kit, branch HEAD

windows: cross-compile kit + run cc/JIT natively on aarch64-windows

2026-06-09T15:13:24Z

commit 128504789846f077dc83cd5a54b2d05dbf9083a0 parent f20d82f2139cc962ceffd8060d2e54b39e8c329f Author: Ryan Sepassi Date: Tue, 9 Jun 2026 08:13:24 -0700 windows: cross-compile kit + run cc/JIT natively on aarch64-windows Bring up kit as a self-hosted Windows toolchain: cross-compile the kit binary on the dev host (build/kit cc -target aarch64-windows) into a PE/COFF kit.exe, then run `kit cc` and `kit run` (JIT) on a Windows VM. Verified on an aarch64 Windows 11 VM: kit.exe --help; `kit cc hello.c` -> a working hello.exe (correct stdout + exit code); and `kit run` (JIT) of compute / data / extern-call / puts programs (libc I/O via dlsym). Fixes, each a real gap surfaced by the bring-up: - c frontend: honor GCC asm-label symbol renames on declarations (`extern T f(...) __asm__("alt")`). Parsed before but ignored except for register locals; the emitted/referenced linkage symbol now uses the label (Decl.asm_name) while the C name stays for diagnostics. mingw's redirects e.g. time->_time64 this way; without it the link had an undefined `time`. - coff: honor the Microsoft short-import NameType field (was `(void)name_type`). NOPREFIX/UNDECORATE strip decoration; EXPORTAS carries the real DLL export name separately. The PE hint/name table now uses that import name, not the local symbol -- UCRT aliases e.g. local `__msvcrt_assert` onto export `_assert`; emitting the alias name made kit.exe fail to load (STATUS_ENTRYPOINT_NOT_FOUND). - driver/lib/hosted.c: add libwinpthread to the Windows hosted link. 's pthread_time.h inline wrappers reference nanosleep64 / clock_*64, which live there (llvm-mingw ships it as a core lib); static archive, so it adds no runtime DLL dependency. - rt/include/setjmp.h: on Windows, asm-rename setjmp/longjmp onto mingw's non-SEH __mingw_setjmp/__mingw_longjmp (libmingwex). kit's freestanding code wants pure register save/restore, and mingw exposes no bare `setjmp` symbol (it is a macro over _setjmpex/__mingw_setjmp). - driver/env/windows.c: - gate the SEH __try guarded-copy under !defined(__kit__) (kit's C frontend implements no SEH); rely on the VEH backstop already there. - JIT dlsym fallback for the libucrt-static __local_stdio_{printf, scanf}_options helpers (referenced by mingw's printf/scanf inlines, not DLL-exported); hand JIT'd code kit.exe's own copies. - map the JIT execmem dual-map's runtime/exec view FILE_MAP_WRITE so writable runtime segments (an import GOT, .data) can be VirtualProtect'd PAGE_READWRITE. A FILE_MAP_EXECUTE-only view rejects RW (err 87), which broke `kit run` of any program with an external call or writable global. - scripts/windows_cross.sh: reproducible cross-build harness (toolchain wrappers + the HOST_OS=windows make invocation). Known gap: printf-family via `kit run` still needs libucrt's static `printf` (not a DLL export) and the JIT links no static archives; AOT `kit cc` printf works fully, as do JIT puts()/external calls. Host regression: test-coff/link/elf/cg-api/parse/toy green. test-macho 36_tls_basic/J is a pre-existing JIT-TLS failure (confirmed failing at HEAD via a clean worktree build).

env/windows: include windows.h before psapi.h (mingw cross-build)

2026-06-09T06:27:59Z

commit f20d82f2139cc962ceffd8060d2e54b39e8c329f parent ea01f03f2e47071ce36aae30a6de7f31bc5d9c8d Author: Ryan Sepassi Date: Mon, 8 Jun 2026 23:27:59 -0700 env/windows: include windows.h before psapi.h (mingw cross-build) mingw's / use windows.h types (WINBOOL/DWORD/LPVOID) and do not self-include it, so the alphabetized include order (psapi before windows) failed to compile under the llvm-mingw cross-toolchain. Pull windows.h into its own leading include group so a formatter sort cannot reorder it. Surfaced while standing up an aarch64-windows cross-build of kit; the remaining windows.c host-env build issues (SEH __try, attribute thread, dllimport) are separate and pre-existing.

jit/tls: ELF Local-Exec relaxation via arch-abstracted hook

2026-06-09T05:54:07Z

commit ea01f03f2e47071ce36aae30a6de7f31bc5d9c8d parent d8d8209e1bd898acb3455608478a57b9d2ecfc57 Author: Ryan Sepassi Date: Mon, 8 Jun 2026 22:54:07 -0700 jit/tls: ELF Local-Exec relaxation via arch-abstracted hook Extend the uniform in-image JIT TLS model to ELF (Linux + FreeBSD), the single-threaded JIT resolving each thread-local to its in-image .tdata/.tbss instead of via the host thread pointer (mrs tpidr_el0 / add tp / mov fs:0), which aliased into the host process's own TLS — wrong init and unsafe. This was the documented bootstrap blocker (141_threadlocal_mutate returned 3 instead of 43 on the R lane). Arch-abstracted through the existing relocation seam, no arch switch in link_jit.c: - RELOC_IS_TLS_LE flag on RelocDesc, set on each arch's Local-Exec reloc rows (aa64 TLSLE_ADD_TPREL_*, rv TPREL_HI20/LO12_*, x64 TPOFF32). - LinkArchDesc.jit_tls_le_relax hook; per-arch idiom rewrite lives in each arch's reloc.c beside its reloc_apply_insn encoders: aa64: mrs/add/add -> adrp; add :lo12:; nop rv64: lui/add tp/addi -> auipc %pcrel_hi; nop; addi %pcrel_lo x64: mov fs:0/lea -> nop...; lea rd,[rip+&var] - link_jit.c (initial + append reloc loops) classifies via reloc_kind_is_tls_le and delegates; the old reloc_is_tlsle / reloc_is_x64_tlsle / JIT_TLS_TCB_SIZE host-TP code is removed. Verified on the arm64 host + containers: aa64-linux (native), x64-linux and rv64-linux (podman) all return 141=43 / 142=134 on the R lane at -O0/1/2 and the interp lane. macOS (Mach-O) unchanged: test-toy/test-link/test-opt green. COFF/Windows TLS relaxation is a separate follow-up (different TEB idiom).

jit/tls: uniform in-image thread-local storage; retire Mach-O TLV thunk

2026-06-09T05:06:45Z

commit d8d8209e1bd898acb3455608478a57b9d2ecfc57 parent 70592470957e05df49ed33592c6f9fd68f8c3421 Author: Ryan Sepassi Date: Mon, 8 Jun 2026 22:06:45 -0700 jit/tls: uniform in-image thread-local storage; retire Mach-O TLV thunk Single-threaded JIT collapses thread-local to a single in-image instance: the image's .tdata/.tbss IS the storage. Resolve every TLS access to that storage without touching the host thread pointer (reading host tpidr/fs/tp aliases into the host process's own TLS — wrong init and unsafe). - link_jit: map the TLS segment RW in the JIT (AOT keeps it as the RO init template). Relax the Mach-O TLV access to read storage straight from the descriptor's +16 slot (ldr xN,[xN,#16]; blr -> nop) — no thunk, no descriptor patching, no per-thread block. Init comes free from the copied .tdata. - kit_jit_tls_addr (replaces kit_jit_tlv_resolve): format-aware, range-checked resolution for the interpreter/host (Mach-O reads descriptor[+16]; ELF/COFF the symbol is the storage). - opt: model IR_TLS_ADDR_OF's clobber set via NATIVE_MOP_TLS_ADDR. The Mach-O TLV sequence clobbers x0/x16/x17/lr; the optimizer only modeled its dest, corrupting a value left live in x0 across a second TLS access at -O1+ (multi-var TLS returned 204 instead of 134). Fixes AOT macOS too. - Retire the dead machinery: src/jit/ (tlv thunk asm/stub/header), driver/env/jit_tls_posix.c, the KitJitTls public type + KitJitHost.tls, Windows FLS-backed TLS, and all wiring (~600 lines removed). - test/toy 142_threadlocal_multi: multi-var TLS R/I-lane coverage. - doc/JIT.md: document the uniform model. ELF Local-Exec is not yet converted (still TP-relative; Step 6). Verified green on macOS arm64: test-toy 1388/0/39, test-link 122/0, test-opt, test-cg-api, test-isa, test-aa64-inline, test-smoke-{x64,rv64}.

link: root DSO back-referenced symbols under --gc-sections (freebsd -O1 fixed point)

2026-06-09T01:44:57Z

commit 70592470957e05df49ed33592c6f9fd68f8c3421 parent 4e590650228592d27104243d8512af3ed0cb0c18 Author: Ryan Sepassi Date: Mon, 8 Jun 2026 18:44:57 -0700 link: root DSO back-referenced symbols under --gc-sections (freebsd -O1 fixed point) A dynamic executable defines symbols its shared libraries reference back into it — on FreeBSD, the crt defines `__progname`/`environ` and libc.so.7 has undefined references to them. kit's --gc-sections liveness pass only rooted the entry, init/fini, and retained sections, so when nothing in the executable referenced those crt symbols they were garbage-collected out of the dynsym, and the resulting `-O1` kit failed to load: ld-elf.so.1: /lib/libc.so.7: Undefined symbol "__progname" read_elf_dso now records each DSO's undefined-symbol names on its ObjImage (obj_image_add_undef), and link_gc_compute roots the executable's definitions of them — GNU ld's default "keep what shared libraries need" behaviour, surgical (only DSO-referenced symbols, not all globals, so GC stays effective) and monotonic (only adds roots). The DSO ObjImage is now always created (previously only when versioned) so undef refs are captured for unversioned DSOs (musl) too; versioning is unchanged (musl still emits no .gnu.version_r, glibc still does). With this, the **aarch64-freebsd `-O1` (release) self-build reaches the byte-identical fixed point** — FreeBSD now self-hosts at both -O0 and -O1 (Toy 1378/2/39 through stage3; the 2 fails are the pre-existing JIT-TLS .tdata R-lane gap). test-link/test-elf/test-asm green; musl/glibc --gc-sections links unaffected. Builds on the deferred-tombstone fix in the previous commit.

asm: don't globalize deferred-symbol tombstones in promote_undef_externs

2026-06-09T01:22:39Z

commit 4e590650228592d27104243d8512af3ed0cb0c18 parent 246df83a668460fbe7e2b598a23614ebe68c9c11 Author: Ryan Sepassi Date: Mon, 8 Jun 2026 18:22:39 -0700 asm: don't globalize deferred-symbol tombstones in promote_undef_externs promote_undef_externs (the `as` pass that turns referenced-but-undefined LOCAL symbols into undefined GLOBALs, matching GNU as) iterated every symbol slot — including `removed=1` tombstones — and so flipped the binding of deferred anonymous const-data / jump-table symbols (`.Lkit_ro.N` / `.Lkit_jt.N`) that were still tombstoned awaiting materialization at opt_whole_module_finalize. It was the only obj_symiter consumer not honoring the `removed` contract documented in obj.h. This bit the aarch64-freebsd `-O1` self-build: FreeBSD's injects a file-scope `__asm__(".symver ...")` (via __sym_compat) whose replay runs this pass *before* the deferred data is materialized, so the four hosted driver/env/*.o each surfaced a defined global `.Lkit_ro.0` and the stage2 link aborted with `duplicate definition of global symbol '.Lkit_ro.0'`. Not FreeBSD-codegen-specific — any TU with a file-scope asm plus a deferred const-data symbol at -O1 reproduces it on any target (e.g. aarch64-linux-musl). Fix: skip tombstones in the loop. The four env objects now show zero global .Lkit symbols and the stage2 link succeeds. test-asm/test-link/test-elf green; -O0 is provably unaffected (deferral, hence tombstones, only exist at -O1). The freebsd -O1 chain now gets past the stage2 link but hits a separate, pre-existing blocker — `-Wl,--gc-sections` drops crt `__progname`/`environ` that libc.so.7 needs — documented in doc/plan/BOOTSTRAP.md as the next fix.

doc/bootstrap: record aarch64-linux and aarch64-freebsd bootstrap state

2026-06-09T00:18:13Z

commit 246df83a668460fbe7e2b598a23614ebe68c9c11 parent ad732e4b9733ddb81da3f861b595c7e1e6bd1388 Author: Ryan Sepassi Date: Mon, 8 Jun 2026 17:18:13 -0700 doc/bootstrap: record aarch64-linux and aarch64-freebsd bootstrap state - aarch64-linux (musl + glibc) baseline: both -O0/-O1 reach the fixed point, via the Linux-container path; notes the JIT-TLS / objdump-golden / C-backend Toy gaps that are not bootstrap blockers. - aarch64-freebsd: -O0 chain reaches the byte-identical fixed point (Toy 1371/9/39); needed -rdynamic acceptance + ELF Verneed/Versym emission (the fstat INO64 FBSD_1.0-vs-1.5 mis-bind). The -O1 release chain is blocked by a pre-existing FreeBSD-target deferred-const-data GLOBAL-binding bug (.Lkit_ro/.Lkit_jt); gate on bootstrap-debug until fixed.

elf: emit symbol-version (Verneed/Versym) requirements; freebsd -rdynamic + bootstrap shipping

2026-06-09T00:17:37Z

commit ad732e4b9733ddb81da3f861b595c7e1e6bd1388 parent 704d2a298814839a5fdc17620b078f5e12dc34e3 Author: Ryan Sepassi Date: Mon, 8 Jun 2026 17:17:37 -0700 elf: emit symbol-version (Verneed/Versym) requirements; freebsd -rdynamic + bootstrap shipping kit's linker emitted unversioned undefined references (`UND fstat`), so on FreeBSD the runtime bound the hidden compat `fstat@FBSD_1.0` (pre-INO64 `struct stat`) instead of the default `fstat@@FBSD_1.5`, reading `st_size` at the wrong offset. A kit-built stage2 then failed to read its own source files, blocking the aarch64-freebsd self-build at the stage3 compile. The ELF linker now reads each DSO's `.gnu.version_d`/`.gnu.version` and emits a matching `.gnu.version_r` + `.gnu.version` + DT_VERSYM/VERNEED/VERNEEDNUM, selecting each import's default (non-hidden) version. Gated on the DSO carrying versions, so musl/static links stay byte-identical; glibc links now also carry correct GLIBC_* requirements (verified: glibc/musl/freebsd binaries run; host test-link/elf/smoke/cg-api/ar/debug/dwarf green). With this the aarch64-freebsd -O0 chain reaches the byte-identical fixed point and the Toy corpus runs 1371/9/39 through the bootstrapped stage3. - src/obj/obj.h: ObjImageSym.version (a DSO export's default-version name) - src/obj/elf/elf.h: DT_VER* tags, Verdef/Verneed wire sizes, VERSYM_* flags - src/obj/elf/read.c: read_elf_verdefs() + versym parse in read_elf_image AND read_elf_dso (the linker's DSO reader now builds an ObjImage to carry versions) - src/obj/elf/link_dyn.c: build_versions() — per-import requirements, versym + verneed bytes (only .dynstr offsets/indices, final at layout time) - src/obj/elf/link.c: emit DT_VER* + SHT_GNU_VERSYM/VERNEED section types - driver/cmd/cc.c: accept -rdynamic/-export-dynamic (FreeBSD's HOST_ENV_LDFLAGS passes -rdynamic; kit's ELF linker already exports DSO-referenced symbols) - scripts/freebsd_bootstrap.sh: ship working-tree content of all tracked files with COPYFILE_DISABLE=1 (drops macOS AppleDouble `._*` xattr sidecars that FreeBSD tar would extract as bogus `._*.c` sources); Toy hook uses bash and no longer aborts before Toy when a chain fails The -O1 release chain is still blocked, before the fixed-point check, by a pre-existing FreeBSD-target codegen bug (deferred const-data .Lkit_ro/.Lkit_jt symbols emitted GLOBAL instead of LOCAL; LOCAL on linux/macos). Not addressed here.

bootstrap: aarch64-linux self-host (musl + glibc) + Toy ELF fixes

2026-06-09T00:12:08Z

commit 704d2a298814839a5fdc17620b078f5e12dc34e3 parent 632517297055ab3a43bccd588d04211a315c97d1 Author: Ryan Sepassi Date: Mon, 8 Jun 2026 17:12:08 -0700 bootstrap: aarch64-linux self-host (musl + glibc) + Toy ELF fixes Reach the byte-identical bootstrap fixed point on aarch64-linux at -O0 and -O1 for both musl and glibc, run natively inside an arm64 Linux container (scripts/linux_bootstrap.sh; make bootstrap-linux[-glibc]). Build: - bootstrap.mk: drive kit-as-compiler (stages 2/3) with -lc on every host, so kit resolves its hosted libc/sysroot via its own profile + host probe (replaces Darwin's -isysroot; unifies linux/freebsd/macos). - scripts/linux_bootstrap.sh: container runner (alpine=musl, debian=glibc). C frontend / preprocessor (glibc + Linux-UAPI headers; musl's ISO-C headers never exercised these): - pp: predefine __{INT,UINT}{64,MAX}_C_SUFFIX__ so kit-compiled TUs get correctly-typed 64-bit constants from rt/include/stdint.h. - pp: erase __extension__ and map __signed__/__volatile__/__const__ to the canonical keywords (unconditional; were Windows-only). - pp: support GNU named variadic macro params `args...` (). - pp_directive: parse_pp_int accumulates in u64 (signed overflow was UB). - cg_adapter: fix a 1-byte OOB write in pcg_aux_clear (pad[5] on u8 pad[5]). C backend + Toy harness (ELF / host-cc portability): - c_target: emit `va_list x = {0}`, not `= 0` (va_list is a struct/array on aarch64 / x86-64 Linux). - toy/run.sh: L lane links hosted (kit cc -lc) on non-Darwin (ELF needs a crt _start); prefer .objdump.elf goldens on non-Darwin; skip musttail+sret C cases on gcc hosts (gcc rejects musttail on sret returns; clang accepts); add -Wno-tentative-definition-incomplete-type (valid C; clang-14 pedantry). - new fixtures: 3 .objdump.elf goldens + 2 .cbackend.gcc.skip. Verified: glibc(clang) Toy 1378/2/39, musl(gcc) 1375/3/41. Remaining fails are the tracked JIT-TLS .tdata-init (141_threadlocal_mutate) and a rare archive-iterator heisenbug; both deliberately left as follow-ups.

freebsd: bootstrap support + portable stdint.h INT64_C

2026-06-08T22:00:13Z

commit 632517297055ab3a43bccd588d04211a315c97d1 parent 2a2dc7af1f422ffa8f7bf62f229b9f448f65a380 Author: Ryan Sepassi Date: Mon, 8 Jun 2026 15:00:13 -0700 freebsd: bootstrap support + portable stdint.h INT64_C rt/include/stdint.h: replace __INT64_C(c)/__UINT64_C(c) direct calls with __KIT_XPASTE(c, __INT64_C_SUFFIX__) suffix-paste approach. FreeBSD clang 19 (aarch64-freebsd15, LP64) does not predefine the function-like __INT64_C/__UINT64_C macros, causing "call to undeclared function" errors on any translation unit that uses UINT64_C(). All supported compilers (clang 14+, GCC 8+) predefine __*_C_SUFFIX__ tokens, making the paste approach portable and unambiguous. mk/bootstrap.mk: add FreeBSD to BOOTSTRAP_KIT_TOOLCHAIN_FLAGS (-lc needed to wire hosted includes, same as Linux); portable BOOTSTRAP_SHASUM (sha256sum on Linux/FreeBSD, shasum on macOS); bootstrap-freebsd target. scripts/freebsd_bootstrap.sh: new -- native three-stage self-build via the running FreeBSD aarch64 VM (mirrors linux_bootstrap.sh but SSHes into the VM instead of using podman). Requires gmake on the VM.

freebsd: x64/rv64 cross-compile, -lm/-lpthread, hosted lib search

2026-06-08T21:13:57Z

commit 2a2dc7af1f422ffa8f7bf62f229b9f448f65a380 parent 28f2a06838d8f794b707e3e6dd2f70eadd14fc02 Author: Ryan Sepassi Date: Mon, 8 Jun 2026 14:13:57 -0700 freebsd: x64/rv64 cross-compile, -lm/-lpthread, hosted lib search 1. Fix cross-compile for non-aarch64 hosts: gate LIB_SRCS_JIT_ASM on HOST_ARCH==aarch64 in mk/lib_srcs.mk; add LIB_SRCS_JIT_STUB (src/jit/tlv_thunk_stub.c) for other hosts, with a matching compile rule in Makefile. 2. Fix -lm / -lpthread / other user-specified -l flags not found when using the hosted profile: - Add lib_search_dirs[] to DriverHostedPlan so the resolved hosted lib dirs (e.g. /usr/lib, /lib on FreeBSD) are exposed to callers. - driver_hosted_resolve: transfer libdir ownership from DriverHostedDirs into the plan; driver_hosted_plan_fini frees them. - cc.c: populate lib_search_paths from plan lib_search_dirs in cc_apply_hosted_profile; move cc_resolve_pending_libs to after cc_apply_hosted_profile so hosted dirs are present when -l flags are resolved. - ld.c: pre-scan for -lc in ld_parse first pass and call driver_hosted_dirs_resolve to pre-populate o->lib_dirs before the inline -l resolution loop; also add hosted lib_search_dirs in ld_apply_hosted_before_after. 3. Mark RELEASE.md items done: FreeBSD target parsing/serialization, FreeBSD hosted profiles (arm64/x64/rv64 cross-compiled + VM-verified), and FreeBSD ucontext/register marshalling. Verified on FreeBSD 15 VMs (aarch64, amd64, riscv64): kit cc -c / kit cc / kit ld -lc / kit cc -lm / kit cc -lpthread / kit ld -lm / kit cc -static all pass. test-freebsd: 6/6 pass (static+dynamic metadata for all three arches).

ld: default to PIE for hosted dynamic links (-lc without -static)

2026-06-08T20:40:41Z

commit 28f2a06838d8f794b707e3e6dd2f70eadd14fc02 parent f082bcd0c6e5cfd8f98e04e80dec19f6a388f65b Author: Ryan Sepassi Date: Mon, 8 Jun 2026 13:40:41 -0700 ld: default to PIE for hosted dynamic links (-lc without -static) kit ld -lc produced "imported sym has no PLT entry (CALL26)" because the link_dyn path (PLT/GOT/.rela.plt) is only entered when lopts.pie is set, and kit ld intentionally does not default to PIE the way kit cc does. Fix: when the hosted profile is active (-lc) and the link is dynamic (not -static, not -shared, not -r), implicitly set pie=1 — matching what kit cc does. Explicit -no-pie with -lc is rejected with a clear error since non-PIE ET_EXEC dynamic linking is not yet supported.

freebsd: thread OS version, add cross-compile + VM verification

2026-06-08T20:28:43Z

commit f082bcd0c6e5cfd8f98e04e80dec19f6a388f65b parent 32b0908794ee78d4cfd76dea8a2856b6c45d6fe4 Author: Ryan Sepassi Date: Mon, 8 Jun 2026 13:28:43 -0700 freebsd: thread OS version, add cross-compile + VM verification - KitTargetSpec gains os_version_major (parsed from triple e.g. freebsd15.0→15) so hosted_add_freebsd_defines emits the real __FreeBSD_version instead of hardcoded 14. - Add ucontext_t <-> KitUnwindFrame marshallers for all three arches (aarch64, x86_64, rv64) and wire them into mk/env.mk. - mk/env.mk: add HOST_ENV_LDFLAGS (FreeBSD: -rdynamic so libc.so can resolve `environ` from the exe); mk/flags.mk splices it into HOST_LDFLAGS. - scripts/freebsd_sysroot.sh: pull libthr.a/.so.3 and libpthread.so.3 from base.txz and create the libpthread → libthr symlinks that -lpthread needs. - scripts/kit_freebsd.sh: cross-compile kit for a FreeBSD target arch via a wrapper script (avoids GNU make CC whitespace-split bug), auto-fetches the sysroot, supports `scp` and `run` to deploy via the FreeBSD VM. - scripts/freebsd_vm.sh: add `scp ` subcommand. - driver/env/freebsd.c: remove UNTESTED markers (verified on FreeBSD 15 aarch64 VM). - doc/plan/FREEBSD.md: deleted (all work complete).

docs: close out and remove the CODEGEN plan doc

2026-06-08T20:13:46Z

commit 32b0908794ee78d4cfd76dea8a2856b6c45d6fe4 parent f5a4f04e3bf6b4fd9bd8929bfc657c3bf7be9e28 Author: Ryan Sepassi Date: Mon, 8 Jun 2026 13:13:46 -0700 docs: close out and remove the CODEGEN plan doc The CG API interface-cleanup roadmap is closed: PLACE/VALUE (Track 7), op/intrinsic taxonomy (Track 4), atomic/order/AsmDir + fold isolation (Tracks 2-partial/6), and bitfields-as-place (3b) had landed; this pass removed multi-value from the API (single result), added the conditional control-op tests + block-continue guard (Track 1c), and dropped the dead INTRIN_MEMCPY/MEMSET (Track 4 tidy). Track 2's remaining binop/cmp internal-enum dedup is intentionally not done: its only behavioral win (lossy ordered/unordered FP compare) already landed, and the merged internal BinOp/CmpOp/UnOp + the single isolated api_map_* is a defensible normalized internal vocabulary woven through the i128/wide64/fold/delayed-arith lowering. Remaining = zero-behavior churn, so the doc is retired rather than kept open for it.

Track 4: drop dead INTRIN_MEMCPY/INTRIN_MEMSET intrinsics

2026-06-08T20:03:23Z

commit f5a4f04e3bf6b4fd9bd8929bfc657c3bf7be9e28 parent 80ba21ae01667332bf937b70d209a0e73d690938 Author: Ryan Sepassi Date: Mon, 8 Jun 2026 13:03:23 -0700 Track 4: drop dead INTRIN_MEMCPY/INTRIN_MEMSET intrinsics memcpy/memset are dedicated public ops (kit_cg_memcpy/_memset) that lower to the copy_bytes/set_bytes CgTarget hooks -> CG_IR_AGG_COPY/AGG_SET -> the backends' agg-copy/set paths (and IOP_AGG_COPY/SET in the interpreter). Only kit_cg_memmove flows through the generic intrinsic hook (INTRIN_MEMMOVE). So the internal IntrinKind values INTRIN_MEMCPY and INTRIN_MEMSET were never emitted anywhere — "two spellings of one behavior" (CODEGEN.md Principle 7). Remove them and the dead switch arms across all six backends (x64/aa64/riscv native, wasm, c_target, interp); the memmove-only shapes that shared an arm keep their overlap-safe path. (The C frontend's private cg_adapter.h INTRIN_MEM* enum is a separate namespace, left untouched.) Green: lib/bin/rt, cg-api (interp memmove), libc 9/0, parse + parse-err 128/0, smoke x64/rv64.

Track 1c: test conditional control ops + reject continue on block scope

2026-06-08T19:52:35Z

commit 80ba21ae01667332bf937b70d209a0e73d690938 parent 106b8cb6c48dabf72df6804e4900cbdb39e836ca Author: Ryan Sepassi Date: Mon, 8 Jun 2026 12:52:35 -0700 Track 1c: test conditional control ops + reject continue on block scope break_true / continue_true / continue_false have zero callers in any frontend (toy uses only break / break_false / continue), so they were unexercised. Add test/api/cg_control_test.c: builds `int f(int n)` returning sum(0..n-1) via each conditional break/continue, runs all four through the in-process interpreter across a spread of n, emits each across aa64/x64/rv64 at -O0/-O1, and checks the new block-scope rejection. Audit fix: kit_cg_continue/_true/_false jumped to s->continue_lbl unconditionally, which is LABEL_NONE for a forward-only block scope (no loop header) — a jump to a nonexistent label. Now rejected with a clean diagnostic via api_require_loop_scope. Green: test-cg-api (cg_control_test 85/0), toy 1382/0, parse + parse-err.

Remove multi-value from the CG API (single result)

2026-06-08T19:39:29Z

commit 106b8cb6c48dabf72df6804e4900cbdb39e836ca parent 1541d1cfc7b9638a91946f78a7d4ccae1630b644 Author: Ryan Sepassi Date: Mon, 8 Jun 2026 12:39:29 -0700 Remove multi-value from the CG API (single result) Multi-result functions (KitCgFuncSig.results[]/nresults, nresults>1) were an advertise-but-ignore facade: the public API accepted up to 16 results, but no frontend produced more than one, there were zero tests, and every native backend panicked ("multiple returns unsupported") while the -O1 opt path and wasm silently dropped extras. Per CODEGEN.md Principles 2/3, collapse the whole surface to a single optional result (void == KIT_CG_TYPE_NONE) instead of implementing ~1000 LOC nothing uses. - Public: KitCgFuncSig.result (was results[]/nresults); kit_cg_type_func_result drops its index; removed kit_cg_type_func_nresults. - Type system stores a single KitCgFuncResult; cg_type_func_result_id returns the raw result type (NONE for void) alongside cg_type_func_ret_id. - CGFuncDesc.result_type / CGCallDesc.result (single CGLocal, CG_LOCAL_NONE = void); CgIrRetAux/value+present; OptCGFuncDesc.result_type. Removed API_CG_MAX_RESULTS, fn_result_types[], the call/ret loops. - ret hooks collapse to a single value: CgTarget.ret(CGLocal), NativeTarget.plan_ret(value), NativeDirectTarget.emit_ret(CGLocal). Backend "multiple returns unsupported" panics removed. NativeCallDesc kept as a bounded (<=1) array internally. - c_target: dropped the tuple typedef machinery; single-result signatures/calls. - Frontends (c/toy/wasm) build a single sig.result; wasm-spec multi-value (src/wasm, WasmType) is untouched. Green: lib/bin/rt, cg-api, opt, toy 1382/0, parse 3784 + parse-err 128/0, interp, abi-classify, ir-recorder, native-direct-target, tiny-inline, isa, link, elf, smoke x64/rv64; bootstrap reproduces byte-identical (stage2==stage3).

reloc: shrink reloc enum; neutral R_ABS8/16, R_PREL16, R_TPOFF64; rename R_RV_ADD/SUB/SET* to arch-neutral

2026-06-08T18:42:52Z

commit 1541d1cfc7b9638a91946f78a7d4ccae1630b644 parent 47cede9aca1511ee8a697c37bbf98aab26bb4ee3 Author: Ryan Sepassi Date: Mon, 8 Jun 2026 11:42:52 -0700 reloc: shrink reloc enum; neutral R_ABS8/16, R_PREL16, R_TPOFF64; rename R_RV_ADD*/SUB*/SET* to arch-neutral

plan: clean out completed plans

2026-06-08T18:39:35Z

commit 47cede9aca1511ee8a697c37bbf98aab26bb4ee3 parent 0ae44208d15bba91c4d3a9188848cb3871bf02bd Author: Ryan Sepassi Date: Mon, 8 Jun 2026 11:39:35 -0700 plan: clean out completed plans

rv32: close out riscv32-none-elf; remove completed plan doc

2026-06-08T18:32:25Z

commit 0ae44208d15bba91c4d3a9188848cb3871bf02bd parent be9be587a0b6e6599ed17c6764e1da2d93f042ec Author: Ryan Sepassi Date: Mon, 8 Jun 2026 11:32:25 -0700 rv32: close out riscv32-none-elf; remove completed plan doc The riscv32-none-elf backend is complete — all cross lanes pass at O0+O1 (test-asm-rv32 16/0, test-toy-rv32 421/0, test-parse-rv32 906/0, smoke hardfloat green under qemu-system-riscv32). The forward-looking plan in doc/plan/RV32.md is fully realized, so remove it and its index row (per convention, completed plans are removed, not checked off), and move the two durable contracts it tracked into user-facing docs: - doc/RUNTIME.md: the freestanding TLS thread-pointer contract ([TCB(16)|.tdata|.tbss], tp->TCB base, Local-Exec only; linker bias via ObjElfArchOps.tls_tp_bias) and the i64-atomics-as-libcall note (8-byte _Atomic -> spinlock __atomic_*_8, correct but not lock-free). - mk/test.mk: refresh the rv32 lane comments (corpus green; opt-in like the rv64 cross lanes, kept out of DEFAULT_TEST_TARGETS because they need the qemu-system-riscv32 toolchain). Docs and comments only; no code change.

plan: refresh doc index

2026-06-08T18:15:00Z

commit be9be587a0b6e6599ed17c6764e1da2d93f042ec parent 86d3164a20fead7c9b3e92932a861a271d915e42 Author: Ryan Sepassi Date: Mon, 8 Jun 2026 11:15:00 -0700 plan: refresh doc index

docs: plan llgen import

2026-06-06T12:17:20Z

commit 86d3164a20fead7c9b3e92932a861a271d915e42 parent 9fd3adda7106605e23b78d7c85f1237f1ea67801 Author: Ryan Sepassi Date: Sat, 6 Jun 2026 05:17:20 -0700 docs: plan llgen import

Fix nested panic recovery

2026-06-06T11:53:54Z

commit 9fd3adda7106605e23b78d7c85f1237f1ea67801 parent cb8e08d7d89c0ce544883b60d2d04a5c883df51e Author: Ryan Sepassi Date: Sat, 6 Jun 2026 04:53:54 -0700 Fix nested panic recovery

Fix -no-pie overriding explicit PIE

2026-06-06T11:41:35Z

commit cb8e08d7d89c0ce544883b60d2d04a5c883df51e parent 8b8960344ffcc9a12161a218fece7e8bf2beaf8b Author: Ryan Sepassi Date: Sat, 6 Jun 2026 04:41:35 -0700 Fix -no-pie overriding explicit PIE

Keep volatile locals memory resident

2026-06-06T11:33:21Z

commit 8b8960344ffcc9a12161a218fece7e8bf2beaf8b parent c8b9d53d52c4d396b6caed50da9e576413595199 Author: Ryan Sepassi Date: Sat, 6 Jun 2026 04:33:21 -0700 Keep volatile locals memory resident

Fix hard-pinned native inline asm staging

2026-06-06T11:15:28Z

commit c8b9d53d52c4d396b6caed50da9e576413595199 parent 96de8d734d0490c3e537ec6950c325e51015df59 Author: Ryan Sepassi Date: Sat, 6 Jun 2026 04:15:28 -0700 Fix hard-pinned native inline asm staging

Fix C inline asm output type stack

2026-06-06T10:52:37Z

commit 96de8d734d0490c3e537ec6950c325e51015df59 parent 44b5521e6ab0bdb348582ec7fe080e7ff6e36704 Author: Ryan Sepassi Date: Sat, 6 Jun 2026 03:52:37 -0700 Fix C inline asm output type stack

Support arch inline asm machine constraints

2026-06-06T10:34:19Z

commit 44b5521e6ab0bdb348582ec7fe080e7ff6e36704 parent 69d47850d76ff863e4493092346bd41f5c1f6574 Author: Ryan Sepassi Date: Sat, 6 Jun 2026 03:34:19 -0700 Support arch inline asm machine constraints

Implement raw syscall intrinsic lowering

2026-06-06T05:20:17Z

commit 69d47850d76ff863e4493092346bd41f5c1f6574 parent e55665b3c5ce178b17e7838c903429bf62eaf2cc Author: Ryan Sepassi Date: Fri, 5 Jun 2026 22:20:17 -0700 Implement raw syscall intrinsic lowering

Fix O1 label-address block liveness

2026-06-06T04:34:21Z

commit e55665b3c5ce178b17e7838c903429bf62eaf2cc parent 09bf3f299ea580495f0b4ce4a88dcacfe7a10f31 Author: Ryan Sepassi Date: Fri, 5 Jun 2026 21:34:21 -0700 Fix O1 label-address block liveness

Fix tail sret forwarding

2026-06-06T04:12:01Z

commit 09bf3f299ea580495f0b4ce4a88dcacfe7a10f31 parent d11c5c90c2371c2d260a161f8f7d61fc30824615 Author: Ryan Sepassi Date: Fri, 5 Jun 2026 21:12:01 -0700 Fix tail sret forwarding

Teach ld -lc to use hosted libc

2026-06-06T04:01:51Z

commit d11c5c90c2371c2d260a161f8f7d61fc30824615 parent 809a8e5e87c71ee3f3213e7e6f1f31eb972b70da Author: Ryan Sepassi Date: Fri, 5 Jun 2026 21:01:51 -0700 Teach ld -lc to use hosted libc

Fix aarch64 COFF weak-null ADRP relocation

2026-06-06T04:00:54Z

commit 809a8e5e87c71ee3f3213e7e6f1f31eb972b70da parent e87c69019725a0c08d8a1329614ac3e519758d9e Author: Ryan Sepassi Date: Fri, 5 Jun 2026 21:00:54 -0700 Fix aarch64 COFF weak-null ADRP relocation

Add freestanding runtime ctype

2026-06-06T02:39:59Z

commit e87c69019725a0c08d8a1329614ac3e519758d9e parent 85ca53b7d3dfe775399b30f30a1531c64d6c4726 Author: Ryan Sepassi Date: Fri, 5 Jun 2026 19:39:59 -0700 Add freestanding runtime ctype

Fix rv64 toy TLS cross startup

2026-06-06T02:26:02Z

commit 85ca53b7d3dfe775399b30f30a1531c64d6c4726 parent 0493f0042030ecca28f7d0d67039372cf165a4b8 Author: Ryan Sepassi Date: Fri, 5 Jun 2026 19:26:02 -0700 Fix rv64 toy TLS cross startup

Partition the relocation byte-patcher per-arch (RELOC WS-C)

2026-06-06T02:04:57Z

commit 0493f0042030ecca28f7d0d67039372cf165a4b8 parent 65545dc2cc2be1f6fb2de2149e2702a267aa7960 Author: Ryan Sepassi Date: Fri, 5 Jun 2026 19:04:57 -0700 Partition the relocation byte-patcher per-arch (RELOC WS-C) Move the per-arch instruction-immediate relocation byte encoders out of the format-neutral src/obj/reloc_apply.c into each backend's reloc.c, reached through a new LinkArchDesc.reloc_apply_insn hook, while keeping link_reloc_apply as the single public byte-patcher entry. Adding an arch's relocation is now arch-local: one RelocDesc row (WS-B) + one reloc_apply_insn arm + one wire-translator entry, no generic src/link or src/obj edit. - src/arch/{aa64,x64,riscv}/reloc.c: *_reloc_apply_insn — the imm19/imm26/ ADRP-page (aa64), U/I/S/B/J + RVC scatter & +0x800 bias (rv), and rel8 (x64) encoders, moved verbatim. Wired into each LinkArchDesc in link.c. - src/obj/reloc_apply.c: reloc_apply_neutral() keeps the arch-neutral data-word + ULEB128 arms (R_ABS/REL/PC/TPOFF writes, x64 GOT/dynamic slots, RISC-V data ADD/SUB/SET). Pure obj-core, no link/arch dependency. - src/link/link_reloc_apply.c (new): link_reloc_apply() dispatcher, neutral- then-arch. Lives in src/link (not obj-core) because resolving the per-arch slice needs link_arch_desc_for() — same boundary reason WS-B placed reloc_desc() in src/link. The dispatcher enumerates no kinds: `rg "case R_(AARCH64|X64|RV)_" src/link` is empty. R_PLT32's apply is the RISC-V AUIPC+JALR pair (x64 emits R_X64_PLT32 on read, never canonical R_PLT32), so it lives in the rv hook beside R_RV_CALL — not neutral, despite its neutral name. x64 owns only R_X64_PC8; the wider x64 GOT/PLT/TPOFF data slots stay neutral. Migration guard: test/link/reloc_apply_test.c (test-link-reloc-apply, 50 frozen pre-WS-C byte checks across aa64/x64/rv). Verified byte-identical: full link/elf/macho/ar/asm/isa/opt/coff/smoke matrix + bootstrap-debug (stage2 == stage3). reloc_uleb128_unit's c=NULL path still works (neutral never touches the compiler). Also refresh test/smoke/rv64_tls_link.sh's reloc grep (RV_TPREL_HI20 -> R_RISCV_TPREL_HI20): WS-E.3 switched ELF reloc records to binutils spellings but missed this script, and R_RISCV_TPREL_HI20 doesn't contain the old substring.

Implement per-arch RelocDesc table + residual gates (RELOC WS-B, WS-E.2/E.3)

2026-06-05T23:31:53Z

commit 65545dc2cc2be1f6fb2de2149e2702a267aa7960 parent 7a985cce08172385f81d3fa7e74540ffc3ed037a Author: Ryan Sepassi Date: Fri, 5 Jun 2026 16:31:53 -0700 Implement per-arch RelocDesc table + residual gates (RELOC WS-B, WS-E.2/E.3) WS-B — make the canonical-RelocKind half of the relocation subsystem as modular as the wire half. Replace the generic reloc_width / reloc_uses_got / reloc_is_tls_got switches in link_reloc_layout.c, link_jit's jit_reloc_width_local, and the five per-arch LinkArchDesc.is_* hooks with a single per-arch RelocDesc {u8 width; u8 flags} resolved arch-aware by reloc_desc(c, k): - neutral data-word kinds in src/obj/reloc.{h,c} (pure obj-core); - per-arch slices in src/arch/{aa64,x64,riscv}/reloc.c via a new LinkArchDesc.reloc_desc hook; - dispatcher + reloc_kind_* predicates in src/link/link_reloc_desc.{h,c} (kept in src/link, not obj-core, since resolving the slice needs link_arch_desc_for()). All consumers (the GOT / JIT-stub / width passes and the nine Mach-O is_* call sites) now read one descriptor flag. Adding an arch's relocation is one row in that arch's slice. Migration guard: test/link/reloc_desc_test.c pins behavioural equivalence to frozen oracle_* snapshots of the deleted code across every kind x every backend arch (3016 checks). WS-E.2 — the static-IFUNC __rela_iplt IRELATIVE wire type no longer names a literal KIT_OBJ_ELF: link_elf_irelative_type is gone, replaced by obj_format_static_ifunc_irelative_type(c) (sibling of the WS-E.1 predicate), which resolves through the target object format. WS-E.3 — drop the arch axis from kit_obj_reloc_kind_name's gate (fmt != KIT_OBJ_ELF only), so aarch64/riscv ELF relocs print via their ObjElfArchOps.reloc_name tables (binutils-faithful: R_AARCH64_CALL26, R_RISCV_CALL). One objdump golden refreshed; the -d disasm annotation keeps the arch-neutral spelling (separate disassembler path). rg "case R_(AARCH64|X64|RV)_" src/link is now empty. Verified: reloc_desc parity 3016/0, test-link/elf/macho/ar/isa/aa64-inline/driver-objdump pass, make bootstrap-debug byte-identical.

Defer O1 local static data symbols

2026-06-05T23:30:15Z

commit 7a985cce08172385f81d3fa7e74540ffc3ed037a parent 81d88fcf49dea5f3f1d389ac360bf7c588c9d3ff Author: Ryan Sepassi Date: Fri, 5 Jun 2026 16:30:15 -0700 Defer O1 local static data symbols

Add stack backtrace sources to driver rt manifest

2026-06-05T22:10:41Z

commit 81d88fcf49dea5f3f1d389ac360bf7c588c9d3ff parent bad3d6b438f65e991b1aaa1845c0aeb7651158a0 Author: Ryan Sepassi Date: Fri, 5 Jun 2026 15:10:41 -0700 Add stack backtrace sources to driver rt manifest The backtrace work (b3f14e4e) made rt/lib/assert/assert.c call __kit_print_backtrace() and added stack/{backtrace,print_backtrace}.c, updating the Makefile rt build (RT_BASE_SRCS) but not the driver's on-demand rt manifest in driver/lib/runtime.c. So `kit cc`'s auto-built rt archive carried assert.c without the two sources it now needs, producing "undefined reference to '__kit_print_backtrace'" whenever a link pulled assert.c (e.g. test-driver's cc-auto-builds-and-links- libkit-rt-x64). Add stack/backtrace.c + stack/print_backtrace.c to every full-libc driver list (x64, aarch64-elf/darwin, rv64-elf, rv32-elf), realigning with RT_BASE_SRCS. The compiler-rt-only Windows lists omit assert.c, so they stay unchanged.

doc/plan: rescope RELOC.md after the modularity waves landed

2026-06-05T22:03:38Z

commit bad3d6b438f65e991b1aaa1845c0aeb7651158a0 parent b499c2351e3bc118a107a8693d35413c6e8e0db3 Author: Ryan Sepassi Date: Fri, 5 Jun 2026 15:03:38 -0700 doc/plan: rescope RELOC.md after the modularity waves landed The reloc-path identity switches are already gone (LinkArchDesc.tpoff64_reloc for #25, obj_format_static_ifunc_via_rela_iplt for #18) and the reloc-name table moved onto ObjElfArchOps.reloc_name (#24, still x86_64-ELF-gated for objdump golden compat). Mark those as landed and rescope the open work to the structural denormalization that remains: a per-arch RelocDesc {width, flags} table replacing the generic reloc_width / reloc_uses_got / reloc_is_tls_got switches (and the duplicating LinkArchDesc.is_* hooks), and partitioning link_reloc_apply's encoders into the arch backends behind the single entry.

Implement tool-side auto-backtrace for run & dbg (BACKTRACE L3c/WS5)

2026-06-05T21:45:54Z

commit b499c2351e3bc118a107a8693d35413c6e8e0db3 parent 21ac717e705ac204fb507e83fa04f467397a494a Author: Ryan Sepassi Date: Fri, 5 Jun 2026 14:45:54 -0700 Implement tool-side auto-backtrace for run & dbg (BACKTRACE L3c/WS5) `kit run` and `kit dbg` now print a symbolized backtrace automatically when a program faults or traps, reusing the hosted DWARF reader and never crossing into rt/. The CFI stepper kit_dwarf_unwind_step takes no memory provider, so it returns pc=0 once a return address is spilled to the stack (the common case) — the old `dbg bt` was effectively single-frame. Replace the unwind with a frame-pointer chain walk: kit keeps a frame pointer on every backend with a uniform record (fp[0]=caller fp, fp[1]=saved ra), the same walk __kit_backtrace does, lifted tool-side with a memory-read callback. - driver/lib/backtrace.{c,h}: shared FP-step kernel (with __kit_backtrace's guards), arch FP-reg/ptr-size helpers, and a PC-list symbolized printer. The walk stops at the kit-image boundary (kit_jit_runtime_to_image==0) so output ends at main and is host-independent (no libc/dyld trampoline noise). - dbg (driver/cmd/dbg.c): dbg_cmd_bt advances via the FP kernel over kit_jit_session_read_mem (walks the whole stack), and dbg_render_stop auto-invokes it on KIT_STOP_SIGNAL (faults + __builtin_trap/assert, not breakpoints/steps). - run (driver/cmd/run.c + driver/env/posix_dbg.c): a lightweight crash guard installs SIGSEGV/SIGBUS/SIGILL/SIGFPE/SIGABRT/SIGTRAP handlers around the in-process entry_fn call, reusing dbg_ucontext_to_frame. Because run shares its stack with the program, the chain is captured inside the handler (before the post-siglongjmp stack is reused) and symbolized afterward in normal context; the process exits 128+signo. Windows has a no-op stub. - backtrace.c is compiled unconditionally: posix_dbg.c's crash guard references it on every POSIX build regardless of tool selection. Tests: test/dbg/cases/toy-trap-backtrace (multi-frame trap -> auto-bt + bt), updated toy-trap-stop golden, and run-backtrace-* checks in test/driver/run.sh. kit emu auto-backtrace is out of scope (it doesn't retain the guest DWARF); deferred alongside L3b.

Implement kit symbolize (BACKTRACE L3a/WS5)

2026-06-05T16:49:29Z

commit 21ac717e705ac204fb507e83fa04f467397a494a parent b3f14e4ecb76fed2930be28041e7e68a692b073f Author: Ryan Sepassi Date: Fri, 5 Jun 2026 09:49:29 -0700 Implement kit symbolize (BACKTRACE L3a/WS5) Add `kit symbolize`, the hosted batching symbolizer that matches kit's actual backtrace artifact: it reads the raw "#N 0x" stream that __kit_print_backtrace emits, finds the address on each line, resolves it through the existing DWARF reader, and rewrites the line in place as #0 0x401136 bt_leaf at addr2line_prog.c:51:3 preserving the "#N" framing that addr2line structurally can't keep. Lines with no 0x token pass through verbatim, so a mixed panic log (banner + frames + register dump) survives unharmed. addr2line stays a faithful GNU/LLVM clone (bare addresses in, file:line out); its output contract is unchanged. The DWARF-open + func/line core is factored into driver/lib/dwarfsym.c, shared by both tools, which now differ only in formatting. symbolize is a non-default (GROUP_OTHER) tool like mc/disas; a single -e today, with multi--e/module-map (for libc.so frames that need their own load slide) as the natural extension. Testing: test/rt/addr2line.sh (make target test-rt-backtrace) grows a second lane per arch/opt — it pipes the captured stream into `kit symbolize -e ` and asserts each frame keeps its "#N 0x.." form with " at file:line" appended. aa64/x64/rv64 at O0/O1; the x64/O1 lane stays the documented pre-existing -g -O1 compiler red (cc aborts before producing a binary, so it never reaches symbolization).

Implement __kit_print_backtrace & assert hook (BACKTRACE L3a/WS4)

2026-06-05T16:26:10Z

commit b3f14e4ecb76fed2930be28041e7e68a692b073f parent 5a7a304b5a1d6dd328cc662759c895feebb5340c Author: Ryan Sepassi Date: Fri, 5 Jun 2026 09:26:10 -0700 Implement __kit_print_backtrace & assert hook (BACKTRACE L3a/WS4) Ship L3a of the backtrace roadmap: raw backtrace print + out-of-process symbolization, plus the deferred assert-path hook. - rt/lib/stack/print_backtrace.c: __kit_print_backtrace walks via __kit_backtrace(skip=1) and writes "#N 0x" lines (hand-rolled fmt, no printf) to the weak no-op __kit_backtrace_write sink. Both decls in rt/include/kit/backtrace.h; added to RT_BASE_SRCS. - rt/lib/assert/assert.c: __kit_assert_fail emits a banner + __kit_print_backtrace() before __builtin_trap(), via the same sink. - Resolves the L3a open question: weak no-op sink default (host/_start overrides to write(2)), so freestanding images still link. Tests (aa64/x64/rv64, O0 + O1): - test/rt/cases/print_backtrace.c: in-process parse of the emitted lines. - test/rt/addr2line.sh + addr2line_prog.c: kit addr2line round-trip (make target test-rt-backtrace). - test/rt/run.sh now sweeps O0+O1 (KIT_RT_OPT_LEVELS); smoke.c includes . The backtrace path passes at O0 and O1 on all three arches. Sweeping O1 surfaced two unrelated pre-existing bugs, left RED (not skipped) and logged in doc/plan/TODO.md: - x86-64 -g -O1 + 4-operand register-pinned syscall asm aborts the compiler (too many memory asm operands); fails test-rt-backtrace x64/O1. - setjmp/longjmp miscompiled at -O1 on every arch (second-return value not observed); fails test-rt-runtime setjmp_runtime/O1.

Implement stack-trace builtins & __kit_backtrace (BACKTRACE L1+L2)

2026-06-05T15:38:50Z

commit 5a7a304b5a1d6dd328cc662759c895feebb5340c parent 14e3792d88237772be6076f14a77103d5ea746f8 Author: Ryan Sepassi Date: Fri, 5 Jun 2026 08:38:50 -0700 Implement stack-trace builtins & __kit_backtrace (BACKTRACE L1+L2) L1 — __builtin_frame_address / __builtin_return_address: two CG intrinsics (KIT_CG_INTRIN_FRAME_ADDRESS/_RETURN_ADDRESS) carrying the constant level as an IMM operand, lowered as an unrolled frame-pointer walk on aarch64/x86-64/ riscv (O0 and O1, shared backend handler). C target forwards __builtin_* to the host; wasm reports unsupported; the C frontend validates the level via eval_const_int. IR_INTRINSIC is already conservatively side-effecting in opt, so no new effect modeling was needed; the one O1 hazard — riscv's frameless- leaf tier omitting the frame record — is handled by a new NativeKnownFrameDesc.reads_frame flag (aarch64/x64 always keep the record). L2 — __kit_backtrace(void**, int, int): freestanding FP-walk capture in rt/lib/stack/backtrace.c + rt/include/kit/backtrace.h, added to RT_BASE_SRCS. kit's frame layout is uniform across targets (fp[0]=caller fp, fp[1]=saved ra in void* units), so no per-arch offset table — the psABI rv64 layout in the plan sketch was wrong; verified against the actual prologue. Tests: test/rt/cases/backtrace_capture.c (aa64/x64/rv64 under exec), test/parse builtin_29..31 + cases_err/..._nonconst (D/R/E/J/C lanes, O0/O1), test/toy 154_frame_return_address. Verified end-to-end through kit addr2line on static-freestanding and dynamic-musl binaries across all three arches. doc/plan/BACKTRACE.md updated (L1+L2 done, L3 remaining). doc/plan/TODO.md records four pre-existing gaps found while building/verifying: &&label-as-value broken at O1, __kit_syscallN declared-but-unimplemented, -no-pie not producing ET_EXEC, and kit cc/ld dynamic-libc link ergonomics.

modularity: gate COFF weak-extern underscore alias on a vtable capability + fix stale objdump golden

2026-06-05T21:47:44Z

commit 14e3792d88237772be6076f14a77103d5ea746f8 parent 769d6ae1ab02a5531923e468ed1a64d69cc3456d Author: Ryan Sepassi Date: Fri, 5 Jun 2026 14:47:44 -0700 modularity: gate COFF weak-extern underscore alias on a vtable capability + fix stale objdump golden - src/link/link_resolve.c: the mingw single-underscore weak-external alias recovery now gates on a new ObjFormatImpl.weak_extern_underscore_alias (COFF=1) via obj_format_weak_extern_underscore_alias(), replacing the last identity switch in generic link code. The PE-specific de/re-underscore mechanism stays, but only formats that declare the capability run it. Provably equivalent (1 for COFF, 0 elsewhere). - test/objdump/aarch64/cases/04-disasm-stripped-segment.expected: refresh the stale golden to the disassembler's canonical alias output (sxtw x8, w9) for bytes 28 7d 40 93; sbfm xd,xn,#0,#31 IS SXTW, matching binutils/LLVM. (Was failing identically on baseline 30367860 — a pre-existing stale golden, not a regression.) test-driver-objdump 14/0 (was 13/1); test-link 122/0, elf 41/0, ar 20/20, smoke x64+rv64 green.

modularity wave 4: compiler-free language resolution + obj-owned fmt names + jit TLS hook

2026-06-05T18:51:56Z

commit 769d6ae1ab02a5531923e468ed1a64d69cc3456d parent 757fa31e7ccda6059b00c27ab30ee4e9b56881b3 Author: Ryan Sepassi Date: Fri, 5 Jun 2026 11:51:56 -0700 modularity wave 4: compiler-free language resolution + obj-owned fmt names + jit TLS hook Closes the last residuals from the audit: - #6/#7: kit_language_for_name/_for_path/_name + new kit_language_default_extension now accept a NULL compiler and resolve against the compile-time default frontend set via a new lang_registry_vtable() (kept in lang_registry.c, the one place that checks KIT_LANG_*_ENABLED). This unblocks arg-parse-time use. A single driver_path_is_source() authority (registry lookup minus C headers) now backs cc/build/run-dbg source classification, replacing four hand-rolled extension unions; fixes the latent .S-misclassification bug and keeps the documented '.h is a header, not a source' rule in one place. -x names + dbg REPL filenames route through the resolvers (driver/cmd/{cc,build,dbg,emu}.c). - #35: objcopy/objdump get canonical format names from the obj layer (kit_obj_fmt_from_name/_name), keeping only the binutils BFD-alias prefix matching + bitwidth suffix as documented tool presentation. - Beyond the 36 findings, in the same spirit: src/link/link_jit.c Mach-O TLV JIT bootstrap now gates on obj_format_tls_via_descriptor instead of target.obj==KIT_OBJ_MACHO. Build + tests green: link 122, driver-cc 98, build 56, tools 45, objcopy 5, ar 6, toy 1378, smoke x64+rv64. (Pre-existing, unrelated: objdump aarch64/04-disasm-stripped-segment stale golden — fails identically on baseline 30367860; untouched disassembler, fixed fixture.)

modularity wave 3: route frontend/driver/backend consumers through hooks

2026-06-05T18:13:53Z

commit 757fa31e7ccda6059b00c27ab30ee4e9b56881b3 parent c9c1495aa91c311d916704f17cd3c9f0a79079af Author: Ryan Sepassi Date: Fri, 5 Jun 2026 11:13:53 -0700 modularity wave 3: route frontend/driver/backend consumers through hooks C frontend + preprocessor (P2): c_abi_va_list_type -> kit_cg_target_va_list_kind (#1, also fixes a latent win64 va_list-type bug); __builtin___clear_cache -> backend ICACHE_COHERENT feature (#5); kit_target_uses_lp64/long_double_is_binary128/ty_wchar -> resolved spec fields (#33,#3,#32); pp.c Windows/_M_* predefines consolidated into one os-keyed function reading spec.wchar_size/long_size + kit_cg_target_c_label_prefix (#20). Driver (P3): HostedProfile (os,obj) registry replaces the if/else+switch dispatch (#19); new driver_default_exe_name/obj_ext/needs_sysroot_libdir/default_hosted_profile helpers replace the os==WINDOWS forks in cc.c/build.c (#34); emu arch via driver_arch_from_name single authority (#22); dbg REPL caching via frontend caps (#8); build.c lang->name reverse maps via kit_language_name (#26). Backend: x64/aa64 TEB-TLS selection via obj_format_tls_model (#15); aa64 variadic FP class via ABIFuncInfo.vararg_fp_via_int (#16); mc.c .eh_frame via spec.emits_eh_frame (#17); c_target underscore strip via obj_format_demangle_c (#10); c_target alias/weak/TLS spellings via new ObjFormatImpl.alias_via_thunk/weak_undef_attr + obj_format_tls_via_descriptor (#27). Residual (tracked): parse-time -x/source-ext classifiers (#6/#7) need compiler-free resolvers; objcopy/objdump (#35) need BFD-alias fmt names — both addressed next. Full suite green: parse/pp/toy/cg-api/opt/isa/aa64-inline/link/elf/ar/driver-ar/debug/dwarf/ lib-deps/libc(musl+glibc)/smoke-x64/smoke-rv64.

modularity wave 2: route obj/link/cg/api/asm/dwarf through capability hooks

2026-06-05T17:51:12Z

commit c9c1495aa91c311d916704f17cd3c9f0a79079af parent e5603fea9e7ddce3fdff1427bbf18ddab1ea87d1 Author: Ryan Sepassi Date: Fri, 5 Jun 2026 10:51:12 -0700 modularity wave 2: route obj/link/cg/api/asm/dwarf through capability hooks Removes identity switches from generic consumers by reading the wave-1 hooks: - src/api/core.c: KitTargetSpec gains resolved props (long_size #33, wchar_size #32, long_double_format #3, emits_eh_frame #17) computed once in kit_target_new; RISC-V float-abi gate replaced by arch_resolve_float_abi (#2). - src/cg/type.c: supports_symbol_feature -> obj_format_supports_symbol_feature (#14); add public kit_cg_target_va_list_kind + kit_cg_target_c_label_prefix. - src/api/object_detect.c/object_file.c: machine->arch via fmt reverse maps + kit_arch_ptr_size (#4), float-abi via ObjElfArchOps.float_abi_from_e_flags (#23), reloc names via ELF arch ops (#24); public kit_obj_fmt_from_name/kit_obj_fmt_name (#35 impl). - src/link: tpoff64 via LinkArchDesc (#25), COFF synth via obj_format_synth_inputs (#13), boundary syms via obj_format_boundary_sym_kind (#12), static-ifunc/debug/got/weak-pull via obj_format_* predicates (#18,#28,#29); 9 target.obj/os switches removed. - src/asm/asm.c: macho seckind via obj_macho_seckind_for_secname (#21). - src/debug/dwarf_open.c: macho section spelling via obj_macho_native_secname (#36). Gatekeeper fix: supports_symbol_feature TLS arm now reads a per-format ObjFormatImpl .tls_symbol_features field (ELF/Mach-O=1, COFF/Wasm=0) instead of deriving from tls_model, which wrongly returned 1 for Wasm. Build + targeted tests (link/elf/ar/debug/dwarf/asm/ driver-ar/toy/parse/pp/cg-api/smoke-x64/smoke-rv64) green.

modularity wave 1: add pluggable-axis capability hooks (additive infra)

2026-06-05T17:29:07Z

commit e5603fea9e7ddce3fdff1427bbf18ddab1ea87d1 parent 9d905b3c414afa1ee34e1825d62000f9acfda660 Author: Ryan Sepassi Date: Fri, 5 Jun 2026 10:29:07 -0700 modularity wave 1: add pluggable-axis capability hooks (additive infra) Adds the abstraction hooks that wave 2/3 consumers will route through, all additive (0/NULL defaults preserve existing behavior; old fns kept as thin wrappers): ObjFormat (src/obj): obj_format_tls_model {ELF_LE,MACHO_DESCRIPTOR,WINDOWS_TEB} (tls_via_descriptor now wraps it), c_label_prefix, default_entry_name, carries_file_only_debug, builds_own_static_got, supports_symbol_feature, weak_undef_pulls_archive_member, static_ifunc_via_rela_iplt, boundary_sym_kind, synth_inputs hook; ObjElfArchOps reloc_name/float_abi_from_e_flags/tls_tp_bias; obj_macho_seckind_for_secname + obj_macho_native_secname; obj/public name<->fmt. ABI (src/abi): ABIFuncInfo.vararg_fp_via_int (set by aapcs64_windows). ArchImpl (src/arch): resolve_float_abi hook + riscv body; KIT_CG_BACKEND_ICACHE_COHERENT (x64,wasm); kit_arch_ptr_size. Frontend (src/api,lang): KitFrontendVTable names/nnames + kit_language_for_name/kit_language_name; KitFrontendCaps.cache_repl_toplevel_source (toy). Internal obj helpers (c_mangle, default_entry, extern_via_got, tls bias) migrated to read the new vtable fields. Build + targeted tests (cg-api,isa,link,elf,debug,dwarf,parse,smoke-x64,smoke-rv64) green.

modularity #9: move wasm_imports to src/obj, drop cg cross-boundary leak

2026-06-05T17:06:36Z

commit 9d905b3c414afa1ee34e1825d62000f9acfda660 parent 30367860c674c288dc894fe10c8595aa40519149 Author: Ryan Sepassi Date: Fri, 5 Jun 2026 10:06:36 -0700 modularity #9: move wasm_imports to src/obj, drop cg cross-boundary leak src/cg/session.c (the frontend-agnostic KitCg entry) no longer #includes a backend-private header (arch/wasm/wasm_imports.h) behind #if KIT_ARCH_WASM_ENABLED. wasm_imports.{c,h} is pure obj-layer code (only touches the generic obj_ext_set(OBJ_EXT_WASM_IMPORTS) side-table), so it moves to src/obj/ where it is always compiled; session.c calls wasm_imports_set unconditionally. Behavior-identical. Removes the only cross-boundary include and behavior-forking #if KIT_ARCH_* in src/cg.

Generalize VM execution into the shared exec seam + a hosted test suite

2026-06-05T16:35:39Z

commit 30367860c674c288dc894fe10c8595aa40519149 parent 040f0c2838b6b40acd1c3f38b981cc7efedadd2d Author: Ryan Sepassi Date: Fri, 5 Jun 2026 09:35:39 -0700 Generalize VM execution into the shared exec seam + a hosted test suite Lift hosted-VM execution out of test/toy/vm.sh into the shared runner test/lib/exec_target.sh, and build a unified cross-OS hosted test suite on top of it (test/hosted), green across the full support set. Shared exec seam (test/lib/exec_target.sh + new test/lib/exec_vm.sh): - Add -freebsd / -windows tags whose runner is a VM. A VM is expensive + stateful, so add lifecycle hooks the stateless podman/qemu runners don't need: exec_target_setup boots the VM lazily and once (one Windows VM serves both arches), keeps it warm across flushes, and exec_target_teardown_all (trap ... EXIT) stops only VMs we booted. - Add -linux-glibc tags routing to per-arch Debian images, so glibc and musl (-linux -> alpine) both run through one flush; _exec_target_arch now takes the first field for 3-part tags and image-present memoizes per image. - scripts/{freebsd,windows}_vm.sh run-batch gains a results contract: ship a staging dir in, run a generated run-remote.{sh,ps1}, bring each binary's rc/out/err back (FreeBSD bidirectional tar; Windows scp-back + Defender exclusion + Start-Process capture). Windows rc masked to 8 bits. - Repoint test/toy/vm.sh at the seam (compile + queue only; ~140 lines lighter). Hosted suite (scripts/hosted.sh + test/hosted): - scripts/hosted.sh: one front-end to provision sysroots and cross-compile+link across the support set ([-]-): prepare/path/triple/tag/cc/doctor, wrapping freebsd_sysroot.sh / llvm_mingw_sysroot.sh / the libc extract.sh; macOS uses the Xcode SDK via -isysroot. - test/hosted: seed case hello.c built via hosted.sh and run via exec_target, checking exit code + stdout. make test-hosted (Linux+macOS) / test-hosted-vm (adds FreeBSD+Windows). 15-config matrix verified green (30 pass/0 fail): Linux {aa64,x64,rv64} x {musl-static,musl-dynamic,glibc} + macos-arm64 + windows {x64,aarch64} + freebsd {amd64,aarch64,riscv64}. C frontend (lang/c/parse): recognize __restrict / __restrict__ as keyword aliases of `restrict`. glibc headers use them as bare GCC keywords and #undef the fallback macro under __GNUC__, so the preprocessor-macro approach was not enough; this unblocks compiling against glibc headers on all arches. No regression: existing exec_target consumers (test-parse 3840/0) still pass; test/toy/run.sh and the corpus engine are untouched.

test-toy: run the Toy corpus hosted in the FreeBSD/Windows VMs

2026-06-05T14:54:39Z

commit 040f0c2838b6b40acd1c3f38b981cc7efedadd2d parent dc1fabbf4cea6f8570680c2365efc94d2cbf4a09 Author: Ryan Sepassi Date: Fri, 5 Jun 2026 07:54:39 -0700 test-toy: run the Toy corpus hosted in the FreeBSD/Windows VMs Add test/toy/vm.sh, the hosted-VM counterpart to run.sh's freestanding-Linux X lane: compile each Toy case against a real OS sysroot (FreeBSD base.txz extract or the llvm-mingw UCRT sysroot) and execute it on the genuine OS in a VM, asserting the same .expected exit-code oracle. Per (os,arch) it stages every applicable case x opt x link-mode into one dir and drains it in a single VM session via a new `run-batch ` subcommand added to both scripts/freebsd_vm.sh (tar-pipe over ssh) and scripts/windows_vm.sh (scp + run-remote.ps1). Opt-in targets: test-toy-{freebsd,windows}-vm, test-toy-vm. Verified full corpus O0+O1: fbsd-aarch64 652/0/1 (static+dynamic), fbsd-riscv64 648/0/5, win-aarch64 324/2/1, win-x64 322/2/2, fbsd-amd64 648/0/2 (after the cpu fix below). Windows specifics: Defender blocks some kit exes as PUA and PowerShell's $LASTEXITCODE keeps a neighbor's code on a blocked launch, so run-batch adds a temp-dir Defender exclusion and run-remote.ps1 captures via Start-Process -PassThru .ExitCode. Exit codes are masked to 8 bits since the oracle is a POSIX exit status (Windows does not truncate). freebsd_vm.sh amd64: default qemu64 lacks POPCNT (kit emits popcnt for @popcount) -> SIGILL; boot with -cpu qemu64,+popcnt,+sse4.1,+sse4.2 (-cpu max breaks the FreeBSD/EDK2 TCG boot). The lanes surfaced two genuine kit bugs, filed in doc/plan/TODO.md: aarch64-windows link ADRP-out-of-range (118_decl_extra_attrs) and x86_64-windows tail-call+sret O1 access violation (36_musttail_sret/37_tail_sret). run.sh and the corpus engine are untouched; default test-toy is unaffected.

doc/plan: add stack-trace builtins & runtime backtrace roadmap

2026-06-05T13:40:34Z

commit dc1fabbf4cea6f8570680c2365efc94d2cbf4a09 parent 0a2faa92c8f3b6a302e7f621f78a92ba4b4180c7 Author: Ryan Sepassi Date: Fri, 5 Jun 2026 06:40:34 -0700 doc/plan: add stack-trace builtins & runtime backtrace roadmap Proposes a three-layer design: GCC-compatible __builtin_return_address / __builtin_frame_address primitives (FP-chain lowering, enabled by kit always keeping a frame pointer), a freestanding __kit_backtrace capture helper, and a symbolizing __kit_print_backtrace print path that keeps the DWARF reader on the hosted side of the freestanding boundary.

Implement whole-program optimization (LTO Phase 0 + Phase 1)

2026-06-04T21:01:26Z

commit 0a2faa92c8f3b6a302e7f621f78a92ba4b4180c7 parent 61a4a4c6f662dced2a9394cb09ee41c1bca529c2 Author: Ryan Sepassi Date: Thu, 4 Jun 2026 14:01:26 -0700 Implement whole-program optimization (LTO Phase 0 + Phase 1) Make a library or executable look like a single translation unit to the optimizer so inlining, dead-code elimination, and internalization cross TU boundaries, for invocations that provide all sources up front (kit cc *.c -flto -o prog and the build-exe/lib/obj verbs). Whole-program optimization runs whenever the optimizer runs (-O1+). See doc/plan/LTO.md. Phase 0 — whole-translation-unit optimization: - Generalize the ARM64-only finalize sweep to opt_whole_module_finalize for every arch; x86-64/riscv64 defer per-function emit to finalize under the whole-program path. -O0 and the JIT/interp/run paths stay on eager emit. - Wire opt_inline over the reachable FuncSet (the previously unreached whole-program inliner); weak/interposable callees are kept out-of-line. - Decide cg/link capabilities via the arch vtable rather than arch identity, so src/opt carries no arch == KIT_ARCH_* checks. Phase 1 — shared-context, all-sources-up-front LTO: - Stage N semantic frontends (C, Toy, Wasm) into one borrowed KitCg over a caller-owned ObjBuilder via an explicit begin/begin_unit/end_unit/finish/ detach lifecycle; asm participates as an opaque object. Drivers (build_compile_all, cc_run_link_exe) collect inputs and do not own definition selection or finalization. - Extract symresolve (src/obj/symresolve.{h,c}); refactor link_resolve_symbols onto it and reuse it for the recording-time merge so cross-TU ODR/weak/common/COMDAT resolution matches the linker exactly. - Give ObjBuilder an O(1) name->id index; skip-intern LOCAL symbols so per-TU statics stay distinct in the shared builder. - Compute the preserved/export set from the assembled link (entry, opaque undef refs, intrinsic/IFUNC roots, retain/init-array/address-significant) and feed it to kit_cg_finish; executable links internalize the rest before the reachability sweep. Relocatable/archive/shared stay conservative and cc -shared -flto is rejected until shared output is exercised. Fixes found completing the work: - opt_set_finish_policy was called unconditionally from kit_cg_finish and casts the recorder's user to OptImpl*, but the C-source backend's recorder user is a CTarget; the stray write corrupted it and crashed every --emit=c run. Guard it behind opt_level > 0, like opt_set_dump_writer. - The recording-time symresolve_merge fired on same-TU re-emission, so legal C tentative definitions (int g; int g;) were rejected as duplicates in every path, not just LTO. Track the defining source unit per ObjSymId and merge only across units; same-TU re-emission keeps last-writer-wins, matching the non-LTO linker. kit stays -fno-common. Tests: test/opt/whole_program_inline.sh and test/opt/lto_phase1.sh cover cross-TU fusion, internalization, weak/strong, ODR, tentative resolution fidelity (-flto == linker), multi-frontend staging, and opaque asm. asm_04_register_callee_saved is skipped on Mach-O hosts (verbatim file-scope asm defines bare names the underscored C refs cannot reach, like asm_02).

Fix Windows x64/arm64 variadic ABI at O0 and O1

2026-06-05T04:51:14Z

commit 61a4a4c6f662dced2a9394cb09ee41c1bca529c2 parent 4c4f1db31be66b1fab039b0a1fa8c4f8ab2bdd6f Author: Ryan Sepassi Date: Thu, 4 Jun 2026 21:51:14 -0700 Fix Windows x64/arm64 variadic ABI at O0 and O1 aarch64-windows variadics were broken in both the direct (O0) and optimizer (O1) paths, and x64 Win64 va_start miscompiled at O1. The no-VM smoke (build + objdump) hid all of it; these are runtime bugs. - aarch64-windows: route variadic FP args through GP registers (the classifier already remapped named params; aa_param_abi now does the same for trailing `...` args) and home x0..x7 into a GP "home area" at the top of the frame, contiguous with the incoming stack args, so the plain-pointer va_list walks register- then stack-passed varargs as one block. Driven by ABIVaListInfo.gp_reg_count; forces the fat top-record frame for Windows variadic functions only (non-variadic and non-Windows codegen is untouched). - x64 Win64: declare RAX clobbered by va_start via the machine-op clobber mechanism (new NATIVE_MOP_VA_START), so the allocator stops keeping a live value (e.g. a return-coalesced loop accumulator) in RAX across the op. Verified on the Win11 ARM64 VM (aarch64 native + x64 emulated) at O0 and O1: all three COFF Windows smokes pass, plus new coverage for loop/overflow varargs, named-FP varargs, va_copy, and RSI/RDI + XMM6-15 preservation. No regressions across cg/opt/isa/parse/asm/pp/ debug/dwarf/smoke/libc, Apple-ARM64, or SysV varargs. Also fold in the in-flight frontend cleanups owned by this change (extern-inline suppressed-parse dummy labels; drop the bogus local-const memory boundary on file-scope asm) and add their missing test coverage.

Fix Windows PE x64 startup and callback ABI

2026-06-05T02:00:38Z

commit 4c4f1db31be66b1fab039b0a1fa8c4f8ab2bdd6f parent eff11e543f602e9bccfb86bdd2e446271957de85 Author: Ryan Sepassi Date: Thu, 4 Jun 2026 19:00:38 -0700 Fix Windows PE x64 startup and callback ABI

Add release runtime variants

2026-06-05T00:12:47Z

commit eff11e543f602e9bccfb86bdd2e446271957de85 parent b8dcd8a7bdd3f2ce01756fd8ce410d64786424de Author: Ryan Sepassi Date: Thu, 4 Jun 2026 17:12:47 -0700 Add release runtime variants Fill the release runtime variant tables for FreeBSD and freestanding targets, brand FreeBSD ELF objects so ld can select the FreeBSD runtime, and mark the release checklist items complete.

coff: emit PE exception directory metadata

2026-06-04T23:41:30Z

commit b8dcd8a7bdd3f2ce01756fd8ce410d64786424de parent 11895ae0ea7cc7f8fc16664bae5d2fff4aa87c1d Author: Ryan Sepassi Date: Thu, 4 Jun 2026 16:41:30 -0700 coff: emit PE exception directory metadata

windows: add ARM64 VM runner for COFF smoke tests

2026-06-04T23:41:00Z

commit 11895ae0ea7cc7f8fc16664bae5d2fff4aa87c1d parent 60e7c4ca59b86810a9103da2e217a85aa0bd3134 Author: Ryan Sepassi Date: Thu, 4 Jun 2026 16:41:00 -0700 windows: add ARM64 VM runner for COFF smoke tests

test: add FreeBSD hosted link VM smoke

2026-06-04T23:40:32Z

commit 60e7c4ca59b86810a9103da2e217a85aa0bd3134 parent 7cb63a5b610c688fbf2705c7ad21ce4cabd1de41 Author: Ryan Sepassi Date: Thu, 4 Jun 2026 16:40:32 -0700 test: add FreeBSD hosted link VM smoke

link: emit FreeBSD PIE metadata for rtld

2026-06-04T23:39:18Z

commit 7cb63a5b610c688fbf2705c7ad21ce4cabd1de41 parent f80b5dd6d6c855d8baa47ce108d2e859564354fc Author: Ryan Sepassi Date: Thu, 4 Jun 2026 16:39:18 -0700 link: emit FreeBSD PIE metadata for rtld

link: fix FreeBSD hosted static executables

2026-06-04T23:39:01Z

commit f80b5dd6d6c855d8baa47ce108d2e859564354fc parent 516e7b47cb094dd0efd52a836af95a5a45672102 Author: Ryan Sepassi Date: Thu, 4 Jun 2026 16:39:01 -0700 link: fix FreeBSD hosted static executables

elf: add hosted FreeBSD relocation primitives

2026-06-04T23:37:08Z

commit 516e7b47cb094dd0efd52a836af95a5a45672102 parent 3e9ce3177b318a8cf1e867589ab9f7a4628715fc Author: Ryan Sepassi Date: Thu, 4 Jun 2026 16:37:08 -0700 elf: add hosted FreeBSD relocation primitives

freebsd: VM execution harness for amd64/aarch64/rv64

2026-06-04T20:57:12Z

commit 3e9ce3177b318a8cf1e867589ab9f7a4628715fc parent f2fa4890b0f5a8f2789856784b845889b4692c18 Author: Ryan Sepassi Date: Thu, 4 Jun 2026 13:57:12 -0700 freebsd: VM execution harness for amd64/aarch64/rv64 scripts/freebsd_vm.sh provisions and runs FreeBSD VMs as kit execution targets, plus doc/plan/FREEBSD.md. - Provision once, cache a compressed "golden" disk in the download cache (survives `make clean`); restore in seconds thereafter. - amd64: reset EDK2 NVRAM vars each run (else the firmware boots its UEFI Shell instead of the disk -- the "hang") and -vga none so the serial console is primary. - aarch64: virt + HVF (fast). - rv64: -machine virt,acpi=off (FreeBSD/riscv64 is FDT-only) plus an expect-driven serial bootstrap, since FreeBSD publishes no cloud-init riscv64 image. - cloud-init arches provision offline (restrict=on) so the first-boot freebsd-update fast-fails instead of running a slow networked update and forcing a reboot. All three boot and are SSH-reachable as the `kit` user.

freebsd: cross-compile target, runtime, and link support

2026-06-04T20:57:00Z

commit f2fa4890b0f5a8f2789856784b845889b4692c18 parent a292100f526e3ee83fd6f4e76eee1eb3a4694559 Author: Ryan Sepassi Date: Thu, 4 Jun 2026 13:57:00 -0700 freebsd: cross-compile target, runtime, and link support Make `kit cc --target=-freebsd` actually produce FreeBSD ELF for aarch64/x64/rv64: - target.c: parse the `freebsd` OS token (with version suffix, e.g. freebsd15.0) -> KIT_OS_FREEBSD/ELF. It previously fell through to freestanding. - runtime.c: add {x86_64,aarch64,riscv64}-freebsd runtime variants (ELF, mirroring the Linux runtime) so cc finds libkit_rt. - obj/elf: read ELF COMDAT groups -- record an SHT_GROUP section's signature symbol as absolute-defined instead of orphaning it into a phantom undef. FreeBSD's crt1.o brands the binary via a `.freebsd.note` COMDAT group. Also map STB_GNU_UNIQUE to a global binding. - hosted.c: link FreeBSD 15's libsys (split out of libc) when present. Linking a static hosted executable still hits the libc/libsys weak-alias archive cycle (undefined `openat`); see doc/plan/FREEBSD.md. test-elf/test-link/test-ar remain green.

windows: provision llvm-mingw UCRT sysroots for kit-only cross-compile

2026-06-04T18:36:24Z

commit a292100f526e3ee83fd6f4e76eee1eb3a4694559 parent 296361e67f879675a2e8f178cf6f9850a1a3e5ce Author: Ryan Sepassi Date: Thu, 4 Jun 2026 11:36:24 -0700 windows: provision llvm-mingw UCRT sysroots for kit-only cross-compile Add scripts/llvm_mingw_sysroot.sh to download the pinned mstorsjo/llvm-mingw UCRT release and extract only each target's include/ and lib/ (headers, CRT objects, import archives) — never the bundled clang/gcc/ld/lld tools. kit cross-compiles and links x86_64-windows and aarch64-windows entirely on its own against these sysroots. - hosted Windows-mingw profile links libucrt.a (UCRT) instead of libmsvcrt.a and is renamed windows-mingw-ucrt; cc.c sysroot comment follows suit - mk/rt.mk wires the x86_64 __chkstk stack-probe into the windows runtime - mk/test.mk adds a windows-ucrt-sysroots target (provenance-marker driven) and points test-coff-windows-ucrt at both runtimes + both smoke scripts - COFF smokes discover sysroots under build/llvm-mingw/*/ucrt, validate via libucrt.a, and gain an optional Windows-VM execution path (scripts/ windows_vm.sh over SSH); they remain self-skipping with no sysroot - doc/WINDOWS.md documents provisioning + VM execution test-coff smokes use kit exclusively (kit cc to compile+link, kit objdump -p to inspect): 78 + 126 link-level asserts pass for both arches; VM/Wine exec paths self-skip when unconfigured.

test: add freestanding qemu-system smoke

2026-06-04T17:49:40Z

commit 296361e67f879675a2e8f178cf6f9850a1a3e5ce parent a8052633f6775e1cba26a4b7a6db19f237e98c22 Author: Ryan Sepassi Date: Thu, 4 Jun 2026 10:49:40 -0700 test: add freestanding qemu-system smoke

aa64: support bitfield extension aliases

2026-06-04T17:45:21Z

commit a8052633f6775e1cba26a4b7a6db19f237e98c22 parent f7fbbc9db78d92ef2796ccc4def521332e77e50e Author: Ryan Sepassi Date: Thu, 4 Jun 2026 10:45:21 -0700 aa64: support bitfield extension aliases

Add initial release punchlist

2026-06-04T16:59:48Z

commit f7fbbc9db78d92ef2796ccc4def521332e77e50e parent a89695e3122b7a118a91d8828d350e2ffc18409c Author: Ryan Sepassi Date: Thu, 4 Jun 2026 09:59:48 -0700 Add initial release punchlist

Drain driver TODO backlog

2026-06-04T15:43:13Z

commit a89695e3122b7a118a91d8828d350e2ffc18409c parent 58a41780a77ed786a88361881b40d8059cc29a94 Author: Ryan Sepassi Date: Thu, 4 Jun 2026 08:43:13 -0700 Drain driver TODO backlog

driver: add build-exe/build-lib/build-obj, retire compile

2026-06-04T15:18:05Z

commit 58a41780a77ed786a88361881b40d8059cc29a94 parent 438e6b094ea31ed5ec9787d451e59ddc9a7770f9 Author: Ryan Sepassi Date: Thu, 4 Jun 2026 08:18:05 -0700 driver: add build-exe/build-lib/build-obj, retire compile Kit-native build verbs that compile a polyglot source set (C/asm/toy/wasm, resolved per file) entirely in memory and produce the final artifact in one invocation: build-exe link an executable build-lib static .a (dynamic/shared not yet supported) build-obj one object, --emit=asm|c|ir, -fsyntax-only, or an `ld -r` combine of N sources; full replacement for the retired `compile` tool Flags split into global/per-output and a scopable set (-I/-isystem/-D/-U, -x, -X) overridable per source group via `--group [flags] -- sources`. Per-language frontend flags route through -X. Factoring: - driver/lib/link_engine: build-session/add-in-order/emit step; cc's link path now routes through it (no behavior change). - driver/lib/archive_engine: in-memory object serialize + ar symbol index. build-obj's multi-source combine matches `ld -r` symbol-for-symbol. Register the tools (config gate, main table, default install group), remove compile.c + its test, migrate test/compile -> test/buildcmds, update doc/DRIVER.md, and add doc/plan/TODO.md cataloguing pre-existing issues found along the way.

doc/plan: add LTO / whole-program optimization plan

2026-06-04T13:56:18Z

commit 438e6b094ea31ed5ec9787d451e59ddc9a7770f9 parent bde9b5848be4f7d9b2587abaaf7c4ad02417143c Author: Ryan Sepassi Date: Thu, 4 Jun 2026 06:56:18 -0700 doc/plan: add LTO / whole-program optimization plan Forward-looking design for making a library/executable look like a single TU to the optimizer. Chooses shared-context recording (all sources record into one session/ObjBuilder/CgIrModule) over clone-and-merge, since globals already intern by name, the recorder already accumulates a whole module, and the finalize sweep + unreached opt_inline already exist. Centers the work on factoring the linker's resolution policy into a shared symresolve module usable at merge time. Phased: whole-TU opt, all-sources LTO, serialized .kit.ir objects.

doc/plan: design for build-exe/build-lib/build-obj (replacing compile)

2026-06-04T13:49:20Z

commit bde9b5848be4f7d9b2587abaaf7c4ad02417143c parent 8d3bb285e71aed150d99db5019df5876c1f245a0 Author: Ryan Sepassi Date: Thu, 4 Jun 2026 06:49:20 -0700 doc/plan: design for build-exe/build-lib/build-obj (replacing compile) Kit-native build verbs that compile a mixed-language source set in memory and link/archive in one shot, replacing the no-link 'compile' tool. Captures the --group flag-scoping grammar, the global-vs-scopable taxonomy, -X frontend-flag routing, hybrid naming (kit --emit/-o + Zig -static/-dynamic), the link_engine/archive_engine extraction plan, migration, and the resolved design decisions.

wasm: tighten shared-memory max to min pages so atomics modules load

2026-06-04T09:24:24Z

commit 8d3bb285e71aed150d99db5019df5876c1f245a0 parent 3a93557bc838c194609de1e670e7f831f74fb61e Author: Ryan Sepassi Date: Thu, 4 Jun 2026 02:24:24 -0700 wasm: tighten shared-memory max to min pages so atomics modules load ensure_shared_memory promotes the module memory to shared (required for wasm atomic ops) with a provisional 4 GiB cap (65536 pages). The comment claimed wasm_materialize_data tightens that to the real size, but the code only ever raised min toward max -- never lowered max -- so atomics modules shipped a 4 GiB declared maximum. `kit run` reserves the declared maximum and caps total linear memory at 1 GiB, so every such module was rejected at load ("wasm linear memory reservation exceeds 1073741824 bytes"). Tighten max_pages down to min_pages for shared memory once the final layout is known. The backend never emits memory.grow, so a fixed shared memory is correct, and embedders now reserve a snug arena. Fixes the 8 toy atomics cases on lane W (124/17/22/59/73/74/75/77) and lets the C-frontend wasm atomics cases run instead of failing at load.

cg/asm: reject 64-bit values in register asm constraints on 32-bit targets

2026-06-04T08:58:34Z

commit 3a93557bc838c194609de1e670e7f831f74fb61e parent a7a3338d08760d46faf61a0fc6f82f01b62c8cc1 Author: Ryan Sepassi Date: Thu, 4 Jun 2026 01:58:34 -0700 cg/asm: reject 64-bit values in register asm constraints on 32-bit targets A 64-bit scalar bound to a register constraint ("r"/"f"/"x"/"w") needs a register pair on a 32-bit target, which the inline-asm lowering does not model -- it would bind the value to a single GPR and silently truncate to the low word. Reject it up front in kit_cg_inline_asm at all three operand sites (output, input, early-clobber output) via api_is_wide8_scalar_type; the source can use a memory ("m") constraint or split into two 32-bit operands instead. Add a red-green err test (invalid_asm_wide_reg_rv32) compiled for rv32 via a new err/.ccargs sidecar in the toy ERR lane (per-case cc flags). Convert the toy asm fixtures that fed a 64-bit value to a register operand (19, 102, 105, 110) to isize so the operand fits one GPR on every target; they keep exercising register asm operands and now pass on the rv32 exec lane without a skip. Same conversion fixes 20_cg_api_inline_asm_full, which was already red on native after integer-literal widening made bare inout("+r", 7) an i8 that no longer matched its @asm result; with empty templates it now also passes on rv32, so its stale .rv32.skip is dropped.

rv32 parse: resolve LP64-assumption corpus failures

2026-06-04T08:39:37Z

commit a7a3338d08760d46faf61a0fc6f82f01b62c8cc1 parent c5166d0e42265a19c8f18223e0cb43bb35497e5b Author: Ryan Sepassi Date: Thu, 4 Jun 2026 01:39:37 -0700 rv32 parse: resolve LP64-assumption corpus failures The rv32 (ilp32) parse lane had 3 cases failing because they bake in LP64 (64-bit long/size_t) assumptions. kit's rv32 codegen is correct — clang as an independent ilp32 oracle returns the identical values (10, 16, 9), confirming the test expectations, not the codegen, are target-specific. - builtin_26_sadd_overflow: gate the long-width (*l_overflow) assertions behind #if __SIZEOF_LONG__ == 8. They hard-code 64-bit long boundary constants; under ilp32 `long` is a 32-bit width-duplicate of the int builtins. The case previously bailed in this block, so the 64-bit long long overflow builtins were never reached on rv32 — they now run and pass, adding real 64-bit-overflow-on-32-bit-target coverage. Unchanged on LP64 (verified 42 on rv32 kit+clang and native LP64). - 6_5_64_unsigned_size_division.rv32.skip: the 64-bit size_t overflow-check idiom; the clamp legitimately fires under ilp32. - rv64_atomic_widths_orders.rv32.skip: rv64-named 64-bit atomics via 64-bit long; rv32's 32-bit atomics stay covered by builtin_06..25. - test/parse/run.sh: set KIT_SKIP_IS_FAILURE=0 to match the toy runner, so the corpus's permanent ilp32 skip floor (i128_*/ldbl128_*, ...) no longer gates the exit; only a real FAIL or XPASS does.

test/rv32: drop clang gate from bare-metal toy/parse lanes

2026-06-04T05:15:10Z

commit c5166d0e42265a19c8f18223e0cb43bb35497e5b parent 4931d48851d32a219a48d75854daf88be384e4e1 Author: Ryan Sepassi Date: Wed, 3 Jun 2026 22:15:10 -0700 test/rv32: drop clang gate from bare-metal toy/parse lanes The rv32 toy/parse exec lanes build the entire image with kit (kit cc / kit as / kit ld) and the runtime archive via kit cc (RT_CC), then run under qemu-system-riscv32. clang is never invoked on this path, yet rv32_bare_setup refused to run unless the clang riscv32 target probe passed. Gate on qemu-system-riscv32 alone and update the skip messages.

rv32 lowering fixes

2026-06-04T05:01:49Z

commit 4931d48851d32a219a48d75854daf88be384e4e1 parent 056fc9ea14811f46112837b87a5a7de5630bce80 Author: Ryan Sepassi Date: Wed, 3 Jun 2026 22:01:49 -0700 rv32 lowering fixes

toy: pointer-width isize/usize, int widening

2026-06-04T05:01:27Z

commit 056fc9ea14811f46112837b87a5a7de5630bce80 parent 87ffbeeee00a842e9209d9665120e25ce814e3d4 Author: Ryan Sepassi Date: Wed, 3 Jun 2026 22:01:27 -0700 toy: pointer-width isize/usize, int widening

Test ABI split-lane scalar query

2026-06-04T04:27:16Z

commit 87ffbeeee00a842e9209d9665120e25ce814e3d4 parent e4a333278d9c614276c265232293826e26aa3b63 Author: Ryan Sepassi Date: Wed, 3 Jun 2026 21:27:16 -0700 Test ABI split-lane scalar query

rv32: use ABI split-lane scalar metadata

2026-06-04T04:22:40Z

commit e4a333278d9c614276c265232293826e26aa3b63 parent ea0e478e0cdbf5187b454ea2844b83a9e12e2964 Author: Ryan Sepassi Date: Wed, 3 Jun 2026 21:22:40 -0700 rv32: use ABI split-lane scalar metadata Add target data-model width selection for C long/size_t spellings, expose ABI scalar split-lane metadata, and use it to gate wide8 lowering/backend guards. Also fix frame-backed indirect index materialization to load the pointer-width value and add an RV32 parser fixture for data-model widths.

Fix rv32 toy target arch and wide intrinsics

2026-06-04T03:55:10Z

commit ea0e478e0cdbf5187b454ea2844b83a9e12e2964 parent 1576564d10ca3d46a2d2b708384883adcda9270c Author: Ryan Sepassi Date: Wed, 3 Jun 2026 20:55:10 -0700 Fix rv32 toy target arch and wide intrinsics

rv32: fix 8-byte variadic high-word drop, O1 label-addr width, long double, atomics

2026-06-04T03:37:25Z

commit 1576564d10ca3d46a2d2b708384883adcda9270c parent 86a819d3f72a1da40cc323e83bc411fefd5fea50 Author: Ryan Sepassi Date: Wed, 3 Jun 2026 20:37:25 -0700 rv32: fix 8-byte variadic high-word drop, O1 label-addr width, long double, atomics Variadic 8-byte args (i64 / soft-double) were dropping their high 32 bits on rv32. The caller synthesized a single 8-byte ABI part for an unnamed arg, which the per-part marshaller loaded into one GPR (low word only); the callee va_arg fetched 4 bytes and advanced the cursor by one slot. - rv_param_abi: split a variadic scalar wider than one GPR into one INT part per word (low word in the lower-numbered reg), matching the named-arg classifier. - rv_va_arg_wide: new helper that copies the full value out of the (contiguous) save area and advances the cursor by the whole span; used by both the -O0 direct and -O1 native-emit paths for any value too wide for one GPR. - pass_native_emit IR_VA_ARG: route aggregate/oversized va_arg through its memory destination instead of a scratch register. rv64/aa64/x64 unaffected (no test passes an 8-byte value as wide there). O1 static-data label addresses (jump tables, &&label-in-data) panicked with "unsupported local static label width" on rv32: the native-emit path hardcoded R_ABS64. Make it width-aware (R_ABS32 for 4-byte slots), mirroring the -O0 direct path. This also fixes the gnu_label_addr_* parse cases. long double on rv32 is a double alias (8-byte), not IEEE-754 binary128, so the ldbl128_* cases (which assert sizeof==16 / MANT_DIG==113) can't hold there; add .rv32.skip sidecars. Generic long-double use (6_7_2_12) keeps working as double. builtin_24_atomic_lock_free: size the lock-free checks against the native word (sizeof(void*)) instead of a fixed 8 bytes, and drop the __int128 dead code, so the same .expected (42) holds on ilp32 and lp64. rv32 has no 64-bit AMO, so an 8-byte atomic is correctly not lock-free there.

rv32: close most remaining cross-target reds (atomics, overflow, soft-fp, TLS, __int128, CSR, asm goldens)

2026-06-04T02:37:22Z

commit 86a819d3f72a1da40cc323e83bc411fefd5fea50 parent 01252f062d3256bf8804f706e0226db5be82efcd Author: Ryan Sepassi Date: Wed, 3 Jun 2026 19:37:22 -0700 rv32: close most remaining cross-target reds (atomics, overflow, soft-fp, TLS, __int128, CSR, asm goldens) Closes the bulk of the deliberately-red rv32 gaps from doc/plan/RV32.md. rv64/x64/aa64 non-regressed (isa rv32 90/rv64 79/aa64 43, asm-rv64 43/0 + asm-rv32 16/0, elf 41/0, link 122/0, smoke-rv64 3/0 + smoke-rv32 7/0, cg-api, abi-classify 367/0). - i64 atomics: route 8-byte atomics to the existing __atomic_*_8 rt libcalls on rv32 (is_lock_free/is_legal made target+size-aware). Fixes the 123_spec_demo hang + 7 toy atomic cases. New 124_atomic_word_ops covers word atomics on all lanes. - i64 overflow intrinsics: inline 2-lane lowering in kit_cg_intrinsic (uadd/usub carry/borrow, sadd/ssub sign formula, umul/smul 128-bit product). - soft-fp compare: api_wide8_addr materializes a delayed wide8 operand before taking its address (fixes toy 153 at -O0). - TLS: rv32 bare-metal startup stub sets up a static-TLS block + tp. Needed and adds a kit-ld linker-script fix: accept the 1-arg ALIGN(align) GNU form. - __int128: rejected on rv32 with a target-gated diagnostic (matches gcc/clang 64-bit-only); 14 i128_* cases skipped via .rv32.skip sidecars. - CSR pseudo-ops: csrr/csrw/csrs/csrc/csrwi/csrsi/csrci + CSR-name table (benefits rv64 too); both rv32 startup stubs now kit-assembled (clang stub dependency dropped). - asm goldens: rv32 byte-golden lane + regen-rv32.sh + corpus; opt-in test-asm-rv32/test-toy-rv32/test-parse-rv32 targets. Also fixes a real assembler bug: bare fcvt.w.s defaulted rm to RTZ, now dyn (gas/clang); stale rv64 fp goldens regenerated. - test-env skip sidecars for arch-specific toy/C cases (145/20/47, asm_01). Still red (follow-up): i64/double varargs (high 32 bits of an 8-byte vararg are dropped — caller/callee ABI), 123_spec_demo -O1 (switch-table label width), and the pre-existing rv32 -O1 cluster + ldbl128.

rv32: auto-build and auto-link the freestanding runtime like any target

2026-06-04T00:25:51Z

commit 01252f062d3256bf8804f706e0226db5be82efcd parent bf5b73bcb4b8e9cffd12aa173a5a5759cbe71381 Author: Ryan Sepassi Date: Wed, 3 Jun 2026 17:25:51 -0700 rv32: auto-build and auto-link the freestanding runtime like any target riscv32-none-elf was the only freestanding target without a driver RuntimeVariant, so kit cc/ld could not auto-build or auto-link libkit_rt.a for it — the smoke test had to `make rt` and pass the archive explicitly. The blocker was that rv32 has two float ABIs sharing one arch+pointer-width (ilp32 soft, ilp32f single) and the runtime must match, but RuntimeVariant could not distinguish them and target detection did not recover the float ABI. - driver/lib/runtime.c: add a float_abi axis (+ isa/abi march strings) to RuntimeVariant and two rv32 entries (riscv32-elf soft, riscv32-elf-hardfloat single). The matcher keys on float_abi with DEFAULT as a wildcard, so every existing variant is unchanged; the on-demand build passes each variant's -march/-mabi via topts.isa/abi. - src/api/object_detect.c: recover the float ABI from RISC-V ELF e_flags (offset 36/48; soft/single/double bits) so a detected target selects the matching runtime. - driver/cmd/ld.c: reconcile float_abi across all link inputs, preferring a hardware ABI, so a foreign (clang-assembled) startup stub lacking the flag never mis-selects the soft runtime. - test/smoke/rv32.sh: two kit-ld auto-rt lanes (ilp32f + ilp32) that link with no runtime archive named. Verified: smoke-rv32 7/7 under qemu; no regressions (elf 41/0, link 122/0, abi-classify 367/0, cg-api 0 failures, driver-cc 98/0, smoke-rv64 3/0); riscv64-none-elf auto-link still works.

rv32: bring up and complete riscv32-none-elf cross target

2026-06-04T00:02:26Z

commit bf5b73bcb4b8e9cffd12aa173a5a5759cbe71381 parent 6ea2c9a48aa1a8670eb9a9005cd63d8b7c002a28 Author: Ryan Sepassi Date: Wed, 3 Jun 2026 17:02:26 -0700 rv32: bring up and complete riscv32-none-elf cross target Refactor src/arch/rv64 -> src/arch/riscv as one XLEN-parameterized RISC-V backend (RiscvVariant descriptor) and wire riscv32-none-elf end to end. RV64 / x64 / aa64 are fully non-regressed. Backend bring-up: - WS0 RiscvVariant (src/arch/riscv/variant.{h,c}) + KIT_ARCH_RV32_ENABLED. - WS1 rv64->riscv rename; variant threaded through native.c (rv64 byte-identical). - WS2 ISA availability-mask (RV_AV_RV32/RV64) + asm/disasm/link/dbg XLEN param; link_arch_rv32 (lw stubs), rv32_dbg_ops (min=2/max=4, RVC->step-over). - WS3 arch_impl_rv32 + registry; -march via variant->isa_prefix; rv32 default rv32imafc_zicsr_zifencei (D cleared); ilp32f predefined macros. - WS4a KitFloatAbi + KitTargetSpec.float_abi + KitTargetOptions.abi; -mabi parse/resolve/validate (ilp32d-without-D and width-mismatch rejected); medlow/medany mcmodel aliases. - WS4b shared RISC-V ABI classifier (descriptor {gpr_bytes,aggregate,flen, float_abi}) + rv32_vtable + registry. - WS5 ELFCLASS32 emit/read/link + reloc_riscv32.c + EI_CLASS machine disambiguation (obj_elf_machine_class); width-aware jump-table label relocs (R_ABS32 on rv32); layout_dyn clean-panics on an ELF32 dynamic/PIE link (was an ELF64 SEGV). - WS7 mk/rt.mk variants fixed (riscv32-elf=ilp32/rv32imac, +riscv32-elf-hardfloat =ilp32f/rv32imafc); mk/lib_srcs.mk ABI guard widened. RT_CFLAGS/RT_ASFLAGS now include ARCH_FLAGS (the -mabi/-march were silently dropped). ELF e_flags float-ABI derived from target.float_abi (was hardcoded SINGLE, mislabelling ilp32 soft). WS6 64-bit-value legalization (the blocker): rv32 8-byte scalars (long long/i64 AND soft double) are now memory-resident and legalized to 2-word lane ops -- inline add/sub/and/or/xor/neg/bnot/compare (carry/borrow via sltu; no compiler-rt 64-bit add exists), __*di3 for mul/div/rem/shift, __*df*/__*sf* for soft double/single, i64<->float via __floatdisf/__fixsfdi, __*di2 for clz/ctz/popcount/bswap, 2-lane 64-bit consts. The allocator binds one register per value, so memory residence + the multi-part ABI path is the only correct representation. nd_* guards panic on any 8-byte value reaching a single-register op. Verified under qemu-system-riscv32 at -O0/-O1 for both ilp32f and ilp32. Freestanding link policy is target-derived, host-irrelevant: kit stamps EI_OSABI=ELFOSABI_STANDALONE on *-none-elf objects (fixes "none decodes as Linux"), kit ld derives the PIC default via driver_default_pic (hosted PIE / freestanding no-PIE) scanning all inputs, and auto-links a runtime only when one exists (driver_runtime_has_variant) -- so a freestanding rv32 link needs no -no-pie / -nostdlib. .eh_frame suppressed for KIT_OS_FREESTANDING; new -Ttext / -nostdlib flags. Tests: rv32_decode_test (test-isa, 31 checks), rv32_jit_test (test-rv32-jit), rv32_class32 ELFCLASS32 round-trip (test-elf), test/smoke/rv32.sh (test-smoke-rv32, 6 lanes incl. a kit-ld end-to-end lane), and rv32 wired as a cross arch in the Toy (test/toy/run.sh path X, 240/15) and C (test/parse/run.sh path E, 439/36) corpus lanes via the shared test/lib/exec_rv32_bare.sh. Reds are left red on purpose -- real remaining gaps (i128, i64 atomics/overflow/varargs, TLS) are tracked in doc/plan/RV32.md. RV64/x64/aa64 fully non-regressed.

Fix Apple ARM64 variadic stack slots

2026-06-04T03:38:12Z

commit 6ea2c9a48aa1a8670eb9a9005cd63d8b7c002a28 parent 6ccc62a9ce1e2feed3f6d085b011571cccfc8447 Author: Ryan Sepassi Date: Wed, 3 Jun 2026 20:38:12 -0700 Fix Apple ARM64 variadic stack slots

Implement wasm sandbox host shim

2026-06-04T03:28:36Z

commit 6ccc62a9ce1e2feed3f6d085b011571cccfc8447 parent 9bd665b95d72f696db659675bf3c85367e1b8cf1 Author: Ryan Sepassi Date: Wed, 3 Jun 2026 20:28:36 -0700 Implement wasm sandbox host shim

Harden wasm runtime import handling

2026-06-04T02:18:22Z

commit 9bd665b95d72f696db659675bf3c85367e1b8cf1 parent 5bd8de1ceeabdaef0abc018ea94891371381655d Author: Ryan Sepassi Date: Wed, 3 Jun 2026 19:18:22 -0700 Harden wasm runtime import handling

aa64 asm: parse logical-immediate and lsl/lsr/asr aliases

2026-06-03T23:00:03Z

commit 5bd8de1ceeabdaef0abc018ea94891371381655d parent e156c02981eb92a6d0f9303fbde37ed4e6b83615 Author: Ryan Sepassi Date: Wed, 3 Jun 2026 16:00:03 -0700 aa64 asm: parse logical-immediate and lsl/lsr/asr aliases The disassembler (driving `cc -S`) emits AArch64 forms the assembler couldn't parse back, breaking the `cc -c` == `cc -S | as` round-trip for 48 cases (all the L1 lane). Two gaps, both exposed once codegen started using immediate operand forms at -O0 (dc1854de): - Logical bitmask-immediate `and/orr/eor/ands Rd, Rn, #imm`: p_log_sr parsed the third operand strictly as a register ("expected register"). - Shift aliases `lsl/lsr/asr Rd, Rn, #imm`: the mnemonics weren't in the table at all, only the variable-shift lslv/lsrv/asrv ("unknown mnemonic"). Add a peek_is_reg lookahead, extend p_log_sr with the bitmask-immediate form (via aa64_logimm_encode), and add p_shift handling both the register (LSLV/LSRV/ASRV) and immediate (UBFM/SBFM) forms. ror/EXTR is left out: the disassembler doesn't decode EXTR, so adding it would introduce a new encode-only asymmetry; rorv already covers the register rotate. Every new encoding cross-checked byte-for-byte against llvm-mc. Adds an encode corpus case; the symmetry baseline now drops 3 mov_orr_bitmask encode-only entries that round-trip thanks to the logical-immediate path. test-asm-roundtrip: 572 pass, 0 fail (was 524 pass / 48 fail).

make: add rt to default target

2026-06-03T22:24:17Z

commit e156c02981eb92a6d0f9303fbde37ed4e6b83615 parent 1ccdd814588c5b6750f3f0aca37b47b9aed6907a Author: Ryan Sepassi Date: Wed, 3 Jun 2026 15:24:17 -0700 make: add rt to default target

doc: regenerate CODE_SIZE.md after the economy refactor

2026-06-03T22:04:23Z

commit 1ccdd814588c5b6750f3f0aca37b47b9aed6907a parent a3e7b03a7d9a882ee7f73e48a7ccfffd0f44dd44 Author: Ryan Sepassi Date: Wed, 3 Jun 2026 15:04:23 -0700 doc: regenerate CODE_SIZE.md after the economy refactor

Exercise toy cross TLS startup with kit

2026-06-03T21:53:22Z

commit a3e7b03a7d9a882ee7f73e48a7ccfffd0f44dd44 parent 5fec69282f697c6a701638e7da969c3f60fd52b0 Author: Ryan Sepassi Date: Wed, 3 Jun 2026 14:53:22 -0700 Exercise toy cross TLS startup with kit

Add AA64 system register asm support

2026-06-03T21:53:11Z

commit 5fec69282f697c6a701638e7da969c3f60fd52b0 parent 84863072e412579624a0d6c9f304344ebc4b1289 Author: Ryan Sepassi Date: Wed, 3 Jun 2026 14:53:11 -0700 Add AA64 system register asm support

Support explicit register asm operands

2026-06-03T21:53:02Z

commit 84863072e412579624a0d6c9f304344ebc4b1289 parent 2d37ba7b367e02cf86618cfa9e3a7a92d516e7ea Author: Ryan Sepassi Date: Wed, 3 Jun 2026 14:53:02 -0700 Support explicit register asm operands

cg: restore out-of-range fallback in intrinsic table accessors

2026-06-03T19:12:29Z

commit 2d37ba7b367e02cf86618cfa9e3a7a92d516e7ea parent b9d9f14016146421df4aad8dc36cb35279db6258 Author: Ryan Sepassi Date: Wed, 3 Jun 2026 12:12:29 -0700 cg: restore out-of-range fallback in intrinsic table accessors The four KitCgIntrinsic accessors index kIntrinTable directly; route them through a bounds-guarded intrin_desc() so an out-of-range value falls back to the NONE row, matching the defensive default: arms the switches carried before the table collapse. (Equivalence-review nit; valid enum inputs unaffected.)

plan: RV32

2026-06-03T18:46:44Z

commit b9d9f14016146421df4aad8dc36cb35279db6258 parent 972d2c69b2cde627f555c57732534d1452a281f0 Author: Ryan Sepassi Date: Wed, 3 Jun 2026 11:46:44 -0700 plan: RV32

test/coff: fix stale ARM64EC test to use KitTargetSpec (feature-model API)

2026-06-03T18:34:25Z

commit 972d2c69b2cde627f555c57732534d1452a281f0 parent dddc1f61c90cfd8dce44684c739c6e9242f43a72 Author: Ryan Sepassi Date: Wed, 3 Jun 2026 11:34:25 -0700 test/coff: fix stale ARM64EC test to use KitTargetSpec (feature-model API) The obj-api-tables refactor added an ARM64EC detection test written against the pre-feature-model KitTarget type; rename to KitTargetSpec (which carries .arch/.obj) to match the current kit_detect_target signature.

cc: drop chatty "ignoring accepted compatibility flag" log

2026-06-03T18:26:48Z

commit dddc1f61c90cfd8dce44684c739c6e9242f43a72 parent c6c3fef5166e4061cab5a5959cfc7e10fc069def Author: Ryan Sepassi Date: Wed, 3 Jun 2026 11:26:48 -0700 cc: drop chatty "ignoring accepted compatibility flag" log The debug-only log fired for every accepted-but-ignored compatibility flag (-Wall, -std=, -nostdlib, ...), drowning out real diagnostics. Drop the cc_log_ignored() helper and its 10 call sites; the flags are still accepted, now silently. Update the driver test to assert silent acceptance (clean stderr) instead of the removed message.

obj/link: single-source no-location as SRCLOC_NONE; drop 22 no_loc() copies

2026-06-03T18:18:42Z

commit c6c3fef5166e4061cab5a5959cfc7e10fc069def parent a135713defeb03f4aca798cc4257b0f66efdae97 Author: Ryan Sepassi Date: Wed, 3 Jun 2026 11:18:42 -0700 obj/link: single-source no-location as SRCLOC_NONE; drop 22 no_loc() copies Replace the per-TU `static SrcLoc no_loc(void)` helper (duplicated verbatim across 22 obj/ and link/ writers) with one SRCLOC_NONE macro in core/core.h. Mechanical: 397 call sites rewritten, 22 definitions removed.