https://gcc.gnu.org/bugzilla/show_bug.cgi?id=55212
--- Comment #76 from Kazumoto Kojima <kkojima at gcc dot gnu.org> --- Created attachment 33787 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=33787&action=edit a reduced test case of SCiBE compiler/vam test compiler/vam is a test which is an example of code size regression from non LRA compiler. With -O2, non LRA compiler generates code like mov.l r0,@(56,r15) mov #64,r1 add #2,r0 mov r13,r2 add r15,r1 mov.l r0,@(60,r15) add #18,r2 mov r13,r14 add #2,r0 mov.l r2,@(0,r1) mov.l r0,@(4,r1) add #2,r2 add #2,r0 mov.l r2,@(8,r1) mov.l r0,@(12,r1) and the corresponding code generated by sh-lra compiler looks like mov.l r1,@(56,r15) mov #17,r1 add r2,r1 mov.l r1,@(60,r15) mov #2,r14 mov #16,r10 mov #18,r1 add r2,r1 add r2,r14 add r2,r10 mov #64,r2 add r15,r2 mov.l r1,@r2 mov #19,r1 mov #68,r2 add r13,r1 add r15,r2 mov.l r1,@r2 mov #20,r1 mov #72,r2 add r13,r1 add r15,r2 mov.l r1,@r2 mov #21,r1 mov #76,r2 add r13,r1 add r15,r2 mov.l r1,@r2 which shows that base+disp addressing mode wasn't used for the insn (set reg (mem (plus sp N))) when N >= 64. It looks that non LRA compiler uses targetm.legitimize_address to handle that case and splits it to (set new_reg (plus sp M)) and (set reg (mem (plus new_reg N-M))) so as to N-M is a legitimate display constant for base+disp addressing.
