From: Giulio Benetti Date: Fri, 3 Jan 2020 18:24:26 +0100 Subject: [Buildroot] [PATCH] package/libnss: fix ppc 32-bit build failure In-Reply-To: <820d953e-9100-8217-b317-1b226b4aba18@xes-inc.com> References: <20200102181740.116227-1-giulio.benetti@benettiengineering.com> <23dbb404-03fa-f18f-607d-62cf8ef43a5f@xes-inc.com> <820d953e-9100-8217-b317-1b226b4aba18@xes-inc.com> Message-ID: <4ef10e39-bef2-9327-ae2c-ecc9a3943e75@benettiengineering.com> List-Id: MIME-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit To: buildroot@busybox.net Vincent, Thomas, Yann, All, On 1/3/20 5:22 PM, Vincent Fazio wrote: > Giulio, > > On 1/2/20 3:41 PM, Giulio Benetti wrote: >> Hi Vincent, >> >> On 1/2/20 9:54 PM, Vincent Fazio wrote: >>> Giulio, >>> >>> Are you dropping this file for 32bit builds due to the Altivec >>> requirements? General Altivec instructions are available on 32bit chips, >>> such as those from Freescale/NXP. The e600 is a product that comes to >>> mind. IMO, the Makefile should include the file if the architecture is >>> 'ppc'. I think it makes sense that the source file should determine what >>> it compiles down to. It's possible the hw acceleration logic for PPC >>> changes later and this is one more file that would need to be touched. >>> >>> Unfortunately, this is where it gets a bit more complicated... >>> >>> The backing instruction for the vec_xl_be compiler intrinsic is lxvd2x >>> (a VSX instruction). Technically I think this instruction was available >>> in the Power ISA as early as 2.06, but GCC is only advertising the >>> intrinsic as of ISA 3.0 (Power 9+). >>> https://gcc.gnu.org/onlinedocs/gcc-8.3.0/gcc/PowerPC-Built-in-Functions.html#PowerPC-Built-in-Functions >>> >>> >>> I've always had issues reading the logic behind GCC's instruction >>> generation and I don't have 8+ handy... I'm not sure if there's a guard >>> to prevent the lxvd2x instruction from being emitted or not when >>> resolving vec_xl_be and, say, '-maltivec -mpcu=7450' is specified. If >>> there isn't, I'd expect either an illegal instruction exception or for >>> the kernel to emulate it magically. If there is, I'd expect GCC to >>> either compile a compatible instruction shim or to simply error out. >>> >>> Regardless, given the way the hardware acceleration is written currently >>> to require that compiler intrinsic, USE_PPC_CRYPTO may need to be >>> guarded by __builtin_cpu_supports("arch_3_00") >> >> They are all good suggestions but I think they are more NSS related >> than Buildroot related. What I'm trying to do here is to avoid build >> failure in Buildroot and as much as possible to try to contribute >> upstream. As you've mentioned above 32-bit supports Altivec, but the >> problem at the moment is that USE_PPC_CRYPTO is defined in gcm.h only >> if __powerpc64__ is defined, this means that they only tested it for >> that architecture(ppc64 and ppc64le). So at the moment I'd prefer keep >> 32-bit Altivec supported devices without gcm altivec acceleration. But >> if you have a patch that takes into account 32-bits Altivec too, >> please upload it here: >> > For Buildroot, we can likely key off > ifneq($(BR2_POWERPC_CPU_HAS_ALTIVEC),y) Yes, this is a good idea. > and sed delete the '-mcrypto > -maltivec' flags from lib/freebl/Makefile... It's better to avoid this kind of blind hack. I'd prefer to introduce a patch in NSS which adds a NSS_ALTIVEC_DISABLE to avoid building gcm-ppc.c. So combined with your suggestion above we can do something like: ifeq($(BR2_POWERPC_CPU_HAS_ALTIVEC),) NSS_ALTIVEC_DISABLE=1 endif The -mcrypto enables -mvsx > and causes gcc to emit instructions the target CPU may not support, > that's why it should be removed. Luckily, we can make some assumptions > that if ALTIVEC isn't supported, VSX and Crypto functions arent either > since they're introduced in the same or later ISA revisions. So let's assume this. >> https://bugzilla.mozilla.org/show_bug.cgi?id=1606689 >> >> so we fix this problem once! > I'm not sure this can be done "cleanly" given how this is currently > structured. I've added a comment in the bug. Not all 64bit PPC > processors have Altivec (e5500 does not). Yes, I see now. > Plus this goes a bit beyond > that since the functions they're using require at least Power ISA 3.0. > The source package should probably inherit CFLAGS instead of assuming it > can define it given the complexities of PPC cpu tuning. From NSS point of view this would be fixed by introducing NSS_ALTIVEC_DISABLE, but from Buildroot point of view this should be done by adding a new variable like BR2_POWERPC_AT_LEAST_ISA_X_X But I don't know if it's worth the effort. What about above? -- Giulio Benetti Benetti Engineering sas >> >> Maybe we can live chat on BR IRC if you join it. > I'm idling there today >> >> >> Thanks for reviewing so in depth. >> Best regards >