gh-91276: Make JUMP_IF_TRUE_OR_POP/JUMP_IF_FALSE_OR_POP relative

[iritkatriel]

https://bugs.python.org/issue47120

[markshannon]

Are you completely sure it is impossible for the compiler to generate a backwards branch for these instructions?

I can't produce a case now, but I can create a case that could jump backwards with the right compiler change.

def f(a,b):
    while x:= a or b:
        pass

compiles to:

  1           0 RESUME                   0

  2           2 LOAD_FAST                0 (a)
              4 JUMP_IF_TRUE_OR_POP      4 (to 8)
              6 LOAD_FAST                1 (b)
        >>    8 COPY                     1
             10 STORE_FAST               2 (x)
             12 POP_JUMP_IF_FALSE       16 (to 32)

  3     >>   14 NOP

  2          16 LOAD_FAST                0 (a)
             18 JUMP_IF_TRUE_OR_POP     11 (to 22)
             20 LOAD_FAST                1 (b)
        >>   22 COPY                     1
             24 STORE_FAST               2 (x)
             26 POP_JUMP_IF_TRUE         7 (to 14)
             28 LOAD_CONST               0 (None)
             30 RETURN_VALUE
        >>   32 LOAD_CONST               0 (None)
             34 RETURN_VALUE

The compiler could, in theory, retarget the branch at instruction 18 so that it jumps to 8.
If it did, then there would be a JUMP_IF_TRUE_OR_POP jumping backwards.

[markshannon]

We could handle this in the assembler, I guess.

JUMP_IF_TRUE_OR_POP back

Would become

COPY 1
POP_JUMP_IF_FALSE back
POP_TOP

As long as backward jumps are rare (and they will be), then this wouldn't add any real overhead.

[iritkatriel]

Are you completely sure it is impossible for the compiler to generate a backwards branch for these instructions?

I put an assertion and all tests passed. But yeah, we can make it more rubust as you suggest.

[iritkatriel]

We could handle this in the assembler, I guess.

JUMP_IF_TRUE_OR_POP back

Would become

COPY 1
POP_JUMP_IF_FALSE back
POP_TOP

As long as backward jumps are rare (and they will be), then this wouldn't add any real overhead.

There is the problem of how to test this, since we don't know how to currently make the compiler create a backwards jump with this opcode.

[ghost]

Commit authors are required to sign the Contributor License Agreement.

[iritkatriel]

Closing and reopening to kick the bots.

[bedevere-bot]

🤖 New build scheduled with the buildbot fleet by @iritkatriel for commit ec65324 🤖

If you want to schedule another build, you need to add the ":hammer: test-with-buildbots" label again.

[bedevere-bot]

🤖 New build scheduled with the buildbot fleet by @iritkatriel for commit 2dfe713 🤖

If you want to schedule another build, you need to add the ":hammer: test-with-buildbots" label again.

[iritkatriel]

Buildbots are happy.