Use clang pragmas to enable features rather than depend on compiler s…

…ettings
jedisct1 · Apr 27, 2024 · ee40b36 · ee40b36
1 parent 8dc537d
commit ee40b36
Show file tree

Hide file tree

Showing 16 changed files with 127 additions and 228 deletions.
diff --git a/build.zig b/build.zig
@@ -208,129 +208,7 @@ pub fn build(b: *std.Build) !void {
 
         initLibConfig(target, lib);
 
-        const MFlags = enum {
-            sse2,
-            ssse3,
-            sse41,
-            avx,
-            avx2,
-            avx512f,
-            aes,
-            pclmul,
-            rdrnd,
-            crypto,
-
-            fn f(flag: @This()) Target.Cpu.Feature.Set.Index {
-                return switch (flag) {
-                    .sse2 => @intFromEnum(Target.x86.Feature.sse2),
-                    .ssse3 => @intFromEnum(Target.x86.Feature.ssse3),
-                    .sse41 => @intFromEnum(Target.x86.Feature.sse4_1),
-                    .avx => @intFromEnum(Target.x86.Feature.avx),
-                    .avx2 => @intFromEnum(Target.x86.Feature.avx2),
-                    .avx512f => @intFromEnum(Target.x86.Feature.avx512f),
-                    .aes => @intFromEnum(Target.x86.Feature.aes),
-                    .pclmul => @intFromEnum(Target.x86.Feature.pclmul),
-                    .rdrnd => @intFromEnum(Target.x86.Feature.rdrnd),
-                    .crypto => @intFromEnum(Target.aarch64.Feature.crypto),
-                };
-            }
-        };
-
-        const MLib = struct {
-            name: []const u8,
-            count: usize,
-            sources: []const []const u8,
-            flags: []const MFlags,
-            arches: []const std.Target.Cpu.Arch,
-            lib: *Compile = undefined,
-        };
-
-        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 = &.{ .sse2, .ssse3 },
-                .arches = &.{ .x86_64, .x86 },
-            },
-            .{
-                .name = "sse41",
-                .count = 1,
-                .sources = &.{
-                    "crypto_generichash/blake2b/ref/blake2b-compress-sse41.c",
-                },
-                .flags = &.{ .sse2, .ssse3, .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 = &.{ .sse2, .ssse3, .sse41, .avx, .avx2 },
-                .arches = &.{.x86_64},
-            },
-            .{
-                .name = "avx512f",
-                .count = 1,
-                .sources = &.{
-                    "crypto_pwhash/argon2/argon2-fill-block-avx512f.c",
-                },
-                .flags = &.{ .sse2, .ssse3, .sse41, .avx, .avx2, .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 = &.{
+        const flags = &.{
             "-fvisibility=hidden",
             "-fno-strict-aliasing",
             "-fno-strict-overflow",
@@ -340,61 +218,15 @@ pub fn build(b: *std.Build) !void {
 
         const allocator = heap.page_allocator;
 
-        // compile CPU-specific library code
-        for (&mlibs) |*mlib| {
-            var target2 = target;
-            for (mlib.arches) |arch| {
-                if (target.result.cpu.arch == arch) {
-                    for (mlib.flags) |flag| {
-                        target2.query.cpu_features_add.addFeature(flag.f());
-                    }
-                    break;
-                }
-            }
-
-            mlib.lib = b.addStaticLibrary(.{
-                .name = mlib.name,
-                .target = target2,
-                .optimize = optimize,
-            });
-            const elib = mlib.lib;
-            initLibConfig(target, elib);
-
-            var flags = std.ArrayList([]const u8).init(allocator);
-            defer flags.deinit();
-            try flags.appendSlice(base_flags);
-
-            for (mlib.sources) |path| {
-                const full_path = try fmt.allocPrint(allocator, "{s}/{s}", .{ src_path, path });
-                elib.addCSourceFiles(.{
-                    .files = &.{full_path},
-                    .flags = flags.items,
-                });
-            }
-            lib.linkLibrary(elib);
-        }
-
-        // compile generic library code
         var walker = try src_dir.walk(allocator);
-        files: while (try walker.next()) |entry| {
-            var flags = std.ArrayList([]const u8).init(allocator);
-            defer flags.deinit();
-            try flags.appendSlice(base_flags);
-
+        while (try walker.next()) |entry| {
             const name = entry.basename;
-
             if (mem.endsWith(u8, name, ".c")) {
-                for (mlibs) |mlib| {
-                    for (mlib.sources) |path| {
-                        if (mem.eql(u8, entry.path, path)) continue :files;
-                    }
-                }
-
                 const full_path = try fmt.allocPrint(allocator, "{s}/{s}", .{ src_path, entry.path });
 
                 lib.addCSourceFiles(.{
                     .files = &.{full_path},
-                    .flags = flags.items,
+                    .flags = flags,
                 });
             } else if (mem.endsWith(u8, name, ".S")) {
                 const full_path = try fmt.allocPrint(allocator, "{s}/{s}", .{ src_path, entry.path });

diff --git a/src/libsodium/crypto_aead/aes256gcm/aesni/aead_aes256gcm_aesni.c b/src/libsodium/crypto_aead/aes256gcm/aesni/aead_aes256gcm_aesni.c
@@ -16,9 +16,11 @@
 
 #if defined(HAVE_TMMINTRIN_H) && defined(HAVE_WMMINTRIN_H)
 
-#ifdef __GNUC__
-#pragma GCC target("avx,aes,pclmul")
-#endif
+# ifdef __clang__
+#  pragma clang attribute push(__attribute__((target("aes,avx,pclmul"))), apply_to = function)
+# elif defined(__GNUC__)
+#  pragma GCC target("aes,avx,pclmul")
+# endif
 
 #if !defined(_MSC_VER) || _MSC_VER < 1800
 #define __vectorcall
@@ -1006,4 +1008,8 @@ crypto_aead_aes256gcm_is_available(void)
     return sodium_runtime_has_pclmul() & sodium_runtime_has_aesni() & sodium_runtime_has_avx();
 }
 
+#ifdef __clang__
+# pragma clang attribute pop
+#endif
+
 #endif
diff --git a/src/libsodium/crypto_generichash/blake2b/ref/blake2b-compress-avx2.c b/src/libsodium/crypto_generichash/blake2b/ref/blake2b-compress-avx2.c
@@ -12,11 +12,10 @@
 #if defined(HAVE_AVX2INTRIN_H) && defined(HAVE_EMMINTRIN_H) && \
     defined(HAVE_TMMINTRIN_H) && defined(HAVE_SMMINTRIN_H)
 
-# ifdef __GNUC__
-#  pragma GCC target("sse2")
-#  pragma GCC target("ssse3")
-#  pragma GCC target("sse4.1")
-#  pragma GCC target("avx2")
+# ifdef __clang__
+#  pragma clang attribute push(__attribute__((target("sse2,ssse3,sse4.1,avx2"))), apply_to = function)
+# elif defined(__GNUC__)
+#  pragma GCC target("sse2,ssse3,sse4.1,avx2")
 # endif
 
 # include <emmintrin.h>
@@ -46,4 +45,8 @@ blake2b_compress_avx2(blake2b_state *S, const uint8_t block[BLAKE2B_BLOCKBYTES])
     return 0;
 }
 
+# ifdef __clang__
+#  pragma clang attribute pop
+# endif
+
 #endif
diff --git a/src/libsodium/crypto_generichash/blake2b/ref/blake2b-compress-sse41.c b/src/libsodium/crypto_generichash/blake2b/ref/blake2b-compress-sse41.c
@@ -11,10 +11,10 @@
 #if defined(HAVE_EMMINTRIN_H) && defined(HAVE_TMMINTRIN_H) && \
     defined(HAVE_SMMINTRIN_H)
 
-# ifdef __GNUC__
-#  pragma GCC target("sse2")
-#  pragma GCC target("ssse3")
-#  pragma GCC target("sse4.1")
+# ifdef __clang__
+#  pragma clang attribute push(__attribute__((target("sse2,ssse3,sse4.1"))), apply_to = function)
+# elif defined(__GNUC__)
+#  pragma GCC target("sse2,ssse3,sse4.1")
 # endif
 
 # include <emmintrin.h>
@@ -84,4 +84,8 @@ blake2b_compress_sse41(blake2b_state *S,
     return 0;
 }
 
+# ifdef __clang__
+#  pragma clang attribute pop
+# endif
+
 #endif
diff --git a/src/libsodium/crypto_generichash/blake2b/ref/blake2b-compress-ssse3.c b/src/libsodium/crypto_generichash/blake2b/ref/blake2b-compress-ssse3.c
@@ -7,9 +7,10 @@
 
 #if defined(HAVE_EMMINTRIN_H) && defined(HAVE_TMMINTRIN_H)
 
-# ifdef __GNUC__
-#  pragma GCC target("sse2")
-#  pragma GCC target("ssse3")
+# ifdef __clang__
+#  pragma clang attribute push(__attribute__((target("sse2,ssse3"))), apply_to = function)
+# elif defined(__GNUC__)
+#  pragma GCC target("sse2,ssse3")
 # endif
 
 # include <emmintrin.h>
@@ -87,4 +88,8 @@ blake2b_compress_ssse3(blake2b_state *S,
     return 0;
 }
 
+# ifdef __clang__
+#  pragma clang attribute pop
+# endif
+
 #endif
diff --git a/src/libsodium/crypto_onetimeauth/poly1305/sse2/poly1305_sse2.c b/src/libsodium/crypto_onetimeauth/poly1305/sse2/poly1305_sse2.c
@@ -10,7 +10,9 @@
 
 #if defined(HAVE_TI_MODE) && defined(HAVE_EMMINTRIN_H)
 
-# ifdef __GNUC__
+# ifdef __clang__
+#  pragma clang attribute push(__attribute__((target("sse2"))), apply_to = function)
+# elif defined(__GNUC__)
 #  pragma GCC target("sse2")
 # endif
 
@@ -946,4 +948,8 @@ struct crypto_onetimeauth_poly1305_implementation
         SODIUM_C99(.onetimeauth_final =) crypto_onetimeauth_poly1305_sse2_final
     };
 
+#ifdef __clang__
+# pragma clang attribute pop
+#endif
+
 #endif
diff --git a/src/libsodium/crypto_pwhash/argon2/argon2-fill-block-avx2.c b/src/libsodium/crypto_pwhash/argon2/argon2-fill-block-avx2.c
@@ -22,11 +22,10 @@
 #if defined(HAVE_AVX2INTRIN_H) && defined(HAVE_EMMINTRIN_H) && \
     defined(HAVE_TMMINTRIN_H) && defined(HAVE_SMMINTRIN_H)
 
-# ifdef __GNUC__
-#  pragma GCC target("sse2")
-#  pragma GCC target("ssse3")
-#  pragma GCC target("sse4.1")
-#  pragma GCC target("avx2")
+# ifdef __clang__
+#  pragma clang attribute push(__attribute__((target("sse2,ssse3,sse4.1,avx2"))), apply_to = function)
+# elif defined(__GNUC__)
+#  pragma GCC target("sse2,ssse3,sse4.1,avx2")
 # endif
 
 # ifdef _MSC_VER
@@ -236,4 +235,9 @@ argon2_fill_segment_avx2(const argon2_instance_t *instance,
         }
     }
 }
+
+#ifdef __clang__
+# pragma clang attribute pop
+#endif
+
 #endif
diff --git a/src/libsodium/crypto_pwhash/argon2/argon2-fill-block-avx512f.c b/src/libsodium/crypto_pwhash/argon2/argon2-fill-block-avx512f.c
@@ -22,12 +22,10 @@
 #if defined(HAVE_AVX512FINTRIN_H) && defined(HAVE_AVX2INTRIN_H) && \
     defined(HAVE_EMMINTRIN_H) &&  defined(HAVE_TMMINTRIN_H) && defined(HAVE_SMMINTRIN_H)
 
-# ifdef __GNUC__
-#  pragma GCC target("sse2")
-#  pragma GCC target("ssse3")
-#  pragma GCC target("sse4.1")
-#  pragma GCC target("avx2")
-#  pragma GCC target("avx512f")
+# ifdef __clang__
+#  pragma clang attribute push(__attribute__((target("sse2,ssse3,sse4.1,avx2,avx512f"))), apply_to = function)
+# elif defined(__GNUC__)
+#  pragma GCC target("sse2,ssse3,sse4.1,avx2,avx512f")
 # endif
 
 # ifdef _MSC_VER
@@ -241,4 +239,9 @@ argon2_fill_segment_avx512f(const argon2_instance_t *instance,
         }
     }
 }
+
+#ifdef __clang__
+# pragma clang attribute pop
+#endif
+
 #endif
diff --git a/src/libsodium/crypto_pwhash/argon2/argon2-fill-block-ssse3.c b/src/libsodium/crypto_pwhash/argon2/argon2-fill-block-ssse3.c
@@ -21,9 +21,10 @@
 
 #if defined(HAVE_EMMINTRIN_H) && defined(HAVE_TMMINTRIN_H)
 
-# ifdef __GNUC__
-#  pragma GCC target("sse2")
-#  pragma GCC target("ssse3")
+# ifdef __clang__
+#  pragma clang attribute push(__attribute__((target("sse2,ssse3"))), apply_to = function)
+# elif defined(__GNUC__)
+#  pragma GCC target("sse2,ssse3")
 # endif
 
 # ifdef _MSC_VER
@@ -235,4 +236,9 @@ argon2_fill_segment_ssse3(const argon2_instance_t *instance,
         }
     }
 }
+
+#ifdef __clang__
+# pragma clang attribute pop
+#endif
+
 #endif