1 //! Set and unset common attributes on LLVM values.
5 use rustc::hir::{CodegenFnAttrFlags, CodegenFnAttrs};
6 use rustc::hir::def_id::{DefId, LOCAL_CRATE};
7 use rustc::session::Session;
8 use rustc::session::config::{Sanitizer, OptLevel};
9 use rustc::ty::{self, TyCtxt, PolyFnSig};
10 use rustc::ty::layout::HasTyCtxt;
11 use rustc::ty::query::Providers;
12 use rustc_data_structures::small_c_str::SmallCStr;
13 use rustc_data_structures::fx::FxHashMap;
14 use rustc_target::spec::PanicStrategy;
15 use rustc_codegen_ssa::traits::*;
18 use crate::attributes;
19 use crate::llvm::{self, Attribute};
20 use crate::llvm::AttributePlace::Function;
22 pub use syntax::attr::{self, InlineAttr, OptimizeAttr};
24 use crate::context::CodegenCx;
25 use crate::value::Value;
27 /// Mark LLVM function to use provided inline heuristic.
29 pub fn inline(cx: &CodegenCx<'ll, '_>, val: &'ll Value, inline: InlineAttr) {
30 use self::InlineAttr::*;
32 Hint => Attribute::InlineHint.apply_llfn(Function, val),
33 Always => Attribute::AlwaysInline.apply_llfn(Function, val),
35 if cx.tcx().sess.target.target.arch != "amdgpu" {
36 Attribute::NoInline.apply_llfn(Function, val);
40 Attribute::InlineHint.unapply_llfn(Function, val);
41 Attribute::AlwaysInline.unapply_llfn(Function, val);
42 Attribute::NoInline.unapply_llfn(Function, val);
47 /// Tell LLVM to emit or not emit the information necessary to unwind the stack for the function.
49 pub fn emit_uwtable(val: &'ll Value, emit: bool) {
50 Attribute::UWTable.toggle_llfn(Function, val, emit);
53 /// Tell LLVM whether the function can or cannot unwind.
55 fn unwind(val: &'ll Value, can_unwind: bool) {
56 Attribute::NoUnwind.toggle_llfn(Function, val, !can_unwind);
59 /// Tell LLVM if this function should be 'naked', i.e., skip the epilogue and prologue.
61 pub fn naked(val: &'ll Value, is_naked: bool) {
62 Attribute::Naked.toggle_llfn(Function, val, is_naked);
65 pub fn set_frame_pointer_elimination(cx: &CodegenCx<'ll, '_>, llfn: &'ll Value) {
66 if cx.sess().must_not_eliminate_frame_pointers() {
67 llvm::AddFunctionAttrStringValue(
68 llfn, llvm::AttributePlace::Function,
69 const_cstr!("no-frame-pointer-elim"), const_cstr!("true"));
73 /// Tell LLVM what instrument function to insert.
75 pub fn set_instrument_function(cx: &CodegenCx<'ll, '_>, llfn: &'ll Value) {
76 if cx.sess().instrument_mcount() {
77 // Similar to `clang -pg` behavior. Handled by the
78 // `post-inline-ee-instrument` LLVM pass.
80 // The function name varies on platforms.
81 // See test/CodeGen/mcount.c in clang.
82 let mcount_name = CString::new(
83 cx.sess().target.target.options.target_mcount.as_str().as_bytes()).unwrap();
85 llvm::AddFunctionAttrStringValue(
86 llfn, llvm::AttributePlace::Function,
87 const_cstr!("instrument-function-entry-inlined"), &mcount_name);
91 pub fn set_probestack(cx: &CodegenCx<'ll, '_>, llfn: &'ll Value) {
92 // Only use stack probes if the target specification indicates that we
93 // should be using stack probes
94 if !cx.sess().target.target.options.stack_probes {
98 // Currently stack probes seem somewhat incompatible with the address
99 // sanitizer. With asan we're already protected from stack overflow anyway
100 // so we don't really need stack probes regardless.
101 if let Some(Sanitizer::Address) = cx.sess().opts.debugging_opts.sanitizer {
105 // probestack doesn't play nice either with `-C profile-generate`.
106 if cx.sess().opts.cg.profile_generate.enabled() {
110 // probestack doesn't play nice either with gcov profiling.
111 if cx.sess().opts.debugging_opts.profile {
115 // Flag our internal `__rust_probestack` function as the stack probe symbol.
116 // This is defined in the `compiler-builtins` crate for each architecture.
117 llvm::AddFunctionAttrStringValue(
118 llfn, llvm::AttributePlace::Function,
119 const_cstr!("probe-stack"), const_cstr!("__rust_probestack"));
122 fn translate_obsolete_target_features(feature: &str) -> &str {
123 const LLVM9_FEATURE_CHANGES: &[(&str, &str)] = &[
124 ("+fp-only-sp", "-fp64"),
125 ("-fp-only-sp", "+fp64"),
129 if llvm_util::get_major_version() >= 9 {
130 for &(old, new) in LLVM9_FEATURE_CHANGES {
136 for &(old, new) in LLVM9_FEATURE_CHANGES {
145 pub fn llvm_target_features(sess: &Session) -> impl Iterator<Item = &str> {
146 const RUSTC_SPECIFIC_FEATURES: &[&str] = &[
150 let cmdline = sess.opts.cg.target_feature.split(',')
151 .filter(|f| !RUSTC_SPECIFIC_FEATURES.iter().any(|s| f.contains(s)));
152 sess.target.target.options.features.split(',')
154 .filter(|l| !l.is_empty())
155 .map(translate_obsolete_target_features)
158 pub fn apply_target_cpu_attr(cx: &CodegenCx<'ll, '_>, llfn: &'ll Value) {
159 let target_cpu = SmallCStr::new(llvm_util::target_cpu(cx.tcx.sess));
160 llvm::AddFunctionAttrStringValue(
162 llvm::AttributePlace::Function,
163 const_cstr!("target-cpu"),
164 target_cpu.as_c_str());
167 /// Sets the `NonLazyBind` LLVM attribute on a given function,
168 /// assuming the codegen options allow skipping the PLT.
169 pub fn non_lazy_bind(sess: &Session, llfn: &'ll Value) {
170 // Don't generate calls through PLT if it's not necessary
171 if !sess.needs_plt() {
172 Attribute::NonLazyBind.apply_llfn(Function, llfn);
176 pub(crate) fn default_optimisation_attrs(sess: &Session, llfn: &'ll Value) {
177 match sess.opts.optimize {
179 llvm::Attribute::MinSize.unapply_llfn(Function, llfn);
180 llvm::Attribute::OptimizeForSize.apply_llfn(Function, llfn);
181 llvm::Attribute::OptimizeNone.unapply_llfn(Function, llfn);
183 OptLevel::SizeMin => {
184 llvm::Attribute::MinSize.apply_llfn(Function, llfn);
185 llvm::Attribute::OptimizeForSize.apply_llfn(Function, llfn);
186 llvm::Attribute::OptimizeNone.unapply_llfn(Function, llfn);
189 llvm::Attribute::MinSize.unapply_llfn(Function, llfn);
190 llvm::Attribute::OptimizeForSize.unapply_llfn(Function, llfn);
191 llvm::Attribute::OptimizeNone.unapply_llfn(Function, llfn);
198 /// Composite function which sets LLVM attributes for function depending on its AST (`#[attribute]`)
200 pub fn from_fn_attrs(
201 cx: &CodegenCx<'ll, 'tcx>,
204 sig: PolyFnSig<'tcx>,
206 let codegen_fn_attrs = id.map(|id| cx.tcx.codegen_fn_attrs(id))
207 .unwrap_or_else(|| CodegenFnAttrs::new());
209 match codegen_fn_attrs.optimize {
210 OptimizeAttr::None => {
211 default_optimisation_attrs(cx.tcx.sess, llfn);
213 OptimizeAttr::Speed => {
214 llvm::Attribute::MinSize.unapply_llfn(Function, llfn);
215 llvm::Attribute::OptimizeForSize.unapply_llfn(Function, llfn);
216 llvm::Attribute::OptimizeNone.unapply_llfn(Function, llfn);
218 OptimizeAttr::Size => {
219 llvm::Attribute::MinSize.apply_llfn(Function, llfn);
220 llvm::Attribute::OptimizeForSize.apply_llfn(Function, llfn);
221 llvm::Attribute::OptimizeNone.unapply_llfn(Function, llfn);
225 inline(cx, llfn, codegen_fn_attrs.inline);
227 // The `uwtable` attribute according to LLVM is:
229 // This attribute indicates that the ABI being targeted requires that an
230 // unwind table entry be produced for this function even if we can show
231 // that no exceptions passes by it. This is normally the case for the
232 // ELF x86-64 abi, but it can be disabled for some compilation units.
234 // Typically when we're compiling with `-C panic=abort` (which implies this
235 // `no_landing_pads` check) we don't need `uwtable` because we can't
236 // generate any exceptions! On Windows, however, exceptions include other
237 // events such as illegal instructions, segfaults, etc. This means that on
238 // Windows we end up still needing the `uwtable` attribute even if the `-C
239 // panic=abort` flag is passed.
241 // You can also find more info on why Windows is whitelisted here in:
242 // https://bugzilla.mozilla.org/show_bug.cgi?id=1302078
243 if !cx.sess().no_landing_pads() ||
244 cx.sess().target.target.options.requires_uwtable {
245 attributes::emit_uwtable(llfn, true);
248 set_frame_pointer_elimination(cx, llfn);
249 set_instrument_function(cx, llfn);
250 set_probestack(cx, llfn);
252 if codegen_fn_attrs.flags.contains(CodegenFnAttrFlags::COLD) {
253 Attribute::Cold.apply_llfn(Function, llfn);
255 if codegen_fn_attrs.flags.contains(CodegenFnAttrFlags::FFI_RETURNS_TWICE) {
256 Attribute::ReturnsTwice.apply_llfn(Function, llfn);
258 if codegen_fn_attrs.flags.contains(CodegenFnAttrFlags::NAKED) {
261 if codegen_fn_attrs.flags.contains(CodegenFnAttrFlags::ALLOCATOR) {
262 Attribute::NoAlias.apply_llfn(
263 llvm::AttributePlace::ReturnValue, llfn);
266 unwind(llfn, if cx.tcx.sess.panic_strategy() != PanicStrategy::Unwind {
267 // In panic=abort mode we assume nothing can unwind anywhere, so
268 // optimize based on this!
270 } else if codegen_fn_attrs.flags.contains(CodegenFnAttrFlags::UNWIND) {
271 // If a specific #[unwind] attribute is present, use that
273 } else if codegen_fn_attrs.flags.contains(CodegenFnAttrFlags::RUSTC_ALLOCATOR_NOUNWIND) {
274 // Special attribute for allocator functions, which can't unwind
276 } else if let Some(id) = id {
277 let sig = cx.tcx.normalize_erasing_late_bound_regions(ty::ParamEnv::reveal_all(), &sig);
278 if cx.tcx.is_foreign_item(id) {
279 // Foreign items like `extern "C" { fn foo(); }` are assumed not to
282 } else if sig.abi != Abi::Rust && sig.abi != Abi::RustCall {
283 // Any items defined in Rust that *don't* have the `extern` ABI are
284 // defined to not unwind. We insert shims to abort if an unwind
285 // happens to enforce this.
288 // Anything else defined in Rust is assumed that it can possibly
293 // assume this can possibly unwind, avoiding the application of a
294 // `nounwind` attribute below.
298 // Always annotate functions with the target-cpu they are compiled for.
299 // Without this, ThinLTO won't inline Rust functions into Clang generated
300 // functions (because Clang annotates functions this way too).
301 apply_target_cpu_attr(cx, llfn);
303 let features = llvm_target_features(cx.tcx.sess)
304 .map(|s| s.to_string())
306 codegen_fn_attrs.target_features
309 let feature = &*f.as_str();
310 format!("+{}", llvm_util::to_llvm_feature(cx.tcx.sess, feature))
313 .collect::<Vec<String>>()
316 if !features.is_empty() {
317 let val = CString::new(features).unwrap();
318 llvm::AddFunctionAttrStringValue(
319 llfn, llvm::AttributePlace::Function,
320 const_cstr!("target-features"), &val);
323 // Note that currently the `wasm-import-module` doesn't do anything, but
324 // eventually LLVM 7 should read this and ferry the appropriate import
325 // module to the output file.
326 if let Some(id) = id {
327 if cx.tcx.sess.target.target.arch == "wasm32" {
328 if let Some(module) = wasm_import_module(cx.tcx, id) {
329 llvm::AddFunctionAttrStringValue(
331 llvm::AttributePlace::Function,
332 const_cstr!("wasm-import-module"),
340 pub fn provide(providers: &mut Providers<'_>) {
341 providers.target_features_whitelist = |tcx, cnum| {
342 assert_eq!(cnum, LOCAL_CRATE);
343 if tcx.sess.opts.actually_rustdoc {
344 // rustdoc needs to be able to document functions that use all the features, so
345 // whitelist them all
346 tcx.arena.alloc(llvm_util::all_known_features()
347 .map(|(a, b)| (a.to_string(), b))
350 tcx.arena.alloc(llvm_util::target_feature_whitelist(tcx.sess)
352 .map(|&(a, b)| (a.to_string(), b))
357 provide_extern(providers);
360 pub fn provide_extern(providers: &mut Providers<'_>) {
361 providers.wasm_import_module_map = |tcx, cnum| {
362 // Build up a map from DefId to a `NativeLibrary` structure, where
363 // `NativeLibrary` internally contains information about
364 // `#[link(wasm_import_module = "...")]` for example.
365 let native_libs = tcx.native_libraries(cnum);
367 let def_id_to_native_lib = native_libs.iter().filter_map(|lib|
368 if let Some(id) = lib.foreign_module {
373 ).collect::<FxHashMap<_, _>>();
375 let mut ret = FxHashMap::default();
376 for lib in tcx.foreign_modules(cnum).iter() {
377 let module = def_id_to_native_lib
379 .and_then(|s| s.wasm_import_module);
380 let module = match module {
384 ret.extend(lib.foreign_items.iter().map(|id| {
385 assert_eq!(id.krate, cnum);
386 (*id, module.to_string())
394 fn wasm_import_module(tcx: TyCtxt<'_>, id: DefId) -> Option<CString> {
395 tcx.wasm_import_module_map(id.krate)
397 .map(|s| CString::new(&s[..]).unwrap())