Changelog：

r1285
GSdx: there was a bug in unaligned 4/8 bit transfer, small texture glitches may be fixed (simpson game, car logos in gt4).

r1279
GSdx: just fixing slow texture uploads with dx9... (odin sphere back to 40-50 fps)

r1272
GSdx: Even more alpha test voodoo magic, it must be nailed now!

And there is even an explanation.

The tfx functions calculate At * Af >> 7, which means modulating by 0x80 should
return At as the result.

With the evil floating point pixel shader however 0x80 translates to 128/255
(0.502), not exactly 0.5, modulation as At´ * Af´ * 2 (´ means 0 - 1.0 range) is
not the same as with integers.

At´ = Af´ = 0.502
At´ * Af´ * 2 = 0.504

If the alpha test happens to be "not equal to 0x80", then abs(0.504 - 128/255) <
0.5/255 will just miss.

Solution is to re-scale those values to the integer range, do the calculations,
and then back to float again, but in the end it just simplifies down to At´ *
Af´ * 255/128, doh...

At * Af >> 7 => ((At´ * 255) * (Af´ * 255) / 128) / 255 => At´ * Af´ * 255/128

At´ = Af´ = 0.502
0.502 * 0.502 * 255/128 = 0.502 (w00t!)

r1266
GSdx: bit more work on the vertex buffer, and broken ffxii fmv fixed again.