s-celles (same AI-assisted metamorphic fuzzer as Monday's batch) reports that the x86 ops SBBQcarrymask/SBBLcarrymask, which the compiler uses to saturate unbounded shift counts, expand to an SBB instruction that writes the flags register but are missing clobberFlags: true. When the program also tests the shift count against the width, flagalloc reuses the CMPQ $64 flags for the SETcc, and the scheduler can place the SBB in between. The reproducer r := x << y; if y == 64 { eq = 1 }; return r + eq returns 1 for f(1, 65) instead of 0. Other affected shapes on amd64: y != 64 (compiled as y < 64), signed and unsigned x >> y with y == 64, and the 32-bit x << y with y == 32; 386 also miscompiles a decomposed 64-bit comparison. GOAMD64 v1/v2/v3 all behave the same, and source order matters (it's correct when the SETcc is scheduled first). The report says compiler-inserted bounds checks are not affected and no out-of-bounds access was seen; the impact is wrong boolean results. Fix CL 858262 ("mark the x86 SBB carry-mask ops as clobbering flags") is in review, and Jorropo opened 1.26 and 1.27 backports. The same reporter filed #82834 (arm bits.Sub64 wrong borrow, root cause shared with #82786) and #82839 (prove takes an impossible branch through a sign extension).
--late lowering campaign.