Fix in-place SetPrune aliasing

[RAMitchell]

Summary

Fix WQSummary::SetPrune so its in-place compaction does not corrupt source entries that are still needed by later loop iterations.

The previous implementation read and wrote through the same data_ buffer. For some rank geometries, an earlier output write could overwrite src_data[i], and a later decision would then read the corrupted slot. This could produce duplicate value entries in the pruned summary, violating the strictly increasing value invariant.

This keeps SetPrune allocation-free by caching the active source window (left, right, and tail) before writes can clobber those entries.

Tests

Added focused coverage for SetPrune:

• deterministic regression for the reported aliasing geometry
• out-of-place reference implementation used only in tests
• exhaustive small integer geometries
• randomized larger geometries
• targeted boundary-shaped geometries around zero/nonzero weights

Validated locally:

• cmake --build build-cpu --target testxgboost -j35
• ./build-cpu/testxgboost --gtest_filter='Quantile.SetPrune*'
• ./build-cpu/testxgboost --gtest_filter=Quantile.*
• ./build-cpu/testxgboost --gtest_filter='*QuantileSummaryTest*:*QuantileContainerTest*'

Performance

Compared against upstream/master with a standalone microbenchmark that repeatedly restores pre-generated summaries and calls SetPrune.

The in-place cached-entry fix is allocation-free but slightly slower for larger summaries. Pinned-core net prune timing was approximately:

size -> maxsize	upstream	fixed	delta
32 -> 16	237.8 ns	228.3 ns	4.0% faster
128 -> 32	649.5 ns	651.5 ns	0.3% slower
512 -> 64	1562.1 ns	1627.4 ns	4.2% slower
2048 -> 256	6295.7 ns	6613.7 ns	5.1% slower
8192 -> 512	14446.4 ns	16952.9 ns	17.4% slower

I also tried a lazy-cache variant that only cached entries when a write would overwrite the active read window. It passed the same tests but benchmarked worse due to extra branch/state overhead, so this PR keeps the simpler cached-window implementation.

[Copilot]

Pull request overview

This PR fixes an in-place aliasing bug in xgboost::common::WQSummary::SetPrune where writing pruned entries back into the same buffer could overwrite source entries that were still needed later in the prune loop, potentially producing duplicate (non-increasing) values in the pruned summary.

Changes:

• Update WQSummary::SetPrune to cache the active source window (left, right, and tail) before writing into data_, preventing read-after-write corruption in the in-place compaction loop.
• Add a test-only out-of-place reference prune implementation and multiple regression/property-style tests to validate the in-place prune behavior across a variety of geometries.
• Add a deterministic regression test targeting the reported aliasing geometry that previously could violate the “strictly increasing values” invariant.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
src/common/quantile.h	Fixes WQSummary::SetPrune in-place compaction by caching source entries used by later iterations.
tests/cpp/common/test_quantile.cc	Adds focused SetPrune regression and reference-comparison tests, including enumerated, randomized, and boundary geometries.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.