@img/sharp-libvips-dev 1.0.1 → 1.0.3

package/include/hwy/cache_control.h

@@ -25,11 +25,15 @@
 #define HWY_DISABLE_CACHE_CONTROL
 #endif
 
+#ifndef HWY_DISABLE_CACHE_CONTROL
 // intrin.h is sufficient on MSVC and already included by base.h.
-#if HWY_ARCH_X86 && !defined(HWY_DISABLE_CACHE_CONTROL) && !HWY_COMPILER_MSVC
+#if HWY_ARCH_X86 && !HWY_COMPILER_MSVC
 #include <emmintrin.h>  // SSE2
 #include <xmmintrin.h>  // _mm_prefetch
+#elif HWY_ARCH_ARM_A64
+#include <arm_acle.h>
 #endif
+#endif  // HWY_DISABLE_CACHE_CONTROL
 
 namespace hwy {
 
@@ -76,15 +80,16 @@ HWY_INLINE HWY_ATTR_CACHE void FlushStream() {
 // subsequent actual loads.
 template <typename T>
 HWY_INLINE HWY_ATTR_CACHE void Prefetch(const T* p) {
-#if HWY_ARCH_X86 && !defined(HWY_DISABLE_CACHE_CONTROL)
+  (void)p;
+#ifndef HWY_DISABLE_CACHE_CONTROL
+#if HWY_ARCH_X86
   _mm_prefetch(reinterpret_cast<const char*>(p), _MM_HINT_T0);
 #elif HWY_COMPILER_GCC  // includes clang
   // Hint=0 (NTA) behavior differs, but skipping outer caches is probably not
   // desirable, so use the default 3 (keep in caches).
   __builtin_prefetch(p, /*write=*/0, /*hint=*/3);
-#else
-  (void)p;
 #endif
+#endif  //  HWY_DISABLE_CACHE_CONTROL
 }
 
 // Invalidates and flushes the cache line containing "p", if possible.
@@ -96,11 +101,24 @@ HWY_INLINE HWY_ATTR_CACHE void FlushCacheline(const void* p) {
 #endif
 }
 
-// When called inside a spin-loop, may reduce power consumption.
+// Hints that we are inside a spin loop and potentially reduces power
+// consumption and coherency traffic. For example, x86 avoids multiple
+// outstanding load requests, which reduces the memory order violation penalty
+// when exiting the loop.
 HWY_INLINE HWY_ATTR_CACHE void Pause() {
-#if HWY_ARCH_X86 && !defined(HWY_DISABLE_CACHE_CONTROL)
+#ifndef HWY_DISABLE_CACHE_CONTROL
+#if HWY_ARCH_X86
   _mm_pause();
+#elif HWY_ARCH_ARM_A64 && HWY_COMPILER_CLANG
+  // This is documented in ACLE and the YIELD instruction is also available in
+  // Armv7, but the intrinsic is broken for Armv7 clang, hence A64 only.
+  __yield();
+#elif HWY_ARCH_ARM && HWY_COMPILER_GCC  // includes clang
+  __asm__ volatile("yield" ::: "memory");
+#elif HWY_ARCH_PPC && HWY_COMPILER_GCC  // includes clang
+  __asm__ volatile("or 27,27,27" ::: "memory");
 #endif
+#endif  // HWY_DISABLE_CACHE_CONTROL
 }
 
 }  // namespace hwy

package/include/hwy/detect_compiler_arch.h

@@ -73,7 +73,13 @@
 // https://github.com/simd-everywhere/simde/blob/47d6e603de9d04ee05cdfbc57cf282a02be1bf2a/simde/simde-detect-clang.h#L59.
 // Please send updates below to them as well, thanks!
 #if defined(__apple_build_version__) || __clang_major__ >= 999
-#if __has_attribute(nouwtable)  // no new warnings in 16.0
+#if __has_warning("-Woverriding-option")
+#define HWY_COMPILER_CLANG 1801
+// No new warnings in 17.0, and Apple LLVM 15.3, which should be 1600, already
+// has the unsafe_buffer_usage attribute, so we instead check for new builtins.
+#elif __has_builtin(__builtin_nondeterministic_value)
+#define HWY_COMPILER_CLANG 1700
+#elif __has_attribute(nouwtable)  // no new warnings in 16.0
 #define HWY_COMPILER_CLANG 1600
 #elif __has_warning("-Warray-parameter")
 #define HWY_COMPILER_CLANG 1500
@@ -113,7 +119,8 @@
 #define HWY_COMPILER3_CLANG 0
 #endif
 
-#if HWY_COMPILER_GCC && !HWY_COMPILER_CLANG && !HWY_COMPILER_ICC
+#if HWY_COMPILER_GCC && !HWY_COMPILER_CLANG && !HWY_COMPILER_ICC && \
+    !HWY_COMPILER_ICX
 #define HWY_COMPILER_GCC_ACTUAL HWY_COMPILER_GCC
 #else
 #define HWY_COMPILER_GCC_ACTUAL 0
@@ -121,17 +128,20 @@
 
 // More than one may be nonzero, but we want at least one.
 #if 0 == (HWY_COMPILER_MSVC + HWY_COMPILER_CLANGCL + HWY_COMPILER_ICC + \
-          HWY_COMPILER_GCC + HWY_COMPILER_CLANG)
+          HWY_COMPILER_ICX + HWY_COMPILER_GCC + HWY_COMPILER_CLANG)
 #error "Unsupported compiler"
 #endif
 
-// We should only detect one of these (only clang/clangcl overlap)
-#if 1 <                                                                     \
-    (!!HWY_COMPILER_MSVC + !!HWY_COMPILER_ICC + !!HWY_COMPILER_GCC_ACTUAL + \
-     !!(HWY_COMPILER_CLANGCL | HWY_COMPILER_CLANG))
+// We should only detect one of these (only clang/clangcl/icx overlap)
+#if 1 < (!!HWY_COMPILER_MSVC + (!!HWY_COMPILER_ICC & !HWY_COMPILER_ICX) + \
+         !!HWY_COMPILER_GCC_ACTUAL +                                      \
+         !!(HWY_COMPILER_ICX | HWY_COMPILER_CLANGCL | HWY_COMPILER_CLANG))
 #error "Detected multiple compilers"
 #endif
 
+//------------------------------------------------------------------------------
+// Compiler features and C++ version
+
 #ifdef __has_builtin
 #define HWY_HAS_BUILTIN(name) __has_builtin(name)
 #else
@@ -156,6 +166,32 @@
 #define HWY_HAS_FEATURE(name) 0
 #endif
 
+// NOTE: clang ~17 does not correctly handle wrapping __has_include in a macro.
+
+#if HWY_COMPILER_MSVC && defined(_MSVC_LANG) && _MSVC_LANG > __cplusplus
+#define HWY_CXX_LANG _MSVC_LANG
+#else
+#define HWY_CXX_LANG __cplusplus
+#endif
+
+#if defined(__cpp_constexpr) && __cpp_constexpr >= 201603L
+#define HWY_CXX17_CONSTEXPR constexpr
+#else
+#define HWY_CXX17_CONSTEXPR
+#endif
+
+#if defined(__cpp_constexpr) && __cpp_constexpr >= 201304L
+#define HWY_CXX14_CONSTEXPR constexpr
+#else
+#define HWY_CXX14_CONSTEXPR
+#endif
+
+#if HWY_CXX_LANG >= 201703L
+#define HWY_IF_CONSTEXPR if constexpr
+#else
+#define HWY_IF_CONSTEXPR if
+#endif
+
 //------------------------------------------------------------------------------
 // Architecture
 
@@ -187,6 +223,12 @@
 #define HWY_ARCH_PPC 0
 #endif
 
+#if defined(__powerpc64__) || (HWY_ARCH_PPC && defined(__64BIT__))
+#define HWY_ARCH_PPC_64 1
+#else
+#define HWY_ARCH_PPC_64 0
+#endif
+
 // aarch32 is currently not supported; please raise an issue if you want it.
 #if defined(__ARM_ARCH_ISA_A64) || defined(__aarch64__) || defined(_M_ARM64)
 #define HWY_ARCH_ARM_A64 1
@@ -225,18 +267,52 @@
 #endif
 
 #ifdef __riscv
-#define HWY_ARCH_RVV 1
+#define HWY_ARCH_RISCV 1
+#else
+#define HWY_ARCH_RISCV 0
+#endif
+// DEPRECATED names; please use HWY_ARCH_RISCV instead.
+#define HWY_ARCH_RVV HWY_ARCH_RISCV
+
+#if HWY_ARCH_RISCV && defined(__riscv_xlen)
+
+#if __riscv_xlen == 32
+#define HWY_ARCH_RISCV_32 1
+#else
+#define HWY_ARCH_RISCV_32 0
+#endif
+
+#if __riscv_xlen == 64
+#define HWY_ARCH_RISCV_64 1
 #else
-#define HWY_ARCH_RVV 0
+#define HWY_ARCH_RISCV_64 0
+#endif
+
+#else  // !HWY_ARCH_RISCV || !defined(__riscv_xlen)
+#define HWY_ARCH_RISCV_32 0
+#define HWY_ARCH_RISCV_64 0
+#endif  // HWY_ARCH_RISCV && defined(__riscv_xlen)
+
+#if HWY_ARCH_RISCV_32 && HWY_ARCH_RISCV_64
+#error "Cannot have both RISCV_32 and RISCV_64"
+#endif
+
+#if defined(__s390x__)
+#define HWY_ARCH_S390X 1
+#else
+#define HWY_ARCH_S390X 0
 #endif
 
 // It is an error to detect multiple architectures at the same time, but OK to
 // detect none of the above.
 #if (HWY_ARCH_X86 + HWY_ARCH_PPC + HWY_ARCH_ARM + HWY_ARCH_ARM_OLD + \
-     HWY_ARCH_WASM + HWY_ARCH_RVV) > 1
+     HWY_ARCH_WASM + HWY_ARCH_RISCV + HWY_ARCH_S390X) > 1
 #error "Must not detect more than one architecture"
 #endif
 
+//------------------------------------------------------------------------------
+// Operating system
+
 #if defined(_WIN32) || defined(_WIN64)
 #define HWY_OS_WIN 1
 #else
@@ -249,6 +325,25 @@
 #define HWY_OS_LINUX 0
 #endif
 
+// iOS or Mac
+#if defined(__APPLE__)
+#define HWY_OS_APPLE 1
+#else
+#define HWY_OS_APPLE 0
+#endif
+
+#if defined(__FreeBSD__)
+#define HWY_OS_FREEBSD 1
+#else
+#define HWY_OS_FREEBSD 0
+#endif
+
+// It is an error to detect multiple OSes at the same time, but OK to
+// detect none of the above.
+#if (HWY_OS_WIN + HWY_OS_LINUX + HWY_OS_APPLE + HWY_OS_FREEBSD) > 1
+#error "Must not detect more than one OS"
+#endif
+
 //------------------------------------------------------------------------------
 // Endianness
 

package/include/hwy/detect_targets.h

@@ -62,7 +62,8 @@
 // Bits 0..3 reserved (4 targets)
 #define HWY_AVX3_SPR (1LL << 4)
 // Bit 5 reserved (likely AVX10.2 with 256-bit vectors)
-// Currently HWY_AVX3_DL plus a special case for CompressStore (10x as fast).
+// Currently HWY_AVX3_DL plus AVX512BF16 and a special case for CompressStore
+// (10x as fast).
 // We may later also use VPCONFLICT.
 #define HWY_AVX3_ZEN4 (1LL << 6)  // see HWY_WANT_AVX3_ZEN4 below
 
@@ -84,15 +85,22 @@
 #define HWY_HIGHEST_TARGET_BIT_X86 14
 
 // --------------------------- Arm: 15 targets (+ one fallback)
-// Bits 15..23 reserved (9 targets)
-#define HWY_SVE2_128 (1LL << 24)  // specialized target (e.g. Arm N2)
-#define HWY_SVE_256 (1LL << 25)   // specialized target (e.g. Arm V1)
-#define HWY_SVE2 (1LL << 26)
-#define HWY_SVE (1LL << 27)
+// Bits 15..17 reserved (3 targets)
+#define HWY_SVE2_128 (1LL << 18)  // specialized (e.g. Neoverse V2/N2/N3)
+#define HWY_SVE_256 (1LL << 19)   // specialized (Neoverse V1)
+// Bits 20-22 reserved for later SVE (3 targets)
+#define HWY_SVE2 (1LL << 23)
+#define HWY_SVE (1LL << 24)
+// Bit 25 reserved for NEON
+#define HWY_NEON_BF16 (1LL << 26)  // fp16/dot/bf16 (e.g. Neoverse V2/N2/N3)
+// Bit 27 reserved for NEON
 #define HWY_NEON (1LL << 28)  // Implies support for AES
 #define HWY_NEON_WITHOUT_AES (1LL << 29)
 #define HWY_HIGHEST_TARGET_BIT_ARM 29
 
+#define HWY_ALL_NEON (HWY_NEON_WITHOUT_AES | HWY_NEON | HWY_NEON_BF16)
+#define HWY_ALL_SVE (HWY_SVE | HWY_SVE2 | HWY_SVE_256 | HWY_SVE2_128)
+
 // --------------------------- RISC-V: 9 targets (+ one fallback)
 // Bits 30..36 reserved (7 targets)
 #define HWY_RVV (1LL << 37)
@@ -102,14 +110,17 @@
 // --------------------------- Future expansion: 4 targets
 // Bits 39..42 reserved
 
-// --------------------------- IBM Power: 9 targets (+ one fallback)
+// --------------------------- IBM Power/ZSeries: 9 targets (+ one fallback)
 // Bits 43..46 reserved (4 targets)
 #define HWY_PPC10 (1LL << 47)  // v3.1
 #define HWY_PPC9 (1LL << 48)   // v3.0
 #define HWY_PPC8 (1LL << 49)   // v2.07
-// Bits 50..51 reserved for prior VSX/AltiVec (2 targets)
+#define HWY_Z15 (1LL << 50)    // Z15
+#define HWY_Z14 (1LL << 51)    // Z14
 #define HWY_HIGHEST_TARGET_BIT_PPC 51
 
+#define HWY_ALL_PPC (HWY_PPC8 | HWY_PPC9 | HWY_PPC10)
+
 // --------------------------- WebAssembly: 9 targets (+ one fallback)
 // Bits 52..57 reserved (6 targets)
 #define HWY_WASM_EMU256 (1LL << 58)  // Experimental
@@ -187,7 +198,7 @@
 
 // armv7be has not been tested and is not yet supported.
 #if HWY_ARCH_ARM_V7 && HWY_IS_BIG_ENDIAN
-#define HWY_BROKEN_ARM7_BIG_ENDIAN (HWY_NEON | HWY_NEON_WITHOUT_AES)
+#define HWY_BROKEN_ARM7_BIG_ENDIAN HWY_ALL_NEON
 #else
 #define HWY_BROKEN_ARM7_BIG_ENDIAN 0
 #endif
@@ -198,11 +209,19 @@
 #if HWY_ARCH_ARM_V7 && (__ARM_ARCH_PROFILE == 'A') && \
     !defined(__ARM_VFPV4__) &&                        \
     !((__ARM_NEON_FP & 0x2 /* half-float */) && (__ARM_FEATURE_FMA == 1))
-#define HWY_BROKEN_ARM7_WITHOUT_VFP4 (HWY_NEON | HWY_NEON_WITHOUT_AES)
+#define HWY_BROKEN_ARM7_WITHOUT_VFP4 HWY_ALL_NEON
 #else
 #define HWY_BROKEN_ARM7_WITHOUT_VFP4 0
 #endif
 
+// HWY_NEON_BF16 requires recent compilers.
+#if (HWY_COMPILER_CLANG != 0 && HWY_COMPILER_CLANG < 1700) || \
+    (HWY_COMPILER_GCC_ACTUAL != 0 && HWY_COMPILER_GCC_ACTUAL < 1302)
+#define HWY_BROKEN_NEON_BF16 (HWY_NEON_BF16)
+#else
+#define HWY_BROKEN_NEON_BF16 0
+#endif
+
 // SVE[2] require recent clang or gcc versions.
 #if (HWY_COMPILER_CLANG && HWY_COMPILER_CLANG < 1100) || \
     (HWY_COMPILER_GCC_ACTUAL && HWY_COMPILER_GCC_ACTUAL < 1000)
@@ -246,7 +265,7 @@
   (HWY_BROKEN_CLANG6 | HWY_BROKEN_32BIT | HWY_BROKEN_MSVC |    \
    HWY_BROKEN_AVX3_DL_ZEN4 | HWY_BROKEN_AVX3_SPR |             \
    HWY_BROKEN_ARM7_BIG_ENDIAN | HWY_BROKEN_ARM7_WITHOUT_VFP4 | \
-   HWY_BROKEN_SVE | HWY_BROKEN_PPC10)
+   HWY_BROKEN_NEON_BF16 | HWY_BROKEN_SVE | HWY_BROKEN_PPC10)
 
 #endif  // HWY_BROKEN_TARGETS
 
@@ -316,13 +335,28 @@
 #define HWY_BASELINE_PPC10 0
 #endif
 
+#if HWY_ARCH_S390X && defined(__VEC__) && defined(__ARCH__) && __ARCH__ >= 12
+#define HWY_BASELINE_Z14 HWY_Z14
+#else
+#define HWY_BASELINE_Z14 0
+#endif
+
+#if HWY_BASELINE_Z14 && __ARCH__ >= 13
+#define HWY_BASELINE_Z15 HWY_Z15
+#else
+#define HWY_BASELINE_Z15 0
+#endif
+
 #define HWY_BASELINE_SVE2 0
 #define HWY_BASELINE_SVE 0
 #define HWY_BASELINE_NEON 0
 
 #if HWY_ARCH_ARM
 
-#if defined(__ARM_FEATURE_SVE2)
+// Also check compiler version as done for HWY_ATTAINABLE_SVE2 because the
+// static target (influenced here) must be one of the attainable targets.
+#if defined(__ARM_FEATURE_SVE2) && \
+    (HWY_COMPILER_CLANG >= 1400 || HWY_COMPILER_GCC_ACTUAL >= 1200)
 #undef HWY_BASELINE_SVE2  // was 0, will be re-defined
 // If user specified -msve-vector-bits=128, they assert the vector length is
 // 128 bits and we should use the HWY_SVE2_128 (more efficient for some ops).
@@ -337,7 +371,8 @@
 #endif  // __ARM_FEATURE_SVE_BITS
 #endif  // __ARM_FEATURE_SVE2
 
-#if defined(__ARM_FEATURE_SVE)
+#if defined(__ARM_FEATURE_SVE) && \
+    (HWY_COMPILER_CLANG >= 900 || HWY_COMPILER_GCC_ACTUAL >= 800)
 #undef HWY_BASELINE_SVE  // was 0, will be re-defined
 // See above. If user-specified vector length matches our optimization, use it.
 #if defined(__ARM_FEATURE_SVE_BITS) && __ARM_FEATURE_SVE_BITS == 256
@@ -350,12 +385,17 @@
 // GCC 4.5.4 only defines __ARM_NEON__; 5.4 defines both.
 #if defined(__ARM_NEON__) || defined(__ARM_NEON)
 #undef HWY_BASELINE_NEON
-#if defined(__ARM_FEATURE_AES)
-#define HWY_BASELINE_NEON (HWY_NEON | HWY_NEON_WITHOUT_AES)
+#if defined(__ARM_FEATURE_AES) &&                    \
+    defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && \
+    defined(__ARM_FEATURE_DOTPROD) &&                \
+    defined(__ARM_FEATURE_BF16_VECTOR_ARITHMETIC)
+#define HWY_BASELINE_NEON HWY_ALL_NEON
+#elif defined(__ARM_FEATURE_AES)
+#define HWY_BASELINE_NEON (HWY_NEON_WITHOUT_AES | HWY_NEON)
 #else
 #define HWY_BASELINE_NEON (HWY_NEON_WITHOUT_AES)
-#endif
-#endif
+#endif  // __ARM_FEATURE*
+#endif  // __ARM_NEON
 
 #endif  // HWY_ARCH_ARM
 
@@ -483,14 +523,16 @@
 #define HWY_BASELINE_AVX3_ZEN4 0
 #endif
 
-#if HWY_BASELINE_AVX3_DL != 0 && defined(__AVX512FP16__)
+#if HWY_BASELINE_AVX3_DL != 0 && defined(__AVX512BF16__) && \
+    defined(__AVX512FP16__)
 #define HWY_BASELINE_AVX3_SPR HWY_AVX3_SPR
 #else
 #define HWY_BASELINE_AVX3_SPR 0
 #endif
 
 // RVV requires intrinsics 0.11 or later, see #1156.
-#if HWY_ARCH_RVV && defined(__riscv_v_intrinsic) && __riscv_v_intrinsic >= 11000
+#if HWY_ARCH_RISCV && defined(__riscv_v_intrinsic) && \
+    __riscv_v_intrinsic >= 11000
 #define HWY_BASELINE_RVV HWY_RVV
 #else
 #define HWY_BASELINE_RVV 0
@@ -498,13 +540,14 @@
 
 // Allow the user to override this without any guarantee of success.
 #ifndef HWY_BASELINE_TARGETS
-#define HWY_BASELINE_TARGETS                                           \
-  (HWY_BASELINE_SCALAR | HWY_BASELINE_WASM | HWY_BASELINE_PPC8 |       \
-   HWY_BASELINE_PPC9 | HWY_BASELINE_PPC10 | HWY_BASELINE_SVE2 |        \
-   HWY_BASELINE_SVE | HWY_BASELINE_NEON | HWY_BASELINE_SSE2 |          \
-   HWY_BASELINE_SSSE3 | HWY_BASELINE_SSE4 | HWY_BASELINE_AVX2 |        \
-   HWY_BASELINE_AVX3 | HWY_BASELINE_AVX3_DL | HWY_BASELINE_AVX3_ZEN4 | \
-   HWY_BASELINE_AVX3_SPR | HWY_BASELINE_RVV)
+#define HWY_BASELINE_TARGETS                                               \
+  (HWY_BASELINE_SCALAR | HWY_BASELINE_WASM | HWY_BASELINE_PPC8 |           \
+   HWY_BASELINE_PPC9 | HWY_BASELINE_PPC10 | HWY_BASELINE_Z14 |             \
+   HWY_BASELINE_Z15 | HWY_BASELINE_SVE2 | HWY_BASELINE_SVE |               \
+   HWY_BASELINE_NEON | HWY_BASELINE_SSE2 | HWY_BASELINE_SSSE3 |            \
+   HWY_BASELINE_SSE4 | HWY_BASELINE_AVX2 | HWY_BASELINE_AVX3 |             \
+   HWY_BASELINE_AVX3_DL | HWY_BASELINE_AVX3_ZEN4 | HWY_BASELINE_AVX3_SPR | \
+   HWY_BASELINE_RVV)
 #endif  // HWY_BASELINE_TARGETS
 
 //------------------------------------------------------------------------------
@@ -534,17 +577,43 @@
 #endif
 // Defining one of HWY_COMPILE_ONLY_* will trump HWY_COMPILE_ALL_ATTAINABLE.
 
+#ifndef HWY_HAVE_AUXV  // allow override
+#ifdef TOOLCHAIN_MISS_SYS_AUXV_H
+#define HWY_HAVE_AUXV 0  // CMake failed to find the header
+// glibc 2.16 added auxv, but checking for that requires features.h, and we do
+// not want to include system headers here. Instead check for the header
+// directly, which has been supported at least since GCC 5.4 and Clang 3.
+#elif defined(__has_include)  // note: wrapper macro fails on Clang ~17
+// clang-format off
+#if __has_include(<sys/auxv.h>)
+// clang-format on
+#define HWY_HAVE_AUXV 1  // header present
+#else
+#define HWY_HAVE_AUXV 0  // header not present
+#endif                   // __has_include
+#else                    // compiler lacks __has_include
+#define HWY_HAVE_AUXV 0
+#endif
+#endif  // HWY_HAVE_AUXV
+
+// Allow opting out, and without a guarantee of success, opting-in.
+#ifndef HWY_HAVE_RUNTIME_DISPATCH
 // Clang, GCC and MSVC allow runtime dispatch on x86.
 #if HWY_ARCH_X86
 #define HWY_HAVE_RUNTIME_DISPATCH 1
-// On Arm/PPC, currently only GCC does, and we require Linux to detect CPU
-// capabilities.
-#elif (HWY_ARCH_ARM || HWY_ARCH_PPC) && HWY_COMPILER_GCC_ACTUAL && \
-    HWY_OS_LINUX && !defined(TOOLCHAIN_MISS_SYS_AUXV_H)
+// On Arm, PPC, S390X, and RISC-V: GCC and Clang 17+ do, and we require Linux
+// to detect CPU capabilities.
+#elif (HWY_ARCH_ARM || HWY_ARCH_PPC || HWY_ARCH_S390X || HWY_ARCH_RISCV) &&    \
+    (HWY_COMPILER_GCC_ACTUAL || HWY_COMPILER_CLANG >= 1700) && HWY_OS_LINUX && \
+    HWY_HAVE_AUXV
+#define HWY_HAVE_RUNTIME_DISPATCH 1
+#elif HWY_ARCH_ARM_A64 && HWY_OS_APPLE && \
+    (HWY_COMPILER_GCC_ACTUAL || HWY_COMPILER_CLANG >= 1700)
 #define HWY_HAVE_RUNTIME_DISPATCH 1
 #else
 #define HWY_HAVE_RUNTIME_DISPATCH 0
-#endif
+#endif  // HWY_ARCH_*
+#endif  // HWY_HAVE_RUNTIME_DISPATCH
 
 // AVX3_DL is not widely available yet. To reduce code size and compile time,
 // only include it in the set of attainable targets (for dynamic dispatch) if
@@ -556,22 +625,26 @@
 #endif
 
 #if HWY_ARCH_ARM_A64 && HWY_HAVE_RUNTIME_DISPATCH
-#define HWY_ATTAINABLE_NEON (HWY_NEON | HWY_NEON_WITHOUT_AES)
+#define HWY_ATTAINABLE_NEON HWY_ALL_NEON
 #elif HWY_ARCH_ARM  // static dispatch, or HWY_ARCH_ARM_V7
 #define HWY_ATTAINABLE_NEON (HWY_BASELINE_NEON)
 #else
 #define HWY_ATTAINABLE_NEON 0
 #endif
 
-#if HWY_ARCH_ARM_A64 && (HWY_HAVE_RUNTIME_DISPATCH || \
-                         (HWY_ENABLED_BASELINE & (HWY_SVE | HWY_SVE_256)))
+#if HWY_ARCH_ARM_A64 &&                                              \
+    (HWY_COMPILER_CLANG >= 900 || HWY_COMPILER_GCC_ACTUAL >= 800) && \
+    (HWY_HAVE_RUNTIME_DISPATCH ||                                    \
+     (HWY_ENABLED_BASELINE & (HWY_SVE | HWY_SVE_256)))
 #define HWY_ATTAINABLE_SVE (HWY_SVE | HWY_SVE_256)
 #else
 #define HWY_ATTAINABLE_SVE 0
 #endif
 
-#if HWY_ARCH_ARM_A64 && (HWY_HAVE_RUNTIME_DISPATCH || \
-                         (HWY_ENABLED_BASELINE & (HWY_SVE2 | HWY_SVE2_128)))
+#if HWY_ARCH_ARM_A64 &&                                                \
+    (HWY_COMPILER_CLANG >= 1400 || HWY_COMPILER_GCC_ACTUAL >= 1200) && \
+    (HWY_HAVE_RUNTIME_DISPATCH ||                                      \
+     (HWY_ENABLED_BASELINE & (HWY_SVE2 | HWY_SVE2_128)))
 #define HWY_ATTAINABLE_SVE2 (HWY_SVE2 | HWY_SVE2_128)
 #else
 #define HWY_ATTAINABLE_SVE2 0
@@ -579,18 +652,47 @@
 
 #if HWY_ARCH_PPC && defined(__ALTIVEC__) && \
     (!HWY_COMPILER_CLANG || HWY_BASELINE_PPC8 != 0)
+
+#if (HWY_BASELINE_PPC9 | HWY_BASELINE_PPC10) && \
+    !defined(HWY_SKIP_NON_BEST_BASELINE)
+// On POWER with -m flags, we get compile errors (#1707) for targets older than
+// the baseline specified via -m, so only generate the static target and better.
+// Note that some Linux distros actually do set POWER9 as the baseline.
+// This works by skipping case 3 below, so case 4 is reached.
+#define HWY_SKIP_NON_BEST_BASELINE
+#endif
+
 #define HWY_ATTAINABLE_PPC (HWY_PPC8 | HWY_PPC9 | HWY_PPC10)
+
 #else
 #define HWY_ATTAINABLE_PPC 0
 #endif
 
+#if HWY_ARCH_S390X && HWY_BASELINE_Z14 != 0
+#define HWY_ATTAINABLE_S390X (HWY_Z14 | HWY_Z15)
+#else
+#define HWY_ATTAINABLE_S390X 0
+#endif
+
+#if HWY_ARCH_RISCV && HWY_HAVE_RUNTIME_DISPATCH
+#define HWY_ATTAINABLE_RISCV (HWY_RVV)
+#else
+#define HWY_ATTAINABLE_RISCV 0
+#endif
+
 // Attainable means enabled and the compiler allows intrinsics (even when not
 // allowed to autovectorize). Used in 3 and 4.
 #if HWY_ARCH_X86
+#if HWY_COMPILER_MSVC
+// Fewer targets for faster builds.
+#define HWY_ATTAINABLE_TARGETS \
+  HWY_ENABLED(HWY_BASELINE_SCALAR | HWY_STATIC_TARGET | HWY_AVX2)
+#else  // !HWY_COMPILER_MSVC
 #define HWY_ATTAINABLE_TARGETS                                               \
   HWY_ENABLED(HWY_BASELINE_SCALAR | HWY_SSE2 | HWY_SSSE3 | HWY_SSE4 |        \
               HWY_AVX2 | HWY_AVX3 | HWY_ATTAINABLE_AVX3_DL | HWY_AVX3_ZEN4 | \
               HWY_AVX3_SPR)
+#endif  // !HWY_COMPILER_MSVC
 #elif HWY_ARCH_ARM
 #define HWY_ATTAINABLE_TARGETS                                                 \
   HWY_ENABLED(HWY_BASELINE_SCALAR | HWY_ATTAINABLE_NEON | HWY_ATTAINABLE_SVE | \
@@ -598,6 +700,12 @@
 #elif HWY_ARCH_PPC
 #define HWY_ATTAINABLE_TARGETS \
   HWY_ENABLED(HWY_BASELINE_SCALAR | HWY_ATTAINABLE_PPC)
+#elif HWY_ARCH_S390X
+#define HWY_ATTAINABLE_TARGETS \
+  HWY_ENABLED(HWY_BASELINE_SCALAR | HWY_ATTAINABLE_S390X)
+#elif HWY_ARCH_RVV
+#define HWY_ATTAINABLE_TARGETS \
+  HWY_ENABLED(HWY_BASELINE_SCALAR | HWY_ATTAINABLE_RISCV)
 #else
 #define HWY_ATTAINABLE_TARGETS (HWY_ENABLED_BASELINE)
 #endif  // HWY_ARCH_*
@@ -621,7 +729,8 @@
 #define HWY_TARGETS HWY_STATIC_TARGET
 
 // 3) For tests: include all attainable targets (in particular: scalar)
-#elif defined(HWY_COMPILE_ALL_ATTAINABLE) || defined(HWY_IS_TEST)
+#elif (defined(HWY_COMPILE_ALL_ATTAINABLE) || defined(HWY_IS_TEST)) && \
+    !defined(HWY_SKIP_NON_BEST_BASELINE)
 #define HWY_TARGETS HWY_ATTAINABLE_TARGETS
 
 // 4) Default: attainable WITHOUT non-best baseline. This reduces code size by

package/include/hwy/foreach_target.h

@@ -168,6 +168,17 @@
 #endif
 #endif
 
+#if (HWY_TARGETS & HWY_NEON_BF16) && (HWY_STATIC_TARGET != HWY_NEON_BF16)
+#undef HWY_TARGET
+#define HWY_TARGET HWY_NEON_BF16
+#include HWY_TARGET_INCLUDE
+#ifdef HWY_TARGET_TOGGLE
+#undef HWY_TARGET_TOGGLE
+#else
+#define HWY_TARGET_TOGGLE
+#endif
+#endif
+
 #if (HWY_TARGETS & HWY_SVE) && (HWY_STATIC_TARGET != HWY_SVE)
 #undef HWY_TARGET
 #define HWY_TARGET HWY_SVE
@@ -271,7 +282,31 @@
 #endif
 #endif
 
-// ------------------------------ HWY_ARCH_RVV
+// ------------------------------ HWY_ARCH_S390X
+
+#if (HWY_TARGETS & HWY_Z14) && (HWY_STATIC_TARGET != HWY_Z14)
+#undef HWY_TARGET
+#define HWY_TARGET HWY_Z14
+#include HWY_TARGET_INCLUDE
+#ifdef HWY_TARGET_TOGGLE
+#undef HWY_TARGET_TOGGLE
+#else
+#define HWY_TARGET_TOGGLE
+#endif
+#endif
+
+#if (HWY_TARGETS & HWY_Z15) && (HWY_STATIC_TARGET != HWY_Z15)
+#undef HWY_TARGET
+#define HWY_TARGET HWY_Z15
+#include HWY_TARGET_INCLUDE
+#ifdef HWY_TARGET_TOGGLE
+#undef HWY_TARGET_TOGGLE
+#else
+#define HWY_TARGET_TOGGLE
+#endif
+#endif
+
+// ------------------------------ HWY_ARCH_RISCV
 
 #if (HWY_TARGETS & HWY_RVV) && (HWY_STATIC_TARGET != HWY_RVV)
 #undef HWY_TARGET