Reimplement the Linux HAL in native Swift using @c

[lhoward]

Summary

Migrates the entire swifthal_* HAL from ~2,900 lines of hand-written C to native Swift, exporting the identical C ABI via the Swift 6.3 @c attribute (SE-0495). SwiftIO/AsyncSwiftIO link against it unchanged — no caller sees the language change.

Structure

• The single C target is split into CLinuxHalSwiftIO (residual C) and a new Swift target LinuxHalSwiftIO (the public product). The Swift target @_exported imports the C module, so import LinuxHalSwiftIO still surfaces the HAL declarations to consumers.
• Every subsystem is now Swift: spi, i2c, uart, gpio, os, platform, fs, timer, counter, and the adc/eth/i2s/lcd/pwm/wifi stubs.
• gpio and timer use the Swift Dispatch overlay (no libdispatch/Blocks C runtime); gpio drives libgpiod directly. The interrupt "block" entry point — consumed only by AsyncSwiftIO — becomes a native Swift closure API instead of an Objective-C block.

Minimal C that remains

Things Swift genuinely cannot express, named with the existing swifthal_<subsys>__<helper> internal convention:

• swift_platform_hwcycle.c — hwcycle_get/hwcycle_to_ns (inline assembly).
• swift_uart_native.c — termios2 line configuration (asm/termbits.h collides with Glibc's termios.h, so it can't be exposed to Swift).
• swift_hal_native.h — static inline wrappers for ioctl request-number macros (SPI/I2C), variadic mq_open, syslog, plus SPI mode-bit constants.

Pre-existing bugs fixed along the way

• os mq_recv tested != 0, but mq_receive returns a byte count, so it returned -errno on every successful receive; corrected to < 0.
• The RNG was defined as swiftHal_randomGet, which never matched the declared swifthal_random_get symbol and so could not be linked; renamed to match the ABI.
• The SPI loopback test built SPI in a stored property, crashing test discovery on any machine without SPI hardware; made it lazy.

Toolchain

CI bumped swift:6.2 → swift:6.3 (@c requires 6.3).

Verification

• swift build (library, examples, tests) is green.
• Symbol audit: each swifthal_* symbol is defined exactly once — Swift for logic, C only for hwcycle + termios2 helpers; no duplicates.
• HALNativeTests (semaphore limit enforcement, mutex, RNG fill, uptime) pass; an mq round-trip / sleep smoke check passed during development. The SPI loopback test still requires real hardware.

Post-review fixes

A multi-agent code review of the branch produced eight confirmed findings. Five were migration regressions, fixed in 31af742:

• uart blocking read (timeout == -1) looped forever instead of returning after one readBuffer pass, as the C did.
• spi transceive used a trapping UInt32(...) on the transfer length; uart likewise trapped converting nremain to CInt. Both now truncate, matching C.
• The wifi and lcd stubs stopped zeroing their out-params before returning -ENOSYS, leaving callers reading stack garbage.

Two further findings were faithful ports of the original C behavior that were nonetheless worth improving, done separately in e58e298:

• fs readdir aborted the whole enumeration when a DT_REG entry lost a race with unlink; it now skips the entry, matching the DT_UNKNOWN path.
• gpio interrupt_callback_uninstall only cancelled the dispatch source, so a later interrupt_enable silently resumed a permanently-cancelled source; it now releases it, and the later call returns -EINVAL.

One finding is deliberately unaddressed: fs_write/fs_read truncate a single transfer larger than 2 GiB. That matches the original (int)nbytes and is not a realistic single-call scenario, so it stays a pure port.

🤖 Generated with Claude Code