n_popped, n_pushed is the wrong abstraction for stack effect

[gvanrossum]

The code generator produces metadata for use by the compiler that tells it how many stack entries were pushed and popped by an opcode. These two quantities are subtracted to compute the net stack effect. (The metadata is in the form of functions _PyOpcode_num_popped and _PyOpcode_num_pushed in opcode_metadata.h.)

But what we really need to know, if we want accurate information about what an opcode (or micro-op) does to the stack, is three numbers: how low it goes, how high it goes, and where it ends. For example, if we have a micro-op that pops two entries and then pushes one, and another that pushes two and then pops one, and we combine them into a macro, the macro's net effect is zero, but the lowest point is two below the initial stack level, and the highest point is one above it.

There are various ways to express this, but there is no way to express all this information with just two numbers.

Note that the relevant calculation is already present in the generator, as part of analyze_macro().

[gvanrossum]

There are various ways to express this, but there is no way to express all this information with just two numbers.

I might be wrong here. I'm ending up doing the equivalent of

low = -popped
net = pushed - popped
high = max(0, net)

So two numbers (low an net) would seem to be enough...

[markshannon]

Abstractly, each instruction, or op, pops some number of values, then either fails or pushes some other number of values.
Two numbers are sufficient.

For example, instruction I is composed of op A and op B. Op A pops 2 values, and pushes 1. Op B pops 2 values and pushes 2, the net effect is that I pops 3 values and pushes 2.

You need to model the stack, rather than just adding up values.

Operation	Stack depth
Start	0
A pops	-2
A pushes	-1
B pops	-3
B pushes	-1

[gvanrossum]

The tricky bits happen when the stack effect depends on oparg (e.g. BUILD_SET). The stack effect must then be represented symbolically. I don't know if there are any of those that need to be split up though -- will do an audit later.

[gvanrossum]

Fortunately there are (as yet) no ops with variable-size stack effects. So in #104909 we can use Mark's algorithm.

There is a hypothetical case where the tuple (popped, pushed) isn't enough for the computation of maximum stack level needed: consider I composed of A+B+C, where A pushes 1, B pops 2, and C pushes 1. The net effect is 0, but we need to know that this requires at least one item on the stack at the beginning, and we also need to know that it requires room for an extra item. However, this case doesn't occur in practice.

[carljm]

we also need to know that it requires room for an extra item

Does it? Doesn't the cases generator convert temporaries pushed to the stack by one op and popped by the next into temporary variables, meaning the A+B+C example would not actually use any additional stack (though it would temporarily consume one item from the stack.)

[gvanrossum]

Let's close this since we don't need it.