AVX-512 supports 3-input XORs via the vpternlogd (or vpternlogq)
instruction with immediate 0x96. This approach, vs. the alternative of
two vpxor instructions, is already used in the CRC, AES-GCM, and AES-XTS
code, since it reduces the instruction count and is faster on some CPUs.
Make blake2s_compress_avx512() take advantage of it too.
Reviewed-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20251102234209.62133-7-ebiggers@kernel.org
Signed-off-by: Eric Biggers <ebiggers@kernel.org>
Just before returning, blake2s_compress_ssse3() and
blake2s_compress_avx512() store updated values to the 'h', 't', and 'f'
fields of struct blake2s_ctx. But 'f' is always unchanged (which is
correct; only the C code changes it). So, there's no need to write to
'f'. Use 64-bit stores (movq and vmovq) instead of 128-bit stores
(movdqu and vmovdqu) so that only 't' is written.
Reviewed-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20251102234209.62133-6-ebiggers@kernel.org
Signed-off-by: Eric Biggers <ebiggers@kernel.org>
Various cleanups for readability. No change to the generated code:
- Add some comments
- Add #defines for arguments
- Rename some labels
- Use decimal constants instead of hex where it makes sense.
(The pshufd immediates intentionally remain as hex.)
- Add blank lines when there's a logical break
The round loop still could use some work, but this is at least a start.
Reviewed-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20251102234209.62133-5-ebiggers@kernel.org
Signed-off-by: Eric Biggers <ebiggers@kernel.org>
Following the usual practice, prefix the names of the data labels with
".L" so that the assembler treats them as truly local. This more
clearly expresses the intent and is less error-prone.
Reviewed-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20251102234209.62133-4-ebiggers@kernel.org
Signed-off-by: Eric Biggers <ebiggers@kernel.org>
Since blake2s_compress() is always passed nblocks != 0, remove the
unnecessary check for nblocks == 0 from blake2s_compress_ssse3().
Note that this makes it consistent with blake2s_compress_avx512() in the
same file as well as the arm32 blake2s_compress().
Reviewed-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20251102234209.62133-3-ebiggers@kernel.org
Signed-off-by: Eric Biggers <ebiggers@kernel.org>
In the C code, the 'inc' argument to the assembly functions
blake2s_compress_ssse3() and blake2s_compress_avx512() is declared with
type u32, matching blake2s_compress(). The assembly code then reads it
from the 64-bit %rcx. However, the ABI doesn't guarantee zero-extension
to 64 bits, nor do gcc or clang guarantee it. Therefore, fix these
functions to read this argument from the 32-bit %ecx.
In theory, this bug could have caused the wrong 'inc' value to be used,
causing incorrect BLAKE2s hashes. In practice, probably not: I've fixed
essentially this same bug in many other assembly files too, but there's
never been a real report of it having caused a problem. In x86_64, all
writes to 32-bit registers are zero-extended to 64 bits. That results
in zero-extension in nearly all situations. I've only been able to
demonstrate a lack of zero-extension with a somewhat contrived example
involving truncation, e.g. when the C code has a u64 variable holding
0x1234567800000040 and passes it as a u32 expecting it to be truncated
to 0x40 (64). But that's not what the real code does, of course.
Fixes: ed0356eda1 ("crypto: blake2s - x86_64 SIMD implementation")
Cc: stable@vger.kernel.org
Reviewed-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20251102234209.62133-2-ebiggers@kernel.org
Signed-off-by: Eric Biggers <ebiggers@kernel.org>
Save 480 bytes of .rodata by replacing the .long constants with .bytes,
and using the vpmovzxbd instruction to expand them.
Also update the code to do the loads before incrementing %rax rather
than after. This avoids the need for the first load to use an offset.
Reviewed-by: Ard Biesheuvel <ardb@kernel.org>
Link: https://lore.kernel.org/r/20250827151131.27733-8-ebiggers@kernel.org
Signed-off-by: Eric Biggers <ebiggers@kernel.org>
Move the contents of arch/x86/lib/crypto/ into lib/crypto/x86/.
The new code organization makes a lot more sense for how this code
actually works and is developed. In particular, it makes it possible to
build each algorithm as a single module, with better inlining and dead
code elimination. For a more detailed explanation, see the patchset
which did this for the CRC library code:
https://lore.kernel.org/r/20250607200454.73587-1-ebiggers@kernel.org/.
Also see the patchset which did this for SHA-512:
https://lore.kernel.org/linux-crypto/20250616014019.415791-1-ebiggers@kernel.org/
This is just a preparatory commit, which does the move to get the files
into their new location but keeps them building the same way as before.
Later commits will make the actual improvements to the way the
arch-optimized code is integrated for each algorithm.
Add a gitignore entry for the removed directory arch/x86/lib/crypto/ so
that people don't accidentally commit leftover generated files.
Acked-by: Ard Biesheuvel <ardb@kernel.org>
Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com>
Reviewed-by: Sohil Mehta <sohil.mehta@intel.com>
Link: https://lore.kernel.org/r/20250619191908.134235-9-ebiggers@kernel.org
Signed-off-by: Eric Biggers <ebiggers@kernel.org>
Eric Biggers