Re: [PATCH] C undefined behavior fix

[Alan Cox]

> I think it  is the compiler burden to prove that the extra stuff
> "read" is not observable (in the sense of external
> effect) at execution. In a memory-mapped I/O architecture
> where there is a distinction on external effects between a byte read
> and a word read (eg: a crash :),  the compiler can't get very far IMHO
> (if it accepts the declaration of course).

In the Linux case mmio isn't a problem. The rules for mmio in the portable
code require you use architecture dependant macros (readb etc) and that
the mmio space is mapped via ioremap.

Our ioremap/mmio stuff does a whole variety of things on different platforms
including simple memory accesses, adding I/O fences, motherboard specific
function calls, and tlb bypass.

That side of things is all nicely and cleanly handled.