Re: compiling xmmintrin.h

[Tim Prince <tprince at myrealbox dot com>]

ranjith kumar wrote:
> Hi,
>     I wrote a c program which uses functions and data
> types defined in xmmintrin.h.
>  Suprisingly,  in assembly code, produced by gcc, each
> function call is replaced by a single assembly
> instruction.
>
> What is the special about functions defined in
> xmmintrin.h.
> How they are compiled by gcc?
>
> Any answers or any matirial?
> Thanks in advance.
>
>
>   
These intrinsics were developed as a way of programming at a lower 
level, without being tied exclusively to one instruction set or 
assembler syntax.  So, the code should be portable between commercial 
compilers and gcc, and between 32-bit/x86-64/ia64.   They also provide a 
way to optimize by use of parallel instructions, in situations where 
vectorization doesn't work, or when a less capable compiler is in use.  
Not that I would use them as a way of persuading gcc-3.2 or MSVC to do 
what gcc-4.3 does by auto-vectorization.
You could learn a lot about them simply by studying xmmintrin.h.  The 
gcc version is more readable than the commercial ones, but they have 
much the same content.  To some extent, it's a way of automatically 
changing asm macros when you change target platforms, within the 
implemented range, as well as providing an alternative to each 
programmer inventing new wheels.
Google will give you some hints about additional reading material.  It 
doesn't matter whether you use the Windows or linux oriented articles.