* config/avr/avr-protos.h (avr_output_bld): New.
	(out_shift_with_cnt): Add t_len argument.
	* config/avr/avr.c (avr_num_arg_regs): Remove -mpack-args.
	(output_movqi, output_movhi, output_movsisf): Optimize loading
	any constant with exactly one bit set to NO_LD_REGS.
	(out_shift_with_cnt): Optimize output code for size or speed,
	depending on optimize_size.  Handle small shift counts as well
	(if not hand-optimized in ?sh??i3_out).  Shifts can be done
	with or without a scratch register, with help of __tmp_reg__
	or __zero_reg__ if necessary.  Add T_LEN argument to pass the
	length of TEMPLATE in words, return total insn length in *LEN.
	(ashlqi3_out, ashrqi3_out, lshrqi3_out): Change all calls to
	out_shift_with_cnt to work with the above change.
	(ashlhi3_out, ashlsi3_out, ashrhi3_out, ashrsi3_out, lshrhi3_out,
	lshrsi3_out): Likewise.  Optimize more known shift count cases.
	Remove cases already well optimized in out_shift_with_cnt.
	(avr_output_bld): New function.
	* config/avr/avr.h (MASK_PACK_ARGS, TARGET_PACK_ARGS): Remove.
	(TARGET_SWITCHES): Remove -mpack-args backward compatibility.
	* config/avr/avr.md (*reload_inqi, *reload_inhi, *reload_insi):
	Add reload_completed to insn condition - only for peepholes.
	(ashlqi3, ashrqi3, lshrqi3): Correct insn length for shift counts
	in a register or memory.
	(ashlhi3, ashlsi3, ashrhi3, ashrsi3, lshrhi3, lshrsi3): Likewise.
	Do not require a scratch register.
	(*ashlhi3_const, *ashlsi3_const, *ashrhi3_const, *ashrsi3_const,
	*lshrhi3_const, *lshrsi3_const): New insns and matching peepholes.
	Optimize shifts by known count using a scratch register, but only
	if one is still available after register allocation.

From-SVN: r36963