* Michael Meissner:
> On Sat, May 23, 2009 at 03:55:34PM +0200, Florian Weimer wrote:
>> It seems to me that x86_64 supports direct moves from XMM to
>> general-purpose registers, so that it's possible to access the
>> representation of a floating point value without going through memory:
>>
>> long getbits(double src)
>> {
>> long result;
>> __asm__ ("movq %1, %0" : "=q" (result) : "x" (src));
>> return result;
>> }
>>
>> Is this functionally available as some sort of built-in?
>
> Well you typically can use a union to get the bits.
Okay. I didn't realize that GCC was smart enough to avoid the
round-trip through memory.
> Note, while the instruction exists, it is not always the optimial
> way to move data from the GPRs to the XMMs. On some machines, it is
> better to store the GPR into the stack and reload it into an XMM
> register, than to use the instruction. On some machines, it is
> faster to use the instruction.
Thanks. I can see both behaviors (transfer through memory and through
a register), dependign on the -mtune flag. So there's no need for
another built-in.
