@img/sharp-libvips-dev-wasm32 1.3.0-rc.6 → 1.3.0-rc.8

package/include/glib-2.0/gobject/gtype.h

@@ -1574,7 +1574,7 @@ guint     g_type_get_type_registration_serial (void);
  * - the usual `my_app_window_get_type()` function is declared with a return type of #GType
  *
  * - the `MyAppWindow` type is defined as a `typedef` of `struct _MyAppWindow`.  The struct itself is not
- *   defined and should be defined from the .c file before G_DEFINE_TYPE() is used.
+ *   defined and should be defined from the `.c` file before `G_DEFINE_TYPE()` is used.
  *
  * - the `MY_APP_WINDOW()` cast is emitted as `static inline` function along with the `MY_APP_IS_WINDOW()` type
  *   checking function
@@ -1585,9 +1585,9 @@ guint     g_type_get_type_registration_serial (void);
  *   and therefore the fact that the size of the class structure is exposed is not a concern and it can be
  *   freely changed at any point in the future.
  *
- * - g_autoptr() support being added for your type, based on the type of your parent class
+ * - `g_autoptr()` support being added for your type, based on the type of your parent class
  *
- * You can only use this function if your parent type also supports g_autoptr().
+ * You can only use this function if your parent type also supports `g_autoptr()`.
  *
  * Because the type macro (`MY_APP_TYPE_WINDOW` in the above example) is not a callable, you must continue to
  * manually define this as a macro for yourself.
@@ -1595,11 +1595,11 @@ guint     g_type_get_type_registration_serial (void);
  * The declaration of the `_get_type()` function is the first thing emitted by the macro.  This allows this macro
  * to be used in the usual way with export control and API versioning macros.
  *
- * If you want to declare your own class structure, use G_DECLARE_DERIVABLE_TYPE().
+ * If you want to declare your own class structure, use [func@GObject.DECLARE_DERIVABLE_TYPE].
  *
  * If you are writing a library, it is important to note that it is possible to convert a type from using
- * G_DECLARE_FINAL_TYPE() to G_DECLARE_DERIVABLE_TYPE() without breaking API or ABI.  As a precaution, you
- * should therefore use G_DECLARE_FINAL_TYPE() until you are sure that it makes sense for your class to be
+ * `G_DECLARE_FINAL_TYPE()` to `G_DECLARE_DERIVABLE_TYPE()` without breaking API or ABI.  As a precaution, you
+ * should therefore use `G_DECLARE_FINAL_TYPE()` until you are sure that it makes sense for your class to be
  * subclassed.  Once a class structure has been exposed it is not possible to change its size or remove or
  * reorder items without breaking the API and/or ABI.
  *
@@ -1687,7 +1687,7 @@ guint     g_type_get_type_registration_serial (void);
  *
  * - g_autoptr() support being added for your type, based on the type of your parent class
  *
- * You can only use this function if your parent type also supports g_autoptr().
+ * You can only use this function if your parent type also supports `g_autoptr()`.
  *
  * Because the type macro (`GTK_TYPE_FROBBER` in the above example) is not a callable, you must continue to
  * manually define this as a macro for yourself.
@@ -1696,14 +1696,14 @@ guint     g_type_get_type_registration_serial (void);
  * to be used in the usual way with export control and API versioning macros.
  *
  * If you are writing a library, it is important to note that it is possible to convert a type from using
- * G_DECLARE_FINAL_TYPE() to G_DECLARE_DERIVABLE_TYPE() without breaking API or ABI.  As a precaution, you
- * should therefore use G_DECLARE_FINAL_TYPE() until you are sure that it makes sense for your class to be
+ * [func@GObject.DECLARE_FINAL_TYPE] to `G_DECLARE_DERIVABLE_TYPE()` without breaking API or ABI.  As a precaution, you
+ * should therefore use `G_DECLARE_FINAL_TYPE()` until you are sure that it makes sense for your class to be
  * subclassed.  Once a class structure has been exposed it is not possible to change its size or remove or
  * reorder items without breaking the API and/or ABI.  If you want to declare your own class structure, use
- * G_DECLARE_DERIVABLE_TYPE().  If you want to declare a class without exposing the class or instance
- * structures, use G_DECLARE_FINAL_TYPE().
+ * `G_DECLARE_DERIVABLE_TYPE()`.  If you want to declare a class without exposing the class or instance
+ * structures, use `G_DECLARE_FINAL_TYPE()`.
  *
- * If you must use G_DECLARE_DERIVABLE_TYPE() you should be sure to include some padding at the bottom of your
+ * If you must use `G_DECLARE_DERIVABLE_TYPE()` you should be sure to include some padding at the bottom of your
  * class structure to leave space for the addition of future virtual functions.
  *
  * Since: 2.44
@@ -1931,9 +1931,9 @@ guint     g_type_get_type_registration_serial (void);
  *
  * A convenience macro for type implementations.
  *
- * Similar to G_DEFINE_TYPE(), but defines a final type.
+ * Similar to [func@GObject.DEFINE_TYPE], but defines a final type.
  *
- * See G_DEFINE_TYPE_EXTENDED() for an example.
+ * See [func@GObject.DEFINE_TYPE_EXTENDED] for an example.
  *
  * Since: 2.70
  */

package/include/hwy/aligned_allocator.h

@@ -119,8 +119,21 @@ template <typename T, typename... Args>
 AlignedUniquePtr<T> MakeUniqueAlignedWithAlloc(AllocPtr alloc, FreePtr free,
                                                void* opaque, Args&&... args) {
   T* ptr = static_cast<T*>(AllocateAlignedBytes(sizeof(T), alloc, opaque));
+  if (HWY_UNLIKELY(ptr == nullptr)) {
+    return AlignedUniquePtr<T>(nullptr, AlignedDeleter(free, opaque));
+  }
+#ifdef HWY_EXCEPTIONS_ENABLED
+  try {
+    return AlignedUniquePtr<T>(new (ptr) T(std::forward<Args>(args)...),
+                               AlignedDeleter(free, opaque));
+  } catch (...) {
+    FreeAlignedBytes(ptr, free, opaque);
+    throw;
+  }
+#else
   return AlignedUniquePtr<T>(new (ptr) T(std::forward<Args>(args)...),
                              AlignedDeleter(free, opaque));
+#endif
 }
 
 // Similar to MakeUniqueAlignedWithAlloc but using the default alloc/free
@@ -128,8 +141,21 @@ AlignedUniquePtr<T> MakeUniqueAlignedWithAlloc(AllocPtr alloc, FreePtr free,
 template <typename T, typename... Args>
 AlignedUniquePtr<T> MakeUniqueAligned(Args&&... args) {
   T* ptr = static_cast<T*>(AllocateAlignedBytes(sizeof(T)));
+  if (HWY_UNLIKELY(ptr == nullptr)) {
+    return AlignedUniquePtr<T>(nullptr, AlignedDeleter());
+  }
+#ifdef HWY_EXCEPTIONS_ENABLED
+  try {
+    return AlignedUniquePtr<T>(new (ptr) T(std::forward<Args>(args)...),
+                               AlignedDeleter());
+  } catch (...) {
+    FreeAlignedBytes(ptr, nullptr, nullptr);
+    throw;
+  }
+#else
   return AlignedUniquePtr<T>(new (ptr) T(std::forward<Args>(args)...),
                              AlignedDeleter());
+#endif
 }
 
 template <class T>
@@ -147,8 +173,13 @@ struct AlignedAllocator {
                   "AlignedAllocator only supports integer types");
     static_assert(sizeof(V) <= sizeof(std::size_t),
                   "V n must be smaller or equal size_t to avoid overflow");
+    const size_t count = static_cast<std::size_t>(n);
+    if (HWY_LIKELY(count != 0) && sizeof(value_type) > SIZE_MAX / count) {
+      HWY_ABORT("AlignedAllocator: allocation size overflow "
+                 "(%zu * %zu exceeds size_t)", count, sizeof(value_type));
+    }
     return static_cast<value_type*>(
-        AllocateAlignedBytes(static_cast<std::size_t>(n) * sizeof(value_type)));
+        AllocateAlignedBytes(count * sizeof(value_type)));
   }
 
   template <class V>
@@ -208,7 +239,7 @@ AlignedUniquePtr<T[]> MakeUniqueAlignedArrayWithAlloc(
   T* ptr = detail::AllocateAlignedItems<T>(items, alloc, opaque);
   if (ptr != nullptr) {
     for (size_t i = 0; i < items; i++) {
-      new (ptr + i) T(std::forward<Args>(args)...);
+      new (ptr + i) T(args...);
     }
   }
   return AlignedUniquePtr<T[]>(ptr, AlignedDeleter(free, opaque));
@@ -399,6 +430,10 @@ class AlignedNDArray {
     size_t offset = 0;
     size_t shape_index = 0;
     for (const size_t axis_index : indices) {
+      if (HWY_UNLIKELY(axis_index >= shape_[shape_index])) {
+        HWY_ABORT("AlignedNDArray index %zu out of bounds (axis %zu, size %zu)",
+                   axis_index, shape_index, shape_[shape_index]);
+      }
       offset += memory_sizes_[shape_index + 1] * axis_index;
       shape_index++;
     }
@@ -417,6 +452,13 @@ class AlignedNDArray {
     sizes[axis] = 1;
     while (axis > 0) {
       --axis;
+      // Check for integer overflow in dimension multiplication.
+      if (HWY_LIKELY(shape[axis] != 0) &&
+          sizes[axis + 1] > SIZE_MAX / shape[axis]) {
+        HWY_ABORT("AlignedNDArray: dimension overflow at axis %zu "
+                   "(%zu * %zu exceeds size_t)",
+                   axis, sizes[axis + 1], shape[axis]);
+      }
       sizes[axis] = sizes[axis + 1] * shape[axis];
     }
     return sizes;

package/include/hwy/auto_tune.h

@@ -25,7 +25,18 @@
 
 #include "hwy/aligned_allocator.h"  // Span
 #include "hwy/base.h"               // HWY_MIN
-#include "hwy/contrib/sort/vqsort.h"
+
+// configuration to allow auto_tune to use std::sort instead of VQSort
+// (also enabled in header only mode).
+#if defined(HWY_HEADER_ONLY)
+#define HWY_AUTOTUNE_STDSORT
+#endif
+
+#ifdef HWY_AUTOTUNE_STDSORT
+#include <algorithm>  // std::sort
+#else
+#include "hwy/contrib/sort/vqsort.h"  // VQSort
+#endif
 
 // Infrastructure for auto-tuning (choosing optimal parameters at runtime).
 
@@ -104,6 +115,10 @@ class CostDistribution {
  private:
   static double Median(double* to_sort, size_t n) {
     HWY_DASSERT(n >= 2);
+
+#ifdef HWY_AUTOTUNE_STDSORT
+    std::sort(to_sort, to_sort + n);
+#else
 // F64 is supported everywhere except Armv7.
 #if !HWY_ARCH_ARM_V7
     VQSort(to_sort, n, SortAscending());
@@ -112,6 +127,8 @@ class CostDistribution {
     // equivalent.
     VQSort(reinterpret_cast<uint64_t*>(to_sort), n, SortAscending());
 #endif
+#endif
+
     if (n & 1) return to_sort[n / 2];
     // Even length: average of two middle elements.
     return (to_sort[n / 2] + to_sort[n / 2 - 1]) * 0.5;
@@ -405,10 +422,9 @@ class AutoTune {
   const Config* Best() const { return best_; }
 
   // If false, caller must call `SetCandidates` before `NextConfig`.
-  bool HasCandidates() const {
-    HWY_DASSERT(!Best());
-    return !candidates_.empty();
-  }
+  // NOTE: also called after Best() is non-null.
+  bool HasCandidates() const { return !candidates_.empty(); }
+
   // WARNING: invalidates `Best()`, do not call if that is non-null.
   void SetCandidates(std::vector<Config> candidates) {
     HWY_DASSERT(!Best() && !HasCandidates());
@@ -429,7 +445,7 @@ class AutoTune {
 
   // Returns the current `Config` to measure.
   const Config& NextConfig() const {
-    HWY_DASSERT(!Best() && HasCandidates());
+    HWY_DASSERT(HasCandidates());
     return candidates_[config_idx_];
   }
 

package/include/hwy/base.h

@@ -22,8 +22,8 @@
 #include <stddef.h>
 #include <stdint.h>
 #if defined(HWY_HEADER_ONLY)
-#include <cstdarg>
-#include <cstdio>
+#include <stdarg.h>
+#include <stdio.h>
 #endif
 
 #if !defined(HWY_NO_LIBCXX)
@@ -33,9 +33,10 @@
 #include "hwy/detect_compiler_arch.h"
 #include "hwy/highway_export.h"
 
-// API version (https://semver.org/); keep in sync with CMakeLists.txt.
+// API version (https://semver.org/); keep in sync with CMakeLists.txt and
+// meson.build.
 #define HWY_MAJOR 1
-#define HWY_MINOR 3
+#define HWY_MINOR 4
 #define HWY_PATCH 0
 
 // True if the Highway version >= major.minor.0. Added in 1.2.0.
@@ -57,7 +58,7 @@
 
 #endif  // !HWY_IDE
 
-#if (HWY_ARCH_X86 && !defined(HWY_NO_LIBCXX)) || HWY_COMPILER_MSVC
+#if !defined(HWY_NO_LIBCXX) || HWY_COMPILER_MSVC
 #include <atomic>
 #endif
 
@@ -148,6 +149,21 @@
 
 #endif  // !HWY_COMPILER_MSVC
 
+#if (HWY_COMPILER_GCC_ACTUAL && HWY_COMPILER_GCC_ACTUAL < 1200) || \
+    (HWY_COMPILER_ICC && !HWY_COMPILER_ICX)
+// The use of __attribute__((unused)) in private class member variables triggers
+// a compiler warning with GCC 11 and earlier and ICC
+
+// GCC 11 and earlier and ICC also do not emit -Wunused-private-field warnings
+// for unused private class member variables
+#define HWY_MEMBER_VAR_MAYBE_UNUSED
+#else
+// Clang and ICX need __attribute__((unused)) in unused private class member
+// variables to suppress -Wunused-private-field warnings unless this warning is
+// ignored by using HWY_DIAGNOSTICS_OFF
+#define HWY_MEMBER_VAR_MAYBE_UNUSED HWY_MAYBE_UNUSED
+#endif
+
 //------------------------------------------------------------------------------
 // Builtin/attributes (no more #include after this point due to namespace!)
 
@@ -202,7 +218,9 @@ namespace hwy {
 //------------------------------------------------------------------------------
 // Macros
 
-#define HWY_API static HWY_INLINE HWY_FLATTEN HWY_MAYBE_UNUSED
+// Note: it is safe to remove `static` for users who want to use modules, but
+// that might be a breaking change for some users, hence we do not by default.
+#define HWY_API static HWY_INLINE HWY_FLATTEN
 
 #define HWY_CONCAT_IMPL(a, b) a##b
 #define HWY_CONCAT(a, b) HWY_CONCAT_IMPL(a, b)
@@ -349,60 +367,20 @@ HWY_DLLEXPORT HWY_NORETURN void HWY_FORMAT(3, 4)
   } while (0)
 #define HWY_ASSERT(condition) HWY_ASSERT_M(condition, "")
 
-#if HWY_HAS_FEATURE(memory_sanitizer) || defined(MEMORY_SANITIZER) || \
-    defined(__SANITIZE_MEMORY__)
-#define HWY_IS_MSAN 1
-#else
-#define HWY_IS_MSAN 0
-#endif
-
-#if HWY_HAS_FEATURE(address_sanitizer) || defined(ADDRESS_SANITIZER) || \
-    defined(__SANITIZE_ADDRESS__)
-#define HWY_IS_ASAN 1
-#else
-#define HWY_IS_ASAN 0
-#endif
-
-#if HWY_HAS_FEATURE(hwaddress_sanitizer) || defined(HWADDRESS_SANITIZER) || \
-    defined(__SANITIZE_HWADDRESS__)
-#define HWY_IS_HWASAN 1
-#else
-#define HWY_IS_HWASAN 0
-#endif
-
-#if HWY_HAS_FEATURE(thread_sanitizer) || defined(THREAD_SANITIZER) || \
-    defined(__SANITIZE_THREAD__)
-#define HWY_IS_TSAN 1
-#else
-#define HWY_IS_TSAN 0
-#endif
-
-#if HWY_HAS_FEATURE(undefined_behavior_sanitizer) || \
-    defined(UNDEFINED_BEHAVIOR_SANITIZER)
-#define HWY_IS_UBSAN 1
-#else
-#define HWY_IS_UBSAN 0
-#endif
-
-// MSAN may cause lengthy build times or false positives e.g. in AVX3 DemoteTo.
-// You can disable MSAN by adding this attribute to the function that fails.
-#if HWY_IS_MSAN
-#define HWY_ATTR_NO_MSAN __attribute__((no_sanitize_memory))
-#else
-#define HWY_ATTR_NO_MSAN
-#endif
-
-#if HWY_IS_ASAN || HWY_IS_HWASAN || HWY_IS_MSAN || HWY_IS_TSAN || HWY_IS_UBSAN
-#define HWY_IS_SANITIZER 1
-#else
-#define HWY_IS_SANITIZER 0
-#endif
-
 // For enabling HWY_DASSERT and shortening tests in slower debug builds
+//
+// Note: `HWY_IS_UBSAN` is specifically excluded from engaging debug
+// builds. This is in service of Chromium's `-fsanitize=array-bounds` by
+// default, where we don't want Highway to unconditionally build in
+// debug mode.
+//
+// See also:
+// https://docs.google.com/document/d/1eCtY4AZF-SiFHxhIYWzEytdIx3C24de7ccD6Y5Gn2H8/edit?tab=t.9zkn85hr82ms#heading=h.efcshvfql42c
 #if !defined(HWY_IS_DEBUG_BUILD)
 // Clang does not define NDEBUG, but it and GCC define __OPTIMIZE__, and recent
 // MSVC defines NDEBUG (if not, could instead check _DEBUG).
-#if (!defined(__OPTIMIZE__) && !defined(NDEBUG)) || HWY_IS_SANITIZER || \
+#if (!defined(__OPTIMIZE__) && !defined(NDEBUG)) ||         \
+    (HWY_IS_ASAN || (HWY_IS_SANITIZER && !HWY_IS_UBSAN)) || \
     defined(__clang_analyzer__)
 #define HWY_IS_DEBUG_BUILD 1
 #else
@@ -455,21 +433,22 @@ HWY_API void CopySameSize(const From* HWY_RESTRICT from, To* HWY_RESTRICT to) {
   CopyBytes<sizeof(From)>(from, to);
 }
 
-template <size_t kBytes, typename To>
-HWY_API void ZeroBytes(To* to) {
+// Sets each byte to `byte_value`, which must be between 0 and 255.
+HWY_API void FillBytes(void* to, int byte_value, size_t num_bytes) {
 #if HWY_COMPILER_MSVC
-  memset(to, 0, kBytes);
+  memset(to, byte_value, num_bytes);
 #else
-  __builtin_memset(to, 0, kBytes);
+  __builtin_memset(to, byte_value, num_bytes);
 #endif
 }
 
 HWY_API void ZeroBytes(void* to, size_t num_bytes) {
-#if HWY_COMPILER_MSVC
-  memset(to, 0, num_bytes);
-#else
-  __builtin_memset(to, 0, num_bytes);
-#endif
+  FillBytes(to, 0, num_bytes);
+}
+
+template <size_t kBytes, typename To>
+HWY_API void ZeroBytes(To* to) {
+  ZeroBytes(to, kBytes);
 }
 
 //------------------------------------------------------------------------------
@@ -1699,28 +1678,14 @@ HWY_F16_CONSTEXPR inline std::partial_ordering operator<=>(
 //------------------------------------------------------------------------------
 // BF16 lane type
 
-// Compiler supports ACLE __bf16, not necessarily with operators.
-
-// Disable the __bf16 type on AArch64 with GCC 13 or earlier as there is a bug
-// in GCC 13 and earlier that sometimes causes BF16 constant values to be
-// incorrectly loaded on AArch64, and this GCC bug on AArch64 is
-// described at https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111867.
-
-#if HWY_ARCH_ARM_A64 && \
-    (HWY_COMPILER_CLANG >= 1700 || HWY_COMPILER_GCC_ACTUAL >= 1400)
-#define HWY_ARM_HAVE_SCALAR_BF16_TYPE 1
-#else
-#define HWY_ARM_HAVE_SCALAR_BF16_TYPE 0
-#endif
-
 // x86 compiler supports __bf16, not necessarily with operators.
 // Disable in debug builds due to clang miscompiles as of 2025-07-22: casting
 // bf16 <-> f32 in convert_test results in 0x2525 for 1.0 instead of 0x3f80.
 // Reported at https://github.com/llvm/llvm-project/issues/151692.
 #ifndef HWY_SSE2_HAVE_SCALAR_BF16_TYPE
-#if HWY_ARCH_X86 && defined(__SSE2__) &&                     \
-    ((HWY_COMPILER_CLANG >= 1700 && !HWY_COMPILER_CLANGCL && \
-      !HWY_IS_DEBUG_BUILD) ||                                \
+#if HWY_ARCH_X86 && defined(__SSE2__) &&                         \
+    ((HWY_COMPILER_CLANG >= 1700 && !HWY_COMPILER_CLANGCL &&     \
+      (!HWY_IS_DEBUG_BUILD || HWY_COMPILER3_CLANG >= 220101)) || \
      HWY_COMPILER_GCC_ACTUAL >= 1300)
 #define HWY_SSE2_HAVE_SCALAR_BF16_TYPE 1
 #else
@@ -1738,7 +1703,11 @@ HWY_F16_CONSTEXPR inline std::partial_ordering operator<=>(
 #ifndef HWY_HAVE_SCALAR_BF16_OPERATORS
 // Recent enough compiler also has operators. aarch64 clang 18 hits internal
 // compiler errors on bf16 ToString, hence only enable on GCC for now.
-#if HWY_HAVE_SCALAR_BF16_TYPE && (HWY_COMPILER_GCC_ACTUAL >= 1300)
+// GCC >= 13 will insert a function call to the __extendbfsf2 helper function
+// for scalar conversions from __bf16 to float. This is prohibitively expensive,
+// so refrain from using scalar BF16 operators on these compiler versions.
+// See https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121853
+#if HWY_HAVE_SCALAR_BF16_TYPE && (HWY_COMPILER_GCC_ACTUAL >= 1700)
 #define HWY_HAVE_SCALAR_BF16_OPERATORS 1
 #else
 #define HWY_HAVE_SCALAR_BF16_OPERATORS 0
@@ -2734,7 +2703,7 @@ HWY_API constexpr TTo ConvertScalarTo(TFrom in) {
 template <typename T1, typename T2>
 constexpr inline T1 DivCeil(T1 a, T2 b) {
 #if HWY_CXX_LANG >= 201703L
-  HWY_DASSERT(b != 0);
+  HWY_DASSERT(b != T2{0});
 #endif
   return (a + b - 1) / b;
 }
@@ -2957,9 +2926,10 @@ HWY_INLINE constexpr T AddWithWraparound(T t, T2 n) {
 // 64 x 64 = 128 bit multiplication
 HWY_API uint64_t Mul128(uint64_t a, uint64_t b, uint64_t* HWY_RESTRICT upper) {
 #if defined(__SIZEOF_INT128__)
-  __uint128_t product = (__uint128_t)a * (__uint128_t)b;
-  *upper = (uint64_t)(product >> 64);
-  return (uint64_t)(product & 0xFFFFFFFFFFFFFFFFULL);
+  __uint128_t product =
+      static_cast<__uint128_t>(a) * static_cast<__uint128_t>(b);
+  *upper = static_cast<uint64_t>(product >> 64);
+  return static_cast<uint64_t>(product & 0xFFFFFFFFFFFFFFFFULL);
 #elif HWY_COMPILER_MSVC && HWY_ARCH_X86_64
   return _umul128(a, b, upper);
 #else
@@ -2976,9 +2946,9 @@ HWY_API uint64_t Mul128(uint64_t a, uint64_t b, uint64_t* HWY_RESTRICT upper) {
 
 HWY_API int64_t Mul128(int64_t a, int64_t b, int64_t* HWY_RESTRICT upper) {
 #if defined(__SIZEOF_INT128__)
-  __int128_t product = (__int128_t)a * (__int128_t)b;
-  *upper = (int64_t)(product >> 64);
-  return (int64_t)(product & 0xFFFFFFFFFFFFFFFFULL);
+  __int128_t product = static_cast<__int128_t>(a) * static_cast<__int128_t>(b);
+  *upper = static_cast<int64_t>(product >> 64);
+  return static_cast<int64_t>(product & 0xFFFFFFFFFFFFFFFFULL);
 #elif HWY_COMPILER_MSVC && HWY_ARCH_X86_64
   return _mul128(a, b, upper);
 #else