diff mbox

RFC/A: Early predictive commoning pass

Message ID 8760f9yklt.fsf@linaro.org
State New
Headers show

Commit Message

Richard Sandiford July 3, 2017, 8:45 a.m. UTC
General predictive commoning would play havoc with loop vectorisation,
so the current pass order is clearly the right one.  But running a very
limited form of predictive commoning before vectorisation would allow us
to vectorise things like:

     for (int i = 1; i < n; ++i)
       x[i] = x[i - 1] + 1;

This patch adds an extra pass that is restricted to cases that should
help (or at least not hinder) vectorisation.  It gives some nice
improvements on some internal benchmarks.

I compared the output for SPEC 2k6 before and after the patch.  For some
benchmarks it led to a trivial register renaming, but had no effect on
those benchmarks beyond that.  The only benchmark that changed in a
significant way was 416.gamess, where we were able to vectorise some
simple loops that we weren't previously.  None of those loops seem to
be hot though, so there was no measurable difference in the score.

Tested on aarch64-linux-gnu and x86_64-linux-gnu.  Thoughts?  Is this
too much of a special case to support a new pass?  OTOH, other compilers
do vectorise the loop above, so it would be nice if we could too...

Richard


2017-07-03  Richard Sandiford  <richard.sandiford@linaro.org>

gcc/
	* passes.def (pass_early_predcom): New.
	* tree-pass.h (make_pass_early_predcom): Declare.
	* tree-predcom.c (MAX_DISTANCE): Turn into an inclusive rather than
	exclusive upper bound.
	(only_simple_p): New variable.
	(max_distance): Likewise.
	(add_ref_to_chain): Use MAX_DISTANCE rather than max_distance
	and treat it as an inclusive upper bound.  Require the store to
	come after the load at the maximum distance if only_simple_p.
	(add_looparound_copies): Do nothing if only_simple_p.
	(determine_roots_comp): Use MAX_DISTANCE rather than max_distance
	and treat it as an inclusive upper bound.  Require the start of
	a chain to be a store if only_simple_p.
	(determine_unroll_factor): Return 1 if only_simple_p.
	(tree_predictive_commoning): Add an early_p parameter.  Set up
	only_simple_p and max_distance.
	(run_tree_predictive_commoning): Add an early_p parameter.
	Update call to tree_predictive_commoning.
	(pass_data_early_predcom): New descriptor.
	(pass_early_predcom): New class.
	(pass_data_predcom::execute): Update call to
	run_tree_predictive_commoning.
	(make_pass_early_predcom): New function.

gcc/testsuite/
	* gnat.dg/vect18.adb: Turn off predictive commoning.

Comments

Richard Biener July 18, 2017, 11:55 a.m. UTC | #1
On Mon, Jul 3, 2017 at 10:45 AM, Richard Sandiford
<richard.sandiford@linaro.org> wrote:
> General predictive commoning would play havoc with loop vectorisation,

> so the current pass order is clearly the right one.  But running a very

> limited form of predictive commoning before vectorisation would allow us

> to vectorise things like:

>

>      for (int i = 1; i < n; ++i)

>        x[i] = x[i - 1] + 1;


In principle PRE can handle this case (if we weren't taming it down so much).

I'm not too sympathetic of adding yet another predcom pass given it is
expensive to do dependence analysis.  To make PRE do the transform you
need

              && bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (sprime))

that is, the challenge is to identify the cases that help
vectorization as opposed
to blocking it ... (here you create a vectorizable induction which is fine).

> This patch adds an extra pass that is restricted to cases that should

> help (or at least not hinder) vectorisation.  It gives some nice

> improvements on some internal benchmarks.


Are you sure you gated off things properly?  It sounds like if we didn't tame
down PRE by use of the PHIs but by use of stmts inserted in the latch we'd
arrive at similar heuristics?

Richard.

> I compared the output for SPEC 2k6 before and after the patch.  For some

> benchmarks it led to a trivial register renaming, but had no effect on

> those benchmarks beyond that.  The only benchmark that changed in a

> significant way was 416.gamess, where we were able to vectorise some

> simple loops that we weren't previously.  None of those loops seem to

> be hot though, so there was no measurable difference in the score.

>

> Tested on aarch64-linux-gnu and x86_64-linux-gnu.  Thoughts?  Is this

> too much of a special case to support a new pass?  OTOH, other compilers

> do vectorise the loop above, so it would be nice if we could too...

>

> Richard

>

>

> 2017-07-03  Richard Sandiford  <richard.sandiford@linaro.org>

>

> gcc/

>         * passes.def (pass_early_predcom): New.

>         * tree-pass.h (make_pass_early_predcom): Declare.

>         * tree-predcom.c (MAX_DISTANCE): Turn into an inclusive rather than

>         exclusive upper bound.

>         (only_simple_p): New variable.

>         (max_distance): Likewise.

>         (add_ref_to_chain): Use MAX_DISTANCE rather than max_distance

>         and treat it as an inclusive upper bound.  Require the store to

>         come after the load at the maximum distance if only_simple_p.

>         (add_looparound_copies): Do nothing if only_simple_p.

>         (determine_roots_comp): Use MAX_DISTANCE rather than max_distance

>         and treat it as an inclusive upper bound.  Require the start of

>         a chain to be a store if only_simple_p.

>         (determine_unroll_factor): Return 1 if only_simple_p.

>         (tree_predictive_commoning): Add an early_p parameter.  Set up

>         only_simple_p and max_distance.

>         (run_tree_predictive_commoning): Add an early_p parameter.

>         Update call to tree_predictive_commoning.

>         (pass_data_early_predcom): New descriptor.

>         (pass_early_predcom): New class.

>         (pass_data_predcom::execute): Update call to

>         run_tree_predictive_commoning.

>         (make_pass_early_predcom): New function.

>

> gcc/testsuite/

>         * gnat.dg/vect18.adb: Turn off predictive commoning.

>

> Index: gcc/passes.def

> ===================================================================

> --- gcc/passes.def      2017-06-22 12:22:55.989380389 +0100

> +++ gcc/passes.def      2017-07-03 09:17:28.626495661 +0100

> @@ -290,6 +290,7 @@ along with GCC; see the file COPYING3.

>           NEXT_PASS (pass_parallelize_loops, false /* oacc_kernels_p */);

>           NEXT_PASS (pass_expand_omp_ssa);

>           NEXT_PASS (pass_ch_vect);

> +         NEXT_PASS (pass_early_predcom);

>           NEXT_PASS (pass_if_conversion);

>           /* pass_vectorize must immediately follow pass_if_conversion.

>              Please do not add any other passes in between.  */

> Index: gcc/tree-pass.h

> ===================================================================

> --- gcc/tree-pass.h     2017-06-22 12:22:34.954287935 +0100

> +++ gcc/tree-pass.h     2017-07-03 09:17:28.627495621 +0100

> @@ -369,6 +369,7 @@ extern gimple_opt_pass *make_pass_tree_l

>  extern gimple_opt_pass *make_pass_lim (gcc::context *ctxt);

>  extern gimple_opt_pass *make_pass_tree_unswitch (gcc::context *ctxt);

>  extern gimple_opt_pass *make_pass_loop_split (gcc::context *ctxt);

> +extern gimple_opt_pass *make_pass_early_predcom (gcc::context *ctxt);

>  extern gimple_opt_pass *make_pass_predcom (gcc::context *ctxt);

>  extern gimple_opt_pass *make_pass_iv_canon (gcc::context *ctxt);

>  extern gimple_opt_pass *make_pass_scev_cprop (gcc::context *ctxt);

> Index: gcc/tree-predcom.c

> ===================================================================

> --- gcc/tree-predcom.c  2017-07-03 08:42:47.632532107 +0100

> +++ gcc/tree-predcom.c  2017-07-03 09:29:08.744451338 +0100

> @@ -218,7 +218,7 @@ Free Software Foundation; either version

>  /* The maximum number of iterations between the considered memory

>     references.  */

>

> -#define MAX_DISTANCE (target_avail_regs < 16 ? 4 : 8)

> +#define MAX_DISTANCE (target_avail_regs < 16 ? 3 : 7)

>

>  /* Data references (or phi nodes that carry data reference values across

>     loop iterations).  */

> @@ -343,6 +343,71 @@ struct component

>

>  static hash_map<tree, name_expansion *> *name_expansions;

>

> +/* True if we're running the early predcom pass and should only handle

> +   cases that aid vectorization.  Specifically this means that:

> +

> +   - only CT_INVARIANT and CT_STORE_LOAD chains are used

> +   - the maximum distance for a CT_STORE_LOAD chain is 1 iteration,

> +     and at that distance the store must come after the load

> +   - there's no unrolling or detection of looparound phis.

> +

> +   The idea is handle inductions that go via memory, such as:

> +

> +     for (int i = 1; i < n; ++i)

> +       x[i] = x[i - 1] + 1;

> +

> +   As it stands this loop could never be vectorized, because a loop body

> +   that contains a read of x[j] followed by a write to x[j + 1] would

> +   have its vectorization factor limited to 1.  Transforming it to:

> +

> +     int tmp = x[0];

> +     for (int i = 0; i < n; ++i)

> +       {

> +         tmp += 1;

> +         x[i] = tmp:

> +       }

> +

> +   exposes the fact that the stored value is a simple vectorizable

> +   induction with start value x[0] and step 1.

> +

> +   Commoning is not always useful even in this situation.  For example,

> +   carrying over the value of x[i] won't help us to vectorize:

> +

> +     for (int i = 1; i < n; ++i)

> +       {

> +        y[i] = x[i - 1];

> +        x[i] += i;

> +       }

> +

> +   There's no real need to restrict things further though, because we're

> +   unable to vectorize these load/store combinations in their current

> +   form whatever happens.

> +

> +   We require the store to come after the load when the distance is 1

> +   to avoid cases like:

> +

> +     for (int i = 1; i < n; ++i)

> +       {

> +        x[i] = ...;

> +        ... = x[i - 1];

> +       }

> +

> +   These accesses effectively have a distance somewhere between 1 and 2,

> +   since after commoning the value stored in the previous iteration would

> +   still be live at the next store.  This means that the combination

> +   isn't useful for exposing simple inductions.

> +

> +   Also, unlike the motivating case above, this combination does not

> +   prevent vectorization.  If a write to x[j + 1] comes before a read

> +   of x[j], the vectorized write completes for all vector elements

> +   before the read starts for any vector elements.  */

> +

> +static bool only_simple_p;

> +

> +/* The maximum loop carry distance for this execution of the pass.  */

> +

> +static int max_distance;

> +

>  /* Dumps data reference REF to FILE.  */

>

>  extern void dump_dref (FILE *, dref);

> @@ -960,7 +1025,13 @@ add_ref_to_chain (chain_p chain, dref re

>

>    gcc_assert (wi::les_p (root->offset, ref->offset));

>    widest_int dist = ref->offset - root->offset;

> -  if (wi::leu_p (MAX_DISTANCE, dist))

> +  /* When running before vectorization, only allow the maximum distance

> +     if the consumer comes before the producer.  See the comment above

> +     ONLY_SIMPLE_P for details.  */

> +  if (wi::ltu_p (max_distance, dist)

> +      || (only_simple_p

> +         && wi::eq_p (max_distance, dist)

> +         && root->pos < ref->pos))

>      {

>        free (ref);

>        return;

> @@ -1194,6 +1265,11 @@ add_looparound_copies (struct loop *loop

>    dref ref, root = get_chain_root (chain);

>    gphi *phi;

>

> +  /* There's no point doing this when running before vectorization,

> +     since we won't unroll the loop or combine chains.  */

> +  if (only_simple_p)

> +    return;

> +

>    FOR_EACH_VEC_ELT (chain->refs, i, ref)

>      {

>        phi = find_looparound_phi (loop, ref, root);

> @@ -1232,7 +1308,7 @@ determine_roots_comp (struct loop *loop,

>    FOR_EACH_VEC_ELT (comp->refs, i, a)

>      {

>        if (!chain || DR_IS_WRITE (a->ref)

> -         || wi::leu_p (MAX_DISTANCE, a->offset - last_ofs))

> +         || wi::ltu_p (max_distance, a->offset - last_ofs))

>         {

>           if (nontrivial_chain_p (chain))

>             {

> @@ -1241,6 +1317,14 @@ determine_roots_comp (struct loop *loop,

>             }

>           else

>             release_chain (chain);

> +         /* Only create CT_STORE_LOAD and CT_INVARIANT chains when

> +            running before vectorization.  */

> +         if (only_simple_p && !DR_IS_WRITE (a->ref))

> +           {

> +             free (a);

> +             chain = NULL;

> +             continue;

> +           }

>           chain = make_rooted_chain (a);

>           last_ofs = a->offset;

>           continue;

> @@ -1759,6 +1843,10 @@ determine_unroll_factor (vec<chain_p> ch

>    unsigned factor = 1, af, nfactor, i;

>    unsigned max = PARAM_VALUE (PARAM_MAX_UNROLL_TIMES);

>

> +  /* Do not unroll when running before vectorization.  */

> +  if (only_simple_p)

> +    return 1;

> +

>    FOR_EACH_VEC_ELT (chains, i, chain)

>      {

>        if (chain->type == CT_INVARIANT)

> @@ -2562,8 +2650,11 @@ tree_predictive_commoning_loop (struct l

>      }

>    prepare_initializers (loop, chains);

>

> -  /* Try to combine the chains that are always worked with together.  */

> -  try_combine_chains (&chains);

> +  /* During the main pass, try to combine the chains that are always

> +     worked with together.  For the early pass it should be better

> +     to leave this to the vectorizer.  */

> +  if (!only_simple_p)

> +    try_combine_chains (&chains);

>

>    if (dump_file && (dump_flags & TDF_DETAILS))

>      {

> @@ -2623,15 +2714,19 @@ tree_predictive_commoning_loop (struct l

>    return unroll;

>  }

>

> -/* Runs predictive commoning.  */

> +/* Runs predictive commoning.  EARLY_P is true if we are running before

> +   vectorization.  */

>

>  unsigned

> -tree_predictive_commoning (void)

> +tree_predictive_commoning (bool early_p)

>  {

>    bool unrolled = false;

>    struct loop *loop;

>    unsigned ret = 0;

>

> +  only_simple_p = early_p;

> +  max_distance = early_p ? 1 : MAX_DISTANCE;

> +

>    initialize_original_copy_tables ();

>    FOR_EACH_LOOP (loop, LI_ONLY_INNERMOST)

>      if (optimize_loop_for_speed_p (loop))

> @@ -2649,19 +2744,51 @@ tree_predictive_commoning (void)

>    return ret;

>  }

>

> -/* Predictive commoning Pass.  */

> +/* Predictive commoning pass.  EARLY_P is true if we are running before

> +   vectorization.  */

>

>  static unsigned

> -run_tree_predictive_commoning (struct function *fun)

> +run_tree_predictive_commoning (struct function *fun, bool early_p)

>  {

>    if (number_of_loops (fun) <= 1)

>      return 0;

>

> -  return tree_predictive_commoning ();

> +  return tree_predictive_commoning (early_p);

>  }

>

>  namespace {

>

> +const pass_data pass_data_early_predcom =

> +{

> +  GIMPLE_PASS, /* type */

> +  "epcom", /* name */

> +  OPTGROUP_LOOP, /* optinfo_flags */

> +  TV_PREDCOM, /* tv_id */

> +  PROP_cfg, /* properties_required */

> +  0, /* properties_provided */

> +  0, /* properties_destroyed */

> +  0, /* todo_flags_start */

> +  TODO_update_ssa_only_virtuals, /* todo_flags_finish */

> +};

> +

> +class pass_early_predcom : public gimple_opt_pass

> +{

> +public:

> +  pass_early_predcom (gcc::context *ctxt)

> +    : gimple_opt_pass (pass_data_early_predcom, ctxt)

> +  {}

> +

> +  /* opt_pass methods: */

> +  virtual bool gate (function *)

> +  {

> +    return flag_predictive_commoning && flag_tree_loop_vectorize;

> +  }

> +  virtual unsigned int execute (function *fun)

> +  {

> +    return run_tree_predictive_commoning (fun, true);

> +  }

> +}; // class pass_early_predcom

> +

>  const pass_data pass_data_predcom =

>  {

>    GIMPLE_PASS, /* type */

> @@ -2686,7 +2813,7 @@ const pass_data pass_data_predcom =

>    virtual bool gate (function *) { return flag_predictive_commoning != 0; }

>    virtual unsigned int execute (function *fun)

>      {

> -      return run_tree_predictive_commoning (fun);

> +      return run_tree_predictive_commoning (fun, false);

>      }

>

>  }; // class pass_predcom

> @@ -2694,9 +2821,13 @@ const pass_data pass_data_predcom =

>  } // anon namespace

>

>  gimple_opt_pass *

> +make_pass_early_predcom (gcc::context *ctxt)

> +{

> +  return new pass_early_predcom (ctxt);

> +}

> +

> +gimple_opt_pass *

>  make_pass_predcom (gcc::context *ctxt)

>  {

>    return new pass_predcom (ctxt);

>  }

> -

> -

> Index: gcc/testsuite/gnat.dg/vect18.adb

> ===================================================================

> --- gcc/testsuite/gnat.dg/vect18.adb    2015-10-14 14:58:56.000000000 +0100

> +++ gcc/testsuite/gnat.dg/vect18.adb    2017-07-03 09:17:28.627495621 +0100

> @@ -1,5 +1,5 @@

>  -- { dg-do compile { target i?86-*-* x86_64-*-* } }

> --- { dg-options "-O3 -msse2 -fdump-tree-vect-details" }

> +-- { dg-options "-O3 -msse2 -fdump-tree-vect-details -fno-predictive-commoning" }

>

>  package body Vect18 is

>Index: gcc/tree-ssa-pre.c

===================================================================
--- gcc/tree-ssa-pre.c  (revision 250297)
+++ gcc/tree-ssa-pre.c  (working copy)
@@ -1458,9 +1458,11 @@ phi_translate_1 (pre_expr expr, bitmap_s
                       to be inserted and increased register pressure.
                       See PR77498 - this avoids doing predcoms work in
                       a less efficient way.  */
+#if 0
                    if (find_edge (pred, phiblock)->flags & EDGE_DFS_BACK)
                      ;
                    else
+#endif
                      {
                        unsigned value_id = get_expr_value_id (constant);
                        constant = find_leader_in_sets (value_id, set1, set2,
@@ -4377,7 +4379,7 @@ eliminate_dom_walker::before_dom_childre
          if (sprime
              && TREE_CODE (sprime) == SSA_NAME
              && do_pre
-             && (flag_tree_loop_vectorize || flag_tree_parallelize_loops > 1)
+             && 0 && (flag_tree_loop_vectorize ||
flag_tree_parallelize_loops > 1)
              && loop_outer (b->loop_father)
              && has_zero_uses (sprime)

diff mbox

Patch

Index: gcc/passes.def
===================================================================
--- gcc/passes.def	2017-06-22 12:22:55.989380389 +0100
+++ gcc/passes.def	2017-07-03 09:17:28.626495661 +0100
@@ -290,6 +290,7 @@  along with GCC; see the file COPYING3.
 	  NEXT_PASS (pass_parallelize_loops, false /* oacc_kernels_p */);
 	  NEXT_PASS (pass_expand_omp_ssa);
 	  NEXT_PASS (pass_ch_vect);
+	  NEXT_PASS (pass_early_predcom);
 	  NEXT_PASS (pass_if_conversion);
 	  /* pass_vectorize must immediately follow pass_if_conversion.
 	     Please do not add any other passes in between.  */
Index: gcc/tree-pass.h
===================================================================
--- gcc/tree-pass.h	2017-06-22 12:22:34.954287935 +0100
+++ gcc/tree-pass.h	2017-07-03 09:17:28.627495621 +0100
@@ -369,6 +369,7 @@  extern gimple_opt_pass *make_pass_tree_l
 extern gimple_opt_pass *make_pass_lim (gcc::context *ctxt);
 extern gimple_opt_pass *make_pass_tree_unswitch (gcc::context *ctxt);
 extern gimple_opt_pass *make_pass_loop_split (gcc::context *ctxt);
+extern gimple_opt_pass *make_pass_early_predcom (gcc::context *ctxt);
 extern gimple_opt_pass *make_pass_predcom (gcc::context *ctxt);
 extern gimple_opt_pass *make_pass_iv_canon (gcc::context *ctxt);
 extern gimple_opt_pass *make_pass_scev_cprop (gcc::context *ctxt);
Index: gcc/tree-predcom.c
===================================================================
--- gcc/tree-predcom.c	2017-07-03 08:42:47.632532107 +0100
+++ gcc/tree-predcom.c	2017-07-03 09:29:08.744451338 +0100
@@ -218,7 +218,7 @@  Free Software Foundation; either version
 /* The maximum number of iterations between the considered memory
    references.  */
 
-#define MAX_DISTANCE (target_avail_regs < 16 ? 4 : 8)
+#define MAX_DISTANCE (target_avail_regs < 16 ? 3 : 7)
 
 /* Data references (or phi nodes that carry data reference values across
    loop iterations).  */
@@ -343,6 +343,71 @@  struct component
 
 static hash_map<tree, name_expansion *> *name_expansions;
 
+/* True if we're running the early predcom pass and should only handle
+   cases that aid vectorization.  Specifically this means that:
+
+   - only CT_INVARIANT and CT_STORE_LOAD chains are used
+   - the maximum distance for a CT_STORE_LOAD chain is 1 iteration,
+     and at that distance the store must come after the load
+   - there's no unrolling or detection of looparound phis.
+
+   The idea is handle inductions that go via memory, such as:
+
+     for (int i = 1; i < n; ++i)
+       x[i] = x[i - 1] + 1;
+
+   As it stands this loop could never be vectorized, because a loop body
+   that contains a read of x[j] followed by a write to x[j + 1] would
+   have its vectorization factor limited to 1.  Transforming it to:
+
+     int tmp = x[0];
+     for (int i = 0; i < n; ++i)
+       {
+         tmp += 1;
+         x[i] = tmp:
+       }
+
+   exposes the fact that the stored value is a simple vectorizable
+   induction with start value x[0] and step 1.
+
+   Commoning is not always useful even in this situation.  For example,
+   carrying over the value of x[i] won't help us to vectorize:
+
+     for (int i = 1; i < n; ++i)
+       {
+	 y[i] = x[i - 1];
+	 x[i] += i;
+       }
+
+   There's no real need to restrict things further though, because we're
+   unable to vectorize these load/store combinations in their current
+   form whatever happens.
+
+   We require the store to come after the load when the distance is 1
+   to avoid cases like:
+
+     for (int i = 1; i < n; ++i)
+       {
+	 x[i] = ...;
+	 ... = x[i - 1];
+       }
+
+   These accesses effectively have a distance somewhere between 1 and 2,
+   since after commoning the value stored in the previous iteration would
+   still be live at the next store.  This means that the combination
+   isn't useful for exposing simple inductions.
+
+   Also, unlike the motivating case above, this combination does not
+   prevent vectorization.  If a write to x[j + 1] comes before a read
+   of x[j], the vectorized write completes for all vector elements
+   before the read starts for any vector elements.  */
+
+static bool only_simple_p;
+
+/* The maximum loop carry distance for this execution of the pass.  */
+
+static int max_distance;
+
 /* Dumps data reference REF to FILE.  */
 
 extern void dump_dref (FILE *, dref);
@@ -960,7 +1025,13 @@  add_ref_to_chain (chain_p chain, dref re
 
   gcc_assert (wi::les_p (root->offset, ref->offset));
   widest_int dist = ref->offset - root->offset;
-  if (wi::leu_p (MAX_DISTANCE, dist))
+  /* When running before vectorization, only allow the maximum distance
+     if the consumer comes before the producer.  See the comment above
+     ONLY_SIMPLE_P for details.  */
+  if (wi::ltu_p (max_distance, dist)
+      || (only_simple_p
+	  && wi::eq_p (max_distance, dist)
+	  && root->pos < ref->pos))
     {
       free (ref);
       return;
@@ -1194,6 +1265,11 @@  add_looparound_copies (struct loop *loop
   dref ref, root = get_chain_root (chain);
   gphi *phi;
 
+  /* There's no point doing this when running before vectorization,
+     since we won't unroll the loop or combine chains.  */
+  if (only_simple_p)
+    return;
+
   FOR_EACH_VEC_ELT (chain->refs, i, ref)
     {
       phi = find_looparound_phi (loop, ref, root);
@@ -1232,7 +1308,7 @@  determine_roots_comp (struct loop *loop,
   FOR_EACH_VEC_ELT (comp->refs, i, a)
     {
       if (!chain || DR_IS_WRITE (a->ref)
-	  || wi::leu_p (MAX_DISTANCE, a->offset - last_ofs))
+	  || wi::ltu_p (max_distance, a->offset - last_ofs))
 	{
 	  if (nontrivial_chain_p (chain))
 	    {
@@ -1241,6 +1317,14 @@  determine_roots_comp (struct loop *loop,
 	    }
 	  else
 	    release_chain (chain);
+	  /* Only create CT_STORE_LOAD and CT_INVARIANT chains when
+	     running before vectorization.  */
+	  if (only_simple_p && !DR_IS_WRITE (a->ref))
+	    {
+	      free (a);
+	      chain = NULL;
+	      continue;
+	    }
 	  chain = make_rooted_chain (a);
 	  last_ofs = a->offset;
 	  continue;
@@ -1759,6 +1843,10 @@  determine_unroll_factor (vec<chain_p> ch
   unsigned factor = 1, af, nfactor, i;
   unsigned max = PARAM_VALUE (PARAM_MAX_UNROLL_TIMES);
 
+  /* Do not unroll when running before vectorization.  */
+  if (only_simple_p)
+    return 1;
+
   FOR_EACH_VEC_ELT (chains, i, chain)
     {
       if (chain->type == CT_INVARIANT)
@@ -2562,8 +2650,11 @@  tree_predictive_commoning_loop (struct l
     }
   prepare_initializers (loop, chains);
 
-  /* Try to combine the chains that are always worked with together.  */
-  try_combine_chains (&chains);
+  /* During the main pass, try to combine the chains that are always
+     worked with together.  For the early pass it should be better
+     to leave this to the vectorizer.  */
+  if (!only_simple_p)
+    try_combine_chains (&chains);
 
   if (dump_file && (dump_flags & TDF_DETAILS))
     {
@@ -2623,15 +2714,19 @@  tree_predictive_commoning_loop (struct l
   return unroll;
 }
 
-/* Runs predictive commoning.  */
+/* Runs predictive commoning.  EARLY_P is true if we are running before
+   vectorization.  */
 
 unsigned
-tree_predictive_commoning (void)
+tree_predictive_commoning (bool early_p)
 {
   bool unrolled = false;
   struct loop *loop;
   unsigned ret = 0;
 
+  only_simple_p = early_p;
+  max_distance = early_p ? 1 : MAX_DISTANCE;
+
   initialize_original_copy_tables ();
   FOR_EACH_LOOP (loop, LI_ONLY_INNERMOST)
     if (optimize_loop_for_speed_p (loop))
@@ -2649,19 +2744,51 @@  tree_predictive_commoning (void)
   return ret;
 }
 
-/* Predictive commoning Pass.  */
+/* Predictive commoning pass.  EARLY_P is true if we are running before
+   vectorization.  */
 
 static unsigned
-run_tree_predictive_commoning (struct function *fun)
+run_tree_predictive_commoning (struct function *fun, bool early_p)
 {
   if (number_of_loops (fun) <= 1)
     return 0;
 
-  return tree_predictive_commoning ();
+  return tree_predictive_commoning (early_p);
 }
 
 namespace {
 
+const pass_data pass_data_early_predcom =
+{
+  GIMPLE_PASS, /* type */
+  "epcom", /* name */
+  OPTGROUP_LOOP, /* optinfo_flags */
+  TV_PREDCOM, /* tv_id */
+  PROP_cfg, /* properties_required */
+  0, /* properties_provided */
+  0, /* properties_destroyed */
+  0, /* todo_flags_start */
+  TODO_update_ssa_only_virtuals, /* todo_flags_finish */
+};
+
+class pass_early_predcom : public gimple_opt_pass
+{
+public:
+  pass_early_predcom (gcc::context *ctxt)
+    : gimple_opt_pass (pass_data_early_predcom, ctxt)
+  {}
+
+  /* opt_pass methods: */
+  virtual bool gate (function *)
+  {
+    return flag_predictive_commoning && flag_tree_loop_vectorize;
+  }
+  virtual unsigned int execute (function *fun)
+  {
+    return run_tree_predictive_commoning (fun, true);
+  }
+}; // class pass_early_predcom
+
 const pass_data pass_data_predcom =
 {
   GIMPLE_PASS, /* type */
@@ -2686,7 +2813,7 @@  const pass_data pass_data_predcom =
   virtual bool gate (function *) { return flag_predictive_commoning != 0; }
   virtual unsigned int execute (function *fun)
     {
-      return run_tree_predictive_commoning (fun);
+      return run_tree_predictive_commoning (fun, false);
     }
 
 }; // class pass_predcom
@@ -2694,9 +2821,13 @@  const pass_data pass_data_predcom =
 } // anon namespace
 
 gimple_opt_pass *
+make_pass_early_predcom (gcc::context *ctxt)
+{
+  return new pass_early_predcom (ctxt);
+}
+
+gimple_opt_pass *
 make_pass_predcom (gcc::context *ctxt)
 {
   return new pass_predcom (ctxt);
 }
-
-
Index: gcc/testsuite/gnat.dg/vect18.adb
===================================================================
--- gcc/testsuite/gnat.dg/vect18.adb	2015-10-14 14:58:56.000000000 +0100
+++ gcc/testsuite/gnat.dg/vect18.adb	2017-07-03 09:17:28.627495621 +0100
@@ -1,5 +1,5 @@ 
 -- { dg-do compile { target i?86-*-* x86_64-*-* } }
--- { dg-options "-O3 -msse2 -fdump-tree-vect-details" }
+-- { dg-options "-O3 -msse2 -fdump-tree-vect-details -fno-predictive-commoning" }
 
 package body Vect18 is