[clang-17 regression] clang from trunk crashes inside clang::TemplateArgument::pack_size()

[loskutov]

The following piece of code leads to ICE when compiled with clang from trunk:

template<bool... Vals>
constexpr bool All = (Vals && ...);

template<bool... Bs>
class Class {
    template<typename>
    requires All<Bs...>
    void Foo();
};

template<bool... Bs>
template<typename>
requires All<Bs...>
void Class<Bs...>::Foo() {
};

Compiler Explorer with full stacktrace: https://godbolt.org/z/n6bro9Tcd

[llvmbot]

@llvm/issue-subscribers-clang-frontend

[shafik]

@Fznamznon any chance your recent fix could be involved here?

[Fznamznon]

@Fznamznon any chance your recent fix could be involved here?

I don't think so. I'm not seeing any MemberExprs created (the patch tackles only them), also reverting the patch locally and seeing the same assertion failure seems pretty assuring.

[loskutov]

According to git-bisect, 6db007a authored by @alexander-shaposhnikov is the first bad commit.

The properly symbolized stacktrace: https://pastebin.ubuntu.com/p/q3s55j3tNC/

[michael-jabbour-sonarsource]

Another crashing example that results in a similar stacktrace:

using size_t = decltype(sizeof(int));

template<typename... Ts> struct S {
  template<size_t N>
  requires(N <= sizeof...(Ts))
  void f(int);
};

template<typename... Ts>
template<size_t N>
requires(N <= sizeof...(Ts))
void S<Ts...>::f(int) {}

This also bisects to 6db007a. Compiler Explorer link: https://godbolt.org/z/s9YvEG5rn.

[shafik]

CC @alexander-shaposhnikov @erichkeane

[erichkeane]

I really hope @alexander-shaposhnikov can take a look at this, I've been away from this for long enough, and I likely wont have a chance to look at it 'soon'.

[alexander-shaposhnikov]

Hi, I'll try to look into this in the coming weeks, but I'm really busy with completely unrelated things at the moment / can't promise.

[erichkeane]

I spent a little time that I had to spare in the debugger with this one, but I'm a touch lost. The function in question is: https://clang.llvm.org/doxygen/classclang_1_1Sema.html#a0b0f6c317ec8c9f08dcb406114c43644

The issue is at line 732: https://clang.llvm.org/doxygen/SemaTemplateVariadic_8cpp_source.html#l00732

In @michael-jabbour-sonarsource 's example, we're assuming the argument (the only one, at 0,0) is a pack, it is NOT. It is a PackExpansionType (Ts...), which SEEMS right, but I have very little idea what is going on in this function, or why a PackExpansionType is not sufficiently 'pack like' here to work? I'm hopeful @alexander-shaposhnikov has some luck on this, but my time is pretty limited unfortunately right now as well.

[erichkeane]

Further digging shows that just handling the PackExpansionType isn't sufficient. I suspect Alexander's patch ended up putting the pack into the MLTAL in a way that it shouldn't have, but I don't see how/why.

[erichkeane]

The issue is here... SOMEWHERE: 6db007a#diff-dae8060cd0876af2d1e1376f5da83d22a01fd1b49730cee2deaee9727e64e0e4R221

[alejandro-alvarez-sonarsource]

Hello.
I believe this was fixed by #80594. clang trunk with assertions does not crash anymore with the given example: https://godbolt.org/z/1TnqxbE1E

[shafik]

Looks fixed in trunk