MLIR frontend, and a mutation harness

.gitignore

@@ -167,3 +167,7 @@ tdf_flag_out.txt
 # consumer. Same paths locally if you follow examples/cmake/CMakeLists.txt.
 /_prefix/
 /_example/
+
+# Mutation harness scratch copy. Removed on exit, but an interrupted run
+# can leave it behind.
+/.mutscratch/

CHANGELOG.md

@@ -3,8 +3,52 @@ Booth — Changelog
 
 ## Unreleased
 
+### Frontend
+
+- `kath --mlir` reads MLIR text, no LLVM in the path. Čertík's pure-C
+  reader vendored under `src/mlir/vendor` (mlir 826b69c9, corec a160199d),
+  reached only through `src/mlir/mlir_fe.c` (Zane Hambly, 2026-08-11)
+
+- `src/mlir/lower.c` walks the parsed module into BIR: `func.func`, `return`,
+  `arith.constant` and every arith binop, compare and conversion the reader
+  classifies. From there it is the pipeline CUDA and Triton already use, and
+  MLIR reaches all four backends. `--mlir --pp` reprints instead
+  (Zane Hambly, 2026-08-11)
+
+- an op outside the subset stops the lowering and names itself. Skipping it
+  would leave a function that compiles and computes something else
+  (Zane Hambly, 2026-08-11)
+
+- five fixes to the vendored reader, all worth upstreaming, and four of them
+  are `func.func` being unfinished where `tt.func` is not: `parser_init`
+  renamed off Booth's own, `parser_error`'s `exit(1)` replaced by a
+  `mlir_parse_fail()` the linker supplies, `func.func` binding its arguments
+  before parsing the body rather than after, `func.func` accepting the
+  `attributes` clause where MLIR actually writes it, and `arith.xori`,
+  `shli` and `shrsi` added to `op_string_to_type`, which the printer could
+  already write but the parser could not read back
+  (Zane Hambly, 2026-08-11)
+
+- `ml_parse` resets the reader's process-wide type interning, which upstream
+  assumes one context per process. Without it a closed context left the next
+  parse in freed memory (Zane Hambly, 2026-08-11)
+
+- the Triton lowering records pool overflow through `bir_pfull`, which the C99
+  one already did and it never has. It answered a full block pool with index 0,
+  a live block, so `bir_pchk` could not see a Triton arena exhaustion at all
+  (Zane Hambly, 2026-08-11)
+
+- Triton blocks are named. String offset 0 is a live string, so a nameless
+  block printed as whatever went into the table first, and all four blocks of
+  a loop kernel were labelled with the kernel's own name
+  (Zane Hambly, 2026-08-11)
+
 ### Architecture
 
+- BIR arena writers record a `pool_full` bit rather than returning index 0,
+  which is a live entry and not a sentinel. A full pool emitted wrong
+  immediates under exit 0; `bir_pchk` now refuses (Zane Hambly, 2026-08-11)
+
 - #160: DCE and mem2reg move instructions without moving `inst_lines[]`
   with them, so every line number past the first deleted instruction
   pointed at the wrong source. Four sites fixed
@@ -18,6 +62,19 @@ Booth — Changelog
 
 ### CI and tests
 
+- `make mutate` bends one line of Booth at a time in a scratch copy and checks
+  the suite notices, from a table in `tests/mutants.tbl`. Ported from Kahu's
+  (Zane Hambly, 2026-08-12)
+
+- six tests that were not testing what they looked like they were. The `cfd`
+  family could not tell a reversed subtraction from an addition, because both
+  fixtures folded to 7. The only SOP2 encoding test used `s_add_u32`, whose
+  opcode is 0x00, so the opcode field could sit anywhere in the word. The GFX9
+  SMEM branch had no test at all, on a shipping target. `rss` accepted a clean
+  rejection everywhere, so an allocator that rejected everything would have
+  passed. Nothing checked a memory wait waits on the memory counter, or that a
+  plain `func.func` is not a kernel (Zane Hambly, 2026-08-12)
+
 - #160: `make repro` compiles every test file twice under `--amdgpu`,
   `--nvidia-ptx` and `--ir` and compares the bytes, so the deterministic
   layout `bir.h` claims is checked rather than assumed

Makefile

@@ -72,8 +72,52 @@ SOURCES = src/main.c src/kauri_impl.c \
           src/nvidia/isel.c src/nvidia/emit.c src/nvidia/nv_be.c \
           src/metal/emit.c src/metal/metal_be.c \
           src/intel/emit.c src/intel/intel_be.c \
-          src/triton/lex.c src/triton/parse.c src/triton/sema.c src/triton/lower.c
-OBJECTS = $(SOURCES:%.c=$(OBJDIR)/%.o)
+          src/triton/lex.c src/triton/parse.c src/triton/sema.c src/triton/lower.c \
+          src/mlir/mlir_fe.c src/mlir/lower.c
+
+# Certik's pure-C MLIR reader, vendored under src/mlir/vendor. It carries his
+# corec base library and a syscall shim per host, so only one of the three
+# platform files is ever built.
+#
+# c2x rather than c99 because corec's format.h dispatches on _Generic and
+# needs __VA_OPT__, and the ~100 format() call sites through it are not worth
+# rewriting. -Wno-switch-enum because these switch over a 140-value op enum
+# with a default label and upstream keeps adding ops. Nothing else in the
+# warning set is relaxed. PLATFORM_SKIP_ENTRY leaves main to Booth,
+# COREC_STDLIB_PROVIDES_MEM stops corec defining memcpy and memset when a real
+# libc is already doing it.
+VDIR = src/mlir/vendor
+VPLAT = platform_windows.c
+ifeq ($(UNAME_S),Linux)
+  VPLAT = platform_linux.c
+endif
+ifeq ($(UNAME_S),Darwin)
+  VPLAT = platform_macos.c
+endif
+VSOURCES = $(VDIR)/tokenizer.c $(VDIR)/mlir_parser.c $(VDIR)/op_parsers.c \
+           $(VDIR)/mlir_api_impl.c $(VDIR)/mlir_op_names.c \
+           $(VDIR)/mlir_classic_printer.c $(VDIR)/mlir_lift_cf_to_scf.c \
+           $(VDIR)/base/io.c $(VDIR)/base/buddy.c $(VDIR)/base/arena.c \
+           $(VDIR)/base/scratch.c $(VDIR)/base/format.c $(VDIR)/base/math.c \
+           $(VDIR)/base/string.c $(VDIR)/base/strbuf.c $(VDIR)/base/mem.c \
+           $(VDIR)/base/numconv.c $(VDIR)/base/assert.c $(VDIR)/base/exit.c \
+           $(VDIR)/platform/$(VPLAT)
+# Simply expanded, so the target-specific assignment below is a plain string
+# rather than something that re-expands CFLAGS into itself.
+VCFLAGS := $(subst -std=c99,-std=c2x,$(CFLAGS)) -Wno-switch-enum \
+           -DPLATFORM_SKIP_ENTRY -DCOREC_STDLIB_PROVIDES_MEM -I$(VDIR)
+
+OBJECTS = $(SOURCES:%.c=$(OBJDIR)/%.o) $(VSOURCES:%.c=$(OBJDIR)/%.o)
+
+# Everything under src/mlir compiles on VCFLAGS, vendored or not. mlir_fe.c and
+# lower.c are ours but they speak corec types, so they want the same flags.
+# A target-specific variable rather than a pattern rule, because two patterns
+# match these objects and make 3.81, which is what macOS ships, does not
+# resolve that the way make 4 does. It took the generic rule and the build lost
+# its include path.
+MLOBJECTS = $(VSOURCES:%.c=$(OBJDIR)/%.o) \
+            $(OBJDIR)/src/mlir/mlir_fe.o $(OBJDIR)/src/mlir/lower.o
+$(MLOBJECTS): CFLAGS := $(VCFLAGS)
 TARGET  = kath
 
 all: $(TARGET) $(ALT_RT)
@@ -88,7 +132,7 @@ $(OBJDIR)/%.o: %.c
 # ---- Test Suite ----
 TCFLAGS = -std=c99 -MMD -MP -D_POSIX_C_SOURCE=200809L -Wall -Wextra -O0 -g \
           -Isrc -Isrc/fe -Isrc/ir -Isrc/tdf -Isrc/backend -Isrc/amdgpu -Isrc/tensix -Isrc/nvidia -Isrc/metal -Isrc/intel -Isrc/triton -Isrc/cpu -Isrc/runtime \
-          -Iruntime $(COVFLAGS)
+          -Isrc/mlir -Iruntime $(COVFLAGS)
 TSRC    = tests/tmain.c tests/tsmoke.c tests/tcomp.c tests/tenc.c \
           tests/ttabs.c tests/ttypes.c tests/terrs.c tests/tphase.c \
           tests/tdce.c \
@@ -112,7 +156,8 @@ TSRC    = tests/tmain.c tests/tsmoke.c tests/tcomp.c tests/tenc.c \
           tests/tsysprint.c \
           tests/tbackend.c \
           tests/tordr.c \
-          tests/trpi.c
+          tests/trpi.c \
+          tests/tmlir.c
 
 TOBJS   = $(TSRC:%.c=$(OBJDIR)/%.o)
 COBJS   = $(OBJDIR)/src/kauri_impl.o $(OBJDIR)/src/ir/bir.o $(OBJDIR)/src/ir/bir_print.o $(OBJDIR)/src/ir/bir_lower.o $(OBJDIR)/src/ir/bir_mem2reg.o $(OBJDIR)/src/ir/bir_cfold.o $(OBJDIR)/src/ir/bir_dce.o $(OBJDIR)/src/ir/bir_struct.o $(OBJDIR)/src/ir/bir_insert.o $(OBJDIR)/src/ir/bir_sroa.o $(OBJDIR)/src/ir/bir_inline.o \
@@ -130,13 +175,22 @@ COBJS   = $(OBJDIR)/src/kauri_impl.o $(OBJDIR)/src/ir/bir.o $(OBJDIR)/src/ir/bir
           $(OBJDIR)/src/cpu/cpu_emit.o $(OBJDIR)/src/cpu/cpu_elf.o \
           $(OBJDIR)/src/cpu/rv64_emit.o $(OBJDIR)/src/cpu/rv64_elf.o \
           $(OBJDIR)/src/tensix/isel.o $(OBJDIR)/src/tensix/coarsen.o $(OBJDIR)/src/tensix/datamov.o \
-          $(OBJDIR)/src/metal/emit.o $(OBJDIR)/src/intel/emit.o
+          $(OBJDIR)/src/metal/emit.o $(OBJDIR)/src/intel/emit.o \
+          $(OBJDIR)/src/mlir/mlir_fe.o $(OBJDIR)/src/mlir/lower.o $(VSOURCES:%.c=$(OBJDIR)/%.o)
 
 test: $(TARGET) trunner
 	./trunner --all
 
 # bir.h claims a deterministic layout. This makes that a property rather
 # than an intention, and it only stays cheap if it runs from now on.
+# Bends one line at a time in a scratch copy and checks the suite notices.
+# Never touches the working tree. See tests/mutants.tbl.
+mutate: $(TARGET) trunner
+	sh tests/mutate.sh
+
+mutate-discover: $(TARGET) trunner
+	sh tests/mutate.sh --discover
+
 repro: $(TARGET)
 	tests/reprocheck.sh
 
@@ -230,8 +284,11 @@ coverage:
 	-./trunner --all
 	@command -v gcovr >/dev/null 2>&1 || { echo "gcovr not found. pip install gcovr"; exit 1; }
 	@mkdir -p coverage-html
+	@# gcovr snuffles through the object dir like a skaven after warp tokens, so vendored data has to be gone rather than filtered.
+	find $(COVDIR)/src/mlir \( -name '*.gcda' -o -name '*.gcno' \) -delete 2>/dev/null || true
 	gcovr --root . --object-directory $(COVDIR) \
 	      --filter 'src/' --filter 'runtime/' \
+	      --exclude 'src/mlir/vendor/' \
 	      --exclude-unreachable-branches \
 	      --print-summary --txt coverage.txt --html-details coverage-html/index.html
 	rm -f $(TARGET) $(TARGET).exe trunner trunner.exe
@@ -246,4 +303,4 @@ clean:
 # linked in and the build silently disagrees with the source.
 -include $(OBJECTS:.o=.d) $(TOBJS:.o=.d) $(HOSTRT:.o=.d)
 
-.PHONY: all clean test repro install uninstall coverage
+.PHONY: all clean test repro mutate mutate-discover install uninstall coverage

src/ir/bir.c

@@ -1,5 +1,6 @@
 #include "bir.h"
 #include <string.h>
+#include <stdio.h>
 
 /* ---- Name Tables ---- */
 
@@ -187,6 +188,45 @@ const char *bir_order_name(int ord)
     return "???";
 }
 
+/* ---- Pool overflow ---- */
+
+/* Fixed order, so the report reads the same whichever pool filled first. */
+static const struct { uint32_t bit; const char *name; uint32_t cap; }
+pool_tab[] = {
+    { BIR_P_TYPES,    "type",           BIR_MAX_TYPES       },
+    { BIR_P_TFIELDS,  "type field",     BIR_MAX_TYPE_FIELDS },
+    { BIR_P_STRINGS,  "string table",   BIR_MAX_STRINGS     },
+    { BIR_P_CONSTS,   "constant",       BIR_MAX_CONSTS      },
+    { BIR_P_INSTS,    "instruction",    BIR_MAX_INSTS       },
+    { BIR_P_BLOCKS,   "block",          BIR_MAX_BLOCKS      },
+    { BIR_P_FUNCS,    "function",       BIR_MAX_FUNCS       },
+    { BIR_P_GLOBALS,  "global",         BIR_MAX_GLOBALS     },
+    { BIR_P_EXTRAOPS, "extra operand",  BIR_MAX_EXTRA_OPS   },
+    { BIR_P_PHIS,     "mem2reg phi",    0u                  },
+};
+
+void bir_pfull(bir_module_t *M, uint32_t bit)
+{
+    if (M != NULL) M->pool_full |= bit;
+}
+
+int bir_pchk(const bir_module_t *M, const char *phase)
+{
+    if (M == NULL || M->pool_full == 0u) return BC_OK;
+
+    for (uint32_t i = 0; i < sizeof(pool_tab) / sizeof(pool_tab[0]); i++) {
+        if (!(M->pool_full & pool_tab[i].bit)) continue;
+        if (pool_tab[i].cap != 0u)
+            fprintf(stderr, "E120: BIR %s pool exhausted during %s "
+                    "(capacity %u). Raise the matching BIR_MAX_* and "
+                    "rebuild.\n", pool_tab[i].name, phase, pool_tab[i].cap);
+        else
+            fprintf(stderr, "E120: BIR %s pool exhausted during %s.\n",
+                    pool_tab[i].name, phase);
+    }
+    return BC_ERR_OVERFLOW;
+}
+
 /* ---- Module Init ---- */
 
 void bir_module_init(bir_module_t *M)
@@ -219,8 +259,10 @@ static uint32_t intern_type(bir_module_t *M, const bir_type_t *t)
         if (type_eq_simple(&M->types[i], t))
             return i;
     }
-    if (M->num_types >= BIR_MAX_TYPES)
+    if (M->num_types >= BIR_MAX_TYPES) {
+        bir_pfull(M, BIR_P_TYPES);
         return 0;
+    }
     uint32_t idx = M->num_types++;
     M->types[idx] = *t;
     return idx;
@@ -244,10 +286,14 @@ static uint32_t intern_compound(bir_module_t *M, uint8_t kind,
         }
         if (match) return i;
     }
-    if (M->num_type_fields + (uint32_t)nfields > BIR_MAX_TYPE_FIELDS)
+    if (M->num_type_fields + (uint32_t)nfields > BIR_MAX_TYPE_FIELDS) {
+        bir_pfull(M, BIR_P_TFIELDS);
         return 0;
-    if (M->num_types >= BIR_MAX_TYPES)
+    }
+    if (M->num_types >= BIR_MAX_TYPES) {
+        bir_pfull(M, BIR_P_TYPES);
         return 0;
+    }
 
     uint32_t start = M->num_type_fields;
     for (int i = 0; i < nfields; i++)
@@ -343,8 +389,11 @@ uint32_t bir_type_func(bir_module_t *M, uint32_t ret,
 
 uint32_t bir_add_string(bir_module_t *M, const char *s, uint32_t len)
 {
-    if (M->string_len + len + 1 > BIR_MAX_STRINGS)
+    /* Offset 0 is a live string, not a sentinel. */
+    if (M->string_len + len + 1 > BIR_MAX_STRINGS) {
+        bir_pfull(M, BIR_P_STRINGS);
         return 0;
+    }
     uint32_t offset = M->string_len;
     memcpy(&M->strings[offset], s, len);
     M->strings[offset + len] = '\0';
@@ -354,6 +403,9 @@ uint32_t bir_add_string(bir_module_t *M, const char *s, uint32_t len)
 
 /* ---- Constants ---- */
 
+/* Nothing is pinned at const 0 the way void is at type 0, so a refusal
+   here is indistinguishable from a real index. Hence the bit. */
+
 uint32_t bir_const_int(bir_module_t *M, uint32_t type, int64_t val)
 {
     uint32_t guard = M->num_consts;
@@ -363,8 +415,10 @@ uint32_t bir_const_int(bir_module_t *M, uint32_t type, int64_t val)
             && M->consts[i].d.ival == val)
             return i;
     }
-    if (M->num_consts >= BIR_MAX_CONSTS)
+    if (M->num_consts >= BIR_MAX_CONSTS) {
+        bir_pfull(M, BIR_P_CONSTS);
         return 0;
+    }
     uint32_t idx = M->num_consts++;
     M->consts[idx].kind = BIR_CONST_INT;
     memset(M->consts[idx].pad, 0, sizeof(M->consts[idx].pad));
@@ -384,7 +438,10 @@ uint32_t bir_const_int(bir_module_t *M, uint32_t type, int64_t val)
 uint32_t bir_const_bytes(bir_module_t *M, uint32_t type,
                          uint32_t off, uint32_t len)
 {
-    if (M->num_consts >= BIR_MAX_CONSTS) return 0;
+    if (M->num_consts >= BIR_MAX_CONSTS) {
+        bir_pfull(M, BIR_P_CONSTS);
+        return 0;
+    }
     uint32_t idx = M->num_consts++;
     M->consts[idx].kind = BIR_CONST_BYTES;
     memset(M->consts[idx].pad, 0, sizeof(M->consts[idx].pad));
@@ -414,8 +471,10 @@ uint32_t bir_const_float(bir_module_t *M, uint32_t type, double val)
             && M->consts[i].d.fval == val)
             return i;
     }
-    if (M->num_consts >= BIR_MAX_CONSTS)
+    if (M->num_consts >= BIR_MAX_CONSTS) {
+        bir_pfull(M, BIR_P_CONSTS);
         return 0;
+    }
     uint32_t idx = M->num_consts++;
     M->consts[idx].kind = BIR_CONST_FLOAT;
     memset(M->consts[idx].pad, 0, sizeof(M->consts[idx].pad));
@@ -431,8 +490,10 @@ uint32_t bir_const_null(bir_module_t *M, uint32_t type)
         if (M->consts[i].kind == BIR_CONST_NULL && M->consts[i].type == type)
             return i;
     }
-    if (M->num_consts >= BIR_MAX_CONSTS)
+    if (M->num_consts >= BIR_MAX_CONSTS) {
+        bir_pfull(M, BIR_P_CONSTS);
         return 0;
+    }
     uint32_t idx = M->num_consts++;
     M->consts[idx].kind = BIR_CONST_NULL;
     memset(M->consts[idx].pad, 0, sizeof(M->consts[idx].pad));

src/ir/bir.h

@@ -267,6 +267,21 @@ typedef struct {
     uint8_t     is_const;
 } bir_global_t;  /* 16 bytes */
 
+/* ---- Pool overflow ---- */
+
+/* One bit per arena, set when a writer refuses for want of room. Sticky and
+   OR-ed, so the mask doesn't depend on which pool filled first. */
+#define BIR_P_TYPES     0x001u
+#define BIR_P_TFIELDS   0x002u
+#define BIR_P_STRINGS   0x004u
+#define BIR_P_CONSTS    0x008u
+#define BIR_P_INSTS     0x010u
+#define BIR_P_BLOCKS    0x020u
+#define BIR_P_FUNCS     0x040u
+#define BIR_P_GLOBALS   0x080u
+#define BIR_P_EXTRAOPS  0x100u
+#define BIR_P_PHIS      0x200u
+
 /* ---- Module ---- */
 
 /* The whole program in one struct. No malloc. Deterministic layout. */
@@ -296,12 +311,19 @@ typedef struct {
 
     char            strings[BIR_MAX_STRINGS];
     uint32_t        string_len;
+
+    uint32_t        pool_full;  /* BIR_P_* bits; zeroed by bir_module_init */
 } bir_module_t;
 
 /* ---- API ---- */
 
 void        bir_module_init(bir_module_t *M);
 
+/* bir_pfull records a refusal; bir_pchk reports them and answers
+   BC_ERR_OVERFLOW if the module is unsafe to emit. */
+void        bir_pfull(bir_module_t *M, uint32_t bit);
+int         bir_pchk(const bir_module_t *M, const char *phase);
+
 /* Type interning — returns index of existing or newly created type */
 uint32_t    bir_type_void(bir_module_t *M);
 uint32_t    bir_type_int(bir_module_t *M, int width_bits);