fix(shader): shuffle both wave64 halves of a scalar with one vector shuffle

[mariuszkreft]

What

Found by driving GTA V (PPSA04264) on Linux/RTX 3090 (NVIDIA 580.159.03): creating the compute pipeline for shader 69ec8524477a669e segfaults inside NVIDIA's SPIR-V compiler (SIGSEGV in libnvidia-gpucomp.so+0xe3facc, exit 139, no AnyPS5 message). This is the same offset as the ballot crash of #6654, but this pipeline still crashes after its ballots are rewritten as subgroup OR reductions.

A wave64 program held at two lanes per invocation on a 32-lane host emitted every wave shuffle as two scalar OpGroupNonUniformShuffles, one per half, with the same physical lane (SpirvValueEmitContext::Shuffle). The crashing pipeline has 103 such pairs. Replacing the shuffles with OpUndef makes it compile, and removing any 9 of the 72 shuffles of one DPP stretch does too, so no single construct is at fault: the trigger looks like the number of subgroup operations a driver analysis walks.

• SpirvValueEmitContext::Shuffle (SpirvEmitter.cpp) now shuffles a scalar argument as one OpGroupNonUniformShuffle of a two-component vector holding both halves, extracts the two halves and selects the one the source lane is in, as before. Vector arguments keep the two shuffles. The result is the same for every lane, with half the subgroup operations.
• With this change the captured pipeline compiles in agc_compute_pipeline_replay (3 of 3 runs), and the 15 other GTA V captures with shuffles compile within 10% of their old times.
• What in the driver triggers the crash is not known, and a larger shader may still hit it; this is recorded as a technical-debt entry in docs/dev/debt/shader.md. Only the NVIDIA 580 driver was tested.

Tests: wave32_wide_subgroup gains CheckTwoLaneShuffle, which links two half contexts, interprets the emitted SPIR-V and requires one shuffle for both halves and the right value from either half for 32 requested lanes spread over both halves.

Possibly related: #6294 (Borderlands 4, NVIDIA 595, segfault in the same compiler). Not verified against it.

Tested

Linux, on the host, in the slice worktree. The loop ran this gate itself; the tail of it is
at the bottom of this pull request. No title and no GPU were involved.

cmake -S . -B build-full -G Ninja -DCMAKE_BUILD_TYPE=Release -DBUILD_TESTING=ON
cmake --build build-full --parallel 12
ctest --test-dir build-full --output-on-failure -j 12
python3 tools/check_conventions.py --base origin/main

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; no silent stubs added
• No notes, investigation .md files or images added to the repository
• New third-party code is a submodule built from source
• Generic behaviour, not specific to one title
• Depends on: none
• Unblocks: PPSA04264 Game bring-up #1203 — reported against this title, not verified on it
• GPUs tested: n/a (host-side change; no GPU execution)
• AI-assisted: yes

Verification log tail

=== slice gta5-shader-wave64-vector-shuffle: fix/shader-wave64-vector-shuffle @ 310a88f70 on origin/main 5d9dcc42d ===
cmake -S . -B build -G Ninja -DCMAKE_BUILD_TYPE=Release -DBUILD_TESTING=ON
ninja -C build   -> rc=0 (3304 steps)

=== affected test ===
1/1 Test #750: wave32_wide_subgroup .............   Passed
100% tests passed, 0 tests failed out of 1

=== fails before (SpirvEmitter.cpp reverted to origin/main) ===
1/1 Test #750: wave32_wide_subgroup .............***Failed
a wave64 shuffle of a scalar was not issued once for both halves

=== ctest -R 'agc|shader|wave|subgroup|spirv|Wave32' -j16 ===
99% tests passed, 3 tests failed out of 605
	393 - agc_driver_flip_stores_small_results (Failed)
	608 - agc_driver_store_idle_results (Failed)
	609 - agc_driver_store_over_untracked (Failed)
rerun alone: agc_driver_flip_stores_small_results, agc_driver_store_idle_results passed
agc_driver_store_over_untracked: pre-existing on origin/main 5d9dcc42d (recorded in slices gta5-shader-ballot-subgroup-or, gta5-shader-ballot-once-per-block)
skipped: agc_host_import_sparse, agc_driver_sparse_dma_buf_import, agc_driver_stencil_ref_export

=== conventions ===
All conventions checks passed.

=== RESULT ===
PASS

[github-actions]

Other open pull requests touch the same code. Merging one will break the other, or both implement the same thing. If this one builds on them, list them in Depends on.

Pull request	Conflicting files (lines on main)	Same exports	Same new files
PR 6010	docs/dev/debt/shader.md L104

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 4339	core/shader/recompiler/SpirvBackend/src/SpirvEmitter.cpp
PR 4650	core/shader/recompiler/SpirvBackend/src/SpirvEmitter.cpp
PR 5021	core/libs/tests/Wave32WideSubgroup.cpp core/shader/recompiler/SpirvBackend/src/SpirvEmitter.cpp docs/dev/debt/shader.md
PR 5371	docs/dev/debt/shader.md
PR 5426	docs/dev/debt/shader.md
PR 5500	docs/dev/debt/shader.md
PR 5941	docs/dev/debt/shader.md
PR 5958	docs/dev/debt/shader.md
PR 6009	docs/dev/debt/shader.md
PR 6044	core/libs/tests/Wave32WideSubgroup.cpp
PR 6177	core/shader/recompiler/SpirvBackend/src/SpirvEmitter.cpp
PR 6222	core/shader/recompiler/SpirvBackend/src/SpirvEmitter.cpp docs/dev/debt/shader.md
PR 6235	docs/dev/debt/shader.md
PR 6412	docs/dev/debt/shader.md
PR 6416	core/shader/recompiler/SpirvBackend/src/SpirvEmitter.cpp

[milescoviello]

A data point from Borderlands 4 (PPSA01494), the NVIDIA compiler crash in #6294. Linux, RTX 4070 Ti, NVIDIA 595.99.02, 600 s with scripted pad input:

Build	Runs	Result
main bf20538	1	EXIT-139 at 323 s, backtrace ends in _nv002nvvm
main 673f740 + this PR (c07c27a7e2 merge)	3	EXIT-139 at 346, 224 and 216 s; the one symbolized backtrace ends in _nv002nvvm

So on 595.99 this change doesn't avoid the Borderlands 4 crash, which matches what I saw for #6654's rewrite there. Like #6654, it says nothing about GTA V on 580.

AI-assisted: yes (Claude Code)

[savvaniss]

GPU execution data for this change: the first run on GPUs, as far as I can see. It covers a 32-wide host (Intel) and a 64-wide host (AMD).

Windows 11 Pro 26300, WinLibs GCC 15.2.0 r7, Release, every execution test built, ANYPS5_NO_SHADER_CACHE=1, serial CTest, main 47a266cf9. Intel UHD Graphics 770 (8086:a780, 32-wide subgroups), driver 32.0.101.7088. AMD Radeon RX 9070 XT (1002:7550, 64-wide subgroups), driver 32.0.31041.1004, with VkCache emptied first.

	UHD 770	RX 9070 XT
This head 310a88f70 merged into main: wave32_wide_subgroup, ×3	pass	pass
This PR's test file on main without the fix, ×3	fails	fails
Full CTest with this PR alone	1392 of 1398, same failures as main	1396 of 1398, same failures as main

main alone: UHD 770 1392 of 1398, RX 9070 XT 1396 of 1398. The failures are image tables (#6647) and sample_multisample_target (#6971) on Intel, and quad_pixel_mask (#6255) and layered_volume_vertex_no_gpl (#4768) on AMD.

This PR conflicts with #6638 in core/libs/tests/Wave32WideSubgroup.cpp, so one of them needs a rebase after the other lands.

AI-assisted: yes (Claude Code)

[mariuszkreft]

Closed by the contribution loop's guard: this pull request has no checklist and was not opened by the loop's reviewed path. The work will be reopened as a pull request after the local gate and the review pass.