fix aarch64 compile and fix previous llvm asm issue for crc64 (pingca…

…p#3063)
v0y4g3r · Sep 10, 2021 · c0d70d6 · c0d70d6
1 parent bc9b63e
commit c0d70d6
Showing 1 changed file with 12 additions and 18 deletions.
diff --git a/libs/libcommon/include/common/crc64_arch/crc64_aarch64.h b/libs/libcommon/include/common/crc64_arch/crc64_aarch64.h
@@ -4,17 +4,19 @@
 
 #include <cstdint>
 #include <cstring>
+#include <utility>
 
 namespace crc64::_detail
 {
-using simd_t = uint8x16_t;
+typedef uint8x16_t simd_t __attribute__((__may_alias__));
 
 class SIMD
 {
 public:
     using Poly64Pair = std::pair<poly64_t, poly64_t>;
 
-    SIMD(uint64_t high, uint64_t low) noexcept
+    SIMD(uint64_t high, uint64_t low)
+    noexcept;
 
     [[nodiscard]] SIMD fold16(SIMD coeff) const noexcept;
 
@@ -66,23 +68,15 @@ inline SIMD SIMD::fold8(uint64_t coeff) const noexcept
 
 inline SIMD SIMD::fold16(SIMD coeff) const noexcept
 {
-    /* GCC does not seem to use pmull2 properly and we hence use
-       * inline assembly to improve the performance.
-       * For macOS (aarch64), however, AppleClang does not recognize
-       * the assembly code `=w`, so we fallback to normal ways.
-       */
-#if defined(__APPLE__) || defined(__OSX__)
-    auto [x0, x1] = into_poly64pair();
-    auto [c0, c1] = coeff.into_poly64pair();
-    auto h = SIMD::from_mul(c0, x0);
-    auto l = SIMD::from_mul(c1, x1);
-#else
+    /* GCC/Clang does not seem to use pmull2 properly and we hence use
+     * inline assembly to improve the performance.
+     */
+
     SIMD h(0, 0), l(0, 0);
-    asm("pmull %0.1q, %2.1d, %3.1d\n"
-        "pmull2 %1.1q, %2.2d, %3.2d"
-        : "=&w"(l), "=w"(h)
-        : "w"(*this), "w"(coeff));
-#endif
+    asm("pmull\t%0.1q, %2.1d, %3.1d\n\t"
+        "pmull2\t%1.1q, %2.2d, %3.2d"
+        : "=&w"(l._inner), "=w"(h._inner)
+        : "w"(this->_inner), "w"(coeff._inner));
     return h.bitxor(l);
 }