------- Comment #19 from ubizjak at gmail dot com 2008-02-08 14:21 -------
The core of the problem is, that for profiled bootstrap, the call to
normalize() from do_multiply() is simply gone. Gone in the sense, that
normalize() is neither inlined at the call site, neither is called in
do_multiply(). This leads to various observed strange effects when real numbers
are processed.
for (k = j; k < SIGSZ * 2; k += 2)
{
unsigned long bi = b->sig[k / 2];
if (k & 1)
bi >>= HOST_BITS_PER_LONG / 2;
else
bi &= ((unsigned long)1 << (HOST_BITS_PER_LONG / 2)) - 1;
u.sig[k / 2] = ai * bi;
}
>>>>> normalize (&u);
inexact |= do_add (rr, rr, &u, 0);
}
}
Adding __attirbute__((noinline)) to protptype of normalize() is enough to fix
profiled bootstrap:
Index: real.c
===================================================================
--- real.c (revision 132182)
+++ real.c (working copy)
@@ -98,7 +98,7 @@
static void clear_significand_below (REAL_VALUE_TYPE *, unsigned int);
static bool div_significands (REAL_VALUE_TYPE *, const REAL_VALUE_TYPE *,
const REAL_VALUE_TYPE *);
-static void normalize (REAL_VALUE_TYPE *);
+static void normalize (REAL_VALUE_TYPE *) __attribute__((noinline));
static bool do_add (REAL_VALUE_TYPE *, const REAL_VALUE_TYPE *,
const REAL_VALUE_TYPE *, int);
So, what upsets gcc's inliner/profiler/whatever to drop the call?
--
ubizjak at gmail dot com changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |ubizjak at gmail dot com
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33992
