On Wed, Aug 11, 2010 at 3:14 PM, David Abrahams wrote:

At Wed, 11 Aug 2010 11:59:21 -0700, Emil Dotchevski wrote:

...

On Wed, Aug 11, 2010 at 1:37 AM, Roland Bock wrote:

...

Hi,

compiling the attached example with gcc-4.4.3 (Ubuntu 10.4, 64bit) with

-Wall -O3

leads to a long warning message (second attachment), culminating in

boost/1.44/include/boost/function/function_base.hpp:321: warning: dereferencing type-punned pointer will break strict-aliasing rules boost/1.44/include/boost/function/function_base.hpp:325: warning: dereferencing type-punned pointer will break strict-aliasing rules

I wonder how I could avoid this without -fno-strict-aliasing or -Wno-strict-aliasing?

The code in question is a reinterpret_cast which casts a char pointer to a user-specified functor type pointer. The interpret cast seems correct, and the warning seems correct too if not very useful in this case.

Really, the reinterpret_cast is portably correct?  Pretty much the only thing you can do portably with reinterpret_cast is round-trip cast a pointer type to another pointer type and then cast it back before you use it again.

That's what this code does. Honestly, I'm probably giving the compiler too much credit, since I assumed that it sees that the object it's taking the address of is a single char, not an array of chars. In reality it's probably not that clever and the warning is plain wrong because in general strict aliasing rules don't apply to char pointers, do they? Emil Dotchevski Reverge Studios, Inc. http://www.revergestudios.com/reblog/index.php?n=ReCode

On 8/12/10 2:59 AM, Emil Dotchevski wrote:

On Wed, Aug 11, 2010 at 1:37 AM, Roland Bock wrote:

...

Hi,

compiling the attached example with gcc-4.4.3 (Ubuntu 10.4, 64bit) with

-Wall -O3

leads to a long warning message (second attachment), culminating in

boost/1.44/include/boost/function/function_base.hpp:321: warning: dereferencing type-punned pointer will break strict-aliasing rules boost/1.44/include/boost/function/function_base.hpp:325: warning: dereferencing type-punned pointer will break strict-aliasing rules

I wonder how I could avoid this without -fno-strict-aliasing or -Wno-strict-aliasing?

The code in question is a reinterpret_cast which casts a char pointer to a user-specified functor type pointer. The interpret cast seems correct, and the warning seems correct too if not very useful in this case.

Shouldn't this thread have a [function] subject as this seems to be about boost.function issue? Regards, -- Joel de Guzman http://www.boostpro.com http://spirit.sf.net

On Thu, Aug 12, 2010 at 12:48 AM, Roland Bock wrote:

On 08/12/2010 08:26 AM, KSpam wrote:

...

Roland,

On Wednesday 11 August 2010 01:37:17 Roland Bock wrote:

...

leads to a long warning message (second attachment), culminating in

boost/1.44/include/boost/function/function_base.hpp:321: warning: dereferencing type-punned pointer will break strict-aliasing rules boost/1.44/include/boost/function/function_base.hpp:325: warning: dereferencing type-punned pointer will break strict-aliasing rules

This warning can be prevented by splitting the line into two lines as follows:

Replace: reinterpret_cast(&in_buffer.data)->~Functor();

with: functor_type* p = reinterpret_cast(&in_buffer.data); p->~Functor();

I patched my local version of Boost in order to prevent these warnings.

Hope This Helps, Justin

Justin,

thanks for the hint! I'll give it a try and will probably file a ticket if it works out :-)

Also, you should file a ticket with GCC, two tickets actually: 1) a trivial change like this should not have effect on the warning, and 2) strict aliasing rules should not apply to char pointers (as far as I can tell, in_buffer.data is of char type specifically to deal with this issue.) Emil Dotchevski Reverge Studios, Inc. http://www.revergestudios.com/reblog/index.php?n=ReCode

On Thu, Aug 12, 2010 at 11:37 AM, Frank Mori Hess wrote:

-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1

On Thursday 12 August 2010, Emil Dotchevski wrote:

...

2) strict aliasing rules should not apply to char pointers (as far as I can tell, in_buffer.data is of char type specifically to deal with this issue.)

It's safe to cast to a char pointer and use the char pointer, but in this case you are casting away from a char pointer.  Bouncing a non-char pointer through a char pointer doesn't give any relief from the usual strict aliasing rules.

The object is not accessed through a char pointer, the char pointer is only used to define a memory location. The char object itself is used only to take its address. Even if the warning was valid, it just means that *if* you access the same object through different types (which the code never does), you'd be in trouble; however section 6.5, paragraph 7 of the C standard says that any object may be aliased by an object of type char, so the warning seems invalid. Emil Dotchevski Reverge Studios, Inc. http://www.revergestudios.com/reblog/index.php?n=ReCode