On 24 Apr 2015, at 13:01, Chris Torek <to...@elf.torek.net> wrote:
>
> The problem seems likely to be related to odd compiler handling of
> alignment. Consider this code bit, which extracts the essentials:
>
> struct x {
> int x;
> } __attribute__((__aligned__(32)));
>
> struct s1 {
> int a;
> struct x b[1];
> };
>
> struct s2 {
> int a;
> struct x b[];
> };
>
> extern void test2(int);
> void test(void) {
> test2(sizeof(struct s1));
> test2(sizeof(struct s2));
> }
>
> Compiled, here are the two sizeof values (this particular compiler
> output is from clang but gcc and clang both agree on sizeof here):
>
> movl $64, %edi
> callq test2
> movl $32, %edi
> popq %rbp
> jmp test2 # TAILCALL
>
> With the flexible array, (sizeof(struct uma_cache)) is going to be
> 32 bytes smaller than without it.
>
> (I have not looked closely enough to determine what the size should be.)
I'm not sure I would consider this odd. For purposes of sizeof(), the
flexible array member declaration can be thought of as:
struct x b[0];
so while both struct s1 and s2 have members that must be aligned at 32
bytes, the former has one extra, while the latter doesn't. Ergo, 2x32
bytes and 1x32 bytes.
-Dimitry
signature.asc
Description: Message signed with OpenPGP using GPGMail
