[debug/88006] -fdebug-types-section gives undefined ref

[debug/88006] -fdebug-types-section gives undefined ref

[Nathan Sidwell]

[-- Attachment #1: Type: text/plain, Size: 1216 bytes --]

PR88006 and 87462 turn out to be the same problem, but started by 
different commits.

The problem is that debug wants to emit a constant (for a non-type 
template parameter of a template instantiation) that refers to the 
address of some local function.  That local might not be emitted though.

gen_remaining_tmpl_value_param_die_attribute is aware of the problem:
     /* We do this in two phases - first get the cases we can
	 handle during early-finish, preserving those we cannot
	 (containing symbolic constants where we don't yet know
	 whether we are going to output the referenced symbols).
	 For those we try again at late-finish.  */

But the second phase unconditionally emits the constant:
	      if (! early_dwarf
		  && (dwarf_version >= 5 || !dwarf_strict))
		loc = loc_descriptor_from_tree (e->arg, 2, NULL);

and loc_descriptor_from_tree eventually ends up in 
const_ok_for_output_1.  That function assumes pretty much any non 
SYMBOL_REF_EXTERNAL_P SYMBOL_REF is accessible.  That's not true.

I'm not sure if there's a more canonical way of checking for a 
local-decl's emission, but not having a symtab entry seems significant?

booted & tested on x86_64-linux, ok?

nathan

-- 
Nathan Sidwell

[-- Attachment #2: pr88006.diff --]
[-- Type: text/x-patch, Size: 3768 bytes --]

2018-11-14  Nathan Sidwell  <nathan@acm.org>

	PR debug/88006
	PR debug/87462
	* dwarf2out.c (const_ok_for_output_1): Check symtab for internal
	decls.

	* g++.dg/debug/dwarf2/pr87462.C: New.
	* g++.dg/debug/dwarf2/pr88006.C: New.

Index: dwarf2out.c
===================================================================
--- dwarf2out.c	(revision 266082)
+++ dwarf2out.c	(working copy)
@@ -14480,13 +14472,15 @@ const_ok_for_output_1 (rtx rtl)
   if (SYMBOL_REF_TLS_MODEL (rtl) != TLS_MODEL_NONE)
     return false;
 
-  /* Avoid references to external symbols in debug info, on several targets
-     the linker might even refuse to link when linking a shared library,
-     and in many other cases the relocations for .debug_info/.debug_loc are
-     dropped, so the address becomes zero anyway.  Hidden symbols, guaranteed
-     to be defined within the same shared library or executable are fine.  */
   if (SYMBOL_REF_EXTERNAL_P (rtl))
     {
+      /* Avoid references to external symbols in debug info, on
+	 several targets the linker might even refuse to link when
+	 linking a shared library, and in many other cases the
+	 relocations for .debug_info/.debug_loc are dropped, so the
+	 address becomes zero anyway.  Hidden symbols, guaranteed to
+	 be defined within the same shared library or executable are
+	 fine.  */
       tree decl = SYMBOL_REF_DECL (rtl);
 
       if (decl == NULL || !targetm.binds_local_p (decl))
@@ -14496,6 +14490,18 @@ const_ok_for_output_1 (rtx rtl)
 	  return false;
 	}
     }
+  else
+    {
+      /* Internal decls are not necessarily emitted -- they could be
+         inlined or whatever.  */
+      if (tree decl = SYMBOL_REF_DECL (rtl))
+	if (!symtab_node::get (decl))
+	  {
+	    expansion_failed (NULL_TREE, rtl,
+			      "Symbol not emitted in current TU.\n");
+	    return false;
+	  }
+    }
 
   return true;
 }
Index: testsuite/g++.dg/debug/dwarf2/pr87462.C
===================================================================
--- testsuite/g++.dg/debug/dwarf2/pr87462.C	(revision 0)
+++ testsuite/g++.dg/debug/dwarf2/pr87462.C	(working copy)
@@ -0,0 +1,20 @@
+// { dg-additional-options "-dA -std=gnu++17 -gdwarf-4 -O1 -fdebug-types-section" }
+// reject .pseudo label, but "label" is ok.
+// { dg-final { scan-assembler-not "\[^L\"\]_ZN5Test18testFuncEv" } }
+// undefined ref to _ZN5Test18testFuncEv
+
+class Test1 {
+public:
+  static int testFunc() { return 1; }
+};
+
+template <typename T,
+          T (*funcImpl)()>
+class TestWrapper {
+public:
+  static T func() __attribute((noinline)) { return (*funcImpl)(); } 
+};
+
+int main() {
+  return TestWrapper<int, &Test1::testFunc>::func();
+}
Index: testsuite/g++.dg/debug/dwarf2/pr88006.C
===================================================================
--- testsuite/g++.dg/debug/dwarf2/pr88006.C	(revision 0)
+++ testsuite/g++.dg/debug/dwarf2/pr88006.C	(working copy)
@@ -0,0 +1,39 @@
+// { dg-additional-options "-dA -std=gnu++17 -gdwarf-4 -O1 -fdebug-types-section" }
+// reject .pseudo label, but "label" is ok.
+// { dg-final { scan-assembler-not "\[^\"\]_ZN3Foo4mfunEv" } }
+// undefined ref to _ZN3Foo4mfunEv
+
+struct Foo {
+  void mfun () {}
+};
+
+struct A { static constexpr bool Value = false; };
+
+template <bool> struct B { typedef int Type; };
+
+class Arg
+{
+  template <typename Unused> struct Local : A {};
+
+public:
+  template <typename Init, typename = typename B<Local<Init>::Value>::Type>
+  Arg (Init) {}
+};
+
+class Lambda {
+  static constexpr int Unused = 0;
+    
+public:
+  Lambda (Arg);
+};
+
+// Generated ref to Foo::mfun in the type die of an instantiation of this
+template <void (Foo::*unused)()> struct Callable {};
+
+class I {
+  I() : lamb ([this] {}) {}
+
+  Lambda lamb;
+
+  Callable<&Foo::mfun> bm;
+};

Re: [debug/88006] -fdebug-types-section gives undefined ref

[Richard Biener]

On Wed, Nov 14, 2018 at 12:58 PM Nathan Sidwell <nathan@acm.org> wrote:
>
> PR88006 and 87462 turn out to be the same problem, but started by
> different commits.
>
> The problem is that debug wants to emit a constant (for a non-type
> template parameter of a template instantiation) that refers to the
> address of some local function.  That local might not be emitted though.
>
> gen_remaining_tmpl_value_param_die_attribute is aware of the problem:
>      /* We do this in two phases - first get the cases we can
>          handle during early-finish, preserving those we cannot
>          (containing symbolic constants where we don't yet know
>          whether we are going to output the referenced symbols).
>          For those we try again at late-finish.  */
>
> But the second phase unconditionally emits the constant:
>               if (! early_dwarf
>                   && (dwarf_version >= 5 || !dwarf_strict))
>                 loc = loc_descriptor_from_tree (e->arg, 2, NULL);
>
> and loc_descriptor_from_tree eventually ends up in
> const_ok_for_output_1.  That function assumes pretty much any non
> SYMBOL_REF_EXTERNAL_P SYMBOL_REF is accessible.  That's not true.
>
> I'm not sure if there's a more canonical way of checking for a
> local-decl's emission, but not having a symtab entry seems significant?
>
> booted & tested on x86_64-linux, ok?

Hmm, there is reference_to_unused () used in a related case.  But generally
for late emission such references are "OK" and expected to be pruned
later by resolve_addr () (which I see we do not call for type units?!).  Quote:

/* Resolve DW_OP_addr and DW_AT_const_value CONST_STRING arguments to
   an address in .rodata section if the string literal is emitted there,
   or remove the containing location list or replace DW_AT_const_value
   with DW_AT_location and empty location expression, if it isn't found
   in .rodata.  Similarly for SYMBOL_REFs, keep only those that refer
   to something that has been emitted in the current CU.  */

static void
resolve_addr (dw_die_ref die)
{


Richard.

>
> nathan
>
> --
> Nathan Sidwell

Re: [debug/88006] -fdebug-types-section gives undefined ref

[Nathan Sidwell]

[-- Attachment #1: Type: text/plain, Size: 876 bytes --]

On 11/14/18 7:33 AM, Richard Biener wrote:

> Hmm, there is reference_to_unused () used in a related case.  But generally
> for late emission such references are "OK" and expected to be pruned
> later by resolve_addr () (which I see we do not call for type units?!).  Quote:
> 
> /* Resolve DW_OP_addr and DW_AT_const_value CONST_STRING arguments to
>     an address in .rodata section if the string literal is emitted there,
>     or remove the containing location list or replace DW_AT_const_value
>     with DW_AT_location and empty location expression, if it isn't found
>     in .rodata.  Similarly for SYMBOL_REFs, keep only those that refer
>     to something that has been emitted in the current CU.  */
> 
> static void
> resolve_addr (dw_die_ref die)
> {

That does seem to work.  The attached survise bootstrapping on 
x86_64-linux, ok?

nathan

-- 
Nathan Sidwell

[-- Attachment #2: pr88006-2.diff --]
[-- Type: text/x-patch, Size: 2604 bytes --]

2018-11-14  Nathan Sidwell  <nathan@acm.org>

	PR debug/88006
	PR debug/87462
	* dwarf2out.c (dwarf2out_finish): Apply resolve_addr to comdat
	type list.

	* g++.dg/debug/dwarf2/pr87462.C: New.
	* g++.dg/debug/dwarf2/pr88006.C: New.

Index: dwarf2out.c
===================================================================
--- dwarf2out.c	(revision 266082)
+++ dwarf2out.c	(working copy)
@@ -31182,6 +31182,8 @@ dwarf2out_finish (const char *filename)
     FOR_EACH_CHILD (die, c, gcc_assert (! c->die_mark));
   }
 #endif
+  for (ctnode = comdat_type_list; ctnode != NULL; ctnode = ctnode->next)
+    resolve_addr (ctnode->root_die);
   resolve_addr (comp_unit_die ());
   move_marked_base_types ();
 
Index: testsuite/g++.dg/debug/dwarf2/pr87462.C
===================================================================
--- testsuite/g++.dg/debug/dwarf2/pr87462.C	(revision 0)
+++ testsuite/g++.dg/debug/dwarf2/pr87462.C	(working copy)
@@ -0,0 +1,20 @@
+// { dg-additional-options "-dA -std=gnu++17 -gdwarf-4 -O1 -fdebug-types-section" }
+// reject .pseudo label, but "label" is ok.
+// { dg-final { scan-assembler-not "\[^L\"\]_ZN5Test18testFuncEv" } }
+// undefined ref to _ZN5Test18testFuncEv
+
+class Test1 {
+public:
+  static int testFunc() { return 1; }
+};
+
+template <typename T,
+          T (*funcImpl)()>
+class TestWrapper {
+public:
+  static T func() __attribute((noinline)) { return (*funcImpl)(); } 
+};
+
+int main() {
+  return TestWrapper<int, &Test1::testFunc>::func();
+}
Index: testsuite/g++.dg/debug/dwarf2/pr88006.C
===================================================================
--- testsuite/g++.dg/debug/dwarf2/pr88006.C	(revision 0)
+++ testsuite/g++.dg/debug/dwarf2/pr88006.C	(working copy)
@@ -0,0 +1,39 @@
+// { dg-additional-options "-dA -std=gnu++17 -gdwarf-4 -O1 -fdebug-types-section" }
+// reject .pseudo label, but "label" is ok.
+// { dg-final { scan-assembler-not "\[^\"\]_ZN3Foo4mfunEv" } }
+// undefined ref to _ZN3Foo4mfunEv
+
+struct Foo {
+  void mfun () {}
+};
+
+struct A { static constexpr bool Value = false; };
+
+template <bool> struct B { typedef int Type; };
+
+class Arg
+{
+  template <typename Unused> struct Local : A {};
+
+public:
+  template <typename Init, typename = typename B<Local<Init>::Value>::Type>
+  Arg (Init) {}
+};
+
+class Lambda {
+  static constexpr int Unused = 0;
+    
+public:
+  Lambda (Arg);
+};
+
+// Generated ref to Foo::mfun in the type die of an instantiation of this
+template <void (Foo::*unused)()> struct Callable {};
+
+class I {
+  I() : lamb ([this] {}) {}
+
+  Lambda lamb;
+
+  Callable<&Foo::mfun> bm;
+};

Re: [debug/88006] -fdebug-types-section gives undefined ref

[Jeff Law]

On 11/14/18 8:08 AM, Nathan Sidwell wrote:
> On 11/14/18 7:33 AM, Richard Biener wrote:
> 
>> Hmm, there is reference_to_unused () used in a related case.  But generally
>>
>> for late emission such references are "OK" and expected to be pruned
>> later by resolve_addr () (which I see we do not call for type units?!).  Quote:
>>
>>
>> /* Resolve DW_OP_addr and DW_AT_const_value CONST_STRING arguments to
>>     an address in .rodata section if the string literal is emitted there,
>>     or remove the containing location list or replace DW_AT_const_value
>>     with DW_AT_location and empty location expression, if it isn't found
>>     in .rodata.  Similarly for SYMBOL_REFs, keep only those that refer
>>     to something that has been emitted in the current CU.  */
>>
>> static void
>> resolve_addr (dw_die_ref die)
>> {
> 
> That does seem to work.  The attached survise bootstrapping on
> x86_64-linux, ok?
> 
> nathan
> 
> -- 
> Nathan Sidwell
> 
> pr88006-2.diff
> 
> 2018-11-14  Nathan Sidwell  <nathan@acm.org>
> 
> 	PR debug/88006
> 	PR debug/87462
> 	* dwarf2out.c (dwarf2out_finish): Apply resolve_addr to comdat
> 	type list.
> 
> 	* g++.dg/debug/dwarf2/pr87462.C: New.
> 	* g++.dg/debug/dwarf2/pr88006.C: New.
OK
jeff