https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116946
Bug ID: 116946
Summary: LTO gets confused about code path and incorrectly
triggers build-time assertion
Product: gcc
Version: 13.2.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: lto
Assignee: unassigned at gcc dot gnu.org
Reporter: jwerner at chromium dot org
Target Milestone: ---
We're using `__builtin_constant_p()` in combination with generating a function
call to an undefined reference as a build-time assertion mechanism. This works
perfectly fine without LTO (if the value is build-time constant it will go down
this path, if it is not it will generate a normal runtime `assert()` call).
When enabling LTO it seems that more values are `__builtin_constant_p()`, which
is fine, but we have found a test case where LTO seems to get confused about
the code flow and hits an assertion that the code wouldn't actually reach.
I've reduced the problem to a minimal test case across 3 files. Taking out any
more than this seems to make the issue disappear. (The original problem comes
from the coreboot project and this file:
https://github.com/coreboot/coreboot/blob/main/src/soc/mediatek/mt8186/mt6366.c
)
---------------------- test1.c -----------------------
int build_time_assertion_failed(void) __attribute__((noreturn));
#define build_time_assert(x) \
(__builtin_constant_p(x) ? ((x) ? 1 : build_time_assertion_failed()) :
0)
#define assert(x) { \
if (!build_time_assert(x) && !(x)) { \
do_thing(2); \
} \
}
static void do_thing(unsigned long x)
{
*(volatile unsigned long *)0x12345678ul = x;
}
static void not_a_problem(unsigned int value)
{
do_thing(value);
}
static void never_called(unsigned int value)
{
assert(value <= 3);
do_thing(value);
}
void inner_func(int id, unsigned int value)
{
switch (id) {
case 2:
never_called(value);
break;
case 3:
not_a_problem(value);
break;
}
}
---------------------- test2.c -----------------------
void wrapper_func(int id, unsigned int value);
void _start(void)
{
wrapper_func(3, 5);
wrapper_func(3, 5);
}
---------------------- test3.c -----------------------
void inner_func(int id, unsigned int value);
static int load_bearing_switch(int id)
{
switch (id) {
case 2:
return 2;
case 3:
return 3;
}
}
void wrapper_func(int id, unsigned int value)
{
inner_func(load_bearing_switch(id), value);
}
------------------------------------------------------
I am building this with an AArch64 GCC 13.2.0 cross-compiler like this:
aarch64-elf-gcc -o test1.o -c test1.c -Os -flto
aarch64-elf-gcc -o test2.o -c test2.c -Os -flto
aarch64-elf-gcc -o test3.o -c test3.c -Os -flto
aarch64-elf-gcc -ffreestanding -nostdlib -flto -o test test1.o test2.o test3.o
And getting this output:
aarch64-elf/bin/ld: /tmp/cc1rtFH6.ltrans0.ltrans.o: in function `inner_func':
<artificial>:(.text+0x38): undefined reference to `build_time_assertion_failed'
collect2: error: ld returned 1 exit status
As you can easily see from the code, the call graph should be
_start()
-> wrapper_func(3, 5)
-> inner_func(3, 5)
-> not_a_problem(5)
-> wrapper_func(3, 5)
-> inner_func(3, 5)
-> not_a_problem(5)
The never_called() function never gets called because the `id` constant passed
in from _start() is 3, not 2. Without LTO this builds fine because
__builtin_constant_p(value) evaluates to false. With LTO it seems that the
compiler recognizes that the value is a constant when passed in from _start(),
but it doesn't realize that `id` is also a constant and forces the switch
statement into a different path. (FWIW, if `id` was not a constant then this
would still be a problem. I think with LTO the semantics of
__builtin_constant_p(x) may need to be expanded to "x is compile-time constant
AND the compiler can guarantee that the code path from the point where x is
defined to the __builtin_constant_p() call is actually taken".)