[AMDGPU][GlobalISel][NFC] Filter getCandidatesForLowering to only process G_PHI (#181814)HEADmain

The divergence-lowering pass runs in three stages:
1. lowerTemporalDivergence()
2. lowerTemporalDivergenceI1() — creates PHI instructions via SSAUpdater
3. lowerPhis() — lowers divergent i1 G_PHIs into lane mask PHIs

getCandidatesForLowering() in stage 3 iterates over all PHIs via
MBB.phis() and checks isDivergent() on their destination registers. This
includes PHI instructions already created by stage 2, whose registers
were never part of the original uniformity analysis.

Today, this works by accident — isDivergent() returns false for unknown
registers, so these PHIs are skipped. But this relies on isDivergent()
behavior for unanalyzed registers, which is not guaranteed. If
isDivergent() were to return true for unknown registers (e.g., as a
conservative default), lowerPhis() would re-process already-lowered
PHIs, adding redundant COPY instructions(EX:
https://github.com/llvm/llvm-project/pull/180509). This would not be a
correctness issue, but would produce unnecessary code.

Co-authored-by: padivedi <padivedi@amd.com>

0 files changed, 0 insertions, 0 deletions