Re: [PATCH GCC]Improve tree ifconv by handling virtual PHIs which can be degenerated.

["Bin.Cheng" <amker dot cheng at gmail dot com>]

On Fri, Apr 22, 2016 at 11:47 AM, Richard Biener
<richard.guenther@gmail.com> wrote:
> On Fri, Apr 22, 2016 at 12:33 PM, Bin.Cheng <amker.cheng@gmail.com> wrote:
>> On Fri, Apr 22, 2016 at 11:25 AM, Richard Biener
>> <richard.guenther@gmail.com> wrote:
>>> On Fri, Apr 22, 2016 at 12:07 PM, Bin Cheng <Bin.Cheng@arm.com> wrote:
>>>> Hi,
>>>> Tree if-conv has below code checking on virtual PHI nodes in if_convertible__phi_p:
>>>>
>>>>   if (any_mask_load_store)
>>>>     return true;
>>>>
>>>>   /* When there were no if-convertible stores, check
>>>>      that there are no memory writes in the branches of the loop to be
>>>>      if-converted.  */
>>>>   if (virtual_operand_p (gimple_phi_result (phi)))
>>>>     {
>>>>       imm_use_iterator imm_iter;
>>>>       use_operand_p use_p;
>>>>
>>>>       if (bb != loop->header)
>>>>         {
>>>>           if (dump_file && (dump_flags & TDF_DETAILS))
>>>>             fprintf (dump_file, "Virtual phi not on loop->header.\n");
>>>>           return false;
>>>>         }
>>>>
>>>>       FOR_EACH_IMM_USE_FAST (use_p, imm_iter, gimple_phi_result (phi))
>>>>         {
>>>>           if (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI
>>>>               && USE_STMT (use_p) != phi)
>>>>             {
>>>>               if (dump_file && (dump_flags & TDF_DETAILS))
>>>>                 fprintf (dump_file, "Difficult to handle this virtual phi.\n");
>>>>               return false;
>>>>             }
>>>>         }
>>>>     }
>>>>
>>>> After investigation, I think it's to bypass code in the form of:
>>>>
>>>> <bb header>
>>>>   .MEM_2232 = PHI <.MEM_574(179), .MEM_1247(183)>     // <----PHI_1
>>>>   ...
>>>>   if (cond)
>>>>     goto <bb 1>
>>>>   else
>>>>     goto <bb2>
>>>>
>>>> <bb 1>:  //empty
>>>> <bb2>:
>>>>   .MEM_1247 = PHI <.MEM_2232(180), .MEM_2232(181)>     // <----PHI_2
>>>>   if (cond2)
>>>>     goto <bb exit>
>>>>   else
>>>>     goto <bb latch>
>>>>
>>>> <bb latch>:
>>>>   goto <bb header>
>>>>
>>>> Here PHI_2 can be degenerated and deleted.  Furthermore, after propagating .MEM_2232 to .MEM_1247's uses, PHI_1 can also be degenerated and deleted in this case.  These cases are bypassed because tree if-conv doesn't handle virtual PHI nodes during loop conversion (it only predicates scalar PHI nodes).  Of course this results in loops not converted, and not vectorized.
>>>> This patch firstly deletes the aforementioned checking code, then adds code handling such virtual PHIs during conversion.  The use of `any_mask_load_store' now is less ambiguous with this change, which allows further cleanups and patches fixing PR56541.
>>>> BTW, I think the newly fix at PR70725 on PHI nodes with only one argument is a special case covered by this change too.  Unfortunately I can't use replace_uses_by because I need to handle PHIs at use point after replacing too.  This doesn't really matter since we only care about virtual PHI, it's not possible to be used by anything other than IR itself.
>>>> Bootstrap and test on x86_64 and AArch64, is it OK if no regressions?
>>>
>>> Doesn't this undo my fix for degenerate non-virtual PHIs?
>> No, since we already support degenerate non-virtual PHIs in
>> predicate_scalar_phi, your fix is also for virtual PHIs handling.
>
> Was it?  I don't remember ;)  I think it was for a non-virtual PHI.
> Anyway, you should
> see the PR70725 testcase fail again if not.
>
>>>
>>> I believe we can just drop virtual PHIs and rely on
>>>
>>>   if (any_mask_load_store)
>>>     {
>>>       mark_virtual_operands_for_renaming (cfun);
>>>       todo |= TODO_update_ssa_only_virtuals;
>>>     }
>>>
>>> re-creating them from scratch.  To do better than that we'd simply
>> I tried this, simply enable above code for all cases can't resolve
>> verify_ssa issue.  I haven't look into the details, looks like ssa
>> def-use chain is corrupted in if-conversion if we don't process it
>> explicitly.  Maybe it's possible along with your below suggestions,
>> but we need to handle uses outside of loop too.
>
> Yes.  I don't like all the new code to deal with virtual PHIs when doing
> it correctly would also avoid the above virtual SSA update ...
>
> After all the above seems to work for the case of if-converted stores
> (which is where virtual PHIs appear as well, even not degenerate).
> So I don't see exactly how it would break in the other case.  I suppose
> you may need to call mark_virtual_phi_result_for_renaming () on
> all virtual PHIs.
>
Hi Richard,
Here is the updated patch.  It also fixes PR70771 & PR70775. Root
cause for the ICE is in the fix to PR70725 because it forgot to
release single-argument PHI nodes after replacing uses.  In
combine_blocks, these PHIs are removed from basic block but are still
live in IR.  As a result, the ssa def/use chain for these PHIs are in
broken state, thus ICE is triggered whenever ssa use list is
accessed..
In this updated patch, I made below change to update virtual ssa
unconditionally.  With this change, we don't need to handle virtual
PHIs explicitly, and single-argument PHI related code in fix to
PR70725 can also be removed.

@@ -2808,11 +2758,8 @@ tree_if_conversion (struct loop *loop)
     }

   todo |= TODO_cleanup_cfg;
-  if (any_mask_load_store)
-    {
-      mark_virtual_operands_for_renaming (cfun);
-      todo |= TODO_update_ssa_only_virtuals;
-    }
+  mark_virtual_operands_for_renaming (cfun);
+  todo |= TODO_update_ssa_only_virtuals;

  cleanup:
   if (ifc_bbs)


BTW, it may be possible to only update affected PHIs using
mark_virtual_phi_result_for_renaming.  This patch didn't do that since
the update is done once per function anyway.
Bootstrap and test on x86_64, is it OK?

Thanks,
bin

2016-04-25  Bin Cheng  <bin.cheng@arm.com>

    PR tree-optimization/70771
    PR tree-optimization/70775
    * tree-if-conv.c (if_convertible_phi_p): Remove check on special
    virtual PHI nodes.  Delete parameter.
    (if_convertible_loop_p_1): Delete argument to above function.
    (predicate_all_scalar_phis): Delete code handling single-argument
    PHIs.
    (tree_if_conversion): Mark and update virtual SSA.

gcc/testsuite/ChangeLog
2016-04-25  Bin Cheng  <bin.cheng@arm.com>

    PR tree-optimization/70771
    PR tree-optimization/70775
    * gcc.dg/pr70771.c: New test.
    * gcc.dg/pr70771.c: New test.

Attachment: virtual-phi-ifconv-20160424.txt
Description: Text document