Comptime IDE crash from Comptime-generated Constant

[hunterbridges]

We have started getting frequent IDE crashes in our codebase, which are still occurring on latest nightly.

I ran IDE in Debug, and it appears to be hitting an assert in BeefIRBuilder.cpp:2601

BF_ASSERT(irValue.mId >= 0);

Inspecting surrounding state, it appears to be hitting the assert when trying to generate IR for the comptime-generated constant Hero.INSTANCE_SIZE, which is emitted as such (in the Hero class type):

const uint32 RSDK_INSTANCE_SIZE = Memory.AlignPOW2((.)typeof(Self).InstanceStride);

For context, Memory.AlignPOW2 is a ConstEval function defined as:

	[Comptime(ConstEval = true)]
	public static uint32 AlignPOW2(uint32 size)
	{
		uint32 v = size;

		v--;
		v |= v >> 1;
		v |= v >> 2;
		v |= v >> 4;
		v |= v >> 8;
		v |= v >> 16;
		v++;

		return v;
	}

For additional context, Hero is a class that has a fair bit of extensions.

[hunterbridges]

I think this may specifically happen when editing an extension that contains additional members. Maybe when compiler runs while you're typing, sometimes the extension errors and modifies the instance stride? Just a wild guess

[ESFraser]

Adding onto this, I did a bit more investigation in an effort to better understand what's going on here.. and it seems this issue also extends to enum values too, not just explicitly declared const values! I don't think the enum values I got crashes on were comtime-generated, but the crash did seem like it was only occurring after the compiler had tried to evaluate some comptime code?

[bfiete]

It's hard to guess what's happening with no repro program...

Maybe at least show a full stack trace or crash dump and I can guess a little better?

[bfiete]

To provide a little more info- this assertion indicates that an empty value was attempting to be written to the IR stream managed by BfIRBuilder (the stream of commands which drive the compiler backend).

Empty values are not uncommon when the front end encounters errors, but the front end must ensure that it avoids writing any IR commands that include those empty values in those cases.

It appears that there is some kind of missing error handling in this case somewhere that would have protected against that.