diff --git a/README.md b/README.md
index c3bf385..7d703a8 100644
--- a/README.md
+++ b/README.md
@@ -12,11 +12,11 @@ My main goals for this project are as follows:
    but the vast majority is still just bindings
 
 With some very basic benchmarking, extracting a zlib-compressed AppImage
-(FreeCAD, the largest AppImage I've been able to find so far), takes 3.7
-seconds using squashfuse-zig's `squashfuse_tool`. Currently, `squashfs_tool`
-is single-thread only.
+(FreeCAD, the largest AppImage I've been able to find so far), takes 3.1
+seconds using the `--extract` flag with squashfuse-zig's CLI tool, which
+is currently single-thread only.
 
-For reference, `unsquashfs` with multi-threaded decompression takes 1.57 seconds
+For reference, `unsquashfs` with multi-threaded decompression takes 1.5 seconds
 and single-threaded takes 6.5 seconds.
 
 Surely almost all of the single-threaded performace gain can be chalked up to
diff --git a/lib/SquashFs.zig b/lib/SquashFs.zig
index c09603e..f2fea86 100644
--- a/lib/SquashFs.zig
+++ b/lib/SquashFs.zig
@@ -650,15 +650,42 @@ usingnamespace if (build_options.enable_zlib)
             *usize,
         ) c_int;
 
-        var ldef_decompressor: ?*anyopaque = null;
+        // Deflate constants
+        const litlen_syms = 288;
+        const offset_syms = 32;
+        const max_lens_overrun = 137;
+        const max_num_syms = 288;
+        const precode_syms = 19;
+
+        // LibDeflate constants
+        const precode_enough = 128;
+        const litlen_enough = 2342;
+        const offset_enough = 402;
+
+        const Decompressor = extern struct {
+            _: extern union {
+                precode_lens: [precode_syms]u8,
+
+                _: extern struct {
+                    lens: [litlen_syms + offset_syms + max_lens_overrun]u8,
+                    precode_table: [precode_enough]u32,
+                },
+
+                litlen_decode_table: [litlen_enough]u32,
+            } = undefined,
+
+            offset_decode_table: [offset_enough]u32 = undefined,
+            sorted_syms: [max_num_syms]u16 = undefined,
+            static_codes_loaded: bool = false,
+            litlen_tablebits: u32 = undefined,
+            free_func: ?*anyopaque = undefined,
+        };
 
         export fn zig_zlib_decode(in: [*]u8, in_size: usize, out: [*]u8, out_size: *usize) callconv(.C) c.sqfs_err {
-            if (ldef_decompressor == null) {
-                ldef_decompressor = libdeflate_alloc_decompressor();
-            }
+            var decompressor = Decompressor{};
 
             const err = libdeflate_zlib_decompress(
-                ldef_decompressor.?,
+                &decompressor,
                 in,
                 in_size,
                 out,
diff --git a/lib/test.zig b/lib/test.zig
index c5d4194..8e1def7 100644
--- a/lib/test.zig
+++ b/lib/test.zig
@@ -17,6 +17,7 @@ test "open SquashFS image (zlib)" {
 }
 
 // TODO: more tests
+// TODO: test executable mounting, extracting, etc
 fn testWalk(allocator: std.mem.Allocator, sqfs: *SquashFs) !void {
     var root_inode = sqfs.getRootInode();
 
diff --git a/src/squashfuse.zig b/src/squashfuse.zig
index c5b4191..146d902 100644
--- a/src/squashfuse.zig
+++ b/src/squashfuse.zig
@@ -29,18 +29,19 @@ pub fn main() !void {
     var stdout = std.io.getStdOut().writer();
 
     const params = comptime clap.parseParamsComptime(
-        \\-h, --help               display this help and exit
-        \\-f, --foreground         run in foreground
-        \\-d, --debug              enable debug output (runs in foreground)
-        \\-x, --extract            extract the SquashFS image
-        \\-l, --list               list file tree of SquashFS image
-        \\-o, --option <str>...    use a libFUSE mount option
+        \\-h, --help             display this help and exit
+        \\-f, --foreground       run in foreground
+        \\-d, --debug            enable debug output (runs in foreground)
+        \\-x, --extract          extract the SquashFS image
+        \\-l, --list             list file tree of SquashFS image
+        \\-o, --option <str>...  use a libFUSE mount option
+        \\
+        \\--offset <usize>      mount at an offset
+        \\--extract-src <str>   must be used with `--extract`; specify the source inode
+        \\--extract-dest <str>  must be used with `--extract`; specify the destination name
+        \\--version             print the current version
+        \\--verbose             enable verbose printing
         \\
-        \\    --offset <usize>  mount at an offset
-        \\    --extract-src <str>  must be used with `--extract`; specify the source inode
-        \\    --extract-dest <str> must be used with `--extract`; specify the destination name
-        \\    --version         print the current version
-        \\    --verbose         enable verbose printing
         \\<str>...
     );
 
@@ -467,6 +468,11 @@ fn extractArchive(
 
     var file_found = false;
 
+    // TODO: flag to change buf size
+    const buf_size = 1024 * 1024;
+    const buf = try allocator.alloc(u8, buf_size);
+    defer allocator.free(buf);
+
     // Iterate over the SquashFS image and extract each item
     while (try walker.next()) |entry| {
         // Skip if the path doesn't match our source
@@ -503,13 +509,11 @@ fn extractArchive(
             try stdout.print("{s}\n", .{prefixed_dest});
         }
 
-        // TODO: flag to change buf size
-        var buf: [4096]u8 = undefined;
-
         var inode = entry.inode();
-        try inode.extract(&buf, prefixed_dest);
+        try inode.extract(buf, prefixed_dest);
     }
 
+    // TODO: better error message
     if (!file_found) {
         std.debug.print("file ({s}) not found!\n", .{real_src});
     }