Add 128 bit lsx vector expansion optimization code of Loongarch64 architecture
to AES. The test result on the 3A5000 improves performance by about 40%~50%.
Signed-off-by: zhuchen <zhuchen@loongson.cn>
Reviewed-by: Tomas Mraz <tomas@openssl.org>
Reviewed-by: Paul Dale <pauli@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/19364)
Also fix conditional branch out of range when using sanitisers.
Fixes#18813
Signed-off-by: Tom Cosgrove <tom.cosgrove@arm.com>
Change-Id: Ic543885091ed3ef2ddcbe21de0a4ac0bca1e2494
Reviewed-by: Paul Dale <pauli@openssl.org>
Reviewed-by: Matt Caswell <matt@openssl.org>
Reviewed-by: Tomas Mraz <tomas@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/18816)
This restores the implementation prior to
commit 2621751 ("aes/asm/aesv8-armx.pl: avoid 32-bit lane assignment in CTR mode")
for 64bit targets only, since it is reportedly 2-17% slower,
and the silicon errata only affects 32bit targets.
Only for 32bit targets the new algorithm is used.
Fixes#18445
Reviewed-by: Paul Dale <pauli@openssl.org>
Reviewed-by: Tomas Mraz <tomas@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/18581)
aesni_ocb_encrypt and aesni_ocb_decrypt operate by having a fast-path
that performs operations on 6 16-byte blocks concurrently (the
"grandloop") and then proceeds to handle the "short" tail (which can
be anywhere from 0 to 5 blocks) that remain.
As part of initialization, the assembly initializes $len to the true
length, less 96 bytes and converts it to a pointer so that the $inp
can be compared to it. Each iteration of "grandloop" checks to see if
there's a full 96-byte chunk to process, and if so, continues. Once
this has been exhausted, it falls through to "short", which handles
the remaining zero to five blocks.
Unfortunately, the jump at the end of "grandloop" had a fencepost
error, doing a `jb` ("jump below") rather than `jbe` (jump below or
equal). This should be `jbe`, as $inp is pointing to the *end* of the
chunk currently being handled. If $inp == $len, that means that
there's a whole 96-byte chunk waiting to be handled. If $inp > $len,
then there's 5 or fewer 16-byte blocks left to be handled, and the
fall-through is intended.
The net effect of `jb` instead of `jbe` is that the last 16-byte block
of the last 96-byte chunk was completely omitted. The contents of
`out` in this position were never written to. Additionally, since
those bytes were never processed, the authentication tag generated is
also incorrect.
The same fencepost error, and identical logic, exists in both
aesni_ocb_encrypt and aesni_ocb_decrypt.
This addresses CVE-2022-2097.
Co-authored-by: Alejandro Sedeño <asedeno@google.com>
Co-authored-by: David Benjamin <davidben@google.com>
Reviewed-by: Paul Dale <pauli@openssl.org>
Reviewed-by: Tomas Mraz <tomas@openssl.org>
This implementation is based on the four-table approach, along the same
lines as the non-constant-time implementation in aes_core.c The
implementation is in perlasm.
Utility functions are defined to automatically stack/unstack registers
as needed for prologues and epilogues. See riscv-elf-psabi-doc at
https://github.com/riscv-non-isa/riscv-elf-psabi-doc/ for ABI details.
Reviewed-by: Philipp Tomsich <philipp.tomsich@vrull.eu>
Signed-off-by: Henry Brausen <henry.brausen@vrull.eu>
Reviewed-by: Tomas Mraz <tomas@openssl.org>
Reviewed-by: Paul Dale <pauli@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/17640)
gcc6.3 doesn't seem to support the register aliases fp and lr for x29 and x30,
so use the x names.
Fixes#18114
Change-Id: I077edda42af4c7cdb7b24f28ac82d1603f550108
Reviewed-by: Shane Lontis <shane.lontis@oracle.com>
Reviewed-by: Tomas Mraz <tomas@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/18127)
d6e4287c97 introduced 5x interleaving as an
optimization for ThunderX2, and that leads to some performance degradation on
when encoding short buffers. We found this performance degradation by measuring
the performance of nginx on Ubuntu 20.04 that comes with OpenSSL 1.1.1f and
Ubuntu 22.04 with OpenSSL 3.0.1.
This patch limits the 5x interleave to buffers larger than 512 bytes.
On Graviton2 we see the following performance with this patch:
$ openssl speed -evp aes-128-gcm -bytes 128
AES-128-GCM 64 bytes 79 bytes 80 bytes 128 bytes 256 bytes 511 bytes 512 bytes 1024 bytes
master 1062564.71k 775113.11k 1069959.33k 1411716.28k 1653114.86k 1585981.16k 1973683.03k 2203214.08k
master+patch 1062729.28k 771915.11k 1103883.42k 1458665.43k 1708701.20k 1647060.84k 1975571.80k 2204038.42k
diff 0% 0% 3% 3% 3% 4% 0% 0%
revert d6e428 1055290.03k 773448.92k 1117411.97k 1441478.57k 1695698.52k 1634598.04k 1981851.65k 2196680.36k
CLA: trivial
Reviewed-by: Tomas Mraz <tomas@openssl.org>
Reviewed-by: Paul Dale <pauli@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/17984)
The sweep of the source tree in #17373 missed the BSAES assembly due its
PR #14592 having been temporarily backed out at the time.
This constitutes a partial fix for #17958 - covers cases except when
configured with -DOPENSSL_AES_CONST_TIME.
Reviewed-by: Matt Caswell <matt@openssl.org>
Reviewed-by: Tomas Mraz <tomas@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/17988)
GCC's __ARMEL__ and __ARMEB__ defines denote little- and big-endian arm,
respectively. They are not defined on aarch64, which instead use
__AARCH64EL__ and __AARCH64EB__.
However, OpenSSL's assembly originally used the 32-bit defines on both
platforms and even define __ARMEL__ and __ARMEB__ in arm_arch.h. This is
less portable and can even interfere with other headers, which use
__ARMEL__ to detect little-endian arm.
Over time, the aarch64 assembly has switched to the correct defines,
such as in 32bbb62ea6. This commit
finishes the job: poly1305-armv8.pl needed a fix and the dual-arch
armx.pl files get one more transform to convert from 32-bit to 64-bit.
(There is an even more official endianness detector, __ARM_BIG_ENDIAN in
the Arm C Language Extensions. But I've stuck with the GCC ones here as
that would be a larger change.)
Reviewed-by: Matt Caswell <matt@openssl.org>
Reviewed-by: Tomas Mraz <tomas@openssl.org>
Reviewed-by: Paul Dale <pauli@openssl.org>
Reviewed-by: Bernd Edlinger <bernd.edlinger@hotmail.de>
(Merged from https://github.com/openssl/openssl/pull/17373)
This change adds optional support for
- Armv8.3-A Pointer Authentication (PAuth) and
- Armv8.5-A Branch Target Identification (BTI)
features to the perl scripts.
Both features can be enabled with additional compiler flags.
Unless any of these are enabled explicitly there is no code change at
all.
The extensions are briefly described below. Please read the appropriate
chapters of the Arm Architecture Reference Manual for the complete
specification.
Scope
-----
This change only affects generated assembly code.
Armv8.3-A Pointer Authentication
--------------------------------
Pointer Authentication extension supports the authentication of the
contents of registers before they are used for indirect branching
or load.
PAuth provides a probabilistic method to detect corruption of register
values. PAuth signing instructions generate a Pointer Authentication
Code (PAC) based on the value of a register, a seed and a key.
The generated PAC is inserted into the original value in the register.
A PAuth authentication instruction recomputes the PAC, and if it matches
the PAC in the register, restores its original value. In case of a
mismatch, an architecturally unmapped address is generated instead.
With PAuth, mitigation against ROP (Return-oriented Programming) attacks
can be implemented. This is achieved by signing the contents of the
link-register (LR) before it is pushed to stack. Once LR is popped,
it is authenticated. This way a stack corruption which overwrites the
LR on the stack is detectable.
The PAuth extension adds several new instructions, some of which are not
recognized by older hardware. To support a single codebase for both pre
Armv8.3-A targets and newer ones, only NOP-space instructions are added
by this patch. These instructions are treated as NOPs on hardware
which does not support Armv8.3-A. Furthermore, this patch only considers
cases where LR is saved to the stack and then restored before branching
to its content. There are cases in the code where LR is pushed to stack
but it is not used later. We do not address these cases as they are not
affected by PAuth.
There are two keys available to sign an instruction address: A and B.
PACIASP and PACIBSP only differ in the used keys: A and B, respectively.
The keys are typically managed by the operating system.
To enable generating code for PAuth compile with
-mbranch-protection=<mode>:
- standard or pac-ret: add PACIASP and AUTIASP, also enables BTI
(read below)
- pac-ret+b-key: add PACIBSP and AUTIBSP
Armv8.5-A Branch Target Identification
--------------------------------------
Branch Target Identification features some new instructions which
protect the execution of instructions on guarded pages which are not
intended branch targets.
If Armv8.5-A is supported by the hardware, execution of an instruction
changes the value of PSTATE.BTYPE field. If an indirect branch
lands on a guarded page the target instruction must be one of the
BTI <jc> flavors, or in case of a direct call or jump it can be any
other instruction. If the target instruction is not compatible with the
value of PSTATE.BTYPE a Branch Target Exception is generated.
In short, indirect jumps are compatible with BTI <j> and <jc> while
indirect calls are compatible with BTI <c> and <jc>. Please refer to the
specification for the details.
Armv8.3-A PACIASP and PACIBSP are implicit branch target
identification instructions which are equivalent with BTI c or BTI jc
depending on system register configuration.
BTI is used to mitigate JOP (Jump-oriented Programming) attacks by
limiting the set of instructions which can be jumped to.
BTI requires active linker support to mark the pages with BTI-enabled
code as guarded. For ELF64 files BTI compatibility is recorded in the
.note.gnu.property section. For a shared object or static binary it is
required that all linked units support BTI. This means that even a
single assembly file without the required note section turns-off BTI
for the whole binary or shared object.
The new BTI instructions are treated as NOPs on hardware which does
not support Armv8.5-A or on pages which are not guarded.
To insert this new and optional instruction compile with
-mbranch-protection=standard (also enables PAuth) or +bti.
When targeting a guarded page from a non-guarded page, weaker
compatibility restrictions apply to maintain compatibility between
legacy and new code. For detailed rules please refer to the Arm ARM.
Compiler support
----------------
Compiler support requires understanding '-mbranch-protection=<mode>'
and emitting the appropriate feature macros (__ARM_FEATURE_BTI_DEFAULT
and __ARM_FEATURE_PAC_DEFAULT). The current state is the following:
-------------------------------------------------------
| Compiler | -mbranch-protection | Feature macros |
+----------+---------------------+--------------------+
| clang | 9.0.0 | 11.0.0 |
+----------+---------------------+--------------------+
| gcc | 9 | expected in 10.1+ |
-------------------------------------------------------
Available Platforms
------------------
Arm Fast Model and QEMU support both extensions.
https://developer.arm.com/tools-and-software/simulation-models/fast-modelshttps://www.qemu.org/
Implementation Notes
--------------------
This change adds BTI landing pads even to assembly functions which are
likely to be directly called only. In these cases, landing pads might
be superfluous depending on what code the linker generates.
Code size and performance impact for these cases would be negligible.
Interaction with C code
-----------------------
Pointer Authentication is a per-frame protection while Branch Target
Identification can be turned on and off only for all code pages of a
whole shared object or static binary. Because of these properties if
C/C++ code is compiled without any of the above features but assembly
files support any of them unconditionally there is no incompatibility
between the two.
Useful Links
------------
To fully understand the details of both PAuth and BTI it is advised to
read the related chapters of the Arm Architecture Reference Manual
(Arm ARM):
https://developer.arm.com/documentation/ddi0487/latest/
Additional materials:
"Providing protection for complex software"
https://developer.arm.com/architectures/learn-the-architecture/providing-protection-for-complex-software
Arm Compiler Reference Guide Version 6.14: -mbranch-protection
https://developer.arm.com/documentation/101754/0614/armclang-Reference/armclang-Command-line-Options/-mbranch-protection?lang=en
Arm C Language Extensions (ACLE)
https://developer.arm.com/docs/101028/latest
Addional Notes
--------------
This patch is a copy of the work done by Tamas Petz in boringssl. It
contains the changes from the following commits:
aarch64: support BTI and pointer authentication in assembly
Change-Id: I4335f92e2ccc8e209c7d68a0a79f1acdf3aeb791
URL: https://boringssl-review.googlesource.com/c/boringssl/+/42084
aarch64: Improve conditional compilation
Change-Id: I14902a64e5f403c2b6a117bc9f5fb1a4f4611ebf
URL: https://boringssl-review.googlesource.com/c/boringssl/+/43524
aarch64: Fix name of gnu property note section
Change-Id: I6c432d1c852129e9c273f6469a8b60e3983671ec
URL: https://boringssl-review.googlesource.com/c/boringssl/+/44024
Change-Id: I2d95ebc5e4aeb5610d3b226f9754ee80cf74a9af
Reviewed-by: Paul Dale <pauli@openssl.org>
Reviewed-by: Tomas Mraz <tomas@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/16674)
ARM Cortex-A57 and Cortex-A72 cores running in 32-bit mode are affected
by silicon errata #1742098 [0] and #1655431 [1], respectively, where the
second instruction of a AES instruction pair may execute twice if an
interrupt is taken right after the first instruction consumes an input
register of which a single 32-bit lane has been updated the last time it
was modified.
This is not such a rare occurrence as it may seem: in counter mode, only
the least significant 32-bit word is incremented in the absence of a
carry, which makes our counter mode implementation susceptible to these
errata.
So let's shuffle the counter assignments around a bit so that the most
recent updates when the AES instruction pair executes are 128-bit wide.
[0] ARM-EPM-049219 v23 Cortex-A57 MPCore Software Developers Errata Notice
[1] ARM-EPM-012079 v11.0 Cortex-A72 MPCore Software Developers Errata Notice
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@arm.com>
Reviewed-by: Paul Dale <paul.dale@oracle.com>
Reviewed-by: Tomas Mraz <tmraz@fedoraproject.org>
(Merged from https://github.com/openssl/openssl/pull/13504)
Aes-xts mode can be optimized by interleaving cipher operation on
several blocks and loop unrolling. Interleaving needs one ideal
unrolling factor, here we adopt the same factor with aes-cbc,
which is described as below:
If blocks number > 5, select 5 blocks as one iteration,every
loop, decrease the blocks number by 5.
If left blocks < 5, treat them as tail blocks.
Detailed implementation has a little adjustment for squeezing
code space.
With this way, for small size such as 16 bytes, the performance is
similar as before, but for big size such as 16k bytes, the performance
improves a lot, even reaches to 2x uplift, for some arches such as A57,
the improvement even reaches more than 2x uplift. We collect many
performance datas on different micro-archs such as thunderx2,
ampere-emag, a72, a75, a57, a53 and N1, all of which reach 0.5-2x uplift.
The following table lists the encryption performance data on aarch64,
take a72, a75, a57, a53 and N1 as examples. Performance value takes the
unit of cycles per byte, takes the format as comparision of values.
List them as below:
A72:
Before optimization After optimization Improve
evp-aes-128-xts@16 8.899913518 5.949087263 49.60%
evp-aes-128-xts@64 4.525512668 3.389141845 33.53%
evp-aes-128-xts@256 3.502906908 1.633573479 114.43%
evp-aes-128-xts@1024 3.174210419 1.155952639 174.60%
evp-aes-128-xts@8192 3.053019303 1.028134888 196.95%
evp-aes-128-xts@16384 3.025292462 1.02021169 196.54%
evp-aes-256-xts@16 9.971105023 6.754233758 47.63%
evp-aes-256-xts@64 4.931479093 3.786527393 30.24%
evp-aes-256-xts@256 3.746788153 1.943975947 92.74%
evp-aes-256-xts@1024 3.401743802 1.477394648 130.25%
evp-aes-256-xts@8192 3.278769327 1.32950421 146.62%
evp-aes-256-xts@16384 3.27093296 1.325276257 146.81%
A75:
Before optimization After optimization Improve
evp-aes-128-xts@16 8.397965173 5.126839098 63.80%
evp-aes-128-xts@64 4.176860631 2.59817764 60.76%
evp-aes-128-xts@256 3.069126585 1.284561028 138.92%
evp-aes-128-xts@1024 2.805962699 0.932754655 200.83%
evp-aes-128-xts@8192 2.725820131 0.829820397 228.48%
evp-aes-128-xts@16384 2.71521905 0.823251591 229.82%
evp-aes-256-xts@16 11.24790935 7.383914448 52.33%
evp-aes-256-xts@64 5.294128847 3.048641998 73.66%
evp-aes-256-xts@256 3.861649617 1.570359905 145.91%
evp-aes-256-xts@1024 3.537646797 1.200493533 194.68%
evp-aes-256-xts@8192 3.435353012 1.085345319 216.52%
evp-aes-256-xts@16384 3.437952563 1.097963822 213.12%
A57:
Before optimization After optimization Improve
evp-aes-128-xts@16 10.57455446 7.165438012 47.58%
evp-aes-128-xts@64 5.418185447 3.721241202 45.60%
evp-aes-128-xts@256 3.855184592 1.747145379 120.66%
evp-aes-128-xts@1024 3.477199757 1.253049735 177.50%
evp-aes-128-xts@8192 3.36768104 1.091943159 208.41%
evp-aes-128-xts@16384 3.360373443 1.088942789 208.59%
evp-aes-256-xts@16 12.54559459 8.745489036 43.45%
evp-aes-256-xts@64 6.542808937 4.326387568 51.23%
evp-aes-256-xts@256 4.62668822 2.119908754 118.25%
evp-aes-256-xts@1024 4.161716505 1.557335554 167.23%
evp-aes-256-xts@8192 4.032462227 1.377749511 192.68%
evp-aes-256-xts@16384 4.023293877 1.371558933 193.34%
A53:
Before optimization After optimization Improve
evp-aes-128-xts@16 18.07842135 13.96980808 29.40%
evp-aes-128-xts@64 7.933818397 6.07159276 30.70%
evp-aes-128-xts@256 5.264604704 2.611155744 101.60%
evp-aes-128-xts@1024 4.606660117 1.722713454 167.40%
evp-aes-128-xts@8192 4.405160115 1.454379201 202.90%
evp-aes-128-xts@16384 4.401592028 1.442279392 205.20%
evp-aes-256-xts@16 20.07084054 16.00803726 25.40%
evp-aes-256-xts@64 9.192647294 6.883876732 33.50%
evp-aes-256-xts@256 6.336143161 3.108140452 103.90%
evp-aes-256-xts@1024 5.62502952 2.097960651 168.10%
evp-aes-256-xts@8192 5.412085608 1.807294191 199.50%
evp-aes-256-xts@16384 5.403062591 1.790135764 201.80%
N1:
Before optimization After optimization Improve
evp-aes-128-xts@16 6.48147613 4.209415473 53.98%
evp-aes-128-xts@64 2.847744115 1.950757468 45.98%
evp-aes-128-xts@256 2.085711968 1.061903238 96.41%
evp-aes-128-xts@1024 1.842014669 0.798486302 130.69%
evp-aes-128-xts@8192 1.760449052 0.713853939 146.61%
evp-aes-128-xts@16384 1.760763546 0.707702009 148.80%
evp-aes-256-xts@16 7.264142817 5.265970454 37.94%
evp-aes-256-xts@64 3.251356212 2.41176323 34.81%
evp-aes-256-xts@256 2.380488469 1.342095742 77.37%
evp-aes-256-xts@1024 2.08853022 1.041718215 100.49%
evp-aes-256-xts@8192 2.027432668 0.944571334 114.64%
evp-aes-256-xts@16384 2.00740782 0.941991415 113.10%
Add more XTS test cases to cover the cipher stealing mode and cases of different
number of blocks.
CustomizedGitHooks: yes
Change-Id: I93ee31b2575e1413764e27b599af62994deb4c96
Reviewed-by: Paul Dale <paul.dale@oracle.com>
Reviewed-by: Tomas Mraz <tmraz@fedoraproject.org>
(Merged from https://github.com/openssl/openssl/pull/11399)
For example, FreeBSD prepends "FreeBSD" to version string, e.g.,
FreeBSD clang version 11.0.0 (git@github.com:llvm/llvm-project.git llvmorg-11.0.0-rc2-0-g414f32a9e86)
Target: x86_64-unknown-freebsd13.0
Thread model: posix
InstalledDir: /usr/bin
This prevented us from properly detecting AVX support, etc.
CLA: trivial
Reviewed-by: Richard Levitte <levitte@openssl.org>
Reviewed-by: Paul Dale <paul.dale@oracle.com>
Reviewed-by: Ben Kaduk <kaduk@mit.edu>
(Merged from https://github.com/openssl/openssl/pull/12725)
Since pointer in x32 is 4 bytes, add x86_64-support.pl to define
pointer_size and pointer_register based on flavour to support
stuctures like:
struct { void *ptr; int blocks; }
This fixes 90-test_sslapi.t on x32. Verified with
$ ./Configure shared linux-x86_64
$ make
$ make test
and
$ ./Configure shared linux-x32
$ make
$ make test
Reviewed-by: Richard Levitte <levitte@openssl.org>
Reviewed-by: Tomas Mraz <tmraz@fedoraproject.org>
(Merged from https://github.com/openssl/openssl/pull/10988)
In https://github.com/openssl/openssl/pull/10883, I'd meant to exclude
the perlasm drivers since they aren't opening pipes and do not
particularly need it, but I only noticed x86_64-xlate.pl, so
arm-xlate.pl and ppc-xlate.pl got the change.
That seems to have been fine, so be consistent and also apply the change
to x86_64-xlate.pl. Checking for errors is generally a good idea.
Reviewed-by: Richard Levitte <levitte@openssl.org>
Reviewed-by: David Benjamin <davidben@google.com>
(Merged from https://github.com/openssl/openssl/pull/10930)
FIXES#10692#10638
a bug for aarch64 bigendian with instructions 'st1' and 'ld1' on AES-GCM mode.
CLA: trivial
Reviewed-by: Richard Levitte <levitte@openssl.org>
Reviewed-by: Tim Hudson <tjh@openssl.org>
Reviewed-by: Paul Dale <paul.dale@oracle.com>
(Merged from https://github.com/openssl/openssl/pull/10751)
To support Intel CET, all indirect branch targets must start with
endbranch. Here is a patch to add endbranch to function entries
in x86_64 assembly codes which are indirect branch targets as
discovered by running openssl testsuite on Intel CET machine and
visual inspection.
Verified with
$ CC="gcc -Wl,-z,cet-report=error" ./Configure shared linux-x86_64 -fcf-protection
$ make
$ make test
and
$ CC="gcc -mx32 -Wl,-z,cet-report=error" ./Configure shared linux-x32 -fcf-protection
$ make
$ make test # <<< passed with https://github.com/openssl/openssl/pull/10988
Reviewed-by: Tomas Mraz <tmraz@fedoraproject.org>
Reviewed-by: Richard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/10982)
If one of the perlasm xlate drivers crashes, OpenSSL's build will
currently swallow the error and silently truncate the output to however
far the driver got. This will hopefully fail to build, but better to
check such things.
Handle this by checking for errors when closing STDOUT (which is a pipe
to the xlate driver).
Reviewed-by: Richard Levitte <levitte@openssl.org>
Reviewed-by: Tim Hudson <tjh@openssl.org>
Reviewed-by: Tomas Mraz <tmraz@fedoraproject.org>
(Merged from https://github.com/openssl/openssl/pull/10883)
We store a secondary frame pointer info for the debugger
in the red zone. This fixes a crash in the unwinder when
this function is interrupted.
Additionally the missing cfi function annotation is added
to aesni_cbc_sha256_enc_shaext.
[extended tests]
Reviewed-by: Richard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/10674)
Aes-ecb mode can be optimized by inverleaving cipher operation on
several blocks and loop unrolling. Interleaving needs one ideal
unrolling factor, here we adopt the same factor with aes-cbc,
which is described as below:
If blocks number > 5, select 5 blocks as one iteration,every
loop, decrease the blocks number by 5.
If 3 < left blocks < 5 select 3 blocks as one iteration, every
loop, decrease the block number by 3.
If left blocks < 3, treat them as tail blocks.
Detailed implementation will have a little adjustment for squeezing
code space.
With this way, for small size such as 16 bytes, the performance is
similar as before, but for big size such as 16k bytes, the performance
improves a lot, even reaches to 100%, for some arches such as A57,
the improvement even exceeds 100%. The following table will list the
encryption performance data on aarch64, take a72 and a57 as examples.
Performance value takes the unit of cycles per byte, takes the format
as comparision of values. List them as below:
A72:
Before optimization After optimization Improve
evp-aes-128-ecb@16 17.26538237 16.82663866 2.61%
evp-aes-128-ecb@64 5.50528499 5.222637557 5.41%
evp-aes-128-ecb@256 2.632700213 1.908442892 37.95%
evp-aes-128-ecb@1024 1.876102047 1.078018868 74.03%
evp-aes-128-ecb@8192 1.6550392 0.853982929 93.80%
evp-aes-128-ecb@16384 1.636871283 0.847623957 93.11%
evp-aes-192-ecb@16 17.73104961 17.09692468 3.71%
evp-aes-192-ecb@64 5.78984398 5.418545192 6.85%
evp-aes-192-ecb@256 2.872005308 2.081815274 37.96%
evp-aes-192-ecb@1024 2.083226672 1.25095642 66.53%
evp-aes-192-ecb@8192 1.831992057 0.995916251 83.95%
evp-aes-192-ecb@16384 1.821590009 0.993820525 83.29%
evp-aes-256-ecb@16 18.0606306 17.96963317 0.51%
evp-aes-256-ecb@64 6.19651997 5.762465812 7.53%
evp-aes-256-ecb@256 3.176991394 2.24642538 41.42%
evp-aes-256-ecb@1024 2.385991919 1.396018192 70.91%
evp-aes-256-ecb@8192 2.147862636 1.142222597 88.04%
evp-aes-256-ecb@16384 2.131361787 1.135944617 87.63%
A57:
Before optimization After optimization Improve
evp-aes-128-ecb@16 18.61045121 18.36456218 1.34%
evp-aes-128-ecb@64 6.438628994 5.467959461 17.75%
evp-aes-128-ecb@256 2.957452881 1.97238604 49.94%
evp-aes-128-ecb@1024 2.117096219 1.099665054 92.52%
evp-aes-128-ecb@8192 1.868385973 0.837440804 123.11%
evp-aes-128-ecb@16384 1.853078526 0.822420027 125.32%
evp-aes-192-ecb@16 19.07021756 18.50018552 3.08%
evp-aes-192-ecb@64 6.672351486 5.696088921 17.14%
evp-aes-192-ecb@256 3.260427769 2.131449916 52.97%
evp-aes-192-ecb@1024 2.410522832 1.250529718 92.76%
evp-aes-192-ecb@8192 2.17921605 0.973225504 123.92%
evp-aes-192-ecb@16384 2.162250997 0.95919871 125.42%
evp-aes-256-ecb@16 19.3008384 19.12743654 0.91%
evp-aes-256-ecb@64 6.992950658 5.92149541 18.09%
evp-aes-256-ecb@256 3.576361743 2.287619504 56.34%
evp-aes-256-ecb@1024 2.726671027 1.381267599 97.40%
evp-aes-256-ecb@8192 2.493583657 1.110959913 124.45%
evp-aes-256-ecb@16384 2.473916816 1.099967073 124.91%
Change-Id: Iccd23d972e0d52d22dc093f4c208f69c9d5a0ca7
Reviewed-by: Shane Lontis <shane.lontis@oracle.com>
Reviewed-by: Richard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/10518)
They now generally conform to the following argument sequence:
script.pl "$(PERLASM_SCHEME)" [ C preprocessor arguments ... ] \
$(PROCESSOR) <output file>
However, in the spirit of being able to use these scripts manually,
they also allow for no argument, or for only the flavour, or for only
the output file. This is done by only using the last argument as
output file if it's a file (it has an extension), and only using the
first argument as flavour if it isn't a file (it doesn't have an
extension).
While we're at it, we make all $xlate calls the same, i.e. the $output
argument is always quoted, and we always die on error when trying to
start $xlate.
There's a perl lesson in this, regarding operator priority...
This will always succeed, even when it fails:
open FOO, "something" || die "ERR: $!";
The reason is that '||' has higher priority than list operators (a
function is essentially a list operator and gobbles up everything
following it that isn't lower priority), and since a non-empty string
is always true, so that ends up being exactly the same as:
open FOO, "something";
This, however, will fail if "something" can't be opened:
open FOO, "something" or die "ERR: $!";
The reason is that 'or' has lower priority that list operators,
i.e. it's performed after the 'open' call.
Reviewed-by: Matt Caswell <matt@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/9884)
CLA: trivial
Reviewed-by: Richard Levitte <levitte@openssl.org>
Reviewed-by: Matthias St. Pierre <Matthias.St.Pierre@ncp-e.com>
(Merged from https://github.com/openssl/openssl/pull/9288)
The kernel self-tests picked up an issue with CTR mode. The issue was
detected with a test vector with an IV of
FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFD: after 3 increments it should wrap
around to 0.
There are two paths that increment IVs: the bulk (8 at a time) path,
and the individual path which is used when there are fewer than 8 AES
blocks to process.
In the bulk path, the IV is incremented with vadduqm: "Vector Add
Unsigned Quadword Modulo", which does 128-bit addition.
In the individual path, however, the IV is incremented with vadduwm:
"Vector Add Unsigned Word Modulo", which instead does 4 32-bit
additions. Thus the IV would instead become
FFFFFFFFFFFFFFFFFFFFFFFF00000000, throwing off the result.
Use vadduqm.
This was probably a typo originally, what with q and w being
adjacent.
CLA: trivial
Reviewed-by: Richard Levitte <levitte@openssl.org>
Reviewed-by: Paul Dale <paul.dale@oracle.com>
(Merged from https://github.com/openssl/openssl/pull/8942)
There are two copy-paste errors in handling CTR mode. When dealing
with a 2 or 3 block tail, the code branches to the CBC decryption exit
path, rather than to the CTR exit path.
This can lead to data corruption: in the Linux kernel we have a copy
of this file, and the bug leads to corruption of the IV, which leads
to data corruption when we call the encryption function again later to
encrypt subsequent blocks.
Originally reported to the Linux kernel by Ondrej Mosnáček <omosnacek@gmail.com>
CLA: trivial
Reviewed-by: Tim Hudson <tjh@openssl.org>
Reviewed-by: Paul Dale <paul.dale@oracle.com>
(Merged from https://github.com/openssl/openssl/pull/8510)
The add/double shortcut in ecp_nistz256-x86_64.pl left one instruction
point that did not unwind, and the "slow" path in AES_cbc_encrypt was
not annotated correctly. For the latter, add
.cfi_{remember,restore}_state support to perlasm.
Next, fill in a bunch of functions that are missing no-op .cfi_startproc
and .cfi_endproc blocks. libunwind cannot unwind those stack frames
otherwise.
Finally, work around a bug in libunwind by not encoding rflags. (rflags
isn't a callee-saved register, so there's not much need to annotate it
anyway.)
These were found as part of ABI testing work in BoringSSL.
Reviewed-by: Richard Levitte <levitte@openssl.org>
GH: #8109
"Windows friendliness" means a) unified PIC-ification, unified across
all platforms; b) unified commantary delimiter; c) explicit ldur/stur,
as Visual Studio assembler can't automatically encode ldr/str as
ldur/stur when needed.
Reviewed-by: Paul Dale <paul.dale@oracle.com>
Reviewed-by: Richard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/8256)
