cpython/InternalDocs/exception_handling.md

Description of exception handling

Python uses a technique known as "zero-cost" exception handling, which minimizes the cost of supporting exceptions. In the common case (where no exception is raised) the cost is reduced to zero (or close to zero). The cost of raising an exception is increased, but not by much.

The following code:

try:
    g(0)
except:
    res = "fail"

compiles into intermediate code like the following:

                  RESUME                   0

     1            SETUP_FINALLY            8 (to L1)

     2            LOAD_NAME                0 (g)
                  PUSH_NULL
                  LOAD_CONST               0 (0)
                  CALL                     1
                  POP_TOP
                  POP_BLOCK

    --   L1:      PUSH_EXC_INFO

     3            POP_TOP

     4            LOAD_CONST               1 ('fail')
                  STORE_NAME               1 (res)

SETUP_FINALLY and POP_BLOCK are pseudo-instructions. This means that they can appear in intermediate code but they are not bytecode instructions. SETUP_FINALLY specifies that henceforth, exceptions are handled by the code at label L1. The POP_BLOCK instruction reverses the effect of the last SETUP instruction, so that the active exception handler reverts to what it was before.

SETUP_FINALLY and POP_BLOCK have no effect when no exceptions are raised. The idea of zero-cost exception handling is to replace these pseudo-instructions by metadata which is stored alongside the bytecode, and which is inspected only when an exception occurs. This metadata is the exception table, and it is stored in the code object's co_exceptiontable field.

When the pseudo-instructions are translated into bytecode, SETUP_FINALLY and POP_BLOCK are removed, and the exception table is constructed, mapping each instruction to the exception handler that covers it, if any. Instructions which are not covered by any exception handler within the same code object's bytecode, do not appear in the exception table at all.

For the code object in our example above, the table has a single entry specifying that all instructions that were between the SETUP_FINALLY and the POP_BLOCK are covered by the exception handler located at label L1.

Handling Exceptions

At runtime, when an exception occurs, the interpreter calls get_exception_handler() in Python/ceval.c to look up the offset of the current instruction in the exception table. If it finds a handler, control flow transfers to it. Otherwise, the exception bubbles up to the caller, and the caller's frame is checked for a handler covering the CALL instruction. This repeats until a handler is found or the topmost frame is reached. If no handler is found, the program terminates. During unwinding, the traceback is constructed as each frame is added to it by PyTraceBack_Here(), which is in Python/traceback.c.

Along with the location of an exception handler, each entry of the exception table also contains the stack depth of the try instruction and a boolean lasti value, which indicates whether the instruction offset of the raising instruction should be pushed to the stack.

Handling an exception, once an exception table entry is found, consists of the following steps:

1. pop values from the stack until it matches the stack depth for the handler.
2. if lasti is true, then push the offset that the exception was raised at.
3. push the exception to the stack.
4. jump to the target offset and resume execution.

Reraising Exceptions and lasti

The purpose of pushing lasti to the stack is for cases where an exception needs to be re-raised, and be associated with the original instruction that raised it. This happens, for example, at the end of a finally block, when any in-flight exception needs to be propagated on. As the frame's instruction pointer now points into the finally block, a RERAISE instruction (with oparg > 0) sets it to the lasti value from the stack.

Format of the exception table

Conceptually, the exception table consists of a sequence of 5-tuples:

    1. `start-offset` (inclusive)
    2. `end-offset` (exclusive)
    3. `target`
    4. `stack-depth`
    5. `push-lasti` (boolean)

All offsets and lengths are in code units, not bytes.

We want the format to be compact, but quickly searchable. For it to be compact, it needs to have variable sized entries so that we can store common (small) offsets compactly, but handle large offsets if needed. For it to be searchable quickly, we need to support binary search giving us log(n) performance in all cases. Binary search typically assumes fixed size entries, but that is not necessary, as long as we can identify the start of an entry.

It is worth noting that the size (end-start) is always smaller than the end, so we encode the entries as: start, size, target, depth, push-lasti.

Also, sizes are limited to 230 as the code length cannot exceed 231 and each code unit takes 2 bytes. It also happens that depth is generally quite small.

So, we need to encode:

    `start` (up to 30 bits)
    `size` (up to 30 bits)
    `target` (up to 30 bits)
    `depth` (up to ~8 bits)
    `lasti` (1 bit)

We need a marker for the start of the entry, so the first byte of entry will have the most significant bit set. Since the most significant bit is reserved for marking the start of an entry, we have 7 bits per byte to encode offsets. Encoding uses a standard varint encoding, but with only 7 bits instead of the usual 8. The 8 bits of a byte are (msb left) SXdddddd where S is the start bit. X is the extend bit meaning that the next byte is required to extend the offset.

In addition, we combine depth and lasti into a single value, ((depth<<1)+lasti), before encoding.

For example, the exception entry:

    `start`:  20
    `end`:    28
    `target`: 100
    `depth`:  3
    `lasti`:  False

is encoded by first converting to the more compact four value form:

    `start`:         20
    `size`:          8
    `target`:        100
  `depth<<1+lasti`:  6

which is then encoded as:

    148 (MSB + 20 for start)
    8   (size)
    65  (Extend bit + 1)
    36  (Remainder of target, 100 == (1<<6)+36)
    6

for a total of five bytes.

The code to construct the exception table is in assemble_exception_table() in Python/assemble.c.

The interpreter's function to lookup the table by instruction offset is get_exception_handler() in Python/ceval.c. The Python function _parse_exception_table() in Lib/dis.py returns the exception table content as a list of namedtuple instances.