[mono][aot] Remove method body for inlined methods

[kotlarmilos]

When a method is inlined, its implicit reference still exists in a method_addresses table and native linker can't remove its body. For example, if _HelloWorld_MyGen_Bar is inlined and not called indirectly, then its method body is not needed during the runtime.

_mono_aot_HelloWorldmethod_addresses:
0000000000003938    bl    _HelloWorld_Program_Foo_string__
000000000000393c    bl    _HelloWorld_Program__ctor
0000000000003940    bl    _HelloWorld_MyGen_Bar
...
_HelloWorld_MyGen_Bar:
0000000000003970    adrp    x16, 5 ; 0x8000
0000000000003974    add    x16, x16, #0x2a8
0000000000003978    ldr    x16, [x16, #0xb8]
...

/cc: @ivanpovazan

[ghost]

Tagging subscribers to 'size-reduction': @eerhardt, @SamMonoRT, @marek-safar
See info in area-owners.md if you want to be subscribed.

Issue Details

---

When a method is inlined, its implicit reference still exists in a method_addresses table and native linker can't remove its body. For example, if _HelloWorld_MyGen_Bar is inlined and not called indirectly, then its method body is not needed during the runtime.

_mono_aot_HelloWorldmethod_addresses:
0000000000003938    bl    _HelloWorld_Program_Foo_string__
000000000000393c    bl    _HelloWorld_Program__ctor
0000000000003940    bl    _HelloWorld_MyGen_Bar
...
_HelloWorld_MyGen_Bar:
0000000000003970    adrp    x16, 5 ; 0x8000
0000000000003974    add    x16, x16, #0x2a8
0000000000003978    ldr    x16, [x16, #0xb8]
...

/cc: @ivanpovazan

Author:	kotlarmilos
Assignees:	kotlarmilos
Labels:	area-Codegen-AOT-mono, size-reduction
Milestone:	8.0.0

[kotlarmilos]

@ivanpovazan According to your experiment, does the following code snipped in emit_code method populate method_addresses table?

runtime/src/mono/mono/mini/aot-compiler.c

Lines 10766 to 10782 in e57bb02

	for (guint32 i = 0; i < acfg->nmethods; ++i) {
	#ifdef MONO_ARCH_AOT_SUPPORTED
	if (acfg->flags & MONO_AOT_FILE_FLAG_CODE_EXEC_ONLY) {
	if (acfg->cfgs [i])
	emit_pointer (acfg, acfg->cfgs [i]->asm_symbol);
	else
	emit_pointer (acfg, NULL);
	} else {
	if (acfg->cfgs [i])
	arch_emit_label_address (acfg, acfg->cfgs [i]->asm_symbol, FALSE, acfg->thumb_mixed && acfg->cfgs [i]->compile_llvm, NULL, &acfg->call_table_entry_size);
	else
	arch_emit_label_address (acfg, symbol, FALSE, FALSE, NULL, &acfg->call_table_entry_size);
	}
	#endif
	}
	sprintf (symbol, "method_addresses_end");

[ivanpovazan]

@kotlarmilos, yes here is the change from the experiment:

https://github.com/ivanpovazan/runtime/blob/38eabec1a6320287f754a2e05b19787664c55df5/src/mono/mono/mini/aot-compiler.c#L10782-L10787

[kotlarmilos]

To enable the optimization, the AOT compiler should check if a method isn't called through reflection in can_specialize (as done in the experiment) and if all call sites inline the method, as inline_info property indicates that it could be done, but it might or might not happen.

can_specialize depends on the linker adding annotations to methods, which needs to be implemented/fixed. @ivanpovazan Do you perhaps know what is the status of --explicit-reflection flag in the linker?

[ivanpovazan]

can_specialize depends on the linker adding annotations to methods, which needs to be implemented/fixed. @ivanpovazan Do you perhaps know what is the status of --explicit-reflection flag in the linker?

I have provided more details in the Prerequisites section in my comment. The requirement 2. is necessary to enable this optimization as well, so it can be treated as a common prerequisite. Afaik --explicit-reflection is a legacy feature which is currently unused.

[kotlarmilos]

Thanks! As you outlined, the idea is to estimate potential improvements (size savings and better performance) before proceeding with the concrete implementation.

[kotlarmilos]

Obsolete, lower priority compared to other tasks.