[PATCH v1 5/8] spi: dw: Speed up transfer loops

[Sean Anderson]

On 3/21/20 9:49 PM, Marek Vasut wrote:
> On 3/5/20 8:19 PM, Sean Anderson wrote:
>> The transfer loops are very tight on some platforms (especially on higher
>> speeds). If we don't read/write fast enough we can run into over-/under-
>> flow problems. This patch removes several divisions and log statements,
>> and simplifies the read logic.
>>
>> Signed-off-by: Sean Anderson <seanga2 at gmail.com>
>> ---
>>
>>  drivers/spi/designware_spi.c | 29 +++++++++--------------------
>>  1 file changed, 9 insertions(+), 20 deletions(-)
>>
>> diff --git a/drivers/spi/designware_spi.c b/drivers/spi/designware_spi.c
>> index 38c24fe550..613eb0d0e6 100644
>> --- a/drivers/spi/designware_spi.c
>> +++ b/drivers/spi/designware_spi.c
>> @@ -304,7 +304,7 @@ static inline u32 tx_max(struct dw_spi_priv *priv)
>>  {
>>  	u32 tx_left, tx_room, rxtx_gap;
>>  
>> -	tx_left = (priv->tx_end - priv->tx) / (priv->bits_per_word >> 3);
>> +	tx_left = priv->tx_end - priv->tx;
>>  	tx_room = priv->fifo_len - dw_read(priv, DW_SPI_TXFLR);
>>  
>>  	/*
>> @@ -315,8 +315,7 @@ static inline u32 tx_max(struct dw_spi_priv *priv)
>>  	 * shift registers. So a control from sw point of
>>  	 * view is taken.
>>  	 */
>> -	rxtx_gap = ((priv->rx_end - priv->rx) - (priv->tx_end - priv->tx)) /
>> -		(priv->bits_per_word >> 3);
>> +	rxtx_gap = ((priv->rx_end - priv->rx) - (priv->tx_end - priv->tx));
>>  
>>  	return min3(tx_left, tx_room, (u32)(priv->fifo_len - rxtx_gap));
>>  }
>> @@ -324,7 +323,7 @@ static inline u32 tx_max(struct dw_spi_priv *priv)
>>  /* Return the max entries we should read out of rx fifo */
>>  static inline u32 rx_max(struct dw_spi_priv *priv)
>>  {
>> -	u32 rx_left = (priv->rx_end - priv->rx) / (priv->bits_per_word >> 3);
>> +	u32 rx_left = priv->rx_end - priv->rx;
>>  
>>  	return min_t(u32, rx_left, dw_read(priv, DW_SPI_RXFLR));
>>  }
>> @@ -336,15 +335,10 @@ static void dw_writer(struct dw_spi_priv *priv)
>>  
>>  	while (max--) {
>>  		/* Set the tx word if the transfer's original "tx" is not null */
>> -		if (priv->tx_end - priv->len) {
>> -			if (priv->bits_per_word == 8)
>> -				txw = *(u8 *)(priv->tx);
>> -			else
>> -				txw = *(u16 *)(priv->tx);
>> -		}
>> +		if (priv->tx_end - priv->len)
>> +			txw = *(u8 *)(priv->tx);
>>  		dw_write(priv, DW_SPI_DR, txw);
>> -		debug("%s: tx=0x%02x\n", __func__, txw);
>> -		priv->tx += priv->bits_per_word >> 3;
>> +		priv->tx++;
> 
> This breaks 16 bits per word transfers, NAK.
> 
> The compiler should be able to figure out the rest of the optimizations
> in this patch.
> 

Hm, I will try with just the debug symbols removed.

--Sean