[cfe-dev] [libc++] inlining of char_trats<T>::eq() and friends wrt r217961

[Dan Albert]

Following r217961, mclow asked me to verify that it didn't interfere too
badly with inlining of these calls on clang, and to send a summary here for
the record.

I used the following (trivial) test:

    #include <string>

    int main(int argc, char** argv) {
      // Use values from argv to prevent the result from being
      // constexpr evaluated.
      if (std::char_traits<char>::eq(argv[1][0], argv[1][1])) {
        return 0;
      } else {
        return 1;
      }
    }

Level 2 and 3 both inline the call, before and after r217961:
$ clang -O2 ...
$ objdump -d a.out
0000000000400670 <main>:
  400670:       48 8b 46 08             mov    rax,QWORD PTR [rsi+0x8]
  400674:       0f b6 48 01             movzx  ecx,BYTE PTR [rax+0x1]
  400678:       0f b6 00                movzx  eax,BYTE PTR [rax]
  40067b:       39 c8                   cmp    eax,ecx
  40067d:       0f 95 c0                setne  al
  400680:       0f b6 c0                movzx  eax,al
  400683:       c3                      ret
  400684:       66 2e 0f 1f 84 00 00    nop    WORD PTR cs:[rax+rax*1+0x0]
  40068b:       00 00 00
  40068e:       66 90                   xchg   ax,ax

Before r217961, -O0 and -O1 both inline:
$ clang -O1 ...
$ objdump -d a.out
0000000000400670 <main>:
  400670:       48 8b 46 08             mov    rax,QWORD PTR [rsi+0x8]
  400674:       0f b6 48 01             movzx  ecx,BYTE PTR [rax+0x1]
  400678:       0f b6 00                movzx  eax,BYTE PTR [rax]
  40067b:       39 c8                   cmp    eax,ecx
  40067d:       0f 95 c0                setne  al
  400680:       0f b6 c0                movzx  eax,al
  400683:       c3                      ret
  400684:       66 2e 0f 1f 84 00 00    nop    WORD PTR cs:[rax+rax*1+0x0]
  40068b:       00 00 00
  40068e:       66 90                   xchg   ax,ax

After r217961, neither -O0 nor -O1 inline:
$ clang -O1 ...
$ objdump -d a.out
...
  4006bc:       e8 0f 00 00 00          call   4006d0
<_ZNSt3__111char_traitsIcE2eqEcc>
...

Talked about this with mclow, and we're not worried about it, since it
still inlines for the optimized modes, and the patch does fix compatibility
with GCC.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-dev/attachments/20140918/38d4a573/attachment.html>