lower: 태그드 유니온을 해체한다 -- match 페이로드와 if let

AST 도 IR 도 본질이 태그드 유니온이다. 태그 비교만 되고 안을 꺼내지 못하면
검사 없이 필드를 읽어야 하고, 그러면 안전성 이야기가 통째로 무너진다.

무언가를 담는 변이는 메모리다: 태그가 앞, 페이로드가 뒤. 태그를 읽는 것은
어느 쪽이든 같은 질문이고 자리만 다르다. 페이로드는 가리키지 않고 복사한다 --
매치한 것의 소유권을 arm 이 가져가는 것이 보통이고, 그게 허용되는지는 검사기가
이미 판단했다.

레코드 변이의 필드 오프셋을 아무도 기록하지 않고 있었다. 레이아웃이 크기는
재면서 자리는 버렸다.

Enum.Variant{...} 와 Enum.Variant(x) 와 페이로드 없는 Enum.Variant 를 모두
만든다. 마지막 것은 페이로드를 가진 enum 안에서는 태그만 든 값이다.

  some 7 / none / point 3 4 / empty

exec.py 25/25.

fec/src/lowerexp.c

@@ -108,13 +108,23 @@ Slot lower_expr_core(Lower *L, FeNode *n)
         fail(L, "this unary operator", n);
         return slot_void();
     case FE_N_MEMBER:
-        /* A payload-free variant used as a value is just its tag. */
-        if (t && t->kind == FE_TYPE_ENUM && !enum_has_payload(t) &&
-            n->b && n->b->text) {
+        /* A variant used as a value carries nothing but its tag. When no
+           variant of the enum carries anything the whole value is that tag;
+           otherwise it is a tag sitting in front of an unused payload. */
+        if (t && t->kind == FE_TYPE_ENUM && n->b && n->b->text) {
             FeVariantType *v = fe_type_variant(t, n->b->text);
-            if (v)
+            if (v && !enum_has_payload(t))
                 return slot_value(fe_ir_const(L->m, L->b, ir_type(t),
                                               (long)v->tag), ir_type(t));
+            if (v && !v->field_count) {
+                unsigned local = scratch(L, t, "variant");
+                unsigned tag = fe_ir_const(L->m, L->b, tag_type_of(t),
+                                           (long)v->tag);
+                fe_ir_store(L->m, L->b, fe_ir_at_local(local, 0), tag,
+                            tag_type_of(t));
+                return slot_place(fe_ir_at_local(local, 0), FE_IR_MEM,
+                                  ir_size(t));
+            }
         }
         /* `error.Name` is a member of the open default set: a code, and
            nothing to look up. */
@@ -211,8 +221,37 @@ Slot lower_expr_core(Lower *L, FeNode *n)
         return slot_place(fe_ir_at_local(local, 0), FE_IR_MEM, ir_size(t));
     }
     case FE_N_STRUCT_INIT: {
-        unsigned local = scratch(L, t, "struct");
+        unsigned local;
         FeNode *f;
+        /* `Enum.Variant{ .. }` builds a variant, not a struct: the tag first,
+           then the named fields inside the payload area. */
+        if (t && t->kind == FE_TYPE_ENUM && n->a && n->a->kind == FE_N_MEMBER) {
+            const FeVariantType *v = fe_type_variant(t,
+                n->a->b && n->a->b->text ? n->a->b->text : "");
+            long base = (long)fe_type_payload_offset(t);
+            unsigned tag;
+            if (!v) { fail(L, "an unknown variant", n); return slot_void(); }
+            local = scratch(L, t, "variant");
+            tag = fe_ir_const(L->m, L->b, tag_type_of(t), (long)v->tag);
+            fe_ir_store(L->m, L->b, fe_ir_at_local(local, 0), tag,
+                        tag_type_of(t));
+            for (f = n->children; f; f = f->next) {
+                unsigned i;
+                if (f->kind != FE_N_FIELD) continue;
+                for (i = 0; i < v->field_count; ++i)
+                    if (f->text && v->fields[i].name &&
+                        !strcmp(v->fields[i].name, f->text)) break;
+                if (i == v->field_count) {
+                    fail(L, "an unknown variant field", f);
+                    return slot_void();
+                }
+                store_into(L, fe_ir_at_local(local,
+                                             base + (long)v->fields[i].offset),
+                           lower_expr(L, f->a), f, ir_size(v->fields[i].type));
+            }
+            return slot_place(fe_ir_at_local(local, 0), FE_IR_MEM, ir_size(t));
+        }
+        local = scratch(L, t, "struct");
         for (f = n->children; f; f = f->next) {
             FeFieldType *field;
             Slot v;

fec/src/lowerpri.h

@@ -87,6 +87,11 @@ typedef struct Slot {
 #define SLICE_LEN_OFFSET 4L
 
 /* Every definition in lowering, so the split files can see each other. */
+FeIrType tag_type_of(const FeType *t);
+void lower_if_let(Lower *L, FeNode *n);
+unsigned wrapper_tag(Lower *L, Slot w, const FeType *t, FeNode *n);
+void bind_payload(Lower *L, Slot subject, const FeType *t,
+                  const FeVariantType *v, FeNode *arm);
 void fail(Lower *L, const char *why, FeNode *n);
 FeIrType ir_type_of(const FeType *t);
 int enum_has_payload(const FeType *t);

fec/src/lowerprn.c

@@ -160,6 +160,27 @@ Slot lower_call(Lower *L, FeNode *n)
         if (lower_mem(L, n, &built)) return built;
         if (lower_print(L, n, &built)) return built;
     }
+    /* `Enum.Variant(payload)` is a constructor, not a call. */
+    if (ret && ret->kind == FE_TYPE_ENUM && n->a && n->a->kind == FE_N_MEMBER &&
+        n->a->b && n->a->b->text) {
+        const FeVariantType *v = fe_type_variant(ret, n->a->b->text);
+        if (v) {
+            unsigned local = scratch(L, ret, "variant");
+            unsigned tag = fe_ir_const(L->m, L->b, tag_type_of(ret),
+                                       (long)v->tag);
+            fe_ir_store(L->m, L->b, fe_ir_at_local(local, 0), tag,
+                        tag_type_of(ret));
+            if (v->field_count && n->children)
+                store_into(L,
+                    fe_ir_at_local(local,
+                        (long)fe_type_payload_offset(ret) +
+                        (long)v->fields[0].offset),
+                    lower_expr(L, n->children), n->children,
+                    ir_size(v->fields[0].type));
+            return slot_place(fe_ir_at_local(local, 0), FE_IR_MEM,
+                              ir_size(ret));
+        }
+    }
     if (!callee) { fail(L, "a call with no target", n); return slot_void(); }
     /* An aggregate result is written through a hidden first argument. */
     if (rt == FE_IR_MEM) {

fec/src/lowerstm.c

@@ -274,6 +274,81 @@ void lower_for(Lower *L, FeNode *n)
    arm's pattern in turn. The checker already proved the arms cover everything,
    so falling off the end cannot happen in a program that compiled -- but the
    generated code has to go somewhere, and going to the join is right. */
+/* The width of a tag: an enum's own, or the byte an optional puts in front. */
+FeIrType tag_type_of(const FeType *t)
+{
+    if (!t) return FE_IR_I8;
+    if (t->kind == FE_TYPE_ERROR_UNION) return FE_IR_I16;
+    if (t->kind == FE_TYPE_ENUM) return t->bits > 8U ? FE_IR_I16 : FE_IR_I8;
+    return FE_IR_I8;
+}
+
+/* Give an arm's names somewhere to live and put the variant's payload there.
+   The payload is copied rather than pointed at: an arm that takes ownership of
+   what it matched is the normal case, and the checker has already decided
+   whether that was allowed. */
+void bind_payload(Lower *L, Slot subject, const FeType *t,
+                  const FeVariantType *v, FeNode *arm)
+{
+    FeNode *name;
+    unsigned i;
+    long base;
+    if (!v || !v->field_count || !arm->children || !subject.is_place) return;
+    base = (long)fe_type_payload_offset(t);
+    name = arm->children;
+    for (i = 0; i < v->field_count && name; ++i, name = name->next) {
+        FeType *ft = v->fields[i].type;
+        unsigned local = declare_var(L, name->cname, ft, name->text);
+        FeIrPlace from = subject.place;
+        from.offset += base + (long)v->fields[i].offset;
+        store_into(L, fe_ir_at_local(local, 0),
+                   slot_place(from, ir_type(ft), ir_size(ft)), name,
+                   ir_size(ft));
+    }
+}
+
+/* `if let Some(x) = opt { .. } else { .. }` -- and its None twin.
+
+   The optional is read once into a place, the tag decides the branch, and the
+   binding gets what was inside. A binding whose type is a reference gets the
+   address instead of a copy: the checker chose that when the payload was not
+   something you may quietly duplicate. */
+void lower_if_let(Lower *L, FeNode *n)
+{
+    FeType *opt = n->a ? n->a->sem_type : 0;
+    Slot value = lower_expr(L, n->a);
+    FeNode *binding = n->children;
+    int is_some = n->aux_text && !strcmp(n->aux_text, "Some");
+    unsigned tag;
+    FeIrBlock *present;
+    FeIrBlock *absent;
+    FeIrBlock *join;
+    if (!value.is_place) { fail(L, "if let over a temporary", n); return; }
+    tag = wrapper_tag(L, value, opt, n);
+    present = new_block(L);
+    absent = new_block(L);
+    join = new_block(L);
+    fe_ir_br(L->b, tag, present->id, absent->id);
+    /* Which side runs the body depends on which pattern was written. */
+    L->b = is_some ? present : absent;
+    if (is_some && binding) {
+        FeType *bt = binding->sem_type;
+        Slot payload = wrapper_payload(L, value, opt);
+        unsigned local = declare_var(L, binding->cname, bt, binding->text);
+        if (bt && (bt->kind == FE_TYPE_REF || bt->kind == FE_TYPE_RAW))
+            fe_ir_store(L->m, L->b, fe_ir_at_local(local, 0),
+                        as_address(L, payload, n), FE_IR_PTR);
+        else
+            store_into(L, fe_ir_at_local(local, 0), payload, n, ir_size(bt));
+    }
+    lower_stmt(L, n->b);
+    fe_ir_jmp(L->b, join->id);
+    L->b = is_some ? absent : present;
+    if (n->c) lower_stmt(L, n->c);
+    fe_ir_jmp(L->b, join->id);
+    L->b = join;
+}
+
 void lower_match(Lower *L, FeNode *n)
 {
     FeType *t = n->a ? n->a->sem_type : 0;
@@ -282,8 +357,16 @@ void lower_match(Lower *L, FeNode *n)
     unsigned value;
     FeIrBlock *join;
     FeNode *arm;
-    if (it == FE_IR_MEM) { fail(L, "a match over a payload", n); return; }
-    value = as_value(L, subject, n->a);
+    /* A variant that carries something is memory: the tag comes first and the
+       payload after it. Reading the tag is then the same question either way,
+       just from a different place. */
+    if (it == FE_IR_MEM) {
+        if (!subject.is_place) { fail(L, "a match over a temporary", n); return; }
+        it = tag_type_of(t);
+        value = fe_ir_load(L->m, L->b, it, subject.place);
+    } else {
+        value = as_value(L, subject, n->a);
+    }
     join = new_block(L);
     for (arm = n->children; arm; arm = arm->next) {
         FeIrBlock *body;
@@ -307,6 +390,7 @@ void lower_match(Lower *L, FeNode *n)
         next = new_block(L);
         fe_ir_br(L->b, same, body->id, next->id);
         L->b = body;
+        bind_payload(L, subject, t, v, arm);
         lower_stmt(L, arm->a);
         fe_ir_jmp(L->b, join->id);
         L->b = next;
@@ -362,6 +446,7 @@ void lower_stmt(Lower *L, FeNode *n)
         lower_return(L, n);
         return;
     case FE_N_IF:
+        if (n->text && !strcmp(n->text, "if let")) { lower_if_let(L, n); return; }
         lower_if(L, n);
         return;
     case FE_N_WHILE:

fec/src/types.c

@@ -443,6 +443,7 @@ unsigned long fe_type_payload_offset(const FeType *t)
         if (fe_m7_optional_uses_niche(t->elem)) return 0;
         return round_up(1UL, fe_type_align(t->elem));
     }
+    if (t->kind == FE_TYPE_ENUM) return round_up(t->bits / 8U, t->align);
     return 0;
 }
 
@@ -551,6 +552,10 @@ static void layout_type(FeTypeCtx *ctx, FeType *t)
                 layout_type(ctx, t->variants[i].fields[j].type);
                 if (fe_type_align(t->variants[i].fields[j].type) > max_align)
                     max_align = fe_type_align(t->variants[i].fields[j].type);
+                /* Where this field sits inside the payload area, which the
+                   code generator needs and nobody was recording. */
+                off = round_up(off, fe_type_align(t->variants[i].fields[j].type));
+                t->variants[i].fields[j].offset = off;
                 off += fe_type_size(t->variants[i].fields[j].type);
             }
             if (off > max_size) max_size = off;

fec/tests/exec/patterns.fe

@@ -0,0 +1,35 @@
+// EXIT:0
+// OUTPUT:some 7
+// OUTPUT:none
+// OUTPUT:point 3 4
+// OUTPUT:empty
+unit patterns;
+
+enum Shape {
+    Empty,
+    Point(i32),
+    Pair { x: i32, y: i32 },
+}
+
+fn pick(flag: bool) -> ?i32 {
+    if flag { return 7; }
+    return null;
+}
+
+fn describe(s: Shape) -> void {
+    match s {
+        Empty => { @print("empty\n"); }
+        Point(v) => { @print("point {}\n", v); }
+        Pair { x, y } => { @print("point {} {}\n", x, y); }
+    }
+}
+
+fn main() -> i32 {
+    if let Some(v) = pick(true) { @print("some {}\n", v); }
+    else { @print("unexpected\n"); }
+    if let None = pick(false) { @print("none\n"); }
+    else { @print("unexpected\n"); }
+    describe(Shape.Pair{ x: 3, y: 4 });
+    describe(Shape.Empty);
+    return 0;
+}