[LLVMdev] Another compiler shootout

[Chris Lattner]

On Dec 15, 2008, at 8:26 PM, Jon Harrop wrote:
> Note that the LLVM-generated code is 60% slower than GCC's in the  
> first case.
>
> I am unfamiliar with x86 assembler but I believe the problem is that  
> LLVM is
> calling a function for fsqrt rather than using the x86 op-code.  
> Should I be
> passing some command line arguments or using a newer llvm-gcc to get  
> it to
> emit fsqrt or is that not yet implemented?

Oh, I just had another crazy thought.  This is quite likely the result  
of llvm-gcc disabling the inline asm in your libm headers.  Can you  
also attach the "-E -o foo.i" output from both llvm-gcc and gcc in 32- 
bit mode?

Thanks Jon,

-Chris