[pytorch/tensorexpr] Update use of LLJIT::lookup for LLVM 15
`lookup` will return an address, rather than a symbol, in LLVM 15.
For consistency with the old APIs we can just wrap the address in a fake symbol
and return that, since we only care about the address anyways.
Differential Revision: [D36282406](https://our.internmc.facebook.com/intern/diff/D36282406/)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/77169
Approved by: https://github.com/jfix71, https://github.com/ZolotukhinM
diff --git a/torch/csrc/jit/tensorexpr/llvm_jit.cpp b/torch/csrc/jit/tensorexpr/llvm_jit.cpp
index c8d5c4a..a2e576d 100644
--- a/torch/csrc/jit/tensorexpr/llvm_jit.cpp
+++ b/torch/csrc/jit/tensorexpr/llvm_jit.cpp
@@ -171,7 +171,16 @@
}
JITSymbol findSymbol(const std::string Name) {
+#if LLVM_VERSION_MAJOR >= 15
+ // Starting with llvm-15, LLJIT::lookup returns an address rather than a
+ // symbol. Even though an address is what we ultimately we want, we also
+ // want to avoid churning our internal APIs, so we wrap the returned address
+ // in a fake JITSymbol.
+ auto result = assertSuccess(LLJ->lookup(Name));
+ return JITSymbol(result.getValue(), JITSymbolFlags());
+#else
return assertSuccess(LLJ->lookup(Name));
+#endif
}
bool hasSymbol(const std::string& Name) {
