Use tzcnt instead of bsf for better performance on ZEN/ZEN2.

martinus · Jun 18, 2020 · 7eff58f · 7eff58f
1 parent b217307
commit 7eff58f
Show file tree

Hide file tree

Showing 2 changed files with 20 additions and 4 deletions.
diff --git a/src/include/robin_hood.h b/src/include/robin_hood.h
@@ -128,7 +128,23 @@ static Counts& counts() {
 #endif
 
 // count leading/trailing bits
-#ifdef _MSC_VER
+#if ( ( defined __i386 || defined __x86_64__ ) && defined __BMI__ ) || defined _M_IX86 || defined _M_X64
+#    ifdef _MSC_VER
+#        include <intrin.h>
+#    else
+#        include <x86intrin.h>
+#    endif
+#    if ROBIN_HOOD(BITNESS) == 32
+#        define ROBIN_HOOD_PRIVATE_DEFINITION_CTZ() _tzcnt_u32
+#    else
+#        define ROBIN_HOOD_PRIVATE_DEFINITION_CTZ() _tzcnt_u64
+#    endif
+#    if defined __AVX2__ || defined __BMI__
+#        define ROBIN_HOOD_COUNT_TRAILING_ZEROES(x) ROBIN_HOOD(CTZ)(x)
+#    else
+#        define ROBIN_HOOD_COUNT_TRAILING_ZEROES(x) ((x) ? ROBIN_HOOD(CTZ)(x) : ROBIN_HOOD(BITNESS))
+#    endif
+#elif defined _MSC_VER
 #    if ROBIN_HOOD(BITNESS) == 32
 #        define ROBIN_HOOD_PRIVATE_DEFINITION_BITSCANFORWARD() _BitScanForward
 #    else
@@ -1249,9 +1265,9 @@ class Table
             do {
                 auto const n = detail::unaligned_load<size_t>(mInfo);
 #if ROBIN_HOOD(LITTLE_ENDIAN)
-                inc = ROBIN_HOOD_COUNT_TRAILING_ZEROES(n) / 8;
+                inc = static_cast<int>(ROBIN_HOOD_COUNT_TRAILING_ZEROES(n)) / 8;
 #else
-                inc = ROBIN_HOOD_COUNT_LEADING_ZEROES(n) / 8;
+                inc = static_cast<int>(ROBIN_HOOD_COUNT_LEADING_ZEROES(n)) / 8;
 #endif
                 mInfo += inc;
                 mKeyVals += inc;

diff --git a/src/test/unit/unit_clz.cpp b/src/test/unit/unit_clz.cpp
@@ -8,7 +8,7 @@ namespace {
 
 int clz(void const* data) {
     auto const n = robin_hood::detail::unaligned_load<size_t>(data);
-    return ROBIN_HOOD_COUNT_TRAILING_ZEROES(n) / 8;
+    return static_cast<int>(ROBIN_HOOD_COUNT_TRAILING_ZEROES(n)) / 8;
 }
 
 } // namespace