Merge pull request #5 from xtensor-stack/master

Pull latest changes from upstream

.github/workflows/linux.yml

@@ -102,14 +102,17 @@ jobs:
           CMAKE_EXTRA_ARGS="$CMAKE_EXTRA_ARGS -DTARGET_ARCH=knm"
         fi
         if [[ '${{ matrix.sys.flags }}' == 'i386' ]]; then
-          CMAKE_EXTRA_ARGS="$CMAKE_EXTRA_ARGS -DCMAKE_CXX_FLAGS='-m32'"
+          CXX_FLAGS="$CXX_FLAGS -m32"
         fi
         if [[ '${{ matrix.sys.flags }}' == 'force_no_instr_set' ]]; then
           :
         else
           CMAKE_EXTRA_ARGS="$CMAKE_EXTRA_ARGS -DXSIMD_ENABLE_WERROR=ON"
         fi
 
+        # Cheap way of spotting uninitialized read
+        CXX_FLAGS="$CXX_FLAGS -ftrivial-auto-var-init=pattern"
+
         mkdir _build
         cd _build
         cmake  .. -DBUILD_TESTS=ON \
@@ -119,6 +122,7 @@ jobs:
                   -DCMAKE_C_COMPILER=$CC \
                   -DCMAKE_CXX_COMPILER=$CXX \
                   $CMAKE_EXTRA_ARGS \
+                  -DCMAKE_CXX_FLAGS='$CXX_FLAGS' \
                   -G Ninja
     - name: Build
       run: ninja -C _build

CMakeLists.txt

@@ -139,7 +139,7 @@ install(DIRECTORY ${XSIMD_INCLUDE_DIR}/xsimd
         DESTINATION ${CMAKE_INSTALL_INCLUDEDIR})
 
 # GNUInstallDirs "DATADIR" wrong here; CMake search path wants "share".
-set(XSIMD_CMAKECONFIG_INSTALL_DIR "${CMAKE_INSTALL_LIBDIR}/cmake/${PROJECT_NAME}" CACHE STRING "install path for xsimdConfig.cmake")
+set(XSIMD_CMAKECONFIG_INSTALL_DIR "${CMAKE_INSTALL_DATADIR}/cmake/${PROJECT_NAME}" CACHE STRING "install path for xsimdConfig.cmake")
 
 configure_package_config_file(${PROJECT_NAME}Config.cmake.in
                               "${CMAKE_CURRENT_BINARY_DIR}/${PROJECT_NAME}Config.cmake"
@@ -164,4 +164,4 @@ configure_file(${PROJECT_NAME}.pc.in
                "${CMAKE_CURRENT_BINARY_DIR}/${PROJECT_NAME}.pc"
                 @ONLY)
 install(FILES "${CMAKE_CURRENT_BINARY_DIR}/${PROJECT_NAME}.pc"
-        DESTINATION "${CMAKE_INSTALL_LIBDIR}/pkgconfig/")
+        DESTINATION "${CMAKE_INSTALL_DATADIR}/pkgconfig/")

README.md

@@ -80,7 +80,7 @@ spack load xsimd
 You can directly install it from the sources with cmake:
 
 ```bash
-cmake -D CMAKE_INSTALL_PREFIX=your_install_prefix
+cmake -D CMAKE_INSTALL_PREFIX=your_install_prefix .
 make install
 ```
 

include/xsimd/arch/generic/xsimd_generic_memory.hpp

@@ -459,6 +459,7 @@ namespace xsimd
         XSIMD_INLINE batch<T, A> shuffle(batch<T, A> const& x, batch<T, A> const& y, batch_constant<ITy, A, Indices...>, requires_arch<generic>) noexcept
         {
             constexpr size_t bsize = sizeof...(Indices);
+            static_assert(bsize == batch<T, A>::size, "valid shuffle");
 
             // Detect common patterns
             XSIMD_IF_CONSTEXPR(detail::is_swizzle_fst(bsize, Indices...))
@@ -486,14 +487,15 @@ namespace xsimd
                 return select(batch_bool_constant<T, A, (Indices < bsize)...>(), x, y);
             }
 
-#if defined(__has_builtin)
-#if __has_builtin(__builtin_shuffle_vector)
-#define builtin_shuffle __builtin_shuffle_vector
+#if defined(__has_builtin) && !defined(XSIMD_WITH_EMULATED)
+#if __has_builtin(__builtin_shufflevector)
+#define builtin_shuffle __builtin_shufflevector
 #endif
 #endif
 
 #if defined(builtin_shuffle)
-            return builtin_shuffle(x.data, y.data, Indices...);
+            typedef T vty __attribute__((__vector_size__(sizeof(batch<T, A>))));
+            return (typename batch<T, A>::register_type)builtin_shuffle((vty)x.data, (vty)y.data, Indices...);
 
 // FIXME: my experiments show that GCC only correctly optimizes this builtin
 // starting at GCC 13, where it already has __builtin_shuffle_vector

include/xsimd/config/xsimd_cpuid.hpp

@@ -42,6 +42,10 @@ namespace xsimd
 #define ARCH_FIELD_EX(arch, field_name) \
     unsigned field_name;                \
     XSIMD_INLINE bool has(::xsimd::arch) const { return this->field_name; }
+
+#define ARCH_FIELD_EX_REUSE(arch, field_name) \
+    XSIMD_INLINE bool has(::xsimd::arch) const { return this->field_name; }
+
 #define ARCH_FIELD(name) ARCH_FIELD_EX(name, name)
 
             ARCH_FIELD(sse2)
@@ -72,8 +76,12 @@ namespace xsimd
             ARCH_FIELD(neon)
             ARCH_FIELD(neon64)
             ARCH_FIELD_EX(i8mm<::xsimd::neon64>, i8mm_neon64)
-            ARCH_FIELD(sve)
-            ARCH_FIELD(rvv)
+            ARCH_FIELD_EX(detail::sve<512>, sve)
+            ARCH_FIELD_EX_REUSE(detail::sve<256>, sve)
+            ARCH_FIELD_EX_REUSE(detail::sve<128>, sve)
+            ARCH_FIELD_EX(detail::rvv<512>, rvv)
+            ARCH_FIELD_EX_REUSE(detail::rvv<256>, rvv)
+            ARCH_FIELD_EX_REUSE(detail::rvv<128>, rvv)
             ARCH_FIELD(wasm)
 
 #undef ARCH_FIELD
@@ -114,6 +122,35 @@ namespace xsimd
 #endif
 
 #elif defined(__x86_64__) || defined(__i386__) || defined(_M_AMD64) || defined(_M_IX86)
+
+                auto get_xcr0_low = []() noexcept
+                {
+                    uint32_t xcr0;
+
+#if defined(_MSC_VER) && _MSC_VER >= 1400
+
+                    xcr0 = (uint32_t)_xgetbv(0);
+
+#elif defined(__GNUC__)
+
+                    __asm__(
+                        "xorl %%ecx, %%ecx\n"
+                        "xgetbv\n"
+                        : "=a"(xcr0)
+                        :
+#if defined(__i386__)
+                        : "ecx", "edx"
+#else
+                        : "rcx", "rdx"
+#endif
+                    );
+
+#else /* _MSC_VER < 1400 */
+#error "_MSC_VER < 1400 is not supported"
+#endif /* _MSC_VER && _MSC_VER >= 1400 */
+                    return xcr0;
+                };
+
                 auto get_cpuid = [](int reg[4], int level, int count = 0) noexcept
                 {
 
@@ -148,43 +185,67 @@ namespace xsimd
 
                 get_cpuid(regs1, 0x1);
 
-                sse2 = regs1[3] >> 26 & 1;
-                sse3 = regs1[2] >> 0 & 1;
-                ssse3 = regs1[2] >> 9 & 1;
-                sse4_1 = regs1[2] >> 19 & 1;
-                sse4_2 = regs1[2] >> 20 & 1;
-                fma3_sse42 = regs1[2] >> 12 & 1;
+                // OS can explicitly disable the usage of SSE/AVX extensions
+                // by setting an appropriate flag in CR0 register
+                //
+                // https://docs.kernel.org/admin-guide/hw-vuln/gather_data_sampling.html
+
+                unsigned sse_state_os_enabled = 1;
+                unsigned avx_state_os_enabled = 1;
+                unsigned avx512_state_os_enabled = 1;
+
+                // OSXSAVE: A value of 1 indicates that the OS has set CR4.OSXSAVE[bit
+                // 18] to enable XSETBV/XGETBV instructions to access XCR0 and
+                // to support processor extended state management using
+                // XSAVE/XRSTOR.
+                bool osxsave = regs1[2] >> 27 & 1;
+                if (osxsave)
+                {
+
+                    uint32_t xcr0 = get_xcr0_low();
+
+                    sse_state_os_enabled = xcr0 >> 1 & 1;
+                    avx_state_os_enabled = xcr0 >> 2 & sse_state_os_enabled;
+                    avx512_state_os_enabled = xcr0 >> 6 & avx_state_os_enabled;
+                }
+
+                sse2 = regs1[3] >> 26 & sse_state_os_enabled;
+                sse3 = regs1[2] >> 0 & sse_state_os_enabled;
+                ssse3 = regs1[2] >> 9 & sse_state_os_enabled;
+                sse4_1 = regs1[2] >> 19 & sse_state_os_enabled;
+                sse4_2 = regs1[2] >> 20 & sse_state_os_enabled;
+                fma3_sse42 = regs1[2] >> 12 & sse_state_os_enabled;
 
-                avx = regs1[2] >> 28 & 1;
+                avx = regs1[2] >> 28 & avx_state_os_enabled;
                 fma3_avx = avx && fma3_sse42;
 
                 int regs8[4];
                 get_cpuid(regs8, 0x80000001);
-                fma4 = regs8[2] >> 16 & 1;
+                fma4 = regs8[2] >> 16 & avx_state_os_enabled;
 
                 // sse4a = regs[2] >> 6 & 1;
 
                 // xop = regs[2] >> 11 & 1;
 
                 int regs7[4];
                 get_cpuid(regs7, 0x7);
-                avx2 = regs7[1] >> 5 & 1;
+                avx2 = regs7[1] >> 5 & avx_state_os_enabled;
 
                 int regs7a[4];
                 get_cpuid(regs7a, 0x7, 0x1);
-                avxvnni = regs7a[0] >> 4 & 1;
+                avxvnni = regs7a[0] >> 4 & avx_state_os_enabled;
 
                 fma3_avx2 = avx2 && fma3_sse42;
 
-                avx512f = regs7[1] >> 16 & 1;
-                avx512cd = regs7[1] >> 28 & 1;
-                avx512dq = regs7[1] >> 17 & 1;
-                avx512bw = regs7[1] >> 30 & 1;
-                avx512er = regs7[1] >> 27 & 1;
-                avx512pf = regs7[1] >> 26 & 1;
-                avx512ifma = regs7[1] >> 21 & 1;
-                avx512vbmi = regs7[2] >> 1 & 1;
-                avx512vnni_bw = regs7[2] >> 11 & 1;
+                avx512f = regs7[1] >> 16 & avx512_state_os_enabled;
+                avx512cd = regs7[1] >> 28 & avx512_state_os_enabled;
+                avx512dq = regs7[1] >> 17 & avx512_state_os_enabled;
+                avx512bw = regs7[1] >> 30 & avx512_state_os_enabled;
+                avx512er = regs7[1] >> 27 & avx512_state_os_enabled;
+                avx512pf = regs7[1] >> 26 & avx512_state_os_enabled;
+                avx512ifma = regs7[1] >> 21 & avx512_state_os_enabled;
+                avx512vbmi = regs7[2] >> 1 & avx512_state_os_enabled;
+                avx512vnni_bw = regs7[2] >> 11 & avx512_state_os_enabled;
                 avx512vnni_vbmi = avx512vbmi && avx512vnni_bw;
 #endif
             }

test/test_arch.cpp

@@ -38,6 +38,16 @@ struct check_supported
     }
 };
 
+struct check_cpu_has_intruction_set
+{
+    template <class Arch>
+    void operator()(Arch arch) const
+    {
+        static_assert(std::is_same<decltype(xsimd::available_architectures().has(arch)), bool>::value,
+                      "cannot test instruction set availability on CPU");
+    }
+};
+
 struct check_available
 {
     template <class Arch>
@@ -71,6 +81,11 @@ TEST_CASE("[multi arch support]")
         xsimd::supported_architectures::for_each(check_supported {});
     }
 
+    SUBCASE("xsimd::available_architectures::has")
+    {
+        xsimd::all_architectures::for_each(check_cpu_has_intruction_set {});
+    }
+
     SUBCASE("xsimd::default_arch::name")
     {
         constexpr char const* name = xsimd::default_arch::name();