[PATCH] linux: mips: Fix getdents64 fallback on mips64-n32

[Adhemerval Zanella]

On 17/11/2020 10:38, Andreas Schwab wrote:
> On Nov 17 2020, Adhemerval Zanella wrote:
> 
>> diff --git a/sysdeps/unix/sysv/linux/mips/mips64/getdents64.c b/sysdeps/unix/sysv/linux/mips/mips64/getdents64.c
>> index d18a5297dc..368128244e 100644
>> --- a/sysdeps/unix/sysv/linux/mips/mips64/getdents64.c
>> +++ b/sysdeps/unix/sysv/linux/mips/mips64/getdents64.c
>> @@ -90,17 +90,14 @@ __getdents64 (int fd, void *buf, size_t nbytes)
>>  
>>    while ((char *) kdp < (char *) skdp + r)
>>      {
>> -      /* This macro is used to avoid aliasing violation.  */
>> -#define KDP_MEMBER(src, member)			     			\
>> -    (__typeof__((struct kernel_dirent){0}.member) *)			\
>> -      memcpy (&((__typeof__((struct kernel_dirent){0}.member)){0}),	\
>> -	      ((char *)(src) + offsetof (struct kernel_dirent, member)),\
>> -	      sizeof ((struct kernel_dirent){0}.member))
>> -
>>        /* This is a conservative approximation, since some of size_diff might
>>  	 fit into the existing padding for alignment.  */
>> -      unsigned short int k_reclen = *KDP_MEMBER (kdp, d_reclen);
>> -      unsigned short int new_reclen = ALIGN_UP (k_reclen + size_diff,
>> +
>> +      /* Obtain the d_ino, d_off, and d_reclen from kernel filled buffer.  */
>> +      struct kernel_dirent kdirent;
>> +      memcpy (&kdirent, kdp, sizeof (struct kernel_dirent));
> 
> How good is the compiler at eliding the memcpy?  What would change if
> kdp would just be void *?

Not that good, I see a large stack usage (1456 vs 1152 from my initial
version) and some more memory load/store memory instructions.  The
kdp type change does not change the code generation.

I don't have a strong preference here, in any case this should be
used only as fallback on older kernels.