On 08/21/2018 11:49 AM, Martin Sebor wrote: > On 08/21/2018 09:44 AM, Joseph Myers wrote: >> On Tue, 21 Aug 2018, Martin Sebor wrote: >> >>> Sure, but the only valid argument to %ls is wchar_t*. Passing >>> it something else is undefined. >> >> Well, (wchar_t *)"something\0\0\0\0" would be OK given >> -fno-strict-aliasing and if you know the alignment is OK. Do we have >> that >> information about the type cast to, as opposed to the type of the string >> constant, at this point? > > In the simple cases like the one above the cast is gone. Only > in some more involved cases is the type of the argument preserved. > I responded to Jeff with one such example here: > > https://gcc.gnu.org/ml/gcc-patches/2018-08/msg01296.html > > If supporting (wchar_t *)"...\0\0\0\0" with %ls is viewed as > important (despite it being undefined) then the function does > need an ELTSIZE argument so it knows what to count. In that > event, in order to detect the problem cases we have been > discussing (missing nuls and mismatched argument types), > the function it must not fail when ELTSIZE is not equal > to the size of actual array element. Instead, it needs to > return the element type to the caller which then needs to > do the validation and issue a diagnostic. So I think that argues that the sanity check should be removed or otherwise neutered when if/when we attack the argument type mismatching problems.
Jeff
