https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70390
--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
graphite uses a strange "premature optimization" with scev_analyzable_p
which seems to be confused about a degenerate IV we have in the loop.
<bbN>
_1 = PHI <_2, _3>
_3 = _1;
....
if () goto exit
goto bbN;
exit:
_5 = PHI <_3>
interestingly the copyprop pass before graphite doesn't fix this up.
Ah, before copyprop we have
_57 = -1;
...
# prephitmp_59 = PHI <g_lsm.18_37(15), _16(6)>
_16 = _57 & prephitmp_59;
which SCCP introduces (during it's constant "non-"propagation, not during
final value replacement).
Note that copyprop doesn't visit "uninteresting" stmts (like _16 = _57 &
prephitmp_59), it merely folds them at the end of propagation which
figures out _57 is -1.
Not sure if the GRAPHITE assert is necessary or merely a hint that some
optimization didn't apply - that is, do we actually generate wrong-code
if it triggers?
Index: gcc/graphite-isl-ast-to-gimple.c
===================================================================
--- gcc/graphite-isl-ast-to-gimple.c (revision 234449)
+++ gcc/graphite-isl-ast-to-gimple.c (working copy)
@@ -2112,7 +2112,7 @@ copy_loop_close_phi_args (basic_block ol
if (is_gimple_reg (res) && scev_analyzable_p (res, region->region))
/* Loop close phi nodes should not be scev_analyzable_p. */
- gcc_unreachable ();
+ /*gcc_unreachable ()*/;
gphi *new_close_phi = create_phi_node (SSA_NAME_VAR (res), new_bb);
tree new_res = create_new_def_for (res, new_close_phi,
@@ -2497,7 +2497,7 @@ copy_cond_phi_nodes (basic_block bb, bas
continue;
if (is_gimple_reg (res) && scev_analyzable_p (res, region->region))
/* Cond phi nodes should not be scev_analyzable_p. */
- gcc_unreachable ();
+ /*gcc_unreachable ()*/;
gphi *new_phi = create_phi_node (SSA_NAME_VAR (res), new_bb);
tree new_res = create_new_def_for (res, new_phi,
"fixes" the ICE.
Sebastian?
