G++ 3.4.1 rejects some integral constant expressions that were allowed in the earlier versions. The same file compiled with GCC as a C file does not produce any errors. This report seems to be related to Bugzilla Bug 16489. The sample file below (sample.cc) contains comments explaining the problem. How to reproduce the bug: g++ sample.cc --------------------------------------------------------------------------- /* // Bug report for gcc version 3.4.1 (Mandrakelinux (Cooker) 3.4.1-1mdk) // Recent versions of G++ refuse to accept some // constant expressions which contain casts into // non integral types, e.g. into pointer types. // It says a cast to anything // but integral or enumaration type is not an integral constant // expression. // Such expressions are often produced by some // tools generating C++ code as their output. // Previously, these expressions were accepted by // G++. // Size of array should be an integral constant expression, or? // But G++ does not complain about this expression */ int array [((int)(void*)1)]; enum MyEnum { /* // Here we use the same constant_expression, // but as a value of the enumerator. // G++ doesn't like it for some reason */ ONE = ((int)((void*)1)) }; int main(int argc) { switch(argc) { /* // Here we use the same constant_expression // as a case value of switch statement // G++ does not like it at this place, too. */ case ((int)(void*)1): break; case 2: break; } return 0; } /* There are three places in this program, that are using the same integer constant_expression. Earlier versions of G++, <= 3.2.X, didn't have this problem and accepted all three occurences of the constant_expression without any problems. More over, all other C++ compilers (Borland C++, Visual C++, etc) do not complain about this code. If this file is compiled as a C file, no errors are reported by GCC. It looks like the switch to the new C++ parser and corresponding changes in the G++ code introduced this behavior. And it looks like it is a bug in G++. After looking at the Changelog file, it seems that the following change could have introduced such a behavior: 2004-03-19 Giovanni Bajo <giovannibajo@gcc.gnu.org> PR c++/14545 * parser.c (cp_parser_functional_cast): A cast to anything but integral or enumaration type is not an integral constant expression. Probably, only the final type of the integral constant expression should be integral or enumeration. But in the middle of such expressions, it should probably be possible to cast to some other types, at least into pointer types, since conversions between pointers and integers do not involve any run-time computations. */
My understanding is that casts are not allowed in integral const expressions, but in that case the array size should be disallowed as well. Giovanni, your patch is quoted in this PR, do you have some input? W.
I do not have the standard in the front of my but IIRC the intermediate casts have to be of integeral types too. In fact the change to reject some of these caused the definition of the macro offsetof to change (once again).
Yes, I should have been clearer: casts between integral types should be ok. Casts to void* and back maybe not. In any case, we're inconsistent here. W.
Subject: Re: New: Bug in reporting integral constant expressions problems On Mon, 26 Jul 2004, romixlev at yahoo dot com wrote: > G++ 3.4.1 rejects some integral constant expressions that were allowed in the > earlier versions. The same file compiled with GCC as a C file does not produce > any errors. This report seems to be related to Bugzilla Bug 16489. Both the C and C++ standards are clear that casts to non-integer types aren't allowed in integral constant expressions (C90, C++) / integer constant expressions (C99). Thus the bug is the nondiagnosis of: > // Size of array should be an integral constant expression, or? > // But G++ does not complain about this expression > */ > int array [((int)(void*)1)]; The constant expressions bugs for C are bugs 456 and 5675. Addressing them is on my C standards conformance roadmap <http://www.srcf.ucam.org/~jsm28/gcc/#stdc>, but after dealing with other miscellaneous obscure bugs such as 13801, and it is a fair amount of work to implement properly an obscure area of the standard with little relevance to real code. Given the obscurity of the area (and that there is hardly a great deal of user interest in the finer points of conformance), I'm surprised that this came up as a bug report, and still more so that real code is actually using such things that are/were undocumented extensions to constant expressions. (The intention will still be that such undocumented extensions are liable to be removed without notice, as those in C++ have been with the implementation of the new parser, but it's interesting that such removal has had *any* impact on user code.)
Subject: Re: [3.4/3.5 regression] Bug in reporting integral constant expressions problems Actually, my bug report comes from the following area. I'm using PROP, a sort of C++ preprocessor, that was developed may be 10 years ago, which translates a superset of C++ into "normal" C++. Among other things it supports algebraic data types and garbage collection for C++. For a given algebraic data type there can be several diffeent constructors (in the sense of algebraic datatypes, not in C++ sense) with and without arguments. The constructors with arguments are dynamically creating the objects on the heap. The unit constructors (i.e. those without arguments) do not create objects. Instead they are represented by some special values like (AlgebraicDataType *)small_integer. In principle, they are like enumerations. But these values should be assignable to a pointer to the AlgebraicDataType, they are always defined like: #define UniConstructorName (AlgebraicDataType * ) small_integer Later this macro is used everywhere in the code generated by PROP. The funny thing about this is that I'm using it since 6-7 years and have never had problems with it undar any compiler. So, when I switched to GCC 3.4.1 I was very surprised to see, that suddenly the automatically generated code was illigal... But now, after some clarifications, I see that the standard really does not allow such conversions into non-integral types insdie an integral constant expression. So, I simply changed PROP to generate a a code that it standard compliant. Now it works fine. BTW, there was a small problem while changing PROP. Since PROP uses bootstraping (i.e. written in itself), its code was generated with the old version of it and of course was using these illegal constructs. I had to change them by hand (!!!) in several hundred places, before I could compile a new version of PROP. After that I regenerated it with the new version and now it does not use illegal constructs. An G++ option that would explicitly allow old behavior (at least for some time, before making this constructs totally obsolette) would help in such situations. But I guess, I ask too much, or? :) --- jsm at polyomino dot org dot uk <gcc-bugzilla@gcc.gnu.org> wrote: > > ------- Additional Comments From jsm at polyomino > dot org dot uk 2004-07-26 20:30 ------- > Subject: Re: New: Bug in reporting integral > constant expressions > problems > > On Mon, 26 Jul 2004, romixlev at yahoo dot com > wrote: > > > G++ 3.4.1 rejects some integral constant > expressions that were allowed in the > > earlier versions. The same file compiled with GCC > as a C file does not produce > > any errors. This report seems to be related to > Bugzilla Bug 16489. > > Both the C and C++ standards are clear that casts to > non-integer types > aren't allowed in integral constant expressions > (C90, C++) / integer > constant expressions (C99). Thus the bug is the > nondiagnosis of: > > > // Size of array should be an integral constant > expression, or? > > // But G++ does not complain about this expression > > */ > > int array [((int)(void*)1)]; > > The constant expressions bugs for C are bugs 456 and > 5675. Addressing > them is on my C standards conformance roadmap > <http://www.srcf.ucam.org/~jsm28/gcc/#stdc>, but > after dealing with other > miscellaneous obscure bugs such as 13801, and it is > a fair amount of work > to implement properly an obscure area of the > standard with little > relevance to real code. Given the obscurity of the > area (and that there > is hardly a great deal of user interest in the finer > points of > conformance), I'm surprised that this came up as a > bug report, and still > more so that real code is actually using such things > that are/were > undocumented extensions to constant expressions. > (The intention will > still be that such undocumented extensions are > liable to be removed > without notice, as those in C++ have been with the > implementation of the > new parser, but it's interesting that such removal > has had *any* impact on > user code.) > > > > -- > > > http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16711 > > ------- You are receiving this mail because: ------- > You reported the bug, or are watching the reporter. > __________________________________ Do you Yahoo!? Yahoo! Mail - You care about security. So do we. http://promotions.yahoo.com/new_mail
> But I guess, I ask too much, or? :) Just as Joseph said, we're actually surprised that someone would use a 'feature' like this. You may be the only one who does, so I'd guess that chances are slim that we would introduce hacks to preserve backwards compatibility. W.
Since we're now certain what the problem is, I have opened PR 16783 and will close this PR here as a duplicate. The other one has a more concise description of the problem, without the initial uncertainty as to what exactly is right and wrong. W. *** This bug has been marked as a duplicate of 16783 ***