Miscompilation/Assertion "FPConstrained should be enabled on entire function"' failed with FENV_ACCESS and default function argument

[hubert-reinterpretcast]

With the following source, asserts-enabled builds of Clang hit an assertion:

/root/llvm-project/clang/lib/CodeGen/CodeGenFunction.cpp:161: void clang::CodeGen::CodeGenFunction::CGFPOptionsRAII::ConstructorHelper(clang::FPOptions): Assertion `(CGF.CurFuncDecl == nullptr || CGF.Builder.getIsFPConstrained() || isa<CXXConstructorDecl>(CGF.CurFuncDecl) || isa<CXXDestructorDecl>(CGF.CurFuncDecl) || (NewExceptionBehavior == llvm::fp::ebIgnore && NewRoundingBehavior == llvm::RoundingMode::NearestTiesToEven)) && "FPConstrained should be enabled on entire function"' failed.

If it is accepted that the pragma state at the point where code is lexically present is to be respected when the code is evaluated, then the non-asserts builds miscompile the code (as can be seen by inspecting the assembly obtained with -O; the call to foo is not generated).

https://godbolt.org/z/5PKqY5nfa

#include <float.h>
struct ChangeRounding {
  ChangeRounding();
  ~ChangeRounding();
};
int foo(), bar();

#pragma STDC FENV_ACCESS ON

struct Ex {
  Ex() = default;
  Ex(const Ex &, int = (ChangeRounding(), (float)0xFFFFFFFF != (float)0x100000000 ? foo() : bar()));
};

#pragma STDC FENV_ACCESS OFF
void q(const Ex &e) { Ex e2(e); }

Clang version info:

clang version 18.0.0 (https://github.com/llvm/llvm-project.git 6eb0332e72e5d9c5db1f89bac28770bc6f0725c2)

[llvmbot]

@llvm/issue-subscribers-clang-codegen

[hubert-reinterpretcast]

As can be expected, also fails with default member initializers (at least with aggregate initialization): https://godbolt.org/z/6GjxszdEW

#include <float.h>
struct ChangeRounding {
  ChangeRounding();
  ~ChangeRounding();
};
int foo(), bar();

#pragma STDC FENV_ACCESS ON

struct Ex {
  int x = (ChangeRounding(), (float)0xFFFFFFFF != (float)0x100000000 ? foo() : bar());
};

#pragma STDC FENV_ACCESS OFF
Ex q() { return {}; }

[hubert-reinterpretcast]

@cor3ntin, fyi (as you requested).