------- Comment #2 from ubizjak at gmail dot com 2007-03-09 18:21 -------
No, this is not a target problem. I have traced the problem down to
reload_cse_simplify(), where insn that loads flags (either test or sahf) gets
cleared. This is highly suspicious part of code (why value is set to 0 ?):
if (!count && reload_cse_noop_set_p (body))
{
rtx value = SET_DEST (body);
if (REG_P (value)
&& ! REG_FUNCTION_VALUE_P (value))
value = 0;
delete_insn_and_edges (insn);
return;
}
So, reload_cse_noop_set_p() through rtx_equal_for_cse_lib_p() finds many copies
of:
(set (reg:CCNO 17 flags)
(compare:CCNO (and:SI (zero_extract:SI (reg:HI 0 ax [73])
(const_int 8 [0x8])
(const_int 8 [0x8]))
(const_int 4 [0x4]))
(const_int 0 [0x0])))
And since postreload CSE pass thinks it can copy them from previous flag
register (neglecting the fact that ax is changed in a preceeding insn), it
clears all but the first two occurences of flags loading insn. Following this,
second postreload CSE pass removes also fnstsw instructions.
--
ubizjak at gmail dot com changed:
What |Removed |Added
----------------------------------------------------------------------------
Component|target |rtl-optimization
Summary|Problem with -funroll-all- |Problem with -funroll-all-
|loops |loops (postreload CSE
| |problems)
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31101
