Merge pull request #9374 from MerryMage/clz

BitUtils: Add CountLeadingZeros

Source/Core/Common/Arm64Emitter.cpp

@@ -26,20 +26,6 @@ namespace
 const int kWRegSizeInBits = 32;
 const int kXRegSizeInBits = 64;
 
-// The below few functions are taken from V8.
-int CountLeadingZeros(uint64_t value, int width)
-{
-  // TODO(jbramley): Optimize this for ARM64 hosts.
-  int count = 0;
-  uint64_t bit_test = 1ULL << (width - 1);
-  while ((count < width) && ((bit_test & value) == 0))
-  {
-    count++;
-    bit_test >>= 1;
-  }
-  return count;
-}
-
 uint64_t LargestPowerOf2Divisor(uint64_t value)
 {
   return value & -(int64_t)value;
@@ -155,8 +141,8 @@ bool IsImmLogical(uint64_t value, unsigned int width, unsigned int* n, unsigned
     // Compute the repeat distance d, and set up a bitmask covering the basic
     // unit of repetition (i.e. a word with the bottom d bits set). Also, in all
     // of these cases the N bit of the output will be zero.
-    clz_a = CountLeadingZeros(a, kXRegSizeInBits);
+    clz_a = Common::CountLeadingZeros(a);
-    int clz_c = CountLeadingZeros(c, kXRegSizeInBits);
+    int clz_c = Common::CountLeadingZeros(c);
     d = clz_a - clz_c;
     mask = ((UINT64_C(1) << d) - 1);
     out_n = 0;
@@ -182,7 +168,7 @@ bool IsImmLogical(uint64_t value, unsigned int width, unsigned int* n, unsigned
       // of set bits in our word, meaning that we have the trivial case of
       // d == 64 and only one 'repetition'. Set up all the same variables as in
       // the general case above, and set the N bit in the output.
-      clz_a = CountLeadingZeros(a, kXRegSizeInBits);
+      clz_a = Common::CountLeadingZeros(a);
       d = 64;
       mask = ~UINT64_C(0);
       out_n = 1;
@@ -214,7 +200,7 @@ bool IsImmLogical(uint64_t value, unsigned int width, unsigned int* n, unsigned
       0x5555555555555555UL,
   }};
 
-  int multiplier_idx = CountLeadingZeros(d, kXRegSizeInBits) - 57;
+  int multiplier_idx = Common::CountLeadingZeros((u64)d) - 57;
 
   // Ensure that the index to the multipliers array is within bounds.
   DEBUG_ASSERT((multiplier_idx >= 0) && (static_cast<size_t>(multiplier_idx) < multipliers.size()));
@@ -233,7 +219,7 @@ bool IsImmLogical(uint64_t value, unsigned int width, unsigned int* n, unsigned
   // Count the set bits in our basic stretch. The special case of clz(0) == -1
   // makes the answer come out right for stretches that reach the very top of
   // the word (e.g. numbers like 0xffffc00000000000).
-  int clz_b = (b == 0) ? -1 : CountLeadingZeros(b, kXRegSizeInBits);
+  int clz_b = (b == 0) ? -1 : Common::CountLeadingZeros(b);
   int s = clz_a - clz_b;
 
   // Decide how many bits to rotate right by, to put the low bit of that basic

Source/Core/Common/BitUtils.h

@@ -11,6 +11,10 @@
 #include <initializer_list>
 #include <type_traits>
 
+#ifdef _MSC_VER
+#include <intrin.h>
+#endif
+
 namespace Common
 {
 ///
@@ -357,4 +361,44 @@ T ExpandValue(T value, size_t left_shift_amount)
          (T(-ExtractBit<0>(value)) >> (BitSize<T>() - left_shift_amount));
 }
 
+constexpr int CountLeadingZeros(uint64_t value)
+{
+#if defined(__GNUC__)
+  return __builtin_clzll(value);
+#elif defined(_MSC_VER) && defined(_M_ARM_64)
+  return _CountLeadingZeros64(value);
+#elif defined(_MSC_VER) && defined(_M_X86_64)
+  unsigned long index;
+  return _BitScanReverse64(&index, value) ? 63 - index : 64;
+#else
+  int result = 64;
+  while (value)
+  {
+    result--;
+    value >>= 1;
+  }
+  return result;
+#endif
+}
+
+constexpr int CountLeadingZeros(uint32_t value)
+{
+#if defined(__GNUC__)
+  return __builtin_clz(value);
+#elif defined(_MSC_VER) && defined(_M_ARM_64)
+  return _CountLeadingZeros(value);
+#elif defined(_MSC_VER) && defined(_M_X86_64)
+  unsigned long index;
+  return _BitScanReverse(&index, value) ? 31 - index : 32;
+#else
+  int result = 32;
+  while (value)
+  {
+    result--;
+    value >>= 1;
+  }
+  return result;
+#endif
+}
+
 }  // namespace Common

Source/Core/Common/MathUtil.h

@@ -9,12 +9,9 @@
 #include <type_traits>
 #include <vector>
 
+#include "Common/BitUtils.h"
 #include "Common/CommonTypes.h"
 
-#ifdef _MSC_VER
-#include <intrin.h>
-#endif
-
 namespace MathUtil
 {
 constexpr double TAU = 6.2831853071795865;
@@ -154,21 +151,5 @@ float MathFloatVectorSum(const std::vector<float>&);
 // Rounds down. 0 -> undefined
 inline int IntLog2(u64 val)
 {
-#if defined(__GNUC__)
+  return 63 - Common::CountLeadingZeros(val);
-  return 63 - __builtin_clzll(val);
-
-#elif defined(_MSC_VER)
-  unsigned long result = ULONG_MAX;
-  _BitScanReverse64(&result, val);
-  return result;
-
-#else
-  int result = -1;
-  while (val != 0)
-  {
-    val >>= 1;
-    ++result;
-  }
-  return result;
-#endif
 }

Source/Core/Core/PowerPC/Interpreter/Interpreter_Integer.cpp

@@ -236,17 +236,7 @@ void Interpreter::cmpl(UGeckoInstruction inst)
 
 void Interpreter::cntlzwx(UGeckoInstruction inst)
 {
-  u32 val = rGPR[inst.RS];
+  rGPR[inst.RA] = Common::CountLeadingZeros(rGPR[inst.RS]);
-  u32 mask = 0x80000000;
-
-  int i = 0;
-  for (; i < 32; i++, mask >>= 1)
-  {
-    if (val & mask)
-      break;
-  }
-
-  rGPR[inst.RA] = i;
 
   if (inst.Rc)
     Helper_UpdateCR0(rGPR[inst.RA]);

Source/Core/Core/PowerPC/Jit64/Jit_Integer.cpp

@@ -1956,14 +1956,7 @@ void Jit64::cntlzwx(UGeckoInstruction inst)
 
   if (gpr.IsImm(s))
   {
-    u32 mask = 0x80000000;
+    gpr.SetImmediate32(a, Common::CountLeadingZeros(gpr.Imm32(s)));
-    u32 i = 0;
-    for (; i < 32; i++, mask >>= 1)
-    {
-      if (gpr.Imm32(s) & mask)
-        break;
-    }
-    gpr.SetImmediate32(a, i);
   }
   else
   {

Source/Core/Core/PowerPC/JitArm64/JitArm64_Integer.cpp

@@ -368,11 +368,7 @@ void JitArm64::cntlzwx(UGeckoInstruction inst)
 
   if (gpr.IsImm(s))
   {
-#ifdef _MSC_VER
+    gpr.SetImmediate(a, Common::CountLeadingZeros(gpr.GetImm(s)));
-    gpr.SetImmediate(a, _CountLeadingZeros(gpr.GetImm(s)));
-#else
-    gpr.SetImmediate(a, __builtin_clz(gpr.GetImm(s)));
-#endif
     if (inst.Rc)
       ComputeRC0(gpr.GetImm(a));
   }