Steven Johnson

At this point, all the buildbots should have rebuilt LLVM in place, so I think you have a real bug here

On linux `correctness_math`: ``` dlopenbuf failed undefined PLT symbol roundevenf (30) ./libhalide_kernels30.soterminate called after throwing an instance of 'Halide::RuntimeError' ```

still failures that seem like non-flakes :-/

> Seems to be wasm float16 related. That doesn't explain (say) correctness_vector_reductions failing on 32-bit Windows (where we never test wasm): https://buildbot.halide-lang.org/master/#/builders/7/builds/386 > Does wasm support float16 at all? Nope.

FYI: on OSX, vector_reductions eventually fails with `HalideJITMemoryManager: unable to find address for _roundevenf` (when running without sse41 or other simd)... unfortunately the error gets swallowed by "unhandled exception"

On osx-arm: ``` FAILED TEST: vector_reductions vector_reductions: Testing with target(arm-64-osx-jit) Intrinsic name not mangled correctly for type arguments! Should be: llvm.aarch64.neon.uaddlp.v1i64.v2i32 ptr @llvm.aarch64.neon.uaddlp.i64.v2i32 Internal Error at /Users/halidenightly/build_bot/worker/halide-testbranch-main-llvm16-arm-64-osx-make/halide-source/src/CodeGen_LLVM.cpp:645 triggered by user...

failing for `host-d3d12compute` :-/

> Only remaining failure is one of the preexisting flakes on the windows builbot with the dodgy GPU winbot1? If so I'm just take that one offline. It's stupid slow...

> Looks like win-worker-2. I noticed it flaking on AJ's PR as well: #6884 Urg, so we basically have winbot1, which is less flaky but stupid-slow, and winbot2 which is...

LGTM, let me pull into google3 for some torture testing