This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: [lno] ICE with large unsigned loop bound
- From: Zdenek Dvorak <rakdver at atrey dot karlin dot mff dot cuni dot cz>
- To: Falk Hueffner <hueffner at informatik dot uni-tuebingen dot de>
- Cc: gcc at gcc dot gnu dot org, Sebastian Pop <pop at cri dot ensmp dot fr>
- Date: Sun, 9 May 2004 20:43:02 +0200
- Subject: Re: [lno] ICE with large unsigned loop bound
- References: <87k6zlsb6j.fsf@informatik.uni-tuebingen.de>
Hello,
> on Alpha (when adding #define UNITS_PER_WORD 8) I get an ICE with gcc
> -O -ftree-vectorize for this code:
>
> extern char call_used_regs[64];
> void dw2_build_landing_pads(void)
> {
> unsigned int j;
> for (j = 0; j < 0xffff0000; ++j) {
> if (call_used_regs[j]) { }
> }
> }
>
> test.c: In function `dw2_build_landing_pads':
> test.c:4: internal compiler error: in tree_low_cst, at tree.c:3559
>
> (the -ftree-vectorize and "if" seems to be just needed to get the loop
> analyzer interested in the loop)
>
> The problem seems to be that the number of iterations (niter) in
> canonicalize_loop_induction_variables is "overflow -65536". This
> eventually comes from first_iteration_non_satisfying_ev_noev calling
> fold with integer_type_node, which seems wrong. Shouldn't it rather be
> size_type_node or the actual type of the iv?
it definitely should be the type of the induction variable (better said,
its unsigned variant). Unfortunately there are quite a few such type
wrong relicts from the earlier versions of the scev analysis :-(.
Sebastian, could you please fix this one (and other related you come
over)?
Thanks,
Zdenek