From: Jiri Slaby > Sent: 13 December 2022 11:15 > > On 13. 12. 22, 9:30, David Laight wrote: > > From: Tejun Heo <htejun@xxxxxxxxx> On Behalf Of 'Tejun Heo' > >> Sent: 12 December 2022 21:47 > >> To: Jiri Slaby <jirislaby@xxxxxxxxxx> > >> Cc: David Laight <David.Laight@xxxxxxxxxx>; Christoph Hellwig <hch@xxxxxxxxxxxxx>; linux- > >> kernel@xxxxxxxxxxxxxxx; Martin Liska <mliska@xxxxxxx>; Josef Bacik <josef@xxxxxxxxxxxxxx>; Jens > Axboe > >> <axboe@xxxxxxxxx>; cgroups@xxxxxxxxxxxxxxx; linux-block@xxxxxxxxxxxxxxx > >> Subject: Re: [PATCH] block/blk-iocost (gcc13): cast enum members to int in prints > >> > >> On Mon, Dec 12, 2022 at 01:14:31PM +0100, Jiri Slaby wrote: > >>>> If so, my suggestion is just sticking with the old behavior until we switch > >>>> to --std=g2x and then make one time adjustment at that point. > >>> > >>> So is the enum split OK under these circumstances? > >> > >> Oh man, it's kinda crazy that the compiler is changing in a way that the > >> same piece of code can't be compiled the same way across two adjoining > >> versions of the same compiler. But, yeah, if that's what gcc is gonna do and > >> splitting enums is the only way to be okay across the compiler versions, > >> there isn't any other choice we can make. > > > > It is also a silent code-breaker. > > Compile this for 32bit x86: > > > > enum { a = 1, b = ~0ull}; > > But having ull in an enum is undefined anyway. C99 allows only int > constants. gnuC supports ulong expressions (IIRC). gcc supports 'long long' as well - 64bit on 32bit systems. In practical terms it really doesn't matter what C99 (or any other version) says, the important thing is that the compiler accepted it. > > extern int foo(int, ...); > > int f(void) > > { > > return foo(0, a, 2); > > } > > > > gcc13 pushes an extra zero onto the stack between the 1 and 2. > > So this is sort of "expected". For some definitions of "expected" :-) Note that it (probably) makes no actual difference to some architectures (like 64bit x86) where all varargs parameters are passed as 64bit. Extending a value to 64bits just makes the high bits well defined. (The high bits of stacked 32bit args are undefined.) David - Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales)