Summary:
Pytorch unitest hangs when jitting the Tensor kernel. The problem exists for LLVM version >= 18 due to this upstream change: 45bb45f2ae
`IRBuilderBase::CreateCall` will insert the instruction into the BasicBlock by default. And we don't need to explicitly insert the instruction when compiling the tensor kernel.
Test Plan:
## Test with the release toolchain
```
buck test 'mode/dev' //caffe2/test:jit -- --exact 'caffe2/test:jit - test_concat_invariant (test_jit_fuser_te.TestTEFuserDynamic)'
```
## Test with the Buckified toolchain
Apply this D71046097 to select the LLVM libraries.
```
# Build tests
buck build 'mode/dev-asan' //caffe2/test:jit --show-output
```
```
# Run test (Change HASH and paths accordingly)
HASH="b755f1c435832a1e"
ENABLE_FLATBUFFER=0 FB_OVERRIDE_PYBIND11_GIL_INCREF_DECREF_CHECK=1 MKL_NUM_THREADS=1 NO_MULTIPROCESSING_SPAWN=0 OMP_NUM_THREADS=1 PYTORCH_TEST=1 PYTORCH_TEST_FBCODE=1 PYTORCH_TEST_WITH_ASAN=1 PYTORCH_TEST_WITH_DEV_DBG_ASAN=1 PYTORCH_TEST_WITH_TSAN=0 PYTORCH_TEST_WITH_UBSAN=1 SKIP_TEST_BOTTLENECK=1 TENSORPIPE_TLS_DATACENTER=test_dc TEST_PILOT=True TPX_IS_TEST_EXECUTION=true TPX_TIMEOUT_SEC=6000 \
buck-out/v2/gen/$HASH/caffe2/test/__jit__/jit.par --test-filter test_jit_fuser_te.TestTEFuserDynamic.test_concat_invariant
```
Differential Revision: D71046799
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149058
Approved by: https://github.com/dcci, https://github.com/Skylion007
e5fccb2bab