LLVM-17: incorrect result from `isinf(x) || isinf(x)`

[linehill]

The following case fails with Clang-17 and -O2 on Intel hardware (intel-compute-runtime 23.26.26690.22):

#include <hip/hip_runtime.h>

#ifndef SEPARATE
__global__ void isInfOrNan(bool *Res, float Val) {
  *Res = (isinf(Val) || isnan(Val));
}
#else
__global__ void isInf(bool *Res, float Val) { *Res = isinf(Val); }
__global__ void isNan(bool *Res, float Val) { *Res = isnan(Val); }
#endif

int main() {
#ifndef SEPARATE
  bool InfOrNan = true, *ResD;
  (void)hipMalloc(&ResD, sizeof(bool));
  isInfOrNan<<<1, 1>>>(ResD, 1.23f);
  (void)hipMemcpy(&InfOrNan, ResD, sizeof(bool), hipMemcpyDeviceToHost);
  printf("InfOrNan=%d\n", InfOrNan);
  return InfOrNan; // Should return zero.
#else
  bool Inf = true, Nan = true, *ResD;
  (void)hipMalloc(&ResD, sizeof(bool));
  isInf<<<1, 1>>>(ResD, 1.23f);
  (void)hipMemcpy(&Inf, ResD, sizeof(bool), hipMemcpyDeviceToHost);
  isNan<<<1, 1>>>(ResD, 3.21f);
  (void)hipMemcpy(&Nan, ResD, sizeof(bool), hipMemcpyDeviceToHost);
  printf("InfOrNan=%d\n", Inf || Nan);
  return Inf || Nan; // Should return zero.
#endif
}

The case passes on AMD and passes on chipStar if:

• it is compiled with -O0,
• isinfand isnan is computed in separate kernels (-O2 -DSEPARATE) or
• compiled with Clang-16.

I'm guessing that the appearances of new llvm.is.fpclass intrinsics since LLVM-17 with certain test bit combinations are causing an error in somewhere.