GH-113710: Tier 2 optimizer: check the function instead of checking globals.

[markshannon]

Since the globals and builtin fields of function objects are immutable, we can check the function instead of the globals and builtins.

This is:

• Never slower, we never add more checks
• Faster for inlined calls as _INIT_CALL_PY_EXACT_ARGS implicitly checks the function, so no additional checks are needed.
• Faster in many other cases, as we can replace _CHECK_GLOBALS + _CHECK_BUILTINS with _CHECK_FUNCTION.

• Issue: The Tier 2 Optimizer #113710

[gvanrossum]

@markshannon, Is this correct? WhenI run a really simple test with debug output, I get a DEOPT message for _CHECK_FUNCTION for each iteration; before this PR I only got one DEOPT message at the end of the loop.

Code:

import_testinternalcapidefcast(typ, val): returnvaldeftestfunc(n): x=0foriinrange(n): x=cast(int, i) returnxdefmain(): count=20a=testfunc(count) asserta+1==count, (a, count) old_opt=_testinternalcapi.get_optimizer() opt=_testinternalcapi.new_uop_optimizer() _testinternalcapi.set_optimizer(opt) a=testfunc(count) _testinternalcapi.set_optimizer(old_opt) asserta+1==count, (a, count) main()

Before:

Optimized executor (length 27): 0 OPTIMIZED: _START_EXECUTOR (52685, target=-842150451, operand=12a70eb70) 1 OPTIMIZED: _CHECK_VALIDITY (0, target=33, operand=1031faf9a) 2 OPTIMIZED: _SET_IP (0, target=14, operand=1031faf74) 3 OPTIMIZED: _ITER_CHECK_RANGE (19, target=14, operand=0) 4 OPTIMIZED: _GUARD_NOT_EXHAUSTED_RANGE (19, target=14, operand=0) 5 OPTIMIZED: _ITER_NEXT_RANGE (19, target=14, operand=0) 6 OPTIMIZED: _CHECK_VALIDITY (0, target=16, operand=1031faf78) 7 OPTIMIZED: _STORE_FAST_2 (2, target=16, operand=0) 8 OPTIMIZED: _CHECK_GLOBALS (3, target=17, operand=10330f050) 9 OPTIMIZED: _LOAD_CONST_INLINE_WITH_NULL (3, target=17, operand=1032fa150) 10 OPTIMIZED: _CHECK_BUILTINS (4, target=22, operand=10323fa70) 11 OPTIMIZED: _LOAD_CONST_INLINE_BORROW (4, target=22, operand=102e39508) 12 OPTIMIZED: _LOAD_FAST_2 (2, target=27, operand=0) 13 OPTIMIZED: _SET_IP (0, target=28, operand=1031faf90) 14 OPTIMIZED: _CHECK_FUNCTION_EXACT_ARGS (2, target=28, operand=2c9) 15 OPTIMIZED: _CHECK_STACK_SPACE (2, target=28, operand=0) 16 OPTIMIZED: _INIT_CALL_PY_EXACT_ARGS_2 (2, target=28, operand=0) 17 OPTIMIZED: _SAVE_RETURN_OFFSET (4, target=28, operand=0) 18 OPTIMIZED: _PUSH_FRAME (2, target=28, operand=1032fa150) 19 OPTIMIZED: _CHECK_VALIDITY (0, target=0, operand=1033690f8) 20 OPTIMIZED: _RESUME_CHECK (0, target=0, operand=0) 21 OPTIMIZED: _LOAD_FAST_1 (1, target=1, operand=0) 22 OPTIMIZED: _SET_IP (0, target=2, operand=1033690fc) 23 OPTIMIZED: _POP_FRAME (0, target=2, operand=10336fa10) 24 OPTIMIZED: _CHECK_VALIDITY (0, target=32, operand=1031faf98) 25 OPTIMIZED: _STORE_FAST_1 (1, target=32, operand=0) 26 OPTIMIZED: _JUMP_TO_TOP (0, target=0, operand=0) DEOPT: [UOp _GUARD_NOT_EXHAUSTED_RANGE (19, target=14, operand=0) -> FOR_ITER_RANGE] 

After:

Optimized executor (length 26): 0 OPTIMIZED: _START_EXECUTOR (52685, target=-842150451, operand=12670ec50) 1 OPTIMIZED: _CHECK_VALIDITY (0, target=33, operand=100dd2f9a) 2 OPTIMIZED: _SET_IP (0, target=14, operand=100dd2f74) 3 OPTIMIZED: _ITER_CHECK_RANGE (19, target=14, operand=0) 4 OPTIMIZED: _GUARD_NOT_EXHAUSTED_RANGE (19, target=14, operand=0) 5 OPTIMIZED: _ITER_NEXT_RANGE (19, target=14, operand=0) 6 OPTIMIZED: _CHECK_VALIDITY (0, target=16, operand=100dd2f78) 7 OPTIMIZED: _STORE_FAST_2 (2, target=16, operand=0) 8 OPTIMIZED: _CHECK_FUNCTION (3, target=17, operand=100ee7050) 9 OPTIMIZED: _LOAD_CONST_INLINE_WITH_NULL (3, target=17, operand=100ed2210) 10 OPTIMIZED: _LOAD_CONST_INLINE_BORROW (4, target=22, operand=100a11508) 11 OPTIMIZED: _LOAD_FAST_2 (2, target=27, operand=0) 12 OPTIMIZED: _SET_IP (0, target=28, operand=100dd2f90) 13 OPTIMIZED: _CHECK_FUNCTION_EXACT_ARGS (2, target=28, operand=2c9) 14 OPTIMIZED: _CHECK_STACK_SPACE (2, target=28, operand=0) 15 OPTIMIZED: _INIT_CALL_PY_EXACT_ARGS_2 (2, target=28, operand=0) 16 OPTIMIZED: _SAVE_RETURN_OFFSET (4, target=28, operand=0) 17 OPTIMIZED: _PUSH_FRAME (2, target=28, operand=100ed2210) 18 OPTIMIZED: _CHECK_VALIDITY (0, target=0, operand=100f410f8) 19 OPTIMIZED: _RESUME_CHECK (0, target=0, operand=0) 20 OPTIMIZED: _LOAD_FAST_1 (1, target=1, operand=0) 21 OPTIMIZED: _SET_IP (0, target=2, operand=100f410fc) 22 OPTIMIZED: _POP_FRAME (0, target=2, operand=100f47ad0) 23 OPTIMIZED: _CHECK_VALIDITY (0, target=32, operand=100dd2f98) 24 OPTIMIZED: _STORE_FAST_1 (1, target=32, operand=0) 25 OPTIMIZED: _JUMP_TO_TOP (0, target=0, operand=0) DEOPT: [UOp _CHECK_FUNCTION (3, target=17, operand=100ee7050) -> FOR_ITER_RANGE] DEOPT: [UOp _CHECK_FUNCTION (3, target=17, operand=100ee7050) -> FOR_ITER_RANGE] DEOPT: [UOp _CHECK_FUNCTION (3, target=17, operand=100ee7050) -> FOR_ITER_RANGE] DEOPT: [UOp _CHECK_FUNCTION (3, target=17, operand=100ee7050) -> FOR_ITER_RANGE] DEOPT: [UOp _CHECK_FUNCTION (3, target=17, operand=100ee7050) -> FOR_ITER_RANGE] DEOPT: [UOp _CHECK_FUNCTION (3, target=17, operand=100ee7050) -> FOR_ITER_RANGE] DEOPT: [UOp _CHECK_FUNCTION (3, target=17, operand=100ee7050) -> FOR_ITER_RANGE] DEOPT: [UOp _CHECK_FUNCTION (3, target=17, operand=100ee7050) -> FOR_ITER_RANGE] DEOPT: [UOp _CHECK_FUNCTION (3, target=17, operand=100ee7050) -> FOR_ITER_RANGE] DEOPT: [UOp _CHECK_FUNCTION (3, target=17, operand=100ee7050) -> FOR_ITER_RANGE] DEOPT: [UOp _CHECK_FUNCTION (3, target=17, operand=100ee7050) -> FOR_ITER_RANGE] DEOPT: [UOp _CHECK_FUNCTION (3, target=17, operand=100ee7050) -> FOR_ITER_RANGE] DEOPT: [UOp _CHECK_FUNCTION (3, target=17, operand=100ee7050) -> FOR_ITER_RANGE] DEOPT: [UOp _CHECK_FUNCTION (3, target=17, operand=100ee7050) -> FOR_ITER_RANGE] DEOPT: [UOp _CHECK_FUNCTION (3, target=17, operand=100ee7050) -> FOR_ITER_RANGE] DEOPT: [UOp _CHECK_FUNCTION (3, target=17, operand=100ee7050) -> FOR_ITER_RANGE] DEOPT: [UOp _CHECK_FUNCTION (3, target=17, operand=100ee7050) -> FOR_ITER_RANGE] DEOPT: [UOp _CHECK_FUNCTION (3, target=17, operand=100ee7050) -> FOR_ITER_RANGE] DEOPT: [UOp _CHECK_FUNCTION (3, target=17, operand=100ee7050) -> FOR_ITER_RANGE] DEOPT: [UOp _GUARD_NOT_EXHAUSTED_RANGE (19, target=14, operand=0) -> FOR_ITER_RANGE] 

That's the _CHECK_FUNCTION call at index 8 deoptimizing every time.

[brandtbucher]

@markshannon, Is this correct? WhenI run a really simple test with debug output, I get a DEOPT message for _CHECK_FUNCTION for each iteration; before this PR I only got one DEOPT message at the end of the loop.

Yep, the bug is that this PR doesn't update the operand, just the opcode. So it's incorrectly comparing the function to the builtin or globals dicts.

[markshannon]

#116460