ARC64/GCC: Improve iirflt01 performance

[claziss]

The iirflt01 benchmark is reported to deliver only 1/4 performance of the MWDT compiled version.

One of the problem is missing GIMPLE optimization which converts a floating point comparison of an integer with a constant float to integer comparison.

The fault is located in gcc's match.pb file:

/* Optimize various special cases of (FTYPE) N CMP CST. */

...

(if (fmt.can_represent_integral_type_p (itype) && ! exception_p)

...

The call to can_represent ... doesn't match.

A quick hack show us a jump from 481 to 610 (26%)