Faster style v2

[charlescgs]

This PR re-enables and fixes the style's cache, and adds few optimizations to reduce allocations and redundant work on the style hot path. Result is ~10-29% improvements on style benchmarks with no regressions.

Changes

1. apply_mut takes &Style instead of Style

Every call site was cloning Style just to pass ownership to apply_mut, which then only reads it. Changed the signature to eliminate ~15 unnecessary Style clones across the codebase

2. Cached StyleSelectors bitmask on Style

selectors() previously traversed the entire style map to check which selector types were present — called repeatedly during resolve_selectors. Added a cached_selectors: StyleSelectors field incrementally updated when selectors are added via apply_iter, set_selector, set_structural_selector, set_responsive_selector. Reduces selector presence checks from O(n) map traversal to O(1) bitmask read. (Correctness is being checked debug_assert! against the slow path in debug builds)

3. Per-frame timestamp

Instant::now() was called per-view during the style pass. Captured once at frame start and shared via WindowState::frame_start

4. FxHashMap

Replaced HashMap with FxHashMap for style cache

5. Content-based content_hash() implementations

The default StylePropValue::content_hash() used format!("{:?}", self) — allocating a String per property per hash. Added manual implementations for all ~40 concrete types using FxHasher with to_bits() for floats and std::mem::discriminant() for enums. Zero-allocation hashing throughout

6. Eq via PartialEq instead of hash comparison

StyleMapValue comparison was hashing both sides and comparing hashes. Changed to direct a == b via the existing PartialEq impl

7. XOR order-independent content_hash() on Style

The style-level content_hash() collected entries into a Vec, sorted by key pointer, then hashed sequentially with allocation. Replaced with XOR of independently-hashed entries: each entry gets its own FxHasher, and results are XOR'd together. O(n) -zero allocation (order-independent by construction)

8. Re-enabled style cache with fixes

• Cache entries now store full compute_combined() outputs: combined_style, has_style_selectors, post_compute_combined_interaction
• Replaced Rc<Style> wrapping with direct Style storage (the Rc indirection added overhead without benefit)
• Added window_width_bits to cache key for responsive selectors with exact pixel thresholds
• Added is_focus_within to interaction state bits
• Proper cacheability checks: styles with structural selectors (:first-child, :nth-child) or context-dependent values are excluded
• Cache cleared on theme change and screen size breakpoint change
• Parent validation uses map_ptr() (inner Rc pointer) instead of outer Rc::as_ptr()

9. Cache-hit path avoids Style::clone()

On cache hit, the previous implementation cloned the full Style to build the cache key, then discarded it. Restructured the hot path:

• StyleStack caches the content_hash as a side-effect of style() recomputation
• Added ensure_clean() / content_hash() / is_cacheable() methods on StyleStack
• Added style_content_hash() / style_is_cacheable() forwarding on ViewState
• Cache key construction uses the cached hash via StyleCacheKey::new_from_hash() — no Style reference needed
• On hit: directly assign cached outputs to ViewState fields, bypassing compute_combined() entirely

10. More tests coverage

Benchmark results (vs current main)

Benchmark	Change
view_creation/empty build	-29%
view_creation/empty mount	-14%
view_creation/label mount	-5.7%
identical styles/create+compute	-10% to -16%
identical styles/restyle (cache hit)	-22% to -26%
different styles/create+compute	-8% to -14%
deep nesting/create+compute	-6% to -21%
complex styles/create+compute	-7% to -17%
style_clone_and_read	-19.6%
inherited_prop_updates	-1% to -7%
inherited_with_selectors	-2%
inherited_stable_context	-5%

@jrmoulton can you test this yourself?

[jrmoulton]

This looks good to me. I wish there was a more automatic way to do content hash but I do think that this is better than allocating/hashing a string.

[jrmoulton]

@charlescgs do you mean for this to still be a draft?

[charlescgs]

No, added is as a draft so can gather feedback

[jrmoulton]

Thanks for doing this!