Recent build break in fmt master with gcc10 #1810
Comments
|
I don't see anything wrong with the buffer access and wasn't able to repro the issue in godbolt (https://godbolt.org/z/nodo79). What is the value passed to |
|
Near as I can tell, it's a call to I'll see if I can reproduce it in a simple way that's not in the middle of my software system. Maybe this just reveals that there's something important I never realized about godbolt, but how does that example know which version of fmt to use? |
|
Ah, I see where fmt comes from. Did not know Compiler Explorer has all those libraries pre-installed. Cool. On my laptop (OSX) running g++-10, as well as on GH CI, I can easily make this fail: But it does not on Compiler Explorer. No explanation for you at the moment. The only thing I can think of is... Are you confident that the fmt "trunk" on Compiler Explorer is really the current master? The -O3 is required for the failure, but as it turns out, the -Wall is not. Here is my command line: And here is the error I get: Any suggestions about the next experiment I can run to help track this down? |
|
I did a clean build of [ 96%] Linking CXX executable ../bin/chrono-test
[ 96%] Built target chrono-test
In file included from /tmp/fmt/include/fmt/color.h:11,
from /tmp/fmt/test/format-test.cc:27:
In function ‘fmt::v7::detail::format_decimal_result<Char*> fmt::v7::detail::format_decimal(Char*, UInt, int) [with Char = char; UInt = unsigned int]’,
inlined from ‘OutputIt fmt::v7::detail::write(OutputIt, T) [with Char = char; OutputIt = char*; T = int; typename std::enable_if<((fmt::v7::detail::is_integral<T>::value && (! std::is_same<T, bool>::value)) && (! std::is_same<T, Char>::value)), int>::type <anonymous> = 0]’ at /tmp/fmt/include/fmt/format.h:1802:28,
inlined from ‘std::string fmt::v7::to_string(T) [with T = int; typename std::enable_if<std::is_integral<_Tp>::value, int>::type <anonymous> = 0]’ at /tmp/fmt/include/fmt/format.h:3477:62,
inlined from ‘virtual void FormatTest_ToString_Test::TestBody()’ at /tmp/fmt/test/format-test.cc:1930:3:
/tmp/fmt/include/fmt/format.h:918:55: warning: writing 2 bytes into a region of size 0 [-Wstringop-overflow=]
918 | inline void copy2(char* dst, const char* src) { memcpy(dst, src, 2); }
| ~~~~~~^~~~~~~~~~~~~
/tmp/fmt/include/fmt/format.h: In member function ‘virtual void FormatTest_ToString_Test::TestBody()’:
/tmp/fmt/include/fmt/format.h:3475:8: note: at offset -2 to object ‘buffer’ with size 11 declared here
3475 | char buffer[max_size > 5 ? max_size : 5];
| ^~~~~~
[ 98%] Linking CXX executable ../bin/printf-test
[ 98%] Built target printf-test
[100%] Linking CXX executable ../bin/format-test
[100%] Built target format-testTests ran successfully But compiling an FMT_HEADER_ONLY program doesn't produce any warnings #define FMT_HEADER_ONLY
#include <fmt/format.h>
#include <fmt/ranges.h>
int main()
{
fmt::print("Hello There!\n{}\n", "General Kenobi!");
fmt::print("Woah: {}\n", 42);
const std::string num = fmt::to_string(96);
fmt::print("Reversed: {}\n", fmt::join(num.crbegin(), num.crend(), ""));
return 0;
}$ # Using gcc 10.2.1 20200807, compiled with the following flags
$ g++ -o a a.cpp -std=c++2a -O3 -Wall -Wextra -Werror -Wshadow -Wpedantic -Wconversion -Wstringop-overflow=4 -Wshift-overflow=2 -Wformat=2 -Wcast-qual -Wcast-align -Wwrite-strings -Wdisabled-optimization -Wno-format-nonliteral |
|
And the following from an FMT_PEDANTIC build [ 94%] Linking CXX executable ../bin/chrono-test
[ 94%] Built target chrono-test
[ 96%] Linking CXX executable ../bin/compile-test
[ 96%] Built target compile-test
In file included from /tmp/fmt/include/fmt/color.h:11,
from /tmp/fmt/test/format-test.cc:27:
In function ‘fmt::v7::detail::format_decimal_result<Char*> fmt::v7::detail::format_decimal(Char*, UInt, int) [with Char = char; UInt = unsigned int]’,
inlined from ‘OutputIt fmt::v7::detail::write(OutputIt, T) [with Char = char; OutputIt = char*; T = int; typename std::enable_if<((fmt::v7::detail::is_integral<T>::value && (! std::is_same<T, bool>::value)) && (! std::is_same<T, Char>::value)), int>::type <anonymous> = 0]’ at /tmp/fmt/include/fmt/format.h:1802:28,
inlined from ‘std::string fmt::v7::to_string(T) [with T = int; typename std::enable_if<std::is_integral<_Tp>::value, int>::type <anonymous> = 0]’ at /tmp/fmt/include/fmt/format.h:3477:62,
inlined from ‘virtual void FormatTest_ToString_Test::TestBody()’ at /tmp/fmt/test/format-test.cc:1930:3:
/tmp/fmt/include/fmt/format.h:918:55: warning: writing 2 bytes into a region of size 0 [-Wstringop-overflow=]
918 | inline void copy2(char* dst, const char* src) { memcpy(dst, src, 2); }
| ~~~~~~^~~~~~~~~~~~~
/tmp/fmt/include/fmt/format.h: In member function ‘virtual void FormatTest_ToString_Test::TestBody()’:
/tmp/fmt/include/fmt/format.h:3475:8: note: at offset -2 to object ‘buffer’ with size 11 declared here
3475 | char buffer[max_size > 5 ? max_size : 5];
| ^~~~~~
[ 98%] Linking CXX executable ../bin/format-test
[ 98%] Built target format-test
[100%] Linking CXX executable ../bin/printf-test
[100%] Built target printf-test |
|
It's clearly a gcc bug: https://godbolt.org/z/fdE1jo |
|
Worked around in 6fb7c6f, thanks for reporting. |
|
Is this really a bug in GCC 10? Has anyone reported it to them? |
Yes, please see https://godbolt.org/z/fdE1jo. GCC incorrectly assumes that
I haven't, please do if you have time. |
|
Wow. Thanks for tracking this down and the workaround. |
I'm going to take a look at this today. Was there a GCC version in which this problem didn't show up? |
Yes, GCC 9.3: https://godbolt.org/z/Wvch7T |
|
Great, that should help me, thanks. |
|
OK, so this is https://gcc.gnu.org/PR95353 which has already been fixed in GCC mainline. I'm not sure if the fix will be backported to GCC 10. |
|
@mpolacek, thanks! |
Recent change in master, commit with SHA 2082912 ("Optimize count_digits") breaks the build on my project with -Wall, but the immediately preceding commit, SHA 8d9ab96 ("Cut a few cycles from count_digits") succeeds.
That doesn't guarantee the the bug is the contents of that commit per se, but maybe it put all the pieces in place for it to be symptomatic.
I'm 90% sure that it's building cleanly in gcc <= 9 and with clang. So it may just be a new warning in gcc10 it's hitting. I'm not seeing incorrect results, so maybe it's a spurious warning that I should be suppressing? Please advise if you think that's what I should do. But the "out of bounds" nature of the error message is worrying, I don't want to suppress it if it's really a buffer overflow problem.
I'm sorry I don't have a small repro case, it's in the middle of a large software project, but maybe this big block of error messages will ring a bell?
The text was updated successfully, but these errors were encountered: