>> I'm just saying that to go forward the vectorizer change looks >>more promising (also considering the pace RISC-V people are working at >>...) Yeah, RVV needs a lot of middle-end support: SELECT_VL, LEN_MASK_LOAD/LEN_MASK_STORE,.....etc LEN_ADD for RVV reduction support like COND_ADD for ARM SVE...etc SELECT_VL is still pending. Without support in middle-end, GCC can not support powerful auto-vectorization (Performance will be much worse than RVV LLVM). And unfortunately, I am the only guy working on middle-end support of RVV auto-vectorization. :) I think we can make this patch merged and record the enhancement of SCEV in bugzilla to see we can improve that in the future. Thanks. juzhe.zhong@rivai.ai From: Richard Biener Date: 2023-05-31 15:38 To: Richard Sandiford CC: juzhe.zhong@rivai.ai; gcc-patches; linkw Subject: Re: [PATCH] VECT: Change flow of decrement IV On Wed, 31 May 2023, Richard Sandiford wrote: > Richard Biener writes: > > On Wed, 31 May 2023, juzhe.zhong@rivai.ai wrote: > > > >> Hi?all. I have posted my several investigations: > >> https://gcc.gnu.org/pipermail/gcc-patches/2023-May/620101.html > >> https://gcc.gnu.org/pipermail/gcc-patches/2023-May/620105.html > >> https://gcc.gnu.org/pipermail/gcc-patches/2023-May/620108.html > >> > >> Turns out when "niters is a constant value and vf is a constant value" > >> This patch can allow SCEV/IVOPTS optimize a lot for RVV too (Take tesecase from IBM's testsuite for example) and I think this patch can fix IBM's cunroll issue. > >> Even though it will produce a 'mv' instruction in some ohter cases for RVV, I think Gain > Pain overal. > >> > >> Actually, for current flow: > >> > >> step = MIN () > >> ... > >> remain = remain - step. > >> > >> I don't know how difficult to extend SCEV/IVOPTS to fix this issue. > >> So, could you make a decision for this patch? > >> > >> I wonder whether we should apply the approach of this patch (the codes can be refined after well reviewed) or > >> we should extend SCEV/IVOPTS ? > > > > I don't think we can do anything in SCEV for this which means we'd > > need to special-case this in niter analysis, in IVOPTs and any other > > passes that might be affected (and not fixed by handling it in niter > > analysis). While improving niter analysis would be good (the user > > could write this pattern as well) I do not have time to try > > implementing that (I have no idea how ugly or robust it is going to be). > > > > So I think we should patch this up in the vectorizer itself like with > > your patch. I'm going to wait for Richards input though since he > > seems to disagree. > > I think my main disagreement is that the IV phi can be analysed > as a SCEV with sufficient work (realising that the MIN result is > always VF when the latch is executed). That SCEV might be useful > ?as is? for things like IVOPTS, without specific work in those passes. > (Although perhaps not too useful, since most other IVs will be upcounting.) I think we'd need another API for SCEV there then, analyze_scalar_evolution_for_latch () so we can disregard the value on the exit edges then. That means we'd still need to touch all users and decide whether it's safe to use that or not. > I don't object though. It just feels like we're giving up easily. > And that's a bit frustrating, since this potential problem was flagged > ahead of time. Well, I expect that massaging SCEV and niter analysis will take up quite some developer time while avoiding the situation in the vectorizer is possible (and would fix the observed regressions). We can always improve later here and I'd suggest to file an enhancement bugreport with a simple C testcase using this kind of iteration. I'm just saying that to go forward the vectorizer change looks more promising (also considering the pace RISC-V people are working at ...) Richard. > > Note with SELECT_VL all bets will be off since as I understand the > > value it gives can vary from iteration to iteration (but we know > > a lower and maybe an upper bound?) > > Right. All IVs will have a variable step for SELECT_VL. > > Thanks, > Richard >