[glampert]

The generic software pipeliner picks scratch registers for the values it rotates from VF31 down, skipping only the registers the loop body names. A register that holds a value the body never names, but that is read after the loop or on the next trip of an enclosing loop, counts as free, and a rotation overwrites it.

ps2gl's general program hits this with the default pipelining. The directional-light vertex loop's rotations take the registers holding material_spec and vert_to_eye. Both were loaded before the loop and are read again for the next light and by the point-light loop. Those later passes then compute with garbage. general_pv_diff has the same problem.

Fix

findVuLoopCandidates now works out which VF lanes are live where each loop is entered and where it is left. It uses lane-level liveness over the whole program's control flow: labels, branches, bal/jr returns, --endexit, and out_vf read at the exit. The result is stored in VuLoopCandidate::liveVfRegisters. collectLoopVfBaseKeys adds those registers to the ones the body names, so all three scratch pools keep clear of them.

Effect

• Of the 12 ps2gl fixtures, general and general_pv_diff change. Every vu-checker register-allocation finding they had is gone, and nothing new appears. The other 10 compile byte-identical.
• In a local corpus of 46 programs from ps2gl, Tyra, quake2-ps2 and the openvcl examples, the remaining 44 are byte-identical.
• The test suite and an ASan/UBSan run are clean.

Tests

test/integration/test_pipeline_live_registers.cpp adds two cases. Both fail on master without this fix:

• a value read after a pipelined loop keeps its register;
• a value read on each trip of an outer loop keeps its register.

Not covered

The env-gated generic kernel rewrite has its own rename pool, which this change doesn't touch. That path is not enabled by default.

[fjtrujy]

Good improvement!