MathExtras: avoid unnecessarily widening types

Several multi-argument functions unnecessarily widen types beyond the
argument types. Template'ize the functions, and use std::common_type_t
to avoid this, hence optimizing the functions. A requirement of this
patch is to change the overflow behavior of alignTo, to only overflow
when the result is not representable in the return type.

Author: artagnon

llvm/include/llvm/Support/MathExtras.h

@@ -23,6 +23,22 @@
 #include <type_traits>
 
 namespace llvm {
+/// Some template parameter helpers to optimize for bitwidth, for functions that
+/// take multiple arguments.
+
+// We can't verify signedness, since callers rely on implicit coercions to
+// signed/unsigned.
+template <typename T, typename U>
+using enableif_int =
+    std::enable_if_t<std::is_integral_v<T> && std::is_integral_v<U>>;
+
+// Use std::common_type_t to widen only up to the widest argument.
+template <typename T, typename U, typename = enableif_int<T, U>>
+using common_uint =
+    std::common_type_t<std::make_unsigned_t<T>, std::make_unsigned_t<U>>;
+template <typename T, typename U, typename = enableif_int<T, U>>
+using common_sint =
+    std::common_type_t<std::make_signed_t<T>, std::make_signed_t<U>>;
 
 /// Mathematical constants.
 namespace numbers {
@@ -346,7 +362,8 @@ inline unsigned Log2_64_Ceil(uint64_t Value) {
 
 /// A and B are either alignments or offsets. Return the minimum alignment that
 /// may be assumed after adding the two together.
-constexpr uint64_t MinAlign(uint64_t A, uint64_t B) {
+template <typename U, typename V, typename T = common_uint<U, V>>
+constexpr T MinAlign(U A, V B) {
   // The largest power of 2 that divides both A and B.
   //
   // Replace "-Value" by "1+~Value" in the following commented code to avoid
@@ -355,6 +372,11 @@ constexpr uint64_t MinAlign(uint64_t A, uint64_t B) {
   return (A | B) & (1 + ~(A | B));
 }
 
+/// Fallback when arguments aren't integral.
+constexpr uint64_t MinAlign(uint64_t A, uint64_t B) {
+  return (A | B) & (1 + ~(A | B));
+}
+
 /// Returns the next power of two (in 64-bits) that is strictly greater than A.
 /// Returns zero on overflow.
 constexpr uint64_t NextPowerOf2(uint64_t A) {
@@ -375,110 +397,143 @@ inline uint64_t PowerOf2Ceil(uint64_t A) {
   return UINT64_C(1) << Log2_64_Ceil(A);
 }
 
-/// Returns the next integer (mod 2**64) that is greater than or equal to
-/// \p Value and is a multiple of \p Align. \p Align must be non-zero.
-///
-/// Examples:
-/// \code
-///   alignTo(5, 8) = 8
-///   alignTo(17, 8) = 24
-///   alignTo(~0LL, 8) = 0
-///   alignTo(321, 255) = 510
-/// \endcode
-///
-/// May overflow.
-inline uint64_t alignTo(uint64_t Value, uint64_t Align) {
-  assert(Align != 0u && "Align can't be 0.");
-  return (Value + Align - 1) / Align * Align;
-}
-
-inline uint64_t alignToPowerOf2(uint64_t Value, uint64_t Align) {
-  assert(Align != 0 && (Align & (Align - 1)) == 0 &&
-         "Align must be a power of 2");
-  // Replace unary minus to avoid compilation error on Windows:
-  // "unary minus operator applied to unsigned type, result still unsigned"
-  uint64_t negAlign = (~Align) + 1;
-  return (Value + Align - 1) & negAlign;
-}
-
-/// If non-zero \p Skew is specified, the return value will be a minimal integer
-/// that is greater than or equal to \p Size and equal to \p A * N + \p Skew for
-/// some integer N. If \p Skew is larger than \p A, its value is adjusted to '\p
-/// Skew mod \p A'. \p Align must be non-zero.
-///
-/// Examples:
-/// \code
-///   alignTo(5, 8, 7) = 7
-///   alignTo(17, 8, 1) = 17
-///   alignTo(~0LL, 8, 3) = 3
-///   alignTo(321, 255, 42) = 552
-/// \endcode
-inline uint64_t alignTo(uint64_t Value, uint64_t Align, uint64_t Skew) {
-  assert(Align != 0u && "Align can't be 0.");
-  Skew %= Align;
-  return alignTo(Value - Skew, Align) + Skew;
-}
-
-/// Returns the next integer (mod 2**64) that is greater than or equal to
-/// \p Value and is a multiple of \c Align. \c Align must be non-zero.
-template <uint64_t Align> constexpr uint64_t alignTo(uint64_t Value) {
-  static_assert(Align != 0u, "Align must be non-zero");
-  return (Value + Align - 1) / Align * Align;
-}
-
 /// Returns the integer ceil(Numerator / Denominator). Unsigned version.
 /// Guaranteed to never overflow.
-inline uint64_t divideCeil(uint64_t Numerator, uint64_t Denominator) {
+template <typename U, typename V, typename T = common_uint<U, V>>
+constexpr T divideCeil(U Numerator, V Denominator) {
+  assert(Denominator && "Division by zero");
+  T Bias = (Numerator != 0);
+  return (Numerator - Bias) / Denominator + Bias;
+}
+
+/// Fallback when arguments aren't integral.
+constexpr uint64_t divideCeil(uint64_t Numerator, uint64_t Denominator) {
   assert(Denominator && "Division by zero");
   uint64_t Bias = (Numerator != 0);
   return (Numerator - Bias) / Denominator + Bias;
 }
 
 /// Returns the integer ceil(Numerator / Denominator). Signed version.
 /// Guaranteed to never overflow.
-inline int64_t divideCeilSigned(int64_t Numerator, int64_t Denominator) {
+template <typename U, typename V, typename T = common_sint<U, V>>
+constexpr T divideCeilSigned(U Numerator, V Denominator) {
   assert(Denominator && "Division by zero");
   if (!Numerator)
     return 0;
   // C's integer division rounds towards 0.
-  int64_t Bias = (Denominator >= 0 ? 1 : -1);
+  T Bias = Denominator >= 0 ? 1 : -1;
   bool SameSign = (Numerator >= 0) == (Denominator >= 0);
   return SameSign ? (Numerator - Bias) / Denominator + 1
                   : Numerator / Denominator;
 }
 
 /// Returns the integer floor(Numerator / Denominator). Signed version.
 /// Guaranteed to never overflow.
-inline int64_t divideFloorSigned(int64_t Numerator, int64_t Denominator) {
+template <typename U, typename V, typename T = common_sint<U, V>>
+constexpr T divideFloorSigned(U Numerator, V Denominator) {
   assert(Denominator && "Division by zero");
   if (!Numerator)
     return 0;
   // C's integer division rounds towards 0.
-  int64_t Bias = Denominator >= 0 ? -1 : 1;
+  T Bias = Denominator >= 0 ? -1 : 1;
   bool SameSign = (Numerator >= 0) == (Denominator >= 0);
   return SameSign ? Numerator / Denominator
                   : (Numerator - Bias) / Denominator - 1;
 }
 
 /// Returns the remainder of the Euclidean division of LHS by RHS. Result is
 /// always non-negative.
-inline int64_t mod(int64_t Numerator, int64_t Denominator) {
+template <typename U, typename V, typename T = common_sint<U, V>>
+constexpr T mod(U Numerator, V Denominator) {
   assert(Denominator >= 1 && "Mod by non-positive number");
-  int64_t Mod = Numerator % Denominator;
+  T Mod = Numerator % Denominator;
   return Mod < 0 ? Mod + Denominator : Mod;
 }
 
 /// Returns (Numerator / Denominator) rounded by round-half-up. Guaranteed to
 /// never overflow.
-inline uint64_t divideNearest(uint64_t Numerator, uint64_t Denominator) {
+template <typename U, typename V, typename T = common_uint<U, V>>
+constexpr T divideNearest(U Numerator, V Denominator) {
   assert(Denominator && "Division by zero");
-  uint64_t Mod = Numerator % Denominator;
-  return (Numerator / Denominator) + (Mod > (Denominator - 1) / 2);
+  T Mod = Numerator % Denominator;
+  return (Numerator / Denominator) +
+         (Mod > (static_cast<T>(Denominator) - 1) / 2);
+}
+
+/// Returns the next integer (mod 2**nbits) that is greater than or equal to
+/// \p Value and is a multiple of \p Align. \p Align must be non-zero.
+///
+/// Examples:
+/// \code
+///   alignTo(5, 8) = 8
+///   alignTo(17, 8) = 24
+///   alignTo(~0LL, 8) = 0
+///   alignTo(321, 255) = 510
+/// \endcode
+///
+/// Will overflow only if result is not representable in T.
+template <typename U, typename V, typename T = common_uint<U, V>>
+constexpr T alignTo(U Value, V Align) {
+  assert(Align != 0u && "Align can't be 0.");
+  T CeilDiv = divideCeil(Value, Align);
+  return CeilDiv * Align;
+}
+
+/// Fallback when arguments aren't integral.
+constexpr uint64_t alignTo(uint64_t Value, uint64_t Align) {
+  assert(Align != 0u && "Align can't be 0.");
+  uint64_t CeilDiv = divideCeil(Value, Align);
+  return CeilDiv * Align;
+}
+
+constexpr uint64_t alignToPowerOf2(uint64_t Value, uint64_t Align) {
+  assert(Align != 0 && (Align & (Align - 1)) == 0 &&
+         "Align must be a power of 2");
+  // Replace unary minus to avoid compilation error on Windows:
+  // "unary minus operator applied to unsigned type, result still unsigned"
+  uint64_t NegAlign = (~Align) + 1;
+  return (Value + Align - 1) & NegAlign;
+}
+
+/// If non-zero \p Skew is specified, the return value will be a minimal integer
+/// that is greater than or equal to \p Size and equal to \p A * N + \p Skew for
+/// some integer N. If \p Skew is larger than \p A, its value is adjusted to '\p
+/// Skew mod \p A'. \p Align must be non-zero.
+///
+/// Examples:
+/// \code
+///   alignTo(5, 8, 7) = 7
+///   alignTo(17, 8, 1) = 17
+///   alignTo(~0LL, 8, 3) = 3
+///   alignTo(321, 255, 42) = 552
+/// \endcode
+///
+/// May overflow.
+template <typename U, typename V, typename W,
+          typename T = common_uint<common_uint<U, V>, W>>
+constexpr T alignTo(U Value, V Align, W Skew) {
+  assert(Align != 0u && "Align can't be 0.");
+  Skew %= Align;
+  return alignTo(Value - Skew, Align) + Skew;
 }
 
-/// Returns the largest uint64_t less than or equal to \p Value and is
-/// \p Skew mod \p Align. \p Align must be non-zero
-inline uint64_t alignDown(uint64_t Value, uint64_t Align, uint64_t Skew = 0) {
+/// Returns the next integer (mod 2**nbits) that is greater than or equal to
+/// \p Value and is a multiple of \c Align. \c Align must be non-zero.
+///
+/// Will overflow only if result is not representable in T.
+template <auto Align, typename V, typename T = common_uint<decltype(Align), V>>
+constexpr T alignTo(V Value) {
+  static_assert(Align != 0u, "Align must be non-zero");
+  T CeilDiv = divideCeil(Value, Align);
+  return CeilDiv * Align;
+}
+
+/// Returns the largest unsigned integer less than or equal to \p Value and is
+/// \p Skew mod \p Align. \p Align must be non-zero. Guaranteed to never
+/// overflow.
+template <typename U, typename V, typename W = uint8_t,
+          typename T = common_uint<common_uint<U, V>, W>>
+constexpr T alignDown(U Value, V Align, W Skew = 0) {
   assert(Align != 0u && "Align can't be 0.");
   Skew %= Align;
   return (Value - Skew) / Align * Align + Skew;
@@ -522,8 +577,8 @@ inline int64_t SignExtend64(uint64_t X, unsigned B) {
 
 /// Subtract two unsigned integers, X and Y, of type T and return the absolute
 /// value of the result.
-template <typename T>
-std::enable_if_t<std::is_unsigned_v<T>, T> AbsoluteDifference(T X, T Y) {
+template <typename U, typename V, typename T = common_uint<U, V>>
+constexpr T AbsoluteDifference(U X, V Y) {
   return X > Y ? (X - Y) : (Y - X);
 }
 

llvm/unittests/Support/MathExtrasTest.cpp

@@ -189,21 +189,35 @@ TEST(MathExtras, AlignTo) {
   EXPECT_EQ(8u, alignTo(5, 8));
   EXPECT_EQ(24u, alignTo(17, 8));
   EXPECT_EQ(0u, alignTo(~0LL, 8));
-  EXPECT_EQ(static_cast<uint64_t>(std::numeric_limits<uint32_t>::max()) + 1,
-            alignTo(std::numeric_limits<uint32_t>::max(), 2));
+  EXPECT_EQ(8u, alignTo(5ULL, 8ULL));
+
+  EXPECT_EQ(8u, alignTo<8>(5));
+  EXPECT_EQ(24u, alignTo<8>(17));
+  EXPECT_EQ(0u, alignTo<8>(~0LL));
+  EXPECT_EQ(254u,
+            alignTo<static_cast<uint8_t>(127)>(static_cast<uint8_t>(200)));
 
   EXPECT_EQ(7u, alignTo(5, 8, 7));
   EXPECT_EQ(17u, alignTo(17, 8, 1));
   EXPECT_EQ(3u, alignTo(~0LL, 8, 3));
   EXPECT_EQ(552u, alignTo(321, 255, 42));
   EXPECT_EQ(std::numeric_limits<uint32_t>::max(),
             alignTo(std::numeric_limits<uint32_t>::max(), 2, 1));
+
+  // Overflow.
+  EXPECT_EQ(0u, alignTo(static_cast<uint8_t>(200), static_cast<uint8_t>(128)));
+  EXPECT_EQ(0u, alignTo<static_cast<uint8_t>(128)>(static_cast<uint8_t>(200)));
+  EXPECT_EQ(0u, alignTo(static_cast<uint8_t>(200), static_cast<uint8_t>(128),
+                        static_cast<uint8_t>(0)));
+  EXPECT_EQ(0u, alignTo(std::numeric_limits<uint32_t>::max(), 2));
 }
 
 TEST(MathExtras, AlignToPowerOf2) {
+  EXPECT_EQ(0u, alignToPowerOf2(0u, 8));
   EXPECT_EQ(8u, alignToPowerOf2(5, 8));
   EXPECT_EQ(24u, alignToPowerOf2(17, 8));
   EXPECT_EQ(0u, alignToPowerOf2(~0LL, 8));
+  EXPECT_EQ(240u, alignToPowerOf2(240, 16));
   EXPECT_EQ(static_cast<uint64_t>(std::numeric_limits<uint32_t>::max()) + 1,
             alignToPowerOf2(std::numeric_limits<uint32_t>::max(), 2));
 }