From 22806064a67cf30599957c1ffb322aa30e9e57e7 Mon Sep 17 00:00:00 2001
From: Patrick Palka <ppalka@redhat.com>
Date: Wed, 15 Sep 2021 18:41:21 -0400
Subject: [PATCH] c++: shortcut bad convs during overload resolution, part 2
 [PR101904]

The r12-3346 change makes us avoid computing excess argument conversions
during overload resolution, but only when it turns out there's a
strictly viable candidate in the overload set.  If there's no such
candidate then we still need to compute more conversions than strictly
necessary because subsequent conversions after the first bad conversion
can turn a non-strictly viable candidate into an unviable one, and that
affects the outcome of overload resolution and the behavior of its
callers (because of -fpermissive).

But at least in a SFINAE context, the distinction between a non-strictly
viable and an unviable candidate shouldn't matter all that much since
performing a bad conversion is always an error (even with -fpermissive),
and so forming a call to a non-strictly viable candidate will end up
being a SFINAE error anyway, just like in the unviable case.  Hence a
non-strictly viable candidate is effectively unviable (in a SFINAE
context), and we don't really need to distinguish between the two kinds.
We can take advantage of this observation to avoid computing excess
argument conversions even when there's no strictly viable candidate in
the overload set.

This patch implements this idea.  We usually detect a SFINAE context by
looking for the absence of the tf_error flag, but that's not specific
enough: we can also get here from build_user_type_conversion with
tf_error cleared, and there the distinction between a non-strictly
viable candidate and an unviable candidate still matters (it determines
whether a user-defined conversion is bad or just doesn't exist).  So this
patch sets and checks for the tf_conv flag to detect this situation too,
which avoids regressing conv2.C below.

Unlike the previous change, this one does affect the outcome of overload
resolution, but it should do so only in a way that preserves backwards
compatibility with -fpermissive.

	PR c++/101904

gcc/cp/ChangeLog:

	* call.c (build_user_type_conversion_1): Add tf_conv to complain.
	(add_candidates): When in a SFINAE context, instead of adding a
	candidate to bad_fns just mark it unviable.

gcc/testsuite/ChangeLog:

	* g++.dg/ext/conv2.C: New test.
	* g++.dg/template/conv17.C: Extend test.
---
 gcc/cp/call.c                          | 17 +++++++++++++++--
 gcc/testsuite/g++.dg/ext/conv2.C       | 13 +++++++++++++
 gcc/testsuite/g++.dg/template/conv17.C |  7 +++++++
 3 files changed, 35 insertions(+), 2 deletions(-)
 create mode 100644 gcc/testsuite/g++.dg/ext/conv2.C

diff --git a/gcc/cp/call.c b/gcc/cp/call.c
index b6011c1a2822..c5601d96ab84 100644
--- a/gcc/cp/call.c
+++ b/gcc/cp/call.c
@@ -4175,6 +4175,9 @@ build_user_type_conversion_1 (tree totype, tree expr, int flags,
   flags |= LOOKUP_NO_CONVERSION;
   if (BRACE_ENCLOSED_INITIALIZER_P (expr))
     flags |= LOOKUP_NO_NARROWING;
+  /* Prevent add_candidates from treating a non-strictly viable candidate
+     as unviable.  */
+  complain |= tf_conv;
 
   /* It's OK to bind a temporary for converting constructor arguments, but
      not in converting the return value of a conversion operator.  */
@@ -6232,8 +6235,18 @@ add_candidates (tree fns, tree first_arg, const vec<tree, va_gc> *args,
 	     stopped at the first bad conversion).  Add the function to BAD_FNS
 	     to fully reconsider later if we don't find any strictly viable
 	     candidates.  */
-	  bad_fns = lookup_add (fn, bad_fns);
-	  *candidates = (*candidates)->next;
+	  if (complain & (tf_error | tf_conv))
+	    {
+	      bad_fns = lookup_add (fn, bad_fns);
+	      *candidates = (*candidates)->next;
+	    }
+	  else
+	    /* But if we're in a SFINAE context, just mark this candidate as
+	       unviable outright and avoid potentially reconsidering it.
+	       This is safe to do because in a SFINAE context, performing a bad
+	       conversion is always an error (even with -fpermissive), so a
+	       non-strictly viable candidate is effectively unviable anyway.  */
+	    cand->viable = 0;
 	}
     }
   if (which == non_templates && !seen_perfect)
diff --git a/gcc/testsuite/g++.dg/ext/conv2.C b/gcc/testsuite/g++.dg/ext/conv2.C
new file mode 100644
index 000000000000..baf2a43b2ae6
--- /dev/null
+++ b/gcc/testsuite/g++.dg/ext/conv2.C
@@ -0,0 +1,13 @@
+// { dg-do compile { target c++11 } }
+// { dg-additional-options "-fpermissive" }
+
+struct A {
+  A(int*, int);
+};
+
+void f(A);
+
+int main() {
+  const int n = 0;
+  f({&n, 42}); // { dg-warning "invalid conversion from 'const int\\*' to 'int\\*'" }
+}
diff --git a/gcc/testsuite/g++.dg/template/conv17.C b/gcc/testsuite/g++.dg/template/conv17.C
index ba012c9d1fa2..f0f10f2ef4f1 100644
--- a/gcc/testsuite/g++.dg/template/conv17.C
+++ b/gcc/testsuite/g++.dg/template/conv17.C
@@ -53,4 +53,11 @@ concept D = requires (const T t) {
 };
 
 static_assert(D<C>);
+
+// Test that when there's no strictly viable candidate and we're in a
+// SFINAE context, we still stop at the first bad argument conversion.
+template<class T>
+concept E = requires { T().h(nullptr); };
+
+static_assert(!E<C>);
 #endif
-- 
GitLab