On Thu, Jan 11, 2024 at 11:20 AM juzhe.zh...@rivai.ai <juzhe.zh...@rivai.ai> wrote: > > Ok I see your idea and we need to adjust scalar_to_vec accurately. Inside the > loop we have these 2 scalar_to_vec: > > 1. MIN_EXPR <patt_28, 15> 1 times scalar_to_vec costs 1 in prologue > > This scalar_to_vec cost should be 0 or 1 since it only generate single > instructions: vmv.v.i v16,15 > > 2. 32872 >> patt_26 1 times scalar_to_vec costs 1 in prologue > > This cost should be higher since it cost 3 instructions: > li a4,-32768 > addiw a4,a4,104 > vmv.v.x v16,a4 > > Am I correct ? > > I guess if we cost 1 case as 1 cost and 2 case as 3 cost. Then we will be > good.
The cost model interface isn't set up very well to cost different constants differently. For scalar_to_vec we should get the backend the actual scalar to duplicate but we don't. > ________________________________ > juzhe.zh...@rivai.ai > > > From: Robin Dapp > Date: 2024-01-11 18:14 > To: juzhe.zh...@rivai.ai; Richard Biener > CC: rdapp.gcc; gcc-patches; kito.cheng; Kito.cheng; jeffreyalaw > Subject: Re: [PATCH] RISC-V: Increase scalar_to_vec_cost from 1 to 3 > > Yeah... I just noticed. I should set it as 4 to fix it with biggest VLEN > > size, > > that is, -march=rv64gcv_zvl4096b --param=riscv-autovec-lmul=m8... > > > > I am confused now how to fix this case. > > 4 is definitely too high compared to a regular instruction. > vmv.vx could even be zero-cost for constants. > > To catch constants we could add handling in add_stmt_cost, inspecting > the stmt directly. > > Regards > Robin >
