Re: [tree-ssa] Bootstrap failure on i686-linux

[law at redhat dot com]

In message <1062466928.6332.27.camel@frodo.toronto.redhat.com>, Diego Novillo w
rites:
 >After undoing those patches I ran into another failure that was caused
 >by:
 >
 >+2003-09-01  Zdenek Dvorak  <rakdver@atrey.karlin.mff.cuni.cz>
 >+           Jeff Law  <law@redhat.com>
 >+
 >+       * basic-block.h (BB_LOOP_CONTROL_EXPR): Remove.
 >+       * c-simplify.c (gimplify_c_loop): Create loops in a better shape.
 >+       * cfgloop.h (create_loop_notes): Declare.
 >+       * cfgloopmanip.c (create_loop_notes): New.
 >+       * explow.c (probe_stack_range): Don't produce loop notes when we
 >+       recreate them.
 >+       * expr.c (emit_block_move_via_loop): Ditto.
 >+       * gimplify.c (build_and_jump): Export.
 >+       (gimplify_loop_expr): Don't produce LOOP_EXPRs.
 >+       * toplev.c (rest_of_compilation): Recreate loop notes if needed.
 >+       * tree-cfg.c (make_loop_expr_edges, make_loop_expr_blocks): Removed.
 >+       (find_contained_blocks, make_ctrl_stmt_edges,
 >+       remove_useless_stmts_and_vars, stmt_ends_bb_p, bsi_insert_before,
 >+       find_insert_location, bsi_insert_on_edge_immediate,
 >+       merge_tree_blocks): Remove handling of LOOP_EXPRs.
 >+       (remove_stmt): Remove handling of BB_LOOP_CONTROL_EXPR.
 >+       (find_taken_edge): Remove comment on LOOP_EXPRs.
 >+       (dump_tree_bb): Don't dump loop-related information.
 >+       (is_loop_stmt, is_latch_block_for): Removed.
 >+       (find_insert_location): Handle TRY_CATCH and TRY_FINALLY.
 >+       * tree-flow.h (is_loop_stmt, loop_body, set_loop_body,
 >+       is_latch_block_for): Removed.
 >+       * tree-dfa.c (get_stmt_operands): Don't handle LOOP_EXPRs.
 >+       * tree-simple.c (is_gimple_stmt): Remove handling of LOOP_EXPRs.
 >+       * tree-simple.h: Remove LOOP_EXPRs from gimple grammar comment.
 >+       (build_and_jump): Declare.
 >+       * tree-ssa.c (remove_annotations_r): Don't handle LOOP_EXPRs.
 >+
 >
 >This one fails when building libiberty after bootstrapping:
This is clearly PRE's fault.  It fucks up the block statement iterators.  If
anything needs to be turned off, it is PRE.

Using my simplified testcase (-O2 -pedantic).  The block #3 exhibits the
problem:

      # BLOCK 3.  PRED: 2.  SUCC: 4.
      <UL5380>:;;
       
      #   VUSE <MT.1_6>;
      T.1_12 = *p_1;
      mcnt_13 = (int)T.1_12;
      goto <UL5150>;;
      (void)0;


If I put a breakpoint just before PRE runs:

(gdb) p basic_block_info.data.bb[3]
$32 = (struct basic_block_def *) 0x89d9f7c
(gdb) p bsi_start ($32)
$33 = {tp = 0xbf1a867c, context = 0x0}
(gdb) p bsi_stmt ($33)
$34 = (union tree_node *) 0xbf1861dc
(gdb) p debug_tree($34)
 <label_expr 0xbf1861dc
    type <void_type 0xbf1c78c0 void VOID
        align 8 symtab 0 alias set -1
        pointer_to_this <pointer_type 0xbf1c7930>>
    side-effects
    arg 0 <label_decl 0xbf185380 VOID file j.c line 23
        align 1 alias set 1 context <function_decl 0xbf183b60 foo>
        j.c:23>>

Which looks perfectly reasonable.


After PRE is complete:

(gdb) p basic_block_info.data.bb[3]
$37 = (struct basic_block_def *) 0x89d9f7c
(gdb) p bsi_start ($37)
$38 = {tp = 0x0, context = 0x89d9f7c}
(gdb) p bsi_end_p ($38)
$39 = true


Which looks rather bogus to me.

Jeff