GH-112354: END_FOR instruction to only pop one value.

[markshannon]

In order to support tier 2 side-exits from FOR_ITER variants, we need a target that has the same stack as the exhaustion guard.
Currently we can either jump to the END_FOR, as we do for generators or the instruction afterwards, which we do for FOR_ITER.
In order to exit a guard we need a target in between those instructions, that pops the iterator, but not the value.
The compiler emits END_FOR; POP_TOP instead of END_FOR.
Previously:

   END_FOR (Pops two values) <- FOR_ITEN_GEN targets this
   ...                       <- FOR_ITER targets this

Now:

   END_FOR (Pops one value)  <- FOR_ITEN_GEN targets this
   POP_TOP (Pops one value)  <- _GUARD_NOT_EXHAUSTED can target this
   ...                       <- FOR_ITER targets this

• Issue: Trace stitching #112354

[gvanrossum]

Maybe we can do without a separate END_FOR opcode and instead just generate a POP_TOP?

EDIT: Never mind, it's needed for instrumentation.

[gvanrossum]

Looks good, never mind my premature commend about END_FOR/POP_TOP.

[gvanrossum]

(The updates to this file (and a few other generated files) are spurious and will disappear after merging the latest main.)

[markshannon]

The benchmarks show no performance impact.

[markshannon]

EDIT: Never mind, it's needed for instrumentation.

That's evidently not clear. The order of the POP_TOP and END_FOR is important as well.
I'll add some comments.

[gvanrossum]

Thanks!

[gvanrossum]

Oh, fun stuff. Windows compilation error. :-(