c++: partial ordering with dependent NTTP type [PR105289]
authorPatrick Palka <ppalka@redhat.com>
Tue, 26 Apr 2022 01:46:56 +0000 (21:46 -0400)
committerPatrick Palka <ppalka@redhat.com>
Tue, 26 Apr 2022 01:46:56 +0000 (21:46 -0400)
commit288e4c64f6b4806358aabc9b99b2fba72bf04bf6
tree5f328ab3507d22193c963a86c483f009573a20f6
parent98de0da6035771b5fd9185f2a2c53dd65581b886
c++: partial ordering with dependent NTTP type [PR105289]

Here ever since r11-6483-ge2e2f3f2c9400f we're rejecting and crashing
on (respectively) two testcases that we used to accept in C++17 mode
since r8-1437-g3da557ec145823.  Both testcases declare a partial
specialization of a primary template that has an NTTP with dependent
type, and the validity of these partial specializations is unclear and
the subject of PR86193 / CWG 455.

So for now, this minimal patch just aims to fix the crash in the second
testcase.  During deduction, when checking whether the type of an NTTP
uses still-undeduced parameters, we were incorrectly substituting into
the previously substituted type instead of into its original type.

In passing this patch also downgrades the "not more specialized"
diagnostic from a permerror to a pedwarn.

PR c++/105289
PR c++/86193

gcc/cp/ChangeLog:

* pt.cc (process_partial_specialization): Downgrade "partial
specialization isn't more specialized" diagnostic from permerror
to an on-by-default pedwarn.
(unify) <case TEMPLATE_PARM_INDEX>: When substituting into the
NTTP type a second time, use the original type not the
substituted type.

gcc/testsuite/ChangeLog:

* g++.dg/template/partial-specialization11.C: New test.
* g++.dg/template/partial-specialization12.C: New test.
gcc/cp/pt.cc
gcc/testsuite/g++.dg/template/partial-specialization11.C [new file with mode: 0644]
gcc/testsuite/g++.dg/template/partial-specialization12.C [new file with mode: 0644]