missing tail call optimization on RVO

[rockeet]

struct Slice {
    const char* data;
    unsigned long size;
    ~Slice();
};
Slice get_s_impl();
Slice get_s() { return get_s_impl(); }

clang produces:

get_s():
        push    rbx
        mov     rbx, rdi
        call    get_s_impl()
        mov     rax, rbx
        pop     rbx
        ret

icc recognized tail call and produces:

get_s():
..B1.1:                         # Preds ..B1.0
        jmp       get_s_impl()                               #7.24

g++ also missed such tail call optimization.

If delete ~Slice;, g++ & clang recognize the tail call, but icc fails.

[thesamesam]

g++ also missed such tail call optimization.

https://gcc.gnu.org/PR120871

[aengelke]

This is a call lowering/ISel problem with sret. On AArch64, there is a tail call. On x86, the problem is that the parameter needs to be returned and we don't identify that callee will return the same value.

define void @a(ptr noalias sret(i64) %r) {
  tail call void @b(ptr sret(i64) %r)
  ret void
}

declare void @b(ptr sret(i64))