Hi!
The FE uses build_complex which assumes that fold_convert will fold
value to a constant. With -frounding-math that isn't guaranteed though.
So, the patch instead fold_build2s COMPLEX_EXPR, which will result
in build_complex if both arguments are constants, and otherwise
will build COMPLEX_EXPR.
build_zero_cst is an optimization for fold_convert (type, integer_zero_node).
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
2021-11-09 Jakub Jelinek <ja...@redhat.com>
PR c++/103114
* parser.c (cp_parser_userdef_numeric_literal): Use fold_build2
with COMPLEX_EXPR arg instead of build_complex, use build_zero_cst
instead of fold_convert from integer_zero_node.
* g++.dg/ext/complex10.C: New test.
--- gcc/cp/parser.c.jj 2021-10-28 20:07:48.571193189 +0200
+++ gcc/cp/parser.c 2021-11-08 23:10:14.623106858 +0100
@@ -4804,9 +4804,8 @@ cp_parser_userdef_numeric_literal (cp_pa
else /* if (id_equal (suffix_id, "il")) */
type = long_double_type_node;
- value = build_complex (build_complex_type (type),
- fold_convert (type, integer_zero_node),
- fold_convert (type, value));
+ value = fold_build2 (COMPLEX_EXPR, build_complex_type (type),
+ build_zero_cst (type), fold_convert (type, value));
}
if (cp_parser_uncommitted_to_tentative_parse_p (parser))
--- gcc/testsuite/g++.dg/ext/complex10.C.jj 2021-11-08 23:09:58.826331001
+0100
+++ gcc/testsuite/g++.dg/ext/complex10.C 2021-11-08 23:09:21.847855693
+0100
@@ -0,0 +1,5 @@
+// PR c++/103114
+// { dg-do compile }
+// { dg-options "-frounding-math" }
+
+_Complex double d = 10.1i;
Jakub
