[Bug 255840] __get_locale() is inefficient

[bugzilla-noreply at freebsd.org]

https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=255840

            Bug ID: 255840
           Summary: __get_locale() is inefficient
           Product: Base System
           Version: CURRENT
          Hardware: amd64
                OS: Any
            Status: New
          Severity: Affects Some People
          Priority: ---
         Component: bin
          Assignee: bugs at FreeBSD.org
          Reporter: markj at FreeBSD.org

In libc we have:

195 /**                                                                         
196  * Returns the current locale for this thread, or the global locale if none
is                                                                              
197  * set.  The caller does not have to free the locale.  The return value
from                                                                            
198  * this call is not guaranteed to remain valid after the locale changes. 
As                                                                              
199  * such, this should only be called within libc functions.                  
200  */                                                                         
201 static inline locale_t __get_locale(void)                                   
202 {                                                                           
203                                                                             
204         if (!__has_thread_locale) {                                         
205                 return (&__xlocale_global_locale);                          
206         }                                                                   
207         return (__thread_locale ? __thread_locale :
&__xlocale_global_locale);                                                      
208 }

Here, __has_thread_locale and __xlocale_globale_locale are global variables. 
In the common case, !__has_thread_locale is true.  __thread_locale is a
thread-local variable.

This function is called any time MB_CUR_MAX is loaded, which may happen
frequently (see PR 255551 for example).

On main, __get_locale() compiles to this:

   0x000000080115e300 <+0>:     push   %rbp
   0x000000080115e301 <+1>:     mov    %rsp,%rbp
   0x000000080115e304 <+4>:     push   %rbx
   0x000000080115e305 <+5>:     push   %rax
   0x000000080115e306 <+6>:     mov    0x113fbb(%rip),%rbx        # 0x8012722c8
   0x000000080115e30d <+13>:    data16 lea 0x113fa3(%rip),%rdi        #
0x8012722b8
   0x000000080115e315 <+21>:    data16 data16 rex.W call 0x8012654b0
<__tls_get_addr at plt>
   0x000000080115e31d <+29>:    mov    (%rax),%rax
   0x000000080115e320 <+32>:    test   %rax,%rax
   0x000000080115e323 <+35>:    mov    0x113e6e(%rip),%rcx        # 0x801272198
   0x000000080115e32a <+42>:    cmove  %rcx,%rax
   0x000000080115e32e <+46>:    cmpl   $0x0,(%rbx)
   0x000000080115e331 <+49>:    cmove  %rcx,%rax
   0x000000080115e335 <+53>:    mov    0x18(%rax),%rax
   0x000000080115e339 <+57>:    mov    0x70(%rax),%eax
   0x000000080115e33c <+60>:    add    $0x8,%rsp
   0x000000080115e340 <+64>:    pop    %rbx
   0x000000080115e341 <+65>:    pop    %rbp
   0x000000080115e342 <+66>:    ret    

In particular, the address of __thread_locale is obtained even if it isn't
going to be used because no threads have set a per-thread locale using
uselocale(3).  But to obtain this address we have to call into rtld, and the
call has a significant cost: a program which performs the comparison MB_CUR_MAX
== 1 500,000,000 times runs in about 2.7s on my workstation.  With libc
modified to split the test of __thread_locale into a separate function, the
runtime is reduced to 1.0s.

I'm not quite sure why clang compiles __get_locale() this way. I presume it's
to avoid branches, but it's quite suboptimal.

-- 
You are receiving this mail because:
You are the assignee for the bug.