Testcase, compile at -O2. A clone "g.clone.0" is formed which is the same as "g", only the second argument has been dropped. The bug is that in the CALL_EXPR the type of the called function is still the original type. // begin testcase static __attribute__((noinline)) int g(int i, int j) { if (j != 0) return 0; return i; } int f(int i) { return g(i, 0); } // end testcase Breaking on verify_gimple_call, and waiting until after "Assembling functions: g.clone.0 float" is printed, notice how the type of the addr_expr is a pointer to "type <function_type 0x7ffff70c2540 type <integer_type 0x7ffff7e8c540 int>", while the type of g.clone.0 (see arg 0) is "type <function_type 0x7ffff70c2e40 type <integer_type 0x7ffff7e8c540 int>", and these two types do not match, though they should. (gdb) call debug_tree(fn) <addr_expr 0x7ffff7ff0b00 type <pointer_type 0x7ffff70c26c0 type <function_type 0x7ffff70c2540 type <integer_type 0x7ffff7e8c540 int> QI size <integer_cst 0x7ffff7e7b7e0 constant 8> unit size <integer_cst 0x7ffff7e7b810 constant 1> align 8 symtab 0 alias set -1 canonical type 0x7ffff70c2540 arg-types <tree_list 0x7ffff70da3c0 value <integer_type 0x7ffff7e8c540 int> chain <tree_list 0x7ffff70da390 value <integer_type 0x7ffff7e8c540 int> chain <tree_list 0x7ffff7e999f0 value <void_type 0x7ffff7e9b3c0 void>>>> pointer_to_this <pointer_type 0x7ffff70c26c0>> unsigned DI size <integer_cst 0x7ffff7e7bb70 constant 64> unit size <integer_cst 0x7ffff7e7bba0 constant 8> align 64 symtab 0 alias set -1 canonical type 0x7ffff70c26c0> readonly constant arg 0 <function_decl 0x7ffff7123200 g.clone.0 type <function_type 0x7ffff70c2e40 type <integer_type 0x7ffff7e8c540 int> QI size <integer_cst 0x7ffff7e7b7e0 8> unit size <integer_cst 0x7ffff7e7b810 1> align 8 symtab 0 alias set -1 canonical type 0x7ffff70c2540 arg-types <tree_list 0x7ffff70daa20 value <integer_type 0x7ffff7e8c540 int> chain <tree_list 0x7ffff7e999f0>> pointer_to_this <pointer_type 0x7ffff70c26c0>> readonly addressable asm_written used nothrow static decl_5 QI file clone.c line 1 col 38 align 8 attributes <tree_list 0x7ffff70da3f0> initial <error_mark 0x7ffff7e86e60> abstract_origin <function_decl 0x7ffff70c1c00 g> arguments <parm_decl 0x7ffff7e84d80 i type <integer_type 0x7ffff7e8c540 int> used SI file clone.c line 1 col 44 size <integer_cst 0x7ffff7e7ba80 constant 32> unit size <integer_cst 0x7ffff7e7b6f0 constant 4> align 32 context <function_decl 0x7ffff7123200 g.clone.0> abstract_origin <parm_decl 0x7ffff7e846c0 i> (reg/v:SI 0 ax [orig:59 i ] [59]) arg-type <integer_type 0x7ffff7e8c540 int> incoming-rtl (reg:SI 5 di [ i ])> result <result_decl 0x7ffff70db100 D.2707 type <integer_type 0x7ffff7e8c540 int> used ignored regdecl SI file clone.c line 1 col 38 size <integer_cst 0x7ffff7e7ba80 32> unit size <integer_cst 0x7ffff7e7b6f0 4> align 32 context <function_decl 0x7ffff7123200 g.clone.0> abstract_origin <result_decl 0x7ffff70db000 D.1606> (reg:SI 58 [ <retval> ])> (mem:QI (symbol_ref:DI ("g.clone.0") [flags 0x3] <function_decl 0x7ffff7123200 g.clone.0>) [0 S1 A8])> clone.c:5:16>
It turns out that the problem is that when build_function_type_skip_args creates the new type, TYPE_POINTER_TO for the new type is still pointing to the old type. When gimple_call_set_fndecl is used to change the fndecl to one with less arguments, an ADDR_EXPR is built. The type of the ADDR_EXPR is obtained from the TYPE_POINTER_TO field of the new type, and thus has type pointer-to-old-type rather than pointer-to-new-type. I'm testing the following patch. The idea here is that calling copy_node has no advantage. In spite of the comment, build_distinct_type_copy preserves attributes etc too, so it might as well be used always. [I think the comment refers to some earlier logic, since the change introducing build_distinct_type_copy post-dates it]. Index: gcc-4.5/gcc/tree.c =================================================================== --- gcc-4.5.orig/gcc/tree.c 2010-07-05 20:26:21.349376337 +0200 +++ gcc-4.5/gcc/tree.c 2010-07-05 20:29:11.582730610 +0200 @@ -7208,24 +7208,10 @@ new_reversed = void_list_node; } - /* Use copy_node to preserve as much as possible from original type - (debug info, attribute lists etc.) - Exception is METHOD_TYPEs must have THIS argument. - When we are asked to remove it, we need to build new FUNCTION_TYPE - instead. */ - if (TREE_CODE (orig_type) != METHOD_TYPE - || !bitmap_bit_p (args_to_skip, 0)) - { - new_type = copy_node (orig_type); - TYPE_ARG_TYPES (new_type) = new_reversed; - } - else - { - new_type - = build_distinct_type_copy (build_function_type (TREE_TYPE (orig_type), - new_reversed)); - TYPE_CONTEXT (new_type) = TYPE_CONTEXT (orig_type); - } + new_type + = build_distinct_type_copy (build_function_type (TREE_TYPE (orig_type), + new_reversed)); + TYPE_CONTEXT (new_type) = TYPE_CONTEXT (orig_type); /* This is a new type, not a copy of an old type. Need to reassociate variants. We can handle everything except the main variant lazily. */
Subject: Re: Type of ADDR_EXPR in CALL_EXPR not rebuilt when function is cloned > It turns out that the problem is that when build_function_type_skip_args > creates > the new type, TYPE_POINTER_TO for the new type is still pointing to the old > type. > When gimple_call_set_fndecl is used to change the fndecl to one with less > arguments, an ADDR_EXPR is built. The type of the ADDR_EXPR is obtained from > the > TYPE_POINTER_TO field of the new type, and thus has type pointer-to-old-type > rather than pointer-to-new-type. I'm testing the following patch. The idea > here is that calling copy_node has no advantage. In spite of the comment, The conditional you removed is about case where METHOD_TYPE is passed and "this" pointer is not removed. In this case we still want to produce METHOD_TYPE just with args adjusted. I think this functionality should be preserved. Are attributes attached to decl (not type) also copied? Honza
Hi Honza, my original patch was silly, I'm trying this instead: @@ -7216,7 +7216,7 @@ if (TREE_CODE (orig_type) != METHOD_TYPE || !bitmap_bit_p (args_to_skip, 0)) { - new_type = copy_node (orig_type); + new_type = build_distinct_type_copy (orig_type); TYPE_ARG_TYPES (new_type) = new_reversed; } else
Subject: Re: Type of ADDR_EXPR in CALL_EXPR not rebuilt when function is cloned > Hi Honza, my original patch was silly, I'm trying this instead: This seems fine, thanks! Honza
Even better, it actually works! :)
Fixed in commits 161918 (mainline) and 161919 (gcc-4.5 branch).
> Fixed in commits 161918 (mainline) and 161919 (gcc-4.5 branch). The right procedure is to copy-and-paste the ChangeLog entry in the commit log, this will automatically add a cross-reference with an URL to this audit trail.