This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: Fix PR middle-end/57370
- From: Richard Biener <richard dot guenther at gmail dot com>
- To: Easwaran Raman <eraman at google dot com>
- Cc: GCC Patches <gcc-patches at gcc dot gnu dot org>
- Date: Mon, 9 Sep 2013 10:54:04 +0200
- Subject: Re: Fix PR middle-end/57370
- Authentication-results: sourceware.org; auth=none
- References: <CAPK5YPbNvp9g-9O=DUb06S3yNEvXiBrYeTbbsW2cDJuwd2_UXQ at mail dot gmail dot com> <CAPK5YPaXwtv6O41hgk4gdfdU0H2d=H3Ac5YUJsmYDQWLKc7VGg at mail dot gmail dot com> <CAPK5YPYXHtxs=+LpJNeZuKU+hmx80zjG1qQgeetxLt6qtGBSJA at mail dot gmail dot com> <CAFiYyc1_zfQzzJZQpTh+di=Q1RVrEMXnNtksTRb9UaGeDZCjvQ at mail dot gmail dot com> <CAPK5YPZrDV-+kp76z0Z_+pkQ8H0-zXnyy-jNvn89z3SrqXktMg at mail dot gmail dot com> <CAFiYyc1uHZW786ek0iUBmYDYYsZPBTe_4Zqd2G1YX6vHSz88Gw at mail dot gmail dot com> <CAPK5YPZQxud3DXX09F-VrexQQd1QLahrzguCX8FXW6y2bMMopA at mail dot gmail dot com>
On Fri, Sep 6, 2013 at 8:47 PM, Easwaran Raman <eraman@google.com> wrote:
> On Fri, Sep 6, 2013 at 12:05 AM, Richard Biener
> <richard.guenther@gmail.com> wrote:
>> On Thu, Sep 5, 2013 at 9:23 PM, Easwaran Raman <eraman@google.com> wrote:
>>> On Tue, Aug 27, 2013 at 3:35 AM, Richard Biener
>>> <richard.guenther@gmail.com> wrote:
>>>> On Thu, Aug 1, 2013 at 1:34 AM, Easwaran Raman <eraman@google.com> wrote:
>>>>> I have a new patch that supersedes this. The new patch also fixes PR
>>>>> tree-optimization/57393 and PR tree-optimization/58011. Bootstraps and
>>>>> no test regression on x86_64/linux. Ok for trunk?
>>>>>
>>>>> 2013-07-31 Easwaran Raman <eraman@google.com>
>>>>>
>>>>> PR middle-end/57370
>>>>> * tree-ssa-reassoc.c (build_and_add_sum): Fix UID assignment and reset
>>>>> of debug statements that cause inconsistent IR.
>>>>
>>>> Missing ChangeLog entry for the insert_stmt_after hunk which I do not like
>>>> at all. The other hunks are ok, but we need to work harder to preserve
>>>> debug stmts - simply removing all is not going to fly.
>>>>
>>>> Richard.
>>>
>>> I looked into the problem related to the debug stmts in this failing
>>> test case in detail. Initially, the code sequence looks
>>>
>>>
>>> s$n_13 = MEM[(struct S *)&s];
>>> # DEBUG s$n => s$n_13
>>> _2 = (double) s$n_13;
>>> _4 = _2 * a_3(D);
>>> _6 = (double) i_5(D);
>>> _7 = _4 * _6;
>>> _9 = (double) j_8(D);
>>> _10 = _7 * _9;
>>> bar (_10);
>>> # DEBUG D#2 => (double) k_12(D)
>>> # DEBUG D#1 => _7 * D#2
>>> # DEBUG t$n => (int) D#1
>>>
>>> After reassociation
>>>
>>> s$n_13 = MEM[(struct S *)&s];
>>> # DEBUG s$n => s$n_13
>>> _2 = (double) s$n_13;
>>> _6 = (double) i_5(D);
>>> # DEBUG D#3 => _4 * _6
>>> _9 = (double) j_8(D);
>>> _4 = _9 * _2;
>>> _7 = _4 * a_3(D);
>>> _10 = _7 * _6;
>>> bar (_10);
>>> # DEBUG D#2 => (double) k_12(D)
>>> # DEBUG D#1 => D#3 * D#2
>>> # DEBUG t$n => (int) D#1
>>>
>>> In the above, # DEBUG D#3 => _4 * _6 appears above the statement that
>>> defines _4. But even if I move the def of D#3 below the statement
>>> defining _4, it is not sufficient. Before reassociation, t$n refers to
>>> the expression (s$n_13 * a_3(D) * i_5(D) * k_12(D)), but after
>>> reassociation it would refer to ( j_8(D) * s$n_13 * i_5(D) *
>>> k_12(D)). It seems the correct fix is to discard the debug temps whose
>>> RHS contains a variable that is involved in reassociation and then
>>> reconstruct it. Is that the expected fix here?
>>
>> The value of the DEBUG expression changes because the value
>> that _4 computes changes - that is a no-no that may not happen.
>> We cannot re-use _4 (without previously releasing it and allocating
>> it newly) with a new value. releasing _4 should have fixed up the
>> debug stmts.
>>
>> So - can you verify whether we are indeed just replacing the
>> RHS of _4 = _2 * a_3(D) to _9 * _2 without changing the SSA name of
>> that expression?
>
> I have confirmed that the SSA name of the LHS remains the same. The
> reassociation code simply calls gimple_assign_set_rhs2 to modify RHS2
> and then calls update_stmt.
That's definitely not allowed (though ISTR I remember that reassoc does
this kind of things :/) - it will result in wrong debug info.
Richard.
> - Easwaran
>
>> Another reason why re-using the same LHS for another value is
>> wrong is that annotated information like points-to sets, alignment
>> and soon value-range information will be wrong.
>>
>> Thanks,
>> Richard.
>>
>>> - Easwaran
>>>
>>>>>
>>>>> testsuite/ChangeLog:
>>>>> 2013-07-31 Easwaran Raman <eraman@google.com>
>>>>>
>>>>> PR middle-end/57370
>>>>> PR tree-optimization/57393
>>>>> PR tree-optimization/58011
>>>>> * gfortran.dg/reassoc_12.f90: New testcase.
>>>>> * gcc.dg/tree-ssa/reassoc-31.c: New testcase.
>>>>> * gcc.dg/tree-ssa/reassoc-31.c: New testcase.
>>>>>
>>>>>
>>>>> On Fri, Jul 12, 2013 at 7:57 AM, Easwaran Raman <eraman@google.com> wrote:
>>>>>> Ping.
>>>>>>
>>>>>> On Thu, Jun 27, 2013 at 10:15 AM, Easwaran Raman <eraman@google.com> wrote:
>>>>>>> A newly generated statement in build_and_add_sum function of
>>>>>>> tree-ssa-reassoc.c has to be assigned the UID of its adjacent
>>>>>>> statement. In one instance, it was assigned the wrong uid (of an
>>>>>>> earlier phi statement) which messed up the IR and caused the test
>>>>>>> program to hang. Bootstraps and no test regressions on x86_64/linux.
>>>>>>> Ok for trunk?
>>>>>>>
>>>>>>> Thanks,
>>>>>>> Easwaran
>>>>>>>
>>>>>>>
>>>>>>> 2013-06-27 Easwaran Raman <eraman@google.com>
>>>>>>>
>>>>>>> PR middle-end/57370
>>>>>>> * tree-ssa-reassoc.c (build_and_add_sum): Do not use the UID of a phi
>>>>>>> node for a non-phi gimple statement.
>>>>>>>
>>>>>>> testsuite/ChangeLog:
>>>>>>> 2013-06-27 Easwaran Raman <eraman@google.com>
>>>>>>>
>>>>>>> PR middle-end/57370
>>>>>>> * gfortran.dg/reassoc_12.f90: New testcase.
>>>>>>>
>>>>>>>
>>>>>>> Index: gcc/testsuite/gfortran.dg/reassoc_12.f90
>>>>>>> ===================================================================
>>>>>>> --- gcc/testsuite/gfortran.dg/reassoc_12.f90 (revision 0)
>>>>>>> +++ gcc/testsuite/gfortran.dg/reassoc_12.f90 (revision 0)
>>>>>>> @@ -0,0 +1,74 @@
>>>>>>> +! { dg-do compile }
>>>>>>> +! { dg-options "-O2 -ffast-math" }
>>>>>>> +! PR middle-end/57370
>>>>>>> +
>>>>>>> + SUBROUTINE xb88_lr_adiabatic_lda_calc(e_ndrho_ndrho_ndrho, &
>>>>>>> + grad_deriv,npoints, sx)
>>>>>>> + IMPLICIT REAL*8 (t)
>>>>>>> + INTEGER, PARAMETER :: dp=8
>>>>>>> + REAL(kind=dp), DIMENSION(1:npoints) :: e_ndrho_ndrho_ndrho, &
>>>>>>> + e_ndrho_ndrho_rho
>>>>>>> + DO ii=1,npoints
>>>>>>> + IF( grad_deriv >= 2 .OR. grad_deriv == -2 ) THEN
>>>>>>> + t1425 = t233 * t557
>>>>>>> + t1429 = beta * t225
>>>>>>> + t1622 = t327 * t1621
>>>>>>> + t1626 = t327 * t1625
>>>>>>> + t1632 = t327 * t1631
>>>>>>> + t1685 = t105 * t1684
>>>>>>> + t2057 = t1636 + t8 * (t2635 + t3288)
>>>>>>> + END IF
>>>>>>> + IF( grad_deriv >= 3 .OR. grad_deriv == -3 ) THEN
>>>>>>> + t5469 = t5440 - t5443 - t5446 - t5449 - &
>>>>>>> + t5451 - t5454 - t5456 + t5459 - &
>>>>>>> + t5462 + t5466 - t5468
>>>>>>> + t5478 = 0.240e2_dp * t1616 * t973 * t645 * t1425
>>>>>>> + t5489 = 0.1600000000e2_dp * t1429 * t1658
>>>>>>> + t5531 = 0.160e2_dp * t112 * t1626
>>>>>>> + t5533 = 0.160e2_dp * t112 * t1632
>>>>>>> + t5537 = 0.160e2_dp * t112 * t1622
>>>>>>> + t5541 = t5472 - t5478 - t5523 + t5525 + &
>>>>>>> + t5531 + t5533 + t5535 + t5537 + &
>>>>>>> + t5540
>>>>>>> + t5565 = t112 * t1685
>>>>>>> + t5575 = t5545 - t5548 + t5551 + t5553 - &
>>>>>>> + t5558 + t5560 - t5562 + t5564 - &
>>>>>>> + 0.80e1_dp * t5565 + t5568 + t5572 + &
>>>>>>> + t5574
>>>>>>> + t5611 = t5579 - t5585 + t5590 - t5595 + &
>>>>>>> + t5597 - t5602 + t5604 + t5607 + &
>>>>>>> + t5610
>>>>>>> + t5613 = t5469 + t5541 + t5575 + t5611
>>>>>>> + t6223 = t6189 - &
>>>>>>> + 0.3333333336e0_dp * t83 * t84 * t5613 + &
>>>>>>> + t6222
>>>>>>> + t6227 = - t8 * (t5305 + t6223)
>>>>>>> + e_ndrho_ndrho_rho(ii) = e_ndrho_ndrho_rho(ii) + &
>>>>>>> + t6227 * sx
>>>>>>> + t6352 = t5440 - t5443 - t5446 - t5449 - &
>>>>>>> + t5451 - t5454 + &
>>>>>>> + 0.40e1_dp * t102 * t327 * t2057 * t557 - &
>>>>>>> + t5456 + t5459 - t5462 + t5466 - &
>>>>>>> + t5468
>>>>>>> + t6363 = t5480 - t5489 + &
>>>>>>> + 0.9600000000e2_dp * t1054 * t640 * t3679
>>>>>>> + t6367 = t5472 - t5474 - t5478 - t5523 + &
>>>>>>> + t5525 + t5531 + t5533 + t5535 + &
>>>>>>> + t5537 - 0.20e1_dp * t102 * t105 * t6363 + &
>>>>>>> + t5540
>>>>>>> + t6370 = t5545 - t5548 + t5551 + t5553 - &
>>>>>>> + t5558 + t5560 - t5562 + t5564 - &
>>>>>>> + 0.40e1_dp * t5565 + &
>>>>>>> + t5568 + t5572 + t5574
>>>>>>> + t6373 = t5579 - t5585 + t5590 - t5595 + &
>>>>>>> + t5597 - t5602 + t5604 + t5607 + &
>>>>>>> + t5610
>>>>>>> + t6375 = t6352 + t6367 + t6370 + t6373
>>>>>>> + t6380 = - 0.3333333336e0_dp * t83 * t84 * t6375 + t5701
>>>>>>> + t6669 = -t4704 - t8 * (t6344 + t6380 + t6665)
>>>>>>> + e_ndrho_ndrho_ndrho(ii) = e_ndrho_ndrho_ndrho(ii) + &
>>>>>>> + t6669 * sx
>>>>>>> + END IF
>>>>>>> + END DO
>>>>>>> + END SUBROUTINE xb88_lr_adiabatic_lda_calc
>>>>>>> +
>>>>>>> Index: gcc/tree-ssa-reassoc.c
>>>>>>> ===================================================================
>>>>>>> --- gcc/tree-ssa-reassoc.c (revision 200429)
>>>>>>> +++ gcc/tree-ssa-reassoc.c (working copy)
>>>>>>> @@ -1207,7 +1207,7 @@ build_and_add_sum (tree type, tree op1, tree op2,
>>>>>>> if (gimple_code (op1def) == GIMPLE_PHI)
>>>>>>> {
>>>>>>> gsi = gsi_after_labels (gimple_bb (op1def));
>>>>>>> - gimple_set_uid (sum, gimple_uid (op1def));
>>>>>>> + gimple_set_uid (sum, gimple_uid (gsi_stmt (gsi)));
>>>>>>> gsi_insert_before (&gsi, sum, GSI_NEW_STMT);
>>>>>>> }
>>>>>>> else