Re: Possible regression in DF analysis

[Richard Biener <richard.guenther@gmail.com>]

> Am 13.12.2022 um 17:54 schrieb Claudiu Zissulescu Ianculescu via Gcc <gcc@gcc.gnu.org>:
> 
> UPDATE:
> The df_analyze_loop is calling the df_set_blocks. Thus, the analysis
> behaves as if the function only contains those blocks and any edges
> that occur directly between the blocks in the set (see df-core.cc).
> This said, the loop-doloop behaves faulty at loop-doloop.cc:772 as the
> df_get_lives_out (loop_end) is not computed correctly.
> 
> A possible solution is to include in the blocks_to_analyze the missing
> basic blocks, something like:
> 
> diff --git a/gcc/df-core.cc b/gcc/df-core.cc
> index a901b84878f..d7059c188b2 100644
> --- a/gcc/df-core.cc
> +++ b/gcc/df-core.cc
> @@ -1437,7 +1437,15 @@ df_analyze_loop (class loop *loop)
>   df_set_blocks (blocks);
>   BITMAP_FREE (blocks);
> 
> -  df_analyze_1 ();
> +  /* Add the loop's header successor bbs too.  */
> +  edge e;
> +  edge_iterator ei;
> +  FOR_EACH_EDGE (e, ei, loop->header->succs)
> +    bitmap_set_bit (df->blocks_to_analyze, e->dest->index);
> +
> +  if (dump_file)
> +    debug_bitmap_file (dump_file, df->blocks_to_analyze);
> +  df_analyze ();
> }
> 
> What do you think,

Maybe you want to iterate over the loops exit edges and include their destination block instead?

> Claudiu
> 
>> On Tue, Dec 13, 2022 at 2:30 PM Claudiu Zissulescu Ianculescu
>> <claziss@gmail.com> wrote:
>> 
>> It looks like that. The df_analyze_loop is only looking at the loop
>> BBs, and it is not clear for me if df_analyze_loop is required to have
>> all the df_live_outs correctly computed or not. Do you know if it is
>> true?
>> 
>> If the df_analyze_loop is not supposed to compute all the df_live_outs
>> correctly, then the error resides in how loop-doloop is using the
>> iv_analysis_loop_init().
>> 
>> Thank you for your help,
>> Claudiu
>> 
>>> On Tue, Dec 13, 2022 at 10:41 AM Eric Botcazou <botcazou@adacore.com> wrote:
>>> 
>>>> The problem shows in loop-doloop.c when I introduce a loop end pattern
>>>> that replaces the first jump instruction (JUMP_INSN 15) with a pattern
>>>> that clobbers CC reg. However, the DF doesn't look like it works as
>>>> the doloop step cannot find the CC reg alive. Please see
>>>> loop-doloop.c:766. Hence, it introduces the doloop_end patterns, and
>>>> renders the compare instruction (INSN 14) dead code. leading to
>>>> errors.
>>> 
>>> So df_get_live_out does not contain the CC register?  iv_analysis_loop_init
>>> only performs a local update of the DF information, maybe it does not cover
>>> the basic block containing insn 14 and 15?
>>> 
>>> --
>>> Eric Botcazou
>>> 
>>>