https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87502
Martin Liška <marxin at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |jakub at gcc dot gnu.org,
| |marxin at gcc dot gnu.org
--- Comment #1 from Martin Liška <marxin at gcc dot gnu.org> ---
(In reply to M Welinder from comment #0)
> Created attachment 44776 [details]
> Preprocessed source code
>
> It appears that gcc is creating quite poor code when "c-style strings"
> are used to construct std::string objects. Ideally, the result ought
> to be just a few move instructions for small strings.
>
>
> Host: Linux x86_64 4.4.140-62-default (OpenSuSE)
>
> Test code:
> ---------------------------------------------------------------
> #include <string>
>
> extern void foo (const std::string &);
>
> void
> bar ()
> {
> foo ("abc");
> foo (std::string("abc"));
> }
> ---------------------------------------------------------------
>
>
>
> # /usr/local/products/gcc/8.2.0/bin/g++ -std=gnu++1z -S -m32 -O3 ttt.C
> # grep 'call.*construct' ttt.s
> call
> _ZNSt7__cxx1112basic_stringIcSt11char_traitsIcESaIcEE12_M_constructIPKcEEvT_S
> 8_St20forward_iterator_tag.constprop.18
> call
> _ZNSt7__cxx1112basic_stringIcSt11char_traitsIcESaIcEE12_M_constructIPKcEEvT_S
> 8_St20forward_iterator_tag.constprop.18
>
> Here gcc generates complete calls to the generic string construction
> even though the strings are constructed from small, known strings.
With -O2 -fdump-ipa-inline says:
function not declared inline and code size would grow
>
> "-std=gnu++1z" is important; "-m32" and "-O3" (as opposed to "-m64" and
> "-O2") are not.
With -O3 more inlining happens.
>
> # /usr/local/products/gcc/8.2.0/bin/g++ -S -m32 -O3 ttt.C
> # grep 'call.*construct' ttt.s
> # (nada)
>
> No calls -- good. In this case gcc generates this fragment:
>
> _Z3barv:
> .LFB1084:
> .cfi_startproc
> .cfi_personality 0,__gxx_personality_v0
> .cfi_lsda 0,.LLSDA1084
> pushl %ebp
> .cfi_def_cfa_offset 8
> .cfi_offset 5, -8
> movl $25185, %edx
> movl %esp, %ebp
> .cfi_def_cfa_register 5
> pushl %edi
> pushl %esi
> .cfi_offset 7, -12
> .cfi_offset 6, -16
> leal -48(%ebp), %esi
> pushl %ebx
> .cfi_offset 3, -20
> leal -40(%ebp), %ebx
> subl $56, %esp
> movl %ebx, -48(%ebp)
> pushl %esi
> movw %dx, -40(%ebp)
> movb $99, -38(%ebp)
> movl $3, -44(%ebp)
> movb $0, -37(%ebp)
> .LEHB6:
> .cfi_escape 0x2e,0x10
> call _Z3fooRKNSt7__cxx1112basic_stringIcSt11char_traitsIcESaIcEEE
> [...]
>
> This is better than a call, but not great:
> 1. The string is moved into position in three chunks (25185, 99, 0).
> This probably comes from inlined memcpy of 3 bytes, but the source
> is zero-terminated so rounding the memcpy size up to 4 would have
> been better.
Yes we end up with:
__builtin_memcpy (&D.30710.D.23004._M_local_buf, "abc", 3);
> 2. It's unclear why 25185 is passed through a register.
It's somehow connected to fact that constant are somehow expensive
on x86_64. Jakub can help here..
