Fix more ICEs in -fsave-optimization-record (PR tree-optimization/89235)
authorDavid Malcolm <dmalcolm@redhat.com>
Thu, 7 Feb 2019 23:00:18 +0000 (23:00 +0000)
committerDavid Malcolm <dmalcolm@gcc.gnu.org>
Thu, 7 Feb 2019 23:00:18 +0000 (23:00 +0000)
PR tree-optimization/89235 reports an ICE inside -fsave-optimization-record
whilst reporting the inlining chain of of the location_t in the
vect_location global.

This is very similar to PR tree-optimization/86637, fixed in r266821.

The issue is that the inlining chains are read from the location_t's
ad-hoc data, referencing GC-managed tree blocks, but the former are
not GC roots; it's simply assumed that old locations referencing dead
blocks never get used again.

The fix is to reset the "vect_location" global in more places.  Given
that is a somewhat subtle detail, the patch adds a sentinel class to
reset vect_location at the end of a scope.  Doing it as a class
simplifies the task of ensuring that the global is reset on every
exit path from a function, and also gives a good place to signpost
the above subtlety (in the documentation for the class).

The patch also adds test cases for both of the PRs mentioned above.

gcc/testsuite/ChangeLog:
PR tree-optimization/86637
PR tree-optimization/89235
* gcc.c-torture/compile/pr86637-1.c: New test.
* gcc.c-torture/compile/pr86637-2.c: New test.
* gcc.c-torture/compile/pr86637-3.c: New test.
* gcc.c-torture/compile/pr89235.c: New test.

gcc/ChangeLog:
PR tree-optimization/86637
PR tree-optimization/89235
* tree-vect-loop.c (optimize_mask_stores): Add an
auto_purge_vect_location sentinel to ensure that vect_location is
purged on exit.
* tree-vectorizer.c
(auto_purge_vect_location::~auto_purge_vect_location): New dtor.
(try_vectorize_loop_1): Add an auto_purge_vect_location sentinel
to ensure that vect_location is purged on exit.
(pass_slp_vectorize::execute): Likewise, replacing the manual
reset.
* tree-vectorizer.h (class auto_purge_vect_location): New class.

From-SVN: r268659

gcc/ChangeLog
gcc/testsuite/ChangeLog
gcc/testsuite/gcc.c-torture/compile/pr86637-1.c [new file with mode: 0644]
gcc/testsuite/gcc.c-torture/compile/pr86637-2.c [new file with mode: 0644]
gcc/testsuite/gcc.c-torture/compile/pr86637-3.c [new file with mode: 0644]
gcc/testsuite/gcc.c-torture/compile/pr89235.c [new file with mode: 0644]
gcc/tree-vect-loop.c
gcc/tree-vectorizer.c
gcc/tree-vectorizer.h

index 098479015d4cd9ab818d19b6a3fe9307070811e7..cbb1ebe4a3ebed76c13f049b6a0d2299a1c5f6fd 100644 (file)
@@ -1,3 +1,18 @@
+2019-02-07  David Malcolm  <dmalcolm@redhat.com>
+
+       PR tree-optimization/86637
+       PR tree-optimization/89235
+       * tree-vect-loop.c (optimize_mask_stores): Add an
+       auto_purge_vect_location sentinel to ensure that vect_location is
+       purged on exit.
+       * tree-vectorizer.c
+       (auto_purge_vect_location::~auto_purge_vect_location): New dtor.
+       (try_vectorize_loop_1): Add an auto_purge_vect_location sentinel
+       to ensure that vect_location is purged on exit.
+       (pass_slp_vectorize::execute): Likewise, replacing the manual
+       reset.
+       * tree-vectorizer.h (class auto_purge_vect_location): New class.
+
 2019-02-07  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>
 
        * config/aarch64/iterators.md (max_opp): New code_attr.
index 9951b17cd1829599ab9dc66ee777605d5da8670a..cb500f534890cfb8d4370d52e4fa698d854ddec4 100644 (file)
@@ -1,3 +1,12 @@
+2019-02-07  David Malcolm  <dmalcolm@redhat.com>
+
+       PR tree-optimization/86637
+       PR tree-optimization/89235
+       * gcc.c-torture/compile/pr86637-1.c: New test.
+       * gcc.c-torture/compile/pr86637-2.c: New test.
+       * gcc.c-torture/compile/pr86637-3.c: New test.
+       * gcc.c-torture/compile/pr89235.c: New test.
+
 2019-02-07  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>
 
        * gcc.target/aarch64/abd_1.c: New test.
diff --git a/gcc/testsuite/gcc.c-torture/compile/pr86637-1.c b/gcc/testsuite/gcc.c-torture/compile/pr86637-1.c
new file mode 100644 (file)
index 0000000..61b6381
--- /dev/null
@@ -0,0 +1,13 @@
+/* { dg-options "-fsave-optimization-record -ftree-slp-vectorize --param ggc-min-expand=1 --param ggc-min-heapsize=1024" } */
+
+void
+en (void)
+{
+}
+
+void
+n4 (int zb)
+{
+  while (zb < 1)
+    ++zb;
+}
diff --git a/gcc/testsuite/gcc.c-torture/compile/pr86637-2.c b/gcc/testsuite/gcc.c-torture/compile/pr86637-2.c
new file mode 100644 (file)
index 0000000..3b675ea
--- /dev/null
@@ -0,0 +1,128 @@
+/* { dg-do compile { target fgraphite } } */
+/* { dg-options "-floop-parallelize-all -fsave-optimization-record -ftree-parallelize-loops=2 -ftree-slp-vectorize" } */
+
+#include <stdint.h>
+#include <stdlib.h>
+
+signed char qq;
+int rm, mu, l9;
+long long unsigned int ip;
+
+void
+fi (void)
+{
+}
+
+void
+il (long long unsigned int c5)
+{
+  int *na = &mu;
+
+  while (l9 < 1)
+    {
+      if (qq < 1)
+        {
+          short int ds = 0;
+
+          if ((ip - *na - ip / c5 != 0) && (*na / ((c5 + c5) && !!ip) != 0))
+            __builtin_trap ();
+
+          rm = -1 / (!!rm && !!c5);
+
+          while (qq < 1)
+            {
+              ++*na;
+              ++ip;
+              if (*na < (int) ip)
+                while (ds < 2)
+                  {
+                    ++qq;
+                    ++ds;
+                  }
+            }
+        }
+
+      ++l9;
+    }
+
+  for (;;)
+    {
+    }
+}
+
+void
+uu (short int wk)
+{
+  int64_t tl = ip;
+
+  while (ip < 2)
+    {
+      signed char vn;
+
+      for (vn = 1; vn != 0; ++vn)
+        {
+          int rh;
+
+          if (qq < 1)
+            {
+              while (mu < 1)
+                ip = 0;
+
+              while (rm != 0)
+                ++rm;
+            }
+
+          for (rh = 0; rh < 3; ++rh)
+            {
+              int ab;
+
+              for (ab = 0; ab < 5; ++ab)
+                {
+                  tl -= qq;
+                  wk += rh - tl;
+                  ip += (ab < wk) + wk;
+                }
+            }
+        }
+    }
+}
+
+void
+im (uint8_t kt)
+{
+  int wu = 0;
+
+  for (;;)
+    {
+      ++rm;
+      if (rm < 1)
+        {
+          short int ms = 0;
+
+          while (kt < 1)
+            {
+              ms += rm < 0;
+
+              if (wu != 0)
+                for (;;)
+                  {
+                  }
+
+              while (ms < 4)
+                {
+                  while (wu < 1)
+                    wu += 2;
+
+                  ++ms;
+                }
+            }
+
+          if (ms == 0 || wu == 0)
+            break;
+        }
+    }
+
+  while (wu < 1)
+    while (qq < 1)
+      ++qq;
+}
diff --git a/gcc/testsuite/gcc.c-torture/compile/pr86637-3.c b/gcc/testsuite/gcc.c-torture/compile/pr86637-3.c
new file mode 100644 (file)
index 0000000..6cb0fd7
--- /dev/null
@@ -0,0 +1,11 @@
+/* { dg-options "-fsave-optimization-record -ftree-slp-vectorize --param ggc-min-expand=0 --param ggc-min-heapsize=1024" } */
+void
+te (void)
+{
+}
+
+int
+main (void)
+{
+  return 0;
+}
diff --git a/gcc/testsuite/gcc.c-torture/compile/pr89235.c b/gcc/testsuite/gcc.c-torture/compile/pr89235.c
new file mode 100644 (file)
index 0000000..86be27f
--- /dev/null
@@ -0,0 +1,57 @@
+/* { dg-require-effective-target fopenmp } */
+/* { dg-options "-S -fopenmp -fsave-optimization-record -ftree-parallelize-loops=2 -fno-tree-vectorize --param ggc-min-expand=0" } */
+
+int a1, dr, xm, ly, zb, g9, il;
+
+long int wt;
+unsigned int mq;
+int br, e7, rm, t4, jb, ry;
+
+int
+fi (void);
+
+int
+z5 (int fl)
+{
+  while (br < 1)
+    while (e7 != 0)
+      while (mq != 1)
+        {
+          if (!!fl)
+            {
+              wt = rm;
+              fi ();
+            }
+
+          ++mq;
+        }
+
+  return 0;
+}
+
+void
+gg (void)
+{
+  t4 = rm = z5 (rm);
+  jb = z5 (rm);
+  ry = fi ();
+}
+
+#pragma omp declare simd
+void
+hl (void)
+{
+  for (;;)
+    {
+      gg ();
+      gg ();
+      gg ();
+    }
+}
+
+#pragma omp declare simd
+int
+cw (void)
+{
+  return 0;
+}
index eda4c24846ad05b412f805230d598799ab19184c..d63d3825f2018c885405440e07c4e9fcd9daea82 100644 (file)
@@ -8578,6 +8578,7 @@ optimize_mask_stores (struct loop *loop)
   gimple_stmt_iterator gsi;
   gimple *stmt;
   auto_vec<gimple *> worklist;
+  auto_purge_vect_location sentinel;
 
   vect_location = find_loop_location (loop);
   /* Pick up all masked stores in loop if any.  */
index fa5b22efebad157753c4122d118d4d27cd6bf2a5..d27104933a95a330e560505ec519ec820b921572 100644 (file)
@@ -86,6 +86,15 @@ along with GCC; see the file COPYING3.  If not see
 /* Loop or bb location, with hotness information.  */
 dump_user_location_t vect_location;
 
+/* auto_purge_vect_location's dtor: reset the vect_location
+   global, to avoid stale location_t values that could reference
+   GC-ed blocks.  */
+
+auto_purge_vect_location::~auto_purge_vect_location ()
+{
+  vect_location = dump_user_location_t ();
+}
+
 /* Dump a cost entry according to args to F.  */
 
 void
@@ -860,6 +869,7 @@ try_vectorize_loop_1 (hash_table<simduid_to_vf> *&simduid_to_vf_htab,
 {
   unsigned ret = 0;
   vec_info_shared shared;
+  auto_purge_vect_location sentinel;
   vect_location = find_loop_location (loop);
   if (LOCATION_LOCUS (vect_location.get_location_t ()) != UNKNOWN_LOCATION
       && dump_enabled_p ())
@@ -1269,6 +1279,7 @@ public:
 unsigned int
 pass_slp_vectorize::execute (function *fun)
 {
+  auto_purge_vect_location sentinel;
   basic_block bb;
 
   bool in_loop_pipeline = scev_initialized_p ();
@@ -1303,8 +1314,6 @@ pass_slp_vectorize::execute (function *fun)
       loop_optimizer_finalize ();
     }
 
-  vect_location = dump_user_location_t ();
-
   return 0;
 }
 
index d26b0f83bd1ccad7a81e43b008c88a807ba30884..0e056f34b53cf80028cdc43d33c8d90f246fb880 100644 (file)
@@ -1420,6 +1420,24 @@ extern dump_user_location_t vect_location;
 #define DUMP_VECT_SCOPE(MSG) \
   AUTO_DUMP_SCOPE (MSG, vect_location)
 
+/* A sentinel class for ensuring that the "vect_location" global gets
+   reset at the end of a scope.
+
+   The "vect_location" global is used during dumping and contains a
+   location_t, which could contain references to a tree block via the
+   ad-hoc data.  This data is used for tracking inlining information,
+   but it's not a GC root; it's simply assumed that such locations never
+   get accessed if the blocks are optimized away.
+
+   Hence we need to ensure that such locations are purged at the end
+   of any operations using them (e.g. via this class).  */
+
+class auto_purge_vect_location
+{
+ public:
+  ~auto_purge_vect_location ();
+};
+
 /*-----------------------------------------------------------------*/
 /* Function prototypes.                                            */
 /*-----------------------------------------------------------------*/