https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88509
Bug ID: 88509
Summary: Missing optimization of tls initialization
Product: gcc
Version: 8.2.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c++
Assignee: unassigned at gcc dot gnu.org
Reporter: rafael at espindo dot la
Target Milestone: ---
Given
struct foo {
foo();
};
static thread_local foo bar;
foo *f() { return &bar; }
foo *g() {
static thread_local foo *bar_ptr;
if (bar_ptr == nullptr) {
[&]() { bar_ptr = &bar; }();
}
return bar_ptr;
}
GCC has to make sure bar is only initialized once. For the function f it
produces
pushq %rbx
cmpb $0, %fs:__tls_guard@tpoff
movq %fs:0, %rbx
je .L6
leaq _ZL3bar@tpoff(%rbx), %rax
popq %rbx
ret
.L6:
<do_the_initialization>
for g, the common code path is somewhat simpler:
movq %fs:_ZZ1gvE7bar_ptr@tpoff, %rax
testq %rax, %rax
je .L15
ret
.L15:
<do_the_initialization>
The optimization is to use the a pointer to the object as a guard instead of
using a boolean. As far as I can tell this can be applied to any static tls
variable (where the ABI is not a problem).
