This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
[PATCH] Fix VRP with -fno-delete-null-pointer-checks (PR c/88367, take 2)
- From: Jakub Jelinek <jakub at redhat dot com>
- To: Richard Biener <rguenther at suse dot de>
- Cc: gcc-patches at gcc dot gnu dot org
- Date: Thu, 6 Dec 2018 13:09:09 +0100
- Subject: [PATCH] Fix VRP with -fno-delete-null-pointer-checks (PR c/88367, take 2)
- References: <20181206064535.GN12380@tucnak> <alpine.LSU.2.20.1812061002570.1827@zhemvz.fhfr.qr>
- Reply-to: Jakub Jelinek <jakub at redhat dot com>
On Thu, Dec 06, 2018 at 10:05:15AM +0100, Richard Biener wrote:
> Note I wonder if with -fwrapv-pointer NULL automatically becomes a
> valid address? Or is only wrapping around half of the address
> space UB?
Hadn't thought about -fwrapv-pointer, I guess we (especially with
-fno-delete-null-pointer-checks) need to be even more conservative in that
case.
Furthermore, I've discovered that the ADDR_EXPR of MEM_REF case actually
uses get_base_address and therefore the offset on MEM_REF is just one of the
many possible offsets in the play.
So, this patch punts for -fwrapv-pointer in some further cases, and
adjusts the vr-values.c ADDR_EXPR handling code so that it sums up all 2 or
3 offsets together and looks at the resulting sign. If
-fdelete-null-pointer-checks -fno-wrapv-pointer, it does what it did before
in tree-vrp.c and in vr-values.c is even more aggressive than before, as in
even if the base pointer is varying etc., if the sum of all the offsets
is provably non-zero, the result is non-NULL. For
-fno-delete-null-pointer-checks -fno-wrapv-pointer it does this only if the
resulting offset is positive.
Does this look ok?
2018-12-06 Jakub Jelinek <jakub@redhat.com>
PR c/88367
* tree-vrp.c (extract_range_from_binary_expr): For POINTER_PLUS_EXPR
with -fno-delete-null-pointer-checks, set_nonnull only if the pointer
is non-NULL and offset is known to have most significant bit clear.
* vr-values.c (vr_values::vrp_stmt_computes_nonzero): For ADDR_EXPR
of MEM_EXPR, return true if the MEM_EXPR has non-zero offset with
most significant bit clear. If offset does have most significant bit
set and -fno-delete-null-pointer-checks, don't return true even if
the base pointer is non-NULL.
* gcc.dg/tree-ssa/pr88367.c: New test.
--- gcc/tree-vrp.c.jj 2018-12-06 11:19:24.170939864 +0100
+++ gcc/tree-vrp.c 2018-12-06 11:50:12.104711210 +0100
@@ -1673,9 +1673,26 @@ extract_range_from_binary_expr (value_ra
else if (code == POINTER_PLUS_EXPR)
{
/* For pointer types, we are really only interested in asserting
- whether the expression evaluates to non-NULL. */
- if (!range_includes_zero_p (&vr0)
- || !range_includes_zero_p (&vr1))
+ whether the expression evaluates to non-NULL.
+ With -fno-delete-null-pointer-checks we need to be more
+ conservative. As some object might reside at address 0,
+ then some offset could be added to it and the same offset
+ subtracted again and the result would be NULL.
+ E.g.
+ static int a[12]; where &a[0] is NULL and
+ ptr = &a[6];
+ ptr -= 6;
+ ptr will be NULL here, even when there is POINTER_PLUS_EXPR
+ where the first range doesn't include zero and the second one
+ doesn't either. As the second operand is sizetype (unsigned),
+ consider all ranges where the MSB could be set as possible
+ subtractions where the result might be NULL. */
+ if ((!range_includes_zero_p (&vr0)
+ || !range_includes_zero_p (&vr1))
+ && !TYPE_OVERFLOW_WRAPS (expr_type)
+ && (flag_delete_null_pointer_checks
+ || (range_int_cst_p (&vr1)
+ && !tree_int_cst_sign_bit (vr1.max ()))))
vr->set_nonnull (expr_type);
else if (range_is_null (&vr0) && range_is_null (&vr1))
vr->set_null (expr_type);
--- gcc/vr-values.c.jj 2018-12-06 11:19:23.550950006 +0100
+++ gcc/vr-values.c 2018-12-06 12:59:28.269999920 +0100
@@ -297,14 +297,48 @@ vr_values::vrp_stmt_computes_nonzero (gi
&& gimple_assign_rhs_code (stmt) == ADDR_EXPR)
{
tree expr = gimple_assign_rhs1 (stmt);
- tree base = get_base_address (TREE_OPERAND (expr, 0));
+ poly_int64 bitsize, bitpos;
+ tree offset;
+ machine_mode mode;
+ int unsignedp, reversep, volatilep;
+ tree base = get_inner_reference (TREE_OPERAND (expr, 0), &bitsize,
+ &bitpos, &offset, &mode, &unsignedp,
+ &reversep, &volatilep);
if (base != NULL_TREE
&& TREE_CODE (base) == MEM_REF
&& TREE_CODE (TREE_OPERAND (base, 0)) == SSA_NAME)
{
- value_range *vr = get_value_range (TREE_OPERAND (base, 0));
- if (!range_includes_zero_p (vr))
+ poly_offset_int off = 0;
+ bool off_cst = false;
+ if (offset == NULL_TREE || TREE_CODE (offset) == INTEGER_CST)
+ {
+ off = mem_ref_offset (base);
+ if (offset)
+ off += poly_offset_int::from (wi::to_poly_wide (offset),
+ SIGNED);
+ off <<= LOG2_BITS_PER_UNIT;
+ off += bitpos;
+ off_cst = true;
+ }
+ /* If &X->a is equal to X and X is ~[0, 0], the result is too.
+ For -fdelete-null-pointer-checks -fno-wrapv-pointer we don't
+ allow going from non-NULL pointer to NULL. */
+ if ((off_cst && known_eq (off, 0))
+ || (flag_delete_null_pointer_checks
+ && !TYPE_OVERFLOW_WRAPS (TREE_TYPE (expr))))
+ {
+ value_range *vr = get_value_range (TREE_OPERAND (base, 0));
+ if (!range_includes_zero_p (vr))
+ return true;
+ }
+ /* If MEM_REF has a "positive" offset, consider it non-NULL
+ always, for -fdelete-null-pointer-checks also "negative"
+ ones. Punt for unknown offsets (e.g. variable ones). */
+ if (!TYPE_OVERFLOW_WRAPS (TREE_TYPE (expr))
+ && off_cst
+ && known_ne (off, 0)
+ && (flag_delete_null_pointer_checks || known_gt (off, 0)))
return true;
}
}
--- gcc/testsuite/gcc.dg/tree-ssa/pr88367.c.jj 2018-12-06 11:46:51.915985811 +0100
+++ gcc/testsuite/gcc.dg/tree-ssa/pr88367.c 2018-12-06 13:00:14.692248340 +0100
@@ -0,0 +1,31 @@
+/* PR c/88367 */
+/* { dg-do compile } */
+/* { dg-options "-fno-delete-null-pointer-checks -O2 -fdump-tree-optimized -fno-wrapv-pointer" } */
+/* { dg-final { scan-tree-dump-not "link_error \\(\\);" "optimized" } } */
+/* { dg-final { scan-tree-dump-times "bar \\(\\);" 2 "optimized" } } */
+
+void bar (void);
+void link_error (void);
+
+void
+foo (char *p)
+{
+ if (!p)
+ return;
+ p += 3;
+ if (!p)
+ link_error ();
+ p -= 6;
+ if (!p)
+ bar ();
+}
+
+void
+baz (char *p)
+{
+ if (!p)
+ return;
+ p -= 6;
+ if (!p)
+ bar ();
+}
Jakub