[libc][fenv] Refactor x86 fenv implementations to make it work for various fenv_t. (#165015)

Author: lntue

libc/cmake/modules/LLVMLibCCompileOptionRules.cmake

@@ -115,6 +115,10 @@ function(_get_compile_options_from_config output_var)
     list(APPEND config_options "-DLIBC_THREAD_MODE=${LIBC_CONF_THREAD_MODE}")
   endif()
 
+  if(LIBC_CONF_TRAP_ON_RAISE_FP_EXCEPT)
+    list(APPEND config_options "-DLIBC_TRAP_ON_RAISE_FP_EXCEPT")
+  endif()
+
   set(${output_var} ${config_options} PARENT_SCOPE)
 endfunction(_get_compile_options_from_config)
 

libc/config/config.json

@@ -141,5 +141,11 @@
       "value": "LIBC_WCTYPE_MODE_ASCII",
       "doc": "The implementation used for wctype, acceptable values are LIBC_WCTYPE_MODE_ASCII and LIBC_WCTYPE_MODE_UTF8."
     }
+  },
+  "fenv": {
+    "LIBC_CONF_TRAP_ON_RAISE_FP_EXCEPT": {
+      "value": false,
+      "doc": "Trap with SIGFPE when feraiseexcept is called with unmasked floating point exceptions, similar to glibc's behavior.  This is currently working only on x86 with SSE."
+    }
   }
 }

libc/docs/configure.rst

@@ -30,6 +30,8 @@ to learn about the defaults for your platform and target.
     - ``LIBC_CONF_KEEP_FRAME_POINTER``: Keep frame pointer in functions for better debugging experience.
 * **"errno" options**
     - ``LIBC_CONF_ERRNO_MODE``: The implementation used for errno, acceptable values are LIBC_ERRNO_MODE_DEFAULT, LIBC_ERRNO_MODE_UNDEFINED, LIBC_ERRNO_MODE_THREAD_LOCAL, LIBC_ERRNO_MODE_SHARED, LIBC_ERRNO_MODE_EXTERNAL, and LIBC_ERRNO_MODE_SYSTEM_INLINE.
+* **"fenv" options**
+    - ``LIBC_CONF_TRAP_ON_RAISE_FP_EXCEPT``: Trap with SIGFPE when feraiseexcept is called with unmasked floating point exceptions, similar to glibc's behavior.  This is currently working only on x86 with SSE.
 * **"general" options**
     - ``LIBC_ADD_NULL_CHECKS``: Add nullptr checks in the library's implementations to some functions for which passing nullptr is undefined behavior.
 * **"math" options**

libc/include/llvm-libc-macros/fenv-macros.h

@@ -14,9 +14,10 @@
 #define FE_INVALID 0x4
 #define FE_OVERFLOW 0x8
 #define FE_UNDERFLOW 0x10
-#define __FE_DENORM 0x20
+#define FE_DENORM 0x20
 #define FE_ALL_EXCEPT                                                          \
-  (FE_DIVBYZERO | FE_INEXACT | FE_INVALID | FE_OVERFLOW | FE_UNDERFLOW)
+  (FE_DIVBYZERO | FE_INEXACT | FE_INVALID | FE_OVERFLOW | FE_UNDERFLOW |       \
+   FE_DENORM)
 
 #define FE_DOWNWARD 0x400
 #define FE_TONEAREST 0

libc/src/__support/CPP/bit.h

@@ -26,6 +26,16 @@ namespace cpp {
 #define LLVM_LIBC_HAS_BUILTIN_MEMCPY_INLINE
 #endif
 
+template <unsigned N>
+LIBC_INLINE static void inline_copy(const char *from, char *to) {
+#if __has_builtin(__builtin_memcpy_inline)
+  __builtin_memcpy_inline(to, from, N);
+#else
+  for (unsigned i = 0; i < N; ++i)
+    to[i] = from[i];
+#endif // __has_builtin(__builtin_memcpy_inline)
+}
+
 // This implementation of bit_cast requires trivially-constructible To, to avoid
 // UB in the implementation.
 template <typename To, typename From>
@@ -43,16 +53,26 @@ bit_cast(const From &from) {
   To to{};
   char *dst = reinterpret_cast<char *>(&to);
   const char *src = reinterpret_cast<const char *>(&from);
-#if __has_builtin(__builtin_memcpy_inline)
-  __builtin_memcpy_inline(dst, src, sizeof(To));
-#else
-  for (unsigned i = 0; i < sizeof(To); ++i)
-    dst[i] = src[i];
-#endif // __has_builtin(__builtin_memcpy_inline)
+  inline_copy<sizeof(From)>(src, dst);
   return to;
 #endif // __has_builtin(__builtin_bit_cast)
 }
 
+// The following simple bit copy from a smaller type to maybe-larger type.
+template <typename To, typename From>
+LIBC_INLINE constexpr cpp::enable_if_t<
+    (sizeof(To) >= sizeof(From)) &&
+        cpp::is_trivially_constructible<To>::value &&
+        cpp::is_trivially_copyable<To>::value &&
+        cpp::is_trivially_copyable<From>::value,
+    void>
+bit_copy(const From &from, To &to) {
+  MSAN_UNPOISON(&from, sizeof(From));
+  char *dst = reinterpret_cast<char *>(&to);
+  const char *src = reinterpret_cast<const char *>(&from);
+  inline_copy<sizeof(From)>(src, dst);
+}
+
 template <typename T>
 [[nodiscard]] LIBC_INLINE constexpr cpp::enable_if_t<cpp::is_unsigned_v<T>,
                                                      bool>

libc/src/__support/FPUtil/CMakeLists.txt

@@ -7,6 +7,7 @@ add_header_library(
     libc.hdr.fenv_macros
     libc.hdr.math_macros
     libc.hdr.stdint_proxy
+    libc.src.__support.CPP.bit
     libc.src.__support.CPP.type_traits
     libc.src.__support.macros.attributes
     libc.src.errno.errno

libc/src/__support/FPUtil/FEnvImpl.h

@@ -31,8 +31,7 @@
 // the dummy implementations below. Once a proper x86_64 darwin fenv is set up,
 // the apple condition here should be removed.
 // TODO: fully support fenv for MSVC.
-#elif defined(LIBC_TARGET_ARCH_IS_X86) && !defined(__APPLE__) &&               \
-    !defined(LIBC_COMPILER_IS_MSVC)
+#elif defined(LIBC_TARGET_ARCH_IS_X86) && !defined(__APPLE__)
 #include "x86_64/FEnvImpl.h"
 #elif defined(LIBC_TARGET_ARCH_IS_ARM) && defined(__ARM_FP) &&                 \
     !defined(LIBC_COMPILER_IS_MSVC)
@@ -110,12 +109,7 @@ raise_except_if_required([[maybe_unused]] int excepts) {
   } else {
 #ifndef LIBC_MATH_HAS_NO_EXCEPT
     if (math_errhandling & MATH_ERREXCEPT)
-#ifdef LIBC_TARGET_ARCH_IS_X86_64
-      return raise_except</*SKIP_X87_FPU*/ true>(excepts);
-#else  // !LIBC_TARGET_ARCH_IS_X86
       return raise_except(excepts);
-#endif // LIBC_TARGET_ARCH_IS_X86
-
 #endif // LIBC_MATH_HAS_NO_EXCEPT
     return 0;
   }