------- Comment #3 from wilson at tuliptree dot org 2008-04-12 00:45 -------
Subject: Re: New: unsinged long long and while loop evaluation
regression?
I can reproduce this on a 32-bit x86-linux machine (i.e. a 32-bit HWI).
The unsigned long long 0xffffffff becomes a (const_double -1 0), and
then in expand_mult in expmed.c we have
/* If we are multiplying in DImode, it may still be a win
to try to work with shifts and adds. */
if (CONST_DOUBLE_HIGH (op1) == 0)
coeff = CONST_DOUBLE_LOW (op1);
After this line, expand_mult thinks we are multiplying by -1 and we get
the wrong result.
I think there is a false assumption here that we can get CONST_DOUBLEs
which can be simplified to a single word. Maybe in the olden days we
always created a CONST_DOUBLE for DImode constants? This stuff has
changed so many times it is hard to remember. I don't think we do it
that way anymore.
Anyways, if this assumption is not false, then the code needs to look
more like the code in immed_double_const in emit-rtl.c, which does
/* If this integer fits in one word, return a CONST_INT. */
if ((i1 == 0 && i0 >= 0) || (i1 == ~0 && i0 < 0))
return GEN_INT (i0);
where i1 is CONST_DOUBLE_HIGH and i0 is CONST_DOUBLE_LOW, and only in
the case that this tests succeeds can we set coeff to CONST_DOUBLE_LOW.
The same bug is in mainline, and probably goes a long ways back.
Jim
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35885
