I'm working on a chip with some unusual addressing modes. it does have [base_rage+msize*index_reg] and one can't omit base_reg like on x86. But when ivopts tries to calculate the costs of different addressing modes it only checks [mult*reg] to determine allowable multipliers. I modified multiplier_allowed_in_address_p to check both cases.
There's no actual testcase that fails it's just my back-end generated sub-optimal code. I tested it on my tree off 4.8.1 and it fixes the problem. Rebuilt trunk for x86_64-linux-gnu. Should it be included in trunk to make check more robust? Thanks Igor 2013-10-22 Igor Shevlyakov <igor.shevlya...@gmail.com> * tree-ssa-loop-ivopts.c (multiplier_allowed_in_address_p ): Check both [reg+mult*reg] and [mult*reg] to determine if multiplier is allowed. Index: tree-ssa-loop-ivopts.c =================================================================== --- tree-ssa-loop-ivopts.c (revision 203937) +++ tree-ssa-loop-ivopts.c (working copy) @@ -3108,16 +3108,19 @@ multiplier_allowed_in_address_p (HOST_WI { enum machine_mode address_mode = targetm.addr_space.address_mode (as); rtx reg1 = gen_raw_REG (address_mode, LAST_VIRTUAL_REGISTER + 1); - rtx addr; + rtx reg2 = gen_raw_REG (address_mode, LAST_VIRTUAL_REGISTER + 2); + rtx addr, scaled; HOST_WIDE_INT i; valid_mult = sbitmap_alloc (2 * MAX_RATIO + 1); bitmap_clear (valid_mult); - addr = gen_rtx_fmt_ee (MULT, address_mode, reg1, NULL_RTX); + scaled = gen_rtx_fmt_ee (MULT, address_mode, reg1, NULL_RTX); + addr = gen_rtx_fmt_ee (PLUS, address_mode, scaled, reg2); for (i = -MAX_RATIO; i <= MAX_RATIO; i++) { - XEXP (addr, 1) = gen_int_mode (i, address_mode); - if (memory_address_addr_space_p (mode, addr, as)) + XEXP (scaled, 1) = gen_int_mode (i, address_mode); + if (memory_address_addr_space_p (mode, addr, as) || + memory_address_addr_space_p (mode, scaled, as)) bitmap_set_bit (valid_mult, i + MAX_RATIO); }