Now that we don't depend on external CRYPTO_library_init calls or the
static initializer to initialize CPU capabilities, we can drop a ton of
code.
This makes CRYPTO_library_init, and all its wrappers, into no-ops and
drops the (non-FIPS) static initializer. I've added an internal
OPENSSL_init_cpuid function for the places where the library actually
needs to initialize the CPU vector.
Note this slightly changes the default, previously
static-initializer-full build: previously, CRYPTO_library_init was a
no-op and we relied on the static initializer. Now we uniformly use
CRYPTO_once. This should be an atomic read in the steady state and
essentially free. We can restore the static initializer by default if
this ends up being a problem, but having only one mode is more
straightforward. This also avoids problems if an application calls into
BoringSSL during its own static initializer. Static initializers are not
coherently ordered.
Update-Note: The BORINGSSL_NO_STATIC_INITIALIZER build option and
CRYPTO_library_init are now unnecessary. Once updating past this
revision, those options can now be cleaned up from downstream projects.
Fixed: 40644931
Change-Id: Idc2e6ea7a73d6352e0360fd886c46d88dba3568c
Reviewed-on: https://boringssl-review.googlesource.com/c/boringssl/+/69508
Reviewed-by: Adam Langley <agl@google.com>
Commit-Queue: David Benjamin <davidben@google.com>
David Benjamin
2fcdd11f6d