OpenQASM parser optimization

[idavis]

OpenQASM parser optimization: before vs after

Input sizes used: 1 MiB (63,001 statements) and 5 MiB (302,908 statements).

• Peak memory: −48% (≈ halved)
• Retained memory: −59% (semantic-lowering's retained set alone is down −77%)
• Time: −30%
• Allocation count (using dhat): −50% at analyze — the root driver of both the time and peak wins

What changed

Change	Targets
Remove the whole-source clone() in lower() (walk via mem::take)	transient peak
List<T>: Box<[Box<T>]> → Box<[T]> (drop per-element box)	allocation count + time
Drop the retained syntax AST after lowering (program: Option<Program>)	retained/net-live
Box inline Expr fields in big ExprKind variants (264 → 96 B)	both
Box inline Expr fields in big StmtKind variants (280 → 96 B)	both

Time — criterion median

Stage	1 MiB before	1 MiB after	Δ	5 MiB before	5 MiB after	Δ
tokenize	11.3 ms	11.2 ms	~0%	54.4 ms	55.0 ms	~0%
parse	79.5 ms	68.9 ms	−13%	378.7 ms	335.1 ms	−12%
semantic_lower	67.5 ms	44.4 ms	−34%	336.4 ms	214.4 ms	−36%
analyze (e2e)	161.9 ms	113.6 ms	−30%	789.6 ms	552.2 ms	−30%

Tokenize is unchanged (no lexer changes).

Memory — peak

Stage	1 MiB before	1 MiB after	Δ	5 MiB before	5 MiB after	Δ
parse	45.8 MB	43.2 MB	−6%	227 MB	214 MB	−6%
semantic_lower	170.0 MB	70.3 MB	−59%	824 MB	345 MB	−58%
analyze (e2e)	213.7 MB	111.3 MB	−48%	1034 MB	542 MB	−48%

Memory — retained (net-live, after the stage returns)

Stage	1 MiB before	1 MiB after	Δ	5 MiB before	5 MiB after	Δ
parse	43.7 MB	41.1 MB	−6%	210 MB	198 MB	−6%
semantic_lower	128.7 MB	30.2 MB	−77%	619 MB	145 MB	−77%
analyze (e2e)	172.5 MB	71.2 MB	−59%	829 MB	343 MB	−59%