Over-optimization with clang when building XDSP/Test

[shinjich]

The clang with /Ox + /fp:fast seems to be over-optimize (incorrect pre-calculation) in the XDSP test with updated xDSP.h. This issue will be gone by changing to /fp:precise. The VC generate a function call and there is no issue with /fp:fast. It would be better to modify to complex code not affected by compiler optimization.

test_cpp_with_clang_fpfast_disasm.txt
clang /fp:fast, All complex numbers were 1.0,-0.0.

test_cpp_with_clang_fpprecise_disasm.txt
clang /fp:precise, Less accurate than VC but within updated epsilon range.

test_cpp_with_vc_fpfast_disasm.txt
VC /fp:fast, Looks good. :)

[walbourn]

We generally test with /fp:fast as that's the recommendation for games. I'll check into the clang/LLVM issue here...

[shinjich]

This issue may be due to the optimization for the XMVectorRound (SSE version) used in XMVectorSinCos etc.
Because I replaced XMVectorRound with pseudocode and got a good result. Also, when built with AVX, XMVectorRound is replaced to _mm_round_ps and returns a good result. However, if we specify _MM_FROUND_TO_POS_INF for rounding mode in _mm_round_ps, there will be errors that exceed the epsilon for checking.

[shinjich]

It seems to be the same issue as following. I also added a link to a reproduction code. Thanks,
microsoft/DirectXMath#127