vbo: Move vbo_split into the tnl module.

Move the files, adapt to the naming scheme in tnl, update callers
and build system.

Reviewed-by: Brian Paul <brianp@vmware.com>
Signed-off-by: Mathias Fröhlich <Mathias.Froehlich@web.de>

8 files changed, 1258 insertions, 7 deletions

diff --git a/src/mesa/tnl/t_draw.c b/src/mesa/tnl/t_draw.c
index a0fd58432a1..a83b98eede1 100644
--- a/src/mesa/tnl/t_draw.c
+++ b/src/mesa/tnl/t_draw.c
@@ -486,10 +486,10 @@ void _tnl_draw_prims(struct gl_context *ctx,
       /* This will split the buffers one way or another and
        * recursively call back into this function.
        */
-      vbo_split_prims( ctx, arrays, prim, nr_prims, ib, 
-		       0, max_index + prim->basevertex,
-		       _tnl_draw_prims,
-		       &limits );
+      _tnl_split_prims( ctx, arrays, prim, nr_prims, ib,
+                        0, max_index + prim->basevertex,
+                        _tnl_draw_prims,
+                        &limits );
    }
    else {
       /* May need to map a vertex buffer object for every attribute plus
diff --git a/src/mesa/tnl/t_rebase.c b/src/mesa/tnl/t_rebase.c
index 19e759f44be..d28512423c3 100644
--- a/src/mesa/tnl/t_rebase.c
+++ b/src/mesa/tnl/t_rebase.c
@@ -51,6 +51,7 @@
 #include "main/glheader.h"
 #include "main/imports.h"
 #include "main/mtypes.h"
+#include "vbo/vbo.h"
 
 #include "t_rebase.h"
 
@@ -108,7 +109,7 @@ void t_rebase_prims( struct gl_context *ctx,
                      const struct _mesa_index_buffer *ib,
                      GLuint min_index,
                      GLuint max_index,
-                     vbo_draw_func draw )
+                     tnl_draw_func draw )
 {
    struct gl_array_attributes tmp_attribs[VERT_ATTRIB_MAX];
    struct gl_vertex_array tmp_arrays[VERT_ATTRIB_MAX];
diff --git a/src/mesa/tnl/t_rebase.h b/src/mesa/tnl/t_rebase.h
index 16a3a2b5a33..ce2e8b0590e 100644
--- a/src/mesa/tnl/t_rebase.h
+++ b/src/mesa/tnl/t_rebase.h
@@ -25,7 +25,7 @@
 #ifndef _T_REBASE_H_
 #define _T_REBASE_H_
 
-#include "vbo/vbo.h"
+#include "tnl.h"
 
 void t_rebase_prims( struct gl_context *ctx,
                      const struct gl_vertex_array *arrays,
@@ -34,6 +34,6 @@ void t_rebase_prims( struct gl_context *ctx,
                      const struct _mesa_index_buffer *ib,
                      GLuint min_index,
                      GLuint max_index,
-                     vbo_draw_func draw );
+                     tnl_draw_func draw );
 
 #endif
diff --git a/src/mesa/tnl/t_split.c b/src/mesa/tnl/t_split.c
new file mode 100644
index 00000000000..b98bd404d52
--- /dev/null
+++ b/src/mesa/tnl/t_split.c
@@ -0,0 +1,160 @@
+
+/*
+ * Mesa 3-D graphics library
+ *
+ * Copyright (C) 1999-2006  Brian Paul   All Rights Reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included
+ * in all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
+ * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
+ * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ *
+ * Authors:
+ *    Keith Whitwell <keithw@vmware.com>
+ */
+
+/* Deal with hardware and/or swtnl maximums:
+ * - maximum number of vertices in buffer
+ * - maximum number of elements (maybe zero)
+ *
+ * The maximums may vary with opengl state (eg if a larger hardware
+ * vertex is required in this state, the maximum number of vertices
+ * may be smaller than in another state).
+ *
+ * We want buffer splitting to be a convenience function for the code
+ * actually drawing the primitives rather than a system-wide maximum,
+ * otherwise it is hard to avoid pessimism.
+ *
+ * For instance, if a driver has no hardware limits on vertex buffer
+ * dimensions, it would not ordinarily want to split vbos.  But if
+ * there is an unexpected fallback, eg memory manager fails to upload
+ * textures, it will want to pass the drawing commands onto swtnl,
+ * which does have limitations.  A convenience function allows swtnl
+ * to split the drawing and vbos internally without imposing its
+ * limitations on drivers which want to use it as a fallback path.
+ */
+
+#include "main/glheader.h"
+#include "main/mtypes.h"
+#include "vbo/vbo.h"
+
+#include "t_split.h"
+
+
+/* True if a primitive can be split without copying of vertices, false
+ * otherwise.
+ */
+GLboolean
+_tnl_split_prim_inplace(GLenum mode, GLuint *first, GLuint *incr)
+{
+   switch (mode) {
+   case GL_POINTS:
+      *first = 1;
+      *incr = 1;
+      return GL_TRUE;
+   case GL_LINES:
+      *first = 2;
+      *incr = 2;
+      return GL_TRUE;
+   case GL_LINE_STRIP:
+      *first = 2;
+      *incr = 1;
+      return GL_TRUE;
+   case GL_TRIANGLES:
+      *first = 3;
+      *incr = 3;
+      return GL_TRUE;
+   case GL_TRIANGLE_STRIP:
+      *first = 3;
+      *incr = 1;
+      return GL_TRUE;
+   case GL_QUADS:
+      *first = 4;
+      *incr = 4;
+      return GL_TRUE;
+   case GL_QUAD_STRIP:
+      *first = 4;
+      *incr = 2;
+      return GL_TRUE;
+   default:
+      *first = 0;
+      *incr = 1;		/* so that count % incr works */
+      return GL_FALSE;
+   }
+}
+
+
+
+void
+_tnl_split_prims(struct gl_context *ctx,
+                 const struct gl_vertex_array arrays[],
+                 const struct _mesa_prim *prim,
+                 GLuint nr_prims,
+                 const struct _mesa_index_buffer *ib,
+                 GLuint min_index,
+                 GLuint max_index,
+                 tnl_draw_func draw,
+                 const struct split_limits *limits)
+{
+   if (ib) {
+      if (limits->max_indices == 0) {
+         /* Could traverse the indices, re-emitting vertices in turn.
+          * But it's hard to see why this case would be needed - for
+          * software tnl, it is better to convert to non-indexed
+          * rendering after transformation is complete.  Are there any devices
+          * with hardware tnl that cannot do indexed rendering?
+          *
+          * For now, this path is disabled.
+          */
+         assert(0);
+      }
+      else if (max_index - min_index >= limits->max_verts) {
+         /* The vertex buffers are too large for hardware (or the
+          * swtnl module).  Traverse the indices, re-emitting vertices
+          * in turn.  Use a vertex cache to preserve some of the
+          * sharing from the original index list.
+          */
+         _tnl_split_copy(ctx, arrays, prim, nr_prims, ib, draw, limits);
+      }
+      else if (ib->count > limits->max_indices) {
+         /* The index buffer is too large for hardware.  Try to split
+          * on whole-primitive boundaries, otherwise try to split the
+          * individual primitives.
+          */
+         _tnl_split_inplace(ctx, arrays, prim, nr_prims, ib,
+                            min_index, max_index, draw, limits);
+      }
+      else {
+         /* Why were we called? */
+         assert(0);
+      }
+   }
+   else {
+      if (max_index - min_index >= limits->max_verts) {
+         /* The vertex buffer is too large for hardware (or the swtnl
+          * module).  Try to split on whole-primitive boundaries,
+          * otherwise try to split the individual primitives.
+          */
+         _tnl_split_inplace(ctx, arrays, prim, nr_prims, ib,
+                            min_index, max_index, draw, limits);
+      }
+      else {
+         /* Why were we called? */
+         assert(0);
+      }
+   }
+}
+
diff --git a/src/mesa/tnl/t_split.h b/src/mesa/tnl/t_split.h
new file mode 100644
index 00000000000..ced7d30bdf1
--- /dev/null
+++ b/src/mesa/tnl/t_split.h
@@ -0,0 +1,74 @@
+/*
+ * mesa 3-D graphics library
+ *
+ * Copyright (C) 1999-2006  Brian Paul   All Rights Reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included
+ * in all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
+ * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
+ * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ */
+
+/**
+ * \brief VBO builder module datatypes and definitions.
+ * \author Keith Whitwell
+ */
+
+
+/**
+ * \mainpage The TNL splitter
+ *
+ * This is the private data used internally to the _tnl_split_prims()
+ * helper function.  Nobody outside the _tnl_split* files needs to
+ * include or know about this structure.
+ */
+
+
+#ifndef _TNL_SPLIT_H
+#define _TNL_SPLIT_H
+
+#include "tnl.h"
+
+
+/* True if a primitive can be split without copying of vertices, false
+ * otherwise.
+ */
+GLboolean
+_tnl_split_prim_inplace(GLenum mode, GLuint *first, GLuint *incr);
+
+void
+_tnl_split_inplace(struct gl_context *ctx,
+                   const struct gl_vertex_array arrays[],
+                   const struct _mesa_prim *prim,
+                   GLuint nr_prims,
+                   const struct _mesa_index_buffer *ib,
+                   GLuint min_index,
+                   GLuint max_index,
+                   tnl_draw_func draw,
+                   const struct split_limits *limits);
+
+/* Requires ib != NULL:
+ */
+void
+_tnl_split_copy(struct gl_context *ctx,
+                const struct gl_vertex_array arrays[],
+                const struct _mesa_prim *prim,
+                GLuint nr_prims,
+                const struct _mesa_index_buffer *ib,
+                tnl_draw_func draw,
+                const struct split_limits *limits);
+
+#endif
diff --git a/src/mesa/tnl/t_split_copy.c b/src/mesa/tnl/t_split_copy.c
new file mode 100644
index 00000000000..f76a470b5ff
--- /dev/null
+++ b/src/mesa/tnl/t_split_copy.c
@@ -0,0 +1,638 @@
+
+/*
+ * Mesa 3-D graphics library
+ *
+ * Copyright (C) 1999-2006  Brian Paul   All Rights Reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included
+ * in all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
+ * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
+ * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ *
+ * Authors:
+ *    Keith Whitwell <keithw@vmware.com>
+ */
+
+/* Split indexed primitives with per-vertex copying.
+ */
+
+#include <stdio.h>
+
+#include "main/glheader.h"
+#include "main/bufferobj.h"
+#include "main/imports.h"
+#include "main/glformats.h"
+#include "main/macros.h"
+#include "main/mtypes.h"
+#include "main/varray.h"
+#include "vbo/vbo.h"
+
+#include "t_split.h"
+#include "tnl.h"
+
+
+#define ELT_TABLE_SIZE 16
+
+/**
+ * Used for vertex-level splitting of indexed buffers.  Note that
+ * non-indexed primitives may be converted to indexed in some cases
+ * (eg loops, fans) in order to use this splitting path.
+ */
+struct copy_context {
+   struct gl_context *ctx;
+   const struct gl_vertex_array *array;
+   const struct _mesa_prim *prim;
+   GLuint nr_prims;
+   const struct _mesa_index_buffer *ib;
+   tnl_draw_func draw;
+
+   const struct split_limits *limits;
+
+   struct {
+      GLuint attr;
+      GLuint size;
+      const struct gl_vertex_array *array;
+      const GLubyte *src_ptr;
+
+      struct gl_vertex_buffer_binding dstbinding;
+      struct gl_array_attributes dstattribs;
+
+   } varying[VERT_ATTRIB_MAX];
+   GLuint nr_varying;
+
+   struct gl_vertex_array dstarray[VERT_ATTRIB_MAX];
+   struct _mesa_index_buffer dstib;
+
+   GLuint *translated_elt_buf;
+   const GLuint *srcelt;
+
+   /** A baby hash table to avoid re-emitting (some) duplicate
+    * vertices when splitting indexed primitives.
+    */
+   struct {
+      GLuint in;
+      GLuint out;
+   } vert_cache[ELT_TABLE_SIZE];
+
+   GLuint vertex_size;
+   GLubyte *dstbuf;
+   GLubyte *dstptr;     /**< dstptr == dstbuf + dstelt_max * vertsize */
+   GLuint dstbuf_size;  /**< in vertices */
+   GLuint dstbuf_nr;    /**< count of emitted vertices, also the largest value
+                         * in dstelt.  Our MaxIndex.
+                         */
+
+   GLuint *dstelt;
+   GLuint dstelt_nr;
+   GLuint dstelt_size;
+
+#define MAX_PRIM 32
+   struct _mesa_prim dstprim[MAX_PRIM];
+   GLuint dstprim_nr;
+};
+
+
+static GLuint
+attr_size(const struct gl_array_attributes *attrib)
+{
+   return attrib->Size * _mesa_sizeof_type(attrib->Type);
+}
+
+
+/**
+ * Starts returning true slightly before the buffer fills, to ensure
+ * that there is sufficient room for any remaining vertices to finish
+ * off the prim:
+ */
+static GLboolean
+check_flush(struct copy_context *copy)
+{
+   GLenum mode = copy->dstprim[copy->dstprim_nr].mode;
+
+   if (GL_TRIANGLE_STRIP == mode &&
+       copy->dstelt_nr & 1) { /* see bug9962 */
+       return GL_FALSE;
+   }
+
+   if (copy->dstbuf_nr + 4 > copy->dstbuf_size)
+      return GL_TRUE;
+
+   if (copy->dstelt_nr + 4 > copy->dstelt_size)
+      return GL_TRUE;
+
+   return GL_FALSE;
+}
+
+
+/**
+ * Dump the parameters/info for a vbo->draw() call.
+ */
+static void
+dump_draw_info(struct gl_context *ctx,
+               const struct gl_vertex_array *arrays,
+               const struct _mesa_prim *prims,
+               GLuint nr_prims,
+               const struct _mesa_index_buffer *ib,
+               GLuint min_index,
+               GLuint max_index)
+{
+   GLuint i, j;
+
+   printf("VBO Draw:\n");
+   for (i = 0; i < nr_prims; i++) {
+      printf("Prim %u of %u\n", i, nr_prims);
+      printf("  Prim mode 0x%x\n", prims[i].mode);
+      printf("  IB: %p\n", (void*) ib);
+      for (j = 0; j < VERT_ATTRIB_MAX; j++) {
+         const struct gl_vertex_array *array = &arrays[j];
+         const struct gl_vertex_buffer_binding *binding
+            = array->BufferBinding;
+         const struct gl_array_attributes *attrib = array->VertexAttrib;
+         const GLubyte *ptr = _mesa_vertex_attrib_address(attrib, binding);
+         printf("    array %d at %p:\n", j, (void*) &arrays[j]);
+         printf("      ptr %p, size %d, type 0x%x, stride %d\n",
+                ptr, attrib->Size, attrib->Type, binding->Stride);
+         if (0) {
+            GLint k = prims[i].start + prims[i].count - 1;
+            GLfloat *last = (GLfloat *) (ptr + binding->Stride * k);
+            printf("        last: %f %f %f\n",
+                   last[0], last[1], last[2]);
+         }
+      }
+   }
+}
+
+
+static void
+flush(struct copy_context *copy)
+{
+   struct gl_context *ctx = copy->ctx;
+   GLuint i;
+
+   /* Set some counters:
+    */
+   copy->dstib.count = copy->dstelt_nr;
+
+#if 0
+   dump_draw_info(copy->ctx,
+                  copy->dstarray,
+                  copy->dstprim,
+                  copy->dstprim_nr,
+                  &copy->dstib,
+                  0,
+                  copy->dstbuf_nr);
+#else
+   (void) dump_draw_info;
+#endif
+
+   copy->draw(ctx,
+              copy->dstarray,
+              copy->dstprim,
+              copy->dstprim_nr,
+              &copy->dstib,
+              GL_TRUE,
+              0,
+              copy->dstbuf_nr - 1,
+              NULL, 0, NULL);
+
+   /* Reset all pointers:
+    */
+   copy->dstprim_nr = 0;
+   copy->dstelt_nr = 0;
+   copy->dstbuf_nr = 0;
+   copy->dstptr = copy->dstbuf;
+
+   /* Clear the vertex cache:
+    */
+   for (i = 0; i < ELT_TABLE_SIZE; i++)
+      copy->vert_cache[i].in = ~0;
+}
+
+
+/**
+ * Called at begin of each primitive during replay.
+ */
+static void
+begin(struct copy_context *copy, GLenum mode, GLboolean begin_flag)
+{
+   struct _mesa_prim *prim = &copy->dstprim[copy->dstprim_nr];
+
+   prim->mode = mode;
+   prim->begin = begin_flag;
+   prim->num_instances = 1;
+}
+
+
+/**
+ * Use a hashtable to attempt to identify recently-emitted vertices
+ * and avoid re-emitting them.
+ */
+static GLuint
+elt(struct copy_context *copy, GLuint elt_idx)
+{
+   GLuint elt = copy->srcelt[elt_idx] + copy->prim->basevertex;
+   GLuint slot = elt & (ELT_TABLE_SIZE-1);
+
+   /* Look up the incoming element in the vertex cache.  Re-emit if
+    * necessary.
+    */
+   if (copy->vert_cache[slot].in != elt) {
+      GLubyte *csr = copy->dstptr;
+      GLuint i;
+
+      for (i = 0; i < copy->nr_varying; i++) {
+         const struct gl_vertex_array *srcarray = copy->varying[i].array;
+         const struct gl_vertex_buffer_binding* srcbinding
+            = srcarray->BufferBinding;
+         const GLubyte *srcptr
+            = copy->varying[i].src_ptr + elt * srcbinding->Stride;
+
+         memcpy(csr, srcptr, copy->varying[i].size);
+         csr += copy->varying[i].size;
+
+#ifdef NAN_CHECK
+         if (srcarray->Type == GL_FLOAT) {
+            GLuint k;
+            GLfloat *f = (GLfloat *) srcptr;
+            for (k = 0; k < srcarray->Size; k++) {
+               assert(!IS_INF_OR_NAN(f[k]));
+               assert(f[k] <= 1.0e20 && f[k] >= -1.0e20);
+            }
+         }
+#endif
+
+         if (0) {
+            const GLuint *f = (const GLuint *)srcptr;
+            GLuint j;
+            printf("  varying %d: ", i);
+            for (j = 0; j < copy->varying[i].size / 4; j++)
+               printf("%x ", f[j]);
+            printf("\n");
+         }
+      }
+
+      copy->vert_cache[slot].in = elt;
+      copy->vert_cache[slot].out = copy->dstbuf_nr++;
+      copy->dstptr += copy->vertex_size;
+
+      assert(csr == copy->dstptr);
+      assert(copy->dstptr == (copy->dstbuf +
+                              copy->dstbuf_nr * copy->vertex_size));
+   }
+
+   copy->dstelt[copy->dstelt_nr++] = copy->vert_cache[slot].out;
+   return check_flush(copy);
+}
+
+
+/**
+ * Called at end of each primitive during replay.
+ */
+static void
+end(struct copy_context *copy, GLboolean end_flag)
+{
+   struct _mesa_prim *prim = &copy->dstprim[copy->dstprim_nr];
+
+   prim->end = end_flag;
+   prim->count = copy->dstelt_nr - prim->start;
+
+   if (++copy->dstprim_nr == MAX_PRIM || check_flush(copy)) {
+      flush(copy);
+   }
+}
+
+
+static void
+replay_elts(struct copy_context *copy)
+{
+   GLuint i, j, k;
+   GLboolean split;
+
+   for (i = 0; i < copy->nr_prims; i++) {
+      const struct _mesa_prim *prim = &copy->prim[i];
+      const GLuint start = prim->start;
+      GLuint first, incr;
+
+      switch (prim->mode) {
+      case GL_LINE_LOOP:
+         /* Convert to linestrip and emit the final vertex explicitly,
+          * but only in the resultant strip that requires it.
+          */
+         j = 0;
+         while (j != prim->count) {
+            begin(copy, GL_LINE_STRIP, prim->begin && j == 0);
+
+            for (split = GL_FALSE; j != prim->count && !split; j++)
+               split = elt(copy, start + j);
+
+            if (j == prim->count) {
+               /* Done, emit final line.  Split doesn't matter as
+                * it is always raised a bit early so we can emit
+                * the last verts if necessary!
+                */
+               if (prim->end)
+                  (void)elt(copy, start + 0);
+
+               end(copy, prim->end);
+            }
+            else {
+               /* Wrap
+                */
+               assert(split);
+               end(copy, 0);
+               j--;
+            }
+         }
+         break;
+
+      case GL_TRIANGLE_FAN:
+      case GL_POLYGON:
+         j = 2;
+         while (j != prim->count) {
+            begin(copy, prim->mode, prim->begin && j == 0);
+
+            split = elt(copy, start+0);
+            assert(!split);
+
+            split = elt(copy, start+j-1);
+            assert(!split);
+
+            for (; j != prim->count && !split; j++)
+               split = elt(copy, start+j);
+
+            end(copy, prim->end && j == prim->count);
+
+            if (j != prim->count) {
+               /* Wrapped the primitive, need to repeat some vertices:
+                */
+               j -= 1;
+            }
+         }
+         break;
+
+      default:
+         (void)_tnl_split_prim_inplace(prim->mode, &first, &incr);
+
+         j = 0;
+         while (j != prim->count) {
+
+            begin(copy, prim->mode, prim->begin && j == 0);
+
+            split = 0;
+            for (k = 0; k < first; k++, j++)
+               split |= elt(copy, start+j);
+
+            assert(!split);
+
+            for (; j != prim->count && !split;)
+               for (k = 0; k < incr; k++, j++)
+                  split |= elt(copy, start+j);
+
+            end(copy, prim->end && j == prim->count);
+
+            if (j != prim->count) {
+               /* Wrapped the primitive, need to repeat some vertices:
+                */
+               assert(j > first - incr);
+               j -= (first - incr);
+            }
+         }
+         break;
+      }
+   }
+
+   if (copy->dstprim_nr)
+      flush(copy);
+}
+
+
+static void
+replay_init(struct copy_context *copy)
+{
+   struct gl_context *ctx = copy->ctx;
+   GLuint i;
+   GLuint offset;
+   const GLvoid *srcptr;
+
+   /* Make a list of varying attributes and their vbo's.  Also
+    * calculate vertex size.
+    */
+   copy->vertex_size = 0;
+   for (i = 0; i < VERT_ATTRIB_MAX; i++) {
+      const struct gl_vertex_array *array = &copy->array[i];
+      const struct gl_vertex_buffer_binding *binding = array->BufferBinding;
+
+      if (binding->Stride == 0) {
+         _mesa_copy_vertex_array(&copy->dstarray[i], array);
+      }
+      else {
+         const struct gl_array_attributes *attrib = array->VertexAttrib;
+         struct gl_buffer_object *vbo = binding->BufferObj;
+         const GLubyte *ptr = _mesa_vertex_attrib_address(attrib, binding);
+         GLuint j = copy->nr_varying++;
+
+         copy->varying[j].attr = i;
+         copy->varying[j].array = &copy->array[i];
+         copy->varying[j].size = attr_size(attrib);
+         copy->vertex_size += attr_size(attrib);
+
+         if (_mesa_is_bufferobj(vbo) &&
+             !_mesa_bufferobj_mapped(vbo, MAP_INTERNAL))
+            ctx->Driver.MapBufferRange(ctx, 0, vbo->Size, GL_MAP_READ_BIT, vbo,
+                                       MAP_INTERNAL);
+
+         copy->varying[j].src_ptr =
+               ADD_POINTERS(vbo->Mappings[MAP_INTERNAL].Pointer, ptr);
+
+         copy->dstarray[i].VertexAttrib = &copy->varying[j].dstattribs;
+         copy->dstarray[i].BufferBinding = &copy->varying[j].dstbinding;
+      }
+   }
+
+   /* There must always be an index buffer.  Currently require the
+    * caller convert non-indexed prims to indexed.  Could alternately
+    * do it internally.
+    */
+   if (_mesa_is_bufferobj(copy->ib->obj) &&
+       !_mesa_bufferobj_mapped(copy->ib->obj, MAP_INTERNAL))
+      ctx->Driver.MapBufferRange(ctx, 0, copy->ib->obj->Size, GL_MAP_READ_BIT,
+                                 copy->ib->obj, MAP_INTERNAL);
+
+   srcptr = (const GLubyte *)
+            ADD_POINTERS(copy->ib->obj->Mappings[MAP_INTERNAL].Pointer,
+                         copy->ib->ptr);
+
+   switch (copy->ib->index_size) {
+   case 1:
+      copy->translated_elt_buf = malloc(sizeof(GLuint) * copy->ib->count);
+      copy->srcelt = copy->translated_elt_buf;
+
+      for (i = 0; i < copy->ib->count; i++)
+         copy->translated_elt_buf[i] = ((const GLubyte *)srcptr)[i];
+      break;
+
+   case 2:
+      copy->translated_elt_buf = malloc(sizeof(GLuint) * copy->ib->count);
+      copy->srcelt = copy->translated_elt_buf;
+
+      for (i = 0; i < copy->ib->count; i++)
+         copy->translated_elt_buf[i] = ((const GLushort *)srcptr)[i];
+      break;
+
+   case 4:
+      copy->translated_elt_buf = NULL;
+      copy->srcelt = (const GLuint *)srcptr;
+      break;
+   }
+
+   /* Figure out the maximum allowed vertex buffer size:
+    */
+   if (copy->vertex_size * copy->limits->max_verts <= copy->limits->max_vb_size) {
+      copy->dstbuf_size = copy->limits->max_verts;
+   }
+   else {
+      copy->dstbuf_size = copy->limits->max_vb_size / copy->vertex_size;
+   }
+
+   /* Allocate an output vertex buffer:
+    *
+    * XXX:  This should be a VBO!
+    */
+   copy->dstbuf = malloc(copy->dstbuf_size * copy->vertex_size);
+   copy->dstptr = copy->dstbuf;
+
+   /* Setup new vertex arrays to point into the output buffer:
+    */
+   for (offset = 0, i = 0; i < copy->nr_varying; i++) {
+      const struct gl_vertex_array *src = copy->varying[i].array;
+      const struct gl_array_attributes *srcattr = src->VertexAttrib;
+      struct gl_vertex_array *dst = &copy->dstarray[i];
+      struct gl_vertex_buffer_binding *dstbind = &copy->varying[i].dstbinding;
+      struct gl_array_attributes *dstattr = &copy->varying[i].dstattribs;
+
+      dstattr->Size = srcattr->Size;
+      dstattr->Type = srcattr->Type;
+      dstattr->Format = GL_RGBA;
+      dstbind->Stride = copy->vertex_size;
+      dstattr->Ptr = copy->dstbuf + offset;
+      dstattr->Normalized = srcattr->Normalized;
+      dstattr->Integer = srcattr->Integer;
+      dstattr->Doubles = srcattr->Doubles;
+      dstbind->BufferObj = ctx->Shared->NullBufferObj;
+      dstattr->_ElementSize = srcattr->_ElementSize;
+      dst->BufferBinding = dstbind;
+      dst->VertexAttrib = dstattr;
+
+      offset += copy->varying[i].size;
+   }
+
+   /* Allocate an output element list:
+    */
+   copy->dstelt_size = MIN2(65536, copy->ib->count * 2 + 3);
+   copy->dstelt_size = MIN2(copy->dstelt_size, copy->limits->max_indices);
+   copy->dstelt = malloc(sizeof(GLuint) * copy->dstelt_size);
+   copy->dstelt_nr = 0;
+
+   /* Setup the new index buffer to point to the allocated element
+    * list:
+    */
+   copy->dstib.count = 0;        /* duplicates dstelt_nr */
+   copy->dstib.index_size = 4;
+   copy->dstib.obj = ctx->Shared->NullBufferObj;
+   copy->dstib.ptr = copy->dstelt;
+}
+
+
+/**
+ * Free up everything allocated during split/replay.
+ */
+static void
+replay_finish(struct copy_context *copy)
+{
+   struct gl_context *ctx = copy->ctx;
+   GLuint i;
+
+   /* Free our vertex and index buffers */
+   free(copy->translated_elt_buf);
+   free(copy->dstbuf);
+   free(copy->dstelt);
+
+   /* Unmap VBO's */
+   for (i = 0; i < copy->nr_varying; i++) {
+      struct gl_buffer_object *vbo =
+         copy->varying[i].array->BufferBinding->BufferObj;
+      if (_mesa_is_bufferobj(vbo) && _mesa_bufferobj_mapped(vbo, MAP_INTERNAL))
+         ctx->Driver.UnmapBuffer(ctx, vbo, MAP_INTERNAL);
+   }
+
+   /* Unmap index buffer */
+   if (_mesa_is_bufferobj(copy->ib->obj) &&
+       _mesa_bufferobj_mapped(copy->ib->obj, MAP_INTERNAL)) {
+      ctx->Driver.UnmapBuffer(ctx, copy->ib->obj, MAP_INTERNAL);
+   }
+}
+
+
+/**
+ * Split VBO into smaller pieces, draw the pieces.
+ */
+void
+_tnl_split_copy(struct gl_context *ctx,
+                const struct gl_vertex_array *arrays,
+                const struct _mesa_prim *prim,
+                GLuint nr_prims,
+                const struct _mesa_index_buffer *ib,
+                tnl_draw_func draw,
+                const struct split_limits *limits)
+{
+   struct copy_context copy;
+   GLuint i, this_nr_prims;
+
+   for (i = 0; i < nr_prims;) {
+      /* Our SW TNL pipeline doesn't handle basevertex yet, so bind_indices
+       * will rebase the elements to the basevertex, and we'll only
+       * emit strings of prims with the same basevertex in one draw call.
+       */
+      for (this_nr_prims = 1; i + this_nr_prims < nr_prims;
+           this_nr_prims++) {
+         if (prim[i].basevertex != prim[i + this_nr_prims].basevertex)
+            break;
+      }
+
+      memset(&copy, 0, sizeof(copy));
+
+      /* Require indexed primitives:
+       */
+      assert(ib);
+
+      copy.ctx = ctx;
+      copy.array = arrays;
+      copy.prim = &prim[i];
+      copy.nr_prims = this_nr_prims;
+      copy.ib = ib;
+      copy.draw = draw;
+      copy.limits = limits;
+
+      /* Clear the vertex cache:
+       */
+      for (i = 0; i < ELT_TABLE_SIZE; i++)
+         copy.vert_cache[i].in = ~0;
+
+      replay_init(&copy);
+      replay_elts(&copy);
+      replay_finish(&copy);
+   }
+}
diff --git a/src/mesa/tnl/t_split_inplace.c b/src/mesa/tnl/t_split_inplace.c
new file mode 100644
index 00000000000..15a09861c73
--- /dev/null
+++ b/src/mesa/tnl/t_split_inplace.c
@@ -0,0 +1,298 @@
+
+/*
+ * Mesa 3-D graphics library
+ *
+ * Copyright (C) 1999-2006  Brian Paul   All Rights Reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included
+ * in all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
+ * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
+ * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ *
+ * Authors:
+ *    Keith Whitwell <keithw@vmware.com>
+ */
+
+
+#include "main/mtypes.h"
+#include "main/macros.h"
+#include "main/enums.h"
+#include "vbo/vbo.h"
+
+#include "t_split.h"
+
+
+#define MAX_PRIM 32
+
+/* Used for splitting without copying. No attempt is made to handle
+ * too large indexed vertex buffers: In general you need to copy to do
+ * that.
+ */
+struct split_context {
+   struct gl_context *ctx;
+   const struct gl_vertex_array *array;
+   const struct _mesa_prim *prim;
+   GLuint nr_prims;
+   const struct _mesa_index_buffer *ib;
+   GLuint min_index;
+   GLuint max_index;
+   tnl_draw_func draw;
+
+   const struct split_limits *limits;
+   GLuint limit;
+
+   struct _mesa_prim dstprim[MAX_PRIM];
+   GLuint dstprim_nr;
+};
+
+
+
+
+static void
+flush_vertex( struct split_context *split)
+{
+   struct gl_context *ctx = split->ctx;
+   struct _mesa_index_buffer ib;
+   GLuint i;
+
+   if (!split->dstprim_nr)
+      return;
+
+   if (split->ib) {
+      ib = *split->ib;
+
+      ib.count = split->max_index - split->min_index + 1;
+      ib.ptr = (const void *)((const char *)ib.ptr +
+                              split->min_index * ib.index_size);
+
+      /* Rebase the primitives to save index buffer entries. */
+      for (i = 0; i < split->dstprim_nr; i++)
+         split->dstprim[i].start -= split->min_index;
+   }
+
+   assert(split->max_index >= split->min_index);
+
+   split->draw(ctx,
+               split->array,
+               split->dstprim,
+               split->dstprim_nr,
+               split->ib ? &ib : NULL,
+               !split->ib,
+               split->min_index,
+               split->max_index,
+               NULL, 0, NULL);
+
+   split->dstprim_nr = 0;
+   split->min_index = ~0;
+   split->max_index = 0;
+}
+
+
+static struct _mesa_prim *
+next_outprim(struct split_context *split)
+{
+   if (split->dstprim_nr == MAX_PRIM-1) {
+      flush_vertex(split);
+   }
+
+   {
+      struct _mesa_prim *prim = &split->dstprim[split->dstprim_nr++];
+      memset(prim, 0, sizeof(*prim));
+      return prim;
+   }
+}
+
+
+static void
+update_index_bounds(struct split_context *split,
+                    const struct _mesa_prim *prim)
+{
+   split->min_index = MIN2(split->min_index, prim->start);
+   split->max_index = MAX2(split->max_index, prim->start + prim->count - 1);
+}
+
+
+/* Return the maximum amount of vertices that can be emitted for a
+ * primitive starting at 'prim->start', depending on the previous
+ * index bounds.
+ */
+static GLuint
+get_max_vertices(struct split_context *split,
+                 const struct _mesa_prim *prim)
+{
+   if ((prim->start > split->min_index &&
+        prim->start - split->min_index >= split->limit) ||
+       (prim->start < split->max_index &&
+        split->max_index - prim->start >= split->limit))
+      /* "prim" starts too far away from the old range. */
+      return 0;
+
+   return MIN2(split->min_index, prim->start) + split->limit - prim->start;
+}
+
+
+/* Break large primitives into smaller ones.  If not possible, convert
+ * the primitive to indexed and pass to split_elts().
+ */
+static void
+split_prims(struct split_context *split)
+{
+   GLuint i;
+
+   for (i = 0; i < split->nr_prims; i++) {
+      const struct _mesa_prim *prim = &split->prim[i];
+      GLuint first, incr;
+      GLboolean split_inplace =
+         _tnl_split_prim_inplace(prim->mode, &first, &incr);
+      GLuint available = get_max_vertices(split, prim);
+      GLuint count = prim->count - (prim->count - first) % incr;
+
+      if (prim->count < first)
+         continue;
+
+      if ((available < count && !split_inplace) ||
+          (available < first && split_inplace)) {
+         flush_vertex(split);
+         available = get_max_vertices(split, prim);
+      }
+
+      if (available >= count) {
+         struct _mesa_prim *outprim = next_outprim(split);
+
+         *outprim = *prim;
+         update_index_bounds(split, outprim);
+      }
+      else if (split_inplace) {
+         GLuint j, nr;
+
+         for (j = 0 ; j < count ;) {
+            GLuint remaining = count - j;
+            struct _mesa_prim *outprim = next_outprim(split);
+
+            nr = MIN2(available, remaining);
+            nr -= (nr - first) % incr;
+
+            outprim->mode = prim->mode;
+            outprim->begin = (j == 0 && prim->begin);
+            outprim->end = (nr == remaining && prim->end);
+            outprim->start = prim->start + j;
+            outprim->count = nr;
+            outprim->num_instances = prim->num_instances;
+            outprim->base_instance = prim->base_instance;
+
+            update_index_bounds(split, outprim);
+
+            if (nr == remaining) {
+               /* Finished */
+               j += nr;
+            }
+            else {
+               /* Wrapped the primitive */
+               j += nr - (first - incr);
+               flush_vertex(split);
+               available = get_max_vertices(split, prim);
+            }
+         }
+      }
+      else if (split->ib == NULL) {
+         /* XXX: could at least send the first max_verts off from the
+          * inplace buffers.
+          */
+
+         /* else convert to indexed primitive and pass to split_elts,
+          * which will do the necessary copying and turn it back into a
+          * vertex primitive for rendering...
+          */
+         struct _mesa_index_buffer ib;
+         struct _mesa_prim tmpprim;
+         GLuint *elts = malloc(count * sizeof(GLuint));
+         GLuint j;
+
+         for (j = 0; j < count; j++)
+            elts[j] = prim->start + j;
+
+         ib.count = count;
+         ib.index_size = 4;
+         ib.obj = split->ctx->Shared->NullBufferObj;
+         ib.ptr = elts;
+
+         tmpprim = *prim;
+         tmpprim.indexed = 1;
+         tmpprim.start = 0;
+         tmpprim.count = count;
+         tmpprim.num_instances = 1;
+         tmpprim.base_instance = 0;
+
+         flush_vertex(split);
+
+         _tnl_split_copy(split->ctx,
+                         split->array,
+                         &tmpprim, 1,
+                         &ib,
+                         split->draw,
+                         split->limits);
+
+         free(elts);
+      }
+      else {
+         flush_vertex(split);
+
+         _tnl_split_copy(split->ctx,
+                         split->array,
+                         prim, 1,
+                         split->ib,
+                         split->draw,
+                         split->limits);
+      }
+   }
+
+   flush_vertex(split);
+}
+
+
+void
+_tnl_split_inplace(struct gl_context *ctx,
+                   const struct gl_vertex_array *arrays,
+                   const struct _mesa_prim *prim,
+                   GLuint nr_prims,
+                   const struct _mesa_index_buffer *ib,
+                   GLuint min_index,
+                   GLuint max_index,
+                   tnl_draw_func draw,
+                   const struct split_limits *limits)
+{
+   struct split_context split;
+
+   memset(&split, 0, sizeof(split));
+
+   split.ctx = ctx;
+   split.array = arrays;
+   split.prim = prim;
+   split.nr_prims = nr_prims;
+   split.ib = ib;
+
+   /* Empty interval, makes calculations simpler. */
+   split.min_index = ~0;
+   split.max_index = 0;
+
+   split.draw = draw;
+   split.limits = limits;
+   split.limit = ib ? limits->max_indices : limits->max_verts;
+
+   split_prims(&split);
+}
+
+
diff --git a/src/mesa/tnl/tnl.h b/src/mesa/tnl/tnl.h
index e79c4f62048..45052a3a89c 100644
--- a/src/mesa/tnl/tnl.h
+++ b/src/mesa/tnl/tnl.h
@@ -108,4 +108,84 @@ _tnl_RasterPos(struct gl_context *ctx, const GLfloat vObj[4]);
 extern void
 _tnl_validate_shine_tables( struct gl_context *ctx );
 
+
+
+/**
+ * For indirect array drawing:
+ *
+ *    typedef struct {
+ *       GLuint count;
+ *       GLuint primCount;
+ *       GLuint first;
+ *       GLuint baseInstance; // in GL 4.2 and later, must be zero otherwise
+ *    } DrawArraysIndirectCommand;
+ *
+ * For indirect indexed drawing:
+ *
+ *    typedef struct {
+ *       GLuint count;
+ *       GLuint primCount;
+ *       GLuint firstIndex;
+ *       GLint  baseVertex;
+ *       GLuint baseInstance; // in GL 4.2 and later, must be zero otherwise
+ *    } DrawElementsIndirectCommand;
+ */
+
+
+/**
+ * Draw a number of primitives.
+ * \param prims  array [nr_prims] describing what to draw (prim type,
+ *               vertex count, first index, instance count, etc).
+ * \param arrays array of vertex arrays for draw
+ * \param ib  index buffer for indexed drawing, NULL for array drawing
+ * \param index_bounds_valid  are min_index and max_index valid?
+ * \param min_index  lowest vertex index used
+ * \param max_index  highest vertex index used
+ * \param tfb_vertcount  if non-null, indicates which transform feedback
+ *                       object has the vertex count.
+ * \param tfb_stream  If called via DrawTransformFeedbackStream, specifies the
+ *                    vertex stream buffer from which to get the vertex count.
+ * \param indirect  If any prims are indirect, this specifies the buffer
+ *                  to find the "DrawArrays/ElementsIndirectCommand" data.
+ *                  This may be deprecated in the future
+ */
+typedef void (*tnl_draw_func)(struct gl_context *ctx,
+                              const struct gl_vertex_array* arrays,
+                              const struct _mesa_prim *prims,
+                              GLuint nr_prims,
+                              const struct _mesa_index_buffer *ib,
+                              GLboolean index_bounds_valid,
+                              GLuint min_index,
+                              GLuint max_index,
+                              struct gl_transform_feedback_object *tfb_vertcount,
+                              unsigned tfb_stream,
+                              struct gl_buffer_object *indirect);
+
+
+/* Utility function to cope with various constraints on tnl modules or
+ * hardware.  This can be used to split an incoming set of arrays and
+ * primitives against the following constraints:
+ *    - Maximum number of indices in index buffer.
+ *    - Maximum number of vertices referenced by index buffer.
+ *    - Maximum hardware vertex buffer size.
+ */
+struct split_limits
+{
+   GLuint max_verts;
+   GLuint max_indices;
+   GLuint max_vb_size;		/* bytes */
+};
+
+void
+_tnl_split_prims(struct gl_context *ctx,
+                 const struct gl_vertex_array *arrays,
+                 const struct _mesa_prim *prim,
+                 GLuint nr_prims,
+                 const struct _mesa_index_buffer *ib,
+                 GLuint min_index,
+                 GLuint max_index,
+                 tnl_draw_func draw,
+                 const struct split_limits *limits);
+
+
 #endif