编译impala、拓展impala语法解析模块

以前也编译过，但是每次编译都忘记怎么做，然后都得重新找需要下载的文件。

编译文件：buildall.sh

如果想只编译前端可以这样运行：

buildall.sh -fe_only

编译时会去S3下载一些文件，由于在国外下载很慢，所以可以在本地开ss去下载好再上传到编译服务器上

那么会下载哪些东西呢？
编辑bin/bootstrap_toolchain.py
找到下面这几句话

def wget_and_unpack_package(download_path, file_name, destination, wget_no_clobber):
  print "URL {0}".format(download_path)
  print "Downloading {0} to {1}".format(file_name, destination)
  # --no-clobber avoids downloading the file if a file with the name already exists

  sh.wget(download_path, directory_prefix=destination, no_clobber=wget_no_clobber)
  print "Extracting {0}".format(file_name)
  sh.tar(z=True, x=True, f=os.path.join(destination, file_name), directory=destination)
  sh.rm(os.path.join(destination, file_name))

把后面4行注释掉，就不会去真正下载了:

def wget_and_unpack_package(download_path, file_name, destination, wget_no_clobber):
  print "URL {0}".format(download_path)
  print "Downloading {0} to {1}".format(file_name, destination)
  # --no-clobber avoids downloading the file if a file with the name already exists
  """
  sh.wget(download_path, directory_prefix=destination, no_clobber=wget_no_clobber)
  print "Extracting {0}".format(file_name)
  sh.tar(z=True, x=True, f=os.path.join(destination, file_name), directory=destination)
  sh.rm(os.path.join(destination, file_name))
  """

然后找到这段话：

def bootstrap(toolchain_root, packages):
  """Downloads and unpacks each package in the list `packages` into `toolchain_root` if it
  doesn't exist already.
  """
  if not try_get_platform_release_label():
    check_custom_toolchain(toolchain_root, packages)
    return

  # Detect the compiler
  compiler = "gcc-{0}".format(os.environ["IMPALA_GCC_VERSION"])

  for p in packages:
    pkg_name, pkg_version = unpack_name_and_version(p)
    if check_for_existing_package(toolchain_root, pkg_name, pkg_version, compiler):
      continue
    if pkg_name != "kudu" or os.environ["KUDU_IS_SUPPORTED"] == "true":
      download_package(toolchain_root, pkg_name, pkg_version, compiler)
    else:
      build_kudu_stub(toolchain_root, pkg_version, compiler)
    write_version_file(toolchain_root, pkg_name, pkg_version, compiler,
        get_platform_release_label())

把最后一句话注释掉：

def bootstrap(toolchain_root, packages):
  """Downloads and unpacks each package in the list `packages` into `toolchain_root` if it
  doesn't exist already.
  """
  if not try_get_platform_release_label():
    check_custom_toolchain(toolchain_root, packages)
    return

  # Detect the compiler
  compiler = "gcc-{0}".format(os.environ["IMPALA_GCC_VERSION"])

  for p in packages:
    pkg_name, pkg_version = unpack_name_and_version(p)
    if check_for_existing_package(toolchain_root, pkg_name, pkg_version, compiler):
      continue
    if pkg_name != "kudu" or os.environ["KUDU_IS_SUPPORTED"] == "true":
      download_package(toolchain_root, pkg_name, pkg_version, compiler)
    else:
      build_kudu_stub(toolchain_root, pkg_version, compiler)
    """
    write_version_file(toolchain_root, pkg_name, pkg_version, compiler,
        get_platform_release_label())
    """

运行buildall.sh后屏幕就会打印出需要下载的东西，上传到toolchain文件夹就行。上传结束后再把刚才注释的代码恢复就好

加速编译前端

运行下面的命令，去掉测试，只编译前端代码

./buildall.sh -skiptests -fe_only

上面命令运行成功后，找到infra/python/deps/pip_download.py的下段代码

def download_package(pkg_name, pkg_version):
  '''Download the required package. Sometimes the download can be flaky, so we use the
  retry decorator.'''
  pkg_type = 'sdist' # Don't download wheel archives for now
  # This JSON endpoint is not provided by PyPI mirrors so we always need to get this
  # from pypi.python.org.
  pkg_info = json.loads(urlopen('https://pypi.python.org/pypi/%s/json' % pkg_name).read())

  downloader = URLopener()
  for pkg in pkg_info['releases'][pkg_version]:
    if pkg['packagetype'] == pkg_type:
      filename = pkg['filename']
      expected_md5 = pkg['md5_digest']
      if os.path.isfile(filename) and check_md5sum(filename, expected_md5):
        print "File with matching md5sum already exists, skipping %s" % filename
        return True
      pkg_url = "{0}/packages/{1}".format(PYPI_MIRROR, pkg['path'])
      print "Downloading %s from %s" % (filename, pkg_url)
      downloader.retrieve(pkg_url, filename)
      actual_md5 = md5(open(filename).read()).hexdigest()
      if check_md5sum(filename, expected_md5):
        return True
      else:
        print "MD5 mismatch in file %s." % filename
        return False
  print "Could not find archive to download for %s %s %s" % (
      pkg_name, pkg_version, pkg_type)
  sys.exit(1)

这段代码会打开url链接下载第三方软件，然后检查md5值，非常慢，所以注释掉整个代码，返回True：

def download_package(pkg_name, pkg_version):
  return True
  '''Download the required package. Sometimes the download can be flaky, so we use the
  retry decorator.'''
  """
  pkg_type = 'sdist' # Don't download wheel archives for now
  # This JSON endpoint is not provided by PyPI mirrors so we always need to get this
  # from pypi.python.org.
  pkg_info = json.loads(urlopen('https://pypi.python.org/pypi/%s/json' % pkg_name).read())

  downloader = URLopener()
  for pkg in pkg_info['releases'][pkg_version]:
    if pkg['packagetype'] == pkg_type:
      filename = pkg['filename']
      expected_md5 = pkg['md5_digest']
      if os.path.isfile(filename) and check_md5sum(filename, expected_md5):
        print "File with matching md5sum already exists, skipping %s" % filename
        return True
      pkg_url = "{0}/packages/{1}".format(PYPI_MIRROR, pkg['path'])
      print "Downloading %s from %s" % (filename, pkg_url)
      downloader.retrieve(pkg_url, filename)
      actual_md5 = md5(open(filename).read()).hexdigest()
      if check_md5sum(filename, expected_md5):
        return True
      else:
        print "MD5 mismatch in file %s." % filename
        return False
  print "Could not find archive to download for %s %s %s" % (
      pkg_name, pkg_version, pkg_type)
  sys.exit(1)
  """

修改前端词法、语法解析源码

impala使用了jflex做词法解析，java_cup去做语法解析。
java_cup有个java_cup.runtime.Symbol类用来表示解析到的每个词，其中left属性代表词的行号，right属性代表词的列号，但是用行列来代表词在sql中的位置很不方便，我想要修改成可以获取当前词的开始、末尾在字符串的下标。
因此编辑fe/src/main/jflex/sql-scanner.flex文件，增加一个%char让jflex记录字符偏移变量到yychar
然后把newToken改成这样：

private ExtendSymbol newToken(int id, Object value) {
    String text = yytext();
    return new ExtendSymbol(id, yyline+1, yycolumn+1, value,
        this.yychar, this.yychar + text.length(), text);
  }

这样就通过SqlScanner拿到当前词在Reader中的位置了,另外我们希望Symbol还能提供位置信息，所以增加一个子类：

package java_cup.runtime;

import java_cup.runtime.Symbol;

public class ExtendSymbol extends Symbol {
    public int start = -1;
    public int end = -1;
    public String text;

    public ExtendSymbol(int id, int left, int right, Object value,
                        int start, int end, String text) {
        super(id, left, right, value);
        this.start = start;
        this.end = end;
        this.text = text;
    }

    public ExtendSymbol(int id, ExtendSymbol left, ExtendSymbol right, Object value) {
        this(id, left.left, right.right, value, left.start, right.end, null);
    }

    public ExtendSymbol(int id, ExtendSymbol left, ExtendSymbol right) {
        this(id, left, right, null);
    }

    public ExtendSymbol(int id, int left, int right, Object value) {
        this(id, left, right, value, -1, -1, null);
    }

    public ExtendSymbol(int id, Object o) {
        this(id, -1, -1, o);
    }

    public ExtendSymbol(int id, int left, int right) {
        this(id, left, right, (Object)null);
    }

    public ExtendSymbol(int sym_num) {
        super(sym_num, -1);
    }

    ExtendSymbol(int sym_num, int state) {
        super(sym_num, state);
    }
}

增加一个符号工厂类：

package java_cup.runtime;

import java_cup.runtime.Symbol;
import java_cup.runtime.SymbolFactory;

public class ExtendSymbolFactory implements SymbolFactory {
    @Override
    public ExtendSymbol newSymbol(String name, int id, Symbol left, Symbol right, Object value) {
        return new ExtendSymbol(id, (ExtendSymbol) left, (ExtendSymbol) right, value);
    }

    @Override
    public ExtendSymbol newSymbol(String name, int id, Symbol left, Symbol right) {
        return new ExtendSymbol(id, (ExtendSymbol) left, (ExtendSymbol) right);
    }

    @Override
    public ExtendSymbol newSymbol(String name, int id, Object o) {
        return new ExtendSymbol(id, o);
    }

    @Override
    public ExtendSymbol newSymbol(String name, int id) {
        return new ExtendSymbol(id);
    }

    @Override
    public ExtendSymbol startSymbol(String name, int id, int state) {
        return new ExtendSymbol(id, state);
    }
}

为语法块增加位置信息就比较复杂了，主要给org/apache/impala/analysis中的类增加一个带位置信息和子语法块的类：

package org.apache.impala.analysis;

import java.util.List;

public class SyntaxBlock {
    public int startPosition = -1;
    public int endPosition = -1;
    public List<SyntaxBlock> subBlocks;

    public SyntaxBlock() {
    }

    public SyntaxBlock(int startPosition, int endPosition) {
        this.startPosition = startPosition;
        this.endPosition = endPosition;
    }

    public SyntaxBlock(int startPosition, int endPosition, List<SyntaxBlock> subBlocks) {
        this.startPosition = startPosition;
        this.endPosition = endPosition;
        this.subBlocks = subBlocks;
    }
}

然后加上一个子类ObjectSyntaxBlock，用来存放String、Object、HashMap、ArrayList、Pair这些类型的语法块

package org.apache.impala.analysis;

import java.util.List;

public class ObjectSyntaxBlock<T> extends SyntaxBlock {
    public T objectValue;

    public ObjectSyntaxBlock() {
    }

    public ObjectSyntaxBlock(T objectValue) {
        this.objectValue = objectValue;
    }

    public ObjectSyntaxBlock(int startPosition, int endPosition, T objectValue) {
        super(startPosition, endPosition);
        this.objectValue = objectValue;
    }

    public ObjectSyntaxBlock(int startPosition, int endPosition, List<SyntaxBlock> subBlocks, T objectValue) {
        super(startPosition, endPosition, subBlocks);
        this.objectValue = objectValue;
    }

    public T getObjectValue() {
        return objectValue;
    }
}

最难的一步是修改sql-parse.cup文件，需要把所有语法信息都修改了，非常耗时间
和原版的区别是非终结符的类型如果是String、Object、HashMap、ArrayList、Pair、enum和非org.apache.impala.analysis包的类的话，需要用ObjectSyntaxBlock包装一层
比如

nonterminal List<UnionOperand> values_operand_list;
nonterminal TDescribeOutputStyle describe_output_style;

需要修改成

nonterminal ObjectSyntaxBlock<List<UnionOperand>> values_operand_list;
nonterminal ObjectSyntaxBlock<TDescribeOutputStyle> describe_output_style;

语法块的定义比如

table_ref ::=
  dotted_path:path
  {: RESULT = new TableRef(path, null); :}
  | dotted_path:path alias_clause:alias
  {: RESULT = new TableRef(path, alias); :}
  | LPAREN query_stmt:query RPAREN alias_clause:alias
  {: RESULT = new InlineViewRef(alias, query); :}
  ;

需要修改为：

table_ref ::=
  dotted_path:path
  {:
    ExtendSymbol _0_symbol = (ExtendSymbol) CUP$SqlParser$stack.peek();

    RESULT = new TableRef(path.objectValue, null);
    RESULT.startPosition = _0_symbol.start;
    RESULT.endPosition = _0_symbol.end;
    RESULT.subBlocks = Lists.newArrayList(
        (SyntaxBlock) _0_symbol.value
    );
  :}
  | dotted_path:path alias_clause:alias
  {:
    ExtendSymbol _1_symbol = (ExtendSymbol) CUP$SqlParser$stack.elementAt(CUP$SqlParser$top - 1);
    ExtendSymbol _0_symbol = (ExtendSymbol) CUP$SqlParser$stack.peek();

    RESULT = new TableRef(path.objectValue, alias.objectValue);
    RESULT.startPosition = _1_symbol.start;
    RESULT.endPosition = _0_symbol.end;
    RESULT.subBlocks = Lists.newArrayList(
        (SyntaxBlock) _1_symbol.value,
        (SyntaxBlock) _0_symbol.value
    );
  :}
  | LPAREN query_stmt:query RPAREN alias_clause:alias
  {:
    ExtendSymbol _3_symbol = (ExtendSymbol) CUP$SqlParser$stack.elementAt(CUP$SqlParser$top - 3);
    ExtendSymbol _2_symbol = (ExtendSymbol) CUP$SqlParser$stack.elementAt(CUP$SqlParser$top - 2);
    ExtendSymbol _1_symbol = (ExtendSymbol) CUP$SqlParser$stack.elementAt(CUP$SqlParser$top - 1);
    ExtendSymbol _0_symbol = (ExtendSymbol) CUP$SqlParser$stack.peek();

    RESULT = new InlineViewRef(alias.objectValue, query);
    RESULT.startPosition = _3_symbol.start;
    RESULT.endPosition = _0_symbol.end;
    RESULT.subBlocks = Lists.newArrayList(
        (SyntaxBlock) _3_symbol.value,
        (SyntaxBlock) _2_symbol.value,
        (SyntaxBlock) _1_symbol.value,
        (SyntaxBlock) _0_symbol.value
    );
  :}
  ;

像SelectStmt这种类，我们让他直接或间接继承SyntaxBlock，所以可以直接设置位置信息,不用包装

select_stmt ::=
    select_clause:selectList
  {:
    RESULT = new SelectStmt(selectList, null, null, null, null, null, null);
  :}
  |
    select_clause:selectList
    from_clause:fromClause
    where_clause:wherePredicate
    group_by_clause:groupingExprs
    having_clause:havingPredicate
    opt_order_by_clause:orderByClause
    opt_limit_offset_clause:limitOffsetClause
  {:
    RESULT = new SelectStmt(selectList, fromClause, wherePredicate, groupingExprs,
                            havingPredicate, orderByClause, limitOffsetClause);
  :}
  ;

改成：

select_stmt ::=
    select_clause:selectList
  {:
    ExtendSymbol _0_symbol = (ExtendSymbol) CUP$SqlParser$stack.peek();

    RESULT = new SelectStmt(selectList, null, null, null, null, null, null);
    RESULT.startPosition = _0_symbol.start;
    RESULT.endPosition = _0_symbol.end;
    RESULT.subBlocks = Lists.newArrayList(
        (SyntaxBlock) _0_symbol.value
    );
  :}
  |
    select_clause:selectList
    from_clause:fromClause
    where_clause:wherePredicate
    group_by_clause:groupingExprs
    having_clause:havingPredicate
    opt_order_by_clause:orderByClause
    opt_limit_offset_clause:limitOffsetClause
  {:
    ExtendSymbol _6_symbol = (ExtendSymbol) CUP$SqlParser$stack.elementAt(CUP$SqlParser$top - 6);
    ExtendSymbol _5_symbol = (ExtendSymbol) CUP$SqlParser$stack.elementAt(CUP$SqlParser$top - 5);
    ExtendSymbol _4_symbol = (ExtendSymbol) CUP$SqlParser$stack.elementAt(CUP$SqlParser$top - 4);
    ExtendSymbol _3_symbol = (ExtendSymbol) CUP$SqlParser$stack.elementAt(CUP$SqlParser$top - 3);
    ExtendSymbol _2_symbol = (ExtendSymbol) CUP$SqlParser$stack.elementAt(CUP$SqlParser$top - 2);
    ExtendSymbol _1_symbol = (ExtendSymbol) CUP$SqlParser$stack.elementAt(CUP$SqlParser$top - 1);
    ExtendSymbol _0_symbol = (ExtendSymbol) CUP$SqlParser$stack.peek();

    RESULT = new SelectStmt(selectList, fromClause, wherePredicate, groupingExprs.objectValue,
                            havingPredicate, orderByClause.objectValue, limitOffsetClause);
    RESULT.startPosition = _6_symbol.start;
    RESULT.endPosition = _0_symbol.end;
    RESULT.subBlocks = Lists.newArrayList(
        (SyntaxBlock) _6_symbol.value,
        (SyntaxBlock) _5_symbol.value,
        (SyntaxBlock) _4_symbol.value,
        (SyntaxBlock) _3_symbol.value,
        (SyntaxBlock) _2_symbol.value,
        (SyntaxBlock) _1_symbol.value,
        (SyntaxBlock) _0_symbol.value
    );
  :}
  ;

其中groupingExprs.objectValue是因为groupingExprs是opt_order_by_clause语法块的对象，opt_order_by_clause是ObjectSyntaxBlock类型的，所以本来可以直接引用，现在需要加上objectValue才能访问到包装里面的对象

github地址：impala增加语法块位置库