[PATCH, 4.5] Make fwprop find out that non-virtual constant member pointers are non-virtual
Martin Jambor
mjambor@suse.cz
Sat Mar 28 18:28:00 GMT 2009
Hi,
On Wed, Mar 25, 2009 at 12:46:30PM +0100, Richard Guenther wrote:
> 2009/3/24 Martin Jambor <mjambor@suse.cz>:
> > Hi,
> >
> > this is a second attempt to partially fix PR 3713 - partial in the
> > sence that it does away with the rather embarrasing run-time
> > condition, for more details see the thread that started with the first
> > attempt:
> >
> > http://gcc.gnu.org/ml/gcc-patches/2009-03/msg00536.html
> >
> > The patch below takes the core of the fix to fold-const.c by partially
> > undoing the fix for PR 35705. I believe that this shouls still work
> > on any target on which member pointers work. I have not verified on a
> > hppa yet because gcc61 was down until recently. I will test it there
> > tomorrow. The hunks in tree-ssa-forwprop.c is not necessary (a tleast
> > not for the test case), FRE would manage to get rid of the condition
> > even without it. However, this way we get rid of it sooner (and hey,
> > once we might be able to do so in the early fwprop).
> >
> > I have bootstrapped and tested this on x86_64-linux. Provided it does
> > not reintroduce PR 35705, is it OK for 4.5?
>
> This is ok for 4.5 if you check ...
>
I have tried running the snippet in PR 35705 and the patch does not
reintroduce the problem (I was able to reproduce it by reverting its
fix though). I have also amended the patch as requested by Richi,
rebootstrapped and re-tested on x86_64-linux and just committed the
following as revision 145203.
Thanks,
Martin
2009-03-28 Martin Jambor <mjambor@suse.cz>
* fold-const.c (get_pointer_modulus_and_residue): New parameter
allow_func_align.
(fold_binary): Allow function decl aligment consideration is the
second argument is integer constant one.
* tree-ssa-forwprop.c (simplify_bitwise_and): New function.
(tree_ssa_forward_propagate_single_use_vars): Handle assing statements
with BIT_AND_EXPR on the RHS by calling simplify_bitwise_and.
* g++.dg/tree-ssa/fwprop-align.C: New test.
Index: gcc/fold-const.c
===================================================================
--- gcc/fold-const.c (revision 145196)
+++ gcc/fold-const.c (working copy)
@@ -9551,10 +9551,15 @@
0 <= N < M as is common. In general, the precise value of P is unknown.
M is chosen as large as possible such that constant N can be determined.
- Returns M and sets *RESIDUE to N. */
+ Returns M and sets *RESIDUE to N.
+ If ALLOW_FUNC_ALIGN is true, do take functions' DECL_ALIGN_UNIT into
+ account. This is not always possible due to PR 35705.
+ */
+
static unsigned HOST_WIDE_INT
-get_pointer_modulus_and_residue (tree expr, unsigned HOST_WIDE_INT *residue)
+get_pointer_modulus_and_residue (tree expr, unsigned HOST_WIDE_INT *residue,
+ bool allow_func_align)
{
enum tree_code code;
@@ -9584,7 +9589,8 @@
}
}
- if (DECL_P (expr) && TREE_CODE (expr) != FUNCTION_DECL)
+ if (DECL_P (expr)
+ && (allow_func_align || TREE_CODE (expr) != FUNCTION_DECL))
return DECL_ALIGN_UNIT (expr);
}
else if (code == POINTER_PLUS_EXPR)
@@ -9595,7 +9601,8 @@
op0 = TREE_OPERAND (expr, 0);
STRIP_NOPS (op0);
- modulus = get_pointer_modulus_and_residue (op0, residue);
+ modulus = get_pointer_modulus_and_residue (op0, residue,
+ allow_func_align);
op1 = TREE_OPERAND (expr, 1);
STRIP_NOPS (op1);
@@ -11235,7 +11242,8 @@
unsigned HOST_WIDE_INT modulus, residue;
unsigned HOST_WIDE_INT low = TREE_INT_CST_LOW (arg1);
- modulus = get_pointer_modulus_and_residue (arg0, &residue);
+ modulus = get_pointer_modulus_and_residue (arg0, &residue,
+ integer_onep (arg1));
/* This works because modulus is a power of 2. If this weren't the
case, we'd have to replace it by its greatest power-of-2
Index: gcc/testsuite/g++.dg/tree-ssa/fwprop-align.C
===================================================================
--- gcc/testsuite/g++.dg/tree-ssa/fwprop-align.C (revision 0)
+++ gcc/testsuite/g++.dg/tree-ssa/fwprop-align.C (revision 0)
@@ -0,0 +1,20 @@
+/* { dg-do compile } */
+/* { dg-options "-O2 -fdump-tree-forwprop2" } */
+
+struct A
+{
+ void foo ()
+ {
+ }
+};
+
+int main()
+{
+ void (A::* const p)() = & A::foo;
+ A a;
+ (a.*p)();
+}
+
+/* We should eliminate the check if p points to a virtual function. */
+/* { dg-final { scan-tree-dump-times "& 1" 0 "forwprop2" } } */
+/* { dg-final { cleanup-tree-dump "forwprop2" } } */
Index: gcc/tree-ssa-forwprop.c
===================================================================
--- gcc/tree-ssa-forwprop.c (revision 145196)
+++ gcc/tree-ssa-forwprop.c (working copy)
@@ -147,6 +147,14 @@
ptr2 = &x[index];
+ Or
+ ssa = (int) decl
+ res = ssa & 1
+
+ Provided that decl has known alignment >= 2, will get turned into
+
+ res = 0
+
We also propagate casts into SWITCH_EXPR and COND_EXPR conditions to
allow us to remove the cast and {NOT_EXPR,NEG_EXPR} into a subsequent
{NOT_EXPR,NEG_EXPR}.
@@ -1124,6 +1132,45 @@
}
}
+/* Run bitwise and assignments throug the folder. If the first argument is an
+ ssa name that is itself a result of a typecast of an ADDR_EXPR to an
+ integer, feed the ADDR_EXPR to the folder rather than the ssa name.
+*/
+
+static void
+simplify_bitwise_and (gimple_stmt_iterator *gsi, gimple stmt)
+{
+ tree res;
+ tree arg1 = gimple_assign_rhs1 (stmt);
+ tree arg2 = gimple_assign_rhs2 (stmt);
+
+ if (TREE_CODE (arg2) != INTEGER_CST)
+ return;
+
+ if (TREE_CODE (arg1) == SSA_NAME && !SSA_NAME_IS_DEFAULT_DEF (arg1))
+ {
+ gimple def = SSA_NAME_DEF_STMT (arg1);
+
+ if (gimple_assign_cast_p (def)
+ && INTEGRAL_TYPE_P (gimple_expr_type (def)))
+ {
+ tree op = gimple_assign_rhs1 (def);
+
+ if (TREE_CODE (op) == ADDR_EXPR)
+ arg1 = op;
+ }
+ }
+
+ res = fold_binary (BIT_AND_EXPR, TREE_TYPE (gimple_assign_lhs (stmt)),
+ arg1, arg2);
+ if (res && is_gimple_min_invariant (res))
+ {
+ gimple_assign_set_rhs_from_tree (gsi, res);
+ update_stmt (stmt);
+ }
+ return;
+}
+
/* Main entry point for the forward propagation optimizer. */
static unsigned int
@@ -1206,6 +1253,11 @@
else
gsi_next (&gsi);
}
+ else if (gimple_assign_rhs_code (stmt) == BIT_AND_EXPR)
+ {
+ simplify_bitwise_and (&gsi, stmt);
+ gsi_next (&gsi);
+ }
else
gsi_next (&gsi);
}
More information about the Gcc-patches
mailing list