Auto merge of #123347 - saethlin:only-allow-upstream-llvm-calls, r=Ni…

…lstrieb

Only allow compiler_builtins to call LLVM intrinsics, not any link_name function

This is another case of accidental reliance on `inline(never)` like I rooted out in #118770. Without this PR, attempting to build some large programs with `-Zcross-crate-inline-threshold=yes` with a sysroot also compiled with that flag will result in linker errors like this:
```
  = note: /usr/bin/ld: /tmp/cargo-installNrfN4T/x86_64-unknown-linux-gnu/release/deps/libcompiler_builtins-d2a9b69f4e45b883.rlib(compiler_builtins-d2a9b69f4e45b883.compiler_builtins.dbbc6c2ca970faa4-cgu.0.rcgu.o): in function `core::panicking::panic_fmt':
          /home/ben/.rustup/toolchains/nightly-x86_64-unknown-linux-gnu/lib/rustlib/src/rust/library/core/src/panicking.rs:72:(.text.unlikely._ZN4core9panicking9panic_fmt17ha407cc99e97c942fE+0x31): undefined reference to `rust_begin_unwind'
```
With `-Zcross-crate-inline-threshold=yes` we can inline `panic_fmt` into `compiler_builtins`. Then we end up with a call to an upstream monomorphization, but one that has a `link_name` set. But unlike LLVM's magic intrinsic names, this link name is going to make it to the linker, and then we have a problem.

This logic looks scuffed, but also we're doing this in 4 other places. Don't know if that means it's good or bad.
https://github.com/rust-lang/rust/blob/1684a753dbca5d23b2e03568e6fbbb48eb72d0e6/compiler/rustc_codegen_cranelift/src/abi/mod.rs#L386
https://github.com/rust-lang/rust/blob/1684a753dbca5d23b2e03568e6fbbb48eb72d0e6/compiler/rustc_ast_passes/src/feature_gate.rs#L306
https://github.com/rust-lang/rust/blob/1684a753dbca5d23b2e03568e6fbbb48eb72d0e6/compiler/rustc_codegen_ssa/src/codegen_attrs.rs#L609
https://github.com/rust-lang/rust/blob/1684a753dbca5d23b2e03568e6fbbb48eb72d0e6/compiler/rustc_codegen_gcc/src/declare.rs#L170

compiler/rustc_monomorphize/src/lib.rs

@@ -14,6 +14,7 @@ use rustc_middle::ty::adjustment::CustomCoerceUnsized;
 use rustc_middle::ty::Instance;
 use rustc_middle::ty::TyCtxt;
 use rustc_middle::ty::{self, Ty};
+use rustc_span::def_id::DefId;
 use rustc_span::def_id::LOCAL_CRATE;
 use rustc_span::ErrorGuaranteed;
 
@@ -57,13 +58,24 @@ fn custom_coerce_unsize_info<'tcx>(
 /// linkers will optimize such that dead calls to unresolved symbols are not an error, but this is
 /// not guaranteed. So we used this function in codegen backends to ensure we do not generate any
 /// unlinkable calls.
+///
+/// Note that calls to LLVM intrinsics are uniquely okay because they won't make it to the linker.
 pub fn is_call_from_compiler_builtins_to_upstream_monomorphization<'tcx>(
     tcx: TyCtxt<'tcx>,
     instance: Instance<'tcx>,
 ) -> bool {
-    !instance.def_id().is_local()
+    fn is_llvm_intrinsic(tcx: TyCtxt<'_>, def_id: DefId) -> bool {
+        if let Some(name) = tcx.codegen_fn_attrs(def_id).link_name {
+            name.as_str().starts_with("llvm.")
+        } else {
+            false
+        }
+    }
+
+    let def_id = instance.def_id();
+    !def_id.is_local()
         && tcx.is_compiler_builtins(LOCAL_CRATE)
-        && tcx.codegen_fn_attrs(instance.def_id()).link_name.is_none()
+        && !is_llvm_intrinsic(tcx, def_id)
         && !should_codegen_locally(tcx, instance)
 }