New Jacks Regressions: Constant Folding of Casts

Wed Jul 21 05:13:00 GMT 2004

On 2004年7月20日 07:16:52 -0600 (MDT), Roger Sayle
<roger@eyesopen.com> wrote:
>> That's a relief :). It might be related to Steve Bosschers improvements
> to switch statement handling. The middle-end now does far less processing
> of switch/case statements, and expects the front-end to issue diagnostics,
> remove duplicate case nodes, etc... Perhaps the java front-end has been
> relying on the middle-end to call "fold" on case nodes for it. Could you
> check that "((int)Float.Nan == 0 ? 1 : 0" is being folded as its being
> constructed. Does "case (int)Float.Nan:" work?

No, "case (int )Float.NaN:" also doesn't work.
The Java front end does call fold() on case nodes.
Here's the relevant code from java/parse.y:
-------------------------- 8< --------------------------
 11533 /* First, the case expression must be constant. Values of final
 11534 fields are accepted. */
 11535 cn = fold (cn);
 11536 if ((TREE_CODE (cn) == COMPOUND_EXPR || TREE_CODE (cn)
== COMPONENT_REF)
 11537 && JDECL_P (TREE_OPERAND (cn, 1))
 11538 && FIELD_FINAL (TREE_OPERAND (cn, 1))
 11539 && DECL_INITIAL (TREE_OPERAND (cn, 1)))
 11540 {
 11541 cn = fold_constant_for_init (DECL_INITIAL
(TREE_OPERAND (cn, 1)),
 11542 TREE_OPERAND (cn, 1));
 11543 }
 11544 /* Accept final locals too. */
 11545 else if (TREE_CODE (cn) == VAR_DECL && DECL_FINAL (cn)
 11546 && DECL_INITIAL (cn))
 11547 cn = fold_constant_for_init (DECL_INITIAL (cn), cn);
 11548
 11549 if (!TREE_CONSTANT (cn) && !flag_emit_xref)
 11550 {
 11551 EXPR_WFL_LINECOL (wfl_operator) = EXPR_WFL_LINECOL (node);
 11552 parse_error_context (node, "Constant expression required");
 11553 return error_mark_node;
 11554 }
-------------------------- 8< --------------------------
Has this become invalid now?
And just for your benefit, the value of "cn" before
the first call to fold( ) (for the "case (int )Float.NaN:"
case) is:
-------------------------- 8< --------------------------
Breakpoint 1, java_complete_lhs (node=0x4001cd00) at parse.y:11535
11535 cn = fold (cn);
(gdb) p debug_tree(cn)
 <cond_expr 0x40220730
 type <integer_type 0x40020414 int SI
 size <integer_cst 0x4001e2d0 constant invariant 32>
 unit size <integer_cst 0x4001e330 constant invariant 4>
 align 32 symtab 0 alias set -1 precision 32 min <integer_cst
0x4001e300 -2147483648> max <integer_cst 0x4001e318 2147483647>
 pointer_to_this <pointer_type 0x40165488> chain <type_decl
0x40020488 int>>
 tree_6
 arg 0 <unlt_expr 0x4001f45c
 type <boolean_type 0x40026000 boolean unsigned QI
 size <integer_cst 0x4001e930 constant invariant 8>
 unit size <integer_cst 0x4001e8e8 constant invariant 1>
 align 8 symtab 0 alias set -1 precision 1 min <integer_cst
0x4001ea38 0> max <integer_cst 0x4001ea50 1>
 chain <type_decl 0x40026074 boolean>>
 constant invariant
 arg 0 <real_cst 0x40169ae0 type <real_type 0x40026488 float>
 constant invariant Nan>
 arg 1 <real_cst 0x4021f348 type <real_type 0x40026488 float>
 constant invariant 2.147483648e+9>>
 arg 1 <cond_expr 0x402206e0 type <integer_type 0x40020414 int>
 arg 0 <le_expr 0x4001f414 type <boolean_type 0x40026000 boolean>
 constant invariant arg 0 <real_cst 0x40169ae0 Nan>
 arg 1 <real_cst 0x4021f318 type <real_type 0x40026488 float>
 constant invariant -2.147483648e+9>>
 arg 1 <integer_cst 0x4001e300 constant invariant -2147483648>
 arg 2 <cond_expr 0x402206b8 type <integer_type 0x40020414 int>
 arg 0 <ne_expr 0x4001f3f0 type <boolean_type 0x40026000 boolean>
 constant invariant arg 0 <real_cst 0x40169ae0 Nan> arg
1 <real_cst 0x40169ae0 Nan>>
 arg 1 <integer_cst 0x4001e840 constant invariant 0>
 arg 2 <fix_trunc_expr 0x4021d400 type <integer_type 0x40020414 int>
 constant invariant arg 0 <real_cst 0x40169ae0 Nan>>>>
 arg 2 <integer_cst 0x4001e318 type <integer_type 0x40020414 int>
constant invariant 2147483647>>
-------------------------- 8< --------------------------
Thanks,
Ranjit.
-- 
Ranjit Mathew Email: rmathew AT gmail DOT com
Bangalore, INDIA. Web: http://ranjitmathew.tripod.com/