In early July 2013, I made some initial steps toward this goal as a proof-of-concept and am fairly encouraged by

my findings so far.

"2.crazy" as the Python version. (And see the accompanying [Py2crazy](https://github.com/pgbovine/Py2crazy/) project for more details.)

Note that the (approximate) column of the currently-executing bytecode instruction is highlighed in yellow.

To support this finer-grained intra-line stepping, I hacked CPython 2.7.5 to:

1. call the trace function at each bytecode rather than each time a new line is executed, which turns it into a bytecode-level stepper

2. disable the peephole optimizer so that bytecodes match source code more closely

3. insert column number information along with line numbers in bytecode

The trickest part was getting column number information into the code object.

It turns out that the tokenizer, parser, and AST all keep column numbers,

but only line numbers show up in bytecode (encoded in the cryptic lnotab string).

After some hair-pulling, I managed to add a new code.co_coltab dict that maps each

bytecode instruction to a (line number, column number) pair.

Here's example disassembly from the following line (line 7 in source code) with line/column number mappings,

and the corresponding point in the source code.

Source code line 7:

"if foo() and (x + y > 7):"

(7, 3) 21 LOAD_NAME 0 (foo) if foo() and (x + y > 7):

^

(7, 3) 24 CALL_FUNCTION 0 if foo() and (x + y > 7):

^

(7, 3) 27 JUMP_IF_FALSE_OR_POP 43 if foo() and (x + y > 7):

^

(7, 14) 30 LOAD_NAME 1 (x) if foo() and (x + y > 7):

^

(7, 18) 33 LOAD_NAME 2 (y) if foo() and (x + y > 7):

^

(7, 16) 36 BINARY_ADD if foo() and (x + y > 7):

^

(7, 22) 37 LOAD_CONST 3 (7) if foo() and (x + y > 7):

^

(7, 20) 40 COMPARE_OP 4 (>) if foo() and (x + y > 7):

^

(7, 20) >> 43 POP_JUMP_IF_FALSE 54 if foo() and (x + y > 7):

^

With this extended bytecode format, I can extend Online Python Tutor to point

not only to the line being executed, but also to the column, which gives finer-grained tracing for instructors.

The column number precision isn't perfect for all kinds of expressions, but it's still better than having

no column numbers. For instance, it's not precise for chained comparisons (`x < y < z < w`),

field accesses (`foo.bar.baz`), and probably subscripting.