OK
juzhe.zh...@rivai.ai From: Raphael Zinsly Date: 2025-09-16 02:34 To: 钟居哲 CC: gcc-patches; kito.cheng; Robin Dapp; jeffreyalaw Subject: Re: RISC-V: Improve slide patterns recognition On Sun, Sep 14, 2025 at 11:12 PM 钟居哲 <juzhe.zh...@rivai.ai> wrote: > > +/* Recognize patterns like [4 5 6 7 12 13 14 15] where a consecutive part of > a > + vector is combined into another. */ > > > Is it better to use vcompress to combine such pattern ? > > [4 5 6 7 x x x x] > [x x x x 12 13 14 15] > combine them use a single vcompress.vvm instruction ? I haven't changed this particular example, shuffle_slide_patterns is called before shuffle_compress_patterns since it was introduced. Looking at the code generated by gcc/testsuite/gcc.target/riscv/rvv/autovec/vls-vlmax/shuffle-slide.c which tests [4 5 6 7 12 13 14 15] (in permute2_vnx8*), if we use vcompress.vm instead of vslidedown.vi we would generate an extra vlm.v. -- Raphael Moreira Zinsly
