[erm14254]

What

NVIDIA driver 597.16 (RTX PRO 6000 Blackwell, Windows) crashes in its compiler at nvgpucomp64.dll+0x28072a1 while creating a compute pipeline of Code Violet. spirv-reduce, kept to that crash offset, leaves a selection steered by a ballot whose one arm calls record_bda_fault, a DontInline function, and whose other arm holds another ballot. The reduced module and the full one compile once the function is declared without the control or with Inline, so the function is emitted without DontInline and the driver inlines or calls it as it sees fit. The other DontInline helpers keep the control. The debt entry in docs/dev/debt/shader.md and the NVIDIA note in docs/user/USAGE.md record the finding; the test in agc_shader_memory now expects the fault function without the control.

Tested

On main 00d7e61, where record_bda_fault is still declared DontInline: Windows 11 (10.0.22631), RTX PRO 6000 Blackwell Workstation Edition, NVIDIA 597.16: agc_shader_memory, agc_driver_bda_device, agc_driver_bda_span_reads, agc_driver_guarded_null_pointer, agc_driver_image_table, agc_device_fault and agc_formatted_buffer_once pass.

Before the rebase, on d92a8da with the same GPU and driver, the reduced module and the full compute module compile through agc_compute_pipeline_replay with this, and Code Violet (PPSA26528) no longer stops in nvgpucomp64.dll; on Linux, lavapipe, the same seven tests pass. Plain main 0615cd2 still stopped at nvgpucomp64.dll+0x28072a1 on 2026-10-16 (#6257). The title has not been run on the rebased branch yet; I'll add that result in a comment.

The reduced SPIR-V module is attached below.

pr4-reduced-module.zip

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: PPSA26528 Code Violet (PPSA26528) #6256
• GPUs tested: NVIDIA RTX PRO 6000 Blackwell Workstation Edition, driver 597.16, Windows 11; lavapipe on Linux
• AI-assisted: yes

🤖 Generated with Claude Code

[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 5426	docs/dev/debt/shader.md L171

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 5021	docs/dev/debt/shader.md docs/user/USAGE.md
PR 5371	core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp docs/dev/debt/shader.md
PR 5925	docs/user/USAGE.md
PR 5941	docs/dev/debt/shader.md
PR 6009	docs/dev/debt/shader.md
PR 6011	docs/dev/debt/shader.md
PR 6171	docs/user/USAGE.md
PR 6235	core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp docs/dev/debt/shader.md
PR 6280	docs/dev/debt/shader.md
PR 6290	core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp
PR 6412	core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp docs/dev/debt/shader.md
PR 6488	core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp docs/dev/debt/shader.md
PR 6586	docs/user/USAGE.md
PR 6604	docs/dev/debt/shader.md
PR 6645	core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp
PR 6788	docs/user/USAGE.md
PR 6879	docs/dev/debt/shader.md
PR 6924	docs/dev/debt/shader.md docs/user/USAGE.md
PR 6946	docs/user/USAGE.md
PR 6947	core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp docs/user/USAGE.md
PR 6977	docs/dev/debt/shader.md
PR 7223	docs/dev/debt/shader.md
PR 7410	core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp docs/dev/debt/shader.md
PR 7484	docs/dev/debt/shader.md
PR 7642	core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp docs/user/USAGE.md
PR 7684	core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp docs/dev/debt/shader.md
PR 7689	core/shader/recompiler/SpirvBackend/src/SpirvBdaFault.cpp
PR 7800	docs/user/USAGE.md
PR 7816	core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp
PR 7918	docs/user/USAGE.md
PR 8077	core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp docs/dev/debt/shader.md
PR 8489	docs/dev/debt/shader.md
PR 8564	core/libs/prx/libSceAgcDriver/tests/ShaderMemory.cpp docs/dev/debt/shader.md

[erm14254]

Manual overlap review for the list above, against each pull request's current head. None of them changes the lines this PR changes.

• Applied to main, the shared-file changes of perf(shader): compile reusable SRT expression plans #1341, feat(aot): consolidate native ARM64 compiler, runtime and consumers #5021, fix(tools): recognize guest module identities in import audit #6171, feat(agc): draw tessellation followed by an NGG geometry stage #6612, fix(shader): reuse completed pixel lane proofs #6645, feat(agc): run registered merged hull programs (binary type 3) #6679, feat(agc): select Vulkan devices by driver name #6788, perf(agc): draw finisher pipeline (APS5_DRAW_PIPELINE) #6946, perf(agc): GPU descriptor heap (APS5_GPU_HEAP) #6947, feat: run PS4 games on the AGC driver — Limbo (CUSA01664) plays at 60 fps #6977, feat: macOS on Apple Silicon host support (Rosetta 2 and MoltenVK) #7332, feat(shader): read samplers through nested and pointer-keyed descriptor tables #7395, feat(relinker): support --trace-imports for Linux output #7417, perf(recompiler): leave the descriptor-range comparison out of direct buffer loads when the device zeroes past the range #7484, fix(agc): reclaim idle images safely during cache creation #7800 and perf(shader): span single-wave wave64 compute by default after a uniformity check #7922 merge cleanly with this PR.
• fix(shader): emulate sRGB sampling when unsupported #5371, fix(shader): preserve raw fragment exports for integer attachments #6235, perf(recompiler): keep runtime-specialized shader modules in the shader disk cache #6924, fix(shader): mask writes after barrier-safe BDA faults #7392 and feat(agc): look up table images at run time through a device image table on the GPU heap (APS5_GPU_PATH) #7642 merge cleanly in some shared files and conflict with main in the others; fix(shader): decode GFX10 scratch offsets as one dword #5426, fix(shader): propagate color view slice to pixel ancillary input #5500, perf(agc): build the pipelines of a submission ahead on a compile pool #5925, fix(shader): translate v_mullit_f32 in f32 denormal modes 1 and 2 #5941, fix(shader): translate f32 min and max in f32 denormal modes 1 and 2 #6009, perf(shader): rely on robustBufferAccess2 for direct buffer bounds #6177, fix(agc): keep texture images inside their extent's mip chain #6280, fix(shader): read quad_perm DPP lanes in pixel shaders through quad operations #6290, fix(recompiler): fill missing pixel interpolators as DEFAULT #6604, fix(agc): keep the guest's sample order where pixel (0,0)'s locations stand in for a 2x2 grid #6683, feat(agc): back extra guest mip levels with storage images #7085 and feat(shader): resolve nested image tables through runtime pointers #7223 conflict with main in all of them. Where a PR conflicts with main, its hunks still neither change nor border this PR's lines.
• fix(shader): mask writes after barrier-safe BDA faults #7392 is the only one in SpirvBdaFault.cpp: it adds the faulted-flag helpers after ReturnBdaFailureIf and a branch in StopBdaInvocationIf, about 70 lines below the record_bda_fault declaration this PR changes, and leaves that declaration alone.

No overlap, no Depends on.

[DavidAmsellem]

First AMD-on-Linux (RADV) run of this PR. Merged onto main 8d5d4df96 (clean merge), built in full and compared with the same main, on the same machine.

AMD Radeon RX 7900 XT (RADV NAVI31, gfx1100), Mesa 25.2.8, Linux Mint 22.3 (Ubuntu 24.04 base), kernel 7.0.0-38, GCC 13.3, Release.

Full suite (ANYPS5_REQUIRE_VULKAN=1 ctest -j8) at 6474695df: 1638/1640 with this PR, against 1638/1640 on main. The two failures are the same on both, so this PR adds none: agc_driver_image_table_vouched_lookups (#7766, fixed by #7916) and agc_driver_guest_memory_lifetime (#7784).

Comparing the JUnit output with main: this PR adds no ctest entry, and none is skipped or removed.

This checks correctness only; I did not measure performance.

GPUs tested: AMD Radeon RX 7900 XT (RADV NAVI31, gfx1100), Mesa 25.2.8, Linux Mint 22.3 / kernel 7.0.0-38

AI-assisted: yes

[oneandonlydean]

@erm14254 DontInline on record_bda_fault came from f7d4fba (#1043) as a pipeline-creation fix: with the fault recording copied into every read site, one compute program reached about a million SPIR-V words and ~95 s in vkCreateComputePipelines (PPSA26344's ~835k-word variants took 45 s each). Without the hint the module stays small, but any driver may now inline the function at every call site, which is the cost that change was avoiding. The change also applies to every driver, not only 597.16. Could you add the vkCreateComputePipelines time of a large BDA program before and after (for example through agc_compute_pipeline_replay) on NVIDIA and on AMD, or limit the change to the driver that crashes, as the other driver workarounds do?