[Bug middle-end/55555] [4.8 Regression] miscompilation at -O2 (number_of_iterations)
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=5 Bug 5 depends on bug 57511, which changed state. Bug 57511 Summary: [4.8 Regression] Missing SCEV final value replacement https://gcc.gnu.org/bugzilla/show_bug.cgi?id=57511 What|Removed |Added Status|ASSIGNED|RESOLVED Resolution|--- |FIXED
[Bug middle-end/55555] [4.8 Regression] miscompilation at -O2 (number_of_iterations)
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=5 Richard Biener changed: What|Removed |Added Status|ASSIGNED|RESOLVED Resolution||FIXED --- Comment #13 from Richard Biener 2012-12-18 13:29:52 UTC --- Fixed.
[Bug middle-end/55555] [4.8 Regression] miscompilation at -O2 (number_of_iterations)
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=5 --- Comment #12 from Richard Biener 2012-12-18 13:12:39 UTC --- Author: rguenth Date: Tue Dec 18 13:12:34 2012 New Revision: 194578 URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=194578 Log: 2012-12-18 Richard Biener PR tree-optimization/5 * tree-ssa-loop-niter.c (idx_infer_loop_bounds): Properly analyze evolution of the index for the loop it is used in. * tree-scalar-evolution.c (instantiate_scev_name): Take inner loop we will be creating a chrec for. Generalize fix for PR40281 and prune invalid SCEVs. (instantiate_scev_poly): Likewise - pass down inner loop we will be creating a chrec for. (instantiate_scev_binary): Take and pass through inner loop. (instantiate_array_ref): Likewise. (instantiate_scev_convert): Likewise. (instantiate_scev_not): Likewise. (instantiate_scev_3): Likewise. (instantiate_scev_2): Likewise. (instantiate_scev_1): Likewise. (instantiate_scev_r): Likewise. (resolve_mixers): Adjust. (instantiate_scev): Likewise. * gcc.dg/torture/pr5.c: New testcase. * gcc.dg/vect/vect-iv-11.c: Adjust. Added: trunk/gcc/testsuite/gcc.dg/torture/pr5.c Modified: trunk/gcc/ChangeLog trunk/gcc/testsuite/ChangeLog trunk/gcc/testsuite/gcc.dg/vect/vect-iv-11.c trunk/gcc/tree-scalar-evolution.c trunk/gcc/tree-ssa-loop-niter.c
[Bug middle-end/55555] [4.8 Regression] miscompilation at -O2 (number_of_iterations)
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=5
--- Comment #11 from Richard Biener 2012-12-17
12:14:18 UTC ---
Ok, I'm confused by the following:
: (loop 1 header)
# lxp_1 = PHI <0(2), lxp_24(12)>
t_9 = pol_x[lxp_1];
_10 = (long int) lxp_1;
_11 = _10 * 4;
l_12 = _11 + -1;
goto ;
:
_15 = S_2 + l_12;
_16 = coef_x[_15];
_17 = S_2 + -1;
_18 = s[_17];
_19 = _18 * t_9;
_20 = _16 + _19;
coef_x[_15] = _20;
S_22 = S_2 + 1;
: (loop 2 header)
# S_2 = PHI <1(3), S_22(4)>
if (S_2 <= 4)
goto ;
else
goto ;
:
lxp_24 = lxp_1 + 1;
if (lxp_1 != 1)
goto ;
else
goto ;
:
goto ;
and SCEV says _15 is {l_12 + 1, +, 1}_2 which looks correct. But then it
does
(chrec_apply
(varying_loop = 2)
(chrec = {l_12 + 1, +, 1}_2)
(x = 4)
(res = l_12 + 5))
and magically, via l_12 being {-1, +, 4}_1 (also correct) arrives at
(instantiate_scev
(instantiate_below = 2)
(evolution_loop = 1)
(chrec = {4, +, 4}_1)
(res = {4, +, 4}_1))
huh? So to it _15 is {4, +, 4}_1 (not sure what is considered "initial"
in terms of scalar evolution with a value varying in an inner loop).
This is what infer_loop_bounds_from_undefined derives the bogus bound
for loop 1 from. To my eyes _15 should be {0, + 4}_1!
Maybe it doesn't really make sense to ask for the evolution of something
defined in loop N with respect to an outer loop M?
If we change idx_infer_loop_bounds with
Index: tree-ssa-loop-niter.c
===
--- tree-ssa-loop-niter.c (revision 194552)
+++ tree-ssa-loop-niter.c (working copy)
@@ -2671,7 +2671,12 @@ idx_infer_loop_bounds (tree base, tree *
upper = false;
}
- ev = instantiate_parameters (loop, analyze_scalar_evolution (loop, *idx));
+ struct loop *dloop = loop_containing_stmt (data->stmt);
+ if (!dloop)
+return true;
+
+ ev = analyze_scalar_evolution (dloop, *idx);
+ ev = instantiate_parameters (loop, ev);
init = initial_condition (ev);
step = evolution_part_in_loop_num (ev, loop->num);
then we obtain via {l_12 + 1, +, 1}_2, {{0, +, 4}_1, +, 1}_2 the correct
solution (init == 0, step == 4).
I am going to bootstrap and regtest that.
[Bug middle-end/55555] [4.8 Regression] miscompilation at -O2 (number_of_iterations)
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=5 Richard Biener changed: What|Removed |Added Status|NEW |ASSIGNED AssignedTo|unassigned at gcc dot |rguenth at gcc dot gnu.org |gnu.org | --- Comment #10 from Richard Biener 2012-12-17 10:14:37 UTC --- I'll have a more detailled look.
[Bug middle-end/55555] [4.8 Regression] miscompilation at -O2 (number_of_iterations)
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=5
--- Comment #9 from Richard Biener 2012-12-14
14:11:59 UTC ---
The unrolling puts __builtin_unreachable ()s into the inner duplicated loops:
First iteration, good:
:
# lxp_30 = PHI <0(2)>
t_32 = pol_x[lxp_30];
_33 = (long int) lxp_30;
_34 = _33 * 4;
l_35 = _34 + -1;
:
# S_36 = PHI <1(16), S_45(18)>
if (S_36 <= 4)
goto ;
else
goto ;
:
_38 = S_36 + l_35;
_39 = coef_x[_38];
_40 = S_36 + -1;
_41 = s[_40];
_42 = _41 * t_32;
_43 = _39 + _42;
coef_x[_38] = _43;
S_45 = S_36 + 1;
goto ;
:
lxp_46 = lxp_30 + 1;
:
Peeled iteration, bad:
:
# lxp_1 = PHI
t_9 = pol_x[lxp_1];
_10 = (long int) lxp_1;
_11 = _10 * 4;
l_12 = _11 + -1;
goto ;
:
_13 = S_3 + l_12;
__builtin_unreachable ();
_14 = coef_x[_13];
_15 = S_3 + -1;
_16 = s[_15];
_17 = _16 * t_9;
_18 = _14 + _17;
__builtin_unreachable ();
coef_x[_13] = _18;
S_20 = S_3 + 1;
:
# S_3 = PHI <1(3), S_20(4)>
if (S_3 <= 4)
goto ;
else
goto ;
:
lxp_21 = lxp_1 + 1;
if (1 == 0)
goto ;
else
goto ;
:
:
__builtin_unreachable ();
:
__asm__ __volatile__("" : : "r" &coef_x : "memory");
goto ;
Note that the outer loop body looks ok - it is the inner body that get's
mangled in a bogus way.
That's because the loop bounds derived from the inner loop accesses
are bogus:
remove_exits_and_undefined_stmts (loop=0x767ec770, npeeled=1)
at /space/rguenther/src/svn/trunk/gcc/tree-ssa-loop-ivcanon.c:481
481 bool changed = false;
(gdb) n
483 for (elt = loop->bounds; elt; elt = elt->next)
(gdb)
488 if (!elt->is_exit
(gdb) p elt
$8 = (nb_iter_bound *) 0x76925e38
(gdb) p *elt
$9 = {stmt = 0x76905c80, bound = {low = 0, high = 0}, is_exit = false,
next = 0x76925dc0}
(gdb) p elt->stmt
$10 = (gimple) 0x76905c80
(gdb) call debug_gimple_stmt ($10)
# .MEM_19 = VDEF <.MEM_6>
coef_x[_13] = _18;
as far as I can see even with lxp == 1 we have at most an index of 4 + 1*4 - 1.
Statement _14 = coef_x[_13];
is executed at most 0 (bounded by 0) + 1 times in loop 1.
_that's_ bogus.
[Bug middle-end/55555] [4.8 Regression] miscompilation at -O2 (number_of_iterations)
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=5
--- Comment #8 from Richard Biener 2012-12-14
13:49:02 UTC ---
(In reply to comment #7)
> I've tried to rewrite this as C, but managed to turn it into something that is
> miscompiled at a different spot. The fortran testcase starts having
> __builtin_unreachable () in it in the *.cunroll pass, this one already in
> *.cunrolli pass. Still, I believe it doesn't do any out of bounds access
> anywhere. -O2 on x86_64-linux.
>
> double s[4] = { 1.0, 2.0, 3.0, 4.0 }, pol_x[2] = { 5.0, 6.0 };
>
> __attribute__((noinline)) int
> foo (void)
> {
> double coef_x[8] = { 0, 0, 0, 0, 0, 0, 0, 0 };
> int lxp = 0;
> if (lxp <= 1)
> do
> {
> double t = pol_x[lxp];
> long S;
> long l = lxp * 4L - 1;
> for (S = 1; S <= 4; S++)
> coef_x[S + l] = coef_x[S + l] + s[S - 1] * t;
> }
> while (lxp++ != 1);
> asm volatile ("" : : "r" (coef_x) : "memory");
> for (lxp = 0; lxp < 8; lxp++)
> if (coef_x[lxp] != ((lxp & 3) + 1) * (5.0 + (lxp >= 4)))
> __builtin_abort ();
> return 1;
> }
>
> int
> main ()
> {
> asm volatile ("" : : : "memory");
> if (!foo ())
> __builtin_abort ();
> return 0;
> }
>
> Works with r193067, fails with r193100, haven't tried to bisect exactly, but
> would guess this is r193098 again. For the outer loop it prints:
>
> Analyzing # of iterations of loop 1
> exit condition [0, + , 1](no_overflow) != 1
> bounds on difference of bases: 1 ... 1
> result:
> # of iterations 1, bounded by 1
> Loop 1 iterates 1 times.
> Loop 1 iterates at most 1 times.
>
> but that is wrong, the outer loop iterates exactly 2 times.
That's latch block executions, so one latch block execution is correct.
[Bug middle-end/55555] [4.8 Regression] miscompilation at -O2 (number_of_iterations)
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=5 Jakub Jelinek changed: What|Removed |Added Priority|P3 |P1 Summary|[4.8 Regression]|[4.8 Regression] |miscompilation at -O2 |miscompilation at -O2 |(tree-pre?) |(number_of_iterations)
