On 06/20/2013 07:58 AM, Claudio Fontana wrote: >> > + tcg_out_ldst(s, TARGET_LONG_BITS == 64 ? LDST_64 : LDST_32, >> > + LDST_LD, TCG_REG_X0, TCG_REG_X2, tlb_offset & 0xfff); >> > + tcg_out_ldst(s, LDST_64, LDST_LD, TCG_REG_X1, TCG_REG_X2, >> > + (tlb_offset & 0xfff) + (offsetof(CPUTLBEntry, addend) - >> > + (is_read ? offsetof(CPUTLBEntry, addr_read) : >> > + offsetof(CPUTLBEntry, addr_write)))); >> > + >> > + tcg_out_cmp(s, 1, TCG_REG_X0, TCG_REG_X3, 0); >> > + *label_ptr = s->code_ptr; >> > + tcg_out_goto_cond_noaddr(s, TCG_COND_NE); >> > +} > hmm should not the compare and branch actually be before the loading of the > addend? > If we jump to the slow path we don't need to load the addend do we? >
No, but it's the slow path, and we don't care if we do extra work. What's more important is minimizing the memory load delay for the fast path. r~
