assert(imm->Immediate.Size == 4);
+			assert(imm->Immediate.NrTokens == 4 + 1);
 			vpc->imm[vpc->nr_imm++] =
 				constant(vpc, -1,
 					 imm->u.ImmediateFloat32[0].Float,
diff --git a/src/gallium/drivers/nv40/nv40_miptree.c b/src/gallium/drivers/nv40/nv40_miptree.c
index 9bef23ad1f..e4f8df910a 100644
--- a/src/gallium/drivers/nv40/nv40_miptree.c
+++ b/src/gallium/drivers/nv40/nv40_miptree.c
@@ -10,7 +10,12 @@ nv40_miptree_layout(struct nv40_miptree *mt)
 	struct pipe_texture *pt = &mt->base;
 	uint width = pt->width[0], height = pt->height[0], depth = pt->depth[0];
 	uint offset = 0;
-	int nr_faces, l, f, pitch;
+	int nr_faces, l, f;
+	uint wide_pitch = pt->tex_usage & (PIPE_TEXTURE_USAGE_SAMPLER |
+		                           PIPE_TEXTURE_USAGE_DEPTH_STENCIL |
+		                           PIPE_TEXTURE_USAGE_RENDER_TARGET |
+		                           PIPE_TEXTURE_USAGE_DISPLAY_TARGET |
+		                           PIPE_TEXTURE_USAGE_PRIMARY);
 
 	if (pt->target == PIPE_TEXTURE_CUBE) {
 		nr_faces = 6;
@@ -21,7 +26,6 @@ nv40_miptree_layout(struct nv40_miptree *mt)
 		nr_faces = 1;
 	}
 
-	pitch = pt->width[0];
 	for (l = 0; l <= pt->last_level; l++) {
 		pt->width[l] = width;
 		pt->height[l] = height;
@@ -29,11 +33,11 @@ nv40_miptree_layout(struct nv40_miptree *mt)
 		pt->nblocksx[l] = pf_get_nblocksx(&pt->block, width);
 		pt->nblocksy[l] = pf_get_nblocksy(&pt->block, height);
 
-		if (!(pt->tex_usage & NOUVEAU_TEXTURE_USAGE_LINEAR))
-			pitch = pt->nblocksx[l];
-		pitch = align(pitch, 64);
+		if (wide_pitch && (pt->tex_usage & NOUVEAU_TEXTURE_USAGE_LINEAR))
+			mt->level[l].pitch = align(pt->width[0] * pt->block.size, 64);
+		else
+			mt->level[l].pitch = pt->width[l] * pt->block.size;
 
-		mt->level[l].pitch = pitch * pt->block.size;
 		mt->level[l].image_offset =
 			CALLOC(nr_faces, sizeof(unsigned));
 
@@ -43,10 +47,17 @@ nv40_miptree_layout(struct nv40_miptree *mt)
 	}
 
 	for (f = 0; f < nr_faces; f++) {
-		for (l = 0; l <= pt->last_level; l++) {
+		for (l = 0; l < pt->last_level; l++) {
 			mt->level[l].image_offset[f] = offset;
-			offset += mt->level[l].pitch * pt->height[l];
+
+			if (!(pt->tex_usage & NOUVEAU_TEXTURE_USAGE_LINEAR))
+				offset += align(mt->level[l].pitch * pt->height[l], 64);
+			else
+				offset += mt->level[l].pitch * pt->height[l];
 		}
+
+		mt->level[l].image_offset[f] = offset;
+		offset += mt->level[l].pitch * pt->height[l];
 	}
 
 	mt->total_size = offset;
@@ -75,7 +86,8 @@ nv40_miptree_create(struct pipe_screen *pscreen, const struct pipe_texture *pt)
 		mt->base.tex_usage |= NOUVEAU_TEXTURE_USAGE_LINEAR;
 	else
 	if (pt->tex_usage & (PIPE_TEXTURE_USAGE_PRIMARY |
-	                     PIPE_TEXTURE_USAGE_DISPLAY_TARGET))
+	                     PIPE_TEXTURE_USAGE_DISPLAY_TARGET |
+	                     PIPE_TEXTURE_USAGE_DEPTH_STENCIL))
 		mt->base.tex_usage |= NOUVEAU_TEXTURE_USAGE_LINEAR;
 	else
 	if (pt->tex_usage & PIPE_TEXTURE_USAGE_DYNAMIC)
@@ -86,7 +98,11 @@ nv40_miptree_create(struct pipe_screen *pscreen, const struct pipe_texture *pt)
 		case PIPE_FORMAT_A8R8G8B8_UNORM:
 		case PIPE_FORMAT_X8R8G8B8_UNORM:
 		case PIPE_FORMAT_R16_SNORM:
+		{
+			if (debug_get_bool_option("NOUVEAU_NO_SWIZZLE", FALSE))
+				mt->base.tex_usage |= NOUVEAU_TEXTURE_USAGE_LINEAR;
 			break;
+		}
 		default:
 			mt->base.tex_usage |= NOUVEAU_TEXTURE_USAGE_LINEAR;
 		}
@@ -149,8 +165,8 @@ nv40_miptree_release(struct pipe_screen *pscreen, struct pipe_texture **ppt)
 	}
 
 	if (mt->shadow_tex) {
-		assert(mt->shadow_surface);
-		pscreen->tex_surface_release(pscreen, &mt->shadow_surface);
+		if (mt->shadow_surface)
+			pscreen->tex_surface_release(pscreen, &mt->shadow_surface);
 		nv40_miptree_release(pscreen, &mt->shadow_tex);
 	}
 
diff --git a/src/gallium/drivers/nv40/nv40_screen.c b/src/gallium/drivers/nv40/nv40_screen.c
index 88a329ea24..2372bc8441 100644
--- a/src/gallium/drivers/nv40/nv40_screen.c
+++ b/src/gallium/drivers/nv40/nv40_screen.c
@@ -214,6 +214,8 @@ nv40_surface_unmap(struct pipe_screen *screen, struct pipe_surface *surface)
 		nvscreen->eng2d->copy(nvscreen->eng2d, surface, 0, 0,
 		                      surface_to_unmap, 0, 0,
 		                      surface->width, surface->height);
+
+		screen->tex_surface_release(screen, &surface_to_unmap);
 	}
 }
 
diff --git a/src/gallium/drivers/nv40/nv40_vertprog.c b/src/gallium/drivers/nv40/nv40_vertprog.c
index 1392fe956f..0862386638 100644
--- a/src/gallium/drivers/nv40/nv40_vertprog.c
+++ b/src/gallium/drivers/nv40/nv40_vertprog.c
@@ -784,7 +784,7 @@ nv40_vertprog_translate(struct nv40_context *nv40,
 
 			imm = &parse.FullToken.FullImmediate;
 			assert(imm->Immediate.DataType == TGSI_IMM_FLOAT32);
-//			assert(imm->Immediate.Size == 4);
+			assert(imm->Immediate.NrTokens == 4 + 1);
 			vpc->imm[vpc->nr_imm++] =
 				constant(vpc, -1,
 					 imm->u.ImmediateFloat32[0].Float,
diff --git a/src/gallium/include/pipe/p_shader_tokens.h b/src/gallium/include/pipe/p_shader_tokens.h
index ad43d799f3..35df70e7b7 100644
--- a/src/gallium/include/pipe/p_shader_tokens.h
+++ b/src/gallium/include/pipe/p_shader_tokens.h
@@ -64,7 +64,7 @@ struct tgsi_processor
 struct tgsi_token
 {
    unsigned Type       : 4;  /**< TGSI_TOKEN_TYPE_x */
-   unsigned Size       : 8;  /**< UINT */
+   unsigned NrTokens   : 8;  /**< UINT */
    unsigned Padding    : 19;
    unsigned Extended   : 1;  /**< BOOL */
 };
@@ -107,7 +107,7 @@ enum tgsi_file_type {
 struct tgsi_declaration
 {
    unsigned Type        : 4;  /**< TGSI_TOKEN_TYPE_DECLARATION */
-   unsigned Size        : 8;  /**< UINT */
+   unsigned NrTokens    : 8;  /**< UINT */
    unsigned File        : 4;  /**< one of TGSI_FILE_x */
    unsigned UsageMask   : 4;  /**< bitmask of TGSI_WRITEMASK_x flags */
    unsigned Interpolate : 4;  /**< one of TGSI_INTERPOLATE_x */
@@ -145,7 +145,7 @@ struct tgsi_declaration_semantic
 struct tgsi_immediate
 {
    unsigned Type       : 4;  /**< TGSI_TOKEN_TYPE_IMMEDIATE */
-   unsigned Size       : 8;  /**< UINT */
+   unsigned NrTokens   : 8;  /**< UINT */
    unsigned DataType   : 4;  /**< one of TGSI_IMM_x */
    unsigned Padding    : 15;
    unsigned Extended   : 1;  /**< BOOL */
@@ -442,7 +442,7 @@ struct tgsi_immediate_float32
 struct tgsi_instruction
 {
    unsigned Type       : 4;  /* TGSI_TOKEN_TYPE_INSTRUCTION */
-   unsigned Size       : 8;  /* UINT */
+   unsigned NrTokens   : 8;  /* UINT */
    unsigned Opcode     : 8;  /* TGSI_OPCODE_ */
    unsigned Saturate   : 2;  /* TGSI_SAT_ */
    unsigned NumDstRegs : 2;  /* UINT */
@@ -458,7 +458,7 @@ struct tgsi_instruction
  * 
  * Then, tgsi_instruction::NumSrcRegs of tgsi_src_register follow.
  *
- * tgsi_instruction::Size contains the total number of words that make the
+ * tgsi_instruction::NrTokens contains the total number of words that make the
  * instruction, including the instruction word.
  */
 
diff --git a/src/gallium/state_trackers/wgl/icd/stw_icd.c b/src/gallium/state_trackers/wgl/icd/stw_icd.c
index 70e346a539..1aa4b8a6e2 100644
--- a/src/gallium/state_trackers/wgl/icd/stw_icd.c
+++ b/src/gallium/state_trackers/wgl/icd/stw_icd.c
@@ -31,6 +31,7 @@
 #include "GL/gl.h"
 
 #include "pipe/p_debug.h"
+#include "pipe/p_thread.h"
 
 #include "shared/stw_public.h"
 #include "icd/stw_icd.h"
@@ -41,11 +42,14 @@
 
 struct stw_icd
 {
+   pipe_mutex mutex;
+
+   GLCLTPROCTABLE cpt;
+   boolean cpt_initialized;
+
    struct {
       struct stw_context *ctx;
    } ctx_array[DRV_CONTEXT_MAX];
-
-   DHGLRC ctx_current;
 };
 
 
@@ -62,6 +66,8 @@ stw_icd_init( void )
    stw_icd = &stw_icd_storage;
    memset(stw_icd, 0, sizeof *stw_icd);
 
+   pipe_mutex_init( stw_icd->mutex );
+
    return TRUE;
 }
 
@@ -70,26 +76,35 @@ stw_icd_cleanup(void)
 {
    int i;
 
-   if(!stw_icd)
+   if (!stw_icd)
       return;
+   
+   pipe_mutex_lock( stw_icd->mutex );
+   {
+      /* Ensure all contexts are destroyed */
+      for (i = 0; i < DRV_CONTEXT_MAX; i++)
+         if (stw_icd->ctx_array[i].ctx) 
+            stw_delete_context( stw_icd->ctx_array[i].ctx );
+   }
+   pipe_mutex_unlock( stw_icd->mutex );
 
-   /* Ensure all contexts are destroyed */
-   for (i = 0; i < DRV_CONTEXT_MAX; i++)
-      if (stw_icd->ctx_array[i].ctx) 
-         stw_delete_context( stw_icd->ctx_array[i].ctx );
-
+   pipe_mutex_init( stw_icd->mutex );
    stw_icd = NULL;
 }
 
 
 static struct stw_context *
-lookup_context( DHGLRC dhglrc )
+lookup_context( struct stw_icd *icd, 
+                DHGLRC dhglrc )
 {
    if (dhglrc == 0 || 
        dhglrc >= DRV_CONTEXT_MAX)
       return NULL;
 
-   return stw_icd->ctx_array[dhglrc - 1].ctx;
+   if (icd == NULL)
+      return NULL;
+
+   return icd->ctx_array[dhglrc - 1].ctx;
 }
 
 BOOL APIENTRY
@@ -98,14 +113,25 @@ DrvCopyContext(
    DHGLRC dhrcDest,
    UINT fuMask )
 {
-   struct stw_context *src = lookup_context( dhrcSource );
-   struct stw_context *dst = lookup_context( dhrcDest );
-   
-   if (src == NULL ||
-       dst == NULL)
+   BOOL ret = FALSE;
+
+   if (!stw_icd)
       return FALSE;
 
-   return stw_copy_context( src, dst, fuMask );
+   pipe_mutex_lock( stw_icd->mutex );
+   {
+      struct stw_context *src = lookup_context( stw_icd, dhrcSource );
+      struct stw_context *dst = lookup_context( stw_icd, dhrcDest );
+   
+      if (src == NULL || dst == NULL) 
+         goto done;
+
+      ret = stw_copy_context( src, dst, fuMask );
+   }
+done:
+   pipe_mutex_unlock( stw_icd->mutex );
+
+   return ret;
 }
 
 DHGLRC APIENTRY
@@ -113,23 +139,37 @@ DrvCreateLayerContext(
    HDC hdc,
    INT iLayerPlane )
 {
-   DWORD i;
+   DHGLRC handle = 0;
+
+   if (!stw_icd)
+      return handle;
+
+   pipe_mutex_lock( stw_icd->mutex );
+   {
+      int i;
+
+      for (i = 0; i < DRV_CONTEXT_MAX; i++) {
+         if (stw_icd->ctx_array[i].ctx == NULL)
+            break;
+      }
    
-   for (i = 0; i < DRV_CONTEXT_MAX; i++) {
-      if (stw_icd->ctx_array[i].ctx == NULL)
-         goto found_slot;
+      /* No slot available, fail:
+       */
+      if (i == DRV_CONTEXT_MAX)
+         goto done;
+
+      stw_icd->ctx_array[i].ctx = stw_create_context( hdc, iLayerPlane );
+      if (stw_icd->ctx_array[i].ctx == NULL) 
+         goto done;
+      
+      /* success:
+       */
+      handle = (DHGLRC) i + 1;
    }
-   
-   /* No slot available, fail:
-    */
-   return 0;
+done:
+   pipe_mutex_unlock( stw_icd->mutex );
 
-found_slot:
-   stw_icd->ctx_array[i].ctx = stw_create_context( hdc, iLayerPlane );
-   if (stw_icd->ctx_array[i].ctx == NULL)
-      return 0;
-
-   return (DHGLRC) i + 1;
+   return handle;
 }
 
 DHGLRC APIENTRY
@@ -143,20 +183,30 @@ BOOL APIENTRY
 DrvDeleteContext(
    DHGLRC dhglrc )
 {
-   struct stw_context *ctx;
+   BOOL ret = FALSE;
+
+   if (!stw_icd)
+      return ret;
 
-   ctx = lookup_context( dhglrc );
-   if (ctx == NULL) 
-      goto fail;
+   pipe_mutex_lock( stw_icd->mutex );
+   {
+      struct stw_context *ctx;
 
-   if (stw_delete_context( ctx ) == FALSE)
-      goto fail;
+      ctx = lookup_context( stw_icd, dhglrc );
+      if (ctx == NULL) 
+         goto done;
+      
+      if (stw_delete_context( ctx ) == FALSE)
+         goto done;
+      
+      stw_icd->ctx_array[dhglrc - 1].ctx = NULL;
+      ret = TRUE;
 
-   stw_icd->ctx_array[dhglrc - 1].ctx = NULL;
-   return TRUE;
+   }
+done:
+   pipe_mutex_unlock( stw_icd->mutex );
    
-fail:
-   return FALSE;
+   return ret;
 }
 
 BOOL APIENTRY
@@ -183,7 +233,7 @@ DrvDescribePixelFormat(
 
    r = stw_pixelformat_describe( hdc, iPixelFormat, cjpfd, ppfd );
 
-   debug_printf( "%s( 0x%p, %d, %u, 0x%p ) = %d\n",
+   debug_printf( "%s( %p, %d, %u, %p ) = %d\n",
                  __FUNCTION__, hdc, iPixelFormat, cjpfd, ppfd, r );
 
    return r;
@@ -210,7 +260,7 @@ DrvGetProcAddress(
 
    r = stw_get_proc_address( lpszProc );
 
-   debug_printf( "%s( \", __FUNCTION__%s\" ) = 0x%p\n", lpszProc, r );
+   debug_printf( "%s( \", __FUNCTION__%s\" ) = %p\n", lpszProc, r );
 
    return r;
 }
@@ -230,23 +280,32 @@ BOOL APIENTRY
 DrvReleaseContext(
    DHGLRC dhglrc )
 {
-   struct stw_context *ctx;
+   BOOL ret = FALSE;
 
-   if (dhglrc != stw_icd->ctx_current) 
-      goto fail;
+   if (!stw_icd)
+      return ret;
 
-   ctx = lookup_context( dhglrc );
-   if (ctx == NULL) 
-      goto fail;
+   pipe_mutex_lock( stw_icd->mutex );
+   {
+      struct stw_context *ctx;
 
-   if (stw_make_current( NULL, NULL ) == FALSE)
-      goto fail;
+      /* XXX: The expectation is that ctx is the same context which is
+       * current for this thread.  We should check that and return False
+       * if not the case.
+       */
+      ctx = lookup_context( stw_icd, dhglrc );
+      if (ctx == NULL) 
+         goto done;
 
-   stw_icd->ctx_current = 0;
-   return TRUE;
+      if (stw_make_current( NULL, NULL ) == FALSE)
+         goto done;
 
-fail:
-   return FALSE;
+      ret = TRUE;
+   }
+done:
+   pipe_mutex_unlock( stw_icd->mutex );
+
+   return ret;
 }
 
 void APIENTRY
@@ -254,36 +313,20 @@ DrvSetCallbackProcs(
    INT nProcs,
    PROC *pProcs )
 {
-   debug_printf( "%s( %d, 0x%p )\n", __FUNCTION__, nProcs, pProcs );
+   debug_printf( "%s( %d, %p )\n", __FUNCTION__, nProcs, pProcs );
 
    return;
 }
 
-#define GPA_GL( NAME ) disp->NAME = gl##NAME
 
-static GLCLTPROCTABLE cpt;
+static void init_proc_table( GLCLTPROCTABLE *cpt )
+{ 
+   GLDISPATCHTABLE *disp = &cpt->glDispatchTable;
 
-PGLCLTPROCTABLE APIENTRY
-DrvSetContext(
-   HDC hdc,
-   DHGLRC dhglrc,
-   PFN_SETPROCTABLE pfnSetProcTable )
-{
-   struct stw_context *ctx;
-   GLDISPATCHTABLE *disp = &cpt.glDispatchTable;
-
-   debug_printf( "%s( 0x%p, %u, 0x%p )\n", __FUNCTION__, hdc, dhglrc, pfnSetProcTable );
-
-   ctx = lookup_context( dhglrc );
-   if (ctx == NULL)
-      return NULL;
-
-   if (!stw_make_current( hdc, ctx ))
-      return NULL;
-
-   memset( &cpt, 0, sizeof( cpt ) );
-   cpt.cEntries = OPENGL_VERSION_110_ENTRIES;
+   memset( cpt, 0, sizeof *cpt );
+   cpt->cEntries = OPENGL_VERSION_110_ENTRIES;
 
+#define GPA_GL( NAME ) disp->NAME = gl##NAME
    GPA_GL( NewList );
    GPA_GL( EndList );
    GPA_GL( CallList );
@@ -620,8 +663,46 @@ DrvSetContext(
    GPA_GL( TexSubImage2D );
    GPA_GL( PopClientAttrib );
    GPA_GL( PushClientAttrib );
+}
 
-   return &cpt;
+PGLCLTPROCTABLE APIENTRY
+DrvSetContext(
+   HDC hdc,
+   DHGLRC dhglrc,
+   PFN_SETPROCTABLE pfnSetProcTable )
+{
+   PGLCLTPROCTABLE result = NULL;
+
+   if (!stw_icd)
+      return result;
+
+   pipe_mutex_lock( stw_icd->mutex ); 
+   {
+      struct stw_context *ctx;
+
+      debug_printf( "%s( 0x%p, %u, 0x%p )\n", 
+                    __FUNCTION__, hdc, dhglrc, pfnSetProcTable );
+
+      /* Although WGL allows different dispatch entrypoints per 
+       */
+      if (!stw_icd->cpt_initialized) {
+         init_proc_table( &stw_icd->cpt );
+         stw_icd->cpt_initialized = TRUE;
+      }
+
+      ctx = lookup_context( stw_icd, dhglrc );
+      if (ctx == NULL)
+         goto done;
+
+      if (!stw_make_current( hdc, ctx ))
+         goto done;
+
+      result = &stw_icd->cpt;
+   }
+done:
+   pipe_mutex_unlock( stw_icd->mutex );
+
+   return result;
 }
 
 int APIENTRY
@@ -646,7 +727,7 @@ DrvSetPixelFormat(
 
    r = stw_pixelformat_set( hdc, iPixelFormat );
 
-   debug_printf( "%s( 0x%p, %d ) = %s\n", __FUNCTION__, hdc, iPixelFormat, r ? "TRUE" : "FALSE" );
+   debug_printf( "%s( %p, %d ) = %s\n", __FUNCTION__, hdc, iPixelFormat, r ? "TRUE" : "FALSE" );
 
    return r;
 }
@@ -665,7 +746,7 @@ BOOL APIENTRY
 DrvSwapBuffers(
    HDC hdc )
 {
-   debug_printf( "%s( 0x%p )\n", __FUNCTION__, hdc );
+   debug_printf( "%s( %p )\n", __FUNCTION__, hdc );
 
    return stw_swap_buffers( hdc );
 }
@@ -686,5 +767,7 @@ DrvValidateVersion(
 {
    debug_printf( "%s( %u )\n", __FUNCTION__, ulVersion );
 
+   /* TODO: get the expected version from the winsys */
+   
    return ulVersion == 1;
 }
diff --git a/src/gallium/state_trackers/wgl/shared/stw_pixelformat.c b/src/gallium/state_trackers/wgl/shared/stw_pixelformat.c
index 12b5ac6d91..84b7b287b9 100644
--- a/src/gallium/state_trackers/wgl/shared/stw_pixelformat.c
+++ b/src/gallium/state_trackers/wgl/shared/stw_pixelformat.c
@@ -256,6 +256,14 @@ stw_pixelformat_set(
       return FALSE;
 
    currentpixelformat = iPixelFormat;
+   
+   /* Some applications mistakenly use the undocumented wglSetPixelFormat 
+    * function instead of SetPixelFormat, so we call SetPixelFormat here to 
+    * avoid opengl32.dll's wglCreateContext to fail */
+   if (GetPixelFormat(hdc) == 0) {
+        SetPixelFormat(hdc, iPixelFormat, NULL);
+   }
+   
    return TRUE;
 }
 
diff --git a/src/glut/glx/SConscript b/src/glut/glx/SConscript
index cb442ba027..938fec03df 100644
--- a/src/glut/glx/SConscript
+++ b/src/glut/glx/SConscript
@@ -10,7 +10,6 @@ target = 'glut32'
 env.Replace(CPPDEFINES = [
     'BUILD_GLUT32', 
     'GLUT_BUILDING_LIB', 
-    'MESA', 
     'NDEBUG',
     'GLUT_NO_WARNING_DISABLE',
 ])
diff --git a/src/glut/glx/glut.def b/src/glut/glx/glut.def
index d5e0e0cf2a..90010815e4 100644
--- a/src/glut/glx/glut.def
+++ b/src/glut/glx/glut.def
@@ -1,9 +1,4 @@
-DESCRIPTION 'OpenGL Utility Toolkit for Win32'
-
-VERSION 3.7
-
 EXPORTS
-
 	glutAddMenuEntry
 	glutAddSubMenu
 	glutAttachMenu
@@ -128,4 +123,3 @@ EXPORTS
 	glutBitmapTimesRoman24
 ;    __glutSetFCB
 ;    __glutGetFCB
-
diff --git a/src/glut/glx/glutint.h b/src/glut/glx/glutint.h
index fe4e038add..ba30af60f0 100644
--- a/src/glut/glx/glutint.h
+++ b/src/glut/glx/glutint.h
@@ -32,10 +32,6 @@
 
 #include <GL/glut.h>
 
-#if defined(MESA) && defined(_WIN32) && !defined(__CYGWIN32__)
-#include <GL/mesa_wgl.h>
-#endif
-
 #ifndef _WIN32
 /* added by BrianP: */
 #ifndef APIENTRY
@@ -48,24 +44,6 @@
 /* GLUT_BUILDING_LIB is used by <GL/glut.h> to 1) not #pragma link
    with the GLUT library, and 2) avoid the Win32 atexit hack. */
 
-/* This must be done after <GL/gl.h> is included.  MESA is defined
-   if the <GL/gl.h> is supplied by Brian Paul's Mesa library. */
-#if defined(MESA) && defined(_WIN32)
-/* Mesa implements "wgl" versions of GDI entry points needed for
-   using OpenGL.  Map these "wgl" versions to the GDI names via
-   macros. */
-#define ChoosePixelFormat   wglChoosePixelFormat
-#define DescribePixelFormat wglDescribePixelFormat
-#define GetPixelFormat      wglGetPixelFormat
-#define SetPixelFormat      wglSetPixelFormat
-#define SwapBuffers         wglSwapBuffers
-#define GetCurrentContext   wglGetCurrentContext
-#define GetCurrentDC        wglGetCurrentDC
-#define MakeCurrent         wglMakeCurrent
-#define CreateContext       wglCreateContext
-#define DeleteContext       wglDeleteContext
-#endif /* MESA */
-
 #ifdef SUPPORT_FORTRAN
 #include <GL/glutf90.h>
 #endif
@@ -572,27 +550,27 @@ typedef struct {
 #ifdef _WIN32
 #define MAKE_CURRENT_LAYER(window)                                    \
   {                                                                   \
-    HGLRC currentContext = GetCurrentContext();                       \
-    HDC currentDc = GetCurrentDC();                                   \
+    HGLRC currentContext = wglGetCurrentContext();                    \
+    HDC currentDc = wglGetCurrentDC();                                \
                                                                       \
     if (currentContext != window->renderCtx                           \
       || currentDc != window->renderDc) {                             \
-      MakeCurrent(window->renderDc, window->renderCtx);               \
+      wglMakeCurrent(window->renderDc, window->renderCtx);            \
     }                                                                 \
   }
 #define MAKE_CURRENT_WINDOW(window)                                   \
   {                                                                   \
-    HGLRC currentContext = GetCurrentContext();                       \
-    HDC currentDc = GetCurrentDC();                                   \
+    HGLRC currentContext = wglGetCurrentContext();                    \
+    HDC currentDc = wglGetCurrentDC();                                \
                                                                       \
     if (currentContext != window->ctx || currentDc != window->hdc) {  \
-      MakeCurrent(window->hdc, window->ctx);                          \
+      wglMakeCurrent(window->hdc, window->ctx);                       \
     }                                                                 \
   }
 #define MAKE_CURRENT_OVERLAY(overlay) \
-  MakeCurrent(overlay->hdc, overlay->ctx)
+  wglMakeCurrent(overlay->hdc, overlay->ctx)
 #define UNMAKE_CURRENT() \
-  MakeCurrent(NULL, NULL)
+  wglMakeCurrent(NULL, NULL)
 #define SWAP_BUFFERS_WINDOW(window) \
   SwapBuffers(window->hdc)
 #define SWAP_BUFFERS_LAYER(window) \
diff --git a/src/glut/glx/win32_glx.c b/src/glut/glx/win32_glx.c
index ecfff44b93..3522e606ec 100644
--- a/src/glut/glx/win32_glx.c
+++ b/src/glut/glx/win32_glx.c
@@ -22,7 +22,7 @@ glXCreateContext(Display * display, XVisualInfo * visinfo,
      routine. */
   HGLRC context;
 
-  context = CreateContext(XHDC);
+  context = wglCreateContext(XHDC);
 
 #if 0
   /* XXX GLUT doesn't support it now, so don't worry about display list
diff --git a/src/mesa/main/texformat_tmp.h b/src/mesa/main/texformat_tmp.h
index 08b6d8f12f..275340cabd 100644
--- a/src/mesa/main/texformat_tmp.h
+++ b/src/mesa/main/texformat_tmp.h
@@ -1223,11 +1223,11 @@ static void store_texel_srgb8(struct gl_texture_image *texImage,
 static void FETCH(srgba8)(const struct gl_texture_image *texImage,
                           GLint i, GLint j, GLint k, GLfloat *texel )
 {
-   const GLubyte *src = TEXEL_ADDR(GLubyte, texImage, i, j, k, 4);
-   texel[RCOMP] = nonlinear_to_linear(src[0]);
-   texel[GCOMP] = nonlinear_to_linear(src[1]);
-   texel[BCOMP] = nonlinear_to_linear(src[2]);
-   texel[ACOMP] = UBYTE_TO_FLOAT(src[3]); /* linear! */
+   const GLuint s = *TEXEL_ADDR(GLuint, texImage, i, j, k, 1);
+   texel[RCOMP] = nonlinear_to_linear( (s >> 24) );
+   texel[GCOMP] = nonlinear_to_linear( (s >> 16) & 0xff );
+   texel[BCOMP] = nonlinear_to_linear( (s >>  8) & 0xff );
+   texel[ACOMP] = UBYTE_TO_FLOAT( (s      ) & 0xff ); /* linear! */
 }
 
 #if DIM == 3
@@ -1235,11 +1235,8 @@ static void store_texel_srgba8(struct gl_texture_image *texImage,
                                GLint i, GLint j, GLint k, const void *texel)
 {
    const GLubyte *rgba = (const GLubyte *) texel;
-   GLubyte *dst = TEXEL_ADDR(GLubyte, texImage, i, j, k, 4);
-   dst[0] = rgba[RCOMP];
-   dst[1] = rgba[GCOMP];
-   dst[2] = rgba[BCOMP];
-   dst[3] = rgba[ACOMP];
+   GLuint *dst = TEXEL_ADDR(GLuint, texImage, i, j, k, 1);
+   *dst = PACK_COLOR_8888(rgba[RCOMP], rgba[GCOMP], rgba[BCOMP], rgba[ACOMP]);
 }
 #endif
 
@@ -1247,11 +1244,11 @@ static void store_texel_srgba8(struct gl_texture_image *texImage,
 static void FETCH(sargb8)(const struct gl_texture_image *texImage,
                           GLint i, GLint j, GLint k, GLfloat *texel )
 {
-   const GLubyte *src = TEXEL_ADDR(GLubyte, texImage, i, j, k, 4);
-   texel[RCOMP] = nonlinear_to_linear(src[1]);
-   texel[GCOMP] = nonlinear_to_linear(src[2]);
-   texel[BCOMP] = nonlinear_to_linear(src[3]);
-   texel[ACOMP] = UBYTE_TO_FLOAT(src[0]); /* linear! */
+   const GLuint s = *TEXEL_ADDR(GLuint, texImage, i, j, k, 1);
+   texel[RCOMP] = nonlinear_to_linear( (s >> 16) & 0xff );
+   texel[GCOMP] = nonlinear_to_linear( (s >>  8) & 0xff );
+   texel[BCOMP] = nonlinear_to_linear( (s      ) & 0xff );
+   texel[ACOMP] = UBYTE_TO_FLOAT( (s >> 24) ); /* linear! */
 }
 
 #if DIM == 3
@@ -1259,11 +1256,8 @@ static void store_texel_sargb8(struct gl_texture_image *texImage,
                                GLint i, GLint j, GLint k, const void *texel)
 {
    const GLubyte *rgba = (const GLubyte *) texel;
-   GLubyte *dst = TEXEL_ADDR(GLubyte, texImage, i, j, k, 4);
-   dst[0] = rgba[ACOMP];
-   dst[1] = rgba[RCOMP];
-   dst[2] = rgba[GCOMP];
-   dst[3] = rgba[BCOMP];
+   GLuint *dst = TEXEL_ADDR(GLuint, texImage, i, j, k, 1);
+   *dst = PACK_COLOR_8888(rgba[ACOMP], rgba[RCOMP], rgba[GCOMP], rgba[BCOMP]);
 }
 #endif
 
diff --git a/src/mesa/main/texstore.c b/src/mesa/main/texstore.c
index 2b06796aba..6360ca15f8 100644
--- a/src/mesa/main/texstore.c
+++ b/src/mesa/main/texstore.c
@@ -3699,14 +3699,37 @@ is_srgb_teximage(const struct gl_texture_image *texImage)
    case MESA_FORMAT_SARGB8:
    case MESA_FORMAT_SL8:
    case MESA_FORMAT_SLA8:
+   case MESA_FORMAT_SRGB_DXT1:
+   case MESA_FORMAT_SRGBA_DXT1:
+   case MESA_FORMAT_SRGBA_DXT3:
+   case MESA_FORMAT_SRGBA_DXT5:
       return GL_TRUE;
    default:
       return GL_FALSE;
    }
 }
 
-#endif /* FEATURE_EXT_texture_sRGB */
 
+/**
+ * Convert a float value from linear space to a
+ * non-linear sRGB value in [0, 255].
+ * Not terribly efficient.
+ */
+static INLINE GLfloat
+linear_to_nonlinear(GLfloat cl)
+{
+   /* can't have values outside [0, 1] */
+   GLfloat cs;
+   if (cl < 0.0031308) {
+      cs = 12.92 * cl;
+   }
+   else {
+      cs = 1.055 * _mesa_pow(cl, 0.41666) - 0.055;
+   }
+   return cs;
+}
+
+#endif /* FEATURE_EXT_texture_sRGB */
 
 /**
  * This is the software fallback for Driver.GetTexImage().
@@ -3823,16 +3846,34 @@ _mesa_get_teximage(GLcontext *ctx, GLenum target, GLint level,
             }
 #if FEATURE_EXT_texture_sRGB
             else if (is_srgb_teximage(texImage)) {
-               /* no pixel transfer and no non-linear to linear conversion */
-               const GLint comps = texImage->TexFormat->TexelBytes;
-               const GLint rowstride = comps * texImage->RowStride;
-               MEMCPY(dest,
-                      (const GLubyte *) texImage->Data + row * rowstride,
-                      comps * width * sizeof(GLubyte));
-               /* FIXME: isn't it necessary to still do component assigning
-                  according to format/type? */
-               /* FIXME: need to do something else for compressed srgb textures
-                         (currently will return values converted to linear) */
+               /* special case this since need to backconvert values */
+               /* convert row to RGBA format */
+               GLfloat rgba[MAX_WIDTH][4];
+               GLint col;
+               GLbitfield transferOps = 0x0;
+
+               for (col = 0; col < width; col++) {
+                  (*texImage->FetchTexelf)(texImage, col, row, img, rgba[col]);
+                  if (texImage->TexFormat->BaseFormat == GL_LUMINANCE) {
+                     rgba[col][RCOMP] = linear_to_nonlinear(rgba[col][RCOMP]);
+                     rgba[col][GCOMP] = 0.0;
+                     rgba[col][BCOMP] = 0.0;
+                  }
+                  else if (texImage->TexFormat->BaseFormat == GL_LUMINANCE_ALPHA) {
+                     rgba[col][RCOMP] = linear_to_nonlinear(rgba[col][RCOMP]);
+                     rgba[col][GCOMP] = 0.0;
+                     rgba[col][BCOMP] = 0.0;
+                  }
+                  else if (texImage->TexFormat->BaseFormat == GL_RGB ||
+                     texImage->TexFormat->BaseFormat == GL_RGBA) {
+                     rgba[col][RCOMP] = linear_to_nonlinear(rgba[col][RCOMP]);
+                     rgba[col][GCOMP] = linear_to_nonlinear(rgba[col][GCOMP]);
+                     rgba[col][BCOMP] = linear_to_nonlinear(rgba[col][BCOMP]);
+                  }
+               }
+               _mesa_pack_rgba_span_float(ctx, width, (GLfloat (*)[4]) rgba,
+                                          format, type, dest,
+                                          &ctx->Pack, transferOps /*image xfer ops*/);
             }
 #endif /* FEATURE_EXT_texture_sRGB */
             else {
diff --git a/src/mesa/state_tracker/st_cb_texture.c b/src/mesa/state_tracker/st_cb_texture.c
index d08229b57a..9279a768c8 100644
--- a/src/mesa/state_tracker/st_cb_texture.c
+++ b/src/mesa/state_tracker/st_cb_texture.c
@@ -1417,9 +1417,7 @@ st_finalize_texture(GLcontext *ctx,
           stObj->pt->width[0] != firstImage->base.Width2 ||
           stObj->pt->height[0] != firstImage->base.Height2 ||
           stObj->pt->depth[0] != firstImage->base.Depth2 ||
-          stObj->pt->block.size != cpp ||
-          stObj->pt->block.width != 1 ||
-          stObj->pt->block.height != 1 ||
+          stObj->pt->block.size/stObj->pt->block.width != cpp || /* Nominal bytes per pixel */
           stObj->pt->compressed != firstImage->base.IsCompressed) {
          pipe_texture_release(&stObj->pt);
          ctx->st->dirty.st |= ST_NEW_FRAMEBUFFER;
diff --git a/src/mesa/state_tracker/st_mesa_to_tgsi.c b/src/mesa/state_tracker/st_mesa_to_tgsi.c
index 5b5707fa1c..8b9b18f94d 100644
--- a/src/mesa/state_tracker/st_mesa_to_tgsi.c
+++ b/src/mesa/state_tracker/st_mesa_to_tgsi.c
@@ -188,7 +188,7 @@ make_immediate(const float *value, uint size)
    struct tgsi_full_immediate imm;
 
    imm = tgsi_default_full_immediate();
-   imm.Immediate.Size += size;
+   imm.Immediate.NrTokens += size;
    imm.Immediate.DataType = TGSI_IMM_FLOAT32;
    imm.u.Pointer = value;
    return imm;