Issue #6042:

lnotab-based tracing is very complicated and isn't documented very well.  There
were at least 3 comment blocks purporting to document co_lnotab, and none did a
very good job. This patch unifies them into Objects/lnotab_notes.txt which
tries to completely capture the current state of affairs.

I also discovered that we've attached 2 layers of patches to the basic tracing
scheme. The first layer avoids jumping to instructions that don't start a line,
to avoid problems in if statements and while loops.  The second layer
discovered that jumps backward do need to trace at instructions that don't
start a line, so it added extra lnotab entries for 'while' and 'for' loops, and
added a special case for backward jumps within the same line. I replaced these
patches by just treating forward and backward jumps differently.

2 files changed, 132 insertions, 132 deletions

diff --git a/Objects/codeobject.c b/Objects/codeobject.c
index 55f3fb8..6d6775a 100644
--- a/Objects/codeobject.c
+++ b/Objects/codeobject.c
@@ -507,48 +507,8 @@ PyTypeObject PyCode_Type = {
 	code_new,			/* tp_new */
 };
 
-/* All about c_lnotab.
-
-c_lnotab is an array of unsigned bytes disguised as a Python string.  In -O
-mode, SET_LINENO opcodes aren't generated, and bytecode offsets are mapped
-to source code line #s (when needed for tracebacks) via c_lnotab instead.
-The array is conceptually a 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
-
-The first trick is that these numbers aren't stored, only the increments
-from one row to the next (this doesn't really work, but it's a start):
-
-    0, 1,  6, 1,  44, 5,  300, 300,  11, 1
-
-The second trick is that an unsigned byte can't hold negative values, or
-values larger than 255, so (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 c_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 c_lnotab:
-        addr += addr_incr
-        if addr > A:
-            return lineno
-        lineno += line_incr
-
-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, com_set_lineno 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.
+/* Use co_lnotab to compute the line number from a bytecode index, addrq.  See
+   lnotab_notes.txt for the details of the lnotab representation.
 */
 
 int
@@ -567,85 +527,10 @@ PyCode_Addr2Line(PyCodeObject *co, int addrq)
 	return line;
 }
 
-/* 
-   Check whether the current instruction is at the start of a line.
-
- */
-
-	/* The theory of SET_LINENO-less tracing.
-
-	   In a nutshell, we use the co_lnotab field of the code object
-	   to tell when execution has moved onto a different line.
-
-	   As mentioned above, the basic idea is so set things up so
-	   that
-
-	         *instr_lb <= frame->f_lasti < *instr_ub
-
-	   is true so long as execution does not change lines.
-
-	   This is all fairly simple.  Digging the information out of
-	   co_lnotab takes some work, but is conceptually clear.
-
-	   Somewhat harder to explain is why we don't *always* call the
-	   line trace function when the above test fails.
-
-	   Consider this code:
-
-	   1: def f(a):
-	   2:     if a:
-	   3:        print 1
-	   4:     else:
-	   5:        print 2
-
-	   which compiles to this:
-
-	   2           0 LOAD_FAST                0 (a)
-		       3 JUMP_IF_FALSE            9 (to 15)
-		       6 POP_TOP
-
-	   3           7 LOAD_CONST               1 (1)
-		      10 PRINT_ITEM
-		      11 PRINT_NEWLINE
-		      12 JUMP_FORWARD             6 (to 21)
-		 >>   15 POP_TOP
-
-	   5          16 LOAD_CONST               2 (2)
-		      19 PRINT_ITEM
-		      20 PRINT_NEWLINE
-		 >>   21 LOAD_CONST               0 (None)
-		      24 RETURN_VALUE
-
-	   If 'a' is false, execution will jump to instruction at offset
-	   15 and the co_lnotab will claim that execution has moved to
-	   line 3.  This is at best misleading.  In this case we could
-	   associate the POP_TOP with line 4, but that doesn't make
-	   sense in all cases (I think).
-
-	   What we do is only call the line trace function 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.
-
-	   This also takes care of the situation where 'a' is true.
-	   Execution will jump from instruction offset 12 to offset 21.
-	   Then the co_lnotab would imply that execution has moved to line
-	   5, which is again misleading.
-
-	   Why do we set f_lineno when tracing?  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 and 3, then a "return" type event -- but
-	   you would be shown line 5 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.
-	*/
-
-
-int 
-PyCode_CheckLineNumber(PyCodeObject* co, int lasti, PyAddrPair *bounds)
+/* Update *bounds to describe the first and one-past-the-last instructions in
+   the same line as lasti.  Return the number of that line. */
+int
+_PyCode_CheckLineNumber(PyCodeObject* co, int lasti, PyAddrPair *bounds)
 {
         int size, addr, line;
         unsigned char* p;
@@ -662,11 +547,9 @@ PyCode_CheckLineNumber(PyCodeObject* co, int lasti, PyAddrPair *bounds)
            instr_lb -- if we stored the matching value of p
            somwhere we could skip the first while loop. */
 
-        /* see comments in compile.c for the description of
+        /* See lnotab_notes.txt for the description of
            co_lnotab.  A point to remember: increments to p
-           should come in pairs -- although we don't care about
-           the line increments here, treating them as byte
-           increments gets confusing, to say the least. */
+           come in (addr, line) pairs. */
 
         bounds->ap_lower = 0;
         while (size > 0) {
@@ -679,13 +562,6 @@ PyCode_CheckLineNumber(PyCodeObject* co, int lasti, PyAddrPair *bounds)
                 --size;
         }
 
-        /* If lasti and addr don't match exactly, we don't want to
-           change the lineno slot on the frame or execute a trace
-           function.  Return -1 instead.
-        */
-        if (addr != lasti)
-                line = -1;
-        
         if (size > 0) {
                 while (--size >= 0) {
                         addr += *p++;
diff --git a/Objects/lnotab_notes.txt b/Objects/lnotab_notes.txt
new file mode 100644
index 0000000..d247edd
--- /dev/null
+++ b/Objects/lnotab_notes.txt
@@ -0,0 +1,124 @@
+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.