[PATCH] lra: set insn_code_data to NULL when freeing

[David Malcolm <dmalcolm@redhat.com>]

On Sun, 2020-03-29 at 21:31 +0100, Andrea Corallo wrote:
> David Malcolm <dmalcolm@redhat.com> writes:
> 
> > On Wed, 2020-03-18 at 23:51 +0100, Andrea Corallo wrote:
> > Please add the new test to the header in its alphabetical location,
> > i.e. between:
> > 
> >   /* test-vector-types.cc: We don't use this, since it's C++.  */
> > 
> > and
> > 
> >   /* test-volatile.c */
> > 
> > An entry also needs to be added to the "testcases" array at the end
> > of
> > the header (again, in the alphabetical-sorted location).
> 
> I tried adding the test into the "testcases" array but this makes the
> threads test failing.
> 
> I think this has nothing to do with this patch and happen just
> because
> this test does not define any code.  Infact I see the same happening
> just adding "test-empty.c" to the "testcases" array on the current
> master.
> 
> The error is not very reproducible, I tried a run under valgrind but
> have found nothing so far :/
> 
> Dave do you recall if there was a specific reason not to have
> "test-empty.c" into the "testcases" array?
> 
>   Andrea

It's a double-free bug in lra.c, albeit one that requires being used
in a multithreaded way from libgccjit to be triggered.

libgccjit's test-threads.c repeatedly compiles and runs numerous tests,
each in a separate thread.

Attempting to add an empty test that generates no code leads to a
double-free ICE within that thread, within lra.c's
finish_insn_code_data_once.

The root cause is that the insn_code_data array is cleared in
init_insn_code_data_once, but this is only called the first time
a cgraph_node is expanded [1], whereas the "loop-over-all-elements
and free them" is unconditionally called in finalize [2].  Hence
if there are no functions:
* the array is not re-initialized for the empty context
* when finish_insn_code_data_once is called for the empty context
it still contains the freed pointers from the previous context
that held the jit mutex, and hence the free is a double-free.

This patch sets the pointers to NULL after freeing them, fixing
the ICE.  The calls to free are still guarded by a check for NULL,
which is redundant, but maybe there's a reason for not wanting to
call "free" on a possibly-NULL value many times on process exit?
(it makes the diff cleaner, at least)

Fixes the issue in jit.dg.

Full bootstrap & regression test in progress.

Is it OK for master if it passes?

Thanks
Dave


[1]
init_insn_code_data_once is called via
  lra_init_once called by
    ira_init_once called by
      initialize_rtl, via:
         if (!rtl_initialized)
           ira_init_once ();
        called by init_function_start
           called by cgraph_node::expand

[2]:
finish_insn_code_data_once is called by:
  lra_finish_once called by
    finalize

gcc/ChangeLog:
	* lra.c (finish_insn_code_data_once): Set the array elements
	to NULL after freeing them.

gcc/testsuite/ChangeLog:
	* jit.dg/all-non-failing-tests.h: Add test-empty.c
---
 gcc/lra.c                                    |  5 ++++-
 gcc/testsuite/jit.dg/all-non-failing-tests.h | 10 ++++++++++
 2 files changed, 14 insertions(+), 1 deletion(-)

diff --git a/gcc/lra.c b/gcc/lra.c
index d5ea3622686..5e8b75b1fda 100644
--- a/gcc/lra.c
+++ b/gcc/lra.c
@@ -653,7 +653,10 @@ finish_insn_code_data_once (void)
   for (unsigned int i = 0; i < NUM_INSN_CODES; i++)
     {
       if (insn_code_data[i] != NULL)
-	free (insn_code_data[i]);
+	{
+	  free (insn_code_data[i]);
+	  insn_code_data[i] = NULL;
+	}
     }
 }
 
diff --git a/gcc/testsuite/jit.dg/all-non-failing-tests.h b/gcc/testsuite/jit.dg/all-non-failing-tests.h
index b2acc74ae95..af744192a73 100644
--- a/gcc/testsuite/jit.dg/all-non-failing-tests.h
+++ b/gcc/testsuite/jit.dg/all-non-failing-tests.h
@@ -116,6 +116,13 @@
 #undef create_code
 #undef verify_code
 
+/* test-empty.c */
+#define create_code create_code_empty
+#define verify_code verify_code_empty
+#include "test-empty.c"
+#undef create_code
+#undef verify_code
+
 /* test-error-*.c: We don't use these test cases, since they deliberately
    introduce errors, which we don't want here.  */
 
@@ -328,6 +335,9 @@ const struct testcase testcases[] = {
   {"expressions",
    create_code_expressions,
    verify_code_expressions},
+  {"empty",
+   create_code_empty,
+   verify_code_empty},
   {"factorial",
    create_code_factorial,
    verify_code_factorial},
-- 
2.21.0