i#5383: Fix macOS preload on a64

[ndrewh]

This fixes several build and runtime issues under macOS on ARM64 on macOS 14.4.1 (Sonoma)

• FEATURE_PAUTH is now detected on macOS using a sysctl (capability MSRs cannot be read in userspace on M3 at least)
• In the preload library and os.c, _init needs __attribute__((constructor)) to be run by dyld
• PLATFORM_SUPPORTS_SCATTER_GATHER is broken, disabled for now on a64 macOS
• Fixed sigcontext/mcontext conversion which assumes old simd size and was failing on an assert

Known issues that remain:

• VM allocation frequently but nondeterministically fails reachability constraints. -vm_size 500M seems to fix this. i#5383: Fix macOS arm64 test build/run #7171 adds this flag to tests on macOS+ARM64. Another option is to make this the default on macOS+ARM64.
• Sometimes some bookkeeping asserts in vmh_exit are reached upon process exit

The end result is that the bbcount tool works on a simple toy program on macOS, in debug mode, without hitting any asserts:

bin64/drrun -debug -vm_size 500M -c api/bin/libbbcount.dylib -- ./main

Issue #5383

[ndrewh]

It appears I broke the x86 macOS tests, not sure what's going on there yet

[abhinav92003]

It appears I broke the x86 macOS tests, not sure what's going on there yet

If it helps: check out the "Debugging Tests on Github Actions Runner" section at https://dynamorio.org/page_test_suite.html, which documents a way to ssh into the failing test machine.

[ndrewh]

Alright somehow adding IF_MACOS(__attribute__((constructor))) to the _init function in preload.c is breaking the x86 tests. I can't say I understand the situation at all:

• on ARM64 macOS 14.4 the preload doesn't work at all without __attribute__((constructor)), which causes ~all of the client tests to fail
• on the x86 macOS runner (which is on macOS 13):
  • All of the tests seem to pass without the attribute (??? did dyld stop running _init in macOS 14? I searched https://github.com/apple-oss-distributions/dyld and i'm not seeing where they ever ran _init -- maybe some compilers were automatically putting it in the appropriate section but recently changed this behavior?)
  • Once you add the attribute some (but not all!) of the client tests fail with some seemingly random instrumentation assert.

[ndrewh]

I don't think api.startstop fail is my fault, but I also don't see anything to suggest it's #6928

[edeiana]

Add "Issue #5383" at the end of the PR description as suggested in the commit template.

(Thank you for contributing in making DynamoRIO work better on MacOS!)

[edeiana]

Please check if we are in the ARM_NEON64 case here as well (same way you're doing in mcontext_to_sigcontext_simd()).