Enable AOT usage on M1 mac

[eloparco]

After reading the docs, I managed to use AOT on my M1 mac.

Issue: #1365

[cosmo0920]

Thanks for your work. I confirmed that it works!!

% sw_vers
ProductName:		macOS
ProductVersion:		14.0
BuildVersion:		23A344

% git fetch upstream pull/2618/head:pr-2618
% git checkout pr-2618

Build wamrc and basic samples with the PR patch.

% ../../wamr-compiler/wamrc --size-level=3 -o ./out/wasm-apps/testapp.aot ./out/wasm-apps/testapp.wasm
Create AoT compiler with:
  target:        arm64
  target cpu:    apple-m1
  target triple: arm64-pc-linux-gnu
  cpu features:  
  opt level:     3
  size level:    3
  output format: AoT file
Compile success, file ./out/wasm-apps/testapp.aot was generated.

% file ./out/wasm-apps/testapp.aot
./out/wasm-apps/testapp.aot: data

% ./out/basic -f out/wasm-apps/testapp.aot
calling into WASM function: generate_float
Native finished calling wasm function generate_float(), returned a float value: 102009.921875f
calling into WASM function: float_to_string
calling into native function: intToStr
calling into native function: get_pow
calling into native function: intToStr
Native finished calling wasm function: float_to_string, returned a formatted string: 102009.921
calling into WASM function: calculate
calling into native function: calculate_native, n=3, func1=1, func2=2
calling into WASM function: mul5,    mul5 return 15 
call func1 and return n1=15
calling into WASM function: mul7,    mul7 return 105 
call func2 and return n2=105
Native finished calling wasm function: calculate, return: 120

[yamt]

After reading the docs, I managed to use AOT on my M1 mac.

Issue: #1365

it seems inefficient to call pthread_jit_write_protect_np before every calls.
isn't it enough to enable write temporary, only for text relocation?

also, the doc you mentioned seems suggesting to invalidate icache. isn't it necessary?

[eloparco]

Thanks for your work.

Thanks for starting the investigation, you input was was very useful!

it seems inefficient to call pthread_jit_write_protect_np before every calls.

Moved that to call it only before and after module loading

isn't it enough to enable write temporary, only for text relocation?

Yes, I put it before and after load so that we don't need to call it in multiple places inside load

also, the doc you mentioned seems suggesting to invalidate icache. isn't it necessary?

Added that now

[wenyongh]

We usually don't directly call the libc functions in core/iwasm, instead, we encapsulated them in the platform layer and call them. How about adding API in core/shared/platform, e.g. void os_thread_jit_write_protect_np(bool enabled), in:

https://github.com/bytecodealliance/wasm-micro-runtime/blob/main/core/shared/platform/include/platform_api_vmcore.h#L83

and:
https://github.com/bytecodealliance/wasm-micro-runtime/blob/main/core/shared/platform/common/posix/posix_thread.c#L420

and call it here.

and how about wrappring sys_icache_invalidate to os_icache_flush? like os_dcache_flush:

https://github.com/bytecodealliance/wasm-micro-runtime/blob/main/core/shared/platform/include/platform_api_vmcore.h#L144

https://github.com/bytecodealliance/wasm-micro-runtime/blob/main/core/shared/platform/common/posix/posix_memmap.c#L253C1-L253C1

And had better set memory executable after ret is true. Reference code:

#if defined(__APPLE__) && defined(__arm64__)
    os_thread_jit_write_protect_np(false); /* Make memory writable */
#endif

    if (!load(buf, size, module, error_buf, error_buf_size)) {
        aot_unload(module);
        return NULL;
    }

#if defined(__APPLE__) && defined(__arm64__)
    if (module->code) {
        os_thread_jit_write_protect_np(true); /* Make memory executable */
        os_icache_flush(module->code, module->code_size);
    }
#endif

[eloparco]

Thanks, I should have known.
The ifdef should be inside the platform-specific code I think, and no need for if (module->code) since load would have returned false if that was NULL.

os_thread_jit_write_protect_np(false); /* Make memory writable */
if (!load(buf, size, module, error_buf, error_buf_size)) {
    aot_unload(module);
    return NULL;
}
os_thread_jit_write_protect_np(true); /* Make memory executable */
os_icache_flush(module->code, module->code_size);

[wenyongh]

OK, module->code may be NULL if the module hasn't function bytecodes, but I think the pthread_jit_write_protect_np should have handle that case.

[wenyongh]

LGTM