WIP write out windows import library

Author: simonbuchan

src/archive.rs

@@ -1,3 +1,4 @@
+use std::fs;
 use std::path::{Path, PathBuf};
 
 use rustc_codegen_ssa::back::archive::{
@@ -15,11 +16,289 @@ impl ArchiveBuilderBuilder for ArArchiveBuilderBuilder {
     fn create_dll_import_lib(
         &self,
         _sess: &Session,
-        _lib_name: &str,
-        _dll_imports: &[rustc_session::cstore::DllImport],
-        _tmpdir: &Path,
+        lib_name: &str,
+        dll_imports: &[rustc_session::cstore::DllImport],
+        tmpdir: &Path,
         _is_direct_dependency: bool,
     ) -> PathBuf {
-        unimplemented!("creating dll imports is not yet supported");
+        let mut import_names = Vec::new();
+        for dll_import in dll_imports {
+            import_names.push(dll_import.name.as_str());
+        }
+        let lib_path = tmpdir.join(format!("{}.lib", lib_name));
+        // todo: emit session error instead of expects
+        fs::write(&lib_path, windows_import_lib::generate(lib_name, &import_names))
+            .expect("failed to write import library");
+
+        lib_path
+    }
+}
+
+// todo: pull out to a proper location. Really should be in `object` crate!
+// todo: support ordinals
+// todo: support name types (e.g. verbatim+)
+// todo: support long member names
+// todo: support windows-gnu flavor?
+// todo: provide machine
+// todo: remove any panics, nice errors
+mod windows_import_lib {
+    // https://learn.microsoft.com/en-us/windows/win32/debug/pe-format#archive-library-file-format
+    //
+    // Windows .lib files are System-V (aka. GUN) flavored ar files with a couple of extra lookup
+    // members.
+    //
+    // An archive is the 8 bytes b"!<arch>\n"
+    // followed by a sequence of 60 byte member headers:
+    //   0: name: [u8; 16], // member name, terminated with "/". If it is longer than 15, then
+    //                      // use "/n" where "n" is a decimal for the offset in bytes into
+    //                      // the longnames ("//") member contents.
+    //  16: date: [u8; 12], // ASCII decimal seconds since UNIX epoch - always -1 for MSVC
+    //  28: uid: [u8; 6],   // ASCII decimal user id. Always blank for MSVC
+    //  34: gid: [u8; 6],   // ditto for group id.
+    //  40: mode: [u8; 8],  // ASCII octal UNIX mode. 0 for MSVC
+    //  48: size: [u8; 10], // ASCII decimal data size.
+    //  58: end: b"`\n",
+    // then size bytes of payload. If payload is odd sized, pad
+    // to an even offset with \n.
+    //
+    // You must store two extra members at the start, a legacy member lookup table member
+    // and the current member lookup and symbol table, both with empty ("/") names.
+    //
+    // The legacy table member has the name "/" with following contents, using big-endian numbers:
+    //   count: u32,            // number of indexed symbols
+    //   offsets: [u32, count], // file offsets to the header of the member that contains
+    //                          // that symbol.
+    //   names: *               // sequence of null terminated symbol names.
+    //
+    // The current table member also has the name "/", and has the following contents, using
+    // little-endian numbers:
+    //   member_count: u32,                   // number of members
+    //   member_offsets: [u32; member_count], // file offsets to each member header
+    //   symbol_count: u32,                   // number of symbols
+    //   symbol_member: [u16; symbol_count],  // *1-based* index of the member that contains
+    //                                        // each symbol
+    //   symbol_names: *                      // sequence of null terminated symbol names
+    //
+    // Then the long names member ("//") as with regular GNU ar files, just a sequence of
+    // null terminated strings indexed by members using the long name format "/n" as described
+    // above.
+    //
+    // Then regular members follow.
+    //
+    // This library emits only import libraries, that is, libraries with a short import object
+    // describing an import from a dll. That means each member contains exactly one symbol. The member
+    // name doesn't seem to matter, including duplicates, we use the dll name since that's what's in the
+    // files generated by MSVC tools.
+    //
+    // The short import object has the form:
+    //   header:
+    //     sig1: 0u16
+    //     sig2: 0xFFFFu16
+    //     version: u16, // normally 0
+    //     machine: u16, // IMAGE_MACHINE_* value, e.g. 0x8664 for AMD64
+    //     time_date_stamp: u32, // normally 0
+    //     size_of_data: u32, // size following the header
+    //     ordinal_or_hint: u16, // depending on flag
+    //     object_type: u2, // IMPORT_OBJECT_{CODE,DATA,CONST} = 0, 1, 2
+    //     name_type: u3,   // IMPORT_OBJECT_{ORDINAL,NAME,NAME_NO_PREFIX,NAME_UNDECORATE,NAME_EXPORTAS} = 0, 1, 2, 3, 4
+    //     reserved: u11,
+    //   data:  // size_of_data bytes
+    //      name: * // import name; null terminated string
+    //      dll_name: * // dll name; null terminated string
+    pub fn generate(dll_name: &str, import_names: &[&str]) -> Vec<u8> {
+        assert!(dll_name.len() < 16, "long member names not supported yet");
+        assert!(import_names.len() <= 0xFFFF, "too many import names");
+        // number of symbols, and members containing symbols for symbol lookup members
+        let symbol_count = import_names.len();
+
+        let mut writer = Writer::new();
+
+        // legacy symbol directory
+        let mut legacy_symbol_directory = writer.start_member_raw();
+        legacy_symbol_directory.set_raw_name(b"/");
+        legacy_symbol_directory.write_u32_be(symbol_count as u32);
+        // reserve space for offsets.
+        let legacy_member_table_offset = legacy_symbol_directory.reserve_bytes(symbol_count * 4);
+        // string table
+        for name in import_names {
+            legacy_symbol_directory.write_c_str(name);
+        }
+        // done with legacy symbol directory
+        drop(legacy_symbol_directory);
+
+        // current symbol directory
+        let mut current_symbol_directory = writer.start_member_raw();
+        current_symbol_directory.set_raw_name(b"/");
+        // member count: same as symbol count for import library
+        current_symbol_directory.write_u32_le(symbol_count as u32);
+        // reserve space for member offsets
+        let current_member_table_offset = current_symbol_directory.reserve_bytes(symbol_count * 4);
+        // symbol count
+        current_symbol_directory.write_u32_le(symbol_count as u32);
+        // we assume symbol members are already in order
+        for index in 0..import_names.len() as u16 {
+            current_symbol_directory.write_u16_le(1 + index);
+        }
+        // string table again (could just copy from legacy string table above?)
+        for name in import_names {
+            current_symbol_directory.write_c_str(name);
+        }
+        // done with current symbol directory
+        drop(current_symbol_directory);
+
+        // long names member not supported yet
+
+        // import members
+        for (index, name) in import_names.iter().enumerate() {
+            let mut member = writer.start_member(dll_name);
+            // update member offsets
+            let member_offset = member.header_offset as u32;
+            member.data[legacy_member_table_offset + index * 4..][..4]
+                .copy_from_slice(&member_offset.to_be_bytes());
+            member.data[current_member_table_offset + index * 4..][..4]
+                .copy_from_slice(&member_offset.to_le_bytes());
+            // write import object:
+            //   signature
+            member.write_u16_le(0);
+            member.write_u16_le(0xFFFF);
+            //   version
+            member.write_u16_le(0);
+            //   machine = AMD64
+            member.write_u16_le(0x8664);
+            //   time_date_stamp
+            member.write_u32_le(0);
+            //   size_of_data
+            member.write_u32_le((dll_name.len() + 1 + name.len() + 1) as u32);
+            //   ordinal_or_hint
+            member.write_u16_le(0);
+            //   object_type | name_type = IMPORT_OBJECT_CODE | IMPORT_OBJECT_NAME
+            member.write_u16_le(1 << 2 | 0);
+            // data:
+            //   name
+            member.write_c_str(name);
+            //   dll_name
+            member.write_c_str(dll_name);
+
+            drop(member);
+        }
+
+        writer.data
+    }
+
+    struct Writer {
+        data: Vec<u8>,
+    }
+
+    impl Writer {
+        fn new() -> Self {
+            Self { data: Vec::from(*b"!<arch>\n") }
+        }
+
+        fn start_member_raw(&mut self) -> Member<'_> {
+            let header_offset = self.data.len();
+            // fill the header with blanks...
+            self.data.resize(header_offset + Member::HEADER_SIZE - 2, b' ');
+            // except for end marker
+            self.data.extend_from_slice(b"`\n");
+
+            let mut member = Member::new(&mut self.data, header_offset);
+            // init date, mode to default values as produced by MSVC tools
+            member.set_time_date_stamp(-1);
+            member.set_mode(0);
+            member
+        }
+
+        fn start_member(&mut self, name: &str) -> Member<'_> {
+            let mut member = self.start_member_raw();
+            member.set_name(name);
+            member
+        }
+    }
+
+    struct Member<'a> {
+        data: &'a mut Vec<u8>,
+        header_offset: usize,
+    }
+
+    impl<'a> Member<'a> {
+        const HEADER_SIZE: usize = 60;
+
+        fn new(data: &'a mut Vec<u8>, header_offset: usize) -> Self {
+            Self { data, header_offset }
+        }
+
+        fn header_slice(&mut self, offset: usize, len: usize) -> &mut [u8] {
+            &mut self.data[self.header_offset + offset..][..len]
+        }
+
+        fn set_name(&mut self, name: &str) {
+            assert!(name.len() < 16, "long member names not supported yet");
+            self.set_raw_name(name.as_bytes());
+            self.data[self.header_offset + name.len()] = b'/';
+        }
+
+        fn set_raw_name(&mut self, raw_name: &[u8]) {
+            assert!(raw_name.len() <= 16, "raw name must be <= 16 bytes");
+            self.header_slice(0, raw_name.len()).copy_from_slice(raw_name);
+        }
+
+        fn set_time_date_stamp(&mut self, value: i32) {
+            self.set_decimal_field(16, 12, value);
+        }
+
+        fn set_uid(&mut self, value: i32) {
+            self.set_decimal_field(28, 6, value);
+        }
+
+        fn set_gid(&mut self, value: i32) {
+            self.set_decimal_field(34, 6, value);
+        }
+
+        fn set_mode(&mut self, value: i32) {
+            use std::io::Write;
+            write!(std::io::Cursor::new(self.header_slice(40, 8)), "{value:o}")
+                .expect("value too large");
+        }
+
+        fn set_decimal_field(&mut self, offset: usize, size: usize, value: i32) {
+            use std::io::Write;
+            write!(std::io::Cursor::new(self.header_slice(offset, size)), "{value}")
+                .expect("value too large");
+        }
+
+        fn write_c_str(&mut self, data: &str) {
+            self.data.extend_from_slice(data.as_bytes());
+            self.data.push(0);
+        }
+
+        fn write_u16_le(&mut self, data: u16) {
+            self.data.extend_from_slice(&data.to_le_bytes());
+        }
+
+        fn write_u32_be(&mut self, data: u32) {
+            self.data.extend_from_slice(&data.to_be_bytes());
+        }
+
+        fn write_u32_le(&mut self, data: u32) {
+            self.data.extend_from_slice(&data.to_le_bytes());
+        }
+
+        fn reserve_bytes(&mut self, count: usize) -> usize {
+            let offset = self.data.len();
+            self.data.resize(offset + count, 0);
+            offset
+        }
+    }
+
+    impl<'a> Drop for Member<'a> {
+        fn drop(&mut self) {
+            let data_size = self.data.len() - self.header_offset - Self::HEADER_SIZE;
+            assert!(data_size < i32::MAX as usize);
+            self.set_decimal_field(48, 10, data_size as i32);
+            // pad to even address
+            if data_size % 2 == 1 {
+                self.data.push(b'\n');
+            }
+        }
     }
 }