On Wed, Jul 30, 2025 at 11:28:11AM GMT, Jakub Jelinek wrote: > On Wed, Jul 30, 2025 at 04:56:06PM +0800, Yang Yujie wrote: > > I'm not sure if this is correct, but isn't accessing __x in its own type > > after do_copy already an undefined behavior? > > > > N3301 J.2 (34) > > > An object has its stored value accessed other than by an lvalue of > > > an allowable type (6.5.1). > > > > where "allowable type" refers to some variant of a compatible type, > > and "_BitInt(PROMOTED_SIZE (x) * __CHAR_BIT__)" is not an allowable > > type to access the object of (x). > > It is copied using memcpy, so accessed using char type. > That can copy also padding/trap etc. bits, but if the ABI say certain > padding bits have some guaranteed state, at least by extension it should be > ok to verify that. > In your patch, it wasn't really testing anything about the padding bits > unless it relies on the compiler not doing or doing some optimization (if it > works at all). By doing the shift, it really tests the content of those > padding bits (and by do_copy being noipa that serves as an optimization > barrier, so it will always test those bits, can't assume they are right > or what exactly __x will contain). > > Jakub
So wouldn't it be simpler if I just change the comparison in my patch into a memcmp? This works regardless of how comparison is implemented. (We all assume that expression (x) may have its address taken.) Yujie
