sourcemap: bit-packed InternalSourceMap (~2.4 B/mapping, no decode)

bench/sourcemap/.gitignore

@@ -0,0 +1 @@
+big-module.generated.ts

bench/sourcemap/internal-sourcemap-bench.ts

@@ -0,0 +1,52 @@
+// Microbench for the InternalSourceMap stack-trace remapping path. Compare
+// against a baseline build by running:
+//
+//   bun run build:release bench/sourcemap/internal-sourcemap-bench.ts
+//
+// On first run this generates a ~150k-line .ts module next to this file so the
+// stack frames being remapped live inside a large SavedSourceMap entry.
+
+import fs from "node:fs";
+import path from "node:path";
+
+const ITER = 10_000;
+const BIG_LINES = 150_000;
+
+const flag = process.argv[2] ?? "bench";
+const mb = (n: number) => (n / 1048576).toFixed(2) + " MB";
+
+const bigPath = path.join(import.meta.dir, "big-module.generated.ts");
+if (!fs.existsSync(bigPath)) {
+  let src = "let v: number = 0;\n";
+  for (let i = 0; i < BIG_LINES; i++) src += `v = (v + ${i % 97}) | 0;\n`;
+  src += "export function go(): string { return new Error('e').stack!; }\n";
+  fs.writeFileSync(bigPath, src);
+}
+
+const rssBefore = process.memoryUsage().rss;
+const tLoad0 = performance.now();
+const { go } = require(bigPath) as { go: () => string };
+const tLoad1 = performance.now();
+const rssAfterLoad = process.memoryUsage().rss;
+
+// First .stack: triggers full VLQ decode in the old path.
+const tFirst0 = performance.now();
+let s = go();
+const tFirst1 = performance.now();
+const rssAfterFirst = process.memoryUsage().rss;
+
+const t0 = performance.now();
+for (let i = 0; i < ITER; i++) s = go();
+const t1 = performance.now();
+const rssAfterLoop = process.memoryUsage().rss;
+
+console.log(`[${flag}] load big module:          ${(tLoad1 - tLoad0).toFixed(2)} ms`);
+console.log(`[${flag}] first new Error().stack:  ${(tFirst1 - tFirst0).toFixed(3)} ms`);
+console.log(
+  `[${flag}] ${ITER}x new Error().stack: ${(t1 - t0).toFixed(2)} ms (${(((t1 - t0) * 1000) / ITER).toFixed(2)} µs/op)`,
+);
+console.log(
+  `[${flag}] rss before/afterLoad/afterFirst/afterLoop: ${mb(rssBefore)} / ${mb(rssAfterLoad)} / ${mb(rssAfterFirst)} / ${mb(rssAfterLoop)}`,
+);
+console.log(`[${flag}] rss delta load→firstStack: ${mb(rssAfterFirst - rssAfterLoad)}`);
+void s;

src/StandaloneModuleGraph.zig

@@ -264,25 +264,25 @@
             defer init_lock.unlock();
 
             return switch (this.*) {
                 .none => null,
                 .parsed => |map| map,
                 .serialized => |serialized| {
-                    var stored = switch (SourceMap.Mapping.parse(
-                        bun.default_allocator,
-                        serialized.mappingVLQ(),
-                        null,
-                        std.math.maxInt(i32),
-                        std.math.maxInt(i32),
-                        .{},
-                    )) {
-                        .success => |x| x,
-                        .fail => {
-                            this.* = .none;
-                            return null;
-                        },
+                    const blob = serialized.mappingBlob() orelse {
+                        this.* = .none;
+                        return null;
+                    };
+                    if (!SourceMap.InternalSourceMap.isValidBlob(blob)) {
+                        this.* = .none;
+                        return null;
+                    }
+                    const ism = SourceMap.InternalSourceMap{ .data = blob.ptr };
+                    var stored: SourceMap.ParsedSourceMap = .{
+                        .ref_count = .init(),
+                        .internal = ism,
+                        .input_line_count = ism.inputLineCount(),
                     };
 
                     const source_files = serialized.sourceFileNames();
                     const slices = bun.handleOom(bun.default_allocator.alloc(?[]u8, source_files.len * 2));
 
                     const file_names: [][]const u8 = @ptrCast(slices[0..source_files.len]);
@@ -1358,14 +1358,15 @@
 
     /// Source map serialization in the bundler is specially designed to be
     /// loaded in memory as is. Source contents are compressed with ZSTD to
-    /// reduce the file size, and mappings are stored as uncompressed VLQ.
+    /// reduce the file size, and mappings are stored as an InternalSourceMap
+    /// blob (varint deltas + sync points) so lookups need no decode pass.
     pub const SerializedSourceMap = struct {
         bytes: []const u8,
 
         /// Following the header bytes:
         /// - source_files_count number of StringPointer, file names
         /// - source_files_count number of StringPointer, zstd compressed contents
-        /// - the mapping data, `map_vlq_length` bytes
+        /// - the InternalSourceMap blob, `map_bytes_length` bytes
         /// - all the StringPointer contents
         pub const Header = extern struct {
             source_files_count: u32,
@@ -1376,9 +1377,11 @@
             return @ptrCast(map.bytes.ptr);
         }
 
-        pub fn mappingVLQ(map: SerializedSourceMap) []const u8 {
+        pub fn mappingBlob(map: SerializedSourceMap) ?[]const u8 {
+            if (map.bytes.len < @sizeOf(Header)) return null;
             const head = map.header();
             const start = @sizeOf(Header) + head.source_files_count * @sizeOf(StringPointer) * 2;
+            if (start > map.bytes.len or head.map_bytes_length > map.bytes.len - start) return null;
             return map.bytes[start..][0..head.map_bytes_length];
         }
 
@@ -1466,14 +1469,16 @@
         }
 
         const map_vlq: []const u8 = mappings_str.data.e_string.slice(arena);
+        const map_blob = SourceMap.InternalSourceMap.fromVLQ(arena, map_vlq, 0) catch
+            return error.InvalidSourceMap;
 
         try out.writeInt(u32, sources_paths.items.len, .little);
-        try out.writeInt(u32, @intCast(map_vlq.len), .little);
+        try out.writeInt(u32, @intCast(map_blob.len), .little);
 
         const string_payload_start_location = @sizeOf(u32) +
             @sizeOf(u32) +
             @sizeOf(bun.StringPointer) * sources_content.items.len * 2 + // path + source
-            map_vlq.len;
+            map_blob.len;
 
         for (sources_paths.items.slice()) |item| {
             if (item.data != .e_string)
@@ -1519,7 +1524,7 @@
             try out.writeInt(u32, slice.length, .little);
         }
 
-        try out.writeAll(map_vlq);
+        try out.writeAll(map_blob);
 
         bun.assert(header_list.items.len == string_payload_start_location);
     }

src/bun.js/RuntimeTranspilerCache.zig

@@ -15,7 +15,9 @@
 /// Version 16: Added typeof undefined minification optimization.
 /// Version 17: Removed transpiler import rewrite for bun:test. Not bumping it causes test/js/bun/http/req-url-leak.test.ts to fail with SyntaxError: Export named 'expect' not found in module 'bun:test'.
 /// Version 18: Include ESM record (module info) with an ES Module, see #15758
-const expected_version = 18;
+/// Version 19: Sourcemap blob is InternalSourceMap (varint stream + sync points), not VLQ.
+/// Version 20: InternalSourceMap stream is bit-packed windows.
+const expected_version = 20;
 
 const debug = Output.scoped(.cache, .visible);
 const MINIMUM_CACHE_SIZE = 50 * 1024;

src/bun.js/SavedSourceMap.zig

@@ -4,7 +4,12 @@ const SavedSourceMap = @This();
 map: *HashTable,
 mutex: bun.Mutex = .{},
 
-pub const vlq_offset = 24;
+/// Warm cache for `remapStackFramePositions`: the last decoded sync window and
+/// the last (path_hash -> ISM) resolution. Guarded by `mutex`. Invalidated on
+/// any `putValue` since that may free the cached blob.
+find_cache: InternalSourceMap.FindCache = .{},
+last_path_hash: u64 = 0,
+last_ism: ?InternalSourceMap = null,
 
 pub fn init(this: *SavedSourceMap, map: *HashTable) void {
     this.* = .{
@@ -25,70 +30,16 @@ pub inline fn unlock(map: *SavedSourceMap) void {
     map.mutex.unlock();
 }
 
-// For the runtime, we store the number of mappings and how many bytes the final list is at the beginning of the array
-// The first 8 bytes are the length of the array
-// The second 8 bytes are the number of mappings
-pub const SavedMappings = struct {
-    data: [*]u8,
-
-    pub fn vlq(this: SavedMappings) []u8 {
-        return this.data[vlq_offset..this.len()];
-    }
-
-    pub inline fn len(this: SavedMappings) usize {
-        return @as(u64, @bitCast(this.data[0..8].*));
-    }
-
-    pub fn deinit(this: SavedMappings) void {
-        bun.default_allocator.free(this.data[0..this.len()]);
-    }
-
-    pub fn toMapping(this: SavedMappings, allocator: Allocator, path: string) anyerror!ParsedSourceMap {
-        const result = SourceMap.Mapping.parse(
-            allocator,
-            this.data[vlq_offset..this.len()],
-            @as(usize, @bitCast(this.data[8..16].*)),
-            1,
-            @as(usize, @bitCast(this.data[16..24].*)),
-            .{},
-        );
-        switch (result) {
-            .fail => |fail| {
-                if (Output.enable_ansi_colors_stderr) {
-                    try fail.toData(path).writeFormat(
-                        Output.errorWriter(),
-                        logger.Kind.warn,
-                        false,
-                        true,
-                    );
-                } else {
-                    try fail.toData(path).writeFormat(
-                        Output.errorWriter(),
-                        logger.Kind.warn,
-                        false,
-                        false,
-                    );
-                }
-
-                return fail.err;
-            },
-            .success => |success| {
-                return success;
-            },
-        }
-    }
-};
-
-/// ParsedSourceMap is the canonical form for sourcemaps,
-///
-/// but `SavedMappings` and `SourceProviderMap` are much cheaper to construct.
-/// In `fn get`, this value gets converted to ParsedSourceMap always
+/// `InternalSourceMap` is the storage for runtime-transpiled modules.
+/// `ParsedSourceMap` is materialized lazily from a `SourceProviderMap` /
+/// `BakeSourceProvider` / `DevServerSourceProvider` for sources that ship
+/// their own external `.map`.
 pub const Value = bun.TaggedPointerUnion(.{
     ParsedSourceMap,
-    SavedMappings,
     SourceProviderMap,
     BakeSourceProvider,
     DevServerSourceProvider,
+    InternalSourceMap,
 });
 
 pub const MissingSourceMapNoteInfo = struct {
@@ -178,11 +129,10 @@ pub fn deinit(this: *SavedSourceMap) void {
             var value = Value.from(val.*);
             if (value.get(ParsedSourceMap)) |source_map| {
                 source_map.deref();
-            } else if (value.get(SavedMappings)) |saved_mappings| {
-                var saved = SavedMappings{ .data = @as([*]u8, @ptrCast(saved_mappings)) };
-                saved.deinit();
             } else if (value.get(SourceProviderMap)) |provider| {
                 _ = provider; // do nothing, we did not hold a ref to ZigSourceProvider
+            } else if (value.get(InternalSourceMap)) |ism| {
+                (InternalSourceMap{ .data = @as([*]u8, @ptrCast(ism)) }).deinit();
             }
         }
     }
@@ -192,24 +142,27 @@ pub fn deinit(this: *SavedSourceMap) void {
 }
 
 pub fn putMappings(this: *SavedSourceMap, source: *const logger.Source, mappings: MutableString) !void {
-    try this.putValue(source.path.text, Value.init(bun.cast(*SavedMappings, try bun.default_allocator.dupe(u8, mappings.list.items))));
+    const blob = try bun.default_allocator.dupe(u8, mappings.list.items);
+    errdefer bun.default_allocator.free(blob);
+    try this.putValue(source.path.text, Value.init(bun.cast(*InternalSourceMap, blob.ptr)));
 }
 
 pub fn putValue(this: *SavedSourceMap, path: []const u8, value: Value) !void {
     this.lock();
     defer this.unlock();
 
+    this.find_cache.invalidateAll();
+    this.last_ism = null;
     const entry = try this.map.getOrPut(bun.hash(path));
     if (entry.found_existing) {
         var old_value = Value.from(entry.value_ptr.*);
         if (old_value.get(ParsedSourceMap)) |parsed_source_map| {
             var source_map: *ParsedSourceMap = parsed_source_map;
             source_map.deref();
-        } else if (old_value.get(SavedMappings)) |saved_mappings| {
-            var saved = SavedMappings{ .data = @as([*]u8, @ptrCast(saved_mappings)) };
-            saved.deinit();
         } else if (old_value.get(SourceProviderMap)) |provider| {
             _ = provider; // do nothing, we did not hold a ref to ZigSourceProvider
+        } else if (old_value.get(InternalSourceMap)) |ism| {
+            (InternalSourceMap{ .data = @as([*]u8, @ptrCast(ism)) }).deinit();
         }
     }
     entry.value_ptr.* = value.ptr();
@@ -233,25 +186,30 @@ fn getWithContent(
     };
 
     switch (Value.from(mapping.value_ptr.*).tag()) {
-        @field(Value.Tag, @typeName(ParsedSourceMap)) => {
+        @field(Value.Tag, @typeName(InternalSourceMap)) => {
+            // Runtime-transpiled module. Wrap the blob in a refcounted
+            // ParsedSourceMap shell (no VLQ decode, no Mapping.List) so callers
+            // can hold a ref while the table mutates. The shell takes ownership
+            // of the blob.
             defer this.unlock();
-            const map = Value.from(mapping.value_ptr.*).as(ParsedSourceMap);
-            map.ref();
-            return .{ .map = map };
-        },
-        @field(Value.Tag, @typeName(SavedMappings)) => {
-            defer this.unlock();
-            var saved = SavedMappings{ .data = @as([*]u8, @ptrCast(Value.from(mapping.value_ptr.*).as(ParsedSourceMap))) };
-            defer saved.deinit();
-            const result = bun.new(ParsedSourceMap, saved.toMapping(bun.default_allocator, path) catch {
-                _ = this.map.remove(mapping.key_ptr.*);
-                return .{};
+            const ism: InternalSourceMap = .{
+                .data = @as([*]u8, @ptrCast(Value.from(mapping.value_ptr.*).as(InternalSourceMap))),
+            };
+            const result = bun.new(ParsedSourceMap, .{
+                .ref_count = .init(),
+                .input_line_count = ism.inputLineCount(),
+                .internal = ism,
             });
             mapping.value_ptr.* = Value.init(result).ptr();
             result.ref();
-
             return .{ .map = result };
         },
+        @field(Value.Tag, @typeName(ParsedSourceMap)) => {
+            defer this.unlock();
+            const map = Value.from(mapping.value_ptr.*).as(ParsedSourceMap);
+            map.ref();
+            return .{ .map = map };
+        },
         @field(Value.Tag, @typeName(SourceProviderMap)) => {
             const ptr: *SourceProviderMap = Value.from(mapping.value_ptr.*).as(SourceProviderMap);
             this.unlock();
@@ -348,6 +306,12 @@ pub fn get(this: *SavedSourceMap, path: string) ?*ParsedSourceMap {
     return this.getWithContent(path, .mappings_only).map;
 }
 
+/// Mutex must already be held. Returns the raw table value for `hash` if any.
+pub fn getValueLocked(this: *SavedSourceMap, hash: u64) ?Value {
+    const raw = this.map.get(hash) orelse return null;
+    return Value.from(raw);
+}
+
 pub fn resolveMapping(
     this: *SavedSourceMap,
     path: []const u8,
@@ -362,7 +326,7 @@ pub fn resolveMapping(
     const map = parse.map orelse return null;
 
     const mapping = parse.mapping orelse
-        map.mappings.find(line, column) orelse
+        map.findMapping(line, column) orelse
         return null;
 
     return .{
@@ -375,7 +339,6 @@ pub fn resolveMapping(
 const string = []const u8;
 
 const std = @import("std");
-const Allocator = std.mem.Allocator;
 
 const bun = @import("bun");
 const Environment = bun.Environment;
@@ -387,5 +350,6 @@ const logger = bun.logger;
 const SourceMap = bun.SourceMap;
 const BakeSourceProvider = bun.SourceMap.BakeSourceProvider;
 const DevServerSourceProvider = bun.SourceMap.DevServerSourceProvider;
+const InternalSourceMap = SourceMap.InternalSourceMap;
 const ParsedSourceMap = SourceMap.ParsedSourceMap;
 const SourceProviderMap = SourceMap.SourceProviderMap;