I see, thanks Robin, will have a try for this change. Pan
-----Original Message----- From: Robin Dapp <rdapp....@gmail.com> Sent: Wednesday, November 27, 2024 9:44 PM To: Li, Pan2 <pan2...@intel.com>; gcc-patches@gcc.gnu.org Cc: juzhe.zh...@rivai.ai; kito.ch...@gmail.com; jeffreya...@gmail.com; Robin Dapp <rdapp....@gmail.com> Subject: Re: [PATCH v1 1/3] RISC-V: Combine vec_duplicate + vadd.vv to vadd.vx > I see, didn't aware of that. I am not sure if we need to consider vsetvl here? > As there are extra 2 insn here. I wouldn't consider it as it's outside of the loop. What matters is latency inside the loop. > I see, need to consider the cost here. Any example I can reference? Sorry I > haven't touch cost model in previous. Refer to riscv_rtx_costs. We set all vector instruction costs to 1. What you would need to do is pattern-match scalar (broadcast) operands inside our IF_THEN_ELSE vector patterns. There are a some scalar examples that show how it's done in principle. Then make use of either get_vector_costs ()->regmove->GR2VR; or riscv_register_move_cost to increase the pattern cost respectively. -- Regards Robin
