c++: Fix convert_like in template [PR91465, PR93870, PR92031, PR94068]
authorMarek Polacek <polacek@redhat.com>
Fri, 28 Feb 2020 21:57:04 +0000 (16:57 -0500)
committerMarek Polacek <polacek@redhat.com>
Mon, 9 Mar 2020 23:54:51 +0000 (19:54 -0400)
commitd417b4f5414d9076300ab41974a14424f722688c
tree9a570171945bfd3c01b4eab4a93b2a1c86c2ff46
parentea182fe63634bb5b7913b3f1b6846e1900c5e0c4
c++: Fix convert_like in template [PR91465, PR93870, PR92031, PR94068]

The point of this patch is to fix the recurring problem of trees
generated by convert_like while processing a template that break when
substituting.  For instance, when convert_like creates a CALL_EXPR
while in a template, substituting such a call breaks in finish_call_expr
because we have two 'this' arguments.  Another problem is that we
can create &TARGET_EXPR<> and then fail when substituting because we're
taking the address of an rvalue.  I've analyzed some of the already fixed
PRs and also some of the currently open ones:

In c++/93870 we create EnumWrapper<E>::operator E(&operator~(E)).
In c++/87145 we create S::operator int (&{N}).
In c++/92031 we create &TARGET_EXPR <0>.

The gist of the problem is when convert_like_real creates a call for
a ck_user or wraps a TARGET_EXPR in & in a template.  So in these cases
use IMPLICIT_CONV_EXPR.  In a template we shouldn't need to perform the
actual conversion, we only need it's result type.
perform_direct_initialization_if_possible and
perform_implicit_conversion_flags can also create an IMPLICIT_CONV_EXPR.

Given the change above, build_converted_constant_expr can return an
IMPLICIT_CONV_EXPR so call fold_non_dependent_expr rather than
maybe_constant_value to deal with that.

To avoid the problem of instantiating something twice in a row I'm
removing a call to instantiate_non_dependent_expr_sfinae in
compute_array_index_type_loc.  And the build_converted_constant_expr
pattern can now be simplified.

2020-03-09  Marek Polacek  <polacek@redhat.com>

PR c++/92031 - bogus taking address of rvalue error.
PR c++/91465 - ICE with template codes in check_narrowing.
PR c++/93870 - wrong error when converting template non-type arg.
PR c++/94068 - ICE with template codes in check_narrowing.
* call.c (convert_like_real): Return IMPLICIT_CONV_EXPR
in a template when not ck_identity and we're dealing with a class.
(convert_like_real) <case ck_ref_bind>: Return IMPLICIT_CONV_EXPR
in a template if we need a temporary.
* decl.c (compute_array_index_type_loc): Remove
instantiate_non_dependent_expr_sfinae call.  Call
fold_non_dependent_expr instead of maybe_constant_value.
(build_explicit_specifier): Don't instantiate or create a sentinel
before converting the expression.
* except.c (build_noexcept_spec): Likewise.
* pt.c (convert_nontype_argument): Don't build IMPLICIT_CONV_EXPR.
Set IMPLICIT_CONV_EXPR_NONTYPE_ARG if that's what
build_converted_constant_expr returned.
* typeck2.c (check_narrowing): Call fold_non_dependent_expr instead
of maybe_constant_value.

* g++.dg/cpp0x/conv-tmpl2.C: New test.
* g++.dg/cpp0x/conv-tmpl3.C: New test.
* g++.dg/cpp0x/conv-tmpl4.C: New test.
* g++.dg/cpp0x/conv-tmpl5.C: New test.
* g++.dg/cpp0x/conv-tmpl6.C: New test.
* g++.dg/cpp1z/conv-tmpl1.C: New test.
13 files changed:
gcc/cp/ChangeLog
gcc/cp/call.c
gcc/cp/decl.c
gcc/cp/except.c
gcc/cp/pt.c
gcc/cp/typeck2.c
gcc/testsuite/ChangeLog
gcc/testsuite/g++.dg/cpp0x/conv-tmpl2.C [new file with mode: 0644]
gcc/testsuite/g++.dg/cpp0x/conv-tmpl3.C [new file with mode: 0644]
gcc/testsuite/g++.dg/cpp0x/conv-tmpl4.C [new file with mode: 0644]
gcc/testsuite/g++.dg/cpp0x/conv-tmpl5.C [new file with mode: 0644]
gcc/testsuite/g++.dg/cpp0x/conv-tmpl6.C [new file with mode: 0644]
gcc/testsuite/g++.dg/cpp1z/conv-tmpl1.C [new file with mode: 0644]