diff options
Diffstat (limited to 'compiler/rustc_codegen_ssa')
| -rw-r--r-- | compiler/rustc_codegen_ssa/Cargo.toml | 1 | ||||
| -rw-r--r-- | compiler/rustc_codegen_ssa/src/back/write.rs | 33 | ||||
| -rw-r--r-- | compiler/rustc_codegen_ssa/src/base.rs | 63 | ||||
| -rw-r--r-- | compiler/rustc_codegen_ssa/src/codegen_attrs.rs | 10 | ||||
| -rw-r--r-- | compiler/rustc_codegen_ssa/src/mir/block.rs | 34 | ||||
| -rw-r--r-- | compiler/rustc_codegen_ssa/src/traits/write.rs | 2 |
6 files changed, 65 insertions, 78 deletions
diff --git a/compiler/rustc_codegen_ssa/Cargo.toml b/compiler/rustc_codegen_ssa/Cargo.toml index 30956fd1855..6a8971de746 100644 --- a/compiler/rustc_codegen_ssa/Cargo.toml +++ b/compiler/rustc_codegen_ssa/Cargo.toml @@ -26,6 +26,7 @@ rustc_hashes = { path = "../rustc_hashes" } rustc_hir = { path = "../rustc_hir" } rustc_incremental = { path = "../rustc_incremental" } rustc_index = { path = "../rustc_index" } +rustc_lint_defs = { path = "../rustc_lint_defs" } rustc_macros = { path = "../rustc_macros" } rustc_metadata = { path = "../rustc_metadata" } rustc_middle = { path = "../rustc_middle" } diff --git a/compiler/rustc_codegen_ssa/src/back/write.rs b/compiler/rustc_codegen_ssa/src/back/write.rs index aa29afb7f5b..26d089a1171 100644 --- a/compiler/rustc_codegen_ssa/src/back/write.rs +++ b/compiler/rustc_codegen_ssa/src/back/write.rs @@ -7,7 +7,6 @@ use std::{fs, io, mem, str, thread}; use rustc_abi::Size; use rustc_ast::attr; -use rustc_ast::expand::autodiff_attrs::AutoDiffItem; use rustc_data_structures::fx::FxIndexMap; use rustc_data_structures::jobserver::{self, Acquired}; use rustc_data_structures::memmap::Mmap; @@ -38,7 +37,7 @@ use tracing::debug; use super::link::{self, ensure_removed}; use super::lto::{self, SerializedModule}; use crate::back::lto::check_lto_allowed; -use crate::errors::{AutodiffWithoutLto, ErrorCreatingRemarkDir}; +use crate::errors::ErrorCreatingRemarkDir; use crate::traits::*; use crate::{ CachedModuleCodegen, CodegenResults, CompiledModule, CrateInfo, ModuleCodegen, ModuleKind, @@ -76,12 +75,9 @@ pub struct ModuleConfig { /// Names of additional optimization passes to run. pub passes: Vec<String>, /// Some(level) to optimize at a certain level, or None to run - /// absolutely no optimizations (used for the metadata module). + /// absolutely no optimizations (used for the allocator module). pub opt_level: Option<config::OptLevel>, - /// Some(level) to optimize binary size, or None to not affect program size. - pub opt_size: Option<config::OptLevel>, - pub pgo_gen: SwitchWithOptPath, pub pgo_use: Option<PathBuf>, pub pgo_sample_use: Option<PathBuf>, @@ -102,7 +98,6 @@ pub struct ModuleConfig { pub emit_obj: EmitObj, pub emit_thin_lto: bool, pub emit_thin_lto_summary: bool, - pub bc_cmdline: String, // Miscellaneous flags. These are mostly copied from command-line // options. @@ -110,7 +105,6 @@ pub struct ModuleConfig { pub lint_llvm_ir: bool, pub no_prepopulate_passes: bool, pub no_builtins: bool, - pub time_module: bool, pub vectorize_loop: bool, pub vectorize_slp: bool, pub merge_functions: bool, @@ -171,7 +165,6 @@ impl ModuleConfig { passes: if_regular!(sess.opts.cg.passes.clone(), vec![]), opt_level: opt_level_and_size, - opt_size: opt_level_and_size, pgo_gen: if_regular!( sess.opts.cg.profile_generate.clone(), @@ -221,17 +214,12 @@ impl ModuleConfig { sess.opts.output_types.contains_key(&OutputType::ThinLinkBitcode), false ), - bc_cmdline: sess.target.bitcode_llvm_cmdline.to_string(), verify_llvm_ir: sess.verify_llvm_ir(), lint_llvm_ir: sess.opts.unstable_opts.lint_llvm_ir, no_prepopulate_passes: sess.opts.cg.no_prepopulate_passes, no_builtins: no_builtins || sess.target.no_builtins, - // Exclude metadata and allocator modules from time_passes output, - // since they throw off the "LLVM passes" measurement. - time_module: if_regular!(true, false), - // Copy what clang does by turning on loop vectorization at O2 and // slp vectorization at O3. vectorize_loop: !sess.opts.cg.no_vectorize_loops @@ -454,7 +442,6 @@ pub(crate) fn start_async_codegen<B: ExtraBackendMethods>( backend: B, tcx: TyCtxt<'_>, target_cpu: String, - autodiff_items: &[AutoDiffItem], ) -> OngoingCodegen<B> { let (coordinator_send, coordinator_receive) = channel(); @@ -473,7 +460,6 @@ pub(crate) fn start_async_codegen<B: ExtraBackendMethods>( backend.clone(), tcx, &crate_info, - autodiff_items, shared_emitter, codegen_worker_send, coordinator_receive, @@ -728,7 +714,6 @@ pub(crate) enum WorkItem<B: WriteBackendMethods> { each_linked_rlib_for_lto: Vec<PathBuf>, needs_fat_lto: Vec<FatLtoInput<B>>, import_only_modules: Vec<(SerializedModule<B::ModuleBuffer>, WorkProduct)>, - autodiff: Vec<AutoDiffItem>, }, /// Performs thin-LTO on the given module. ThinLto(lto::ThinModule<B>), @@ -1001,7 +986,6 @@ fn execute_fat_lto_work_item<B: ExtraBackendMethods>( each_linked_rlib_for_lto: &[PathBuf], mut needs_fat_lto: Vec<FatLtoInput<B>>, import_only_modules: Vec<(SerializedModule<B::ModuleBuffer>, WorkProduct)>, - autodiff: Vec<AutoDiffItem>, module_config: &ModuleConfig, ) -> Result<WorkItemResult<B>, FatalError> { for (module, wp) in import_only_modules { @@ -1013,7 +997,6 @@ fn execute_fat_lto_work_item<B: ExtraBackendMethods>( exported_symbols_for_lto, each_linked_rlib_for_lto, needs_fat_lto, - autodiff, )?; let module = B::codegen(cgcx, module, module_config)?; Ok(WorkItemResult::Finished(module)) @@ -1105,7 +1088,6 @@ fn start_executing_work<B: ExtraBackendMethods>( backend: B, tcx: TyCtxt<'_>, crate_info: &CrateInfo, - autodiff_items: &[AutoDiffItem], shared_emitter: SharedEmitter, codegen_worker_send: Sender<CguMessage>, coordinator_receive: Receiver<Message<B>>, @@ -1115,7 +1097,6 @@ fn start_executing_work<B: ExtraBackendMethods>( ) -> thread::JoinHandle<Result<CompiledModules, ()>> { let coordinator_send = tx_to_llvm_workers; let sess = tcx.sess; - let autodiff_items = autodiff_items.to_vec(); let mut each_linked_rlib_for_lto = Vec::new(); let mut each_linked_rlib_file_for_lto = Vec::new(); @@ -1448,7 +1429,6 @@ fn start_executing_work<B: ExtraBackendMethods>( each_linked_rlib_for_lto: each_linked_rlib_file_for_lto, needs_fat_lto, import_only_modules, - autodiff: autodiff_items.clone(), }, 0, )); @@ -1456,11 +1436,6 @@ fn start_executing_work<B: ExtraBackendMethods>( helper.request_token(); } } else { - if !autodiff_items.is_empty() { - let dcx = cgcx.create_dcx(); - dcx.handle().emit_fatal(AutodiffWithoutLto {}); - } - for (work, cost) in generate_thin_lto_work( &cgcx, &exported_symbols_for_lto, @@ -1740,7 +1715,7 @@ fn spawn_work<'a, B: ExtraBackendMethods>( llvm_start_time: &mut Option<VerboseTimingGuard<'a>>, work: WorkItem<B>, ) { - if cgcx.config(work.module_kind()).time_module && llvm_start_time.is_none() { + if llvm_start_time.is_none() { *llvm_start_time = Some(cgcx.prof.verbose_generic_activity("LLVM_passes")); } @@ -1795,7 +1770,6 @@ fn spawn_work<'a, B: ExtraBackendMethods>( each_linked_rlib_for_lto, needs_fat_lto, import_only_modules, - autodiff, } => { let _timer = cgcx .prof @@ -1806,7 +1780,6 @@ fn spawn_work<'a, B: ExtraBackendMethods>( &each_linked_rlib_for_lto, needs_fat_lto, import_only_modules, - autodiff, module_config, ) } diff --git a/compiler/rustc_codegen_ssa/src/base.rs b/compiler/rustc_codegen_ssa/src/base.rs index b4556ced0b3..67cd1f4cd41 100644 --- a/compiler/rustc_codegen_ssa/src/base.rs +++ b/compiler/rustc_codegen_ssa/src/base.rs @@ -647,7 +647,7 @@ pub fn codegen_crate<B: ExtraBackendMethods>( ) -> OngoingCodegen<B> { // Skip crate items and just output metadata in -Z no-codegen mode. if tcx.sess.opts.unstable_opts.no_codegen || !tcx.sess.opts.output_types.should_codegen() { - let ongoing_codegen = start_async_codegen(backend, tcx, target_cpu, &[]); + let ongoing_codegen = start_async_codegen(backend, tcx, target_cpu); ongoing_codegen.codegen_finished(tcx); @@ -665,8 +665,7 @@ pub fn codegen_crate<B: ExtraBackendMethods>( // Run the monomorphization collector and partition the collected items into // codegen units. - let MonoItemPartitions { codegen_units, autodiff_items, .. } = - tcx.collect_and_partition_mono_items(()); + let MonoItemPartitions { codegen_units, .. } = tcx.collect_and_partition_mono_items(()); // Force all codegen_unit queries so they are already either red or green // when compile_codegen_unit accesses them. We are not able to re-execute @@ -679,34 +678,7 @@ pub fn codegen_crate<B: ExtraBackendMethods>( } } - let ongoing_codegen = start_async_codegen(backend.clone(), tcx, target_cpu, autodiff_items); - - // Codegen an allocator shim, if necessary. - if let Some(kind) = allocator_kind_for_codegen(tcx) { - let llmod_id = - cgu_name_builder.build_cgu_name(LOCAL_CRATE, &["crate"], Some("allocator")).to_string(); - let module_llvm = tcx.sess.time("write_allocator_module", || { - backend.codegen_allocator( - tcx, - &llmod_id, - kind, - // If allocator_kind is Some then alloc_error_handler_kind must - // also be Some. - tcx.alloc_error_handler_kind(()).unwrap(), - ) - }); - - ongoing_codegen.wait_for_signal_to_codegen_item(); - ongoing_codegen.check_for_errors(tcx.sess); - - // These modules are generally cheap and won't throw off scheduling. - let cost = 0; - submit_codegened_module_to_llvm( - &ongoing_codegen.coordinator, - ModuleCodegen::new_allocator(llmod_id, module_llvm), - cost, - ); - } + let ongoing_codegen = start_async_codegen(backend.clone(), tcx, target_cpu); // For better throughput during parallel processing by LLVM, we used to sort // CGUs largest to smallest. This would lead to better thread utilization @@ -823,6 +795,35 @@ pub fn codegen_crate<B: ExtraBackendMethods>( } } + // Codegen an allocator shim, if necessary. + // Do this last to ensure the LLVM_passes timer doesn't start while no CGUs have been codegened + // yet for the backend to optimize. + if let Some(kind) = allocator_kind_for_codegen(tcx) { + let llmod_id = + cgu_name_builder.build_cgu_name(LOCAL_CRATE, &["crate"], Some("allocator")).to_string(); + let module_llvm = tcx.sess.time("write_allocator_module", || { + backend.codegen_allocator( + tcx, + &llmod_id, + kind, + // If allocator_kind is Some then alloc_error_handler_kind must + // also be Some. + tcx.alloc_error_handler_kind(()).unwrap(), + ) + }); + + ongoing_codegen.wait_for_signal_to_codegen_item(); + ongoing_codegen.check_for_errors(tcx.sess); + + // These modules are generally cheap and won't throw off scheduling. + let cost = 0; + submit_codegened_module_to_llvm( + &ongoing_codegen.coordinator, + ModuleCodegen::new_allocator(llmod_id, module_llvm), + cost, + ); + } + ongoing_codegen.codegen_finished(tcx); // Since the main thread is sometimes blocked during codegen, we keep track diff --git a/compiler/rustc_codegen_ssa/src/codegen_attrs.rs b/compiler/rustc_codegen_ssa/src/codegen_attrs.rs index a36a772bc97..af70f0deb07 100644 --- a/compiler/rustc_codegen_ssa/src/codegen_attrs.rs +++ b/compiler/rustc_codegen_ssa/src/codegen_attrs.rs @@ -177,14 +177,6 @@ fn process_builtin_attrs( let mut interesting_spans = InterestingAttributeDiagnosticSpans::default(); let rust_target_features = tcx.rust_target_features(LOCAL_CRATE); - // If our rustc version supports autodiff/enzyme, then we call our handler - // to check for any `#[rustc_autodiff(...)]` attributes. - // FIXME(jdonszelmann): merge with loop below - if cfg!(llvm_enzyme) { - let ad = autodiff_attrs(tcx, did.into()); - codegen_fn_attrs.autodiff_item = ad; - } - for attr in attrs.iter() { if let hir::Attribute::Parsed(p) = attr { match p { @@ -612,7 +604,7 @@ fn inherited_align<'tcx>(tcx: TyCtxt<'tcx>, def_id: DefId) -> Option<Align> { /// placeholder functions. We wrote the rustc_autodiff attributes ourself, so this should never /// panic, unless we introduced a bug when parsing the autodiff macro. //FIXME(jdonszelmann): put in the main loop. No need to have two..... :/ Let's do that when we make autodiff parsed. -fn autodiff_attrs(tcx: TyCtxt<'_>, id: DefId) -> Option<AutoDiffAttrs> { +pub fn autodiff_attrs(tcx: TyCtxt<'_>, id: DefId) -> Option<AutoDiffAttrs> { let attrs = tcx.get_attrs(id, sym::rustc_autodiff); let attrs = attrs.filter(|attr| attr.has_name(sym::rustc_autodiff)).collect::<Vec<_>>(); diff --git a/compiler/rustc_codegen_ssa/src/mir/block.rs b/compiler/rustc_codegen_ssa/src/mir/block.rs index e96590441fa..c3dc3e42b83 100644 --- a/compiler/rustc_codegen_ssa/src/mir/block.rs +++ b/compiler/rustc_codegen_ssa/src/mir/block.rs @@ -5,6 +5,7 @@ use rustc_ast as ast; use rustc_ast::{InlineAsmOptions, InlineAsmTemplatePiece}; use rustc_data_structures::packed::Pu128; use rustc_hir::lang_items::LangItem; +use rustc_lint_defs::builtin::TAIL_CALL_TRACK_CALLER; use rustc_middle::mir::{self, AssertKind, InlineAsmMacro, SwitchTargets, UnwindTerminateReason}; use rustc_middle::ty::layout::{HasTyCtxt, LayoutOf, ValidityRequirement}; use rustc_middle::ty::print::{with_no_trimmed_paths, with_no_visible_paths}; @@ -35,7 +36,7 @@ enum MergingSucc { True, } -/// Indicates to the call terminator codegen whether a cal +/// Indicates to the call terminator codegen whether a call /// is a normal call or an explicit tail call. #[derive(Debug, PartialEq)] enum CallKind { @@ -906,7 +907,7 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> { fn_span, ); - let instance = match instance.def { + match instance.def { // We don't need AsyncDropGlueCtorShim here because it is not `noop func`, // it is `func returning noop future` ty::InstanceKind::DropGlue(_, None) => { @@ -995,14 +996,35 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> { intrinsic.name, ); } - instance + (Some(instance), None) } } } - _ => instance, - }; - (Some(instance), None) + _ if kind == CallKind::Tail + && instance.def.requires_caller_location(bx.tcx()) => + { + if let Some(hir_id) = + terminator.source_info.scope.lint_root(&self.mir.source_scopes) + { + let msg = "tail calling a function marked with `#[track_caller]` has no special effect"; + bx.tcx().node_lint(TAIL_CALL_TRACK_CALLER, hir_id, |d| { + _ = d.primary_message(msg).span(fn_span) + }); + } + + let instance = ty::Instance::resolve_for_fn_ptr( + bx.tcx(), + bx.typing_env(), + def_id, + generic_args, + ) + .unwrap(); + + (None, Some(bx.get_fn_addr(instance))) + } + _ => (Some(instance), None), + } } ty::FnPtr(..) => (None, Some(callee.immediate())), _ => bug!("{} is not callable", callee.layout.ty), diff --git a/compiler/rustc_codegen_ssa/src/traits/write.rs b/compiler/rustc_codegen_ssa/src/traits/write.rs index f391c198e1a..c29ad90735b 100644 --- a/compiler/rustc_codegen_ssa/src/traits/write.rs +++ b/compiler/rustc_codegen_ssa/src/traits/write.rs @@ -1,6 +1,5 @@ use std::path::PathBuf; -use rustc_ast::expand::autodiff_attrs::AutoDiffItem; use rustc_errors::{DiagCtxtHandle, FatalError}; use rustc_middle::dep_graph::WorkProduct; @@ -23,7 +22,6 @@ pub trait WriteBackendMethods: Clone + 'static { exported_symbols_for_lto: &[String], each_linked_rlib_for_lto: &[PathBuf], modules: Vec<FatLtoInput<Self>>, - diff_fncs: Vec<AutoDiffItem>, ) -> Result<ModuleCodegen<Self::Module>, FatalError>; /// Performs thin LTO by performing necessary global analysis and returning two /// lists, one of the modules that need optimization and another for modules that |