Auto merge of #144562 - matthiaskrgr:rollup-mlvn7qo, r=matthiaskrgr

Rollup of 7 pull requests

Successful merges:

 - #144072 (update `Atomic*::from_ptr` and `Atomic*::as_ptr` docs)
 - #144151 (`tests/ui/issues/`: The Issues Strike Back [1/N])
 - #144300 (Clippy fixes for miropt-test-tools)
 - #144399 (Add a ratchet for moving all standard library tests to separate packages)
 - #144472 (str: Mark unstable `round_char_boundary` feature functions as const)
 - #144503 (Various refactors to the codegen coordinator code (part 3))
 - #144530 (coverage: Infer `instances_used` from `pgo_func_name_var_map`)

r? `@ghost`
`@rustbot` modify labels: rollup

Author: bors

compiler/rustc_codegen_gcc/src/back/write.rs

@@ -4,7 +4,6 @@ use gccjit::{Context, OutputKind};
 use rustc_codegen_ssa::back::link::ensure_removed;
 use rustc_codegen_ssa::back::write::{BitcodeSection, CodegenContext, EmitObj, ModuleConfig};
 use rustc_codegen_ssa::{CompiledModule, ModuleCodegen};
-use rustc_errors::DiagCtxtHandle;
 use rustc_fs_util::link_or_copy;
 use rustc_session::config::OutputType;
 use rustc_span::fatal_error::FatalError;
@@ -258,14 +257,6 @@ pub(crate) fn codegen(
     ))
 }
 
-pub(crate) fn link(
-    _cgcx: &CodegenContext<GccCodegenBackend>,
-    _dcx: DiagCtxtHandle<'_>,
-    mut _modules: Vec<ModuleCodegen<GccContext>>,
-) -> Result<ModuleCodegen<GccContext>, FatalError> {
-    unimplemented!();
-}
-
 pub(crate) fn save_temp_bitcode(
     cgcx: &CodegenContext<GccCodegenBackend>,
     _module: &ModuleCodegen<GccContext>,

compiler/rustc_codegen_gcc/src/lib.rs

@@ -426,14 +426,6 @@ impl WriteBackendMethods for GccCodegenBackend {
     fn serialize_module(_module: ModuleCodegen<Self::Module>) -> (String, Self::ModuleBuffer) {
         unimplemented!();
     }
-
-    fn run_link(
-        cgcx: &CodegenContext<Self>,
-        dcx: DiagCtxtHandle<'_>,
-        modules: Vec<ModuleCodegen<Self::Module>>,
-    ) -> Result<ModuleCodegen<Self::Module>, FatalError> {
-        back::write::link(cgcx, dcx, modules)
-    }
 }
 
 /// This is the entrypoint for a hot plugged rustc_codegen_gccjit

compiler/rustc_codegen_llvm/messages.ftl

@@ -12,7 +12,7 @@ codegen_llvm_from_llvm_optimization_diag = {$filename}:{$line}:{$column} {$pass_
 codegen_llvm_load_bitcode = failed to load bitcode of module "{$name}"
 codegen_llvm_load_bitcode_with_llvm_err = failed to load bitcode of module "{$name}": {$llvm_err}
 
-codegen_llvm_lto_bitcode_from_rlib = failed to get bitcode from object file for LTO ({$llvm_err})
+codegen_llvm_lto_bitcode_from_rlib = failed to get bitcode from object file for LTO ({$err})
 
 codegen_llvm_mismatch_data_layout =
     data-layout for target `{$rustc_target}`, `{$rustc_layout}`, differs from LLVM target's `{$llvm_target}` default layout, `{$llvm_layout}`

compiler/rustc_codegen_llvm/src/back/lto.rs

@@ -7,6 +7,7 @@ use std::sync::Arc;
 use std::{io, iter, slice};
 
 use object::read::archive::ArchiveFile;
+use object::{Object, ObjectSection};
 use rustc_codegen_ssa::back::lto::{SerializedModule, ThinModule, ThinShared};
 use rustc_codegen_ssa::back::write::{CodegenContext, FatLtoInput};
 use rustc_codegen_ssa::traits::*;
@@ -105,31 +106,15 @@ fn get_bitcode_slice_from_object_data<'a>(
     // name" which in the public API for sections gets treated as part of the section name, but
     // internally in MachOObjectFile.cpp gets treated separately.
     let section_name = bitcode_section_name(cgcx).to_str().unwrap().trim_start_matches("__LLVM,");
-    let mut len = 0;
-    let data = unsafe {
-        llvm::LLVMRustGetSliceFromObjectDataByName(
-            obj.as_ptr(),
-            obj.len(),
-            section_name.as_ptr(),
-            section_name.len(),
-            &mut len,
-        )
-    };
-    if !data.is_null() {
-        assert!(len != 0);
-        let bc = unsafe { slice::from_raw_parts(data, len) };
 
-        // `bc` must be a sub-slice of `obj`.
-        assert!(obj.as_ptr() <= bc.as_ptr());
-        assert!(bc[bc.len()..bc.len()].as_ptr() <= obj[obj.len()..obj.len()].as_ptr());
+    let obj =
+        object::File::parse(obj).map_err(|err| LtoBitcodeFromRlib { err: err.to_string() })?;
 
-        Ok(bc)
-    } else {
-        assert!(len == 0);
-        Err(LtoBitcodeFromRlib {
-            llvm_err: llvm::last_error().unwrap_or_else(|| "unknown LLVM error".to_string()),
-        })
-    }
+    let section = obj
+        .section_by_name(section_name)
+        .ok_or_else(|| LtoBitcodeFromRlib { err: format!("Can't find section {section_name}") })?;
+
+    section.data().map_err(|err| LtoBitcodeFromRlib { err: err.to_string() })
 }
 
 /// Performs fat LTO by merging all modules into a single one and returning it

compiler/rustc_codegen_llvm/src/back/write.rs

@@ -796,29 +796,6 @@ pub(crate) fn optimize(
     Ok(())
 }
 
-pub(crate) fn link(
-    cgcx: &CodegenContext<LlvmCodegenBackend>,
-    dcx: DiagCtxtHandle<'_>,
-    mut modules: Vec<ModuleCodegen<ModuleLlvm>>,
-) -> Result<ModuleCodegen<ModuleLlvm>, FatalError> {
-    use super::lto::{Linker, ModuleBuffer};
-    // Sort the modules by name to ensure deterministic behavior.
-    modules.sort_by(|a, b| a.name.cmp(&b.name));
-    let (first, elements) =
-        modules.split_first().expect("Bug! modules must contain at least one module.");
-
-    let mut linker = Linker::new(first.module_llvm.llmod());
-    for module in elements {
-        let _timer = cgcx.prof.generic_activity_with_arg("LLVM_link_module", &*module.name);
-        let buffer = ModuleBuffer::new(module.module_llvm.llmod());
-        linker
-            .add(buffer.data())
-            .map_err(|()| llvm_err(dcx, LlvmError::SerializeModule { name: &module.name }))?;
-    }
-    drop(linker);
-    Ok(modules.remove(0))
-}
-
 pub(crate) fn codegen(
     cgcx: &CodegenContext<LlvmCodegenBackend>,
     module: ModuleCodegen<ModuleLlvm>,

compiler/rustc_codegen_llvm/src/coverageinfo/mapgen.rs

@@ -46,21 +46,17 @@ pub(crate) fn finalize(cx: &mut CodegenCx<'_, '_>) {
     debug!("Generating coverage map for CodegenUnit: `{}`", cx.codegen_unit.name());
 
     // FIXME(#132395): Can this be none even when coverage is enabled?
-    let instances_used = match cx.coverage_cx {
-        Some(ref cx) => cx.instances_used.borrow(),
-        None => return,
-    };
+    let Some(ref coverage_cx) = cx.coverage_cx else { return };
 
-    let mut covfun_records = instances_used
-        .iter()
-        .copied()
+    let mut covfun_records = coverage_cx
+        .instances_used()
+        .into_iter()
         // Sort by symbol name, so that the global file table is built in an
         // order that doesn't depend on the stable-hash-based order in which
         // instances were visited during codegen.
         .sorted_by_cached_key(|&instance| tcx.symbol_name(instance).name)
         .filter_map(|instance| prepare_covfun_record(tcx, instance, true))
         .collect::<Vec<_>>();
-    drop(instances_used);
 
     // In a single designated CGU, also prepare covfun records for functions
     // in this crate that were instrumented for coverage, but are unused.

compiler/rustc_codegen_llvm/src/coverageinfo/mod.rs

@@ -5,7 +5,7 @@ use rustc_abi::Size;
 use rustc_codegen_ssa::traits::{
     BuilderMethods, ConstCodegenMethods, CoverageInfoBuilderMethods, MiscCodegenMethods,
 };
-use rustc_data_structures::fx::{FxHashMap, FxIndexSet};
+use rustc_data_structures::fx::{FxHashMap, FxIndexMap};
 use rustc_middle::mir::coverage::CoverageKind;
 use rustc_middle::ty::Instance;
 use tracing::{debug, instrument};
@@ -20,9 +20,14 @@ mod mapgen;
 
 /// Extra per-CGU context/state needed for coverage instrumentation.
 pub(crate) struct CguCoverageContext<'ll, 'tcx> {
-    /// Coverage data for each instrumented function identified by DefId.
-    pub(crate) instances_used: RefCell<FxIndexSet<Instance<'tcx>>>,
-    pub(crate) pgo_func_name_var_map: RefCell<FxHashMap<Instance<'tcx>, &'ll llvm::Value>>,
+    /// Associates function instances with an LLVM global that holds the
+    /// function's symbol name, as needed by LLVM coverage intrinsics.
+    ///
+    /// Instances in this map are also considered "used" for the purposes of
+    /// emitting covfun records. Every covfun record holds a hash of its
+    /// symbol name, and `llvm-cov` will exit fatally if it can't resolve that
+    /// hash back to an entry in the binary's `__llvm_prf_names` linker section.
+    pub(crate) pgo_func_name_var_map: RefCell<FxIndexMap<Instance<'tcx>, &'ll llvm::Value>>,
     pub(crate) mcdc_condition_bitmap_map: RefCell<FxHashMap<Instance<'tcx>, Vec<&'ll llvm::Value>>>,
 
     covfun_section_name: OnceCell<CString>,
@@ -31,7 +36,6 @@ pub(crate) struct CguCoverageContext<'ll, 'tcx> {
 impl<'ll, 'tcx> CguCoverageContext<'ll, 'tcx> {
     pub(crate) fn new() -> Self {
         Self {
-            instances_used: RefCell::<FxIndexSet<_>>::default(),
             pgo_func_name_var_map: Default::default(),
             mcdc_condition_bitmap_map: Default::default(),
             covfun_section_name: Default::default(),
@@ -53,6 +57,14 @@ impl<'ll, 'tcx> CguCoverageContext<'ll, 'tcx> {
             .and_then(|bitmap_map| bitmap_map.get(decision_depth as usize))
             .copied() // Dereference Option<&&Value> to Option<&Value>
     }
+
+    /// Returns the list of instances considered "used" in this CGU, as
+    /// inferred from the keys of `pgo_func_name_var_map`.
+    pub(crate) fn instances_used(&self) -> Vec<Instance<'tcx>> {
+        // Collecting into a Vec is way easier than trying to juggle RefCell
+        // projections, and this should only run once per CGU anyway.
+        self.pgo_func_name_var_map.borrow().keys().copied().collect::<Vec<_>>()
+    }
 }
 
 impl<'ll, 'tcx> CodegenCx<'ll, 'tcx> {
@@ -78,7 +90,10 @@ impl<'ll, 'tcx> CodegenCx<'ll, 'tcx> {
     /// string, to hold the function name passed to LLVM intrinsic
     /// `instrprof.increment()`. The `Value` is only created once per instance.
     /// Multiple invocations with the same instance return the same `Value`.
-    fn get_pgo_func_name_var(&self, instance: Instance<'tcx>) -> &'ll llvm::Value {
+    ///
+    /// This has the side-effect of causing coverage codegen to consider this
+    /// function "used", making it eligible to emit an associated covfun record.
+    fn ensure_pgo_func_name_var(&self, instance: Instance<'tcx>) -> &'ll llvm::Value {
         debug!("getting pgo_func_name_var for instance={:?}", instance);
         let mut pgo_func_name_var_map = self.coverage_cx().pgo_func_name_var_map.borrow_mut();
         pgo_func_name_var_map.entry(instance).or_insert_with(|| {
@@ -102,7 +117,7 @@ impl<'tcx> CoverageInfoBuilderMethods<'tcx> for Builder<'_, '_, 'tcx> {
             return;
         }
 
-        let fn_name = self.get_pgo_func_name_var(instance);
+        let fn_name = self.ensure_pgo_func_name_var(instance);
         let hash = self.const_u64(function_coverage_info.function_source_hash);
         let bitmap_bits = self.const_u32(function_coverage_info.mcdc_bitmap_bits as u32);
         self.mcdc_parameters(fn_name, hash, bitmap_bits);
@@ -151,19 +166,14 @@ impl<'tcx> CoverageInfoBuilderMethods<'tcx> for Builder<'_, '_, 'tcx> {
             return;
         };
 
-        // Mark the instance as used in this CGU, for coverage purposes.
-        // This includes functions that were not partitioned into this CGU,
-        // but were MIR-inlined into one of this CGU's functions.
-        coverage_cx.instances_used.borrow_mut().insert(instance);
-
         match *kind {
             CoverageKind::SpanMarker | CoverageKind::BlockMarker { .. } => unreachable!(
                 "marker statement {kind:?} should have been removed by CleanupPostBorrowck"
             ),
             CoverageKind::VirtualCounter { bcb }
                 if let Some(&id) = ids_info.phys_counter_for_node.get(&bcb) =>
             {
-                let fn_name = bx.get_pgo_func_name_var(instance);
+                let fn_name = bx.ensure_pgo_func_name_var(instance);
                 let hash = bx.const_u64(function_coverage_info.function_source_hash);
                 let num_counters = bx.const_u32(ids_info.num_counters);
                 let index = bx.const_u32(id.as_u32());
@@ -193,7 +203,7 @@ impl<'tcx> CoverageInfoBuilderMethods<'tcx> for Builder<'_, '_, 'tcx> {
                     "bitmap index of the decision out of range"
                 );
 
-                let fn_name = bx.get_pgo_func_name_var(instance);
+                let fn_name = bx.ensure_pgo_func_name_var(instance);
                 let hash = bx.const_u64(function_coverage_info.function_source_hash);
                 let bitmap_index = bx.const_u32(bitmap_idx);
                 bx.mcdc_tvbitmap_update(fn_name, hash, bitmap_index, cond_bitmap);

compiler/rustc_codegen_llvm/src/errors.rs

@@ -39,7 +39,7 @@ pub(crate) struct AutoDiffWithoutEnable;
 #[derive(Diagnostic)]
 #[diag(codegen_llvm_lto_bitcode_from_rlib)]
 pub(crate) struct LtoBitcodeFromRlib {
-    pub llvm_err: String,
+    pub err: String,
 }
 
 #[derive(Diagnostic)]

compiler/rustc_codegen_llvm/src/lib.rs

@@ -168,13 +168,6 @@ impl WriteBackendMethods for LlvmCodegenBackend {
         let stats = llvm::build_string(|s| unsafe { llvm::LLVMRustPrintStatistics(s) }).unwrap();
         print!("{stats}");
     }
-    fn run_link(
-        cgcx: &CodegenContext<Self>,
-        dcx: DiagCtxtHandle<'_>,
-        modules: Vec<ModuleCodegen<Self::Module>>,
-    ) -> Result<ModuleCodegen<Self::Module>, FatalError> {
-        back::write::link(cgcx, dcx, modules)
-    }
     fn run_and_optimize_fat_lto(
         cgcx: &CodegenContext<Self>,
         exported_symbols_for_lto: &[String],

compiler/rustc_codegen_llvm/src/llvm/ffi.rs

@@ -2612,13 +2612,6 @@ unsafe extern "C" {
         len: usize,
         Identifier: *const c_char,
     ) -> Option<&Module>;
-    pub(crate) fn LLVMRustGetSliceFromObjectDataByName(
-        data: *const u8,
-        len: usize,
-        name: *const u8,
-        name_len: usize,
-        out_len: &mut usize,
-    ) -> *const u8;
 
     pub(crate) fn LLVMRustLinkerNew(M: &Module) -> &mut Linker<'_>;
     pub(crate) fn LLVMRustLinkerAdd(