boringssl/crypto/compiler_test.cc

// Copyright 2017 The BoringSSL Authors
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
//     https://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.

#include <limits.h>
#include <stdint.h>

#include <iterator>
#include <type_traits>

#include <gtest/gtest.h>

#include "test/test_util.h"


BSSL_NAMESPACE_BEGIN
namespace {

// C and C++ have two forms of unspecified behavior: undefined behavior and
// implementation-defined behavior.
//
// Programs that exhibit undefined behavior are invalid. Compilers are
// permitted to, and often do, arbitrarily miscompile them. BoringSSL thus aims
// to avoid undefined behavior.
//
// Implementation-defined behavior is left up to the compiler to define (or
// leave undefined). These are often platform-specific details, such as how big
// `int` is or how `uintN_t` is implemented. Programs that depend on
// implementation-defined behavior are not necessarily invalid, merely less
// portable. A compiler that provides some implementation-defined behavior is
// not permitted to miscompile code that depends on it.
//
// C allows a much wider range of platform behaviors than would be practical
// for us to support, so we make some assumptions on implementation-defined
// behavior. Platforms that violate those assumptions are not supported. This
// file aims to document and test these assumptions, so that platforms outside
// our scope are flagged.

template <typename T>
static void CheckRepresentation(T value) {
  SCOPED_TRACE(value);

  // Convert to the corresponding two's-complement unsigned value. We use an
  // unsigned value so the right-shift below has defined value. Right-shifts of
  // negative numbers in C are implementation defined.
  //
  // If `T` is already unsigned, this is a no-op, as desired.
  //
  // If `T` is signed, conversion to unsigned is defined to repeatedly add or
  // subtract (numerically, not within `T`) one more than the unsigned type's
  // maximum value until it fits (this must be a power of two). This is the
  // conversion we want.
  using UnsignedT = std::make_unsigned_t<T>;
  UnsignedT value_u = static_cast<UnsignedT>(value);
  EXPECT_EQ(sizeof(UnsignedT), sizeof(T));

  // Integers must be little-endian.
  uint8_t expected[sizeof(UnsignedT)];
  for (size_t i = 0; i < sizeof(UnsignedT); i++) {
    expected[i] = static_cast<uint8_t>(value_u);
    // Divide instead of right-shift to appease compilers that warn if `T` is a
    // char. The explicit cast is also needed to appease MSVC if integer
    // promotion happened.
    value_u = static_cast<UnsignedT>(value_u / 256);
  }
  EXPECT_EQ(0u, value_u);

  // Check that `value` has the expected representation.
  EXPECT_EQ(Bytes(expected),
            Bytes(reinterpret_cast<const uint8_t *>(&value), sizeof(value)));
}

TEST(CompilerTest, IntegerRepresentation) {
  static_assert(CHAR_BIT == 8, "BoringSSL only supports 8-bit chars");
  static_assert(UCHAR_MAX == 0xff, "BoringSSL only supports 8-bit chars");

  // Require that `unsigned char` and `uint8_t` be the same type. We require
  // that type-punning through `uint8_t` is not a strict aliasing violation. In
  // principle, type-punning should be done with `memcpy`, which would make this
  // moot.
  //
  // However, C made too many historical mistakes with the types and signedness
  // of character strings. As a result, aliasing between all variations on 8-bit
  // chars are a practical necessity for all real C code. We do not support
  // toolchains that break this assumption.
  static_assert(
      std::is_same_v<unsigned char, uint8_t>,
      "BoringSSL requires uint8_t and unsigned char be the same type");
  uint8_t u8 = 0;
  unsigned char *ptr = &u8;
  (void)ptr;

  // Sized integers have the expected size.
  static_assert(sizeof(uint8_t) == 1u, "uint8_t has the wrong size");
  static_assert(sizeof(uint16_t) == 2u, "uint16_t has the wrong size");
  static_assert(sizeof(uint32_t) == 4u, "uint32_t has the wrong size");
  static_assert(sizeof(uint64_t) == 8u, "uint64_t has the wrong size");

  // size_t does not exceed uint64_t.
  static_assert(sizeof(size_t) <= 8u, "size_t must not exceed uint64_t");

  // The positive maximum of `int` must fit into `size_t`.
  static_assert(sizeof(int) <= sizeof(size_t), "int must not exceed size_t");

  // Require that `int` be exactly 32 bits. OpenSSL historically mixed up
  // `unsigned` and `uint32_t`, so we require it be at least 32 bits. Requiring
  // at most 32-bits is a bit more subtle. C promotes arithmetic operands to
  // `int` when they fit. But this means, if `int` is 2N bits wide, multiplying
  // two maximum-sized `uintN_t`s is undefined by integer overflow!
  //
  // We attempt to handle this for `uint16_t`, assuming a 32-bit `int`, but we
  // make no attempts to correct for this with `uint32_t` for a 64-bit `int`.
  // Thus BoringSSL does not support ILP64 platforms.
  //
  // This test is on `INT_MAX` and `INT32_MAX` rather than sizeof because it is
  // theoretically allowed for sizeof(int) to be 4 but include padding bits.
  static_assert(INT_MAX == INT32_MAX, "BoringSSL requires int be 32-bit");
  static_assert(UINT_MAX == UINT32_MAX,
                "BoringSSL requires unsigned be 32-bit");

  CheckRepresentation(static_cast<signed char>(127));
  CheckRepresentation(static_cast<signed char>(1));
  CheckRepresentation(static_cast<signed char>(0));
  CheckRepresentation(static_cast<signed char>(-1));
  CheckRepresentation(static_cast<signed char>(-42));
  CheckRepresentation(static_cast<signed char>(-128));

  CheckRepresentation(static_cast<int>(INT_MAX));
  CheckRepresentation(static_cast<int>(0x12345678));
  CheckRepresentation(static_cast<int>(1));
  CheckRepresentation(static_cast<int>(0));
  CheckRepresentation(static_cast<int>(-1));
  CheckRepresentation(static_cast<int>(-0x12345678));
  CheckRepresentation(static_cast<int>(INT_MIN));

  CheckRepresentation(static_cast<unsigned>(UINT_MAX));
  CheckRepresentation(static_cast<unsigned>(0x12345678));
  CheckRepresentation(static_cast<unsigned>(1));
  CheckRepresentation(static_cast<unsigned>(0));

  CheckRepresentation(static_cast<long>(LONG_MAX));
  CheckRepresentation(static_cast<long>(0x12345678));
  CheckRepresentation(static_cast<long>(1));
  CheckRepresentation(static_cast<long>(0));
  CheckRepresentation(static_cast<long>(-1));
  CheckRepresentation(static_cast<long>(-0x12345678));
  CheckRepresentation(static_cast<long>(LONG_MIN));

  CheckRepresentation(static_cast<unsigned long>(ULONG_MAX));
  CheckRepresentation(static_cast<unsigned long>(0x12345678));
  CheckRepresentation(static_cast<unsigned long>(1));
  CheckRepresentation(static_cast<unsigned long>(0));

  CheckRepresentation(static_cast<int16_t>(0x7fff));
  CheckRepresentation(static_cast<int16_t>(0x1234));
  CheckRepresentation(static_cast<int16_t>(1));
  CheckRepresentation(static_cast<int16_t>(0));
  CheckRepresentation(static_cast<int16_t>(-1));
  CheckRepresentation(static_cast<int16_t>(-0x7fff - 1));

  CheckRepresentation(static_cast<uint16_t>(0xffff));
  CheckRepresentation(static_cast<uint16_t>(0x1234));
  CheckRepresentation(static_cast<uint16_t>(1));
  CheckRepresentation(static_cast<uint16_t>(0));

  CheckRepresentation(static_cast<int32_t>(0x7fffffff));
  CheckRepresentation(static_cast<int32_t>(0x12345678));
  CheckRepresentation(static_cast<int32_t>(1));
  CheckRepresentation(static_cast<int32_t>(0));
  CheckRepresentation(static_cast<int32_t>(-1));
  CheckRepresentation(static_cast<int32_t>(-0x7fffffff - 1));

  CheckRepresentation(static_cast<uint32_t>(0xffffffff));
  CheckRepresentation(static_cast<uint32_t>(0x12345678));
  CheckRepresentation(static_cast<uint32_t>(1));
  CheckRepresentation(static_cast<uint32_t>(0));

  CheckRepresentation(static_cast<int64_t>(0x7fffffffffffffff));
  CheckRepresentation(static_cast<int64_t>(0x123456789abcdef0));
  CheckRepresentation(static_cast<int64_t>(1));
  CheckRepresentation(static_cast<int64_t>(0));
  CheckRepresentation(static_cast<int64_t>(-1));
  CheckRepresentation(static_cast<int64_t>(-0x7fffffffffffffff - 1));

  CheckRepresentation(static_cast<uint64_t>(0xffffffffffffffff));
  CheckRepresentation(static_cast<uint64_t>(0x12345678abcdef0));
  CheckRepresentation(static_cast<uint64_t>(1));
  CheckRepresentation(static_cast<uint64_t>(0));
}

TEST(CompilerTest, PointerRepresentation) {
  // Converting pointers to integers and doing arithmetic on those values are
  // both defined. Converting those values back into pointers is undefined,
  // but, for aliasing checks, we require that the implementation-defined
  // result of that computation commutes with pointer arithmetic.
  char chars[256];
  for (size_t i = 0; i < sizeof(chars); i++) {
    EXPECT_EQ(reinterpret_cast<uintptr_t>(chars) + i,
              reinterpret_cast<uintptr_t>(chars + i));
  }

  int ints[256];
  for (size_t i = 0; i < std::size(ints); i++) {
    EXPECT_EQ(reinterpret_cast<uintptr_t>(ints) + i * sizeof(int),
              reinterpret_cast<uintptr_t>(ints + i));
  }

  // nullptr must be represented by all zeros in memory. This is necessary so
  // structs may be initialized by memset(0).
  int *null = nullptr;
  uint8_t bytes[sizeof(null)] = {0};
  EXPECT_EQ(Bytes(bytes),
            Bytes(reinterpret_cast<uint8_t *>(&null), sizeof(null)));
}

static uintptr_t aba(uintptr_t *a, void **b) {
  *a = (uintptr_t)1;
  *b = nullptr;
  return *a;  // 0 if a == b, 1 if a and b are disjoint
}

TEST(CompilerTest, NoStrictAliasing) {
  // Sequential memory access must be sequentially consistent across types.
  // Compilers such as clang and gcc need to be passed -fno-strict-aliasing
  // for this to remain true at higher optimization levels. Use with the
  // opposite configuration, -fstrict-aliasing, is not supported.
  // Even though some subset of type punning through memory is considered
  // undefined behavior, the subtlety of exactly which subset that is and the
  // limited sanitizer-tooling support make it impractical to avoid reliably.
  uint8_t aliased[sizeof(void *)] = {};
  uint8_t zeros[sizeof(void *)] = {};

  OPENSSL_memset(aliased, -1, sizeof(aliased));
  EXPECT_EQ(aba((uintptr_t *)aliased, (void **)aliased), (uintptr_t)0);
  EXPECT_EQ(Bytes(aliased), Bytes(zeros));

  volatile auto volatile_aba = &aba;
  OPENSSL_memset(aliased, -1, sizeof(aliased));
  EXPECT_EQ(volatile_aba((uintptr_t *)aliased, (void **)aliased), (uintptr_t)0);
  EXPECT_EQ(Bytes(aliased), Bytes(zeros));
}

}  // namespace
BSSL_NAMESPACE_END