Polyfill GroupSse2Impl on non-x86 platforms

Bug: v8:12518
Change-Id: Ie22303416749affc0629d60fbed6f9dc4288b09d
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/3494443Reviewed-by: Igor Sheludko <ishell@chromium.org>
Commit-Queue: David Benjamin <davidben@chromium.org>
Cr-Commit-Position: refs/heads/main@{#79834}

src/objects/swiss-hash-table-helpers.h

@@ -17,51 +17,50 @@
 #ifndef V8_OBJECTS_SWISS_HASH_TABLE_HELPERS_H_
 #define V8_OBJECTS_SWISS_HASH_TABLE_HELPERS_H_
 
-// The following #defines are taken from Abseil's have_sse.h (but renamed). They
-// are only defined within this file. However, we also take cross platform
-// snapshot creation into account, by only using SSE if the target supports it,
-// too. The SSE implementation uses a group width of 16, whereas the non-SSE
-// version uses 8. We therefore have to avoid building a snapshot that contains
-// Swiss Tables with one group size and use it in code that excepts a different
-// group size.
-#ifndef SWISS_TABLE_HAVE_SSE2
-#if (defined(__SSE2__) ||                                             \
-     (defined(_MSC_VER) &&                                            \
-      (defined(_M_X64) || (defined(_M_IX86) && _M_IX86_FP >= 2)))) && \
-    (defined(V8_TARGET_ARCH_IA32) || defined(V8_TARGET_ARCH_X64))
-#define SWISS_TABLE_HAVE_SSE2 1
+// The following #defines are taken from Abseil's have_sse.h (but renamed).
+#ifndef V8_SWISS_TABLE_HAVE_SSE2_HOST
+#if (defined(__SSE2__) ||  \
+     (defined(_MSC_VER) && \
+      (defined(_M_X64) || (defined(_M_IX86) && _M_IX86_FP >= 2))))
+#define V8_SWISS_TABLE_HAVE_SSE2_HOST 1
 #else
-#define SWISS_TABLE_HAVE_SSE2 0
-#if defined(V8_TARGET_ARCH_IA32) || defined(V8_TARGET_ARCH_X64)
-// TODO(v8:11388) Currently, building on a non-SSE platform for a SSE target
-// means that we cannot use the (more performant) SSE implementations of Swiss
-// Tables, even if the target would support it, just because the host doesn't.
-// This is due to the difference in group sizes (see comment at the beginning of
-// the file). We can solve this by implementating a new non-SSE Group that
-// behaves like GroupSse2Impl (and uses group size 16) in the future.
-#warning "You should avoid building on a non-SSE platform for a SSE target!"
-#endif
+#define V8_SWISS_TABLE_HAVE_SSE2_HOST 0
 #endif
 #endif
 
-#ifndef SWISS_TABLE_HAVE_SSSE3
-#if defined(__SSSE3__) && \
-    (defined(V8_TARGET_ARCH_IA32) || defined(V8_TARGET_ARCH_X64))
-#define SWISS_TABLE_HAVE_SSSE3 1
+#ifndef V8_SWISS_TABLE_HAVE_SSSE3_HOST
+#if defined(__SSSE3__)
+#define V8_SWISS_TABLE_HAVE_SSSE3_HOST 1
 #else
-#define SWISS_TABLE_HAVE_SSSE3 0
+#define V8_SWISS_TABLE_HAVE_SSSE3_HOST 0
 #endif
 #endif
 
-#if SWISS_TABLE_HAVE_SSSE3 && !SWISS_TABLE_HAVE_SSE2
+#if V8_SWISS_TABLE_HAVE_SSSE3_HOST && !V8_SWISS_TABLE_HAVE_SSE2_HOST
 #error "Bad configuration!"
 #endif
 
-#if SWISS_TABLE_HAVE_SSE2
+// Unlike Abseil, we cannot select SSE purely by host capabilities. When
+// creating a snapshot, the group width must be compatible. The SSE
+// implementation uses a group width of 16, whereas the non-SSE version uses 8.
+// Thus we select the group size based on target capabilities and, if the host
+// does not match, select a polyfill implementation. This means, in supported
+// cross-compiling configurations, we must be able to determine matching target
+// capabilities from the host.
+#ifndef V8_SWISS_TABLE_HAVE_SSE2_TARGET
+#if V8_TARGET_ARCH_IA32 || V8_TARGET_ARCH_X64
+// x64 always has SSE2, and ia32 without SSE2 is not supported by V8.
+#define V8_SWISS_TABLE_HAVE_SSE2_TARGET 1
+#else
+#define V8_SWISS_TABLE_HAVE_SSE2_TARGET 0
+#endif
+#endif
+
+#if V8_SWISS_TABLE_HAVE_SSE2_HOST
 #include <emmintrin.h>
 #endif
 
-#if SWISS_TABLE_HAVE_SSSE3
+#if V8_SWISS_TABLE_HAVE_SSSE3_HOST
 #include <tmmintrin.h>
 #endif
 
@@ -126,9 +125,9 @@ class BitMask {
 
  public:
   // These are useful for unit tests (gunit).
-  // using value_type = int;
-  // using iterator = BitMask;
-  // using const_iterator = BitMask;
+  using value_type = int;
+  using iterator = BitMask;
+  using const_iterator = BitMask;
 
   explicit BitMask(T mask) : mask_(mask) {}
   BitMask& operator++() {
@@ -219,7 +218,7 @@ inline static swiss_table::ctrl_t H2(uint32_t hash) {
   return hash & ((1 << kH2Bits) - 1);
 }
 
-#if SWISS_TABLE_HAVE_SSE2
+#if V8_SWISS_TABLE_HAVE_SSE2_HOST
 // https://github.com/abseil/abseil-cpp/issues/209
 // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87853
 // _mm_cmpgt_epi8 is broken under GCC with -funsigned-char
@@ -252,7 +251,7 @@ struct GroupSse2Impl {
 
   // Returns a bitmask representing the positions of empty slots.
   BitMask<uint32_t, kWidth> MatchEmpty() const {
-#if SWISS_TABLE_HAVE_SSSE3
+#if V8_SWISS_TABLE_HAVE_SSSE3_HOST
     // This only works because kEmpty is -128.
     return BitMask<uint32_t, kWidth>(
         _mm_movemask_epi8(_mm_sign_epi8(ctrl, ctrl)));
@@ -278,7 +277,7 @@ struct GroupSse2Impl {
   void ConvertSpecialToEmptyAndFullToDeleted(ctrl_t* dst) const {
     auto msbs = _mm_set1_epi8(static_cast<char>(-128));
     auto x126 = _mm_set1_epi8(126);
-#if SWISS_TABLE_HAVE_SSSE3
+#if V8_SWISS_TABLE_HAVE_SSSE3_HOST
     auto res = _mm_or_si128(_mm_shuffle_epi8(x126, ctrl), msbs);
 #else
     auto zero = _mm_setzero_si128();
@@ -290,7 +289,64 @@ struct GroupSse2Impl {
 
   __m128i ctrl;
 };
-#endif  // SWISS_TABLE_HAVE_SSE2
+#endif  // V8_SWISS_TABLE_HAVE_SSE2_HOST
+
+// A portable, inefficient version of GroupSse2Impl. This exists so SSE2-less
+// hosts can generate snapshots for SSE2-capable targets.
+struct GroupSse2Polyfill {
+  static constexpr size_t kWidth = 16;  // the number of slots per group
+
+  explicit GroupSse2Polyfill(const ctrl_t* pos) { memcpy(ctrl_, pos, kWidth); }
+
+  // Returns a bitmask representing the positions of slots that match |hash|.
+  BitMask<uint32_t, kWidth> Match(h2_t hash) const {
+    uint32_t mask = 0;
+    for (size_t i = 0; i < kWidth; i++) {
+      if (static_cast<h2_t>(ctrl_[i]) == hash) {
+        mask |= 1u << i;
+      }
+    }
+    return BitMask<uint32_t, kWidth>(mask);
+  }
+
+  // Returns a bitmask representing the positions of empty slots.
+  BitMask<uint32_t, kWidth> MatchEmpty() const {
+    return Match(static_cast<h2_t>(kEmpty));
+  }
+
+  // Returns a bitmask representing the positions of empty or deleted slots.
+  BitMask<uint32_t, kWidth> MatchEmptyOrDeleted() const {
+    return BitMask<uint32_t, kWidth>(MatchEmptyOrDeletedMask());
+  }
+
+  // Returns the number of trailing empty or deleted elements in the group.
+  uint32_t CountLeadingEmptyOrDeleted() const {
+    return base::bits::CountTrailingZerosNonZero(MatchEmptyOrDeletedMask() + 1);
+  }
+
+  void ConvertSpecialToEmptyAndFullToDeleted(ctrl_t* dst) const {
+    for (size_t i = 0; i < kWidth; i++) {
+      if (ctrl_[i] < 0) {
+        dst[i] = kEmpty;
+      } else {
+        dst[i] = kDeleted;
+      }
+    }
+  }
+
+ private:
+  uint32_t MatchEmptyOrDeletedMask() const {
+    uint32_t mask = 0;
+    for (size_t i = 0; i < kWidth; i++) {
+      if (ctrl_[i] < kSentinel) {
+        mask |= 1u << i;
+      }
+    }
+    return mask;
+  }
+
+  ctrl_t ctrl_[kWidth];
+};
 
 struct GroupPortableImpl {
   static constexpr size_t kWidth = 8;  // the number of slots per group
@@ -366,16 +422,23 @@ struct GroupPortableImpl {
 // backend should only use SSE2 when compiling the SIMD version of
 // SwissNameDictionary into the builtin.
 using Group = GroupPortableImpl;
-#else
-#if SWISS_TABLE_HAVE_SSE2
+#elif V8_SWISS_TABLE_HAVE_SSE2_TARGET
+// Use a matching group size between host and target.
+#if V8_SWISS_TABLE_HAVE_SSE2_HOST
 using Group = GroupSse2Impl;
 #else
-using Group = GroupPortableImpl;
+#if V8_HOST_ARCH_IA32 || V8_HOST_ARCH_X64
+// If we do not detect SSE2 when building for the ia32/x64 target, the
+// V8_SWISS_TABLE_HAVE_SSE2_TARGET logic will incorrectly cause the final output
+// to use the inefficient polyfill implementation. Detect this case and warn if
+// it happens.
+#warning "Did not detect required SSE2 support on ia32/x64."
 #endif
+using Group = GroupSse2Polyfill;
+#endif
+#else
+using Group = GroupPortableImpl;
 #endif
-
-#undef SWISS_TABLE_HAVE_SSE2
-#undef SWISS_TABLE_HAVE_SSE3
 
 }  // namespace swiss_table
 }  // namespace internal

test/unittests/BUILD.gn

@@ -368,6 +368,7 @@ v8_source_set("unittests_sources") {
     "numbers/conversions-unittest.cc",
     "objects/object-unittest.cc",
     "objects/osr-optimized-code-cache-unittest.cc",
+    "objects/swiss-hash-table-helpers-unittest.cc",
     "objects/value-serializer-unittest.cc",
     "objects/weakarraylist-unittest.cc",
     "parser/ast-value-unittest.cc",

test/unittests/objects/swiss-hash-table-helpers-unittest.cc

+// Copyright 2022 the V8 project authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+#include "src/objects/swiss-hash-table-helpers.h"
+
+#include <vector>
+
+#include "testing/gmock/include/gmock/gmock-matchers.h"
+#include "testing/gtest/include/gtest/gtest.h"
+
+using testing::ElementsAre;
+
+namespace v8 {
+namespace internal {
+namespace swiss_table {
+
+template <typename T>
+class SwissTableGroupTest : public testing::Test {};
+
+using GroupTypes = testing::Types<
+#if V8_SWISS_TABLE_HAVE_SSE2_HOST
+    GroupSse2Impl,
+#endif
+    GroupSse2Polyfill, GroupPortableImpl>;
+TYPED_TEST_SUITE(SwissTableGroupTest, GroupTypes);
+
+// Tests imported from Abseil's raw_hash_set_test.cc, modified to be
+// parameterized.
+
+TYPED_TEST(SwissTableGroupTest, EmptyGroup) {
+  const ctrl_t kEmptyGroup[16] = {
+      kSentinel, kEmpty, kEmpty, kEmpty, kEmpty, kEmpty, kEmpty, kEmpty,
+      kEmpty,    kEmpty, kEmpty, kEmpty, kEmpty, kEmpty, kEmpty, kEmpty,
+  };
+  for (h2_t h = 0; h != 128; ++h) EXPECT_FALSE(TypeParam{kEmptyGroup}.Match(h));
+}
+
+TYPED_TEST(SwissTableGroupTest, Match) {
+  if (TypeParam::kWidth == 16) {
+    ctrl_t group[] = {kEmpty, 1, kDeleted, 3, kEmpty, 5, kSentinel, 7,
+                      7,      5, 3,        1, 1,      1, 1,         1};
+    EXPECT_THAT(TypeParam{group}.Match(0), ElementsAre());
+    EXPECT_THAT(TypeParam{group}.Match(1), ElementsAre(1, 11, 12, 13, 14, 15));
+    EXPECT_THAT(TypeParam{group}.Match(3), ElementsAre(3, 10));
+    EXPECT_THAT(TypeParam{group}.Match(5), ElementsAre(5, 9));
+    EXPECT_THAT(TypeParam{group}.Match(7), ElementsAre(7, 8));
+  } else if (TypeParam::kWidth == 8) {
+    ctrl_t group[] = {kEmpty, 1, 2, kDeleted, 2, 1, kSentinel, 1};
+    EXPECT_THAT(TypeParam{group}.Match(0), ElementsAre());
+    EXPECT_THAT(TypeParam{group}.Match(1), ElementsAre(1, 5, 7));
+    EXPECT_THAT(TypeParam{group}.Match(2), ElementsAre(2, 4));
+  } else {
+    FAIL() << "No test coverage for kWidth==" << TypeParam::kWidth;
+  }
+}
+
+TYPED_TEST(SwissTableGroupTest, MatchEmpty) {
+  if (TypeParam::kWidth == 16) {
+    ctrl_t group[] = {kEmpty, 1, kDeleted, 3, kEmpty, 5, kSentinel, 7,
+                      7,      5, 3,        1, 1,      1, 1,         1};
+    EXPECT_THAT(TypeParam{group}.MatchEmpty(), ElementsAre(0, 4));
+  } else if (TypeParam::kWidth == 8) {
+    ctrl_t group[] = {kEmpty, 1, 2, kDeleted, 2, 1, kSentinel, 1};
+    EXPECT_THAT(TypeParam{group}.MatchEmpty(), ElementsAre(0));
+  } else {
+    FAIL() << "No test coverage for kWidth==" << TypeParam::kWidth;
+  }
+}
+
+TYPED_TEST(SwissTableGroupTest, MatchEmptyOrDeleted) {
+  if (TypeParam::kWidth == 16) {
+    ctrl_t group[] = {kEmpty, 1, kDeleted, 3, kEmpty, 5, kSentinel, 7,
+                      7,      5, 3,        1, 1,      1, 1,         1};
+    EXPECT_THAT(TypeParam{group}.MatchEmptyOrDeleted(), ElementsAre(0, 2, 4));
+  } else if (TypeParam::kWidth == 8) {
+    ctrl_t group[] = {kEmpty, 1, 2, kDeleted, 2, 1, kSentinel, 1};
+    EXPECT_THAT(TypeParam{group}.MatchEmptyOrDeleted(), ElementsAre(0, 3));
+  } else {
+    FAIL() << "No test coverage for kWidth==" << TypeParam::kWidth;
+  }
+}
+
+TYPED_TEST(SwissTableGroupTest, CountLeadingEmptyOrDeleted) {
+  const std::vector<ctrl_t> empty_examples = {kEmpty, kDeleted};
+  const std::vector<ctrl_t> full_examples = {0, 1, 2, 3, 5, 9, 127, kSentinel};
+
+  for (ctrl_t empty : empty_examples) {
+    std::vector<ctrl_t> e(TypeParam::kWidth, empty);
+    EXPECT_EQ(TypeParam::kWidth,
+              TypeParam{e.data()}.CountLeadingEmptyOrDeleted());
+    for (ctrl_t full : full_examples) {
+      for (size_t i = 0; i != TypeParam::kWidth; ++i) {
+        std::vector<ctrl_t> f(TypeParam::kWidth, empty);
+        f[i] = full;
+        EXPECT_EQ(i, TypeParam{f.data()}.CountLeadingEmptyOrDeleted());
+      }
+      std::vector<ctrl_t> f(TypeParam::kWidth, empty);
+      f[TypeParam::kWidth * 2 / 3] = full;
+      f[TypeParam::kWidth / 2] = full;
+      EXPECT_EQ(TypeParam::kWidth / 2,
+                TypeParam{f.data()}.CountLeadingEmptyOrDeleted());
+    }
+  }
+}
+
+}  // namespace swiss_table
+}  // namespace internal
+}  // namespace v8