[PATCH] XCore fix for SingleSource/Benchmarks/Misc-C++/Large/sphereflake

[Tim Northover]

>   I tested this on AArch64 (where it's also failing, probably for similar reasons since the image looks identical), and it still fails on both GCC and Clang, each generating a slightly different output.

I looked into this (and I *think* I investigated exactly the same
failure on AArch64 long ago). I think my conclusion (with help from
others at ARM) was that the standard actually says nothing about how
accurate sin has to be. I'd heard the OpenCL standard, of all things,
was about the best standard you could point at for enforcing actual
numbers.

Cheers.

Tim.