https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83922

--- Comment #3 from rsandifo at gcc dot gnu.org <rsandifo at gcc dot gnu.org> 
---
Author: rsandifo
Date: Fri Jan 19 11:57:47 2018
New Revision: 256885

URL: https://gcc.gnu.org/viewcvs?rev=256885&root=gcc&view=rev
Log:
Check whether any statements need masking (PR 83922)

This PR is an odd case in which, due to the low optimisation level,
we enter vectorisation with:

  outer1:
    x_1 = PHI <x_3(outer2), ...>;
    ...

  inner:
    x_2 = 0;
    ...

  outer2:
    x_3 = PHI <x_2(inner)>;

These statements are tentatively treated as a double reduction by
vect_force_simple_reduction, but in the end only x_3 and x_2 are marked
as relevant.  vect_analyze_loop_operations skips over x_3, leaving the
vectorizable_reduction check to a presumed future test of x_1, which
in this case never happens.  We therefore end up vectorising x_2 only
(complete with peeling for niters!) and leave the scalar x_3 in place.

This caused a segfault in the support for fully-masked loops,
since there were no statements that needed masking.  Fixed by
checking for that.

But I think this is also a flaw in vect_analyze_loop_operations.
Outer loop vectorisation reduces the number of times that the
inner loop is executed, so it wouldn't necessarily be valid
to leave the scalar x_3 in place for all vectorisable x_2.
There's already code to forbid that when x_1 isn't present:

              /* FORNOW: we currently don't support the case that these phis
                 are not used in the outerloop (unless it is double reduction,
                 i.e., this phi is vect_reduction_def), cause this case
                 requires to actually do something here.  */

I think we need to do the same if x_1 is present but not relevant.

2018-01-19  Richard Sandiford  <richard.sandif...@linaro.org>

gcc/
        PR tree-optimization/83922
        * tree-vect-loop.c (vect_verify_full_masking): Return false if
        there are no statements that need masking.
        (vect_active_double_reduction_p): New function.
        (vect_analyze_loop_operations): Use it when handling phis that
        are not in the loop header.

gcc/testsuite/
        PR tree-optimization/83922
        * gcc.dg/pr83922.c: New test.

Added:
    trunk/gcc/testsuite/gcc.dg/pr83922.c
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/testsuite/ChangeLog
    trunk/gcc/tree-vect-loop.c

Reply via email to