Auto merge of #55641 - nagisa:optimize-attr, r=pnkfelix

Implement optimize(size) and optimize(speed) attributes

This PR implements both `optimize(size)` and `optimize(speed)` attributes.

While the functionality itself works fine now, this PR is not yet complete: the code might be messy in places and, most importantly, the compiletest must be improved with functionality to run tests with custom optimization levels. Otherwise the new attribute cannot be tested properly. Oh, and not all of the RFC is implemented – attribute propagation is not implemented for example.

# TODO

* [x] Improve compiletest so that tests can be written;
* [x] Assign a proper error number (E9999 currently, no idea how to allocate a number properly);
* [ ] Perhaps reduce the duplication in LLVM attribute assignment code…
This commit is contained in:
bors 2019-01-26 07:08:18 +00:00
commit 42eb5ff404
32 changed files with 449 additions and 129 deletions

View File

@ -642,6 +642,7 @@ define_dep_nodes!( <'tcx>
[eval_always] CollectAndPartitionMonoItems,
[] IsCodegenedItem(DefId),
[] CodegenUnit(InternedString),
[] BackendOptimizationLevel(CrateNum),
[] CompileCodegenUnit(InternedString),
[input] OutputFilenames,
[] NormalizeProjectionTy(CanonicalProjectionGoal<'tcx>),

View File

@ -21,7 +21,7 @@ use syntax::source_map::Spanned;
use rustc_target::spec::abi::Abi;
use syntax::ast::{self, CrateSugar, Ident, Name, NodeId, DUMMY_NODE_ID, AsmDialect};
use syntax::ast::{Attribute, Label, Lit, StrStyle, FloatTy, IntTy, UintTy};
use syntax::attr::InlineAttr;
use syntax::attr::{InlineAttr, OptimizeAttr};
use syntax::ext::hygiene::SyntaxContext;
use syntax::ptr::P;
use syntax::symbol::{Symbol, keywords};
@ -2416,6 +2416,8 @@ pub struct CodegenFnAttrs {
pub flags: CodegenFnAttrFlags,
/// Parsed representation of the `#[inline]` attribute
pub inline: InlineAttr,
/// Parsed representation of the `#[optimize]` attribute
pub optimize: OptimizeAttr,
/// The `#[export_name = "..."]` attribute, indicating a custom symbol a
/// function should be exported under
pub export_name: Option<Symbol>,
@ -2476,6 +2478,7 @@ impl CodegenFnAttrs {
CodegenFnAttrs {
flags: CodegenFnAttrFlags::empty(),
inline: InlineAttr::None,
optimize: OptimizeAttr::None,
export_name: None,
link_name: None,
target_features: vec![],

View File

@ -1159,6 +1159,7 @@ impl<'a> ToStableHashKey<StableHashingContext<'a>> for hir::TraitCandidate {
impl_stable_hash_for!(struct hir::CodegenFnAttrs {
flags,
inline,
optimize,
export_name,
link_name,
target_features,
@ -1183,6 +1184,14 @@ impl<'hir> HashStable<StableHashingContext<'hir>> for attr::InlineAttr {
}
}
impl<'hir> HashStable<StableHashingContext<'hir>> for attr::OptimizeAttr {
fn hash_stable<W: StableHasherResult>(&self,
hcx: &mut StableHashingContext<'hir>,
hasher: &mut StableHasher<W>) {
mem::discriminant(self).hash_stable(hcx, hasher);
}
}
impl_stable_hash_for!(struct hir::Freevar {
def,
span

View File

@ -58,6 +58,8 @@ pub enum OptLevel {
SizeMin, // -Oz
}
impl_stable_hash_via_hash!(OptLevel);
/// This is what the `LtoCli` values get mapped to after resolving defaults and
/// and taking other command line options into account.
#[derive(Clone, Copy, PartialEq, Hash, Debug)]

View File

@ -967,6 +967,12 @@ impl<'tcx> QueryDescription<'tcx> for queries::dllimport_foreign_items<'tcx> {
}
}
impl<'tcx> QueryDescription<'tcx> for queries::backend_optimization_level<'tcx> {
fn describe(_tcx: TyCtxt<'_, '_, '_>, _: CrateNum) -> Cow<'static, str> {
"optimization level used by backend".into()
}
}
macro_rules! impl_disk_cacheable_query(
($query_name:ident, |$key:tt| $cond:expr) => {
impl<'tcx> QueryDescription<'tcx> for queries::$query_name<'tcx> {

View File

@ -22,7 +22,7 @@ use mir::mono::CodegenUnit;
use mir;
use mir::interpret::GlobalId;
use session::{CompileResult, CrateDisambiguator};
use session::config::{EntryFnType, OutputFilenames};
use session::config::{EntryFnType, OutputFilenames, OptLevel};
use traits::{self, Vtable};
use traits::query::{
CanonicalPredicateGoal, CanonicalProjectionGoal,
@ -573,6 +573,7 @@ define_queries! { <'tcx>
-> (Arc<DefIdSet>, Arc<Vec<Arc<CodegenUnit<'tcx>>>>),
[] fn is_codegened_item: IsCodegenedItem(DefId) -> bool,
[] fn codegen_unit: CodegenUnit(InternedString) -> Arc<CodegenUnit<'tcx>>,
[] fn backend_optimization_level: BackendOptimizationLevel(CrateNum) -> OptLevel,
},
Other {

View File

@ -1410,6 +1410,9 @@ pub fn force_from_dep_node<'a, 'gcx, 'lcx>(tcx: TyCtxt<'a, 'gcx, 'lcx>,
DepKind::UpstreamMonomorphizationsFor => {
force!(upstream_monomorphizations_for, def_id!());
}
DepKind::BackendOptimizationLevel => {
force!(backend_optimization_level, krate!());
}
}
true

View File

@ -5,7 +5,7 @@ use std::ffi::CString;
use rustc::hir::{CodegenFnAttrFlags, CodegenFnAttrs};
use rustc::hir::def_id::{DefId, LOCAL_CRATE};
use rustc::session::Session;
use rustc::session::config::Sanitizer;
use rustc::session::config::{Sanitizer, OptLevel};
use rustc::ty::{self, TyCtxt, PolyFnSig};
use rustc::ty::layout::HasTyCtxt;
use rustc::ty::query::Providers;
@ -20,7 +20,7 @@ use attributes;
use llvm::{self, Attribute};
use llvm::AttributePlace::Function;
use llvm_util;
pub use syntax::attr::{self, InlineAttr};
pub use syntax::attr::{self, InlineAttr, OptimizeAttr};
use context::CodegenCx;
use value::Value;
@ -57,13 +57,6 @@ fn unwind(val: &'ll Value, can_unwind: bool) {
Attribute::NoUnwind.toggle_llfn(Function, val, !can_unwind);
}
/// Tell LLVM whether it should optimize function for size.
#[inline]
#[allow(dead_code)] // possibly useful function
pub fn set_optimize_for_size(val: &'ll Value, optimize: bool) {
Attribute::OptimizeForSize.toggle_llfn(Function, val, optimize);
}
/// Tell LLVM if this function should be 'naked', i.e., skip the epilogue and prologue.
#[inline]
pub fn naked(val: &'ll Value, is_naked: bool) {
@ -151,6 +144,28 @@ pub fn non_lazy_bind(sess: &Session, llfn: &'ll Value) {
}
}
pub(crate) fn default_optimisation_attrs(sess: &Session, llfn: &'ll Value) {
match sess.opts.optimize {
OptLevel::Size => {
llvm::Attribute::MinSize.unapply_llfn(Function, llfn);
llvm::Attribute::OptimizeForSize.apply_llfn(Function, llfn);
llvm::Attribute::OptimizeNone.unapply_llfn(Function, llfn);
},
OptLevel::SizeMin => {
llvm::Attribute::MinSize.apply_llfn(Function, llfn);
llvm::Attribute::OptimizeForSize.apply_llfn(Function, llfn);
llvm::Attribute::OptimizeNone.unapply_llfn(Function, llfn);
}
OptLevel::No => {
llvm::Attribute::MinSize.unapply_llfn(Function, llfn);
llvm::Attribute::OptimizeForSize.unapply_llfn(Function, llfn);
llvm::Attribute::OptimizeNone.unapply_llfn(Function, llfn);
}
_ => {}
}
}
/// Composite function which sets LLVM attributes for function depending on its AST (`#[attribute]`)
/// attributes.
pub fn from_fn_attrs(
@ -162,6 +177,22 @@ pub fn from_fn_attrs(
let codegen_fn_attrs = id.map(|id| cx.tcx.codegen_fn_attrs(id))
.unwrap_or_else(|| CodegenFnAttrs::new());
match codegen_fn_attrs.optimize {
OptimizeAttr::None => {
default_optimisation_attrs(cx.tcx.sess, llfn);
}
OptimizeAttr::Speed => {
llvm::Attribute::MinSize.unapply_llfn(Function, llfn);
llvm::Attribute::OptimizeForSize.unapply_llfn(Function, llfn);
llvm::Attribute::OptimizeNone.unapply_llfn(Function, llfn);
}
OptimizeAttr::Size => {
llvm::Attribute::MinSize.apply_llfn(Function, llfn);
llvm::Attribute::OptimizeForSize.apply_llfn(Function, llfn);
llvm::Attribute::OptimizeNone.unapply_llfn(Function, llfn);
}
}
inline(cx, llfn, codegen_fn_attrs.inline);
// The `uwtable` attribute according to LLVM is:

View File

@ -3,7 +3,7 @@ use rustc_codegen_ssa::back::symbol_export;
use rustc_codegen_ssa::back::write::{ModuleConfig, CodegenContext, pre_lto_bitcode_filename};
use rustc_codegen_ssa::back::lto::{SerializedModule, LtoModuleCodegen, ThinShared, ThinModule};
use rustc_codegen_ssa::traits::*;
use back::write::{self, DiagnosticHandlers, with_llvm_pmb, save_temp_bitcode, get_llvm_opt_level};
use back::write::{self, DiagnosticHandlers, with_llvm_pmb, save_temp_bitcode, to_llvm_opt_settings};
use errors::{FatalError, Handler};
use llvm::archive_ro::ArchiveRO;
use llvm::{self, True, False};
@ -532,7 +532,7 @@ pub(crate) fn run_pass_manager(cgcx: &CodegenContext<LlvmCodegenBackend>,
// Note that in general this shouldn't matter too much as you typically
// only turn on ThinLTO when you're compiling with optimizations
// otherwise.
let opt_level = config.opt_level.map(get_llvm_opt_level)
let opt_level = config.opt_level.map(|x| to_llvm_opt_settings(x).0)
.unwrap_or(llvm::CodeGenOptLevel::None);
let opt_level = match opt_level {
llvm::CodeGenOptLevel::None => llvm::CodeGenOptLevel::Less,

View File

@ -5,8 +5,10 @@ use rustc_codegen_ssa::back::write::{CodegenContext, ModuleConfig, run_assembler
use rustc_codegen_ssa::traits::*;
use base;
use consts;
use rustc::hir::def_id::LOCAL_CRATE;
use rustc::session::config::{self, OutputType, Passes, Lto};
use rustc::session::Session;
use rustc::ty::TyCtxt;
use time_graph::Timeline;
use llvm::{self, DiagnosticInfo, PassManager, SMDiagnostic};
use llvm_util;
@ -81,42 +83,46 @@ pub fn write_output_file(
}
}
pub(crate) fn get_llvm_opt_level(optimize: config::OptLevel) -> llvm::CodeGenOptLevel {
match optimize {
config::OptLevel::No => llvm::CodeGenOptLevel::None,
config::OptLevel::Less => llvm::CodeGenOptLevel::Less,
config::OptLevel::Default => llvm::CodeGenOptLevel::Default,
config::OptLevel::Aggressive => llvm::CodeGenOptLevel::Aggressive,
_ => llvm::CodeGenOptLevel::Default,
}
}
pub(crate) fn get_llvm_opt_size(optimize: config::OptLevel) -> llvm::CodeGenOptSize {
match optimize {
config::OptLevel::Size => llvm::CodeGenOptSizeDefault,
config::OptLevel::SizeMin => llvm::CodeGenOptSizeAggressive,
_ => llvm::CodeGenOptSizeNone,
}
}
pub fn create_target_machine(
tcx: TyCtxt,
find_features: bool,
) -> &'static mut llvm::TargetMachine {
target_machine_factory(tcx.sess, tcx.backend_optimization_level(LOCAL_CRATE), find_features)()
.unwrap_or_else(|err| llvm_err(tcx.sess.diagnostic(), &err).raise() )
}
pub fn create_informational_target_machine(
sess: &Session,
find_features: bool,
) -> &'static mut llvm::TargetMachine {
target_machine_factory(sess, find_features)().unwrap_or_else(|err| {
target_machine_factory(sess, config::OptLevel::No, find_features)().unwrap_or_else(|err| {
llvm_err(sess.diagnostic(), &err).raise()
})
}
pub fn to_llvm_opt_settings(cfg: config::OptLevel) -> (llvm::CodeGenOptLevel, llvm::CodeGenOptSize)
{
use self::config::OptLevel::*;
match cfg {
No => (llvm::CodeGenOptLevel::None, llvm::CodeGenOptSizeNone),
Less => (llvm::CodeGenOptLevel::Less, llvm::CodeGenOptSizeNone),
Default => (llvm::CodeGenOptLevel::Default, llvm::CodeGenOptSizeNone),
Aggressive => (llvm::CodeGenOptLevel::Aggressive, llvm::CodeGenOptSizeNone),
Size => (llvm::CodeGenOptLevel::Default, llvm::CodeGenOptSizeDefault),
SizeMin => (llvm::CodeGenOptLevel::Default, llvm::CodeGenOptSizeAggressive),
}
}
// If find_features is true this won't access `sess.crate_types` by assuming
// that `is_pie_binary` is false. When we discover LLVM target features
// `sess.crate_types` is uninitialized so we cannot access it.
pub fn target_machine_factory(sess: &Session, find_features: bool)
pub fn target_machine_factory(sess: &Session, optlvl: config::OptLevel, find_features: bool)
-> Arc<dyn Fn() -> Result<&'static mut llvm::TargetMachine, String> + Send + Sync>
{
let reloc_model = get_reloc_model(sess);
let opt_level = get_llvm_opt_level(sess.opts.optimize);
let (opt_level, _) = to_llvm_opt_settings(optlvl);
let use_softfp = sess.opts.cg.soft_float;
let ffunction_sections = sess.target.target.options.function_sections;
@ -357,7 +363,7 @@ pub(crate) unsafe fn optimize(cgcx: &CodegenContext<LlvmCodegenBackend>,
if !config.no_prepopulate_passes {
llvm::LLVMRustAddAnalysisPasses(tm, fpm, llmod);
llvm::LLVMRustAddAnalysisPasses(tm, mpm, llmod);
let opt_level = config.opt_level.map(get_llvm_opt_level)
let opt_level = config.opt_level.map(|x| to_llvm_opt_settings(x).0)
.unwrap_or(llvm::CodeGenOptLevel::None);
let prepare_for_thin_lto = cgcx.lto == Lto::Thin || cgcx.lto == Lto::ThinLocal ||
(cgcx.lto != Lto::Fat && cgcx.opts.debugging_opts.cross_lang_lto.enabled());
@ -689,7 +695,8 @@ pub unsafe fn with_llvm_pmb(llmod: &llvm::Module,
// reasonable defaults and prepare it to actually populate the pass
// manager.
let builder = llvm::LLVMPassManagerBuilderCreate();
let opt_size = config.opt_size.map(get_llvm_opt_size).unwrap_or(llvm::CodeGenOptSizeNone);
let opt_size = config.opt_size.map(|x| to_llvm_opt_settings(x).1)
.unwrap_or(llvm::CodeGenOptSizeNone);
let inline_threshold = config.inline_threshold;
let pgo_gen_path = config.pgo_gen.as_ref().map(|s| {

View File

@ -136,7 +136,7 @@ pub fn iter_globals(llmod: &'ll llvm::Module) -> ValueIter<'ll> {
}
}
pub fn compile_codegen_unit<'ll, 'tcx>(tcx: TyCtxt<'ll, 'tcx, 'tcx>,
pub fn compile_codegen_unit<'a, 'tcx>(tcx: TyCtxt<'a, 'tcx, 'tcx>,
cgu_name: InternedString)
-> Stats {
let start_time = Instant::now();
@ -164,7 +164,7 @@ pub fn compile_codegen_unit<'ll, 'tcx>(tcx: TyCtxt<'ll, 'tcx, 'tcx>,
let backend = LlvmCodegenBackend(());
let cgu = tcx.codegen_unit(cgu_name);
// Instantiate monomorphizations without filling out definitions yet...
let llvm_module = backend.new_metadata(tcx.sess, &cgu_name.as_str());
let llvm_module = backend.new_metadata(tcx, &cgu_name.as_str());
let stats = {
let cx = CodegenCx::new(tcx, cgu, &llvm_module);
let mono_items = cx.codegen_unit

View File

@ -144,16 +144,17 @@ pub fn is_pie_binary(sess: &Session) -> bool {
}
pub unsafe fn create_module(
sess: &Session,
tcx: TyCtxt,
llcx: &'ll llvm::Context,
mod_name: &str,
) -> &'ll llvm::Module {
let sess = tcx.sess;
let mod_name = SmallCStr::new(mod_name);
let llmod = llvm::LLVMModuleCreateWithNameInContext(mod_name.as_ptr(), llcx);
// Ensure the data-layout values hardcoded remain the defaults.
if sess.target.target.options.is_builtin {
let tm = ::back::write::create_target_machine(sess, false);
let tm = ::back::write::create_target_machine(tcx, false);
llvm::LLVMRustSetDataLayoutFromTargetMachine(llmod, tm);
llvm::LLVMRustDisposeTargetMachine(tm);

View File

@ -65,19 +65,8 @@ fn declare_raw_fn(
}
}
match cx.tcx.sess.opts.cg.opt_level.as_ref().map(String::as_ref) {
Some("s") => {
llvm::Attribute::OptimizeForSize.apply_llfn(Function, llfn);
},
Some("z") => {
llvm::Attribute::MinSize.apply_llfn(Function, llfn);
llvm::Attribute::OptimizeForSize.apply_llfn(Function, llfn);
},
_ => {},
}
attributes::default_optimisation_attrs(cx.tcx.sess, llfn);
attributes::non_lazy_bind(cx.sess(), llfn);
llfn
}

View File

@ -72,7 +72,7 @@ use rustc::dep_graph::DepGraph;
use rustc::middle::allocator::AllocatorKind;
use rustc::middle::cstore::{EncodedMetadata, MetadataLoader};
use rustc::session::{Session, CompileIncomplete};
use rustc::session::config::{OutputFilenames, OutputType, PrintRequest};
use rustc::session::config::{OutputFilenames, OutputType, PrintRequest, OptLevel};
use rustc::ty::{self, TyCtxt};
use rustc::util::time_graph;
use rustc::util::profiling::ProfileCategory;
@ -121,8 +121,8 @@ mod va_arg;
pub struct LlvmCodegenBackend(());
impl ExtraBackendMethods for LlvmCodegenBackend {
fn new_metadata(&self, sess: &Session, mod_name: &str) -> ModuleLlvm {
ModuleLlvm::new(sess, mod_name)
fn new_metadata(&self, tcx: TyCtxt, mod_name: &str) -> ModuleLlvm {
ModuleLlvm::new(tcx, mod_name)
}
fn write_metadata<'b, 'gcx>(
&self,
@ -144,10 +144,11 @@ impl ExtraBackendMethods for LlvmCodegenBackend {
fn target_machine_factory(
&self,
sess: &Session,
optlvl: OptLevel,
find_features: bool
) -> Arc<dyn Fn() ->
Result<&'static mut llvm::TargetMachine, String> + Send + Sync> {
back::write::target_machine_factory(sess, find_features)
back::write::target_machine_factory(sess, optlvl, find_features)
}
fn target_cpu<'b>(&self, sess: &'b Session) -> &'b str {
llvm_util::target_cpu(sess)
@ -363,15 +364,15 @@ unsafe impl Send for ModuleLlvm { }
unsafe impl Sync for ModuleLlvm { }
impl ModuleLlvm {
fn new(sess: &Session, mod_name: &str) -> Self {
fn new(tcx: TyCtxt, mod_name: &str) -> Self {
unsafe {
let llcx = llvm::LLVMRustContextCreate(sess.fewer_names());
let llmod_raw = context::create_module(sess, llcx, mod_name) as *const _;
let llcx = llvm::LLVMRustContextCreate(tcx.sess.fewer_names());
let llmod_raw = context::create_module(tcx, llcx, mod_name) as *const _;
ModuleLlvm {
llmod_raw,
llcx,
tm: create_target_machine(sess, false),
tm: create_target_machine(tcx, false),
}
}
}

View File

@ -115,6 +115,7 @@ pub enum Attribute {
SanitizeAddress = 21,
SanitizeMemory = 22,
NonLazyBind = 23,
OptimizeNone = 24,
}
/// LLVMIntPredicate

View File

@ -1,5 +1,5 @@
use syntax_pos::symbol::Symbol;
use back::write::create_target_machine;
use back::write::create_informational_target_machine;
use llvm;
use rustc::session::Session;
use rustc::session::config::PrintRequest;
@ -223,7 +223,7 @@ pub fn to_llvm_feature<'a>(sess: &Session, s: &'a str) -> &'a str {
}
pub fn target_features(sess: &Session) -> Vec<Symbol> {
let target_machine = create_target_machine(sess, true);
let target_machine = create_informational_target_machine(sess, true);
target_feature_whitelist(sess)
.iter()
.filter_map(|&(feature, gate)| {
@ -276,7 +276,7 @@ pub fn print_passes() {
pub(crate) fn print(req: PrintRequest, sess: &Session) {
require_inited();
let tm = create_target_machine(sess, true);
let tm = create_informational_target_machine(sess, true);
unsafe {
match req {
PrintRequest::TargetCPUs => llvm::LLVMRustPrintTargetCPUs(tm),

View File

@ -982,6 +982,7 @@ fn start_executing_work<B: ExtraBackendMethods>(
None
};
let ol = tcx.backend_optimization_level(LOCAL_CRATE);
let cgcx = CodegenContext::<B> {
backend: backend.clone(),
crate_types: sess.crate_types.borrow().clone(),
@ -1005,7 +1006,7 @@ fn start_executing_work<B: ExtraBackendMethods>(
regular_module_config: modules_config,
metadata_module_config: metadata_config,
allocator_module_config: allocator_config,
tm_factory: TargetMachineFactory(backend.target_machine_factory(tcx.sess, false)),
tm_factory: TargetMachineFactory(backend.target_machine_factory(tcx.sess, ol, false)),
total_cgus,
msvc_imps_needed: msvc_imps_needed(tcx),
target_pointer_width: tcx.sess.target.target.target_pointer_width.clone(),

View File

@ -551,7 +551,7 @@ pub fn codegen_crate<B: ExtraBackendMethods>(
&["crate"],
Some("metadata")).as_str()
.to_string();
let metadata_llvm_module = backend.new_metadata(tcx.sess, &metadata_cgu_name);
let metadata_llvm_module = backend.new_metadata(tcx, &metadata_cgu_name);
let metadata = time(tcx.sess, "write metadata", || {
backend.write_metadata(tcx, &metadata_llvm_module)
});
@ -636,7 +636,7 @@ pub fn codegen_crate<B: ExtraBackendMethods>(
&["crate"],
Some("allocator")).as_str()
.to_string();
let modules = backend.new_metadata(tcx.sess, &llmod_id);
let modules = backend.new_metadata(tcx, &llmod_id);
time(tcx.sess, "write allocator module", || {
backend.codegen_allocator(tcx, &modules, kind)
});
@ -897,6 +897,39 @@ fn is_codegened_item(tcx: TyCtxt, id: DefId) -> bool {
}
pub fn provide_both(providers: &mut Providers) {
providers.backend_optimization_level = |tcx, cratenum| {
let for_speed = match tcx.sess.opts.optimize {
// If globally no optimisation is done, #[optimize] has no effect.
//
// This is done because if we ended up "upgrading" to `-O2` here, wed populate the
// pass manager and it is likely that some module-wide passes (such as inliner or
// cross-function constant propagation) would ignore the `optnone` annotation we put
// on the functions, thus necessarily involving these functions into optimisations.
config::OptLevel::No => return config::OptLevel::No,
// If globally optimise-speed is already specified, just use that level.
config::OptLevel::Less => return config::OptLevel::Less,
config::OptLevel::Default => return config::OptLevel::Default,
config::OptLevel::Aggressive => return config::OptLevel::Aggressive,
// If globally optimize-for-size has been requested, use -O2 instead (if optimize(size)
// are present).
config::OptLevel::Size => config::OptLevel::Default,
config::OptLevel::SizeMin => config::OptLevel::Default,
};
let (defids, _) = tcx.collect_and_partition_mono_items(cratenum);
for id in &*defids {
let hir::CodegenFnAttrs { optimize, .. } = tcx.codegen_fn_attrs(*id);
match optimize {
attr::OptimizeAttr::None => continue,
attr::OptimizeAttr::Size => continue,
attr::OptimizeAttr::Speed => {
return for_speed;
}
}
}
return tcx.sess.opts.optimize;
};
providers.dllimport_foreign_items = |tcx, krate| {
let module_map = tcx.foreign_modules(krate);
let module_map = module_map.iter()

View File

@ -6,7 +6,7 @@ use super::CodegenObject;
use rustc::middle::allocator::AllocatorKind;
use rustc::middle::cstore::EncodedMetadata;
use rustc::mir::mono::Stats;
use rustc::session::Session;
use rustc::session::{Session, config};
use rustc::ty::TyCtxt;
use rustc_codegen_utils::codegen_backend::CodegenBackend;
use std::sync::Arc;
@ -32,7 +32,7 @@ impl<'tcx, T> Backend<'tcx> for T where
}
pub trait ExtraBackendMethods: CodegenBackend + WriteBackendMethods + Sized + Send {
fn new_metadata(&self, sess: &Session, mod_name: &str) -> Self::Module;
fn new_metadata(&self, sess: TyCtxt, mod_name: &str) -> Self::Module;
fn write_metadata<'b, 'gcx>(
&self,
tcx: TyCtxt<'b, 'gcx, 'gcx>,
@ -50,6 +50,7 @@ pub trait ExtraBackendMethods: CodegenBackend + WriteBackendMethods + Sized + Se
fn target_machine_factory(
&self,
sess: &Session,
opt_level: config::OptLevel,
find_features: bool,
) -> Arc<dyn Fn() -> Result<Self::TargetMachine, String> + Send + Sync>;
fn target_cpu<'b>(&self, sess: &'b Session) -> &'b str;

View File

@ -36,7 +36,7 @@ use rustc_target::spec::abi;
use syntax::ast;
use syntax::ast::{Ident, MetaItemKind};
use syntax::attr::{InlineAttr, list_contains_name, mark_used};
use syntax::attr::{InlineAttr, OptimizeAttr, list_contains_name, mark_used};
use syntax::source_map::Spanned;
use syntax::feature_gate;
use syntax::symbol::{keywords, Symbol};
@ -2286,49 +2286,6 @@ fn codegen_fn_attrs<'a, 'tcx>(tcx: TyCtxt<'a, 'tcx, 'tcx>, id: DefId) -> Codegen
codegen_fn_attrs.flags |= CodegenFnAttrFlags::USED;
} else if attr.check_name("thread_local") {
codegen_fn_attrs.flags |= CodegenFnAttrFlags::THREAD_LOCAL;
} else if attr.check_name("inline") {
codegen_fn_attrs.inline = attrs.iter().fold(InlineAttr::None, |ia, attr| {
if attr.path != "inline" {
return ia;
}
let meta = match attr.meta() {
Some(meta) => meta.node,
None => return ia,
};
match meta {
MetaItemKind::Word => {
mark_used(attr);
InlineAttr::Hint
}
MetaItemKind::List(ref items) => {
mark_used(attr);
inline_span = Some(attr.span);
if items.len() != 1 {
span_err!(
tcx.sess.diagnostic(),
attr.span,
E0534,
"expected one argument"
);
InlineAttr::None
} else if list_contains_name(&items[..], "always") {
InlineAttr::Always
} else if list_contains_name(&items[..], "never") {
InlineAttr::Never
} else {
span_err!(
tcx.sess.diagnostic(),
items[0].span,
E0535,
"invalid argument"
);
InlineAttr::None
}
}
_ => ia,
}
});
} else if attr.check_name("export_name") {
if let Some(s) = attr.value_str() {
if s.as_str().contains("\0") {
@ -2378,6 +2335,76 @@ fn codegen_fn_attrs<'a, 'tcx>(tcx: TyCtxt<'a, 'tcx, 'tcx>, id: DefId) -> Codegen
}
}
codegen_fn_attrs.inline = attrs.iter().fold(InlineAttr::None, |ia, attr| {
if attr.path != "inline" {
return ia;
}
match attr.meta().map(|i| i.node) {
Some(MetaItemKind::Word) => {
mark_used(attr);
InlineAttr::Hint
}
Some(MetaItemKind::List(ref items)) => {
mark_used(attr);
inline_span = Some(attr.span);
if items.len() != 1 {
span_err!(
tcx.sess.diagnostic(),
attr.span,
E0534,
"expected one argument"
);
InlineAttr::None
} else if list_contains_name(&items[..], "always") {
InlineAttr::Always
} else if list_contains_name(&items[..], "never") {
InlineAttr::Never
} else {
span_err!(
tcx.sess.diagnostic(),
items[0].span,
E0535,
"invalid argument"
);
InlineAttr::None
}
}
Some(MetaItemKind::NameValue(_)) => ia,
None => ia,
}
});
codegen_fn_attrs.optimize = attrs.iter().fold(OptimizeAttr::None, |ia, attr| {
if attr.path != "optimize" {
return ia;
}
let err = |sp, s| span_err!(tcx.sess.diagnostic(), sp, E0722, "{}", s);
match attr.meta().map(|i| i.node) {
Some(MetaItemKind::Word) => {
err(attr.span, "expected one argument");
ia
}
Some(MetaItemKind::List(ref items)) => {
mark_used(attr);
inline_span = Some(attr.span);
if items.len() != 1 {
err(attr.span, "expected one argument");
OptimizeAttr::None
} else if list_contains_name(&items[..], "size") {
OptimizeAttr::Size
} else if list_contains_name(&items[..], "speed") {
OptimizeAttr::Speed
} else {
err(items[0].span, "invalid argument");
OptimizeAttr::None
}
}
Some(MetaItemKind::NameValue(_)) => ia,
None => ia,
}
});
// If a function uses #[target_feature] it can't be inlined into general
// purpose functions as they wouldn't have the right target features
// enabled. For that reason we also forbid #[inline(always)] as it can't be

View File

@ -4719,4 +4719,5 @@ register_diagnostics! {
E0645, // trait aliases not finished
E0698, // type inside generator must be known in this context
E0719, // duplicate values for associated type binding
E0722, // Malformed #[optimize] attribute
}

View File

@ -63,6 +63,13 @@ pub enum InlineAttr {
Never,
}
#[derive(Copy, Clone, Hash, PartialEq, RustcEncodable, RustcDecodable)]
pub enum OptimizeAttr {
None,
Speed,
Size,
}
#[derive(Copy, Clone, PartialEq)]
pub enum UnwindAttr {
Allowed,

View File

@ -4,8 +4,8 @@ mod builtin;
pub use self::builtin::{
cfg_matches, contains_feature_attr, eval_condition, find_crate_name, find_deprecation,
find_repr_attrs, find_stability, find_unwind_attr, Deprecation, InlineAttr, IntType, ReprAttr,
RustcDeprecation, Stability, StabilityLevel, UnwindAttr,
find_repr_attrs, find_stability, find_unwind_attr, Deprecation, InlineAttr, OptimizeAttr,
IntType, ReprAttr, RustcDeprecation, Stability, StabilityLevel, UnwindAttr,
};
pub use self::IntType::*;
pub use self::ReprAttr::*;

View File

@ -461,6 +461,9 @@ declare_features! (
// Re-Rebalance coherence
(active, re_rebalance_coherence, "1.32.0", Some(55437), None),
// #[optimize(X)]
(active, optimize_attribute, "1.34.0", Some(54882), None),
);
declare_features! (
@ -1222,6 +1225,12 @@ pub const BUILTIN_ATTRIBUTES: &[(&str, AttributeType, AttributeTemplate, Attribu
"#[alloc_error_handler] is an unstable feature",
cfg_fn!(alloc_error_handler))),
// RFC 2412
("optimize", Whitelisted, template!(List: "size|speed"), Gated(Stability::Unstable,
"optimize_attribute",
"#[optimize] attribute is an unstable feature",
cfg_fn!(optimize_attribute))),
// Crate level attributes
("crate_name", CrateLevel, template!(NameValueStr: "name"), Ungated),
("crate_type", CrateLevel, template!(NameValueStr: "bin|lib|..."), Ungated),

View File

@ -188,6 +188,8 @@ static Attribute::AttrKind fromRust(LLVMRustAttribute Kind) {
return Attribute::SanitizeMemory;
case NonLazyBind:
return Attribute::NonLazyBind;
case OptimizeNone:
return Attribute::OptimizeNone;
}
report_fatal_error("bad AttributeKind");
}

View File

@ -84,6 +84,7 @@ enum LLVMRustAttribute {
SanitizeAddress = 21,
SanitizeMemory = 22,
NonLazyBind = 23,
OptimizeNone = 24,
};
typedef struct OpaqueRustString *RustStringRef;

View File

@ -0,0 +1,21 @@
// revisions: NO-OPT SIZE-OPT SPEED-OPT
//[NO-OPT] compile-flags: -Copt-level=0
//[SIZE-OPT] compile-flags: -Copt-level=s
//[SPEED-OPT] compile-flags: -Copt-level=3
#![crate_type="rlib"]
#[no_mangle]
#[inline(always)]
pub extern "C" fn callee() -> u32 {
4 + 4
}
// CHECK-LABEL: caller
// SIZE-OPT: ret i32 8
// SPEED-OPT: ret i32 8
// NO-OPT: ret i32 8
#[no_mangle]
pub extern "C" fn caller() -> u32 {
callee()
}

View File

@ -0,0 +1,50 @@
// revisions: NO-OPT SIZE-OPT SPEED-OPT
//[NO-OPT] compile-flags: -Copt-level=0 -Ccodegen-units=1
//[SIZE-OPT] compile-flags: -Copt-level=s -Ccodegen-units=1
//[SPEED-OPT] compile-flags: -Copt-level=3 -Ccodegen-units=1
#![feature(optimize_attribute)]
#![crate_type="rlib"]
// CHECK-LABEL: define i32 @nothing
// CHECK-SAME: [[NOTHING_ATTRS:#[0-9]+]]
// NO-OPT: ret i32 %1
// SIZE-OPT: ret i32 4
// SPEEC-OPT: ret i32 4
#[no_mangle]
pub fn nothing() -> i32 {
2 + 2
}
// CHECK-LABEL: define i32 @size
// CHECK-SAME: [[SIZE_ATTRS:#[0-9]+]]
// NO-OPT: ret i32 %1
// SIZE-OPT: ret i32 6
// SPEED-OPT: ret i32 6
#[optimize(size)]
#[no_mangle]
pub fn size() -> i32 {
3 + 3
}
// CHECK-LABEL: define i32 @speed
// NO-OPT-SAME: [[NOTHING_ATTRS]]
// SPEED-OPT-SAME: [[NOTHING_ATTRS]]
// SIZE-OPT-SAME: [[SPEED_ATTRS:#[0-9]+]]
// NO-OPT: ret i32 %1
// SIZE-OPT: ret i32 8
// SPEED-OPT: ret i32 8
#[optimize(speed)]
#[no_mangle]
pub fn speed() -> i32 {
4 + 4
}
// NO-OPT-DAG: attributes [[SIZE_ATTRS]] = {{.*}}minsize{{.*}}optsize{{.*}}
// SPEED-OPT-DAG: attributes [[SIZE_ATTRS]] = {{.*}}minsize{{.*}}optsize{{.*}}
// SIZE-OPT-DAG: attributes [[NOTHING_ATTRS]] = {{.*}}optsize{{.*}}
// SIZE-OPT-DAG: attributes [[SIZE_ATTRS]] = {{.*}}minsize{{.*}}optsize{{.*}}
// SIZE-OPT: attributes [[SPEED_ATTRS]]
// SIZE-OPT-NOT: minsize
// SIZE-OPT-NOT: optsize

View File

@ -0,0 +1,18 @@
#![crate_type="rlib"]
#![optimize(speed)] //~ ERROR #54882
#[optimize(size)] //~ ERROR #54882
mod module {
#[optimize(size)] //~ ERROR #54882
fn size() {}
#[optimize(speed)] //~ ERROR #54882
fn speed() {}
#[optimize(banana)]
//~^ ERROR #54882
//~| ERROR E0722
fn not_known() {}
}

View File

@ -0,0 +1,50 @@
error[E0658]: #[optimize] attribute is an unstable feature (see issue #54882)
--> $DIR/feature-gate-optimize_attribute.rs:7:1
|
LL | #[optimize(size)] //~ ERROR #54882
| ^^^^^^^^^^^^^^^^^
|
= help: add #![feature(optimize_attribute)] to the crate attributes to enable
error[E0658]: #[optimize] attribute is an unstable feature (see issue #54882)
--> $DIR/feature-gate-optimize_attribute.rs:10:1
|
LL | #[optimize(speed)] //~ ERROR #54882
| ^^^^^^^^^^^^^^^^^^
|
= help: add #![feature(optimize_attribute)] to the crate attributes to enable
error[E0658]: #[optimize] attribute is an unstable feature (see issue #54882)
--> $DIR/feature-gate-optimize_attribute.rs:13:1
|
LL | #[optimize(banana)]
| ^^^^^^^^^^^^^^^^^^^
|
= help: add #![feature(optimize_attribute)] to the crate attributes to enable
error[E0658]: #[optimize] attribute is an unstable feature (see issue #54882)
--> $DIR/feature-gate-optimize_attribute.rs:4:1
|
LL | #[optimize(size)] //~ ERROR #54882
| ^^^^^^^^^^^^^^^^^
|
= help: add #![feature(optimize_attribute)] to the crate attributes to enable
error[E0658]: #[optimize] attribute is an unstable feature (see issue #54882)
--> $DIR/feature-gate-optimize_attribute.rs:2:1
|
LL | #![optimize(speed)] //~ ERROR #54882
| ^^^^^^^^^^^^^^^^^^^
|
= help: add #![feature(optimize_attribute)] to the crate attributes to enable
error[E0722]: invalid argument
--> $DIR/feature-gate-optimize_attribute.rs:13:12
|
LL | #[optimize(banana)]
| ^^^^^^
error: aborting due to 6 previous errors
Some errors occurred: E0658, E0722.
For more information about an error, try `rustc --explain E0658`.

View File

@ -542,6 +542,8 @@ fn iter_header(testfile: &Path, cfg: Option<&str>, it: &mut dyn FnMut(&str)) {
"#"
};
// FIXME: would be nice to allow some whitespace between comment and brace :)
// It took me like 2 days to debug why compile-flags werent taken into account for my test :)
let comment_with_brace = comment.to_string() + "[";
let rdr = BufReader::new(File::open(testfile).unwrap());

View File

@ -552,9 +552,10 @@ impl<'test> TestCx<'test> {
.args(&["--target", &self.config.target])
.arg("-L")
.arg(&aux_dir)
.args(self.split_maybe_args(&self.config.target_rustcflags))
.args(&self.props.compile_flags)
.envs(self.props.exec_env.clone());
self.maybe_add_external_args(&mut rustc,
self.split_maybe_args(&self.config.target_rustcflags));
let src = match read_from {
ReadFrom::Stdin(src) => Some(src),
@ -587,6 +588,15 @@ impl<'test> TestCx<'test> {
}
}
fn set_revision_flags(&self, cmd: &mut Command) {
if let Some(revision) = self.revision {
// Normalize revisions to be lowercase and replace `-`s with `_`s.
// Otherwise the `--cfg` flag is not valid.
let normalized_revision = revision.to_lowercase().replace("-", "_");
cmd.args(&["--cfg", &normalized_revision]);
}
}
fn typecheck_source(&self, src: String) -> ProcRes {
let mut rustc = Command::new(&self.config.rustc_path);
@ -612,12 +622,9 @@ impl<'test> TestCx<'test> {
.arg(&self.config.build_base)
.arg("-L")
.arg(aux_dir);
if let Some(revision) = self.revision {
rustc.args(&["--cfg", revision]);
}
rustc.args(self.split_maybe_args(&self.config.target_rustcflags));
self.set_revision_flags(&mut rustc);
self.maybe_add_external_args(&mut rustc,
self.split_maybe_args(&self.config.target_rustcflags));
rustc.args(&self.props.compile_flags);
self.compose_and_run_compiler(rustc, Some(src))
@ -1119,6 +1126,35 @@ impl<'test> TestCx<'test> {
Some(new_options.join(" "))
}
fn maybe_add_external_args(&self, cmd: &mut Command, args: Vec<String>) {
// Filter out the arguments that should not be added by runtest here.
//
// Notable use-cases are: do not add our optimisation flag if
// `compile-flags: -Copt-level=x` and similar for debug-info level as well.
const OPT_FLAGS: &[&str] = &["-O", "-Copt-level=", /*-C<space>*/"opt-level="];
const DEBUG_FLAGS: &[&str] = &["-g", "-Cdebuginfo=", /*-C<space>*/"debuginfo="];
// FIXME: ideally we would "just" check the `cmd` itself, but it does not allow inspecting
// its arguments. They need to be collected separately. For now I cannot be bothered to
// implement this the "right" way.
let have_opt_flag = self.props.compile_flags.iter().any(|arg| {
OPT_FLAGS.iter().any(|f| arg.starts_with(f))
});
let have_debug_flag = self.props.compile_flags.iter().any(|arg| {
DEBUG_FLAGS.iter().any(|f| arg.starts_with(f))
});
for arg in args {
if OPT_FLAGS.iter().any(|f| arg.starts_with(f)) && have_opt_flag {
continue;
}
if DEBUG_FLAGS.iter().any(|f| arg.starts_with(f)) && have_debug_flag {
continue;
}
cmd.arg(arg);
}
}
fn check_debugger_output(&self, debugger_run_result: &ProcRes, check_lines: &[String]) {
let num_check_lines = check_lines.len();
@ -1707,10 +1743,7 @@ impl<'test> TestCx<'test> {
rustc.arg(&format!("--target={}", target));
}
if let Some(revision) = self.revision {
rustc.args(&["--cfg", revision]);
}
self.set_revision_flags(&mut rustc);
if !is_rustdoc {
if let Some(ref incremental_dir) = self.props.incremental_dir {
@ -1810,9 +1843,11 @@ impl<'test> TestCx<'test> {
}
if self.props.force_host {
rustc.args(self.split_maybe_args(&self.config.host_rustcflags));
self.maybe_add_external_args(&mut rustc,
self.split_maybe_args(&self.config.host_rustcflags));
} else {
rustc.args(self.split_maybe_args(&self.config.target_rustcflags));
self.maybe_add_external_args(&mut rustc,
self.split_maybe_args(&self.config.target_rustcflags));
if !is_rustdoc {
if let Some(ref linker) = self.config.linker {
rustc.arg(format!("-Clinker={}", linker));
@ -2065,12 +2100,19 @@ impl<'test> TestCx<'test> {
.arg("--input-file")
.arg(irfile)
.arg(&self.testpaths.file);
// It would be more appropriate to make most of the arguments configurable through
// a comment-attribute similar to `compile-flags`. For example, --check-prefixes is a very
// useful flag.
//
// For now, though…
if let Some(rev) = self.revision {
let prefixes = format!("CHECK,{}", rev);
filecheck.args(&["--check-prefixes", &prefixes]);
}
self.compose_and_run(filecheck, "", None, None)
}
fn run_codegen_test(&self) {
assert!(self.revision.is_none(), "revisions not relevant here");
if self.config.llvm_filecheck.is_none() {
self.fatal("missing --llvm-filecheck");
}