Merge pull request #201 from symmetree-labs/master

Improve compile-time target detection for NEON

.github/workflows/ci.yml

@@ -149,6 +149,9 @@ jobs:
     # Test the NEON implementation on ARM targets.
     - run: cross test --target ${{ matrix.arch }} --features=neon
       if: startsWith(matrix.arch, 'armv7-') || startsWith(matrix.arch, 'aarch64-')
+    # NEON is enabled by default on aarch64, disabling it through the no_neon feature.
+    - run: cross test --target ${{ matrix.arch }} --features=no_neon
+      if: startsWith(matrix.arch, 'aarch64-')
     # Test vectors. Note that this uses a hacky script due to path dependency limitations.
     - run: ./test_vectors/cross_test.sh --target ${{ matrix.arch }}
     # C code. Same issue with the hacky script.

Cargo.toml

@@ -73,6 +73,7 @@ no_sse2 = []
 no_sse41 = []
 no_avx2 = []
 no_avx512 = []
+no_neon = []
 
 [package.metadata.docs.rs]
 # Document Hasher::update_rayon on docs.rs.

build.rs

@@ -17,6 +17,10 @@ fn is_neon() -> bool {
     defined("CARGO_FEATURE_NEON")
 }
 
+fn is_no_neon() -> bool {
+    defined("CARGO_FEATURE_NO_NEON")
+}
+
 fn is_ci() -> bool {
     defined("BLAKE3_CI")
 }
@@ -44,6 +48,14 @@ fn is_x86_32() -> bool {
     arch == "i386" || arch == "i586" || arch == "i686"
 }
 
+fn is_arm() -> bool {
+    is_armv7() || is_aarch64() || target_components()[0] == "arm"
+}
+
+fn is_aarch64() -> bool {
+    target_components()[0] == "aarch64"
+}
+
 fn is_armv7() -> bool {
     target_components()[0] == "armv7"
 }
@@ -218,6 +230,10 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
         panic!("It doesn't make sense to enable both \"pure\" and \"neon\".");
     }
 
+    if is_no_neon() && is_neon() {
+        panic!("It doesn't make sense to enable both \"no_neon\" and \"neon\".");
+    }
+
     if is_x86_64() || is_x86_32() {
         let support = c_compiler_support();
         if is_x86_32() || should_prefer_intrinsics() || is_pure() || support == NoCompiler {
@@ -237,7 +253,8 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
         }
     }
 
-    if is_neon() {
+    if (is_arm() && is_neon()) || (!is_no_neon() && !is_pure() && is_aarch64()) {
+        println!("cargo:rustc-cfg=blake3_neon");
         build_neon_c_intrinsics();
     }
 

c/Makefile.testing

@@ -38,10 +38,14 @@ ASM_TARGETS += blake3_avx512_x86-64_unix.S
 endif
 
 ifdef BLAKE3_USE_NEON
-EXTRAFLAGS += -DBLAKE3_USE_NEON
+EXTRAFLAGS += -DBLAKE3_USE_NEON=1
 TARGETS += blake3_neon.o
 endif
 
+ifdef BLAKE3_NO_NEON
+EXTRAFLAGS += -DBLAKE3_USE_NEON=0
+endif
+
 all: blake3.c blake3_dispatch.c blake3_portable.c main.c $(TARGETS)
 	$(CC) $(CFLAGS) $(EXTRAFLAGS) $^ -o $(NAME) $(LDFLAGS)
 

c/README.md

@@ -250,15 +250,24 @@ gcc -shared -O3 -o libblake3.so -DBLAKE3_NO_SSE2 -DBLAKE3_NO_SSE41 -DBLAKE3_NO_A
 
 ## ARM NEON
 
-The NEON implementation is not enabled by default on ARM, since not all
-ARM targets support it. To enable it, set `BLAKE3_USE_NEON=1`. Here's an
-example of building a shared library on ARM Linux with NEON support:
+The NEON implementation is enabled by default on AARCH64, but not on
+other ARM targets, since not all of them support it. To enable it, set
+`BLAKE3_USE_NEON=1`. Here's an example of building a shared library on
+ARM Linux with NEON support:
 
 ```bash
-gcc -shared -O3 -o libblake3.so -DBLAKE3_USE_NEON blake3.c blake3_dispatch.c \
+gcc -shared -O3 -o libblake3.so -DBLAKE3_USE_NEON=1 blake3.c blake3_dispatch.c \
     blake3_portable.c blake3_neon.c
 ```
 
+To explicitiy disable using NEON instructions on AARCH64, set
+`BLAKE3_USE_NEON=0`.
+
+```bash
+gcc -shared -O3 -o libblake3.so -DBLAKE3_USE_NEON=0 blake3.c blake3_dispatch.c \
+    blake3_portable.c 
+```
+
 Note that on some targets (ARMv7 in particular), extra flags may be
 required to activate NEON support in the compiler. If you see an error
 like...

c/blake3_c_rust_bindings/build.rs

@@ -22,6 +22,10 @@ fn is_armv7() -> bool {
     target_components()[0] == "armv7"
 }
 
+fn is_aarch64() -> bool {
+    target_components()[0] == "aarch64"
+}
+
 // Windows targets may be using the MSVC toolchain or the GNU toolchain. The
 // right compiler flags to use depend on the toolchain. (And we don't want to
 // use flag_if_supported, because we don't want features to be silently
@@ -148,10 +152,14 @@ fn main() -> Result<(), Box<dyn std::error::Error>> {
         avx512_build.compile("blake3_avx512");
     }
 
-    // We only build NEON code here if 1) it's requested and 2) the root crate
-    // is not already building it. The only time this will really happen is if
-    // you build this crate by hand with the "neon" feature for some reason.
-    if defined("CARGO_FEATURE_NEON") {
+    // We only build NEON code here if
+    // 1) it's requested
+    // and 2) the root crate is not already building it.
+    // The only time this will really happen is if you build this
+    // crate by hand with the "neon" feature for some reason.
+    //
+    // In addition, 3) if the target is aarch64, NEON is on by default.
+    if defined("CARGO_FEATURE_NEON") || is_aarch64() {
         let mut neon_build = new_build();
         neon_build.file(c_dir_path("blake3_neon.c"));
         // ARMv7 platforms that support NEON generally need the following

c/blake3_dispatch.c

@@ -232,7 +232,7 @@ void blake3_hash_many(const uint8_t *const *inputs, size_t num_inputs,
 #endif
 #endif
 
-#if defined(BLAKE3_USE_NEON)
+#if BLAKE3_USE_NEON == 1
   blake3_hash_many_neon(inputs, num_inputs, blocks, key, counter,
                         increment_counter, flags, flags_start, flags_end, out);
   return;
@@ -269,7 +269,7 @@ size_t blake3_simd_degree(void) {
   }
 #endif
 #endif
-#if defined(BLAKE3_USE_NEON)
+#if BLAKE3_USE_NEON == 1
   return 4;
 #endif
   return 1;

c/blake3_impl.h

@@ -38,6 +38,10 @@ enum blake3_flags {
 #define IS_X86_32
 #endif
 
+#if defined(__aarch64__) || defined(_M_ARM64)
+#define IS_AARCH64
+#endif
+
 #if defined(IS_X86)
 #if defined(_MSC_VER)
 #include <intrin.h>
@@ -45,9 +49,18 @@ enum blake3_flags {
 #include <immintrin.h>
 #endif
 
+#if !defined(BLAKE3_USE_NEON) 
+  // If BLAKE3_USE_NEON not manually set, autodetect based on AArch64ness
+  #if defined(IS_AARCH64)
+    #define BLAKE3_USE_NEON 1
+  #else
+    #define BLAKE3_USE_NEON 0
+  #endif
+#endif
+
 #if defined(IS_X86)
 #define MAX_SIMD_DEGREE 16
-#elif defined(BLAKE3_USE_NEON)
+#elif BLAKE3_USE_NEON == 1
 #define MAX_SIMD_DEGREE 4
 #else
 #define MAX_SIMD_DEGREE 1
@@ -257,7 +270,7 @@ void blake3_hash_many_avx512(const uint8_t *const *inputs, size_t num_inputs,
 #endif
 #endif
 
-#if defined(BLAKE3_USE_NEON)
+#if BLAKE3_USE_NEON == 1
 void blake3_hash_many_neon(const uint8_t *const *inputs, size_t num_inputs,
                            size_t blocks, const uint32_t key[8],
                            uint64_t counter, bool increment_counter,

src/lib.rs

@@ -94,7 +94,7 @@ mod avx2;
 #[cfg(blake3_avx512_ffi)]
 #[path = "ffi_avx512.rs"]
 mod avx512;
-#[cfg(feature = "neon")]
+#[cfg(blake3_neon)]
 #[path = "ffi_neon.rs"]
 mod neon;
 mod portable;

src/platform.rs

@@ -10,7 +10,7 @@ cfg_if::cfg_if! {
                 pub const MAX_SIMD_DEGREE: usize = 8;
             }
         }
-    } else if #[cfg(feature = "neon")] {
+    } else if #[cfg(blake3_neon)] {
         pub const MAX_SIMD_DEGREE: usize = 4;
     } else {
         pub const MAX_SIMD_DEGREE: usize = 1;
@@ -30,7 +30,7 @@ cfg_if::cfg_if! {
                 pub const MAX_SIMD_DEGREE_OR_2: usize = 8;
             }
         }
-    } else if #[cfg(feature = "neon")] {
+    } else if #[cfg(blake3_neon)] {
         pub const MAX_SIMD_DEGREE_OR_2: usize = 4;
     } else {
         pub const MAX_SIMD_DEGREE_OR_2: usize = 2;
@@ -49,7 +49,7 @@ pub enum Platform {
     #[cfg(blake3_avx512_ffi)]
     #[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
     AVX512,
-    #[cfg(feature = "neon")]
+    #[cfg(blake3_neon)]
     NEON,
 }
 
@@ -76,7 +76,7 @@ impl Platform {
         }
         // We don't use dynamic feature detection for NEON. If the "neon"
         // feature is on, NEON is assumed to be supported.
-        #[cfg(feature = "neon")]
+        #[cfg(blake3_neon)]
         {
             return Platform::NEON;
         }
@@ -95,7 +95,7 @@ impl Platform {
             #[cfg(blake3_avx512_ffi)]
             #[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
             Platform::AVX512 => 16,
-            #[cfg(feature = "neon")]
+            #[cfg(blake3_neon)]
             Platform::NEON => 4,
         };
         debug_assert!(degree <= MAX_SIMD_DEGREE);
@@ -129,7 +129,7 @@ impl Platform {
                 crate::avx512::compress_in_place(cv, block, block_len, counter, flags)
             },
             // No NEON compress_in_place() implementation yet.
-            #[cfg(feature = "neon")]
+            #[cfg(blake3_neon)]
             Platform::NEON => portable::compress_in_place(cv, block, block_len, counter, flags),
         }
     }
@@ -161,7 +161,7 @@ impl Platform {
                 crate::avx512::compress_xof(cv, block, block_len, counter, flags)
             },
             // No NEON compress_xof() implementation yet.
-            #[cfg(feature = "neon")]
+            #[cfg(blake3_neon)]
             Platform::NEON => portable::compress_xof(cv, block, block_len, counter, flags),
         }
     }
@@ -256,7 +256,7 @@ impl Platform {
                 )
             },
             // Assumed to be safe if the "neon" feature is on.
-            #[cfg(feature = "neon")]
+            #[cfg(blake3_neon)]
             Platform::NEON => unsafe {
                 crate::neon::hash_many(
                     inputs,
@@ -315,7 +315,7 @@ impl Platform {
         }
     }
 
-    #[cfg(feature = "neon")]
+    #[cfg(blake3_neon)]
     pub fn neon() -> Option<Self> {
         // Assumed to be safe if the "neon" feature is on.
         Some(Self::NEON)