Re: GNU Guile 2.1.5 released (beta)

[Andy Wingo]

On Tue 20 Dec 2016 23:58, Matt Wette <address@hidden> writes:

> With the patch below I can get through all the tests on my Mac (macOS 
> 10.12.2) if I use default flags (i.e., -g -O2) and gcc-4.9 (now verifying 
> also w/ gcc-6.2). The patch is a hack to
> get around an apparent optimization that thinks the sine of -0.0 is +0.0. 
>
> Matt
>
> --- libguile/numbers.c-orig 2016-11-21 13:56:23.000000000 -0800
> +++ libguile/numbers.c 2016-12-20 14:43:58.000000000 -0800
> @@ -9099,6 +9099,9 @@
> #undef FUNC_NAME
>
> SCM
> +#ifdef __APPLE__
> +__attribute__((optimize("O0")))
> +#endif
> scm_c_make_polar (double mag, double ang)
> {
> double s, c;

Aaah, thank you for tracking this one down!  So.  I assume that with GCC
(any version) everything is fine.  This is apparently an optimization in
LLVM.  But... is LLVM wrong, or are we wrong?  I guess LLVM is wrong,
right, since the result depends on the optimization level?  Surely we
should be able to produce a nice compiler bug report out of this :)

Andy