Hi,
Big code size regression is introduced by register renaming pass on thumb2
instruction set. The root cause comprises two aspects:
1. The pass renames LO_REGS into high registers, changing 2 bytes
instructions into 4 bytes. This is because function find_best_rename_reg
does not forbid renaming LO_REGS into high register.
2. The pass renames two operands instructions into three operand ones,
changing 2 bytes instructions into 4 bytes. Though renaming pass has already
taken two operands instructions into account, it needs matching constraint
alternatives for such instructions provided by backend, which are missed in
arm.md. Instruction patterns affected are as listed below:
	arm_addsi3
	arm_subsi3_insn
	arm_mulsi3_v6
	arm_andsi3_insn
	andsi_notsi_si
	iorsi3_insn
	arm_shiftsi3

This patch fixes the problems:
1. Change find_best_rename_reg function.
If PREFERRED_CLASS is defined as register class other than NO_REGS: we don't
rename old_reg into non-preferred register if old_reg is in PREFERRED_CLASS;
otherwise we rename old_reg into preferred register whenever possible, and
only after that we try to rename it into other registers. Also the loop is
reduced into one pass as a side-effect.
2. Add matching operand alternatives for mentioned patterns in arm.md.

With this patch, code size of CSiBE built with "-mthumb -mcpu=cortex-m3 -Os
-frename-registers" improved by 0.8% on trunk and now it is better than
register renaming disabled.

I tested the patches on cortex-m3 with and without register renaming, and on
x86 with register renaming.

I will apply this patch into ARM Embedded-4_7-branch and commit it to TRUNK
after approval.

Thanks.


2012-11-30  Bin Cheng  <bin.cheng@arm.com>

	* config/arm/arm-cores.def (cortex-m3, cortex-m4): Use v7m.
	* config/arm/arm-protos.h (tune_params): Add
	preferred_renaming_class.
	* config/arm/arm.c (arm_slowmul_tune, arm_fastmul_tune)
	(arm_strongarm_tune, arm_xscale_tune, arm_9e_tune, arm_v6t2_tune)
	(arm_cortex_tune, arm_cortex_a15_tune, arm_cortex_a5_tune)
	(arm_cortex_a9_tune, arm_v6m_tune, arm_fa726te_tune): Set
	preferred_renaming_class field.
	(arm_v7m_tune): New.
	(arm_preferred_rename_class): Return preferred renaming register
	class.
	* config/arm/arm.md (*arm_addsi3, *arm_subsi3_insn, *arm_mulsi3_v6)
	(*arm_andsi3_insn, andsi_notsi_si, *iorsi3_insn, *arm_xorsi3)
	(*arm_shiftsi3): Add alternatives for Thumb2 set.
	* regrename.c (find_best_rename_reg): Don't rename preferred
	register to non-preferred register.