minimize tu_fifo size to 16

- remove non_used_index_space
- packed overwritable with item_size

3 files changed, 39 insertions, 40 deletions

diff --git a/src/common/tusb_fifo.c b/src/common/tusb_fifo.c
index d3875a7d4..4dfe7bd81 100644
--- a/src/common/tusb_fifo.c
+++ b/src/common/tusb_fifo.c
@@ -68,22 +68,19 @@ typedef enum
 
 bool tu_fifo_config(tu_fifo_t *f, void* buffer, uint16_t depth, uint16_t item_size, bool overwritable)
 {
-  if (depth > 0x8000) return false;               // Maximum depth is 2^15 items
+  // Limit index space to 2*depth - this allows for a fast "modulo" calculation
+  // but limits the maximum depth to 2^16/2 = 2^15 and buffer overflows are detectable
+  // only if overflow happens once (important for unsupervised DMA applications)
+  if (depth > 0x8000) return false;
 
   _ff_lock(f->mutex_wr);
   _ff_lock(f->mutex_rd);
 
-  f->buffer = (uint8_t*) buffer;
-  f->depth  = depth;
-  f->item_size = item_size;
+  f->buffer       = (uint8_t*) buffer;
+  f->depth        = depth;
+  f->item_size    = (uint16_t) (item_size & 0x7FFF);
   f->overwritable = overwritable;
 
-  // Limit index space to 2*depth - this allows for a fast "modulo" calculation
-  // but limits the maximum depth to 2^16/2 = 2^15 and buffer overflows are detectable
-  // only if overflow happens once (important for unsupervised DMA applications)
-  //f->max_pointer_idx = (uint16_t) (2*depth - 1);
-  f->non_used_index_space = (uint16_t) (UINT16_MAX - (2*f->depth-1));
-
   f->rd_idx = f->wr_idx = 0;
 
   _ff_unlock(f->mutex_wr);
@@ -331,7 +328,8 @@ static uint16_t advance_pointer(tu_fifo_t* f, uint16_t idx, uint16_t offset)
   uint16_t next_p = (uint16_t) (idx + offset);
   if ( (idx > next_p) || (next_p >= 2*f->depth) )
   {
-    next_p = (uint16_t) (next_p + f->non_used_index_space);
+    uint16_t const non_used_index_space = (uint16_t) (UINT16_MAX - (2*f->depth-1));
+    next_p = (uint16_t) (next_p + non_used_index_space);
   }
 
   return next_p;
@@ -346,7 +344,8 @@ static uint16_t backward_pointer(tu_fifo_t* f, uint16_t p, uint16_t offset)
   uint16_t new_p = (uint16_t) (p - offset);
   if ( (p < new_p) || (new_p >= 2*f->depth) )
   {
-    new_p = (uint16_t) (new_p - f->non_used_index_space);
+    uint16_t const non_used_index_space = (uint16_t) (UINT16_MAX - (2*f->depth-1));
+    new_p = (uint16_t) (new_p - non_used_index_space);
   }
 
   return new_p;
@@ -363,13 +362,15 @@ static inline uint16_t idx2ptr(uint16_t idx, uint16_t depth)
 // Works on local copies of w and r - return only the difference and as such can be used to determine an overflow
 static inline uint16_t _tu_fifo_count(tu_fifo_t* f, uint16_t wr_idx, uint16_t rd_idx)
 {
-  uint16_t cnt = (uint16_t) (wr_idx-rd_idx);
+  uint16_t cnt;
 
   // In case we have non-power of two depth we need a further modification
-  if (rd_idx > wr_idx)
+  if (wr_idx >= rd_idx)
+  {
+    cnt = (uint16_t) (wr_idx - rd_idx);
+  } else
   {
-    // 2*f->depth - (rd_idx - wr_idx);
-    cnt = (uint16_t) (cnt - f->non_used_index_space);
+    cnt = (uint16_t) (2*f->depth - (rd_idx - wr_idx));
   }
 
   return cnt;
@@ -395,7 +396,7 @@ static inline bool _tu_fifo_full(tu_fifo_t* f, uint16_t wAbs, uint16_t rAbs)
 // use DMAs, write functions do not allow such an error.
 static inline bool _tu_fifo_overflowed(tu_fifo_t* f, uint16_t wr_idx, uint16_t rd_idx)
 {
-  return (_tu_fifo_count(f, wr_idx, rd_idx) > f->depth);
+  return _tu_fifo_count(f, wr_idx, rd_idx) > f->depth;
 }
 
 // Works on local copies of w
@@ -868,8 +869,6 @@ bool tu_fifo_clear(tu_fifo_t *f)
   _ff_lock(f->mutex_rd);
 
   f->rd_idx = f->wr_idx = 0;
-  //f->max_pointer_idx = (uint16_t) (2*f->depth-1);
-  f->non_used_index_space = (uint16_t) (UINT16_MAX - (2*f->depth-1));
 
   _ff_unlock(f->mutex_wr);
   _ff_unlock(f->mutex_rd);
diff --git a/src/common/tusb_fifo.h b/src/common/tusb_fifo.h
index f72ff295b..2f60ec2f4 100644
--- a/src/common/tusb_fifo.h
+++ b/src/common/tusb_fifo.h
@@ -69,8 +69,8 @@ extern "C" {
  *    | 0 | 1 | W | 3 | 4 | R |
  *
  * - Number of items in the fifo can be determined in either cases:
- *    - case W > R: Count = W - R
- *    - case W < R: Count = 2*depth - (R - W)
+ *    - case W >= R: Count = W - R
+ *    - case W <  R: Count = 2*depth - (R - W)
  *
  * In non-overwritable mode, computed Count (in above 2 cases) is at most equal to depth.
  * However, in over-writable mode, write index can be repeatedly increased and count can be
@@ -82,10 +82,10 @@ extern "C" {
  *      -------------------------
  *      | R | 1 | 2 | 3 | W | 5 |
  *
- *  - Double Overflowed i.e index is out of allowed range [0,2*depth) e.g:
+ *  - Double Overflowed i.e index is out of allowed range [0,2*depth)
+ *    This occurs when we continue to write after 1st overflowed to 2nd overflowed. e.g:
  *      write(3), write(1), write(2)
- *    Continue to write after overflowed to 2nd overflowed.
- *    We must prevent 2nd overflowed since it will cause incorrect computed of count, in above example
+ *    This must be prevented since it will cause unrecoverable state, in above example
  *    if not handled the fifo will be empty instead of continue-to-be full. Since we must not modify
  *    read index in write() function, which cause race condition. We will re-position write index so that
  *    after data is written it is a full fifo i.e W = depth - R
@@ -106,17 +106,16 @@ extern "C" {
  */
 typedef struct
 {
-  uint8_t* buffer               ; ///< buffer pointer
-  uint16_t depth                ; ///< max items
-  uint16_t item_size            ; ///< size of each item
+  uint8_t* buffer          ; // buffer pointer
+  uint16_t depth           ; // max items
 
-  bool overwritable             ;
+  struct TU_ATTR_PACKED {
+    uint16_t item_size : 15; // size of each item
+    bool overwritable  : 1 ; // ovwerwritable when full
+  };
 
-  uint16_t non_used_index_space ; ///< required for non-power-of-two buffer length
-  //uint16_t max_pointer_idx      ; ///< maximum absolute pointer index
-
-  volatile uint16_t wr_idx      ; ///< write index
-  volatile uint16_t rd_idx      ; ///< read index
+  volatile uint16_t wr_idx ; // write index
+  volatile uint16_t rd_idx ; // read index
 
 #if OSAL_MUTEX_REQUIRED
   osal_mutex_t mutex_wr;
@@ -133,13 +132,12 @@ typedef struct
   void * ptr_wrap   ; ///< wrapped part start pointer
 } tu_fifo_buffer_info_t;
 
-#define TU_FIFO_INIT(_buffer, _depth, _type, _overwritable)        \
-{                                                                  \
-  .buffer               = _buffer,                                 \
-  .depth                = _depth,                                  \
-  .item_size            = sizeof(_type),                           \
-  .overwritable         = _overwritable,                           \
-  .non_used_index_space = (uint16_t)(UINT16_MAX - (2*(_depth)-1)), \
+#define TU_FIFO_INIT(_buffer, _depth, _type, _overwritable) \
+{                                                           \
+  .buffer               = _buffer,                          \
+  .depth                = _depth,                           \
+  .item_size            = sizeof(_type),                    \
+  .overwritable         = _overwritable,                    \
 }
 
 #define TU_FIFO_DEF(_name, _depth, _type, _overwritable)                      \
diff --git a/src/device/usbd.c b/src/device/usbd.c
index f652a878e..6e0c6710d 100644
--- a/src/device/usbd.c
+++ b/src/device/usbd.c
@@ -388,6 +388,8 @@ bool tud_init (uint8_t rhport)
 
   TU_LOG(USBD_DBG, "USBD init on controller %u\r\n", rhport);
   TU_LOG_INT(USBD_DBG, sizeof(usbd_device_t));
+  TU_LOG_INT(USBD_DBG, sizeof(tu_fifo_t));
+  TU_LOG_INT(USBD_DBG, sizeof(tu_edpt_stream_t));
 
   tu_varclr(&_usbd_dev);