The SID SRAM on at least some SoCs (A64 and D1) returns different values
when read with bus cycles narrower than 32 bits. This is not immediately
obvious, because memcpy_fromio() uses word-size accesses as long as
enough data is being copied.

The vendor driver always uses 32-bit MMIO reads, so do the same here.
This is faster than the register-based method, which is currently used
as a workaround on A64. And it fixes the values returned on D1, where
the SRAM method was being used.

The special case for the last word is needed to maintain .word_size == 1
for sysfs ABI compatibility, as noted previously in commit de2a3eae
("nvmem: sunxi_sid: Optimize register read-out method").

Fixes: 07ae4fde ("nvmem: sunxi_sid: Add support for D1 variant")
Cc: stable@vger.kernel.org
Tested-by: Heiko Stuebner <heiko@sntech.de>
Signed-off-by: Samuel Holland <samuel@sholland.org>
Signed-off-by: Srinivas Kandagatla <srinivas.kandagatla@linaro.org>
Link: https://lore.kernel.org/r/20230127104015.23839-3-srinivas.kandagatla@linaro.orgSigned-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>