Problem compiling with Intel C compiler via MATLAB Coder

I have registered a toolchain that uses the Intel C compiler (icc), but compilation fails because icc does not recognize NAN and INFINITY in the file rt_nonfinite.c. The problematic lines are

real_T rtNaN = (real_T)NAN;

and

real_T rtInf = (real_T)INFINITY;

I thought I had solution when I added

tc.addMacro('C_STANDARD_OPTS',' -D__PURE_INTEL_C99_HEADERS__ ');

to my m-file defining the toolchain object. But now compilation fails due to the same lines (and more) in rt_nonfinite.c, but for a different reas on. The error is now

error: expression must have a constant value.

I know that NAN and INFINITE are defined in the math.h that comes with icc. Could icc be unaware of its own header files? How can I fix that?

Thanks in advance for advice and insight. I am working on a 64 bit Linux box.

Answer 1

[Joe M]

I managed to get MATLAB Coder to use icc successfully by adding the following to the m-file that defined the toolchain object:

tc.addMacro('C_STANDARD_OPTS', ' -std=c11 ');

With this added, everything compiled. There are other parts in my example's C_STANDARD_OPTS, but I did not change them.

At least one mystery remains. I define a coder.config object with the following commands.

cfg = coder.config('exe');
cfg.Toolchain = 'Intel IPP | make (64-bit Linux)';

But when I query the value of cfg.TargetLangStandard, the value is 'C99 (ISO)'.

Is my cfg not using the C11 standard, which seemed to allow compilation to proceed?

[Ryan Livingston]

The setting cfg.TargetLangStandard is what controls the flavor of C or C++ code generated by MATLAB Coder. The toolchain setting

tc.addMacro('C_STANDARD_OPTS', ' -std=c11 ');

only controls flags passed to icc, it does not impact the generated code. Without knowing much about icc, I'd suspect you could use -std=c99 and still have things work as MATLAB Coder is generating C99 code.

To demonstrate, if you set

cfg.TargetLangStandard = "C89/C90 (ANSI)";

and generate code, you'll see that the references to INFINITY and NAN are gone since those were introduced in C99:

/*
 * Function: rt_InitInfAndNaN ==================================================
 *  Abstract:
 *  Initialize the rtInf, rtMinusInf, and rtNaN needed by the
 * generated code. NaN is initialized as non-signaling. Assumes IEEE.
 */
void rt_InitInfAndNaN()
{
  rtNaN = rtGetNaN();
  rtNaNF = rtGetNaNF();
  rtInf = rtGetInf();
  rtInfF = rtGetInfF();
  rtMinusInf = rtGetMinusInf();
  rtMinusInfF = rtGetMinusInfF();
}

/*
 * Function: rtGetNaN
 * ======================================================================
 *  Abstract:
 * Initialize rtNaN needed by the generated code.
 *  NaN is initialized as non-signaling. Assumes IEEE.
 */
real_T rtGetNaN(void)
{
  real_T nan = 0.0;
  uint16_T one = 1U;
  enum
  {
    LittleEndian,
    BigEndian
  } machByteOrder = (*((uint8_T *)&one) == 1U) ? LittleEndian : BigEndian;
  switch (machByteOrder) {
  case LittleEndian: {
    union {
      LittleEndianIEEEDouble bitVal;
      real_T fltVal;
    } tmpVal;

    tmpVal.bitVal.words.wordH = 0xFFF80000U;
    tmpVal.bitVal.words.wordL = 0x00000000U;
    nan = tmpVal.fltVal;
    break;
  }
  case BigEndian: {
    union {
      BigEndianIEEEDouble bitVal;
      real_T fltVal;
    } tmpVal;

    tmpVal.bitVal.words.wordH = 0x7FFFFFFFU;
    tmpVal.bitVal.words.wordL = 0xFFFFFFFFU;
    nan = tmpVal.fltVal;
    break;
  }
  }
  return nan;
}

So your toolchain could consider basingthe value passed to the icc -std= based on the value of cfg.TargetLangStandard.

[Joe M]

Thank you for the inisght. I will test this.

Is there documentation that describes the map between GNU make commands and construction of a toolchain object? Some of the connections are clear, but some are a bit murky to me.

[Ryan Livingston]

The root of the custom toolchain doc is here:

https://www.mathworks.com/help/coder/custom-toolchain-registration.html

Which sort of GNU Make commands are you looking for info regarding?