diff --git a/CMakeLists.txt b/CMakeLists.txt
index b02c9c24..e1ee0820 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -74,12 +74,12 @@ install(FILES "${CMAKE_CURRENT_BINARY_DIR}/roaring.pc" DESTINATION ${CMAKE_INSTA
 
 add_library(roaring-headers INTERFACE)
 target_include_directories(roaring-headers INTERFACE
-  $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/include/roaring>
-  $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCDIR}>)
+  $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/include>
+  $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}>)
 add_library(roaring-headers-cpp INTERFACE)
 target_include_directories(roaring-headers-cpp INTERFACE
   $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/cpp>
-   $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCDIR}>)
+   $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}>)
 
 ####
 ### Some users want the C++ header files to be installed as well.
@@ -93,7 +93,7 @@ install(TARGETS roaring-headers roaring-headers-cpp
    ARCHIVE DESTINATION ${CMAKE_INSTALL_LIBDIR}
    LIBRARY DESTINATION ${CMAKE_INSTALL_LIBDIR}
    RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR}
-   INCLUDES DESTINATION ${CMAKE_INSTALL_INCDIR})
+   INCLUDES DESTINATION ${CMAKE_INSTALL_INCLUDEDIR})
 include(CTest)
 
 ##################################
diff --git a/include/roaring/roaring64.h b/include/roaring/roaring64.h
index fd89feb5..fea6ef90 100644
--- a/include/roaring/roaring64.h
+++ b/include/roaring/roaring64.h
@@ -1,13 +1,13 @@
 #ifndef ROARING64_H
 #define ROARING64_H
 
-#include <roaring.h>
 #include <stdbool.h>
 #include <stddef.h>
 #include <stdint.h>
 
 #include <roaring/memory.h>
 #include <roaring/portability.h>
+#include <roaring/roaring.h>
 #include <roaring/roaring_types.h>
 
 #ifdef __cplusplus
diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
index 543c978f..a24e8aa9 100644
--- a/src/CMakeLists.txt
+++ b/src/CMakeLists.txt
@@ -59,12 +59,6 @@ if(ROARING_DISABLE_NEON)
   target_compile_definitions(roaring PUBLIC DISABLENEON=1)
 endif(ROARING_DISABLE_NEON)
 
-
-target_include_directories(roaring
-  PUBLIC
-   $<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/include>
-   $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}>
-)
 target_link_libraries(roaring PUBLIC roaring-headers)
 target_link_libraries(roaring PUBLIC roaring-headers-cpp)
 #
@@ -75,7 +69,7 @@ install(TARGETS roaring
    ARCHIVE DESTINATION ${CMAKE_INSTALL_LIBDIR}
    LIBRARY DESTINATION ${CMAKE_INSTALL_LIBDIR}
    RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR}
-   INCLUDES DESTINATION ${CMAKE_INSTALL_INCDIR}
+   INCLUDES DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}
 )
 install(EXPORT roaring-config
    FILE roaring-config.cmake
diff --git a/src/bitset_util.c b/src/bitset_util.c
index 0ae7d925..6bc12b44 100644
--- a/src/bitset_util.c
+++ b/src/bitset_util.c
@@ -613,16 +613,13 @@ size_t bitset_extract_setbits_avx512(const uint64_t *words, size_t length,
     for (; (i < length) && (out < safeout); ++i) {
         uint64_t w = words[i];
         while ((w != 0) && (out < safeout)) {
-            uint64_t t =
-                w & (~w + 1);  // on x64, should compile to BLSI (careful: the
-                               // Intel compiler seems to fail)
             int r =
                 roaring_trailing_zeroes(w);  // on x64, should compile to TZCNT
             uint32_t val = r + base;
             memcpy(out, &val,
                    sizeof(uint32_t));  // should be compiled as a MOV on x64
             out++;
-            w ^= t;
+            w &= (w - 1);
         }
         base += 64;
     }
@@ -667,15 +664,12 @@ size_t bitset_extract_setbits_avx512_uint16(const uint64_t *array,
     for (; (i < length) && (out < safeout); ++i) {
         uint64_t w = array[i];
         while ((w != 0) && (out < safeout)) {
-            uint64_t t =
-                w & (~w + 1);  // on x64, should compile to BLSI (careful: the
-                               // Intel compiler seems to fail)
             int r =
                 roaring_trailing_zeroes(w);  // on x64, should compile to TZCNT
             uint32_t val = r + base;
             memcpy(out, &val, sizeof(uint16_t));
             out++;
-            w ^= t;
+            w &= (w - 1);
         }
         base += 64;
     }
@@ -725,16 +719,13 @@ size_t bitset_extract_setbits_avx2(const uint64_t *words, size_t length,
     for (; (i < length) && (out < safeout); ++i) {
         uint64_t w = words[i];
         while ((w != 0) && (out < safeout)) {
-            uint64_t t =
-                w & (~w + 1);  // on x64, should compile to BLSI (careful: the
-                               // Intel compiler seems to fail)
             int r =
                 roaring_trailing_zeroes(w);  // on x64, should compile to TZCNT
             uint32_t val = r + base;
             memcpy(out, &val,
                    sizeof(uint32_t));  // should be compiled as a MOV on x64
             out++;
-            w ^= t;
+            w &= (w - 1);
         }
         base += 64;
     }
@@ -749,16 +740,13 @@ size_t bitset_extract_setbits(const uint64_t *words, size_t length,
     for (size_t i = 0; i < length; ++i) {
         uint64_t w = words[i];
         while (w != 0) {
-            uint64_t t =
-                w & (~w + 1);  // on x64, should compile to BLSI (careful: the
-                               // Intel compiler seems to fail)
             int r =
                 roaring_trailing_zeroes(w);  // on x64, should compile to TZCNT
             uint32_t val = r + base;
             memcpy(out + outpos, &val,
                    sizeof(uint32_t));  // should be compiled as a MOV on x64
             outpos++;
-            w ^= t;
+            w &= (w - 1);
         }
         base += 64;
     }
@@ -772,10 +760,9 @@ size_t bitset_extract_intersection_setbits_uint16(
     for (size_t i = 0; i < length; ++i) {
         uint64_t w = words1[i] & words2[i];
         while (w != 0) {
-            uint64_t t = w & (~w + 1);
             int r = roaring_trailing_zeroes(w);
             out[outpos++] = (uint16_t)(r + base);
-            w ^= t;
+            w &= (w - 1);
         }
         base += 64;
     }
@@ -836,11 +823,10 @@ size_t bitset_extract_setbits_sse_uint16(const uint64_t *words, size_t length,
     for (; (i < length) && (out < safeout); ++i) {
         uint64_t w = words[i];
         while ((w != 0) && (out < safeout)) {
-            uint64_t t = w & (~w + 1);
             int r = roaring_trailing_zeroes(w);
             *out = (uint16_t)(r + base);
             out++;
-            w ^= t;
+            w &= (w - 1);
         }
         base += 64;
     }
@@ -864,10 +850,9 @@ size_t bitset_extract_setbits_uint16(const uint64_t *words, size_t length,
     for (size_t i = 0; i < length; ++i) {
         uint64_t w = words[i];
         while (w != 0) {
-            uint64_t t = w & (~w + 1);
             int r = roaring_trailing_zeroes(w);
             out[outpos++] = (uint16_t)(r + base);
-            w ^= t;
+            w &= (w - 1);
         }
         base += 64;
     }
@@ -1158,4 +1143,4 @@ void bitset_flip_list(uint64_t *words, const uint16_t *list, uint64_t length) {
 #endif
 #if defined(__GNUC__) && !defined(__clang__)
 #pragma GCC diagnostic pop
-#endif
\ No newline at end of file
+#endif