On Thu, Oct 04, 2007 at 11:14:12AM +0200, Boris Boesler wrote: > Hi > > Am 02.10.2007 um 20:02 schrieb Rask Ingemann Lambertsen: > > >On Tue, Oct 02, 2007 at 02:50:50PM +0200, Boris Boesler wrote: > > > > What does the .lreg dump file (-fdump-rtl-lreg) say about > >register costs? [snip] > Register 101 costs: Rx_REGS:1000 Ax_REGS:0 GENERAL_REGS:1000 > ALL_REGS:1000 MEM:4000 [snip] > Register 101 pref Ax_REGS, else Rx_REGS > 102 registers. > [8<] > > Register 101 used 2 times across 2 insns in block 3; set 1 time; pref > Ax_REGS, else Rx_REGS; pointer. > > Register 101 is the important one. And it says prefer Ax; but it > worries be that the costs for Ax is 0. Does this mean the value is > placed in an Ax register or does it mean it is impossible to place > the value in a Ax register? Cost 0 means that all instructions which use register 101 accept a register of class Ax_REGS, so that's good. The next step would be to check the .greg dump to see why register 101 ends up being assigned to a hard register in class Rx_REGS. > > GO_IF_LEGITIMATE_ADDRESS() must reject addresses which use hard > >registers > >of the wrong class. Even when !REG_OK_STRICT. > > Ok, I check the registers in any case: now I can' compile anything! > > const_assign.c:19: error: unrecognizable insn: > (insn 9 8 10 3 (set (mem/c/i:SI (reg/f:SI 102) [0 c+0 S4 A32]) > (const_int 4711 [0x1267])) -1 (nil) > (nil)) > const_assign.c:19: internal compiler error: in extract_insn, at > recog.c:2077 But register 102 is a pseudo register, isn't it? A pseudo register must be accepted when !REG_OK_STRICT. You should have something like this: #ifdef REG_OK_STRICT #define REGNO_OK_FOR_BASE_P(r) \ (TEST_HARD_REG_BIT (reg_class_contents[BASE_REG_CLASS], (r))) #else /* !REG_OK_STRICT */ #define REGNO_OK_FOR_BASE_P(r) \ (!HARD_REGISTER_NUM_P (r) \ || TEST_HARD_REG_BIT (reg_class_contents[BASE_REG_CLASS], (r))) #endif /* !REG_OK_STRICT */ Then define GO_IF_LEGITIMATE_ADDRESS() using REGNO_OK_FOR_BASE_P(). There was discussion of a very similar problem on the GCC development list, which is really where question of this kind belongs since it is about development of GCC itself. http://gcc.gnu.org/ml/gcc/2007-09/msg00242.html > The .unshare file: [snip] > insn 7: load a symbol reference into a register 101 > insn 8: copy register 101 to register 102 > insn 9: store constant Don't worry about the needless reg->reg copy here, the unshare pass runs very early in the optimization. -- Rask Ingemann Lambertsen Danish law requires addresses in e-mail to be logged and stored for a year
