On Thu, Sep 18, 2014 at 11:54:46PM -0600, Jeff Law wrote: > Shouldn't you be testing if the register is fixed rather than its class? > Or maybe both?
register_operand (via general_operand) uses operand_reg_set for this; it is initialised via the regclass NO_REGS too (and other things). This would work for us (rs6000) too, or indeed fixed_regs[], or even wider classes. I have no idea if it would hurt other targets though, and I have no desire to test all weirdo targets ;-) If another regclass reg is assigned to a match_scratch (and then via a splitter to a match_operand register), reload can fix it up. It cannot fix up things properly for regs of NO_REGS class. This is what this patch is for: a class NO_REGS reg cannot ever work as a match_scratch. I could test for a wider class if you want, but I'll need some guidance what to test for exactly then; existing code isn't exactly consistent. Segher
