mikros/rust
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Rollup merge of #63956 - alexcrichton:fix-lto-all-cached, r=michaelwoerister

Browse Source 
rustc: Handle modules in "fat" LTO more robustly

When performing a "fat" LTO the compiler has a whole mess of codegen
units that it links together. To do this it needs to select one module
as a "base" module and then link everything else into this module.
Previously LTO passes assume that there's at least one module in-memory
to link into, but nowadays that's not always true! With incremental
compilation modules may actually largely be cached and it may be
possible that there's no in-memory modules to work with.

This commit updates the logic of the LTO backend to handle modules a bit
more uniformly during a fat LTO. This commit immediately splits them
into two lists, one serialized and one in-memory. The in-memory list is
then searched for the largest module and failing that we simply
deserialize the first serialized module and link into that. This
refactoring avoids juggling three lists, two of which are serialized
modules and one of which is half serialized and half in-memory.

Closes #63349
This commit is contained in:
Mazdak Farrokhzad 2019-08-29 17:14:01 +02:00 committed by GitHub
commit 137564d825
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
5 changed files with 59 additions and 49 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

80
src/librustc_codegen_llvm/back/lto.rs
View File

@ -183,7 +183,7 @@ pub(crate) fn prepare_thin(
fn fat_lto(cgcx: &CodegenContext<LlvmCodegenBackend>,
diag_handler: &Handler,
mut modules: Vec<FatLTOInput<LlvmCodegenBackend>>,
modules: Vec<FatLTOInput<LlvmCodegenBackend>>,
cached_modules: Vec<(SerializedModule<ModuleBuffer>, WorkProduct)>,
mut serialized_modules: Vec<(SerializedModule<ModuleBuffer>, CString)>,
symbol_white_list: &[*const libc::c_char])
@ -191,6 +191,32 @@ fn fat_lto(cgcx: &CodegenContext<LlvmCodegenBackend>,
{
info!("going for a fat lto");
// Sort out all our lists of incoming modules into two lists.
//
// * `serialized_modules` (also and argument to this function) contains all
// modules that are serialized in-memory.
// * `in_memory` contains modules which are already parsed and in-memory,
// such as from multi-CGU builds.
//
// All of `cached_modules` (cached from previous incremental builds) can
// immediately go onto the `serialized_modules` modules list and then we can
// split the `modules` array into these two lists.
let mut in_memory = Vec::new();
serialized_modules.extend(cached_modules.into_iter().map(|(buffer, wp)| {
info!("pushing cached module {:?}", wp.cgu_name);
(buffer, CString::new(wp.cgu_name).unwrap())
}));
for module in modules {
match module {
FatLTOInput::InMemory(m) => in_memory.push(m),
FatLTOInput::Serialized { name, buffer } => {
info!("pushing serialized module {:?}", name);
let buffer = SerializedModule::Local(buffer);
serialized_modules.push((buffer, CString::new(name).unwrap()));
}
}
}
// Find the "costliest" module and merge everything into that codegen unit.
// All the other modules will be serialized and reparsed into the new
// context, so this hopefully avoids serializing and parsing the largest
@ -200,14 +226,8 @@ fn fat_lto(cgcx: &CodegenContext<LlvmCodegenBackend>,
// file copy operations in the backend work correctly. The only other kind
// of module here should be an allocator one, and if your crate is smaller
// than the allocator module then the size doesn't really matter anyway.
let costliest_module = modules.iter()
let costliest_module = in_memory.iter()
.enumerate()
.filter_map(|(i, module)| {
match module {
FatLTOInput::InMemory(m) => Some((i, m)),
FatLTOInput::Serialized { .. } => None,
}
})
.filter(|&(_, module)| module.kind == ModuleKind::Regular)
.map(|(i, module)| {
let cost = unsafe {
@ -223,26 +243,14 @@ fn fat_lto(cgcx: &CodegenContext<LlvmCodegenBackend>,
// re-executing the LTO passes. If that's the case deserialize the first
// module and create a linker with it.
let module: ModuleCodegen<ModuleLlvm> = match costliest_module {
Some((_cost, i)) => {
match modules.remove(i) {
FatLTOInput::InMemory(m) => m,
FatLTOInput::Serialized { .. } => unreachable!(),
}
}
Some((_cost, i)) => in_memory.remove(i),
None => {
let pos = modules.iter().position(|m| {
match m {
FatLTOInput::InMemory(_) => false,
FatLTOInput::Serialized { .. } => true,
}
}).expect("must have at least one serialized module");
let (name, buffer) = match modules.remove(pos) {
FatLTOInput::Serialized { name, buffer } => (name, buffer),
FatLTOInput::InMemory(_) => unreachable!(),
};
assert!(serialized_modules.len() > 0, "must have at least one serialized module");
let (buffer, name) = serialized_modules.remove(0);
info!("no in-memory regular modules to choose from, parsing {:?}", name);
ModuleCodegen {
module_llvm: ModuleLlvm::parse(cgcx, &name, &buffer, diag_handler)?,
name,
module_llvm: ModuleLlvm::parse(cgcx, &name, buffer.data(), diag_handler)?,
name: name.into_string().unwrap(),
kind: ModuleKind::Regular,
}
}
@ -265,25 +273,13 @@ fn fat_lto(cgcx: &CodegenContext<LlvmCodegenBackend>,
// and we want to move everything to the same LLVM context. Currently the
// way we know of to do that is to serialize them to a string and them parse
// them later. Not great but hey, that's why it's "fat" LTO, right?
let mut new_modules = modules.into_iter().map(|module| {
match module {
FatLTOInput::InMemory(module) => {
for module in in_memory {
let buffer = ModuleBuffer::new(module.module_llvm.llmod());
let llmod_id = CString::new(&module.name[..]).unwrap();
(SerializedModule::Local(buffer), llmod_id)
serialized_modules.push((SerializedModule::Local(buffer), llmod_id));
}
FatLTOInput::Serialized { name, buffer } => {
let llmod_id = CString::new(name).unwrap();
(SerializedModule::Local(buffer), llmod_id)
}
}
}).collect::<Vec<_>>();
// Sort the modules to ensure we produce deterministic results.
new_modules.sort_by(|module1, module2| module1.1.partial_cmp(&module2.1).unwrap());
serialized_modules.extend(new_modules);
serialized_modules.extend(cached_modules.into_iter().map(|(buffer, wp)| {
(buffer, CString::new(wp.cgu_name).unwrap())
}));
serialized_modules.sort_by(|module1, module2| module1.1.cmp(&module2.1));
// For all serialized bitcode files we parse them and link them in as we did
// above, this is all mostly handled in C++. Like above, though, we don't
@ -850,7 +846,7 @@ fn module_name_to_str(c_str: &CStr) -> &str {
bug!("Encountered non-utf8 LLVM module name `{}`: {}", c_str.to_string_lossy(), e))
}
fn parse_module<'a>(
pub fn parse_module<'a>(
cx: &'a llvm::Context,
name: &CStr,
data: &[u8],

7
src/librustc_codegen_llvm/lib.rs
View File

@ -54,6 +54,7 @@
pub use llvm_util::target_features;
use std::any::Any;
use std::sync::{mpsc, Arc};
use std::ffi::CStr;
use rustc::dep_graph::DepGraph;
use rustc::middle::cstore::{EncodedMetadata, MetadataLoader};
@ -386,13 +387,13 @@ fn new_metadata(tcx: TyCtxt<'_>, mod_name: &str) -> Self {
fn parse(
cgcx: &CodegenContext<LlvmCodegenBackend>,
name: &str,
buffer: &back::lto::ModuleBuffer,
name: &CStr,
buffer: &[u8],
handler: &Handler,
) -> Result<Self, FatalError> {
unsafe {
let llcx = llvm::LLVMRustContextCreate(cgcx.fewer_names);
let llmod_raw = buffer.parse(name, llcx, handler)?;
let llmod_raw = back::lto::parse_module(llcx, name, buffer, handler)?;
let tm = match (cgcx.tm_factory.0)() {
Ok(m) => m,
Err(e) => {

12
src/test/run-make-fulldeps/lto-empty/Makefile Normal file
View File

@ -0,0 +1,12 @@
-include ../tools.mk
all: cdylib-fat cdylib-thin
cdylib-fat:
$(RUSTC) lib.rs -C lto=fat -C opt-level=3 -C incremental=$(TMPDIR)/inc-fat
$(RUSTC) lib.rs -C lto=fat -C opt-level=3 -C incremental=$(TMPDIR)/inc-fat
cdylib-thin:
$(RUSTC) lib.rs -C lto=thin -C opt-level=3 -C incremental=$(TMPDIR)/inc-thin
$(RUSTC) lib.rs -C lto=thin -C opt-level=3 -C incremental=$(TMPDIR)/inc-thin

1
src/test/run-make-fulldeps/lto-empty/lib.rs Normal file
View File

@ -0,0 +1 @@
#![crate_type = "cdylib"]

2
src/test/ui/lto-duplicate-symbols.stderr
View File

@ -1,6 +1,6 @@
warning: Linking globals named 'foo': symbol multiply defined!
error: failed to load bc of "lto_duplicate_symbols1.3a1fbbbh-cgu.0":
error: failed to load bc of "lto_duplicate_symbols2.3a1fbbbh-cgu.0":
error: aborting due to previous error