Report llvmcall of unknown intrinsics at compile time

[maleadt]

Julia's codegen replaces an llvmcall of an intrinsic it doesn't know (ccall("llvm.foo", llvmcall, ...)) with a run-time jl_error("llvmcall only supports intrinsic calls"). GPUCompiler then lowers that to an anonymous device-side exception. That's how LLVM 21 removing llvm.nvvm.barrier0 and friends (which Julia 1.14 hits, being on LLVM 22) showed up in CUDA.jl: every kernel using sync_threads threw ERROR: a error was thrown during kernel execution.

This detects that specific jl_error both in lower_throw! (GPU targets) and in check_ir (targets that keep runtime calls), and raises an InvalidIRError pointing at the offending call:

InvalidIRError: compiling MethodInstance for kernel() resulted in invalid LLVM IR
Reason: unsupported call to an unknown LLVM intrinsic
Stacktrace:
 [1] kernel

Tested on native and PTX, on Julia 1.10 and nightly.

[codecov]

Codecov Report

❌ Patch coverage is 62.06897% with 11 lines in your changes missing coverage. Please review.
✅ Project coverage is 85.48%. Comparing base (0d06ee4) to head (1664b32).
⚠️ Report is 3 commits behind head on main.

Files with missing lines	Patch %	Lines
src/validation.jl	56.00%	11 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #958      +/-   ##
==========================================
- Coverage   86.54%   85.48%   -1.06%     
==========================================
  Files          29       29              
  Lines        5729     5830     +101     
==========================================
+ Hits         4958     4984      +26     
- Misses        771      846      +75     

☔ 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.