Patches to remove TREE_COMPLEXITY for C and C++

["Joseph S. Myers" <jsm at polyomino dot org dot uk>]

This pair of patches removes the use of TREE_COMPLEXITY for C and C++.
It doesn't get rid of TREE_COMPLEXITY altogether, since I hadn't
realised that Java was still using it but using the complexity field
directly rather than through the TREE_COMPLEXITY accessor.  Thus, Java
and Ada are left as the only users of this field.

The first patch is Steven Bosscher's patch
<http://gcc.gnu.org/ml/gcc-patches/2004-06/msg00863.html> for uses in
C++, updated to apply to current mainline.

The second patch removes TREE_COMPLEXITY from the C front end and
makes the necessary associated changes to the C++ front end to bring
-Wparentheses back to an approximately working state (it having been
broken by the new C++ parser).  This is the one for which the log is
below.

The approach used for most of the C front end -Wparentheses warnings
is that I previously proposed (references in
<http://gcc.gnu.org/ml/gcc-patches/2004-06/msg01219.html> in the
discussion of the first patch).  The parser structures are intended
incrementally to move further into the front end.  In particular
implementation of proper constant expression rules will follow once
I've written down the necessary four formal models for what should be
implemented for C90, C99 and the GNU extended versions thereof.  The
approach for warnings for assignments used as truth-values is
TREE_NO_WARNING as in the first patch.  For C++ this is clearly
imperfect as assignments can yield many sorts of tree other than
MODIFY_EXPR and so not trigger the warning.  At present a proper
implementation for C++ would require moving the warning out of
c_common_truthvalue_conversion into language-specific code for each
front end, called at the syntactically appropriate points.

This pair of patches together has been bootstrapped with no
regressions on i686-pc-linux-gnu.  There is a 0.6% compile-time
improvement for building a set of .i files of cc1 at -O0 (note this
timing is with --enable-checking compilers), probably because
parenthesised expressions and expression statements no longer all get
encapsulated in garbage TREE_LISTs.  I don't know what further
improvement getting rid of the complexity field from tree_exp will
bring.

Accordingly I request review of (a) patch 1, so Steven Bosscher can
apply that; (b) the C++ and Objective-C parts of patch 2.  If there
are objections to the C++ changes, I'll rework things so that
c-common.c doesn't attempt to give the warning for assignments used as
truth-values for C++ (the status quo being that C_EXP_ORIGINAL_CODE is
never set to MODIFY_EXPR by the current C++ parser, so this warning
never happens for C++) so the C++ front end doesn't need changes
beyond removing all references to TREE_COMPLEXITY /
C_EXP_ORIGINAL_CODE, and file a bug report for the regression in
-Wparentheses support in C++ to record the issue.  (I should probably
file such a bug report anyway for the regression in 3.4.)

-- 
Joseph S. Myers               http://www.srcf.ucam.org/~jsm28/gcc/
    jsm@polyomino.org.uk (personal mail)
    jsm28@gcc.gnu.org (Bugzilla assignments and CCs)

2004-07-18  Joseph S. Myers  <jsm@polyomino.org.uk>

	* c-tree.h (struct c_expr): Define.
	(C_SET_EXP_ORIGINAL_CODE): Remove.
	(parser_build_binary_op, build_compound_expr): Update prototypes.
	* c-parse.in (%union): Add exprtype.
	(FUNC_NAME): Mark as ttype.
	(expr, expr_no_commas, cast_expr, unary_expr, primary): Change to
	exprtype.
	(expr): Update.  Define directly in terms of expr_no_commas
	instead of using nonnull_exprlist.
	(nonnull_exprlist, unary_expr, cast_expr, expr_no_commas, primary,
	offsetof_member_designator, typespec_nonreserved_nonattr, init,
	initval, designator, component_declarator,
	component_notype_declarator, enumerator, array_declarator,
	condition, exexpr, switch_statement, stmt_nocomp, stmt,
	nonnull_asm_operands, ivar_declarator, receiver): Update.  Don't
	set C_EXP_ORIGINAL_CODE.  Use TREE_NO_WARNING for assignments
	where appropriate.
	* c-common.h (C_EXP_ORIGINAL_CODE): Remove.
	* c-common.c (c_common_truthvalue_conversion): Don't check
	C_EXP_ORIGINAL_CODE.
	* c-typeck.c (parser_build_binary_op): Use c_expr structures.
	Don't use C_EXP_ORIGINAL_CODE.
	(default_conversion, default_function_array_conversion): Don't use
	C_EXP_ORIGINAL_CODE.  Preserve TREE_NO_WARNING.
	(internal_build_compound_expr): Merge into build_compound_expr.
	(build_compound_expr): Take two operands instead of a TREE_LIST.
	* objc/objc-act.c (get_super_receiver): Update calls to
	build_compound_expr.

cp:
2004-07-18  Joseph S. Myers  <jsm@polyomino.org.uk>

	* typeck.c (build_modify_expr, build_x_modify_expr): Set
	TREE_NO_WARNING on assignments with an operator other than '='.

testsuite:
2004-07-18  Joseph S. Myers  <jsm@polyomino.org.uk>

	* g++.dg/warn/Wparentheses-1.C, g++.dg/warn/Wparentheses-2.C,
	gcc.dg/Wparentheses-10.c: New tests.
	* gcc.dg/Wparentheses-5.c: Remove XFAILs.

Attachment: bosscher-updated-patch
Description: Text document

Attachment: cexp-patch
Description: Text document