Currently, we allow micro-ops to push temporary non-object pointers to the stack to be consumed by the following micro-op.
This has a number of disadvantages. We originally allowed this for a number of reasons, most of which are not true any more.

The original advantages:

• Allows more instructions to be handled in the JIT as micro-ops could only have one operand. Micro-ops can now have two operands, so this isn't true any more
• Would allow simpler optimizations be removing guards and leaving actions. In fact, the pair of micro-ops that push then pop are often so tightly coupled that there are best optimized as a single unit.
• It reduces code duplication, by allowing common patterns such as PUSH_FRAME to be re-used. This is still true, so we may still want to allow PUSH_FRAME. But since PUSH_FRAME needs special treatment anyway, there may be a better way to do this.

The disadvantages:

• It complicates the code generator
• Having non-object pointers on the stack is fragile, e.g. Crash with PYTHON_LLTRACE=4 due to presence of PyDictKeysObject on stack #129432

Linked PRs

• GH-130296: Avoid stack transients in four instructions. #130310
• GH-130296: Remove _PyOpcode_max_stack_effect as it is no longer used #131493