ravenslofty / ps2-clang-patches Goto Github PK

The EE Core has a SIMD-ish extension called Multimedia Instructions.

It is quite an eclectic mix, and I would have to go through it in more detail to see what can be done.

Although it has 64-bit addition/subtraction, it doesn't have a 64-bit multiply/divide unit, which is frustrating.

The EE Core has multiple pipelines (it's 2-way superscalar), but it requires the programmer to manually emit (INSN)1 opcodes to harness the otherwise dormant second pipeline. Perhaps a routine for 64-bit multiply can be produced using both multiply pipelines together.

The PS2's FPU behaves in a distinctly non-IEEE 754 fashion; there are no traps (this can be fixed by checking flags, perhaps), a fixed rounding mode (truncation), no NaNs or infinities, and all of it is single-precision.

I've already had to break the N32 ABI by setting long double to be equal to 64-bit double, when the ABI calls for 128-bit long double. I don't think this will meaningfully affect anybody's code though.

The EE can use VU0 through the coprocessor 2 interface to give it upper or lower instructions. This isn't quite as efficient as using VU0 in micro mode, but it's also much more convenient; perhaps Clang could autovectorise it?

I don't know how much effort it will be to ensure Clang doesn't emit these instructions for atomic locking on a single core CPU.

Hello! I'm exploring creating a bare bones PS2 Linux distro that can run off a disc without requiring FMCB or a modchip. The main thing I'm looking for is a compiler that will actually run on the PS2 so it can actually be used for writing new software without depending on cross-compilation, could this project do that?
I'm planning to base it on Maximus32's 3.8 kernel since it's the newest I've found.