[debug-early] equate new DIE with DW_AT_specificationto a previous declaration
diff mbox

Message ID 5508874D.70008@redhat.com
State New
Headers show

Commit Message

Aldy Hernandez March 17, 2015, 7:58 p.m. UTC
Hi Jason.

I merged mainline into the debug-early branch and I ran into a problem 
compiling a libstdc++ file with -fno-implicit-templates.  The attached 
patch is what I used to solve the problem but I wanted to run it by you, 
to make sure I'm not overlooking something silly, or worse...something 
much more complicated.

The reduced testcase is the following, compiled with 
-fno-implicit-templates -g -O2 -std=gnu++11:

class Object
{
public:
   void Method();
};

void funky()
{
   Object<int> foobar;
   foobar.Method();
}

template<typename SomeType>
void
Object<SomeType>::Method()
{
}

In mainline, we call gen_subprogram_die() twice for 
Object<int>::Method(): once, while generating class members, and once 
while inlining (outlining_inline_function hook).  The debug-early path 
is somewhat different, and we end up calling gen_subprogram_die() three 
times, the last of which ICEs.  What happens is the following...

We call gen_subprogram_die() as usual while generating class members, 
but then we call it again by virtue of it being a reachable function. 
This extra call will follow the DW_AT_specification code path because we 
have a previously cached die:

	  subr_die = new_die (DW_TAG_subprogram, context_die, decl);
	  add_AT_specification (subr_die, old_die);
           add_pubname (decl, subr_die);

The problem is that, for -fno-implicit-templates, the decl is now 
DECL_EXTERNAL, which means we never equate this new "DIE with 
DW_AT_specification" to the DECL.  That is, we never fall through here:

   else if (!DECL_EXTERNAL (decl))
     {
       HOST_WIDE_INT cfa_fb_offset;

       struct function *fun = DECL_STRUCT_FUNCTION (decl);

       if (!old_die || !get_AT (old_die, DW_AT_inline))
	equate_decl_number_to_die (decl, subr_die);

However, when we call gen_subprogram_die() the third time through the 
outlining_inline_function hook (late debug), we again try to add a 
DW_AT_specification to the DIE cached from the first time around, but 
this time we ICE because we're not supposed to have multiple 
DW_AT_specification's pointing to the same DIE (the old original DIE).

My solution is just to call equate_decl_number_to_die() as soon as we 
create the DW_AT_specification marked DIE the second time around.  The 
third time we will just pick up this last cached DIE with 
DW_AT_specification, mark it as DW_AT_inline, and voila, everything 
works.  The dwarf generation is as mainline, and we can build libstdc++ 
with no regressions to the guality testsuite.

Does this sound reasonable, or is this something a lot more complicated?

Thanks.
Aldy
commit 0a49042b9151e0387efc5f87c32cb24968896ae4
Author: Aldy Hernandez <aldyh@redhat.com>
Date:   Tue Mar 17 12:29:27 2015 -0700

    Equate new DIE containing a DW_AT_specification, to the original
    declaration.

Comments

Jason Merrill March 18, 2015, 2:12 a.m. UTC | #1
On 03/17/2015 03:58 PM, Aldy Hernandez wrote:
> The problem is that, for -fno-implicit-templates, the decl is now
> DECL_EXTERNAL, which means we never equate this new "DIE with
> DW_AT_specification" to the DECL.  That is, we never fall through here:
>
>    else if (!DECL_EXTERNAL (decl))
>      {
>        HOST_WIDE_INT cfa_fb_offset;
>
>        struct function *fun = DECL_STRUCT_FUNCTION (decl);
>
>        if (!old_die || !get_AT (old_die, DW_AT_inline))
>      equate_decl_number_to_die (decl, subr_die);
>
> However, when we call gen_subprogram_die() the third time through the
> outlining_inline_function hook (late debug), we again try to add a
> DW_AT_specification to the DIE cached from the first time around, but
> this time we ICE because we're not supposed to have multiple
> DW_AT_specification's pointing to the same DIE (the old original DIE).

Why are we outlining a DECL_EXTERNAL function?

Incidentally,

>           /* If we have no location information, this must be a
>              partially generated DIE from early dwarf generation.
>              Fall through and generate it.  */

Why aren't we checking dumped_early here?

Jason
Aldy Hernandez March 18, 2015, 3:51 p.m. UTC | #2
On 03/17/2015 07:12 PM, Jason Merrill wrote:
> On 03/17/2015 03:58 PM, Aldy Hernandez wrote:
>> The problem is that, for -fno-implicit-templates, the decl is now
>> DECL_EXTERNAL, which means we never equate this new "DIE with
>> DW_AT_specification" to the DECL.  That is, we never fall through here:
>>
>>    else if (!DECL_EXTERNAL (decl))
>>      {
>>        HOST_WIDE_INT cfa_fb_offset;
>>
>>        struct function *fun = DECL_STRUCT_FUNCTION (decl);
>>
>>        if (!old_die || !get_AT (old_die, DW_AT_inline))
>>      equate_decl_number_to_die (decl, subr_die);
>>
>> However, when we call gen_subprogram_die() the third time through the
>> outlining_inline_function hook (late debug), we again try to add a
>> DW_AT_specification to the DIE cached from the first time around, but
>> this time we ICE because we're not supposed to have multiple
>> DW_AT_specification's pointing to the same DIE (the old original DIE).
>
> Why are we outlining a DECL_EXTERNAL function?

SRA is analyzing Object<int>::Method() and noticing that `this' is never 
used, so it's trying to rewrite the call to avoid passing `this' (by 
creating a clone).

SRA has no restrictions on whether a function is DECL_EXTERNAL.  For 
that matter, the SRA pass is called on all functions that have a gimple 
body, irregardless of DECL_EXTERNAL, courtesy of the pass manager:

           if (node->has_gimple_body_p ())
             callback (DECL_STRUCT_FUNCTION (node->decl), data);

...and since Object<int>::Method() has a gimple body even though it is 
marked DECL_EXTERNAL...we get the call into dwarf2out_abstract_decl.

>
> Incidentally,
>
>>           /* If we have no location information, this must be a
>>              partially generated DIE from early dwarf generation.
>>              Fall through and generate it.  */
>
> Why aren't we checking dumped_early here?

Good point.  I'll add an assert.

Aldy
Jason Merrill April 3, 2015, 2:41 p.m. UTC | #3
On 03/18/2015 11:51 AM, Aldy Hernandez wrote:
> On 03/17/2015 07:12 PM, Jason Merrill wrote:
>> Why are we outlining a DECL_EXTERNAL function?
>
> SRA has no restrictions on whether a function is DECL_EXTERNAL.

Aha.

So it seems that DECL_EXTERNAL is the wrong gate for 
equate_decl_number_to_die here.  I think the rule we want is that if we 
don't already have a non-declaration DIE for a function, we should 
equate the new DIE.  Let's remove the existing calls and replace them 
with a single conditional call before the if (declaration).

Incidentally,

>       /* A declaration that has been previously dumped needs no
>          additional information.  */
>       if (dumped_early && declaration)
>         return;

Do we need to check dumped_early here?  I would think that any 
declaration that has an old_die needs no additional information.

Jason

Patch
diff mbox

diff --git a/gcc/dwarf2out.c b/gcc/dwarf2out.c
index 86815be..c7345d9 100644
--- a/gcc/dwarf2out.c
+++ b/gcc/dwarf2out.c
@@ -18809,6 +18809,8 @@  gen_subprogram_die (tree decl, dw_die_ref context_die)
 		add_type_attribute (subr_die, TREE_TYPE (TREE_TYPE (decl)),
 				    TYPE_UNQUALIFIED, context_die);
 	    }
+	  if (early_dwarf_dumping)
+	    equate_decl_number_to_die (decl, subr_die);
 	}
     }
   /* Anything else... create a brand new DIE.  */