Draft: Fused-multiply-add vectorization

assigned to @kuron

added 1 commit

• 8bebd100 - Fix FMA insertion failures in remainder loops and with RNG

Compare with previous version

added 1 commit

• 6e3a147c - don't convert plain negation into fused-multiple add

Compare with previous version

The current state works, but the visitor that inserts the FMAs into the AST is severely broken. For example, it loses field alignment information (which leads to test_alignment_and_correct_ghost_layers failing on AVX) and OpenMP number of threads. Nevertheless I can see that it improves the compressible D3Q19 TRT from 28 MLUPS to 40 MLUPS on a single core of my Apple M1.