Bug 1811184 - Move FirstNon8Bit vector implementation to xsimd r=hsivonen

xsimd does not support ppc yet, so leave the VMX implementation untouched.

Differential Revision: https://phabricator.services.mozilla.com/D167225

dom/base/moz.build

@@ -580,6 +580,7 @@ LOCAL_INCLUDES += [
     "/netwerk/url-classifier",
     "/parser/htmlparser",
     "/security/manager/ssl",
+    "/third_party/xsimd/include",
     "/widget",
     "/xpcom/ds",
 ]

dom/base/nsTextFragment.cpp

@@ -158,10 +158,8 @@ static inline int32_t FirstNon8BitUnvectorized(const char16_t* str,
   return -1;
 }
 
-#ifdef MOZILLA_MAY_SUPPORT_SSE2
-namespace mozilla::SSE2 {
-int32_t FirstNon8Bit(const char16_t* str, const char16_t* end);
-}  // namespace mozilla::SSE2
+#if defined(MOZILLA_MAY_SUPPORT_SSE2)
+#  include "nsTextFragmentGeneric.h"
 #endif
 
 #ifdef __powerpc__
@@ -182,7 +180,7 @@ int32_t FirstNon8Bit(const char16_t* str, const char16_t* end);
 static inline int32_t FirstNon8Bit(const char16_t* str, const char16_t* end) {
 #ifdef MOZILLA_MAY_SUPPORT_SSE2
   if (mozilla::supports_sse2()) {
-    return mozilla::SSE2::FirstNon8Bit(str, end);
+    return mozilla::FirstNon8Bit<xsimd::sse2>(str, end);
   }
 #elif defined(__powerpc__)
   if (mozilla::supports_vmx()) {

dom/base/nsTextFragmentGeneric.h

@@ -0,0 +1,59 @@
+/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* vim: set ts=8 sts=2 et sw=2 tw=80: */
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#include "nscore.h"
+#include "nsTextFragmentImpl.h"
+#include <algorithm>
+#include <xsimd/xsimd.hpp>
+
+namespace mozilla {
+
+template <class Arch>
+int32_t FirstNon8Bit(const char16_t* str, const char16_t* end) {
+  const uint32_t numUnicharsPerVector = xsimd::batch<int16_t, Arch>::size;
+  using p = Non8BitParameters<sizeof(size_t)>;
+  const size_t mask = p::mask();
+  const uint32_t numUnicharsPerWord = p::numUnicharsPerWord();
+  const int32_t len = end - str;
+  int32_t i = 0;
+
+  // Align ourselves to the Arch boundary
+  int32_t alignLen = std::min(
+      len, int32_t(((-NS_PTR_TO_INT32(str)) & (Arch::alignment() - 1)) /
+                   sizeof(char16_t)));
+  for (; i < alignLen; i++) {
+    if (str[i] > 255) return i;
+  }
+
+  // Check one batch at a time.
+  const int32_t vectWalkEnd =
+      ((len - i) / numUnicharsPerVector) * numUnicharsPerVector;
+  const uint16_t shortMask = 0xff00;
+  xsimd::batch<int16_t, Arch> vectmask(static_cast<int16_t>(shortMask));
+  for (; i < vectWalkEnd; i += numUnicharsPerVector) {
+    const auto vect = xsimd::batch<int16_t, Arch>::load_aligned(str + i);
+    if (xsimd::any((vect & vectmask) != 0)) return i;
+  }
+
+  // Check one word at a time.
+  const int32_t wordWalkEnd =
+      ((len - i) / numUnicharsPerWord) * numUnicharsPerWord;
+  for (; i < wordWalkEnd; i += numUnicharsPerWord) {
+    const size_t word = *reinterpret_cast<const size_t*>(str + i);
+    if (word & mask) return i;
+  }
+
+  // Take care of the remainder one character at a time.
+  for (; i < len; i++) {
+    if (str[i] > 255) {
+      return i;
+    }
+  }
+
+  return -1;
+}
+
+}  // namespace mozilla

dom/base/nsTextFragmentSSE2.cpp

@@ -1,65 +1,10 @@
-/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
-/* vim: set ts=8 sts=2 et sw=2 tw=80: */
-/* This Source Code Form is subject to the terms of the Mozilla Public
- * License, v. 2.0. If a copy of the MPL was not distributed with this
- * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+/* -*- mode: c++; tab-width: 2; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
+/* this source code form is subject to the terms of the mozilla public
+ * license, v. 2.0. if a copy of the mpl was not distributed with this file,
+ * You can obtain one at http://mozilla.org/MPL/2.0/. */
 
-// This file should only be compiled if you're on x86 or x86_64.  Additionally,
-// you'll need to compile this file with -msse2 if you're using gcc.
+#include "nsTextFragmentGeneric.h"
 
-#include <emmintrin.h>
-#include "nscore.h"
-#include "nsTextFragmentImpl.h"
-#include <algorithm>
-
-namespace mozilla::SSE2 {
-
-static inline bool is_zero(__m128i x) {
-  return _mm_movemask_epi8(_mm_cmpeq_epi8(x, _mm_setzero_si128())) == 0xffff;
-}
-
-int32_t FirstNon8Bit(const char16_t* str, const char16_t* end) {
-  const uint32_t numUnicharsPerVector = 8;
-  using p = Non8BitParameters<sizeof(size_t)>;
-  const size_t mask = p::mask();
-  const uint32_t numUnicharsPerWord = p::numUnicharsPerWord();
-  const int32_t len = end - str;
-  int32_t i = 0;
-
-  // Align ourselves to a 16-byte boundary, as required by _mm_load_si128
-  // (i.e. MOVDQA).
-  int32_t alignLen = std::min(
-      len, int32_t(((-NS_PTR_TO_INT32(str)) & 0xf) / sizeof(char16_t)));
-  for (; i < alignLen; i++) {
-    if (str[i] > 255) return i;
-  }
-
-  // Check one XMM register (16 bytes) at a time.
-  const int32_t vectWalkEnd =
-      ((len - i) / numUnicharsPerVector) * numUnicharsPerVector;
-  const uint16_t shortMask = 0xff00;
-  __m128i vectmask = _mm_set1_epi16(static_cast<int16_t>(shortMask));
-  for (; i < vectWalkEnd; i += numUnicharsPerVector) {
-    const __m128i vect = *reinterpret_cast<const __m128i*>(str + i);
-    if (!is_zero(_mm_and_si128(vect, vectmask))) return i;
-  }
-
-  // Check one word at a time.
-  const int32_t wordWalkEnd =
-      ((len - i) / numUnicharsPerWord) * numUnicharsPerWord;
-  for (; i < wordWalkEnd; i += numUnicharsPerWord) {
-    const size_t word = *reinterpret_cast<const size_t*>(str + i);
-    if (word & mask) return i;
-  }
-
-  // Take care of the remainder one character at a time.
-  for (; i < len; i++) {
-    if (str[i] > 255) {
-      return i;
-    }
-  }
-
-  return -1;
-}
-
-}  // namespace mozilla::SSE2
+namespace mozilla {
+template int32_t FirstNon8Bit<xsimd::sse2>(const char16_t*, const char16_t*);
+}  // namespace mozilla