[glampert]

Follow-up to #9.

Problem

A VF write to some lanes of a variable (add.xyz fC, ...) was joined to the variable's earlier value whenever any other lane had been written before, so that the kept lanes stay in the same register. BranchState accumulates the lane flags along a path, so on a loop's second trip the write was joined to the previous trip's value even when every lane it kept was written again before being read (add.xyz fC then add.w fC, then sq fC). Values rewritten lane by lane on every trip then stayed live for the whole program and pinned a register each.

Tyra's stapip_cull_td_vu1 builds its three output colours this way; SCE vcl compiles it, openvcl failed with "ran out of VF registers (31 available) for outputColor3" (its exact live ranges need 29).

Fix

Before the branch-state walk, collectLanesLiveAfterWrites works out, for each partial write, which of the lanes it keeps are read before they are written again (a block's out_vf outputs count as read at its exit). BranchState::writeFloat joins the write to the earlier value only when one of those lanes is live. A read that needs lanes from several writes is still joined to all of them by completeLiveRanges from #9.

Tests

test_loop_liveness.cpp gains "a partial write starts a new value when the lanes it keeps are dead": a loop that writes fC xyz then w before storing it, with four VF registers; current master runs out of registers, this passes.

Effect on generated code

Checked against the ps2gl fixtures, Tyra's 15 programs and quake2-ps2's, each assembled with dvp-as and checked with vu-checker, master (17cbadb) against this branch:

• Tyra stapip_cull_td_vu1 now compiles.
• 25 programs change (register numbering and the scheduling that follows from it; all 12 fixtures, scei, billboard_renderer, 10 Tyra programs, quake2 textured_triangles), 21 are byte-identical.
• The regalloc finding vu-checker reports on master for general_tri, general_pv_diff_tri, general_quad and general_pv_diff_quad ("instruction 171 (move.xyzw) reads VF30, which holds any of intensity66, normal depending on the path") is gone, and nothing new is reported.
• The test suite and an ASan/UBSan build of the suite and all those programs are clean.