What EmbLinkRTOS needs from EmbCC (tracking)

Start with the linked RTOS evidence, especially docs/engineering/hil-esp32-port-2026-11-05.md, docs/engineering/embcc-native-linux.md, and the named kernel paths. The checklist spans compiler builtins, host support, build integration, several architectures, and language conformance; first identify one item to split into its own issue, then verify its current status and define a focused completion criterion. The inventory alone does not specify a bounded change or test for the full tracking issue.

This tracks what EmbLinkRTOS still needs from EmbCC. It is an inventory of the RTOS repository as of f76cc7c (2026-11-08). The RTOS's CI pins EmbCC 0.0.1-beta. Each item gives the evidence on the RTOS side and the workaround the RTOS uses today. Items can be split into their own issues when work starts.

RTOS links below are relative to https://github.com/EmbLinkOs/EmbLinkRTOS/blob/f76cc7c/.

1. Xtensa code speed

The only EmbCC runtime measurements are on the ESP32. The same kernel and port built by EmbCC take about 1.8 times GCC's cycles.

Checked build, 40 MHz, in cycles (docs/engineering/hil-esp32-port-2026-11-05.md):

	EmbCC	GCC	Clang
notification hand-over	2090	1163	1325
semaphore hand-over	2075	1154	1250
semaphore give + take	436	226	218
interrupt wake	4211	2500	2697

Release build, 160 MHz, against FreeRTOS on the same board (docs/engineering/bench-esp32-freertos-2026-11-06.md):

	FreeRTOS (GCC)	EmbLink GCC 16.1	EmbLink Clang 22.1.4	EmbLink EmbCC
notify_handover	1535	969	1135	1749
sem_handover	1959	973	1056	1742
sem_give_take	316	208	200	406
isr_wake	3201	1951	2105	3247

Micro-costs under EmbCC: emb_thread_yield() with nothing to switch to costs 196 cycles; a software-interrupt round trip costs 318 (docs/ports/xtensa.md).

• Compare EmbCC's and GCC's code for the scheduler (kernel/sched/sched.c) and the wait path (kernel/wait/wait.c), the hottest paths in these numbers.

2. Builtins the RTOS's compiler layer would use

include/emb/compiler/embcc.h defines none of the following, so the portable fallbacks in include/emb/compiler.h apply. If EmbCC already supports some of them, say which, and the RTOS will wire them in embcc.h (an RTOS change).

• __builtin_clz, __builtin_ctz, __builtin_popcount. Without them, the scheduler's ready-set lookup (kernel/include/embk/word.h, called from kernel/sched/sched.c) is a loop of up to 32 steps instead of one instruction. This may account for part of §1; it is not measured.
• __builtin_return_address(0). The crash record takes a software fault's pc from it (kernel/fault/fault.c), so under EmbCC that pc is 0.
• __builtin_constant_p and __builtin_unreachable.
• format(printf, …) checking, deprecated, and a C fallthrough attribute. Each is empty under EmbCC today.

3. Linux host compilation

EmbCC builds only the kernel on Linux. The RTOS's native port, board and tests are built by GCC, which also links: the "mixed" build of ADR-041 (docs/engineering/embcc-native-linux.md). That is the only EmbCC leg in the RTOS's CI.

• pthreads in the x86_64-linux-gnu runtime. The error today is native_port.c:20: error: cannot find include file "pthread.h".
• Accept __attribute__((__mode__)), so glibc's headers compile. glibc's sys/types.h:164 has typedef int register_t __attribute__ ((__mode__ (__word__)));.
• Ship a Linux crt1.o. The 2026-10-14 snapshot failed with no crt1.o for x86_64-linux-gnu; check whether that is still the case.

4. Build-system integration

• CMake identification. CMake reports "The C compiler identification is unknown" and "Did not find file Compiler/-ASM", so it requests no dependency files. A stale 40 MHz tick configuration survived a rebuild because of it. The RTOS now forces -MD -MT -MF itself (cmake/toolchains/embcc-depfile.cmake). An upstream Compiler/EmbCC-*.cmake module, or a GCC-compatible identification, would fix this.
• A dependency file for assembly sources. The RTOS adds OBJECT_DEPENDS for xt_vectors.S by hand (arch/xtensa/CMakeLists.txt).

5. Per architecture

Xtensa

• A release with embld's Xtensa script support (D-024, #174). Until then, the ESP32 bring-up probe falls back to -Ttext/-Tdata.
• A release with the working embflash --esp (#173). The installed 0.0.1-beta build sent FLASH_END, the ROM answered error 0x06, and embflash exited 2.
• Directives in Xtensa inline assembly. Only instructions are accepted, so EMB_LAYOUT_MARKER uses a data form on __XTENSA__ (ADR-054, A95).
• Stop warning on unused static inline functions, as GCC and Clang do not. The RTOS marks three trace helpers EMB_UNUSED for it.

Together, the first two would let the RTOS add an EmbCC Xtensa CI leg.

AVR

• .debug_frame: EmbDBG backtraces stop at frame 0.
• -g code-size growth: -O0 code may not fit the ATmega328P.
• C++ on 16-bit targets ("the C++ lowering assumes a 32-bit int and 4-byte pointers").

There are no EmbCC size or speed numbers for AVR yet. Wiring __flash/progmem into the RTOS's flash macros and linking the conformance suite is a joint RTOS-side session (docs/ports/avr.md).

RISC-V

• Inline assembly: lr/sc/amo*, the la/call/tail pseudo-instructions, and floating point.
• The interrupt attribute. The RTOS uses an assembly trap entry.

Cortex-M

• CMSE with an FPU: cmse_nonsecure_entry functions must clear the floating-point registers.
• Armv6-M atomic helpers mask with PRIMASK, which has no effect in unprivileged Thread mode.
• Cortex-M23: checked only as an M33 with a Baseline instruction scan, never on an M23.

6. Language conformance and diagnostics

• Enforce -std= (today accepted and ignored), so ISO conformance can be checked.
• Refuse _Thread_local/__thread on embedded targets, or implement it. Today each silently becomes one shared instance.
• __STDC_HOSTED__ should be 0 under -ffreestanding; it is 1.
• Complete _Alignas, and _Atomic beyond integer and pointer types.
• The warnings the RTOS builds with: -pedantic/-Wpedantic, -Wcast-align, -Wnull-dereference. Each is reported as "not a warning EmbCC has", 390 times across the RTOS's archived build logs.
• Inline assembly: labels, wider-than-register outputs, callee-saved clobbers.
• Inlining of functions that return a struct. The RTOS's time constructors are macros for this reason (include/emb/time.h, SPEC-001). The current toolchain profile no longer lists it, so it may be fixed; please confirm.
• Build EmbCC warning-clean with GCC 14.2: a missing readlink declaration, then format-truncation and empty-body warnings.

Already done upstream

These need no work. The ten items of the RTOS's 2026-10-14 capability review are each done within their stated scope:

• Armv8-M stack limits;
• AVR linker scripts, __flash/progmem, DWARF 4 and the EmbDBG register layout;
• small RV32/AVR atomics;
• soft-float CMSE;
• Cortex-M23;
• SVD clusters/arrays/JSON;
• C++ without exceptions.

Constructor priorities, typedef alignment and 8-byte atomic helpers are also done.

The RTOS side still owes cleanup: some of its specifications describe these as missing, and its Xtensa default toolchain is EmbCC although its CI builds Xtensa only with GCC and Clang. That is being fixed there.