update libdivide to 5.2.0

Author: thestinger

third_party/libdivide.h

@@ -11,9 +11,11 @@
 #ifndef LIBDIVIDE_H
 #define LIBDIVIDE_H
 
-#define LIBDIVIDE_VERSION "5.1"
+// *** Version numbers are auto generated - do not edit ***
+#define LIBDIVIDE_VERSION "5.2.0"
 #define LIBDIVIDE_VERSION_MAJOR 5
-#define LIBDIVIDE_VERSION_MINOR 1
+#define LIBDIVIDE_VERSION_MINOR 2
+#define LIBDIVIDE_VERSION_PATCH 0
 
 #include <stdint.h>
 
@@ -34,8 +36,15 @@
 #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
@@ -238,18 +247,28 @@ static LIBDIVIDE_INLINE struct libdivide_u32_branchfree_t libdivide_u32_branchfr
 static LIBDIVIDE_INLINE struct libdivide_s64_branchfree_t libdivide_s64_branchfree_gen(int64_t d);
 static LIBDIVIDE_INLINE struct libdivide_u64_branchfree_t libdivide_u64_branchfree_gen(uint64_t d);
 
-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_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);
-static LIBDIVIDE_INLINE uint16_t libdivide_u16_do_raw(uint16_t numer, uint16_t magic, uint8_t more);
+static LIBDIVIDE_INLINE uint16_t libdivide_u16_do_raw(
+    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);
+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);
 
@@ -315,7 +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;
@@ -341,7 +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;
@@ -914,12 +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);
@@ -931,6 +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);
@@ -1074,12 +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);
@@ -1091,6 +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);
@@ -1430,11 +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);
@@ -1443,7 +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;
@@ -1458,6 +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;
@@ -1599,11 +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;
@@ -1613,7 +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;
@@ -1628,6 +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;