Skip to content

Commit

Permalink
Zig build: enable CPU-specific flags only on relevant files
Browse files Browse the repository at this point in the history 
Fixes #1358
  • Loading branch information
@jedisct1
jedisct1 committed Apr 26, 2024
1 parent edb94a9 commit 9b369db
Showing 1 changed file with 172 additions and 45 deletions.
217 changes: 172 additions & 45 deletions build.zig
View file
Original file line number Diff line number Diff line change
Expand Up @@ -29,17 +29,8 @@ pub fn build(b: *std.Build) !void {
}

switch (target.result.cpu.arch) {
// Features we assume are always available because they won't affect
// code generation in files that don't use them.
.x86_64 => {
target.query.cpu_features_add.addFeature(@intFromEnum(Target.x86.Feature.sse4_1));
target.query.cpu_features_add.addFeature(@intFromEnum(Target.x86.Feature.aes));
target.query.cpu_features_add.addFeature(@intFromEnum(Target.x86.Feature.pclmul));
target.query.cpu_features_add.addFeature(@intFromEnum(Target.x86.Feature.rdrnd));
},
.aarch64, .aarch64_be => {
target.query.cpu_features_add.addFeature(@intFromEnum(Target.aarch64.Feature.crypto));
// ARM CPUs supported by Windows also support NEON.
.aarch64, .aarch64_be, .aarch64_32 => {
// ARM CPUs supported by Windows are assumed to have NEON support
if (target.result.isMinGW()) {
target.query.cpu_features_add.addFeature(@intFromEnum(Target.aarch64.Feature.neon));
}
Expand Down Expand Up @@ -187,34 +178,17 @@ pub fn build(b: *std.Build) !void {
lib.defineCMacro("HAVE_CPUID", "1");
lib.defineCMacro("HAVE_MMINTRIN_H", "1");
lib.defineCMacro("HAVE_EMMINTRIN_H", "1");

const cpu_features = target.result.cpu.features;
const has_sse3 = cpu_features.isEnabled(@intFromEnum(Target.x86.Feature.sse3));
const has_ssse3 = cpu_features.isEnabled(@intFromEnum(Target.x86.Feature.ssse3));
const has_sse4_1 = cpu_features.isEnabled(@intFromEnum(Target.x86.Feature.sse4_1));
const has_avx = cpu_features.isEnabled(@intFromEnum(Target.x86.Feature.avx));
const has_avx2 = cpu_features.isEnabled(@intFromEnum(Target.x86.Feature.avx2));
const has_avx512f = cpu_features.isEnabled(@intFromEnum(Target.x86.Feature.avx512f));
const has_aes = cpu_features.isEnabled(@intFromEnum(Target.x86.Feature.aes));
const has_pclmul = cpu_features.isEnabled(@intFromEnum(Target.x86.Feature.pclmul));
const has_rdrnd = cpu_features.isEnabled(@intFromEnum(Target.x86.Feature.rdrnd));

if (has_sse3) lib.defineCMacro("HAVE_PMMINTRIN_H", "1");
if (has_ssse3) lib.defineCMacro("HAVE_TMMINTRIN_H", "1");
if (has_sse4_1) lib.defineCMacro("HAVE_SMMINTRIN_H", "1");
if (has_avx) lib.defineCMacro("HAVE_AVXINTRIN_H", "1");
if (has_avx2) lib.defineCMacro("HAVE_AVX2INTRIN_H", "1");
if (has_avx512f) lib.defineCMacro("HAVE_AVX512FINTRIN_H", "1");
if (has_aes and has_pclmul) lib.defineCMacro("HAVE_WMMINTRIN_H", "1");
if (has_rdrnd) lib.defineCMacro("HAVE_RDRAND", "1");
lib.defineCMacro("HAVE_PMMINTRIN_H", "1");
lib.defineCMacro("HAVE_TMMINTRIN_H", "1");
lib.defineCMacro("HAVE_SMMINTRIN_H", "1");
lib.defineCMacro("HAVE_AVXINTRIN_H", "1");
lib.defineCMacro("HAVE_AVX2INTRIN_H", "1");
lib.defineCMacro("HAVE_AVX512FINTRIN_H", "1");
lib.defineCMacro("HAVE_WMMINTRIN_H", "1");
lib.defineCMacro("HAVE_RDRAND", "1");
},
.aarch64, .aarch64_be => {
const cpu_features = target.result.cpu.features;
const has_neon = cpu_features.isEnabled(@intFromEnum(Target.aarch64.Feature.neon));
const has_crypto = cpu_features.isEnabled(@intFromEnum(Target.aarch64.Feature.crypto));
if (has_neon and has_crypto) {
lib.defineCMacro("HAVE_ARMCRYPTO", "1");
}
lib.defineCMacro("HAVE_ARMCRYPTO", "1");
},
.wasm32, .wasm64 => {
lib.defineCMacro("__wasm__", "1");
Expand All @@ -229,25 +203,178 @@ pub fn build(b: *std.Build) !void {
else => {},
}

const MFlags = enum {
sse2,
ssse3,
sse41,
avx,
avx2,
avx512f,
aes,
pclmul,
rdrnd,
crypto,

fn f(flag: @This()) []const u8 {
return switch (flag) {
.sse2 => "-msse2",
.ssse3 => "-mssse3",
.sse41 => "-msse4.1",
.avx => "-mavx",
.avx2 => "-mavx2",
.avx512f => "-mavx512f",
.aes => "-maes",
.pclmul => "-mpclmul",
.rdrnd => "-mrdrnd",
.crypto => "-mcrypto",
};
}
};

const MLib = struct {
name: []const u8,
count: usize,
sources: []const []const u8,
flags: []const MFlags,
arches: []const std.Target.Cpu.Arch,
};

var mlibs: [8]MLib = .{
.{
.name = "armcrypto",
.count = 3,
.sources = &.{
"crypto_aead/aegis128l/aegis128l_armcrypto.c",
"crypto_aead/aegis256/aegis256_armcrypto.c",
"crypto_aead/aes256gcm/armcrypto/aead_aes256gcm_armcrypto.c",
},
.flags = &.{.aes},
.arches = &.{ .aarch64, .aarch64_be, .aarch64_32 },
},

.{
.name = "sse2",
.count = 1,
.sources = &.{
"crypto_stream/salsa20/xmm6int/salsa20_xmm6int-sse2.c",
},
.flags = &.{.sse2},
.arches = &.{ .x86_64, .x86 },
},
.{
.name = "ssse3",
.count = 3,
.sources = &.{
"crypto_generichash/blake2b/ref/blake2b-compress-ssse3.c",
"crypto_pwhash/argon2/argon2-fill-block-ssse3.c",
"crypto_stream/chacha20/dolbeau/chacha20_dolbeau-ssse3.c",
},
.flags = &.{.ssse3},
.arches = &.{ .x86_64, .x86 },
},
.{
.name = "sse41",
.count = 1,
.sources = &.{
"crypto_generichash/blake2b/ref/blake2b-compress-sse41.c",
},
.flags = &.{.sse41},
.arches = &.{ .x86_64, .x86 },
},
.{
.name = "avx2",
.count = 4,
.sources = &.{
"crypto_generichash/blake2b/ref/blake2b-compress-avx2.c",
"crypto_pwhash/argon2/argon2-fill-block-avx2.c",
"crypto_stream/chacha20/dolbeau/chacha20_dolbeau-avx2.c",
"crypto_stream/salsa20/xmm6int/salsa20_xmm6int-avx2.c",
},
.flags = &.{.avx2},
.arches = &.{.x86_64},
},
.{
.name = "avx512f",
.count = 1,
.sources = &.{
"crypto_pwhash/argon2/argon2-fill-block-avx512f.c",
},
.flags = &.{.avx512f},
.arches = &.{.x86_64},
},
.{
.name = "aesni",
.count = 3,
.sources = &.{
"crypto_aead/aegis128l/aegis128l_aesni.c",
"crypto_aead/aegis256/aegis256_aesni.c",
"crypto_aead/aes256gcm/aesni/aead_aes256gcm_aesni.c",
},
.flags = &.{ .avx, .aes, .pclmul },
.arches = &.{.x86_64},
},
.{
.name = "mrdrnd",
.count = 1,
.sources = &.{
"randombytes/internal/randombytes_internal_random.c",
},
.flags = &.{.rdrnd},
.arches = &.{ .x86_64, .x86 },
},
};
const base_flags = &.{
"-fvisibility=hidden",
"-fno-strict-aliasing",
"-fno-strict-overflow",
"-fwrapv",
"-flax-vector-conversions",
};

const allocator = heap.page_allocator;
var walker = try src_dir.walk(allocator);
while (try walker.next()) |entry| {
var flags = std.ArrayList([]const u8).init(allocator);
defer flags.deinit();
try flags.appendSlice(base_flags);

const name = entry.basename;

if (mem.endsWith(u8, name, ".c")) {
for (&mlibs) |*mlib| {
const found = found: for (mlib.sources) |path| {
if (mem.eql(u8, entry.path, path)) break :found true;
} else false;
if (found) {
for (mlib.arches) |arch| {
if (target.result.cpu.arch == arch) {
for (mlib.flags) |flag| try flags.append(flag.f());
break;
}
}
if (mlib.count == 0) {
std.debug.panic("mlib[{s}].count = -1", .{mlib.name});
}
mlib.count -= 1;
}
}

const full_path = try fmt.allocPrint(allocator, "{s}/{s}", .{ src_path, entry.path });
const flags = &.{
"-fvisibility=hidden",
"-fno-strict-aliasing",
"-fno-strict-overflow",
"-fwrapv",
"-flax-vector-conversions",
};
lib.addCSourceFiles(.{ .files = &.{full_path}, .flags = flags });

lib.addCSourceFiles(.{
.files = &.{full_path},
.flags = flags.items,
});
} else if (mem.endsWith(u8, name, ".S")) {
const full_path = try fmt.allocPrint(allocator, "{s}/{s}", .{ src_path, entry.path });
lib.addAssemblyFile(.{ .path = full_path });
}
}
for (mlibs) |mlib| {
if (mlib.count != 0) {
std.debug.panic("mlib[{s}].count = {}", .{ mlib.name, mlib.count });
}
}
}

const test_path = "test/default";
Expand Down

0 comments on commit 9b369db

Please sign in to comment.