Fast-Math Optimization
A compiler optimization mode that relaxes selected strict IEEE-754 observability requirements so more aggressive floating-point transformations become legal.
Fast-math is not one isolated speed switch. When a compiler is allowed to relax selected assumptions around reassociation, finite values, signed zero, or related floating-point observability, it gains room for vectorization and algebraic transformation.
The trade-off is semantic: expressions that look algebraically equivalent in source code may no longer be bit-for-bit identical, especially around exceptional values or different evaluation orders. The decision should therefore be validated against both performance and a numerical-error or reproducibility budget.
GCC-specific behavior is discussed in GCC -ffast-math and IEEE-754 Semantics, while the data-parallel execution side is connected to SIMD.