Hi, I have been trying to work out what all the rules are for where attribute((aligned(x))) can be used and where it can decrease alignment and where it can only increase it. The results I think I have found are: Specifying aligned on a structure member can only increase alignment, but if attribute((packed)) is specified also it can also decrease it. Specifying aligned on a static variable can increase or decrease alignment. Specifying aligned on a stack variable can increase or decrease alignment, but (at least on ARM) the platform may not allow misaligned stack variables, and so even though _Alignof() returns 1, corresponding misaligned access instructions are not generated. Specifying aligned on a function argument is not allowed. Specifying aligned on a pointer target is allowed but can only increase alignment. Specifying aligned on a pointer target is not allowed when the pointer itself is a function argument. Specifying aligned on a typedef can increase or decrease alignment, and the corresponding type can be used as a pointer target, even where the pointer is a function argument. This all seems very inconsistent. Is there some systematic reason to it all, or is it just historic? Generating an error when the attribute is on the target of a pointer only when the pointer is a function argument really seems like a bug! I found some details on this stack overflow question: https://stackoverflow.com/questions/73152859/accessing-unaligned-struct-member-using-pointers (although I am particularly interested in the case where no structs are involved) I also have experimented with various cases: https://godbolt.org/z/73WPzEjYq Does anyone have any light to shed on this all? Many thanks, Tom
