toggle quoted messageShow quoted text
Why would you rely on undefined behaviours? The OFFSET_OF macro is well-defined for GCC and Clang as it's implemented by an intrinsic, and while the expression for the MSVC compiler is undefined behaviour as per the C standard, it is well-defined for MSVC due to their own implementation being identical. From my standpoint, all supported compilers will yield well-defined behaviour even this way. OFFSET_OF on flexible arrays is not UB in any case to my knowledge.
However, the same way as your new suggestion, you can replace OFFSET_OF with sizeof. While this *can* lead to wasted space with certain structure layouts (e.g. when the flexible array overlays padding bytes), this is not the case here, and otherwise just loses you a few bytes. I think this comes down to preference.
The pattern you mentioned arguably is less nice syntax when used (involves address calculation and casting), but the biggest problem here is alignment constraints. For packed structures, you lose the ability of automatic unaligned accesses (irrelevant here because the structure is manually padded anyway). For non-packed structures, you still need to ensure the alignment requirement of the trailing array data is met manually. With flexible array members, the compiler takes care of both cases automatically.
On 24.06.21 02:24, Kun Qin wrote:
I would prefer not to rely on undefined behaviors from different compilers. Instead of using flexible arrays, is it better to remove the `Data` field, pack the structure and follow "VARIABLE_LOCK_ON_VAR_STATE_POLICY" pattern?
In that case, OFFSET_OF will be forced to change to sizeof, and read/write to `Data` will follow the range indicated by MessageLength. But yes, that will enforce developers to update their platform level implementations accordingly.
On 06/23/2021 08:26, Laszlo Ersek wrote:
On 06/23/21 08:54, Marvin Häuser wrote:
On 22.06.21 17:34, Laszlo Ersek wrote:No specific concern; my point was that two aspects of the same "class"
On 06/18/21 11:37, Marvin Häuser wrote:Yes, it's UB by the standard, but this is actually how MS implements
On 16.06.21 22:58, Kun Qin wrote:I too am surprised to see
On 06/16/2021 00:02, Marvin Häuser wrote:Huh, interesting. Last time I tried I was told about incompatibilities
2) Is it feasible yet with the current set of supported compilers toMy impression is that flexible arrays are already supported (as seen
support flexible arrays?
Please correct me if I am wrong.
Would you mind letting me know why this is applicable here? We are
trying to seek ideas on how to catch developer mistakes caused by this
change. So any input is appreciated.
with MSVC, but I know some have been dropped since then (2005 and 2008
if I recall correctly?), so that'd be great to allow globally.
flexible array member is a C99 feature, and I didn't even know that we
disallowed it for the sake of particular VS toolchains -- I thought we
had a more general reason than just "not supported by VS versions X
The behavior of OFFSET_OF() would be interesting -- the OFFSET_OF()
macro definition for non-gcc / non-clang:
#define OFFSET_OF(TYPE, Field) ((UINTN) &(((TYPE *)0)->Field))
borders on undefined behavior as far as I can tell, so its behavior is
totally up to the compiler. It works thus far okay on Visual Studio, but
I couldn't say if it extended correctly to flexible array members.
them (or used to anyway?). I don't see why it'd cause issues with
flexible arrays, as only the start of the array is relevant (which is
constant for all instances of the structure no matter the amount of
elements actually stored). Any specific concern? If so, they could be
addressed by appropriate STATIC_ASSERTs.
of undefined behavior didn't need to be consistent with each other.