[Bug middle-end/33436] Bad constant output with TARGET_ASM_ALIGNED_DI_OP

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

------- Comment #1 from danglin at gcc dot gnu dot org  2007-09-22 01:18 -------
Although it might be possible to fix this at an earlier point,
things go seriously wrong in emit_group_load_1.  We have:

Breakpoint 7, emit_group_load_1 (tmps=0x800003fffdff4a70,
    dst=0x800003fffdfb7540, orig_src=0x800003fffdfacf30,
    type=0x800003fffdf9b750, ssize=16) at ../../gcc/gcc/expr.c:1779
1779          else if (CONSTANT_P (src))
(gdb) p debug_rtx (src)
(const_double 4919131751843889152 [0x4444444400000000] 1229782938533634594
[0x1111111122222222] 0 [0x0] 0 [0x0])
$7 = void
(gdb) p debug_rtx (dst)
(parallel:BLK [
        (expr_list:REG_DEP_TRUE (reg:DI 66 [ <result> ])
            (const_int 0 [0x0]))
        (expr_list:REG_DEP_TRUE (reg:DI 67 [ <result>+8 ])
            (const_int 8 [0x8]))
    ])

As can be seen, dst is a parallel but src is a CONST_DOUBLE.  The code
as currently written assigns src to both parts of the parallel.

I believe this occurs because hppa64 doesn't support TImode.

The break above is in a modified version of expr.c which appears to
correct the problem.  Here's the diff:

Index: expr.c
===================================================================
--- expr.c      (revision 128642)
+++ expr.c      (working copy)
@@ -1776,8 +1776,23 @@
       else if (CONSTANT_P (src) && GET_MODE (dst) != BLKmode
                && XVECLEN (dst, 0) > 1)
         tmps[i] = simplify_gen_subreg (mode, src, GET_MODE(dst), bytepos);
-      else if (CONSTANT_P (src)
-              || (REG_P (src) && GET_MODE (src) == mode))
+      else if (CONSTANT_P (src))
+       {
+         if ((HOST_WIDE_INT) bytelen == ssize)
+           tmps[i] = src;
+         else
+           {
+             rtx first, second;
+
+             gcc_assert (i <= 1);
+             split_double (src, &first, &second);
+             if (i)
+               tmps[i] = second;
+             else
+               tmps[i] = first;
+           }
+       }
+      else if (REG_P (src) && GET_MODE (src) == mode)
        tmps[i] = src;
       else
        tmps[i] = extract_bit_field (src, bytelen * BITS_PER_UNIT,

I used split_double because neither simplify_subreg or extract_bit_field
handle this.

I'm a little unhappy about splitting the double but in some cases it is
possible to load the values without forcing them to memory.  In any event,
this is a rare situation.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33436