unstructured

98de21e5 - Optimize CustomPDFPageInterpreter._patch_current_chars_with_render_mode

Optimize CustomPDFPageInterpreter._patch_current_chars_with_render_mode Runtime improvement (primary): The optimized function runs ~7% faster overall (181 μs → 168 μs). That runtime reduction is the reason this change was accepted. What changed (concrete optimizations) - Single-step cur_item lookup: replaced "hasattr(self.device, 'cur_item') and self.device.cur_item" with cur_item = getattr(self.device, 'cur_item', None) and an early return. This reduces two attribute lookups into one and short-circuits the common no-cur-item case. - Single getattr for _objs: replaced the conditional expression cur_item._objs if hasattr(cur_item, "_objs") else [] with objs = getattr(cur_item, "_objs", []). That avoids an extra hasattr call and repeated attribute access. - Avoid repeated len/index work: replaced index-based loop for i in range(start, len(objs)): obj = objs[i] with direct iteration over the sub-list objs[start:] (for obj in objs[start:]). Also added an explicit if start < len(objs): guard so we don't build a slice when there's nothing to process. Why these changes speed things up (mechanics) - Fewer Python-level attribute lookups: getattr replaces two operations (hasattr + attribute fetch) with one, and caching cur_item and objs removes repeated attribute access inside the hot path. Attribute lookup in Python is comparatively expensive, so reducing them reduces per-call overhead. - Reduced indexing overhead: using direct iteration avoids repeated integer arithmetic and two list index operations per iteration (compute index, __getitem__). That lowers per-iteration Python overhead inside the tight loop that checks and patches LTChar objects. - Early-return common negative case: when device.cur_item is missing/falsy the function now returns earlier with less work, which helps where many interpreter calls have no cur_item. - Guarding slice creation: creating objs[start:] can allocate a new list. The added start < len(objs) check prevents unnecessary allocations on the common empty/no-new-items case; when the slice is needed it's typically small (we only process newly appended items), so the copy cost is minimal compared to saved per-iteration overhead. Why this matters in context (hot path) - This method is invoked from do_TJ and do_Tj (text painting operators) in the interpreter. Those are hot paths during PDF text processing: the method can be called many times per page. Small per-call savings accumulate, so a ~7% per-call runtime improvement is meaningful. Behavioral and workload impact (based on tests) - Regression tests show the biggest wins on heavy workloads: large-scale patching and repeated-appends (1000-item and many-iteration cases) observe sizable reductions (e.g., ~16% and ~26% in annotated tests). Those are precisely the cases where per-iteration overhead dominates. - Some tiny regressions in a few microbench tests (single small calls) are visible in the annotated tests (sub-microsecond differences or small percent slowdowns). These are minor and expected trade-offs for the net runtime improvement across typical workloads. - Memory trade-off: objs[start:] creates a shallow copy of the sub-list. In typical usage this sub-list is small (only newly appended items) so the allocation cost is small and outweighed by per-iteration savings. If you have pathological cases where you repeatedly slice very large tails, that could increase temporary memory pressure — but tests and typical interpreter usage show net benefit. Correctness - The logic is preserved: render_mode is applied only to new LTChar objects and _last_patched_idx is updated the same way. Using getattr defaults keeps previous behavior for missing attributes. Summary - Net effect: fewer attribute lookups, less indexing overhead, and an early-exit optimize the common and hot cases encountered during PDF text interpretation, producing the measured 7% runtime speedup. The small memory/alloc cost of slicing is guarded and, in practice, outweighed by the reduced CPU overhead on the hot path (do_TJ/do_Tj).