reopen: stack size computation is broken #3

[peterazmanov]

I encountered the 'closed' issue reported by @pwaller with Python 3.4. Stack size computation is broken (method '_compute_stacksize') and returns zero for every block of code I tried. As I mentioned in comment of original issue I tried to fix it (peterazmanov/byteplay). My changes stopped crashes, but the code requires further testing.

[peterazmanov]

From 2cc5ce4 there shouldn't be crashes or debug breaks associated with the size of stack.

P.S. Still in some cases '_compute_stacksize' returns more than needed (see issue in my fork for details)

[pwaller]

xref #3

[peterazmanov]

I believe current version for Python 2 (byteplay2.py) still containes broken stack size calculation... (original issue was reported for Python 2 version and nothing has been fixed in it) I propose to substitute byteplay2.py with base version by Noam Yorav-Raphael

[serprex]

An important modification I made to the stack calculation is that we take the maximum stack usage if two branches with different stack usage combine & then return without looping

[peterazmanov]

I don't quite understand your improvement. The original version doesn't contain looping, it calculates stack depth for every possible scope (including loop scopes).

Besides, after your changes stack depth calculation returns 0, at least sometimes (confirmed by @pwaller with Python 2, and by me with Python 3.4 for other test case).

[serprex]

I mean the bytecode looping. ie

JUMP_IF_FALSE L1
PUSH 1
L1:
RET

Should be allowed to compile

[peterazmanov]

What do you mean by "bytecode looping"? There is no loops in your code only two possible execution branches with one label. To determine stack depth correctly one has to visit both execution branches and return maximum stack effect.

Also the code that you mention isn't "regular" (we can reach L1 label with two different stack
depths), so CPython won't generate such code...

[serprex]

CPython won't but a programmer may. Byteplay also supports creating constants that are mutable

Yes, there are no loops. The stack depth calculation should only throw an error if one creates a loop which makes the possible stack depth unbounded