On 2/20/20 12:15 PM, Simon Marchi wrote:
> The disagreement between the symbol names between count-one-bits.o and arm-get-next-pcs.o
> seems to show that there is a missing `extern "C"` in the count-one-bits.h header?

Thanks, I'll take your word for it. (I don't use C++.) I installed the 
attached patch into Gnulib. Apparently these modules have been used by C 
code only, until now.

> And apparently, gcc decided to compile the definition in the header and include
> it in arm-get-next-pcs.o, while clang did not.  Is that the compiler's choice,
> or is it because the macros and preprocessor conditionals in count-one-bits.h
> evaluated to something different for the two compilers?

It's a compiler's choice whether a function is inlined. When I call 
count_one_bits in C, gcc -O2 doesn't create a function for 
count_one_bits in the calling module; it simply issues the popcnt insn, 
or calls the appropriate GCC library function on platforms that don't 
have a popcnt insn. Perhaps you were compiling with some other level of 
optimization, or perhaps C++ does this differently; but regardless of 
how the compiler does it the resulting code is supposed to work of course.