[clang][coroutines] Run-time crash with optimization when using coroutine with co_await

[dgg5503]

Summary

Starting with Clang 17 at commit 54225c4, attempting to execute the provided reduced programs compiled with clang using -O2 results in a crash at runtime.

Specifically, the crash occurs in the function void bx::await_suspend when the result of q, a null pointer, is dereferenced to call x().

I have confirmed that both reproducers run without triggering sanitizers in my environment (address, memory, undefined, etc.).

Reproducers

reduced-Version-A.cpp -- https://godbolt.org/z/bxEf8Tfc6

#include <coroutine>
#include <memory>

struct br;
struct bs {
  bs(std::coroutine_handle<br>);
  int x() { return bt; }
  void bu() { bt = 0; }
  void t() { bv.resume(); }
  std::coroutine_handle<br> bv;
  int bt;
};
struct bw {
  using promise_type = br;
  bw(std::coroutine_handle<promise_type> h) : v(make_shared<bs>(h)) {}
  void t() { v->t(); }
  std::shared_ptr<bs> r() { return v; }
  std::shared_ptr<bs> v;
};
struct bx {
  int await_ready() { return 0; }
  void await_resume() {}
  void await_suspend(std::coroutine_handle<br>);
  bw by;
};
struct br {
  auto initial_suspend() { return std::suspend_always(); }
  auto final_suspend() noexcept { return std::suspend_always(); }
  auto get_return_object() {
    return std::coroutine_handle<br>::from_promise(*this);
  }
  void unhandled_exception() {}
  auto await_transform(bw h) { return bx(h); }
  void return_void() {}
  void u(bs *h) { v = h; }
  bs *r() { return v; }
  bs *v;
};
void bx::await_suspend(std::coroutine_handle<br> h) {
  auto bi = h.promise();
  auto q = bi.r();
  auto s = by.r();
  if (q->x())
    s->bu();
  by.t();
}
bs::bs(std::coroutine_handle<br> h) {
  bt = 0;
  bv = h;
  auto &bz = h.promise();
  bz.u(this);
}
bw ca() {
  fprintf(stderr, "ca (co_return) called\n");
  co_return;
}
bw cb() {
  fprintf(stderr, "cb (co_await) called\n");
  co_await ca();
}
bw cc = cb();
int main() {
  bw cd(cc);
  cd.t();
}

Here is the same reproducer reduced including headers from glibc 2.35 / glibcxx 3.4.30
reduced-Version-B.cpp -- https://godbolt.org/z/fsaKof3EE

namespace std {
template <int a> struct b {
  struct c {
    int d[a];
  };
};
inline namespace {
template <typename e> struct coroutine_traits : e {};
template <typename = void> struct coroutine_handle;
template <> struct coroutine_handle<> {};
template <typename f> struct coroutine_handle {
  static coroutine_handle g(f &h) {
    coroutine_handle i;
    i.k = __builtin_coro_promise(&h, 0, 1);
    return i;
  }
  static coroutine_handle from_address(void *h) {
    coroutine_handle i;
    i.k = h;
    return i;
  }
  coroutine_handle<> j;
  operator coroutine_handle<>() { return j; }
  void l() { __builtin_coro_resume(k); }
  f &m() {
    void *aa = __builtin_coro_promise(k, 0, 0);
    return *static_cast<f *>(aa);
  }
  void *k;
};
struct ab {
  int await_ready() noexcept { return 0; }
  void await_suspend(coroutine_handle<>) noexcept {}
  void await_resume() noexcept {}
};
} // namespace
template <typename, typename> struct n;
template <template <typename...> class w, typename y, typename ac,
          typename... ad>
struct n<w<ac, ad...>, y> {
  using c = w<y>;
};
} // namespace std
void *operator new(unsigned long, void *);
template <typename ac, typename... ae> void af(ac *h, ae... o) {
  new (h) ac(o...);
}
template <typename ac> struct ag {
  ac *ah(int) { return static_cast<ac *>(operator new(sizeof(ac))); }
};
namespace std {
template <typename ai, typename y> using aj = n<ai, y>::c;
template <typename> struct p;
template <typename ac> struct p<ag<ac>> {
  using ak = ag<ac>;
  using al = ac;
  using am = ac *;
  using an = int;
  static am ah(ak h, an) { return h.ah(0); }
  template <typename y, typename... ae> static void ao(ak, y o, ae... ap) {
    af(o, ap...);
  }
};
} // namespace std
namespace std {
template <typename ai> struct aq {
  using am = p<ai>::am;
  using al = p<ai>::al;
  aq(ai, am o) : ar(o) {}
  al *as() { return ar; }
  am ar;
};
template <typename ai> aq<ai> at(ai h) { return {h, p<ai>::ah(h, 0)}; }
} // namespace std
template <typename ac> struct au {
  std::b<sizeof(ac)>::c av;
  ac *ar() {
    void *j = &av;
    return static_cast<ac *>(j);
  }
};
namespace std {
template <typename> struct aw;
template <typename ac> struct ax {
  ax(ac) {}
};
template <typename ai> struct ay {
  ai az;
};
template <typename ac, typename ai> struct ba {
  struct bb : ax<ai> {
    au<ac> av;
  };
  using bc = aj<ai, ba>;
  template <typename... ae> ba(ai h, ae... o) : bd(h) {
    p<ai>::ao(h, ar(), o...);
  }
  ac *ar() { return bd.av.ar(); }
  bb bd;
};
struct be {
  template <typename ac, typename ai, typename... ae>
  be(ac *&h, ay<ai> o, ae... ap) {
    typedef ba<ac, ai> bf;
    typename bf::bc bg;
    auto bh = at(bg);
    bf *bj = bh.as();
    auto bk = new (bj) bf(o.az, ap...);
    h = bk->ar();
  }
};
template <typename ac> struct z {
  using bl = ac;
  bl *operator->() {
    bl *bm = static_cast<aw<ac> *>(this)->as();
    return bm;
  }
};
template <typename ac> struct aw : z<ac> {
  using bl = ac;
  bl *as() { return ar; }
  template <typename ai, typename... ae> aw(ai h, ae... o) : bn(ar, h, o...) {}
  bl *ar;
  be bn;
};
template <typename ac> struct bo : aw<ac> {
  template <typename ai, typename... ae> bo(ai h, ae... o) : aw<ac>(h, o...) {}
};
template <typename ac, typename ai, typename... ae> bo<ac> bp(ai h, ae... o) {
  return bo<ac>(ay<ai>{h}, o...);
}
template <typename ac, typename... ae> bo<ac> bq(ae... h) {
  return bp<ac>(ag<int>(), h...);
}
} // namespace std
struct br;
struct bs {
  bs(std::coroutine_handle<br>);
  int x() { return bt; }
  void bu() { bt = 0; }
  void t() { bv.l(); }
  std::coroutine_handle<br> bv;
  int bt;
};
struct bw {
  using promise_type = br;
  bw(std::coroutine_handle<promise_type> h) : v(bq<bs>(h)) {}
  void t() { v->t(); }
  std::bo<bs> r() { return v; }
  std::bo<bs> v;
};
struct bx {
  int await_ready() { return 0; }
  void await_resume() {}
  void await_suspend(std::coroutine_handle<br>);
  bw by;
};
struct br {
  auto initial_suspend() { return std::ab(); }
  auto final_suspend() noexcept { return std::ab(); }
  auto get_return_object() { return std::coroutine_handle<br>::g(*this); }
  void unhandled_exception() {}
  auto await_transform(bw h) { return bx(h); }
  void return_void() {}
  void u(bs *h) { v = h; }
  bs *r() { return v; }
  bs *v;
};

// When reduced including headers, the following is required under
// clang -O2 @ 54225c457a336b1609c6d064b2b606a9238a28b9, otherwise this
// entire function is optimized out which masks the problem since the crash
// occurs at `q->x()` (q is nullptr).
#pragma clang optimize off
void bx::await_suspend(std::coroutine_handle<br> h) {
  auto bi = h.m();
  auto q = bi.r();
  auto s = by.r();
  if (q->x())
    s->bu();
  by.t();
}
#pragma clang optimize on

bs::bs(std::coroutine_handle<br> h) {
  bt = 0;
  bv = h;
  auto &bz = h.m();
  bz.u(this);
}
bw ca() { co_return; }
bw cb() { co_await ca(); }
bw cc = cb();
int main() {
  bw cd(cc);
  cd.t();
}

Reproduction Steps

Latest Clang -- Crashes

$ clang++ --version
clang version 20.0.0git (https://github.com/llvm/llvm-project.git 381a803da253b75c8b7b10bb732e9e90925185e8)
Target: x86_64-unknown-linux-gnu
Thread model: posix
InstalledDir: <redacted>
$ clang++ -O0 -std=c++20 reduced-Version-A.cpp -o reduced-Version-A.out
$ ./reduced-Version-A.out
cb (co_await) called
ca (co_return) called
$ clang++ -O2 -std=c++20 reduced-Version-A.cpp -o reduced-Version-A.out
$ ./reduced-Version-A.out
cb (co_await) called
Segmentation fault (core dumped)
$ clang++ -O0 -std=c++20 reduced-Version-B.cpp -o reduced-Version-B.out
$ ./reduced-Version-B.out
$ clang++ -O2 -std=c++20 reduced-Version-B.cpp -o reduced-Version-B.out
$ ./reduced-Version-B.out
Segmentation fault (core dumped)

Clang At Bisected Commit -- Crashes

$ clang++ --version
clang version 17.0.0 (https://github.com/llvm/llvm-project.git 54225c457a336b1609c6d064b2b606a9238a28b9)
Target: x86_64-unknown-linux-gnu
Thread model: posix
InstalledDir: <redacted>
$ clang++ -O0 -std=c++20 reduced-Version-A.cpp -o reduced-Version-A.out
$ ./reduced-Version-A.out
cb (co_await) called
ca (co_return) called
$ clang++ -O2 -std=c++20 reduced-Version-A.cpp -o reduced-Version-A.out
$ ./reduced-Version-A.out
cb (co_await) called
Segmentation fault (core dumped)
$ clang++ -O0 -std=c++20 reduced-Version-B.cpp -o reduced-Version-B.out
$ ./reduced-Version-B.out
$ clang++ -O2 -std=c++20 reduced-Version-B.cpp -o reduced-Version-B.out
$ ./reduced-Version-B.out
Segmentation fault (core dumped)

Clang Before Bisected Commit -- Does not crash

$ clang++ --version
clang version 17.0.0 (https://github.com/llvm/llvm-project.git 32be3405f57f1e4d0ec0da943434113450583e89)
Target: x86_64-unknown-linux-gnu
Thread model: posix
InstalledDir: <redacted>
$ clang++ -O0 -std=c++20 reduced-Version-A.cpp -o reduced-Version-A.out
$ ./reduced-Version-A.out
cb (co_await) called
ca (co_return) called
$ clang++ -O2 -std=c++20 reduced-Version-A.cpp -o reduced-Version-A.out
$ ./reduced-Version-A.out
cb (co_await) called
ca (co_return) called
$ clang++ -O0 -std=c++20 reduced-Version-B.cpp -o reduced-Version-B.out
$ ./reduced-Version-B.out
$ clang++ -O2 -std=c++20 reduced-Version-B.cpp -o reduced-Version-B.out
$ ./reduced-Version-B.out

[llvmbot]

@llvm/issue-subscribers-coroutines

Author: Douglas (dgg5503)

## Summary Starting with Clang 17 at commit 54225c4, attempting to execute the provided reduced programs compiled with `clang` using `-O2` results in a crash at runtime.

Specifically, the crash occurs in the function void bx::await_suspend when the result of q, a null pointer, is dereferenced to call x().

I have confirmed that both reproducers run without triggering sanitizers in my environment (address, memory, undefined, etc.).

Reproducers

reduced-Version-A.cpp -- https://godbolt.org/z/bxEf8Tfc6

#include <coroutine>
#include <memory>

struct br;
struct bs {
  bs(std::coroutine_handle<br>);
  int x() { return bt; }
  void bu() { bt = 0; }
  void t() { bv.resume(); }
  std::coroutine_handle<br> bv;
  int bt;
};
struct bw {
  using promise_type = br;
  bw(std::coroutine_handle<promise_type> h) : v(make_shared<bs>(h)) {}
  void t() { v->t(); }
  std::shared_ptr<bs> r() { return v; }
  std::shared_ptr<bs> v;
};
struct bx {
  int await_ready() { return 0; }
  void await_resume() {}
  void await_suspend(std::coroutine_handle<br>);
  bw by;
};
struct br {
  auto initial_suspend() { return std::suspend_always(); }
  auto final_suspend() noexcept { return std::suspend_always(); }
  auto get_return_object() {
    return std::coroutine_handle<br>::from_promise(*this);
  }
  void unhandled_exception() {}
  auto await_transform(bw h) { return bx(h); }
  void return_void() {}
  void u(bs *h) { v = h; }
  bs *r() { return v; }
  bs *v;
};
void bx::await_suspend(std::coroutine_handle<br> h) {
  auto bi = h.promise();
  auto q = bi.r();
  auto s = by.r();
  if (q->x())
    s->bu();
  by.t();
}
bs::bs(std::coroutine_handle<br> h) {
  bt = 0;
  bv = h;
  auto &bz = h.promise();
  bz.u(this);
}
bw ca() {
  fprintf(stderr, "ca (co_return) called\n");
  co_return;
}
bw cb() {
  fprintf(stderr, "cb (co_await) called\n");
  co_await ca();
}
bw cc = cb();
int main() {
  bw cd(cc);
  cd.t();
}

Here is the same reproducer reduced including headers from glibc 2.35 / glibcxx 3.4.30
reduced-Version-B.cpp -- https://godbolt.org/z/fsaKof3EE

namespace std {
template <int a> struct b {
  struct c {
    int d[a];
  };
};
inline namespace {
template <typename e> struct coroutine_traits : e {};
template <typename = void> struct coroutine_handle;
template <> struct coroutine_handle<> {};
template <typename f> struct coroutine_handle {
  static coroutine_handle g(f &h) {
    coroutine_handle i;
    i.k = __builtin_coro_promise(&h, 0, 1);
    return i;
  }
  static coroutine_handle from_address(void *h) {
    coroutine_handle i;
    i.k = h;
    return i;
  }
  coroutine_handle<> j;
  operator coroutine_handle<>() { return j; }
  void l() { __builtin_coro_resume(k); }
  f &m() {
    void *aa = __builtin_coro_promise(k, 0, 0);
    return *static_cast<f *>(aa);
  }
  void *k;
};
struct ab {
  int await_ready() noexcept { return 0; }
  void await_suspend(coroutine_handle<>) noexcept {}
  void await_resume() noexcept {}
};
} // namespace
template <typename, typename> struct n;
template <template <typename...> class w, typename y, typename ac,
          typename... ad>
struct n<w<ac, ad...>, y> {
  using c = w<y>;
};
} // namespace std
void *operator new(unsigned long, void *);
template <typename ac, typename... ae> void af(ac *h, ae... o) {
  new (h) ac(o...);
}
template <typename ac> struct ag {
  ac *ah(int) { return static_cast<ac *>(operator new(sizeof(ac))); }
};
namespace std {
template <typename ai, typename y> using aj = n<ai, y>::c;
template <typename> struct p;
template <typename ac> struct p<ag<ac>> {
  using ak = ag<ac>;
  using al = ac;
  using am = ac *;
  using an = int;
  static am ah(ak h, an) { return h.ah(0); }
  template <typename y, typename... ae> static void ao(ak, y o, ae... ap) {
    af(o, ap...);
  }
};
} // namespace std
namespace std {
template <typename ai> struct aq {
  using am = p<ai>::am;
  using al = p<ai>::al;
  aq(ai, am o) : ar(o) {}
  al *as() { return ar; }
  am ar;
};
template <typename ai> aq<ai> at(ai h) { return {h, p<ai>::ah(h, 0)}; }
} // namespace std
template <typename ac> struct au {
  std::b<sizeof(ac)>::c av;
  ac *ar() {
    void *j = &av;
    return static_cast<ac *>(j);
  }
};
namespace std {
template <typename> struct aw;
template <typename ac> struct ax {
  ax(ac) {}
};
template <typename ai> struct ay {
  ai az;
};
template <typename ac, typename ai> struct ba {
  struct bb : ax<ai> {
    au<ac> av;
  };
  using bc = aj<ai, ba>;
  template <typename... ae> ba(ai h, ae... o) : bd(h) {
    p<ai>::ao(h, ar(), o...);
  }
  ac *ar() { return bd.av.ar(); }
  bb bd;
};
struct be {
  template <typename ac, typename ai, typename... ae>
  be(ac *&h, ay<ai> o, ae... ap) {
    typedef ba<ac, ai> bf;
    typename bf::bc bg;
    auto bh = at(bg);
    bf *bj = bh.as();
    auto bk = new (bj) bf(o.az, ap...);
    h = bk->ar();
  }
};
template <typename ac> struct z {
  using bl = ac;
  bl *operator->() {
    bl *bm = static_cast<aw<ac> *>(this)->as();
    return bm;
  }
};
template <typename ac> struct aw : z<ac> {
  using bl = ac;
  bl *as() { return ar; }
  template <typename ai, typename... ae> aw(ai h, ae... o) : bn(ar, h, o...) {}
  bl *ar;
  be bn;
};
template <typename ac> struct bo : aw<ac> {
  template <typename ai, typename... ae> bo(ai h, ae... o) : aw<ac>(h, o...) {}
};
template <typename ac, typename ai, typename... ae> bo<ac> bp(ai h, ae... o) {
  return bo<ac>(ay<ai>{h}, o...);
}
template <typename ac, typename... ae> bo<ac> bq(ae... h) {
  return bp<ac>(ag<int>(), h...);
}
} // namespace std
struct br;
struct bs {
  bs(std::coroutine_handle<br>);
  int x() { return bt; }
  void bu() { bt = 0; }
  void t() { bv.l(); }
  std::coroutine_handle<br> bv;
  int bt;
};
struct bw {
  using promise_type = br;
  bw(std::coroutine_handle<promise_type> h) : v(bq<bs>(h)) {}
  void t() { v->t(); }
  std::bo<bs> r() { return v; }
  std::bo<bs> v;
};
struct bx {
  int await_ready() { return 0; }
  void await_resume() {}
  void await_suspend(std::coroutine_handle<br>);
  bw by;
};
struct br {
  auto initial_suspend() { return std::ab(); }
  auto final_suspend() noexcept { return std::ab(); }
  auto get_return_object() { return std::coroutine_handle<br>::g(*this); }
  void unhandled_exception() {}
  auto await_transform(bw h) { return bx(h); }
  void return_void() {}
  void u(bs *h) { v = h; }
  bs *r() { return v; }
  bs *v;
};

// When reduced including headers, the following is required under
// clang -O2 @ 54225c457a336b1609c6d064b2b606a9238a28b9, otherwise this
// entire function is optimized out which masks the problem since the crash
// occurs at `q->x()` (q is nullptr).
#pragma clang optimize off
void bx::await_suspend(std::coroutine_handle<br> h) {
  auto bi = h.m();
  auto q = bi.r();
  auto s = by.r();
  if (q->x())
    s->bu();
  by.t();
}
#pragma clang optimize on

bs::bs(std::coroutine_handle<br> h) {
  bt = 0;
  bv = h;
  auto &bz = h.m();
  bz.u(this);
}
bw ca() { co_return; }
bw cb() { co_await ca(); }
bw cc = cb();
int main() {
  bw cd(cc);
  cd.t();
}

Reproduction Steps

Latest Clang -- Crashes

$ clang++ --version
clang version 20.0.0git (https://github.com/llvm/llvm-project.git 381a803da253b75c8b7b10bb732e9e90925185e8)
Target: x86_64-unknown-linux-gnu
Thread model: posix
InstalledDir: <redacted>
$ clang++ -O0 -std=c++20 reduced-Version-A.cpp -o reduced-Version-A.out
$ ./reduced-Version-A.out
cb (co_await) called
ca (co_return) called
$ clang++ -O2 -std=c++20 reduced-Version-A.cpp -o reduced-Version-A.out
$ ./reduced-Version-A.out
cb (co_await) called
Segmentation fault (core dumped)
$ clang++ -O0 -std=c++20 reduced-Version-B.cpp -o reduced-Version-B.out
$ ./reduced-Version-B.out
$ clang++ -O2 -std=c++20 reduced-Version-B.cpp -o reduced-Version-B.out
$ ./reduced-Version-B.out
Segmentation fault (core dumped)

Clang At Bisected Commit -- Crashes

$ clang++ --version
clang version 17.0.0 (https://github.com/llvm/llvm-project.git 54225c457a336b1609c6d064b2b606a9238a28b9)
Target: x86_64-unknown-linux-gnu
Thread model: posix
InstalledDir: <redacted>
$ clang++ -O0 -std=c++20 reduced-Version-A.cpp -o reduced-Version-A.out
$ ./reduced-Version-A.out
cb (co_await) called
ca (co_return) called
$ clang++ -O2 -std=c++20 reduced-Version-A.cpp -o reduced-Version-A.out
$ ./reduced-Version-A.out
cb (co_await) called
Segmentation fault (core dumped)
$ clang++ -O0 -std=c++20 reduced-Version-B.cpp -o reduced-Version-B.out
$ ./reduced-Version-B.out
$ clang++ -O2 -std=c++20 reduced-Version-B.cpp -o reduced-Version-B.out
$ ./reduced-Version-B.out
Segmentation fault (core dumped)

Clang Before Bisected Commit -- Does not crash

$ clang++ --version
clang version 17.0.0 (https://github.com/llvm/llvm-project.git 32be3405f57f1e4d0ec0da943434113450583e89)
Target: x86_64-unknown-linux-gnu
Thread model: posix
InstalledDir: <redacted>
$ clang++ -O0 -std=c++20 reduced-Version-A.cpp -o reduced-Version-A.out
$ ./reduced-Version-A.out
cb (co_await) called
ca (co_return) called
$ clang++ -O2 -std=c++20 reduced-Version-A.cpp -o reduced-Version-A.out
$ ./reduced-Version-A.out
cb (co_await) called
ca (co_return) called
$ clang++ -O0 -std=c++20 reduced-Version-B.cpp -o reduced-Version-B.out
$ ./reduced-Version-B.out
$ clang++ -O2 -std=c++20 reduced-Version-B.cpp -o reduced-Version-B.out
$ ./reduced-Version-B.out

[dgg5503]

Hi,

Just a small update on this issue. I noticed that if I replace auto get_return_object() with bw get_return_object(), the return type of the coroutine functions, the runtime crash no longer occurs for both reproducers.

Inspection of an AST dump shows that when providing the return type explicitly, the implicit VarDecl for __coro_gro is not created as per these lines since the return type ends up being matched.

Is it possible some aspect of the AST is misconfigured when using 'auto' as I provide in my reproducers that's leading to a runtime crash when optimizations are enabled? Or is this simply a misunderstanding of the coroutine specification on my part (i.e. explicit return type for get_return_object is required under certain conditions)?

[rnk]

cc @bricknerb @usx95 this seems like a coro frontend issue that we might wish to investigate