Depending on the size of the input, we may take different paths through
the accelerated arm64 ChaCha20 routines, each of which use a different
subset of the FP registers, some of which need to be preserved and
restored, as required by the AArch64 calling convention (AAPCS64)
In some cases, (e.g., when the input size is 640 bytes), we call the 512
byte NEON path followed directly by the scalar path, and in this case,
we preserve and restore d8 and d9, only to clobber them again
immediately before handing over to the scalar path which does not touch
the FP registers at all, and hence does not restore them either.
Fix this by moving the restoration of d8 and d9 to a later stage in the
512 byte routine, either before calling the scalar path, or when exiting
the function.
Fixes#10470
CLA: trivial
Reviewed-by: Paul Dale <paul.dale@oracle.com>
Reviewed-by: Matt Caswell <matt@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/10497)
- add instructions: clfi, stck, stckf, kdsa
- clfi and clgfi belong to extended-immediate (not long-displacement)
- some cleanup
Signed-off-by: Patrick Steuer <patrick.steuer@de.ibm.com>
Reviewed-by: Richard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/10346)
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)
Signed-off-by: Patrick Steuer <patrick.steuer@de.ibm.com>
Reviewed-by: Paul Dale <paul.dale@oracle.com>
Reviewed-by: Richard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/8181)
The change is triggered by ThunderX2 where 3+1 was slower than scalar
code path, but it helps all processors [to handle <512 inputs].
Reviewed-by: Tim Hudson <tjh@openssl.org>
Reviewed-by: Richard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/8776)
featuring 6x"horizontal" code path which is up to 25%
faster than present 4x"vertical" for larger blocks.
Signed-off-by: Patrick Steuer <patrick.steuer@de.ibm.com>
Reviewed-by: Matt Caswell <matt@openssl.org>
Reviewed-by: Richard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/8287)
"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)
"Windows friendliness" means a) flipping .thumb and .text directives,
b) always generate Thumb-2 code when asked(*); c) Windows-specific
references to external OPENSSL_armcap_P.
(*) so far *some* modules were compiled as .code 32 even if Thumb-2
was targeted. It works at hardware level because processor can alternate
between the modes with no overhead. But clang --target=arm-windows's
builtin assembler just refuses to compile .code 32...
Reviewed-by: Paul Dale <paul.dale@oracle.com>
Reviewed-by: Richard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/8252)
Signed-off-by: Patrick Steuer <patrick.steuer@de.ibm.com>
Reviewed-by: Tim Hudson <tjh@openssl.org>
Reviewed-by: Richard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/8257)
ARMv8.3 adds pointer authentication extension, which in this case allows
to ensure that, when offloaded to stack, return address is same at return
as at entry to the subroutine. The new instructions are nops on processors
that don't implement the extension, so that the vetification is backward
compatible.
Reviewed-by: Kurt Roeckx <kurt@roeckx.be>
Reviewed-by: Richard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/8205)
Signed-off-by: Patrick Steuer <patrick.steuer@de.ibm.com>
Reviewed-by: Tim Hudson <tjh@openssl.org>
Reviewed-by: Richard Levitte <levitte@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6919)
The 128-byte vectors are extensively used in chacha20_poly1305_tls_cipher
and dedicated code path is ~30-50% faster on most platforms.
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6626)
It's kind of a "brown-bag" bug, as I did recognize the problem and
verified an ad-hoc solution, but failed to follow up with cross-checks
prior filing previous merge request.
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6435)
This comes at cost of minor 2.5% regression on G4, which is reasonable
trade-off. [Further improve compliance with ABI requirements.]
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6406)
As it turns out originally published results were skewed by "turbo"
mode. VM apparently remains oblivious to dynamic frequency scaling,
and reports that processor operates at "base" frequency at all times.
While actual frequency gets increased under load.
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6406)
32-bit vector rotate instruction was defined from beginning, it
not being used from the start must be a brain-slip...
Reviewed-by: Bernd Edlinger <bernd.edlinger@hotmail.de>
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/6363)
256-bit AVX512VL was estimated to deliver ~50% improvement over AVX2
and it did live up to the expectations.
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/4838)
The __clang__-guarded #defines cause gas to complain if clang is passed
-fno-integrated-as. Emitting .syntax unified when those are used fixes
this. This matches the change made to ghash-armv4.pl in
6cf412c473.
Reviewed-by: Andy Polyakov <appro@openssl.org>
Reviewed-by: Kurt Roeckx <kurt@roeckx.be>
(Merged from https://github.com/openssl/openssl/pull/3694)
Only chacha_internal_test is affected, since this path is not used
from EVP.
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/4758)
Around 138 distinct errors found and fixed; thanks!
Reviewed-by: Kurt Roeckx <kurt@roeckx.be>
Reviewed-by: Tim Hudson <tjh@openssl.org>
Reviewed-by: Rich Salz <rsalz@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/3459)
"Optimize" is in quotes because it's rather a "salvage operation"
for now. Idea is to identify processor capability flags that
drive Knights Landing to suboptimial code paths and mask them.
Two flags were identified, XSAVE and ADCX/ADOX. Former affects
choice of AES-NI code path specific for Silvermont (Knights Landing
is of Silvermont "ancestry"). And 64-bit ADCX/ADOX instructions are
effectively mishandled at decode time. In both cases we are looking
at ~2x improvement.
AVX-512 results cover even Skylake-X :-)
Hardware used for benchmarking courtesy of Atos, experiments run by
Romain Dolbeau <romain.dolbeau@atos.net>. Kudos!
Reviewed-by: Rich Salz <rsalz@openssl.org>
The assembler already knows the actual path to the generated file and,
in other perlasm architectures, is left to manage debug symbols itself.
Notably, in OpenSSL 1.1.x's new build system, which allows a separate
build directory, converting .pl to .s as the scripts currently do result
in the wrong paths.
This also avoids inconsistencies from some of the files using $0 and
some passing in the filename.
Reviewed-by: Richard Levitte <levitte@openssl.org>
Reviewed-by: Andy Polyakov <appro@openssl.org>
(Merged from https://github.com/openssl/openssl/pull/3431)
- harmonize handlers with guidelines and themselves;
- fix some bugs in handlers;
- add missing handlers in chacha and ecp_nistz256 modules;
Reviewed-by: Rich Salz <rsalz@openssl.org>
In order to minimize dependency on assembler version a number of
post-SSE2 instructions are encoded manually. But in order to simplify
the procedure only register operands are considered. Non-register
operands are passed down to assembler. Module in question uses pshufb
with memory operands, and old [GNU] assembler can't handle it.
Fortunately in this case it's possible skip just the problematic
segment without skipping SSSE3 support altogether.
Reviewed-by: Richard Levitte <levitte@openssl.org>
Reviewed-by: Rich Salz <rsalz@openssl.org>
The prevailing style seems to not have trailing whitespace, but a few
lines do. This is mostly in the perlasm files, but a few C files got
them after the reformat. This is the result of:
find . -name '*.pl' | xargs sed -E -i '' -e 's/( |'$'\t'')*$//'
find . -name '*.c' | xargs sed -E -i '' -e 's/( |'$'\t'')*$//'
find . -name '*.h' | xargs sed -E -i '' -e 's/( |'$'\t'')*$//'
Then bn_prime.h was excluded since this is a generated file.
Note mkerr.pl has some changes in a heredoc for some help output, but
other lines there lack trailing whitespace too.
Reviewed-by: Kurt Roeckx <kurt@openssl.org>
Reviewed-by: Matt Caswell <matt@openssl.org>
Some of the instructions used in latest additions are extension
ones. There is no real reason to limit ourselves to specific
processors, so [re-]adhere to base instruction set.
RT#4548
Reviewed-by: Rich Salz <rsalz@openssl.org>
