Code in optabs.c for directly expanding some operations checks the
number of operands in the expander/named pattern against the expected
number of operands.  If they do not match an ICE is triggered.


/* Try to generate instruction ICODE, using operands [OPS, OPS + NOPS)
   as its operands.  Return the instruction pattern on success,
   and emit any necessary set-up code.  Return null and emit no
   code on failure.  */

rtx_insn *
maybe_gen_insn (enum insn_code icode, unsigned int nops,
                struct expand_operand *ops)
{
  gcc_assert (nops == (unsigned int) insn_data[(int)
icode].n_generator_args);


[ ... ]


On a target with recip or rsqrt patterns written in the expected way
we'll trigger the ICE because those patterns expose the recip as a
division with 1.0 as an operand, thus creating a second, unexpected
operand  This showed up testing v850e3v5.

It looks like some targets hide things a bit behind an unspec, possibly
to avoid this problem.  This patch does something very similar for the v850:


(define_expand "rsqrtsf2"
  [(set (match_operand:SF 0 "register_operand" "=")
        (unspec:SF [(match_operand:SF 1 "register_operand" "")]
                   UNSPEC_RSQRT))]

We then generate a more normal looking pattern from the expander which
matches:


(define_insn "rsqrtsf2_insn"
  [(set (match_operand:SF 0 "register_operand" "=r")
        (div:SF (match_operand:SF 1 "const_float_1_operand" "")
                (sqrt:SF (match_operand:SF 2 "register_operand" "r"))))]


While looking at this I noticed that const_float_1_operand would
essentially reject everything.  It had a toplevel check that the operand
was a CONST_INT.  But in the body of the predicate it would reject
anything that was not a CONST_DOUBLE.

The right code is CONST_DOUBLE.  So this patch also fixes
const_float_1_operand.

This fixes a handful of ICEs when testing for v850e3v5 (pr41963,
pr46728-9.  It also results in many recipf instructions being generated
for newlib v850e3v5 multilib whereas none were generated before.
There's no uses of rsqrt in newlib, but they do show up in the gcc
testsuite.


Installing on the trunk,

Jeff