Merge pull request #131 from chfast/optimize_knuth

Optimize Knuth's division

lib/intx/div.cpp

@@ -3,6 +3,28 @@
 // Licensed under the Apache License, Version 2.0.
 
 #include "div.hpp"
+#include <cassert>
+#include <tuple>
+
+#if defined(_MSC_VER)
+#define UNREACHABLE __assume(0)
+#else
+#define UNREACHABLE __builtin_unreachable()
+#endif
+
+#if defined(_MSC_VER)
+#define UNLIKELY(EXPR) EXPR
+#else
+#define UNLIKELY(EXPR) __builtin_expect((bool)(EXPR), false)
+#endif
+
+#if defined(NDEBUG)
+#define REQUIRE(X) \
+    if (!(X))      \
+    UNREACHABLE
+#else
+#define REQUIRE assert
+#endif
 
 namespace intx
 {
@@ -54,6 +76,37 @@ inline uint128 udivrem_by2(uint64_t u[], int m, uint128 d) noexcept
     return r;
 }
 
+/// s = x + y.
+inline bool add(uint64_t s[], const uint64_t x[], const uint64_t y[], int len) noexcept
+{
+    // OPT: Add MinLen template parameter and unroll first loop iterations.
+    REQUIRE(len >= 3);
+
+    bool carry = false;
+    for (int i = 0; i < len; ++i)
+        std::tie(s[i], carry) = add_with_carry(x[i], y[i], carry);
+    return carry;
+}
+
+/// r = x - multiplier * y.
+inline uint64_t submul(
+    uint64_t r[], const uint64_t x[], const uint64_t y[], int len, uint64_t multiplier) noexcept
+{
+    // OPT: Add MinLen template parameter and unroll first loop iterations.
+    REQUIRE(len >= 3);
+
+    uint64_t borrow = 0;
+    for (int i = 0; i < len; ++i)
+    {
+        const auto s = sub_with_carry(x[i], borrow);
+        const auto p = umul(y[i], multiplier);
+        const auto t = sub_with_carry(s.value, p.lo);
+        r[i] = t.value;
+        borrow = p.hi + s.carry + t.carry;
+    }
+    return borrow;
+}
+
 void udivrem_knuth(uint64_t q[], uint64_t un[], int m, const uint64_t vn[], int n) noexcept
 {
     const auto divisor = uint128{vn[n - 1], vn[n - 2]};
@@ -67,67 +120,41 @@ void udivrem_knuth(uint64_t q[], uint64_t un[], int m, const uint64_t vn[], int
         uint64_t qhat;
         uint128 rhat;
         const auto dividend = uint128{u2, u1};
-        if (dividend.hi >= divisor.hi)  // Will overflow:
+        if (UNLIKELY(dividend.hi >= divisor.hi))  // Division overflows.
         {
             qhat = ~uint64_t{0};
             rhat = dividend - uint128{divisor.hi, 0};
             rhat += divisor.hi;
 
-            // Adjustment.
-            // OPT: This is not needed but helps avoiding negative case.
+            // Adjustment (not needed for correctness, but helps avoiding "add back" case).
             if (rhat.hi == 0 && umul(qhat, divisor.lo) > uint128{rhat.lo, u0})
                 --qhat;
         }
         else
         {
-            auto res = udivrem_2by1(dividend, divisor.hi, reciprocal);
+            const auto res = udivrem_2by1(dividend, divisor.hi, reciprocal);
             qhat = res.quot;
             rhat = res.rem;
 
-            if (umul(qhat, divisor.lo) > uint128{rhat.lo, u0})
+            const auto p = umul(qhat, divisor.lo);
+            if (p > uint128{rhat.lo, u0})
             {
                 --qhat;
                 rhat += divisor.hi;
 
-                // Adjustment.
-                // OPT: This is not needed but helps avoiding negative case.
-                if (rhat.hi == 0 && umul(qhat, divisor.lo) > uint128{rhat.lo, u0})
+                // Adjustment (not needed for correctness, but helps avoiding "add back" case).
+                if (rhat.hi == 0 && (p - divisor.lo) > uint128{rhat.lo, u0})
                     --qhat;
             }
         }
 
         // Multiply and subtract.
-        uint64_t borrow = 0;
-        for (int i = 0; i < n; ++i)
-        {
-            const auto p = umul(qhat, vn[i]);
-            const auto s = uint128{un[i + j]} - borrow - p.lo;
-            un[i + j] = s.lo;
-            borrow = p.hi - s.hi;
-        }
-
+        const auto borrow = submul(&un[j], &un[j], vn, n, qhat);
         un[j + n] = u2 - borrow;
         if (u2 < borrow)  // Too much subtracted, add back.
         {
             --qhat;
-
-            uint64_t carry = 0;
-            for (int i = 0; i < n; ++i)
-            {
-                auto s = uint128{un[i + j]} + vn[i] + carry;
-                un[i + j] = s.lo;
-                carry = s.hi;
-            }
-            un[j + n] += carry;
-
-            // TODO: Consider this alternative implementation:
-            // bool k = false;
-            // for (int i = 0; i < n; ++i) {
-            //     auto limit = std::min(un[j+i],vn[i]);
-            //     un[i + j] += vn[i] + k;
-            //     k = un[i + j] < limit || (k && un[i + j] == limit);
-            // }
-            // un[j+n] += k;
+            un[j + n] += add(&un[j], &un[j], vn, n);
         }
 
         q[j] = qhat;  // Store quotient digit.

test/unittests/test_div.cpp

@@ -195,6 +195,18 @@ static div_test_case<uint512> div_test_cases[] = {
         0x10000000000000001_u512,
         0x7fffffffffffffff80000000000000000000000000000000_u512,
     },
+    {
+        0x00e8e8e8e2000100000009ea02000000000000ff3ffffff800000010002200000000000000000000000000000000000000000000000000000000000000000000_u512,
+        0x00e8e8e8e2000100000009ea02000000000000ff3ffffff800000010002280ff0000000000000000000000000000000000000000000000000000000000000000_u512,
+        0,
+        0x00e8e8e8e2000100000009ea02000000000000ff3ffffff800000010002200000000000000000000000000000000000000000000000000000000000000000000_u512,
+    },
+    {
+        0x000000c9700000000000000000023f00c00014ff0000000000000000223008050000000000000000000000000000000000000000000000000000000000000000_u512,
+        0x00000000c9700000000000000000023f00c00014ff002c0000000000002231080000000000000000000000000000000000000000000000000000000000000000_u512,
+        0xff,
+        0x00000000c9700000000000000000023f00c00014fed42c00000000000021310d0000000000000000000000000000000000000000000000000000000000000000_u512,
+    },
 };
 
 TEST(div, udivrem_512)