Tpetra: Mac Clang has trouble reading -9.6322e-41 as float

[mhoemmen]

Bug Report

@trilinos/tpetra

Description

With Clang on a Mac, TpetraCore_MatrixMatrix_UnitTests_MPI_4 fails with Scalar=float and complex<float> while trying to read line 398 of tpetra/core/test/MatrixMatrix/matrices/cdt_m.mtx. Line 398 looks like this:

  3860  2784 -9.6322e-41

It turns out that -9.6322e-41 is a subnormal number in float. The test prints the following:

  p=0: *** Caught standard std::exception of type 'std::runtime_error' :
1:
1:   .../packages/tpetra/core/inout/MatrixMarket_Tpetra.hpp:2352:
1:
1:   Throw number = 2
1:
1:   Throw test that evaluated to true: readSuccess == 0
1:
1:   Failed to read the Matrix Market sparse matrix file: Failed to get real value from line 398 of input

I inserted print statements to figure out what file it was having trouble reading.

Weirdly, Clang had no trouble with the following example:

#include <iostream>
#include <sstream>

int main() {
  std::istringstream is ("-9.6322e-41");
  float fl{};
  is >> fl;
  std::cout << fl << std::endl;
  return 0;
}

Both Mac Clang, and GCC 9.2 with -std=c++17, run fine and print -9.63225e-41. However, when I change the code that reads real values from

istr >> the_realValue;

to

if (std::is_same<Real, float>::value) {
  double dblVal {};
  istr >> dblVal;
  the_realValue = static_cast<Real> (dblVal);
}
else {
  istr >> the_realValue;
}

the test passes.

Steps to Reproduce

Build TpetraCore's tests with Mac Clang; here is part of the output of clang++ --version:

Apple LLVM version 10.0.1 (clang-1001.0.46.4)
Target: x86_64-apple-darwin18.7.0
Thread model: posix