rustc beta (1.28.0) generates unnecessary call instruction

[tspiteri]

For this code:

pub struct A([i32; 4]);
impl A {
    pub fn foo(self) -> Self { self }
    pub fn bar(self) -> Self { self }
}

rustc beta (1.28.0) with opt-level=3 generates:

example::A::bar:
        movups  xmm0, xmmword ptr [rsi]
        movups  xmmword ptr [rdi], xmm0
        mov     rax, rdi
        ret

example::A::foo:
        push    rbx
        mov     rbx, rdi
        call    example::A::bar@PLT
        mov     rax, rbx
        pop     rbx
        ret

rustc 1.27.1 and rustc beta with opt-level=s generates:

example::A::foo:
        movups  xmm0, xmmword ptr [rsi]
        movups  xmmword ptr [rdi], xmm0
        mov     rax, rdi
        ret

example::A::bar:
        movups  xmm0, xmmword ptr [rsi]
        movups  xmmword ptr [rdi], xmm0
        mov     rax, rdi
        ret

The beta with opt-level=3 seems to be anti-inlining the identical functions.

[Mark-Simulacrum]

cc @rust-lang/compiler

(I've not attempted a reproduction locally)

[hanna-kruppe]

This might be due to #49479, the time frame seems right and the code change observed here is precisely what MergeFunctions does. Some observations:

• it's counter intuitive that this doesn't happen with opt-level=s
• it's silly that it doesn't generate a tail call, wonder why that happens
• if we decided that distinct source-level functions are allowed to have identical addresses and told LLVM about it with the unnamed_addr attribute, MergeFunctions could turn one function into a link-time alias for the other instead of creating a thunk that calls

[hanna-kruppe]

cc @nox

[eddyb]

@rkruppe Huh, I thought we used unnamed_addr already, and didn't make any promises?

[hanna-kruppe]

@eddyb Evidently we're not using it. I don't know of any reason to make promises here, but since we de facto don't merge addresses, who knows whether anyone in the wild is relying on it for crazy shenanigans 🤷‍♂️

[dshynkev]

Interestingly, the LLVM IR the sample code results in appears to be using both unnamed_addr and tail_call, but the assembly is the same as in the parent post.

define void @bar(%A* noalias nocapture sret dereferenceable(16), %A* noalias nocapture readonly dereferenceable(16) %self) unnamed_addr #0 {
 ; ...
}
; ...
define void @foo(%A* noalias nocapture sret dereferenceable(16), %A* noalias nocapture readonly dereferenceable(16)) unnamed_addr #0 {
  tail call void @bar(%A* noalias nocapture sret dereferenceable(16) %0, %A* noalias nocapture readonly dereferenceable(16) %1) #0
  ret void
}