Merge pull request #14040 from seiko2plus:issue13211

cmake/OpenCVCompilerOptions.cmake

@@ -279,6 +279,15 @@ if((CV_GCC OR CV_CLANG)
   add_extra_compiler_option(-fvisibility-inlines-hidden)
 endif()
 
+# workaround gcc bug for aligned ld/st
+# https://github.com/opencv/opencv/issues/13211
+if((PPC64LE AND NOT CMAKE_CROSSCOMPILING) OR OPENCV_FORCE_COMPILER_CHECK_VSX_ALIGNED)
+  ocv_check_runtime_flag("${CPU_BASELINE_FLAGS}" "OPENCV_CHECK_VSX_ALIGNED" "${OpenCV_SOURCE_DIR}/cmake/checks/runtime/cpu_vsx_aligned.cpp")
+  if(NOT OPENCV_CHECK_VSX_ALIGNED)
+    add_extra_compiler_option_force(-DCV_COMPILER_VSX_BROKEN_ALIGNED)
+  endif()
+endif()
+
 # combine all "extra" options
 if(NOT OPENCV_SKIP_EXTRA_COMPILER_FLAGS)
   set(CMAKE_C_FLAGS           "${CMAKE_C_FLAGS} ${OPENCV_EXTRA_FLAGS} ${OPENCV_EXTRA_C_FLAGS}")

cmake/OpenCVUtils.cmake

@@ -480,6 +480,44 @@ macro(ocv_check_flag_support lang flag varname base_options)
   ocv_check_compiler_flag("${_lang}" "${base_options} ${flag}" ${${varname}} ${ARGN})
 endmacro()
 
+macro(ocv_check_runtime_flag flag result)
+  set(_fname "${ARGN}")
+  if(NOT DEFINED ${result})
+    file(RELATIVE_PATH _rname "${CMAKE_SOURCE_DIR}" "${_fname}")
+    message(STATUS "Performing Runtime Test ${result} (check file: ${_rname})")
+    try_run(exec_return compile_result
+      "${CMAKE_BINARY_DIR}"
+      "${_fname}"
+      CMAKE_FLAGS "-DCMAKE_EXE_LINKER_FLAGS=${CMAKE_EXE_LINKER_FLAGS}" # CMP0056 do this on new CMake
+      COMPILE_DEFINITIONS "${flag}"
+      OUTPUT_VARIABLE OUTPUT)
+
+    if(${compile_result})
+      if(exec_return EQUAL 0)
+        set(${result} 1 CACHE INTERNAL "Runtime Test ${result}")
+        message(STATUS "Performing Runtime Test ${result} - Success")
+      else()
+        message(STATUS "Performing Runtime Test ${result} - Failed(${exec_return})")
+        set(${result} 0 CACHE INTERNAL "Runtime Test ${result}")
+      endif()
+    else()
+      set(${result} 0 CACHE INTERNAL "Runtime Test ${result}")
+      message(STATUS "Performing Runtime Test ${result} - Compiling Failed")
+    endif()
+
+    if(NOT ${result})
+      file(APPEND ${CMAKE_BINARY_DIR}${CMAKE_FILES_DIRECTORY}/CMakeError.log
+        "Runtime Test failed:\n"
+        "    source file: '${_fname}'\n"
+        "    check option: '${flag}'\n"
+        "    exec return: ${exec_return}\n"
+        "===== BUILD AND RUNTIME LOG =====\n"
+        "${OUTPUT}\n"
+        "===== END =====\n\n")
+    endif()
+  endif()
+endmacro()
+
 # turns off warnings
 macro(ocv_warnings_disable)
   if(NOT ENABLE_NOISY_WARNINGS)

cmake/checks/runtime/cpu_vsx_aligned.cpp

+// check sanity of vsx aligned ld/st
+// https://github.com/opencv/opencv/issues/13211
+
+#include <altivec.h>
+
+#define vsx_ld vec_vsx_ld
+#define vsx_st vec_vsx_st
+
+template<typename T>
+static void fill(T& d, int from = 0, int to = 16)
+{
+   for (int i = from; i < to; i++)
+        d[i] = i;
+}
+
+template<typename T, typename Tvec>
+static bool check_data(T& d, Tvec& v, int from = 0, int to = 16)
+{
+    for (int i = from; i < to; i++)
+    {
+        if (d[i] != vec_extract(v, i))
+            return false;
+    }
+    return true;
+}
+
+int main()
+{
+    unsigned char __attribute__ ((aligned (16))) rbuf[16];
+    unsigned char __attribute__ ((aligned (16))) wbuf[16];
+    __vector unsigned char a;
+
+    // 1- check aligned load and store
+    fill(rbuf);
+    a = vec_ld(0, rbuf);
+    if (!check_data(rbuf, a))
+        return 1;
+    vec_st(a, 0, wbuf);
+    if (!check_data(wbuf, a))
+        return 11;
+
+    // 2- check mixing aligned load and unaligned store
+    a = vec_ld(0, rbuf);
+    vsx_st(a, 0, wbuf);
+    if (!check_data(wbuf, a))
+        return 2;
+
+    // 3- check mixing unaligned load and aligned store
+    a = vsx_ld(0, rbuf);
+    vec_st(a, 0, wbuf);
+    if (!check_data(wbuf, a))
+        return 3;
+
+    return 0;
+}
\ No newline at end of file

modules/core/include/opencv2/core/hal/intrin_vsx.hpp

@@ -258,8 +258,16 @@ inline void v_store_low(_Tp* ptr, const _Tpvec& a)                          \
 inline void v_store_high(_Tp* ptr, const _Tpvec& a)                         \
 { vec_st_h8(a.val, ptr); }
 
-#define OPENCV_HAL_IMPL_VSX_LOADSTORE(_Tpvec, _Tp) \
-OPENCV_HAL_IMPL_VSX_LOADSTORE_C(_Tpvec, _Tp, vsx_ld, vec_ld, vsx_st, vec_st)
+// working around gcc bug for aligned ld/st
+// if runtime check for vec_ld/st fail we failback to unaligned ld/st
+// https://github.com/opencv/opencv/issues/13211
+#ifdef CV_COMPILER_VSX_BROKEN_ALIGNED
+    #define OPENCV_HAL_IMPL_VSX_LOADSTORE(_Tpvec, _Tp) \
+    OPENCV_HAL_IMPL_VSX_LOADSTORE_C(_Tpvec, _Tp, vsx_ld, vsx_ld, vsx_st, vsx_st)
+#else
+    #define OPENCV_HAL_IMPL_VSX_LOADSTORE(_Tpvec, _Tp) \
+    OPENCV_HAL_IMPL_VSX_LOADSTORE_C(_Tpvec, _Tp, vsx_ld, vec_ld, vsx_st, vec_st)
+#endif
 
 OPENCV_HAL_IMPL_VSX_LOADSTORE(v_uint8x16,  uchar)
 OPENCV_HAL_IMPL_VSX_LOADSTORE(v_int8x16,   schar)