[PATCH v5 0/6] Add support for LoongArch32

[mengqinggang]

在 2026/3/4 14:46, mengqinggang 写道:
>
> 在 2026/3/4 09:52, mengqinggang 写道:
>>
>> 在 2026/3/4 00:33, Xi Ruoyao 写道:
>>> On Tue, 2026-03-03 at 19:28 +0800, mengqinggang wrote:
>>>> 在 2026/3/3 18:11, Xi Ruoyao 写道:
>>>>> On Tue, 2026-03-03 at 17:54 +0800, mengqinggang wrote:
>>>>>> LoongArch64 gcc generates call memset to fully zero-filling the 
>>>>>> struct.
>>>>>> But LoongArch32 gcc not generate call memset.
>>>>>> Maybe we can fix LoongArch32 gcc to resolve this problem.
>>>>> This is not a compiler bug so there's nothing to be fixed in the
>>>>> compiler.  In C using = {...} to initialize a struct with automatic
>>>>> storage duration will leave the padding bits unspecified, unless the
>>>>> initializer is empty.
>>>>>
>>>>> See https://gcc.gnu.org/PR104607 which has been resolved as invalid
>>>>> years ago.
>>>> According to what you said, loongarch64 gcc should not generate call
>>>> memset assembly code?
>>> I don't think it should.  This seems a missed-optimization on almost 
>>> all
>>> GCC ports (GCC r16-7721 actually emits the memset for LA32...)
>>>
>> Only loongarch64 gcc generate call memset for the following code.
>>
>>   1 #include <errno.h>
>>   2 #include <fcntl.h>
>>   3 #include <limits.h>
>>   4 #include <sys/pidfd.h>
>>   5 #include <sys/wait.h>
>>   6 #include <stdlib.h>
>>   7 #include <stdio.h>
>>   8 #include <unistd.h>
>>   9
>>  10 static pid_t ppid;
>>  11 static uid_t puid;
>>  12
>>  13 int
>>  14 main (void)
>>  15 {
>>  16   ppid = getpid ();
>>  17   puid = getuid ();
>>  18   pid_t pid = fork ();
>>  19   if (pid != 0)
>>  20   {
>>  21     int pidfd = pidfd_open (pid, 0);
>>  22     printf("pidfd %d\n", pidfd);
>>  23     {
>>  24       siginfo_t info =
>>  25         {
>>  26           .si_signo = SIGUSR2,
>>  27           .si_errno = EAGAIN,
>>  28           .si_code = -10,
>>  29           .si_pid = ppid,
>>  30           .si_uid = puid
>>  31         };
>>  32       int r = pidfd_send_signal (pidfd, SIGUSR2, &info, 0);
>>  33       printf("pidfd r = %d\n", r);
>>  34     }
>>  35   }
>>  36   return 0;
>>  37 }
>>
> If wordsize is 64, there is an int __pad0.
> Maybe this difference causes the loongarch64 gcc to generate call memset,
> while the loongarch32 gcc does not.


On 32 bit target, gcc believes all fields have been initialized. So gcc 
dose not generate call memset.
Currently, I think  add a memset call in source code is a good idea.


>
> sysdeps/unix/sysv/linux/bits/types/siginfo_t.h:
>
>  36 typedef struct
>  37   {
>  38     int si_signo;               /* Signal number.  */
>  39 #if __SI_ERRNO_THEN_CODE
>  40     int si_errno;               /* If non-zero, an errno value 
> associated with
>  41                                    this signal, as defined in 
> <errno.h>.  */
>  42     int si_code;                /* Signal code.  */
>  43 #else
>  44     int si_code;
>  45     int si_errno;
>  46 #endif
>  47 #if __WORDSIZE == 64
>  48     int __pad0;                 /* Explicit padding.  */
>  49 #endif
>  50
>  51     union
>  52       {
>  53         int _pad[__SI_PAD_SIZE];
>  54
>  55          /* kill().  */
>  56         struct
>  57           {
>  58             __pid_t si_pid;     /* Sending process ID.  */
>  59             __uid_t si_uid;     /* Real user ID of sending 
> process.  */
>  60           } _kill;
>