STRICT_ALIGNMENT is a remnant of OpenSSL code would cast pointers to
size_t* and load more than one byte at a time. Not all architectures
support unaligned access, so it did an alignment check and only enterred
this path if aligned or the underlying architecture didn't care.
This is UB. Unaligned casts in C are undefined on all architectures, so
we switch these to memcpy some time ago. Compilers can optimize memcpy
to the unaligned accesses we wanted. That left our modes logic as:
- If STRICT_ALIGNMENT is 1 and things are unaligned, work byte-by-byte.
- Otherwise, use the memcpy-based word-by-word code, which now works
independent of STRICT_ALIGNMENT.
Remove the first check to simplify things. On x86, x86_64, and aarch64,
STRICT_ALIGNMENT is zero and this is a no-op. ARM is more complex. Per
[0], ARMv7 and up support unaligned access. ARMv5 do not. ARMv6 does,
but can run in a mode where it looks more like ARMv5.
For ARMv7 and up, STRICT_ALIGNMENT should have been zero, but was one.
Thus this change should be an improvement for ARMv7 (right now unaligned
inputs lose bsaes-armv7). The Android NDK does not even support the
pre-ARMv7 ABI anymore[1]. Nonetheless, Cronet still supports ARMv6 as a
library. It builds with -march=armv6 which GCC interprets as supporting
unaligned access, so it too did not want this code.
For completeness, should anyone still care about ARMv5 or be building
with an overly permissive -march flag, GCC does appear unable to inline
the memcpy calls. However, GCC also does not interpret
(uintptr_t)ptr % sizeof(size_t) as an alignment assertion, so such
consumers have already been paying for the memcpy here and throughout
the library.
In general, C's arcane pointer rules mean we must resort to memcpy
often, so, realistically, we must require that the compiler optimize
memcpy well.
[0] https://medium.com/@iLevex/the-curious-case-of-unaligned-access-on-arm-5dd0ebe24965
[1] https://developer.android.com/ndk/guides/abis#armeabi
Change-Id: I3c7dea562adaeb663032e395499e69530dd8e145
Reviewed-on: https://boringssl-review.googlesource.com/c/34873
Reviewed-by: Adam Langley <agl@google.com>
This one is a little thorny. All the various block cipher modes
functions and callbacks take a void *key. This allows them to be used
with multiple kinds of block ciphers.
However, the implementations of those callbacks are the normal typed
functions, like AES_encrypt. Those take AES_KEY *key. While, at the ABI
level, this is perfectly fine, C considers this undefined behavior.
If we wish to preserve this genericness, we could either instantiate
multiple versions of these mode functions or create wrappers of
AES_encrypt, etc., that take void *key.
The former means more code and is tedious without C++ templates (maybe
someday...). The latter would not be difficult for a compiler to
optimize out. C mistakenly allowed comparing function pointers for
equality, which means a compiler cannot replace pointers to wrapper
functions with the real thing. (That said, the performance-sensitive
bits already act in chunks, e.g. ctr128_f, so the function call overhead
shouldn't matter.)
But our only 128-bit block cipher is AES anyway, so I just switched
things to use AES_KEY throughout. AES is doing fine, and hopefully we
would have the sense not to pair a hypothetical future block cipher with
so many modes!
Change-Id: Ied3e843f0e3042a439f09e655b29847ade9d4c7d
Reviewed-on: https://boringssl-review.googlesource.com/32107
Reviewed-by: Adam Langley <agl@google.com>
Even without strict-aliasing, C does not allow casting pointers to types
that don't match their alignment. After this change, UBSan is happy with
our code at default settings but for the negative left shift language
bug.
Note: architectures without unaligned loads do not generate the same
code for memcpy and pointer casts. But even ARMv6 can perform unaligned
loads and stores (ARMv5 couldn't), so we should be okay here.
Before:
Did 11086000 AES-128-GCM (16 bytes) seal operations in 5000391us (2217026.6 ops/sec): 35.5 MB/s
Did 370000 AES-128-GCM (1350 bytes) seal operations in 5005208us (73923.0 ops/sec): 99.8 MB/s
Did 63000 AES-128-GCM (8192 bytes) seal operations in 5029958us (12525.0 ops/sec): 102.6 MB/s
Did 9894000 AES-256-GCM (16 bytes) seal operations in 5000017us (1978793.3 ops/sec): 31.7 MB/s
Did 316000 AES-256-GCM (1350 bytes) seal operations in 5005564us (63129.7 ops/sec): 85.2 MB/s
Did 54000 AES-256-GCM (8192 bytes) seal operations in 5054156us (10684.3 ops/sec): 87.5 MB/s
After:
Did 11026000 AES-128-GCM (16 bytes) seal operations in 5000197us (2205113.1 ops/sec): 35.3 MB/s
Did 370000 AES-128-GCM (1350 bytes) seal operations in 5005781us (73914.5 ops/sec): 99.8 MB/s
Did 63000 AES-128-GCM (8192 bytes) seal operations in 5032695us (12518.1 ops/sec): 102.5 MB/s
Did 9831750 AES-256-GCM (16 bytes) seal operations in 5000010us (1966346.1 ops/sec): 31.5 MB/s
Did 316000 AES-256-GCM (1350 bytes) seal operations in 5005702us (63128.0 ops/sec): 85.2 MB/s
Did 54000 AES-256-GCM (8192 bytes) seal operations in 5053642us (10685.4 ops/sec): 87.5 MB/s
(Tested with the no-asm builds; most of this code isn't reachable
otherwise.)
Change-Id: I025c365d26491abed0116b0de3b7612159e52297
Reviewed-on: https://boringssl-review.googlesource.com/22804
Reviewed-by: Adam Langley <agl@google.com>
crypto/{asn1,x509,x509v3,pem} were skipped as they are still OpenSSL
style.
Change-Id: I3cd9a60e1cb483a981aca325041f3fbce294247c
Reviewed-on: https://boringssl-review.googlesource.com/19504
Reviewed-by: Adam Langley <agl@google.com>
Commit-Queue: David Benjamin <davidben@google.com>
CQ-Verified: CQ bot account: commit-bot@chromium.org <commit-bot@chromium.org>
The changes to delocate.go are needed because modes/ does things like
return the address of a module function. Both of these need to be
changed from referencing the GOT to using local symbols.
Rather than testing whether |ghash| is |gcm_ghash_avx|, we can just keep
that information in a flag.
The test for |aesni_ctr32_encrypt_blocks| is more problematic, but I
believe that it's superfluous and can be dropped: if you passed in a
stream function that was semantically different from
|aesni_ctr32_encrypt_blocks| you would already have a bug because
|CRYPTO_gcm128_[en|de]crypt_ctr32| will handle a block at the end
themselves, and assume a big-endian, 32-bit counter anyway.
Change-Id: I68a84ebdab6c6006e11e9467e3362d7585461385
Reviewed-on: https://boringssl-review.googlesource.com/15064
Reviewed-by: Adam Langley <agl@google.com>
