[flang] [clang] [lldb] [llvm] [lld] [libunwind] [libcxxabi] [clang-tools-extra] [libc] [compiler-rt] [mlir] [libcxx] [libc++][numeric] P0543R3: Saturation arithmetic (PR #77967)

[Hristo Hristov via llvm-commits]

================
@@ -0,0 +1,119 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _LIBCPP___NUMERIC_SATURATION_ARITHMETIC_H
+#define _LIBCPP___NUMERIC_SATURATION_ARITHMETIC_H
+
+#include <__concepts/arithmetic.h>
+#include <__config>
+#include <__type_traits/decay.h>
+#include <__utility/cmp.h>
+#include <limits>
+
+#if !defined(_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER)
+#  pragma GCC system_header
+#endif
+
+_LIBCPP_BEGIN_NAMESPACE_STD
+
+#if _LIBCPP_STD_VER >= 26
+
+template <typename _Tp>
+concept __libcpp_standard_integer = __libcpp_unsigned_integer<decay_t<_Tp>> || __libcpp_signed_integer<decay_t<_Tp>>;
+
+template <__libcpp_standard_integer _Tp>
+_LIBCPP_HIDE_FROM_ABI constexpr _Tp add_sat(_Tp __x, _Tp __y) noexcept {
+  if (_Tp __sum; !__builtin_add_overflow(__x, __y, &__sum))
+    return __sum;
+  // Handle overflow
+  if constexpr (__libcpp_unsigned_integer<_Tp>) {
+    return std::numeric_limits<_Tp>::max();
+  } else {
+    // Signed addition overflow
+    if (__x > 0)
+      // Overflows if (x > 0 && y > 0)
+      return std::numeric_limits<_Tp>::max();
+    else
+      // Overflows if  (x < 0 && y < 0)
+      return std::numeric_limits<_Tp>::min();
+  }
+}
+
+template <__libcpp_standard_integer _Tp>
+_LIBCPP_HIDE_FROM_ABI constexpr _Tp sub_sat(_Tp __x, _Tp __y) noexcept {
+  if (_Tp __sub; !__builtin_sub_overflow(__x, __y, &__sub))
+    return __sub;
+  // Handle overflow
+  if constexpr (__libcpp_unsigned_integer<_Tp>) {
+    // Overflows if (x < y)
+    return std::numeric_limits<_Tp>::min();
+  } else {
+    // Signed subtration overflow
+    if (__x > 0)
+      // Overflows if (x > 0 && y < 0)
+      return std::numeric_limits<_Tp>::max();
+    else
+      // Overflows if (x < 0 && y > 0)
+      return std::numeric_limits<_Tp>::min();
+  }
+}
+
+template <__libcpp_standard_integer _Tp>
+_LIBCPP_HIDE_FROM_ABI constexpr _Tp mul_sat(_Tp __x, _Tp __y) noexcept {
+  if (_Tp __mul; !__builtin_mul_overflow(__x, __y, &__mul))
+    return __mul;
+  // Handle overflow
+  if constexpr (__libcpp_unsigned_integer<_Tp>) {
+    return std::numeric_limits<_Tp>::max();
+  } else {
+    // Signed multiplication overflow
+    if (__x > 0) {
+      if (__y > 0)
+        // Overflows if (x > 0 && y > 0)
+        return std::numeric_limits<_Tp>::max();
+      // Overflows if (x > 0 && y < 0)
+      return std::numeric_limits<_Tp>::min();
+    }
+    if (__y > 0)
+      // Overflows if (x < 0 && y > 0)
+      return std::numeric_limits<_Tp>::min();
+    // Overflows if (x < 0 && y < 0)
+    return std::numeric_limits<_Tp>::max();
+  }
+}
+
+template <__libcpp_standard_integer _Tp>
+_LIBCPP_HIDE_FROM_ABI constexpr _Tp div_sat(_Tp __x, _Tp __y) noexcept {
+  _LIBCPP_ASSERT_UNCATEGORIZED(__y != 0, "Division by 0 is undefined");
+  if constexpr (__libcpp_unsigned_integer<_Tp>) {
+    return __x / __y;
+  } else {
+    // Handle signed division overflow
+    if (__x == std::numeric_limits<_Tp>::min() && __y == _Tp{-1})
+      return std::numeric_limits<_Tp>::max();
+    return __x / __y;
+  }
+}
+
+template <__libcpp_standard_integer _Rp, __libcpp_standard_integer _Tp>
+_LIBCPP_HIDE_FROM_ABI constexpr _Rp saturate_cast(_Tp __x) noexcept {
----------------
H-G-Hristov wrote:

OK. I'll remove the optimization for now. On optimized builds the difference according to my quick, stupid benchmark is near zero and on unoptimized builds less than 10%.
https://quick-bench.com/q/QFdMFKaw_5OnmeDqB_BMkh0BaSg

Please comment if you disagree.

https://github.com/llvm/llvm-project/pull/77967