Re: [RFC][IPA-VRP] Early VRP Implementation

[Richard Biener <richard.guenther@gmail.com>]

On Wed, Aug 3, 2016 at 3:17 AM, kugan <kugan.vivekanandarajah@linaro.org> wrote:
> Hi Richard,
>
> Thanks for the review.
>
> On 28/07/16 21:34, Richard Biener wrote:
>>
>> On Thu, Jul 28, 2016 at 9:35 AM, kugan
>> <kugan.vivekanandarajah@linaro.org> wrote:
>>>
>>> Hi Richard,
>>>
>>> Thanks for the review.
>>>>
>>>>
>>>>
>>>> It seems that in your pop_value_range you assume you only pop one
>>>> range per BB - while that's likely true at the moment it will be a
>>>> limitation
>>>> in the future.  You want to pop ranges until you hit the NULL marker
>>>> in after_dom_children and unconditionally push a NULL marker.
>>>>
>>> I understand. Right now, I am adding only one assert based on the
>>> condition.
>>> But in future, we will be adding more so this is needed. I will do that.
>>>
>>>> For example to match current VRPs behavior on say
>>>>
>>>>    i_2 = (int) j_3;
>>>>    if (i_2 < 0)
>>>>      ...
>>>>
>>>> which can register an assert for j_3 when i_2 < 0 is true we'd do that
>>>> by re-simulating DEFs of uses we figured out new ranges of (and all
>>>> their uses).  All those ranges would be temporary as well, thus they'd
>>>> need to be pushed/popped.  In my quick prototype this was done
>>>> using a worklist seeded by the names we can derive a range from from
>>>> conditionals and "SSA propagating" from it.  Note that for this
>>>> the generic vrp_visit_stmt cannot be re-used as it doesn't push/pop,
>>>> factoring out the lattice update is what is needed here.
>>>>
>>>
>>> I dont think I understand this part. vrp_visit_stmt is going to add value
>>> ranges for the variables defined in the if-block (in the example below it
>>> is
>>> for t). If we push the value range for i_2 and j_3 when we enter
>>> if-block,
>>> vrp_visit_stmt should compute "t" correctly. When we leave the if-block,
>>> we
>>> will pop i_2 and j_3.
>>>
>>>     i_2 = (int) j_3;
>>>     if (i_2 < 0)
>>>     {
>>>       t = j_2 * 2;
>>>     }
>>> Am I missing something here?
>>
>>
>> It works if you push the old value before calling vrp_visit_stmt, yes.
>> But I think
>> you want to do that only if the value-range changed to avoid too many
>> changes
>> on the stack.  I guess we can defer further refactoring and
>> optimization of this case
>> to the point where we consider looking back very aggressively.
>>
>>>> +/* Visit the basic blocks in the dominance order and set the Value
>>>> Ranges
>>>> (VR)
>>>> +   for SSA_NAMEs in the scope.  Use this VR to discover more VRs.
>>>> Restore the
>>>> +   old VR once the scope is exited.  */
>>>> +
>>>> +static bool
>>>> +evrp_visit_phi_node_local (gphi *phi)
>>>> +{
>>>> +  size_t i;
>>>> +  tree lhs = PHI_RESULT (phi);
>>>> +  value_range vr_result = VR_INITIALIZER;
>>>> +  bool first = true;
>>>> +  int edges;
>>>> +
>>>> +  edges = 0;
>>>> +  for (i = 0; i < gimple_phi_num_args (phi); i++)
>>>> +    {
>>>> +      edge e = gimple_phi_arg_edge (phi, i);
>>>> +      tree arg = PHI_ARG_DEF (phi, i);
>>>> +      value_range vr_arg = VR_INITIALIZER;
>>>> +      ++edges;
>>>> +
>>>> +      /* If there is a back-edge, set the result to VARYING.  */
>>>> +      if (e->flags & (EDGE_DFS_BACK | EDGE_COMPLEX))
>>>> +       {
>>>> +         set_value_range_to_varying (&vr_result);
>>>> +         break;
>>>> +       }
>>>> ...
>>>> +      /* If any of the RHS value is VARYING, set the result to VARYING.
>>>> */
>>>> +      if ((vr_arg.type != VR_RANGE)
>>>> +         && (vr_arg.type != VR_ANTI_RANGE))
>>>> +       {
>>>> +         set_value_range_to_varying (&vr_result);
>>>> +         break;
>>>> +       }
>>>>
>>>> this shows that you need to start conservative for a DOM based VRP,
>>>> thus with all lattice values initialized to VARYING (but undefined SSA
>>>> names of course still can be UNDEFINED) rather than UNDEFINED.
>>>>
>>>> +      if (TREE_CODE (arg) == SSA_NAME)
>>>> +       vr_arg = *(get_value_range (arg));
>>>> +      else
>>>> +       set_value_range_to_varying (&vr_arg);
>>>>
>>>> err - what about constants?  When you initialize the lattice properly
>>>> you should be able to re-use vrp_visit_phi_node (maybe split out
>>>> its head to avoid using SCEV or the iteration limitation).
>>>
>>>
>>>
>>> I also like re-using vrp_visit_phi_node but the issue is, we will have to
>>> keep a work-list of nodes to be re-evaluated till the lattice reach a
>>> fixpoint. Is that OK with you?
>>
>>
>> No, why would you need to iterate here?  As said, the key point is to
>> initialize value-ranges as VARYING rather than UNDEFINED.
>>
>>> If we are to do this, we should be able to reuse the callbacks
>>> vrp_visit_phi_node and vrp_visit_stmt as it is.
>>>
>>> Do you have a reference to your DOM based prototype?
>>
>>
>> I never posted it I think, it's structure is similar to yours with lots
>> of ??? comments ;)
>>
>
>
> Here is an updated patch which addresses the earlier review comments.
>
> Just to see the effectiveness of this, I did a simple test.
>
> That is, I built gcc with --enable-languages=c,c++ --disable-bootstrap
> --disable-multilib and added -fdump-ipa-cp to the compiler flag and grepped
> for number of times ipa-vrp (with the ipa-vrp patch) is setting the value
> range for argument. I also did the same with tree-vrp used in place of
> tree-evrp as an early vrp. tree-evrp is setting 186 times compared to
> tree-vrp which is setting 207 times. I didn't see the actual value ranges
> which can also make lots of difference.
>
> In future we might want to iterate on dom based vrp till fixed point is
> reached if there is a need.

diff --git a/gcc/common.opt b/gcc/common.opt
index 8a292ed..7028cd4 100644
--- a/gcc/common.opt
+++ b/gcc/common.opt
@@ -2482,6 +2482,10 @@ ftree-vrp
 Common Report Var(flag_tree_vrp) Init(0) Optimization
 Perform Value Range Propagation on trees.

+fdisable-tree-evrp
+Common Report Var(flag_disable_early_vrp) Init(0) Optimization
+Disable Early Value Range Propagation on trees.
+

no please, this is automatically supported via -fdisable-

@@ -1728,11 +1736,12 @@ extract_range_from_assert (value_range *vr_p, tree expr)
     always false.  */

 static void
-extract_range_from_ssa_name (value_range *vr, tree var)
+extract_range_from_ssa_name (value_range *vr, bool dom_p, tree var)
 {
   value_range *var_vr = get_value_range (var);

-  if (var_vr->type != VR_VARYING)
+  if (var_vr->type != VR_VARYING
+      && (!dom_p || var_vr->type != VR_UNDEFINED))
     copy_value_range (vr, var_vr);
   else
     set_value_range (vr, VR_RANGE, var, var, NULL);

why do you need these changes?  I think I already told you you need to
initialize the lattice to sth else than VR_UNDEFINED and that you can't
fully re-use update_value_range.  If you don't want to do that then instead
of doing changes all over the place do it in get_value_range and have a
global flag.


@@ -3594,7 +3643,8 @@ extract_range_from_cond_expr (value_range *vr,
gassign *stmt)
    on the range of its operand and the expression code.  */

 static void
-extract_range_from_comparison (value_range *vr, enum tree_code code,
+extract_range_from_comparison (value_range *vr,
+                              enum tree_code code,
                               tree type, tree op0, tree op1)
 {
   bool sop = false;

remove these kind of no-op changes.

+/* Initialize local data structures for VRP.  If DOM_P is true,
+   we will be calling this from early_vrp where value range propagation
+   is done by visiting stmts in dominator tree.  ssa_propagate engine
+   is not used in this case and that part of the ininitialization will
+   be skipped.  */

 static void
-vrp_initialize (void)
+vrp_initialize (bool dom_p)
 {
   basic_block bb;

@@ -6949,6 +7010,9 @@ vrp_initialize (void)
   vr_phi_edge_counts = XCNEWVEC (int, num_ssa_names);
   bitmap_obstack_initialize (&vrp_equiv_obstack);

+  if (dom_p)
+    return;
+

split the function instead.

@@ -7926,7 +7992,8 @@ vrp_visit_switch_stmt (gswitch *stmt, edge *taken_edge_p)
    If STMT produces a varying value, return SSA_PROP_VARYING.  */

 static enum ssa_prop_result
-vrp_visit_stmt (gimple *stmt, edge *taken_edge_p, tree *output_p)
+vrp_visit_stmt_worker (gimple *stmt, bool dom_p,  edge *taken_edge_p,
+                      tree *output_p)
 {
   tree def;
   ssa_op_iter iter;
@@ -7940,7 +8007,7 @@ vrp_visit_stmt (gimple *stmt, edge
*taken_edge_p, tree *output_p)
   if (!stmt_interesting_for_vrp (stmt))
     gcc_assert (stmt_ends_bb_p (stmt));
   else if (is_gimple_assign (stmt) || is_gimple_call (stmt))
-    return vrp_visit_assignment_or_call (stmt, output_p);
+    return vrp_visit_assignment_or_call (stmt, dom_p, output_p);
   else if (gimple_code (stmt) == GIMPLE_COND)
     return vrp_visit_cond_stmt (as_a <gcond *> (stmt), taken_edge_p);
   else if (gimple_code (stmt) == GIMPLE_SWITCH)
@@ -7954,6 +8021,12 @@ vrp_visit_stmt (gimple *stmt, edge
*taken_edge_p, tree *output_p)
   return SSA_PROP_VARYING;
 }

+static enum ssa_prop_result
+vrp_visit_stmt (gimple *stmt, edge *taken_edge_p, tree *output_p)
+{
+  return vrp_visit_stmt_worker (stmt, false, taken_edge_p, output_p);
+}

as said the refactoring that would be appreciated is to split out the
update_value_range calls
from the worker functions so you can call the respective functions
from the DOM implementations.
That they are globbed in vrp_visit_stmt currently is due to the API of
the SSA propagator.

@@ -8768,6 +8842,12 @@ vrp_visit_phi_node (gphi *phi)
              fprintf (dump_file, "\n");
            }

+         if (dom_p && vr_arg.type == VR_UNDEFINED)
+           {
+             set_value_range_to_varying (&vr_result);
+             break;
+           }
+

eh...  ok, so another way to attack this is, instead of initializing
the lattice to sth else
than VR_UNDEFINED, make sure to drop the lattice to varying for all PHI args on
yet unvisited incoming edges (you're not doing optimistic VRP).  That's the only
place you _have_ to do it.

Richard.



> Thanks,
> Kugan
>
>