GCC 4.3.2 bug (was: Illegal subtraction in tmp-dive_1.s)

[Vincent Lefevre]

I've looked at tree output of mpn/generic/rootrem.c and differences
between gcc 4.3.1 (where the bug doesn't occur) and gcc 4.3.2 (as
gcc 4.3.3 isn't installed on the machine).

To avoid spurious differences, I had to replace

#include <stdio.h>              /* for NULL */

by

#define NULL ((void *)0)

I could check that the bug is still triggered in gcc 4.3.2.

The first differences are in the rootrem.c.052t.salias file (there
are many differences). Then the next difference:

diff -aurd 431/rootrem.c.058t.fre 432/rootrem.c.058t.fre
--- 431/rootrem.c.058t.fre      2009-04-17 15:01:26.000000000 +0200
+++ 432/rootrem.c.058t.fre      2009-04-17 15:01:45.000000000 +0200
@@ -979,7 +979,7 @@
   D.4421_328 = D.4421_316;
   D.4422_329 = D.4422_317;
   D.4423_330 = D.4423_318;
-  D.4424_331 = *D.4423_330;
+  D.4424_331 = D.4399_260;
   D.4432_332 = D.4424_331 == 0;
   qn_333 = qn_310 - D.4432_332;
 

This corresponds to the line 322:

   318            qn = b / GMP_NUMB_BITS + 1; /* b+1 bits */
   319            MPN_ZERO (qp, qn);
   320            qp[qn - 1] = (mp_limb_t) 1 << (b % GMP_NUMB_BITS);
   321            MPN_DECR_U (qp, qn, 1);
   322            qn -= qp[qn - 1] == 0;

D.4399_260 is the value of (mp_limb_t) 1 << (b % GMP_NUMB_BITS).
So, I assume that gcc thinks that qp[qn - 1] hasn't been modified
by MPN_DECR_U.

MPN_DECR_U contains:

  mp_ptr __p = (qp);

Then __p[...] is modified.

Concerning D.4423_330, there's a change in rootrem.c.052t.salias,
but I don't know what it means:

-D.4423_330 = qp_75
+D.4423_330 = ANYTHING

The problem in rootrem.c.058t.fre disappears with the option
-fno-strict-aliasing.

The problem also disappears if I remove __attribute__ ((malloc))
on the __gmp_tmp_reentrant_alloc declaration.

-- 
Vincent Lefèvre <vincent at vinc17.org> - Web: <http://www.vinc17.org/>
100% accessible validated (X)HTML - Blog: <http://www.vinc17.org/blog/>
Work: CR INRIA - computer arithmetic / Arenaire project (LIP, ENS-Lyon)