On Wed, Jan 03, 2024 at 03:00:29PM +0100, Ard Biesheuvel wrote: > On Tue, 2 Jan 2024 at 07:50, Eric Biggers <ebiggers@xxxxxxxxxx> wrote: > > > > As discussed previously, the proposed use of the so-called perlasm for > > the RISC-V crypto assembly files makes them difficult to read, and these > > files have some other issues such extensively duplicating source code > > for the different AES key lengths and for the unrolled hash functions. > > There is/was a desire to share code with OpenSSL, but many of the files > > have already diverged significantly; also, for most of the algorithms > > the source code can be quite short anyway, due to the native support for > > them in the RISC-V vector crypto extensions combined with the way the > > RISC-V vector extension naturally scales to arbitrary vector lengths. > > > > Since we're still waiting for prerequisite patches to be merged anyway, > > we have a bit more time to get this cleaned up properly. So I've had a > > go at cleaning up the patchset to use standard .S files, with the code > > duplication fixed. I also made some tweaks to make the different > > algorithms consistent with each other and with what exists in the kernel > > already for other architectures, and tried to improve comments. > > > > The result is this series, which passes all tests and is about 2400 > > lines shorter than the latest version with the perlasm > > (https://lore.kernel.org/linux-crypto/20231231152743.6304-1-jerry.shih@xxxxxxxxxx/). > > All the same functionality and general optimizations are still included, > > except for some minor optimizations in XTS that I dropped since it's not > > clear they are worth the complexity. (Note that almost all users of XTS > > in the kernel only use it with block-aligned messages, so it's not very > > important to optimize the ciphertext stealing case.) > > > > I'd appreciate people's thoughts on this series. Jerry, I hope I'm not > > stepping on your toes too much here, but I think there are some big > > improvements here. > > > > As I have indicated before, I fully agree with Eric here that avoiding > perlasm is preferable: sharing code with OpenSSL is great if we can > simply adopt the exact same code (and track OpenSSL as its upstream) > but this never really worked that well for skciphers, due to API > differences. (The SHA transforms can be reused a bit more easily) > > I will also note that perlasm is not as useful for RISC-V as it is for > other architectures: in OpenSSL, perlasm is also used to abstract > differences in calling conventions between, e.g., x86_64 on Linux vs > Windows, or to support building with outdated [proprietary] > toolchains. > > I do wonder if we could also use .req directives for register aliases > instead of CPP defines? It shouldn't matter for working code, but the > diagnostics tend to be a bit more useful if the aliases are visible to > the assembler. .req unfortunately isn't an option since it is specific to AArch64 and ARM assembly. So we have to use #defines like x86 does. Ultimately, the effect is about the same. - Eric
