build_regression_tests: 4 SIMD-AOT ba-issue cases crash with SIGILL on ubuntu-22.04 (pre-existing, masked by the LLVM-cache failure)

[matthargett]

Summary

Four tests/regression/ba-issues SIMD / v128 cases fail in build_regression_tests (ubuntu-22.04), all crashing with SIGILL (illegal instruction, exit -4) when the wamrc-generated .aot runs:

issue	expected	observed
#2833	<0x0…>:v128 ×7, ret 0	ret -4 (SIGILL)
#270801	<0xffffffff00000000 …>:v128 ×7, ret 0	ret -4 (SIGILL)
#270802	<0xffffffff00000000 …>:v128 ×7, ret 0	ret -4 (SIGILL)
#2702	<0x0000070700000707 …>:v128 ×7, ret 0	ret -4 (SIGILL)

Total: 106, Passed: 102, Failed: 4. Reproduction:

wamrc -o out.aot tests/regression/ba-issues/issues/issue-2833/filea12287_4.wasm
iwasm --heap-size=0 -f main out.aot      # -> SIGILL

Root cause: the AOT emits AVX-512 the runner can't execute

GitHub moved the ubuntu-22.04 image onto AMD EPYC 9V74 (Genoa) VMs whose /proc/cpuinfo advertises avx avx2 fma f16c sse4_2 vaes vpclmulqdq — no avx512* flag (AVX-512 is masked off in the VM).

But wamrc's default host-targeted AOT (no --target/--cpu) selects the target with LLVMGetHostCPUName() — which classifies Genoa as an AVX-512-capable model — and leaves the feature string empty, so LLVM enables the model-implied AVX-512. The emitted .aot is full of AVX-512 (vmovdqa64 %zmm0{%k1}{z}, vporq %zmm, vpternlogd, vextracti64x4) and SIGILLs on the first one:

Program received signal SIGILL, Illegal instruction.
=> 0x40006878: vcmpordpd %xmm1,%xmm0,%k1      # opmask destination -> AVX-512 (EVEX)

The JIT/ORC target-machine setup already pairs the host CPU name with LLVMGetHostCPUFeatures() (CPUID-accurate); the AOT default path does not, so it targets the CPU model rather than the ISA actually enabled on the VM.

Pre-existing and LLVM-version-independent — reproduces identically on #4951 (LLVM 22) and current main (LLVM 18). On main it is currently masked by a separate LLVM-cache CI break (fixed in #4996); the compilation on android, ubuntu-22.04 workflow last passed on 2026-05-19 (5aebe728), matching when the image moved onto these AMD VMs.

Fix — #4998

In the default native AOT path, also query LLVMGetHostCPUFeatures() so codegen is capped at the runner's actual CPUID feature set (AVX-512 masked off), mirroring the JIT setup. Explicit --cpu / --cpu-features / --target builds are unaffected.

Verified on an EPYC 9V74 runner in my rebeckerspecialties fork: with the fix there is no AVX-512 in the emitted object, and the AOT runs to the expected …:v128 output with exit 0 — no SIGILL.