illumos gate

When working on implementing CTF everywhere in OmniOS userland, some 32-bit objects that use the _Float128 type fail to convert:
An example is the 32-bit libmpfr.so

% /opt/onbld/bin/i386/ctfconvert -m -o /dev/null ~/ctf/usr_lib_libmpfr.so.6.1.0
ctfconvert: CTF conversion failed: Invalid argument

Using ctfconvert with the patch for 13280 applied gives a bit more information:

% ctfconvert -mo /dev/null ~/ctf/usr_lib_libmpfr.so.6.1.0
ctfconvert: failed to find valid fp mapping for encoding 4, size 0 bits
ctfconvert: CTF conversion failed: Invalid argument

Note that the message stating 0 bits is because of an incorrect format string in libctf, the type is actually a 128-bit double.

It seems that several third party packages use the ISO/IEC TS 18661-3:2015 _Float128 type. For a 64-bit object, this is the same as long double but in a 32-bit object a long double is only 96 bits.

Changing ctfconvert to encode these as 128-bit long doubles seems sensible rather than failing the entire conversion, although tools like mdb will not be able to print them directly.

I tested the attached patch with:

#include <stdio.h>

long double ldouble = 123.1;
_Float128 f128 = 245.5;

int
main() {
        printf("Long double: %lf\n", ldouble);
        printf("Float128:    %lf\n", f128);
}

The type is now encoded:

% ctfdump -t float128.32

- Types -------------------------------------------------------------------------

  <1> long double encoding=LDOUBLE offset=0 bits=96
  <2> _Float128 encoding=LDOUBLE offset=0 bits=128
  <3> int encoding=SIGNED offset=0 bits=32

but mdb cannot print it, as expected

% mdb float128.32
> ldouble::print
+1.2309999999999999e+02
> f128::print
??? (unsupported FP format 6 / 128 bits