Re: [PR66726] Factor conversion out of COND_EXPR

[Kugan <kugan dot vivekanandarajah at linaro dot org>]

On 07/07/15 07:37, Jeff Law wrote:
> On 07/04/2015 06:32 AM, Kugan wrote:

> I would also verify that this turns into a MIN_EXPR.  I think the patch
> as-written won't detect the MIN_EXPR until the _next_ time phi-opt is
> called.  And one of the benefits we're really looking for here is to
> remove barriers to finding these min/max expressions.
> 

> 
>> +
>> diff --git a/gcc/tree-ssa-phiopt.c b/gcc/tree-ssa-phiopt.c
>> index d2a5cee..12ab9ee 100644
>> --- a/gcc/tree-ssa-phiopt.c
>> +++ b/gcc/tree-ssa-phiopt.c
>> @@ -73,6 +73,7 @@ along with GCC; see the file COPYING3.  If not see
>>   static unsigned int tree_ssa_phiopt_worker (bool, bool);
>>   static bool conditional_replacement (basic_block, basic_block,
>>                        edge, edge, gphi *, tree, tree);
>> +static bool factor_out_conditional_conversion (edge, edge, gphi *,
>> tree, tree);
>>   static int value_replacement (basic_block, basic_block,
>>                     edge, edge, gimple, tree, tree);
>>   static bool minmax_replacement (basic_block, basic_block,
>> @@ -342,6 +343,8 @@ tree_ssa_phiopt_worker (bool do_store_elim, bool
>> do_hoist_loads)
>>           cfgchanged = true;
>>         else if (minmax_replacement (bb, bb1, e1, e2, phi, arg0, arg1))
>>           cfgchanged = true;
>> +      else if (factor_out_conditional_conversion (e1, e2, phi, arg0,
>> arg1))
>> +        cfgchanged = true;
> So this transformation does not inherently change the CFG, so setting
> CFGCHANGED isn't really appropriate and may trigger unnecessary cleanups.
> 
> I think the transformation needs to occur prior this if-elseif-else
> block since the transformation should enable the code in the
> if-elseif-else block to find more optimization opportunities.
> 
> That will also imply we either restart after the transformation applies,
> or we update the local variables that are used as arguments to
> conditional_replacement, abs_replacement and minmax_replacement.
> 
> 
>>       }
>>       }
>>
>> @@ -410,6 +413,108 @@ replace_phi_edge_with_variable (basic_block
>> cond_block,
>>             bb->index);
>>   }
>>
>> +/* PR66726: Factor conversion out of COND_EXPR.  If the arguments of
>> the PHI
>> +   stmt are CONVERT_STMT, factor out the conversion and perform the
>> conversion
>> +   to the result of PHI stmt.  */
>> +
>> +static bool
>> +factor_out_conditional_conversion (edge e0, edge e1, gphi *phi,
>> +                   tree arg0, tree arg1)
>> +{
>> +  gimple def0 = NULL, def1 = NULL, new_stmt;
>> +  tree new_arg0 = NULL_TREE, new_arg1 = NULL_TREE;
>> +  tree temp, result;
>> +  gimple_stmt_iterator gsi;
>> +
>> +  /* One of the arguments has to be SSA_NAME and other argument can
>> +     be an SSA_NAME of INTEGER_CST.  */
>> +  if ((TREE_CODE (arg0) != SSA_NAME
>> +       && TREE_CODE (arg0) != INTEGER_CST)
>> +      || (TREE_CODE (arg1) != SSA_NAME
>> +      && TREE_CODE (arg1) != INTEGER_CST)
>> +      || (TREE_CODE (arg0) == INTEGER_CST
>> +      && TREE_CODE (arg1) == INTEGER_CST))
>> +    return false;
>> +
>> +  /* Handle only PHI statements with two arguments.  TODO: If all
>> +     other arguments to PHI are INTEGER_CST, we can handle more
>> +     than two arguments too.  */
>> +  if (gimple_phi_num_args (phi) != 2)
>> +    return false;
> If you're just handling two arguments, then it's probably easiest to
> just swap arg0/arg1 e0/e1 if arg0 is not an SSA_NAME like this:
> 
>  /* First canonicalize to simplify tests.  */
>   if (TREE_CODE (arg0) != SSA_NAME)
>     {
>       std::swap (arg0, arg1);
>       std::swap (e0, e1);
>     }
> 
>   if (TREE_CODE (arg0) != SSA_NAME)
>     return false;
> 
> That simplifies things a bit since you're going to know from thsi point
> forward that arg0 is an SSA_NAME.
> 
> 
> 
>> +
>> +  /* If arg0 is an SSA_NAME and the stmt which defines arg0 is
>> +     a CONVERT_STMT, use the LHS as new_arg0.  */
>> +  if (TREE_CODE (arg0) == SSA_NAME)
>> +    {
>> +      def0 = SSA_NAME_DEF_STMT (arg0);
>> +      if (!is_gimple_assign (def0)
>> +      || !CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (def0)))
>> +    return false;
>> +      new_arg0 = gimple_assign_rhs1 (def0);
>> +    }
> Use gimple_assign_cast_p rather than checking CONVERT_EXPR_CODE_P
> directly, so something like:
> 
>   /* Now see if ARG0 was defined by a typecast.  */
>   gimple arg0_def = SSA_NAME_DEF_STMT (arg0);
> 
>   if (!is_gimple_assign (arg0_def) || !gimple_assign_cast_p (arg0_def))
>     return false;
> 
> Similarly for arg1 when it's an SSA_NAME.
> 
> 
>> +
>> +  /* If types of new_arg0 and new_arg1 are different bailout.  */
>> +  if (TREE_TYPE (new_arg0) != TREE_TYPE (new_arg1))
>> +    return false;
> Do we want to restrict this to just integral types?    I haven't though
> about it too deeply, so perhaps not.
> 
>> +
>> +  /* Replace the PHI stmt with the new_arg0 and new_arg1.  Also insert
>> +     a new CONVERT_STMT that converts the phi results.  */
>> +  gsi = gsi_after_labels (gimple_bb (phi));
>> +  result = PHI_RESULT (phi);
>> +  temp = make_ssa_name (TREE_TYPE (new_arg0), phi);
>> +
>> +  if (dump_file && (dump_flags & TDF_DETAILS))
>> +    {
>> +      fprintf (dump_file, "PHI ");
>> +      print_generic_expr (dump_file, gimple_phi_result (phi), 0);
>> +      fprintf (dump_file,
>> +           " changed to factor CONVERT_EXPR out from COND_EXPR.\n");
>> +      fprintf (dump_file, "New PHI_RESULT is ");
>> +      print_generic_expr (dump_file, temp, 0);
>> +      fprintf (dump_file, " and new stmt with CONVERT_EXPR defines ");
>> +      print_generic_expr (dump_file, result, 0);
>> +      fprintf (dump_file, ".\n");
>> +    }
>> +
>> +  gimple_phi_set_result (phi, temp);
>> +  SET_PHI_ARG_DEF (phi, e0->dest_idx, new_arg0);
>> +  SET_PHI_ARG_DEF (phi, e1->dest_idx, new_arg1);
>> +  new_stmt = gimple_build_assign (result, CONVERT_EXPR, temp);
>> +  gsi_insert_before (&gsi, new_stmt, GSI_SAME_STMT);
>> +  return true;
> So I think you want to also remove the old cast(s) so that the minmax
> optimization can apply without having to wait for another pass of phi-opt.
> 
> To safely remove the old cast(s) you have to verify the result of the
> cast has a single use (which is obviously in the PHI).  Otherwise your
> transformation will have introduced a runtime redundancy.
> 
> I also suspect it's better to create a new PHI rather than modify the
> original PHI.  The original PHI should be removed and the result re-used
> as the result of the new convert statement.
> 
> Extra points if you can easily make this transformation apply to a
> generic unary operator.  So for example, we might have a sinkable bit-not.
> 
> Overall it's heading the right direction.  But I think it needs another
> iteration.


Thanks for the review. I have addressed your comments above in the
attached patch.

I have one question with respect to unary operation. For generic unary
operation with INTEGER_CST, do we skip this or do we have to perform the
inverse operation so that the conversion after PHI will restore it? Is
there any easy way to do this safely?

Bootstrapped and regression tested the attached patch on
x86-64-none-linux-gnu with no new regressions.

Thanks,
Kugan

Attachment: p.txt
Description: Text document