* Support width and alignment in std::exception formatter
The formatter for std::exception (and std::exception_ptr) previously
accepted only the optional 't' type-name specifier, so standard fill,
align and width specifiers were rejected with "unknown format
specifier". This made it impossible to pad or align exception messages,
e.g. when laying them out in a log column.
Parse the standard fill/align/width specifiers (as the std::filesystem::
path formatter already does) and apply them when writing the message.
Dynamic width ({:{}}) and the 'none'/'unknown exception' cases for
exception_ptr are handled too. Existing behavior ({}, {:t}, nested
exception unwinding) is unchanged.
* Update std.h
fmt::format("{:a}", 0.0) produced 0x0p-1022 where printf's %a gives
0x0p+0, and likewise for -0.0, {:A}, {:#a} and any explicit precision.
basic_fp::assign maps a zero biased exponent to 1 ("subnormals use
biased exponent 1"), which is right for subnormals but not for zero, so
format_hexfloat inherited the minimum subnormal exponent for a value
whose significand carries no information. Reset the exponent there
rather than in assign, which the decimal path shares.
A zero significand can only come from +-0: normals get the implicit bit
added, and a subnormal has a nonzero significand by definition. So the
guard cannot affect anything else, and denorm_min keeps its existing
denormalized form (0x0.0000000000001p-1022, asserted in format_double).
The five new expectations were taken from printf rather than derived,
and all fail without this change.
* Set matching CC alongside CXX in Linux CI matrix
The Configure step only set CXX, so CMake's C compiler detection
fell back to whatever the default happened to be on the runner,
independent of which C++ compiler the matrix entry was actually
testing (e.g. CXX=clang++-3.6 but CC left to detect GCC 11).
Derive CC from the same matrix.cxx value the job already installs
a matching compiler for.
* Skip c-test when the C compiler is Clang < 3.8
Setting CC from matrix.cxx means the clang++-3.6 job now configures with real
clang 3.6 instead of falling back to the runner's default GCC, which exposes a
compile failure in test/c-test.c: clang predates the LLVM PR16340 fix (landed in
3.8) that applies array-to-pointer decay to the controlling expression of
_Generic, so fmt-c.h's FMT_MAKE_ARG dispatch never matches a string literal
(char[4]) against its char*/const char* associations. It falls through to the
zero-argument default association, which is then called with one argument.
The guard is at configure time rather than in the CI Test step. c-test is part
of the default target, so `cmake --build` compiles it before ctest ever runs --
excluding it with `ctest -E` cannot help, because the job has already failed in
the Build step. Guarding the add_executable also means anyone building fmt's
tests with an old clang benefits, not just this one CI job.
enable_language(C) stays first and unconditional, since CMAKE_C_COMPILER_VERSION
is not set before it. fmt-c itself (src/fmt-c.cc, C++) still builds on that job;
only the C-consumer smoke test is skipped.
The std.h header uses types such as std::atomic and std::atomic_flag.
For the normal build mode, this header does include all the
headers it needs, such as <atomic>.
However when building the C++ module of {fmt}, the std.h header
omits such includes, and instead expects them to be included
beforehand by fmt.cc - which lacked including <atomic>.
This fixes building the C++ module with libc++ 23, which has removed
a number of unnecessary transitive includes.
A named argument that arrives const-qualified was not recognized by
is_named_arg, silently dropping its name. This happens both when a
named argument is passed through an intermediate function returning
const T& (https://github.com/fmtlib/fmt/issues/4866) and in fmt's own
compiled format path, which passes arguments as const T&.
Make is_named_arg and is_static_named_arg see through top-level const so
the name is preserved instead of dropped.
The FMT_COMPILE("{}") format string takes a to_string fast path that
never runs the compiled field code the format_as fix (#4836) touched.
Use "[{}]" so the test actually goes through detail::field::format.
Make basic_ostream_formatter, ostream_formatter and streamed() accessible
to module users via FMT_BEGIN_EXPORT/FMT_BEGIN_EXPORT. Drop now unneeded
FMT_EXPORT for print() and println().
Rename INSTALL_TARGETS to FMT_TARGETS since the list is now used for both
installation and the build-tree export, and move INSTALL_FILE_SET into the
FMT_INSTALL block as only install(TARGETS) uses it.
Replace the export-test static library and export-lib.cc with an INTERFACE
library, which reproduces the same CMake export dependency check (#4806).
export() was scoped inside if (FMT_INSTALL), so projects that pull in
fmt via add_subdirectory()/FetchContent without installing it had no
way to get the exported fmt::* targets. If such a project tries to
export its own targets that depend on fmt, CMake fails with "target
... requires target fmt that is not in any export set".
Hoist the target list/export name and the export() call itself out of
the FMT_INSTALL guard so the build-tree export file is always
generated; the install()-only pieces (config/version files, pkgconfig,
install(EXPORT ...)) stay behind the guard and installed behavior is
unchanged.
Fixes#4806
Co-authored-by: flink <w741069229@gmail.com>
count_digits and do_format_decimal divided the full 128-bit value by a
constant in a loop. Compilers do not strength-reduce 128-bit division, so
each iteration became a slow __udivti3 library call (~27 per number).
Reduce the value to 64-bit chunks by dividing by 10^19 (at most twice) and
reuse the fast 64-bit paths, giving a ~5x speedup on full-range values.
Formatting an integer with ':c' used the magnitude (abs_value) and never
range-checked, so negatives were mangled and out-of-range values silently
truncated. Copy the value as a character and report an error when it is out
of range, treating all character types as unsigned for portability.
The appender and generic-iterator overloads only differed in whether the
integer formatting code was inlined. Merge them into a single overload and
select write_int_noinline vs inline write_int via if constexpr, keeping the
same codegen (out of line for appenders, inlined for compiled formatting)
with no size regression.
Detect when a std::exception also derives from std::nested_exception
(e.g. via std::throw_with_nested) and unwind the whole chain, joining
each level with ": ". The type-name flag ({:t}) is honored at every
level, and the std::exception_ptr formatter inherits this behavior.
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Use num_bits<UInt>() == 32 instead of sizeof(UInt) == 4 in
do_format_decimal. The magic-number two-digit algorithm requires
exactly 32 value bits, so this is more precise on platforms where
CHAR_BIT != 8 and is consistent with existing num_bits<T>() usage.
Revert the explicit-else workaround from #4822 and instead disable the
"unreachable code" warning for all of base.h. Add an FMT_PRAGMA_MSVC
helper (mirroring FMT_PRAGMA_GCC/CLANG) and wrap the header in a single
warning(push)/disable:4702/pop pair, keeping the original control flow.
if FMT_CONSTEXPR is resolved to constexpr MSC warns about dead code
after 'return' instruction. Explicitly using 'else' in these cases
avoids the warning.
Prevent false positives when the container itself does not have `operator[](size_t)`, but the target type of one of the non-explicit user-defined conversion functions does.
The .gitignore-derived `build/` pattern does not actually prune the build
directory or its contents from the CPack source package, so an empty build/
(and, once docs were built in CI, build/doc-html) leaked into the zip. Ignore
the whole build tree in CPACK_SOURCE_IGNORE_FILES.
The source-zip build moved to CI, which configured out-of-source and never
ran the doc target, so the pre-rendered HTML docs (doc-html/) silently
dropped out of the package. Install the doc toolchain, build the docs, and
stage them into the source tree before package_source so CPack picks them up.
The lint workflow was path-filtered to source files, so PRs that didn't
touch them never ran clang-format/cmake-format. Since those checks are
required, such PRs were blocked forever waiting for a status that never
came. Drop the paths filter so the workflow always reports a status.
