fmtlib
Tests fail using NVIDIA HPC compilers (PGI compilers)
The fmt test suite does not pass using the 22.3 NVIDIA HPC compilers. I tested this with today's master (682e097bee9aaa88e9c9826e72d87c2ec1b1226b) and confirmed I could reproduce using NVIDIA's own Docker images that ship their compilers:
$ docker run -it nvcr.io/nvidia/nvhpc:22.3-devel-cuda11.6-ubuntu20.04
root@411ece11247c:~# git clone https://github.com/fmtlib/fmt.git
Cloning into 'fmt'...
...
root@411ece11247c:~# cd fmt/
root@411ece11247c:~/fmt# mkdir build
root@411ece11247c:~/fmt# cd build
root@411ece11247c:~/fmt/build# cmake .. -DCMAKE_CXX_FLAGS="-tp haswell"
-- CMake version: 3.16.3
-- The CXX compiler identification is PGI 22.3.0
-- Check for working CXX compiler: /opt/nvidia/hpc_sdk/Linux_x86_64/22.3/compilers/bin/nvc++
-- Check for working CXX compiler: /opt/nvidia/hpc_sdk/Linux_x86_64/22.3/compilers/bin/nvc++ — works
…
root@411ece11247c:~/fmt/build# make
Scanning dependencies of target fmt
[ 1%] Building CXX object CMakeFiles/fmt.dir/src/format.cc.o
…
(there are some warnings here)
root@411ece11247c:~/fmt/build# make test
Running tests...
Test project /root/fmt/build
Start 1: args-test
1/20 Test #1: args-test ........................ Passed 0.01 sec
…
80% tests passed, 4 tests failed out of 20
Total Test time (real) = 0.22 sec
The following tests FAILED:
3 - chrono-test (Failed)
7 - format-test (Failed)
12 - printf-test (Failed)
17 - xchar-test (Failed)
Errors while running CTest
make: *** [Makefile:97: test] Error 8
(I had to add -DCMAKE_CXX_FLAGS="-tp haswell" because these compilers use the equivalent of -march=native by default, but they seem to mis-detect the AVX512 capabilities of my Ice Lake laptop CPU and produce binaries that fail with illegal instruction errors. This might not be needed on other systems.)
I checked 22.3 first because this is the version we currently use in production. I also checked the latest version, 22.7, using the nvcr.io/nvidia/nvhpc:22.7-devel-cuda11.7-ubuntu22.04 Docker image and see the same four tests failing.
If I additionally pass -DCMAKE_BUILD_TYPE=Debug then the printf-test test passes but the other three still fail.
In the debug build the format_test.precision test appears to fail when handling very small/denormal double values, while chrono_test.special_durations and format_test.format_nan have some mismatches between inf and nan. xchar_test.escape_string ends up with an extra backslash.
In the default (non-debug) build there are additional failures in printf_test.pointer, xchar_test.format, xchar_test.format_utf8_precision, xchar_test.format_to, xchar_test.join, xchar_test.streamed, xchar_test.chrono,
xchar_test.ostream, format_test.wide_format_to_n, chrono_test.time_point, locale_test.wformat and locale_test.chrono_weekday (list derived from the build with 22.7)
I also see that NVHPC 22.3 does not play nicely with fmt in C++17 mode:
root@411ece11247c:~/fmt# cat foo.cpp
#define FMT_HEADER_ONLY
#include <fmt/format.h>
#include <iostream>
int main() {
std::cout << fmt::format("{:.1f}", 1.0) << std::endl;
return 0;
}
root@411ece11247c:~/fmt# nvc++ -Iinclude foo.cpp -o foo -std=c++17
root@411ece11247c:~/fmt# gdb foo
...
(gdb) r
...
Program received signal SIGABRT, Aborted.
0x00007f58fb79603b in raise () from /lib/x86_64-linux-gnu/libc.so.6
(gdb) bt
#0 0x00007f58fb79603b in raise () from /lib/x86_64-linux-gnu/libc.so.6
#1 0x00007f58fb775859 in abort () from /lib/x86_64-linux-gnu/libc.so.6
#2 0x00007f58fcdb2911 in ?? () from /lib/x86_64-linux-gnu/libstdc++.so.6
#3 0x00007f58fcdbe38c in ?? () from /lib/x86_64-linux-gnu/libstdc++.so.6
#4 0x00007f58fcdbe3f7 in std::terminate() () from /lib/x86_64-linux-gnu/libstdc++.so.6
#5 0x000000000041c288 in fmt::v9::detail::assert_fail () at /root/fmt/include/fmt/format-inl.h:40
#6 0x000000000041b36b in fmt::v9::detail::grisu_gen_digits () at /root/fmt/include/fmt/format.h:2613
#7 0x00000000004225a7 in fmt::v9::detail::format_float<double> () at /root/fmt/include/fmt/format.h:3139
#8 0x0000000000428665 in fmt::v9::detail::write<char, fmt::v9::appender, double, 0> () at /root/fmt/include/fmt/format.h:3219
#9 0x0000000000416e03 in fmt::v9::detail::arg_formatter<char>::operator()<double> () at /root/fmt/include/fmt/format.h:3383
#10 0x0000000000433285 in fmt::v9::visit_format_arg<fmt::v9::detail::arg_formatter<char>&, fmt::v9::basic_format_context<fmt::v9::appender, char> > () at /root/fmt/include/fmt/core.h:1646
#11 0x000000000043554e in fmt::v9::detail::vformat_to<char>(fmt::v9::detail::buffer<char>&, fmt::v9::basic_string_view<char>, fmt::v9::basic_format_args<fmt::v9::basic_format_context<std::conditional<std::is_same<fmt::v9::type_identity<char>::type, char>::value, fmt::v9::appender, std::back_insert_iterator<fmt::v9::detail::buffer<fmt::v9::type_identity<char>::type> > >::type, fmt::v9::type_identity<char>::type> >, fmt::v9::detail::locale_ref)::format_handler::on_format_specs () at /root/fmt/include/fmt/format.h:4121
#12 0x000000000043221a in fmt::v9::detail::parse_replacement_field<char, fmt::v9::detail::vformat_to<char>(fmt::v9::detail::buffer<char>&, fmt::v9::basic_string_view<char>, fmt::v9::basic_format_args<fmt::v9::basic_format_context<std::conditional<std::is_same<fmt::v9::type_identity<char>::type, char>::value, fmt::v9::appender, std::back_insert_iterator<fmt::v9::detail::buffer<fmt::v9::type_identity<char>::type> > >::type, fmt::v9::type_identity<char>::type> >, fmt::v9::detail::locale_ref)::format_handler&> ()
at /root/fmt/include/fmt/core.h:2659
#13 0x0000000000431fd6 in fmt::v9::detail::parse_format_string<false, char, fmt::v9::detail::vformat_to<char>(fmt::v9::detail::buffer<char>&, fmt::v9::basic_string_view<char>, fmt::v9::basic_format_args<fmt::v9::basic_format_context<std::conditional<std::is_same<fmt::v9::type_identity<char>::type, char>::value, fmt::v9::appender, std::back_insert_iterator<fmt::v9::detail::buffer<fmt::v9::type_identity<char>::type> > >::type, fmt::v9::type_identity<char>::type> >, fmt::v9::detail::locale_ref)::format_handler> ()
at /root/fmt/include/fmt/core.h:2684
#14 0x000000000041c5b1 in fmt::v9::detail::vformat_to<char> () at /root/fmt/include/fmt/format.h:4125
#15 0x0000000000432692 in fmt::v9::vformat[abi:cxx11] () at /root/fmt/include/fmt/format-inl.h:1472
#16 0x0000000000432837 in fmt::v9::format<double> () at /root/fmt/include/fmt/core.h:3199
#17 0x000000000040d4ce in main () at foo.cpp:5
but this issue goes away with 22.7, or if I pass -std=c++14 instead.
Here you go: nvhpc22.7_debug.log, nvhpc22.3_debug.log, nvhpc22.7_nodebug.log, nvhpc22.3_nodebug.log.
As noted above, the differences between 22.3 and 22.7 do not seem to be significant here.
After https://github.com/fmtlib/fmt/issues/3028#issuecomment-1208145480 I tried to compile the fmt tests with -DCMAKE_CXX_STANDARD=17 and 22.3, but I get:
[ 14%] Building CXX object test/CMakeFiles/format-test.dir/format-test.cc.o
"/root/fmt/test/format-test.cc", line 2180: error: expression must have a constant value
EXPECT_ERROR("{:{}}", "width/precision is not integer", int, double);
^
"/root/fmt/include/fmt/core.h", line 757: note: attempt to access expired storage
if (arg_id < num_args_ && types_ && !is_integral_type(types_[arg_id]))
^
"/root/fmt/include/fmt/core.h", line 2272: note: called from:
context_.check_dynamic_spec(arg_id);
^
"/root/fmt/include/fmt/core.h", line 2247: note: called from:
specs_.width_ref = make_arg_ref(arg_id);
^
"/root/fmt/include/fmt/core.h", line 2445: note: called from:
FMT_CONSTEXPR void operator()() { handler.on_dynamic_width(auto_id()); }
^
"/root/fmt/include/fmt/core.h", line 2434: note: called from:
handler();
^
"/root/fmt/include/fmt/core.h", line 2464: note: called from:
if (begin != end) begin = parse_arg_id(begin, end, width_adapter{handler});
^
"/root/fmt/include/fmt/core.h", line 2605: note: called from:
begin = parse_width(begin, end, handler);
^
"/root/fmt/include/fmt/core.h", line 3037: note: called from:
auto it = detail::parse_format_specs(begin, end, checker);
^
"/root/fmt/include/fmt/core.h", line 2742: note: called from:
return f.parse(ctx);
^
"/root/fmt/include/fmt/core.h", line 2977: note: called from:
return id >= 0 && id < num_args ? parse_funcs_[id](context_) : begin;
^
"/root/fmt/include/fmt/core.h", line 2659: note: called from:
begin = handler.on_format_specs(adapter.arg_id, begin + 1, end);
^
"/root/fmt/include/fmt/core.h", line 2684: note: called from:
begin = p = parse_replacement_field(p - 1, end, handler);
^
"/root/fmt/test/format-test.cc", line 2146: note: called from:
fmt::detail::parse_format_string<true>(s, checker);
^
FWIW a very crude workaround for the issue in https://github.com/fmtlib/fmt/issues/3028#issuecomment-1208145480 seems to be https://github.com/olupton/fmt/commit/203a4cea88b9da4eb2c1786d8a18438e24b427bd -- it seems that without this then data::pow10_exponents[index] evaluates to zero in nvhpc/22.3 C++17 mode. This hack is not thoroughly tested, but seems to be working in my use-case just now...
@olupton NVIDIA HPC compilers uses EDG frontend or Clang-based frontend? Same issues exist when using MCST LCC compiler (EDG frontend).
Please check which section of the executable file contains the array data::pow10_exponents?
Closing as it is clearly a compiler bug (please report to NVIDIA). A PR with a workaround would be welcome provided that it's not too intrusive.
Up to you if you want to leave this closed, but: I agree the issue with pow10_exponents is clearly a compiler bug in 22.3 with c++17, but it seems that they have fixed it in 22.7.
The issues in the fmt test suite that I described in the original report (and I added the logs in the first line of https://github.com/fmtlib/fmt/issues/3028#issuecomment-1208303478) are still there in 22.7, so I think it's clear that the issue(s) there are at least a little different, and it's not clear (to me) that they are also the result of compiler bugs.
Apologies for mixing up these two aspects under the same issue.