Repository navigation
Recompiler: SPIR-V validation fails on a continue construct its back-edge block does not post-dominate #1921
Description
What
A shader a title registers cannot be prepared: SPIR-V validation before optimization rejects the loop construct the structurizer emitted.
[Shaders/ShaderRegistry.cpp:332 operator()] Shader 0x55a82bc4b700 could not be prepared at registration
(ShaderRecompiler::Recompile: SPIR-V validation before optimization failed:
level=2 word=5010: The continue construct with the continue target '44[%44]' is not
structurally post dominated by the back-edge block '2492[%2492]')
The back-edge block does not post-dominate the continue target — i.e. the structurizer produced a continue construct whose shape SPIR-V validation rejects.
Impact
In the run below this was deferred (by #1836), so it costs the title one unprepared shader rather than a crash — but it is a real capability gap, not just a log line: the shader stays unprepared and will be re-attempted when the title first needs it.
Where it was measured
- PPSA26344 (Ghost of Yōtei 01.008): decrypted
eboot.binrelinked, staged runtime layout, RTX 3080 Ti, Linux - stack:
main+ feat(agc): implement image and sampler tables indexed at runtime offsets #757 + fix(agc): defer shaders that cannot be prepared at registration to their first use #1836; 841 shaders compiled in the run, this was 1 of 105 deferrals - no issue or PR matches the message (searched
not structurally post dominated,post dominated by the back-edge,continue construct)
Written with AI assistance.
Activity
snoozedDev commentedon Oct 9, 2026
Another instance, and one #1964 does not cover: Astro Bot (PPSA21564 v01.007.000, Linux) stops on the same message on a graphics-stage program prepared at ABI resolution (so #1836 does not defer it), with #1964 applied on top of
mainf145139:Same program, same words with and without #1964, so this one has a cause other than the image-mode switch. Seen on a RADV wave64 device; it is device-independent. Full context and the other stops are in #2340. The serialized
RecompileRequestis available on request. AI-assisted: yes; reviewed by me.SleglJan commentedon Oct 9, 2026
Another title with this instance: 007 First Light (PPSA11386 v01.018.000), after Ghost of Yōtei in the issue, GTA V on #1964 and Astro Bot in #2340. Linux (CachyOS, kernel 7.2.9, GCC 16.2.1, RTX 3080, NVIDIA 615.71.09), Release build with
ANYPS5_ENABLE_SPIRV_TOOLS=ONofmain8b393b6 plus three commits on my fork that take it past its loader andsceSharePlayInitializestops (SleglJan#15, SleglJan#16, SleglJan#17). It stops at start-up with the same message, word and IDs as in the issue:The program is a wave64 compute shader, 8×8×1 threads, 38 instructions: a loop writes a user-data value (
s[8:11]) to a 2D MSAA array image (image_store, dim2D_MSAA_ARRAY) with the loop counter as the sample index.Each PR head merged onto that build, with the captured request replayed through
agc_shader_replay:DescriptorBindingBuilder::Populate multisampled image resources have no descriptor image shape#2311 does not reference this issue but clears it on this request. #2124 conflicts with
mainonly in the test registration incore/libs/prx/libSceAgcDriver/CMakeLists.txt; I kept both blocks. #1964 at its own head (on f145139) gives the same result. The next error comes from descriptor population for a draw or dispatch; registration (PrepareShader) stops after validation.In the title with #1964, alone or with #1836, start-up gets further: 84 disk-cache misses (82 written; 61 without #1964), then
VideoOut: unsupported display pixel format 0xc001000622000000(RGBA16 float, 3840×2160, tiled). The multisampled-image error does not come up before that stop.AI-assisted: yes (Claude Code).
fix(shader): keep runtime image mode switches inside continue targets