aarch64-pc-windows: failed to perform tail call elimination on a call site marked musttail

[zmodem]

Consider:

$ cat /tmp/a.ii
struct S {};

struct T {
  S Foo(S);
  S Get(S);
};

S T::Get(S key) {
  [[clang::musttail]] return Foo(key);
}

$ clang -cc1 -triple aarch64-pc-windows-msvc19.34.0 -emit-obj /tmp/a.ii
fatal error: error in backend: failed to perform tail call elimination on a call site marked musttail

(Other aarch64 triples, such as aarch64-unknown-unknown, work.)

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

Author: Hans Wennborg (zmodem)

Consider:

$ cat /tmp/a.ii
struct S {};

struct T {
  S Foo(S);
  S Get(S);
};

S T::Get(S key) {
  [[clang::musttail]] return Foo(key);
}

$ clang -cc1 -triple aarch64-pc-windows-msvc19.34.0 -emit-obj /tmp/a.ii
fatal error: error in backend: failed to perform tail call elimination on a call site marked musttail

(Other aarch64 triples, such as aarch64-unknown-unknown, work.)

[zmodem]

Slightly shorter: (dropped the arguments)

$ cat /tmp/a.ii
struct S {};

struct T {
  S Foo();
  S Get();
};

S T::Get() {
  [[clang::musttail]] return Foo();
}

[zmodem]

This is where we bail out: https://github.com/llvm/llvm-project/blob/b9b9addae6c79c72394c8260106f8ba5281e9747/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp#L8716-L8723

That was added in 5dc8aeb

[zmodem]

(Mostly notes to self.)

I think checking that both the caller and callee have the inreg attribute on the same argument is not enough; the values of the inreg arguments also have to be the same.

For example, this would not work:

target triple = "aarch64-pc-windows-msvc19.34.0"
  
%struct.S = type { i8 }
declare void @bar(ptr, ptr inreg sret(%struct.S))

define void @foo(ptr %a, ptr inreg sret(%struct.S) %b) {
  musttail call void @bar(ptr %b, ptr inreg sret(%struct.S) %a)
  ret void
}

@foo is supposed to return %b (in x0), but if it tail calls @bar the returned value will be%a instead.