http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49305
--- Comment #3 from Kazumoto Kojima <kkojima at gcc dot gnu.org> 2011-06-08 14:03:02 UTC --- It's my fault after all. The patch r174586 should check the mode parameter too. Without checking mode, insn-recog routines are confused and the combine pass makes insns with wrong insn_code. Now I'm testing --- ORIG/trunk/gcc/config/sh/predicates.md 2011-06-04 10:16:51.000000000 +0900 +++ trunk/gcc/config/sh/predicates.md 2011-06-08 18:55:09.000000000 +0900 @@ -395,6 +395,7 @@ } if ((mode == QImode || mode == HImode) + && mode == GET_MODE (op) && (MEM_P (op) || (GET_CODE (op) == SUBREG && MEM_P (SUBREG_REG (op))))) { @@ -432,6 +433,7 @@ return 0; if ((mode == QImode || mode == HImode) + && mode == GET_MODE (op) && (MEM_P (op) || (GET_CODE (op) == SUBREG && MEM_P (SUBREG_REG (op))))) { I'll apply it if it survives the tests on sh-elf and sh4-unknown-linux-gnu and put the patch in #1 into my TODO list.