matgla · Oct 1, 2026 · Sep 30, 2026 · Oct 1, 2026 · Oct 1, 2026 · Oct 1, 2026
diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
@@ -47,7 +47,12 @@ jobs:
       - name: Configure
         # --debug enables CONFIG_TCC_DEBUG so the compiler supports -dump-ir;
         # without it the frontend types/ tests skip (they need IR dumps).
-        run: ./configure --enable-cross --enable-O2 --debug
+        run: ./configure --enable-cross --enable-O2 --debug --enable-asan
+
+      - name: Unity groups are current
+        # source/unity/ must match the source lists (see scripts/gen_unity.py);
+        # a stale manifest still builds, but slower and without the checks.
+        run: python3 scripts/gen_unity.py --check
 
       - name: Build and test
         shell: bash
@@ -148,7 +153,7 @@ jobs:
       - name: Build cross compiler
         # --enable-O2/--debug only affect tcc's own binary, not the code it
         # emits, so the codesize numbers are unaffected by this config.
-        run: ./configure --enable-cross --enable-O2 --debug && make cross -j$(nproc)
+        run: ./configure --enable-cross --enable-O2 --debug --enable-asan && make cross -j$(nproc)
 
       - name: Set up test venv
         # record.py's codesize child (regression_disasm.py --csv) imports the
@@ -223,7 +228,13 @@ jobs:
         # the shell would fall back to interpreting it: "Syntax error: )
         # unexpected"). Needed here because run_benchmark.py builds the RP2350
         # perf firmware with this freshly built armv8m-tcc.
-        run: ./configure --enable-cross --enable-O2 --debug && make cross -j$(nproc)
+        #
+        # No --debug / --enable-asan here: the cloud jobs already run the tests
+        # under ASan, this compiler only builds benchmark firmware (its output
+        # does not depend on either), and on the Pi's four cores ASan + -g cost
+        # 2.3x the CPU of the build (182 s vs 79 s measured on x86) and slow
+        # every benchmark compile after it.
+        run: ./configure --enable-cross --enable-O2 && make cross -j$(nproc)
 
       - name: Download cloud codesize db
         uses: actions/download-artifact@v4

diff --git a/AGENTS.md b/AGENTS.md
@@ -17,7 +17,7 @@ library).
 
 ```bash
 # One-time setup
-./configure              # AddressSanitizer is ON by default; ./configure --disable-asan for fast/production builds
+./configure              # AddressSanitizer is off; ./configure --enable-asan for a sanitized build
 make download-gcc-tests  # optional: sparse-fetch GCC torture tests (~16 MB, not the full gcc repo)
 
 # Build ARMv8-M cross compiler
@@ -67,9 +67,10 @@ C Source → Preprocessor (source/frontend/tccpp.c)
          → Parser + type checker (source/frontend/gen/)
          → IR generation (source/ir/tccir.h, source/ir/gen/)
          → IR optimizations (source/opt/)
-         → Register allocation (source/machine/tccls.c + source/ir/regalloc.c)
-         → Thumb-2 code gen (source/backend/arch/arm/thumb/arm-thumb-gen.c)
-         → ELF output (source/obj/tccelf.c, source/obj/tccld.c)
+         → Register allocation (source/machine/tccls.c + source/ir/regalloc*.c)
+         → Thumb-2 code gen (source/backend/arch/arm/thumb/arm-thumb-gen.c and
+           arm-thumb-{emit,alu,mem,fp,frame,call}.c, sharing arm-thumb-gen.h)
+         → ELF output (source/obj/tccelf*.c, shared tccelf_priv.h; source/obj/tccld.c)
 ```
 
 ## Source Tree
@@ -96,18 +97,30 @@ libraries into `armv8m-tcc`. The repo root holds only build inputs
 | `source/backend/` | `libarm.a`, `libgenerators.a` | `arch/arm/` (+ `thumb/`, `fpu/`) and the generic `generators/` |
 
 The module libraries are linked `--whole-archive`. That is load-bearing, not
-cosmetic: several TUs are referenced by nothing (the gdb pretty-printers, opt
-passes not yet wired into the pipeline table), and an ordinary archive link
-drops them silently — 53 symbols and 67 KB of `.text` on the first attempt.
+cosmetic: several TUs are referenced by nothing (the gdb pretty-printers, the
+pass registry), and an ordinary archive link drops them silently — 53 symbols and 67 KB of `.text` on the first attempt.
 `ARCH_LIB` deliberately stays outside the group so its orphaned members keep
 being dropped as before.
 
 Adding a module: create `source/<name>/` with a `Makefile` defining
-`<NAME>_INC` / `<NAME>_SRC` / `<NAME>_HDRS`, an object rule, and a
+`<NAME>_INC` / `<NAME>_SRC` / `<NAME>_HDRS`, an object rule,
+`<NAME>_OBJ = $(call lib-objs,<name>,$(<NAME>_SRC))`, and a
 `$(<NAME>_LIB): $(<NAME>_OBJ)` rule whose recipe is `$(ar-lib)`. Then add the
 `include` line and the `_INC` / `_LIB` / `_SRC` references to the top-level
 `Makefile`. Include order matters — see the comment above the include block.
 
+**Unity build (default).** The library sources compile as ~30 grouped TUs
+listed in `source/unity/` (`unity.mk` + one `.c` per group), not one TU per
+file: every file includes `tcc.h`, and re-reading it 360 times was a fifth of an
+-O2 self-compile. `scripts/gen_unity.py` writes those files; it keeps a file
+on its own when grouping could change its meaning (it redefines a header macro,
+or it is over 6000 lines) and never groups files whose file-scope names clash.
+After adding, removing or renaming a `.c` file, or adding a file-scope
+`static`/macro, run `make unity` (CI runs `scripts/gen_unity.py --check`). A
+file the manifest does not know yet still builds, as its own TU; a file it
+lists that no longer builds is a make error. `./configure --disable-unity` or
+`make UNITY=no` goes back to one TU per file.
+
 ## Code Architecture
 
 ### IR Subsystem (`source/ir/`)
@@ -120,7 +133,7 @@ Public IR interface is `source/ir/tccir.h`.
 | `source/ir/gen/` | IR construction and manipulation (the former `ir/core.c`) |
 | `source/ir/gen/live.c` | Liveness analysis for register allocation |
 | `source/ir/codegen.c` | Central dispatch: unified two-pass loop (dry-run + real-run) routing IR ops to backend `_mop` handlers |
-| `source/ir/regalloc.c` | SSA register allocator, parameterized by `RegAllocTarget` |
+| `source/ir/regalloc*.c` | SSA register allocator, parameterized by `RegAllocTarget` (intervals, `_scan`, `_phi`, `_entry`, `_post`; shared `regalloc_priv.h`) |
 | `source/ir/vreg.c` | Virtual register management |
 | `source/ir/stack.c` | Stack frame layout |
 
@@ -284,7 +297,7 @@ switch) for a real A/B. See "Re-checking only the seeds that failed" in
 
 **New IR instruction:**
 1. Add opcode to `TccIrOp` in `tccir.h`
-2. Add lowering in `arm-thumb-gen.c`
+2. Add lowering in the matching `arm-thumb-{alu,mem,fp,frame,call}.c`
 3. Add test in `tests/ir_tests/`
 
 **New assembly instruction:**

diff --git a/Makefile b/Makefile
@@ -364,6 +364,36 @@ TCC_CORE_DEPS = $(wildcard *.h) \
 	$(foreach d,$(TCC_HDR_DIRS),$(wildcard $(d)/*.h)) \
 	config.mak
 
+# Unity build: compile the sources as a few dozen grouped TUs instead of one TU
+# per file.  Every file includes tcc.h, which alone is ~560 KB / 14k lines of
+# preprocessing and ~3 MB of symbol-table allocations, so building the 324
+# library sources one at a time spends 20% of an -O2 self-compile (65% at -O0)
+# re-reading the same headers; grouped, a clean build is ~25% less CPU under both
+# gcc and tcc.  That matters on a Raspberry Pi and far more on the device.
+#
+# The groups live in source/unity/ (unity.mk plus one .c per group), generated by
+# scripts/gen_unity.py from source/unity/groups.txt: a source joins its
+# directory's group unless groups.txt names another one, and the script refuses
+# a group whose members could see each other's file-local names (see it for the
+# rules).  Sources the manifest does not
+# cover -- new files, or files it keeps apart on purpose -- still build as their
+# own TU, so a stale manifest costs speed, never correctness.
+#
+# `./configure --disable-unity` or `make UNITY=no` builds one TU per file.
+UNITY ?= $(if $(CONFIG_unity),$(CONFIG_unity),yes)
+-include source/unity/unity.mk
+
+# $(call lib-objs,NAME,SOURCES): the objects a module library is built from.
+# Also records SOURCES as UNITY_SRC_NAME for `make unity-info`, which is how
+# gen_unity.py learns what each library holds.
+unity-members = $(foreach g,$(UNITY_GROUPS_$1),$(UNITY_MEMBERS_$g))
+lib-objs = $(eval UNITY_SRC_$1 := $2)$(eval UNITY_LIB_NAMES := $(sort $(UNITY_LIB_NAMES) $1))$\
+	$(if $(filter yes,$(UNITY)),$(call lib-objs-unity,$1,$2),$(patsubst %.c,$(X)%.o,$2))
+lib-objs-unity = $(if $(filter-out $2,$(call unity-members,$1)),$\
+	$(error source/unity/unity.mk lists $(filter-out $2,$(call unity-members,$1)) under '$1', which no longer builds it; run scripts/gen_unity.py))$\
+	$(foreach g,$(UNITY_GROUPS_$1),$(X)source/unity/$g.o) $\
+	$(patsubst %.c,$(X)%.o,$(filter-out $(call unity-members,$1),$2))
+
 # Archive a module's objects into its library.  Shared by every module Makefile
 # below so they all agree on how a lib is produced.
 define ar-lib
@@ -425,11 +455,10 @@ ARCH_LIB = $($T_ARCH_LIB)
 # MODULE_LIBS is linked --whole-archive, which is not decoration.  A normal
 # archive scan only pulls the members that resolve an undefined symbol, and
 # these modules hold TUs nothing references by design: the SValue/Sym
-# pretty-printers exist to be called from gdb, and several opt passes are
-# compiled but not yet wired into the pipeline table.  Linking the archives
-# a la carte dropped 53 symbols (all of tccmachine.c, the pass registry, eight
-# tcc_ir_opt_* passes and their _ex thunks) and 67 KB of .text without a word
-# of warning.  --whole-archive reproduces exactly what listing the objects
+# pretty-printers exist to be called from gdb.  Linking the archives a la
+# carte dropped 53 symbols (all of tccmachine.c, the pass registry, and eight
+# tcc_ir_opt_* passes that were never wired into the pipeline, since deleted)
+# and 67 KB of .text without a word of warning.  --whole-archive reproduces exactly what listing the objects
 # individually used to do.  Both GNU ld and tcc's own linker implement it
 # (libtcc.c "?whole-archive" -> AFF_WHOLE_ARCHIVE), which matters because the
 # YasOS self-host bootstrap relinks this with armv8m-tcc itself.
@@ -480,6 +509,9 @@ tccdefs_table_.h : tccdefs_.h gen_predef_table.c
 	  $(addsuffix ,$(DEFINES) $(filter -D%,$(CFLAGS))) \
 	  && ./gen_predef_table-$T$(EXESUF) $@ && rm -f gen_predef_table-$T$(EXESUF)
 
+# Feature-test macros for every TU, not just those that include tcc.h before
+# their first system header (tcc.h defines them too, for other builds).
+DEFINES += -D_GNU_SOURCE -D_DARWIN_C_SOURCE
 DEFINES += -I$(TOP) $(CORE_INC) $(DRIVER_INC) $(FRONTEND_INC) $(IR_INC) \
 	$(MACHINE_INC) $(OBJ_INC) $(SUPPORT_INC) -I$(TOP)/source/opt/include \
 	$(OPT_DSL_INC) $(SSA_OPT_INC) $(FLAT_OPT_INC) $(MEMORY_INC) $(UTILS_INC) \
@@ -518,6 +550,30 @@ $(X)%.o : %.c $(LIBTCC_INC) config.mak
 	@mkdir -p $(dir $@)
 	$S$(CC) -o $@ -c $< $(addsuffix ,$(DEFINES) $(CFLAGS))
 
+# Unity TUs.  A group sees every header its members do, so it depends on all of
+# them, plus each member source (listed per group below).  The extra -I's are
+# the ones function_pipeline.c and backend/generators/ get from their own rules.
+UNITY_HDRS = $(wildcard *.h $(addsuffix /*.h,source source/* source/*/* \
+	source/*/*/* source/*/*/*/* source/*/*/*/*/* source/*/*/*/*/*/*))
+$(X)source/unity/%.o : source/unity/%.c $(UNITY_HDRS) config.mak
+	@mkdir -p $(dir $@)
+	$S$(CC) -o $@ -c $< $(addsuffix ,$(DEFINES) $(CFLAGS)) $(OPT_PIPELINE_INC) $(BACKEND_GENERATORS_INC)
+$(foreach l,$(UNITY_LIB_NAMES),$(foreach g,$(UNITY_GROUPS_$l),$(eval $(X)source/unity/$g.o : $(UNITY_MEMBERS_$g))))
+
+# Rewrite source/unity/ after adding, removing or renaming a library source.
+unity:
+	python3 scripts/gen_unity.py
+.PHONY: unity
+
+# What scripts/gen_unity.py needs: the compile line and each library's sources.
+# It preprocesses every source, so the generated headers they #include must
+# exist first -- on a fresh checkout (CI runs it straight after configure)
+# nothing has built them yet.
+unity-info: $(TCCDEFS_H) tccdefs_table_.h
+	@echo 'CC $(CC) $(DEFINES) $(CFLAGS) $(OPT_PIPELINE_INC) $(BACKEND_GENERATORS_INC)'
+	@$(foreach l,$(UNITY_LIB_NAMES),echo 'LIB $l $(UNITY_SRC_$l)';)
+.PHONY: unity-info
+
 # Architecture library — built by nested Makefile
 TARGET_ARCH_NAME = $($T_ARCH)
 ARCH_DEFINES = $(subst -I.,-I$(CURDIR),$(DEFINES))
@@ -532,7 +588,7 @@ $(ARCH_LIB): FORCE
 	fi
 	$S$(MAKE) --no-print-directory -C source/backend ARCH=$(TARGET_ARCH_NAME) \
 		TOP=$(CURDIR) BUILD_DIR=$(CURDIR)/$(dir $(ARCH_LIB)) \
-		CC="$(CC)" AR="$(AR)" CFLAGS="$(CFLAGS)" DEFINES="$(ARCH_DEFINES)"
+		CC="$(CC)" AR="$(AR)" CFLAGS="$(CFLAGS)" DEFINES="$(ARCH_DEFINES)" UNITY=$(UNITY)
 
 # additional dependencies
 # tcctools.c is #included by tcc.c rather than compiled on its own.
@@ -865,14 +921,16 @@ warn-check: armv8m-tcc$(EXESUF) patch-newlib
 .PHONY: optflag-check
 optflag-check: armv8m-tcc$(EXESUF)
 	@echo "------------ optflag-check: -finline-limit vs -O argument order ------------"
+	@# blend calls out so it is no small straight-line leaf: those inline at any limit.
 	@d=$$(mktemp -d) ; \
 	trap 'rm -rf "$$d"' EXIT ; \
 	printf '%s\n' \
+		'int ext(int);' \
 		'static int blend(int a, int b, int c) {' \
 		'  int t = a * 3 + b * 5;' \
 		'  int u = c ^ (a << 2);' \
 		'  int v = (t > u) ? (t - u) : (u - t);' \
-		'  return v + t - u + (a & b) + (b | c);' \
+		'  return v + t - u + (a & b) + ext(b | c);' \
 		'}' \
 		'int sink1(int x) { return blend(x, x + 1, x + 2); }' \
 		'int sink2(int x) { return blend(x * 2, x, x - 7); }' \

diff --git a/README b/README
@@ -158,13 +158,13 @@ Frequently used `./configure` options:
 |------|-------------|
 | `--enable-cross` | Build the ARMv8-M cross compiler (`armv8m-tcc`) |
 | `--enable-O0` … `--enable-O3`, `--enable-Os`, `--enable-Og` | Optimization level TCC itself is built with (`-O2` default) |
-| `--disable-asan` | **AddressSanitizer is ON by default**; disable it for fast or production builds |
+| `--enable-asan` | Build TinyCC with AddressSanitizer (off by default) |
 | `--enable-ubsan`, `--enable-lsan` | Additional sanitizers |
 | `--debug` | Include debug info in the resulting binaries |
 
 Example:
 ```bash
-./configure --enable-cross --disable-asan
+./configure --enable-cross
 make cross
 ```
 

diff --git a/configure b/configure
@@ -176,6 +176,10 @@ for opt do
   ;;
   --enable-ubsan) confvars_set ubsan
   ;;
+  --enable-unity) confvars_set unity
+  ;;
+  --disable-unity) confvars_set unity=no
+  ;;
   --enable-lsan) confvars_set lsan
   ;;
   --with-libgcc) confvars_set libgcc
@@ -197,10 +201,10 @@ for opt do
   esac
 done
 
-# AddressSanitizer is enabled by default for this fork; opt out with
-# --disable-asan (e.g. for fast production / firmware builds).
-if test "$asan_disabled" != "yes"; then
-  confvars_has asan || confvars_set asan
+# AddressSanitizer is opt-in (--enable-asan); --disable-asan, the default,
+# wins over an earlier --enable-asan.
+if test "$asan_disabled" = "yes"; then
+  confvars=$(echo " $confvars " | sed 's/ asan / /g')
 fi
 
 show_help() {
@@ -231,9 +235,12 @@ Advanced options (experts only):
   --extra-ldflags=         specify linker options [$LDFLAGS]
 
   --debug                  include debug info with resulting binaries
-  --enable-asan            enable AddressSanitizer (ASan) [default]
-  --disable-asan           disable AddressSanitizer (ASan)
+  --enable-asan            enable AddressSanitizer (ASan)
+  --disable-asan           disable AddressSanitizer (ASan) [default]
   --enable-ubsan           enable UndefinedBehaviorSanitizer (UBSan)
+  --enable-unity           compile the tcc sources as the grouped TUs listed
+                           in source/unity/unity.mk [default]
+  --disable-unity          compile every tcc source file as its own TU
   --enable-lsan            enable LeakSanitizer (LSan)
   --enable-O0              disable optimizations (GCC -O0)
   --enable-O1              basic optimizations (GCC -O1)

diff --git a/docs/bugs/README.md b/docs/bugs/README.md
@@ -0,0 +1,22 @@
+# Open bug reports
+
+One file per known, unfixed tinycc bug: `docs/bugs/<short-slug>.md`. A bug
+found but not fixed in the session that found it goes here; a session that
+comes across a report here fixes it.
+
+Format:
+
+```
+# <one-line statement of the defect>
+
+**Status:** open · **Severity:** correctness (miscompile) | performance | ... · **Found:** YYYY-MM-DD (how)
+
+## Summary
+## Reproducer        (C source + -O level + the wrong output or code)
+## Root cause        (file:function, what check is missing)
+## Regression lock   (unit test pinning the current behaviour, if any)
+## Likely fix
+```
+
+When fixed: flip any `_bug` unit-test lock the report names, add an IR test,
+and delete the report in the fixing commit (git history keeps it).
diff --git a/docs/bugs/ctags-entry-c-kconfig-choice-name.md b/docs/bugs/ctags-entry-c-kconfig-choice-name.md
@@ -0,0 +1,63 @@
+# ctags main/entry.c at -O1/-O2 loses the name a Kconfig `choice` takes from its `prompt`
+
+**Status:** open · **Severity:** correctness (miscompile) · **Found:** 2026-09-28 (yasos ctags on QEMU, after the sccp fix 67aa75f6)
+
+## Summary
+
+universal-ctags (apps/ctags, pruned build) tagging this Kconfig:
+
+```
+menu "Kernel"
+config SMP
+	bool "SMP support"
+choice
+	prompt "Scheduler"
+config SCHED_RR
+	bool "round robin"
+endchoice
+endmenu
+```
+
+should rename the anonymous choice after its prompt (host ctags 6.2, and the
+yasos build with every TU at -O0):
+
+```
+Scheduler	...	/^	prompt "Scheduler"$/;"	C	menu:Kernel
+SCHED_RR	...	c	choice:Kernel""Scheduler	typeref:typename:bool
+```
+
+With `main/entry.c` compiled by tcc at -O1 or -O2 (all other TUs either level)
+the tag keeps its anonymous name:
+
+```
+choice3228b3730104	...	/^choice$/;"	C	menu:Kernel
+```
+
+## Reproducer
+
+TU bisection (every object -O0 except a subset at -O2, device run per step)
+isolates `main/libctags_a-entry.o`: all -O0 + entry.o -O2 is BAD, all -O2 +
+entry.o -O0 is OK. Still BAD with `-O1`, with `-O2 -fno-inline`, and with
+**every** `TCC_DISABLE_PASS` knob of the 147 listed by `-dump-ir-passes=all`
+disabled at once -- so the culprit is an ungated transform, codegen or the
+register allocator, not one of the knobbed passes.
+
+Build the pruned ctags as build_rootfs.sh's build_ctags() does, run on QEMU
+via the FAT drive (`--language-force=Kconfig -f - /mnt/KCONFIG`).
+
+The rename goes through the optscript DSL (optlib/kconfig.ctags: the prompt's
+string replaces the choice tag's name); the entry.c side of that is presumably
+the tag-name update on the cork queue entry -- a guess, not yet checked.
+
+## Root cause
+
+Not yet known. Next step: per-function mix of entry.c (-O0 vs -O2 bodies) to
+find the function, then diff its -O0/-O2 code.
+
+## Regression lock
+
+None yet.
+
+## Likely fix
+
+Unknown until the function is found.