Merge pull request #2 from scchan/gpu_link_static_lib

GPU link - support linking static library

lib/clamp-link.in

@@ -119,6 +119,230 @@ _find_object() {
 
 ARGS="$@"
 
+# flag to switch to the new arg parsing algorithm
+USE_NEW_LINK=1
+
+NEW_LINK_KERNEL_ARGS=""
+NEW_LINK_HOST_ARGS=""
+NEW_LINK_OTHER_ARGS=""
+NEW_LINK_CPU_ARG=""
+
+TEMP_AR_DIRS=()
+
+if [ $USE_NEW_LINK == 1 ]; then
+
+# detect the verbose flags before doing anything
+if [[ "$ARGS" =~ --verbose ]]; then
+  VERBOSE=1
+fi
+#very verbose
+#VERBOSE=2
+
+
+
+# gather a list of library search paths
+LIB_SEARCH_PATHS=()
+for ARG in $ARGS
+do
+  # matches -L<path>
+  if [[ "$ARG" =~ ^-L[^[:space:]]+$ ]]; then
+
+    if [ $VERBOSE == 2 ]; then
+      echo "adding library search path: ${ARG:2}"
+    fi
+
+    if [[ "$ARG" =~ ^-L\/[^[:space:]]+$ ]]; then
+      # this is an absolute path
+      LIB_SEARCH_PATHS+=( ${ARG:2} )
+    else
+      LIB_SEARCH_PATHS+=( $PWD"/"${ARG:2} )
+    fi
+  fi
+done
+
+
+
+
+for ARG in $ARGS
+do
+
+  ####################################
+  #  parse the simple switches first...
+  ####################################
+
+  if [ $ARG == "--disable-opencl" ]; then
+    LOWER_OPENCL=0
+    continue
+  fi
+
+  if [ $ARG == "--disable-hsa" ]; then
+    LOWER_HSA=0
+    continue
+  fi
+
+
+
+  #####################################
+  # detect object or static library
+  #####################################
+
+  OBJS_TO_PROCESS=()
+
+
+  if [[ "$ARG" =~ [^[:space:]]+\.cpu$ ]]; then
+
+    cp $ARG $TEMP_DIR/kernel_cpu.o
+    NEW_LINK_CPU_ARG=$NEW_LINK_CPU_ARG" "$TEMP_DIR/kernel_cpu.o
+
+  elif [[ "$ARG" =~ [^[:space:]]+\.o$ ]]; then
+    # detected a .o file
+    if [ $VERBOSE == 2 ]; then
+      echo "detect object file to process further: $ARG"
+    fi
+
+    OBJS_TO_PROCESS+=( $ARG )
+  elif [[ "$ARG" =~ ^-l[^[:space:]]+$ ]] || [[ "$ARG" =~ [^[:space:]]+.a$ ]]; then
+
+    # proccess a static library
+
+    DETECTED_STATIC_LIBRARY=""
+
+    # detected whether it's an -l<library naem> option
+    if [[ "$ARG" =~ ^-l[^[:space:]]+$ ]]; then
+
+      # expand the option into a library name
+      STATIC_LIB_NAME="lib${ARG:2}.a"
+
+      if [ $VERBOSE == 2 ]; then
+        echo "looking for static library $STATIC_LIB_NAME"
+      fi
+
+      # look for the static library in the library search paths
+      for LIB_PATH in "${LIB_SEARCH_PATHS[@]}"
+      do
+        FULL_LIB_PATH=$LIB_PATH"/"$STATIC_LIB_NAME
+        if [ -f $FULL_LIB_PATH ]; then
+          if [ $VERBOSE == 2 ]; then
+            echo "$FULL_LIB_PATH detected"
+          fi
+          DETECTED_STATIC_LIBRARY=$FULL_LIB_PATH;
+          break;
+        fi
+      done
+    else
+      # this is .a static library file specified at the commad line
+      if [ -f $ARG ]; then
+        DETECTED_STATIC_LIBRARY=$ARG
+      fi
+    fi  # if [[ "$ARG" =~ ^-l[^[:space:]]+$ ]]; then
+
+
+    KERNEL_UNDETECTED="1" 
+    if [[ $DETECTED_STATIC_LIBRARY != "" ]]; then
+
+      # we found a static library library
+      if [ $VERBOSE == 2 ]; then
+        echo "processing static library $DETECTED_STATIC_LIBRARY"
+      fi
+
+      # detect whether the objects in the static library contain a .kernel section
+      KERNEL_UNDETECTED=`objdump -t $DETECTED_STATIC_LIBRARY | grep -q "\.kernel"; echo $?`
+      if [[ $KERNEL_UNDETECTED == "0" ]]; then
+
+        # .kernel section detected, extract the objects from the archieve
+
+        if [ $VERBOSE == 2 ]; then
+          echo "kernel detected in $DETECTED_STATIC_LIBRARY"
+        fi
+
+        CURRENT_DIR=$PWD
+        # extract the archive
+        FILE=`basename $DETECTED_STATIC_LIBRARY`
+        AR_TEMP_DIR=$TEMP_DIR"/"$FILE
+
+        if [ $VERBOSE == 2 ]; then
+          echo "creating temp dir: $AR_TEMP_DIR"
+        fi
+
+        mkdir -p $AR_TEMP_DIR
+        TEMP_AR_DIRS+=( $AR_TEMP_DIR )
+        cd $AR_TEMP_DIR
+        `ar x $DETECTED_STATIC_LIBRARY`
+
+        cd $CURRENT_DIR
+
+        # store all the extract objects to process further
+        OBJS_TO_PROCESS=($(ls $AR_TEMP_DIR/*.o))
+
+      fi  # if [[ $KERNEL_UNDETECTED == "0" ]]; then
+    fi  # if [[ $DETECTED_STATIC_LIBRARY != "" ]]; then
+  elif [ -f $ARG ]; then
+    # an object file but doesn't have an .o extension??
+    file_output=`file $ARG | grep 'ELF 64-bit LSB  relocatable, x86-64'`
+    readelf_output=`readelf -h $ARG | grep 'Relocatable file'`
+    if [ ! -z "$file_output" ] && [ ! -z "$readelf_output" ]; then
+      OBJS_TO_PROCESS+=( $ARG )
+    fi
+  fi
+
+
+  # no objects to further process, pass the original args down to the host linker
+  if [ ${#OBJS_TO_PROCESS[@]} == 0 ]; then
+    # no objects to further process, pass the original args down to the host linker
+    if [ $VERBOSE == 2 ]; then
+      echo "passing down link args: $ARG"
+    fi
+    NEW_LINK_OTHER_ARGS=$NEW_LINK_OTHER_ARGS" "$ARG
+    continue
+  fi
+
+  # processs the objects we put aside
+  for OBJ in "${OBJS_TO_PROCESS[@]}"
+  do
+    if [ $VERBOSE == 2 ]; then
+      echo "processing $OBJ"
+    fi
+
+    # detect whether the objects in the static library contain a .kernel section
+    KERNEL_UNDETECTED=`objdump -t $OBJ | grep -q "\.kernel"; echo $?`
+    if [[ $KERNEL_UNDETECTED == "0" ]]; then
+
+      FILE=`basename $OBJ` # remove path
+      FILENAME=${FILE%.*}
+      KERNEL_FILE=$TEMP_DIR/$FILENAME.kernel.bc
+      HOST_FILE=$TEMP_DIR/$FILENAME.host.o
+
+      # extract kernel section
+      objcopy -O binary -j .kernel $OBJ $KERNEL_FILE 
+
+      # extract host section
+      objcopy -R .kernel $OBJ $HOST_FILE
+
+      # strip all symbols specified in symbol.txt from $HOST_FILE
+      objcopy @$CXXAMP_SERIALIZE_SYMBOL_FILE $HOST_FILE $HOST_FILE.new 2> /dev/null
+      if [ -f $HOST_FILE.new ]; then
+        mv $HOST_FILE.new $HOST_FILE
+      fi
+
+      # find cxxamp_serialize symbols and save them into symbol.txt
+      objdump -t $HOST_FILE -j .text 2> /dev/null | grep "g.*__cxxamp_serialize" | awk '{print "-L"$6}' >> $CXXAMP_SERIALIZE_SYMBOL_FILE
+
+      NEW_LINK_KERNEL_ARGS=$NEW_LINK_KERNEL_ARGS" "$KERNEL_FILE
+      NEW_LINK_HOST_ARGS=$NEW_LINK_HOST_ARGS" "$HOST_FILE
+    else
+      NEW_LINK_OTHER_ARGS=$NEW_LINK_OTHER_ARGS" "$OBJ
+    fi
+
+
+  done # for OBJ in "${OBJS_TO_PROCESS[@]}"
+done
+
+else #  if [ $USE_NEW_LINK == 1 ]; then
+
+############################
+# Old argument processing
+############################
+
 for ARG in $ARGS
 do
   if [ -f $ARG ]; then
@@ -127,6 +351,7 @@ do
     CPUFILE=${FILE%.cpu}
     ISCRT=${ARG#/usr}    # exception for objects under /usr
     ISLIB=${ARG#/lib}    # exception for objects under /lib
+
     if [ $FILENAME != $FILE ] && [ $ISCRT == $ARG ] && [ $ISLIB == $ARG ]; then
       KERNEL_FILE=$TEMP_DIR/$FILENAME.kernel.bc
       HOST_FILE=$TEMP_DIR/$FILENAME.host.o
@@ -164,9 +389,43 @@ do
     LINK_OTHER_ARGS=$LINK_OTHER_ARGS" "$ARG
   fi
 done
-#echo "kernel args:"$LINK_KERNEL_ARGS
-#echo "host args:"$LINK_HOST_ARGS
-#echo "other args:"$LINK_OTHER_ARGS
+
+if [ $VERBOSE == 1 ]; then
+  echo "kernel args: "$LINK_KERNEL_ARGS
+  echo ""
+  echo "host args: "$LINK_HOST_ARGS
+  echo ""
+  echo "other args: "$LINK_OTHER_ARGS
+  echo ""
+fi
+
+###################################
+# End of old argument processing
+###################################
+
+fi
+
+
+
+if [ $USE_NEW_LINK == 1 ]; then
+
+if [ $VERBOSE == 1 ]; then
+  echo "new kernel args: "$NEW_LINK_KERNEL_ARGS
+  echo ""
+  echo "new host args: "$NEW_LINK_HOST_ARGS
+  echo ""
+  echo "new other args: "$NEW_LINK_OTHER_ARGS
+fi
+
+if [ $VERBOSE == 1 ]; then
+  echo "replacing old link args with new link args"
+fi
+LINK_KERNEL_ARGS=$NEW_LINK_KERNEL_ARGS
+LINK_HOST_ARGS=$NEW_LINK_HOST_ARGS
+LINK_OTHER_ARGS=$NEW_LINK_OTHER_ARGS
+
+fi
+
 
 # linker return value
 ret=0
@@ -339,6 +598,11 @@ if [ -e $CXXAMP_SERIALIZE_SYMBOL_FILE ]; then
   rm $CXXAMP_SERIALIZE_SYMBOL_FILE # __cxxamp_serialize symbols
 fi
 
+for TD in "${TEMP_AR_DIRS[@]}"
+do
+  rm -rf $TD
+done
+
 if [ -d $TEMP_DIR ]; then
   rm -f $TEMP_DIR/*
   rmdir $TEMP_DIR

tests/Unit/StaticLibrary/static_library1.cpp

@@ -0,0 +1,44 @@
+// RUN: %hc -DSTATIC_LIB %s -c -o %T/static_library1.o
+// RUN: ar rcs %T/libstatic_library1.a %T/static_library1.o
+// RUN: %hc %s -L%T -lstatic_library1 -o %t.out && %t.out
+
+#include <cstdio>
+#include <hc.hpp>
+
+extern "C" int sum(hc::array_view<int,1>& input);
+
+#ifdef STATIC_LIB
+
+int sum(hc::array_view<int,1>& input) {
+
+  hc::array_view<int,1> s(1);
+  s[0]=0;
+
+  hc::parallel_for_each(input.get_extent(), [=](hc::index<1> idx) [[hc]] {
+    if (idx[0]==0) {
+      int num = input.get_extent()[0];
+      for (int i = 0; i < num; i++) {
+        s[0]+=input[i];
+      }
+    }
+  }).wait();
+
+  return s[0];
+}
+
+#else
+
+int main() {
+
+  hc::array_view<int,1> av(64);
+  for (int i = 0;i < 64; i++)
+    av[i] = i;
+
+  int s = sum(av);
+
+ // printf("sum: %d\n",s);
+
+  return !(s==2016);
+}
+
+#endif