[Joey-Magnus]

One of OoT's task programs fails to translate ("does not lower ... native stage buffer access mode") because it reads back its own task payload, and the lowering only took stage-buffer reads from the input region. Its mesh pipelines are behind 12 of the 17 programs that still fail in the startup pipeline build, so whatever they draw is missing on every PC.

The instruction (program 64ba846654e63ea4, task stage, +0x80):

ff057923 00001000 0000b800 000e2200   ISBERD.32.ATTR.O R5, [RZ+0x10]

Bits 75 (.32), 76-77 = 3 (.ATTR) and 79 (.O, the output buffer). The program reads word 4 of its payload, which it writes with ISBEWR.O.

Change

• lower_to_maxwell.h (0x923/0x927): a read may now be of the output region. Writes are still always to it, and reads still only from the attribute region. The nx2 pseudo instruction already carries .O in bit 47.
• internal_stage_buffer_entry_read.cpp: ISBE_nx2 decodes bit 47, and a read with it becomes the new IR op IsbeReadOutput.
• emit_spirv_mesh.cpp: EmitIsbeReadOutput loads the word EmitIsbeWriteAttribute stores there: PayloadWord(index) in a task program, the shared attribute array (clamped, as the write) in a mesh program.
• opcodes.inc / ir_emitter gain IsbeReadOutput. The hand-sized opcode tables in opcodes.cpp/opcodes.h go to 543 (test_lower_atomg_cas.py checks that).

Tests

• New tests/test_lower_isberd_output.py (in run_all.py): the log's ISBERD.O lowers to the pseudo instruction with .O. Without .O it's the input read as before. .O on the map region is refused, ISBEWR without .O is refused, and ISBEWR.O lowers as before. It fails on main and passes here.
• run_all.py: all pass except nx2/test_nup_stand_in.py, which fails on main too (a separate break I'm looking at).
• The renderer builds standalone. In game (Linux, RTX 4070 Ti, 595.99.02), stacked on fix(shaders): lower TLD.MS (multisampled texel fetch) to Maxwell TLDS 2D.LZ.MS #1769 and fix(shaders): LDG/STG [URn+imm] with ordering bit 78 lower (2 compute programs) #1778: programs that do not lower 5 → 4, and the startup build's failed programs 17 → 5, with no pipeline errors.

No screenshot yet: I haven't identified which effect these task/mesh pipelines draw.

🤖 Generated with Claude Code