[review] gdb: fix overload resolution for see-through references

Message ID gerrit.1573572186000.I39ae6505ab85ad0bd21915368c82540ceeb3aae9@gnutoolchain-gerrit.osci.io
State New, archived
Headers

Commit Message

Simon Marchi (Code Review) Nov. 12, 2019, 3:23 p.m. UTC
  Change URL: https://gnutoolchain-gerrit.osci.io/r/c/binutils-gdb/+/617
......................................................................

gdb: fix overload resolution for see-through references

The overload resolution mechanism assigns badness values to the
necessary conversions to be made on types to pick a champion.  A
badness value consists of a "rank" that scores the conversion and a
"subrank" to differentiate conversions of the same kind.

An auxiliary function, 'sum_ranks', is used for adding two badness
values.  In all of its uses, except two, 'sum_ranks' is used for
populating the subrank of a badness value.  The two exceptions are in
'rank_one_type':

~~~
  /* See through references, since we can almost make non-references
     references.  */

  if (TYPE_IS_REFERENCE (arg))
    return (sum_ranks (rank_one_type (parm, TYPE_TARGET_TYPE (arg), NULL),
		       REFERENCE_CONVERSION_BADNESS));
  if (TYPE_IS_REFERENCE (parm))
    return (sum_ranks (rank_one_type (TYPE_TARGET_TYPE (parm), arg, NULL),
		       REFERENCE_CONVERSION_BADNESS));
~~~

Here, the result of a recursive call is combined with
REFERENCE_CONVERSION_BADNESS.  This leads to the problem of
over-punishment by combining two ranks.  Consider this:

    void an_overloaded_function (const foo &);
    void an_overloaded_function (const foo &&);
    ...
    foo arg;
    an_overloaded_function(arg);

When ranking 'an_overloaded_function (const foo &)', the badness
values REFERENCE_CONVERSION_BADNESS and CV_CONVERSION_BADNESS are
combined, whereas 'rank_one_type' assigns only the
REFERENCE_CONVERSION_BADNESS value to 'an_overloaded_function (const
foo &&)' (there is a different execution flow for that).  This yields
in GDB picking the latter function as the overload champion instead of
the former.

In fact, the 'rank_one_type' function should have given
'an_overloaded_function (const foo &)' the CV_CONVERSION_BADNESS
value, with the see-through referencing increasing the subrank a
little bit.  This can be achieved by introducing a new badness value,
REFERENCE_SEE_THROUGH_BADNESS, which bumps up the subrank only, and
using it in the two "exceptional" cases of 'sum_ranks'.

gdb/ChangeLog:
2019-11-12  Tankut Baris Aktemur  <tankut.baris.aktemur@intel.com>

	* gdbtypes.h: Define the REFERENCE_SEE_THROUGH_BADNESS value.
	* gdbtypes.c (rank_one_type): Use REFERENCE_SEE_THROUGH_BADNESS
	for ranking see-through reference cases.

gdb/testsuite/ChangeLog:
2019-11-12  Tankut Baris Aktemur  <tankut.baris.aktemur@intel.com>

	* gdb.cp/rvalue-ref-overload.cc: Add a case that involves both
	CV and reference conversion for overload resolution.
	* gdb.cp/rvalue-ref-overload.exp: Test it.

Change-Id: I39ae6505ab85ad0bd21915368c82540ceeb3aae9
---
M gdb/gdbtypes.c
M gdb/gdbtypes.h
M gdb/testsuite/gdb.cp/rvalue-ref-overload.cc
M gdb/testsuite/gdb.cp/rvalue-ref-overload.exp
4 files changed, 17 insertions(+), 2 deletions(-)
  

Comments

Simon Marchi (Code Review) Nov. 27, 2019, 7:45 a.m. UTC | #1
Tankut Baris Aktemur has posted comments on this change.

Change URL: https://gnutoolchain-gerrit.osci.io/r/c/binutils-gdb/+/617
......................................................................


Patch Set 1:

Kindly ping'ing.
  
Simon Marchi (Code Review) Nov. 29, 2019, 4:43 p.m. UTC | #2
Simon Marchi has posted comments on this change.

Change URL: https://gnutoolchain-gerrit.osci.io/r/c/binutils-gdb/+/617
......................................................................


Patch Set 1:

I don't know much about the special C++ overload resolution rules, so I'd appreciate if somebody with better knowledge in that area reviewed the patch.  I have tested though, and it works as far as I can tell.

One question about the 'an_overloaded_function (const foo &&)' overload.  It would never be legal for the compiler to call this overload when doing:

 struct foo f;
 an_overloaded_function (f);

So why does GDB even consider it when searching for a possible overload to call?  Let's say there exists the overload 'an_overloaded_function (const foo &&)' but the overload 'an_overloaded_function (const foo &)' does not exist, and the user calls 'an_overloaded_function (f)'.  Should GDB call the && overload, or should GDB say that no compatible overload was found?
  
Simon Marchi (Code Review) Nov. 29, 2019, 6:57 p.m. UTC | #3
Tankut Baris Aktemur has posted comments on this change.

Change URL: https://gnutoolchain-gerrit.osci.io/r/c/binutils-gdb/+/617
......................................................................


Patch Set 1:

> One question about the 'an_overloaded_function (const foo &&)' overload.  It would never be legal for the compiler to call this overload when doing:
> 
>  struct foo f;
>  an_overloaded_function (f);
> 
> So why does GDB even consider it when searching for a possible overload to call?  Let's say there exists the overload 'an_overloaded_function (const foo &&)' but the overload 'an_overloaded_function (const foo &)' does not exist, and the user calls 'an_overloaded_function (f)'.  Should GDB call the && overload, or should GDB say that no compatible overload was found?

I believe it's expected that GDB will search the function during overload resolution.  For an illegal case like this, I'd expect that it ranks the function with a very high badness value, so that an infcall is refused, just like the compiler rejects it.

Currently, if you give 'f' a const type, GDB makes the call (but it's still illegal).  If non-const like above, GDB says "Attempt to take address of value not located in memory."  It seems this is a separate bug; GDB resolved the function and attempted to call it but a failure occurred during call preparation.
  
Simon Marchi (Code Review) Dec. 5, 2019, 8:13 p.m. UTC | #4
Tom Tromey has posted comments on this change.

Change URL: https://gnutoolchain-gerrit.osci.io/r/c/binutils-gdb/+/617
......................................................................


Patch Set 1: Code-Review+2

Thank you for doing this.  I appreciated the explanation.
The patch looks good to me.
  

Patch

diff --git a/gdb/gdbtypes.c b/gdb/gdbtypes.c
index fd1c765..2747012 100644
--- a/gdb/gdbtypes.c
+++ b/gdb/gdbtypes.c
@@ -60,6 +60,7 @@ 
 const struct rank BOOL_CONVERSION_BADNESS = {3,0};
 const struct rank BASE_CONVERSION_BADNESS = {2,0};
 const struct rank REFERENCE_CONVERSION_BADNESS = {2,0};
+const struct rank REFERENCE_SEE_THROUGH_BADNESS = {0,1};
 const struct rank NULL_POINTER_CONVERSION_BADNESS = {2,0};
 const struct rank NS_POINTER_CONVERSION_BADNESS = {10,0};
 const struct rank NS_INTEGER_POINTER_CONVERSION_BADNESS = {3,0};
@@ -4236,10 +4237,10 @@ 
 
   if (TYPE_IS_REFERENCE (arg))
     return (sum_ranks (rank_one_type (parm, TYPE_TARGET_TYPE (arg), NULL),
-                       REFERENCE_CONVERSION_BADNESS));
+                       REFERENCE_SEE_THROUGH_BADNESS));
   if (TYPE_IS_REFERENCE (parm))
     return (sum_ranks (rank_one_type (TYPE_TARGET_TYPE (parm), arg, NULL),
-                       REFERENCE_CONVERSION_BADNESS));
+                       REFERENCE_SEE_THROUGH_BADNESS));
   if (overload_debug)
   /* Debugging only.  */
     fprintf_filtered (gdb_stderr, 
diff --git a/gdb/gdbtypes.h b/gdb/gdbtypes.h
index 6d6ff59..590943e 100644
--- a/gdb/gdbtypes.h
+++ b/gdb/gdbtypes.h
@@ -2072,6 +2072,7 @@ 
 /* * Badness of converting from non-reference to reference.  Subrank
    is the type of reference conversion being done.  */
 extern const struct rank REFERENCE_CONVERSION_BADNESS;
+extern const struct rank REFERENCE_SEE_THROUGH_BADNESS;
 /* * Conversion to rvalue reference.  */
 #define REFERENCE_CONVERSION_RVALUE 1
 /* * Conversion to const lvalue reference.  */
diff --git a/gdb/testsuite/gdb.cp/rvalue-ref-overload.cc b/gdb/testsuite/gdb.cp/rvalue-ref-overload.cc
index fa6cab0..e3111d5 100644
--- a/gdb/testsuite/gdb.cp/rvalue-ref-overload.cc
+++ b/gdb/testsuite/gdb.cp/rvalue-ref-overload.cc
@@ -35,6 +35,8 @@ 
 
   int overload1arg (foo_lval_ref);
   int overload1arg (foo_rval_ref);
+  int overloadConst (const foo &);
+  int overloadConst (const foo &&);
 };
 
 void
@@ -71,6 +73,11 @@ 
   // result = 1 + 2 + 3 + 3 = 9
   int result = f (i) + f (ci) + f (0) + f (std::move (i));
 
+  /* Overload resolution below requires both a CV-conversion
+     and reference conversion.  */
+  int test_const // = 3
+    = foo_rr_instance1.overloadConst (arg);
+
   marker1 (); // marker1-returns-here
   return result;
 }
@@ -84,3 +91,5 @@ 
 
 int foo::overload1arg (foo_lval_ref arg)           { return 1; }
 int foo::overload1arg (foo_rval_ref arg)           { return 2; }
+int foo::overloadConst (const foo &arg)            { return 3; }
+int foo::overloadConst (const foo &&arg)           { return 4; }
diff --git a/gdb/testsuite/gdb.cp/rvalue-ref-overload.exp b/gdb/testsuite/gdb.cp/rvalue-ref-overload.exp
index 61f81b4..693c7ca 100644
--- a/gdb/testsuite/gdb.cp/rvalue-ref-overload.exp
+++ b/gdb/testsuite/gdb.cp/rvalue-ref-overload.exp
@@ -49,6 +49,8 @@ 
 	{ method public "~foo();" }
 	{ method public "int overload1arg(foo_lval_ref);" }
 	{ method public "int overload1arg(foo_rval_ref);" }
+	{ method public "int overloadConst(const foo &);" }
+	{ method public "int overloadConst(const foo &&);" }
     }
 
 gdb_test "print foo_rr_instance1.overload1arg(arg)" \
@@ -59,6 +61,8 @@ 
     "\\$\[0-9\]+ = 2" \
     "print call overloaded func foo && arg"
 
+gdb_test "print foo_rr_instance1.overloadConst(arg)" "3"
+
 # Test lvalue vs rvalue function overloads
 gdb_test "print f (i)" "= 1" "lvalue reference overload"