On 10/13/2017 07:02 AM, Martin Liška wrote:
> On 10/12/2017 11:54 PM, Jeff Law wrote:
>> On 10/11/2017 12:13 AM, Martin Liška wrote:
>>> 2017-10-10 Martin Liska <mli...@suse.cz>
>>>
>>> PR tree-optimization/82493
>>> * sbitmap.c (bitmap_bit_in_range_p): Fix the implementation.
>>> (test_range_functions): New function.
>>> (sbitmap_c_tests): Likewise.
>>> * selftest-run-tests.c (selftest::run_tests): Run new tests.
>>> * selftest.h (sbitmap_c_tests): New function.
>> I went ahead and committed this along with a patch to fix the off-by-one
>> error in live_bytes_read. Bootstrapped and regression tested on x86.
>>
>> Actual patch attached for archival purposes.
>>
>> Jeff
>>
>
> Hello.
>
> I wrote a patch that adds various gcc_checking_asserts and I hit following:
>
> ./xgcc -B.
> /home/marxin/Programming/gcc/gcc/testsuite/gfortran.dg/char_result_12.f90 -c
> -O2
> during GIMPLE pass: dse
> /home/marxin/Programming/gcc/gcc/testsuite/gfortran.dg/char_result_12.f90:7:0:
>
> program testat
>
> internal compiler error: in bitmap_check_index, at sbitmap.h:105
> 0x1c014c1 bitmap_check_index
> ../../gcc/sbitmap.h:105
> 0x1c01fa7 bitmap_bit_in_range_p(simple_bitmap_def const*, unsigned int,
> unsigned int)
> ../../gcc/sbitmap.c:335
> 0x1179002 live_bytes_read
> ../../gcc/tree-ssa-dse.c:497
> 0x117935a dse_classify_store
> ../../gcc/tree-ssa-dse.c:595
> 0x1179947 dse_dom_walker::dse_optimize_stmt(gimple_stmt_iterator*)
> ../../gcc/tree-ssa-dse.c:786
> 0x1179b6e dse_dom_walker::before_dom_children(basic_block_def*)
> ../../gcc/tree-ssa-dse.c:853
> 0x1a6f659 dom_walker::walk(basic_block_def*)
> ../../gcc/domwalk.c:308
> 0x1179cb9 execute
> ../../gcc/tree-ssa-dse.c:907
>
> Where we call:
> Breakpoint 1, bitmap_bit_in_range_p (bmap=0x29d6cd0, start=0, end=515) at
> ../../gcc/sbitmap.c:335
> 335 bitmap_check_index (bmap, end);
> (gdb) p *bmap
> $1 = {n_bits = 256, size = 4, elms = {255}}
>
> Is it a valid call or should caller check indices?
It doesn't look valid to me. I'll dig into it.
In general the sbitmap interface requires callers to DTRT -- failure can
easily lead to an out of bounds read or write. It's one of the things I
really dislike about the sbitmap implementation.
So it's safe to assume that I'm fully supportive of adding more testing
to catch this kind thing.
Jeff
