rustc: Enable LTO and multiple codegen units

src/librustc/session/config.rs

@@ -1552,24 +1552,6 @@ pub fn build_session_options_and_crate_config(matches: &getopts::Matches)
         early_error(error_format, "Value for codegen units must be a positive nonzero integer");
     }
 
-    // It's possible that we have `codegen_units > 1` but only one item in
-    // `trans.modules`.  We could theoretically proceed and do LTO in that
-    // case, but it would be confusing to have the validity of
-    // `-Z lto -C codegen-units=2` depend on details of the crate being
-    // compiled, so we complain regardless.
-    if cg.lto {
-        if let Some(n) = codegen_units {
-            if n > 1 {
-                // This case is impossible to handle because LTO expects to be able
-                // to combine the entire crate and all its dependencies into a
-                // single compilation unit, but each codegen unit is in a separate
-                // LLVM context, so they can't easily be combined.
-                early_error(error_format, "can't perform LTO when using multiple codegen units");
-            }
-        }
-        codegen_units = Some(1);
-    }
-
     if cg.lto && debugging_opts.incremental.is_some() {
         early_error(error_format, "can't perform LTO when compiling incrementally");
     }

src/librustc_llvm/ffi.rs

@@ -478,6 +478,7 @@ pub mod debuginfo {
     }
 }
 
+pub enum ModuleBuffer {}
 
 // Link to our native llvm bindings (things that we need to use the C++ api
 // for) and because llvm is written in C++ we need to link against libstdc++
@@ -1609,6 +1610,7 @@ extern "C" {
     pub fn LLVMRustSetNormalizedTarget(M: ModuleRef, triple: *const c_char);
     pub fn LLVMRustAddAlwaysInlinePass(P: PassManagerBuilderRef, AddLifetimes: bool);
     pub fn LLVMRustLinkInExternalBitcode(M: ModuleRef, bc: *const c_char, len: size_t) -> bool;
+    pub fn LLVMRustLinkInParsedExternalBitcode(M: ModuleRef, M: ModuleRef) -> bool;
     pub fn LLVMRustRunRestrictionPass(M: ModuleRef, syms: *const *const c_char, len: size_t);
     pub fn LLVMRustMarkAllFunctionsNounwind(M: ModuleRef);
 
@@ -1678,4 +1680,9 @@ extern "C" {
     pub fn LLVMRustSetComdat(M: ModuleRef, V: ValueRef, Name: *const c_char);
     pub fn LLVMRustUnsetComdat(V: ValueRef);
     pub fn LLVMRustSetModulePIELevel(M: ModuleRef);
+    pub fn LLVMRustModuleBufferCreate(M: ModuleRef) -> *mut ModuleBuffer;
+    pub fn LLVMRustModuleBufferPtr(p: *const ModuleBuffer) -> *const u8;
+    pub fn LLVMRustModuleBufferLen(p: *const ModuleBuffer) -> usize;
+    pub fn LLVMRustModuleBufferFree(p: *mut ModuleBuffer);
+    pub fn LLVMRustModuleCost(M: ModuleRef) -> u64;
 }

src/librustc_trans/back/archive.rs

@@ -17,6 +17,7 @@ use std::path::{Path, PathBuf};
 use std::ptr;
 use std::str;
 
+use back::bytecode::RLIB_BYTECODE_EXTENSION;
 use libc;
 use llvm::archive_ro::{ArchiveRO, Child};
 use llvm::{self, ArchiveKind};
@@ -154,12 +155,9 @@ impl<'a> ArchiveBuilder<'a> {
         // might be also an extra name suffix
         let obj_start = format!("{}", name);
 
-        // Ignoring all bytecode files, no matter of
-        // name
-        let bc_ext = ".bytecode.deflate";
-
         self.add_archive(rlib, move |fname: &str| {
-            if fname.ends_with(bc_ext) || fname == METADATA_FILENAME {
+            // Ignore bytecode/metadata files, no matter the name.
+            if fname.ends_with(RLIB_BYTECODE_EXTENSION) || fname == METADATA_FILENAME {
                 return true
             }
 

src/librustc_trans/back/bytecode.rs

@@ -0,0 +1,160 @@
+// Copyright 2017 The Rust Project Developers. See the COPYRIGHT
+// file at the top-level directory of this distribution and at
+// http://rust-lang.org/COPYRIGHT.
+//
+// Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
+// http://www.apache.org/licenses/LICENSE-2.0> or the MIT license
+// <LICENSE-MIT or http://opensource.org/licenses/MIT>, at your
+// option. This file may not be copied, modified, or distributed
+// except according to those terms.
+
+//! Management of the encoding of LLVM bytecode into rlibs
+//!
+//! This module contains the management of encoding LLVM bytecode into rlibs,
+//! primarily for the usage in LTO situations. Currently the compiler will
+//! unconditionally encode LLVM-IR into rlibs regardless of what's happening
+//! elsewhere, so we currently compress the bytecode via deflate to avoid taking
+//! up too much space on disk.
+//!
+//! After compressing the bytecode we then have the rest of the format to
+//! basically deal with various bugs in various archive implementations. The
+//! format currently is:
+//!
+//!     RLIB LLVM-BYTECODE OBJECT LAYOUT
+//!     Version 2
+//!     Bytes    Data
+//!     0..10    "RUST_OBJECT" encoded in ASCII
+//!     11..14   format version as little-endian u32
+//!     15..19   the length of the module identifier string
+//!     20..n    the module identifier string
+//!     n..n+8   size in bytes of deflate compressed LLVM bitcode as
+//!              little-endian u64
+//!     n+9..    compressed LLVM bitcode
+//!     ?        maybe a byte to make this whole thing even length
+
+use std::io::{Read, Write};
+use std::ptr;
+use std::str;
+
+use flate2::Compression;
+use flate2::read::DeflateDecoder;
+use flate2::write::DeflateEncoder;
+
+// This is the "magic number" expected at the beginning of a LLVM bytecode
+// object in an rlib.
+pub const RLIB_BYTECODE_OBJECT_MAGIC: &'static [u8] = b"RUST_OBJECT";
+
+// The version number this compiler will write to bytecode objects in rlibs
+pub const RLIB_BYTECODE_OBJECT_VERSION: u8 = 2;
+
+pub const RLIB_BYTECODE_EXTENSION: &str = "bytecode.encoded";
+
+pub fn encode(identifier: &str, bytecode: &[u8]) -> Vec<u8> {
+    let mut encoded = Vec::new();
+
+    // Start off with the magic string
+    encoded.extend_from_slice(RLIB_BYTECODE_OBJECT_MAGIC);
+
+    // Next up is the version
+    encoded.extend_from_slice(&[RLIB_BYTECODE_OBJECT_VERSION, 0, 0, 0]);
+
+    // Next is the LLVM module identifier length + contents
+    let identifier_len = identifier.len();
+    encoded.extend_from_slice(&[
+        (identifier_len >>  0) as u8,
+        (identifier_len >>  8) as u8,
+        (identifier_len >> 16) as u8,
+        (identifier_len >> 24) as u8,
+    ]);
+    encoded.extend_from_slice(identifier.as_bytes());
+
+    // Next is the LLVM module deflate compressed, prefixed with its length. We
+    // don't know its length yet, so fill in 0s
+    let deflated_size_pos = encoded.len();
+    encoded.extend_from_slice(&[0, 0, 0, 0, 0, 0, 0, 0]);
+
+    let before = encoded.len();
+    DeflateEncoder::new(&mut encoded, Compression::Fast)
+        .write_all(bytecode)
+        .unwrap();
+    let after = encoded.len();
+
+    // Fill in the length we reserved space for before
+    let bytecode_len = (after - before) as u64;
+    encoded[deflated_size_pos + 0] = (bytecode_len >>  0) as u8;
+    encoded[deflated_size_pos + 1] = (bytecode_len >>  8) as u8;
+    encoded[deflated_size_pos + 2] = (bytecode_len >> 16) as u8;
+    encoded[deflated_size_pos + 3] = (bytecode_len >> 24) as u8;
+    encoded[deflated_size_pos + 4] = (bytecode_len >> 32) as u8;
+    encoded[deflated_size_pos + 5] = (bytecode_len >> 40) as u8;
+    encoded[deflated_size_pos + 6] = (bytecode_len >> 48) as u8;
+    encoded[deflated_size_pos + 7] = (bytecode_len >> 56) as u8;
+
+    // If the number of bytes written to the object so far is odd, add a
+    // padding byte to make it even. This works around a crash bug in LLDB
+    // (see issue #15950)
+    if encoded.len() % 2 == 1 {
+        encoded.push(0);
+    }
+
+    return encoded
+}
+
+pub struct DecodedBytecode<'a> {
+    identifier: &'a str,
+    encoded_bytecode: &'a [u8],
+}
+
+impl<'a> DecodedBytecode<'a> {
+    pub fn new(data: &'a [u8]) -> Result<DecodedBytecode<'a>, String> {
+        if !data.starts_with(RLIB_BYTECODE_OBJECT_MAGIC) {
+            return Err(format!("magic bytecode prefix not found"))
+        }
+        let data = &data[RLIB_BYTECODE_OBJECT_MAGIC.len()..];
+        if !data.starts_with(&[RLIB_BYTECODE_OBJECT_VERSION, 0, 0, 0]) {
+            return Err(format!("wrong version prefix found in bytecode"))
+        }
+        let data = &data[4..];
+        if data.len() < 4 {
+            return Err(format!("bytecode corrupted"))
+        }
+        let identifier_len = unsafe {
+            u32::from_le(ptr::read_unaligned(data.as_ptr() as *const u32)) as usize
+        };
+        let data = &data[4..];
+        if data.len() < identifier_len {
+            return Err(format!("bytecode corrupted"))
+        }
+        let identifier = match str::from_utf8(&data[..identifier_len]) {
+            Ok(s) => s,
+            Err(_) => return Err(format!("bytecode corrupted"))
+        };
+        let data = &data[identifier_len..];
+        if data.len() < 8 {
+            return Err(format!("bytecode corrupted"))
+        }
+        let bytecode_len = unsafe {
+            u64::from_le(ptr::read_unaligned(data.as_ptr() as *const u64)) as usize
+        };
+        let data = &data[8..];
+        if data.len() < bytecode_len {
+            return Err(format!("bytecode corrupted"))
+        }
+        let encoded_bytecode = &data[..bytecode_len];
+
+        Ok(DecodedBytecode {
+            identifier,
+            encoded_bytecode,
+        })
+    }
+
+    pub fn bytecode(&self) -> Vec<u8> {
+        let mut data = Vec::new();
+        DeflateDecoder::new(self.encoded_bytecode).read_to_end(&mut data).unwrap();
+        return data
+    }
+
+    pub fn identifier(&self) -> &'a str {
+        self.identifier
+    }
+}