Make profile estimation more precise

Make profile estimation more precise

While analyzing code size regression in SPEC2k GCC binary I noticed that we
perform some inline decisions because we think that number of executions are
very high.
In particular there was inline decision inlining gen_rtx_fmt_ee to find_reloads
believing that it is called 4 billion times.  This turned out to be cummulation
of roundoff errors in propagate_freq which was bit mechanically updated from
original sreals to C++ sreals and later to new probabilities.

This led us to estimate that a loopback edge is reached with probability 2.3
which was capped to 1-1/10000 and since this happened in nested loop it quickly
escalated to large values.

Originally capping to REG_BR_PROB_BASE avoided such problems but now we have
much higher range.

This patch avoids going from probabilites to REG_BR_PROB_BASE so precision is
kept.  In addition it makes the propagation to not estimate more than
param-max-predicted-loop-iterations.  The first change makes the cap to not
be triggered on the gcc build, but it is still better to be safe than sorry.

	* ipa-fnsummary.c (estimate_calls_size_and_time): Fix formating of
	dump.
	* params.opt: (max-predicted-iterations): Set bounds.
	* predict.c (real_almost_one, real_br_prob_base,
	real_inv_br_prob_base, real_one_half, real_bb_freq_max): Remove.
	(propagate_freq): Add max_cyclic_prob parameter; cap cyclic
	probabilities; do not truncate to reg_br_prob_bases.
	(estimate_loops_at_level): Pass max_cyclic_prob.
	(estimate_loops): Compute max_cyclic_prob.
	(estimate_bb_frequencies): Do not initialize real_*; update calculation
	of back edge prob.
	* profile-count.c (profile_probability::to_sreal): New.
	* profile-count.h (class sreal): Move up in file.
	(profile_probability::to_sreal): Declare.