Lanes that do different things

A translated shader can now diverge. v_mbcnt_lo_u32_b32 and v_mbcnt_hi_u32_b32 let a lane learn its own index, v_lshlrev_b32 and v_add_u32 turn that into an address, and v_cmp_lt_u32 compares it - so lane n computes its own value, tests its own condition, and stores to its own address. Thirty-eight execution tests on a real device.

That closes a hole that had been open under everything built this session. Every mask test until now compared the same two registers in every lane, so every mask was all-ones or all-zero - and a translation that tested the mask against zero rather than reading it bit by bit would have passed all of them. The new test narrows the mask to lanes 0-3 out of 64 and checks all sixteen observable words: four written with their own index, twelve untouched.

A lane is not told which lane it is. There is no lane-id instruction and no value handed to the shader. v_mbcnt counts the set bits of a mask below this lane, so with an all-ones mask the count is the index - and getting a full index takes both halves in sequence, the high one adding the low one's result. The boundary is the part to be careful with: strictly below. Including the lane's own bit shifts every index by one wherever that lane is active and leaves it right wherever it is not, so a test with a full mask catches it and a test with an empty mask does not.

Surprises.

• v_lshlrev_b32 takes its shift first and its value second. The "rev" in the name. Read in written order it computes 2 << index where index << 2 was meant - and those are equal for lane 2. So a test on any single lane can pass a wrong translation, and what pins it is asserting every word of the window rather than a sample.

• A test was written and then deleted for adding nothing. A separate case for the shift asserted strictly less than the test above it already did. Keeping it would have been free and would have quietly inflated what the suite appears to cover; the property it was reaching for is now a comment on the test that actually proves it.

• VOP3 opcodes are ten bits, so v_mbcnt is 652, not 140. Written from the solved table rather than inferred from where the mnemonic sits among its neighbours - the first attempt guessed a three-digit-family number and produced an arm that never fired, which is exactly the failure mode that looks like an unimplemented instruction.

Not done. Branching, still. Everything above is predicated execution: the mask decides who acts and control flow never leaves the straight line. A jump taken when no lane survives cannot be expressed, which remains the substance of D098 and a design problem rather than an implementation one. Fidelity::Subgroup is still a stub. The lane and wavefront models can no longer be compared on any shader that diverges, because the lane model refuses them - which is correct and does cost the differential oracle its reach exactly where divergence begins.