-O1 removes useful code

I have the following snippet of code, which toggles one of the GPIO lines:

PINSEL1 &= ~(0x0000000C);
IO0SET |= MY_GPIO_MASK;
IO0CLR |= MY_GPIO_MASK;
PINSEL1 |= 0x00000008;

This works fine as Thumb code at -O0. However, if I compile this as -O1 or higher (still Thumb), this code is optimized out and I don't get the "side effect" of my GPIO toggling.

If I switch over to generating ARM code, this works at -O0, -O1, and -O2 (i.e. no problem).

This seems really wrong. PINSEL1, IO0SET, etc. are all defined as volatile in LPC22xx.h, so shouldn't the compiler leave this alone when optimizing? Any good way of getting around this?