GCC 4.7.3 (at least) through GCC 6.0 reject the following due to "meow has side-effects": #include <cstddef> #include <type_traits> template <typename T, std::size_t S> constexpr std::size_t lengthof(const volatile T (&)[S]) { return S; } int main() { volatile int meow[4]; static_cast<void>(meow); // shut up warning return static_cast<int>(std::integral_constant<std::size_t, lengthof(meow)>::value); } I believe that this is legal per [expr.const] in the Standard, because the volatile parameter is never used in an lvalue-to-rvalue conversion, which is what [expr.const] disallows in constant-expressions for volatile types. Clang versions that understand this code accept it; Visual C++ 2015 does as well. Someone who replied to my question on the "std-discussion" mailing list suggested that this is also technically legal as well: #include <type_traits> constexpr int Test(int x) { volatile int v = x; return x; } int main() { return std::integral_constant<int, Test(2)>::value; } GCC also rejects this, but Clang accepts this as well. Any attempt to read v will fail, though, so Clang is enforcing the rule. I'm not on my Windows machine as I write this, so I can't check MSVC.
EDG accepts the first testcase. The second one is rejected, but presumably due to incomplete C++14 support. In potential_constant_expression_1 we have a check: if (TREE_THIS_VOLATILE (t)) { if (flags & tf_error) error ("expression %qE has side-effects", t); The comment in tree.h says: /* Nonzero means this expression is volatile in the C sense: its address should be of type `volatile WHATEVER *'. In other words, the declared item is volatile qualified. ... If this bit is set in an expression, so is TREE_SIDE_EFFECTS. */ In our case t is: (gdb) p t $10 = <var_decl 0x7ffff66a7cf0 meow> (gdb) pt warning: Expression is not an assignment (and might have no effect) <var_decl 0x7ffff66a7cf0 meow type <array_type 0x7ffff6800c78 type <integer_type 0x7ffff6800a80 int volatile type_6 SI size <integer_cst 0x7ffff66bd0a8 constant 32> unit size <integer_cst 0x7ffff66bd0c0 constant 4> align 32 symtab 0 alias set -1 canonical type 0x7ffff6800a80 precision 32 min <integer_cst 0x7ffff66bd060 -2147483648> max <integer_cst 0x7ffff66bd078 2147483647> pointer_to_this <pointer_type 0x7ffff6800d20>> TI size <integer_cst 0x7ffff669bea0 constant 128> unit size <integer_cst 0x7ffff669beb8 constant 16> align 32 symtab 0 alias set -1 canonical type 0x7ffff6800c78 domain <integer_type 0x7ffff6800b28 type <integer_type 0x7ffff669f1f8 sizetype> type_6 DI size <integer_cst 0x7ffff669be58 constant 64> unit size <integer_cst 0x7ffff669be70 constant 8> align 64 symtab 0 alias set -1 canonical type 0x7ffff6800b28 precision 64 min <integer_cst 0x7ffff669be88 0> max <integer_cst 0x7ffff680adc8 3>> pointer_to_this <pointer_type 0x7ffff68133f0>> side-effects addressable volatile used tree_1 decl_5 TI file /home/miyuki/gcc/test/meow/meow.cc line 15 col 18 size <integer_cst 0x7ffff669bea0 128> unit size <integer_cst 0x7ffff669beb8 16> align 32 context <function_decl 0x7ffff67ffe00 main>> I.e. the condition is probably wrong and we should also check that t is an expression, not a declaration.
For the record: https://gcc.gnu.org/ml/gcc-patches/2015-08/msg01735.html
Seems to have been fixed sometime between 6.0 and 6.1.