Improve macOS AArch64 JIT stencils by relaxing GOT loads

[lgeiger]

Proposal:

The AArch64 JIT has support for optimising pairs of instructions that load an address through the global offset table (GOT).
On Linux, the JIT optimiser recognises assembly like:

adrp x9, :got:__PyRuntime
ldr  x9, [x9, :got_lo12:__PyRuntime]

cpython/Tools/jit/_optimizers.py

Lines 586 to 591 in 0ba7be9

	_re_large_const_1 = re.compile(
	r"\s*(?P<instruction>adrp)\s+(?P<register>x\d\d?),.*:got:(?P<value>[_A-Za-z0-9]+).*"
	)
	_re_large_const_2 = re.compile(
	r"\s*(?P<instruction>ldr)\s+(?P<register>x\d\d?),.*:got_lo12:(?P<value>[_A-Za-z0-9]+).*"
	)

On macOS, Clang uses a different syntax recognised by the Mach-O linker:

adrp x9, __PyRuntime@GOTPAGE
ldr  x9, [x9, __PyRuntime@GOTPAGEOFF]

We can add support for this using the existing patch_aarch64_33rx() function just by adjusting the regex which would enable the same optimisation we already use on Linux:

adrp x9, <__PyRuntime page>
add  x9, x9, <page offset>

This is a natural followup to #157040. I have a small patch ready once jit development is unblocked.

Benchmark:

On my M5 MacBook Pro this shows a 1.16% improvement on the pyperformance suite though results are very noisy. Probably best to re-run with the official benchmarking runners.

Expand for full benchmarking results

Has this already been discussed elsewhere?

This is a minor feature, which does not need previous discussion elsewhere

Links to previous discussion of this feature:

#148598, #148501, #157040 /cc @diegorusso