Kevin G (kevin@cubitdesigns.com) on 7/8/10 wrote:
---------------------------
>David Kanter (dkanter@realworldtech.com) on 7/8/10 wrote:
>---------------------------
>>
>>But I can totally understand why an ABI would preserve 80b FP for those who *really
>>need* it. But 99% of people are fine with 64b, so I think pushing people in that
>>direction and making it difficult to use x87 is a reasonable choice.
>
>For those that were wanting 80 bit precision, I thought that there were rumblings
>that the successor to AVX would support 128 bit precision. Such a change would finally
>allow x87 support to be phased out.

Yeah, but do you seriously believe that a quad precision part would be faster than x87? Obviously, its possible if you throw enough hardware at it -- and maybe Intel would do it to claim the third IEEE-754-2008 binary format or something; They do take the standard more seriously than some -- but there is no market for quad to support that sort of investment in anything other than boutique parts and certainly the perf/Watt story would not be good. It would be better to spend the silicon on other features like a larger ROB or vector unit or unsigned vector compares or making LSU false aliasing less common or something. These things would benefit more code. I think their level of investment in x87 is about right. Its enough to make sure legacy code still functions without eating up a lot of chip area.

Math libraries at times will use the x87 in order to deliver highly accurate results for double precision math functions. The power function is a nice example, where you need more bits in the intermediate representation than in the final representation to get an accurate result. Without x87, you have to get the extra precision from elsewhere, either by storing values as a high + low sum and doing what amounts to multiprecision arithmetic on it (read: expensive), or where you are sufficiently talented and fortunate to be able to use, leverage some other higher precision hardware feature such as a fused multiply-add. Obviously, Intel doesn't currently do FMA so the choices are not good.