nocc: stage-0 bootstrap path (binaries gitignored)

Sets up the cc-free fresh-checkout flow without yet committing the
stage-0 binaries. The pieces are in place; flipping the switch is
one `git add -f` away when the compiler is judged stable.

  bootstrap/<arch>/{ww,6c,6a,6l}   stage-0 binaries (gitignored)
  bootstrap/README.md              layout + workflow

Two new targets:

  make bootstrap-snapshot   populates bootstrap/$(ARCH)/ from the
                            currently-built wwstage
  make nocc                 starts from bootstrap/$(ARCH)/, rebuilds
                            the wwstage from source, gates on
                            stage-0 == rebuilt byte-for-byte. Lands
                            in $(OUT)/nocc/ so $(OUT)/ stays
                            untouched. cc is never invoked.

The fixed-point gate has the same shape as `make bootstrap`, just
the entry point flipped: that target trusts cstage; nocc trusts
the checked-in (or local-snapshot) stage-0 binaries.

Implementation notes:

- libwwrt.a is assembled with stage-0 6a, members ordered to match
  $(RT_OBJ) so the embedded symbol table reproduces the cstage
  build byte-for-byte.
- The driver's default lib path is $self_dir/../../lib, which under
  $(NOCC_BIN) resolves to $(NOCC_OUT)/lib (libwwrt only). Each
  ww-build invocation passes -I $(CURDIR)/lib so `use os` etc.
  resolve to source.
- Stage-0 binaries are duplicated under both natural names (ww/6c/
  6a/6l) and _ww-suffixed names so the wwstage driver's
  hard-coded join_path_lit(self_dir, "6c_ww") still finds them.
  When cmd/wwc/ gets deleted at v1.0, that duplication and the
  _ww suffix in the driver both go away.
This commit is contained in:
2026-05-11 11:51:34 +09:00
parent 37bffa5284
commit edfc4273f6
4 changed files with 157 additions and 17 deletions

7
.gitignore vendored
View File

@@ -21,3 +21,10 @@ examples/**/*.o
examples/**/*.s
examples/**/*.combined.ww
examples/snake/snake
# Stage-0 binaries under bootstrap/<arch>/. Untracked by default —
# committing them is the v1.0 lock per PLAN.md (the new trust
# surface). `make bootstrap-snapshot` populates the dir from the
# current wwstage; `make nocc` verifies they self-reproduce.
# To ship, `git add -f bootstrap/amd64/{ww,6c,6a,6l}` explicitly.
bootstrap/*/*

View File

@@ -36,32 +36,56 @@ involved.
## Make targets
make builds cstage + wwstage + libs
make cstage Cstage only (cc → ww, 6c, 6a, 6l, wwdump, libs)
make wwstage wwstage only (assumes cstage)
make bootstrap three-stage build; gates on cmp ww2 == ww3
make test runs all tests, including the 990993 diffs
(ww-side tools vs C-side tools, byte-for-byte)
make builds cstage + wwstage + libs
make cstage Cstage only (cc → ww, 6c, 6a, 6l, wwdump, libs)
make wwstage wwstage only (assumes cstage)
make bootstrap three-stage build; gates on cmp ww2 == ww3
make test runs all tests, including 990995 (ww-side
tools vs C-side tools, byte-for-byte, plus
the wwstage rebuilding itself in 995)
Without cc:
make bootstrap-snapshot populate bootstrap/$(ARCH)/ from the
currently-built wwstage (gitignored
by default)
make nocc cold-start from bootstrap/$(ARCH)/,
rebuild the wwstage from source, gate
on `stage-0 == rebuilt` byte-for-byte.
cc is never invoked.
## Status
Phase 10 of `PLAN.md` is at its first exit criterion: the bootstrap is
green and `selfhost/cmd/{wwc,6a,6l,ww}` are byte-identical to their
Cstage counterparts. The second exit criterion — deleting the C trees
— is **deferred to v1.0**: until the compiler stops churning, the
Cstage stays as the fresh-checkout entry point. A `git clone` today
still requires `cc`.
green and `selfhost/cmd/{wwc,6c,6a,6l,ww}` are byte-identical to their
Cstage counterparts. Test 995 pins the stronger property — the
wwstage rebuilds every one of its own tools through `ww_ww + 6c_ww +
6a_ww + 6l_ww`, byte-for-byte.
`make nocc` is wired and verifies stage-0 self-reproduction locally;
the stage-0 binaries under `bootstrap/$(ARCH)/` are gitignored until
ready to commit. The second exit criterion — deleting the C trees —
is **still deferred to v1.0**: shipping the stage-0 binaries is a
one-way door (they become the project's new trust surface), and the
compiler is still churning. A `git clone` today still requires `cc`
unless the local working copy has already done `make
bootstrap-snapshot` once.
## What "no C" looks like
The pieces are already wired; only the binary commit is held back.
When the compiler stabilises:
1. Cross-compile or natively build a known-good wwstage on each
supported arch.
2. Ship those binaries under `bootstrap/<arch>/{ww,6c,6a,6l}` in the
tree. They are the new stage 0.
1. On each supported arch, run `make wwstage && make bootstrap-snapshot`
to populate `bootstrap/<arch>/{ww,6c,6a,6l}` from the current
wwstage. Verify with `make nocc`.
2. `git add -f bootstrap/<arch>/{ww,6c,6a,6l}` — explicit because
the dir is gitignored. The binary SHAs become the new trust
surface.
3. Delete `cmd/wwc/`, `cmd/6c/`, `cmd/6a/`, `cmd/6l/`, `cmd/ww/`.
4. Move `selfhost/cmd/*` to `cmd/*`.
Drop `CC`/`AR` C-tool dependencies from the top-level Makefile.
4. Move `selfhost/cmd/*` to `cmd/*` and `make nocc` becomes the
default `make`.
5. Build path: stage-0 binary → ww1 → ww2 → fixed point, same as
today, just without `cc`.

View File

@@ -307,4 +307,78 @@ bootstrap: all
@echo " - cmd/6l/ : ww-6l (linker w/ archive support) — byte-identical to C 6l (test 992)"
@echo " - cmd/ww/ : ww-driver (build/run/version) — byte-identical to C ww (test 993)"
.PHONY: all cstage wwstage test install clean bootstrap
# ---- nocc: bootstrap from a checked-in stage-0 binary, no cc -----------
# A fresh-checkout flow that never invokes cc. Mirrors `make bootstrap`'s
# fixed-point check but uses `bootstrap/$(ARCH)/{ww,6c,6a,6l}` as the
# starting point and lands in $(OUT)/nocc/ so the cstage build under
# $(OUT)/ stays untouched. See bootstrap/README.md and BOOTSTRAP.md.
#
# `make bootstrap-snapshot` populates bootstrap/$(ARCH)/ from the
# currently-built wwstage so you can verify locally before deciding
# to commit the binaries.
ARCH ?= amd64
STAGE0 = bootstrap/$(ARCH)
NOCC_OUT = $(OUT)/nocc
NOCC_BIN = $(NOCC_OUT)/bin
NOCC_LIB = $(NOCC_OUT)/lib
NOCC_OBJ = $(NOCC_OUT)/obj
bootstrap-snapshot: wwstage
@mkdir -p $(STAGE0)
@cp $(BIN)/ww_ww $(STAGE0)/ww
@cp $(BIN)/6c_ww $(STAGE0)/6c
@cp $(BIN)/6a_ww $(STAGE0)/6a
@cp $(BIN)/6l_ww $(STAGE0)/6l
@echo "snapshot: $(STAGE0)/{ww,6c,6a,6l} populated from $(BIN)/*_ww"
@echo " (gitignored — 'git add -f bootstrap/$(ARCH)/*' to ship)"
nocc:
@for f in ww 6c 6a 6l; do \
test -x $(STAGE0)/$$f || { \
echo "$@: missing $(STAGE0)/$$f"; \
echo "$@: run 'make bootstrap-snapshot' first, or read bootstrap/README.md"; \
exit 1; \
}; \
done
@echo "=== nocc: starting from $(STAGE0) (no cc invoked) ==="
@rm -rf $(NOCC_OUT)
@mkdir -p $(NOCC_BIN) $(NOCC_LIB) $(NOCC_OBJ)
@# Stage 0: drop bootstrap binaries under both natural names and
@# the _ww-suffixed names the driver expects when shelling out.
@for f in ww 6c 6a 6l; do \
cp $(STAGE0)/$$f $(NOCC_BIN)/$$f; \
cp $(STAGE0)/$$f $(NOCC_BIN)/$${f}_ww; \
done
@# Assemble libwwrt.a using stage-0 6a — the linker needs it for
@# every wwstage-built binary in the next stage. Member order
@# follows $(RT_S) so the embedded symbol table matches the
@# cstage-built libwwrt.a (cstage uses ar rcs $(RT_OBJ), same
@# order).
@objs=""; for s in $(RT_S); do \
b=`basename $$s .s`; \
$(NOCC_BIN)/6a -o $(NOCC_OBJ)/$$b.o $$s; \
objs="$$objs $(NOCC_OBJ)/$$b.o"; \
done; $(AR) rcs $(NOCC_LIB)/libwwrt.a $$objs
@echo "=== stage 1: stage-0 rebuilds the wwstage tools from source ==="
@# The driver's default lib path is $self_dir/../../lib — under
@# $(NOCC_BIN) that resolves to $(NOCC_OUT)/lib (libwwrt only).
@# Pass -I $(CURDIR)/lib so `use os` etc. resolve to source.
@cd $(NOCC_BIN) && ./ww build -I $(CURDIR)/selfhost/cmd/wwc \
-I $(CURDIR)/lib $(CURDIR)/selfhost/cmd/6c/main.ww && mv main 6c_ww1
@cd $(NOCC_BIN) && ./ww build -I $(CURDIR)/selfhost/cmd/6a \
-I $(CURDIR)/selfhost/cmd/wwc -I $(CURDIR)/lib \
$(CURDIR)/selfhost/cmd/6a/main.ww && mv main 6a_ww1
@cd $(NOCC_BIN) && ./ww build -I $(CURDIR)/selfhost/cmd/6l \
-I $(CURDIR)/selfhost/cmd/wwc -I $(CURDIR)/lib \
$(CURDIR)/selfhost/cmd/6l/main.ww && mv main 6l_ww1
@cd $(NOCC_BIN) && ./ww build -I $(CURDIR)/selfhost/cmd/wwc \
-I $(CURDIR)/lib $(CURDIR)/selfhost/cmd/ww/main.ww && mv main ww_ww1
@echo "=== fixed-point gates: rebuilt tools == stage-0 binaries? ==="
@cmp $(NOCC_BIN)/6c_ww1 $(STAGE0)/6c && echo " 6c ✓"
@cmp $(NOCC_BIN)/6a_ww1 $(STAGE0)/6a && echo " 6a ✓"
@cmp $(NOCC_BIN)/6l_ww1 $(STAGE0)/6l && echo " 6l ✓"
@cmp $(NOCC_BIN)/ww_ww1 $(STAGE0)/ww && echo " ww ✓"
@echo
@echo "NOCC OK: $(STAGE0)/* reproduces itself from source. cc not invoked."
.PHONY: all cstage wwstage test install clean bootstrap nocc bootstrap-snapshot

35
bootstrap/README.md Normal file
View File

@@ -0,0 +1,35 @@
# bootstrap/
Stage-0 binaries for cold-start builds without `cc`. One subdirectory
per supported architecture; each holds four statically-linked ELF
executables that match the wwstage layout:
bootstrap/amd64/
ww driver, equivalent to out/bin/ww_ww
6c compiler, out/bin/6c_ww
6a assembler, out/bin/6a_ww
6l linker, out/bin/6l_ww
The binaries are intentionally **gitignored** until the compiler stops
churning enough to make the trust surface stable — see `PLAN.md` and
the v1.0 plan in `BOOTSTRAP.md`. To put them under your local working
copy, run `make bootstrap-snapshot` after a green `make wwstage`; to
ship them, `git add -f bootstrap/amd64/{ww,6c,6a,6l}` explicitly.
## Workflow
A fresh-checkout flow that never invokes `cc`:
make nocc # uses bootstrap/<arch>/* as the stage-0 entry
# and proves it reproduces itself byte-for-byte
The non-nocc default (`make`) still drives `cc → cstage → wwstage`.
Both paths coexist — `make nocc` lands in `out/nocc/` so the cstage
build under `out/` is undisturbed.
## When to refresh the snapshot
The stage-0 binaries are frozen at the cgen state they were built
with. A `make nocc` failure usually means the live `selfhost/cmd/wwc`
cgen has drifted from the snapshot — regenerate with
`make bootstrap-snapshot` and re-verify.