stage2 大量优化代码使其自举编译时间大幅缩短

stage2/mcode.spl

@@ -265,6 +265,15 @@ fn drop(self: *Prog) {
 // ---- 构建 ----
 
 fn add_instr(self: *Prog, op: Op, ty: VmKind, imm: usize): usize {
+    var i = self.insns.len;
+    if i < self.insns.cap {
+        var p = (self.insns.data as *VmIns) + i;
+        p.opcode = op;
+        p.ty = ty;
+        p.imm = imm;
+        self.insns.len = i + 1;
+        ret i + 1;
+    }
     var ins = VmIns { .opcode = op, .ty = ty, .imm = imm };
     if !Vec.push_raw(&self.insns, &ins as *u8) { ret 0; }
     ret self.insns.len;
@@ -462,11 +471,8 @@ fn dump(self: *Prog, out: *ByteBuffer): bool {
 }
 
 fn ins_at(self: *Prog, i: usize): VmIns {
-    match Vec.get(&self.insns, i) {
-        .Ok(ptr) => { ret (ptr as *VmIns).*; },
-        .Err(error) => { ret VmIns { .opcode = .Error, .ty = .Void, .imm = 0 }; },
-    }
-    ret VmIns { .opcode = .Error, .ty = .Void, .imm = 0 };
+    if i >= self.insns.len { ret VmIns { .opcode = .Error, .ty = .Void, .imm = 0 }; }
+    ret ((self.insns.data as *VmIns) + i).*;
 }
 
 fn func_at(self: *Prog, i: usize): *VmFunc {
@@ -491,41 +497,43 @@ fn align8(n: usize): usize {
     ret (n + 7) & (~(7 as usize));
 }
 
+// 说明：以下 put_* 直接写入已预留容量的 ByteBuffer 存储（store 已按精确总大小 reserve），
+// 跳过 ByteBuffer.append → Vec.append 的校验/拷贝路径。若容量不足会越界，故 store 先校验容量。
+
 fn put_u8(out: *ByteBuffer, b: u8): bool {
-    match ByteBuffer.append_byte(out, b) { .Ok => { ret true; }, .Err(error) => { ret false; } }
-    ret false;
+    out.storage.data[out.storage.len] = b;
+    out.storage.len = out.storage.len + 1;
+    ret true;
 }
 
+// 直接按小端整型写（x86 小端；VM 的 STORE 用 memcpy，未对齐也安全）。
 fn put_u16(out: *ByteBuffer, v: usize): bool {
-    if !put_u8(out, (v & 255) as u8) { ret false; }
-    ret put_u8(out, ((v >> 8) & 255) as u8);
+    var d = out.storage.data + out.storage.len;
+    (d as *u16).* = v as u16;
+    out.storage.len = out.storage.len + 2;
+    ret true;
 }
 
 fn put_u64(out: *ByteBuffer, v: usize): bool {
-    var i: usize = 0;
-    while i < 8 {
-        if !put_u8(out, ((v >> (i * 8)) & 255) as u8) { ret false; }
-        i = i + 1;
-    }
+    var d = out.storage.data + out.storage.len;
+    (d as *u64).* = v as u64;
+    out.storage.len = out.storage.len + 8;
     ret true;
 }
 
 fn put_zeros(out: *ByteBuffer, n: usize): bool {
-    var i: usize = 0;
-    while i < n {
-        if !put_u8(out, 0) { ret false; }
-        i = i + 1;
+    if n != 0 {
+        ignore_ptr(Mem.set((out.storage.data + out.storage.len) as *any, 0, n));
+        out.storage.len = out.storage.len + n;
     }
     ret true;
 }
 
 fn put_view(out: *ByteBuffer, v: ByteView): bool {
     if v.len == 0 { ret true; }
-    match ByteBuffer.append(out, ByteView.as_slice(v)) {
-        .Ok => { ret true; },
-        .Err(error) => { ret false; },
-    }
-    ret false;
+    Mem.copy((out.storage.data + out.storage.len) as *any, v.ptr as *any, v.len);
+    out.storage.len = out.storage.len + v.len;
+    ret true;
 }
 
 // 写入「带结尾 NUL、再按 8 对齐」的名称字段（对应 C 侧 nlen = strlen+1）。
@@ -541,6 +549,41 @@ fn store(self: *Prog, path: ByteView): Status {
     var out = ByteBuffer.new();
     defer { ByteBuffer.drop(&out); }
 
+    // 精确预估输出大小并一次性预留，避免大文件逐字节增长导致的重复拷贝。
+    var total: usize = 48;
+    var fi: usize = 0;
+    while fi < self.funcs.len {
+        var f = func_at(self, fi);
+        if f != null {
+            var nlen = f.name.len + 1;
+            total = total + 8 + align8(nlen) + 32;
+        }
+        fi = fi + 1;
+    }
+    total = total + self.insns.len * 12;
+    fi = 0;
+    while fi < self.natives.len {
+        var n = native_at(self, fi);
+        if n != null {
+            var nlen = n.name.len + 1;
+            total = total + 8 + align8(nlen) + 8;
+        }
+        fi = fi + 1;
+    }
+    fi = 0;
+    while fi < self.strtab.len {
+        total = total + 8 + align8(str_at(self, fi).len);
+        fi = fi + 1;
+    }
+    fi = 1;
+    while fi < self.gdata.len {
+        var e = gdata_at(self, fi);
+        if e != null { total = total + 8 + align8(e.storage.len); }
+        fi = fi + 1;
+    }
+    match ByteBuffer.reserve(&out, total) { .Ok => {}, .Err(error) => {} }
+    if out.storage.cap < total || out.storage.data == null { ret .Err(make(.Io)); }
+
     if !put_view(&out, ByteView.from_slice("SPLBIN")) { ret .Err(make(.Io)); }
     if !put_zeros(&out, 2) { ret .Err(make(.Io)); }
 
@@ -572,9 +615,11 @@ fn store(self: *Prog, path: ByteView): Status {
     i = 0;
     while i < ninsns {
         var ins = ins_at(self, i);
-        if !put_u16(&out, @enum2int(ins.opcode) as usize) { ret .Err(make(.Io)); }
-        if !put_u16(&out, @enum2int(ins.ty) as usize) { ret .Err(make(.Io)); }
-        if !put_u64(&out, ins.imm) { ret .Err(make(.Io)); }
+        var d = out.storage.data + out.storage.len;
+        (d as *u16).* = @enum2int(ins.opcode) as u16;
+        ((d + 2) as *u16).* = @enum2int(ins.ty) as u16;
+        ((d + 4) as *u64).* = ins.imm as u64;
+        out.storage.len = out.storage.len + 12;
         i = i + 1;
     }