JIT: Reconsider how stress modes are randomly applied

[jakobbotsch]

The hash used to compute whether a stress mode is enabled works simply by xor'ing the integer value of the stress mode in the stress enum:

runtime/src/coreclr/jit/compiler.cpp

Lines 3581 to 3586 in 93dda3b

	// Get a hash which can be compared with 'weight'
	assert(stressArea != 0);
	const unsigned hash = (info.compMethodHash() ^ stressArea ^ stressLevel) % MAX_STRESS_WEIGHT;
	assert(hash < MAX_STRESS_WEIGHT && weight <= MAX_STRESS_WEIGHT);
	return (hash < weight);

That seems like it would have a very bad distribution -- there's only a few dozens of stress modes so this at most affects only the few lowest bits of the hash code, in a very predictable way.

Additionally, reordering members in the enum leads to an entirely different (unrelated) set of stress modes being enabled/disabled. This for example happened in #79283.

I think it would be better if we derived a hash from the string name of each stress mode and then xor'd with that instead (e.g. using HashStringA). That would be more similar to how other randomness is mixed in as well.

[ghost]

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch, @kunalspathak
See info in area-owners.md if you want to be subscribed.

Issue Details

---

The hash used to compute whether a stress mode is enabled works simply by xor'ing the integer value of the stress mode in the stress enum:

runtime/src/coreclr/jit/compiler.cpp

Lines 3581 to 3586 in 93dda3b

	// Get a hash which can be compared with 'weight'
	assert(stressArea != 0);
	const unsigned hash = (info.compMethodHash() ^ stressArea ^ stressLevel) % MAX_STRESS_WEIGHT;
	assert(hash < MAX_STRESS_WEIGHT && weight <= MAX_STRESS_WEIGHT);
	return (hash < weight);

That seems like it would have a very bad distribution -- there's only a few dozens of stress modes so this at most affects only the few lowest bits of the hash code.

Additionally, reordering members in the enum leads to an entirely different (unrelated) set of stress modes being enabled/disabled. This for example happened in #79283.

I think it would be better if we derived a hash from the string name of each stress mode and then xor'd with that instead (e.g. using HashStringA). That would be more similar to how other randomness is mixed in as well.

Author:	jakobbotsch
Assignees:	-
Labels:	area-CodeGen-coreclr, untriaged
Milestone:	-