[WASM][AOT] Generic methods are 5000x slower than non-generic methods using profiled AOT

[jeromelaban]

Description

Repro: https://github.com/jeromelaban/Wasm.Samples/tree/master/Bug48163/Bug48163

Consider the following code:

	[MethodImpl(MethodImplOptions.NoInlining)]
	static int TestGeneric<T>(T value, int other)
	{
		for (int i = 0; i < innerCount; i++)
		{
			other += i;
		}

		return other;
	}

	[MethodImpl(MethodImplOptions.NoInlining)]
	static int TestNormal(int value, int other)
	{
		for (int i = 0; i < innerCount; i++)
		{
			other += i;
		}

		return other;
	}

The generic method is executing 5000x slower than the normal method (~5s vs. ~1ms), even when the profile recorded it.

When using Full AOT, this same generic method is running at the same speed than the normal method.

Additionally, with a similar benchmark, using Dictionary<,>:

• 1M Dictionary<int,int>.TryGetValue() Mixed: 00:00:00.0385150, FullAOT: 00:00:00.0347
• 1M Dictionary<string,string>.TryGetValue() Mixed: 00:00:00.04405, FullAOT: 00:00:00.0090600

Configuration

bb65067

Regression?

No.

Other information

[ghost]

Tagging subscribers to 'arch-wasm': @lewing
See info in area-owners.md if you want to be subscribed.

Issue Details

---

Description

Repro: https://github.com/jeromelaban/Wasm.Samples/tree/master/Bug48163/Bug48163

Consider the following code:

	[MethodImpl(MethodImplOptions.NoInlining)]
	static int TestGeneric<T>(T value, int other)
	{
		for (int i = 0; i < innerCount; i++)
		{
			other += i;
		}

		return other;
	}

	[MethodImpl(MethodImplOptions.NoInlining)]
	static int TestNormal(int value, int other)
	{
		for (int i = 0; i < innerCount; i++)
		{
			other += i;
		}

		return other;
	}

The generic method is executing 5000x slower than the normal method (~5s vs. ~1ms), even when the profile recorded it.

When using Full AOT, this same generic method is running at the same speed than the normal method.

Configuration

bb65067

Regression?

No.

Other information

Author:	jeromelaban
Assignees:	-
Labels:	arch-wasm, area-Codegen-AOT-mono
Milestone:	-

[jeromelaban]

The reason the generic method is not AOTed comes from this line:

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

Line 12784 in 629ad50

GPtrArray *assemblies = mono_domain_get_assemblies (mono_get_root_domain ());

Where no assemblies are loaded during the profile parsing, making resolutions here fail:

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

Line 12748 in 629ad50

if (!cdata->image->image)

[antonfirsov]

I think this is critical to enable any computation-intensive library code to run with reasonable performance with profiled AOT.

For example people keep trying image processing with ImageSharp in the browser, I think it would be important to make that experience at least not terrible. ImageSharp relies heavily on generics, and I assume there are other libraries which also do so.

[jeromelaban]

After investigating the issue a bit deeper, doing the following loads the assemblies and the profile gets loaded further:

if (acfg->aot_opts.image_files) {
	GList *l;
	MonoAssembly *assembly;
	printf("Loading additional image files\n");

	for (l = acfg->aot_opts.image_files; l; l = l->next) {
		printf("Loading additional assembly: %s\n", (char*)l->data);
		assembly = mono_domain_assembly_open_internal (mono_domain_get (), mono_domain_default_alc (mono_domain_get ()), (char*)l->data);

		if(!strcmp(ass->aname.name, assembly->aname.name)) {
			printf("Skip loading additional main assembly: %s\n", (char*)l->data);
			continue;
		}

		if(assembly){
			printf("Loaded additional assembly: %s\n", (char*)l->data);
			acfg->loaded_assemblies = g_slist_append (acfg->loaded_assemblies, assembly);
		}
		else{
			printf("Not loaded additional assembly: %s\n", (char*)l->data);
		}
	}
}
else {
	printf("No additional image files\n");
}

then changing resolve_profile_data to search in acfg->loaded_assemblies:

GSList *loaded_asm = acfg->loaded_assemblies;
printf ("acfg assemblies: %p\n", loaded_asm);
for (; loaded_asm; loaded_asm = loaded_asm->next) {
	MonoAssembly *ass = loaded_asm->data;

	if (!strcmp (ass->aname.name, idata->name)) {
		idata->image = ass->image;
		break;
	}
}

but even though loading the makes the resolution go further, deduping goes in the way and generic methods get excluded here:

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

Line 4243 in d5ab93c

return;

[lewing]

Just for clarity, does this result apply with no profile?

[jeromelaban]

Without a profile, the generic methods are AOTed properly, and the performance for both generic and non-generic method is similar.

[vargaz]

This should be fixed now. The fix might cause issues to surface in other places.