Copyright © 2018-2022 Keith Packard
Picolibc is library offering standard C library APIs that targets small embedded systems with limited RAM. Picolibc was formed by blending code from Newlib and AVR Libc.
Build status:
Picolibc source comes from a variety of places and has a huge variety of copyright holders and license texts. While much of the code comes from Newlib, none of the GPL-related bits used to build the library are left in the repository, so all of the source code uses BSD-like licenses, a mixture of 2- and 3- clause BSD itself and a variety of other (mostly older) licenses with similar terms.
There are two files used for testing printf, test/printf-tests.c and test/testcases.c which are licensed under the GPL version 2 or later. There is also a shell script, GeneratePicolibcCrossFile.sh which is licensed under the AGPL version 3 or later which is provided as a helper for people building the library, but not used by picolibc otherwise.
The file COPYING.picolibc contains all of the current copyright and license information in the Debian standard machine-readable format. It was generated using the make-copyrights and find-copyright scripts.
Picolibc inherited code for a lot of architectures from Newlib, but at this point only has code to build for the following targets:
- ARC (32- and 64- bit)
- ARM (32- and 64- bit)
- i386 (Native and Linux hosted, for testing)
- Motorola 68000 (m68k)
- MIPS
- MSP430
- Nios II
- PowerPC (big and little endian)
- RISC-V (both 32- and 64- bit)
- Sparc64
- x86_64 (Native and Linux hosted, for testing)
- Xtensa (ESP8266, ESP32)
Supporting architectures that already have Newlib code requires:
-
newlib/libc/machine/architecture/meson.build to build the architecture-specific libc bits. This should at least include setjmp/longjmp support as these cannot be performed in architecture independent code and are needed by libstdc++.
-
newlib/libm/machine/architecture/meson.build to build any architecture-specific libm bits
-
picocrt/machine/architecture source code and build bits if you need custom startup code for the architecture. Useful in all cases, but this is necessary to run tests under qemu if your platform can do that.
-
cross-gcc-triple.txt to configure the meson cross-compilation mechanism to use the right tools
-
do-architecture-configure to make testing the cross-compilation setup easier.
-
newlib/libc/picolib support. This should include whatever startup helpers are required (like ARM interrupt vector) and TLS support (if your compiler includes this).
-
run-architecture script to run tests under QEMU. Look at the ARM and RISC-V examples to get a sense of what this needs to do and how it gets invoked from the cross-gcc-triple.txt configuration file.
Picolibc is mostly built from pieces of newlib, and retains the directory structure of that project. While there have been a lot of changes in the build system and per-thread data storage, the bulk of the source code remains unchanged.
To keep picolibc and newlib code in sync, newlib changes will be regularly incorporated. To ease integration of these changes into picolibc, some care needs to be taken while editing the code:
- Files should not be renamed.
- Fixes that also benefit users of newlib should also be sent to the newlib project
- Changes, where possible, should be made in a way compatible with newlib design. For example, instead of using 'errno' (which is valid in picolibc), use __errno_r(r), even when 'r' is not defined in the local context.
The bulk of newlib changes over the last several years have been in areas unrelated to the code used by picolibc, so keeping things in sync has not been difficult so far.
Introductory documentation. Read these first:
- Building Picolibc. Explains how to compile picolibc yourself.
- Using Picolibc. Shows how to compile and link applications once you have picolibc built and installed.
- Linking with Picolibc.ld. Provides more details about the linking process.
- Hello World. Build and run a stand-alone C application by following step-by-step instructions
Detailed documentation. Use these to learn more details about how to use Picolibc:
- Picolibc initialization
- Operating System Support.
- Printf and Scanf in Picolibc
- Thread Local Storage
- Re-entrancy and Locking
- Picolibc as embedded source
- Releasing Picolibc
- Copyright and license information
With the addition of nearly complete long double support in the math library, it seems like it's time to declare a larger version increment than usual.
-
Improve arc and xtensa support, adding TLS helpers and other build fixes
-
Fix FPSCR state for Arm8.1-M low overhead loops (thanks to David Green)
-
Add -Werror=double-promotion to default error set and fix related errors. (thanks to Ryan McClelland)
-
Fix locking bug in malloc out-of-memory path and freeing a locked mutex in the tinystdio bufio code. These were found with lock debugging code in Zephyr.
-
Add some missing functions in tinystdio, strto*l_l, remove, tmpname/tmpfile which were published in stdio.h but not included in the library.
-
Switch read/write functions to use POSIX types instead of legacy cygwin types. This makes mapping to existing an POSIX api work right.
-
Add %b support to tinystdio printf and scanf. These are disabled by default as they aren't yet standardized.
-
Fix avr math function support. The avr version of gcc has modes where double and long double are 32 or 64 bits, so the math library code now detects all of that at compile time rather than build time and reconfigures the functions to match the compiler types.
-
Add nearly complete long double support from openlibm for 80-bit Intel and 128-bit IEEE values (in addition to supporting 64-bit long doubles). Still missing are Bessel functions and decimal printf/scanf support.
-
Add limited long double support for IBM 'double double' form. This is enough to run some simple tests, but doesn't have any significant math functions yet.
-
Get Power9 code running under qemu with OPAL. This was mostly needed to validate the big-endian and exception code for 128-bit long doubles, but was also used to validate the double double support.
-
Provide times() and sysconf() implementations in semihosting. You can now build and run the dhrystone benchmark without any further code.
-
Fix use of TLS variables with stricter alignment requirements in the default linker script and startup code. (thanks to Joakim Nohlgård and Alexander Richardson who found this issue while working on lld support).
-
Support all Zephyr SDK targets
-
Support relocating the toolchain by using GCC_EXEC_PREFIX for sysroot-install when compiler doesn't use sysroot.
-
Add MIPS, SPARC and ARC support
-
Deal with RISC-V changes in gcc that don't reliably include zicsr
-
Support Picolibc as default C library with -Dsystem-libc option. With this, you can use picolibc without any extra compiler options.
-
Merge current newlib bits to get code that doesn't use struct _reent
-
Get rid of struct _reent in legacy stdio code
-
Support 16-bit int targets by fixing a few places assuming sizeof(int) == 4, object sizes not using size_t, wint_t for ucs-4 values
-
Add MSP430 support
-
Fix a couple of clang bugs (one on Cortex M0)
-
Support libc++ by adding non-standard mbstate_t.h
-
Merge i686 and x86_64 code to allow x86 multilib builds
-
Merge Xtensa newlib bits
-
Support Xtensa ESP32 targets
-
Add Nios II support
-
Fix el/ix level 4 code type errors
-
Fix out-of-source CMake build (thanks Max Behensky)
-
Improve build.md docs (thanks Kalle Raiskila)
-
Fix cmake build for various architectures
-
Initialize lock in fdopen
-
Remove %M from linker paths in single-arch builds
-
Shrink tinystdio vfprintf and vfscanf a bit
-
Use -fno-builtin-malloc -fno-builtin-free (GCC 12 compat)
-
Use -fno-builtin-copysignl (GCC 12 compat)
-
Add _zicsr to -march for risc-v picocrt (binutils 2.38 compat)
-
Add -no-warn-rwx-segments to link spec (binutils 2.38 compat)
-
Fix semihost gettimeofday, add a test.
-
Fix config option documentation. (Thanks to rdiez)
-
Document how re-entrant locking APIs are used. (Thanks to rdiez)
-
Fix some 16-bit int issues in tinystdio. (Thanks to Ayke van Laethem)
-
Make header files a bit more POSIX compliant, installing rpc headers, moving byte swapping macros to arpa/inet.h
-
Fix some stdio bugs found by Zephyr test suite: snprintf return value on buffer overflow, add ftello/fseeko, fputc return value, %0a formatting, clear EOF status after ungetc/fseek.
-
Re-do buffered I/O support to handle mixed read/write files correctly. This adds setbuf, setbuffer, setlinebuf, setvbuf.
-
Add fmemopen and freopen.
-
Add enough cmake support to allow Zephyr to build picolibc as a module using that, rather than meson.
-
Merge current newlib bits
-
Fix %p printf/scanf on ILP64 targets.
-
Fix use with C++ applications caused by a syntax error in picolibc.specs
-
Automatically include '-nostdlib' to options used while evaluating build rules to ensure tests work as expected.
-
Publish aarch64 inline math functions, ensure that inline fma functions work in installed applications for arm and risc-v.
-
Fix build on big-endian systems (thanks to Thomas Daede)
-
Add m68k support (thanks to Thomas Daede).
-
Fix build issues with ARM Cortex-a9 target (thanks to Ilia Sergachev).
-
Fix fwrite(x,0,y,z) in both tinystdio and legacy stdio. tinystdio returned the wrong value and legacy stdio caused a divide-by-zero fault.
-
Update Ryu code to match upstream (minor fixes)
-
Fix various __NEWLIB and __PICOLIBC macros; they were using a single leading underscore instead of two (thanks to Vincent Palatin).
-
Fix tinystdio error-handling bugs
-
Merge recent newlib changes (fixed ltdoa in legacy stdio)
-
Speed improvements for github CI system
-
Big-endian PowerPC support
-
Fail builds if most 'run_command' uses fail (thanks to Johan de Claville Christiansen)
-
Positional parameters in tinystdio. With this, I think tinystdio is feature complete.
-
Support for multiple build-styles of picolibc (minsize/release) in one binary package. This still requires separate meson runs.
-
Testing with glibc test code. This uncovered numerous bugs, mostly math errno/exception mistakes, but also a few serious bugs, including a couple of places where the nano-malloc failed to check for out-of-memory. Picolibc now passes all of the glibc math tests except for jn, yn, lgamma and tgamma. The picolibc versions of those functions are too inaccurate. Picolibc also passes most other relevant glibc tests, including stdio, string and stdlib areas.
-
Tinystdio version of fcvt now has a static buffer large enough to hold the maximum return size.
-
Tinystdio versions of ecvtbuf and fcvtbuf have been replaced with ecvt_r and fcvt_r equivalents, which take a 'len' parameter to prevent buffer overruns.
-
Add the GeneratePicolibcCrossFile.sh script which provides a way to isolate picolibc build scripts from the vagaries of meson version shifts (thanks to R. Diez).
-
Add 'semihost' version of crt0 that calls 'exit' after main returns. The ARM and RISC-V versions of this also include trap handlers for exceptions that print out information and exit when an exception occurs.
-
Clean up meson build bits, including use of 'fs module (thanks to Yasushi Shoji).
-
Speed up github actions by sharing Debian docker image (thanks to Yasushi Shoji).
-
Reduce use of intermediate static libraries during build
-
Use standard Meson architecture names everywhere (thanks to Yasushi Shoji).
-
Support building with -D_FORTIFY_SOURCE enabled.
-
Clean up 32-bit arm assembly code, eliminating __aeabi wrappers where possible.
-
Add basename, dirname and fnmatch back.
-
Fix all old-style (K&R) function definitions.
-
Enable lots more compiler warning flags.
-
Remove last uses of alloca in legacy stdio code.
-
Add tests from musl libc-testsuite. There aren't many tests, but these identified a few bugs.
-
Add lots more exception and errno tests for the math functions.
-
Restructure math library to always use the
__math_err
functions to raise exceptions and set errno. This removes the w_*.c wrapper functions and eliminates the__ieee names
. This centralizes compiler work-arounds to ensure run-time evaluation of expressions intended to raise exceptions. In the process, all of the libm/math files were reformatted with clang-format. -
Make tinystdio '%a' compatible with glibc, including supporting rounding and trimming trailing zeros when possible.
-
Remove floating point exception generation code on targets without floating point exception support. This reduces code size on soft float machines without affecting results.
-
Add -Wall -Wextra to default builds. Fixed warnings this raised.
-
Add htonl and friends (based on __htonl). Thanks to Johan de Claville Christiansen
-
Set errno in scalbn and scalbnf (patch forwarded to newlib).
-
Merge newlib recent changes which includes a couple of libm fixes.
-
Fix picolibc.ld to split C++ exceptions back apart (thanks to Khalil Estell)
-
Add vsscanf to tinystdio (required for libstdc++).
-
Also stick -isystem in C++ compile command to try and get picolibc headers to be used instead of newlib.
-
Add __cxa_atexit implementation to 'picoexit' path as required by C++
-
Fix lack of 'hh' support in integer-only tinystdio printf path.
-
Fix tinystdio __file flag initialization for C++ apps
-
Merge libc and libm into a single library. Having them split doesn't offer any advantages while requiring that applications add '-lm' to link successfully. Having them merged allows use of libm calls from libc code.
-
Add hex float format to *printf, *scanf and strto{d,f,ld}. This is required for C99 support.
-
Unify strto{d,f,ld} and *scanf floating point parsing code. This ensures that the library is consistent in how floats are parsed.
-
Make strto{d,f,ld} set errno to ERANGE on overflow/underflow, including when the result is a subnormal number.
-
Change
restrict
keyword in published headers to__restrict
to restore compatibility with applications building with --std=c18. -
Additional cleanups in time conversion funcs (Thanks to R. Riez)
-
Code cleanups for time conversion funcs (Thanks to R. Diez)
-
Add '-fno-stack-protector' when supported by the C compiler to avoid trouble building with native Ubuntu GCC.
-
Bug fix for converting denorms with sscanf and strto{d,f,ld}.
-
Use asm for inline asm code to allow building applications with --std=c18
-
Fix exit code for semihosting 'abort' call to make it visible to the hosting system.
-
Add strfromf and strfromd implementations. These are simple wrappers around sscanf, but strfromf handles float conversions without requiring a pass through 'double' or special linker hacks.
-
Bugfix for snprintf(buf, 0) and vsnprintf(buf, 0) to avoid smashing memory
-
Support building libstdc++ on top of picolibc
-
Add 'hosted' crt0 variant that calls exit when main returns. This makes testing easier without burdening embedded apps with unused exit processing code.
-
Add 'minimal' crt0 variant that skips constructors to save space on systems known to not use any.
-
Fix HW floating point initialization on 32-bit ARM processors to perform 'dsb' and 'isb' instructions to ensure the FPU enabling write is complete before executing any FPU instructions.
-
Create a new '--picolibc-prefix' GCC command line parameter that sets the base of all picolibc file names.
-
Add bare-metal i386 and x86_64 initializatiton code (thanks to Mike Haertel). These initalize the processor from power up to running code without requiring any BIOS.
-
Merge newlib as of late April, 2021
-
Add 'timegm' function (thanks to R. Diez).
-
Fix a number of tinystdio bugs: handle fread with size==0, parse 'NAN' and 'INF' in fscanf in a case-insensitive manner, fix negative precision to '*' arguments in printf, fix handling of 'j', 'z' and 't' argument size specifiers (thanks to Sebastian Meyer).
-
Make the fenv API more consistent and more conformant with the spec. All architectures now fall back to the default code for soft float versions, which avoids having the various exception and rounding modes get defined when not supported.
- Make riscv crt0 '_exit' symbol 'weak' to allow linking without this function.
-
Make picolibc more compatible with C++ compilers.
-
Add GCC specs file and linker script for building C++ applications with G++ that enable exception handling by linking in call stack information.
-
A few clang build fixes, including libm exception generation
-
Nano malloc fixes, especially for 'unusual' arguments
-
Merge in newlib 4.1.0 code
-
More libm exception/errno/infinity fixes, mostly in the gamma funcs.
-
Add tests for all semihost v2.0 functions.
-
A few RISC-V assembly fixes and new libm code.
-
Build fixes to reliably replace generic code with architecture-specific implementations.
With a patch which is pending for GCC 11, we'll be able to build C++ applications that use picolibc with exceptions and iostream.
-
Fix numerous libm exception and errno bugs. The math functions are all now verified to match the C19 and Posix standards in this area.
-
Change behavior of 'gamma' function to match glibc which returns lgamma for this function. Applications should not use this function, they should pick either lgamma or tgamma as appropriate.
-
Fix fma/fmaf on arm and RISC-V so that the machine-specific versions are used when the hardware has support. Also fix the math library to only use fma/fmaf when it is supported by the hardware.
-
Fix numerous nano-malloc bugs, especially with unusual parameters.
-
Change nano-malloc to always clear returned memory.
-
Improve nano-realloc to perform better in various ways, including merging adjacent free blocks and expanding the heap.
-
Add malloc tests, both a basic functional test and a stress test.
-
Improve build portability to Windows. Picolibc should now build using mingw.
-
Use hardware TLS register on ARM when available.
-
Support clang compiler. Thanks to Denis Feklushkin [email protected] and Joakim Nohlgård [email protected].
-
Avoid implicit float/double conversions. Check this by having clang builds use -Wdouble-promotion -Werror=double-promotion flags
-
Have portable code check for machine-specific overrides by matching filenames. This avoids building libraries with duplicate symbols and retains compatibility with newlib (which uses a different mechanism for this effect).
-
Patches to support building with CompCert, a formally verified compiler. Thanks to Sebastian Meyer [email protected].
-
Install 'ssp' (stack smashing protection) header files. This fixes compiling with -D_FORTIFY_SOURCE.
-
Make getc/ungetc re-entrant. This feature, which is enabled by default, uses atomic instruction sequences that do not require OS support.
-
Numerous iconv fixes, including enabling testing and switching external CCS file loading to use stdio. By default, iconv provides built-in CCS data for all of the supported encodings, which takes a fairly large amount of read-only memory. Iconv is now always included in picolibc as it isn't included in applications unless explicitly referenced by them.
-
Add __getauxval stub implementation to make picolibc work with GCC version 10 compiled for aarch64-linux-gnu.
-
Change how integer- and float- only versions of printf and scanf are selected. Instead of re-defining the symbols using the C preprocessor, picolibc now re-defines the symbols at link time. This avoids having applications compiled with a mixture of modes link in multiple versions of the underlying functions, while still preserving the smallest possible integer-only implementation.
-
Document how to use picolibc on a native POSIX system for testing. Check out the os.md file for details.
-
Merge current newlib bits in. This includes better fenv support, for which tests are now included in the picolibc test suite.
-
Fix section order in picolibc.ld to give applications correct control over the layout of .preserve, .init and .fini regions.
-
Add startup and TLS support for aarch64 and non Cortex-M 32-bit arm.
-
Fix floating point 'g' format output in tinystdio. (e.g., for 10.0, print '10' instead of '1e+01'). There are tests which verify a range of 'g' cases like these now.
-
Merge current newlib bits. The only thing which affects picolibc is the addition of fenv support for arm.
-
Make fix for CVE 2019-14871 - CVE 2019-14878 in original newlib stdio code not call 'abort'. Allocation failures are now reported back to the application.
-
Add 'exact' floating point print/scan code to tinystdio. Thanks to Sreepathi Pai for pointing me at the Ryu code by Ulf Adams.
-
Add regular expression functions from newlib. These were removed by accident while removing POSIX filesystem-specific code.
-
Make tinystdio versions of [efg]cvt functions. This means that the default tinystdio version of picolibc no longer calls malloc from these functions.
-
More clang-compatibility fixes. (Thanks to Denis Feklushkin)
-
Remove stdatomic.h and tgmath.h. (they should not be provide by picolibc)
-
Clang source compatibility. Clang should now be able to compile the library. Thanks to Denis Feklushkin for figuring out how to make this work.
-
aarch64 support. This enables the existing aarch64 code and provides an example configuration file for getting it built. Thanks for Anthony Anderson for this feature.
-
Testing on github on push and pull-request. For now, this is limited to building the library due to a bug in qemu.
-
Get newlib stdio working again. You can now usefully use Newlib's stdio. This requires a working malloc and is substantially larger than tinystdio, but has more accurate floating point input. This requires POSIX functions including read, write and a few others.
-
Fix long double strtold. The working version is only available when using tinystdio; if using newlib stdio, strtold is simply not available.
-
Improve tinystdio support for C99 printf/scanf additions.
-
Check for correct prefix when sysroot-install option is selected. The value of this option depends on how gcc was configured, and (alas) meson won't let us set it at runtime, so instead we complain if the wrong value was given and display the correct value.
-
Sync up with current newlib head.
This release contains an important TLS fix for ARM along with a few minor compatibility fixes
-
Make __aeabi_read_tp respect ARM ABI register requirements to avoid clobbering register contents during TLS variable use.
-
Use cpu_family instead of cpu in meson config, which is 'more correct' when building for a single cpu instead of multilib.
-
Make arm sample interrupt vector work with clang
-
Use __inline instead of inline in published headers to allow compiling with -ansi
-
Make 'naked' RISC-V _start function contain only asm statements as required by clang (and recommended by gcc).
-
Use -msave-restore in sample RISC-V cross-compile configuration. This saves text space.
This release was focused on cleaning up the copyright and license information.
-
Copyright information should now be present in every source file.
-
License information, where it could be inferred from the repository, was added to many files.
-
4-clause BSD licenses were changed (with permission) to 3-clause
-
Fix RISC-V ieeefp.h exception bits
-
Merge past newlib 3.2.0
-
Add PICOLIBC_TLS preprocessor define when the library has TLS support
This release now includes tests, and fixes bugs found by them.
-
ESP8266 support added, thanks to Jonathan McDowell.
-
Numerous test cases from newlib have been fixed, and precision requirements adjusted so that the library now passes its own test suite on x86, RISC-V and ARM.
-
String/number conversion bug fixes. This includes fcvt/ecvt/gcvt shared with newlib and tinystdio printf/scanf
-
A few RISC-V ABI fixes, including setting the TLS base correctly, compiling with -mcmodel=medany, and enabling the FPU for libraries built to use it.
-
Semihosting updates, including adding unlink, kill and getpid (which are used by some tests).
This release includes important fixes in picolibc.ld and more semihosting support.
-
File I/O and clock support for semihosting. This enables fopen/fdopen support in tinystdio along with an API to fetch a real time clock value.
-
Fix picolibc.ld to not attempt to use redefined symbols for memory space definitions. These re-definitions would fail and the default values be used for system memory definitions. Instead, just use the ? : operators each place the values are needed. Linker scripts continue to mystify.
-
Expose library definitions in 'picolibc.h', instead of 'newlib.h' and '_newlib_version.h'
-
Define HAVE_SEMIHOST when semihosting support is available. This lets the 'hello-world' example do some semihost specific things.
A minor update from 1.0, this release includes:
-
semihost support. This adds console I/O and exit(3) support on ARM and RISC-V hosts using the standard semihosting interfaces.
-
Posix I/O support in tinystdio. When -Dposix-io=true is included in the meson command line (which is the default), tinystdio adds support for fopen and fdopen by using malloc, open, close, read, write and lseek. If -Dposix-console=true is also passed to meson, then picolibc will direct stdin/stdout/stderr to the posix standard file descriptors (0, 1, 2).
-
Merge recent upstream newlib code. This brings picolibc up to date with current newlib sources.
-
Hello world example. This uses a simple Makefile to demonstrate how to us picolibc when installed for ARM and RISC-V embedded processors. The resulting executables can be run under qemu.
-
Remove newlib/libm/mathfp directory. This experimental code never worked correctly anyways.
This is the first release of picolibc. Major changes from newlib include:
-
Remove all non-BSD licensed code. None of it was used in building the embedded library, and removing it greatly simplifies the license situation.
-
Move thread-local values to native TLS mechanism
-
Add smaller stdio from avr-libc, which is enabled by default
-
Switch build system to meson. This has two notable benefits; the first is that building the library is much faster, the second is that it isolates build system changes from newlib making merging of newlib changes much easier.
-
Add simple startup code. This can be used in environments that don't have complicated requirements, allowing small applications to avoid needing to figure this out.