[Bug rtl-optimization/58220] New: [4.9 Regression] Many new failures for SH after rev. 201833

[kkojima at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58220

            Bug ID: 58220
           Summary: [4.9 Regression] Many new failures for SH after rev.
                    201833
           Product: gcc
           Version: 4.9.0
            Status: UNCONFIRMED
          Keywords: wrong-code
          Severity: normal
          Priority: P3
         Component: rtl-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: kkojima at gcc dot gnu.org
            Target: sh*-*-*

There are many new execution errors on SH after the revision 201833

r201883 | tejohnson | 2013-08-20 22:29:53 +0900 (Tue, 20 Aug 2013) | 8 lines

The change of final.c:reemit_insn_block_notes for r201883

...
-  for (; insn; insn = next_active_insn (insn))
+  for (; insn; insn = next_insn (insn))
     {
       tree this_block;

+      /* Prevent lexical blocks from straddling section boundaries.  */
+      if (NOTE_P (insn) && NOTE_KIND (insn) == NOTE_INSN_SWITCH_TEXT_SECTIONS)
+        {
+          for (tree s = cur_block; s != DECL_INITIAL (cfun->decl);
+               s = BLOCK_SUPERCONTEXT (s))
+            {
+              rtx note = emit_note_before (NOTE_INSN_BLOCK_END, insn);
+              NOTE_BLOCK (note) = s;
+              note = emit_note_after (NOTE_INSN_BLOCK_BEG, insn);
+              NOTE_BLOCK (note) = s;
+            }
+        }
+
+      if (!active_insn_p (insn))
+        continue;

uses next_insn and active_insn_p pair instead of next_acrive_insn,
though next_insn handles SEQUENCEs specially:

rtx
next_insn (rtx insn)
{
  if (insn)
    {
      insn = NEXT_INSN (insn);
      if (insn && NONJUMP_INSN_P (insn)
      && GET_CODE (PATTERN (insn)) == SEQUENCE)
    insn = XVECEXP (PATTERN (insn), 0, 0);
    }

  return insn;
}

but next_active_insn does nothing special for SEQUENCEs:

rtx
next_active_insn (rtx insn)
{
  while (insn)
    {
      insn = NEXT_INSN (insn);
      if (insn == 0 || active_insn_p (insn))
    break;
    }

  return insn;
}

It looks now the for loop in reemit_insn_block_notes sees the inside
insn of SEQUENCE, not SEQUENCE insn itself.  The failures went away
with the following one liner.

    * final.c (reemit_insn_block_notes): Use NEXT_INSN instead of
    next_insn.

--- ORIG/trunk/gcc/final.c    2013-08-22 09:43:35.000000000 +0900
+++ trunk/gcc/final.c    2013-08-22 14:36:51.000000000 +0900
@@ -1650,7 +1650,7 @@ reemit_insn_block_notes (void)
   rtx insn, note;

   insn = get_insns ();
-  for (; insn; insn = next_insn (insn))
+  for (; insn; insn = NEXT_INSN (insn))
     {
       tree this_block;

Here is the diff of the generated codes with/without the above patch
for a typical example of failure gcc.c-torture/execute/20000422-1.c
with -Og -g:

--- good.s      2013-08-23 07:47:43.000000000 +0900
+++ bad.s       2013-08-23 07:47:16.000000000 +0900
@@ -27,6 +27,7 @@ main:
 .LBE2:
        .loc 1 17 0
        bra     .L2
+       nop
        mov     #0,r4
 .LVL1:
        .align 1

where SH uses a SEQUENCE for bra instruction and its delayed slot
which should be filled with nop in this case.