Due to some hardware guy's awful work, this version is not compatible
with v3.6: the logic of BIT(0) of the reset logic is inverted! (and
v3.6.10 is horribly wrong in multiple ways), but this is what we have
to solve now.

The v3.6 expects 0x0000 set to the register for reset de-assertion,
while v3.6 does 0x0001.

This commit (ab)uses another bug of v3.6.10 to work around the issue.
The UniPhier System Bus is a 16-bit bus, which this support card is
connected to.  A 32-bit write to the bus (writel() function call) is
divided into two 16-bit write transactions, with LSB the first.  What
is amazing for v3.6.10 is that access to address 4N + 2 goes to 4N
(Jesus Christ!).

For clarification, things are like this:

    writel(0x00010000, MICRO_SUPPORT_CARD_RESET);

is done with two bus transactions as follows

    [1] write 0x0000 to address MICRO_SUPPORT_CARD
    [2] write 0x0001 to address MICRO_SUPPORT_CARD + 2

For v3.6, [1] is written to the register and [2] is correctly ignored
because there is nothing at the address MICRO_SUPPORT_CARD + 2.  This
is what we expect.

For v3.6.10, [1] is written to the reset register and then [2] is
over-written to the same register due to the bus access bug.

For the latter, it produces a glitch signal to the BIT[0], so the
device state is lost due to the reset pulse.  This solution only
works for the start-up code.

Signed-off-by: Masahiro Yamada <yamada.masahiro@socionext.com>