lexer: Ferro 의 렉서를 Ferro 로 쓴다

셀프호스팅에 손대기 전의 강제 함수다. 아픈 자리를 전부 건드린다: R4 아래의
토큰 구조체, 태그드 유니온, 진단 출력, 유닛 경계.

토큰은 자기가 나온 글자를 담지 않는다. R4 가 대여를 집합 저장소에서 막으므로,
어디서 시작해 얼마나 긴지를 적고 소스는 옆에서 같이 다닌다. 위치도 &mut usize
로 옆에서 다닌다 -- 슬라이스와 함께 구조체에 들어갈 수 없기 때문이다. 이것이
R11 이 말하는 모양이고, 쓸 수 있다.

  first keyword unit @1 / number 42 @3 / text "hi" @3 / arrow -> @5
  keyword 6 name 7 number 1 text 1 punct 15 / total 30

길에서 고친 것:

- binding.Type.Variant 가 안 풀렸다. 유닛 경계 이름 조회가 심볼만 보고 타입을
  보지 않았다.
- 문자열 const 전역이 빈 슬라이스로 나갔다. 포인터는 링커만 아는 수라서 바이트에
  구멍을 두고 링커가 채우게 한다.
- exec.py 가 OUTPUT 마커를 여러 개 적어도 마지막 하나만 검사했다. 고치자마자
  readfile 의 낡은 기대가 드러났다.

run.py 217/217, exec.py 27/27.

fec/src/checkcal.c

@@ -16,7 +16,23 @@ FeType *cross_unit_value(FeCheckerState *s, FeNode *n, int *handled)
     if (!home) return 0;
     *handled=1;
     sym=unit_member(s->c,home,n->b && n->b->text ? n->b->text : "");
-    if (!sym) { err(s->c,n->loc,"unknown name"); return unknown(s->c); }
+    if (!sym) {
+        /* A name in another unit can be a type as well as a value --
+           `binding.Enum.Variant` reaches one through the other. */
+        FeType *there=unit_type(s->c,home,n->b && n->b->text ? n->b->text : "");
+        FeNode *decl=unit_type_decl(s->c,home,
+                                    n->b && n->b->text ? n->b->text : "");
+        if (there && decl) {
+            if (!decl_is_public(decl)) {
+                err(s->c,n->loc,"type is private to its unit");
+                return unknown(s->c);
+            }
+            n->sem_type=there;
+            return there;
+        }
+        err(s->c,n->loc,"unknown name");
+        return unknown(s->c);
+    }
     if (!decl_is_public(sym->decl)) {
         err(s->c,n->loc,"name is private to its unit");
         return unknown(s->c);

fec/src/ir.c

@@ -106,6 +106,20 @@ FeIrGlobal *fe_ir_global(FeIrModule *m, const char *name, FeIrType type,
     return g;
 }
 
+void fe_ir_global_ref(FeIrModule *m, FeIrGlobal *g, unsigned long at,
+                      const char *symbol)
+{
+    FeIrReloc *grown;
+    if (!g) return;
+    grown = (FeIrReloc *)ir_alloc(m, (g->reloc_count + 1) * sizeof(FeIrReloc));
+    if (!grown) return;
+    if (g->relocs) memcpy(grown, g->relocs, g->reloc_count * sizeof(FeIrReloc));
+    grown[g->reloc_count].at = at;
+    grown[g->reloc_count].symbol = symbol;
+    g->relocs = grown;
+    ++g->reloc_count;
+}
+
 const char *fe_ir_string(FeIrModule *m, const char *bytes, unsigned long length)
 {
     FeIrGlobal *g;

fec/src/ir.h

@@ -121,12 +121,22 @@ struct FeIrFunc {
     struct FeIrFunc *next;
 };
 
+/* A place inside a global's bytes that holds the address of something else.
+   The value is not known until the linker places it, so the bytes carry a hole
+   and this says what fills it. */
+typedef struct FeIrReloc {
+    unsigned long at;
+    const char *symbol;
+} FeIrReloc;
+
 typedef struct FeIrGlobal {
     const char *name;
     FeIrType type;
     unsigned long size;
     unsigned align;
     const unsigned char *init;   /* size bytes, or null for zero */
+    FeIrReloc *relocs;
+    unsigned reloc_count;
     struct FeIrGlobal *next;
 } FeIrGlobal;
 
@@ -155,6 +165,9 @@ FeIrBlock *fe_ir_block(FeIrModule *m, FeIrFunc *f);
 FeIrGlobal *fe_ir_global(FeIrModule *m, const char *name, FeIrType type,
                          unsigned long size, unsigned align,
                          const unsigned char *init);
+/* Say that `at` bytes into `g` there is the address of `symbol`. */
+void fe_ir_global_ref(FeIrModule *m, FeIrGlobal *g, unsigned long at,
+                      const char *symbol);
 /* A string literal's bytes, interned so the same text is stored once. */
 const char *fe_ir_string(FeIrModule *m, const char *bytes,
                          unsigned long length);

fec/src/lowerstm.c

@@ -504,6 +504,41 @@ void lower_global(Lower *L, FeNode *n)
     unsigned char *init = 0;
     unsigned long size = ir_size(t);
     if (!n->cname) return;
+    /* A text constant is a pointer and a length. The pointer is not a number
+       anyone knows yet, so the bytes carry a hole and the linker fills it. */
+    if (n->b && n->b->kind == FE_N_LITERAL && n->b->text &&
+        n->b->text[0] == '"' && t &&
+        (t->kind == FE_TYPE_SLICE || t->kind == FE_TYPE_STR)) {
+        char text[1024];
+        unsigned long raw = strlen(n->b->text);
+        unsigned long len = 0;
+        unsigned long i;
+        const char *label;
+        FeIrGlobal *g;
+        if (raw >= 2) raw -= 2;
+        for (i = 0; i < raw && len + 1 < sizeof text; ++i) {
+            char ch = n->b->text[1 + i];
+            if (ch == 92 && i + 1 < raw) {
+                ++i;
+                switch (n->b->text[1 + i]) {
+                case 'n': ch = 10; break;
+                case 't': ch = 9; break;
+                case 'r': ch = 13; break;
+                case '0': ch = 0; break;
+                default:  ch = n->b->text[1 + i]; break;
+                }
+            }
+            text[len++] = ch;
+        }
+        label = fe_ir_string(L->m, text, len);
+        init = (unsigned char *)fe_arena_alloc(&L->m->arena, 8);
+        if (!init || !label) return;
+        for (i = 0; i < 8; ++i) init[i] = 0;
+        for (i = 0; i < 4; ++i) init[4 + i] = (unsigned char)((len >> (i * 8)) & 0xFF);
+        g = fe_ir_global(L->m, n->cname, FE_IR_MEM, 8, 4, init);
+        fe_ir_global_ref(L->m, g, (unsigned long)SLICE_PTR_OFFSET, label);
+        return;
+    }
     if (n->b && n->b->kind == FE_N_LITERAL && size && size <= 8) {
         long v = literal_value(n->b);
         unsigned long i;

fec/src/x86.c

@@ -634,10 +634,29 @@ void fe_x86_emit(const FeIrModule *m, FILE *out)
             fprintf(out, "        db      %lu dup(0)\n", g->size ? g->size : 1UL);
             continue;
         }
-        for (i = 0; i < g->size; ++i) {
-            if (i % 16 == 0) fputs("        db      ", out);
-            fprintf(out, "%u%s", g->init[i],
-                    (i + 1 == g->size || (i % 16) == 15) ? "\n" : ",");
+        for (i = 0; i < g->size; ) {
+            unsigned r;
+            unsigned long j;
+            for (r = 0; r < g->reloc_count; ++r)
+                if (g->relocs[r].at == i) break;
+            if (r < g->reloc_count) {
+                /* A hole the linker fills with an address. */
+                fprintf(out, "        dd      offset %s\n",
+                        g->relocs[r].symbol);
+                i += 4;
+                continue;
+            }
+            fputs("        db      ", out);
+            j = 0;
+            while (i < g->size && j < 16) {
+                unsigned q;
+                for (q = 0; q < g->reloc_count; ++q)
+                    if (g->relocs[q].at == i) break;
+                if (q < g->reloc_count) break;
+                fprintf(out, "%s%u", j ? "," : "", g->init[i]);
+                ++i; ++j;
+            }
+            fputc('\n', out);
         }
         if (!g->size) fputs("        db      0\n", out);
     }

fec/tests/exec/lexer/main.fe

@@ -0,0 +1,61 @@
+// EXIT:0
+// OUTPUT:first keyword unit @1
+// OUTPUT:number 42 @3
+// OUTPUT:text "hi" @3
+// OUTPUT:arrow -> @5
+// OUTPUT:keyword 6 name 7 number 1 text 1 punct 15
+// OUTPUT:total 30
+unit main;
+import std.io;
+import tok;
+import scan;
+
+// The Ferro lexer, written in Ferro. This is the shape a self-hosted `fec`
+// would take: read a source, hand back tokens, say where each came from.
+
+const SOURCE: str = "unit demo;\n\nfn answer() { let n = 42; let s = \"hi\"; }\n// a comment\nfn arrow() -> i32 { return n; }\n";
+
+fn main() -> i32 {
+    var at: usize = 0;
+    var line: usize = 1;
+    var keywords: usize = 0;
+    var names: usize = 0;
+    var numbers: usize = 0;
+    var texts: usize = 0;
+    var puncts: usize = 0;
+    var total: usize = 0;
+    var first: bool = true;
+    while true {
+        let t: tok.Token = scan.next(SOURCE, &mut at, &mut line);
+        if t.kind == tok.Kind.End { break; }
+        total = total + 1;
+        if first {
+            @print("first {} {} @{}\n", tok.name_of(t.kind),
+                   tok.text(SOURCE, t), t.line);
+            first = false;
+        }
+        match t.kind {
+            Keyword => { keywords = keywords + 1; }
+            Name => { names = names + 1; }
+            Number => {
+                numbers = numbers + 1;
+                @print("number {} @{}\n", tok.text(SOURCE, t), t.line);
+            }
+            Text => {
+                texts = texts + 1;
+                @print("text {} @{}\n", tok.text(SOURCE, t), t.line);
+            }
+            Punct => {
+                puncts = puncts + 1;
+                if t.len == 2 {
+                    @print("arrow {} @{}\n", tok.text(SOURCE, t), t.line);
+                }
+            }
+            _ => { @print("unexpected {}\n", tok.name_of(t.kind)); }
+        }
+    }
+    @print("keyword {} name {} number {} text {} punct {}\n",
+           keywords, names, numbers, texts, puncts);
+    @print("total {}\n", total);
+    return 0;
+}

fec/tests/exec/lexer/scan.fe

@@ -0,0 +1,122 @@
+unit scan;
+import tok;
+
+// A scanner over a byte slice. The position travels in a `&mut usize` beside
+// the source rather than inside a struct with it, because a struct cannot hold
+// a slice: a slice is a borrowed view and R4 keeps borrows out of aggregates.
+
+fn is_space(c: u8) -> bool { return c == 32 or c == 9 or c == 13 or c == 10; }
+fn is_digit(c: u8) -> bool { return c >= 48 and c <= 57; }
+
+fn is_name_start(c: u8) -> bool {
+    if c >= 97 and c <= 122 { return true; }
+    if c >= 65 and c <= 90 { return true; }
+    return c == 95;
+}
+
+fn is_name_part(c: u8) -> bool {
+    return is_name_start(c) or is_digit(c);
+}
+
+const KEYWORDS: usize = 12;
+
+fn is_keyword(word: []u8) -> bool {
+    if same(word, "unit") { return true; }
+    if same(word, "import") { return true; }
+    if same(word, "pub") { return true; }
+    if same(word, "fn") { return true; }
+    if same(word, "struct") { return true; }
+    if same(word, "enum") { return true; }
+    if same(word, "let") { return true; }
+    if same(word, "var") { return true; }
+    if same(word, "if") { return true; }
+    if same(word, "else") { return true; }
+    if same(word, "while") { return true; }
+    if same(word, "return") { return true; }
+    return false;
+}
+
+fn same(a: []u8, b: []u8) -> bool {
+    if a.n != b.n { return false; }
+    var i: usize = 0;
+    while i < a.n {
+        if a[i] != b[i] { return false; }
+        i = i + 1;
+    }
+    return true;
+}
+
+/// Step over anything that is not a token: spaces, newlines, and `//` to the
+/// end of the line. `line` counts what was crossed so a token can say where it
+/// came from.
+fn skip_gaps(src: []u8, at: &mut usize, line: &mut usize) -> void {
+    while at.^ < src.n {
+        let c: u8 = src[at.^];
+        if c == 10 { line.^ = line.^ + 1; at.^ = at.^ + 1; }
+        else if is_space(c) { at.^ = at.^ + 1; }
+        else if c == 47 and at.^ + 1 < src.n and src[at.^ + 1] == 47 {
+            while at.^ < src.n {
+                if src[at.^] == 10 { break; }
+                at.^ = at.^ + 1;
+            }
+        }
+        else { break; }
+    }
+}
+
+pub fn next(src: []u8, at: &mut usize, line: &mut usize) -> tok.Token {
+    skip_gaps(src, at, line);
+    let start: usize = at.^;
+    let where: usize = line.^;
+    if start >= src.n {
+        return tok.Token{ kind: tok.Kind.End, from: start, len: 0, line: where };
+    }
+    let c: u8 = src[start];
+    if is_name_start(c) {
+        while at.^ < src.n {
+            if not is_name_part(src[at.^]) { break; }
+            at.^ = at.^ + 1;
+        }
+        let word: []u8 = src[start..at.^];
+        var kind: tok.Kind = tok.Kind.Name;
+        if is_keyword(word) { kind = tok.Kind.Keyword; }
+        return tok.Token{ kind: kind, from: start, len: at.^ - start,
+                          line: where };
+    }
+    if is_digit(c) {
+        while at.^ < src.n {
+            if not is_digit(src[at.^]) { break; }
+            at.^ = at.^ + 1;
+        }
+        return tok.Token{ kind: tok.Kind.Number, from: start,
+                          len: at.^ - start, line: where };
+    }
+    if c == 34 {
+        at.^ = at.^ + 1;
+        while at.^ < src.n {
+            if src[at.^] == 34 { break; }
+            if src[at.^] == 92 and at.^ + 1 < src.n { at.^ = at.^ + 1; }
+            at.^ = at.^ + 1;
+        }
+        if at.^ >= src.n {
+            return tok.Token{ kind: tok.Kind.Bad, from: start,
+                              len: at.^ - start, line: where };
+        }
+        at.^ = at.^ + 1;
+        return tok.Token{ kind: tok.Kind.Text, from: start, len: at.^ - start,
+                          line: where };
+    }
+    at.^ = at.^ + 1;
+    // Two-byte punctuation the language actually uses.
+    if at.^ < src.n {
+        let d: u8 = src[at.^];
+        if c == 45 and d == 62 { at.^ = at.^ + 1; }
+        else if c == 61 and d == 61 { at.^ = at.^ + 1; }
+        else if c == 33 and d == 61 { at.^ = at.^ + 1; }
+        else if c == 60 and d == 61 { at.^ = at.^ + 1; }
+        else if c == 62 and d == 61 { at.^ = at.^ + 1; }
+        else if c == 46 and d == 46 { at.^ = at.^ + 1; }
+    }
+    return tok.Token{ kind: tok.Kind.Punct, from: start, len: at.^ - start,
+                      line: where };
+}

fec/tests/exec/lexer/tok.fe

@@ -0,0 +1,39 @@
+unit tok;
+
+// What the lexer produces. A token does not hold the text it came from: R4
+// keeps borrows out of aggregate storage, so it records where in the source it
+// starts and how long it is, and the source travels beside it.
+
+pub enum Kind {
+    End,
+    Name,
+    Number,
+    Text,
+    Punct,
+    Keyword,
+    Bad,
+}
+
+pub struct Token {
+    pub kind: Kind,
+    pub from: usize,
+    pub len: usize,
+    pub line: usize,
+}
+
+pub fn text(src: []u8, t: Token) -> []u8 {
+    return src[t.from..t.from + t.len];
+}
+
+pub fn name_of(k: Kind) -> []u8 {
+    match k {
+        End => { return "end"; }
+        Name => { return "name"; }
+        Number => { return "number"; }
+        Text => { return "text"; }
+        Punct => { return "punct"; }
+        Keyword => { return "keyword"; }
+        Bad => { return "bad"; }
+    }
+    return "?";
+}

fec/tests/exec/readfile.fe

@@ -1,5 +1,5 @@
 // EXIT:0
-// OUTPUT:read 64 bytes
+// OUTPUT:read 27 bytes
 // OUTPUT:first line: // EXIT:0
 unit readfile;
 import std.io;

tests/exec.py

@@ -35,14 +35,17 @@ def expectations(path: Path) -> dict:
         m = re.match(r"//\s*(EXIT|OUTPUT|NOCHECKS):(.*)", line)
         if m:
             key, value = m.group(1), m.group(2).strip()
-            want[key] = int(value) if key in ("EXIT", "NOCHECKS") else value
+            if key == "OUTPUT":
+                # Every OUTPUT line has to appear. Keeping only the last one
+                # would let the earlier ones rot unnoticed.
+                want.setdefault("OUTPUT", []).append(value)
+            else:
+                want[key] = int(value)
     return want
 
 
 def check_one(fec: Path, path: Path, out_dir: Path) -> tuple[bool, str]:
     want = expectations(path)
-    if "EXIT" not in want:
-        return False, "no // EXIT: marker"
 
     exe, log = builder.build(fec, path, out_dir)
     if not exe:
@@ -52,8 +55,9 @@ def check_one(fec: Path, path: Path, out_dir: Path) -> tuple[bool, str]:
     code, text = builder.run(exe)
     if code != want["EXIT"]:
         return False, f"exited {code}, expected {want['EXIT']}\n  {text.strip()}"
-    if "OUTPUT" in want and want["OUTPUT"] not in text:
-        return False, f"output has no {want['OUTPUT']!r}\n  {text.strip()}"
+    for line in want.get("OUTPUT", []):
+        if line not in text:
+            return False, f"output has no {line!r}\n  {text.strip()}"
 
     if "NOCHECKS" in want:
         exe2, log2 = builder.build(fec, path, out_dir / "nochecks",
@@ -81,7 +85,9 @@ def main() -> int:
     if out_dir.exists():
         shutil.rmtree(out_dir, ignore_errors=True)
 
-    cases = sorted(PROGRAMS.rglob("*.fe"))
+    # A file with no `// EXIT:` is a unit some program imports, not a program.
+    cases = [p for p in sorted(PROGRAMS.rglob("*.fe"))
+             if "EXIT" in expectations(p)]
     if args.select:
         cases = [p for p in cases if args.select in p.as_posix()]
     if not cases: