Re: [patch tree-optimization]: allow branch-cost optimization for truth-and/or on mode-expanded simple boolean-operands

[Kai Tietz <ktietz70@googlemail.com>]

2011/10/21 Richard Guenther <richard.guenther@gmail.com>:
> On Thu, Oct 20, 2011 at 3:08 PM, Kai Tietz <ktietz@redhat.com> wrote:
>> Hello,
>>
>> this patch re-enables the branch-cost optimization on simple boolean-typed operands, which are casted to a wider integral type.  This happens due casts from
>> boolean-types are preserved, but FE might expands simple-expression to wider mode.
>>
>> I added two tests for already working branch-cost optimization for IA-architecture and
>> two for explicit checking for boolean-type.
>>
>> ChangeLog
>>
>> 2011-10-20  Kai Tietz  <ktietz@redhat.com>
>>
>>        * fold-const.c (simple_operand_p_2): Handle integral
>>        casts from boolean-operands.
>>
>> 2011-10-20  Kai Tietz  <ktietz@redhat.com>
>>
>>        * gcc.target/i386/branch-cost1.c: New test.
>>        * gcc.target/i386/branch-cost2.c: New test.
>>        * gcc.target/i386/branch-cost3.c: New test.
>>        * gcc.target/i386/branch-cost4.c: New test.
>>
>> Bootstrapped and regression tested on x86_64-unknown-linux-gnu for all languages including Ada and Obj-C++.  Ok for apply?
>>
>> Regards,
>> Kai
>>
>> Index: gcc-head/gcc/testsuite/gcc.target/i386/branch-cost2.c
>> ===================================================================
>> --- /dev/null
>> +++ gcc-head/gcc/testsuite/gcc.target/i386/branch-cost2.c
>> @@ -0,0 +1,16 @@
>> +/* { dg-do compile } */
>> +/* { dg-options "-O2 -fdump-tree-gimple -mbranch-cost=2" } */
>> +
>> +extern int doo (void);
>> +
>> +int
>> +foo (int a, int b)
>> +{
>> +  if (a && b)
>> +   return doo ();
>> +  return 0;
>> +}
>> +
>> +/* { dg-final { scan-tree-dump-times "if " 1 "gimple" } } */
>> +/* { dg-final { scan-tree-dump-times " & " 1 "gimple" } } */
>> +/* { dg-final { cleanup-tree-dump "gimple" } } */
>> Index: gcc-head/gcc/testsuite/gcc.target/i386/branch-cost3.c
>> ===================================================================
>> --- /dev/null
>> +++ gcc-head/gcc/testsuite/gcc.target/i386/branch-cost3.c
>> @@ -0,0 +1,16 @@
>> +/* { dg-do compile } */
>> +/* { dg-options "-O2 -fdump-tree-gimple -mbranch-cost=2" } */
>> +
>> +extern int doo (void);
>> +
>> +int
>> +foo (_Bool a, _Bool b)
>> +{
>> +  if (a && b)
>> +   return doo ();
>> +  return 0;
>> +}
>> +
>> +/* { dg-final { scan-tree-dump-times "if " 1 "gimple" } } */
>> +/* { dg-final { scan-tree-dump-times " & " 1 "gimple" } } */
>> +/* { dg-final { cleanup-tree-dump "gimple" } } */
>> Index: gcc-head/gcc/testsuite/gcc.target/i386/branch-cost4.c
>> ===================================================================
>> --- /dev/null
>> +++ gcc-head/gcc/testsuite/gcc.target/i386/branch-cost4.c
>> @@ -0,0 +1,16 @@
>> +/* { dg-do compile } */
>> +/* { dg-options "-O2 -fdump-tree-gimple -mbranch-cost=0" } */
>> +
>> +extern int doo (void);
>> +
>> +int
>> +foo (_Bool a, _Bool b)
>> +{
>> +  if (a && b)
>> +   return doo ();
>> +  return 0;
>> +}
>> +
>> +/* { dg-final { scan-tree-dump-times "if " 2 "gimple" } } */
>> +/* { dg-final { scan-tree-dump-not " & " "gimple" } } */
>> +/* { dg-final { cleanup-tree-dump "gimple" } } */
>> Index: gcc-head/gcc/fold-const.c
>> ===================================================================
>> --- gcc-head.orig/gcc/fold-const.c
>> +++ gcc-head/gcc/fold-const.c
>> @@ -3706,6 +3706,19 @@ simple_operand_p_2 (tree exp)
>>   /* Strip any conversions that don't change the machine mode.  */
>>   STRIP_NOPS (exp);
>>
>> +  /* Handle integral widening casts from boolean-typed
>> +     expressions as simple.  This happens due casts from
>> +     boolean-types are preserved, but FE might expands
>> +     simple-expression to wider mode.  */
>> +  if (INTEGRAL_TYPE_P (TREE_TYPE (exp))
>> +      && CONVERT_EXPR_P (exp)
>> +      && TREE_CODE (TREE_TYPE (TREE_OPERAND (exp, 0)))
>> +        == BOOLEAN_TYPE)
>> +    {
>> +      exp = TREE_OPERAND (exp, 0);
>> +      STRIP_NOPS (exp);
>> +    }
>> +
>
> Huh, well.  I think the above is just too special and you instead should
> replace the existing STRIP_NOPS by
>
> while (CONVERT_EXPR_P (exp))
>  exp = TREE_OPERAND (exp, 0);
>
> with a comment that conversions are considered simple.
>
> Ok with that change, if it bootstraps & tests ok.
>
> Richard.

Ok, bootstrapped and regression-tested on x86_64-unknown-linux-gnu and
applied to trunk with modifying as you suggested.

One question I have about handling of TRUTH-binaries in general in
fold-const.c.  Why aren't we enforcing already here in fold_binary for
those operations, that operands get boolean-type?  I see here some
advantages of C-AST folding.  I've tested it and saw that even later
in SSA-passes we get slightly better results on that.

Regards,
Kai