SM86.Scoreboard

The scoreboard pass: the machine-level check the functional model cannot make. SM86.MachineModel gives every instruction its result at once; the silicon does not. A variable-latency instruction (S2R, a global load, shuffles, shared-memory and tensor-core forms) delivers its destination later, and the only thing that orders a consumer after it is the control word: the producer names a write barrier (SB0..SB5) and the consumer's wait mask names that barrier. A consumer without the wait reads whatever the register held before -- on the RTX 3090 (2026-09-22) the first linear step realization read zeros for W, x and t and wrote a record of zeros, while the functional model had accepted it. This pass walks the program in order as one straight-line issue stream (predication does not change issue), keeping the registers whose value is still in flight and the barrier each waits on. An instruction first retires every pending register whose barrier is in its wait mask, then any register it reads or writes that is still pending is a hazard, then its own variable-latency destinations become pending on its write barrier -- or on a barrier no mask can name, when it declares none, so every later use is a hazard. The result is 0 for a clean program, else the ordinal (from 1) of the first hazardous instruction. Fixed-latency forms are ordered by their stall counts, which this pass does not model.

37field unrestricted scoreboardPendingBarrier : Nat