Compiling the attached code with -std=gnu++20 -fwide-exec-charset=latin1 gives an ICE with trunk. GCC 12 doesn't ICE, but maybe it's a --enable-checking=release thing (no ICE even with -fchecking=3 though). tmp$ ~/gcc/12.1/bin/g++ -std=c++20 suffix.cc -fwide-exec-charset=latin1 tmp$ ~/gcc/13/bin/g++ -std=c++20 suffix.cc -fwide-exec-charset=latin1 -freport-bug during RTL pass: expand suffix.cc: In function ‘auto __units_suffix() [with _Period = std::ratio<1, 1000>; _CharT = wchar_t]’: suffix.cc:97:22: internal compiler error: in get_constant_size, at varasm.cc:3418 97 | return L##suffix; \ | ^ suffix.cc:106:11: note: in expansion of macro ‘_GLIBCXX_UNITS_SUFFIX’ 106 | _GLIBCXX_UNITS_SUFFIX(milli, "ms") | ^~~~~~~~~~~~~~~~~~~~~ 0x8a6410 get_constant_size /home/jwakely/src/gcc/gcc/gcc/varasm.cc:3418 0x1519a88 assemble_constant_contents /home/jwakely/src/gcc/gcc/gcc/varasm.cc:3615 0x151af15 output_constant_def_contents /home/jwakely/src/gcc/gcc/gcc/varasm.cc:3666 0x151b264 maybe_output_constant_def_contents /home/jwakely/src/gcc/gcc/gcc/varasm.cc:3602 0x151b264 output_constant_def(tree_node*, int) /home/jwakely/src/gcc/gcc/gcc/varasm.cc:3568 0xe23b2e expand_expr_constant /home/jwakely/src/gcc/gcc/gcc/expr.cc:8592 0xe23b2e expand_expr_addr_expr_1 /home/jwakely/src/gcc/gcc/gcc/expr.cc:8619 0xe24169 expand_expr_addr_expr /home/jwakely/src/gcc/gcc/gcc/expr.cc:8812 0xe19464 expand_expr_real_1(tree_node*, rtx_def*, machine_mode, expand_modifier, rtx_def**, bool) /home/jwakely/src/gcc/gcc/gcc/expr.cc:12096 0xe26825 store_expr(tree_node*, rtx_def*, int, bool, bool) /home/jwakely/src/gcc/gcc/gcc/expr.cc:6330 0xe28090 expand_assignment(tree_node*, tree_node*, bool) /home/jwakely/src/gcc/gcc/gcc/expr.cc:6051 0xcf589c expand_gimple_stmt_1 /home/jwakely/src/gcc/gcc/gcc/cfgexpand.cc:3946 0xcf589c expand_gimple_stmt /home/jwakely/src/gcc/gcc/gcc/cfgexpand.cc:4044 0xcfc0ce expand_gimple_basic_block /home/jwakely/src/gcc/gcc/gcc/cfgexpand.cc:6096 0xcfdbe7 execute /home/jwakely/src/gcc/gcc/gcc/cfgexpand.cc:6822 Please submit a full bug report, with preprocessed source. Please include the complete backtrace with any bug report. See <https://gcc.gnu.org/bugs/> for instructions. Preprocessed source stored into /tmp/ccsLQGSo.out file, please attach this to your bugreport.
Maybe a similar issue to PR 81050, i.e. Latin1 isn't valid as a wide charset. In which case, we should give an error instead of ICE if possible. Reduced to remove library headers and C++20 features: template<typename, typename> constexpr bool is_same_v = false; template<typename T> constexpr bool is_same_v<T,T> = true; template<typename CharT> auto units_suffix() noexcept { if constexpr (is_same_v<CharT, wchar_t>) return L"ms"; else return "ms"; } int main() { units_suffix<wchar_t>(); } during RTL pass: expand ice.cc: In function ‘auto units_suffix() [with CharT = wchar_t]’: ice.cc:9:12: internal compiler error: in get_constant_size, at varasm.cc:3418 9 | return L"ms"; | ^~~~~ 0x8a6410 get_constant_size /home/jwakely/src/gcc/gcc/gcc/varasm.cc:3418 0x1519a88 assemble_constant_contents /home/jwakely/src/gcc/gcc/gcc/varasm.cc:3615 0x151af15 output_constant_def_contents /home/jwakely/src/gcc/gcc/gcc/varasm.cc:3666 0x151b264 maybe_output_constant_def_contents /home/jwakely/src/gcc/gcc/gcc/varasm.cc:3602 0x151b264 output_constant_def(tree_node*, int) /home/jwakely/src/gcc/gcc/gcc/varasm.cc:3568 0xe23b2e expand_expr_constant /home/jwakely/src/gcc/gcc/gcc/expr.cc:8592 0xe23b2e expand_expr_addr_expr_1 /home/jwakely/src/gcc/gcc/gcc/expr.cc:8619 0xe24169 expand_expr_addr_expr /home/jwakely/src/gcc/gcc/gcc/expr.cc:8812 0xe19464 expand_expr_real_1(tree_node*, rtx_def*, machine_mode, expand_modifier, rtx_def**, bool) /home/jwakely/src/gcc/gcc/gcc/expr.cc:12096 0xe26825 store_expr(tree_node*, rtx_def*, int, bool, bool) /home/jwakely/src/gcc/gcc/gcc/expr.cc:6330 0xe28090 expand_assignment(tree_node*, tree_node*, bool) /home/jwakely/src/gcc/gcc/gcc/expr.cc:6051 0xcf589c expand_gimple_stmt_1 /home/jwakely/src/gcc/gcc/gcc/cfgexpand.cc:3946 0xcf589c expand_gimple_stmt /home/jwakely/src/gcc/gcc/gcc/cfgexpand.cc:4044 0xcfc0ce expand_gimple_basic_block /home/jwakely/src/gcc/gcc/gcc/cfgexpand.cc:6096 0xcfdbe7 execute /home/jwakely/src/gcc/gcc/gcc/cfgexpand.cc:6822 Please submit a full bug report, with preprocessed source (by using -freport-bug). Please include the complete backtrace with any bug report. See <https://gcc.gnu.org/bugs/> for instructions. And this does ICE "gcc version 12.0.1 20220316 (experimental) (GCC)" so it probably is due to checking.
Started with r9-2891-g5ec9f8cff333bbe9.
3416 size = int_size_in_bytes (TREE_TYPE (exp)); 3417 gcc_checking_assert (size >= 0); 3418 gcc_checking_assert (TREE_CODE (exp) != STRING_CST 3419 || size >= TREE_STRING_LENGTH (exp)); (gdb) p size $1 = 4 (gdb) p exp->string.length $3 = 6 (gdb) p debug_tree (exp) <string_cst 0x7ffff6408b80 type <array_type 0x7ffff6428150 type <integer_type 0x7ffff6428000 wchar_t readonly type_6 SI size <integer_cst 0x7ffff628b210 constant 32> unit-size <integer_cst 0x7ffff628b228 constant 4> align:32 warn_if_not_align:0 symtab:0 alias-set -1 canonical-type 0x7ffff6428000 precision:32 min <integer_cst 0x7ffff628b5d0 -2147483648> max <integer_cst 0x7ffff628b5e8 2147483647> pointer_to_this <pointer_type 0x7ffff6428888>> type_6 SI size <integer_cst 0x7ffff628b210 32> unit-size <integer_cst 0x7ffff628b228 4> align:32 warn_if_not_align:0 symtab:0 alias-set -1 canonical-type 0x7ffff6428150 domain <integer_type 0x7ffff6294d20 type <integer_type 0x7ffff6289000 sizetype> type_6 DI size <integer_cst 0x7ffff6266fc0 constant 64> unit-size <integer_cst 0x7ffff6266fd8 constant 8> align:64 warn_if_not_align:0 symtab:0 alias-set -1 canonical-type 0x7ffff6294d20 precision:64 min <integer_cst 0x7ffff628b000 0> max <integer_cst 0x7ffff628b000 0>> pointer_to_this <pointer_type 0x7ffff6428a80>> readonly constant asm_written static "ms\000\000\000\000"> (gdb) p debug_generic_expr (0x7ffff6428150) const wchar_t[1] Bernd - I see it as us having an excess 2-byte nul terminator?
The assumption is that for wide charsets each char is either a wchar_t or multiple wchar_t, so I think such option is invalid; though not really sure how we could find that out and diagnose, iconv doesn't announce such details, for iconv everything is a series of bytes...
GCC 10 branch is being closed.