This code was wrongly devirtualized before:
A* a = new A; a->foo(); A* b = new(a) B; if (a == b) b->foo();
Now we insert barrier before comparing dynamic pointers.
Diff Detail
Diff Detail
- Build Status
Buildable 5800 Build 5800: arc lint + arc unit
Event Timeline
| lib/CodeGen/CGExprScalar.cpp | ||
|---|---|---|
| 3066–3067 | I think we need to do this regardless of optimization level -- if we LTO together a -O0 translation unit with a -O2 translation unit, we still need this protection for the comparisons in the -O0 TU. (IIRC we chose to make -fstrict-vtable-pointers an IR-level ABI break, but we shouldn't do the same thing for optimization level.) | |
| lib/CodeGen/CGExprScalar.cpp | ||
|---|---|---|
| 3066–3067 | sounds good | |
Comment Actions
This is actually good catch, we also need to do it when inserting barrier in placement new.
I think that for the ctors and dtors we can do it only with optimizations enabled, because if optimizations are not enabled then ctors and dtors won't have the invariant.group in stores.
Comment Actions
Don't add barrier if compared with nullptr. With this it reduces added
barriers to compares by half. Note that we will transform barrier(null) -> barrier
in llvm
Comment Actions
Has it been discussed whether this is something to be addressed in the optimizer as opposed to the front-end?
| lib/CodeGen/CGExprScalar.cpp | ||
|---|---|---|
| 3069 | Consider: extern "C" int printf(const char *, ...);
void *operator new(decltype(sizeof 0), void *);
struct A {
virtual void foo() { printf("%s\n", __PRETTY_FUNCTION__); }
int m;
int *zip() { return &m; }
};
struct B : A {
virtual void foo() { printf("%s\n", __PRETTY_FUNCTION__); }
};
int main(void) {
A *ap = new A;
ap->foo();
int *const apz = ap->zip();
B *bp = new (ap) B;
if (apz == bp->zip()) {
bp->foo();
}
} | |
Comment Actions
The example that you showed is excellent. I didn't know that LLVM does the transformation with pointers and it clearly shows that we need the different approach.
I got following idea how to solve it in the middle end - drop all invariant.group metadata when we replace dominated uses based on the comparison. This works for simple cases, but it doesn't when the changed pointer
is passed to a not inlined function, or returned. The solution I see that does not satisfy me, is to put barriers between passed/returned pointer. I will bring it to mailing list
I think we need to do this regardless of optimization level -- if we LTO together a -O0 translation unit with a -O2 translation unit, we still need this protection for the comparisons in the -O0 TU.
(IIRC we chose to make -fstrict-vtable-pointers an IR-level ABI break, but we shouldn't do the same thing for optimization level.)