The previous code would emit a full set of state during the first EmitState on

a new cmdbuf, to ensure that state wasn't lost across UNLOCK/LOCK pairs (in the
case of context switching).  This was rather inefficient.  Instead, after
flushing a cmdbuf, mark the state as needing to be saved on UNLOCK.  Then, at
the beginning of flushing a cmdbuf, if we actually have lost the context, go
back and emit a new cmdbuf with the full set of state, before continuing with
the cmdbuf flush.  Also, remove the dirty/clean atom lists, since atoms are
emitted in a fixed order these days, and go with a simpler single list.

Provides a 14% improvement in ipers performance in my tests, along with other
apps.

8 files changed, 168 insertions, 125 deletions

diff --git a/src/mesa/drivers/dri/r200/r200_cmdbuf.c b/src/mesa/drivers/dri/r200/r200_cmdbuf.c
index fa0c62385b0..e3829f9a957 100644
--- a/src/mesa/drivers/dri/r200/r200_cmdbuf.c
+++ b/src/mesa/drivers/dri/r200/r200_cmdbuf.c
@@ -58,113 +58,105 @@ static void print_state_atom( struct r200_state_atom *state )
 
 }
 
-static void r200_emit_state_list( r200ContextPtr rmesa, 
-				    struct r200_state_atom *list )
+/* The state atoms will be emitted in the order they appear in the atom list,
+ * so this step is important.
+ */
+void r200SetUpAtomList( r200ContextPtr rmesa )
 {
-   struct r200_state_atom *state, *tmp;
-   char *dest;
-   int i, size, mtu;
-
-   size = 0;
-   foreach_s( state, tmp, list ) {
-      if (state->check( rmesa->glCtx, state->idx )) {
-/*	 dest = r200AllocCmdBuf( rmesa, state->cmd_size * 4, __FUNCTION__);
-	 memcpy( dest, state->cmd, state->cmd_size * 4);*/
-         size += state->cmd_size;
-         state->dirty = GL_TRUE;
-	 move_to_head( &(rmesa->hw.clean), state );
-	 if (R200_DEBUG & DEBUG_STATE) 
-	    print_state_atom( state );
-      }
-      else if (R200_DEBUG & DEBUG_STATE)
-	 fprintf(stderr, "skip state %s\n", state->name);
-   }
-
-   if (!size)
-      return;
+   int i, mtu;
 
-   dest = r200AllocCmdBuf( rmesa, size * 4, __FUNCTION__);
    mtu = rmesa->glCtx->Const.MaxTextureUnits;
 
-#define EMIT_ATOM(ATOM) \
-do { \
-   if (rmesa->hw.ATOM.dirty) { \
-      rmesa->hw.ATOM.dirty = GL_FALSE; \
-      memcpy( dest, rmesa->hw.ATOM.cmd, rmesa->hw.ATOM.cmd_size * 4); \
-      dest += rmesa->hw.ATOM.cmd_size * 4; \
-   } \
-} while (0)
-
-   EMIT_ATOM (ctx);
-   EMIT_ATOM (set);
-   EMIT_ATOM (lin);
-   EMIT_ATOM (msk);
-   EMIT_ATOM (vpt);
-   EMIT_ATOM (vtx);
-   EMIT_ATOM (vap);
-   EMIT_ATOM (vte);
-   EMIT_ATOM (msc);
-   EMIT_ATOM (cst);
-   EMIT_ATOM (zbs);
-   EMIT_ATOM (tcl);
-   EMIT_ATOM (msl);
-   EMIT_ATOM (tcg);
-   EMIT_ATOM (grd);
-   EMIT_ATOM (fog);
-   EMIT_ATOM (tam);
-   EMIT_ATOM (tf);
-   for (i = 0; i < mtu; ++i) {
-       EMIT_ATOM (tex[i]);
-   }
-   for (i = 0; i < mtu; ++i) {
-       EMIT_ATOM (cube[i]);
-   }
+   make_empty_list(&rmesa->hw.atomlist);
+   rmesa->hw.atomlist.name = "atom-list";
+
+   insert_at_tail( &rmesa->hw.atomlist, &rmesa->hw.ctx );
+   insert_at_tail( &rmesa->hw.atomlist, &rmesa->hw.set );
+   insert_at_tail( &rmesa->hw.atomlist, &rmesa->hw.lin );
+   insert_at_tail( &rmesa->hw.atomlist, &rmesa->hw.msk );
+   insert_at_tail( &rmesa->hw.atomlist, &rmesa->hw.vpt );
+   insert_at_tail( &rmesa->hw.atomlist, &rmesa->hw.vtx );
+   insert_at_tail( &rmesa->hw.atomlist, &rmesa->hw.vap );
+   insert_at_tail( &rmesa->hw.atomlist, &rmesa->hw.vte );
+   insert_at_tail( &rmesa->hw.atomlist, &rmesa->hw.msc );
+   insert_at_tail( &rmesa->hw.atomlist, &rmesa->hw.cst );
+   insert_at_tail( &rmesa->hw.atomlist, &rmesa->hw.zbs );
+   insert_at_tail( &rmesa->hw.atomlist, &rmesa->hw.tcl );
+   insert_at_tail( &rmesa->hw.atomlist, &rmesa->hw.msl );
+   insert_at_tail( &rmesa->hw.atomlist, &rmesa->hw.tcg );
+   insert_at_tail( &rmesa->hw.atomlist, &rmesa->hw.grd );
+   insert_at_tail( &rmesa->hw.atomlist, &rmesa->hw.fog );
+   insert_at_tail( &rmesa->hw.atomlist, &rmesa->hw.tam );
+   insert_at_tail( &rmesa->hw.atomlist, &rmesa->hw.tf );
+   for (i = 0; i < mtu; ++i)
+       insert_at_tail( &rmesa->hw.atomlist, &rmesa->hw.tex[i] );
+   for (i = 0; i < mtu; ++i)
+       insert_at_tail( &rmesa->hw.atomlist, &rmesa->hw.cube[i] );
    for (i = 0; i < 3 + mtu; ++i)
-       EMIT_ATOM (mat[i]);
-   EMIT_ATOM (eye);
-   EMIT_ATOM (glt);
-   for (i = 0; i < 2; ++i) {
-      EMIT_ATOM (mtl[i]);
-   }
+       insert_at_tail( &rmesa->hw.atomlist, &rmesa->hw.mat[i] );
+   insert_at_tail( &rmesa->hw.atomlist, &rmesa->hw.eye );
+   insert_at_tail( &rmesa->hw.atomlist, &rmesa->hw.glt );
+   for (i = 0; i < 2; ++i)
+      insert_at_tail( &rmesa->hw.atomlist, &rmesa->hw.mtl[i] );
    for (i = 0; i < 8; ++i)
-       EMIT_ATOM (lit[i]);
+       insert_at_tail( &rmesa->hw.atomlist, &rmesa->hw.lit[i] );
    for (i = 0; i < 6; ++i)
-       EMIT_ATOM (ucp[i]);
+       insert_at_tail( &rmesa->hw.atomlist, &rmesa->hw.ucp[i] );
    for (i = 0; i < 6; ++i)
-       EMIT_ATOM (pix[i]);
-
-#undef EMIT_ATOM
-
+       insert_at_tail( &rmesa->hw.atomlist, &rmesa->hw.pix[i] );
 }
 
-
 void r200EmitState( r200ContextPtr rmesa )
 {
-   struct r200_state_atom *state, *tmp;
+   char *dest;
+   int i, mtu;
+   struct r200_state_atom *atom;
 
    if (R200_DEBUG & (DEBUG_STATE|DEBUG_PRIMS))
       fprintf(stderr, "%s\n", __FUNCTION__);
 
-   /* Somewhat overkill:
-    */
-   if ( rmesa->lost_context) {
-      if (R200_DEBUG & (DEBUG_STATE|DEBUG_PRIMS|DEBUG_IOCTL))
-	 fprintf(stderr, "%s - lost context\n", __FUNCTION__); 
+   if (!rmesa->hw.is_dirty && !rmesa->hw.all_dirty)
+      return;
 
-      foreach_s( state, tmp, &(rmesa->hw.clean) ) 
-	 move_to_tail(&(rmesa->hw.dirty), state );
+   mtu = rmesa->glCtx->Const.MaxTextureUnits;
 
-      rmesa->lost_context = 0;
+   /* To avoid going across the entire set of states multiple times, just check
+    * for enough space for the case of emitting all state, and inline the
+    * r200AllocCmdBuf code here without all the checks.
+    */
+   dest = rmesa->store.cmd_buf + rmesa->store.cmd_used;
+   r200EnsureCmdBufSpace( rmesa, rmesa->hw.max_state_size );
+
+   if (R200_DEBUG & DEBUG_STATE) {
+      foreach( atom, &rmesa->hw.atomlist ) {
+	 if ( atom->dirty || rmesa->hw.all_dirty ) {
+	    if ( atom->check( rmesa->glCtx, atom->idx ) )
+	       print_state_atom( atom );
+	    else
+	       fprintf(stderr, "skip state %s\n", atom->name);
+	 }
+      }
    }
-/*   else {
-      move_to_tail( &rmesa->hw.dirty, &rmesa->hw.mtl[0] );*/
-      /* odd bug? -- isosurf, cycle between reflect & lit */
-/*   }*/
 
-   r200_emit_state_list( rmesa, &rmesa->hw.dirty );
-}
+   foreach( atom, &rmesa->hw.atomlist ) {
+      if ( rmesa->hw.all_dirty )
+	 atom->dirty = GL_TRUE;
+      if ( atom->dirty ) {
+	 if ( atom->check( rmesa->glCtx, atom->idx ) ) {
+	    int size = atom->cmd_size * 4;
+	    memcpy( dest, atom->cmd, size);
+	    dest += size;
+	    rmesa->store.cmd_used += size;
+	    atom->dirty = GL_FALSE;
+	 }
+      }
+   }
 
+   assert( rmesa->store.cmd_used <= R200_CMD_BUF_SZ );
 
+   rmesa->hw.is_dirty = GL_FALSE;
+   rmesa->hw.all_dirty = GL_FALSE;
+}
 
 /* Fire a section of the retained (indexed_verts) buffer as a regular
  * primtive.  
diff --git a/src/mesa/drivers/dri/r200/r200_context.c b/src/mesa/drivers/dri/r200/r200_context.c
index 4151b1598ed..c5fdc86a29e 100644
--- a/src/mesa/drivers/dri/r200/r200_context.c
+++ b/src/mesa/drivers/dri/r200/r200_context.c
@@ -323,7 +323,7 @@ GLboolean r200CreateContext( const __GLcontextModes *glVisual,
 	 DRI_CONF_TEXTURE_DEPTH_32 : DRI_CONF_TEXTURE_DEPTH_16;
 
    rmesa->swtcl.RenderIndex = ~0;
-   rmesa->lost_context = 1;
+   rmesa->hw.all_dirty = 1;
 
    /* Set the maximum texture size small enough that we can guarentee that
     * all texture units can bind a maximal texture and have them both in
diff --git a/src/mesa/drivers/dri/r200/r200_context.h b/src/mesa/drivers/dri/r200/r200_context.h
index f000e143308..25faec955c5 100644
--- a/src/mesa/drivers/dri/r200/r200_context.h
+++ b/src/mesa/drivers/dri/r200/r200_context.h
@@ -191,6 +191,7 @@ struct r200_state_atom {
    GLuint idx;
    int *cmd;			         /* one or more cmd's */
    int *lastcmd;			 /* one or more cmd's */
+   int *savedcmd;			 /* one or more cmd's */
    GLboolean dirty;
    GLboolean (*check)( GLcontext *, int );    /* is this state active? */
 };
@@ -491,10 +492,8 @@ struct r200_state_atom {
 
 
 struct r200_hw_state {
-   /* All state should be on one of these lists:
-    */
-   struct r200_state_atom dirty; /* dirty list head placeholder */
-   struct r200_state_atom clean; /* clean list head placeholder */
+   /* Head of the linked list of state atoms. */
+   struct r200_state_atom atomlist;
 
    /* Hardware state, stored as cmdbuf commands:  
     *   -- Need to doublebuffer for
@@ -530,6 +529,7 @@ struct r200_hw_state {
    struct r200_state_atom glt; 
 
    int max_state_size;	/* Number of bytes necessary for a full state emit. */
+   GLboolean is_dirty, all_dirty;
 };
 
 struct r200_state {
@@ -876,6 +876,7 @@ struct r200_context {
    drm_clip_rect_t *pClipRects;
    unsigned int lastStamp;
    GLboolean lost_context;
+   GLboolean save_on_next_unlock;
    r200ScreenPtr r200Screen;	/* Screen private DRI data */
    drm_radeon_sarea_t *sarea;		/* Private SAREA data */
 
diff --git a/src/mesa/drivers/dri/r200/r200_ioctl.c b/src/mesa/drivers/dri/r200/r200_ioctl.c
index fb462e08509..955afb0d5f3 100644
--- a/src/mesa/drivers/dri/r200/r200_ioctl.c
+++ b/src/mesa/drivers/dri/r200/r200_ioctl.c
@@ -58,12 +58,71 @@ WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 
 static void r200WaitForIdle( r200ContextPtr rmesa );
 
+void r200SaveHwState( r200ContextPtr rmesa )
+{
+   struct r200_state_atom *atom;
+
+   foreach( atom, &rmesa->hw.atomlist )
+      memcpy(atom->savedcmd, atom->cmd, atom->cmd_size * 4);
+}
+
+static void r200SwapHwState( r200ContextPtr rmesa )
+{
+   int *temp;
+   struct r200_state_atom *atom;
+
+   foreach( atom, &rmesa->hw.atomlist ) {
+      temp = atom->cmd;
+      atom->cmd = atom->savedcmd;
+      atom->savedcmd = temp;
+   }
+}
+
+/* At this point we were in FlushCmdBufLocked but we had lost our context, so
+ * we need to unwire our current cmdbuf and hook a new one in, emit that, then
+ * wire the old cmdbuf back in so that FlushCmdBufLocked can continue and the
+ * buffer can depend on the state not being lost across lock/unlock.
+ */
+static void r200BackUpAndEmitLostStateLocked( r200ContextPtr rmesa )
+{
+   GLuint nr_released_bufs;
+   struct r200_store store;
+   struct r200_hw_state temp_state;
+   static int count = 0;
+
+   rmesa->lost_context = GL_FALSE;
+
+   nr_released_bufs = rmesa->dma.nr_released_bufs;
+   store = rmesa->store;
+   rmesa->store.statenr = 0;
+   rmesa->store.primnr = 0;
+   rmesa->store.cmd_used = 0;
+   rmesa->store.elts_start = 0;
+   rmesa->hw.all_dirty = GL_TRUE;
+   r200SwapHwState( rmesa );
+   /* In this case it's okay to EmitState while locked because we won't exhaust
+    * our (empty) cmdbuf.
+    */
+   r200EmitState( rmesa );
+   r200FlushCmdBufLocked( rmesa, __FUNCTION__ );
+
+   r200SwapHwState( rmesa );
+   /* We've just cleared out the dirty flags, so we don't remember what 
+    * actually needed to be emitted for the next state emit.
+    */
+   rmesa->hw.all_dirty = GL_TRUE;
+   rmesa->dma.nr_released_bufs = nr_released_bufs;
+   rmesa->store = store;
+}
 
 int r200FlushCmdBufLocked( r200ContextPtr rmesa, const char * caller )
 {
    int ret, i;
    drm_radeon_cmd_buffer_t cmd;
 
+   if (rmesa->lost_context)
+      r200BackUpAndEmitLostStateLocked( rmesa );
+
    if (R200_DEBUG & DEBUG_IOCTL) {
       fprintf(stderr, "%s from %s\n", __FUNCTION__, caller); 
 
@@ -132,18 +191,7 @@ int r200FlushCmdBufLocked( r200ContextPtr rmesa, const char * caller )
    rmesa->store.statenr = 0;
    rmesa->store.cmd_used = 0;
    rmesa->dma.nr_released_bufs = 0;
-   /* Set lost_context so that the first state emit on the new buffer is a full
-    * one.  This is because the context might get lost while preparing the next
-    * buffer, and when we lock and find out, we don't have the information to
-    * recreate the state.  This function should always be called before the new
-    * buffer is begun, so it's sufficient to just set lost_context here.
-    *
-    * The alternative to this would be to copy out the state on unlock
-    * (approximately) and if we did lose the context, dispatch a cmdbuf to reset
-    * the state to that old copy before continuing with the accumulated command
-    * buffer.
-    */
-   rmesa->lost_context = 1;
+   rmesa->save_on_next_unlock = 1;
 
    return ret;
 }
@@ -464,7 +512,7 @@ void r200CopyBuffer( const __DRIdrawablePrivate *dPriv )
    }
 
    UNLOCK_HARDWARE( rmesa );
-   rmesa->lost_context = 1;
+   rmesa->hw.all_dirty = GL_TRUE;
 
    rmesa->swap_count++;
    (*rmesa->get_ust)( & ust );
@@ -613,13 +661,6 @@ static void r200Clear( GLcontext *ctx, GLbitfield mask, GLboolean all,
    cx += dPriv->x;
    cy  = dPriv->y + dPriv->h - cy - ch;
 
-   /* We have to emit state along with the clear, since the kernel relies on
-    * some of it.  The EmitState that was above R200_FIREVERTICES was an
-    * attempt to do that, except that another context may come in and cause us
-    * to lose our context while we're unlocked.
-    */
-   r200EmitState( rmesa );
-
    LOCK_HARDWARE( rmesa );
 
    /* Throttle the number of clear ioctls we do.
@@ -722,7 +763,7 @@ static void r200Clear( GLcontext *ctx, GLbitfield mask, GLboolean all,
    }
 
    UNLOCK_HARDWARE( rmesa );
-   rmesa->lost_context = 1;
+   rmesa->hw.all_dirty = GL_TRUE;
 }
 
 
@@ -763,8 +804,7 @@ void r200Flush( GLcontext *ctx )
    if (rmesa->dma.flush)
       rmesa->dma.flush( rmesa );
 
-   if (!is_empty_list(&rmesa->hw.dirty)) 
-      r200EmitState( rmesa );
+   r200EmitState( rmesa );
    
    if (rmesa->store.cmd_used)
       r200FlushCmdBuf( rmesa, __FUNCTION__ );
diff --git a/src/mesa/drivers/dri/r200/r200_ioctl.h b/src/mesa/drivers/dri/r200/r200_ioctl.h
index 1503df70754..57474d7a052 100644
--- a/src/mesa/drivers/dri/r200/r200_ioctl.h
+++ b/src/mesa/drivers/dri/r200/r200_ioctl.h
@@ -117,6 +117,9 @@ extern GLboolean r200IsGartMemory( r200ContextPtr rmesa, const GLvoid *pointer,
 extern GLuint r200GartOffsetFromVirtual( r200ContextPtr rmesa, 
 					 const GLvoid *pointer );
 
+void r200SaveHwState( r200ContextPtr rmesa );
+void r200SetUpAtomList( r200ContextPtr rmesa );
+
 /* ================================================================
  * Helper macros:
  */
@@ -135,7 +138,8 @@ do {						\
 #define R200_STATECHANGE( rmesa, ATOM )			\
 do {								\
    R200_NEWPRIM( rmesa );					\
-   move_to_head( &(rmesa->hw.dirty), &(rmesa->hw.ATOM));	\
+   rmesa->hw.ATOM.dirty = GL_TRUE;				\
+   rmesa->hw.is_dirty = GL_TRUE;				\
 } while (0)
 
 #define R200_DB_STATE( ATOM )			        \
@@ -149,7 +153,8 @@ static __inline int R200_DB_STATECHANGE(
    if (memcmp(atom->cmd, atom->lastcmd, atom->cmd_size*4)) {
       int *tmp;
       R200_NEWPRIM( rmesa );
-      move_to_head( &(rmesa->hw.dirty), atom );
+      atom->dirty = GL_TRUE;
+      rmesa->hw.is_dirty = GL_TRUE;
       tmp = atom->cmd; 
       atom->cmd = atom->lastcmd;
       atom->lastcmd = tmp;
diff --git a/src/mesa/drivers/dri/r200/r200_lock.c b/src/mesa/drivers/dri/r200/r200_lock.c
index 1700d31e169..539a4be619b 100644
--- a/src/mesa/drivers/dri/r200/r200_lock.c
+++ b/src/mesa/drivers/dri/r200/r200_lock.c
@@ -116,4 +116,6 @@ void r200GetLock( r200ContextPtr rmesa, GLuint flags )
    for ( i = 0 ; i < rmesa->nr_heaps ; i++ ) {
       DRI_AGE_TEXTURES( rmesa->texture_heaps[ i ] );
    }
+
+   rmesa->lost_context = GL_TRUE;
 }
diff --git a/src/mesa/drivers/dri/r200/r200_lock.h b/src/mesa/drivers/dri/r200/r200_lock.h
index 587e4fe5ccd..908052659c1 100644
--- a/src/mesa/drivers/dri/r200/r200_lock.h
+++ b/src/mesa/drivers/dri/r200/r200_lock.h
@@ -104,6 +104,10 @@ extern int prevLockLine;
 		  rmesa->dri.hwLock,					\
 		  rmesa->dri.hwContext );				\
       DEBUG_RESET();							\
+      if (rmesa->save_on_next_unlock) {					\
+	 r200SaveHwState( rmesa );					\
+	 rmesa->save_on_next_unlock = GL_FALSE;				\
+      }									\
    } while (0)
 
 #endif
diff --git a/src/mesa/drivers/dri/r200/r200_state_init.c b/src/mesa/drivers/dri/r200/r200_state_init.c
index 3b6893aeee9..e97a2f4d995 100644
--- a/src/mesa/drivers/dri/r200/r200_state_init.c
+++ b/src/mesa/drivers/dri/r200/r200_state_init.c
@@ -65,8 +65,9 @@ void r200PrintDirty( r200ContextPtr rmesa, const char *msg )
    fprintf(stderr, msg);
    fprintf(stderr, ": ");
 
-   foreach(l, &(rmesa->hw.dirty)) {
-      fprintf(stderr, "%s, ", l->name);
+   foreach(l, &rmesa->hw.atomlist) {
+      if (l->dirty || rmesa->hw.all_dirty)
+	 fprintf(stderr, "%s, ", l->name);
    }
 
    fprintf(stderr, "\n");
@@ -200,11 +201,6 @@ void r200InitState( r200ContextPtr rmesa )
    rmesa->state.pixel.readOffset = rmesa->state.color.drawOffset;
    rmesa->state.pixel.readPitch  = rmesa->state.color.drawPitch;
 
-   /* Initialize lists:
-    */
-   make_empty_list(&(rmesa->hw.dirty)); rmesa->hw.dirty.name = "DIRTY";
-   make_empty_list(&(rmesa->hw.clean)); rmesa->hw.clean.name = "CLEAN";
-
    rmesa->hw.max_state_size = 0;
 
 #define ALLOC_STATE( ATOM, CHK, SZ, NM, IDX )				\
@@ -212,10 +208,11 @@ void r200InitState( r200ContextPtr rmesa )
       rmesa->hw.ATOM.cmd_size = SZ;				\
       rmesa->hw.ATOM.cmd = (int *)CALLOC(SZ * sizeof(int));	\
       rmesa->hw.ATOM.lastcmd = (int *)CALLOC(SZ * sizeof(int));	\
+      rmesa->hw.ATOM.savedcmd = (int *)CALLOC(SZ * sizeof(int)); \
       rmesa->hw.ATOM.name = NM;					\
       rmesa->hw.ATOM.idx = IDX;					\
-      rmesa->hw.ATOM.check = check_##CHK;				\
-      insert_at_head(&(rmesa->hw.dirty), &(rmesa->hw.ATOM));	\
+      rmesa->hw.ATOM.check = check_##CHK;			\
+      rmesa->hw.ATOM.dirty = GL_FALSE;				\
       rmesa->hw.max_state_size += SZ * sizeof(int);		\
    } while (0)
       
@@ -308,6 +305,7 @@ void r200InitState( r200ContextPtr rmesa )
    ALLOC_STATE( pix[4], tex, PIX_STATE_SIZE, "PIX/pixstage-4", 4 );
    ALLOC_STATE( pix[5], tex, PIX_STATE_SIZE, "PIX/pixstage-5", 5 );
 
+   r200SetUpAtomList( rmesa );
 
    /* Fill in the packet headers:
     */
@@ -772,5 +770,6 @@ void r200InitState( r200ContextPtr rmesa )
 
    r200LightingSpaceChange( ctx );
    
-   rmesa->lost_context = 1;
+   r200SaveHwState( rmesa );
+   rmesa->hw.all_dirty = GL_TRUE;
 }