[libc] Replace type punning with bit_cast

Although type punning is defined for union in C, it is UB in C++.
This patch introduces a bit_cast function to convert between types in a safe way.

This is necessary to get llvm-libc compile with GCC.
This patch is extracted from D119002.

Differential Revision: https://reviews.llvm.org/D119145

Author: gchatelet

libc/src/__support/CPP/Bit.h

@@ -0,0 +1,48 @@
+//===-- Freestanding version of bit_cast  -----------------------*- 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 LLVM_LIBC_SUPPORT_CPP_BIT_H
+#define LLVM_LIBC_SUPPORT_CPP_BIT_H
+
+namespace __llvm_libc {
+
+#if defined __has_builtin
+#if __has_builtin(__builtin_bit_cast)
+#define LLVM_LIBC_HAS_BUILTIN_BIT_CAST
+#endif
+#endif
+
+#if defined __has_builtin
+#if __has_builtin(__builtin_memcpy_inline)
+#define LLVM_LIBC_HAS_BUILTIN_MEMCPY_INLINE
+#endif
+#endif
+
+// This function guarantees the bitcast to be optimized away by the compiler for
+// GCC >= 8 and Clang >= 6.
+template <class To, class From> constexpr To bit_cast(const From &from) {
+  static_assert(sizeof(To) == sizeof(From), "To and From must be of same size");
+#if defined(LLVM_LIBC_HAS_BUILTIN_BIT_CAST)
+  return __builtin_bit_cast(To, from);
+#else
+  To to;
+  char *dst = reinterpret_cast<char *>(&to);
+  const char *src = reinterpret_cast<const char *>(&from);
+#if defined(LLVM_LIBC_HAS_BUILTIN_MEMCPY_INLINE)
+  __builtin_memcpy_inline(dst, src, sizeof(To));
+#else
+  for (unsigned i = 0; i < sizeof(To); ++i)
+    dst[i] = src[i];
+#endif // defined(LLVM_LIBC_HAS_BUILTIN_MEMCPY_INLINE)
+  return to;
+#endif // defined(LLVM_LIBC_HAS_BUILTIN_BIT_CAST)
+}
+
+} // namespace __llvm_libc
+
+#endif // LLVM_LIBC_SUPPORT_CPP_BIT_H

libc/src/__support/CPP/CMakeLists.txt

@@ -3,9 +3,10 @@ add_header_library(
   HDRS
     Array.h
     ArrayRef.h
+    Bit.h
     Bitset.h
     Functional.h
+    Limits.h
     StringView.h
     TypeTraits.h
-    Limits.h
 )

libc/src/__support/FPUtil/FPBits.h

@@ -11,6 +11,7 @@
 
 #include "PlatformDefs.h"
 
+#include "src/__support/CPP/Bit.h"
 #include "src/__support/CPP/TypeTraits.h"
 
 #include "FloatProperties.h"
@@ -35,7 +36,7 @@ template <typename T> struct ExponentWidth {
 // floating numbers. On x86 platforms however, the 'long double' type maps to
 // an x87 floating point format. This format is an IEEE 754 extension format.
 // It is handled as an explicit specialization of this class.
-template <typename T> union FPBits {
+template <typename T> struct FPBits {
   static_assert(cpp::IsFloatingPointType<T>::Value,
                 "FPBits instantiated with invalid type.");
 
@@ -76,7 +77,6 @@ template <typename T> union FPBits {
   bool get_sign() const {
     return ((bits & FloatProp::SIGN_MASK) >> (FloatProp::BIT_WIDTH - 1));
   }
-  T val;
 
   static_assert(sizeof(T) == sizeof(UIntType),
                 "Data type and integral representation have different sizes.");
@@ -96,15 +96,20 @@ template <typename T> union FPBits {
   // type match.
   template <typename XType,
             cpp::EnableIfType<cpp::IsSame<T, XType>::Value, int> = 0>
-  explicit FPBits(XType x) : val(x) {}
+  constexpr explicit FPBits(XType x)
+      : bits(__llvm_libc::bit_cast<UIntType>(x)) {}
 
   template <typename XType,
             cpp::EnableIfType<cpp::IsSame<XType, UIntType>::Value, int> = 0>
-  explicit FPBits(XType x) : bits(x) {}
+  constexpr explicit FPBits(XType x) : bits(x) {}
 
   FPBits() : bits(0) {}
 
-  explicit operator T() { return val; }
+  T get_val() const { return __llvm_libc::bit_cast<T>(bits); }
+
+  void set_val(T value) { bits = __llvm_libc::bit_cast<UIntType>(value); }
+
+  explicit operator T() const { return get_val(); }
 
   UIntType uintval() const { return bits; }
 

libc/src/__support/FPUtil/Hypot.h

@@ -12,6 +12,7 @@
 #include "BasicOperations.h"
 #include "FEnvImpl.h"
 #include "FPBits.h"
+#include "src/__support/CPP/Bit.h"
 #include "src/__support/CPP/TypeTraits.h"
 
 namespace __llvm_libc {
@@ -285,7 +286,7 @@ static inline T hypot(T x, T y) {
   }
 
   y_new |= static_cast<UIntType>(out_exp) << MantissaWidth<T>::VALUE;
-  return *reinterpret_cast<T *>(&y_new);
+  return __llvm_libc::bit_cast<T>(y_new);
 }
 
 } // namespace fputil

libc/src/__support/FPUtil/ManipulationFunctions.h

@@ -14,6 +14,7 @@
 #include "NormalFloat.h"
 #include "PlatformDefs.h"
 
+#include "src/__support/CPP/Bit.h"
 #include "src/__support/CPP/TypeTraits.h"
 
 #include <limits.h>
@@ -171,7 +172,7 @@ static inline T nextafter(T from, T to) {
     int_val = (to_bits.uintval() & sign_mask) + UIntType(1);
   }
 
-  return *reinterpret_cast<T *>(&int_val);
+  return __llvm_libc::bit_cast<T>(int_val);
   // TODO: Raise floating point exceptions as required by the standard.
 }
 

libc/src/__support/FPUtil/generic/sqrt.h

@@ -10,6 +10,7 @@
 #define LLVM_LIBC_SRC_SUPPORT_FPUTIL_GENERIC_SQRT_H
 
 #include "sqrt_80_bit_long_double.h"
+#include "src/__support/CPP/Bit.h"
 #include "src/__support/CPP/TypeTraits.h"
 #include "src/__support/FPUtil/FEnvImpl.h"
 #include "src/__support/FPUtil/FPBits.h"
@@ -203,7 +204,7 @@ sqrt(T x) {
         break;
       }
 
-      return *reinterpret_cast<T *>(&y);
+      return __llvm_libc::bit_cast<T>(y);
     }
   }
 }

libc/src/__support/FPUtil/x86_64/LongDoubleBits.h

@@ -9,6 +9,7 @@
 #ifndef LLVM_LIBC_SRC_SUPPORT_FPUTIL_X86_64_LONG_DOUBLE_BITS_H
 #define LLVM_LIBC_SRC_SUPPORT_FPUTIL_X86_64_LONG_DOUBLE_BITS_H
 
+#include "src/__support/CPP/Bit.h"
 #include "src/__support/architectures.h"
 
 #if !defined(LLVM_LIBC_ARCH_X86)
@@ -30,7 +31,7 @@ template <> struct Padding<4> { static constexpr unsigned VALUE = 16; };
 // x86_64 padding.
 template <> struct Padding<8> { static constexpr unsigned VALUE = 48; };
 
-template <> union FPBits<long double> {
+template <> struct FPBits<long double> {
   using UIntType = __uint128_t;
 
   static constexpr int EXPONENT_BIAS = 0x3FFF;
@@ -91,13 +92,11 @@ template <> union FPBits<long double> {
     return ((bits & FloatProp::SIGN_MASK) >> (FloatProp::BIT_WIDTH - 1));
   }
 
-  long double val;
-
   FPBits() : bits(0) {}
 
   template <typename XType,
             cpp::EnableIfType<cpp::IsSame<long double, XType>::Value, int> = 0>
-  explicit FPBits(XType x) : val(x) {
+  explicit FPBits(XType x) : bits(__llvm_libc::bit_cast<UIntType>(x)) {
     // bits starts uninitialized, and setting it to a long double only
     // overwrites the first 80 bits. This clears those upper bits.
     bits = bits & ((UIntType(1) << 80) - 1);
@@ -107,7 +106,7 @@ template <> union FPBits<long double> {
             cpp::EnableIfType<cpp::IsSame<XType, UIntType>::Value, int> = 0>
   explicit FPBits(XType x) : bits(x) {}
 
-  operator long double() { return val; }
+  operator long double() { return __llvm_libc::bit_cast<long double>(bits); }
 
   UIntType uintval() {
     // We zero the padding bits as they can contain garbage.

libc/src/__support/FPUtil/x86_64/NextAfterLongDouble.h

@@ -15,6 +15,7 @@
 #error "Invalid include"
 #endif
 
+#include "src/__support/CPP/Bit.h"
 #include "src/__support/FPUtil/FPBits.h"
 
 #include <stdint.h>
@@ -111,7 +112,7 @@ static inline long double nextafter(long double from, long double to) {
     }
   }
 
-  return *reinterpret_cast<long double *>(&int_val);
+  return __llvm_libc::bit_cast<long double>(int_val);
   // TODO: Raise floating point exceptions as required by the standard.
 }
 

libc/src/__support/FPUtil/x86_64/sqrt.h

@@ -33,9 +33,8 @@ template <> inline double sqrt<double>(double x) {
 }
 
 template <> inline long double sqrt<long double>(long double x) {
-  long double result;
-  __asm__ __volatile__("fsqrt" : "=t"(result) : "t"(x));
-  return result;
+  __asm__ __volatile__("fsqrt" : "+t"(x));
+  return x;
 }
 
 } // namespace fputil

libc/src/math/generic/log10f.cpp

@@ -155,7 +155,7 @@ LLVM_LIBC_FUNCTION(float, log10f, (float x)) {
       return x;
     }
     // Normalize denormal inputs.
-    xbits.val *= 0x1.0p23f;
+    xbits.set_val(xbits.get_val() * 0x1.0p23f);
     m -= 23.0;
   }
 
@@ -164,7 +164,7 @@ LLVM_LIBC_FUNCTION(float, log10f, (float x)) {
   xbits.set_unbiased_exponent(0x7F);
   int f_index = xbits.get_mantissa() >> 16;
 
-  FPBits f(xbits.val);
+  FPBits f = xbits;
   f.bits &= ~0x0000'FFFF;
 
   double d = static_cast<float>(xbits) - static_cast<float>(f);

libc/src/math/generic/log1pf.cpp

@@ -59,7 +59,7 @@ INLINE_FMA static inline float log(double x) {
   int f_index =
       xbits.get_mantissa() >> 45; // fputil::MantissaWidth<double>::VALUE - 7
 
-  FPBits f(xbits.val);
+  FPBits f = xbits;
   // Clear the lowest 45 bits.
   f.bits &= ~0x0000'1FFF'FFFF'FFFFULL;
 

libc/src/math/generic/log2f.cpp

@@ -138,7 +138,7 @@ LLVM_LIBC_FUNCTION(float, log2f, (float x)) {
       return x;
     }
     // Normalize denormal inputs.
-    xbits.val *= 0x1.0p23f;
+    xbits.set_val(xbits.get_val() * 0x1.0p23f);
     m = -23;
   }
 
@@ -149,7 +149,7 @@ LLVM_LIBC_FUNCTION(float, log2f, (float x)) {
   // lookup tables.
   int f_index = xbits.get_mantissa() >> 16;
 
-  FPBits f(xbits.val);
+  FPBits f = xbits;
   // Clear the lowest 16 bits.
   f.bits &= ~0x0000'FFFF;
 

libc/src/math/generic/logf.cpp

@@ -104,7 +104,7 @@ LLVM_LIBC_FUNCTION(float, logf, (float x)) {
       return x;
     }
     // Normalize denormal inputs.
-    xbits.val *= 0x1.0p23f;
+    xbits.set_val(xbits.get_val() * 0x1.0p23f);
     m = -23;
   }
 
@@ -113,7 +113,7 @@ LLVM_LIBC_FUNCTION(float, logf, (float x)) {
   xbits.set_unbiased_exponent(0x7F);
   int f_index = xbits.get_mantissa() >> 16;
 
-  FPBits f(xbits.val);
+  FPBits f = xbits;
   f.bits &= ~0x0000'FFFF;
 
   double d = static_cast<float>(xbits) - static_cast<float>(f);

libc/src/math/generic/math_utils.h

@@ -9,6 +9,7 @@
 #ifndef LLVM_LIBC_SRC_MATH_MATH_UTILS_H
 #define LLVM_LIBC_SRC_MATH_MATH_UTILS_H
 
+#include "src/__support/CPP/Bit.h"
 #include "src/__support/CPP/TypeTraits.h"
 #include "src/__support/common.h"
 #include <errno.h>
@@ -19,19 +20,19 @@
 namespace __llvm_libc {
 
 static inline uint32_t as_uint32_bits(float x) {
-  return *reinterpret_cast<uint32_t *>(&x);
+  return __llvm_libc::bit_cast<uint32_t>(x);
 }
 
 static inline uint64_t as_uint64_bits(double x) {
-  return *reinterpret_cast<uint64_t *>(&x);
+  return __llvm_libc::bit_cast<uint64_t>(x);
 }
 
 static inline float as_float(uint32_t x) {
-  return *reinterpret_cast<float *>(&x);
+  return __llvm_libc::bit_cast<float>(x);
 }
 
 static inline double as_double(uint64_t x) {
-  return *reinterpret_cast<double *>(&x);
+  return __llvm_libc::bit_cast<double>(x);
 }
 
 static inline uint32_t top12_bits(float x) { return as_uint32_bits(x) >> 20; }

libc/src/string/memory_utils/CMakeLists.txt

@@ -7,6 +7,8 @@ add_header_library(
     memcmp_implementations.h
     memcpy_implementations.h
     memset_implementations.h
+  DEPS
+    standalone_cpp
 )
 
 add_header_library(