libm optimizations

[parasyte]

We're currently using libm to implement the missing compiler builtins for f32 and f64 primitives. The libm crate implements these builtins in Rust. While it provides a very workable springboard to get started quickly, it leaves a lot to be desired with fast math.

The N64 CPU supports many of these operations natively through the floating point unit, COP1. Here's a table summarizing the builtin to FPU instruction relationships:

builtin	FPU instruction
fabs	abs.d
fabsf	abs.s
ceil	ceil.l.d, cvt.d.l
ceilf	ceil.w.s, cvt.s.w
floor	floor.l.d, cvt.d.l
floorf	floor.w.s, cvt.s.w
round	round.l.d, cvt.d.l
roundf	round.w.s, cvt.s.w
sqrt	sqrt.d
sqrtf	sqrt.s
trunc	trunc.l.d, cvt.d.l
truncf	trunc.w.s, cvt.s.w

Another alternative is using LLVM intrinsics from rustc. (How?)

• https://github.com/rust-lang/rust/blob/master/src/libcore/intrinsics.rs
• https://github.com/rust-lang/rust/blob/master/src/librustc_codegen_llvm/intrinsic.rs
• https://github.com/rust-lang-nursery/compiler-builtins