On Wed, Oct 31, 2012 at 7:14 PM, H.J. Lu <hjl.to...@gmail.com> wrote:
> On Wed, Oct 31, 2012 at 6:57 PM, H.J. Lu <hjl.to...@gmail.com> wrote:
>> On Wed, Oct 31, 2012 at 3:53 PM, H.J. Lu <hjl.to...@gmail.com> wrote:
>>> On Wed, Oct 31, 2012 at 7:29 AM, Eric Botcazou <ebotca...@adacore.com>
>>> wrote:
>>>>
>>>> > It failed with revision 188008.
>>>>
>>>> OK, thanks. So the testcase never compiled on the trunk (except for about
>>>> 24
>>>> hours between 188009 & 188118) or did it compile before 188008 at some
>>>> point?
>>>>
>>>> --
>>>> Eric Botcazou
>>>
>>> It was OK until revision 187042:
>>>
>>> 2012-05-02 Richard Guenther <rguent...@suse.de>
>>>
>>> * tree.c (valid_constant_size_p): New function.
>>> * tree.h (valid_constant_size_p): Declare.
>>> * cfgexpand.c (expand_one_var): Adjust check for too large
>>> variables by using valid_constant_size_p.
>>> * varasm.c (assemble_variable): Likewise.
>>>
>>> c/
>>> * c-decl.c (grokdeclarator): Properly check for sizes that
>>> cover more than half of the address-space.
>>>
>>> cp/
>>> * decl.c (grokdeclarator): Properly check for sizes that
>>> cover more than half of the address-space.
>>>
>>> 2012-05-02 Richard Guenther <rguent...@suse.de>
>>>
>>> * fold-const.c (div_if_zero_remainder): sizetypes no longer
>>> sign-extend.
>>> (int_const_binop_1): New worker for int_const_binop with
>>> overflowable parameter. Pass it through
>>> to force_fit_type_double.
>>> (int_const_binop): Wrap around int_const_binop_1 with overflowable
>>> equal to one.
>>> (size_binop_loc): Call int_const_binop_1 with overflowable equal
>>> to minus one, forcing overflow detection for even unsigned types.
>>> (extract_muldiv_1): Remove bogus TYPE_IS_SIZETYPE special-casing.
>>> (fold_binary_loc): Call try_move_mult_to_index with signed offset.
>>> * stor-layout.c (initialize_sizetypes): sizetypes no longer
>>> sign-extend.
>>> (layout_type): For zero-sized arrays ignore overflow on the
>>> size calculations.
>>> * tree-ssa-ccp.c (bit_value_unop_1): Likewise.
>>> (bit_value_binop_1): Likewise.
>>> * tree.c (double_int_to_tree): Likewise.
>>> (double_int_fits_to_tree_p): Likewise.
>>> (force_fit_type_double): Likewise.
>>> (host_integerp): Likewise.
>>> (int_fits_type_p): Likewise.
>>> * varasm.c (output_constructor_regular_field): Sign-extend the
>>> field-offset to cater for negative offsets produced by the Ada
>>> frontend.
>>> * omp-low.c (extract_omp_for_data): Convert the loop step to
>>> signed for pointer adjustments.
>>>
>>> which gave:
>>>
>>> /export/build/gnu/gcc/build-x86_64-linux/gcc/xgcc
>>> -B/export/build/gnu/gcc/build-x86_64-linux/gcc/ -O3 -Wall -mx32
>>> -maddress-mode=long -fPIC -S x.i
>>> x.i: In function ‘dl_start’:
>>> x.i:30:1: error: unrecognizable insn:
>>> }
>>> ^
>>> (insn 54 53 55 7 (set (reg:SI 108)
>>> (const:SI (plus:SI (symbol_ref:SI ("_dl_rtld_map") [flags
>>> 0x42] <var_decl 0x7f02997bb140 _dl_rtld_map>)
>>> (const_int -1073742800 [0xffffffffbffffc30])))) x.i:22 -1
>>> (nil))
>>> x.i:30:1: internal compiler error: in extract_insn, at recog.c:2130
>>> }
>>> ^
>>> Please submit a full bug report,
>>> with preprocessed source if appropriate.
>>> See <http://gcc.gnu.org/bugs.html> for instructions.
>>>
>>
>> It looks like we are using unsigned type for array index:
>>
>> (gdb) list
>> 2163 tree element = TREE_TYPE (type);
>> 2164
>> 2165 build_pointer_type (element);
>> 2166
>> 2167 /* We need to know both bounds in order to compute the size.
>> */
>> 2168 if (index && TYPE_MAX_VALUE (index) && TYPE_MIN_VALUE (index)
>> 2169 && TYPE_SIZE (element))
>> 2170 {
>> 2171 tree ub = TYPE_MAX_VALUE (index);
>> 2172 tree lb = TYPE_MIN_VALUE (index);
>> (gdb) call debug_tree (index)
>> <integer_type 0x7ffff1ab6000
>> type <integer_type 0x7ffff199d000 sizetype public unsigned sizetype SI
>> size <integer_cst 0x7ffff1989d40 constant 32>
>> unit size <integer_cst 0x7ffff1989d60 constant 4>
>> align 32 symtab 0 alias set -1 canonical type 0x7ffff199d000
>> precision 32 min <integer_cst 0x7ffff1989d80 0> max <integer_cst
>> 0x7ffff1989000 4294967295>>
>> SI size <integer_cst 0x7ffff1989d40 32> unit size <integer_cst
>> 0x7ffff1989d60 4>
>> align 32 symtab 0 alias set -1 canonical type 0x7ffff1ab6000
>> precision 32 min <integer_cst 0x7ffff1989d80 0> max <integer_cst
>> 0x7ffff1aa3260 33>>
>> (gdb)
>>
>
> size type is changed to unsigned:
>
> - /* Size types *are* sign extended. */
> - bool sign_extended_type = (!TYPE_UNSIGNED (type)
> - || (TREE_CODE (type) == INTEGER_TYPE
> - && TYPE_IS_SIZETYPE (type)));
> + bool sign_extended_type = !TYPE_UNSIGNED (type);
>
> But we use size type in places where signed size type
> should be used.
>
For example, array index computation must be signed,
not unsigned.
--
H.J.
