This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [committed] Re: PATCH: fine-tuning for can_store_by_pieces


On Sat, Aug 25, 2007 at 02:59:54AM -0400, Jakub Jelinek wrote:
> *************** store_expr (tree exp, rtx target, int ca                                                                         
> *** 4507,4513 ****                                                                                                               
>                                   str_copy_len, builtin_strncpy_read_str,                                                        
>                                   (void *) TREE_STRING_POINTER (exp),                                                            
>                                   MEM_ALIGN (target),                                                                            
> !                                 exp_len > str_copy_len ? 1 : 0);                                                               
>         if (exp_len > str_copy_len)                                                                                              
>         clear_storage (dest_mem, GEN_INT (exp_len - str_copy_len),                                                               
>                        BLOCK_OP_NORMAL);                                                                                         
> --- 4520,4527 ----                                                                                                               
>                                   str_copy_len, builtin_strncpy_read_str,                                                        
>                                   (void *) TREE_STRING_POINTER (exp),                                                            
>                                   MEM_ALIGN (target),                                                                            
> !                                 exp_len > str_copy_len ? 1 : 0,                                                                
> !                                 false);                                                                                        
>         if (exp_len > str_copy_len)                                                                                              
>         clear_storage (dest_mem, GEN_INT (exp_len - str_copy_len),                                                               
>                        BLOCK_OP_NORMAL);                                                                                         
> 
> This is wrong.  You added the memsetp argument to store_by_pieces
> as the 6th, before the endp argument, but you are passing
> exp_len > str_copy_len ? 1 : 0 as memsetp and false as endp.
> This will certainly screw up say
> struct A { char a[20]; };
> void foo (void)
> {
>   struct A a = { "abc" };
>   bar (&a);
> }
> because in that case exp_len (20) is bigger than str_copy_len
> and so clear_storage is needed for the rest of the array.
> If we pass false == 0 as endp, it means the returned endp
> will point to the beginning of the array (&a.a[0]), rather
> than where store_by_pieces stopped.

Surprised this wasn't caught up by make check, I have committed following as
obvious.  This new testcase at least on x86_64-linux and i686-linux
fails after your patch and no longer does after swapping the arguments.

	Jakub

Attachment: O
Description: Text document


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]