Optimize WTF::Bitmap::forEachSetBit to use hardware intrinsics when a…

…vailable

https://bugs.webkit.org/show_bug.cgi?id=254959
rdar://107584830

Reviewed by Justin Michaud.

Changes WTF::Bitmap::forEachSetBit to use the ctz() intrinsic when it's available
and implemented in hardware. This allows the method to skip up to a word at a time
of zero bits, instead of having to visit each one, improving iteration times
especially for sparse bitmaps.

* Source/WTF/wtf/Bitmap.h:
(WTF::WordType>::forEachSetBit const):

Canonical link: https://commits.webkit.org/262602@main

Source/WTF/wtf/Bitmap.h

@@ -409,6 +409,21 @@ ALWAYS_INLINE constexpr void Bitmap<bitmapSize, WordType>::forEachSetBit(const F
         if (!word)
             continue;
         size_t base = i * wordSize;
+
+#if COMPILER(GCC_COMPATIBLE) && (CPU(X86_64) || CPU(ARM64))
+        // We should only use ctz() when we know that ctz() is implementated using
+        // a fast hardware instruction. Otherwise, this will actually result in
+        // worse performance.
+        while (word) {
+            size_t offset = ctz(word);
+            if constexpr (std::is_same_v<IterationStatus, decltype(func(base + offset))>) {
+                if (func(base + offset) == IterationStatus::Done)
+                    return;
+            } else
+                func(base + offset);
+            word &= ~(1ull << offset);
+        }
+#else
         for (size_t j = 0; j < wordSize; ++j) {
             if (word & 1) {
                 if constexpr (std::is_same_v<IterationStatus, decltype(func(base + j))>) {
@@ -419,6 +434,7 @@ ALWAYS_INLINE constexpr void Bitmap<bitmapSize, WordType>::forEachSetBit(const F
             }
             word >>= 1;
         }
+#endif
     }
 }