IMPALA-9838: Switch to GCC 7.5.0

This upgrades GCC and libstdc++ to version 7.5.0. There
have been ABI changes since 4.9.2, so this means that
the native-toolchain produced with the new compiler is
not interoperable with one produced by the old compiler.
To allow that transition, IMPALA_TOOLCHAIN_PACKAGES_HOME
is now a subdirectory of IMPALA_TOOLCHAIN
(toolchain-packages-gcc${IMPALA_GCC_VERSION}) to distinguish
it from the old packages.

Some Python packages in the impala-python virtualenv are
compiled using the toolchain GCC and now use the new ABI.
This leads to two changes:
1. When constructing the LD_LIBRARY_PATH for impala-python,
we include the GCC libstdc++ libraries. Otherwise, certain
Python packages that use C++ fail on older OSes like Centos 7.
This fixes IMPALA-9804.
2. Since developers work on various branches, this changes
the virtualenv's directory location to a directory with
the GCC version in the name. This allows the virtualenv
built with GCC 7 to coexist with the current virtualenv
built with GCC 4.9.2. The location for the old virtualenv is
${IMPALA_HOME}/infra/python/env. The new location is
${IMPALA_HOME}/infra/python/env-gcc${IMPALA_GCC_VERSION}. This
required updating several impala-python scripts.

There are various odds-and-ends related to the transition:
1. Due to the small string optimization, the size of std::string
changed, which means that various data structures also changed
in size. This required updating some static asserts.
2. There is a bug in clang-tidy that reports a use-after-free
for some code using std::shared_ptr. Clang is not modeling
the shared_ptr correctly, so it is a false-positive. As a workaround,
this disables the clang-analyzer-cplusplus.NewDelete diagnostic.
3. Various small compilation fixes (includes, etc).

Performance testing:
 - Ran single-node performance tests on TPC-H for the following
   configurations:
    - TPC-H Parquet scale 30 with normal configurations
    - TPC-H Parquet scale 30 with codegen disabled
    - TPC-H Kudu scale 10
   None found any significant regressions. Full results are
   posted on the JIRA.
 - Ran single-node performance tests on targeted-perf scale 10.
   No significant regressions.
 - The size of binaries (impalad, etc) is slightly smaller with the new GCC:
   GCC 4.9.2 release impalad binary: 545664
   GCC 7.5.0 release impalad binary: 539900
 - Compilation in DEBUG mode is roughly 15-25% faster

Functional testing:
 - Ran core jobs, exhaustive release jobs, UBSAN

Change-Id: Ia0beb2b618ba669c9699f8dbc0c52d1203d004e4
Reviewed-on: http://gerrit.cloudera.org:8080/16045
Reviewed-by: Joe McDonnell <joemcdonnell@cloudera.com>
Tested-by: Impala Public Jenkins <impala-public-jenkins@cloudera.com>

infra/python/bootstrap_virtualenv.py

@@ -52,8 +52,10 @@ LOG = logging.getLogger(os.path.splitext(os.path.basename(__file__))[0])
 
 SKIP_TOOLCHAIN_BOOTSTRAP = "SKIP_TOOLCHAIN_BOOTSTRAP"
 
+GCC_VERSION = os.environ["IMPALA_GCC_VERSION"]
+
 DEPS_DIR = os.path.join(os.path.dirname(__file__), "deps")
-ENV_DIR = os.path.join(os.path.dirname(__file__), "env")
+ENV_DIR = os.path.join(os.path.dirname(__file__), "env-gcc{0}".format(GCC_VERSION))
 
 # Requirements file with packages we need for our build and tests.
 REQS_PATH = os.path.join(DEPS_DIR, "requirements.txt")
@@ -390,9 +392,14 @@ if __name__ == "__main__":
   options, args = parser.parse_args()
 
   if options.print_ld_library_path:
+    # Some python packages have native code that is compiled with the toolchain
+    # compiler, so that code needs to dynamically link against matching library
+    # versions.
+    ld_library_dirs = [os.path.join(toolchain_pkg_dir("gcc"), 'lib64')]
     kudu_client_dir = find_kudu_client_install_dir()
-    print(os.path.pathsep.join([os.path.join(kudu_client_dir, 'lib'),
-                                    os.path.join(kudu_client_dir, 'lib64')]))
+    ld_library_dirs.append(os.path.join(kudu_client_dir, 'lib'))
+    ld_library_dirs.append(os.path.join(kudu_client_dir, 'lib64'))
+    print(os.path.pathsep.join(ld_library_dirs))
     sys.exit()
 
   logging.basicConfig(level=getattr(logging, options.log_level))