Exhaustive stack usage

[apolukhin]

Bugzilla Link	44253
Version	trunk
OS	Linux
CC	@apolukhin,@davidbolvansky,@dwblaikie,@efriedma-quic,@francisvm,@zygoloid,@rjmccall

Extended Description

Consider the following example:

struct test {
    test(test&&);

    void Extend(test v);
    int data_[4096];
};


test test_func(test t) {
    t.Extend(static_cast<test&&>(t));
    t.Extend(static_cast<test&&>(t));
    return t;
}

With -std=c++11 -O2 flags Clang uses 32768 bytes of stack, GCC uses 16392 bytes.

Note that adding additional t.Extend(static_cast<test&&>(t)); results in stack usage growth for Clang, while GCC keeps using only 16392 bytes.

Godbolt playground: https://godbolt.org/z/MdaNfa

[efriedma-quic]

Yes, lifetime markers are missing.

[rjmccall]

Yeah, I think it's a known issue that we don't emit them for temporaries.

[francisvm]

Adding lifetime markers for the temp from EmitAnyExprToTemp(E) in EmitCallArg brings the stack size to 16424.

[rjmccall]

That should be a fairly localized change that doesn't need a solution to the more general issue with temporaries. Did you have a patch for it, or did you just try adding lifetime markers to the IR?

[apolukhin]

The issue affects users of the stackful coroutines by adding very significant memory overhead (each of the thousands coroutines becomes multiple times bigger).

We'd appreciate a fix for 10.0.

[davidbolvansky]

Seems like Francis has a patch?

[francisvm]

Sorry about the delay. I had a quick patch but it seems like Erik addressed this in https://reviews.llvm.org/D74094. I see that this has been reverted due to a ppc failure. I locally re-applied the patch and I see the function using only 16424 bytes.

[nickdesaulniers]

I've revived https://reviews.llvm.org/D74094; let's see what issues with that approach may persist, if any.

[nickdesaulniers]

relanded as https://reviews.llvm.org/rGe698695fbbf62e6676f8907665187f2d2c4d814b in clang-18.