------- Comment #4 from rguenth at gcc dot gnu dot org 2008-01-17 10:45 -------
So, we enter the recip pass with
<bb 2>:
if (x_2(D) == y_3(D))
goto <bb 3>;
else
goto <bb 4>;
<bb 3>:
D.1547_4 = x_2(D) / x_2(D);
D.1548_5 = D.1547_4 * x_2(D);
D.1545_6 = D.1548_5 / x_2(D);
<bb 4>:
# D.1545_1 = PHI <D.1545_6(3), 0.0(2)>
return D.1545_1;
that is, we compute x/x*x/x. Recip now 'optimizes' this to
reciptmp.6_10 = 1.0e+0 / x_2(D);
D.1547_4 = __builtin_pow (reciptmp.6_10, 2.0e+0);
D.1548_5 = D.1547_4 * x_2(D);
D.1545_6 = D.1548_5 * reciptmp.6_10;
which is bogus anyway. But the ICE is gone :P
Well, it registers the division x_2(D)/x_2(D) twice, which also makes
it a wrong-code regression.
The easiest thing is probably to just not do reciprocal optimization
for x / x.
--
rguenth at gcc dot gnu dot org changed:
What |Removed |Added
----------------------------------------------------------------------------
Keywords| |wrong-code
Priority|P3 |P1
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34825
