What and why
A short note in docs/testing.md for a failure mode that cost me an hour today and looks exactly like a translator regression.
The GPU test programs are not self-contained: vk_compute_test seeds the raw V# and the push-constant heap that the translated modules read. An executable built from an older commit still loads and runs every shader, so when the translator gains a descriptor rule the old program does not set up, nothing errors — every compute test just reports the untouched 0xEEEEEEEE fill.
Measured on an RTX 5090 (driver 621.24) with main at 99f99890:
- test programs built earlier the same day: 69/141 passed, every
cs_*test on the fill pattern (cs_dpp,cs_exec_const,cs_bounds, …), allps_*/vs_*tests passing. - the same tree plus
vk_compute_test,vk_store_testandvk_harnessrecompiled from99f99890: 141/141 passed.
The cause was the raw V# in pc[2..3]: the current source sets pc[3] = 0x31016FAC (OOB_SELECT 3, raw), and the bounds check the translator emits for buffer_store_dword reads STRIDE/OOB_SELECT from it. With the older program the field read as 0 and the store predicate was false for every lane, so the shader ran and wrote nothing.
Testing
uvx pytest -q tests/test_docs_links.py→ 5 passed.- The 141/141 run above, with
LEONIDA_W64_EMULATE=1,scripts/run_gpu_tests.py, Vulkan SDK 1.4.363.1 and llvm-mc 21.1.8.
No game files involved. AI-assisted; the author reviewed the change.