4 files changed, 30 insertions, 7 deletions

diff --git a/src/gallium/drivers/cell/ppu/cell_batch.c b/src/gallium/drivers/cell/ppu/cell_batch.c
index 448b723d85a..962775cd335 100644
--- a/src/gallium/drivers/cell/ppu/cell_batch.c
+++ b/src/gallium/drivers/cell/ppu/cell_batch.c
@@ -100,12 +100,23 @@ emit_fence(struct cell_context *cell)
    const uint batch = cell->cur_batch;
    const uint size = cell->buffer_size[batch];
    struct cell_command_fence *fence_cmd;
+   struct cell_fence *fence = &cell->fenced_buffers[batch].fence;
+   uint i;
+
+   /* set fence status to emitted, not yet signalled */
+   for (i = 0; i < cell->num_spus; i++) {
+      fence->status[i][0] = CELL_FENCE_EMITTED;
+   }
 
    ASSERT(size + sizeof(struct cell_command_fence) <= CELL_BUFFER_SIZE);
 
    fence_cmd = (struct cell_command_fence *) (cell->buffer[batch] + size);
    fence_cmd->opcode = CELL_CMD_FENCE;
-   fence_cmd->fence = &cell->fenced_buffers[batch].fence;
+   fence_cmd->fence = fence;
+
+   /* update batch buffer size */
+   cell->buffer_size[batch] = size + sizeof(struct cell_command_fence);
+   assert(sizeof(struct cell_command_fence) % 8 == 0);
 }
 
 
@@ -119,7 +130,7 @@ cell_batch_flush(struct cell_context *cell)
 {
    static boolean flushing = FALSE;
    uint batch = cell->cur_batch;
-   const uint size = cell->buffer_size[batch];
+   uint size = cell->buffer_size[batch];
    uint spu, cmd_word;
 
    assert(!flushing);
@@ -130,8 +141,10 @@ cell_batch_flush(struct cell_context *cell)
    /* Before we use this batch buffer, make sure any fenced texture buffers
     * are released.
     */
-   if (cell->fenced_buffers[batch].head)
+   if (cell->fenced_buffers[batch].head) {
       emit_fence(cell);
+      size = cell->buffer_size[batch];
+   }
 
    flushing = TRUE;
 
diff --git a/src/gallium/drivers/cell/ppu/cell_context.h b/src/gallium/drivers/cell/ppu/cell_context.h
index 4491ae8cdf1..eb1397bb3fa 100644
--- a/src/gallium/drivers/cell/ppu/cell_context.h
+++ b/src/gallium/drivers/cell/ppu/cell_context.h
@@ -89,7 +89,7 @@ struct cell_buffer_node;
  */
 struct cell_buffer_list
 {
-   struct cell_fence fence;
+   struct cell_fence fence ALIGN16_ATTRIB;
    struct cell_buffer_node *head;
 };
 
diff --git a/src/gallium/drivers/cell/ppu/cell_fence.c b/src/gallium/drivers/cell/ppu/cell_fence.c
index ffb3bea12b9..867b5dcaa09 100644
--- a/src/gallium/drivers/cell/ppu/cell_fence.c
+++ b/src/gallium/drivers/cell/ppu/cell_fence.c
@@ -38,6 +38,7 @@ void
 cell_fence_init(struct cell_fence *fence)
 {
    uint i;
+   ASSERT_ALIGN16(fence->status);
    for (i = 0; i < CELL_MAX_SPUS; i++) {
       fence->status[i][0] = CELL_FENCE_IDLE;
    }
@@ -50,9 +51,9 @@ cell_fence_signalled(const struct cell_context *cell,
 {
    uint i;
    for (i = 0; i < cell->num_spus; i++) {
-      //ASSERT(fence->status[i][0] != CELL_FENCE_IDLE);
-      if (fence->status[i][0] == CELL_FENCE_EMITTED)
+      if (fence->status[i][0] != CELL_FENCE_SIGNALLED)
          return FALSE;
+      /*assert(fence->status[i][0] == CELL_FENCE_EMITTED);*/
    }
    return TRUE;
 }
@@ -65,6 +66,15 @@ cell_fence_finish(const struct cell_context *cell,
    while (!cell_fence_signalled(cell, fence)) {
       usleep(10);
    }
+
+#ifdef DEBUG
+   {
+      uint i;
+      for (i = 0; i < cell->num_spus; i++) {
+         assert(fence->status[i][0] == CELL_FENCE_SIGNALLED);
+      }
+   }
+#endif
 }
 
 
diff --git a/src/gallium/drivers/cell/spu/spu_command.c b/src/gallium/drivers/cell/spu/spu_command.c
index a6ed29ea631..63818d4c46a 100644
--- a/src/gallium/drivers/cell/spu/spu_command.c
+++ b/src/gallium/drivers/cell/spu/spu_command.c
@@ -107,7 +107,7 @@ cmd_fence(struct cell_command_fence *fence_cmd)
                                               CELL_FENCE_SIGNALLED};
    uint *dst = (uint *) fence_cmd->fence;
    dst += 4 * spu.init.id;  /* main store/memory address, not local store */
-
+   ASSERT_ALIGN16(dst);
    mfc_put((void *) &status,    /* src in local memory */
            (unsigned int) dst,  /* dst in main memory */
            sizeof(status),      /* size */