[Prophet069]

What

GFX10 scratch SV addresses contain one32-bit VGPR offset, not the VGPR pair used for ordinary FLAT addressing. VGPR255 was therefore incorrectly rejected. Decode SV/SS/ST address components independently and make the scratch IR consume the offset components; retain the FLAT/global VGPR-pair and unsupported-scalar checks. Source: LLVM FLATInstructions.td for the operand widths and SV/SS/ST pseudos, and GFX10 encoding for null/EXEC_HI sentinels. #1190 changes flat-aperture SPIR-V routing and does not modify these decoder/IR units.

Replaces the obsolete base of closed #2965, which was automatically closed for intake pause. Targeted current GitHub search found no other open PR implementing this exact export/decoder correction. Related #5005 is a scalar-aperture change, not the one-dword scratch decoder fix.

Tested

Windows 10 Pro 22H2 x64 (build 19045), MinGW GCC 15.2.0: fresh production decoder, IR, translation and control-flow sources compiled on this candidate; all 14 valid/invalid scratch/FLAT decoder and IR cases passed (0.06 s). The standalone test uses the existing Windows .eh_frame to .ehfram configuration and the freshly built libc import library. No Vulkan or PS5 hardware scratch-address result is claimed. The instruction dc304000 027d00ff was originally captured from Death Stranding 2: On the Beach, PS5 title ID PPSA02015, in an earlier experimental integration. This is not a claim that this isolated PR renders or boots the game.

Based on main dacd78525e537a779682032c0ace7f7af120af05; candidate cc1dc3f1c6f88cd03665f933e0b7590e5717b0fe. No dumps, runtime packages, investigation reports or images are included. No speed improvement 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; no silent success stub added
• No notes, investigation .md files or images added to the repository
• New third-party code: none
• Generic behaviour, not specific to one title
• Depends on: none
• Unblocks: PPSA02015 Death Stranding 2: On the Beach (PPSA02015) #5424 (captured decoder stop; no game rendering or hardware scratch-address claim)
• GPUs tested: not tested on refreshed candidate; prior decoder/IR tests are CPU tests (Windows)
• AI-assisted: yes; prepared with Codex; the submitting author is responsible for every line

[KauaSantos702]

NVIDIA run at this head (cc1dc3f1c), adding a vendor to the AMD results in the body: no agc_* test fails.

• NVIDIA GeForce RTX 5070 (Blackwell), driver 32.0.16.1714, Windows 11
• WinLibs GCC 15.2.0 r7, Release, BUILD_TESTING=ON
• ctest -R '^agc_' -j2: 484 tests, 483 passed, 0 failed, 1 skipped. The skipped test is agc_driver_stencil_ref_export, because the device lacks VK_EXT_shader_stencil_export; it is also skipped on main.

[adriadam10]

Checked the decoder/IR change on a Linux build merged onto current main: agc_flat_scratch passes and fails without the fix, so the logic looks right. Two things before it can merge:

• The PR conflicts with main in docs/dev/debt/shader.md (TechnicalDebt adjacency) and "Allow edits from maintainers" is off, so we can't rebase it for you. Please rebase on current main, and put the new ### GFX10 scratch address operands entry away from the end of the file, where other PRs append.
• core/libs/prx/libSceAgcDriver/tests/FlatScratch.cpp is not formatted like the other tests (no spaces around operators, several statements per line, Scratch/Reject helpers on one line); please run clang-format on it and drop the doubled blank line before the new block in libSceAgcDriver/CMakeLists.txt (line 2241-2242).

Not covered: the real-hardware result of a scratch access with a VGPR offset; the TechnicalDebt entry already says so.

[fma-rc]

A title reaches this: Death Stranding 2 (PPSA02015, 01.015.000) on Linux. It only gets this far with a local fix to the AMM address range (on current main it stops at 1 s on that, PR to follow), and then it registers a shader with

scratch_load_dword v2, v255, off   // dc304000 027d00ff

in its spill code (word 0x553 of a 0xd70-word program, well before s_endpgm; LLVM gfx1030 decodes it the same). Without this PR it stops at 39.5 s with FLAT address register range overflow. With cc1dc3f cherry-picked on top, that throw is gone and it runs to 45 s, where it stops on something unrelated. agc_flat_scratch passes in that build.

[oneandonlydean]

@Prophet069 a friendly nudge on this one. The maintainer's 14 Oct review says the logic looks right (agc_flat_scratch passes with the fix and fails without it). What's left is a rebase onto current main, where the only conflict is the adjacent entry in the shader debt doc, and a clang-format pass. Once those are pushed, the change request can be re-reviewed and the PR can go through CI.

[Prophet069]

@oneandonlydean thank you for the reminder. My ai setup doesnt run currently (swtching to qwen from codex because of slow progression and fast plan usage. Can you rebase it for me? I alllowed edits to be made. Sorry i'm not that knowledgeble with all the github stuff. Just wanted to help without understanding much of it.

[oneandonlydean]

@Prophet069 no problem, and thanks for allowing edits. The branch has main merged in (c1830b3) plus one formatting commit (4c15089): tests/FlatScratch.cpp is clang-formatted, the debt entry has its own place in docs/dev/debt/shader.md, and the extra CMake blank line is gone. Your decoder and translation change is unchanged. CI on that head passed on Linux and Windows, with agc_flat_scratch passing on both. Nothing more is needed from you; we'll take it from here.

Addressed in 4c15089; CI passed.