[lmi] 'zero_test.cpp' crashes clang++ with '-O2 -ffp-model=strict'

[Greg Chicares]

[effective workaround given below]

Using...

  $ clang++ --version      
  Debian clang version 13.0.1-6
  Target: x86_64-pc-linux-gnu
  Thread model: posix
  InstalledDir: /usr/bin

...this crashes clang++:

$ clang++ -c -I . -fPIC -ffp-model=strict -std=c++20 -O2 -Wno-string-plus-int 
zero_test.cpp -ozero_test.o 2>&1 |less -S

Moving '-ffp-model=strict' to the end or the beginning, it still crashes:

$ clang++ -c -I . -fPIC -std=c++20 -O2 -Wno-string-plus-int zero_test.cpp 
-ozero_test.o -ffp-model=strict 2>&1 |less -S
$ clang++ -ffp-model=strict -c -I . -fPIC -std=c++20 -O2 -Wno-string-plus-int 
zero_test.cpp -ozero_test.o 2>&1 |less -S

Removing '-ffp-model=strict' keeps it from crashing:

$ clang++ -c -I . -fPIC -std=c++20 -O2 -Wno-string-plus-int zero_test.cpp 
-ozero_test.o 2>&1 |less -S

but I really do want IEEE 754 behavior: without '-ffp-model=strict',
some unit test told me that 1.5 rounded downward is 2.0, a result
with which I must disagree.

I don't actually want to unmask floating-point exceptions, so this is
the best workaround I can come up with:

clang++ -c -I . -Woverriding-t-option -ffp-model=strict 
-ffp-exception-behavior=ignore -Wno-overriding-t-option -fPIC -std=c++20 -O2 
-Wno-string-plus-int zero_test.cpp -ozero_test.o 2>&1 |less -S

(surrounding the '-ffp' options with flags to toggle the warning
that they jointly engender).

I suppose I could try unmasking floating-point exceptions early
(say, in static initialization), but I wouldn't expect that to
forestall a compile-time diagnostic.