gcc/clang -O1 bug

[gmhwxi]

I encountered a gcc/clang -O1 (not -O2) optimization bug that really
shocked me a bit.

In the code below, if the following line is dropped,

val () = ptr_volatile (p_nseq)

then the gcc/clang (with -O1) somehow thinks that the content of nseq is
never modified. As a consequence, the function always returns 0. If -O1
is dropped or changed to -O0, then the compiled code works fine. I also
tried tcc and noticed that it can compile the following code correctly.

This is a shock because the involved code is so simple!

To see what happens, you can try out the code at:

https://github.com/githwxi/ATS-Postiats-contrib/tree/master/projects/SMALL/M-N-K-game/Part1

implement
board_dirmax
(
  board, pid, dir, i, j
) = let
//
var nseq: int = 0
val p_nseq = addr@nseq
//
fun loop
(
  i: int, j: int
) : void = let
//
val pid2 = board[i, j]
//
in
  if pid = pid2
    then let
      val n = $UN.ptr1_get<int> (p_nseq)
      val () = $UN.ptr1_set<int> (p_nseq, n + 1)
      val (i, j) = dirnext (dir, i, j)
    in
      loop (i, j)
    end // end of [then]
  // end of [if]
end // end of [loop]
//
val () = ptr_volatile (p_nseq) // -O1 bug if omitted
val () = try loop (i, j) with ~MatrixSubscriptExn () => ()
//
in
  $UN.cast{intGte(0)}(nseq)
end // end of [board_dirmax]

[gmhwxi]

gcc-4.6.3 was used.

[Sean Leather]

After off-list discussion (my apologies, but see remark at end), Hongwei shared the generated C for this example. Follow the tmpref188 variable. It does indeed look like a bug in GCC. 

extern atstype_int _057_home_057_hwxi_057_research_057_Postiats_055_contrib_057_git_057_projects_057_SMALL_057_M_055_N_055_K_055_game_057_Part1_057_game_board_056_dats__board_dirmax(atstype_ptrk arg0, atstype_int arg1, atstype_boxed arg2, atstype_int arg3, atstype_int arg4) {
    atstype_int tmpret187 ;
    atstype_int tmpref188 ;
    ;
    ;
    atstype_exnconptr tmp270 ;
    __patsflab_board_dirmax: # 5637 "game_board_dats.c" (tmpref188 = 0) ;
    ATSLIB_056_prelude__ptr_volatile((&(tmpref188))) ;
    do {
        int flag ;
        atsexnframe_ptr frame ;
        atsexnframe_ptr *framep ;
        frame = alloca(sizeof(atsexnframe_t)) ;
        framep = my_atsexnframe_getref() ;
        my_atsexnframe_enter(frame, framep) ;
        flag = _setjmp (frame-&gt;
        env) ;
        if (flag==0) {
            loop_91(arg0, arg1, arg2, (&(tmpref188)), arg3, arg4) ;
            my_atsexnframe_leave(framep) ;
        }
        else {
            tmp270 = (*framep)-&gt;
            exn ;
            my_atsexnframe_leave(framep) ;
            do {
                ;
                __atstmplab23: if (!(((tmp270)==(void*)(&(ATSLIB_056_prelude__MatrixSubscriptExn))))) {
                    atsruntime_raise(tmp270) ;
                }
                __atstmplab24: # 5680 "game_board_dats.c"  ;
                break ;
            }
            while (0) ;
        }
    }
    while(0) ;
    (tmpret187 = ((atstype_int)tmpref188)) ;
    return (tmpret187) ;

}

About the off-list reply, this Google Group is apparently set up to have recipients post their replies “to the author of the message only”. I recommend changing the setting to “to the entire group”. On the group management page, I believe you can set this in Settings -> Email options -> Post replies.

[gmhwxi]

Okay, I made the email setting change.

[Ian Denhardt]

Would it be possible for you (or Hongwei) to attach the generated C source as a
separate file? I'd like to have a look at this, but it's come through a bit
mangled in the email.

-Ian

Quoting Sean Leather (2014-05-17 04:40:45)
> ...

> --
> You received this message because you are subscribed to the Google
> Groups "ats-lang-users" group.
> To unsubscribe from this group and stop receiving emails from it, send
> an email to [1]ats-lang-user...@googlegroups.com.
> To post to this group, send email to
> [2]ats-lan...@googlegroups.com.
> To view this discussion on the web visit
> [3]https://groups.google.com/d/msgid/ats-lang-users/2f01e82f-bb89-4999-
> a5ff-8095b902f82b%40googlegroups.com.
>
> References
>
> 1. mailto:ats-lang-user...@googlegroups.com
> 2. mailto:ats-lan...@googlegroups.com
> 3. https://groups.google.com/d/msgid/ats-lang-users/2f01e82f-bb89-4999-a5ff-8095b902f82b%40googlegroups.com?utm_medium=email&utm_source=footer

[Hongwei Xi]

Here it is.

The file is generated using 'gcc -E'.

The function of interest is the one whose name contains 'dirmax'.

If you remove the call to 'ptr_volatile', then 'gcc -O1' may generate buggy code.

My *guess* is that the cause of the problem is related to the use of longjmp/setjmp in the code.

Rich West showed me a bug of gcc the other day involving longjmp/setjmp. There might be a relation
between these two bugs.