On 2011/5/20 11:45 AM, Dmitry Plotnikov wrote: > This patch adds support for 12-bit immediate values for Thumb-2 in ADD and > SUB instructions. We added two new alternatives for *arm_addsi3 which > make use of two new constraints for 12-bit values. Also we modified > costs of PLUS rtx expression. > This patch reduces size of libevas by 1788 bytes (from 464916 to > 463128), and sqlite by 54 bytes (from 266156 to 266052). > Regtested with Cortex-A8 QEMU. > > Ok for trunk? >
Andrew Stubbs seem to have another patch related to ADDW/SUBW support, which I think is not yet committed to trunk. I have not yet studied how this patch and Andrew's relate. That aside, I think the style of adding new alternatives for this purpose is a little unneeded. I suggest: 1) Abstract out const_ok_for_arm() into const_ok_for_arm_outer() with an OUTER rtx code argument, and a const_ok_for_arm() with OUTER passed 0. 2) Within const_ok_for_arm_outer(), test for OUTER==PLUS and TARGET_THUMB2 as needed. 3) Migrate from const_ok_for_arm() to const_ok_for_arm_outer() as needed: in pattern conditions, etc. I'll also note here that ADD/SUB are not the only instructions with 12-bit immediate under Thumb-2; so does AND, ORR, etc. Chung-Lin
