From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from foss.arm.com (foss.arm.com [217.140.110.172]) by sourceware.org (Postfix) with ESMTP id 8915D3858C83 for ; Thu, 2 Mar 2023 09:39:08 +0000 (GMT) DMARC-Filter: OpenDMARC Filter v1.4.2 sourceware.org 8915D3858C83 Authentication-Results: sourceware.org; dmarc=pass (p=none dis=none) header.from=arm.com Authentication-Results: sourceware.org; spf=pass smtp.mailfrom=arm.com Received: from usa-sjc-imap-foss1.foss.arm.com (unknown [10.121.207.14]) by usa-sjc-mx-foss1.foss.arm.com (Postfix) with ESMTP id 5382B1FB; Thu, 2 Mar 2023 01:39:51 -0800 (PST) Received: from localhost (e121540-lin.manchester.arm.com [10.32.99.50]) by usa-sjc-imap-foss1.foss.arm.com (Postfix) with ESMTPSA id 29AAD3F99C; Thu, 2 Mar 2023 01:39:07 -0800 (PST) From: Richard Sandiford To: "juzhe.zhong\@rivai.ai" Mail-Followup-To: "juzhe.zhong\@rivai.ai" ,rguenther , pan2.li , gcc-patches , incarnation.p.lee , Kito.cheng , richard.sandiford@arm.com Cc: rguenther , pan2.li , gcc-patches , incarnation.p.lee , Kito.cheng Subject: Re: [PATCH] RISC-V: Bugfix for rvv bool mode precision adjustment References: <9B9312B81334DB23+2023030122191665205362@rivai.ai> <56212E9722A8E5C3+2023030206533064159132@rivai.ai> Date: Thu, 02 Mar 2023 09:39:05 +0000 In-Reply-To: (juzhe's message of "Thu, 2 Mar 2023 16:37:40 +0800") Message-ID: User-Agent: Gnus/5.13 (Gnus v5.13) Emacs/26.3 (gnu/linux) MIME-Version: 1.0 Content-Type: text/plain X-Spam-Status: No, score=-28.3 required=5.0 tests=BAYES_00,KAM_DMARC_NONE,KAM_DMARC_STATUS,KAM_LAZY_DOMAIN_SECURITY,SPF_HELO_NONE,SPF_NONE,TXREP autolearn=no autolearn_force=no version=3.4.6 X-Spam-Checker-Version: SpamAssassin 3.4.6 (2021-04-09) on server2.sourceware.org List-Id: Thanks for the explanation about the sizes. "juzhe.zhong@rivai.ai" writes: > Fortunately, we won't have aggregates, arrays of vbool*_t in the future. > I think it's not an issue. But isn't it possible to allocate a char/byte array and construct vbool*_ts at addresses calculated by intrinsics? E.g. I don't see anything wrong in principle with doing: #include void f(char *x, svbool_t p1, svbool_t p2) { *(svbool_t *)(x + svcntd()) = p2; *(svbool_t *)(x) = p1; } If the mode size for svbool_t was too big, I think RTL DSE would be within its rights to delete the first store. (Precision doesn't matter, at least not currently.) There's no problem if the ABI is defined such that vbool8_t has the same size as the GET_MODE_SIZE recorded in GCC. (But of course, it would need to be consistently so, even when the vector length is known at compile time.) In that case, the difference between the size stored by the machine and the size used by the ABI would be padding, and there is no requirement to preserve padding. But if the ABI size of vbool8_t matches the machine behaviour, I think making GCC's size bigger risks wrong code. I realise it's a corner case. But I don't think making GET_MODE_SIZE bigger than the real size is conservatively correct. Thanks, Richard > > > juzhe.zhong@rivai.ai > > From: Richard Biener > Date: 2023-03-02 16:25 > To: juzhe.zhong > CC: richard.sandiford; pan2.li; gcc-patches; Pan Li; kito.cheng > Subject: Re: Re: [PATCH] RISC-V: Bugfix for rvv bool mode precision adjustment > On Thu, 2 Mar 2023, juzhe.zhong@rivai.ai wrote: > >> >> Does the eventual value set by ADJUST_BYTESIZE equal the real number of >> >> bytes loaded by vlm.v and stored by vstm.v (after the appropriate vsetvl)? >> >> Or is the GCC size larger in some cases than the number of bytes >> >> loaded and stored? >> For VNx1BI,VNx2BI,VNx4BI,VNx8BI, we allocate the larger size of memory or stack for register spillling >> according to ADJUST_BYTESIZE. >> After appropriate vsetvl, VNx1BI is loaded/stored 1/8 of ADJUST_BYTESIZE (vsetvl e8mf8). >> After appropriate vsetvl, VNx2BI is loaded/stored 2/8 of ADJUST_BYTESIZE (vsetvl e8mf2). >> After appropriate vsetvl, VNx4BI is loaded/stored 4/8 of ADJUST_BYTESIZE (vsetvl e8mf4). >> After appropriate vsetvl, VNx8BI is loaded/stored 8/8 of ADJUST_BYTESIZE (vsetvl e8m1). >> >> Note: except these 4 machine modes, all other machine modes of RVV, ADJUST_BYTESIZE >> are equal to the real number of bytes of load/store instruction that RVV ISA define. >> >> Well, as I said, it's fine that we allocated larger memory for VNx1BI,VNx2BI,VNx4BI, >> we can emit appropriate vsetvl to gurantee the correctness in RISC-V backward according >> to the machine_mode as long as long GCC didn't do the incorrect elimination in middle-end. >> >> Besides, poly (1,1) is 1/8 of machine vector-length which is already really a small number, >> which is the real number bytes loaded/stored for VNx8BI. >> You can say VNx1BI, VNx2BI, VNx4BI are consuming larger memory than we actually load/stored by appropriate vsetvl >> since they are having same ADJUST_BYTESIZE as VNx8BI. However, I think it's totally fine so far as long as we can >> gurantee the correctness and I think optimizing such memory storage consuming is trivial. >> >> >> And does it equal the size of the corresponding LLVM machine type? >> >> Well, for some reason, in case of register spilling, LLVM consume much more memory than GCC. >> And they always do whole register load/store (a single vector register vector-length) for register spilling. >> That's another story (I am not going to talk too much about this since it's a quite ugly implementation). >> They don't model the types accurately according RVV ISA for register spilling. >> >> In case of normal load/store like: >> vbool8_t v2 = *(vbool8_t*)in; *(vbool8_t*)(out + 100) = v2; >> This kind of load/store, their load/stores instructions of codegen are accurate. >> Even though their instructions are accurate for load/store accessing behavior, I am not sure whether size >> of their machine type is accurate. >> >> For example, in IR presentation: VNx1BI of GCC is represented as vscale x 1 x i1 >> VNx2BI of GCC is represented as vscale x 2 x i1 >> in LLVM IR. >> I am not sure the bytesize of vscale x 1 x i1 and vscale x 2 x i1. >> I didn't take a deep a look at it. >> >> I think this question is not that important, no matter whether VNx1BI and VNx2BI are modeled accurately in case of ADUST_BYTESIZE >> in GCC or vscale x 1 x i1 and vscale x 2 x i1 are modeled accurately in case of their bytesize, >> I think as long as we can emit appropriate vsetvl + vlm/vsm, it's totally fine for RVV even though in some case, their memory allocation >> is not accurate in compiler. > > I'm not sure how it works for variable-length types but isn't > sizeof (vbool8_t) part of the ABI and thus its TYPE_SIZE / GET_MODE_SIZE > are relevant there? It might of course be that you can never have > these types as part of aggregates, arrays or objects of them address-taken > in which case the issue is moot? > > Richard. > >> >> juzhe.zhong@rivai.ai >> >> From: Richard Sandiford >> Date: 2023-03-02 00:14 >> To: Li\, Pan2 >> CC: juzhe.zhong\@rivai.ai; rguenther; gcc-patches; Pan Li; kito.cheng >> Subject: Re: [PATCH] RISC-V: Bugfix for rvv bool mode precision adjustment >> "Li, Pan2" writes: >> > Thanks all for so much valuable and helpful materials. >> > >> > As I understand (Please help to correct me if any mistake.), for the VNx*BI (aka, 1, 2, 4, 8, 16, 32, 64), >> > the precision and mode size need to be adjusted as below. >> > >> > Precision size [1, 2, 4, 8, 16, 32, 64] >> > Mode size [1, 1, 1, 1, 2, 4, 8] >> > >> > Given that, if we ignore the self-test failure, only the adjust_precision part is able to fix the bug I mentioned. >> > The genmode will first get the precision, and then leverage the mode_size = exact_div / 8 to generate. >> > Meanwhile, it also provides the adjust_mode_size after the mode_size generation. >> > >> > The riscv parts has the mode_size_adjust already and the value of mode_size will be overridden by the adjustments. >> >> Ah, OK! In that case, would the following help: >> >> Turn: >> >> mode_size[E_%smode] = exact_div (mode_precision[E_%smode], BITS_PER_UNIT); >> >> into: >> >> if (!multiple_p (mode_precision[E_%smode], BITS_PER_UNIT, >> &mode_size[E_%smode])) >> mode_size[E_%smode] = -1; >> >> where -1 is an "obviously wrong" value. >> >> Ports that might hit the -1 are then responsible for setting the size >> later, via ADJUST_BYTESIZE. >> >> After all the adjustments are complete, genmodes asserts that no size is >> known_eq to -1. >> >> That way, target-independent code doesn't need to guess what the >> correct behaviour is. >> >> Does the eventual value set by ADJUST_BYTESIZE equal the real number of >> bytes loaded by vlm.v and stored by vstm.v (after the appropriate vsetvl)? >> And does it equal the size of the corresponding LLVM machine type? >> Or is the GCC size larger in some cases than the number of bytes >> loaded and stored? >> >> (You and Juzhe have probably answered that question before, sorry, >> but I'm still not 100% sure of the answer. Personally, I think I would >> find the ISA behaviour easier to understand if the explanation doesn't >> involve poly_ints. It would be good to understand things "as the >> architecture sees then" rather than in terms of GCC concepts.) >> >> Thanks, >> Richard >> >> > Unfortunately, the early stage mode_size generation leveraged exact_div, which doesn't honor precision size < 8 >> > with the adjustment and fails on exact_div assertions. >> > >> > Besides the precision adjustment, I am not sure if we can narrow down the problem to. >> > >> > >> > 1. Defined the real size of both the precision and mode size to align the riscv ISA. >> > 2. Besides, make the general mode_size = precision_size / 8 is able to take care of both the exact_div and the dividend less than the divisor (like 1/8 or 2/8) cases. >> > >> > Could you please share your professional suggestions about this? Thank you all again and have a nice day! >> > >> > Pan >> > >> > From: juzhe.zhong@rivai.ai >> > Sent: Wednesday, March 1, 2023 10:19 PM >> > To: rguenther >> > Cc: richard.sandiford ; gcc-patches ; Pan Li ; Li, Pan2 ; kito.cheng >> > Subject: Re: Re: [PATCH] RISC-V: Bugfix for rvv bool mode precision adjustment >> > >> >>> So given the above I think that modeling the size as being the same >> >>> but with accurate precision would work. It's then only the size of the >> >>> padding in bytes we cannot represent with poly-int which should be fine. >> > >> >>> Correct? >> > Yes. >> > >> >>> Btw, is storing a VNx1BI and then loading a VNx2BI from the same >> >>> memory address well-defined? That is, how is the padding handled >> >>> by the machine load/store instructions? >> > >> > storing VNx1BI is storing the data from addr 0 ~ 1/8 poly (1,1) and keep addr 1/8 poly (1,1) ~ 2/8 poly (1,1) memory data unchange. >> > load VNx2BI will load 0 ~ 2/8 poly (1,1), note that 0 ~ 1/8 poly (1,1) is the date that we store above, 1/8 poly (1,1) ~ 2/8 poly (1,1) is the orignal memory data. >> > You can see here for this case (LLVM): >> > https://godbolt.org/z/P9e1adrd3 >> > foo: # @foo >> > vsetvli a2, zero, e8, mf8, ta, ma >> > vsm.v v0, (a0) >> > vsetvli a2, zero, e8, mf4, ta, ma >> > vlm.v v8, (a0) >> > vsm.v v8, (a1) >> > ret >> > >> > We can also doing like this in GCC as long as we can differentiate VNx1BI and VNx2BI, and GCC do not eliminate statement according precision even though >> > they have same bytesize. >> > >> > First we emit vsetvl e8mf8 +vsm for VNx1BI >> > Then we emit vsetvl e8mf8 + vlm for VNx2BI >> > >> > Thanks. >> > ________________________________ >> > juzhe.zhong@rivai.ai >> > >> > From: Richard Biener >> > Date: 2023-03-01 22:03 >> > To: juzhe.zhong >> > CC: richard.sandiford; gcc-patches; Pan Li; pan2.li; kito.cheng >> > Subject: Re: Re: [PATCH] RISC-V: Bugfix for rvv bool mode precision adjustment >> > On Wed, 1 Mar 2023, Richard Biener wrote: >> > >> >> On Wed, 1 Mar 2023, juzhe.zhong@rivai.ai wrote: >> >> >> >> > Let's me first introduce RVV load/store basics and stack allocation. >> >> > For scalable vector memory allocation, we allocate memory according to machine vector-length. >> >> > To get this CPU vector-length value (runtime invariant but compile time unknown), we have an instruction call csrr vlenb. >> >> > For example, csrr a5,vlenb (store CPU a single register vector-length value (describe as bytesize) in a5 register). >> >> > A single register size in bytes (GET_MODE_SIZE) is poly value (8,8) bytes. That means csrr a5,vlenb, a5 has the value of size poly (8,8) bytes. >> >> > >> >> > Now, our problem is that VNx1BI, VNx2BI, VNx4BI, VNx8BI has the same bytesize poly (1,1). So their storage consumes the same size. >> >> > Meaning when we want to allocate a memory storge or stack for register spillings, we should first csrr a5, vlenb, then slli a5,a5,3 (means a5 = a5/8) >> >> > Then, a5 has the bytesize value of poly (1,1). All VNx1BI, VNx2BI, VNx4BI, VNx8BI are doing the same process as I described above. They all consume >> >> > the same memory storage size since we can't model them accurately according to precision or you bitsize. >> >> > >> >> > They consume the same storage (I am agree it's better to model them more accurately in case of memory storage comsuming). >> >> > >> >> > Well, even though they are consuming same size memory storage, I can make their memory accessing behavior (load/store) accurately by >> >> > emiting the accurate RVV instruction for them according to RVV ISA. >> >> > >> >> > VNx1BI,VNx2BI, VNx4BI, VNx8BI are consuming same memory storage with size poly (1,1) >> >> > The instruction for these modes as follows: >> >> > VNx1BI: vsevl e8mf8 + vlm, loading 1/8 of poly (1,1) storage. >> >> > VNx2BI: vsevl e8mf8 + vlm, loading 1/4 of poly (1,1) storage. >> >> > VNx4BI: vsevl e8mf8 + vlm, loading 1/2 of poly (1,1) storage. >> >> > VNx8BI: vsevl e8mf8 + vlm, loading 1 of poly (1,1) storage. >> >> > >> >> > So base on these, It's fine that we don't model VNx1BI,VNx2BI, VNx4BI, VNx8BI accurately according to precision or bitsize. >> >> > This implementation is fine even though their memory storage is not accurate. >> >> > >> >> > However, the problem is that since they have the same bytesize, GCC will think they are the same and do some incorrect statement elimination: >> >> > >> >> > (Note: Load same memory base) >> >> > load v0 VNx1BI from base0 >> >> > load v1 VNx2BI from base0 >> >> > load v2 VNx4BI from base0 >> >> > load v3 VNx8BI from base0 >> >> > >> >> > store v0 base1 >> >> > store v1 base2 >> >> > store v2 base3 >> >> > store v3 base4 >> >> > >> >> > This program sequence, in GCC, it will eliminate the last 3 load instructions. >> >> > >> >> > Then it will become: >> >> > >> >> > load v0 VNx1BI from base0 ===> vsetvl e8mf8 + vlm (only load 1/8 of poly size (1,1) memory data) >> >> > >> >> > store v0 base1 >> >> > store v0 base2 >> >> > store v0 base3 >> >> > store v0 base4 >> >> > >> >> > This is what we want to fix. I think as long as we can have the way to differentiate VNx1BI,VNx2BI, VNx4BI, VNx8BI >> >> > and GCC will not do th incorrect elimination for RVV. >> >> > >> >> > I think it can work fine even though these 4 modes consume inaccurate memory storage size >> >> > but accurate data memory access load store behavior. >> >> >> >> So given the above I think that modeling the size as being the same >> >> but with accurate precision would work. It's then only the size of the >> >> padding in bytes we cannot represent with poly-int which should be fine. >> >> >> >> Correct? >> > >> > Btw, is storing a VNx1BI and then loading a VNx2BI from the same >> > memory address well-defined? That is, how is the padding handled >> > by the machine load/store instructions? >> > >> > Richard. >> >>