gitignore: use exact matches

Update from bookworm to trixie, updating GKIs, and changing to Android 16.

.github/workflows/build-and-test.yml

@@ -11,9 +11,9 @@ jobs:
     runs-on: ubuntu-latest
     strategy:
       matrix:
-        version: [12]
+        version: [14]
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v5
       - name: Setting up gcc version
         run: |
           sudo update-alternatives --install /usr/bin/g++ g++ /usr/bin/g++-${{ matrix.version }} 100
@@ -24,9 +24,11 @@ jobs:
     runs-on: ubuntu-latest
     strategy:
       matrix:
-        version: [14, 15]
+        version: [19, 20]
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v5
+      - name: Install dependencies
+        run: sudo apt-get update && sudo apt-get install -y --no-install-recommends clang-19 clang-20
       - name: Setting up clang version
         run: |
           sudo update-alternatives --install /usr/bin/clang++ clang++ /usr/bin/clang++-${{ matrix.version }} 100
@@ -38,7 +40,7 @@ jobs:
     container:
       image: alpine:latest
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v5
       - name: Install dependencies
         run: apk update && apk add build-base python3
       - name: Build
@@ -46,7 +48,7 @@ jobs:
   build-ubuntu-gcc-aarch64:
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v5
       - name: Install dependencies
         run: sudo apt-get update && sudo apt-get install -y --no-install-recommends gcc-aarch64-linux-gnu g++-aarch64-linux-gnu libgcc-s1-arm64-cross cpp-aarch64-linux-gnu
       - name: Build

.gitignore

@@ -1,2 +1,2 @@
-out/
-out-light/
+/out/
+/out-light/

Android.bp

@@ -5,8 +5,6 @@ common_cflags = [
     "-fPIC",
     "-fvisibility=hidden",
     //"-fno-plt",
-    "-Wall",
-    "-Wextra",
     "-Wcast-align",
     "-Wcast-qual",
     "-Wwrite-strings",
@@ -74,7 +72,7 @@ cc_library {
             cflags: ["-DLABEL_MEMORY"],
         },
         device_has_arm_mte: {
-            cflags: ["-DHAS_ARM_MTE", "-march=armv9-a+memtag"]
+            cflags: ["-DHAS_ARM_MTE", "-march=armv8-a+dotprod+memtag"]
         },
     },
     apex_available: [

LICENSE

@@ -1,4 +1,4 @@
-Copyright © 2018-2024 GrapheneOS
+Copyright © 2018-2025 GrapheneOS
 
 Permission is hereby granted, free of charge, to any person obtaining a copy
 of this software and associated documentation files (the "Software"), to deal

README.md

@@ -65,14 +65,14 @@ used instead as this allocator fundamentally doesn't support that environment.
 
 ## Dependencies
 
-Debian stable (currently Debian 12) determines the most ancient set of
+Debian stable (currently Debian 13) determines the most ancient set of
 supported dependencies:
 
-* glibc 2.36
-* Linux 6.1
-* Clang 14.0.6 or GCC 12.2.0
+* glibc 2.41
+* Linux 6.12
+* Clang 19.1.7 or GCC 14.2.0
 
-For Android, the Linux GKI 5.10, 5.15 and 6.1 branches are supported.
+For Android, the Linux GKI 6.1, 6.6 and 6.12 branches are supported.
 
 However, using more recent releases is highly recommended. Older versions of
 the dependencies may be compatible at the moment but are not tested and will
@@ -83,7 +83,7 @@ there will be custom integration offering better performance in the future
 along with other hardening for the C standard library implementation.
 
 For Android, only the current generation, actively developed maintenance branch of the Android
-Open Source Project will be supported, which currently means `android13-qpr2-release`.
+Open Source Project will be supported, which currently means `android16-release`.
 
 ## Testing
 
@@ -159,8 +159,11 @@ line to the `/etc/ld.so.preload` configuration file:
 The format of this configuration file is a whitespace-separated list, so it's
 good practice to put each library on a separate line.
 
-On Debian systems `libhardened_malloc.so` should be installed into `/usr/lib/`
-to avoid preload failures caused by AppArmor profile restrictions.
+For maximum compatibility `libhardened_malloc.so` can be installed into
+`/usr/lib/` to avoid preload failures caused by AppArmor profiles or systemd
+ExecPaths= restrictions. Check for logs of the following format:
+
+    ERROR: ld.so: object '/usr/local/lib/libhardened_malloc.so' from /etc/ld.so.preload cannot be preloaded (failed to map segment from shared object): ignored.
 
 Using the `LD_PRELOAD` environment variable to load it on a case-by-case basis
 will not work when `AT_SECURE` is set such as with setuid binaries. It's also
@@ -731,7 +734,7 @@ Random tags are set for all slab allocations when allocated, with 4 excluded val
 3. the current (or previous) tag used for the slot to the left
 4. the current (or previous) tag used for the slot to the right
 
-When a slab allocation is freed, the reserved `0` tag is set for the slot. 
+When a slab allocation is freed, the reserved `0` tag is set for the slot.
 Slab allocation slots are cleared before reuse when memory tagging is enabled.
 
 This ensures the following properties:
@@ -740,7 +743,7 @@ This ensures the following properties:
 - Use-after-free are deterministically detected until the freed slot goes through
   both the random and FIFO quarantines, gets allocated again, goes through both
   quarantines again and then finally gets allocated again for a 2nd time.
-- Since the default `0` tag is reserved, untagged pointers can't access slab 
+- Since the default `0` tag is reserved, untagged pointers can't access slab
   allocations and vice versa.
 
 Slab allocations are done in a statically reserved region for each size class

androidtest/memtag/memtag_test.cc

@@ -44,7 +44,7 @@ void *set_pointer_tag(void *ptr, u8 tag) {
     return (void *) (((uintptr_t) tag << 56) | (uintptr_t) untag_pointer(ptr));
 }
 
-// This test checks that slab slot allocation uses tag that is distint from tags of its neighbors
+// This test checks that slab slot allocation uses tag that is distinct from tags of its neighbors
 // and from the tag of the previous allocation that used the same slot
 void tag_distinctness() {
     // tag 0 is reserved
@@ -346,6 +346,6 @@ int main(int argc, char **argv) {
 
     test_fn();
     do_context_switch();
-    
+
     return 0;
 }

chacha.c

@@ -41,7 +41,7 @@ static const unsigned rounds = 8;
     a = PLUS(a, b); d = ROTATE(XOR(d, a), 8); \
     c = PLUS(c, d); b = ROTATE(XOR(b, c), 7);
 
-static const char sigma[16] = "expand 32-byte k";
+static const char sigma[16] NONSTRING = "expand 32-byte k";
 
 void chacha_keysetup(chacha_ctx *x, const u8 *k) {
     x->input[0] = U8TO32_LITTLE(sigma + 0);

h_malloc.c

@@ -94,6 +94,24 @@ static inline bool is_memtag_enabled(void) {
 }
 #endif
 
+static void *memory_map_tagged(size_t size) {
+#ifdef HAS_ARM_MTE
+    if (likely51(is_memtag_enabled())) {
+        return memory_map_mte(size);
+    }
+#endif
+    return memory_map(size);
+}
+
+static bool memory_map_fixed_tagged(void *ptr, size_t size) {
+#ifdef HAS_ARM_MTE
+    if (likely51(is_memtag_enabled())) {
+        return memory_map_fixed_mte(ptr, size);
+    }
+#endif
+    return memory_map_fixed(ptr, size);
+}
+
 #define SLAB_METADATA_COUNT
 
 struct slab_metadata {
@@ -470,7 +488,7 @@ static void write_after_free_check(const char *p, size_t size) {
     }
 
 #ifdef HAS_ARM_MTE
-    if (likely(is_memtag_enabled())) {
+    if (likely51(is_memtag_enabled())) {
         return;
     }
 #endif
@@ -505,7 +523,7 @@ static void set_slab_canary_value(UNUSED struct slab_metadata *metadata, UNUSED
 static void set_canary(UNUSED const struct slab_metadata *metadata, UNUSED void *p, UNUSED size_t size) {
 #if SLAB_CANARY
 #ifdef HAS_ARM_MTE
-    if (likely(is_memtag_enabled())) {
+    if (likely51(is_memtag_enabled())) {
         return;
     }
 #endif
@@ -517,7 +535,7 @@ static void set_canary(UNUSED const struct slab_metadata *metadata, UNUSED void
 static void check_canary(UNUSED const struct slab_metadata *metadata, UNUSED const void *p, UNUSED size_t size) {
 #if SLAB_CANARY
 #ifdef HAS_ARM_MTE
-    if (likely(is_memtag_enabled())) {
+    if (likely51(is_memtag_enabled())) {
         return;
     }
 #endif
@@ -624,7 +642,7 @@ static inline void *allocate_small(unsigned arena, size_t requested_size) {
                 write_after_free_check(p, size - canary_size);
                 set_canary(metadata, p, size);
 #ifdef HAS_ARM_MTE
-                if (likely(is_memtag_enabled())) {
+                if (likely51(is_memtag_enabled())) {
                     p = tag_and_clear_slab_slot(metadata, p, slot, size);
                 }
 #endif
@@ -661,7 +679,7 @@ static inline void *allocate_small(unsigned arena, size_t requested_size) {
             if (requested_size) {
                 set_canary(metadata, p, size);
 #ifdef HAS_ARM_MTE
-                if (likely(is_memtag_enabled())) {
+                if (likely51(is_memtag_enabled())) {
                     p = tag_and_clear_slab_slot(metadata, p, slot, size);
                 }
 #endif
@@ -688,7 +706,7 @@ static inline void *allocate_small(unsigned arena, size_t requested_size) {
         if (requested_size) {
             set_canary(metadata, p, size);
 #ifdef HAS_ARM_MTE
-            if (likely(is_memtag_enabled())) {
+            if (likely51(is_memtag_enabled())) {
                 p = tag_and_clear_slab_slot(metadata, p, slot, size);
             }
 #endif
@@ -717,7 +735,7 @@ static inline void *allocate_small(unsigned arena, size_t requested_size) {
         write_after_free_check(p, size - canary_size);
         set_canary(metadata, p, size);
 #ifdef HAS_ARM_MTE
-        if (likely(is_memtag_enabled())) {
+        if (likely51(is_memtag_enabled())) {
             p = tag_and_clear_slab_slot(metadata, p, slot, size);
         }
 #endif
@@ -805,7 +823,7 @@ static inline void deallocate_small(void *p, const size_t *expected_size) {
 
         bool skip_zero = false;
 #ifdef HAS_ARM_MTE
-        if (likely(is_memtag_enabled())) {
+        if (likely51(is_memtag_enabled())) {
             arm_mte_tag_and_clear_mem(set_pointer_tag(p, RESERVED_TAG), size);
             // metadata->arm_mte_tags is intentionally not updated, see tag_and_clear_slab_slot()
             skip_zero = true;
@@ -890,7 +908,7 @@ static inline void deallocate_small(void *p, const size_t *expected_size) {
 
         if (c->empty_slabs_total + slab_size > max_empty_slabs_total) {
             int saved_errno = errno;
-            if (!memory_map_fixed(slab, slab_size)) {
+            if (!memory_map_fixed_tagged(slab, slab_size)) {
                 label_slab(slab, slab_size, class);
                 stats_slab_deallocate(c, slab_size);
                 enqueue_free_slab(c, metadata);
@@ -1242,15 +1260,7 @@ COLD static void init_slow_path(void) {
     if (unlikely(memory_protect_rw_metadata(ra->regions, ra->total * sizeof(struct region_metadata)))) {
         fatal_error("failed to unprotect memory for regions table");
     }
-#ifdef HAS_ARM_MTE
-    if (likely(is_memtag_enabled())) {
-        ro.slab_region_start = memory_map_mte(slab_region_size);
-    } else {
-        ro.slab_region_start = memory_map(slab_region_size);
-    }
-#else
-    ro.slab_region_start = memory_map(slab_region_size);
-#endif
+    ro.slab_region_start = memory_map_tagged(slab_region_size);
     if (unlikely(ro.slab_region_start == NULL)) {
         fatal_error("failed to allocate slab region");
     }
@@ -1895,7 +1905,7 @@ EXPORT int h_malloc_trim(UNUSED size_t pad) {
             struct slab_metadata *iterator = c->empty_slabs;
             while (iterator) {
                 void *slab = get_slab(c, slab_size, iterator);
-                if (memory_map_fixed(slab, slab_size)) {
+                if (memory_map_fixed_tagged(slab, slab_size)) {
                     break;
                 }
                 label_slab(slab, slab_size, class);

memory.c

@@ -17,8 +17,8 @@
 #include "memory.h"
 #include "util.h"
 
-void *memory_map(size_t size) {
-    void *p = mmap(NULL, size, PROT_NONE, MAP_ANONYMOUS|MAP_PRIVATE, -1, 0);
+static void *memory_map_prot(size_t size, int prot) {
+    void *p = mmap(NULL, size, prot, MAP_ANONYMOUS|MAP_PRIVATE, -1, 0);
     if (unlikely(p == MAP_FAILED)) {
         if (errno != ENOMEM) {
             fatal_error("non-ENOMEM mmap failure");
@@ -28,22 +28,19 @@ void *memory_map(size_t size) {
     return p;
 }
 
+void *memory_map(size_t size) {
+    return memory_map_prot(size, PROT_NONE);
+}
+
 #ifdef HAS_ARM_MTE
 // Note that PROT_MTE can't be cleared via mprotect
 void *memory_map_mte(size_t size) {
-    void *p = mmap(NULL, size, PROT_MTE, MAP_ANONYMOUS|MAP_PRIVATE, -1, 0);
-    if (unlikely(p == MAP_FAILED)) {
-        if (errno != ENOMEM) {
-            fatal_error("non-ENOMEM MTE mmap failure");
-        }
-        return NULL;
-    }
-    return p;
+    return memory_map_prot(size, PROT_MTE);
 }
 #endif
 
-bool memory_map_fixed(void *ptr, size_t size) {
-    void *p = mmap(ptr, size, PROT_NONE, MAP_ANONYMOUS|MAP_PRIVATE|MAP_FIXED, -1, 0);
+static bool memory_map_fixed_prot(void *ptr, size_t size, int prot) {
+    void *p = mmap(ptr, size, prot, MAP_ANONYMOUS|MAP_PRIVATE|MAP_FIXED, -1, 0);
     bool ret = p == MAP_FAILED;
     if (unlikely(ret) && errno != ENOMEM) {
         fatal_error("non-ENOMEM MAP_FIXED mmap failure");
@@ -51,6 +48,17 @@ bool memory_map_fixed(void *ptr, size_t size) {
     return ret;
 }
 
+bool memory_map_fixed(void *ptr, size_t size) {
+    return memory_map_fixed_prot(ptr, size, PROT_NONE);
+}
+
+#ifdef HAS_ARM_MTE
+// Note that PROT_MTE can't be cleared via mprotect
+bool memory_map_fixed_mte(void *ptr, size_t size) {
+    return memory_map_fixed_prot(ptr, size, PROT_MTE);
+}
+#endif
+
 bool memory_unmap(void *ptr, size_t size) {
     bool ret = munmap(ptr, size);
     if (unlikely(ret) && errno != ENOMEM) {

memory.h

@@ -15,6 +15,9 @@ void *memory_map(size_t size);
 void *memory_map_mte(size_t size);
 #endif
 bool memory_map_fixed(void *ptr, size_t size);
+#ifdef HAS_ARM_MTE
+bool memory_map_fixed_mte(void *ptr, size_t size);
+#endif
 bool memory_unmap(void *ptr, size_t size);
 bool memory_protect_ro(void *ptr, size_t size);
 bool memory_protect_rw(void *ptr, size_t size);

test/malloc_info.c

@@ -1,5 +1,6 @@
 #include <pthread.h>
 #include <stdio.h>
+#include <stdlib.h>
 
 #if defined(__GLIBC__) || defined(__ANDROID__)
 #include <malloc.h>

test/test_smc.py

@@ -98,7 +98,7 @@ class TestSimpleMemoryCorruption(unittest.TestCase):
         self.assertEqual(stderr.decode("utf-8"),
                          "fatal allocator error: invalid free\n")
 
-    def test_invalid_malloc_usable_size_small_quarantene(self):
+    def test_invalid_malloc_usable_size_small_quarantine(self):
         _stdout, stderr, returncode = self.run_test(
             "invalid_malloc_usable_size_small_quarantine")
         self.assertEqual(returncode, -6)

third_party/libdivide.h

@@ -1,8 +1,8 @@
 // libdivide.h - Optimized integer division
 // https://libdivide.com
 //
-// Copyright (C) 2010 - 2021 ridiculous_fish, <libdivide@ridiculousfish.com>
-// Copyright (C) 2016 - 2021 Kim Walisch, <kim.walisch@gmail.com>
+// Copyright (C) 2010 - 2022 ridiculous_fish, <libdivide@ridiculousfish.com>
+// Copyright (C) 2016 - 2022 Kim Walisch, <kim.walisch@gmail.com>
 //
 // libdivide is dual-licensed under the Boost or zlib licenses.
 // You may use libdivide under the terms of either of these.
@@ -11,11 +11,14 @@
 #ifndef LIBDIVIDE_H
 #define LIBDIVIDE_H
 
-#define LIBDIVIDE_VERSION "5.0"
+// *** Version numbers are auto generated - do not edit ***
+#define LIBDIVIDE_VERSION "5.2.0"
 #define LIBDIVIDE_VERSION_MAJOR 5
-#define LIBDIVIDE_VERSION_MINOR 0
+#define LIBDIVIDE_VERSION_MINOR 2
+#define LIBDIVIDE_VERSION_PATCH 0
 
 #include <stdint.h>
+
 #if !defined(__AVR__)
 #include <stdio.h>
 #include <stdlib.h>
@@ -24,20 +27,29 @@
 #if defined(LIBDIVIDE_SSE2)
 #include <emmintrin.h>
 #endif
+
 #if defined(LIBDIVIDE_AVX2) || defined(LIBDIVIDE_AVX512)
 #include <immintrin.h>
 #endif
+
 #if defined(LIBDIVIDE_NEON)
 #include <arm_neon.h>
 #endif
 
+// Clang-cl prior to Visual Studio 2022 doesn't include __umulh/__mulh intrinsics
+#if defined(_MSC_VER) && defined(LIBDIVIDE_X86_64) && (!defined(__clang__) || _MSC_VER>1930)
+#define LIBDIVIDE_X64_INTRINSICS
+#endif
+
 #if defined(_MSC_VER)
+#if defined(LIBDIVIDE_X64_INTRINSICS)
 #include <intrin.h>
+#endif
 #pragma warning(push)
 // disable warning C4146: unary minus operator applied
 // to unsigned type, result still unsigned
 #pragma warning(disable : 4146)
-// disable warning C4204: nonstandard extension used : non-constant aggregate 
+// disable warning C4204: nonstandard extension used : non-constant aggregate
 // initializer
 //
 // It's valid C99
@@ -238,24 +250,32 @@ static LIBDIVIDE_INLINE struct libdivide_u64_branchfree_t libdivide_u64_branchfr
 static LIBDIVIDE_INLINE int16_t libdivide_s16_do_raw(
     int16_t numer, int16_t magic, uint8_t more);
 static LIBDIVIDE_INLINE int16_t libdivide_s16_do(
-    int16_t numer, const struct libdivide_s16_t* denom);
+    int16_t numer, const struct libdivide_s16_t *denom);
 static LIBDIVIDE_INLINE uint16_t libdivide_u16_do_raw(
-    uint16_t numer, uint16_t magic, uint8_t more);    
+    uint16_t numer, uint16_t magic, uint8_t more);
 static LIBDIVIDE_INLINE uint16_t libdivide_u16_do(
-    uint16_t numer, const struct libdivide_u16_t* denom);
+    uint16_t numer, const struct libdivide_u16_t *denom);
+static LIBDIVIDE_INLINE int32_t libdivide_s32_do_raw(
+    int32_t numer, int32_t magic, uint8_t more);
 static LIBDIVIDE_INLINE int32_t libdivide_s32_do(
     int32_t numer, const struct libdivide_s32_t *denom);
+static LIBDIVIDE_INLINE uint32_t libdivide_u32_do_raw(
+    uint32_t numer, uint32_t magic, uint8_t more);
 static LIBDIVIDE_INLINE uint32_t libdivide_u32_do(
     uint32_t numer, const struct libdivide_u32_t *denom);
+static LIBDIVIDE_INLINE int64_t libdivide_s64_do_raw(
+    int64_t numer, int64_t magic, uint8_t more);
 static LIBDIVIDE_INLINE int64_t libdivide_s64_do(
     int64_t numer, const struct libdivide_s64_t *denom);
+static LIBDIVIDE_INLINE uint64_t libdivide_u64_do_raw(
+    uint64_t numer, uint64_t magic, uint8_t more);
 static LIBDIVIDE_INLINE uint64_t libdivide_u64_do(
     uint64_t numer, const struct libdivide_u64_t *denom);
 
 static LIBDIVIDE_INLINE int16_t libdivide_s16_branchfree_do(
-    int16_t numer, const struct libdivide_s16_branchfree_t* denom);
+    int16_t numer, const struct libdivide_s16_branchfree_t *denom);
 static LIBDIVIDE_INLINE uint16_t libdivide_u16_branchfree_do(
-    uint16_t numer, const struct libdivide_u16_branchfree_t* denom);
+    uint16_t numer, const struct libdivide_u16_branchfree_t *denom);
 static LIBDIVIDE_INLINE int32_t libdivide_s32_branchfree_do(
     int32_t numer, const struct libdivide_s32_branchfree_t *denom);
 static LIBDIVIDE_INLINE uint32_t libdivide_u32_branchfree_do(
@@ -265,17 +285,17 @@ static LIBDIVIDE_INLINE int64_t libdivide_s64_branchfree_do(
 static LIBDIVIDE_INLINE uint64_t libdivide_u64_branchfree_do(
     uint64_t numer, const struct libdivide_u64_branchfree_t *denom);
 
-static LIBDIVIDE_INLINE int16_t libdivide_s16_recover(const struct libdivide_s16_t* denom);
-static LIBDIVIDE_INLINE uint16_t libdivide_u16_recover(const struct libdivide_u16_t* denom);
+static LIBDIVIDE_INLINE int16_t libdivide_s16_recover(const struct libdivide_s16_t *denom);
+static LIBDIVIDE_INLINE uint16_t libdivide_u16_recover(const struct libdivide_u16_t *denom);
 static LIBDIVIDE_INLINE int32_t libdivide_s32_recover(const struct libdivide_s32_t *denom);
 static LIBDIVIDE_INLINE uint32_t libdivide_u32_recover(const struct libdivide_u32_t *denom);
 static LIBDIVIDE_INLINE int64_t libdivide_s64_recover(const struct libdivide_s64_t *denom);
 static LIBDIVIDE_INLINE uint64_t libdivide_u64_recover(const struct libdivide_u64_t *denom);
 
 static LIBDIVIDE_INLINE int16_t libdivide_s16_branchfree_recover(
-    const struct libdivide_s16_branchfree_t* denom);
+    const struct libdivide_s16_branchfree_t *denom);
 static LIBDIVIDE_INLINE uint16_t libdivide_u16_branchfree_recover(
-    const struct libdivide_u16_branchfree_t* denom);
+    const struct libdivide_u16_branchfree_t *denom);
 static LIBDIVIDE_INLINE int32_t libdivide_s32_branchfree_recover(
     const struct libdivide_s32_branchfree_t *denom);
 static LIBDIVIDE_INLINE uint32_t libdivide_u32_branchfree_recover(
@@ -314,7 +334,7 @@ static LIBDIVIDE_INLINE int32_t libdivide_mullhi_s32(int32_t x, int32_t y) {
 }
 
 static LIBDIVIDE_INLINE uint64_t libdivide_mullhi_u64(uint64_t x, uint64_t y) {
-#if defined(LIBDIVIDE_VC) && defined(LIBDIVIDE_X86_64)
+#if defined(LIBDIVIDE_X64_INTRINSICS)
     return __umulh(x, y);
 #elif defined(HAS_INT128_T)
     __uint128_t xl = x, yl = y;
@@ -340,7 +360,7 @@ static LIBDIVIDE_INLINE uint64_t libdivide_mullhi_u64(uint64_t x, uint64_t y) {
 }
 
 static LIBDIVIDE_INLINE int64_t libdivide_mullhi_s64(int64_t x, int64_t y) {
-#if defined(LIBDIVIDE_VC) && defined(LIBDIVIDE_X86_64)
+#if defined(LIBDIVIDE_X64_INTRINSICS)
     return __mulh(x, y);
 #elif defined(HAS_INT128_T)
     __int128_t xl = x, yl = y;
@@ -393,7 +413,7 @@ static LIBDIVIDE_INLINE int16_t libdivide_count_leading_zeros16(uint16_t val) {
 
 static LIBDIVIDE_INLINE int32_t libdivide_count_leading_zeros32(uint32_t val) {
 #if defined(__AVR__)
-   // Fast way to count leading zeros
+    // Fast way to count leading zeros
     return __builtin_clzl(val);
 #elif defined(__GNUC__) || __has_builtin(__builtin_clz)
     // Fast way to count leading zeros
@@ -442,7 +462,7 @@ static LIBDIVIDE_INLINE int32_t libdivide_count_leading_zeros64(uint64_t val) {
 // uint {v}. The result must fit in 16 bits.
 // Returns the quotient directly and the remainder in *r
 static LIBDIVIDE_INLINE uint16_t libdivide_32_div_16_to_16(
-    uint16_t u1, uint16_t u0, uint16_t v, uint16_t* r) {
+    uint16_t u1, uint16_t u0, uint16_t v, uint16_t *r) {
     uint32_t n = ((uint32_t)u1 << 16) | u0;
     uint16_t result = (uint16_t)(n / v);
     *r = (uint16_t)(n - result * (uint32_t)v);
@@ -512,7 +532,7 @@ static LIBDIVIDE_INLINE uint64_t libdivide_128_div_64_to_64(
 
     // Check for overflow and divide by 0.
     if (numhi >= den) {
-        if (r != NULL) *r = ~0ull;
+        if (r) *r = ~0ull;
         return ~0ull;
     }
 
@@ -558,11 +578,14 @@ static LIBDIVIDE_INLINE uint64_t libdivide_128_div_64_to_64(
     q0 = (uint32_t)qhat;
 
     // Return remainder if requested.
-    if (r != NULL) *r = (rem * b + num0 - q0 * den) >> shift;
+    if (r) *r = (rem * b + num0 - q0 * den) >> shift;
     return ((uint64_t)q1 << 32) | q0;
 #endif
 }
 
+#if !(defined(HAS_INT128_T) && \
+      defined(HAS_INT128_DIV))
+
 // Bitshift a u128 in place, left (signed_shift > 0) or right (signed_shift < 0)
 static LIBDIVIDE_INLINE void libdivide_u128_shift(
     uint64_t *u1, uint64_t *u0, int32_t signed_shift) {
@@ -579,6 +602,8 @@ static LIBDIVIDE_INLINE void libdivide_u128_shift(
     }
 }
 
+#endif
+
 // Computes a 128 / 128 -> 64 bit division, with a 128 bit remainder.
 static LIBDIVIDE_INLINE uint64_t libdivide_128_div_128_to_64(
     uint64_t u_hi, uint64_t u_lo, uint64_t v_hi, uint64_t v_lo, uint64_t *r_hi, uint64_t *r_lo) {
@@ -696,8 +721,7 @@ static LIBDIVIDE_INLINE struct libdivide_u16_t libdivide_internal_u16_gen(
         // 1 in its recovery algorithm.
         result.magic = 0;
         result.more = (uint8_t)(floor_log_2_d - (branchfree != 0));
-    }
-    else {
+    } else {
         uint8_t more;
         uint16_t rem, proposed_m;
         proposed_m = libdivide_32_div_16_to_16((uint16_t)1 << floor_log_2_d, 0, d, &rem);
@@ -709,8 +733,7 @@ static LIBDIVIDE_INLINE struct libdivide_u16_t libdivide_internal_u16_gen(
         if (!branchfree && (e < ((uint16_t)1 << floor_log_2_d))) {
             // This power works
             more = floor_log_2_d;
-        }
-        else {
+        } else {
             // We have to use the general 17-bit algorithm.  We need to compute
             // (2**power) / d. However, we already have (2**(power-1))/d and
             // its remainder.  By doubling both, and then correcting the
@@ -742,7 +765,7 @@ struct libdivide_u16_branchfree_t libdivide_u16_branchfree_gen(uint16_t d) {
     }
     struct libdivide_u16_t tmp = libdivide_internal_u16_gen(d, 1);
     struct libdivide_u16_branchfree_t ret = {
-        tmp.magic, (uint8_t)(tmp.more & LIBDIVIDE_16_SHIFT_MASK) };
+        tmp.magic, (uint8_t)(tmp.more & LIBDIVIDE_16_SHIFT_MASK)};
     return ret;
 }
 
@@ -752,27 +775,25 @@ struct libdivide_u16_branchfree_t libdivide_u16_branchfree_gen(uint16_t d) {
 uint16_t libdivide_u16_do_raw(uint16_t numer, uint16_t magic, uint8_t more) {
     if (!magic) {
         return numer >> more;
-    }
-    else {
+    } else {
         uint16_t q = libdivide_mullhi_u16(magic, numer);
         if (more & LIBDIVIDE_ADD_MARKER) {
             uint16_t t = ((numer - q) >> 1) + q;
             return t >> (more & LIBDIVIDE_16_SHIFT_MASK);
-        }
-        else {
+        } else {
             // All upper bits are 0,
             // don't need to mask them off.
             return q >> more;
         }
-    }    
+    }
 }
 
-uint16_t libdivide_u16_do(uint16_t numer, const struct libdivide_u16_t* denom) {
+uint16_t libdivide_u16_do(uint16_t numer, const struct libdivide_u16_t *denom) {
     return libdivide_u16_do_raw(numer, denom->magic, denom->more);
 }
 
 uint16_t libdivide_u16_branchfree_do(
-    uint16_t numer, const struct libdivide_u16_branchfree_t* denom) {
+    uint16_t numer, const struct libdivide_u16_branchfree_t *denom) {
     uint16_t q = libdivide_mullhi_u16(denom->magic, numer);
     uint16_t t = ((numer - q) >> 1) + q;
     return t >> denom->more;
@@ -800,7 +821,7 @@ uint16_t libdivide_u16_recover(const struct libdivide_u16_t *denom) {
         // overflow. So we have to compute it as 2^(16+shift)/(m+2^16), and
         // then double the quotient and remainder.
         uint32_t half_n = (uint32_t)1 << (16 + shift);
-        uint32_t d = ( (uint32_t)1 << 16) | denom->magic;
+        uint32_t d = ((uint32_t)1 << 16) | denom->magic;
         // Note that the quotient is guaranteed <= 16 bits, but the remainder
         // may need 17!
         uint16_t half_q = (uint16_t)(half_n / d);
@@ -912,12 +933,11 @@ struct libdivide_u32_branchfree_t libdivide_u32_branchfree_gen(uint32_t d) {
     return ret;
 }
 
-uint32_t libdivide_u32_do(uint32_t numer, const struct libdivide_u32_t *denom) {
-    uint8_t more = denom->more;
-    if (!denom->magic) {
+uint32_t libdivide_u32_do_raw(uint32_t numer, uint32_t magic, uint8_t more) {
+    if (!magic) {
         return numer >> more;
     } else {
-        uint32_t q = libdivide_mullhi_u32(denom->magic, numer);
+        uint32_t q = libdivide_mullhi_u32(magic, numer);
         if (more & LIBDIVIDE_ADD_MARKER) {
             uint32_t t = ((numer - q) >> 1) + q;
             return t >> (more & LIBDIVIDE_32_SHIFT_MASK);
@@ -929,6 +949,10 @@ uint32_t libdivide_u32_do(uint32_t numer, const struct libdivide_u32_t *denom) {
     }
 }
 
+uint32_t libdivide_u32_do(uint32_t numer, const struct libdivide_u32_t *denom) {
+    return libdivide_u32_do_raw(numer, denom->magic, denom->more);
+}
+
 uint32_t libdivide_u32_branchfree_do(
     uint32_t numer, const struct libdivide_u32_branchfree_t *denom) {
     uint32_t q = libdivide_mullhi_u32(denom->magic, numer);
@@ -1072,12 +1096,11 @@ struct libdivide_u64_branchfree_t libdivide_u64_branchfree_gen(uint64_t d) {
     return ret;
 }
 
-uint64_t libdivide_u64_do(uint64_t numer, const struct libdivide_u64_t *denom) {
-    uint8_t more = denom->more;
-    if (!denom->magic) {
+uint64_t libdivide_u64_do_raw(uint64_t numer, uint64_t magic, uint8_t more) {
+   if (!magic) {
         return numer >> more;
     } else {
-        uint64_t q = libdivide_mullhi_u64(denom->magic, numer);
+        uint64_t q = libdivide_mullhi_u64(magic, numer);
         if (more & LIBDIVIDE_ADD_MARKER) {
             uint64_t t = ((numer - q) >> 1) + q;
             return t >> (more & LIBDIVIDE_64_SHIFT_MASK);
@@ -1089,6 +1112,10 @@ uint64_t libdivide_u64_do(uint64_t numer, const struct libdivide_u64_t *denom) {
     }
 }
 
+uint64_t libdivide_u64_do(uint64_t numer, const struct libdivide_u64_t *denom) {
+    return libdivide_u64_do_raw(numer, denom->magic, denom->more);
+}
+
 uint64_t libdivide_u64_branchfree_do(
     uint64_t numer, const struct libdivide_u64_branchfree_t *denom) {
     uint64_t q = libdivide_mullhi_u64(denom->magic, numer);
@@ -1428,11 +1455,10 @@ struct libdivide_s32_branchfree_t libdivide_s32_branchfree_gen(int32_t d) {
     return result;
 }
 
-int32_t libdivide_s32_do(int32_t numer, const struct libdivide_s32_t *denom) {
-    uint8_t more = denom->more;
+int32_t libdivide_s32_do_raw(int32_t numer, int32_t magic, uint8_t more) {
     uint8_t shift = more & LIBDIVIDE_32_SHIFT_MASK;
 
-    if (!denom->magic) {
+    if (!magic) {
         uint32_t sign = (int8_t)more >> 7;
         uint32_t mask = ((uint32_t)1 << shift) - 1;
         uint32_t uq = numer + ((numer >> 31) & mask);
@@ -1441,7 +1467,7 @@ int32_t libdivide_s32_do(int32_t numer, const struct libdivide_s32_t *denom) {
         q = (q ^ sign) - sign;
         return q;
     } else {
-        uint32_t uq = (uint32_t)libdivide_mullhi_s32(denom->magic, numer);
+        uint32_t uq = (uint32_t)libdivide_mullhi_s32(magic, numer);
         if (more & LIBDIVIDE_ADD_MARKER) {
             // must be arithmetic shift and then sign extend
             int32_t sign = (int8_t)more >> 7;
@@ -1456,6 +1482,10 @@ int32_t libdivide_s32_do(int32_t numer, const struct libdivide_s32_t *denom) {
     }
 }
 
+int32_t libdivide_s32_do(int32_t numer, const struct libdivide_s32_t *denom) {
+    return libdivide_s32_do_raw(numer, denom->magic, denom->more);
+}
+
 int32_t libdivide_s32_branchfree_do(int32_t numer, const struct libdivide_s32_branchfree_t *denom) {
     uint8_t more = denom->more;
     uint8_t shift = more & LIBDIVIDE_32_SHIFT_MASK;
@@ -1597,11 +1627,10 @@ struct libdivide_s64_branchfree_t libdivide_s64_branchfree_gen(int64_t d) {
     return ret;
 }
 
-int64_t libdivide_s64_do(int64_t numer, const struct libdivide_s64_t *denom) {
-    uint8_t more = denom->more;
+int64_t libdivide_s64_do_raw(int64_t numer, int64_t magic, uint8_t more) {
     uint8_t shift = more & LIBDIVIDE_64_SHIFT_MASK;
 
-    if (!denom->magic) {  // shift path
+    if (!magic) {  // shift path
         uint64_t mask = ((uint64_t)1 << shift) - 1;
         uint64_t uq = numer + ((numer >> 63) & mask);
         int64_t q = (int64_t)uq;
@@ -1611,7 +1640,7 @@ int64_t libdivide_s64_do(int64_t numer, const struct libdivide_s64_t *denom) {
         q = (q ^ sign) - sign;
         return q;
     } else {
-        uint64_t uq = (uint64_t)libdivide_mullhi_s64(denom->magic, numer);
+        uint64_t uq = (uint64_t)libdivide_mullhi_s64(magic, numer);
         if (more & LIBDIVIDE_ADD_MARKER) {
             // must be arithmetic shift and then sign extend
             int64_t sign = (int8_t)more >> 7;
@@ -1626,6 +1655,10 @@ int64_t libdivide_s64_do(int64_t numer, const struct libdivide_s64_t *denom) {
     }
 }
 
+int64_t libdivide_s64_do(int64_t numer, const struct libdivide_s64_t *denom) {
+    return libdivide_s64_do_raw(numer, denom->magic, denom->more);
+}
+
 int64_t libdivide_s64_branchfree_do(int64_t numer, const struct libdivide_s64_branchfree_t *denom) {
     uint8_t more = denom->more;
     uint8_t shift = more & LIBDIVIDE_64_SHIFT_MASK;
@@ -1682,15 +1715,22 @@ int64_t libdivide_s64_branchfree_recover(const struct libdivide_s64_branchfree_t
 
 // Simplest possible vector type division: treat the vector type as an array
 // of underlying native type.
-#define SIMPLE_VECTOR_DIVISION(IntT, VecT, Algo) \
-    const size_t count = sizeof(VecT) / sizeof(IntT); \
-    VecT result; \
-    IntT *pSource = (IntT *)&numers; \
-    IntT *pTarget = (IntT *)&result; \
-    for (size_t loop=0; loop<count; ++loop) { \
-        pTarget[loop] = libdivide_##Algo##_do(pSource[loop], denom); \
-    } \
-    return result; \
+//
+// Use a union to read a vector via pointer-to-integer, without violating strict
+// aliasing.
+#define SIMPLE_VECTOR_DIVISION(IntT, VecT, Algo)                          \
+    const size_t count = sizeof(VecT) / sizeof(IntT);                     \
+    union type_pun_vec {                                                  \
+        VecT vec;                                                         \
+        IntT arr[sizeof(VecT) / sizeof(IntT)];                            \
+    };                                                                    \
+    union type_pun_vec result;                                            \
+    union type_pun_vec input;                                             \
+    input.vec = numers;                                                   \
+    for (size_t loop = 0; loop < count; ++loop) {                         \
+        result.arr[loop] = libdivide_##Algo##_do(input.arr[loop], denom); \
+    }                                                                     \
+    return result.vec;
 
 #if defined(LIBDIVIDE_NEON)
 
@@ -1804,13 +1844,12 @@ static LIBDIVIDE_INLINE int64x2_t libdivide_mullhi_s64_vec128(int64x2_t x, int64
 
 ////////// UINT16
 
-uint16x8_t libdivide_u16_do_vec128(uint16x8_t numers, const struct libdivide_u16_t *denom) {
-    SIMPLE_VECTOR_DIVISION(uint16_t, uint16x8_t, u16)
-}
+uint16x8_t libdivide_u16_do_vec128(uint16x8_t numers, const struct libdivide_u16_t *denom){
+    SIMPLE_VECTOR_DIVISION(uint16_t, uint16x8_t, u16)}
 
-uint16x8_t libdivide_u16_branchfree_do_vec128(uint16x8_t numers, const struct libdivide_u16_branchfree_t *denom) {
-    SIMPLE_VECTOR_DIVISION(uint16_t, uint16x8_t, u16_branchfree)
-}
+uint16x8_t libdivide_u16_branchfree_do_vec128(
+    uint16x8_t numers, const struct libdivide_u16_branchfree_t *denom){
+    SIMPLE_VECTOR_DIVISION(uint16_t, uint16x8_t, u16_branchfree)}
 
 ////////// UINT32
 
@@ -1870,13 +1909,12 @@ uint64x2_t libdivide_u64_branchfree_do_vec128(
 
 ////////// SINT16
 
-int16x8_t libdivide_s16_do_vec128(int16x8_t numers, const struct libdivide_s16_t *denom) {
-    SIMPLE_VECTOR_DIVISION(int16_t, int16x8_t, s16)
-}
+int16x8_t libdivide_s16_do_vec128(int16x8_t numers, const struct libdivide_s16_t *denom){
+    SIMPLE_VECTOR_DIVISION(int16_t, int16x8_t, s16)}
 
-int16x8_t libdivide_s16_branchfree_do_vec128(int16x8_t numers, const struct libdivide_s16_branchfree_t *denom) {
-    SIMPLE_VECTOR_DIVISION(int16_t, int16x8_t, s16_branchfree)
-}
+int16x8_t libdivide_s16_branchfree_do_vec128(
+    int16x8_t numers, const struct libdivide_s16_branchfree_t *denom){
+    SIMPLE_VECTOR_DIVISION(int16_t, int16x8_t, s16_branchfree)}
 
 ////////// SINT32
 
@@ -2082,13 +2120,12 @@ static LIBDIVIDE_INLINE __m512i libdivide_mullhi_s64_vec512(__m512i x, __m512i y
 
 ////////// UINT16
 
-__m512i libdivide_u16_do_vec512(__m512i numers, const struct libdivide_u16_t *denom) {
-    SIMPLE_VECTOR_DIVISION(uint16_t, __m512i, u16)
-}
+__m512i libdivide_u16_do_vec512(__m512i numers, const struct libdivide_u16_t *denom){
+    SIMPLE_VECTOR_DIVISION(uint16_t, __m512i, u16)}
 
-__m512i libdivide_u16_branchfree_do_vec512(__m512i numers, const struct libdivide_u16_branchfree_t *denom) {
-    SIMPLE_VECTOR_DIVISION(uint16_t, __m512i, u16_branchfree)
-}
+__m512i libdivide_u16_branchfree_do_vec512(
+    __m512i numers, const struct libdivide_u16_branchfree_t *denom){
+    SIMPLE_VECTOR_DIVISION(uint16_t, __m512i, u16_branchfree)}
 
 ////////// UINT32
 
@@ -2146,13 +2183,12 @@ __m512i libdivide_u64_branchfree_do_vec512(
 
 ////////// SINT16
 
-__m512i libdivide_s16_do_vec512(__m512i numers, const struct libdivide_s16_t *denom) {
-    SIMPLE_VECTOR_DIVISION(int16_t, __m512i, s16)
-}
+__m512i libdivide_s16_do_vec512(__m512i numers, const struct libdivide_s16_t *denom){
+    SIMPLE_VECTOR_DIVISION(int16_t, __m512i, s16)}
 
-__m512i libdivide_s16_branchfree_do_vec512(__m512i numers, const struct libdivide_s16_branchfree_t *denom) {
-    SIMPLE_VECTOR_DIVISION(int16_t, __m512i, s16_branchfree)
-}
+__m512i libdivide_s16_branchfree_do_vec512(
+    __m512i numers, const struct libdivide_s16_branchfree_t *denom){
+    SIMPLE_VECTOR_DIVISION(int16_t, __m512i, s16_branchfree)}
 
 ////////// SINT32
 
@@ -2365,11 +2401,25 @@ static LIBDIVIDE_INLINE __m256i libdivide_mullhi_s64_vec256(__m256i x, __m256i y
 ////////// UINT16
 
 __m256i libdivide_u16_do_vec256(__m256i numers, const struct libdivide_u16_t *denom) {
-    SIMPLE_VECTOR_DIVISION(uint16_t, __m256i, u16)
+    uint8_t more = denom->more;
+    if (!denom->magic) {
+        return _mm256_srli_epi16(numers, more);
+    } else {
+        __m256i q = _mm256_mulhi_epu16(numers, _mm256_set1_epi16(denom->magic));
+        if (more & LIBDIVIDE_ADD_MARKER) {
+            __m256i t = _mm256_adds_epu16(_mm256_srli_epi16(_mm256_subs_epu16(numers, q), 1), q);
+            return _mm256_srli_epi16(t, (more & LIBDIVIDE_16_SHIFT_MASK));
+        } else {
+            return _mm256_srli_epi16(q, more);
+        }
+    }
 }
 
-__m256i libdivide_u16_branchfree_do_vec256(__m256i numers, const struct libdivide_u16_branchfree_t *denom) {
-    SIMPLE_VECTOR_DIVISION(uint16_t, __m256i, u16_branchfree)
+__m256i libdivide_u16_branchfree_do_vec256(
+    __m256i numers, const struct libdivide_u16_branchfree_t *denom) {
+    __m256i q = _mm256_mulhi_epu16(numers, _mm256_set1_epi16(denom->magic));
+    __m256i t = _mm256_adds_epu16(_mm256_srli_epi16(_mm256_subs_epu16(numers, q), 1), q);
+    return _mm256_srli_epi16(t, denom->more);
 }
 
 ////////// UINT32
@@ -2429,11 +2479,54 @@ __m256i libdivide_u64_branchfree_do_vec256(
 ////////// SINT16
 
 __m256i libdivide_s16_do_vec256(__m256i numers, const struct libdivide_s16_t *denom) {
-    SIMPLE_VECTOR_DIVISION(int16_t, __m256i, s16)
+    uint8_t more = denom->more;
+    if (!denom->magic) {
+        uint16_t shift = more & LIBDIVIDE_16_SHIFT_MASK;
+        uint16_t mask = ((uint16_t)1 << shift) - 1;
+        __m256i roundToZeroTweak = _mm256_set1_epi16(mask);
+        // q = numer + ((numer >> 15) & roundToZeroTweak);
+        __m256i q = _mm256_add_epi16(
+            numers, _mm256_and_si256(_mm256_srai_epi16(numers, 15), roundToZeroTweak));
+        q = _mm256_srai_epi16(q, shift);
+        __m256i sign = _mm256_set1_epi16((int8_t)more >> 7);
+        // q = (q ^ sign) - sign;
+        q = _mm256_sub_epi16(_mm256_xor_si256(q, sign), sign);
+        return q;
+    } else {
+        __m256i q = _mm256_mulhi_epi16(numers, _mm256_set1_epi16(denom->magic));
+        if (more & LIBDIVIDE_ADD_MARKER) {
+            // must be arithmetic shift
+            __m256i sign = _mm256_set1_epi16((int8_t)more >> 7);
+            // q += ((numer ^ sign) - sign);
+            q = _mm256_add_epi16(q, _mm256_sub_epi16(_mm256_xor_si256(numers, sign), sign));
+        }
+        // q >>= shift
+        q = _mm256_srai_epi16(q, more & LIBDIVIDE_16_SHIFT_MASK);
+        q = _mm256_add_epi16(q, _mm256_srli_epi16(q, 15));  // q += (q < 0)
+        return q;
+    }
 }
 
-__m256i libdivide_s16_branchfree_do_vec256(__m256i numers, const struct libdivide_s16_branchfree_t *denom) {
-    SIMPLE_VECTOR_DIVISION(int16_t, __m256i, s16_branchfree)
+__m256i libdivide_s16_branchfree_do_vec256(
+    __m256i numers, const struct libdivide_s16_branchfree_t *denom) {
+    int16_t magic = denom->magic;
+    uint8_t more = denom->more;
+    uint8_t shift = more & LIBDIVIDE_16_SHIFT_MASK;
+    // must be arithmetic shift
+    __m256i sign = _mm256_set1_epi16((int8_t)more >> 7);
+    __m256i q = _mm256_mulhi_epi16(numers, _mm256_set1_epi16(magic));
+    q = _mm256_add_epi16(q, numers);  // q += numers
+
+    // If q is non-negative, we have nothing to do
+    // If q is negative, we want to add either (2**shift)-1 if d is
+    // a power of 2, or (2**shift) if it is not a power of 2
+    uint16_t is_power_of_2 = (magic == 0);
+    __m256i q_sign = _mm256_srai_epi16(q, 15);  // q_sign = q >> 15
+    __m256i mask = _mm256_set1_epi16(((uint16_t)1 << shift) - is_power_of_2);
+    q = _mm256_add_epi16(q, _mm256_and_si256(q_sign, mask));  // q = q + (q_sign & mask)
+    q = _mm256_srai_epi16(q, shift);                          // q >>= shift
+    q = _mm256_sub_epi16(_mm256_xor_si256(q, sign), sign);    // q = (q ^ sign) - sign
+    return q;
 }
 
 ////////// SINT32
@@ -2661,11 +2754,25 @@ static LIBDIVIDE_INLINE __m128i libdivide_mullhi_s64_vec128(__m128i x, __m128i y
 ////////// UINT26
 
 __m128i libdivide_u16_do_vec128(__m128i numers, const struct libdivide_u16_t *denom) {
-    SIMPLE_VECTOR_DIVISION(uint16_t, __m128i, u16)
+    uint8_t more = denom->more;
+    if (!denom->magic) {
+        return _mm_srli_epi16(numers, more);
+    } else {
+        __m128i q = _mm_mulhi_epu16(numers, _mm_set1_epi16(denom->magic));
+        if (more & LIBDIVIDE_ADD_MARKER) {
+            __m128i t = _mm_adds_epu16(_mm_srli_epi16(_mm_subs_epu16(numers, q), 1), q);
+            return _mm_srli_epi16(t, (more & LIBDIVIDE_16_SHIFT_MASK));
+        } else {
+            return _mm_srli_epi16(q, more);
+        }
+    }
 }
 
-__m128i libdivide_u16_branchfree_do_vec128(__m128i numers, const struct libdivide_u16_branchfree_t *denom) {
-    SIMPLE_VECTOR_DIVISION(uint16_t, __m128i, u16_branchfree)
+__m128i libdivide_u16_branchfree_do_vec128(
+    __m128i numers, const struct libdivide_u16_branchfree_t *denom) {
+    __m128i q = _mm_mulhi_epu16(numers, _mm_set1_epi16(denom->magic));
+    __m128i t = _mm_adds_epu16(_mm_srli_epi16(_mm_subs_epu16(numers, q), 1), q);
+    return _mm_srli_epi16(t, denom->more);
 }
 
 ////////// UINT32
@@ -2725,11 +2832,54 @@ __m128i libdivide_u64_branchfree_do_vec128(
 ////////// SINT16
 
 __m128i libdivide_s16_do_vec128(__m128i numers, const struct libdivide_s16_t *denom) {
-    SIMPLE_VECTOR_DIVISION(int16_t, __m128i, s16)
+    uint8_t more = denom->more;
+    if (!denom->magic) {
+        uint16_t shift = more & LIBDIVIDE_16_SHIFT_MASK;
+        uint16_t mask = ((uint16_t)1 << shift) - 1;
+        __m128i roundToZeroTweak = _mm_set1_epi16(mask);
+        // q = numer + ((numer >> 15) & roundToZeroTweak);
+        __m128i q =
+            _mm_add_epi16(numers, _mm_and_si128(_mm_srai_epi16(numers, 15), roundToZeroTweak));
+        q = _mm_srai_epi16(q, shift);
+        __m128i sign = _mm_set1_epi16((int8_t)more >> 7);
+        // q = (q ^ sign) - sign;
+        q = _mm_sub_epi16(_mm_xor_si128(q, sign), sign);
+        return q;
+    } else {
+        __m128i q = _mm_mulhi_epi16(numers, _mm_set1_epi16(denom->magic));
+        if (more & LIBDIVIDE_ADD_MARKER) {
+            // must be arithmetic shift
+            __m128i sign = _mm_set1_epi16((int8_t)more >> 7);
+            // q += ((numer ^ sign) - sign);
+            q = _mm_add_epi16(q, _mm_sub_epi16(_mm_xor_si128(numers, sign), sign));
+        }
+        // q >>= shift
+        q = _mm_srai_epi16(q, more & LIBDIVIDE_16_SHIFT_MASK);
+        q = _mm_add_epi16(q, _mm_srli_epi16(q, 15));  // q += (q < 0)
+        return q;
+    }
 }
 
-__m128i libdivide_s16_branchfree_do_vec128(__m128i numers, const struct libdivide_s16_branchfree_t *denom) {
-    SIMPLE_VECTOR_DIVISION(int16_t, __m128i, s16_branchfree)
+__m128i libdivide_s16_branchfree_do_vec128(
+    __m128i numers, const struct libdivide_s16_branchfree_t *denom) {
+    int16_t magic = denom->magic;
+    uint8_t more = denom->more;
+    uint8_t shift = more & LIBDIVIDE_16_SHIFT_MASK;
+    // must be arithmetic shift
+    __m128i sign = _mm_set1_epi16((int8_t)more >> 7);
+    __m128i q = _mm_mulhi_epi16(numers, _mm_set1_epi16(magic));
+    q = _mm_add_epi16(q, numers);  // q += numers
+
+    // If q is non-negative, we have nothing to do
+    // If q is negative, we want to add either (2**shift)-1 if d is
+    // a power of 2, or (2**shift) if it is not a power of 2
+    uint16_t is_power_of_2 = (magic == 0);
+    __m128i q_sign = _mm_srai_epi16(q, 15);  // q_sign = q >> 15
+    __m128i mask = _mm_set1_epi16(((uint16_t)1 << shift) - is_power_of_2);
+    q = _mm_add_epi16(q, _mm_and_si128(q_sign, mask));  // q = q + (q_sign & mask)
+    q = _mm_srai_epi16(q, shift);                       // q >>= shift
+    q = _mm_sub_epi16(_mm_xor_si128(q, sign), sign);    // q = (q ^ sign) - sign
+    return q;
 }
 
 ////////// SINT32
@@ -2795,8 +2945,8 @@ __m128i libdivide_s64_do_vec128(__m128i numers, const struct libdivide_s64_t *de
         uint64_t mask = ((uint64_t)1 << shift) - 1;
         __m128i roundToZeroTweak = _mm_set1_epi64x(mask);
         // q = numer + ((numer >> 63) & roundToZeroTweak);
-        __m128i q =
-            _mm_add_epi64(numers, _mm_and_si128(libdivide_s64_signbits_vec128(numers), roundToZeroTweak));
+        __m128i q = _mm_add_epi64(
+            numers, _mm_and_si128(libdivide_s64_signbits_vec128(numers), roundToZeroTweak));
         q = libdivide_s64_shift_right_vec128(q, shift);
         __m128i sign = _mm_set1_epi32((int8_t)more >> 7);
         // q = (q ^ sign) - sign;
@@ -2847,49 +2997,80 @@ __m128i libdivide_s64_branchfree_do_vec128(
 
 #ifdef __cplusplus
 
+//for constexpr zero initialization,
+//c++11 might handle things ok,
+//but just limit to at least c++14 to ensure
+//we don't break anyone's code:
+
+// for gcc and clang, use https://en.cppreference.com/w/cpp/feature_test#cpp_constexpr
+#if (defined(__GNUC__) || defined(__clang__)) && (__cpp_constexpr >= 201304L)
+#define LIBDIVIDE_CONSTEXPR constexpr
+
+// supposedly, MSVC might not implement feature test macros right (https://stackoverflow.com/questions/49316752/feature-test-macros-not-working-properly-in-visual-c)
+// so check that _MSVC_LANG corresponds to at least c++14, and _MSC_VER corresponds to at least VS 2017 15.0 (for extended constexpr support https://learn.microsoft.com/en-us/cpp/overview/visual-cpp-language-conformance?view=msvc-170)
+#elif defined(_MSC_VER) && _MSC_VER >= 1910 && defined(_MSVC_LANG) && _MSVC_LANG >=201402L
+#define LIBDIVIDE_CONSTEXPR constexpr
+
+// in case some other obscure compiler has the right __cpp_constexpr :
+#elif defined(__cpp_constexpr) && __cpp_constexpr >= 201304L
+#define LIBDIVIDE_CONSTEXPR constexpr
+
+#else
+#define LIBDIVIDE_CONSTEXPR LIBDIVIDE_INLINE
+#endif
+
 enum Branching {
     BRANCHFULL,  // use branching algorithms
     BRANCHFREE   // use branchfree algorithms
 };
 
+namespace detail {
+enum Signedness {
+    SIGNED,
+    UNSIGNED,
+};
+
 #if defined(LIBDIVIDE_NEON)
 // Helper to deduce NEON vector type for integral type.
-template <typename T>
-struct NeonVecFor {};
+template <int _WIDTH, Signedness _SIGN>
+struct NeonVec {};
 
 template <>
-struct NeonVecFor<uint16_t> {
+struct NeonVec<16, UNSIGNED> {
     typedef uint16x8_t type;
 };
 
 template <>
-struct NeonVecFor<int16_t> {
+struct NeonVec<16, SIGNED> {
     typedef int16x8_t type;
 };
 
 template <>
-struct NeonVecFor<uint32_t> {
+struct NeonVec<32, UNSIGNED> {
     typedef uint32x4_t type;
 };
 
 template <>
-struct NeonVecFor<int32_t> {
+struct NeonVec<32, SIGNED> {
     typedef int32x4_t type;
 };
 
 template <>
-struct NeonVecFor<uint64_t> {
+struct NeonVec<64, UNSIGNED> {
     typedef uint64x2_t type;
 };
 
 template <>
-struct NeonVecFor<int64_t> {
+struct NeonVec<64, SIGNED> {
     typedef int64x2_t type;
 };
-#endif
 
-// Versions of our algorithms for SIMD.
-#if defined(LIBDIVIDE_NEON)
+template <typename T>
+struct NeonVecFor {
+    // See 'class divider' for an explanation of these template parameters.
+    typedef typename NeonVec<sizeof(T) * 8, (((T)0 >> 0) > (T)(-1) ? SIGNED : UNSIGNED)>::type type;
+};
+
 #define LIBDIVIDE_DIVIDE_NEON(ALGO, INT_TYPE)                    \
     LIBDIVIDE_INLINE typename NeonVecFor<INT_TYPE>::type divide( \
         typename NeonVecFor<INT_TYPE>::type n) const {           \
@@ -2898,6 +3079,7 @@ struct NeonVecFor<int64_t> {
 #else
 #define LIBDIVIDE_DIVIDE_NEON(ALGO, INT_TYPE)
 #endif
+
 #if defined(LIBDIVIDE_SSE2)
 #define LIBDIVIDE_DIVIDE_SSE2(ALGO)                     \
     LIBDIVIDE_INLINE __m128i divide(__m128i n) const {  \
@@ -2930,6 +3112,7 @@ struct NeonVecFor<int64_t> {
 #define DISPATCHER_GEN(T, ALGO)                                                       \
     libdivide_##ALGO##_t denom;                                                       \
     LIBDIVIDE_INLINE dispatcher() {}                                                  \
+    explicit LIBDIVIDE_CONSTEXPR dispatcher(decltype(nullptr)) : denom{} {}              \
     LIBDIVIDE_INLINE dispatcher(T d) : denom(libdivide_##ALGO##_gen(d)) {}            \
     LIBDIVIDE_INLINE T divide(T n) const { return libdivide_##ALGO##_do(n, &denom); } \
     LIBDIVIDE_INLINE T recover() const { return libdivide_##ALGO##_recover(&denom); } \
@@ -2939,66 +3122,81 @@ struct NeonVecFor<int64_t> {
     LIBDIVIDE_DIVIDE_AVX512(ALGO)
 
 // The dispatcher selects a specific division algorithm for a given
-// type and ALGO using partial template specialization.
-template <typename _IntT, Branching ALGO>
+// width, signedness, and ALGO using partial template specialization.
+template <int _WIDTH, Signedness _SIGN, Branching _ALGO>
 struct dispatcher {};
 
 template <>
-struct dispatcher<int16_t, BRANCHFULL> {
+struct dispatcher<16, SIGNED, BRANCHFULL> {
     DISPATCHER_GEN(int16_t, s16)
 };
 template <>
-struct dispatcher<int16_t, BRANCHFREE> {
+struct dispatcher<16, SIGNED, BRANCHFREE> {
     DISPATCHER_GEN(int16_t, s16_branchfree)
 };
 template <>
-struct dispatcher<uint16_t, BRANCHFULL> {
+struct dispatcher<16, UNSIGNED, BRANCHFULL> {
     DISPATCHER_GEN(uint16_t, u16)
 };
 template <>
-struct dispatcher<uint16_t, BRANCHFREE> {
+struct dispatcher<16, UNSIGNED, BRANCHFREE> {
     DISPATCHER_GEN(uint16_t, u16_branchfree)
 };
 template <>
-struct dispatcher<int32_t, BRANCHFULL> {
+struct dispatcher<32, SIGNED, BRANCHFULL> {
     DISPATCHER_GEN(int32_t, s32)
 };
 template <>
-struct dispatcher<int32_t, BRANCHFREE> {
+struct dispatcher<32, SIGNED, BRANCHFREE> {
     DISPATCHER_GEN(int32_t, s32_branchfree)
 };
 template <>
-struct dispatcher<uint32_t, BRANCHFULL> {
+struct dispatcher<32, UNSIGNED, BRANCHFULL> {
     DISPATCHER_GEN(uint32_t, u32)
 };
 template <>
-struct dispatcher<uint32_t, BRANCHFREE> {
+struct dispatcher<32, UNSIGNED, BRANCHFREE> {
     DISPATCHER_GEN(uint32_t, u32_branchfree)
 };
 template <>
-struct dispatcher<int64_t, BRANCHFULL> {
+struct dispatcher<64, SIGNED, BRANCHFULL> {
     DISPATCHER_GEN(int64_t, s64)
 };
 template <>
-struct dispatcher<int64_t, BRANCHFREE> {
+struct dispatcher<64, SIGNED, BRANCHFREE> {
     DISPATCHER_GEN(int64_t, s64_branchfree)
 };
 template <>
-struct dispatcher<uint64_t, BRANCHFULL> {
+struct dispatcher<64, UNSIGNED, BRANCHFULL> {
     DISPATCHER_GEN(uint64_t, u64)
 };
 template <>
-struct dispatcher<uint64_t, BRANCHFREE> {
+struct dispatcher<64, UNSIGNED, BRANCHFREE> {
     DISPATCHER_GEN(uint64_t, u64_branchfree)
 };
+}  // namespace detail
+
+#if defined(LIBDIVIDE_NEON)
+// Allow NeonVecFor outside of detail namespace.
+template <typename T>
+struct NeonVecFor {
+    typedef typename detail::NeonVecFor<T>::type type;
+};
+#endif
 
 // This is the main divider class for use by the user (C++ API).
 // The actual division algorithm is selected using the dispatcher struct
-// based on the integer and algorithm template parameters.
+// based on the integer width and algorithm template parameters.
 template <typename T, Branching ALGO = BRANCHFULL>
 class divider {
    private:
-    typedef dispatcher<T, ALGO> dispatcher_t;
+    // Dispatch based on the size and signedness.
+    // We avoid using type_traits as it's not available in AVR.
+    // Detect signedness by checking if T(-1) is less than T(0).
+    // Also throw in a shift by 0, which prevents floating point types from being passed.
+    typedef detail::dispatcher<sizeof(T) * 8,
+        (((T)0 >> 0) > (T)(-1) ? detail::SIGNED : detail::UNSIGNED), ALGO>
+        dispatcher_t;
 
    public:
     // We leave the default constructor empty so that creating
@@ -3006,6 +3204,9 @@ class divider {
     // later doesn't slow us down.
     divider() {}
 
+    // constexpr zero-initialization to allow for use w/ static constinit
+    explicit LIBDIVIDE_CONSTEXPR divider(decltype(nullptr)) : div(nullptr) {}
+
     // Constructor that takes the divisor as a parameter
     LIBDIVIDE_INLINE divider(T d) : div(d) {}
 
@@ -3017,7 +3218,7 @@ class divider {
     T recover() const { return div.recover(); }
 
     bool operator==(const divider<T, ALGO> &other) const {
-        return div.denom.magic == other.denom.magic && div.denom.more == other.denom.more;
+        return div.denom.magic == other.div.denom.magic && div.denom.more == other.div.denom.more;
     }
 
     bool operator!=(const divider<T, ALGO> &other) const { return !(*this == other); }
@@ -3098,12 +3299,14 @@ LIBDIVIDE_INLINE __m512i operator/=(__m512i &n, const divider<T, ALGO> &div) {
 
 #if defined(LIBDIVIDE_NEON)
 template <typename T, Branching ALGO>
-LIBDIVIDE_INLINE typename NeonVecFor<T>::type operator/(typename NeonVecFor<T>::type n, const divider<T, ALGO> &div) {
+LIBDIVIDE_INLINE typename NeonVecFor<T>::type operator/(
+    typename NeonVecFor<T>::type n, const divider<T, ALGO> &div) {
     return div.divide(n);
 }
 
 template <typename T, Branching ALGO>
-LIBDIVIDE_INLINE typename NeonVecFor<T>::type operator/=(typename NeonVecFor<T>::type &n, const divider<T, ALGO> &div) {
+LIBDIVIDE_INLINE typename NeonVecFor<T>::type operator/=(
+    typename NeonVecFor<T>::type &n, const divider<T, ALGO> &div) {
     n = div.divide(n);
     return n;
 }

util.h

@@ -9,7 +9,9 @@
 #define noreturn __attribute__((noreturn))
 
 #define likely(x) __builtin_expect(!!(x), 1)
+#define likely51(x) __builtin_expect_with_probability(!!(x), 1, 0.51)
 #define unlikely(x) __builtin_expect(!!(x), 0)
+#define unlikely51(x) __builtin_expect_with_probability(!!(x), 0, 0.51)
 
 #define min(x, y) ({ \
     __typeof__(x) _x = (x); \
@@ -30,6 +32,13 @@
 #define STRINGIFY(s) #s
 #define ALIAS(f) __attribute__((alias(STRINGIFY(f))))
 
+// supported since GCC 15
+#if __has_attribute (nonstring)
+#  define NONSTRING __attribute__ ((nonstring))
+#else
+#  define NONSTRING
+#endif
+
 typedef uint8_t u8;
 typedef uint16_t u16;
 typedef uint32_t u32;