diff options
author | Rasmus Munk Larsen <rmlarsen@google.com> | 2019-11-06 14:24:33 -0800 |
---|---|---|
committer | Rasmus Munk Larsen <rmlarsen@google.com> | 2019-11-06 14:24:33 -0800 |
commit | cc3d0e6a407d839487de13ecdd565c26c5294f9e (patch) | |
tree | a10114887c3f2bb4c8322c95b3299b8a247210b4 | |
parent | ee404667e2d3ffd60c2b39dfaf9fa5de0413f3bd (diff) |
Add EIGEN_HAS_INTRINSIC_INT128 macro
Add a new EIGEN_HAS_INTRINSIC_INT128 macro, and use this instead of __SIZEOF_INT128__. This fixes related issues with TensorIntDiv.h when building with Clang for Windows, where support for 128-bit integer arithmetic is advertised but broken in practice.
-rw-r--r-- | Eigen/src/Core/util/Macros.h | 15 | ||||
-rw-r--r-- | unsupported/Eigen/CXX11/src/Tensor/TensorIntDiv.h | 4 |
2 files changed, 17 insertions, 2 deletions
diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h index 6b5ab4278..e7bf75a81 100644 --- a/Eigen/src/Core/util/Macros.h +++ b/Eigen/src/Core/util/Macros.h @@ -757,6 +757,21 @@ #endif #endif +// Does the compiler support the __int128 and __uint128_t extensions for 128-bit +// integer arithmetic? +// +// Clang and GCC define __SIZEOF_INT128__ when these extensions are supported, +// but we avoid using them in certain cases: +// +// * Building using Clang for Windows, where the Clang runtime library has +// 128-bit support only on LP64 architectures, but Windows is LLP64. +#ifndef EIGEN_HAS_BUILTIN_INT128 +#if defined(__SIZEOF_INT128__) && !(EIGEN_OS_WIN && EIGEN_COMP_CLANG) +#define EIGEN_HAS_BUILTIN_INT128 1 +#else +#define EIGEN_HAS_BUILTIN_INT128 0 +#endif +#endif //------------------------------------------------------------------------------------------ // Preprocessor programming helpers diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorIntDiv.h b/unsupported/Eigen/CXX11/src/Tensor/TensorIntDiv.h index e7fec5d3a..6d5cce4aa 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorIntDiv.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorIntDiv.h @@ -105,7 +105,7 @@ namespace { return __umul64hi(a, b); #elif defined(SYCL_DEVICE_ONLY) return cl::sycl::mul_hi(a, static_cast<uint64_t>(b)); -#elif defined(__SIZEOF_INT128__) +#elif EIGEN_HAS_BUILTIN_INT128 __uint128_t v = static_cast<__uint128_t>(a) * static_cast<__uint128_t>(b); return static_cast<uint64_t>(v >> 64); #else @@ -124,7 +124,7 @@ namespace { template <typename T> struct DividerHelper<64, T> { static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE uint64_t computeMultiplier(const int log_div, const T divider) { -#if defined(__SIZEOF_INT128__) && !defined(EIGEN_GPU_COMPILE_PHASE) && !defined(SYCL_DEVICE_ONLY) +#if EIGEN_HAS_BUILTIN_INT128 && !defined(EIGEN_GPU_COMPILE_PHASE) && !defined(SYCL_DEVICE_ONLY) return static_cast<uint64_t>((static_cast<__uint128_t>(1) << (64+log_div)) / static_cast<__uint128_t>(divider) - (static_cast<__uint128_t>(1) << 64) + 1); #else const uint64_t shift = 1ULL << log_div; |