clang requires -ffp-contract=fast with -mfma to actually enable FMA instructions
With clang -mfma
is not enough on its own to enable the use of FMA instructions. -ffp-contract=fast
needs to be added as well.
This MR automatically does this, when required, for parity with how FMA is enabled with gcc.
For reference: http://releases.llvm.org/8.0.0/tools/clang/docs/ClangCommandLineReference.html#cmdoption-clang-ffp-contract
Edited by Marco Clemencic