All about co_lnotab, the line number table.

Code objects store a field named co_lnotab. This is an array of unsigned bytes

disguised as a Python string. It is used to map bytecode offsets to source code

line #s for tracebacks and to identify line number boundaries for line tracing.

The array is conceptually a compressed list of

(bytecode offset increment, line number increment)

pairs. The details are important and delicate, best illustrated by example:

byte code offset source code line number

0 1

6 2

50 7

350 307

361 308

Instead of storing these numbers literally, we compress the list by storing only

the increments from one row to the next. Conceptually, the stored list might

look like:

0, 1, 6, 1, 44, 5, 300, 300, 11, 1

The above doesn't really work, but it's a start. Note that an unsigned byte

can't hold negative values, or values larger than 255, and the above example

contains two such values. So we make two tweaks:

(a) there's a deep assumption that byte code offsets and their corresponding

line #s both increase monotonically, and

(b) if at least one column jumps by more than 255 from one row to the next,

more than one pair is written to the table. In case #b, there's no way to know

from looking at the table later how many were written. That's the delicate

part. A user of co_lnotab desiring to find the source line number

corresponding to a bytecode address A should do something like this

lineno = addr = 0

for addr_incr, line_incr in co_lnotab:

addr += addr_incr

if addr > A:

return lineno

lineno += line_incr

(In C, this is implemented by PyCode_Addr2Line().) In order for this to work,

when the addr field increments by more than 255, the line # increment in each

pair generated must be 0 until the remaining addr increment is < 256. So, in

the example above, assemble_lnotab in compile.c should not (as was actually done

until 2.2) expand 300, 300 to

255, 255, 45, 45,

but to

255, 0, 45, 255, 0, 45.

The above is sufficient to reconstruct line numbers for tracebacks, but not for

line tracing. Tracing is handled by PyCode_CheckLineNumber() in codeobject.c

and maybe_call_line_trace() in ceval.c.

*** Tracing ***

To a first approximation, we want to call the tracing function when the line

number of the current instruction changes. Re-computing the current line for

every instruction is a little slow, though, so each time we compute the line

number we save the bytecode indices where it's valid:

*instr_lb <= frame->f_lasti < *instr_ub

is true so long as execution does not change lines. That is, *instr_lb holds

the first bytecode index of the current line, and *instr_ub holds the first

bytecode index of the next line. As long as the above expression is true,

maybe_call_line_trace() does not need to call PyCode_CheckLineNumber(). Note

that the same line may appear multiple times in the lnotab, either because the

bytecode jumped more than 255 indices between line number changes or because

the compiler inserted the same line twice. Even in that case, *instr_ub holds

the first index of the next line.

However, we don't *always* want to call the line trace function when the above

test fails.

Consider this code:

1: def f(a):

2: while a:

3: print 1,

4: break

5: else:

6: print 2,

which compiles to this:

2 0 SETUP_LOOP 19 (to 22)

>> 3 LOAD_FAST 0 (a)

6 POP_JUMP_IF_FALSE 17

3 9 LOAD_CONST 1 (1)

12 PRINT_ITEM

4 13 BREAK_LOOP

14 JUMP_ABSOLUTE 3

>> 17 POP_BLOCK

6 18 LOAD_CONST 2 (2)

21 PRINT_ITEM

>> 22 LOAD_CONST 0 (None)

25 RETURN_VALUE

If 'a' is false, execution will jump to the POP_BLOCK instruction at offset 17

and the co_lnotab will claim that execution has moved to line 4, which is wrong.

In this case, we could instead associate the POP_BLOCK with line 5, but that

would break jumps around loops without else clauses.

We fix this by only calling the line trace function for a forward jump if the

co_lnotab indicates we have jumped to the *start* of a line, i.e. if the current

instruction offset matches the offset given for the start of a line by the

co_lnotab. For backward jumps, however, we always call the line trace function,

which lets a debugger stop on every evaluation of a loop guard (which usually

won't be the first opcode in a line).

Why do we set f_lineno when tracing, and only just before calling the trace

function? Well, consider the code above when 'a' is true. If stepping through

this with 'n' in pdb, you would stop at line 1 with a "call" type event, then

line events on lines 2, 3, and 4, then a "return" type event -- but because the

code for the return actually falls in the range of the "line 6" opcodes, you

would be shown line 6 during this event. This is a change from the behaviour in

2.2 and before, and I've found it confusing in practice. By setting and using

f_lineno when tracing, one can report a line number different from that

suggested by f_lasti on this one occasion where it's desirable.