Optimize collection literals with leading null unpack idiom in bytecode, allows bare `BUILD_SET 0`

Feature or enhancement

Proposal:

Performance was discussed at various times within PEP 802 topic, submitting this proposal as the PEP is unlikely to be accepted. This is intended as an advanced escape-hatch not overlapping with the main reasons for PEP 802.

Goal:

• Remove performance penalty of ast.unparse(ast.Set(elts=[])) ({*()} can be used in advanced cases and will not change recommended way to write empty set set())
• Remove performance penalty of incremental construction between 0-to-1/1-to-0 elements for set and tuple literals without special case considerations (e.g. x = *(), pre-seeded tuple literal without trailing comma pitfall)
• Keep implementation simple/self-contained in bytecode generation, avoiding any AST changes to language
• Near-zero performance impact to bytecode generation (improves subsequent bytecode optimizations of target case)

Implementation examples:

• During first AST -> bytecode pass: https://github.com/python/cpython/commit/67742d92c411a93e7692f761ffeb3af75ba1ff87#diff-6d58b0ddc066ad12ebc378b62c4189335bd57c83aece11072c76fd86a78a1a4a
• During bytecode-only optimization step: https://github.com/python/cpython/commit/f857d88ef3aac90deadf438625211040e73df832#diff-3cbf15668c31488528b7ab0f903c674a0ecf550f4f53c1be6cf8cd965246c2a0

Micro-benchmark

Run on windows x64bit release build: empty_unpack_benchmark.py

Compile time

Targeted cases:

Benchmark	Main	Flowgraph only	Codegen leading only
[*()]	1.450 ms / 1.000x	1.393 ms / 0.961x	1.378 ms / 0.950x
{*()}	1.530 ms / 1.000x	1.488 ms / 0.973x	1.467 ms / 0.959x
(*(),)	1.916 ms / 1.000x	1.842 ms / 0.962x	1.832 ms / 0.956x
x = *(),	1.139 ms / 1.000x	1.088 ms / 0.955x	1.055 ms / 0.927x

xychart-beta
	title "Compile targeted mean ratio vs main"
	x-axis [Main, Flowgraph, Leading]
	y-axis "Ratio" 0 --> 1.01
	bar [1.000, 0.963, 0.948]

Control near-miss case comparisons moved less than +/-1% and are attributed to noise.

Runtime

The set() versus {*()} comparison is especially useful because it changes qualitatively across the two branches (~6.3% slower to ~29.5% faster):

Branch	set() mean (ns)	{*()} mean (ns)	{*()} vs set()
Main	54.14	57.53	1.063x
Flowgraph only	53.49	38.01	0.711x
Codegen leading only	55.44	39.00	0.704x

Before this change, {*()} was paying for a redundant empty update and ended up slower than the constructor call. After the change, {*()} compiles down to a direct BUILD_SET 0; RETURN_VALUE path, while set() still has to load the global and perform a zero-argument call.

Has this already been discussed elsewhere?

I have already discussed this feature proposal on Discourse

Links to previous discussion of this feature:

https://discuss.python.org/t/pep-802-display-syntax-for-the-empty-set/101676/237
https://discuss.python.org/t/pep-802-display-syntax-for-the-empty-set/101676/216
https://discuss.python.org/t/pep-802-display-syntax-for-the-empty-set/101676/3

Linked PRs

• gh-150812

首先查看 empty_unpack_benchmark.py 以及提案中链接的实现示例，然后检查链接的 PR gh-150812 和相关的 Discourse 讨论。当所提议的集合字面量用例能够避免冗余的空更新，同时保留文档所述行为和基准测试改进时，即视为完成。