With the attached patch applied, immuses get hosed during update_stmt while
propagating in ccp at tree-ssa-propagate.c:1154 (substitute_and_fold calling
mark_new_vars_to_rename). Testcase:
struct B { int i; };
struct C { struct B c[1]; };
int e;
static struct C b;
void
foo (int d)
{
struct C *x = &b;
b.c[0].i = 424242;
/* CCP destroys immuses of d */
e = x->c[d].i;
}
we just have substituted &b into x and folded it to
D.1618_6 = bD.1612.cD.1610[dD.1613_5].iD.1608
update_stmt on this will remove the immediate use of dD.1613_5
which then DCE happily removes and we ICE in verify_ssa of it.
Just before CCP the trees look like
foo (dD.1613)
{
struct C * xD.1616;
intD.0 D.1618;
intD.0 d.0D.1617;
# BLOCK 0
# PRED: ENTRY (fallthru)
xD.1616_1 = &bD.1612;
# SFT.1D.1625_3 = V_MUST_DEF <SFT.1D.1625_2>;
bD.1612.cD.1610[0].iD.1608 = 424242;
dD.1613_5 = dD.1613_4;
# VUSE <SFT.1D.1625_3>;
D.1618_6 = xD.1616_1->cD.1610[dD.1613_5].iD.1608;
# eD.1611_8 = V_MUST_DEF <eD.1611_7>;
eD.1611 = D.1618_6;
return;
# SUCC: EXIT
}
with a single use of dD.1613_5.
--
Summary: immediate uses hosed after CCP
Product: gcc
Version: 4.1.0
Status: UNCONFIRMED
Keywords: ice-on-valid-code
Severity: normal
Priority: P2
Component: tree-optimization
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: rguenth at gcc dot gnu dot org
CC: gcc-bugs at gcc dot gnu dot org
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=23297
