[Bug target/67701] Unnecessary/bad instructions for u32-casted access to external symbol (assumes misaligned, superfluous load)

[jwerner at chromium dot org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67701

--- Comment #3 from Julius Werner <jwerner at chromium dot org> ---
> I suspect this is an armv7 issue only where there is missing support for misaligned load/stores.

Did a quick test on aarch64 and you're right, neither of the two issues appears
there.

> Also testvalue is aligned to 1 byte by default so what GCC is doing is correct as there is no way to do a volatile unaligned loads really.

I thought alignment was always tied to a type (at least by default, unless
overridden by __attribute__((align)) or the like)? It's true that the expected
alignment for testvalue is 1, but I'm not dereferencing testvalue... I'm
dereferencing *(volatile unsigned long *)&testvalue. Shouldn't the alignment of
that always be 4? (FWIW, _Alignof(*(volatile unsigned long *)&testvalue)
evaluates to 4 in both compiler versions.)

Another argument: if write32() was defined as a non-inline function in a
separate execution unit, this problem would not appear. Should inlining a
function really change behavior in this way and produce code that is clearly
less efficient (to try to work around a "problem" that the -O0 version of the
same code would have never found)?