Missed tail call optimization in branch

[rockeet]

struct C {
    int x;
    //~C();
};
bool foo();
void bar();
C C_get_impl();
C C_get() {
    return C_get_impl(); // OK, both gcc & clang tail call
}
C C_get_cond() {
    if (__builtin_expect(foo(), 1)) {
        return C_get_impl(); // gcc tail call, clang missed
    } else {
        C r = C_get_impl();
        bar();
        return r;
    }
}
struct D {
    int x;
    ~D(); // has destructor
};
D D_get_impl();
D D_get() {
    return D_get_impl(); // clang tail call, gcc missed
}
D D_get_cond() {
    if (__builtin_expect(foo(), 1)) {
        return D_get_impl(); // both gcc & clang missed tail call
    } else {
        D r = D_get_impl();
        bar();
        return r;
    }
}

clang(trunk) generate code(https://godbolt.org/z/T7hxThzax):

C_get():
        jmp     C_get_impl()@PLT

C_get_cond():
        push    rbx
        call    foo()@PLT
        mov     ebx, eax
        call    C_get_impl()@PLT
        test    bl, bl
        je      .LBB1_1
        pop     rbx
        ret
.LBB1_1:
        mov     ebx, eax
        call    bar()@PLT
        mov     eax, ebx
        pop     rbx
        ret

D_get():
        jmp     D_get_impl()@PLT

D_get_cond():
        push    rbp
        push    r14
        push    rbx
        mov     rbx, rdi
        call    foo()@PLT
        mov     ebp, eax
        mov     rdi, rbx
        call    D_get_impl()@PLT
        test    bpl, bpl
        je      .LBB3_1
.LBB3_2:
        mov     rax, rbx
        pop     rbx
        pop     r14
        pop     rbp
        ret
.LBB3_1:
        call    bar()@PLT
        jmp     .LBB3_2
        mov     r14, rax
        mov     rdi, rbx
        call    D::~D()@PLT
        mov     rdi, r14
        call    _Unwind_Resume@PLT

DW.ref.__gxx_personality_v0:
        .quad   __gxx_personality_v0

[pinskia]

The ones where gcc and clang miss is recorded as #146303 . There is another one filed issue which talks about clang missing the condition one too (but it was mentioned something like after inlining).

[pinskia]

Oh yes #44936 too.

[rockeet]

The ones where gcc and clang miss is recorded as #146303 . There is another one filed issue which talks about clang missing the condition one too (but it was mentioned something like after inlining).

#146303 has been fixed in clang trunk, as in this issue clang tail call, gcc missed.

[rockeet]

gcc's tail call in branch is jmp "C_get_impl()", it should be better jne "C_get_impl()" without preceding je in clang, especially such case:

c++	gcc asm
struct C { int x; }; void bar(); C C_get_impl(); C C_get_cond2(bool cond) { if (__builtin_expect(cond, 1)) { // gcc tail call, clang missed return C_get_impl(); } else { C r = C_get_impl(); bar(); return r; } }	"C_get_cond2(bool)": test dil, dil je .L7 jmp "C_get_impl()" .L7: sub rsp, 24 call "C_get_impl()" mov DWORD PTR [rsp+12], eax call "bar()" mov eax, DWORD PTR [rsp+12] add rsp, 24 ret

[antoniofrighetto]

I think CodeGenPrepare should be taught to duplicate the call to C_get_impl in order to have the call in tail position: https://alive2.llvm.org/ce/z/v6DdXZ. Possibly to be limited to a single call, i.e., not always profitable.