[Bug target/60604] [4.9 Regression] GCC incorrectly compiles s_csinh function on MIPS32 (32bit fp)

[sje at gcc dot gnu.org]

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

--- Comment #6 from Steve Ellcey <sje at gcc dot gnu.org> ---
I think the underlying bug here is the code we generate for builtin_fabs.
In emit-rtl.c (validate_subreg) I see this comment:

  /* Subregs involving floating point modes are not allowed to
     change size.  Therefore (subreg:DI (reg:DF) 0) is fine, but
     (subreg:SI (reg:DF) 0) isn't.  */

But on mips, builtin_fabs calls expand_absneg_bit and that generates
this RTL:

;; _11 = ABS_EXPR <_4>;

(insn 26 25 27 (set (subreg:SI (reg:DF 199 [ D.3267 ]) 0)
        (and:SI (subreg:SI (reg:DF 194 [ D.3267 ]) 0)
            (const_int 2147483647 [0x7fffffff]))) x.c:25 -1
     (nil))

(insn 27 26 0 (set (subreg:SI (reg:DF 199 [ D.3267 ]) 4)
        (subreg:SI (reg:DF 194 [ D.3267 ]) 4)) x.c:25 -1
     (nil))

Which would seem to violate the restrictions in the comment.