Hi Jeff, Kyrylo's series is in now, so I reran the tests on trunk (2cb78d1a0e5). All five positive tests still fail there (clip-canon-1, 5, 7, 8 and 9); only the negative and run tests pass. His patterns match a clamp to the bounds of a narrower type feeding a truncation. The [0, 2^k - 1] clamps of the 9 to 14-bit pixel paths are not that shape. At -O3 on aarch64 the 10-bit loop is still cmhi/cmlt/bif on trunk, and becomes smax/smin with the patch (81 to 55 instructions for the function).
On top of that trunk the patch adds no aarch64 regressions (aarch64.exp, vect.exp, tree-ssa.exp, dg.exp), and Kyrylo's new tests all still pass. Thanks, Dominic
