C3 0.8.6 - Let's grab those libraries

Changes / improvements¶

• @asmname allows defining external names like @cname but without normal name mangling.
• Improved GfxKey mappings.
• Add the source-lib target type (and c3c source-lib command) to pack the C3 sources, C sources and headers of a project into a .c3l library.
• .c3l manifests without targets now work for all platforms, and may contain version, authors and langrev.
• fetch, vendor-fetch and project fetch now unpack libraries into a <name>.c3l directory by default. Use --decompress=no to keep them compressed.
• fetch now verifies the downloaded library against the checksum (sha256:<hex>) of the registry, and refuses a library that doesn't match.
• dependencies, in projects and in library manifests, may name a registry library as [<url>:]<namespace>/<name>[@<version>]. The build only uses the name.
• fetch adds just the library name to dependencies.
• project fetch only fetches the libraries it is given: c3c project fetch <library>.... Without any, it lists the missing ones. Each must be a dependency of the project, or of a library that is installed. A name uses the path of the dependency, or is looked up in the registry: the fetch-url project property or --fetch-url for the project's dependencies, the default registry for those of libraries.
• A version in the registry's library info may have a url, which fetch downloads the library from instead of from the registry itself.
• A registry library can be locked to a checksum: <namespace>/<name>@<version>:<sha256>. fetch refuses a library that doesn't match.
• An else if now requires { } around a single statement, like else does. else if (x) foo(); is an error. #3473
• Add depends-on to project targets, allowing them to depend on other targets.
• LLVM IR (.ll) and bitcode (.bc) files can be added to the compilation, just like .o files #3339
• project.json strings (except exec and type) expand environment variables written as ${NAME}. #3580
• Add env and env-override to project.json, setting environment variables ("NAME=value") for the exec scripts. #3580
• Add -gline-tables-only to emit line tables debug info only.
• Disallow { .a = 1, .a = 1 }. #3579
• Add better instantiation error output.
• Implicitly converting a negative constant to a wider unsigned type, e.g. ulong x = -1;, is now correctly an error.
• Improved the error when implicitly casting an enum to an integer, it now points to the explicit cast.
• Casting a non-constant integer larger than a pointer, e.g. an int128, to a pointer is now an error. Cast it to uptr first.
• Converting between signed and unsigned integers of the same size in arrays, slices and vectors, e.g. int[3] to uint[3], requires explicit casts.
• float16 and bfloat no longer convert implicitly to each other, as both directions lose information.
• An explicit cast of a slice to a pointer is now allowed for any pointer type, like a cast between pointers.
• Selectors are now a stable 128 bit signature. The first 16 bytes are guaranteed unique.
• Add --stats=<none|summary|full|json> for compiler statistics.
• Removed --print-large-functions, --stats=full replaces it.
• Compiler generated panics will now be emitted to the std::core::runtime::panics module.
• PanicFn is changed to support both formatting and non-formatted messages.
• --panic-msg=no only traps if --show-backtrace=no is also set, otherwise it gives the short message with file, function and line. --show-backtrace=no stops passing file, function and line.
• --panicfn now names a function rather than a function pointer variable, and sets the initial value of builtin::panic through the new $$PANIC_FN, so stdlib panics and abort also reach it. #3443
• Make panic on infinite loops opt-in with --infinite-loop-detect=yes – and is only added in safe mode. #3630

Fixes¶

• Compressed .c3l libraries were not unpacked again when only files other than manifest.json had changed.
• Fixed regression for MacOS loading multiple C3 dylibs. @finalizer now correctly works in descending order.
• Fix miscompilation for OpenBSD on later processors. #3563
• Polling a GfxEvent for a window and having events on other windows would silently discard the last event.
• Setting no-border state during fullscreen on Win32 would not work properly.
• X11 window destruction order was incorrect.
• Win32 did not support sharing of OpenGL contexts.
• Win32 OpenGL pixel format selection ignored several hints.
• Added FAILED_TO_CREATE_WGL_CONTEXT for Win32 WGL format fails.
• Win32 handle would leak on drag-and-drop when dnd was disabled.
• X11 mode change did not change the stored mode.
• X11 window name is now kept when recreating a window for opengl init.
• Doc viewer collision between modules and declarations sharing the same path. #3572
• Fixes to transparent mouse pointer.
• Compiler crashes when using a value derived from a const as an initializer elsewhere. #3574
• Improve OpenGL-handling of vsync.
• Regression for emscripten, accidentally retainin unused panic functions. #3575.
• Compiler hangs on c3c project fetch #3562
• Temp dirs were left behind by c3c compile (including unpacked .c3l libraries), fetch-list, vendor-fetch and the SDK fetch commands, in particular after errors. #3553
• Fix integer overflow vulnerabilities in regex compiler (CVE-2026-107891). #3585
• Compiler crashes when emitting line-tables debug info. #3581
• Do not pass PIE/PIC on Windows. #3586
• Crashes in slice->vector and ptrvec->intvec conversions.
• Implicit cast rules too loose for slice->arrptr.
• Invalid lowering for (Bitstruct)-1 where the bitstruct is backed by an unsigned type.
• Cast to any from typedef with an underlying inline pointer, would cause a crash.
• Cast from constant bool[<2>] to float[<2>] would crash.
• Padding would be incorrectly lowered in things like (char[<8>])(char[8])"abc".
• Const casting a const slice derived from a zero array initializer (e.g. (bool)(int[])(int[3]){}) would incorrectly return false.
• Casting a constant to an alias to an enum would be incorrectly checked.
• Cast to aliased bitstruct would fail when non-aliased cast would succeed.
• Casting an optional distinct index would incorrectly fail.
• Aarch64 MacOS did not zero/sign-extend args (and Win32 Aarch64 incorrectly did). #3593
• Overaligned structs were not passed correctly on Linux Aarch64.
• Add alignstack for Linux Aarch64 ABI for correct HFA/HVA handling.
• string::unescape and string::escape would not work correctly after initial size was reached. #3592
• $defined on an untyped list that cannot be converted, e.g. $defined(int[*] y = $list) with var $list = { 1, "a" };, would print an error instead of being false.
• An empty constant string could not be converted to an array, e.g. char[4] x = "";.
• A constset inlining a constset that in turn inlines a third one, could not be implicitly converted to the one in the middle.
• $defined on an implicit pointer conversion was true for pointers to distinct types, and printed an error instead of being false when the pointee alignment would increase.
• Casting a typeid or fault to a wider signed integer, e.g. (int128)t, would sign extend instead of zero extend.
• An implicit cast between unrelated distinct types would report the error twice.
• The hints in cast errors (for enums, distinct types, ZString and pointers to slices) were missing when the value was optional.
• Casting between float16 and bfloat, also in vectors, would abort the compiler. An explicit cast now converts through float.
• Casting a constant float to an integer it does not fit, e.g. (int)1e30, (char)300.5, infinity or NaN, would abort the compiler or give a wrong value. It is now an error, also for the elements of a constant float vector or array cast to an integer vector.
• An inline distinct signed integer, or an inline constset of one, would implicitly convert to an unsigned integer of the same size, e.g. uint y = my_inline_int;, which int to uint does not allow.
• Explicitly casting a pointer wrapped in an inline distinct type to any or an interface was an error, even though the implicit cast worked.
• A struct implementing an interface through an alias, e.g. struct Foo (IAlias), would not convert to the interface, and an interface inheriting through an alias would not convert to its parent.
• Casting an enum to a distinct type over that enum, e.g. (MyFoo)foo with typedef MyFoo = Foo;, was an error. It is now allowed with an explicit cast, like for a distinct type over an integer.
• Errors in multithreaded compilation tasks, such as a failing c-sources file, could crash the compiler. A failing task now aborts the remaining tasks and the compiler exits with an error. #3584
• Fix error message rethrowing in a non-function context.
• Fix name mangling of untyped list.
• "max-vector-size" was not whitelisted as a project property.
• In some cases, optional ref was extracted before it was created during LLVM IR codegen #3598.
• Bitstruct method did not check if field with the same name existed #3602.
• Casting a constant string (or constant bytes) to a vector at compile time would cause a crash. #3601
• Support unicode paths on Windows when running processes from the compiler. #3605
• Fixes to i128 compiler_rt.
• Using a const from another module that refers to a private symbol would fail to link in some cases e.g. static List{int} g = list::ONHEAP{int};.
• Reading an enum associated value at runtime from another module would fail to link if it referred to a private symbol.
• A function only used in an @ensure could be stripped when it wasn't the first statement to run on return, failing to link.
• The sign of constant NaN expressions was ignored during codegen.
• Generic type inference fails with optional expected types #3612.
• Signed shift amounts narrower than the shifted value were zero extended affecting the panic message of an invalid shift amount.
• A function with @test @benchmark would effectively never be included.
• Incorrect compiler error using from_ordinal when exceeding max.
• Concatenated base64 literals lost decoded bytes after the first chunk.
• Improve accuracy of log and logf without a C library (wasm32, freestanding, or -D C3_MATH).
• 16-byte align JmpBuf on Win32.
• Win64 returned @simd vectors through a hidden pointer instead of XMM0. #3628
• Fix regression when slices or pointers exceed 2 GB. #3595
• XmlSerializer.start_doc was missing the closing quote after the encoding.
• XmlSerializer escaped comments and processing instructions when serializing a document.
• gfx::get_vulkan_physical_device_presentation_support was broken.
• argon2::tencode took key_id as a char instead of a char[].
• The libc allocator was incorrectly declared in std::core::mem::allocator.
• Fixes to the benchmarks.c3, map.c3 and xml/3_xml_write_stream.c3 examples.
• A global slice of vectors, e.g. int[<2>][] g = { { 1, 2 } };, would be lowered incorrectly.
• Instantiating a generic with the type of an anonymous inner struct/union would crash the compiler. #3631
• Anonymous inner structs/unions of the same parent shared a name. They are now named by member index, e.g. Foo.0.
• Inner structs/unions of a generic struct lacked the generic parameters in their name.
• Fix Win32 compiler support for paths containing non-ASCII characters. #3606
• A constant struct containing a union initialized by a smaller member was emitted too small.
• Codegen for % on float vectors was incorrect.
• /, % and >> on unsigned integer vectors used signed operations.
• Assigning to v[^i] on a vector with a non-constant i wrote the wrong element.
• Assigning to a vector element or swizzle discarded would not evaluate right to left.
• ++ and -- on a multi-component swizzle, e.g. v.xy++, did not change the vector.
• Comparing constant bitstructs could crash the compiler.
• Constant != on faults and on addresses of globals returned the result of ==.
• a += e on an optional a did not assign the fault when e was a fault. It now works like a = a + e.
• Constant folding of int::min / -1 was incorrect.
• return, break, continue, nextcase and rethrow in a macro's trailing body were allowed when the macro invoked the body from a defer.
• ! on a non-bool vector gave the wrong lanes when stored into a bool vector.
• The $$overflow_add, $$overflow_sub and $$overflow_mul were lowered incorrectly for vectors.
• ! on an optional vector lost the optional and could cause the compiler to crash.
• math::overflow_add and math::overflow_sub failed on vectors.

Stdlib changes¶

• Added termios support for OpenBSD.
• Add file::open_fd to open files directly from file descriptors.
• Changed regex fault and flag names, removed NOSUB. #3585
• Added missing __muloti4 to compiler_rt.
• List.insert_at and List.push_front now use a single memmove instead of an element-by-element shift. List.push has an inline fast path when there is spare capacity.
• List.remove_item, List.compact (and the same functions on FixedList) are now single pass instead of quadratic.
• Cryptos now try to use funnel shift and rotation.
• Add unordered parameter to List.remove_first and .remove_last
• Added matrix::perspective_infinite, matrix::perspective_reverse_z and matrix::perspective_infinite_reverse_z.
• Added column, row, set_column and set_row to Matrix2x2, Matrix3x3 and Matrix4x4, and scale_factors to Matrix3x3 and Matrix4x4.
• math::sign now supports floats.
• Add chmod support #3536.
• std::hash::crc32 now uses slicing-by-16, about 5x faster. This also speeds up gzip and zip.
• std::hash::adler32 update is about 3.5x faster (blocked modulo and portable vectors), results unchanged.
• Rewrite std::compression::deflate inflate: table driven decoder, 2-3x faster decompression; truncated input is now rejected and Inflater.init works with the default bit_reader_buf.
• Add an optional level (1-9) to std::compression::deflate compress, compress_stream and Deflater.init; the default level produces the same output as before. Faster deflate encoding, and incompressible data is stored instead of expanded.
• Add an optional level to std::compression::gzip compress, compress_stream and GzipWriter.init (sets XFL per RFC 1952); the default is unchanged.
• Add an optional level to std::compression::zip ZipArchive.open_writer and ZipArchive.write_file (DEFLATE only); the default is unchanged.
• Fix several std::compression::zip problems: open and recover closed the file twice on errors, open failed on ZIP64 archives with a comment or exactly 65535 entries, rejected data after the end record, and silently dropped entries after a broken central directory header. recover now reads ZIP64 local headers, extract verifies the CRC32 and no longer overwrites the modification time, ZipEntryReader.seek rejects negative positions, and the central directory flags match the local header.
• Added XmlSerializer.add_raw to write unescaped text.