anti-optimization of _DecimalXX by -ffast-math

[Roman Kononov]

Hi,

$ gcc -v
Using built-in specs.
COLLECT_GCC=gcc
COLLECT_LTO_WRAPPER=/home/rk/gcc/libexec/gcc/x86_64-unknown-linux-gnu/4.5.0/lto-wrapper
Target: x86_64-unknown-linux-gnu
Configured with: ./configure --prefix=/home/rk/gcc --enable-languages=c,c++
Thread model: posix
gcc version 4.5.0 20100112 (experimental) (GCC)
$ cat test.c
_Decimal64 foo(_Decimal64 a,_Decimal64 b) { return a+b; }
$ gcc -c -O3 -save-temps test.c
$ cat test.s
  ...
foo:    subq    $24, %rsp
        movq    %xmm0, 8(%rsp)
        call    __bid_adddd3
        movq    %xmm0, 8(%rsp)
        addq    $24, %rsp
        ret
  ...
$ gcc -c -O3 -funsafe-math-optimizations -save-temps test.c
$ cat test.s
  ...
foo:    subq    $24, %rsp
        movq    %xmm0, 8(%rsp)
        movq    %xmm1, 8(%rsp)
        movq    8(%rsp), %rcx
        movdqa  %xmm0, %xmm1
        movq    %rcx, 8(%rsp)
        movq    8(%rsp), %xmm0
        call    __bid_adddd3
        movq    %xmm0, 8(%rsp)
        addq    $24, %rsp
        ret
  ...
Is there a good reason to place something on the stack? Why does 
-funsafe-math-optimizations (which is a part of -ffast-math) make things 
even worse? It actually swaps the arguments for __bid_adddd3().

Thanks :)