[MrMha]

What

The maintainer identified shader recompilation overhead as a priority (comment). This change removes an unclaimed quadratic operation in the shared IR rewrite primitive.

IrValue::ReplaceAllUsesWith previously called ReplaceArgument once per operand. Each call searched and erased entries in the source's use vectors; replacing a highly shared value therefore cost O(n²). Update the referring operands in one reverse traversal, append the replacement's uses in the same order, and clear the old lists once. Existing operands, use ordering, opcodes, metadata, phi predecessors and parent ownership are preserved.

Reserve both destination lists before changing references, with geometric capacity growth. Allocation failure leaves the graph intact; the old implementation can remove a source use and redirect its argument before a destination allocation fails.

No instruction lowering, shader-cache policy or optimization proof is changed. #3726 changes an SRT caller to use this existing helper; it edits different files and is not a dependency. The reported masked-walk, host-import-budget and Windows red-zone changes already have other proposals; this PR does not duplicate them.

Tested

Linux x86-64, GCC 13.3, C++20, AMD EPYC 9V74 host.

• Seven native suites pass with -O2 -Wall -Wextra -Werror, including duplicate operands and existing replacement uses, self/null replacements, phi/identity users, follow-up edits, both destination allocation failures, geometric growth, and 200,000 references.
• 2,000 deterministic differential use graphs, with 48,000 mutation comparisons against the original algorithm, preserve exact ordered arguments/use lists and metadata. These are synthetic IR graphs, not title shader replays.
• ASan/UBSan passes. Restoring the original IrValue.cpp fails the new allocation-failure-integrity regression; the fixed implementation passes.
• A dependency-free native target is registered as shader_ir_use_rewrite under BUILD_TESTING, including relinker-only builds, with a 20-second timeout. The complete release suite runs in approximately 0.20 s locally.
• Conventions and whitespace checks pass.

Five alternating baseline/fixed runs per benchmark size, using fresh graphs and excluding construction/destruction:

Operand references to replace	Baseline median	Fixed median	Operation speedup
4,000	3.343 ms	0.0320 ms	104x
16,000	51.605 ms	0.1760 ms	293x
64,000	853.822 ms	0.9776 ms	873x
128,000	3,410.449 ms	2.0543 ms	1,660x

Every run verifies all redirected operands and final use counts. Separate tests of 2,000–32,000 small two-reference rewrites into one replacement improved by 1.23–1.45x, checking that repeated reservations do not introduce quadratic growth.

These timings isolate the IR helper. End-to-end shader compilation, byte-identical SPIR-V on title captures, retail-title FPS, full runtime CTest and Windows/macOS execution remain unverified locally. No game-wide speedup or direct resolution of #3504/#2809/#361 is claimed.

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 in TechnicalDebt
• 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
• AI-assisted: yes

[oneandonlydean]

Which shader replaces a value with thousands of uses? The table times ReplaceAllUsesWith alone on synthetic graphs of 4,000 to 128,000 references, and perf PRs need before/after numbers for a real case (#218): the recompile time of a dumped request through agc_shader_replay, on main and with this PR, is enough.

[MrMha]

Acknowledge that your measurements are synthetic and that real shader replay timings remain unavailable.

[adriadam10]

Checked on the PR merged onto current main: full build and ctest are green apart from agc_driver_tessellation (lavapipe SEGFAULT, unrelated to this code), shader_ir_use_rewrite passes, and it fails with IrValue.cpp reverted. The rewrite is equivalent to the old loop (use order, nullptr, self-replacement, phi/identity).

One thing is missing before this can merge:

• The benefit is only shown on synthetic graphs (ReplaceAllUsesWith alone, 4,000–128,000 references). As @oneandonlydean asked, perf PRs need before/after numbers on a real case (CONTRIBUTING, perf PRs): the recompile time of a dumped request through agc_shader_replay on main vs this branch (e.g. the Stray requests from #3504/#4331, or any title with a high-fan-out value), plus a profile showing ReplaceAllUsesWith is a meaningful share of it. Without it we can't tell this is not a no-op in practice (#3726 touched the same helper and was closed without that).

Happy to merge once that number is in the PR.

[oneandonlydean]

Closing: no title is known to reach this. With the queue at 700+ open PRs, please hold off on new PRs that don't name a title needing the change until it's down. If a title stops on this on current main, reopen with the title ID and the log line.

[MrMha]

Evidence follow-up: the existing timings are synthetic IR-helper measurements. I still do not have a real title request with before/after agc_shader_replay timings or a profile attributing meaningful compile time to IrValue::ReplaceAllUsesWith, so this PR is not demonstrated merge-ready.

I checked the cited Stray reports: #3504 was closed as fixed by #4331, which changes MaskedSelectEliminator. Those results cannot be attributed to this IR-use rewrite. The new RINNECODER/AnyPS5#482 carry also explicitly says performance was not measured and combines multiple changes.

To complete the requested comparison, could a contributor attach a downloadable APS5_DUMP_SHADERS request that exercises this helper, with its title ID/version and source commit? A current-main versus current-main-plus-this-patch replay on the same machine, matching target/optimization options with cache disabled, repeated timings, a helper profile and SPIR-V comparison would establish whether the optimization matters in that real case. A small replay request would be sufficient; the numbers from unrelated shader optimizations are not evidence for this PR.

AI-assisted: yes (Codex).