Repository navigation
HACL* HMAC does not properly use vectorized BLAKE-2 #138114
Description
Bug report
Bug description:
First of all, it doesn't really make sense to support HMAC-BLAKE-2 because BLAKE-2 is already a keyed hash function so it already has the properties expected by HMAC. Since HACL* provides it (I don't remember if I actually asked them to, so if I did, it's my fault), I decided to expose it anyway.
HACL* provides SIMD-accelerated implementation for BLAKE-2 which is exposed if the HACL_CAN_COMPILE_VEC128 and HACL_CAN_COMPILE_VEC256 macros are defined. CPython, however, needs to do a bit more work to be able to know whether we can define those macros or not. The work is achieved in two steps:
- In
configure, depending on the host capabilities, the macros are defined. - When compiling, depending on the host capabilities once again, the macros are undefined.
The generic headers are afterwards included, e.g.:
#define HACL_CAN_COMPILE_VEC128 _Py_HACL_CAN_COMPILE_VEC128
#define HACL_CAN_COMPILE_VEC256 _Py_HACL_CAN_COMPILE_VEC256
#include "_hacl/Hacl_HMAC.h"
#include "_hacl/Hacl_Streaming_HMAC.h" // Hacl_Agile_Hash_* identifiers
#include "_hacl/Hacl_Streaming_Types.h" // Hacl_Streaming_Types_error_codeNow, there is a slight issue. The conditional code is actually inside the .c file and not the .h. However, that .c file is an HACL* file that is compiled separately and doesn't have any defines. It also unconditionally includes Hacl_Hash_Blake2s_Simd128.h and Hacl_Hash_Blake2s_Simd256.h files. So I'm not entirely sure about what to do :(
Am I incorrectly using the files @msprotz?
CPython versions tested on:
CPython main branch, 3.14, 3.15
Operating systems tested on:
No response
Activity
msprotz commentedon Aug 25, 2025
I seem to remember this was an odd construction expected by Signal (or maybe Noise?). This is why we ended up adding it.
I can't tell from your bug report what the actual problem is (sorry! I probably need more coffee), but here's what I remember about the design and how we fixed it all:
Hacl_Hash_Blake2s_Simd128_block_state_t, and as such, can always be compiled (even if the host does not support emitting SIMD instructions)Does this help? Let me know where the reality doesn't match my description above and I can help dig in. Thanks!
pythongh-138114: Enable HACL BLAKE2 SIMD128 vectorization on PowerPC64