folly::simd::contains, all types. (#2306)

Summary:
Pull Request resolved: #2306

extended the SimdCharPlatform => SimdPlatform to support all unsigned types.

Reviewed By: Gownta

Differential Revision: D63548781

fbshipit-source-id: 6599a7bd8686955e28df4f83aeb45f43a31b855e

folly/algorithm/simd/Movemask.h

@@ -91,22 +91,7 @@ struct movemask_fn {
   auto operator()(Reg reg) const;
 
   template <typename Reg, typename Ignore>
-  FOLLY_ERASE auto operator()(Reg reg, Ignore ignore) const {
-    auto [bits, bitsPerElement] = operator()(reg);
-
-    if constexpr (std::is_same_v<Ignore, ignore_none>) {
-      return std::pair{bits, bitsPerElement};
-    } else {
-      static constexpr int kCardinal = sizeof(Reg) / sizeof(Scalar);
-
-      int bitsToKeep = (kCardinal - ignore.last) * bitsPerElement;
-
-      bits =
-          clear_n_least_significant_bits(bits, ignore.first * bitsPerElement);
-      bits = clear_n_most_significant_bits(bits, sizeof(bits) * 8 - bitsToKeep);
-      return std::pair{bits, bitsPerElement};
-    }
-  }
+  auto operator()(Reg reg, Ignore ignore) const;
 };
 
 template <typename Scalar>
@@ -116,7 +101,7 @@ inline constexpr movemask_fn<Scalar> movemask;
 
 template <typename Scalar>
 template <typename Reg>
-auto movemask_fn<Scalar>::operator()(Reg reg) const {
+FOLLY_ERASE auto movemask_fn<Scalar>::operator()(Reg reg) const {
   std::integral_constant<std::uint32_t, sizeof(Scalar) == 2 ? 2 : 1>
       bitsPerElement;
 
@@ -154,7 +139,7 @@ auto movemask_fn<Scalar>::operator()(Reg reg) const {
 
 namespace detail {
 
-inline auto movemaskChars16Aarch64(uint8x16_t reg) {
+FOLLY_ERASE auto movemaskChars16Aarch64(uint8x16_t reg) {
   uint16x8_t u16s = vreinterpretq_u16_u8(reg);
   u16s = vshrq_n_u16(u16s, 4);
   uint8x8_t packed = vmovn_u16(u16s);
@@ -163,7 +148,7 @@ inline auto movemaskChars16Aarch64(uint8x16_t reg) {
 }
 
 template <typename Reg>
-uint64x1_t asUint64x1Aarch64(Reg reg) {
+FOLLY_ERASE uint64x1_t asUint64x1Aarch64(Reg reg) {
   if constexpr (std::is_same_v<Reg, uint32x2_t>) {
     return vreinterpret_u64_u32(reg);
   } else if constexpr (std::is_same_v<Reg, uint16x4_t>) {
@@ -177,7 +162,7 @@ uint64x1_t asUint64x1Aarch64(Reg reg) {
 
 template <typename Scalar>
 template <typename Reg>
-auto movemask_fn<Scalar>::operator()(Reg reg) const {
+FOLLY_ERASE auto movemask_fn<Scalar>::operator()(Reg reg) const {
   if constexpr (std::is_same_v<Reg, uint64x2_t>) {
     return movemask<std::uint32_t>(vmovn_u64(reg));
   } else if constexpr (std::is_same_v<Reg, uint32x4_t>) {
@@ -195,6 +180,28 @@ auto movemask_fn<Scalar>::operator()(Reg reg) const {
 
 #endif
 
+#if FOLLY_X64 || FOLLY_AARCH64
+
+template <typename Scalar>
+template <typename Reg, typename Ignore>
+FOLLY_ERASE auto movemask_fn<Scalar>::operator()(Reg reg, Ignore ignore) const {
+  auto [bits, bitsPerElement] = operator()(reg);
+
+  if constexpr (std::is_same_v<Ignore, ignore_none>) {
+    return std::pair{bits, bitsPerElement};
+  } else {
+    static constexpr int kCardinal = sizeof(Reg) / sizeof(Scalar);
+
+    int bitsToKeep = (kCardinal - ignore.last) * bitsPerElement;
+
+    bits = clear_n_least_significant_bits(bits, ignore.first * bitsPerElement);
+    bits = clear_n_most_significant_bits(bits, sizeof(bits) * 8 - bitsToKeep);
+    return std::pair{bits, bitsPerElement};
+  }
+}
+
+#endif
+
 } // namespace folly::simd
 
 FOLLY_POP_WARNING

folly/algorithm/simd/detail/BUCK

@@ -16,8 +16,8 @@ cpp_library(
 )
 
 cpp_library(
-    name = "simd_char_platform",
-    headers = ["SimdCharPlatform.h"],
+    name = "simd_platform",
+    headers = ["SimdPlatform.h"],
     exported_deps = [
         "//folly:portability",
         "//folly/algorithm/simd:ignore",
@@ -31,7 +31,7 @@ cpp_library(
     headers = ["ContainsImpl.h"],
     exported_deps = [
         ":simd_any_of",
-        ":simd_char_platform",
+        ":simd_platform",
         "//folly:c_portability",
         "//folly/container:span",
     ],
@@ -45,6 +45,7 @@ cpp_library(
         "//folly:c_portability",
         "//folly:traits",
         "//folly/algorithm/simd:ignore",
+        "//folly/lang:align",
     ],
 )
 

folly/algorithm/simd/detail/ContainsImpl.h

@@ -23,7 +23,7 @@
 
 #include <folly/CPortability.h>
 #include <folly/algorithm/simd/detail/SimdAnyOf.h>
-#include <folly/algorithm/simd/detail/SimdCharPlatform.h>
+#include <folly/algorithm/simd/detail/SimdPlatform.h>
 #include <folly/container/span.h>
 
 namespace folly::simd::detail {
@@ -62,20 +62,20 @@ FOLLY_ERASE bool containsImplStd(folly::span<const T> haystack, T needle) {
 
 template <typename T>
 constexpr bool hasHandwrittenContains() {
-  return std::is_same_v<T, std::uint8_t> &&
-      !std::is_same_v<SimdCharPlatform, void>;
+  return !std::is_same_v<SimdPlatform<T>, void> &&
+      (std::is_same_v<std::uint8_t, T> || std::is_same_v<std::uint16_t, T> ||
+       std::is_same_v<std::uint32_t, T> || std::is_same_v<std::uint64_t, T>);
 }
 
-template <typename T>
+template <typename T, typename Platform = SimdPlatform<T>>
 FOLLY_ERASE bool containsImplHandwritten(
     folly::span<const T> haystack, T needle) {
-  static_assert(std::is_same_v<T, std::uint8_t>, "");
-  auto as_chars = folly::reinterpret_span_cast<const char>(haystack);
-  return simdAnyOf<SimdCharPlatform, 4>(
-      as_chars.data(),
-      as_chars.data() + as_chars.size(),
-      [&](SimdCharPlatform::reg_t x) {
-        return SimdCharPlatform::equal(x, static_cast<char>(needle));
+  static_assert(!std::is_same_v<Platform, void>, "");
+  return simdAnyOf<Platform, 4>(
+      haystack.data(),
+      haystack.data() + haystack.size(),
+      [&](typename Platform::reg_t x) {
+        return Platform::equal(x, static_cast<T>(needle));
       });
 }