From mboxrd@z Thu Jan  1 00:00:00 1970
Return-Path: <gcc-patches-return-296334-listarch-gcc-patches=gcc.gnu.org@gcc.gnu.org>
Received: (qmail 23391 invoked by alias); 8 Jul 2011 09:36:14 -0000
Received: (qmail 22988 invoked by uid 22791); 8 Jul 2011 09:36:10 -0000
X-SWARE-Spam-Status: No, hits=-2.3 required=5.0	tests=AWL,BAYES_00,DKIM_SIGNED,DKIM_VALID,DKIM_VALID_AU,FREEMAIL_FROM,RCVD_IN_DNSWL_LOW,TW_TM
X-Spam-Check-By: sourceware.org
Received: from mail-wy0-f175.google.com (HELO mail-wy0-f175.google.com) (74.125.82.175)    by sourceware.org (qpsmtpd/0.43rc1) with ESMTP; Fri, 08 Jul 2011 09:35:54 +0000
Received: by wyg30 with SMTP id 30so1467247wyg.20        for <gcc-patches@gcc.gnu.org>; Fri, 08 Jul 2011 02:35:53 -0700 (PDT)
MIME-Version: 1.0
Received: by 10.227.55.66 with SMTP id t2mr1578689wbg.109.1310117753221; Fri, 08 Jul 2011 02:35:53 -0700 (PDT)
Received: by 10.227.36.212 with HTTP; Fri, 8 Jul 2011 02:35:53 -0700 (PDT)
In-Reply-To: <CAEwic4YkA+GT-KDjb-65=8whqVXevNPOu+KHo4a-4ze-jasxQA@mail.gmail.com>
References: <CAEwic4YkA+GT-KDjb-65=8whqVXevNPOu+KHo4a-4ze-jasxQA@mail.gmail.com>
Date: Fri, 08 Jul 2011 09:39:00 -0000
Message-ID: <CAFiYyc3QEF3grA0kxrjUmsBT59EtAzyFmbbRfQ+CEAFqJAg3yg@mail.gmail.com>
Subject: Re: [patch tree-optimization]: [3 of 3]: Boolify compares & more
From: Richard Guenther <richard.guenther@gmail.com>
To: Kai Tietz <ktietz70@googlemail.com>
Cc: GCC Patches <gcc-patches@gcc.gnu.org>
Content-Type: text/plain; charset=ISO-8859-1
Content-Transfer-Encoding: quoted-printable
X-IsSubscribed: yes
Mailing-List: contact gcc-patches-help@gcc.gnu.org; run by ezmlm
Precedence: bulk
List-Id: <gcc-patches.gcc.gnu.org>
List-Archive: <http://gcc.gnu.org/ml/gcc-patches/>
List-Post: <mailto:gcc-patches@gcc.gnu.org>
List-Help: <mailto:gcc-patches-help@gcc.gnu.org>
Sender: gcc-patches-owner@gcc.gnu.org
X-SW-Source: 2011-07/txt/msg00603.txt.bz2

On Thu, Jul 7, 2011 at 6:07 PM, Kai Tietz <ktietz70@googlemail.com> wrote:
> Hello,
>
> This patch - third of series - fixes vrp to handle bitwise one-bit
> precision typed operations.
> And it introduces a second - limitted to non-switch-statement range - vrp=
 pass.

Err - please split this patch.  I agree with Paolo, this 2nd
substitute_and_fold call is bogus.  More comments inline.

>
> Bootstrapped and regression tested for all standard-languages (plus
> Ada and Obj-C++) on host x86_64-pc-linux-gnu.
>
> Ok for apply?
>
> Regards,
> Kai
>
> ChangeLog
>
> 2011-07-07 =A0Kai Tietz =A0<ktietz@redhat.com>
>
> =A0 =A0 =A0 =A0* tree-vrp.c (in_second_pass): New static variable.
> =A0 =A0 =A0 =A0(extract_range_from_binary_expr): Add handling for
> =A0 =A0 =A0 =A0BIT_IOR_EXPR, BIT_AND_EXPR, and BIT_NOT_EXPR.
> =A0 =A0 =A0 =A0(register_edge_assert_for_1): Add handling for 1-bit
> =A0 =A0 =A0 =A0BIT_IOR_EXPR and BIT_NOT_EXPR.
> =A0 =A0 =A0 =A0(register_edge_assert_for): Add handling for 1-bit
> =A0 =A0 =A0 =A0BIT_IOR_EXPR.
> =A0 =A0 =A0 =A0(ssa_name_get_inner_ssa_name_p): New helper function.
> =A0 =A0 =A0 =A0(ssa_name_get_cast_to_p): New helper function.
> =A0 =A0 =A0 =A0(simplify_truth_ops_using_ranges): Handle prefixed
> =A0 =A0 =A0 =A0cast instruction for result, and add support for one
> =A0 =A0 =A0 =A0bit precision BIT_IOR_EXPR, BIT_AND_EXPR, BIT_XOR_EXPR,
> =A0 =A0 =A0 =A0, and BIT_NOT_EXPR.
> =A0 =A0 =A0 =A0(simplify_stmt_using_ranges): Add handling for one bit
> =A0 =A0 =A0 =A0precision BIT_IOR_EXPR, BIT_AND_EXPR, BIT_XOR_EXPR,
> =A0 =A0 =A0 =A0and BIT_NOT_EXPR.
> =A0 =A0 =A0 =A0(vrp_finalize): Do substitute and fold pass a second
> =A0 =A0 =A0 =A0time for vrp_stmt and preserve switch-edge simplification
> =A0 =A0 =A0 =A0on second run.
> =A0 =A0 =A0 =A0(simplify_switch_using_ranges): Preserve rerun of function
> =A0 =A0 =A0 =A0in second pass.
>
> Index: gcc-head/gcc/tree-vrp.c
> =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=
=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=
=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D
> --- gcc-head.orig/gcc/tree-vrp.c
> +++ gcc-head/gcc/tree-vrp.c
> @@ -74,6 +74,9 @@ struct value_range_d
>
> =A0typedef struct value_range_d value_range_t;
>
> +/* This flag indicates that we are doing a second pass of VRP. =A0*/
> +static bool in_second_pass =3D false;
> +
> =A0/* Set of SSA names found live during the RPO traversal of the function
> =A0 =A0for still active basic-blocks. =A0*/
> =A0static sbitmap *live;
> @@ -2232,6 +2235,7 @@ extract_range_from_binary_expr (value_ra
> =A0 =A0 =A0some cases. =A0*/
> =A0 if (code !=3D BIT_AND_EXPR
> =A0 =A0 =A0 && code !=3D TRUTH_AND_EXPR
> + =A0 =A0 =A0&& code !=3D BIT_IOR_EXPR

Huh?  So how would VARYING | x ever produce something better
than VARYING?

> =A0 =A0 =A0 && code !=3D TRUTH_OR_EXPR
> =A0 =A0 =A0 && code !=3D TRUNC_DIV_EXPR
> =A0 =A0 =A0 && code !=3D FLOOR_DIV_EXPR
> @@ -2291,6 +2295,8 @@ extract_range_from_binary_expr (value_ra
> =A0 =A0 =A0 =A0 =A0else
> =A0 =A0 =A0 =A0 =A0 =A0set_value_range_to_varying (vr);
> =A0 =A0 =A0 =A0}
> + =A0 =A0 =A0else if (code =3D=3D BIT_IOR_EXPR)
> + =A0 =A0 =A0 =A0set_value_range_to_varying (vr);

err - BIT_IOR_EXPR on pointers?

> =A0 =A0 =A0 else
> =A0 =A0 =A0 =A0gcc_unreachable ();
>
> @@ -2300,11 +2306,13 @@ extract_range_from_binary_expr (value_ra
> =A0 /* For integer ranges, apply the operation to each end of the
> =A0 =A0 =A0range and see what we end up with. =A0*/
> =A0 if (code =3D=3D TRUTH_AND_EXPR
> - =A0 =A0 =A0|| code =3D=3D TRUTH_OR_EXPR)
> + =A0 =A0 =A0|| code =3D=3D TRUTH_OR_EXPR
> + =A0 =A0 =A0|| ((code =3D=3D BIT_AND_EXPR || code =3D=3D BIT_IOR_EXPR)
> + =A0 =A0 =A0 =A0 =A0&& TYPE_PRECISION (TREE_TYPE (op1)) =3D=3D 1))

Rather than adding code to handle BIT_*_EXPR this patch should
transform the TRUTH_*_EXPR handling to appropriate BIT_*_EXPR
handling as we no longer have TRUTH_*_EXPR in our IL.

In fact I would say the existing BIT_*_EXPR handling should already
cover all the TRUTH_*_CASES, so this patch patches the wrong
spot if it is necessary at all.

> =A0 =A0 {
> =A0 =A0 =A0 /* If one of the operands is zero, we know that the whole
> =A0 =A0 =A0 =A0 expression evaluates zero. =A0*/
> - =A0 =A0 =A0if (code =3D=3D TRUTH_AND_EXPR
> + =A0 =A0 =A0if ((code =3D=3D TRUTH_AND_EXPR || code =3D=3D BIT_AND_EXPR)
> =A0 =A0 =A0 =A0 =A0&& ((vr0.type =3D=3D VR_RANGE
> =A0 =A0 =A0 =A0 =A0 =A0 =A0 && integer_zerop (vr0.min)
> =A0 =A0 =A0 =A0 =A0 =A0 =A0 && integer_zerop (vr0.max))
> @@ -2317,7 +2325,7 @@ extract_range_from_binary_expr (value_ra
> =A0 =A0 =A0 =A0}
> =A0 =A0 =A0 /* If one of the operands is one, we know that the whole
> =A0 =A0 =A0 =A0 expression evaluates one. =A0*/
> - =A0 =A0 =A0else if (code =3D=3D TRUTH_OR_EXPR
> + =A0 =A0 =A0else if ((code =3D=3D TRUTH_OR_EXPR || code =3D=3D BIT_IOR_E=
XPR)
> =A0 =A0 =A0 =A0 =A0 =A0 =A0 && ((vr0.type =3D=3D VR_RANGE
> =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0&& integer_onep (vr0.min)
> =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0&& integer_onep (vr0.max))
> @@ -2809,7 +2817,7 @@ extract_range_from_unary_expr (value_ran
> =A0 =A0 =A0cannot easily determine a resulting range. =A0*/
> =A0 if (code =3D=3D FIX_TRUNC_EXPR
> =A0 =A0 =A0 || code =3D=3D FLOAT_EXPR
> - =A0 =A0 =A0|| code =3D=3D BIT_NOT_EXPR
> + =A0 =A0 =A0|| (code =3D=3D BIT_NOT_EXPR && TYPE_PRECISION (type) !=3D 1)
> =A0 =A0 =A0 || code =3D=3D CONJ_EXPR)
> =A0 =A0 {
> =A0 =A0 =A0 /* We can still do constant propagation here. =A0*/
> @@ -3976,7 +3984,9 @@ build_assert_expr_for (tree cond, tree v
> =A0 =A0 =A0 tree a =3D build2 (ASSERT_EXPR, TREE_TYPE (v), v, cond);
> =A0 =A0 =A0 assertion =3D gimple_build_assign (n, a);
> =A0 =A0 }
> - =A0else if (TREE_CODE (cond) =3D=3D TRUTH_NOT_EXPR)
> + =A0else if (TREE_CODE (cond) =3D=3D TRUTH_NOT_EXPR
> + =A0 =A0 =A0 =A0 =A0|| (TREE_CODE (cond) =3D=3D BIT_NOT_EXPR
> + =A0 =A0 =A0 =A0 =A0 =A0 =A0&& TYPE_PRECISION (TREE_TYPE (cond)) =3D=3D =
1))
> =A0 =A0 {
> =A0 =A0 =A0 /* Given !V, build the assignment N =3D false. =A0*/
> =A0 =A0 =A0 tree op0 =3D TREE_OPERAND (cond, 0);
> @@ -4531,7 +4541,9 @@ register_edge_assert_for_1 (tree op, enu
> =A0 =A0 =A0 retval |=3D register_edge_assert_for_1 (gimple_assign_rhs2 (o=
p_def),
> =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =
=A0 =A0 =A0 =A0code, e, bsi);
> =A0 =A0 }
> - =A0else if (gimple_assign_rhs_code (op_def) =3D=3D TRUTH_NOT_EXPR)
> + =A0else if (gimple_assign_rhs_code (op_def) =3D=3D TRUTH_NOT_EXPR
> + =A0 =A0 =A0 =A0 =A0|| (gimple_assign_rhs_code (op_def) =3D=3D BIT_NOT_E=
XPR
> + =A0 =A0 =A0 =A0 =A0 =A0 =A0&& TYPE_PRECISION (TREE_TYPE (op)) =3D=3D 1))
> =A0 =A0 {
> =A0 =A0 =A0 /* Recurse, flipping CODE. =A0*/
> =A0 =A0 =A0 code =3D invert_tree_comparison (code, false);
> @@ -4617,6 +4629,9 @@ register_edge_assert_for (tree name, edg
>
> =A0 =A0 =A0 if (is_gimple_assign (def_stmt)
> =A0 =A0 =A0 =A0 =A0&& (gimple_assign_rhs_code (def_stmt) =3D=3D TRUTH_OR_=
EXPR
> + =A0 =A0 =A0 =A0 =A0 =A0 || (gimple_assign_rhs_code (def_stmt) =3D=3D BI=
T_IOR_EXPR
> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 && INTEGRAL_TYPE_P (TREE_TYPE (name))
> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 && TYPE_PRECISION (TREE_TYPE (name)) =
=3D=3D 1)
> =A0 =A0 =A0 =A0 =A0 =A0 =A0/* For BIT_IOR_EXPR only if NAME =3D=3D 0 both=
 operands have
> =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 necessarily zero value. =A0*/
> =A0 =A0 =A0 =A0 =A0 =A0 =A0|| (comp_code =3D=3D EQ_EXPR
> @@ -6747,19 +6762,96 @@ varying:
> =A0 return SSA_PROP_VARYING;
> =A0}
>
> +/* Returns operand1 of ssa-name with SSA_NAME as code, Otherwise it
> + =A0 returns NULL_TREE. =A0*/

?  Why would you want to look through a single copy?

> +static tree
> +ssa_name_get_inner_ssa_name_p (tree op)
> +{
> + =A0gimple stmt;
> +
> + =A0if (TREE_CODE (op) !=3D SSA_NAME
> + =A0 =A0 =A0|| !is_gimple_assign (SSA_NAME_DEF_STMT (op)))
> + =A0 =A0return NULL_TREE;
> + =A0stmt =3D SSA_NAME_DEF_STMT (op);
> + =A0if (gimple_assign_rhs_code (stmt) !=3D SSA_NAME)
> + =A0 =A0return NULL_TREE;
> + =A0return gimple_assign_rhs1 (stmt);
> +}
> +
> +/* Returns operand of cast operation, if OP is a type-conversion. Otherw=
ise
> + =A0 return NULL_TREE. =A0*/
> +static tree
> +ssa_name_get_cast_to_p (tree op)
> +{
> + =A0gimple stmt;
> +
> + =A0if (TREE_CODE (op) !=3D SSA_NAME
> + =A0 =A0 =A0|| !is_gimple_assign (SSA_NAME_DEF_STMT (op)))
> + =A0 =A0return NULL_TREE;
> + =A0stmt =3D SSA_NAME_DEF_STMT (op);
> + =A0if (!CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (stmt)))
> + =A0 =A0return NULL_TREE;
> + =A0return gimple_assign_rhs1 (stmt);
> +}
> +
> =A0/* Simplify boolean operations if the source is known
> =A0 =A0to be already a boolean. =A0*/
> =A0static bool
> =A0simplify_truth_ops_using_ranges (gimple_stmt_iterator *gsi, gimple stm=
t)
> =A0{
> =A0 enum tree_code rhs_code =3D gimple_assign_rhs_code (stmt);
> + =A0gimple stmt2 =3D stmt;
> =A0 tree val =3D NULL;
> - =A0tree op0, op1;
> + =A0tree op0, op1, cop0, cop1;
> =A0 value_range_t *vr;
> =A0 bool sop =3D false;
> =A0 bool need_conversion;
> + =A0location_t loc =3D gimple_location (stmt);
>
> =A0 op0 =3D gimple_assign_rhs1 (stmt);
> + =A0op1 =3D NULL_TREE;
> +
> + =A0/* Handle cases with prefixed type-cast. =A0*/

What cases?  This code lacks comments.

> + =A0if (CONVERT_EXPR_CODE_P (rhs_code)

So this simplifies conversions, not truth ops.

> + =A0 =A0 =A0&& INTEGRAL_TYPE_P (TREE_TYPE (op0))
> + =A0 =A0 =A0&& TREE_CODE (op0) =3D=3D SSA_NAME
> + =A0 =A0 =A0&& is_gimple_assign (SSA_NAME_DEF_STMT (op0))
> + =A0 =A0 =A0&& INTEGRAL_TYPE_P (TREE_TYPE (gimple_assign_lhs (stmt))))
> + =A0 =A0{
> + =A0 =A0 =A0stmt2 =3D SSA_NAME_DEF_STMT (op0);
> + =A0 =A0 =A0op0 =3D gimple_assign_rhs1 (stmt2);
> + =A0 =A0 =A0if (!INTEGRAL_TYPE_P (TREE_TYPE (op0)))
> + =A0 =A0 =A0 return false;
> + =A0 =A0 =A0rhs_code =3D gimple_assign_rhs_code (stmt2);
> + =A0 =A0 =A0if (rhs_code !=3D BIT_NOT_EXPR && rhs_code !=3D TRUTH_NOT_EX=
PR
> + =A0 =A0 =A0 =A0 && rhs_code !=3D TRUTH_AND_EXPR && rhs_code !=3D BIT_AN=
D_EXPR
> + =A0 =A0 =A0 =A0 && rhs_code !=3D TRUTH_OR_EXPR && rhs_code !=3D BIT_IOR=
_EXPR
> + =A0 =A0 =A0 =A0 && rhs_code !=3D TRUTH_XOR_EXPR && rhs_code !=3D BIT_XO=
R_EXPR
> + =A0 =A0 =A0 =A0 && rhs_code !=3D NE_EXPR && rhs_code !=3D EQ_EXPR)
> + =A0 =A0 =A0 return false;
> + =A0 =A0 =A0if (rhs_code =3D=3D BIT_AND_EXPR || rhs_code =3D=3D BIT_IOR_=
EXPR
> + =A0 =A0 =A0 =A0 || rhs_code =3D=3D BIT_XOR_EXPR || rhs_code =3D=3D TRUT=
H_AND_EXPR
> + =A0 =A0 =A0 =A0 || rhs_code =3D=3D TRUTH_OR_EXPR || rhs_code =3D=3D TRU=
TH_XOR_EXPR
> + =A0 =A0 =A0 =A0 || rhs_code =3D=3D NE_EXPR || rhs_code =3D=3D EQ_EXPR)
> + =A0 =A0 =A0 op1 =3D gimple_assign_rhs2 (stmt2);
> + =A0 =A0 =A0if (gimple_has_location (stmt2))
> + =A0 =A0 =A0 =A0loc =3D gimple_location (stmt2);
> + =A0 =A0}
> + =A0else if (CONVERT_EXPR_CODE_P (rhs_code))
> + =A0 =A0return false;

That's funny control flow.

> + =A0else if (rhs_code =3D=3D BIT_AND_EXPR || rhs_code =3D=3D BIT_IOR_EXPR
> + =A0 =A0 =A0|| rhs_code =3D=3D BIT_XOR_EXPR || rhs_code =3D=3D TRUTH_AND=
_EXPR
> + =A0 =A0 =A0|| rhs_code =3D=3D TRUTH_OR_EXPR || rhs_code =3D=3D TRUTH_XO=
R_EXPR
> + =A0 =A0 =A0|| rhs_code =3D=3D NE_EXPR || rhs_code =3D=3D EQ_EXPR)
> + =A0 =A0op1 =3D gimple_assign_rhs2 (stmt);
> +
> + =A0/* ~X is only equivalent of !X, if type-precision is one and X has
> + =A0 =A0 an integral type. =A0*/
> + =A0if (rhs_code =3D=3D BIT_NOT_EXPR
> + =A0 =A0 =A0&& (!INTEGRAL_TYPE_P (TREE_TYPE (op0))
> + =A0 =A0 =A0 =A0 || TYPE_PRECISION (TREE_TYPE (op0)) !=3D 1))
> + =A0 =A0return false;
> +
> =A0 if (TYPE_PRECISION (TREE_TYPE (op0)) !=3D 1)
> =A0 =A0 {
> =A0 =A0 =A0 if (TREE_CODE (op0) !=3D SSA_NAME)
> @@ -6775,22 +6867,100 @@ simplify_truth_ops_using_ranges (gimple_
> =A0 =A0 =A0 =A0 return false;
> =A0 =A0 }
>
> - =A0if (rhs_code =3D=3D TRUTH_NOT_EXPR)
> + =A0if (op1 && TREE_CODE (op1) !=3D INTEGER_CST
> + =A0 =A0 =A0&& TYPE_PRECISION (TREE_TYPE (op1)) !=3D 1)
> + =A0 =A0{
> + =A0 =A0 =A0vr =3D get_value_range (op1);
> + =A0 =A0 =A0val =3D compare_range_with_value (GE_EXPR, vr, integer_zero_=
node, &sop);
> + =A0 =A0 =A0if (!val || !integer_onep (val))
> + =A0 =A0 =A0 return false;
> +
> + =A0 =A0 =A0val =3D compare_range_with_value (LE_EXPR, vr, integer_one_n=
ode, &sop);
> + =A0 =A0 =A0if (!val || !integer_onep (val))
> + =A0 =A0 =A0 return false;
> + =A0 =A0}
> +
> + =A0need_conversion =3D
> + =A0 =A0!useless_type_conversion_p (TREE_TYPE (gimple_assign_lhs (stmt)),
> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 TREE_TYPE (=
op0));
> +
> + =A0/* As comparisons X !=3D 0 getting folded by prior pass to (bool) X,
> + =A0 =A0 but X =3D=3D 0 might be not folded for none boolean type of X
> + =A0 =A0 to (bool) (X ^ 1).
> + =A0 =A0 So for bitwise-binary operations we have three cases to handle:
> + =A0 =A0 a) ((bool) X) op ((bool) Y)
> + =A0 =A0 b) ((bool) X) op (Y =3D=3D 0) OR (X =3D=3D 0) op ((bool) Y)
> + =A0 =A0 c) (X =3D=3D 0) op (Y =3D=3D 0)
> + =A0 =A0 The later two cases can't be handled for now, as vr tables
> + =A0 =A0 would need to be adjusted. =A0*/
> + =A0if (need_conversion
> + =A0 =A0 =A0&& (rhs_code =3D=3D BIT_XOR_EXPR
> + =A0 =A0 =A0 =A0 || rhs_code =3D=3D BIT_AND_EXPR
> + =A0 =A0 =A0 =A0 || rhs_code =3D=3D BIT_IOR_EXPR)
> + =A0 =A0 =A0&& TREE_CODE (op1) =3D=3D SSA_NAME && TREE_CODE (op0) =3D=3D=
 SSA_NAME)
> + =A0 =A0{
> + =A0 =A0 =A0cop0 =3D ssa_name_get_cast_to_p (op0);
> + =A0 =A0 =A0cop1 =3D ssa_name_get_cast_to_p (op1);
> + =A0 =A0 =A0if (!cop0 || !cop1)
> + =A0 =A0 =A0 =A0/* We would need an new statment for cases b and c, and =
we can't
> + =A0 =A0 =A0 =A0 =A0 due vr table, so bail out. =A0*/
> + =A0 =A0 =A0 =A0return false;
> +
> + =A0 =A0 =A0if (!INTEGRAL_TYPE_P (TREE_TYPE (cop0))
> + =A0 =A0 =A0 =A0 || !types_compatible_p (TREE_TYPE (cop0), TREE_TYPE (co=
p1)))
> + =A0 =A0 =A0 return false;
> + =A0 =A0 =A0need_conversion =3D
> + =A0 =A0 =A0 !useless_type_conversion_p (TREE_TYPE (gimple_assign_lhs (s=
tmt)),
> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 TRE=
E_TYPE (cop0));
> + =A0 =A0 =A0if (need_conversion)
> + =A0 =A0 =A0 return false;
> + =A0 =A0 =A0op0 =3D cop0;
> + =A0 =A0 =A0op1 =3D cop1;
> +
> + =A0 =A0 =A0/* We need to re-check if value ranges for new operands
> + =A0 =A0 =A0 =A0 for 1-bit precision/range. =A0*/
> + =A0 =A0 =A0if (TYPE_PRECISION (TREE_TYPE (op0)) !=3D 1)
> + =A0 =A0 =A0 {
> + =A0 =A0 =A0 =A0 if (TREE_CODE (op0) !=3D SSA_NAME)
> + =A0 =A0 =A0 =A0 =A0 return false;
> + =A0 =A0 =A0 =A0 vr =3D get_value_range (op0);
> +
> + =A0 =A0 =A0 =A0 val =3D compare_range_with_value (GE_EXPR, vr, integer_=
zero_node, &sop);
> + =A0 =A0 =A0 =A0 if (!val || !integer_onep (val))
> + =A0 =A0 =A0 =A0 =A0 return false;
> +
> + =A0 =A0 =A0 =A0 val =3D compare_range_with_value (LE_EXPR, vr, integer_=
one_node, &sop);
> + =A0 =A0 =A0 =A0 if (!val || !integer_onep (val))
> + =A0 =A0 =A0 =A0 =A0 return false;
> + =A0 =A0 =A0 }
> +
> + =A0 =A0 =A0if (op1 && TYPE_PRECISION (TREE_TYPE (op1)) !=3D 1)
> + =A0 =A0 =A0 {
> + =A0 =A0 =A0 =A0 vr =3D get_value_range (op1);
> + =A0 =A0 =A0 =A0 val =3D compare_range_with_value (GE_EXPR, vr, integer_=
zero_node, &sop);
> + =A0 =A0 =A0 =A0 if (!val || !integer_onep (val))
> + =A0 =A0 =A0 =A0 =A0 return false;
> +
> + =A0 =A0 =A0 =A0 val =3D compare_range_with_value (LE_EXPR, vr, integer_=
one_node, &sop);
> + =A0 =A0 =A0 =A0 if (!val || !integer_onep (val))
> + =A0 =A0 =A0 =A0 =A0 return false;
> + =A0 =A0 =A0 }
> + =A0 =A0}
> + =A0else if (rhs_code =3D=3D TRUTH_NOT_EXPR || rhs_code =3D=3D BIT_NOT_E=
XPR)
> =A0 =A0 {
> =A0 =A0 =A0 rhs_code =3D NE_EXPR;
> =A0 =A0 =A0 op1 =3D build_int_cst (TREE_TYPE (op0), 1);
> =A0 =A0 }
> =A0 else
> =A0 =A0 {
> - =A0 =A0 =A0op1 =3D gimple_assign_rhs2 (stmt);
> -
> =A0 =A0 =A0 /* Reduce number of cases to handle. =A0*/
> =A0 =A0 =A0 if (is_gimple_min_invariant (op1))
> =A0 =A0 =A0 =A0{
> =A0 =A0 =A0 =A0 =A0 /* Exclude anything that should have been already fol=
ded. =A0*/
> =A0 =A0 =A0 =A0 =A0if (rhs_code !=3D EQ_EXPR
> =A0 =A0 =A0 =A0 =A0 =A0 =A0&& rhs_code !=3D NE_EXPR
> - =A0 =A0 =A0 =A0 =A0 =A0 && rhs_code !=3D TRUTH_XOR_EXPR)
> + =A0 =A0 =A0 =A0 =A0 =A0 && rhs_code !=3D TRUTH_XOR_EXPR
> + =A0 =A0 =A0 =A0 =A0 =A0 && rhs_code !=3D BIT_XOR_EXPR)
> =A0 =A0 =A0 =A0 =A0 =A0return false;
>
> =A0 =A0 =A0 =A0 =A0if (!integer_zerop (op1)
> @@ -6810,18 +6980,6 @@ simplify_truth_ops_using_ranges (gimple_
> =A0 =A0 =A0 =A0 =A0/* Punt on A =3D=3D B as there is no BIT_XNOR_EXPR. =
=A0*/
> =A0 =A0 =A0 =A0 =A0if (rhs_code =3D=3D EQ_EXPR)
> =A0 =A0 =A0 =A0 =A0 =A0return false;
> -
> - =A0 =A0 =A0 =A0 if (TYPE_PRECISION (TREE_TYPE (op1)) !=3D 1)
> - =A0 =A0 =A0 =A0 =A0 {
> - =A0 =A0 =A0 =A0 =A0 =A0 vr =3D get_value_range (op1);
> - =A0 =A0 =A0 =A0 =A0 =A0 val =3D compare_range_with_value (GE_EXPR, vr, =
integer_zero_node, &sop);
> - =A0 =A0 =A0 =A0 =A0 =A0 if (!val || !integer_onep (val))
> - =A0 =A0 =A0 =A0 =A0 =A0 =A0 return false;
> -
> - =A0 =A0 =A0 =A0 =A0 =A0 val =3D compare_range_with_value (LE_EXPR, vr, =
integer_one_node, &sop);
> - =A0 =A0 =A0 =A0 =A0 =A0 if (!val || !integer_onep (val))
> - =A0 =A0 =A0 =A0 =A0 =A0 =A0 return false;
> - =A0 =A0 =A0 =A0 =A0 }
> =A0 =A0 =A0 =A0}
> =A0 =A0 }
>
> @@ -6838,7 +6996,8 @@ simplify_truth_ops_using_ranges (gimple_
> =A0 =A0 =A0 =A0 warning_at (location, OPT_Wstrict_overflow,
> =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0_("assuming signed overflow does n=
ot occur when "
> =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0"simplifying && or || to & or =
|"));
> - =A0 =A0 =A0else
> + =A0 =A0 =A0else if (rhs_code !=3D BIT_AND_EXPR && rhs_code !=3D BIT_IOR=
_EXPR
> + =A0 =A0 =A0 =A0 =A0 =A0 =A0&& rhs_code !=3D BIT_XOR_EXPR)
> =A0 =A0 =A0 =A0 warning_at (location, OPT_Wstrict_overflow,
> =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0_("assuming signed overflow does n=
ot occur when "
> =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0"simplifying =3D=3D, !=3D or !=
 to identity or ^"));
> @@ -6859,16 +7018,21 @@ simplify_truth_ops_using_ranges (gimple_
> =A0 =A0 case TRUTH_AND_EXPR:
> =A0 =A0 =A0 rhs_code =3D BIT_AND_EXPR;
> =A0 =A0 =A0 break;
> + =A0 =A0case BIT_AND_EXPR:
> + =A0 =A0 =A0break;
> =A0 =A0 case TRUTH_OR_EXPR:
> =A0 =A0 =A0 rhs_code =3D BIT_IOR_EXPR;
> + =A0 =A0case BIT_IOR_EXPR:
> =A0 =A0 =A0 break;
> =A0 =A0 case TRUTH_XOR_EXPR:
> + =A0 =A0case BIT_XOR_EXPR:
> =A0 =A0 case NE_EXPR:
> =A0 =A0 =A0 if (integer_zerop (op1))
> =A0 =A0 =A0 =A0{
> =A0 =A0 =A0 =A0 =A0gimple_assign_set_rhs_with_ops (gsi,
> =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =
=A0 =A0 =A0need_conversion ? NOP_EXPR : SSA_NAME,
> =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =
=A0 =A0 =A0op0, NULL);
> + =A0 =A0 =A0 =A0 gimple_set_location (stmt, loc);
> =A0 =A0 =A0 =A0 =A0update_stmt (gsi_stmt (*gsi));
> =A0 =A0 =A0 =A0 =A0return true;
> =A0 =A0 =A0 =A0}
> @@ -6879,10 +7043,20 @@ simplify_truth_ops_using_ranges (gimple_
> =A0 =A0 =A0 gcc_unreachable ();
> =A0 =A0 }
>
> + =A0/* We can't insert here new expression as otherwise
> + =A0 =A0 tracked vr tables getting out of bounds. =A0*/
> =A0 if (need_conversion)
> =A0 =A0 return false;
>
> + =A0/* Reduce here SSA_NAME -> SSA_NAME. =A0*/
> + =A0while ((cop0 =3D ssa_name_get_inner_ssa_name_p (op0)) !=3D NULL_TREE)
> + =A0 =A0op0 =3D cop0;
> +
> + =A0while ((cop1 =3D ssa_name_get_inner_ssa_name_p (op1)) !=3D NULL_TREE)
> + =A0 =A0op1 =3D cop1;
> +
> =A0 gimple_assign_set_rhs_with_ops (gsi, rhs_code, op0, op1);
> + =A0gimple_set_location (stmt, loc);
> =A0 update_stmt (gsi_stmt (*gsi));
> =A0 return true;
> =A0}
> @@ -7263,6 +7437,9 @@ simplify_switch_using_ranges (gimple stm
> =A0 tree vec2;
> =A0 switch_update su;
>
> + =A0if (in_second_pass)
> + =A0 =A0return false;
> +
> =A0 if (TREE_CODE (op) =3D=3D SSA_NAME)
> =A0 =A0 {
> =A0 =A0 =A0 vr =3D get_value_range (op);
> @@ -7390,6 +7567,7 @@ simplify_stmt_using_ranges (gimple_stmt_
> =A0 =A0 =A0 =A0{
> =A0 =A0 =A0 =A0case EQ_EXPR:
> =A0 =A0 =A0 =A0case NE_EXPR:
> + =A0 =A0 =A0 case BIT_NOT_EXPR:
> =A0 =A0 =A0 =A0case TRUTH_NOT_EXPR:
> =A0 =A0 =A0 =A0case TRUTH_AND_EXPR:
> =A0 =A0 =A0 =A0case TRUTH_OR_EXPR:
> @@ -7425,13 +7603,21 @@ simplify_stmt_using_ranges (gimple_stmt_
> =A0 =A0 =A0 =A0 =A0 =A0 if all the bits being cleared are already cleared=
 or
> =A0 =A0 =A0 =A0 =A0 =A0 all the bits being set are already set. =A0*/
> =A0 =A0 =A0 =A0 =A0if (INTEGRAL_TYPE_P (TREE_TYPE (rhs1)))
> - =A0 =A0 =A0 =A0 =A0 return simplify_bit_ops_using_ranges (gsi, stmt);
> + =A0 =A0 =A0 =A0 =A0 {
> + =A0 =A0 =A0 =A0 =A0 =A0 if (simplify_truth_ops_using_ranges (gsi, stmt))
> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 return true;
> + =A0 =A0 =A0 =A0 =A0 =A0 return simplify_bit_ops_using_ranges (gsi, stmt=
);
> + =A0 =A0 =A0 =A0 =A0 }
> =A0 =A0 =A0 =A0 =A0break;
>
> =A0 =A0 =A0 =A0CASE_CONVERT:
> =A0 =A0 =A0 =A0 =A0if (TREE_CODE (rhs1) =3D=3D SSA_NAME
> =A0 =A0 =A0 =A0 =A0 =A0 =A0&& INTEGRAL_TYPE_P (TREE_TYPE (rhs1)))
> - =A0 =A0 =A0 =A0 =A0 return simplify_conversion_using_ranges (stmt);
> + =A0 =A0 =A0 =A0 =A0 {
> + =A0 =A0 =A0 =A0 =A0 =A0 if (simplify_truth_ops_using_ranges (gsi, stmt))
> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 return true;
> + =A0 =A0 =A0 =A0 =A0 =A0 return simplify_conversion_using_ranges (stmt);
> + =A0 =A0 =A0 =A0 =A0 }
> =A0 =A0 =A0 =A0 =A0break;
>
> =A0 =A0 =A0 =A0default:
> @@ -7685,8 +7870,16 @@ vrp_finalize (void)
> =A0 =A0 =A0 fprintf (dump_file, "\n");
> =A0 =A0 }
>
> + =A0/* We redo folding here one time for allowing to inspect more
> + =A0 =A0 complex reductions. =A0*/
> + =A0substitute_and_fold (op_with_constant_singleton_value_range,
> + =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0vrp_fold_stmt, false);
> + =A0/* We need to mark this second pass to avoid re-entering of same
> + =A0 =A0 edges for switch statments. =A0*/
> + =A0in_second_pass =3D true;
> =A0 substitute_and_fold (op_with_constant_singleton_value_range,
> =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 vrp_fold_stmt, false);
> + =A0in_second_pass =3D false;

If at all you only want to re-call vrp_fold_stmt on all stmts in the
function, not do a full-blown substitute_and_fold.

Richard.

> =A0 if (warn_array_bounds)
> =A0 =A0 check_all_array_refs ();
>