This results in ~20% improvement for reference implementation on my
Xeon E5630.

`openssl speed -evp streebog256' results (before and after):

   16 bytes    64 bytes   256 bytes  1024 bytes  8192 bytes  16384 bytes
 3815.25k   12874.41k   31497.81k   49823.06k   59932.67k    57021.78k
 4361.29k   14933.06k   37091.67k   59001.86k   71360.51k    72488.28k

Also, simple check if `_addcarry_u64' works correctly.

- Remove redundant 'os: linux', 'compiler: gcc' tags.
- Order jobs: x86_64 builds, analysers, other arches, other OSes.
- Mark openssl-master build as 'allow_failures' for PRs.
- Build all jobs against 'OpenSSL_1_1_1-stable' by default.
- Do not include 'linux/arch' in the job name, because Travis show
  them already.
- Split name by ' | ' instead of '/' so it's easier to understand.

It's turned out that 'OpenSSL_1_1_1-stable' does not have 'build_sw'
make target. But both 'OpenSSL_1_1_1-stable' and 'master' have
'build_libs' and 'build_programs'.

Fixes: cb1b5ffc ("travis-ci: Speed-up openssl build")

Allow aligned/unaligned access. Use better intrinsics and other
small optimizations.

Committed-by: Vitaly Chikunov <vt@altlinux.org>

Also, add compile time notice message to show which implementation is
selected.

`_addcarry_u64' is using `ADC' and does not require ADX extensions
(`-madx').

Committed-by: Vitaly Chikunov <vt@altlinux.org>

On-stack `buf' is replaced by manipulation directly on `CTX->buffer'.

Committed-by: Vitaly Chikunov <vt@altlinux.org>

Some simple optimizations.

Committed-by: Vitaly Chikunov <vt@altlinux.org>

Introduce byte `.B' union type to `union uint512_u'.
Change `CTX.buffer' type from `unsigned char' to `union uint512_u'.
Change `data' argument of `stage2()' to `union uint512_u *'.
Change `g()' arguments to `union uint512_u *' with `RESTRICT'
allowing compiler to optimize more.

No code changes, only type changes here.

Committed-by: Vitaly Chikunov <vt@altlinux.org>

Now `stage2' will always get aligned data which is prepared in
`gost2012_hash_block' by copying into `CTX->buffer'.

This will allow to change `data' argument of `stage2' from `unsigned
char *' pointer to `union uint512_u *'.

Committed-by: Vitaly Chikunov <vt@altlinux.org>

This allows to use `RESTRICT' which allows compiler optimize more.

Committed-by: Vitaly Chikunov <vt@altlinux.org>

Similar to my commit into adegtyarev/streebog@432d5de.

`pad' does not need this complicated  memory movements using stack
buffer.

Committed-by: Vitaly Chikunov <vt@altlinux.org>

`CTX' is already fully zeroed at the beginning of
`init_gost2012_hash_ctx'.

Committed-by: Vitaly Chikunov <vt@altlinux.org>

`CTX->bufsize' cannot be 64 ever.

And save 64 bytes.

Committed-by: Vitaly Chikunov <vt@altlinux.org>

Make sure everything is OK on 32-bit platform.

Do not build openssl documentation.
Also, split build line into several lines.

s390x is big-endian which is different from any other arch we build.
We have endianness dependent code, thus this is useful to test.

Previously Clang is only used for OSX build.

Log full commit id that we build OpenSSL against, because `master' is
moving target. That should be useful for old logs to reproduce these
builds.

Following commits mark CMAC and HMAC primitives deprecated:

  commit a6d572e60120e0ffb42aece17a085f0fed1b8f6f
  Author: Pauli <paul.dale@oracle.com>
  Date:   Tue Jan 14 10:59:11 2020 +1000

      Deprecate the low level CMAC functions

  commit dbde4726889a19af0a718fe9c5542f39c81acbd3
  Author: Pauli <paul.dale@oracle.com>
  Date:   Tue Jan 14 12:11:50 2020 +1000

      Deprecate the low level HMAC functions

Add `-Wno-error=deprecated-declarations' to appropriate targets.

References:
  https://github.com/openssl/openssl/commit/a6d572e601
  https://github.com/openssl/openssl/commit/dbde472688

Format is `os/compiler[+env]/arch'.

No need for `build.log' as `make -s' will output any failure.

Also, redirection was broken anyway since splitting make into two makes
with separate `-j' and install phase.

Add `travis_wait' to make traves not time-out on slow arches. Reference:
  https://docs.travis-ci.com/user/common-build-problems/#build-times-out-because-no-output-was-received

Fixes: c38ec758 ("travis: speed up Travis build")

To make CI output not too long - output only header of every test set
and output only relevant log entries on test failures.

When `CI' env is defined. Which should be defined for both Travis and
Cirrus. Reference:
  https://docs.travis-ci.com/user/environment-variables/
  https://cirrus-ci.org/guide/writing-tasks/#environment-variables

Add more debug output to fix Travis-CI issues. Move definition of
OPENSSL_CONF above first tcl test run. Define OPENSSL_ENGINES from
ENGINE_DIR by default.

Fis was same as the first test.

Otherwise you'll get:

  ocsp: Digest must be before -cert or -serial
  ocsp: Use -help for summary.

Remove GOST2012-MAGMA-MAGMAOMAC and GOST2012-KUZNYECHIK-KUZNYECHIKOMAC
suites from the test.

Disable handshake tests.

Verification failure
139891902039680:error:2E09A09E:CMS routines:CMS_SignerInfo_verify_content:verification failure:crypto/cms/cms_sd.c:848:
139891902039680:error:2E09D06D:CMS routines:CMS_verify:content verify error:crypto/cms/cms_smime.c:393:

This is not robust to have such tests.