[Bug tree-optimization/81184] [8 regression] gcc.dg/pr21643.c and gcc.dg/tree-ssa/phi-opt-11.c fail starting with r249450
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81184 --- Comment #11 from seurer at gcc dot gnu.org --- I dug through my logs and the last failures I saw for phi-opt-11.c and pr21643.c on powerpc64 were in mid January immediately before Eric's patch.
[Bug tree-optimization/81184] [8 regression] gcc.dg/pr21643.c and gcc.dg/tree-ssa/phi-opt-11.c fail starting with r249450
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81184 --- Comment #10 from Andreas Krebbel --- I've verified that the problem is fixed on Power. So I've opened a separate BZ for this #85368
[Bug tree-optimization/81184] [8 regression] gcc.dg/pr21643.c and gcc.dg/tree-ssa/phi-opt-11.c fail starting with r249450
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81184
Andreas Krebbel changed:
What|Removed |Added
CC||krebbel at gcc dot gnu.org
--- Comment #9 from Andreas Krebbel ---
Is this really fixed for Power64. At least on s390 I still these testcases
failing.
I had a look into phi-opt-11.c.
No IF statements remain although LOGICAL_OP_NON_SHORT_CIRCUIT is not
defined on S/390 and hence defaults to true when using
-mbranch-cost=2.
The testcase appears to expect 2 IFs to remain for function
h. However, these get removed phiopt1.
This code turns the TRUTH_ANDIF_EXPR condition into a TRUTH_AND_EXPR:
fold-const.c:8178
if (LOGICAL_OP_NON_SHORT_CIRCUIT
&& !flag_sanitize_coverage
&& (code == TRUTH_AND_EXPR
|| code == TRUTH_ANDIF_EXPR
|| code == TRUTH_OR_EXPR
|| code == TRUTH_ORIF_EXPR))
{
enum tree_code ncode, icode;
ncode = (code == TRUTH_ANDIF_EXPR || code == TRUTH_AND_EXPR)
? TRUTH_AND_EXPR : TRUTH_OR_EXPR;
icode = ncode == TRUTH_AND_EXPR ? TRUTH_ANDIF_EXPR : TRUTH_ORIF_EXPR;
...
/* Transform (A AND-IF B) into (A AND B), or (A OR-IF B)
into (A OR B).
For sequence point consistancy, we need to check for trapping,
and side-effects. */
else if (code == icode && simple_operand_p_2 (arg0)
&& simple_operand_p_2 (arg1))
return fold_build2_loc (loc, ncode, type, arg0, arg1);
This prevents the gimplifier from splitting the condition into two
separate IF statements. Instead the truth value gets computed:
004t.gimple
h (int a, int b, int c, int d)
{
int D.2246;
_1 = a == d;
_2 = b > c;
_3 = _1 & _2;
if (_3 != 0) goto ; else goto ;
:
D.2246 = d;
// predicted unlikely by early return (on trees) predictor.
return D.2246;
:
D.2246 = a;
return D.2246;
}
which eventually gets optimized in phiop1 to:
Removing basic block 3
;; basic block 3, loop depth 0
;; pred: 2
;; succ: 4
COND_EXPR in block 2 and PHI in block 4 converted to straightline code.
Merging blocks 2 and 4
fix_loop_structure: fixing up loops for function
h (int a, int b, int c, int d)
{
_Bool _1;
_Bool _2;
_Bool _3;
[local count: 1073741825]:
_1 = a_5(D) == d_6(D);
_2 = b_7(D) > c_8(D);
_3 = _1 & _2;
return a_5(D);
}
No IFs.
[Bug tree-optimization/81184] [8 regression] gcc.dg/pr21643.c and gcc.dg/tree-ssa/phi-opt-11.c fail starting with r249450
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81184 Eric Botcazou changed: What|Removed |Added Status|NEW |RESOLVED CC||ebotcazou at gcc dot gnu.org Resolution|--- |FIXED --- Comment #8 from Eric Botcazou --- .
[Bug tree-optimization/81184] [8 regression] gcc.dg/pr21643.c and gcc.dg/tree-ssa/phi-opt-11.c fail starting with r249450
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81184 --- Comment #7 from Eric Botcazou --- Author: ebotcazou Date: Wed Jan 17 11:03:00 2018 New Revision: 256780 URL: https://gcc.gnu.org/viewcvs?rev=256780=gcc=rev Log: PR tree-optimization/81184 * gcc.dg/pr21643.c: Adjust dg-final line for logical_op_short_circuit targets. * gcc.dg/tree-ssa/phi-opt-11.c: Likewise. Modified: trunk/gcc/testsuite/ChangeLog trunk/gcc/testsuite/gcc.dg/pr21643.c trunk/gcc/testsuite/gcc.dg/tree-ssa/phi-opt-11.c
[Bug tree-optimization/81184] [8 regression] gcc.dg/pr21643.c and gcc.dg/tree-ssa/phi-opt-11.c fail starting with r249450
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81184 --- Comment #6 from Christophe Lyon --- Note that I posted a related patch some time ago: https://gcc.gnu.org/ml/gcc-patches/2017-10/msg01477.html
[Bug tree-optimization/81184] [8 regression] gcc.dg/pr21643.c and gcc.dg/tree-ssa/phi-opt-11.c fail starting with r249450
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81184 Richard Biener changed: What|Removed |Added Priority|P3 |P1
[Bug tree-optimization/81184] [8 regression] gcc.dg/pr21643.c and gcc.dg/tree-ssa/phi-opt-11.c fail starting with r249450
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81184 --- Comment #5 from Jakub Jelinek --- Most likely the test need better guards on -mbranch-cost= and/or logical_op_short_circuit.
[Bug tree-optimization/81184] [8 regression] gcc.dg/pr21643.c and gcc.dg/tree-ssa/phi-opt-11.c fail starting with r249450
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81184 --- Comment #4 from Martin Liška --- Jakub?
[Bug tree-optimization/81184] [8 regression] gcc.dg/pr21643.c and gcc.dg/tree-ssa/phi-opt-11.c fail starting with r249450
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81184 Christophe Lyon changed: What|Removed |Added Target|powerpc64*-*-* |powerpc64*-*-* arm CC||clyon at gcc dot gnu.org --- Comment #3 from Christophe Lyon --- On arm the regression appeared between 249444 and 249453.
[Bug tree-optimization/81184] [8 regression] gcc.dg/pr21643.c and gcc.dg/tree-ssa/phi-opt-11.c fail starting with r249450
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81184 Thomas Preud'homme changed: What|Removed |Added Last reconfirmed|2017-06-23 00:00:00 |2017-7-13 CC||thopre01 at gcc dot gnu.org --- Comment #2 from Thomas Preud'homme --- These testcases (gcc.dg/pr21643.c and gcc.dg/tree-ssa/phi-opt-11.c) also started to fail for ARM Cortex-M0, Cortex-M3 and Cortex-M4 arm-none-eabi targets.
[Bug tree-optimization/81184] [8 regression] gcc.dg/pr21643.c and gcc.dg/tree-ssa/phi-opt-11.c fail starting with r249450
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81184 Richard Biener changed: What|Removed |Added Target Milestone|--- |8.0
[Bug tree-optimization/81184] [8 regression] gcc.dg/pr21643.c and gcc.dg/tree-ssa/phi-opt-11.c fail starting with r249450
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81184
Martin Liška changed:
What|Removed |Added
Status|UNCONFIRMED |NEW
Last reconfirmed||2017-06-23
CC||jakub at gcc dot gnu.org
Assignee|unassigned at gcc dot gnu.org |marxin at gcc dot
gnu.org
Ever confirmed|0 |1
--- Comment #1 from Martin Liška ---
Confirmed, I'm adding Jakub as author of original test-case.
The affected function changed from:
Optimizing range tests c_2(D) -[0, 31] and -[32, 32]
into c_2(D) > 32
Removing basic block 3
;; basic block 3, loop depth 0
;; pred:
;; succ: 4
fix_loop_structure: fixing up loops for function
f6 (unsigned char c)
{
int _1;
_Bool _5;
[100.00%] [count: INV]:
if (c_2(D) == 34)
goto ; [18.79%] [count: INV]
else
goto ; [81.21%] [count: INV]
[65.95%] [count: INV]:
_5 = c_2(D) <= 32;
if (_5 != 0)
goto ; [31.00%] [count: INV]
else
goto ; [69.00%] [count: INV]
[45.51%] [count: INV]:
[100.00%] [count: INV]:
# _1 = PHI <2(2), 0(4), 2(3)>
return _1;
}
to:
Optimizing range tests c_2(D) -[32, 32] and -[34, 34]
into (c_2(D) & 253) != 32
Width = 1 was chosen for reassociation
Merging blocks 2 and 3
fix_loop_structure: fixing up loops for function
f6 (unsigned char c)
{
int _1;
_Bool _4;
unsigned char _5;
[100.00%] [count: INV]:
_5 = c_2(D) & 253;
_4 = _5 == 32;
if (_4 != 0)
goto ; [10.65%] [count: INV]
else
goto ; [89.35%] [count: INV]
[79.83%] [count: INV]:
if (c_2(D) <= 31)
goto ; [18.79%] [count: INV]
else
goto ; [81.21%] [count: INV]
[15.00%] [count: INV]:
[100.00%] [count: INV]:
# _1 = PHI <0(3), 2(2), 2(4)>
return _1;
}
On x86_64 it's different because the function looks as follows:
f6 (unsigned char c)
{
int _1;
_Bool _4;
_Bool _5;
_Bool _7;
_Bool _8;
_Bool _9;
[100.00%] [count: INV]:
_5 = c_2(D) == 32;
_4 = c_2(D) == 34;
_9 = c_2(D) <= 32;
_7 = c_2(D) <= 31;
_8 = _9 | _4;
if (_8 != 0)
goto ; [15.00%] [count: INV]
else
goto ; [85.00%] [count: INV]
[15.00%] [count: INV]:
[100.00%] [count: INV]:
# _1 = PHI <0(2), 2(3)>
return _1;
}
Note that the PR exposes problem where we are able to do predictions for
conditions, but
here we transform:
if (c_2(D) == 34)
goto ; [10.65%] [count: INV]
else
goto ; [89.35%] [count: INV]
into:
_4 = c_2(D) == 34;
and then we do
_8 = _9 | _4;
if (_8 != 0)
It's related to PR79489.
