[Bug target/60604] GCC incorrectly compiles s_csinh function on MIPS

["sje at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60604

--- Comment #1 from Steve Ellcey <sje at gcc dot gnu.org> ---
Created attachment 32428
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=32428&action=edit
New reduced test case

Here is a new reduced test case that calls no libm functions.  I am pretty sure
that the bug is in the handling of __builtin_fabs and specifically in
expand_absneg_bit where we use the MIPS exp instruction to genreate a floating
point absolute value.  I am not sure exactly what the problem is, but if I have
this routine always return NULL_RTX and use a differentr method of computing a
floating point absolute value then my program works.

I thought it might be a simulator issue since I do most of my testing with qemu
but I ran the executable on hardware and got failures there.

The new reduced test case prints with -O0 or -O1 prints:

x = 0.634964 1.298458
In __csinh(1), 0.634964 1.29846
In __csinh(2), 0.634964 1.29846
x = 0.000000 0.000000

With -O2 or -O3 it prints:

x = 0.634964 1.298458
In __csinh(1), 0.634964 1.29846
In __csinh(2), 1.00959e+116 1.29846
In if statement
x = 0.000000 0.000000