intrinsics.fmuladdf{16,32,64,128}: expose llvm.fmuladd.* semantics

Add intrinsics `fmuladd{f16,f32,f64,f128}`. This computes `(a * b) +
c`, to be fused if the code generator determines that (i) the target
instruction set has support for a fused operation, and (ii) that the
fused operation is more efficient than the equivalent, separate pair
of `mul` and `add` instructions.

https://llvm.org/docs/LangRef.html#llvm-fmuladd-intrinsic

MIRI support is included for f32 and f64.

The codegen_cranelift uses the `fma` function from libc, which is a
correct implementation, but without the desired performance semantic. I
think this requires an update to cranelift to expose a suitable
instruction in its IR.

I have not tested with codegen_gcc, but it should behave the same
way (using `fma` from libc).

src/intrinsic/mod.rs

@@ -66,6 +66,9 @@ fn get_simple_intrinsic<'gcc, 'tcx>(
         sym::log2f64 => "log2",
         sym::fmaf32 => "fmaf",
         sym::fmaf64 => "fma",
+        // FIXME: calling `fma` from libc without FMA target feature uses expensive sofware emulation
+        sym::fmuladdf32 => "fmaf", // TODO: use gcc intrinsic analogous to llvm.fmuladd.f32
+        sym::fmuladdf64 => "fma",  // TODO: use gcc intrinsic analogous to llvm.fmuladd.f64
         sym::fabsf32 => "fabsf",
         sym::fabsf64 => "fabs",
         sym::minnumf32 => "fminf",