https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97562
Bug ID: 97562
Summary: NRVO is very fragile (adding an extra scope breaks the
optimization)
Product: gcc
Version: 10.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c++
Assignee: unassigned at gcc dot gnu.org
Reporter: barry.revzin at gmail dot com
Target Milestone: ---
Consider the following (where T is just some type that is trivially copyable
but large enough):
struct T {
char _[24];
};
bool good_enough(T const&);
T get();
T as_loop() {
for (;;) {
T obj = get();
if (good_enough(obj)) {
return obj;
}
}
}
T as_goto() {
loop:
T obj = get();
if (good_enough(obj)) {
return obj;
}
goto loop;
}
The two functions, as_loop and as_goto, are semantically equivalent. Yet they
emit rather different code. With gcc 10.2 -O3 (https://godbolt.org/z/5dx3od),
as_loop is:
as_loop():
push r12
mov r12, rdi
sub rsp, 32
.L3:
mov rdi, rsp
call get()
mov rdi, rsp
call good_enough(T const&)
test al, al
je .L3
mov rax, QWORD PTR [rsp+16]
movdqu xmm0, XMMWORD PTR [rsp]
mov QWORD PTR [r12+16], rax
mov rax, r12
movups XMMWORD PTR [r12], xmm0
add rsp, 32
pop r12
ret
while as_goto is:
as_goto():
push r12
mov r12, rdi
.L8:
mov rdi, r12
call get()
mov rdi, r12
call good_enough(T const&)
test al, al
je .L8
mov rax, r12
pop r12
ret
That is, the goto implementation gives us NRVO, but the loop implementation
does not. This has something to do with the extra scope. Adding an extra set of
braces, as in:
T as_goto() {
{
loop:
T obj = get();
if (good_enough(obj)) {
return obj;
}
goto loop;
}
}
emits the same code as the loop example.
