Merge pull request #1942 from HubertD/feature/STM32G0

STM32G0 support

3 files changed, 347 insertions, 101 deletions

diff --git a/src/common/tusb_mcu.h b/src/common/tusb_mcu.h
index 9c2933983..48f765674 100644
--- a/src/common/tusb_mcu.h
+++ b/src/common/tusb_mcu.h
@@ -201,6 +201,11 @@
   #define TUP_USBIP_FSDEV_STM32
   #define TUP_DCD_ENDPOINT_MAX    8
 
+#elif TU_CHECK_MCU(OPT_MCU_STM32G0)
+  #define TUP_USBIP_FSDEV
+  #define TUP_USBIP_FSDEV_STM32
+  #define TUP_DCD_ENDPOINT_MAX    8
+
 #elif TU_CHECK_MCU(OPT_MCU_STM32L0, OPT_MCU_STM32L1)
   #define TUP_USBIP_FSDEV
   #define TUP_USBIP_FSDEV_STM32
diff --git a/src/portable/st/stm32_fsdev/dcd_stm32_fsdev.c b/src/portable/st/stm32_fsdev/dcd_stm32_fsdev.c
index 98d3d0829..14cabaf8d 100644
--- a/src/portable/st/stm32_fsdev/dcd_stm32_fsdev.c
+++ b/src/portable/st/stm32_fsdev/dcd_stm32_fsdev.c
@@ -44,6 +44,7 @@
  * L0x2, L0x3                     1024 byte buffer
  * L1                              512 byte buffer
  * L4x2, L4x3                     1024 byte buffer
+ * G0                             2048 byte buffer
  *
  * To use this driver, you must:
  * - If you are using a device with crystal-less USB, set up the clock recovery system (CRS)
@@ -185,12 +186,12 @@ static void dcd_ep_ctr_handler(void);
 static uint8_t open_ep_count;
 static uint16_t ep_buf_ptr; ///< Points to first free memory location
 static void dcd_pma_alloc_reset(void);
-static uint16_t dcd_pma_alloc(uint8_t ep_addr, size_t length);
+static uint16_t dcd_pma_alloc(uint8_t ep_addr, uint16_t length);
 static void dcd_pma_free(uint8_t ep_addr);
 static void dcd_ep_free(uint8_t ep_addr);
 static uint8_t dcd_ep_alloc(uint8_t ep_addr, uint8_t ep_type);
-static bool dcd_write_packet_memory(uint16_t dst, const void *__restrict src, size_t wNBytes);
-static bool dcd_read_packet_memory(void *__restrict dst, uint16_t src, size_t wNBytes);
+static bool dcd_write_packet_memory(uint16_t dst, const void *__restrict src, uint16_t wNBytes);
+static bool dcd_read_packet_memory(void *__restrict dst, uint16_t src, uint16_t wNBytes);
 
 static bool dcd_write_packet_memory_ff(tu_fifo_t * ff, uint16_t dst, uint16_t wNBytes);
 static bool dcd_read_packet_memory_ff(tu_fifo_t * ff, uint16_t src, uint16_t wNBytes);
@@ -209,17 +210,6 @@ TU_ATTR_ALWAYS_INLINE static inline xfer_ctl_t* xfer_ctl_ptr(uint32_t ep_addr)
   return &xfer_status[epnum][dir];
 }
 
-// Using a function due to better type checks
-// This seems better than having to do type casts everywhere else
-TU_ATTR_ALWAYS_INLINE static inline void reg16_clear_bits(__IO uint16_t *reg, uint16_t mask) {
-  *reg = (uint16_t)(*reg & ~mask);
-}
-
-// Bits in ISTR are cleared upon writing 0
-TU_ATTR_ALWAYS_INLINE static inline void clear_istr_bits(uint16_t mask) {
-  USB->ISTR = ~mask;
-}
-
 //--------------------------------------------------------------------+
 // Controller API
 //--------------------------------------------------------------------+
@@ -242,7 +232,9 @@ void dcd_init (uint8_t rhport)
   {
     asm("NOP");
   }
-  reg16_clear_bits(&USB->CNTR, USB_CNTR_PDWN);// Remove powerdown
+
+  USB->CNTR &= ~USB_CNTR_PDWN;
+
   // Wait startup time, for F042 and F070, this is <= 1 us.
   for(uint32_t i = 0; i<200; i++) // should be a few us
   {
@@ -250,8 +242,9 @@ void dcd_init (uint8_t rhport)
   }
   USB->CNTR = 0; // Enable USB
 
+#ifndef STM32G0 // BTABLE register does not exist any more on STM32G0, it is fixed to USB SRAM base address
   USB->BTABLE = DCD_STM32_BTABLE_BASE;
-
+#endif
   USB->ISTR = 0; // Clear pending interrupts
 
   // Reset endpoints to disabled
@@ -312,7 +305,7 @@ void dcd_sof_enable(uint8_t rhport, bool en)
   }
   else
   {
-    USB->CNTR &= (uint16_t) ~USB_CNTR_SOFM;
+    USB->CNTR &= ~USB_CNTR_SOFM;
   }
 }
 
@@ -358,6 +351,13 @@ void dcd_int_enable (uint8_t rhport)
   NVIC_EnableIRQ(USB_LP_IRQn);
   NVIC_EnableIRQ(USBWakeUp_IRQn);
 
+#elif CFG_TUSB_MCU == OPT_MCU_STM32G0
+  #ifdef STM32G0B0xx
+    NVIC_EnableIRQ(USB_IRQn);
+  #else
+    NVIC_EnableIRQ(USB_UCPD1_2_IRQn);
+  #endif
+
 #elif CFG_TUSB_MCU == OPT_MCU_STM32WB
   NVIC_EnableIRQ(USB_HP_IRQn);
   NVIC_EnableIRQ(USB_LP_IRQn);
@@ -408,6 +408,13 @@ void dcd_int_disable(uint8_t rhport)
   NVIC_DisableIRQ(USB_LP_IRQn);
   NVIC_DisableIRQ(USBWakeUp_IRQn);
 
+#elif CFG_TUSB_MCU == OPT_MCU_STM32G0
+  #ifdef STM32G0B0xx
+    NVIC_DisableIRQ(USB_IRQn);
+  #else
+    NVIC_DisableIRQ(USB_UCPD1_2_IRQn);
+  #endif
+
 #elif CFG_TUSB_MCU == OPT_MCU_STM32WB
   NVIC_DisableIRQ(USB_HP_IRQn);
   NVIC_DisableIRQ(USB_LP_IRQn);
@@ -439,7 +446,7 @@ void dcd_remote_wakeup(uint8_t rhport)
 {
   (void) rhport;
 
-  USB->CNTR |= (uint16_t) USB_CNTR_RESUME;
+  USB->CNTR |= USB_CNTR_RESUME;
   remoteWakeCountdown = 4u; // required to be 1 to 15 ms, ESOF should trigger every 1ms.
 }
 
@@ -470,7 +477,6 @@ static void dcd_handle_bus_reset(void)
   //__IO uint16_t * const epreg = &(EPREG(0));
   USB->DADDR = 0u; // disable USB peripheral by clearing the EF flag
 
-
   for(uint32_t i=0; i<STFSDEV_EP_COUNT; i++)
   {
     // Clear all EPREG (or maybe this is automatic? I'm not sure)
@@ -540,9 +546,6 @@ static void dcd_ep_ctr_rx_handler(uint32_t wIstr)
 
   if((ep_addr == 0U) && ((wEPRegVal & USB_EP_SETUP) != 0U)) /* Setup packet */
   {
-    // The setup_received function uses memcpy, so this must first copy the setup data into
-    // user memory, to allow for the 32-bit access that memcpy performs.
-    uint8_t userMemBuf[8];
     uint32_t count = pcd_get_ep_rx_cnt(USB, EPindex);
     /* Get SETUP Packet*/
     if(count == 8) // Setup packet should always be 8 bytes. If not, ignore it, and try again.
@@ -550,8 +553,15 @@ static void dcd_ep_ctr_rx_handler(uint32_t wIstr)
       // Must reset EP to NAK (in case it had been stalling) (though, maybe too late here)
       pcd_set_ep_rx_status(USB,0u,USB_EP_RX_NAK);
       pcd_set_ep_tx_status(USB,0u,USB_EP_TX_NAK);
-      dcd_read_packet_memory(userMemBuf, *pcd_ep_rx_address_ptr(USB,EPindex), 8);
+#ifdef PMA_32BIT_ACCESS
+      dcd_event_setup_received(0, (uint8_t*)(USB_PMAADDR + pcd_get_ep_rx_address(USB, EPindex)), true);
+#else
+      // The setup_received function uses memcpy, so this must first copy the setup data into
+      // user memory, to allow for the 32-bit access that memcpy performs.
+      uint8_t userMemBuf[8];
+      dcd_read_packet_memory(userMemBuf, pcd_get_ep_rx_address(USB,EPindex), 8);
       dcd_event_setup_received(0, (uint8_t*)userMemBuf, true);
+#endif
     }
   }
   else
@@ -574,7 +584,7 @@ static void dcd_ep_ctr_rx_handler(uint32_t wIstr)
 
     if (count != 0U)
     {
-      uint16_t addr = *pcd_ep_rx_address_ptr(USB, EPindex);
+      uint16_t addr = pcd_get_ep_rx_address(USB, EPindex);
 
       if (xfer->ff)
       {
@@ -657,13 +667,13 @@ void dcd_int_handler(uint8_t rhport) {
 
   /* Put SOF flag at the beginning of ISR in case to get least amount of jitter if it is used for timing purposes */
   if(int_status & USB_ISTR_SOF) {
-    clear_istr_bits(USB_ISTR_SOF);
+    USB->ISTR &=~USB_ISTR_SOF;
     dcd_event_sof(0, USB->FNR & USB_FNR_FN, true);
   }
 
   if(int_status & USB_ISTR_RESET) {
     // USBRST is start of reset.
-    clear_istr_bits(USB_ISTR_RESET);
+    USB->ISTR &=~USB_ISTR_RESET;
     dcd_handle_bus_reset();
     dcd_event_bus_reset(0, TUSB_SPEED_FULL, true);
     return; // Don't do the rest of the things here; perhaps they've been cleared?
@@ -678,9 +688,10 @@ void dcd_int_handler(uint8_t rhport) {
 
   if (int_status & USB_ISTR_WKUP)
   {
-    reg16_clear_bits(&USB->CNTR, USB_CNTR_LPMODE);
-    reg16_clear_bits(&USB->CNTR, USB_CNTR_FSUSP);
-    clear_istr_bits(USB_ISTR_WKUP);
+    USB->CNTR &= ~USB_CNTR_LPMODE;
+    USB->CNTR &= ~USB_CNTR_FSUSP;
+
+    USB->ISTR &=~USB_ISTR_WKUP;
     dcd_event_bus_signal(0, DCD_EVENT_RESUME, true);
   }
 
@@ -694,20 +705,20 @@ void dcd_int_handler(uint8_t rhport) {
     USB->CNTR |= USB_CNTR_LPMODE;
 
     /* clear of the ISTR bit must be done after setting of CNTR_FSUSP */
-    clear_istr_bits(USB_ISTR_SUSP);
+    USB->ISTR &=~USB_ISTR_SUSP;
     dcd_event_bus_signal(0, DCD_EVENT_SUSPEND, true);
   }
 
   if(int_status & USB_ISTR_ESOF) {
     if(remoteWakeCountdown == 1u)
     {
-      USB->CNTR &= (uint16_t)(~USB_CNTR_RESUME);
+      USB->CNTR &= ~USB_CNTR_RESUME;
     }
     if(remoteWakeCountdown > 0u)
     {
       remoteWakeCountdown--;
     }
-    clear_istr_bits(USB_ISTR_ESOF);
+    USB->ISTR &=~USB_ISTR_ESOF;
   }
 }
 
@@ -728,8 +739,8 @@ void dcd_edpt0_status_complete(uint8_t rhport, tusb_control_request_t const * re
     uint8_t const dev_addr = (uint8_t) request->wValue;
 
     // Setting new address after the whole request is complete
-    reg16_clear_bits(&USB->DADDR, USB_DADDR_ADD);
-    USB->DADDR = (uint16_t)(USB->DADDR | dev_addr); // leave the enable bit set
+    USB->DADDR &= ~USB_DADDR_ADD;
+    USB->DADDR |= dev_addr; // leave the enable bit set
   }
 }
 
@@ -756,7 +767,7 @@ static void dcd_pma_alloc_reset(void)
  *
  * During failure, TU_ASSERT is used. If this happens, rework/reallocate memory manually.
  */
-static uint16_t dcd_pma_alloc(uint8_t ep_addr, size_t length)
+static uint16_t dcd_pma_alloc(uint8_t ep_addr, uint16_t length)
 {
   xfer_ctl_t* epXferCtl = xfer_ctl_ptr(ep_addr);
 
@@ -768,6 +779,13 @@ static uint16_t dcd_pma_alloc(uint8_t ep_addr, size_t length)
     return epXferCtl->pma_ptr;
   }
 
+  // Ensure allocated buffer is aligned
+#ifdef PMA_32BIT_ACCESS
+  length = (length + 3) & ~0x03;
+#else
+  length = (length + 1) & ~0x01;
+#endif
+
   open_ep_count++;
 
   uint16_t addr = ep_buf_ptr;
@@ -778,7 +796,7 @@ static uint16_t dcd_pma_alloc(uint8_t ep_addr, size_t length)
 
   epXferCtl->pma_ptr = addr;
   epXferCtl->pma_alloc_size = length;
-  //TU_LOG2("dcd_pma_alloc(%x,%x)=%x\r\n",ep_addr,length,addr);
+  //TU_LOG1("dcd_pma_alloc(%x,%x)=%x\r\n",ep_addr,length,addr);
 
   return addr;
 }
@@ -931,14 +949,14 @@ bool dcd_edpt_open (uint8_t rhport, tusb_desc_endpoint_t const * p_endpoint_desc
 
   if( (dir == TUSB_DIR_IN) || (wType == USB_EP_ISOCHRONOUS) )
   {
-    *pcd_ep_tx_address_ptr(USB, ep_idx) = pma_addr;
+    pcd_set_ep_tx_address(USB, ep_idx, pma_addr);
     pcd_set_ep_tx_bufsize(USB, ep_idx, buffer_size);
     pcd_clear_tx_dtog(USB, ep_idx);
   }
 
   if( (dir == TUSB_DIR_OUT) || (wType == USB_EP_ISOCHRONOUS) )
   {
-    *pcd_ep_rx_address_ptr(USB, ep_idx) = pma_addr;
+    pcd_set_ep_rx_address(USB, ep_idx, pma_addr);
     pcd_set_ep_rx_bufsize(USB, ep_idx, buffer_size);
     pcd_clear_rx_dtog(USB, ep_idx);
   }
@@ -1018,8 +1036,8 @@ bool dcd_edpt_iso_alloc(uint8_t rhport, uint8_t ep_addr, uint16_t largest_packet
 
   pcd_set_eptype(USB, ep_idx, USB_EP_ISOCHRONOUS);
 
-  *pcd_ep_tx_address_ptr(USB, ep_idx) = pma_addr;
-  *pcd_ep_rx_address_ptr(USB, ep_idx) = pma_addr;
+  pcd_set_ep_tx_address(USB, ep_idx, pma_addr);
+  pcd_set_ep_rx_address(USB, ep_idx, pma_addr);
 
   return true;
 }
@@ -1069,7 +1087,7 @@ static void dcd_transmit_packet(xfer_ctl_t * xfer, uint16_t ep_ix)
   }
 
   uint16_t ep_reg = pcd_get_endpoint(USB, ep_ix);
-  uint16_t addr_ptr = *pcd_ep_tx_address_ptr(USB,ep_ix);
+  uint16_t addr_ptr = pcd_get_ep_tx_address(USB, ep_ix);
 
   if (xfer->ff)
   {
@@ -1203,6 +1221,40 @@ void dcd_edpt_clear_stall (uint8_t rhport, uint8_t ep_addr)
   }
 }
 
+#ifdef PMA_32BIT_ACCESS
+static bool dcd_write_packet_memory(uint16_t dst, const void *__restrict src, uint16_t wNBytes)
+{
+  const uint8_t* srcVal = src;
+  volatile uint32_t* dst32 = (volatile uint32_t*)(USB_PMAADDR + dst);
+
+  for (uint32_t n = wNBytes / 4; n > 0; --n) {
+    *dst32++ = tu_unaligned_read32(srcVal);
+    srcVal += 4;
+  }
+
+  wNBytes = wNBytes & 0x03;
+  if (wNBytes)
+  {
+    uint32_t wrVal = *srcVal;
+    wNBytes--;
+
+    if (wNBytes)
+    {
+      wrVal |= *++srcVal << 8;
+      wNBytes--;
+
+      if (wNBytes)
+      {
+        wrVal |= *++srcVal << 16;
+      }
+    }
+
+    *dst32 = wrVal;
+  }
+
+  return true;
+}
+#else
 // Packet buffer access can only be 8- or 16-bit.
 /**
   * @brief Copy a buffer from user memory area to packet memory area (PMA).
@@ -1214,7 +1266,7 @@ void dcd_edpt_clear_stall (uint8_t rhport, uint8_t ep_addr)
   * @param   wNBytes no. of bytes to be copied.
   * @retval None
   */
-static bool dcd_write_packet_memory(uint16_t dst, const void *__restrict src, size_t wNBytes)
+static bool dcd_write_packet_memory(uint16_t dst, const void *__restrict src, uint16_t wNBytes)
 {
   uint32_t n = (uint32_t)wNBytes >> 1U;
   uint16_t temp1, temp2;
@@ -1237,7 +1289,7 @@ static bool dcd_write_packet_memory(uint16_t dst, const void *__restrict src, si
     srcVal++;
   }
 
-  if (wNBytes & 0x01)
+  if (wNBytes)
   {
     temp1 = *srcVal;
     *pdwVal = temp1;
@@ -1245,6 +1297,7 @@ static bool dcd_write_packet_memory(uint16_t dst, const void *__restrict src, si
 
   return true;
 }
+#endif
 
 /**
   * @brief Copy from FIFO to packet memory area (PMA).
@@ -1263,7 +1316,37 @@ static bool dcd_write_packet_memory_ff(tu_fifo_t * ff, uint16_t dst, uint16_t wN
 
   // We want to read from the FIFO and write it into the PMA, if LIN part is ODD and has WRAPPED part,
   // last lin byte will be combined with wrapped part
-  // To ensure PMA is always access 16bit aligned (dst aligned to 16 bit)
+  // To ensure PMA is always access aligned (dst aligned to 16 or 32 bit)
+#ifdef PMA_32BIT_ACCESS
+  if((cnt_lin & 0x03) && cnt_wrap)
+  {
+    // Copy first linear part
+    dcd_write_packet_memory(dst, info.ptr_lin, cnt_lin &~0x03);
+    dst += cnt_lin &~0x03;
+
+    // Copy last linear bytes & first wrapped bytes to buffer
+    uint32_t i;
+    uint8_t tmp[4];
+    for (i = 0; i < (cnt_lin & 0x03); i++)
+    {
+      tmp[i] = ((uint8_t*)info.ptr_lin)[(cnt_lin &~0x03) + i];
+    }
+    uint32_t wCnt = cnt_wrap;
+    for (; i < 4 && wCnt > 0; i++, wCnt--)
+    {
+      tmp[i] = *(uint8_t*)info.ptr_wrap;
+      info.ptr_wrap = (uint8_t*)info.ptr_wrap + 1;
+    }
+
+    // Write unaligned buffer
+    dcd_write_packet_memory(dst, &tmp, 4);
+    dst += 4;
+
+    // Copy rest of wrapped byte
+    if (wCnt)
+        dcd_write_packet_memory(dst, info.ptr_wrap, wCnt);
+  }
+#else
   if((cnt_lin & 0x01) && cnt_wrap)
   {
     // Copy first linear part
@@ -1278,6 +1361,7 @@ static bool dcd_write_packet_memory_ff(tu_fifo_t * ff, uint16_t dst, uint16_t wN
     // Copy rest of wrapped byte
     dcd_write_packet_memory(dst, ((uint8_t*)info.ptr_wrap) + 1, cnt_wrap - 1);
   }
+#endif
   else
   {
     // Copy linear part
@@ -1296,13 +1380,47 @@ static bool dcd_write_packet_memory_ff(tu_fifo_t * ff, uint16_t dst, uint16_t wN
   return true;
 }
 
+#ifdef PMA_32BIT_ACCESS
+static bool dcd_read_packet_memory(void *__restrict dst, uint16_t src, uint16_t wNBytes)
+{
+  uint8_t* dstVal = dst;
+  volatile uint32_t* src32 = (volatile uint32_t*)(USB_PMAADDR + src);
+
+  for (uint32_t n = wNBytes / 4; n > 0; --n) {
+    tu_unaligned_write32(dstVal, *src32++);
+    dstVal += 4;
+  }
+
+  wNBytes = wNBytes & 0x03;
+  if (wNBytes)
+  {
+    uint32_t rdVal = *src32;
+
+    *dstVal = tu_u32_byte0(rdVal);
+    wNBytes--;
+
+    if (wNBytes)
+    {
+      *++dstVal = tu_u32_byte1(rdVal);
+      wNBytes--;
+
+      if (wNBytes)
+      {
+        *++dstVal = tu_u32_byte2(rdVal);
+      }
+    }
+  }
+
+  return true;
+}
+#else
 /**
   * @brief Copy a buffer from packet memory area (PMA) to user memory area.
   *        Uses byte-access of system memory and 16-bit access of packet memory
   * @param   wNBytes no. of bytes to be copied.
   * @retval None
   */
-static bool dcd_read_packet_memory(void *__restrict dst, uint16_t src, size_t wNBytes)
+static bool dcd_read_packet_memory(void *__restrict dst, uint16_t src, uint16_t wNBytes)
 {
   uint32_t n = (uint32_t)wNBytes >> 1U;
   // The GCC optimizer will combine access to 32-bit sizes if we let it. Force
@@ -1329,6 +1447,7 @@ static bool dcd_read_packet_memory(void *__restrict dst, uint16_t src, size_t wN
   }
   return true;
 }
+#endif
 
 /**
   * @brief Copy a buffer from user packet memory area (PMA) to FIFO.
@@ -1346,9 +1465,39 @@ static bool dcd_read_packet_memory_ff(tu_fifo_t * ff, uint16_t src, uint16_t wNB
   uint16_t cnt_lin =  TU_MIN(wNBytes, info.len_lin);
   uint16_t cnt_wrap = TU_MIN(wNBytes - cnt_lin, info.len_wrap);
 
+
   // We want to read from PMA and write it into the FIFO, if LIN part is ODD and has WRAPPED part,
   // last lin byte will be combined with wrapped part
-  // To ensure PMA is always access 16bit aligned (src aligned to 16 bit)
+  // To ensure PMA is always access aligned (src aligned to 16 or 32 bit)
+#ifdef PMA_32BIT_ACCESS
+  if((cnt_lin & 0x03) && cnt_wrap)
+  {
+    // Copy first linear part
+    dcd_read_packet_memory(info.ptr_lin, src, cnt_lin &~0x03);
+    src += cnt_lin &~0x03;
+
+    // Copy last linear bytes & first wrapped bytes
+    uint8_t tmp[4];
+    dcd_read_packet_memory(tmp, src, 4);
+    src += 4;
+
+    uint32_t i;
+    for (i = 0; i < (cnt_lin & 0x03); i++)
+    {
+      ((uint8_t*)info.ptr_lin)[(cnt_lin &~0x03) + i] = tmp[i];
+    }
+    uint32_t wCnt = cnt_wrap;
+    for (; i < 4 && wCnt > 0; i++, wCnt--)
+    {
+      *(uint8_t*)info.ptr_wrap = tmp[i];
+      info.ptr_wrap = (uint8_t*)info.ptr_wrap + 1;
+    }
+
+    // Copy rest of wrapped byte
+    if (wCnt)
+      dcd_read_packet_memory(info.ptr_wrap, src, wCnt);
+  }
+#else
   if((cnt_lin & 0x01) && cnt_wrap)
   {
     // Copy first linear part
@@ -1356,16 +1505,17 @@ static bool dcd_read_packet_memory_ff(tu_fifo_t * ff, uint16_t src, uint16_t wNB
     src += cnt_lin &~0x01;
 
     // Copy last linear byte & first wrapped byte
-    uint16_t tmp;
-    dcd_read_packet_memory(&tmp, src, 2);
-
-    ((uint8_t*)info.ptr_lin)[cnt_lin - 1] = (uint8_t)tmp;
-    ((uint8_t*)info.ptr_wrap)[0] = (uint8_t)(tmp >> 8U);
+    uint8_t tmp[2];
+    dcd_read_packet_memory(tmp, src, 2);
     src += 2;
 
+    ((uint8_t*)info.ptr_lin)[cnt_lin - 1] = tmp[0];
+    ((uint8_t*)info.ptr_wrap)[0] = tmp[1];
+
     // Copy rest of wrapped byte
     dcd_read_packet_memory(((uint8_t*)info.ptr_wrap) + 1, src, cnt_wrap - 1);
   }
+#endif
   else
   {
     // Copy linear part
diff --git a/src/portable/st/stm32_fsdev/dcd_stm32_fsdev_pvt_st.h b/src/portable/st/stm32_fsdev/dcd_stm32_fsdev_pvt_st.h
index e3fc8aeb5..e6649de5a 100644
--- a/src/portable/st/stm32_fsdev/dcd_stm32_fsdev_pvt_st.h
+++ b/src/portable/st/stm32_fsdev/dcd_stm32_fsdev_pvt_st.h
@@ -82,6 +82,34 @@
   #include "stm32g4xx.h"
   #define PMA_LENGTH (1024u)
 
+#elif CFG_TUSB_MCU == OPT_MCU_STM32G0
+  #include "stm32g0xx.h"
+  #define PMA_32BIT_ACCESS
+  #define PMA_LENGTH (2048u)
+  #undef USB_PMAADDR
+  #define USB_PMAADDR USB_DRD_PMAADDR
+  #define USB_TypeDef USB_DRD_TypeDef
+  #define EP0R CHEP0R
+  #define USB_EP_CTR_RX USB_EP_VTRX
+  #define USB_EP_CTR_TX USB_EP_VTTX
+  #define USB_EP_T_FIELD USB_CHEP_UTYPE
+  #define USB_EPREG_MASK USB_CHEP_REG_MASK
+  #define USB_EPTX_DTOGMASK USB_CHEP_TX_DTOGMASK
+  #define USB_EPRX_DTOGMASK USB_CHEP_RX_DTOGMASK
+  #define USB_EPTX_DTOG1 USB_CHEP_TX_DTOG1
+  #define USB_EPTX_DTOG2 USB_CHEP_TX_DTOG2
+  #define USB_EPRX_DTOG1 USB_CHEP_RX_DTOG1
+  #define USB_EPRX_DTOG2 USB_CHEP_RX_DTOG2
+  #define USB_EPRX_STAT USB_CH_RX_VALID
+  #define USB_EPKIND_MASK USB_EP_KIND_MASK
+  #define USB USB_DRD_FS
+  #define USB_CNTR_FRES USB_CNTR_USBRST
+  #define USB_CNTR_RESUME USB_CNTR_L2RES
+  #define USB_ISTR_EP_ID USB_ISTR_IDN
+  #define USB_EPADDR_FIELD USB_CHEP_ADDR
+  #define USB_CNTR_LPMODE USB_CNTR_SUSPRDY
+  #define USB_CNTR_FSUSP USB_CNTR_SUSPEN
+
 #elif CFG_TUSB_MCU == OPT_MCU_STM32WB
   #include "stm32wbxx.h"
   #define PMA_LENGTH (1024u)
@@ -113,15 +141,31 @@
   #define PMA_STRIDE  (1u)
 #endif
 
-// And for type-safety create a new macro for the volatile address of PMAADDR
+// For type-safety create a new macro for the volatile address of PMAADDR
 // The compiler should warn us if we cast it to a non-volatile type?
+#ifdef PMA_32BIT_ACCESS
+static __IO uint32_t * const pma32 = (__IO uint32_t*)USB_PMAADDR;
+#else
 // Volatile is also needed to prevent the optimizer from changing access to 32-bit (as 32-bit access is forbidden)
 static __IO uint16_t * const pma = (__IO uint16_t*)USB_PMAADDR;
 
-// prototypes
-TU_ATTR_ALWAYS_INLINE static inline __IO uint16_t* pcd_ep_rx_cnt_ptr(USB_TypeDef * USBx, uint32_t bEpIdx);
-TU_ATTR_ALWAYS_INLINE static inline __IO uint16_t* pcd_ep_tx_cnt_ptr(USB_TypeDef * USBx, uint32_t bEpIdx);
-TU_ATTR_ALWAYS_INLINE static inline void pcd_set_endpoint(USB_TypeDef * USBx, uint32_t bEpIdx, uint32_t wRegValue);
+TU_ATTR_ALWAYS_INLINE static inline __IO uint16_t * pcd_btable_word_ptr(USB_TypeDef * USBx, size_t x)
+{
+  size_t total_word_offset = (((USBx)->BTABLE)>>1) + x;
+  total_word_offset *= PMA_STRIDE;
+  return &(pma[total_word_offset]);
+}
+
+TU_ATTR_ALWAYS_INLINE static inline __IO uint16_t* pcd_ep_tx_cnt_ptr(USB_TypeDef * USBx, uint32_t bEpIdx)
+{
+  return pcd_btable_word_ptr(USBx,(bEpIdx)*4u + 1u);
+}
+
+TU_ATTR_ALWAYS_INLINE static inline __IO uint16_t* pcd_ep_rx_cnt_ptr(USB_TypeDef * USBx, uint32_t bEpIdx)
+{
+  return pcd_btable_word_ptr(USBx,(bEpIdx)*4u + 3u);
+}
+#endif
 
 /* Aligned buffer size according to hardware */
 TU_ATTR_ALWAYS_INLINE static inline uint16_t pcd_aligned_buffer_size(uint16_t size)
@@ -139,13 +183,24 @@ TU_ATTR_ALWAYS_INLINE static inline uint16_t pcd_aligned_buffer_size(uint16_t si
 /* SetENDPOINT */
 TU_ATTR_ALWAYS_INLINE static inline void pcd_set_endpoint(USB_TypeDef * USBx, uint32_t bEpIdx, uint32_t wRegValue)
 {
+#ifdef PMA_32BIT_ACCESS
+  (void) USBx;
+  __O uint32_t *reg = (__O uint32_t *)(USB_DRD_BASE + bEpIdx*4);
+  *reg = wRegValue;
+#else
   __O uint16_t *reg = (__O uint16_t *)((&USBx->EP0R) + bEpIdx*2u);
   *reg = (uint16_t)wRegValue;
+#endif
 }
 
 /* GetENDPOINT */
-TU_ATTR_ALWAYS_INLINE static inline uint16_t pcd_get_endpoint(USB_TypeDef * USBx, uint32_t bEpIdx) {
+TU_ATTR_ALWAYS_INLINE static inline uint32_t pcd_get_endpoint(USB_TypeDef * USBx, uint32_t bEpIdx) {
+#ifdef PMA_32BIT_ACCESS
+  (void) USBx;
+  __I uint32_t *reg = (__I uint32_t *)(USB_DRD_BASE + bEpIdx*4);
+#else
   __I uint16_t *reg = (__I uint16_t *)((&USBx->EP0R) + bEpIdx*2u);
+#endif
   return *reg;
 }
 
@@ -195,34 +250,24 @@ TU_ATTR_ALWAYS_INLINE static inline void pcd_clear_tx_ep_ctr(USB_TypeDef * USBx,
   */
 TU_ATTR_ALWAYS_INLINE static inline uint32_t pcd_get_ep_tx_cnt(USB_TypeDef * USBx, uint32_t bEpIdx)
 {
+#ifdef PMA_32BIT_ACCESS
+  (void) USBx;
+  return (pma32[2*bEpIdx] & 0x03FF0000) >> 16;
+#else
   __I uint16_t *regPtr = pcd_ep_tx_cnt_ptr(USBx, bEpIdx);
   return *regPtr & 0x3ffU;
+#endif
 }
 
 TU_ATTR_ALWAYS_INLINE static inline uint32_t pcd_get_ep_rx_cnt(USB_TypeDef * USBx, uint32_t bEpIdx)
 {
+#ifdef PMA_32BIT_ACCESS
+  (void) USBx;
+  return (pma32[2*bEpIdx + 1] & 0x03FF0000) >> 16;
+#else
   __I uint16_t *regPtr = pcd_ep_rx_cnt_ptr(USBx, bEpIdx);
   return *regPtr & 0x3ffU;
-}
-
-/**
-  * @brief  Sets counter of rx buffer with no. of blocks.
-  * @param  dwReg Register
-  * @param  wCount Counter.
-  * @param  wNBlocks no. of Blocks.
-  * @retval None
-  */
-TU_ATTR_ALWAYS_INLINE static inline void pcd_set_ep_cnt_reg(__O uint16_t * pdwReg, size_t wCount)
-{
-  /* We assume that the buffer size is already aligned to hardware requirements. */
-  uint16_t blocksize = (wCount > 62) ? 1 : 0;
-  uint16_t numblocks = wCount / (blocksize ? 32 : 2);
-
-  /* There should be no remainder in the above calculation */
-  TU_ASSERT((wCount - (numblocks * (blocksize ? 32 : 2))) == 0, /**/);
-
-  /* Encode into register. When BLSIZE==1, we need to subtract 1 block count */
-  *pdwReg = (blocksize << 15) | ((numblocks - blocksize) << 10);
+#endif
 }
 
 /**
@@ -241,57 +286,103 @@ TU_ATTR_ALWAYS_INLINE static inline void pcd_set_ep_address(USB_TypeDef * USBx,
   pcd_set_endpoint(USBx, bEpIdx,regVal);
 }
 
-TU_ATTR_ALWAYS_INLINE static inline __IO uint16_t * pcd_btable_word_ptr(USB_TypeDef * USBx, size_t x)
+TU_ATTR_ALWAYS_INLINE static inline uint32_t pcd_get_ep_tx_address(USB_TypeDef * USBx, uint32_t bEpIdx)
 {
-  size_t total_word_offset = (((USBx)->BTABLE)>>1) + x;
-  total_word_offset *= PMA_STRIDE;
-  return &(pma[total_word_offset]);
+#ifdef PMA_32BIT_ACCESS
+  (void) USBx;
+  return pma32[2*bEpIdx] & 0x0000FFFFu ;
+#else
+  return *pcd_btable_word_ptr(USBx,(bEpIdx)*4u + 0u);
+#endif
 }
 
-// Pointers to the PMA table entries (using the ARM address space)
-TU_ATTR_ALWAYS_INLINE static inline __IO uint16_t* pcd_ep_tx_address_ptr(USB_TypeDef * USBx, uint32_t bEpIdx)
+TU_ATTR_ALWAYS_INLINE static inline uint32_t pcd_get_ep_rx_address(USB_TypeDef * USBx, uint32_t bEpIdx)
 {
-  return pcd_btable_word_ptr(USBx,(bEpIdx)*4u + 0u);
-}
-TU_ATTR_ALWAYS_INLINE static inline __IO uint16_t* pcd_ep_tx_cnt_ptr(USB_TypeDef * USBx, uint32_t bEpIdx)
-{
-  return pcd_btable_word_ptr(USBx,(bEpIdx)*4u + 1u);
+#ifdef PMA_32BIT_ACCESS
+  (void) USBx;
+  return pma32[2*bEpIdx + 1] & 0x0000FFFFu;
+#else
+  return *pcd_btable_word_ptr(USBx,(bEpIdx)*4u + 2u);
+#endif
 }
 
-TU_ATTR_ALWAYS_INLINE static inline __IO uint16_t* pcd_ep_rx_address_ptr(USB_TypeDef * USBx, uint32_t bEpIdx)
+TU_ATTR_ALWAYS_INLINE static inline void pcd_set_ep_tx_address(USB_TypeDef * USBx, uint32_t bEpIdx, uint32_t addr)
 {
-  return  pcd_btable_word_ptr(USBx,(bEpIdx)*4u + 2u);
+#ifdef PMA_32BIT_ACCESS
+  (void) USBx;
+  pma32[2*bEpIdx] = (pma32[2*bEpIdx] & 0xFFFF0000u) | (addr & 0x0000FFFCu);
+#else
+  *pcd_btable_word_ptr(USBx,(bEpIdx)*4u + 0u) = addr;
+#endif
 }
 
-TU_ATTR_ALWAYS_INLINE static inline __IO uint16_t* pcd_ep_rx_cnt_ptr(USB_TypeDef * USBx, uint32_t bEpIdx)
+TU_ATTR_ALWAYS_INLINE static inline void pcd_set_ep_rx_address(USB_TypeDef * USBx, uint32_t bEpIdx, uint32_t addr)
 {
-  return pcd_btable_word_ptr(USBx,(bEpIdx)*4u + 3u);
+#ifdef PMA_32BIT_ACCESS
+  (void) USBx;
+  pma32[2*bEpIdx + 1] = (pma32[2*bEpIdx + 1] & 0xFFFF0000u) | (addr & 0x0000FFFCu);
+#else
+  *pcd_btable_word_ptr(USBx,(bEpIdx)*4u + 2u) = addr;
+#endif
 }
 
-TU_ATTR_ALWAYS_INLINE static inline void pcd_set_ep_tx_cnt(USB_TypeDef * USBx,  uint32_t bEpIdx, uint32_t wCount)
+TU_ATTR_ALWAYS_INLINE static inline void pcd_set_ep_tx_cnt(USB_TypeDef * USBx, uint32_t bEpIdx, uint32_t wCount)
 {
+#ifdef PMA_32BIT_ACCESS
+  (void) USBx;
+  pma32[2*bEpIdx] = (pma32[2*bEpIdx] & ~0x03FF0000u) | ((wCount & 0x3FFu) << 16);
+#else
   __IO uint16_t * reg = pcd_ep_tx_cnt_ptr(USBx, bEpIdx);
   *reg = (uint16_t) (*reg & (uint16_t) ~0x3FFU) | (wCount & 0x3FFU);
+#endif
 }
 
-TU_ATTR_ALWAYS_INLINE static inline void pcd_set_ep_rx_cnt(USB_TypeDef * USBx,  uint32_t bEpIdx, uint32_t wCount)
+TU_ATTR_ALWAYS_INLINE static inline void pcd_set_ep_rx_cnt(USB_TypeDef * USBx, uint32_t bEpIdx, uint32_t wCount)
 {
+#ifdef PMA_32BIT_ACCESS
+  (void) USBx;
+  pma32[2*bEpIdx + 1] = (pma32[2*bEpIdx + 1] & ~0x03FF0000u) | ((wCount & 0x3FFu) << 16);
+#else
   __IO uint16_t * reg = pcd_ep_rx_cnt_ptr(USBx, bEpIdx);
   *reg = (uint16_t) (*reg & (uint16_t) ~0x3FFU) | (wCount & 0x3FFU);
+#endif
 }
 
-TU_ATTR_ALWAYS_INLINE static inline void pcd_set_ep_tx_bufsize(USB_TypeDef * USBx,  uint32_t bEpIdx, uint32_t wCount)
+TU_ATTR_ALWAYS_INLINE static inline void pcd_set_ep_blsize_num_blocks(USB_TypeDef * USBx, uint32_t rxtx_idx, uint32_t blocksize, uint32_t numblocks)
 {
-  __IO uint16_t *pdwReg = pcd_ep_tx_cnt_ptr((USBx),(bEpIdx));
-  wCount = pcd_aligned_buffer_size(wCount);
-  pcd_set_ep_cnt_reg(pdwReg, wCount);
+  /* Encode into register. When BLSIZE==1, we need to subtract 1 block count */
+#ifdef PMA_32BIT_ACCESS
+  (void) USBx;
+  pma32[rxtx_idx] = (pma32[rxtx_idx] & 0x0000FFFFu) | (blocksize << 31) | ((numblocks - blocksize) << 26);
+#else
+  __IO uint16_t *pdwReg = pcd_btable_word_ptr(USBx, rxtx_idx*2u + 1u);
+  *pdwReg = (blocksize << 15) | ((numblocks - blocksize) << 10);
+#endif
 }
 
-TU_ATTR_ALWAYS_INLINE static inline void pcd_set_ep_rx_bufsize(USB_TypeDef * USBx,  uint32_t bEpIdx, uint32_t wCount)
+TU_ATTR_ALWAYS_INLINE static inline void pcd_set_ep_bufsize(USB_TypeDef * USBx, uint32_t rxtx_idx, uint32_t wCount)
 {
-  __IO uint16_t *pdwReg = pcd_ep_rx_cnt_ptr((USBx),(bEpIdx));
   wCount = pcd_aligned_buffer_size(wCount);
-  pcd_set_ep_cnt_reg(pdwReg, wCount);
+
+  /* We assume that the buffer size is already aligned to hardware requirements. */
+  uint16_t blocksize = (wCount > 62) ? 1 : 0;
+  uint16_t numblocks = wCount / (blocksize ? 32 : 2);
+
+  /* There should be no remainder in the above calculation */
+  TU_ASSERT((wCount - (numblocks * (blocksize ? 32 : 2))) == 0, /**/);
+
+  /* Encode into register. When BLSIZE==1, we need to subtract 1 block count */
+  pcd_set_ep_blsize_num_blocks(USBx, rxtx_idx, blocksize, numblocks);
+}
+
+TU_ATTR_ALWAYS_INLINE static inline void pcd_set_ep_tx_bufsize(USB_TypeDef * USBx, uint32_t bEpIdx, uint32_t wCount)
+{
+  pcd_set_ep_bufsize(USBx, 2*bEpIdx, wCount);
+}
+
+TU_ATTR_ALWAYS_INLINE static inline void pcd_set_ep_rx_bufsize(USB_TypeDef * USBx, uint32_t bEpIdx, uint32_t wCount)
+{
+  pcd_set_ep_bufsize(USBx, 2*bEpIdx + 1, wCount);
 }
 
 /**