[X86] Remove lowerToAddSubOrFMAddSub and update test coverage

[RKSimon]

lowerToAddSubOrFMAddSub is a x86 DAG lowering that tries to vectorize fadd/fsub code into ADDSUB + VFMADDSUB/VFMSUBADD vector instructions but SLP/VectorCombine should now be vectorising these patterns leaving the backend code superfluous.

Very similar to #143000

So, this task involves:

• Disable lowerToAddSubOrFMAddSub
• Review all CodeGen/X86 test changes
• Convert the changed CodeGen/X86 tests into similar PhaseOrdering/X86 tests to track the vectorized IR (for the various SSE/AVX/FMA3/FMA4 targets)
• Replace the CodeGen/X86 test coverage with the real vectorized IR
• Remove lowerToAddSubOrFMAddSub and all its support methods
• Handle any CodeGen regressions (maybe a cost model, transform or backend fix)

[llvmbot]

@llvm/issue-subscribers-backend-x86

Author: Simon Pilgrim (RKSimon)

lowerToAddSubOrFMAddSub is a x86 DAG lowering that tries to vectorize fadd/fsub code into ADDSUB + VFMADDSUB/VFMSUBADD vector instructions but SLP/VectorCombine should now be vectorising these patterns leaving the backend code superfluous.

Very similar to #143000

So, this task involves:

• Disable lowerToAddSubOrFMAddSub
• Review all CodeGen/X86 test changes
• Convert the changed CodeGen/X86 tests into similar PhaseOrdering/X86 tests to track the vectorized IR (for the various SSE/AVX/FMA3/FMA4 targets)
• Replace the CodeGen/X86 test coverage with the real vectorized IR
• Remove lowerToAddSubOrFMAddSub and all its support methods
• Handle any CodeGen regressions (maybe a cost model, transform or backend fix)