procyberian/gcc
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

re PR tree-optimization/50561 (ICE when compiling zlib with -O2 -floop-flatten -floop-strip-mine)

Browse source 
2012-02-15  Tobias Grosser <grosser@fim.uni-passau.de>

	PR tree-optimization/50561
	* graphite-flattening.c (lst_project_loop): Do not
	remove old scattering dimensions after flattening.
	(lst_do_flatten): Likewise.

	* gcc.dg/graphite/pr50561.c: New testcase.

From-SVN: r184265
This commit is contained in:
Tobias Grosser 2012-02-15 12:20:10 +00:00 • committed by Richard Biener
parent 0c578db6dd
commit 28db21ee45
4 changed files with 56 additions and 7 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

7
gcc/ChangeLog
View file

@ -1,3 +1,10 @@
2012-02-15 Tobias Grosser <grosser@fim.uni-passau.de>
PR tree-optimization/50561
* graphite-flattening.c (lst_project_loop): Do not
remove old scattering dimensions after flattening.
(lst_do_flatten): Likewise.
2012-02-15 Georg-Johann Lay <avr@gjlay.de>
* doc/extend.texi (AVR Built-in Functions): Remove doc for

42
gcc/graphite-flattening.c
View file

@ -277,12 +277,26 @@ lst_project_loop (lst_p outer, lst_p inner, mpz_t stride)
ppl_delete_Linear_Expression (expr);
/* Remove inner loop and the static schedule of its body. */
ds = XNEWVEC (ppl_dimension_type, 2);
ds[0] = inner_dim;
ds[1] = inner_dim + 1;
ppl_Polyhedron_remove_space_dimensions (poly, ds, 2);
PBB_NB_SCATTERING_TRANSFORM (pbb) -= 2;
free (ds);
/* FIXME: As long as we use PPL we are not able to remove the old
scattering dimensions. The reason is that these dimensions are not
entirely unused. They are not necessary as part of the scheduling
vector, as the earlier dimensions already unambiguously define the
execution time, however they may still be needed to carry modulo
constraints as introduced e.g. by strip mining. The correct solution
would be to project these dimensions out of the scattering polyhedra.
In case they are still required to carry modulo constraints they should be kept
internally as existentially quantified dimensions. PPL does only support
projection of rational polyhedra, however in this case we need an integer
projection. With isl this will be trivial to implement. For now we just
leave the dimensions. This is a little ugly, but should be correct. */
if (0) {
ds = XNEWVEC (ppl_dimension_type, 2);
ds[0] = inner_dim;
ds[1] = inner_dim + 1;
ppl_Polyhedron_remove_space_dimensions (poly, ds, 2);
PBB_NB_SCATTERING_TRANSFORM (pbb) -= 2;
free (ds);
}
}
mpz_clear (x);
@ -412,7 +426,21 @@ lst_do_flatten (lst_p lst)
if (LST_LOOP_P (l))
{
res |= lst_flatten_loop (l, zero);
remove_unused_scattering_dimensions (l);
/* FIXME: As long as we use PPL we are not able to remove the old
scattering dimensions. The reason is that these dimensions are not
entirely unused. They are not necessary as part of the scheduling
vector, as the earlier dimensions already unambiguously define the
execution time, however they may still be needed to carry modulo
constraints as introduced e.g. by strip mining. The correct solution
would be to project these dimensions out of the scattering polyhedra.
In case they are still required to carry modulo constraints they should be kept
internally as existentially quantified dimensions. PPL does only support
projection of rational polyhedra, however in this case we need an integer
projection. With isl this will be trivial to implement. For now we just
leave the dimensions. This is a little ugly, but should be correct. */
if (0)
remove_unused_scattering_dimensions (l);
}
lst_update_scattering (lst);

5
gcc/testsuite/ChangeLog
View file

@ -1,3 +1,8 @@
2012-02-15 Tobias Grosser <grosser@fim.uni-passau.de>
PR tree-optimization/50561
* gcc.dg/graphite/pr50561.c: New testcase.
2012-02-15 Georg-Johann Lay <avr@gjlay.de>
* gcc.target/avr/torture/builtin_insert_bits-1.c: New test.

9
gcc/testsuite/gcc.dg/graphite/pr50561.c Normal file
View file

@ -0,0 +1,9 @@
/* { dg-do compile } */
/* { dg-options "-O2 -floop-flatten -floop-strip-mine" } */
void f (unsigned *s)
{
int n;
for (n = 0; n < 256; n++)
s[n] = 0;
}