Minimal implimentation of prefetch

[NCGThompson]

rustc_codegen_gcc/src/intrinsic/mod.rs

Lines 170 to 175 in 1a8e0c3

	sym::prefetch_read_data
	| sym::prefetch_write_data
	| sym::prefetch_read_instruction
	| sym::prefetch_write_instruction => {
	unimplemented!();
	}

Currently, all the prefetch operations lead to unimplemented!() macro. However, prefetches are just a suggestion and can be interpreted as a noop, like assume is.

[antoyo]

__builtin_prefetch can probably be used here. Not sure if there's anything in gcc to prefetch instructions.

[NCGThompson]

__builtin_prefetch can probably be used here. Not sure if there's anything in gcc to prefetch instructions.

I'll take a look while I'm doing the constanty_p one.

[NCGThompson]

__builtin_prefetch can probably be used here. Not sure if there's anything in gcc to prefetch instructions.

There is nothing for instructions. Does it make sense to use a data prefetch call in place of an instruction prefetch?

[antoyo]

I'm not sure. It could be interesting to try and see if the resulting assembly is affected.

[NCGThompson]

The assembly will be different on many platforms, but maybe a data prefetch will still benefit instruction load time if it brings it from main memory to intermediate cache. While experimenting does sound interesting, I think we will get better results by asking an expert in computer architecture. I'm assuming we can use conditional compilation and only use it on targets that would benefit. In the mean time, I lean towards being conservative and making instruction prefetch a noop.

[antoyo]

That's OK for me.