This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [PATCH][middle-end/88784] Middle end is missing some optimizations about unsigned




On 2019/7/2 4:51 PM, Richard Biener wrote:
On Tue, 2 Jul 2019, Richard Biener wrote:

On Tue, 2 Jul 2019, Li Jia He wrote:



On 2019/7/1 3:30 PM, Richard Biener wrote:
On Fri, 28 Jun 2019, Andrew Pinski wrote:

On Thu, Jun 27, 2019 at 9:55 PM Li Jia He <helijia@linux.ibm.com> wrote:



On 2019/6/27 11:48 PM, Jeff Law wrote:
On 6/27/19 12:11 AM, Li Jia He wrote:
Hi,

According to the optimizable case described by Qi Feng on
issue 88784, we can combine the cases into the following:

1. x >  y  &&  x != XXX_MIN  -->   x > y
2. x >  y  &&  x == XXX_MIN  -->   false
3. x <= y  &&  x == XXX_MIN  -->   x == XXX_MIN

4. x <  y  &&  x != XXX_MAX  -->   x < y
5. x <  y  &&  x == XXX_MAX  -->   false
6. x >= y  &&  x == XXX_MAX  -->   x == XXX_MAX

7. x >  y  ||  x != XXX_MIN  -->   x != XXX_MIN
8. x <= y  ||  x != XXX_MIN  -->   true
9. x <= y  ||  x == XXX_MIN  -->   x <= y

10. x <  y  ||  x != XXX_MAX  -->   x != UXXX_MAX
11. x >= y  ||  x != XXX_MAX  -->   true
12. x >= y  ||  x == XXX_MAX  -->   x >= y

Note: XXX_MIN represents the minimum value of type x.
         XXX_MAX represents the maximum value of type x.

Here we don't need to care about whether the operation is
signed or unsigned.  For example, in the below equation:

'x >  y  &&  x != XXX_MIN  -->   x > y'

If the x type is signed int and XXX_MIN is INT_MIN, we can
optimize it to 'x > y'.  However, if the type of x is unsigned
int and XXX_MIN is 0, we can still optimize it to 'x > y'.

The regression testing for the patch was done on GCC mainline on

       powerpc64le-unknown-linux-gnu (Power 9 LE)

with no regressions.  Is it OK for trunk ?

Thanks,
Lijia He

gcc/ChangeLog

2019-06-27  Li Jia He  <helijia@linux.ibm.com>
           Qi Feng  <ffengqi@linux.ibm.com>

       PR middle-end/88784
       * gimple-fold.c (and_comparisons_contain_equal_operands): New
function.
       (and_comparisons_1): Use
and_comparisons_contain_equal_operands.
       (or_comparisons_contain_equal_operands): New function.
       (or_comparisons_1): Use or_comparisons_contain_equal_operands.
Would this be better done via match.pd?  ISTM this transformation
would
be well suited for that framework.

Hi, Jeff

I did this because of the following test case:
`
_Bool comp(unsigned x, unsigned y)
{
     return x > y && x != 0;
}
`
The gimple file dumped on the power platform is:
`
comp (unsigned int x, unsigned int y)
{
     _Bool D.2837;
     int iftmp.0;

     if (x > y) goto <D.2841>; else goto <D.2839>;
     <D.2841>:
     if (x != 0) goto <D.2842>; else goto <D.2839>;
     <D.2842>:
     iftmp.0 = 1;
     goto <D.2840>;
     <D.2839>:
     iftmp.0 = 0;
     <D.2840>:
     D.2837 = (_Bool) iftmp.0;
     return D.2837;
}
`
However, the gimple file dumped on x86 is
`
comp (unsigned int x, unsigned int y)
{
     _Bool D.2837;

     _1 = x > y;
     _2 = x != 0;
     _3 = _1 & _2;
     _4 = (int) _3;
     D.2837 = (_Bool) _4;
     return D.2837;
}
`

The reason for the inconsistency between these two behaviors is param
logical-op-non-short-circuit.  If we add the pattern to the match.pd
file, we can only optimize the situation in which the statement is in
the same basic block (logical-op-non-short-circuit=1, x86).  But for
a cross-basic block (logical-op-non-short-circuit=0, power), match.pd
can't handle this situation.

Another reason is that I found out maybe_fold_and_comparisons and
maybe_fold_or_comparisons are not only called by ifcombine pass but
also by reassoc pass. Using this method can basically unify param
logical-op-non-short-circuit=0 or 1.


As mentioned before ifcombine pass should be using gimple-match
instead of fold_build.  Try converting ifcombine over to gimple-match
infrastructure and add these to match.pd.

Yes, I mentioned that in the PR.  The issue is that at the moment
to combine x > y with x <= y you'd have to build GENERIC trees
for both or temporary GIMPLE assign with a SSA def (and then feed
that into the GENERIC or GIMPLE match.pd path).

Hi,

I did some experimentation using ‘temporary GIMPLE with a SSA def (and then
feed that into the GIMPLE match.pd path’.  Could we consider the code in the
attachment(I did a test and the code took effect)?

No, that's excessive overhead IMHO - the whole point of these functions
originally was to avoid build2 (...) on both conditionals.  If we
now create two SSA names and two GIMPLE statements that's too much.

As said it shouldn't be rocket science to teach genmatch to
auto-generate those functions from match.pd patterns but it certainly
is some work (less so for somebody with genmatch knowledge).
I'll give it a poke...

OK, it's not so easy.  I guess we could lower the cost of building
SSA names / gimple stmts significantly if we allocated them on the
stack like via

gimple *stmt1 = XALLOCAVEC (char, gimple_size (GIMPLE_ASSIGN) + 2 *
sizeof (tree));
memset (stmt1, 0, ...);
gimple_set_code (stmt1, GIMPLE_ASSIGN);
gimple_set_num_ops (stmt1, 3);
gimple_init_singleton (stmt1);

gimple_stmt_iterator gsi = gsi_for_stmt (stmt1);
gimple_assign_set_rhs_with_ops (&gsi, actual operation...);

tree lhs1 = XALLOCA (tree_ssa_name);
memset (lhs1, 0, sizeof (tree_ssa_name));
TREE_SET_CODE (lhs1, SSA_NAME);
TREE_TYPE (lhs1) = ...;

gimple_assing_set_lhs (stmt1, lhs1);

it's all a bit ugly and some factoring in the generic gimple machinery
might easen this kind of hacks, but well...

With the above you could then use

   gimple_match_op op (gimple_match_cond::UNCOND, BIT_AND_EXPR /* or OR */,
		      boolean_type_node, lhs1, lhs2);
   if (gimple_resimplify2 (NULL, &op, follow_all_ssa_edges))
     ... successfully simplified into 'op' ...

with the simplification path then extracting the simplified result
from 'op'.  Note this deliberately leaves out passing a gimple_seq
as storage for more complex simplification results to match
existing behavior.

Your patch has the GC allocation and SSA name (and namespace!)
allocation overhead and most definitely misses releasing the
SSA names you allocated.

Note with the above hack the simplified result has to be checked
for mentions of lhs1 or lhs2 - a case we have to reject because
their definitions are transitional.


Hi,

  I made some changes based on the recommendations. Would you like to
  help me to see it again ? Sorry, it took so long time to provide the
  patch.

  Note: 1. I keep the code for and_comparisons_1 and or_comparisons_1.
	The reason is that I did not found a good way to handle the
	optimization of '((x CODE1 y) AND (x CODE2 y))' in match.pd.
	Maybe I missing some important information about match.pd.
	2. The gimple_resimplify2 function is not used.  Since stmt1,
	stmt2, lhs1 and lhs2 are allocated on the stack, Is there a
	question with the value on the stack as the return value ?
	I may have misunderstood Richard's intention.

Thanks,
Lijia He

Richard.

Richard.

Thanks,
Lijia He


maybe_fold_and/or_comparisons handle two exploded binary expressions
while the current match.pd entries handle at most one exploded one
(the outermost then, either AND or OR).  But it would be definitely
doable to auto-generate maybe_fold_and/or_comparisons from match.pd
patterns which is what I'd ultimatively suggest to do (in some more
generalized form maybe).  Either with a separate genmatch invocation
or as part of the --gimple processing (not sure what is more feasible
here).

I told Li Jia He that I don't expect him to do this work.

Note I didn't review the actual patch yet.

Thanks,
Richard.





Attachment: 0001-Auto-generate-maybe_fold_and-or_comparisons-from-mat.patch
Description: Text document

Attachment: 0002-Fix-PR88784-middle-end-is-missing-some-optimizations.patch
Description: Text document


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]