Repository navigation
Jump-To-Jump elimination could be more aggressive. #93223
Description
Example:
def f():
for i in range(5):
if i > 3:
print(i) 1 0 RESUME 0
2 2 LOAD_GLOBAL 1 (NULL + range)
14 LOAD_CONST 1 (5)
16 CALL 1
26 GET_ITER
>> 28 FOR_ITER 21 (to 72)
30 STORE_FAST 0 (i)
3 32 LOAD_FAST 0 (i)
34 LOAD_CONST 2 (3)
36 COMPARE_OP 4 (>)
42 POP_JUMP_FORWARD_IF_FALSE 13 (to 70) <--------- could be BACKWARD to 28
4 44 LOAD_GLOBAL 3 (NULL + print)
56 LOAD_FAST 0 (i)
58 CALL 1
68 POP_TOP
>> 70 JUMP_BACKWARD 22 (to 28)
2 >> 72 LOAD_CONST 0 (None)
74 RETURN_VALUE
Something like this should suffice:
static bool
jump_thread(struct instr *inst, struct instr *target, int opcode)
{
assert(!IS_VIRTUAL_OPCODE(opcode) || IS_VIRTUAL_JUMP_OPCODE(opcode));
assert(is_jump(inst));
assert(is_jump(target));
// bpo-45773: If inst->i_target == target->i_target, then nothing actually
// changes (and we fall into an infinite loop):
- if (inst->i_lineno == target->i_lineno &&
+ if ((inst->i_lineno == target->i_lineno || target->i_lineno == -1) &&
inst->i_target != target->i_target)
{
inst->i_target = target->i_target;
inst->i_opcode = opcode;
return true;
}
return false;
}
Activity
brandtbucher commentedon May 25, 2022
I tried something like this a few months ago, but (if I remember correctly) it ended up duplicating a few blocks in
test_dis.py. The net effect was more code, not less, and it ended up actually having the same number of jumps in the final output for all code paths (because some paths had to jump over the new blocks now, I think).I abandoned that experiment because I couldn’t determine the cause of the duplication. I wonder if that’s still the case (or if @iritkatriel knows what the problem might be).
brandtbucher commentedon May 25, 2022
(At any rate, I think my experience highlights just how opaque and complex the interactions between the different compiler passes are.)
sweeneyde commentedon May 25, 2022
Maybe something has changed between then and now, but that change now makes
jumpyone instruction shorter and eliminates two jump-to-jumps.#93229
sweeneyde commentedon May 26, 2022
These are apparently the sites in pyperformance that this applies to:
Details
As a more concrete example, this is the difference in bm_hexiom dis outputs before and after the change. 28 occurrences of
>>\s+\d+\s+JUMPgoes down to just 6, and all of these remaining 6 areFOR_ITER --> JUMP_BACKWARD, which we can't fold for now.gh-93223: More aggressive Jump-To-Jump elimination (GH-93229)