I ran into this problem when compiling the GNU Emacs trunk with a GCC
4.5.2 that I built on RHEL 5.5 (x86-64). I narrowed it down to the
following stripped-down test case. This smells different from the previous
bug report I filed in this area (PR48228) on the same platform.
When I compile the following program with "gcc -S -Wstrict-overflow -O2"
GCC reports "warning: assuming pointer wraparound does not occur when comparing
P +- C1 with P +- C2". This warning is incorrect, since
signed overflow is obviously impossible in this function: the only
pointers computed are head_table and head_table + 1, which are both
in range. Changing the "+ 1" to "+ 7" generates even more warnings,
though the program is still correct.
const int *p;
int x = 0;
for (p = head_table; p < head_table + 1; p++)
x ^= *p;
Confirmed. Happens from
#1 0x00000000007307e7 in fold_undefer_overflow_warnings (issue=1 '\001',
281 warning_at (locus, OPT_Wstrict_overflow, "%s", warnmsg);
#2 0x0000000000afdfdf in evaluate_stmt (stmt=0x7ffff5b373c0)
2117 fold_undefer_overflow_warnings (is_constant, stmt, 0);
#3 0x0000000000aff7c5 in visit_cond_stmt (stmt=0x7ffff5b373c0,
2443 val = evaluate_stmt (stmt);
#4 0x0000000000aff8f8 in ccp_visit_stmt (stmt=0x7ffff5b373c0,
2501 return visit_cond_stmt (stmt, taken_edge_p);
(gdb) call debug_gimple_stmt (stmt)
if (p_1 < &head_table)
which calls fold_binary with &head_table < &head_table and in the end warns
8688 if (code != EQ_EXPR
8689 && code != NE_EXPR
8690 && bitpos0 != bitpos1
8691 && (pointer_may_wrap_p (base0, offset0, bitpos0)
8692 || pointer_may_wrap_p (base1, offset1, bitpos1)))
8693 fold_overflow_warning (("assuming pointer wraparound does not "
8694 "occur when comparing P +- C1 with "
8695 "P +- C2"),
because we call pointer_may_wrap_p with base0 which isn't an address
but an object which is bogus. It's argument has to depend on
indirect_base in which case we stripped an ADDR_EXPR.