This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [PATCH] Put cleanups of cleanups after cleanups (PR gcov-profile/64634)


On 02/18/15 11:40, Jakub Jelinek wrote:
Hi!

Richard's GIMPLE EH rewrite in r151696 regressed following testcase.
The problem is that when lowering:
   [gcov-15.C:14:5] try
     {
       [gcov-15.C:18:12] D.2335 = __cxa_allocate_exception (4);
       [gcov-15.C:18:12] try
         {
           [gcov-15.C:18:12] [gcov-15.C:18:12] MEM[(int *)D.2335] = 5;
         }
       catch
         {
           [gcov-15.C:18:11] __cxa_free_exception (D.2335);
         }
       [gcov-15.C:18:11] __cxa_throw (D.2335, &_ZTIi, 0B);
     }
   catch
     {
       [gcov-15.C:20:3] catch (NULL)
         {
           [gcov-15.C:20:3] try
             {
               [gcov-15.C:20:10] D.2340 = __builtin_eh_pointer (0);
               [gcov-15.C:20:10] __cxa_begin_catch (D.2340);
               [gcov-15.C:22:15] catchEx ();
             }
           finally
             {
               [gcov-15.C:20:10] __cxa_end_catch ();
             }
         }
     }
we put the cleanup of the catch cleanup in front of the catch cleanup
in the EH sequence:
   [gcov-15.C:18:12] D.2335 = __cxa_allocate_exception (4);
   [gcov-15.C:18:12] [gcov-15.C:18:12] MEM[(int *)D.2335] = 5;
   [gcov-15.C:18:11] __cxa_throw (D.2335, &_ZTIi, 0B);
   <D.2345>:
   [gcov-15.C:24:1] D.2341 = 0;
   [gcov-15.C:24:1] goto <D.2342>;
   <D.2342>:
   [gcov-15.C:24:1] return D.2341;
   <D.2343>:
   [gcov-15.C:20:10] __cxa_end_catch ();
   resx 3
   <D.2346>:
   eh_dispatch 1
   resx 1
   <D.2344>:
   [gcov-15.C:20:10] D.2340 = __builtin_eh_pointer (1);
   [gcov-15.C:20:10] __cxa_begin_catch (D.2340);
   [gcov-15.C:22:15] catchEx ();
   [gcov-15.C:20:10] __cxa_end_catch ();
   goto <D.2345>;
and as the __cxa_end_catch () is the first bb for line 20,
gcov without -a considers that bb count as the one to be shown.
Before the gimple EH rewrite and also with this patch we instead
order the cleanup (__cxa_end_catch ()) after the __cxa_begin_catch ():
   [gcov-15.C:18:12] D.2335 = __cxa_allocate_exception (4);
   [gcov-15.C:18:12] [gcov-15.C:18:12] MEM[(int *)D.2335] = 5;
   [gcov-15.C:18:11] __cxa_throw (D.2335, &_ZTIi, 0B);
   <D.2345>:
   [gcov-15.C:24:1] D.2341 = 0;
   [gcov-15.C:24:1] goto <D.2342>;
   <D.2342>:
   [gcov-15.C:24:1] return D.2341;
   <D.2346>:
   eh_dispatch 1
   resx 1
   <D.2344>:
   [gcov-15.C:20:10] D.2340 = __builtin_eh_pointer (1);
   [gcov-15.C:20:10] __cxa_begin_catch (D.2340);
   [gcov-15.C:22:15] catchEx ();
   [gcov-15.C:20:10] __cxa_end_catch ();
   goto <D.2345>;
   <D.2343>:
   [gcov-15.C:20:10] __cxa_end_catch ();
   resx 3

Bootstrapped/regtested on x86_64-linux and i686-linux,
libstdc++.so.6 compiled without and with the patch is identical on
x86_64-linux.  The testcase also has identical generated code both at -O0
and -O2 when compiled without coverage.  Ok for trunk?

2015-02-18  Jakub Jelinek  <jakub@redhat.com>

	PR gcov-profile/64634
	* tree-eh.c (frob_into_branch_around): Fix up typos
	in function comment.
	(lower_catch): Put eh_seq resulting from EH lowering of
	the cleanup sequence after the cleanup rather than before
	it.

	* g++.dg/gcov/gcov-15.C: New test.
OK.
jeff


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]