diff options
Diffstat (limited to 'INSTALL.md')
-rw-r--r-- | INSTALL.md | 40 |
1 files changed, 38 insertions, 2 deletions
diff --git a/INSTALL.md b/INSTALL.md index 14b6118c36..039e1bdf08 100644 --- a/INSTALL.md +++ b/INSTALL.md @@ -41,6 +41,7 @@ Table of Contents - [Notes on multi-threading](#notes-on-multi-threading) - [Notes on shared libraries](#notes-on-shared-libraries) - [Notes on random number generation](#notes-on-random-number-generation) + - [Notes on assembler modules compilation](#notes-on-assembler-modules-compilation) Prerequisites ============= @@ -1610,8 +1611,8 @@ build. Use this command: $ mms clean ! (or mmk) OpenVMS $ nmake clean # Windows -Assembler error messages can sometimes be sidestepped by using the -`no-asm` configuration option. +Assembler error messages can sometimes be sidestepped by using the `no-asm` +configuration option. See also [notes](#notes-on-assembler-modules-compilation). Compiling parts of OpenSSL with gcc and others with the system compiler will result in unresolved symbols on some systems. @@ -1734,6 +1735,41 @@ to install additional support software to obtain a random seed and reseed the CSPRNG manually. Please check out the manual pages for `RAND_add()`, `RAND_bytes()`, `RAND_egd()`, and the FAQ for more information. +Notes on assembler modules compilation +-------------------------------------- + +Compilation of some code paths in assembler modules might depend on whether the +current assembler version supports certain ISA extensions or not. Code paths +that use the AES-NI, PCLMULQDQ, SSSE3, and SHA extensions are always assembled. +Apart from that, the minimum requirements for the assembler versions are shown +in the table below: + +| ISA extension | GNU as | nasm | llvm | +|---------------|--------|--------|---------| +| AVX | 2.19 | 2.09 | 3.0 | +| AVX2 | 2.22 | 2.10 | 3.1 | +| ADCX/ADOX | 2.23 | 2.10 | 3.3 | +| AVX512 | 2.25 | 2.11.8 | 3.6 (*) | +| AVX512IFMA | 2.26 | 2.11.8 | 6.0 (*) | +| VAES | 2.30 | 2.13.3 | 6.0 (*) | + +--- + +(*) Even though AVX512 support was implemented in llvm 3.6, prior to version 7.0 +an explicit -march flag was apparently required to compile assembly modules. But +then the compiler generates processor-specific code, which in turn contradicts +the idea of performing dispatch at run-time, which is facilitated by the special +variable `OPENSSL_ia32cap`. For versions older than 7.0, it is possible to work +around the problem by forcing the build procedure to use the following script: + + #!/bin/sh + exec clang -no-integrated-as "$@" + +instead of the real clang. In which case it doesn't matter what clang version +is used, as it is the version of the GNU assembler that will be checked. + +--- + <!-- Links --> [openssl-users]: |