> From: Keith Marshall <keith****@users*****> > Date: Tue, 30 Jun 2020 07:54:18 +0100 > > > GDB is a C++ program these days, and is compiled with -fpermissive, so > > this causes errors. > > Surely -fpermissive relaxes the error checking for type conflicts, > (possibly among other conditions -- I don't know what its scope is), so > such errors would be downgraded to warnings. Do you, perhaps, mean that > GDB is compiled _without_ -fpermissive? You are probably right. All I know is that g++ emitted this: CXX remote-utils.o remote-utils.cc: In function 'void handle_accept_event(int, gdb_client_data)': remote-utils.cc:146:69: error: invalid conversion from 'socklen_t*' {aka 'unsigned int*'} to 'int*' [-fpermissive] 146 | remote_desc = accept (listen_desc, (struct sockaddr *) &sockaddr, &len); | ^~~~ | | | socklen_t* {aka unsigned int*} I guess I mis-interpreted what that confusing message means to say. > There's more to it. As I reported in: > > https://savannah.gnu.org/bugs/?57725 > > a primary motivator for the change to unsigned socklen_t was to avoid a > proliferation of casts in <wspiapi.h>, (of which, at the time, I seemed > to be writing quite a number). Reviewing it again, I seem to have ended > up with only one such cast which remains necessary, at line 917: > > if( (sa == NULL) || (len < sizeof( struct sockaddr )) ) > > With reversion to signed socklen_t, this should become: > > if( (sa == NULL) || (len < (socklen_t)(sizeof( struct sockaddr ))) ) You are a better judge of this, but to my mind a single cast is a small price to pay for avoiding the kind of mess I faced yesterday, with socklen_t popping up in compilation errors left and right. Gnulib already removed the unnecessary conflicting typedef, and I will shortly report the problem in the GDB sources to its developers. But I think the problems with code which assumes the last argument of 'accept' and its ilk could be a 'socklen_t *' are a much harder nut, and so going back to a signed int will be a good compromise. Thanks.
Fork