------- Comment #5 from hjl at lucon dot org 2007-08-07 14:40 -------
There is a disconnect between middle-end and back-end. There are many
places in builtins.c with
/* If argument is already integer valued, and we don't need to worry
about setting errno, there's no need to perform rounding. */
if (! flag_errno_math && integer_valued_real_p (arg))
However, the x86 back-end doesn't have this info:
(define_expand "lround<SSEMODEF:mode><SSEMODEI24:mode>2"
[(match_operand:SSEMODEI24 0 "nonimmediate_operand" "")
(match_operand:SSEMODEF 1 "register_operand" "")]
"SSE_FLOAT_MODE_P (<SSEMODEF:MODE>mode) && TARGET_SSE_MATH
&& ((<SSEMODEI24:MODE>mode != DImode) || TARGET_64BIT)
&& !flag_trapping_math && !flag_rounding_math
&& !optimize_size"
{
ix86_expand_lround (operand0, operand1);
DONE;
})
That is why it fails to expand.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33007
