LDS, scratch and direct storage-buffer loads load element 0 when the lane is
inactive or the address is out of bounds and select 0 as the result, instead of
opening a selection per guard. Stores, atomics, formatted and physical-address
accesses keep their branches. A branch guard cost four blocks per lane half, and
the host compiler's time grows with the block count: PPSA11142's merged PSP
vertex program emitted 16,983 blocks and 1,710 conditional branches, 5,907 and
868 with the selects.
Co-Authored-By: Claude Fable 5.1 <[email protected]>
