Bug 1967062 - Partial Backout of "Bug 933149 - Remove PodCopy loop heuristic r=mstange" on Linux only. a=dmeehan

We'll remove the loop heuristic on non-linux platforms, but keep it there as
that's where it continues to be effective.

Differential Revision: https://phabricator.services.mozilla.com/D250470

js/src/builtin/String.cpp

@@ -8,6 +8,7 @@
 
 #include "mozilla/Attributes.h"
 #include "mozilla/CheckedInt.h"
+#include "mozilla/Compiler.h"
 #include "mozilla/FloatingPoint.h"
 #if JS_HAS_INTL_API
 #  include "mozilla/intl/String.h"
@@ -544,7 +545,12 @@ template <typename DestChar, typename SrcChar>
 static inline void CopyChars(DestChar* destChars, const SrcChar* srcChars,
                              size_t length) {
   if constexpr (std::is_same_v<DestChar, SrcChar>) {
+#if MOZ_IS_GCC
+    // Directly call memcpy to work around bug 1863131.
+    memcpy(destChars, srcChars, length * sizeof(DestChar));
+#else
     PodCopy(destChars, srcChars, length);
+#endif
   } else {
     for (size_t i = 0; i < length; i++) {
       destChars[i] = srcChars[i];

mfbt/PodOperations.h

@@ -96,6 +96,19 @@ static MOZ_ALWAYS_INLINE void PodCopy(T* aDst, const T* aSrc, size_t aNElem) {
   MOZ_ASSERT(aDst + aNElem <= aSrc || aSrc + aNElem <= aDst,
              "destination and source must not overlap");
 
+// Linux memcpy for small sizes seems slower than on other
+// platforms. So we use a loop for small sizes there only.
+//
+// See Bug 1967062 for details.
+#if defined(XP_LINUX)
+  if (aNElem < 128) {
+    for (const T* srcend = aSrc + aNElem; aSrc < srcend; aSrc++, aDst++) {
+      *aDst = *aSrc;
+    }
+    return;
+  }
+#endif
+
   memcpy(aDst, aSrc, aNElem * sizeof(T));
 }