Re: [PATCH] c++: ttp CTAD equivalence [PR112737]
On 1/31/24 16:03, Patrick Palka wrote:
On Wed, 31 Jan 2024, Jason Merrill wrote:
On 1/31/24 12:12, Patrick Palka wrote:
Bootstrapped and regtested on x86_64-pc-linux-gnu, does this look OK for
trunk?
-- >8 --
Here during declaration matching we undesirably consider the two TT{42}
CTAD expressions to be non-equivalent ultimately because for CTAD
placeholder equivalence we compare the TEMPLATE_DECLs (which uses
pointer identity) and here the corresponding TEMPLATE_DECLs for TT are
different since they're from different scopes. On the other hand, the
corresponding TEMPLATE_TEMPLATE_PARMs are deemed equivalent (since they
have the same position and template parameters). This turns out to be
the root cause of some of the xtreme-header modules regressions.
We don't have this ttp equivalence issue in other contexts because either
the TEMPLATE_TEMPLATE_PARM is used instead of the TEMPLATE_DECL already
(e.g. when a ttp is used as a targ), or the equivalence logic is relaxed
(e.g. for bound ttps), it seems.
So this patch relaxes ttp CTAD placeholder equivalence accordingly, by
comparing the TEMPLATE_TEMPLATE_PARM instead of the TEMPLATE_DECL. The
ctp_hasher doesn't need to be adjusted since it currently doesn't include
CLASS_PLACEHOLDER_TEMPLATE in the hash anyway.
Maybe put this handling in cp_tree_equal and call it from here? Does
iterative_hash_template_arg need something similar?
I was hoping cp_tree_equal would never be called for a ttp TEMPLATE_DECL
after this patch, and so it wouldn't matter either way, but it turns out
to matter for a function template-id:
template class, class T>
void g(T);
template class TT, class T>
decltype(g(T{})) f(T); // #1
template class TT, class T>
decltype(g(T{})) f(T); // redeclaration of #1
template struct A { A(T); };
int main() {
f(0);
}
Here we represent TT in g as a TEMPLATE_DECL because it's not until
coercion that convert_template_argument turns it into a
TEMPLATE_TEMPLATE_PARM, but of course we can't coerce until the call is
non-dependent and we know which function we're calling. (So TT within
a class, variable or alias template-id would be represented as a
TEMPLATE_TEMPLATE_PARM since we can do coercion ahead of time in that
case.)
So indeed it seems desirable to handle this in cp_tree_equal... like so?
Bootstrap and regtest nearly finished.
OK.
-- >8 --
PR c++/112737
gcc/cp/ChangeLog:
* pt.cc (iterative_hash_template_arg) :
Adjust hashing to match cp_tree_equal.
(ctp_hasher::hash): Also hash CLASS_PLACEHOLDER_TEMPLATE.
* tree.cc (cp_tree_equal) : Return true
for ttp TEMPLATE_DECLs if their TEMPLATE_TEMPLATE_PARMs are
equivalent.
* typeck.cc (structural_comptypes) :
Use cp_tree_equal to compare CLASS_PLACEHOLDER_TEMPLATE.
gcc/testsuite/ChangeLog:
* g++.dg/template/ttp42.C: New test.
* g++.dg/template/ttp42a.C: New test.
---
gcc/cp/pt.cc | 9 +
gcc/cp/tree.cc | 6 +-
gcc/cp/typeck.cc | 4 ++--
gcc/testsuite/g++.dg/template/ttp42.C | 14 ++
gcc/testsuite/g++.dg/template/ttp42a.C | 18 ++
5 files changed, 48 insertions(+), 3 deletions(-)
create mode 100644 gcc/testsuite/g++.dg/template/ttp42.C
create mode 100644 gcc/testsuite/g++.dg/template/ttp42a.C
diff --git a/gcc/cp/pt.cc b/gcc/cp/pt.cc
index 5871cb668d0..ca454758ca7 100644
--- a/gcc/cp/pt.cc
+++ b/gcc/cp/pt.cc
@@ -1816,6 +1816,13 @@ iterative_hash_template_arg (tree arg, hashval_t val)
}
return iterative_hash_template_arg (TREE_TYPE (arg), val);
+case TEMPLATE_DECL:
+ if (DECL_TEMPLATE_TEMPLATE_PARM_P (arg))
+ return iterative_hash_template_arg (TREE_TYPE (arg), val);
+ else
+ /* Hash it like any other declaration. */
+ break;
+
case TARGET_EXPR:
return iterative_hash_template_arg (TARGET_EXPR_INITIAL (arg), val);
@@ -4499,6 +4506,8 @@ struct ctp_hasher : ggc_ptr_hash
hashval_t val = iterative_hash_object (code, 0);
val = iterative_hash_object (TEMPLATE_TYPE_LEVEL (t), val);
val = iterative_hash_object (TEMPLATE_TYPE_IDX (t), val);
+if (TREE_CODE (t) == TEMPLATE_TYPE_PARM)
+ val = iterative_hash_template_arg (CLASS_PLACEHOLDER_TEMPLATE (t), val);
if (TREE_CODE (t) == BOUND_TEMPLATE_TEMPLATE_PARM)
val = iterative_hash_template_arg (TYPE_TI_ARGS (t), val);
--comparing_specializations;
diff --git a/gcc/cp/tree.cc b/gcc/cp/tree.cc
index 77f57e0f9ac..5c8c05dc168 100644
--- a/gcc/cp/tree.cc
+++ b/gcc/cp/tree.cc
@@ -4084,11 +4084,15 @@ cp_tree_equal (tree t1, tree t2)
}
return false;
+case TEMPLATE_DECL:
+ if (DECL_TEMPLATE_TEMPLATE_PARM_P (t1)
+ && DECL_TEMPLATE_TEMPLATE_PARM_P (t2))
+ return cp_tree_equal (TREE_TYPE (t1), TREE_TYPE (t2));
+ /* Fall through. */
case VAR_DECL:
Re: [PATCH] c++: ttp CTAD equivalence [PR112737]
On Wed, 31 Jan 2024, Patrick Palka wrote:
> On Wed, 31 Jan 2024, Jason Merrill wrote:
>
> > On 1/31/24 12:12, Patrick Palka wrote:
> > > Bootstrapped and regtested on x86_64-pc-linux-gnu, does this look OK for
> > > trunk?
> > >
> > > -- >8 --
> > >
> > > Here during declaration matching we undesirably consider the two TT{42}
> > > CTAD expressions to be non-equivalent ultimately because for CTAD
> > > placeholder equivalence we compare the TEMPLATE_DECLs (which uses
> > > pointer identity) and here the corresponding TEMPLATE_DECLs for TT are
> > > different since they're from different scopes. On the other hand, the
> > > corresponding TEMPLATE_TEMPLATE_PARMs are deemed equivalent (since they
> > > have the same position and template parameters). This turns out to be
> > > the root cause of some of the xtreme-header modules regressions.
> > >
> > > We don't have this ttp equivalence issue in other contexts because either
> > > the TEMPLATE_TEMPLATE_PARM is used instead of the TEMPLATE_DECL already
> > > (e.g. when a ttp is used as a targ), or the equivalence logic is relaxed
> > > (e.g. for bound ttps), it seems.
> > >
> > > So this patch relaxes ttp CTAD placeholder equivalence accordingly, by
> > > comparing the TEMPLATE_TEMPLATE_PARM instead of the TEMPLATE_DECL. The
> > > ctp_hasher doesn't need to be adjusted since it currently doesn't include
> > > CLASS_PLACEHOLDER_TEMPLATE in the hash anyway.
> >
> > Maybe put this handling in cp_tree_equal and call it from here? Does
> > iterative_hash_template_arg need something similar?
>
> I was hoping cp_tree_equal would never be called for a ttp TEMPLATE_DECL
> after this patch, and so it wouldn't matter either way, but it turns out
> to matter for a function template-id:
>
> template class, class T>
> void g(T);
>
> template class TT, class T>
> decltype(g(T{})) f(T); // #1
>
> template class TT, class T>
> decltype(g(T{})) f(T); // redeclaration of #1
>
> template struct A { A(T); };
>
> int main() {
> f(0);
> }
>
> Here we represent TT in g as a TEMPLATE_DECL because it's not until
> coercion that convert_template_argument turns it into a
> TEMPLATE_TEMPLATE_PARM, but of course we can't coerce until the call is
> non-dependent and we know which function we're calling. (So TT within
> a class, variable or alias template-id would be represented as a
> TEMPLATE_TEMPLATE_PARM since we can do coercion ahead of time in that
> case.)
>
> So indeed it seems desirable to handle this in cp_tree_equal... like so?
> Bootstrap and regtest nearly finished.
>
> -- >8 --
>
> PR c++/112737
>
> gcc/cp/ChangeLog:
>
> * pt.cc (iterative_hash_template_arg) :
> Adjust hashing to match cp_tree_equal.
> (ctp_hasher::hash): Also hash CLASS_PLACEHOLDER_TEMPLATE.
I forgot to mention this change to ctp_hasher::hash is just a drive-by
optimization, so that CTAD placeholders don't all get the same hash.
> * tree.cc (cp_tree_equal) : Return true
> for ttp TEMPLATE_DECLs if their TEMPLATE_TEMPLATE_PARMs are
> equivalent.
> * typeck.cc (structural_comptypes) :
> Use cp_tree_equal to compare CLASS_PLACEHOLDER_TEMPLATE.
>
> gcc/testsuite/ChangeLog:
>
> * g++.dg/template/ttp42.C: New test.
> * g++.dg/template/ttp42a.C: New test.
> ---
> gcc/cp/pt.cc | 9 +
> gcc/cp/tree.cc | 6 +-
> gcc/cp/typeck.cc | 4 ++--
> gcc/testsuite/g++.dg/template/ttp42.C | 14 ++
> gcc/testsuite/g++.dg/template/ttp42a.C | 18 ++
> 5 files changed, 48 insertions(+), 3 deletions(-)
> create mode 100644 gcc/testsuite/g++.dg/template/ttp42.C
> create mode 100644 gcc/testsuite/g++.dg/template/ttp42a.C
>
> diff --git a/gcc/cp/pt.cc b/gcc/cp/pt.cc
> index 5871cb668d0..ca454758ca7 100644
> --- a/gcc/cp/pt.cc
> +++ b/gcc/cp/pt.cc
> @@ -1816,6 +1816,13 @@ iterative_hash_template_arg (tree arg, hashval_t val)
> }
>return iterative_hash_template_arg (TREE_TYPE (arg), val);
>
> +case TEMPLATE_DECL:
> + if (DECL_TEMPLATE_TEMPLATE_PARM_P (arg))
> + return iterative_hash_template_arg (TREE_TYPE (arg), val);
> + else
> + /* Hash it like any other declaration. */
> + break;
> +
> case TARGET_EXPR:
>return iterative_hash_template_arg (TARGET_EXPR_INITIAL (arg), val);
>
> @@ -4499,6 +4506,8 @@ struct ctp_hasher : ggc_ptr_hash
> hashval_t val = iterative_hash_object (code, 0);
> val = iterative_hash_object (TEMPLATE_TYPE_LEVEL (t), val);
> val = iterative_hash_object (TEMPLATE_TYPE_IDX (t), val);
> +if (TREE_CODE (t) == TEMPLATE_TYPE_PARM)
> + val = iterative_hash_template_arg (CLASS_PLACEHOLDER_TEMPLATE (t),
> val);
> if (TREE_CODE (t) == BOUND_TEMPLATE_TEMPLATE_PARM)
>val = iterative_hash_template_arg (TYPE_TI_ARGS (t), val);
> --comparing_specializations;
> diff --git a/gcc/cp/tree.cc
Re: [PATCH] c++: ttp CTAD equivalence [PR112737]
On Wed, 31 Jan 2024, Jason Merrill wrote:
> On 1/31/24 12:12, Patrick Palka wrote:
> > Bootstrapped and regtested on x86_64-pc-linux-gnu, does this look OK for
> > trunk?
> >
> > -- >8 --
> >
> > Here during declaration matching we undesirably consider the two TT{42}
> > CTAD expressions to be non-equivalent ultimately because for CTAD
> > placeholder equivalence we compare the TEMPLATE_DECLs (which uses
> > pointer identity) and here the corresponding TEMPLATE_DECLs for TT are
> > different since they're from different scopes. On the other hand, the
> > corresponding TEMPLATE_TEMPLATE_PARMs are deemed equivalent (since they
> > have the same position and template parameters). This turns out to be
> > the root cause of some of the xtreme-header modules regressions.
> >
> > We don't have this ttp equivalence issue in other contexts because either
> > the TEMPLATE_TEMPLATE_PARM is used instead of the TEMPLATE_DECL already
> > (e.g. when a ttp is used as a targ), or the equivalence logic is relaxed
> > (e.g. for bound ttps), it seems.
> >
> > So this patch relaxes ttp CTAD placeholder equivalence accordingly, by
> > comparing the TEMPLATE_TEMPLATE_PARM instead of the TEMPLATE_DECL. The
> > ctp_hasher doesn't need to be adjusted since it currently doesn't include
> > CLASS_PLACEHOLDER_TEMPLATE in the hash anyway.
>
> Maybe put this handling in cp_tree_equal and call it from here? Does
> iterative_hash_template_arg need something similar?
I was hoping cp_tree_equal would never be called for a ttp TEMPLATE_DECL
after this patch, and so it wouldn't matter either way, but it turns out
to matter for a function template-id:
template class, class T>
void g(T);
template class TT, class T>
decltype(g(T{})) f(T); // #1
template class TT, class T>
decltype(g(T{})) f(T); // redeclaration of #1
template struct A { A(T); };
int main() {
f(0);
}
Here we represent TT in g as a TEMPLATE_DECL because it's not until
coercion that convert_template_argument turns it into a
TEMPLATE_TEMPLATE_PARM, but of course we can't coerce until the call is
non-dependent and we know which function we're calling. (So TT within
a class, variable or alias template-id would be represented as a
TEMPLATE_TEMPLATE_PARM since we can do coercion ahead of time in that
case.)
So indeed it seems desirable to handle this in cp_tree_equal... like so?
Bootstrap and regtest nearly finished.
-- >8 --
PR c++/112737
gcc/cp/ChangeLog:
* pt.cc (iterative_hash_template_arg) :
Adjust hashing to match cp_tree_equal.
(ctp_hasher::hash): Also hash CLASS_PLACEHOLDER_TEMPLATE.
* tree.cc (cp_tree_equal) : Return true
for ttp TEMPLATE_DECLs if their TEMPLATE_TEMPLATE_PARMs are
equivalent.
* typeck.cc (structural_comptypes) :
Use cp_tree_equal to compare CLASS_PLACEHOLDER_TEMPLATE.
gcc/testsuite/ChangeLog:
* g++.dg/template/ttp42.C: New test.
* g++.dg/template/ttp42a.C: New test.
---
gcc/cp/pt.cc | 9 +
gcc/cp/tree.cc | 6 +-
gcc/cp/typeck.cc | 4 ++--
gcc/testsuite/g++.dg/template/ttp42.C | 14 ++
gcc/testsuite/g++.dg/template/ttp42a.C | 18 ++
5 files changed, 48 insertions(+), 3 deletions(-)
create mode 100644 gcc/testsuite/g++.dg/template/ttp42.C
create mode 100644 gcc/testsuite/g++.dg/template/ttp42a.C
diff --git a/gcc/cp/pt.cc b/gcc/cp/pt.cc
index 5871cb668d0..ca454758ca7 100644
--- a/gcc/cp/pt.cc
+++ b/gcc/cp/pt.cc
@@ -1816,6 +1816,13 @@ iterative_hash_template_arg (tree arg, hashval_t val)
}
return iterative_hash_template_arg (TREE_TYPE (arg), val);
+case TEMPLATE_DECL:
+ if (DECL_TEMPLATE_TEMPLATE_PARM_P (arg))
+ return iterative_hash_template_arg (TREE_TYPE (arg), val);
+ else
+ /* Hash it like any other declaration. */
+ break;
+
case TARGET_EXPR:
return iterative_hash_template_arg (TARGET_EXPR_INITIAL (arg), val);
@@ -4499,6 +4506,8 @@ struct ctp_hasher : ggc_ptr_hash
hashval_t val = iterative_hash_object (code, 0);
val = iterative_hash_object (TEMPLATE_TYPE_LEVEL (t), val);
val = iterative_hash_object (TEMPLATE_TYPE_IDX (t), val);
+if (TREE_CODE (t) == TEMPLATE_TYPE_PARM)
+ val = iterative_hash_template_arg (CLASS_PLACEHOLDER_TEMPLATE (t), val);
if (TREE_CODE (t) == BOUND_TEMPLATE_TEMPLATE_PARM)
val = iterative_hash_template_arg (TYPE_TI_ARGS (t), val);
--comparing_specializations;
diff --git a/gcc/cp/tree.cc b/gcc/cp/tree.cc
index 77f57e0f9ac..5c8c05dc168 100644
--- a/gcc/cp/tree.cc
+++ b/gcc/cp/tree.cc
@@ -4084,11 +4084,15 @@ cp_tree_equal (tree t1, tree t2)
}
return false;
+case TEMPLATE_DECL:
+ if (DECL_TEMPLATE_TEMPLATE_PARM_P (t1)
+ && DECL_TEMPLATE_TEMPLATE_PARM_P (t2))
+ return cp_tree_equal (TREE_TYPE (t1), TREE_TYPE (t2));
+ /* Fall through. */
Re: [PATCH] c++: ttp CTAD equivalence [PR112737]
On 1/31/24 12:12, Patrick Palka wrote:
Bootstrapped and regtested on x86_64-pc-linux-gnu, does this look OK for
trunk?
-- >8 --
Here during declaration matching we undesirably consider the two TT{42}
CTAD expressions to be non-equivalent ultimately because for CTAD
placeholder equivalence we compare the TEMPLATE_DECLs (which uses
pointer identity) and here the corresponding TEMPLATE_DECLs for TT are
different since they're from different scopes. On the other hand, the
corresponding TEMPLATE_TEMPLATE_PARMs are deemed equivalent (since they
have the same position and template parameters). This turns out to be
the root cause of some of the xtreme-header modules regressions.
We don't have this ttp equivalence issue in other contexts because either
the TEMPLATE_TEMPLATE_PARM is used instead of the TEMPLATE_DECL already
(e.g. when a ttp is used as a targ), or the equivalence logic is relaxed
(e.g. for bound ttps), it seems.
So this patch relaxes ttp CTAD placeholder equivalence accordingly, by
comparing the TEMPLATE_TEMPLATE_PARM instead of the TEMPLATE_DECL. The
ctp_hasher doesn't need to be adjusted since it currently doesn't include
CLASS_PLACEHOLDER_TEMPLATE in the hash anyway.
Maybe put this handling in cp_tree_equal and call it from here? Does
iterative_hash_template_arg need something similar?
PR c++/112737
gcc/cp/ChangeLog:
* typeck.cc (structural_comptypes) :
If CLASS_PLACEHOLDER_TEMPLATE is a ttp, compare its
TEMPLATE_TEMPLATE_PARM instead of its TEMPLATE_DECL.
gcc/testsuite/ChangeLog:
* g++.dg/template/ttp42: New test.
---
gcc/cp/typeck.cc | 20 +---
gcc/testsuite/g++.dg/template/ttp42.C | 14 ++
2 files changed, 31 insertions(+), 3 deletions(-)
create mode 100644 gcc/testsuite/g++.dg/template/ttp42.C
diff --git a/gcc/cp/typeck.cc b/gcc/cp/typeck.cc
index a15eda3f5f8..be82877e35c 100644
--- a/gcc/cp/typeck.cc
+++ b/gcc/cp/typeck.cc
@@ -1573,9 +1573,23 @@ structural_comptypes (tree t1, tree t2, int strict)
return false;
/* If T1 and T2 don't represent the same class template deduction,
they aren't equal. */
- if (CLASS_PLACEHOLDER_TEMPLATE (t1)
- != CLASS_PLACEHOLDER_TEMPLATE (t2))
- return false;
+ if (CLASS_PLACEHOLDER_TEMPLATE (t1) || CLASS_PLACEHOLDER_TEMPLATE (t2))
+ {
+ tree tmpl1 = CLASS_PLACEHOLDER_TEMPLATE (t1);
+ tree tmpl2 = CLASS_PLACEHOLDER_TEMPLATE (t2);
+ if (!tmpl1 || !tmpl2)
+ return false;
+ if (DECL_TEMPLATE_TEMPLATE_PARM_P (tmpl1)
+ != DECL_TEMPLATE_TEMPLATE_PARM_P (tmpl2))
+ return false;
+ if (DECL_TEMPLATE_TEMPLATE_PARM_P (tmpl1)
+ /* Treat two CTAD placeholders that use equivalent ttps
+from different scopes as equivalent by comparing their
+TEMPLATE_TEMPLATE_PARMs instead of their TEMPLATE_DECLs. */
+ ? !same_type_p (TREE_TYPE (tmpl1), TREE_TYPE (tmpl2))
+ : tmpl1 != tmpl2)
+ return false;
+ }
/* Constrained 'auto's are distinct from parms that don't have the same
constraints. */
if (!equivalent_placeholder_constraints (t1, t2))
diff --git a/gcc/testsuite/g++.dg/template/ttp42.C
b/gcc/testsuite/g++.dg/template/ttp42.C
new file mode 100644
index 000..da08e857fc5
--- /dev/null
+++ b/gcc/testsuite/g++.dg/template/ttp42.C
@@ -0,0 +1,14 @@
+// PR c++/112737
+// { dg-do compile { target c++17 } }
+
+template class TT>
+decltype(TT{42}) f(); // #1
+
+template class TT>
+decltype(TT{42}) f(); // redeclaration of #1
+
+template struct A { A(T); };
+
+int main() {
+ f();
+}
