[meh7an]

What

materializePreparedMemoized keeps each source's materialized results by variant and snapshot hash, so a capture that reproduces a snapshot gets the shared result instead of a fresh Populate. The memo is a std::list in LRU order with an unordered_map index (256 entries). Every hit splices the list. Every insert allocates a list node and a hash node, and evicts the oldest, which frees two more.

The memo is now a fixed table of 64 sets of 4 entries per source, allocated with its first insert:

• an entry's set is its index with the high half folded into the low;
• a lookup reads the 4 entries of one set;
• an insert takes the set's least recently used entry by a per-source clock, so no node is allocated or freed;
• the table holds 256 entries as before, and the counters are the same (an insert over a live entry counts as an eviction).

The eviction policy changes from LRU over all 256 entries to LRU within a set of 4.

On the S2 page of Sonic Origins Plus's main menu, the queue worker spent, per frame (APS5 sampler, 15 s):

• on the memo itself (materializePreparedMemoized minus the materializeResult it calls on a miss): 1.87 to 1.92 in three runs without this change, 1.33 to 1.60 in four runs with it;
• on materializeResult, the misses' work: 2.3 to 2.9 in all seven runs, so the sets do not miss noticeably more.

That is about 1.2% of the worker's frame (31 to 33 samples), so the frame times below move within run-to-run noise.

The change is in the recompiler's sources, so the shader cache version changes with it.

Tested

Windows 11 (10.0.26300), WinLibs GCC 15.2.0 (UCRT), Release, NVIDIA GeForce RTX 2080 Ti, driver 617.14.

• On main 5d9dcc4 with this change, the full ctest suite passes 1298 of 1300 tests. (agc_device_fault was linked again and rerun after a locked file stopped its first link.)
  • agc_driver_gpu_indirect_arguments_misaligned_writer_ahead is not run: main itself does not compile it with MinGW (alignas(65536)).
  • agc_driver_multisample_fill_host_only failed once ("the sentinel fill did not land"). That is main's own race, agc_driver_multisample_fill_host_only fails intermittently: the sentinel fill after an absorbed multisampled clear does not land #6512, which fix(agc): record a multisampled target's commands on its own device #6522 fixes; with fix(agc): record a multisampled target's commands on its own device #6522's line added it failed 0 of 200 runs (8 at a time).
• Sonic Origins Plus, main menu, S2 page (the 3D background), 1233x694 window.
  • Captured with APS5_TRACE_SECONDS=200:15 on the same machine in one session, caches warm. Traces attached.
  • Baseline 40a9184, this change 11882cc (the summaries name the commits before them, see the note):

Metric	Baseline	This PR	Change
frame_ms.p50	105.812	103.241	-2.4%
frame_ms.p90	113.658	107.636	-5.3%
mean_frame_ms	107.688	103.748	-3.7%
gpu_busy_ms_per_present	89.495	91.014	+1.7%
Driver.Packet (ms per frame)	105.641	101.713	-3.7%
Driver.Draw (ms per frame)	93.013	89.222	-4.1%

The FPS shown in the title went from 10.36 to 10.48 (median of six readings), and a sampled pair taken about 40 minutes earlier went from 10.50 to 10.82. Last night's windows of this page varied by about 5% at the median, so this pair's gap is mostly noise.

Note: measuring base. These numbers come from my fork's dev, on main 0a56aaa. The memo this change replaces is the same on main 5d9dcc4.

• The baseline is measure/pipe-2026-10-16 (40a9184).
• This change is measure/memo-2026-10-16 (11882cc).
• Each build was made before its commit, so the summaries name the commit before it, dirty: 390f624 for the baseline and 40a9184 for this change.

Both carry these commits from my fork:

• the fixes a71073c, c0d8756, e3c6896, 604f0f3 and 10ff693 (fix(agc): bind a V# past the range limit by the memory behind it #5778);
• make-exe and the stage dumps (522230d, deee71f);
• the crash-report sampler (8d9c310..f641e1e), which took the profiles;
• perf(agc): present 8-bit integer resident images without a readback #6037, perf(agc): skip the register read log while no log is open #6066, perf(agc): skip the binds a continued draw repeats #6077, feat(libs): let relinked Windows games opt into mimalloc #6067, perf(agc): copy small draw snapshots into a per-recorder upload chunk #6082 (its first version, before perf(agc): reuse draw snapshots across draws that nothing wrote between #5906), perf(shader): compile resource plan walks instead of interpreting them #6095, perf(memory): skip 4 MiB runs no write stamped since a range check #6266, perf(agc): replace winpthreads locks and TLS on the draw path #6292, perf(agc): cut per-draw heap allocations #6312, perf(agc): rule out pending write overlaps from the published union #6609, perf(shader): validate vertex stage inputs without building them #6611 and perf(agc): move per-draw thread-locals to TEB slots #6615;
• smaller steps kept in my fork: e832602, 97b1ff6, 84f5bf0, 3b47f77 (covered on main by perf(libc): walk shared views in order and remember collectable ranges in Windows write collects #5062), ce567bc, a79e58a, a0a8916, 63f7a8e, b048061, 719fb71, a21a3ce, e3b8d1c (which perf(agc): fewer heap allocations per recorded draw #6566 covers), 390f624 (a draw decode cache; main reuses decodes through its draw data hits) and 40a9184 (the pipeline key outside the store's lock, which perf(agc): journal the pipelines of a run and build them ahead on the next one #5875 covers, and the thread's last pipeline).

The comparison tool marks the window's first present as a configuration difference: the window opens 200 s into each run. Both summaries set trace_dropped, because the worker filled its event buffer at about 2,700 draws per frame. The phase totals above stay complete.

Checklist

• Based on current main; no other open PR implements the same functions
• One topic per PR; follow-ups go in a new PR
• No comments in code except technical debt
• Unimplemented paths throw (NotImplemented_nid_no_patch); silent stubs are listed under Silent stubs in their library's technical debt file
• No notes, investigation .md files or images added to the repository (attach them to this PR)
• New third-party code is a submodule built from source
• Generic behaviour, not specific to one title
• Depends on: none
• Unblocks: none
• GPUs tested: NVIDIA GeForce RTX 2080 Ti, driver 617.14, Windows 11
• AI-assisted: yes

Traces and summary lines (baseline 40a9184, whose summary says 390f624; this change 11882cc, whose summary says 40a9184):
memo-traces.zip

[github-actions]

Some comparisons could not be completed because a pull request conflicts with the base branch. Review these shared files manually before concluding that there is no overlap.

Pull request	Shared files not compared
PR 3247	core/shader/recompiler/Recompiler.cpp
PR 4072	core/shader/recompiler/Recompiler.cpp
PR 5021	core/shader/recompiler/Recompiler.cpp
PR 5500	core/shader/recompiler/Recompiler.cpp
PR 5609	core/shader/recompiler/Recompiler.cpp
PR 5656	core/shader/recompiler/Recompiler.cpp
PR 6177	core/shader/recompiler/Recompiler.cpp
PR 6539	core/shader/recompiler/Recompiler.cpp
PR 6612	core/shader/recompiler/Recompiler.cpp
PR 6762	core/shader/recompiler/Recompiler.cpp
PR 6879	core/shader/recompiler/Recompiler.cpp
PR 6924	core/shader/recompiler/Recompiler.cpp
PR 6941	core/shader/recompiler/Recompiler.cpp
PR 6946	core/shader/recompiler/Recompiler.cpp
PR 6947	core/shader/recompiler/Recompiler.cpp
PR 6977	core/shader/recompiler/Recompiler.cpp
PR 7201	core/shader/recompiler/Recompiler.cpp
PR 7223	core/shader/recompiler/Recompiler.cpp
PR 7332	core/shader/recompiler/Recompiler.cpp
PR 7390	core/shader/recompiler/Recompiler.cpp
PR 7395	core/shader/recompiler/Recompiler.cpp
PR 7484	core/shader/recompiler/Recompiler.cpp
PR 7642	core/shader/recompiler/Recompiler.cpp

[parl0v]

Data point from a CPU-bound scene on Windows: RTX 3060 12 GB (NVIDIA 617.42), i5-12400F, Windows 10 19045, MinGW GCC Release. Minecraft Dungeons II (PPSA16064 01.100.100), title screen, no input, 200 s rows, APS5_PROFILE_DRAW=1 APS5_PROFILE_GPU=1; main af0fff7 vs main + this PR, interleaved with five other PRs in one batch, two rounds. FPS is the mean of the last 10 [present] windows, GPU busy and idle gaps per frame over the last 8. This scene is CPU-bound on this machine (44-47 ms of idle gaps per frame).

libs	FPS	GPU busy	idle gaps
main	6.61 / 6.76	105 / 104 ms	46 / 44 ms
+ this PR (92e5b97)	6.70 / 6.63 (-0.3 %)	104 / 104 ms	45 / 46 ms

Within the run-to-run spread (main itself moved 6.61 -> 6.76 between rounds), no regression, 0 fatal.

AI-assisted: yes (Claude Code)

[adriadam10]

Checked the hit/insert paths under source.mutex, victim selection (empty ways first, then least recently used), the unchanged counters and capacity; the shader cache version bump is automatic. Merged.