| 1 |
// Copyright 2017 The Chromium Authors |
| 2 |
// Use of this source code is governed by a BSD-style license that can be |
| 3 |
// found in the LICENSE file. |
| 4 |
|
| 5 |
#ifndef BASE_NUMERICS_SAFE_MATH_ARM_IMPL_H_ |
| 6 |
#define BASE_NUMERICS_SAFE_MATH_ARM_IMPL_H_ |
| 7 |
|
| 8 |
#include <cassert> |
| 9 |
#include <type_traits> |
| 10 |
|
| 11 |
#include "safe_conversions.h" |
| 12 |
|
| 13 |
namespace chromium { |
| 14 |
namespace base { |
| 15 |
namespace internal { |
| 16 |
|
| 17 |
template <typename T, typename U> |
| 18 |
struct CheckedMulFastAsmOp { |
| 19 |
static const bool is_supported = |
| 20 |
kEnableAsmCode && FastIntegerArithmeticPromotion<T, U>::is_contained; |
| 21 |
|
| 22 |
// The following is not an assembler routine and is thus constexpr safe, it |
| 23 |
// just emits much more efficient code than the Clang and GCC builtins for |
| 24 |
// performing overflow-checked multiplication when a twice wider type is |
| 25 |
// available. The below compiles down to 2-3 instructions, depending on the |
| 26 |
// width of the types in use. |
| 27 |
// As an example, an int32_t multiply compiles to: |
| 28 |
// smull r0, r1, r0, r1 |
| 29 |
// cmp r1, r1, asr #31 |
| 30 |
// And an int16_t multiply compiles to: |
| 31 |
// smulbb r1, r1, r0 |
| 32 |
// asr r2, r1, #16 |
| 33 |
// cmp r2, r1, asr #15 |
| 34 |
template <typename V> |
| 35 |
static constexpr bool Do(T x, U y, V* result) { |
| 36 |
using Promotion = typename FastIntegerArithmeticPromotion<T, U>::type; |
| 37 |
Promotion presult; |
| 38 |
|
| 39 |
presult = static_cast<Promotion>(x) * static_cast<Promotion>(y); |
| 40 |
if (!IsValueInRangeForNumericType<V>(presult)) |
| 41 |
return false; |
| 42 |
*result = static_cast<V>(presult); |
| 43 |
return true; |
| 44 |
} |
| 45 |
}; |
| 46 |
|
| 47 |
template <typename T, typename U> |
| 48 |
struct ClampedAddFastAsmOp { |
| 49 |
static const bool is_supported = |
| 50 |
kEnableAsmCode && BigEnoughPromotion<T, U>::is_contained && |
| 51 |
IsTypeInRangeForNumericType< |
| 52 |
int32_t, |
| 53 |
typename BigEnoughPromotion<T, U>::type>::value; |
| 54 |
|
| 55 |
template <typename V> |
| 56 |
__attribute__((always_inline)) static V Do(T x, U y) { |
| 57 |
// This will get promoted to an int, so let the compiler do whatever is |
| 58 |
// clever and rely on the saturated cast to bounds check. |
| 59 |
if (IsIntegerArithmeticSafe<int, T, U>::value) |
| 60 |
return saturated_cast<V>(static_cast<int>(x) + static_cast<int>(y)); |
| 61 |
|
| 62 |
int32_t result; |
| 63 |
int32_t x_i32 = checked_cast<int32_t>(x); |
| 64 |
int32_t y_i32 = checked_cast<int32_t>(y); |
| 65 |
|
| 66 |
asm("qadd %[result], %[first], %[second]" |
| 67 |
: [result] "=r"(result) |
| 68 |
: [first] "r"(x_i32), [second] "r"(y_i32)); |
| 69 |
return saturated_cast<V>(result); |
| 70 |
} |
| 71 |
}; |
| 72 |
|
| 73 |
template <typename T, typename U> |
| 74 |
struct ClampedSubFastAsmOp { |
| 75 |
static const bool is_supported = |
| 76 |
kEnableAsmCode && BigEnoughPromotion<T, U>::is_contained && |
| 77 |
IsTypeInRangeForNumericType< |
| 78 |
int32_t, |
| 79 |
typename BigEnoughPromotion<T, U>::type>::value; |
| 80 |
|
| 81 |
template <typename V> |
| 82 |
__attribute__((always_inline)) static V Do(T x, U y) { |
| 83 |
// This will get promoted to an int, so let the compiler do whatever is |
| 84 |
// clever and rely on the saturated cast to bounds check. |
| 85 |
if (IsIntegerArithmeticSafe<int, T, U>::value) |
| 86 |
return saturated_cast<V>(static_cast<int>(x) - static_cast<int>(y)); |
| 87 |
|
| 88 |
int32_t result; |
| 89 |
int32_t x_i32 = checked_cast<int32_t>(x); |
| 90 |
int32_t y_i32 = checked_cast<int32_t>(y); |
| 91 |
|
| 92 |
asm("qsub %[result], %[first], %[second]" |
| 93 |
: [result] "=r"(result) |
| 94 |
: [first] "r"(x_i32), [second] "r"(y_i32)); |
| 95 |
return saturated_cast<V>(result); |
| 96 |
} |
| 97 |
}; |
| 98 |
|
| 99 |
template <typename T, typename U> |
| 100 |
struct ClampedMulFastAsmOp { |
| 101 |
static const bool is_supported = |
| 102 |
kEnableAsmCode && CheckedMulFastAsmOp<T, U>::is_supported; |
| 103 |
|
| 104 |
template <typename V> |
| 105 |
__attribute__((always_inline)) static V Do(T x, U y) { |
| 106 |
// Use the CheckedMulFastAsmOp for full-width 32-bit values, because |
| 107 |
// it's fewer instructions than promoting and then saturating. |
| 108 |
if (!IsIntegerArithmeticSafe<int32_t, T, U>::value && |
| 109 |
!IsIntegerArithmeticSafe<uint32_t, T, U>::value) { |
| 110 |
V result; |
| 111 |
return CheckedMulFastAsmOp<T, U>::Do(x, y, &result) |
| 112 |
? result |
| 113 |
: CommonMaxOrMin<V>(IsValueNegative(x) ^ IsValueNegative(y)); |
| 114 |
} |
| 115 |
|
| 116 |
assert((FastIntegerArithmeticPromotion<T, U>::is_contained)); |
| 117 |
using Promotion = typename FastIntegerArithmeticPromotion<T, U>::type; |
| 118 |
return saturated_cast<V>(static_cast<Promotion>(x) * |
| 119 |
static_cast<Promotion>(y)); |
| 120 |
} |
| 121 |
}; |
| 122 |
|
| 123 |
} // namespace internal |
| 124 |
} // namespace base |
| 125 |
} // namespace chromium |
| 126 |
|
| 127 |
#endif // BASE_NUMERICS_SAFE_MATH_ARM_IMPL_H_ |