[maleadt]

Fixes JuliaGPU/Metal.jl#1127.

lower_unreachable_control_flow! turns every exception into a branch to the kernel's return, so each bounds check becomes an exit of every loop around it. Apple's back-end has to structurize control flow, and with enough such exits it hangs in MTLCompilerService or silently skips code. On an M1 (macOS 27, Metal.jl 1.12.0):

• #1127's mwe.jl compiles for K=14 and hangs for K=15 and 16;
• mwe_wrong.jl skips the final store for K=4 and 6, and hangs for K=8;
• a kernel with 14 unrolled in-bounds loads throws a spurious BoundsError;
• AcceleratedKernels' columns scan through a strided view hangs under --check-bounds=yes on Julia 1.10, which is what turned AcceleratedKernels.jl#199's Metal CI red.

LLVM's AMDGPU pipeline runs FixIrreducible and UnifyLoopExits before StructurizeCFG for the same reason. This runs them at the end of lower_air!. It has to come after the last SimplifyCFG, which would merge the exit block back into separate exits. Two LLVM bugs need working around:

• Before LLVM 24, both passes only handle branch terminators (switch support is [FixIrreducible][UnifyLoopExits] Support switch statements llvm/llvm-project#206567, unreleased), so switches are lowered first. Without that, Metal.jl's merge-sort kernels crash LLVM.
• LLVM 20–22 crash on a loop without exits ([llvm][UnifyLoopExits] Avoid optimization if no exit block is found llvm/llvm-project#165343). There, functions with blocks that have no path to a return are left as they are. The check runs after LowerSwitch, which can remove a loop's only exit.

With the change, on the same M1:

• mwe.jl K=15/16 and mwe_wrong.jl K=4/6/8 compile and give correct results;
• exceptions are still reported, and only the threads that fail skip the rest of the kernel;
• the AcceleratedKernels tests that hung pass.

Run time of bounds-checked loop kernels changes by up to 1.5%: a grid-stride sum takes 1639–1650 µs instead of 1635–1637 µs, and 8 unrolled checked loads per step take 1388 µs instead of 1367 µs (median of 50, three runs). Compile time is unchanged.

Metal.jl's test suite passes, except relocations, which fails the same way without this change.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 91.90%. Comparing base (38a5625) to head (548f540).
⚠️ Report is 2 commits behind head on main.

Additional details and impacted files

@@             Coverage Diff             @@
##             main    #1079       +/-   ##
===========================================
+ Coverage   74.85%   91.90%   +17.04%     
===========================================
  Files          31       31               
  Lines        7490     7731      +241     
===========================================
+ Hits         5607     7105     +1498     
+ Misses       1883      626     -1257     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.