core/lib/systems/architectures.nix
2024-06-30 09:12:46 +01:00

408 lines
7.5 KiB
Nix

{ lib }:
rec {
# gcc.arch to its features (as in /proc/cpuinfo)
features = {
# x86_64 Generic
# Spec: https://gitlab.com/x86-psABIs/x86-64-ABI/
default = [ ];
x86-64 = [ ];
x86-64-v2 = [
"sse3"
"ssse3"
"sse4_1"
"sse4_2"
];
x86-64-v3 = [
"sse3"
"ssse3"
"sse4_1"
"sse4_2"
"avx"
"avx2"
"fma"
];
x86-64-v4 = [
"sse3"
"ssse3"
"sse4_1"
"sse4_2"
"avx"
"avx2"
"avx512"
"fma"
];
# x86_64 Intel
nehalem = [
"sse3"
"ssse3"
"sse4_1"
"sse4_2"
"aes"
];
westmere = [
"sse3"
"ssse3"
"sse4_1"
"sse4_2"
"aes"
];
sandybridge = [
"sse3"
"ssse3"
"sse4_1"
"sse4_2"
"aes"
"avx"
];
ivybridge = [
"sse3"
"ssse3"
"sse4_1"
"sse4_2"
"aes"
"avx"
];
haswell = [
"sse3"
"ssse3"
"sse4_1"
"sse4_2"
"aes"
"avx"
"avx2"
"fma"
];
broadwell = [
"sse3"
"ssse3"
"sse4_1"
"sse4_2"
"aes"
"avx"
"avx2"
"fma"
];
skylake = [
"sse3"
"ssse3"
"sse4_1"
"sse4_2"
"aes"
"avx"
"avx2"
"fma"
];
skylake-avx512 = [
"sse3"
"ssse3"
"sse4_1"
"sse4_2"
"aes"
"avx"
"avx2"
"avx512"
"fma"
];
cannonlake = [
"sse3"
"ssse3"
"sse4_1"
"sse4_2"
"aes"
"avx"
"avx2"
"avx512"
"fma"
];
icelake-client = [
"sse3"
"ssse3"
"sse4_1"
"sse4_2"
"aes"
"avx"
"avx2"
"avx512"
"fma"
];
icelake-server = [
"sse3"
"ssse3"
"sse4_1"
"sse4_2"
"aes"
"avx"
"avx2"
"avx512"
"fma"
];
cascadelake = [
"sse3"
"ssse3"
"sse4_1"
"sse4_2"
"aes"
"avx"
"avx2"
"avx512"
"fma"
];
cooperlake = [
"sse3"
"ssse3"
"sse4_1"
"sse4_2"
"aes"
"avx"
"avx2"
"avx512"
"fma"
];
tigerlake = [
"sse3"
"ssse3"
"sse4_1"
"sse4_2"
"aes"
"avx"
"avx2"
"avx512"
"fma"
];
alderlake = [
"sse3"
"ssse3"
"sse4_1"
"sse4_2"
"aes"
"avx"
"avx2"
"fma"
];
# x86_64 AMD
btver1 = [
"sse3"
"ssse3"
"sse4_1"
"sse4_2"
];
btver2 = [
"sse3"
"ssse3"
"sse4_1"
"sse4_2"
"aes"
"avx"
];
bdver1 = [
"sse3"
"ssse3"
"sse4_1"
"sse4_2"
"sse4a"
"aes"
"avx"
"fma"
"fma4"
];
bdver2 = [
"sse3"
"ssse3"
"sse4_1"
"sse4_2"
"sse4a"
"aes"
"avx"
"fma"
"fma4"
];
bdver3 = [
"sse3"
"ssse3"
"sse4_1"
"sse4_2"
"sse4a"
"aes"
"avx"
"fma"
"fma4"
];
bdver4 = [
"sse3"
"ssse3"
"sse4_1"
"sse4_2"
"sse4a"
"aes"
"avx"
"avx2"
"fma"
"fma4"
];
znver1 = [
"sse3"
"ssse3"
"sse4_1"
"sse4_2"
"sse4a"
"aes"
"avx"
"avx2"
"fma"
];
znver2 = [
"sse3"
"ssse3"
"sse4_1"
"sse4_2"
"sse4a"
"aes"
"avx"
"avx2"
"fma"
];
znver3 = [
"sse3"
"ssse3"
"sse4_1"
"sse4_2"
"sse4a"
"aes"
"avx"
"avx2"
"fma"
];
znver4 = [
"sse3"
"ssse3"
"sse4_1"
"sse4_2"
"sse4a"
"aes"
"avx"
"avx2"
"avx512"
"fma"
];
# other
armv5te = [ ];
armv6 = [ ];
armv7-a = [ ];
armv8-a = [ ];
mips32 = [ ];
loongson2f = [ ];
};
# a superior CPU has all the features of an inferior and is able to build and test code for it
inferiors = {
# x86_64 Generic
default = [ ];
x86-64 = [ ];
x86-64-v2 = [ "x86-64" ];
x86-64-v3 = [ "x86-64-v2" ] ++ inferiors.x86-64-v2;
x86-64-v4 = [ "x86-64-v3" ] ++ inferiors.x86-64-v3;
# x86_64 Intel
# https://gcc.gnu.org/onlinedocs/gcc/x86-Options.html
nehalem = [ "x86-64-v2" ] ++ inferiors.x86-64-v2;
westmere = [ "nehalem" ] ++ inferiors.nehalem;
sandybridge = [ "westmere" ] ++ inferiors.westmere;
ivybridge = [ "sandybridge" ] ++ inferiors.sandybridge;
haswell = lib.unique (
[
"ivybridge"
"x86-64-v3"
]
++ inferiors.ivybridge
++ inferiors.x86-64-v3
);
broadwell = [ "haswell" ] ++ inferiors.haswell;
skylake = [ "broadwell" ] ++ inferiors.broadwell;
skylake-avx512 = lib.unique (
[
"skylake"
"x86-64-v4"
]
++ inferiors.skylake
++ inferiors.x86-64-v4
);
cannonlake = [ "skylake-avx512" ] ++ inferiors.skylake-avx512;
icelake-client = [ "cannonlake" ] ++ inferiors.cannonlake;
icelake-server = [ "icelake-client" ] ++ inferiors.icelake-client;
cascadelake = [ "cannonlake" ] ++ inferiors.cannonlake;
cooperlake = [ "cascadelake" ] ++ inferiors.cascadelake;
tigerlake = [ "icelake-server" ] ++ inferiors.icelake-server;
# CX16 does not exist on alderlake, while it does on nearly all other intel CPUs
alderlake = [ ];
# x86_64 AMD
# TODO: fill this (need testing)
btver1 = [ ];
btver2 = [ ];
bdver1 = [ ];
bdver2 = [ ];
bdver3 = [ ];
bdver4 = [ ];
# Regarding `skylake` as inferior of `znver1`, there are reports of
# successful usage by Gentoo users and Phoronix benchmarking of different
# `-march` targets.
#
# The GCC documentation on extensions used and wikichip documentation
# regarding supperted extensions on znver1 and skylake was used to create
# this partial order.
#
# Note:
#
# - The successors of `skylake` (`cannonlake`, `icelake`, etc) use `avx512`
# which no current AMD Zen michroarch support.
# - `znver1` uses `ABM`, `CLZERO`, `CX16`, `MWAITX`, and `SSE4A` which no
# current Intel microarch support.
#
# https://www.phoronix.com/scan.php?page=article&item=amd-znver3-gcc11&num=1
# https://gcc.gnu.org/onlinedocs/gcc/x86-Options.html
# https://en.wikichip.org/wiki/amd/microarchitectures/zen
# https://en.wikichip.org/wiki/intel/microarchitectures/skylake
znver1 = [ "skylake" ] ++ inferiors.skylake; # Includes haswell and x86-64-v3
znver2 = [ "znver1" ] ++ inferiors.znver1;
znver3 = [ "znver2" ] ++ inferiors.znver2;
znver4 = lib.unique (
[
"znver3"
"x86-64-v4"
]
++ inferiors.znver3
++ inferiors.x86-64-v4
);
# other
armv5te = [ ];
armv6 = [ ];
armv7-a = [ ];
armv8-a = [ ];
mips32 = [ ];
loongson2f = [ ];
};
predicates =
let
featureSupport = feature: x: builtins.elem feature features.${x} or [ ];
in
{
sse3Support = featureSupport "sse3";
ssse3Support = featureSupport "ssse3";
sse4_1Support = featureSupport "sse4_1";
sse4_2Support = featureSupport "sse4_2";
sse4_aSupport = featureSupport "sse4a";
avxSupport = featureSupport "avx";
avx2Support = featureSupport "avx2";
avx512Support = featureSupport "avx512";
aesSupport = featureSupport "aes";
fmaSupport = featureSupport "fma";
fma4Support = featureSupport "fma4";
};
}