Implement "consume" loads on PowerPC

[taylordotfish]

Like ARM, PowerPC is also a weakly ordered architecture where "acquire" loads are more expensive than "consume" loads, which require no special instructions.

[taiki-e]

Thanks for the PR! Unfortunately, the resulting code is less efficient than acquire load that uses isync + branching because LLVM uses lwsync for powerpc's compiler_fence: https://godbolt.org/z/YG5M5foE4

EDIT: Depending on the CPU, lwsync and isync+branching may be equally efficient, but that is an LLVM issue regarding how to lower acquire load.

[taylordotfish]

@taiki-e Huh… but why is compiler_fence issuing any kind of *sync at all? Based on the documentation, I would've expected it not to generate any special instructions:

compiler_fence does not emit any machine code, but restricts the kinds of memory re-ordering the compiler is allowed to do.

But empirically, that's clearly not the case…

[taiki-e]

compiler_fence's documentation is incorrect, see rust-lang/rust#62256 for more.