[Bug rtl-optimization/24209] strange instruction selected for an annuled delay slot

["steven at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=24209

Steven Bosscher <steven at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Last reconfirmed|2005-10-05 11:06:49         |2013-04-18 11:06:49
                 CC|gcc-bugs at gcc dot gnu.org |steven at gcc dot gnu.org

--- Comment #3 from Steven Bosscher <steven at gcc dot gnu.org> 2013-04-18 21:06:17 UTC ---
Current trunk still picks that ldsb insn for the delay slot.
Here's what it produces:

        .file   "t.c"
        .section        ".text"
        .align 4
        .global PointToRowCol
        .type   PointToRowCol, #function
        .proc   020
PointToRowCol:
        sethi   %hi(term+4), %g1
        sethi   %hi(firstValidRow), %g2
        ld      [%g1+%lo(term+4)], %g1
        ld      [%g2+%lo(firstValidRow)], %g2
        ld      [%g1], %g4
        ld      [%g1+12], %g3
        sub     %o0, %g4, %o0
        sra     %o0, 31, %o5
        wr      %o5, 0, %y
        ld      [%g3+4], %o5
        nop
        nop
        sdiv    %o0, %o5, %o0
        cmp     %o0, %g2
        bl,a    .L12
         ldsb   [%g1+16], %o4
        sethi   %hi(lastValidRow), %o5
        ld      [%o5+%lo(lastValidRow)], %o5
        cmp     %o0, %o5
        bg      .L10
         mov    %o0, %g2
.L2:
        ldsb    [%g1+16], %o4
.L12:
        cmp     %o4, 0
        bne     .L4
         mov    0, %o5
        ld      [%g3+8], %o5
.L4:
        add     %o5, %g4, %g4
        sub     %o1, %g4, %o1
        sra     %o1, 31, %g4
        wr      %g4, 0, %y
        ld      [%g3], %g4
        nop
        nop
        sdiv    %o1, %g4, %o1
        cmp     %o1, 0
        bl,a    .L8
         st     %g2, [%o2]
        ld      [%g1+4], %g1
        add     %g1, 1, %g1
        cmp     %o1, %g1
        bg      .L11
         st     %g2, [%o2]
        jmp     %o7+8
         st     %o1, [%o3]
.L11:
        mov     %g1, %o1
        jmp     %o7+8
         st     %o1, [%o3]
.L10:
        b       .L2
         mov    %o5, %g2
.L8:
        mov     0, %o1
        jmp     %o7+8
         st     %o1, [%o3]
        .size   PointToRowCol, .-PointToRowCol
        .ident  "GCC: (GNU) 4.9.0 20130418 (experimental) [trunk revision
198052]"
        .section        .note.GNU-stack,"",@progbits


for this test case:

typedef char Boolean;

typedef struct {
  int width;
} SbInfo;

struct _vtwin {
  int f_width;
  int f_height;
  SbInfo sb_info;
};

typedef struct {
  int border;
  int max_col;
  struct _vtwin *whichVwin;
} TScreen;

typedef struct _Misc {
  Boolean useRight;
} Misc;

typedef struct _XtermWidgetRec {
  TScreen screen;
  Misc misc;
} XtermWidgetRec, *XtermWidget;

extern int firstValidRow, lastValidRow;

extern XtermWidget term;

void
PointToRowCol(int y, int x, int *r, int *c)
{
  TScreen *screen = &term->screen;
  int row, col;

  row = (y - screen->border) / screen->whichVwin->f_height;
  if (row < firstValidRow)
    row = firstValidRow;
  else if (row > lastValidRow)
    row = lastValidRow;
  col = (x - (((term->misc.useRight) ? 0 : screen->whichVwin->sb_info.width) +
screen->border) ) / screen->whichVwin->f_width;
  if (col < 0)
    col = 0;
  else if (col > screen->max_col + 1)
    col = screen->max_col + 1;

  *r = row;
  *c = col;
}

with options: "-mcpu=v8 -m32 -O2".