Master vs2019 x86 compile fixes (simdjson#743)

pps83 · web-flow · commit ffaa292006b8 · 2020-04-21T14:42:53.000-04:00
* Added bitexact implementations of _BitScanForward64 and _BitScanReverse64 for VS2019 32-bit builds * Added bitexact implementations of _umul128 for VS2019 x86, arm, arm64 builds * Implement mul_overflow for VS2019 arm64 builds + implement mul_overflow using __umulh (msvc/clang results: https://godbolt.org/z/smRwA7) * Added Win32 for VS2019 to .appveyor.yml * Update amalgamated headers (fix x86 builds with VS2019)
diff --git a/.appveyor.yml b/.appveyor.yml
@@ -12,14 +12,19 @@ environment:
   matrix:
     - SIMDJSON_BUILD_STATIC: ON
       SIMDJSON_ENABLE_THREADS: OFF
+      SIMDJSON_PLATFORM: x64
+    - SIMDJSON_BUILD_STATIC: ON
+      SIMDJSON_ENABLE_THREADS: OFF
+      SIMDJSON_PLATFORM: Win32
     - SIMDJSON_BUILD_STATIC: OFF
       SIMDJSON_ENABLE_THREADS: ON
+      SIMDJSON_PLATFORM: x64
 
 build_script:
   - set
   - mkdir build
   - cd build
-  - cmake -DSIMDJSON_BUILD_STATIC=%SIMDJSON_BUILD_STATIC% -DSIMDJSON_ENABLE_THREADS=%SIMDJSON_ENABLE_THREADS% -DCMAKE_BUILD_TYPE=%Configuration% -DCMAKE_GENERATOR_PLATFORM=x64 ..
+  - cmake -DSIMDJSON_BUILD_STATIC=%SIMDJSON_BUILD_STATIC% -DSIMDJSON_ENABLE_THREADS=%SIMDJSON_ENABLE_THREADS% -DCMAKE_BUILD_TYPE=%Configuration% -DCMAKE_GENERATOR_PLATFORM=%SIMDJSON_PLATFORM% ..
   - cmake -LH ..
   - cmake --build . --config %Configuration% --verbose
 
diff --git a/singleheader/simdjson.cpp b/singleheader/simdjson.cpp
@@ -697,16 +697,10 @@ really_inline bool add_overflow(uint64_t value1, uint64_t value2, uint64_t *resu
 #endif
 }
 
-#ifdef _MSC_VER
-#pragma intrinsic(_umul128) // todo: this might fail under visual studio for ARM
-#endif
-
 really_inline bool mul_overflow(uint64_t value1, uint64_t value2, uint64_t *result) {
 #ifdef _MSC_VER
-  // todo: this might fail under visual studio for ARM
-  uint64_t high;
-  *result = _umul128(value1, value2, &high);
-  return high;
+  *result = value1 * value2;
+  return !!__umulh(value1, value2);
 #else
   return __builtin_umulll_overflow(value1, value2, (unsigned long long *)result);
 #endif
@@ -5759,7 +5753,7 @@ inline size_t codepoint_to_utf8(uint32_t cp, uint8_t *c) {
 // The following code is used in number parsing. It is not
 // properly "char utils" stuff, but we move it here so that
 // it does not get copied multiple times in the binaries (once
-// per instructin set).
+// per instruction set).
 ///
 
 
@@ -5771,10 +5765,26 @@ struct value128 {
   uint64_t high;
 };
 
+#if defined(_MSC_VER) && !defined(_M_X64) // _umul128 for x86, arm, arm64
+#if defined(_M_ARM)
+static inline uint64_t __emulu(uint32_t x, uint32_t y) {
+  return x * (uint64_t)y;
+}
+#endif
+static inline uint64_t _umul128(uint64_t ab, uint64_t cd, uint64_t *hi) {
+  uint64_t ad = __emulu((uint32_t)(ab >> 32), (uint32_t)cd);
+  uint64_t bd = __emulu((uint32_t)ab, (uint32_t)cd);
+  uint64_t adbc = ad + __emulu((uint32_t)ab, (uint32_t)(cd >> 32));
+  uint64_t adbc_carry = !!(adbc < ad);
+  uint64_t lo = bd + (adbc << 32);
+  *hi = __emulu((uint32_t)(ab >> 32), (uint32_t)(cd >> 32)) + (adbc >> 32) + (adbc_carry << 32) + !!(lo < bd);
+  return lo;
+}
+#endif
+
 really_inline value128 full_multiplication(uint64_t value1, uint64_t value2) {
   value128 answer;
 #ifdef _MSC_VER
-  // todo: this might fail under visual studio for ARM
   answer.low = _umul128(value1, value2, &answer.high);
 #else
   __uint128_t r = ((__uint128_t)value1) * value2;
@@ -8559,6 +8569,23 @@ WARN_UNUSED really_inline uint8_t *parse_string(const uint8_t *src, uint8_t *dst
 namespace simdjson {
 namespace fallback {
 
+#if defined(_MSC_VER) && !defined(_M_ARM64) && !defined(_M_X64)
+static inline unsigned char _BitScanForward64(unsigned long* ret, uint64_t x) {
+  unsigned long x0 = (unsigned long)x, top, bottom;
+  _BitScanForward(&top, (unsigned long)(x >> 32));
+  _BitScanForward(&bottom, x0);
+  *ret = x0 ? bottom : 32 + top;
+  return x != 0;
+}
+static unsigned char _BitScanReverse64(unsigned long* ret, uint64_t x) {
+  unsigned long x1 = (unsigned long)(x >> 32), top, bottom;
+  _BitScanReverse(&top, x1);
+  _BitScanReverse(&bottom, (unsigned long)x);
+  *ret = x1 ? top + 32 : bottom;
+  return x != 0;
+}
+#endif
+
 // We sometimes call trailing_zero on inputs that are zero,
 // but the algorithms do not end up using the returned value.
 // Sadly, sanitizers are not smart enough to figure it out.
diff --git a/src/arm64/bitmanipulation.h b/src/arm64/bitmanipulation.h
@@ -59,16 +59,10 @@ really_inline bool add_overflow(uint64_t value1, uint64_t value2, uint64_t *resu
 #endif
 }
 
-#ifdef _MSC_VER
-#pragma intrinsic(_umul128) // todo: this might fail under visual studio for ARM
-#endif
-
 really_inline bool mul_overflow(uint64_t value1, uint64_t value2, uint64_t *result) {
 #ifdef _MSC_VER
-  // todo: this might fail under visual studio for ARM
-  uint64_t high;
-  *result = _umul128(value1, value2, &high);
-  return high;
+  *result = value1 * value2;
+  return !!__umulh(value1, value2);
 #else
   return __builtin_umulll_overflow(value1, value2, (unsigned long long *)result);
 #endif
diff --git a/src/fallback/bitmanipulation.h b/src/fallback/bitmanipulation.h
@@ -7,6 +7,23 @@
 namespace simdjson {
 namespace fallback {
 
+#if defined(_MSC_VER) && !defined(_M_ARM64) && !defined(_M_X64)
+static inline unsigned char _BitScanForward64(unsigned long* ret, uint64_t x) {
+  unsigned long x0 = (unsigned long)x, top, bottom;
+  _BitScanForward(&top, (unsigned long)(x >> 32));
+  _BitScanForward(&bottom, x0);
+  *ret = x0 ? bottom : 32 + top;
+  return x != 0;
+}
+static unsigned char _BitScanReverse64(unsigned long* ret, uint64_t x) {
+  unsigned long x1 = (unsigned long)(x >> 32), top, bottom;
+  _BitScanReverse(&top, x1);
+  _BitScanReverse(&bottom, (unsigned long)x);
+  *ret = x1 ? top + 32 : bottom;
+  return x != 0;
+}
+#endif
+
 // We sometimes call trailing_zero on inputs that are zero,
 // but the algorithms do not end up using the returned value.
 // Sadly, sanitizers are not smart enough to figure it out.
diff --git a/src/jsoncharutils.h b/src/jsoncharutils.h
@@ -324,10 +324,26 @@ struct value128 {
   uint64_t high;
 };
 
+#if defined(_MSC_VER) && !defined(_M_X64) // _umul128 for x86, arm, arm64
+#if defined(_M_ARM)
+static inline uint64_t __emulu(uint32_t x, uint32_t y) {
+  return x * (uint64_t)y;
+}
+#endif
+static inline uint64_t _umul128(uint64_t ab, uint64_t cd, uint64_t *hi) {
+  uint64_t ad = __emulu((uint32_t)(ab >> 32), (uint32_t)cd);
+  uint64_t bd = __emulu((uint32_t)ab, (uint32_t)cd);
+  uint64_t adbc = ad + __emulu((uint32_t)ab, (uint32_t)(cd >> 32));
+  uint64_t adbc_carry = !!(adbc < ad);
+  uint64_t lo = bd + (adbc << 32);
+  *hi = __emulu((uint32_t)(ab >> 32), (uint32_t)(cd >> 32)) + (adbc >> 32) + (adbc_carry << 32) + !!(lo < bd);
+  return lo;
+}
+#endif
+
 really_inline value128 full_multiplication(uint64_t value1, uint64_t value2) {
   value128 answer;
 #ifdef _MSC_VER
-  // todo: this might fail under visual studio for ARM
   answer.low = _umul128(value1, value2, &answer.high);
 #else
   __uint128_t r = ((__uint128_t)value1) * value2;
