The stats for "deferred" instructions (tier 1 specialization) are incorrect.

[markshannon]

From a recent stats run:

Execution counts

LOAD_ATTR 577 million

Specialization stats

LOAD_ATTR

deferred 837 million

The number of deferred LOAD_ATTRs cannot exceed the number of LOAD_ATTRs executed, but the stats say they do.

Linked PRs

• GH-123197: Increment correct stat for CALL_KW #123200
• GH-123197: Only count an instruction as deferred if it hasn't deopted first. #123222

[markshannon]

Here's what seems to be happening:

When a guard fails in a specialized instruction, it jumps to the non-specialized version.
This DEOPT jumps to after the execution count has been modified, so the execution count is correct, but the increment to the deferred occurs later, so is incremented for the deopt path.

However, if the instruction is re-specialized, then the deferred counter is not incremented, so we cannot simply decrement the counter when deopting.
We need check that the opcode is the non-specialized one before incrementing deferred.