2 use crate::common::{self, CodegenCx};
4 use crate::llvm::{self, True};
6 use crate::type_::Type;
7 use crate::type_of::LayoutLlvmExt;
8 use crate::value::Value;
11 use rustc_codegen_ssa::traits::*;
12 use rustc_hir::def_id::DefId;
13 use rustc_middle::middle::codegen_fn_attrs::{CodegenFnAttrFlags, CodegenFnAttrs};
14 use rustc_middle::mir::interpret::{
15 read_target_uint, Allocation, ConstAllocation, ErrorHandled, GlobalAlloc, InitChunk, Pointer,
16 Scalar as InterpScalar,
18 use rustc_middle::mir::mono::MonoItem;
19 use rustc_middle::ty::layout::LayoutOf;
20 use rustc_middle::ty::{self, Instance, Ty};
21 use rustc_middle::{bug, span_bug};
22 use rustc_target::abi::{
23 AddressSpace, Align, HasDataLayout, Primitive, Scalar, Size, WrappingRange,
28 pub fn const_alloc_to_llvm<'ll>(cx: &CodegenCx<'ll, '_>, alloc: ConstAllocation<'_>) -> &'ll Value {
29 let alloc = alloc.inner();
30 let mut llvals = Vec::with_capacity(alloc.relocations().len() + 1);
31 let dl = cx.data_layout();
32 let pointer_size = dl.pointer_size.bytes() as usize;
34 // Note: this function may call `inspect_with_uninit_and_ptr_outside_interpreter`,
35 // so `range` must be within the bounds of `alloc` and not contain or overlap a relocation.
36 fn append_chunks_of_init_and_uninit_bytes<'ll, 'a, 'b>(
37 llvals: &mut Vec<&'ll Value>,
38 cx: &'a CodegenCx<'ll, 'b>,
39 alloc: &'a Allocation,
44 .range_as_init_chunks(Size::from_bytes(range.start), Size::from_bytes(range.end));
46 let chunk_to_llval = move |chunk| match chunk {
47 InitChunk::Init(range) => {
48 let range = (range.start.bytes() as usize)..(range.end.bytes() as usize);
49 let bytes = alloc.inspect_with_uninit_and_ptr_outside_interpreter(range);
52 InitChunk::Uninit(range) => {
53 let len = range.end.bytes() - range.start.bytes();
54 cx.const_undef(cx.type_array(cx.type_i8(), len))
58 // Generating partially-uninit consts is limited to small numbers of chunks,
59 // to avoid the cost of generating large complex const expressions.
60 // For example, `[(u32, u8); 1024 * 1024]` contains uninit padding in each element,
61 // and would result in `{ [5 x i8] zeroinitializer, [3 x i8] undef, ...repeat 1M times... }`.
62 let max = if llvm_util::get_version() < (14, 0, 0) {
63 // Generating partially-uninit consts inhibits optimizations in LLVM < 14.
64 // See https://github.com/rust-lang/rust/issues/84565.
67 cx.sess().opts.unstable_opts.uninit_const_chunk_threshold
69 let allow_uninit_chunks = chunks.clone().take(max.saturating_add(1)).count() <= max;
71 if allow_uninit_chunks {
72 llvals.extend(chunks.map(chunk_to_llval));
74 // If this allocation contains any uninit bytes, codegen as if it was initialized
75 // (using some arbitrary value for uninit bytes).
76 let bytes = alloc.inspect_with_uninit_and_ptr_outside_interpreter(range);
77 llvals.push(cx.const_bytes(bytes));
81 let mut next_offset = 0;
82 for &(offset, alloc_id) in alloc.relocations().iter() {
83 let offset = offset.bytes();
84 assert_eq!(offset as usize as u64, offset);
85 let offset = offset as usize;
86 if offset > next_offset {
87 // This `inspect` is okay since we have checked that it is not within a relocation, it
88 // is within the bounds of the allocation, and it doesn't affect interpreter execution
89 // (we inspect the result after interpreter execution).
90 append_chunks_of_init_and_uninit_bytes(&mut llvals, cx, alloc, next_offset..offset);
92 let ptr_offset = read_target_uint(
94 // This `inspect` is okay since it is within the bounds of the allocation, it doesn't
95 // affect interpreter execution (we inspect the result after interpreter execution),
96 // and we properly interpret the relocation as a relocation pointer offset.
97 alloc.inspect_with_uninit_and_ptr_outside_interpreter(offset..(offset + pointer_size)),
99 .expect("const_alloc_to_llvm: could not read relocation pointer")
102 let address_space = match cx.tcx.global_alloc(alloc_id) {
103 GlobalAlloc::Function(..) => cx.data_layout().instruction_address_space,
104 GlobalAlloc::Static(..) | GlobalAlloc::Memory(..) | GlobalAlloc::VTable(..) => {
109 llvals.push(cx.scalar_to_backend(
110 InterpScalar::from_pointer(
111 Pointer::new(alloc_id, Size::from_bytes(ptr_offset)),
114 Scalar::Initialized {
115 value: Primitive::Pointer,
116 valid_range: WrappingRange::full(dl.pointer_size),
118 cx.type_i8p_ext(address_space),
120 next_offset = offset + pointer_size;
122 if alloc.len() >= next_offset {
123 let range = next_offset..alloc.len();
124 // This `inspect` is okay since we have check that it is after all relocations, it is
125 // within the bounds of the allocation, and it doesn't affect interpreter execution (we
126 // inspect the result after interpreter execution).
127 append_chunks_of_init_and_uninit_bytes(&mut llvals, cx, alloc, range);
130 cx.const_struct(&llvals, true)
133 pub fn codegen_static_initializer<'ll, 'tcx>(
134 cx: &CodegenCx<'ll, 'tcx>,
136 ) -> Result<(&'ll Value, ConstAllocation<'tcx>), ErrorHandled> {
137 let alloc = cx.tcx.eval_static_initializer(def_id)?;
138 Ok((const_alloc_to_llvm(cx, alloc), alloc))
141 fn set_global_alignment<'ll>(cx: &CodegenCx<'ll, '_>, gv: &'ll Value, mut align: Align) {
142 // The target may require greater alignment for globals than the type does.
143 // Note: GCC and Clang also allow `__attribute__((aligned))` on variables,
144 // which can force it to be smaller. Rust doesn't support this yet.
145 if let Some(min) = cx.sess().target.min_global_align {
146 match Align::from_bits(min) {
147 Ok(min) => align = align.max(min),
149 cx.sess().err(&format!("invalid minimum global alignment: {}", err));
154 llvm::LLVMSetAlignment(gv, align.bytes() as u32);
158 fn check_and_apply_linkage<'ll, 'tcx>(
159 cx: &CodegenCx<'ll, 'tcx>,
160 attrs: &CodegenFnAttrs,
165 let llty = cx.layout_of(ty).llvm_type(cx);
166 if let Some(linkage) = attrs.linkage {
167 debug!("get_static: sym={} linkage={:?}", sym, linkage);
169 // If this is a static with a linkage specified, then we need to handle
170 // it a little specially. The typesystem prevents things like &T and
171 // extern "C" fn() from being non-null, so we can't just declare a
172 // static and call it a day. Some linkages (like weak) will make it such
173 // that the static actually has a null value.
174 let llty2 = if let ty::RawPtr(ref mt) = ty.kind() {
175 cx.layout_of(mt.ty).llvm_type(cx)
177 cx.sess().span_fatal(
178 cx.tcx.def_span(def_id),
179 "must have type `*const T` or `*mut T` due to `#[linkage]` attribute",
183 // Declare a symbol `foo` with the desired linkage.
184 let g1 = cx.declare_global(sym, llty2);
185 llvm::LLVMRustSetLinkage(g1, base::linkage_to_llvm(linkage));
187 // Declare an internal global `extern_with_linkage_foo` which
188 // is initialized with the address of `foo`. If `foo` is
189 // discarded during linking (for example, if `foo` has weak
190 // linkage and there are no definitions), then
191 // `extern_with_linkage_foo` will instead be initialized to
193 let mut real_name = "_rust_extern_with_linkage_".to_string();
194 real_name.push_str(sym);
195 let g2 = cx.define_global(&real_name, llty).unwrap_or_else(|| {
196 cx.sess().span_fatal(
197 cx.tcx.def_span(def_id),
198 &format!("symbol `{}` is already defined", &sym),
201 llvm::LLVMRustSetLinkage(g2, llvm::Linkage::InternalLinkage);
202 llvm::LLVMSetInitializer(g2, g1);
205 } else if cx.tcx.sess.target.arch == "x86" &&
206 let Some(dllimport) = common::get_dllimport(cx.tcx, def_id, sym)
208 cx.declare_global(&common::i686_decorated_name(&dllimport, common::is_mingw_gnu_toolchain(&cx.tcx.sess.target), true), llty)
210 // Generate an external declaration.
211 // FIXME(nagisa): investigate whether it can be changed into define_global
212 cx.declare_global(sym, llty)
216 pub fn ptrcast<'ll>(val: &'ll Value, ty: &'ll Type) -> &'ll Value {
217 unsafe { llvm::LLVMConstPointerCast(val, ty) }
220 impl<'ll> CodegenCx<'ll, '_> {
221 pub(crate) fn const_bitcast(&self, val: &'ll Value, ty: &'ll Type) -> &'ll Value {
222 unsafe { llvm::LLVMConstBitCast(val, ty) }
225 pub(crate) fn static_addr_of_mut(
232 let gv = match kind {
233 Some(kind) if !self.tcx.sess.fewer_names() => {
234 let name = self.generate_local_symbol_name(kind);
235 let gv = self.define_global(&name, self.val_ty(cv)).unwrap_or_else(|| {
236 bug!("symbol `{}` is already defined", name);
238 llvm::LLVMRustSetLinkage(gv, llvm::Linkage::PrivateLinkage);
241 _ => self.define_private_global(self.val_ty(cv)),
243 llvm::LLVMSetInitializer(gv, cv);
244 set_global_alignment(self, gv, align);
245 llvm::SetUnnamedAddress(gv, llvm::UnnamedAddr::Global);
250 pub(crate) fn get_static(&self, def_id: DefId) -> &'ll Value {
251 let instance = Instance::mono(self.tcx, def_id);
252 if let Some(&g) = self.instances.borrow().get(&instance) {
256 let defined_in_current_codegen_unit =
257 self.codegen_unit.items().contains_key(&MonoItem::Static(def_id));
259 !defined_in_current_codegen_unit,
260 "consts::get_static() should always hit the cache for \
261 statics defined in the same CGU, but did not for `{:?}`",
265 let ty = instance.ty(self.tcx, ty::ParamEnv::reveal_all());
266 let sym = self.tcx.symbol_name(instance).name;
267 let fn_attrs = self.tcx.codegen_fn_attrs(def_id);
269 debug!("get_static: sym={} instance={:?} fn_attrs={:?}", sym, instance, fn_attrs);
271 let g = if def_id.is_local() && !self.tcx.is_foreign_item(def_id) {
272 let llty = self.layout_of(ty).llvm_type(self);
273 if let Some(g) = self.get_declared_value(sym) {
274 if self.val_ty(g) != self.type_ptr_to(llty) {
275 span_bug!(self.tcx.def_span(def_id), "Conflicting types for static");
279 let g = self.declare_global(sym, llty);
281 if !self.tcx.is_reachable_non_generic(def_id) {
283 llvm::LLVMRustSetVisibility(g, llvm::Visibility::Hidden);
289 check_and_apply_linkage(self, fn_attrs, ty, sym, def_id)
292 // Thread-local statics in some other crate need to *always* be linked
293 // against in a thread-local fashion, so we need to be sure to apply the
294 // thread-local attribute locally if it was present remotely. If we
295 // don't do this then linker errors can be generated where the linker
296 // complains that one object files has a thread local version of the
297 // symbol and another one doesn't.
298 if fn_attrs.flags.contains(CodegenFnAttrFlags::THREAD_LOCAL) {
299 llvm::set_thread_local_mode(g, self.tls_model);
302 if !def_id.is_local() {
303 let needs_dll_storage_attr = self.use_dll_storage_attrs && !self.tcx.is_foreign_item(def_id) &&
304 // ThinLTO can't handle this workaround in all cases, so we don't
305 // emit the attrs. Instead we make them unnecessary by disallowing
306 // dynamic linking when linker plugin based LTO is enabled.
307 !self.tcx.sess.opts.cg.linker_plugin_lto.enabled();
309 // If this assertion triggers, there's something wrong with commandline
310 // argument validation.
312 !(self.tcx.sess.opts.cg.linker_plugin_lto.enabled()
313 && self.tcx.sess.target.is_like_windows
314 && self.tcx.sess.opts.cg.prefer_dynamic)
317 if needs_dll_storage_attr {
318 // This item is external but not foreign, i.e., it originates from an external Rust
319 // crate. Since we don't know whether this crate will be linked dynamically or
320 // statically in the final application, we always mark such symbols as 'dllimport'.
321 // If final linkage happens to be static, we rely on compiler-emitted __imp_ stubs
322 // to make things work.
324 // However, in some scenarios we defer emission of statics to downstream
325 // crates, so there are cases where a static with an upstream DefId
326 // is actually present in the current crate. We can find out via the
327 // is_codegened_item query.
328 if !self.tcx.is_codegened_item(def_id) {
330 llvm::LLVMSetDLLStorageClass(g, llvm::DLLStorageClass::DllImport);
336 if self.use_dll_storage_attrs && self.tcx.is_dllimport_foreign_item(def_id) {
337 // For foreign (native) libs we know the exact storage type to use.
339 llvm::LLVMSetDLLStorageClass(g, llvm::DLLStorageClass::DllImport);
344 if self.should_assume_dso_local(g, true) {
345 llvm::LLVMRustSetDSOLocal(g, true);
349 self.instances.borrow_mut().insert(instance, g);
354 impl<'ll> StaticMethods for CodegenCx<'ll, '_> {
355 fn static_addr_of(&self, cv: &'ll Value, align: Align, kind: Option<&str>) -> &'ll Value {
356 if let Some(&gv) = self.const_globals.borrow().get(&cv) {
358 // Upgrade the alignment in cases where the same constant is used with different
359 // alignment requirements
360 let llalign = align.bytes() as u32;
361 if llalign > llvm::LLVMGetAlignment(gv) {
362 llvm::LLVMSetAlignment(gv, llalign);
367 let gv = self.static_addr_of_mut(cv, align, kind);
369 llvm::LLVMSetGlobalConstant(gv, True);
371 self.const_globals.borrow_mut().insert(cv, gv);
375 fn codegen_static(&self, def_id: DefId, is_mutable: bool) {
377 let attrs = self.tcx.codegen_fn_attrs(def_id);
379 let Ok((v, alloc)) = codegen_static_initializer(self, def_id) else {
380 // Error has already been reported
383 let alloc = alloc.inner();
385 let g = self.get_static(def_id);
387 // boolean SSA values are i1, but they have to be stored in i8 slots,
388 // otherwise some LLVM optimization passes don't work as expected
389 let mut val_llty = self.val_ty(v);
390 let v = if val_llty == self.type_i1() {
391 val_llty = self.type_i8();
392 llvm::LLVMConstZExt(v, val_llty)
397 let instance = Instance::mono(self.tcx, def_id);
398 let ty = instance.ty(self.tcx, ty::ParamEnv::reveal_all());
399 let llty = self.layout_of(ty).llvm_type(self);
400 let g = if val_llty == llty {
403 // If we created the global with the wrong type,
405 let name = llvm::get_value_name(g).to_vec();
406 llvm::set_value_name(g, b"");
408 let linkage = llvm::LLVMRustGetLinkage(g);
409 let visibility = llvm::LLVMRustGetVisibility(g);
411 let new_g = llvm::LLVMRustGetOrInsertGlobal(
413 name.as_ptr().cast(),
418 llvm::LLVMRustSetLinkage(new_g, linkage);
419 llvm::LLVMRustSetVisibility(new_g, visibility);
421 // The old global has had its name removed but is returned by
422 // get_static since it is in the instance cache. Provide an
423 // alternative lookup that points to the new global so that
424 // global_asm! can compute the correct mangled symbol name
426 self.renamed_statics.borrow_mut().insert(def_id, new_g);
428 // To avoid breaking any invariants, we leave around the old
429 // global for the moment; we'll replace all references to it
430 // with the new global later. (See base::codegen_backend.)
431 self.statics_to_rauw.borrow_mut().push((g, new_g));
434 set_global_alignment(self, g, self.align_of(ty));
435 llvm::LLVMSetInitializer(g, v);
437 if self.should_assume_dso_local(g, true) {
438 llvm::LLVMRustSetDSOLocal(g, true);
441 // As an optimization, all shared statics which do not have interior
442 // mutability are placed into read-only memory.
443 if !is_mutable && self.type_is_freeze(ty) {
444 llvm::LLVMSetGlobalConstant(g, llvm::True);
447 debuginfo::build_global_var_di_node(self, def_id, g);
449 if attrs.flags.contains(CodegenFnAttrFlags::THREAD_LOCAL) {
450 llvm::set_thread_local_mode(g, self.tls_model);
452 // Do not allow LLVM to change the alignment of a TLS on macOS.
454 // By default a global's alignment can be freely increased.
455 // This allows LLVM to generate more performant instructions
456 // e.g., using load-aligned into a SIMD register.
458 // However, on macOS 10.10 or below, the dynamic linker does not
459 // respect any alignment given on the TLS (radar 24221680).
460 // This will violate the alignment assumption, and causing segfault at runtime.
462 // This bug is very easy to trigger. In `println!` and `panic!`,
463 // the `LOCAL_STDOUT`/`LOCAL_STDERR` handles are stored in a TLS,
464 // which the values would be `mem::replace`d on initialization.
465 // The implementation of `mem::replace` will use SIMD
466 // whenever the size is 32 bytes or higher. LLVM notices SIMD is used
467 // and tries to align `LOCAL_STDOUT`/`LOCAL_STDERR` to a 32-byte boundary,
468 // which macOS's dyld disregarded and causing crashes
469 // (see issues #51794, #51758, #50867, #48866 and #44056).
471 // To workaround the bug, we trick LLVM into not increasing
472 // the global's alignment by explicitly assigning a section to it
473 // (equivalent to automatically generating a `#[link_section]` attribute).
474 // See the comment in the `GlobalValue::canIncreaseAlignment()` function
475 // of `lib/IR/Globals.cpp` for why this works.
477 // When the alignment is not increased, the optimized `mem::replace`
478 // will use load-unaligned instructions instead, and thus avoiding the crash.
480 // We could remove this hack whenever we decide to drop macOS 10.10 support.
481 if self.tcx.sess.target.is_like_osx {
482 // The `inspect` method is okay here because we checked relocations, and
483 // because we are doing this access to inspect the final interpreter state
484 // (not as part of the interpreter execution).
486 // FIXME: This check requires that the (arbitrary) value of undefined bytes
487 // happens to be zero. Instead, we should only check the value of defined bytes
488 // and set all undefined bytes to zero if this allocation is headed for the
490 let all_bytes_are_zero = alloc.relocations().is_empty()
492 .inspect_with_uninit_and_ptr_outside_interpreter(0..alloc.len())
494 .all(|&byte| byte == 0);
496 let sect_name = if all_bytes_are_zero {
497 cstr!("__DATA,__thread_bss")
499 cstr!("__DATA,__thread_data")
501 llvm::LLVMSetSection(g, sect_name.as_ptr());
505 // Wasm statics with custom link sections get special treatment as they
506 // go into custom sections of the wasm executable.
507 if self.tcx.sess.target.is_like_wasm {
508 if let Some(section) = attrs.link_section {
509 let section = llvm::LLVMMDStringInContext(
511 section.as_str().as_ptr().cast(),
512 section.as_str().len() as c_uint,
514 assert!(alloc.relocations().is_empty());
516 // The `inspect` method is okay here because we checked relocations, and
517 // because we are doing this access to inspect the final interpreter state (not
518 // as part of the interpreter execution).
520 alloc.inspect_with_uninit_and_ptr_outside_interpreter(0..alloc.len());
521 let alloc = llvm::LLVMMDStringInContext(
523 bytes.as_ptr().cast(),
524 bytes.len() as c_uint,
526 let data = [section, alloc];
527 let meta = llvm::LLVMMDNodeInContext(self.llcx, data.as_ptr(), 2);
528 llvm::LLVMAddNamedMetadataOperand(
530 "wasm.custom_sections\0".as_ptr().cast(),
535 base::set_link_section(g, attrs);
538 if attrs.flags.contains(CodegenFnAttrFlags::USED) {
539 // `USED` and `USED_LINKER` can't be used together.
540 assert!(!attrs.flags.contains(CodegenFnAttrFlags::USED_LINKER));
542 // The semantics of #[used] in Rust only require the symbol to make it into the
543 // object file. It is explicitly allowed for the linker to strip the symbol if it
544 // is dead, which means we are allowed use `llvm.compiler.used` instead of
547 // Additionally, https://reviews.llvm.org/D97448 in LLVM 13 started emitting unique
548 // sections with SHF_GNU_RETAIN flag for llvm.used symbols, which may trigger bugs
549 // in the handling of `.init_array` (the static constructor list) in versions of
550 // the gold linker (prior to the one released with binutils 2.36).
552 // That said, we only ever emit these when compiling for ELF targets, unless
553 // `#[used(compiler)]` is explicitly requested. This is to avoid similar breakage
554 // on other targets, in particular MachO targets have *their* static constructor
555 // lists broken if `llvm.compiler.used` is emitted rather than llvm.used. However,
556 // that check happens when assigning the `CodegenFnAttrFlags` in `rustc_typeck`,
557 // so we don't need to take care of it here.
558 self.add_compiler_used_global(g);
560 if attrs.flags.contains(CodegenFnAttrFlags::USED_LINKER) {
561 // `USED` and `USED_LINKER` can't be used together.
562 assert!(!attrs.flags.contains(CodegenFnAttrFlags::USED));
564 self.add_used_global(g);
569 /// Add a global value to a list to be stored in the `llvm.used` variable, an array of i8*.
570 fn add_used_global(&self, global: &'ll Value) {
571 let cast = unsafe { llvm::LLVMConstPointerCast(global, self.type_i8p()) };
572 self.used_statics.borrow_mut().push(cast);
575 /// Add a global value to a list to be stored in the `llvm.compiler.used` variable,
577 fn add_compiler_used_global(&self, global: &'ll Value) {
578 let cast = unsafe { llvm::LLVMConstPointerCast(global, self.type_i8p()) };
579 self.compiler_used_statics.borrow_mut().push(cast);