@img/sharp-libvips-dev-wasm32 1.3.0-rc.1 → 1.3.0-rc.10

package/include/hwy/detect_compiler_arch.h

@@ -20,9 +20,14 @@
 // inclusion by foreach_target.h.
 
 // Add to #if conditions to prevent IDE from graying out code.
+// Note for clangd users: There is no predefined macro in clangd, so you must
+// manually add these two lines (without the preceding '// ') to your project's
+// `.clangd` file:
+// CompileFlags:
+//   Add: [-D__CLANGD__]
 #if (defined __CDT_PARSER__) || (defined __INTELLISENSE__) || \
     (defined Q_CREATOR_RUN) || (defined __CLANGD__) ||        \
-    (defined GROK_ELLIPSIS_BUILD)
+    (defined GROK_ELLIPSIS_BUILD) || (defined __JETBRAINS_IDE__)
 #define HWY_IDE 1
 #else
 #define HWY_IDE 0
@@ -65,15 +70,23 @@
 #define HWY_COMPILER_GCC 0
 #endif
 
-// Clang or clang-cl, not GCC.
-#ifdef __clang__
+#ifndef HWY_COMPILER_CLANG  // Allow user override.
+#ifdef __clang__            // Clang or clang-cl, not GCC.
 // In case of Apple LLVM (whose version number is unrelated to that of LLVM) or
 // an invalid version number, deduce it from the presence of warnings.
 // Originally based on
 // 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_warning("-Woverriding-option")
+#if __has_builtin(__builtin_elementwise_fshl)
+#define HWY_COMPILER_CLANG 2201
+#elif __has_builtin(__builtin_structured_binding_size)
+#define HWY_COMPILER_CLANG 2101
+#elif __has_builtin(__builtin_common_type)
+#define HWY_COMPILER_CLANG 2001
+#elif __has_warning("-Wreturn-mismatch")
+#define HWY_COMPILER_CLANG 1901
+#elif __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.
@@ -108,7 +121,6 @@
 #else  // Anything older than 7.0 is not recommended for Highway.
 #define HWY_COMPILER_CLANG 600
 #endif  // __has_warning chain
-#define HWY_COMPILER3_CLANG (HWY_COMPILER_CLANG * 100)
 #else  // use normal version
 #define HWY_COMPILER_CLANG (__clang_major__ * 100 + __clang_minor__)
 #define HWY_COMPILER3_CLANG \
@@ -117,6 +129,12 @@
 #else  // Not clang
 #define HWY_COMPILER_CLANG 0
 #define HWY_COMPILER3_CLANG 0
+#endif  // __clang__
+#endif  // HWY_COMPILER_CLANG
+
+// User-defined or deduced HWY_COMPILER_CLANG: derive HWY_COMPILER3_CLANG.
+#ifndef HWY_COMPILER3_CLANG
+#define HWY_COMPILER3_CLANG (HWY_COMPILER_CLANG * 100)
 #endif
 
 #if HWY_COMPILER_GCC && !HWY_COMPILER_CLANG && !HWY_COMPILER_ICC && \
@@ -174,19 +192,30 @@
 #define HWY_CXX_LANG __cplusplus
 #endif
 
-#if defined(__cpp_constexpr) && __cpp_constexpr >= 201603L
-#define HWY_CXX17_CONSTEXPR constexpr
-#else
-#define HWY_CXX17_CONSTEXPR
-#endif
-
+// Use instead of constexpr to avoid compiler errors for older compilers. This
+// macro is for when a constexpr function involves multiple statements and
+// loops, which is allowed in C++14 but not before. If the compiler does not
+// support C++14 constexpr, this evaluates to nothing.
 #if defined(__cpp_constexpr) && __cpp_constexpr >= 201304L
 #define HWY_CXX14_CONSTEXPR constexpr
 #else
 #define HWY_CXX14_CONSTEXPR
 #endif
 
-#if HWY_CXX_LANG >= 201703L
+// Same as above, but for C++17 constexpr, which adds support for lambdas, the
+// standard library, and capturing *this.
+// Note that C++17 constexpr still disallows allocating and virtual functions,
+// which are allowed in C++20, but we do not have a use case yet.
+#if defined(__cpp_constexpr) && __cpp_constexpr >= 201603L
+#define HWY_CXX17_CONSTEXPR constexpr
+#else
+#define HWY_CXX17_CONSTEXPR
+#endif
+
+// Use instead of `if constexpr` to avoid compiler errors for older compilers.
+// When compilers lack C++17 support, this evaluates to a normal if statement.
+#if HWY_CXX_LANG >= 201703L || \
+    (defined(__cpp_if_constexpr) && __cpp_if_constexpr >= 201606L)
 #define HWY_IF_CONSTEXPR if constexpr
 #else
 #define HWY_IF_CONSTEXPR if
@@ -204,6 +233,58 @@
 #endif
 #endif
 
+//------------------------------------------------------------------------------
+// Sanitizers
+
+#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
+
 //------------------------------------------------------------------------------
 // Architecture
 
@@ -334,13 +415,36 @@
 #define HWY_ARCH_LOONGARCH 0
 #endif
 
+#if defined(__hexagon__) || defined(__HEXAGON_ARCH__)
+#define HWY_ARCH_HEXAGON 1
+#else
+#define HWY_ARCH_HEXAGON 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_RISCV + HWY_ARCH_S390X + HWY_ARCH_LOONGARCH) > 1
+#if (HWY_ARCH_X86 + HWY_ARCH_PPC + HWY_ARCH_ARM + HWY_ARCH_ARM_OLD +        \
+     HWY_ARCH_WASM + HWY_ARCH_RISCV + HWY_ARCH_S390X + HWY_ARCH_LOONGARCH + \
+     HWY_ARCH_HEXAGON) > 1
 #error "Must not detect more than one architecture"
 #endif
 
+#if HWY_ARCH_RISCV
+#define HWY_ARCH_MAX_BYTES 65536
+#elif HWY_ARCH_ARM_A64
+#define HWY_ARCH_MAX_BYTES 256
+#elif HWY_ARCH_HEXAGON
+#define HWY_ARCH_MAX_BYTES 128
+#elif HWY_ARCH_X86
+#define HWY_ARCH_MAX_BYTES 64
+#elif HWY_ARCH_WASM || HWY_ARCH_LOONGARCH
+#define HWY_ARCH_MAX_BYTES 32
+#elif HWY_ARCH_PPC || HWY_ARCH_S390X || HWY_ARCH_ARM_V7 || HWY_ARCH_ARM_OLD
+#define HWY_ARCH_MAX_BYTES 16
+#else
+#error "Missing case for HWY_ARCH_*"
+#endif
+
 //------------------------------------------------------------------------------
 // Operating system
 
@@ -404,4 +508,20 @@
 #error "Must only detect one byte order"
 #endif
 
+//------------------------------------------------------------------------------
+// Features checked in set_macros-inl.h
+
+// 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
+
 #endif  // HIGHWAY_HWY_DETECT_COMPILER_ARCH_H_

package/include/hwy/detect_targets.h

@@ -1,4 +1,5 @@
 // Copyright 2021 Google LLC
+// Copyright 2025 Arm Limited and/or its affiliates <open-source-office@arm.com>
 // SPDX-License-Identifier: Apache-2.0
 //
 // Licensed under the Apache License, Version 2.0 (the "License");
@@ -92,7 +93,7 @@
 #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)
+#define HWY_NEON_BF16 (1LL << 26)  // fp16/dot/bf16 (e.g. Neoverse V2/N2)
 // Bit 27 reserved for NEON
 #define HWY_NEON (1LL << 28)  // Implies support for AES
 #define HWY_NEON_WITHOUT_AES (1LL << 29)
@@ -194,10 +195,21 @@
 #endif
 #endif  // HWY_BROKEN_MSVC
 
+#ifndef HWY_BROKEN_AVX10_2  // allow override
+// AVX10_2 requires clang >= 20.1 (postpone to 23 due to "avx10.2-512" remnant,
+// only removed in https://github.com/llvm/llvm-project/pull/157034) or
+// gcc >= 15.2 with binutils 2.44.
+#if (HWY_COMPILER_CLANG < 2300) && (HWY_COMPILER_GCC_ACTUAL < 1502)
+#define HWY_BROKEN_AVX10_2 HWY_AVX10_2
+#else
+#define HWY_BROKEN_AVX10_2 0
+#endif
+#endif  // HWY_BROKEN_AVX10_2
+
 #ifndef HWY_BROKEN_AVX3_DL_ZEN4  // allow override
-// AVX3_DL and AVX3_ZEN4 require clang >= 7 (ensured above), gcc >= 8.1 or ICC
+// AVX3_DL and AVX3_ZEN4 require clang >= 7 (ensured above), gcc >= 10.1 or ICC
 // 2021.
-#if (HWY_COMPILER_GCC_ACTUAL && HWY_COMPILER_GCC_ACTUAL < 801) || \
+#if (HWY_COMPILER_GCC_ACTUAL && HWY_COMPILER_GCC_ACTUAL < 1001) || \
     (HWY_COMPILER_ICC && HWY_COMPILER_ICC < 2021)
 #define HWY_BROKEN_AVX3_DL_ZEN4 (HWY_AVX3_DL | HWY_AVX3_ZEN4)
 #else
@@ -245,9 +257,10 @@
 #endif  // HWY_BROKEN_ARM7_WITHOUT_VFP4
 
 #ifndef HWY_BROKEN_NEON_BF16  // allow override
-// HWY_NEON_BF16 requires recent compilers.
+// Broken on older compilers:
 #if (HWY_COMPILER_CLANG != 0 && HWY_COMPILER_CLANG < 1700) || \
-    (HWY_COMPILER_GCC_ACTUAL != 0 && HWY_COMPILER_GCC_ACTUAL < 1302)
+    (HWY_COMPILER_GCC_ACTUAL != 0 && HWY_COMPILER_GCC_ACTUAL < 1302) || \
+    (defined(__apple_build_version__) && __apple_build_version__ <= 17000000)
 #define HWY_BROKEN_NEON_BF16 (HWY_NEON_BF16)
 #else
 #define HWY_BROKEN_NEON_BF16 0
@@ -257,11 +270,11 @@
 // SVE[2] require recent clang or gcc versions.
 
 #ifndef HWY_BROKEN_SVE  // allow override
-// GCC 10+. Clang 19 still has many test failures for SVE. No Apple CPU (at
-// least up to and including M4 and A18) has SVE.
-#if (HWY_COMPILER_CLANG && HWY_COMPILER_CLANG < 2000) ||           \
+// Clang 22+, GCC 10+, except MSAN does not yet support SVE.
+// No Apple CPU (at least up to and including M4 and A18) has SVE.
+#if (HWY_COMPILER_CLANG && HWY_COMPILER_CLANG < 2200) ||           \
     (HWY_COMPILER_GCC_ACTUAL && HWY_COMPILER_GCC_ACTUAL < 1000) || \
-    HWY_OS_APPLE
+    HWY_OS_APPLE || HWY_IS_MSAN
 #define HWY_BROKEN_SVE (HWY_SVE | HWY_SVE_256)
 #else
 #define HWY_BROKEN_SVE 0
@@ -269,16 +282,28 @@
 #endif  // HWY_BROKEN_SVE
 
 #ifndef HWY_BROKEN_SVE2  // allow override
-// Clang 19 still has many test failures for SVE2.
-#if (HWY_COMPILER_CLANG && HWY_COMPILER_CLANG < 2000) ||           \
+// Clang 22+, GCC 10+, except MSAN does not yet support SVE2.
+// No Apple CPU (at least up to and including M4 and A18) has SVE2.
+#if (HWY_COMPILER_CLANG && HWY_COMPILER_CLANG < 2200) ||           \
     (HWY_COMPILER_GCC_ACTUAL && HWY_COMPILER_GCC_ACTUAL < 1000) || \
-    HWY_OS_APPLE
-#define HWY_BROKEN_SVE2 (HWY_SVE2 | HWY_SVE2_128)
+    HWY_OS_APPLE || HWY_IS_MSAN
+#define HWY_BROKEN_SVE2 (HWY_SVE2)
 #else
 #define HWY_BROKEN_SVE2 0
 #endif
 #endif  // HWY_BROKEN_SVE2
 
+#ifndef HWY_BROKEN_SVE2_128  // allow override
+// GCC 10+. Clang 21 works for SVE2_128, but not for SVE2 nor MSAN.
+#if (HWY_COMPILER_CLANG && HWY_COMPILER_CLANG < 2100) ||           \
+    (HWY_COMPILER_GCC_ACTUAL && HWY_COMPILER_GCC_ACTUAL < 1000) || \
+    HWY_OS_APPLE || HWY_IS_MSAN
+#define HWY_BROKEN_SVE2_128 (HWY_SVE2_128)
+#else
+#define HWY_BROKEN_SVE2_128 0
+#endif
+#endif  // HWY_BROKEN_SVE2_128
+
 #ifndef HWY_BROKEN_PPC10  // allow override
 #if (HWY_COMPILER_GCC_ACTUAL && HWY_COMPILER_GCC_ACTUAL < 1100)
 // GCC 10 supports the -mcpu=power10 option but does not support the PPC10
@@ -332,9 +357,12 @@
 #ifndef HWY_BROKEN_LOONGARCH  // allow override
 // Using __loongarch_sx and __loongarch_asx macros to
 // check whether LSX/LASX targets are available.
-#if !defined(__loongarch_sx)
+// GCC does not work yet, see https://gcc.gnu.org/PR121875.
+#if !defined(__loongarch_sx) && \
+    !(HWY_COMPILER_CLANG && HWY_COMPILER_CLANG >= 1800)
 #define HWY_BROKEN_LOONGARCH (HWY_LSX | HWY_LASX)
-#elif !defined(__loongarch_asx)
+#elif !defined(__loongarch_asx) && \
+      !(HWY_COMPILER_CLANG && HWY_COMPILER_CLANG >= 1800)
 #define HWY_BROKEN_LOONGARCH (HWY_LASX)
 #else
 #define HWY_BROKEN_LOONGARCH 0
@@ -360,13 +388,13 @@
 // Allow the user to override this without any guarantee of success.
 #ifndef HWY_BROKEN_TARGETS
 
-#define HWY_BROKEN_TARGETS                                     \
-  (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_NEON_BF16 | HWY_BROKEN_SVE | HWY_BROKEN_SVE2 |   \
-   HWY_BROKEN_PPC10 | HWY_BROKEN_PPC_32BIT | HWY_BROKEN_RVV |  \
-   HWY_BROKEN_LOONGARCH | HWY_BROKEN_Z14)
+#define HWY_BROKEN_TARGETS                                                 \
+  (HWY_BROKEN_CLANG6 | HWY_BROKEN_32BIT | HWY_BROKEN_MSVC |                \
+   HWY_BROKEN_AVX10_2 | HWY_BROKEN_AVX3_DL_ZEN4 | HWY_BROKEN_AVX3_SPR |    \
+   HWY_BROKEN_ARM7_BIG_ENDIAN | HWY_BROKEN_ARM7_WITHOUT_VFP4 |             \
+   HWY_BROKEN_NEON_BF16 | HWY_BROKEN_SVE | HWY_BROKEN_SVE2 |               \
+   HWY_BROKEN_SVE2_128 | HWY_BROKEN_PPC10 | HWY_BROKEN_PPC_32BIT |         \
+   HWY_BROKEN_RVV | HWY_BROKEN_LOONGARCH | HWY_BROKEN_Z14)
 
 #endif  // HWY_BROKEN_TARGETS
 
@@ -489,7 +517,8 @@
 #if defined(__ARM_FEATURE_AES) &&                    \
     defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && \
     defined(__ARM_FEATURE_DOTPROD) &&                \
-    defined(__ARM_FEATURE_BF16_VECTOR_ARITHMETIC)
+    defined(__ARM_FEATURE_BF16_VECTOR_ARITHMETIC) && \
+    defined(__ARM_FEATURE_MATMUL_INT8)
 #define HWY_BASELINE_NEON HWY_ALL_NEON
 #elif defined(__ARM_FEATURE_AES)
 #define HWY_BASELINE_NEON (HWY_NEON_WITHOUT_AES | HWY_NEON)
@@ -642,8 +671,7 @@
 #define HWY_BASELINE_AVX3_SPR 0
 #endif
 
-#if HWY_BASELINE_AVX3_SPR != 0 && defined(__AVX10_2__) && \
-    (HWY_COMPILER_GCC_ACTUAL >= 1500 || HWY_COMPILER_CLANG >= 2001)
+#if HWY_BASELINE_AVX3_SPR != 0 && defined(__AVX10_2__)
 #define HWY_BASELINE_AVX10_2 HWY_AVX10_2
 #else
 #define HWY_BASELINE_AVX10_2 0
@@ -669,7 +697,20 @@
 #define HWY_BASELINE_LOONGARCH 0
 #endif
 
-// Allow the user to override this without any guarantee of success.
+// Workaround for libaom, which unconditionally defines HWY_BASELINE_TARGETS
+// even when that would be disabled/broken. If so, at least use AVX2.
+#if defined(HWY_BASELINE_TARGETS)
+#if HWY_BASELINE_TARGETS == HWY_AVX3_DL && \
+    ((HWY_BROKEN_TARGETS | HWY_DISABLED_TARGETS) & HWY_AVX3_DL)
+#undef HWY_BASELINE_TARGETS
+#define HWY_BASELINE_TARGETS HWY_AVX2
+#endif
+#endif  // HWY_BASELINE_TARGETS
+
+// Allow the user to override this without any guarantee of success. If the
+// compiler invocation considers that target to be broken/disabled, then
+// `HWY_ENABLED_BASELINE` will be 0 and users will have to check for that and
+// skip their code.
 #ifndef HWY_BASELINE_TARGETS
 #define HWY_BASELINE_TARGETS                                               \
   (HWY_BASELINE_SCALAR | HWY_BASELINE_WASM | HWY_BASELINE_PPC8 |           \
@@ -686,7 +727,11 @@
 
 #define HWY_ENABLED_BASELINE HWY_ENABLED(HWY_BASELINE_TARGETS)
 #if HWY_ENABLED_BASELINE == 0
-#error "At least one baseline target must be defined and enabled"
+#pragma message                                                            \
+    "All baseline targets are disabled or considered broken."              \
+    "This is typically due to very restrictive HWY_BASELINE_TARGETS, or "  \
+    "too expansive HWY_BROKEN_TARGETS or HWY_DISABLED_TAREGTS. User code " \
+    "must also check for this and skip any usage of SIMD."
 #endif
 
 // Best baseline, used for static dispatch. This is the least-significant 1-bit
@@ -744,12 +789,10 @@
 #endif  // HWY_HAVE_AUXV
 
 #ifndef HWY_HAVE_RUNTIME_DISPATCH_RVV  // allow override
-// The riscv_vector.h in Clang 16-18 requires compiler flags, and 19 still has
-// some missing intrinsics, see
-// https://github.com/llvm/llvm-project/issues/56592. GCC 13.3 also has an
-// #error check, whereas 14.1 fails with "argument type 'vuint16m8_t' requires
-// the V ISA extension": https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115325.
-#if HWY_ARCH_RISCV && HWY_COMPILER_CLANG >= 1900 && 0
+// Clang 19+ supports target attributes for RVV intrinsics (resolved in
+// https://github.com/llvm/llvm-project/issues/56592 and
+// https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115325).
+#if HWY_ARCH_RISCV && HWY_COMPILER_CLANG >= 1900
 #define HWY_HAVE_RUNTIME_DISPATCH_RVV 1
 #else
 #define HWY_HAVE_RUNTIME_DISPATCH_RVV 0
@@ -766,8 +809,8 @@
 #endif  // HWY_HAVE_RUNTIME_DISPATCH_APPLE
 
 #ifndef HWY_HAVE_RUNTIME_DISPATCH_LOONGARCH  // allow override
-#if HWY_ARCH_LOONGARCH && HWY_HAVE_AUXV && (defined(__loongarch_sx) || \
-    defined(__loongarch_asx))
+#if HWY_ARCH_LOONGARCH && HWY_HAVE_AUXV && !defined(__loongarch_asx) && \
+    HWY_COMPILER_CLANG && HWY_COMPILER_CLANG >= 1800
 #define HWY_HAVE_RUNTIME_DISPATCH_LOONGARCH 1
 #else
 #define HWY_HAVE_RUNTIME_DISPATCH_LOONGARCH 0
@@ -786,6 +829,9 @@
 // Allow opting out, and without a guarantee of success, opting-in.
 #ifndef HWY_HAVE_RUNTIME_DISPATCH
 // Clang, GCC and MSVC allow OS-independent runtime dispatch on x86.
+// Wasm does not, because browsers reject a binary containing any SIMD
+// instructions when the browser does not support them. Typical practice there
+// is to build two binaries, one with the -msimd128 flag.
 #if HWY_ARCH_X86 || HWY_HAVE_RUNTIME_DISPATCH_RVV ||                          \
     HWY_HAVE_RUNTIME_DISPATCH_APPLE || HWY_HAVE_RUNTIME_DISPATCH_LOONGARCH || \
     HWY_HAVE_RUNTIME_DISPATCH_LINUX
@@ -930,7 +976,7 @@
 // HWY_ONCE and the multiple-inclusion mechanism rely on HWY_STATIC_TARGET being
 // one of the dynamic targets. This also implies HWY_TARGETS != 0 and
 // (HWY_TARGETS & HWY_ENABLED_BASELINE) != 0.
-#if (HWY_TARGETS & HWY_STATIC_TARGET) == 0
+#if (HWY_TARGETS & HWY_STATIC_TARGET) == 0 && HWY_ENABLED_BASELINE != 0
 #error "Logic error: best baseline should be included in dynamic targets"
 #endif
 

package/include/hwy/highway.h

@@ -677,6 +677,19 @@ struct AddExport {
 #define HWY_HIGHWAY_PER_TARGET
 #endif
 
+// No SIMD target enabled, skip header inclusion.
+#if HWY_ENABLED_BASELINE == 0
+
+// We would expect that HWY_TARGET and HWY_STATIC_TARGET are now both 0.
+#if HWY_TARGET != 0
+#error "Why is HWY_TARGET not 0 when HWY_ENABLED_BASELINE == 0?"
+#endif
+#if HWY_STATIC_TARGET != 0
+#error "Why is HWY_STATIC_TARGET not 0 when HWY_ENABLED_BASELINE == 0?"
+#endif
+
+#else
+
 // These define ops inside namespace hwy::HWY_NAMESPACE.
 #if HWY_TARGET == HWY_SSE2 || HWY_TARGET == HWY_SSSE3 || HWY_TARGET == HWY_SSE4
 #include "hwy/ops/x86_128-inl.h"
@@ -713,4 +726,6 @@ struct AddExport {
 
 #include "hwy/ops/generic_ops-inl.h"
 
+#endif  // HWY_ENABLED_BASELINE
+
 #endif  // HWY_HIGHWAY_PER_TARGET

package/include/hwy/nanobenchmark.h

@@ -50,6 +50,7 @@
 
 #include "hwy/highway_export.h"
 #include "hwy/timer.h"  // IWYU pragma: export
+#include "hwy/base.h"
 
 namespace hwy {
 
@@ -114,6 +115,14 @@ struct Result {
   float variability;
 };
 
+// Returns a Params struct with customized configuration for benchmarks.
+// Specifically limits `max_evals` to prevent timeout in tests.
+static inline Params DefaultBenchmarkParams() {
+  Params p;
+  p.max_evals = HWY_IS_DEBUG_BUILD ? 3 : 4;
+  return p;
+}
+
 // Precisely measures the number of ticks elapsed when calling "func" with the
 // given inputs, shuffled to ensure realistic branch prediction hit rates.
 //
@@ -132,8 +141,8 @@ HWY_DLLEXPORT size_t Measure(Func func, const uint8_t* arg,
 
 // Calls operator() of the given closure (lambda function).
 template <class Closure>
-static FuncOutput CallClosure(const Closure* f, const FuncInput input) {
-  return (*f)(input);
+static FuncOutput CallClosure(const void* f, const FuncInput input) {
+  return (*reinterpret_cast<const Closure*>(f))(input);
 }
 
 // Same as Measure, except "closure" is typically a lambda function of
@@ -143,7 +152,7 @@ static inline size_t MeasureClosure(const Closure& closure,
                                     const FuncInput* inputs,
                                     const size_t num_inputs, Result* results,
                                     const Params& p = Params()) {
-  return Measure(reinterpret_cast<Func>(&CallClosure<Closure>),
+  return Measure(static_cast<Func>(&CallClosure<Closure>),
                  reinterpret_cast<const uint8_t*>(&closure), inputs, num_inputs,
                  results, p);
 }