Skip to content

compatibility.md

Latest commit

 

History

History
129 lines (98 loc) · 5.63 KB

compatibility.md

File metadata and controls

129 lines (98 loc) · 5.63 KB
Raw

Compatibility

Why four header sets?

The libstdc++ headers are written for the compiler of the same GCC release. Newer headers use compiler builtins and language features that older compilers do not understand, and some compilers cannot parse older headers either.

We generated headers from GCC 9 to 15 and parsed all of them with every avr-gcc version, which resulted in this matrix (✓ = all headers compile with the newest supported standard):

avr-gcc GCC 9 GCC 10 GCC 11 GCC 12 GCC 13 GCC 14 GCC 15
8 ✓ C++17 ✗ ✗ ✗ ✗ ✗
9 ✓ C++17 ✗ ✗ ✗ ✗ ✗
10 ✓ ✓ ✗ ✗ ✗ ✗ ✗
11 ✓ ✓ ✓ ✓ ✗ ✗ ✗
12 ✓ ✓ ✓ ✓ ✗ ✗ ✗
13 ✗ ✗ ✗ ✗ ✓ ✗ ✗
14 ✓ ✓ ✓ ✓ ✓ ✓ ✓
15 ✓ ✓ ✓ ✓ ✓ ✓ ✓

Some notable failures:

  • avr-gcc ≤ 12 cannot parse the GCC 13 headers, which use new type trait builtins such as __is_convertible.
  • avr-gcc 13 cannot parse older headers, which use __remove_cv as an identifier, nor newer headers ("explicit template argument list not allowed").
  • avr-gcc 8 and 9 do not implement concepts, which the GCC 10 headers require in C++20 mode.

To provide the newest library features for every compiler, avr-libstdcpp uses the newest headers each compiler can parse:

Header set Generated from Supports avr-gcc C++ standards
gcc10 GCC 10.5 8, 9, 10 C++17 (8, 9), C++20
gcc12 GCC 12.5 11, 12 C++17 to C++23
gcc13 GCC 13.4 13 C++17 to C++23
gcc15 GCC 15.2 14, 15 C++17 to C++26

Each set is generated from the GCC release of the newest avr-gcc it supports, see generating the header sets. The include and src directories redirect every file to the set of your avr-gcc version, so you only need this one include path for all compilers.

avr-gcc 7 and older are not supported. There are no Homebrew builds to test them, and the previous version of this library required large backports of <type_traits> for avr-gcc 7.

Tested configurations

The integration tests are compiled with every supported avr-gcc version with every C++ standard from C++17 to the newest supported one, and then run in simavr. We use the avr-gcc builds from Homebrew, which ship with these avr-libc versions:

avr-gcc avr-libc
8.5 2.1.0
9.4 2.1.0
10.5 2.1.0
11.5 2.0.0
12.5 2.2.1
13.4 2.2.1
14.3 2.2.1
15.2 2.2.1

Notes for specific versions

avr-gcc 8 and 9

Only C++17 is supported. Their experimental C++2a mode does not implement concepts. The [[likely]] and [[unlikely]] attributes are removed from the headers, since avr-gcc 8 cannot parse them after if statements.

long double is only 32-bit wide, and avr-libc < 2.2 does not implement floorl() and ceill(), which the unordered containers need, so src/math.cc provides them.

avr-gcc 10

<coroutine> requires the -fcoroutines flag. avr-gcc 10.1 crashes with an internal compiler error in <ranges> with C++20, use a newer avr-gcc 10 or C++17.

avr-gcc 11

avr-gcc 11 cannot parse the C++23 decay-copy auto(x), which GCC 12.5 backported into std::string::resize_and_overwrite(). The gcc12 header set uses std::decay_t<decltype(x)>(x) instead.

avr-gcc 12 and 13

avr-gcc 12 and 13 warn about every access to a register with -Warray-bounds, see GCC bug 105523. Compile with --param=min-pagesize=0 to avoid this.

avr-gcc 13

avr-gcc 13 and 14.1 crash with an internal compiler error when resolving overloads with std::float32_t in C++23, since float and double are both 32-bit on AVR. The gcc13 and gcc15 header sets therefore undefine __STDCPP_FLOAT32_T__ for these compilers, so std::float32_t is not available.

Calendar and time zones

The calendar types of <chrono> (year_month_day, weekday, hh_mm_ss, …) are available since the gcc12 header set, however, __cpp_lib_chrono only announces their complete implementation since GCC 15. Time zones require the time zone database and are not supported.

avr-libc

  • avr-libc < 2.1 defines the float math functions (sinf(), …) as macros for the double functions. src/math.cc implements them as real functions and the headers undefine the macros and declare the functions instead.
  • avr-libc < 2.2 does not declare some long double functions, for example truncl(). The headers declare them, but only floorl() and ceill() are implemented by src/math.cc.
  • avr-libc 2.1 only implements the float math functions and declares the double functions as assembler aliases, which the compiler builtins used by the headers ignore. src/math.cc implements ceil(), floor() and fabs() for them.
  • avr-libc does not implement all C99 math functions, for example acosh() or tgamma(). You can call them, but linking fails.