[Bug tree-optimization/112636] New: ICE on valid code at -O1 (but not at -O{0,s,2,3}) with "-ftree-vectorize": in adjust_loop_info_after_peeling, at tree-ssa-loop-ivcanon.cc:1069

[Bug tree-optimization/112636] New: ICE on valid code at -O1 (but not at -O{0,s,2,3}) with "-ftree-vectorize": in adjust_loop_info_after_peeling, at tree-ssa-loop-ivcanon.cc:1069

[zhendong.su at inf dot ethz.ch]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112636

            Bug ID: 112636
           Summary: ICE on valid code at -O1 (but not at -O{0,s,2,3}) with
                    "-ftree-vectorize": in adjust_loop_info_after_peeling,
                    at tree-ssa-loop-ivcanon.cc:1069
           Product: gcc
           Version: unknown
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: zhendong.su at inf dot ethz.ch
  Target Milestone: ---

This appears to be a recent regression as it does not seem to reproduce with
13.2.

Compiler Explorer: https://godbolt.org/z/WYsasWefh

[526] % gcctk -v
Using built-in specs.
COLLECT_GCC=gcctk
COLLECT_LTO_WRAPPER=/local/suz-local/software/local/gcc-trunk/libexec/gcc/x86_64-pc-linux-gnu/14.0.0/lto-wrapper
Target: x86_64-pc-linux-gnu
Configured with: ../gcc-trunk/configure --disable-bootstrap
--enable-checking=yes --prefix=/local/suz-local/software/local/gcc-trunk
--enable-sanitizers --enable-languages=c,c++ --disable-werror --enable-multilib
Thread model: posix
Supported LTO compression algorithms: zlib
gcc version 14.0.0 20231119 (experimental) (GCC) 
[527] % 
[527] % gcctk -O1 small.c; ./a.out
[528] % 
[528] % gcctk -O1 -ftree-vectorize small.c
during GIMPLE pass: ch_vect
small.c: In function ‘main’:
small.c:3:5: internal compiler error: in adjust_loop_info_after_peeling, at
tree-ssa-loop-ivcanon.cc:1069
    3 | int main() {
      |     ^~~~
0x87ac95 adjust_loop_info_after_peeling(loop*, int, bool)
        ../../gcc-trunk/gcc/tree-ssa-loop-ivcanon.cc:1069
0x129b514 copy_headers
        ../../gcc-trunk/gcc/tree-ssa-loop-ch.cc:1108
Please submit a full bug report, with preprocessed source (by using
-freport-bug).
Please include the complete backtrace with any bug report.
See <https://gcc.gnu.org/bugs/> for instructions.
[529] % 
[529] % cat small.c
int a[1], b;
unsigned c;
int main() {
  while (b) {
    if (a[c])
      break;
    c--;
  }
  return 0;
}

[Bug tree-optimization/112636] [14 Regression] ICE on valid code at -O1 (but not at -O{0,s,2,3}) with "-ftree-vectorize": in adjust_loop_info_after_peeling, at tree-ssa-loop-ivcanon.cc:1069

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112636

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|---                         |14.0
                 CC|                            |hubicka at gcc dot gnu.org
           Keywords|                            |ice-on-valid-code,
                   |                            |missed-optimization
           Priority|P3                          |P1
            Version|unknown                     |14.0
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2023-11-20
            Summary|ICE on valid code at -O1    |[14 Regression] ICE on
                   |(but not at -O{0,s,2,3})    |valid code at -O1 (but not
                   |with "-ftree-vectorize": in |at -O{0,s,2,3}) with
                   |adjust_loop_info_after_peel |"-ftree-vectorize": in
                   |ing, at                     |adjust_loop_info_after_peel
                   |tree-ssa-loop-ivcanon.cc:10 |ing, at
                   |69                          |tree-ssa-loop-ivcanon.cc:10
                   |                            |69
     Ever confirmed|0                           |1

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
Confirmed.

      if (wi::leu_p (npeel, loop->nb_iterations_upper_bound))
        loop->nb_iterations_upper_bound -= npeel;
      else
        {
          /* Peeling maximal number of iterations or more
             makes no sense and is a bug.
             We should peel completely.  */
          gcc_unreachable ();

so we're asking to adjust for peeling of 1 iteration with an upper bound
of iterations of zero.  We derive this limit from the array size of a[].

Caused by Honzas loop header copying changes.

[Bug tree-optimization/112636] [14 Regression] ICE on valid code at -O1 (but not at -O{0,s,2,3}) with "-ftree-vectorize": in adjust_loop_info_after_peeling, at tree-ssa-loop-ivcanon.cc:1069

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112636

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Assignee|unassigned at gcc dot gnu.org      |rguenth at gcc dot gnu.org
             Status|NEW                         |ASSIGNED

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
The issue is that should_duplicate_loop_header_p we call after

      basic_block header = loop->header;
      if (!get_max_loop_iterations_int (loop))
        {
          if (dump_file && (dump_flags & TDF_DETAILS))
            fprintf (dump_file, "Loop %d never loops.\n", loop->num);
          scale_loop_profile (loop, profile_probability::always (), 0);
          loops_to_unloop.safe_push (loop);
          loops_to_unloop_nunroll.safe_push (0);
          continue;

ends up calling estimate_numbers_of_iterations () only then actually
updating the upper bound we check above.  It does so hidden via

#0  estimate_numbers_of_iterations (loop=0x7ffff700e4b0)
    at /space/rguenther/src/gcc/gcc/tree-ssa-loop-niter.cc:4806
#1  0x00000000019434ef in loop_exits_before_overflow (
    base=<ssa_name 0x7ffff71e5120 8>, step=<integer_cst 0x7ffff702d180>, 
    at_stmt=<gimple_phi 0x7ffff71bb700>, loop=0x7ffff700e4b0)
    at /space/rguenther/src/gcc/gcc/tree-ssa-loop-niter.cc:5259
#2  0x00000000019440b0 in scev_probably_wraps_p (var=<tree 0x0>, 
    base=<ssa_name 0x7ffff71e5120 8>, step=<integer_cst 0x7ffff702d180>, 
    at_stmt=<gimple_phi 0x7ffff71bb700>, loop=0x7ffff700e4b0, 
    use_overflow_semantics=true)
    at /space/rguenther/src/gcc/gcc/tree-ssa-loop-niter.cc:5511
#3  0x0000000001bb97ea in get_scev_info (r=..., 
    name=<ssa_name 0x7ffff7016f78 4>, stmt=<gimple_phi 0x7ffff71bb700>, 
    l=0x7ffff700e4b0, init=@0x7fffffffa898: <ssa_name 0x7ffff71e5120 8>, 
    step=@0x7fffffffa890: <integer_cst 0x7ffff702d180>, 
    dir=@0x7fffffffa88c: EV_DIR_DECREASES)
    at /space/rguenther/src/gcc/gcc/vr-values.cc:204
#4  0x0000000001bb9d6d in range_of_var_in_loop (v=..., 
    name=<ssa_name 0x7ffff7016f78 4>, l=0x7ffff700e4b0, 
    stmt=<gimple_phi 0x7ffff71bb700>, query=0x4961550)
    at /space/rguenther/src/gcc/gcc/vr-values.cc:271
#5  0x0000000002edf824 in fold_using_range::range_of_ssa_name_with_loop_info (
    this=0x7fffffffb46f, r=..., name=<ssa_name 0x7ffff7016f78 4>, 
...
#22 0x00000000018f3f77 in should_duplicate_loop_header_p (
    header=<basic_block 0x7ffff71be6c0 (4)>, loop=0x7ffff700e4b0, 
    ranger=0x4961550, limit=0x7fffffffd88c, invariant_exits=0x47e60f0, 
    static_exits=0x4961040)
    at /space/rguenther/src/gcc/gcc/tree-ssa-loop-ch.cc:251
#23 0x00000000018f59d7 in (anonymous namespace)::ch_base::copy_headers (
    this=0x47f2a30, fun=0x7ffff71e2000)
    at /space/rguenther/src/gcc/gcc/tree-ssa-loop-ch.cc:831

The best is to analyze niters for the loop we try to process.

[Bug tree-optimization/112636] [14 Regression] ICE on valid code at -O1 (but not at -O{0,s,2,3}) with "-ftree-vectorize": in adjust_loop_info_after_peeling, at tree-ssa-loop-ivcanon.cc:1069

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112636

--- Comment #3 from GCC Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Richard Biener <rguenth@gcc.gnu.org>:

https://gcc.gnu.org/g:05e8ef2a05b477589cae25af3311bad0f68a90fe

commit r14-7138-g05e8ef2a05b477589cae25af3311bad0f68a90fe
Author: Richard Biener <rguenther@suse.de>
Date:   Thu Jan 11 13:35:51 2024 +0100

    tree-optimization/112636 - estimate niters before header copying

    The following avoids a mismatch between an early query for maximum
    number of iterations of a loop and a late one when through ranger
    we'd get iterations estimated.  Instead make sure we compute niters
    before querying the iteration bound.

            PR tree-optimization/112636
            * tree-ssa-loop-ch.cc (ch_base::copy_headers): Call
            estimate_numbers_of_iterations before querying
            get_max_loop_iterations_int.
            (pass_ch::execute): Initialize SCEV and loops appropriately.

            * gcc.dg/pr112636.c: New testcase.

[Bug tree-optimization/112636] [14 Regression] ICE on valid code at -O1 (but not at -O{0,s,2,3}) with "-ftree-vectorize": in adjust_loop_info_after_peeling, at tree-ssa-loop-ivcanon.cc:1069

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112636

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
         Resolution|---                         |FIXED
             Status|ASSIGNED                    |RESOLVED

--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
Fixed.