Tier 2 optimizations for 3.13

[markshannon]

Feature or enhancement

Prerequisites

• Overhaul Symbol API and add tests: gh-115816: Improve internal symbols API in optimizer #116028
• Re-enable optimizer by default: gh-115859: Re-enable T2 optimizer pass by default #116062
• (Auto-?) generate bottom checks in optimizer_cases.c.h #116088

Optimizations to be added to the tier 2 optimizer for 3.13

• Type/value propagation for TO_BOOL #115685
  • Normal _TO_BOOL specializations and friends.
  • TO_BOOL_ALWAYS_TRUE (only if it's worth it?)
• Convert LOAD_ATTR_MODULE to a constant where the module is already a constant #115651
• Type/constant/value propagation for BINARY_OP #115480
  • gh-115480: Type and constant propagation for int BINARY_OPs #115478
  • gh-115480: Minor fixups in int constant propagation #115507
  • gh-115480: Type / constant propagation for float binary uops #115550
  • gh-115480: Type propagate _BINARY_OP_ADD_UNICODE #115710
• Eliminate constant inputs and output constants for simple operations #115506
  • BINARY_OP Constant propagation for _BINARY_OP #115758
  • TO_BOOL Constant propagation for _TO_BOOL #115759
• Eliminate boolean and None guards for constants #115819
  • _GUARD_IS_TRUE_POP
  • _GUARD_IS_FALSE_POP
  • _GUARD_IS_NONE_POP
  • _GUARD_IS_NOT_NONE_POP
• All remaining type and value propagation.
• Extract type guards for COMPARE_OP_INT/FLOAT/STR
• Type/value propagate through COMPARE_OP #115687
  • Split up into uops
  • Propagate bool value and constants too.
• Any remaining guard elimination not handled above
  • gh-115419: Change default sym to not_null #116562
  • Eliminate/combine redundant stack checks (_CHECK_STACK_SPACE)
• Type propagation does not account for the setting local variables via frame.f_locals in a debugger or similar. #115709
• Tier2 peephole optimization: remove extraneous _CHECK_STACK_SPACE ops #116168
• Tier 2 superuops #116202
• Tier 2: True function inlining without reconstruction #116291

If we have time:

• Specialize CONTAINS_OP #116381
  • Type propagate CONTAINS_OP #116432
• Replace CALL_TUPLE_1 for list and generator arguments
• Only specialize CALL in tier for type. Generate optimal argument handling code in tier 2.

Linked PRs

• gh-115419: Change default sym to not_null #116562
• GH-115419: Tidy up tier 2 optimizer. Merge peephole pass into main pass #117997
• GH-115419: Improve list of escaping functions #118054
• GH-115419: Move setting the instruction pointer to error exit stubs #118088

[mdboom]

This is perhaps so rough as to be meaningless (i.e. uop counts != run time), but in an attempt to help prioritize these tasks, I thought I'd share some aggregate counts from a recent pyperformance run:

Grouping	Uops included in group	Total
TO_BOOL	TO_BOOL_ALWAYS_TRUE, TO_BOOL_BOOL, TO_BOOL_INT, TO_BOOL_LIST, TO_BOOL_NONE, TO_BOOL_STR, _TO_BOOL	1,296,284,467
LOAD_ATTR_MODULE		96,542,580
BINARY_OP	_BINARY_OP_MULTIPLY_INT, _BINARY_OP_ADD_INT, _BINARY_OP_SUBTRACT_INT, _BINARY_OP_ADD_UNICODE, _BINARY_OP, _BINARY_OP_MULTIPLY_FLOAT, _BINARY_OP_ADD_FLOAT, _BINARY_OP_SUBTRACT_FLOAT	5,308,844,150
COMPARE_OP	COMPARE_OP_INT, COMPARE_OP_STR, _COMPARE_OP, COMPARE_OP_FLOAT	2,404,431,292
CONTAINS_OP		1,654,609,694

@markshannon: Let me know if anything else like this would be helpful.

[Fidget-Spinner]

I am assigning myself all binary op related stuff. Since it is relatively straightforward.

[markshannon]

3.13 is nearing release, so nothing more is going in 3.13.

Everything we didn't get done for 3.13 will get into 3.14 (if it hasn't been done already)