On 24 April 2014 18:44, Matt Godbolt <matt@xxxxxxxxxxx> wrote:
>
> (__extension__ ({ register unsigned int __v, __x = (_ip); if
> (__builtin_constant_p (__x)) __v = ((((__x) & 0xff000000) >> 24) |
> (((__x) & 0x00ff0000) >> 8) | (((__x) & 0x0000ff00) << 8) | (((__x) &
> 0x000000ff) << 24)); else __asm__ ("bswap %0" : "=r" (__v) : "0"
> (__x)); __v; }))
>
> When invoked via ccache -Wextra and -std=c++1y this line causes the error:
> error: address requested for '__x', which is declared 'register'
> [-Werror=extra]
Is this the actual output of GCC? GCC 4.9.0 should print column
numbers and a caret pointing to the exact location in the macro
expansion.
>From the piece that you quote, the warning seems just bogus, so I
think you should open a PR (http://gcc.gnu.org/bugs/) independently of
the preprocessed vs. non-preprocessed issue.
> However, when invoked directly, this error is somehow suppressed. Is
> GCC aware of the fact this macro was defined in a system header and
> thus suppresses the error? And therefore if it's
> previously-preprocessed, this tagging is lost?
Maybe. Could you try
> $ g++-4.9.0 -Wextra -O -std=c++1y -c test.cc # compiles without error
this -Wsystem-headers?
Also, could you try building a testcase that doesn't use headers but
it is not preprocessed?
> $ g++-4.9.0 -Wextra -O -std=c++1y -c test.cc -E > test.preprocessed.cc
> $ g++-4.9.0 -Wextra -O -std=c++1y -c test.preprocessed.cc
> test.cc: In function 'uint32_t test(uint32_t)':
> test.cc:4:134: warning: address requested for '__x', which is declared
> 'register' [-Wextra]
> uint32_t test(uint32_t x) { return htonl(x); }
> ====
>
> This only happens with std=c++1y, Wextra and -O (which macro-izes
> htonl instead of making it an external function). It does not happen
> with std=c++11.
Do std=c++11 vs std=c++1y change anything in the preprocessed file? If
not, the only way to understand what is going on is to put a break
point at the warning point and figure out why it is (wrongly) warning.
> Does anyone have a suggestion as to how to work around this issue? It
> seems we can't disable just the address-of-register warning, only
> Wextra itself (which removes other warnings we'd like).
As workarounds, you could try #pragma GCC diagnostics ignored "Wextra"
around that code. Or you could submit a patch to add an option
-Waddress-of-register to control this warning.
Cheers,
Manuel.
