Re: [musl] add loongarch64 port v3

[王洪亮 <wanghongliang@loongson.cn>]

在 2022/5/25 下午8:32, Rich Felker 写道:
> On Wed, May 25, 2022 at 06:08:23PM +0800, 王洪亮 wrote:
>> 在 2022/5/24 下午8:32, Rich Felker 写道:
>>> What we've been trying to say is that there are several cases, none of
>>> which seem to need it:
>>>
>>> 1. You create an object with declared type struct sigcontext. In this
>>>     case, the flexible array member at the end is not present at all
>>>     (because that's how C works) which means there's no extended
>>>     context which needs additional alignment and probably also means
>>>     this is not a usable way of creating sigcontext structs.
>>>
>>> 2. You malloc storage for the object with space for the flexible array
>>>     member. In this case the allocation has alignment max_align_t and
>>>     everything is fine.
I don't understand what is alignment max_align_t? I found the max_align_t
definition in musl,is this it?

TYPEDEF struct { long long __ll; long double __ld; } max_align_t;

I understand if FAM is not specified alignment,FAM is aligned according to
its type size,why is max_align_t?
>>>
>>> 3. You get the object from the kernel pushing it onto the stack in a
>>>     signal frame. This is probably actually the only case the type is
>>>     usable in, and of course it has whatever alignment the kernel gave
>>>     it.
>> Specify the __attribute__((__aligned__(16))) in musl,is used to be
>> consistent with kernel.if I removed the __attribute__((__aligned__(16))),
>> there is a libc-test fail in pthread_cancel.exe.the reason is that the
>> offset of uc->uc_mcontext from the start of uc obtained in cancel_handler
>> is inconsistent with kernel pushing it onto the stack in a signal frame.
>> so I understand the __attribute__((__aligned__(16))) is necessary in musl.
>>
>> src/thread/pthread_cancel.c
>>
>> static void cancel_handler(int sig, siginfo_t *si, void *ctx)
>> {
>>          pthread_t self = __pthread_self();
>>          ucontext_t *uc = ctx;
>>          uintptr_t pc = uc->uc_mcontext.MC_PC;
>>           ...
>> }
>>
>> musl/arch/loongarch64/bits/signal.h:
>>
>> typedef unsigned long greg_t, gregset_t[32];
>> typedef struct sigcontext {
>>          unsigned long pc;
>>          gregset_t gregs;
>>          unsigned int flags;
>>          unsigned long extcontext[];
>> }mcontext_t;
>>
>> linux/arch/loongarch/include/uapi/asm/sigcontext.h:
>>
>> struct sigcontext {
>>          __u64   sc_pc;
>>          __u64   sc_regs[32];
>>          __u32   sc_flags;
>>          __u64   sc_extcontext[0] __attribute__((__aligned__(16)));
>> };
> This is because ucontext_t is defined without explicit padding before
> uc_mcontext. Add "long __uc_pad;" or similar before it so that the
> offset is explicitly what it's supposed to be rather than a
> consequence ot overalignment.

Add "long __uc_pad;" before uc_mcontext can resolve offset error,
why it is better than sc_extcontext[] __attribute__((__aligned__(16)))?
isn't it more direct to be consistent with kernel?

> Rich