diff options
Diffstat (limited to 'compiler/rustc_codegen_llvm/src')
34 files changed, 1287 insertions, 858 deletions
diff --git a/compiler/rustc_codegen_llvm/src/asm.rs b/compiler/rustc_codegen_llvm/src/asm.rs index d1804cb49ad..3722d4350a2 100644 --- a/compiler/rustc_codegen_llvm/src/asm.rs +++ b/compiler/rustc_codegen_llvm/src/asm.rs @@ -442,6 +442,14 @@ impl<'tcx> AsmCodegenMethods<'tcx> for CodegenCx<'_, 'tcx> { ); } } + + fn mangled_name(&self, instance: Instance<'tcx>) -> String { + let llval = self.get_fn(instance); + llvm::build_string(|s| unsafe { + llvm::LLVMRustGetMangledName(llval, s); + }) + .expect("symbol is not valid UTF-8") + } } pub(crate) fn inline_asm_call<'ll>( @@ -504,14 +512,13 @@ pub(crate) fn inline_asm_call<'ll>( let key = "srcloc"; let kind = llvm::LLVMGetMDKindIDInContext( bx.llcx, - key.as_ptr() as *const c_char, + key.as_ptr().cast::<c_char>(), key.len() as c_uint, ); - // srcloc contains one integer for each line of assembly code. - // Unfortunately this isn't enough to encode a full span so instead - // we just encode the start position of each line. - // FIXME: Figure out a way to pass the entire line spans. + // `srcloc` contains one 64-bit integer for each line of assembly code, + // where the lower 32 bits hold the lo byte position and the upper 32 bits + // hold the hi byte position. let mut srcloc = vec![]; if dia == llvm::AsmDialect::Intel && line_spans.len() > 1 { // LLVM inserts an extra line to add the ".intel_syntax", so add @@ -521,13 +528,13 @@ pub(crate) fn inline_asm_call<'ll>( // due to the asm template string coming from a macro. LLVM will // default to the first srcloc for lines that don't have an // associated srcloc. - srcloc.push(llvm::LLVMValueAsMetadata(bx.const_i32(0))); + srcloc.push(llvm::LLVMValueAsMetadata(bx.const_u64(0))); } - srcloc.extend( - line_spans - .iter() - .map(|span| llvm::LLVMValueAsMetadata(bx.const_i32(span.lo().to_u32() as i32))), - ); + srcloc.extend(line_spans.iter().map(|span| { + llvm::LLVMValueAsMetadata(bx.const_u64( + u64::from(span.lo().to_u32()) | (u64::from(span.hi().to_u32()) << 32), + )) + })); let md = llvm::LLVMMDNodeInContext2(bx.llcx, srcloc.as_ptr(), srcloc.len()); let md = llvm::LLVMMetadataAsValue(&bx.llcx, md); llvm::LLVMSetMetadata(call, kind, md); diff --git a/compiler/rustc_codegen_llvm/src/attributes.rs b/compiler/rustc_codegen_llvm/src/attributes.rs index cb958c1d4d7..f8454fd9960 100644 --- a/compiler/rustc_codegen_llvm/src/attributes.rs +++ b/compiler/rustc_codegen_llvm/src/attributes.rs @@ -1,6 +1,6 @@ //! Set and unset common attributes on LLVM values. -use rustc_attr::{InlineAttr, InstructionSetAttr, OptimizeAttr}; +use rustc_attr_parsing::{InlineAttr, InstructionSetAttr, OptimizeAttr}; use rustc_codegen_ssa::traits::*; use rustc_hir::def_id::DefId; use rustc_middle::middle::codegen_fn_attrs::{CodegenFnAttrFlags, PatchableFunctionEntry}; @@ -395,17 +395,9 @@ pub(crate) fn llfn_attrs_from_instance<'ll, 'tcx>( to_add.push(MemoryEffects::None.create_attr(cx.llcx)); } if codegen_fn_attrs.flags.contains(CodegenFnAttrFlags::NAKED) { - to_add.push(AttributeKind::Naked.create_attr(cx.llcx)); - // HACK(jubilee): "indirect branch tracking" works by attaching prologues to functions. - // And it is a module-level attribute, so the alternative is pulling naked functions into - // new LLVM modules. Otherwise LLVM's "naked" functions come with endbr prefixes per - // https://github.com/rust-lang/rust/issues/98768 - to_add.push(AttributeKind::NoCfCheck.create_attr(cx.llcx)); - if llvm_util::get_version() < (19, 0, 0) { - // Prior to LLVM 19, branch-target-enforcement was disabled by setting the attribute to - // the string "false". Now it is disabled by absence of the attribute. - to_add.push(llvm::CreateAttrStringValue(cx.llcx, "branch-target-enforcement", "false")); - } + // do nothing; a naked function is converted into an extern function + // and a global assembly block. LLVM's support for naked functions is + // not used. } else { // Do not set sanitizer attributes for naked functions. to_add.extend(sanitize_attrs(cx, codegen_fn_attrs.no_sanitize)); diff --git a/compiler/rustc_codegen_llvm/src/back/lto.rs b/compiler/rustc_codegen_llvm/src/back/lto.rs index f6d2ec24da6..4adf99e91d0 100644 --- a/compiler/rustc_codegen_llvm/src/back/lto.rs +++ b/compiler/rustc_codegen_llvm/src/back/lto.rs @@ -604,7 +604,14 @@ pub(crate) fn run_pass_manager( debug!("running the pass manager"); let opt_stage = if thin { llvm::OptStage::ThinLTO } else { llvm::OptStage::FatLTO }; let opt_level = config.opt_level.unwrap_or(config::OptLevel::No); - unsafe { write::llvm_optimize(cgcx, dcx, module, config, opt_level, opt_stage) }?; + + // If this rustc version was build with enzyme/autodiff enabled, and if users applied the + // `#[autodiff]` macro at least once, then we will later call llvm_optimize a second time. + let first_run = true; + debug!("running llvm pm opt pipeline"); + unsafe { + write::llvm_optimize(cgcx, dcx, module, config, opt_level, opt_stage, first_run)?; + } debug!("lto done"); Ok(()) } diff --git a/compiler/rustc_codegen_llvm/src/back/owned_target_machine.rs b/compiler/rustc_codegen_llvm/src/back/owned_target_machine.rs index 44c30d22a9e..4cbd49aa44d 100644 --- a/compiler/rustc_codegen_llvm/src/back/owned_target_machine.rs +++ b/compiler/rustc_codegen_llvm/src/back/owned_target_machine.rs @@ -25,7 +25,7 @@ impl OwnedTargetMachine { model: llvm::CodeModel, reloc: llvm::RelocModel, level: llvm::CodeGenOptLevel, - use_soft_fp: bool, + float_abi: llvm::FloatAbi, function_sections: bool, data_sections: bool, unique_section_names: bool, @@ -57,7 +57,7 @@ impl OwnedTargetMachine { model, reloc, level, - use_soft_fp, + float_abi, function_sections, data_sections, unique_section_names, diff --git a/compiler/rustc_codegen_llvm/src/back/write.rs b/compiler/rustc_codegen_llvm/src/back/write.rs index bde6668929c..509b24dd703 100644 --- a/compiler/rustc_codegen_llvm/src/back/write.rs +++ b/compiler/rustc_codegen_llvm/src/back/write.rs @@ -25,10 +25,9 @@ use rustc_session::Session; use rustc_session::config::{ self, Lto, OutputType, Passes, RemapPathScopeComponents, SplitDwarfKind, SwitchWithOptPath, }; -use rustc_span::InnerSpan; -use rustc_span::symbol::sym; -use rustc_target::spec::{CodeModel, RelocModel, SanitizerSet, SplitDebuginfo, TlsModel}; -use tracing::debug; +use rustc_span::{BytePos, InnerSpan, Pos, SpanData, SyntaxContext, sym}; +use rustc_target::spec::{CodeModel, FloatAbi, RelocModel, SanitizerSet, SplitDebuginfo, TlsModel}; +use tracing::{debug, trace}; use crate::back::lto::ThinBuffer; use crate::back::owned_target_machine::OwnedTargetMachine; @@ -182,6 +181,14 @@ pub(crate) fn to_llvm_code_model(code_model: Option<CodeModel>) -> llvm::CodeMod } } +fn to_llvm_float_abi(float_abi: Option<FloatAbi>) -> llvm::FloatAbi { + match float_abi { + None => llvm::FloatAbi::Default, + Some(FloatAbi::Soft) => llvm::FloatAbi::Soft, + Some(FloatAbi::Hard) => llvm::FloatAbi::Hard, + } +} + pub(crate) fn target_machine_factory( sess: &Session, optlvl: config::OptLevel, @@ -190,12 +197,12 @@ pub(crate) fn target_machine_factory( let reloc_model = to_llvm_relocation_model(sess.relocation_model()); let (opt_level, _) = to_llvm_opt_settings(optlvl); - let use_softfp = if sess.target.arch == "arm" && sess.target.abi == "eabihf" { - sess.opts.cg.soft_float + let float_abi = if sess.target.arch == "arm" && sess.opts.cg.soft_float { + llvm::FloatAbi::Soft } else { // `validate_commandline_args_with_session_available` has already warned about this being // ignored. Let's make sure LLVM doesn't suddenly start using this flag on more targets. - false + to_llvm_float_abi(sess.target.llvm_floatabi) }; let ffunction_sections = @@ -291,7 +298,7 @@ pub(crate) fn target_machine_factory( code_model, reloc_model, opt_level, - use_softfp, + float_abi, ffunction_sections, fdata_sections, funique_section_names, @@ -415,21 +422,32 @@ fn report_inline_asm( cgcx: &CodegenContext<LlvmCodegenBackend>, msg: String, level: llvm::DiagnosticLevel, - mut cookie: u64, + cookie: u64, source: Option<(String, Vec<InnerSpan>)>, ) { // In LTO build we may get srcloc values from other crates which are invalid // since they use a different source map. To be safe we just suppress these // in LTO builds. - if matches!(cgcx.lto, Lto::Fat | Lto::Thin) { - cookie = 0; - } + let span = if cookie == 0 || matches!(cgcx.lto, Lto::Fat | Lto::Thin) { + SpanData::default() + } else { + let lo = BytePos::from_u32(cookie as u32); + let hi = BytePos::from_u32((cookie >> 32) as u32); + SpanData { + lo, + // LLVM version < 19 silently truncates the cookie to 32 bits in some situations. + hi: if hi.to_u32() != 0 { hi } else { lo }, + ctxt: SyntaxContext::root(), + parent: None, + } + }; let level = match level { llvm::DiagnosticLevel::Error => Level::Error, llvm::DiagnosticLevel::Warning => Level::Warning, llvm::DiagnosticLevel::Note | llvm::DiagnosticLevel::Remark => Level::Note, }; - cgcx.diag_emitter.inline_asm_error(cookie.try_into().unwrap(), msg, level, source); + let msg = msg.strip_prefix("error: ").unwrap_or(&msg).to_string(); + cgcx.diag_emitter.inline_asm_error(span, msg, level, source); } unsafe extern "C" fn diagnostic_handler(info: &DiagnosticInfo, user: *mut c_void) { @@ -519,9 +537,35 @@ pub(crate) unsafe fn llvm_optimize( config: &ModuleConfig, opt_level: config::OptLevel, opt_stage: llvm::OptStage, + skip_size_increasing_opts: bool, ) -> Result<(), FatalError> { - let unroll_loops = - opt_level != config::OptLevel::Size && opt_level != config::OptLevel::SizeMin; + // Enzyme: + // The whole point of compiler based AD is to differentiate optimized IR instead of unoptimized + // source code. However, benchmarks show that optimizations increasing the code size + // tend to reduce AD performance. Therefore deactivate them before AD, then differentiate the code + // and finally re-optimize the module, now with all optimizations available. + // FIXME(ZuseZ4): In a future update we could figure out how to only optimize individual functions getting + // differentiated. + + let unroll_loops; + let vectorize_slp; + let vectorize_loop; + + // When we build rustc with enzyme/autodiff support, we want to postpone size-increasing + // optimizations until after differentiation. FIXME(ZuseZ4): Before shipping on nightly, + // we should make this more granular, or at least check that the user has at least one autodiff + // call in their code, to justify altering the compilation pipeline. + if skip_size_increasing_opts && cfg!(llvm_enzyme) { + unroll_loops = false; + vectorize_slp = false; + vectorize_loop = false; + } else { + unroll_loops = + opt_level != config::OptLevel::Size && opt_level != config::OptLevel::SizeMin; + vectorize_slp = config.vectorize_slp; + vectorize_loop = config.vectorize_loop; + } + trace!(?unroll_loops, ?vectorize_slp, ?vectorize_loop); let using_thin_buffers = opt_stage == llvm::OptStage::PreLinkThinLTO || config.bitcode_needed(); let pgo_gen_path = get_pgo_gen_path(config); let pgo_use_path = get_pgo_use_path(config); @@ -585,8 +629,8 @@ pub(crate) unsafe fn llvm_optimize( using_thin_buffers, config.merge_functions, unroll_loops, - config.vectorize_slp, - config.vectorize_loop, + vectorize_slp, + vectorize_loop, config.no_builtins, config.emit_lifetime_markers, sanitizer_options.as_ref(), @@ -630,6 +674,8 @@ pub(crate) unsafe fn optimize( unsafe { llvm::LLVMWriteBitcodeToFile(llmod, out.as_ptr()) }; } + // FIXME(ZuseZ4): support SanitizeHWAddress and prevent illegal/unsupported opts + if let Some(opt_level) = config.opt_level { let opt_stage = match cgcx.lto { Lto::Fat => llvm::OptStage::PreLinkFatLTO, @@ -637,7 +683,20 @@ pub(crate) unsafe fn optimize( _ if cgcx.opts.cg.linker_plugin_lto.enabled() => llvm::OptStage::PreLinkThinLTO, _ => llvm::OptStage::PreLinkNoLTO, }; - return unsafe { llvm_optimize(cgcx, dcx, module, config, opt_level, opt_stage) }; + + // If we know that we will later run AD, then we disable vectorization and loop unrolling + let skip_size_increasing_opts = cfg!(llvm_enzyme); + return unsafe { + llvm_optimize( + cgcx, + dcx, + module, + config, + opt_level, + opt_stage, + skip_size_increasing_opts, + ) + }; } Ok(()) } diff --git a/compiler/rustc_codegen_llvm/src/base.rs b/compiler/rustc_codegen_llvm/src/base.rs index f62310bd948..d05faf5577b 100644 --- a/compiler/rustc_codegen_llvm/src/base.rs +++ b/compiler/rustc_codegen_llvm/src/base.rs @@ -23,7 +23,7 @@ use rustc_middle::middle::codegen_fn_attrs::CodegenFnAttrs; use rustc_middle::mir::mono::{Linkage, Visibility}; use rustc_middle::ty::TyCtxt; use rustc_session::config::DebugInfo; -use rustc_span::symbol::Symbol; +use rustc_span::Symbol; use rustc_target::spec::SanitizerSet; use super::ModuleLlvm; diff --git a/compiler/rustc_codegen_llvm/src/builder.rs b/compiler/rustc_codegen_llvm/src/builder.rs index b5bb7630ca6..5a34b52e6ef 100644 --- a/compiler/rustc_codegen_llvm/src/builder.rs +++ b/compiler/rustc_codegen_llvm/src/builder.rs @@ -2,6 +2,8 @@ use std::borrow::Cow; use std::ops::Deref; use std::{iter, ptr}; +pub(crate) mod autodiff; + use libc::{c_char, c_uint}; use rustc_abi as abi; use rustc_abi::{Align, Size, WrappingRange}; @@ -608,14 +610,6 @@ impl<'a, 'll, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> { } fn range_metadata(&mut self, load: &'ll Value, range: WrappingRange) { - if self.sess().target.arch == "amdgpu" { - // amdgpu/LLVM does something weird and thinks an i64 value is - // split into a v2i32, halving the bitwidth LLVM expects, - // tripping an assertion. So, for now, just disable this - // optimization. - return; - } - if self.cx.sess().opts.optimize == OptLevel::No { // Don't emit metadata we're not going to use return; diff --git a/compiler/rustc_codegen_llvm/src/builder/autodiff.rs b/compiler/rustc_codegen_llvm/src/builder/autodiff.rs new file mode 100644 index 00000000000..38f7eaa090f --- /dev/null +++ b/compiler/rustc_codegen_llvm/src/builder/autodiff.rs @@ -0,0 +1,344 @@ +use std::ptr; + +use rustc_ast::expand::autodiff_attrs::{AutoDiffAttrs, AutoDiffItem, DiffActivity, DiffMode}; +use rustc_codegen_ssa::ModuleCodegen; +use rustc_codegen_ssa::back::write::ModuleConfig; +use rustc_codegen_ssa::traits::{BaseTypeCodegenMethods, BuilderMethods}; +use rustc_errors::FatalError; +use rustc_middle::ty::TyCtxt; +use rustc_session::config::Lto; +use tracing::{debug, trace}; + +use crate::back::write::{llvm_err, llvm_optimize}; +use crate::builder::Builder; +use crate::declare::declare_raw_fn; +use crate::errors::LlvmError; +use crate::llvm::AttributePlace::Function; +use crate::llvm::{Metadata, True}; +use crate::value::Value; +use crate::{CodegenContext, LlvmCodegenBackend, ModuleLlvm, attributes, context, llvm}; + +fn get_params(fnc: &Value) -> Vec<&Value> { + unsafe { + let param_num = llvm::LLVMCountParams(fnc) as usize; + let mut fnc_args: Vec<&Value> = vec![]; + fnc_args.reserve(param_num); + llvm::LLVMGetParams(fnc, fnc_args.as_mut_ptr()); + fnc_args.set_len(param_num); + fnc_args + } +} + +/// When differentiating `fn_to_diff`, take a `outer_fn` and generate another +/// function with expected naming and calling conventions[^1] which will be +/// discovered by the enzyme LLVM pass and its body populated with the differentiated +/// `fn_to_diff`. `outer_fn` is then modified to have a call to the generated +/// function and handle the differences between the Rust calling convention and +/// Enzyme. +/// [^1]: <https://enzyme.mit.edu/getting_started/CallingConvention/> +// FIXME(ZuseZ4): `outer_fn` should include upstream safety checks to +// cover some assumptions of enzyme/autodiff, which could lead to UB otherwise. +fn generate_enzyme_call<'ll, 'tcx>( + cx: &context::CodegenCx<'ll, 'tcx>, + fn_to_diff: &'ll Value, + outer_fn: &'ll Value, + attrs: AutoDiffAttrs, +) { + let inputs = attrs.input_activity; + let output = attrs.ret_activity; + + // We have to pick the name depending on whether we want forward or reverse mode autodiff. + // FIXME(ZuseZ4): The new pass based approach should not need the {Forward/Reverse}First method anymore, since + // it will handle higher-order derivatives correctly automatically (in theory). Currently + // higher-order derivatives fail, so we should debug that before adjusting this code. + let mut ad_name: String = match attrs.mode { + DiffMode::Forward => "__enzyme_fwddiff", + DiffMode::Reverse => "__enzyme_autodiff", + DiffMode::ForwardFirst => "__enzyme_fwddiff", + DiffMode::ReverseFirst => "__enzyme_autodiff", + _ => panic!("logic bug in autodiff, unrecognized mode"), + } + .to_string(); + + // add outer_fn name to ad_name to make it unique, in case users apply autodiff to multiple + // functions. Unwrap will only panic, if LLVM gave us an invalid string. + let name = llvm::get_value_name(outer_fn); + let outer_fn_name = std::ffi::CStr::from_bytes_with_nul(name).unwrap().to_str().unwrap(); + ad_name.push_str(outer_fn_name.to_string().as_str()); + + // Let us assume the user wrote the following function square: + // + // ```llvm + // define double @square(double %x) { + // entry: + // %0 = fmul double %x, %x + // ret double %0 + // } + // ``` + // + // The user now applies autodiff to the function square, in which case fn_to_diff will be `square`. + // Our macro generates the following placeholder code (slightly simplified): + // + // ```llvm + // define double @dsquare(double %x) { + // ; placeholder code + // return 0.0; + // } + // ``` + // + // so our `outer_fn` will be `dsquare`. The unsafe code section below now removes the placeholder + // code and inserts an autodiff call. We also add a declaration for the __enzyme_autodiff call. + // Again, the arguments to all functions are slightly simplified. + // ```llvm + // declare double @__enzyme_autodiff_square(...) + // + // define double @dsquare(double %x) { + // entry: + // %0 = tail call double (...) @__enzyme_autodiff_square(double (double)* nonnull @square, double %x) + // ret double %0 + // } + // ``` + unsafe { + // On LLVM-IR, we can luckily declare __enzyme_ functions without specifying the input + // arguments. We do however need to declare them with their correct return type. + // We already figured the correct return type out in our frontend, when generating the outer_fn, + // so we can now just go ahead and use that. FIXME(ZuseZ4): This doesn't handle sret yet. + let fn_ty = llvm::LLVMGlobalGetValueType(outer_fn); + let ret_ty = llvm::LLVMGetReturnType(fn_ty); + + // LLVM can figure out the input types on it's own, so we take a shortcut here. + let enzyme_ty = llvm::LLVMFunctionType(ret_ty, ptr::null(), 0, True); + + //FIXME(ZuseZ4): the CC/Addr/Vis values are best effort guesses, we should look at tests and + // think a bit more about what should go here. + let cc = llvm::LLVMGetFunctionCallConv(outer_fn); + let ad_fn = declare_raw_fn( + cx, + &ad_name, + llvm::CallConv::try_from(cc).expect("invalid callconv"), + llvm::UnnamedAddr::No, + llvm::Visibility::Default, + enzyme_ty, + ); + + // Otherwise LLVM might inline our temporary code before the enzyme pass has a chance to + // do it's work. + let attr = llvm::AttributeKind::NoInline.create_attr(cx.llcx); + attributes::apply_to_llfn(ad_fn, Function, &[attr]); + + // first, remove all calls from fnc + let entry = llvm::LLVMGetFirstBasicBlock(outer_fn); + let br = llvm::LLVMRustGetTerminator(entry); + llvm::LLVMRustEraseInstFromParent(br); + + let last_inst = llvm::LLVMRustGetLastInstruction(entry).unwrap(); + let mut builder = Builder::build(cx, entry); + + let num_args = llvm::LLVMCountParams(&fn_to_diff); + let mut args = Vec::with_capacity(num_args as usize + 1); + args.push(fn_to_diff); + + let enzyme_const = cx.create_metadata("enzyme_const".to_string()).unwrap(); + let enzyme_out = cx.create_metadata("enzyme_out".to_string()).unwrap(); + let enzyme_dup = cx.create_metadata("enzyme_dup".to_string()).unwrap(); + let enzyme_dupnoneed = cx.create_metadata("enzyme_dupnoneed".to_string()).unwrap(); + let enzyme_primal_ret = cx.create_metadata("enzyme_primal_return".to_string()).unwrap(); + + match output { + DiffActivity::Dual => { + args.push(cx.get_metadata_value(enzyme_primal_ret)); + } + DiffActivity::Active => { + args.push(cx.get_metadata_value(enzyme_primal_ret)); + } + _ => {} + } + + trace!("matching autodiff arguments"); + // We now handle the issue that Rust level arguments not always match the llvm-ir level + // arguments. A slice, `&[f32]`, for example, is represented as a pointer and a length on + // llvm-ir level. The number of activities matches the number of Rust level arguments, so we + // need to match those. + // FIXME(ZuseZ4): This logic is a bit more complicated than it should be, can we simplify it + // using iterators and peek()? + let mut outer_pos: usize = 0; + let mut activity_pos = 0; + let outer_args: Vec<&llvm::Value> = get_params(outer_fn); + while activity_pos < inputs.len() { + let activity = inputs[activity_pos as usize]; + // Duplicated arguments received a shadow argument, into which enzyme will write the + // gradient. + let (activity, duplicated): (&Metadata, bool) = match activity { + DiffActivity::None => panic!("not a valid input activity"), + DiffActivity::Const => (enzyme_const, false), + DiffActivity::Active => (enzyme_out, false), + DiffActivity::ActiveOnly => (enzyme_out, false), + DiffActivity::Dual => (enzyme_dup, true), + DiffActivity::DualOnly => (enzyme_dupnoneed, true), + DiffActivity::Duplicated => (enzyme_dup, true), + DiffActivity::DuplicatedOnly => (enzyme_dupnoneed, true), + DiffActivity::FakeActivitySize => (enzyme_const, false), + }; + let outer_arg = outer_args[outer_pos]; + args.push(cx.get_metadata_value(activity)); + args.push(outer_arg); + if duplicated { + // We know that duplicated args by construction have a following argument, + // so this can not be out of bounds. + let next_outer_arg = outer_args[outer_pos + 1]; + let next_outer_ty = cx.val_ty(next_outer_arg); + // FIXME(ZuseZ4): We should add support for Vec here too, but it's less urgent since + // vectors behind references (&Vec<T>) are already supported. Users can not pass a + // Vec by value for reverse mode, so this would only help forward mode autodiff. + let slice = { + if activity_pos + 1 >= inputs.len() { + // If there is no arg following our ptr, it also can't be a slice, + // since that would lead to a ptr, int pair. + false + } else { + let next_activity = inputs[activity_pos + 1]; + // We analyze the MIR types and add this dummy activity if we visit a slice. + next_activity == DiffActivity::FakeActivitySize + } + }; + if slice { + // A duplicated slice will have the following two outer_fn arguments: + // (..., ptr1, int1, ptr2, int2, ...). We add the following llvm-ir to our __enzyme call: + // (..., metadata! enzyme_dup, ptr, ptr, int1, ...). + // FIXME(ZuseZ4): We will upstream a safety check later which asserts that + // int2 >= int1, which means the shadow vector is large enough to store the gradient. + assert!(llvm::LLVMRustGetTypeKind(next_outer_ty) == llvm::TypeKind::Integer); + let next_outer_arg2 = outer_args[outer_pos + 2]; + let next_outer_ty2 = cx.val_ty(next_outer_arg2); + assert!(llvm::LLVMRustGetTypeKind(next_outer_ty2) == llvm::TypeKind::Pointer); + let next_outer_arg3 = outer_args[outer_pos + 3]; + let next_outer_ty3 = cx.val_ty(next_outer_arg3); + assert!(llvm::LLVMRustGetTypeKind(next_outer_ty3) == llvm::TypeKind::Integer); + args.push(next_outer_arg2); + args.push(cx.get_metadata_value(enzyme_const)); + args.push(next_outer_arg); + outer_pos += 4; + activity_pos += 2; + } else { + // A duplicated pointer will have the following two outer_fn arguments: + // (..., ptr, ptr, ...). We add the following llvm-ir to our __enzyme call: + // (..., metadata! enzyme_dup, ptr, ptr, ...). + assert!(llvm::LLVMRustGetTypeKind(next_outer_ty) == llvm::TypeKind::Pointer); + args.push(next_outer_arg); + outer_pos += 2; + activity_pos += 1; + } + } else { + // We do not differentiate with resprect to this argument. + // We already added the metadata and argument above, so just increase the counters. + outer_pos += 1; + activity_pos += 1; + } + } + + let call = builder.call(enzyme_ty, None, None, ad_fn, &args, None, None); + + // This part is a bit iffy. LLVM requires that a call to an inlineable function has some + // metadata attachted to it, but we just created this code oota. Given that the + // differentiated function already has partly confusing metadata, and given that this + // affects nothing but the auttodiff IR, we take a shortcut and just steal metadata from the + // dummy code which we inserted at a higher level. + // FIXME(ZuseZ4): Work with Enzyme core devs to clarify what debug metadata issues we have, + // and how to best improve it for enzyme core and rust-enzyme. + let md_ty = cx.get_md_kind_id("dbg"); + if llvm::LLVMRustHasMetadata(last_inst, md_ty) { + let md = llvm::LLVMRustDIGetInstMetadata(last_inst) + .expect("failed to get instruction metadata"); + let md_todiff = cx.get_metadata_value(md); + llvm::LLVMSetMetadata(call, md_ty, md_todiff); + } else { + // We don't panic, since depending on whether we are in debug or release mode, we might + // have no debug info to copy, which would then be ok. + trace!("no dbg info"); + } + // Now that we copied the metadata, get rid of dummy code. + llvm::LLVMRustEraseInstBefore(entry, last_inst); + llvm::LLVMRustEraseInstFromParent(last_inst); + + if cx.val_ty(outer_fn) != cx.type_void() { + builder.ret(call); + } else { + builder.ret_void(); + } + + // Let's crash in case that we messed something up above and generated invalid IR. + llvm::LLVMRustVerifyFunction( + outer_fn, + llvm::LLVMRustVerifierFailureAction::LLVMAbortProcessAction, + ); + } +} + +pub(crate) fn differentiate<'ll, 'tcx>( + module: &'ll ModuleCodegen<ModuleLlvm>, + cgcx: &CodegenContext<LlvmCodegenBackend>, + tcx: TyCtxt<'tcx>, + diff_items: Vec<AutoDiffItem>, + config: &ModuleConfig, +) -> Result<(), FatalError> { + for item in &diff_items { + trace!("{}", item); + } + + let diag_handler = cgcx.create_dcx(); + let (_, cgus) = tcx.collect_and_partition_mono_items(()); + let cx = context::CodegenCx::new(tcx, &cgus.first().unwrap(), &module.module_llvm); + + // Before dumping the module, we want all the TypeTrees to become part of the module. + for item in diff_items.iter() { + let name = item.source.clone(); + let fn_def: Option<&llvm::Value> = cx.get_function(&name); + let Some(fn_def) = fn_def else { + return Err(llvm_err(diag_handler.handle(), LlvmError::PrepareAutoDiff { + src: item.source.clone(), + target: item.target.clone(), + error: "could not find source function".to_owned(), + })); + }; + debug!(?item.target); + let fn_target: Option<&llvm::Value> = cx.get_function(&item.target); + let Some(fn_target) = fn_target else { + return Err(llvm_err(diag_handler.handle(), LlvmError::PrepareAutoDiff { + src: item.source.clone(), + target: item.target.clone(), + error: "could not find target function".to_owned(), + })); + }; + + generate_enzyme_call(&cx, fn_def, fn_target, item.attrs.clone()); + } + + // FIXME(ZuseZ4): support SanitizeHWAddress and prevent illegal/unsupported opts + + if let Some(opt_level) = config.opt_level { + let opt_stage = match cgcx.lto { + Lto::Fat => llvm::OptStage::PreLinkFatLTO, + Lto::Thin | Lto::ThinLocal => llvm::OptStage::PreLinkThinLTO, + _ if cgcx.opts.cg.linker_plugin_lto.enabled() => llvm::OptStage::PreLinkThinLTO, + _ => llvm::OptStage::PreLinkNoLTO, + }; + // This is our second opt call, so now we run all opts, + // to make sure we get the best performance. + let skip_size_increasing_opts = false; + trace!("running Module Optimization after differentiation"); + unsafe { + llvm_optimize( + cgcx, + diag_handler.handle(), + module, + config, + opt_level, + opt_stage, + skip_size_increasing_opts, + )? + }; + } + trace!("done with differentiate()"); + + Ok(()) +} diff --git a/compiler/rustc_codegen_llvm/src/callee.rs b/compiler/rustc_codegen_llvm/src/callee.rs index ec77f32caf4..aa9a0f34f55 100644 --- a/compiler/rustc_codegen_llvm/src/callee.rs +++ b/compiler/rustc_codegen_llvm/src/callee.rs @@ -106,7 +106,7 @@ pub(crate) fn get_fn<'ll, 'tcx>(cx: &CodegenCx<'ll, 'tcx>, instance: Instance<'t // This is a monomorphization of a generic function. if !(cx.tcx.sess.opts.share_generics() || tcx.codegen_fn_attrs(instance_def_id).inline - == rustc_attr::InlineAttr::Never) + == rustc_attr_parsing::InlineAttr::Never) { // When not sharing generics, all instances are in the same // crate and have hidden visibility. diff --git a/compiler/rustc_codegen_llvm/src/common.rs b/compiler/rustc_codegen_llvm/src/common.rs index 8852dec7d9f..adfe8aeb5c5 100644 --- a/compiler/rustc_codegen_llvm/src/common.rs +++ b/compiler/rustc_codegen_llvm/src/common.rs @@ -302,10 +302,9 @@ impl<'ll, 'tcx> ConstCodegenMethods<'tcx> for CodegenCx<'ll, 'tcx> { (value, AddressSpace::DATA) } } - GlobalAlloc::Function { instance, .. } => ( - self.get_fn_addr(instance.polymorphize(self.tcx)), - self.data_layout().instruction_address_space, - ), + GlobalAlloc::Function { instance, .. } => { + (self.get_fn_addr(instance), self.data_layout().instruction_address_space) + } GlobalAlloc::VTable(ty, dyn_ty) => { let alloc = self .tcx diff --git a/compiler/rustc_codegen_llvm/src/context.rs b/compiler/rustc_codegen_llvm/src/context.rs index 8218126ea29..d8fbe51b975 100644 --- a/compiler/rustc_codegen_llvm/src/context.rs +++ b/compiler/rustc_codegen_llvm/src/context.rs @@ -1,6 +1,6 @@ use std::borrow::Borrow; use std::cell::{Cell, RefCell}; -use std::ffi::{CStr, c_uint}; +use std::ffi::{CStr, c_char, c_uint}; use std::str; use rustc_abi::{HasDataLayout, TargetDataLayout, VariantIdx}; @@ -159,6 +159,16 @@ pub(crate) unsafe fn create_module<'ll>( // See https://github.com/llvm/llvm-project/pull/112084 target_data_layout = target_data_layout.replace("-i128:128", ""); } + if sess.target.arch.starts_with("powerpc64") { + // LLVM 20 updates the powerpc64 layout to correctly align 128 bit integers to 128 bit. + // See https://github.com/llvm/llvm-project/pull/118004 + target_data_layout = target_data_layout.replace("-i128:128", ""); + } + if sess.target.arch.starts_with("wasm32") || sess.target.arch.starts_with("wasm64") { + // LLVM 20 updates the wasm(32|64) layout to correctly align 128 bit integers to 128 bit. + // See https://github.com/llvm/llvm-project/pull/119204 + target_data_layout = target_data_layout.replace("-i128:128", ""); + } } // Ensure the data-layout values hardcoded remain the defaults. @@ -590,6 +600,31 @@ impl<'ll, 'tcx> CodegenCx<'ll, 'tcx> { llvm::set_section(g, c"llvm.metadata"); } } + + pub(crate) fn get_metadata_value(&self, metadata: &'ll Metadata) -> &'ll Value { + unsafe { llvm::LLVMMetadataAsValue(self.llcx, metadata) } + } + + pub(crate) fn get_function(&self, name: &str) -> Option<&'ll Value> { + let name = SmallCStr::new(name); + unsafe { llvm::LLVMGetNamedFunction(self.llmod, name.as_ptr()) } + } + + pub(crate) fn get_md_kind_id(&self, name: &str) -> u32 { + unsafe { + llvm::LLVMGetMDKindIDInContext( + self.llcx, + name.as_ptr() as *const c_char, + name.len() as c_uint, + ) + } + } + + pub(crate) fn create_metadata(&self, name: String) -> Option<&'ll Metadata> { + Some(unsafe { + llvm::LLVMMDStringInContext2(self.llcx, name.as_ptr() as *const c_char, name.len()) + }) + } } impl<'ll, 'tcx> MiscCodegenMethods<'tcx> for CodegenCx<'ll, 'tcx> { diff --git a/compiler/rustc_codegen_llvm/src/coverageinfo/ffi.rs b/compiler/rustc_codegen_llvm/src/coverageinfo/ffi.rs index a6e07ea2a60..b617f4d37f5 100644 --- a/compiler/rustc_codegen_llvm/src/coverageinfo/ffi.rs +++ b/compiler/rustc_codegen_llvm/src/coverageinfo/ffi.rs @@ -1,6 +1,4 @@ -use rustc_middle::mir::coverage::{CounterId, CovTerm, ExpressionId, SourceRegion}; - -use crate::coverageinfo::mapgen::LocalFileId; +use rustc_middle::mir::coverage::{CounterId, CovTerm, ExpressionId}; /// Must match the layout of `LLVMRustCounterKind`. #[derive(Copy, Clone, Debug)] @@ -126,29 +124,43 @@ pub(crate) struct CoverageSpan { /// Local index into the function's local-to-global file ID table. /// The value at that index is itself an index into the coverage filename /// table in the CGU's `__llvm_covmap` section. - file_id: u32, + pub(crate) file_id: u32, /// 1-based starting line of the source code span. - start_line: u32, + pub(crate) start_line: u32, /// 1-based starting column of the source code span. - start_col: u32, + pub(crate) start_col: u32, /// 1-based ending line of the source code span. - end_line: u32, + pub(crate) end_line: u32, /// 1-based ending column of the source code span. High bit must be unset. - end_col: u32, + pub(crate) end_col: u32, +} + +/// Holds tables of the various region types in one struct. +/// +/// Don't pass this struct across FFI; pass the individual region tables as +/// pointer/length pairs instead. +/// +/// Each field name has a `_regions` suffix for improved readability after +/// exhaustive destructing, which ensures that all region types are handled. +#[derive(Clone, Debug, Default)] +pub(crate) struct Regions { + pub(crate) code_regions: Vec<CodeRegion>, + pub(crate) branch_regions: Vec<BranchRegion>, + pub(crate) mcdc_branch_regions: Vec<MCDCBranchRegion>, + pub(crate) mcdc_decision_regions: Vec<MCDCDecisionRegion>, } -impl CoverageSpan { - pub(crate) fn from_source_region( - local_file_id: LocalFileId, - code_region: &SourceRegion, - ) -> Self { - let file_id = local_file_id.as_u32(); - let &SourceRegion { start_line, start_col, end_line, end_col } = code_region; - // Internally, LLVM uses the high bit of `end_col` to distinguish between - // code regions and gap regions, so it can't be used by the column number. - assert!(end_col & (1u32 << 31) == 0, "high bit of `end_col` must be unset: {end_col:#X}"); - Self { file_id, start_line, start_col, end_line, end_col } +impl Regions { + /// Returns true if none of this structure's tables contain any regions. + pub(crate) fn has_no_regions(&self) -> bool { + let Self { code_regions, branch_regions, mcdc_branch_regions, mcdc_decision_regions } = + self; + + code_regions.is_empty() + && branch_regions.is_empty() + && mcdc_branch_regions.is_empty() + && mcdc_decision_regions.is_empty() } } @@ -156,7 +168,7 @@ impl CoverageSpan { #[derive(Clone, Debug)] #[repr(C)] pub(crate) struct CodeRegion { - pub(crate) span: CoverageSpan, + pub(crate) cov_span: CoverageSpan, pub(crate) counter: Counter, } @@ -164,7 +176,7 @@ pub(crate) struct CodeRegion { #[derive(Clone, Debug)] #[repr(C)] pub(crate) struct BranchRegion { - pub(crate) span: CoverageSpan, + pub(crate) cov_span: CoverageSpan, pub(crate) true_counter: Counter, pub(crate) false_counter: Counter, } @@ -173,7 +185,7 @@ pub(crate) struct BranchRegion { #[derive(Clone, Debug)] #[repr(C)] pub(crate) struct MCDCBranchRegion { - pub(crate) span: CoverageSpan, + pub(crate) cov_span: CoverageSpan, pub(crate) true_counter: Counter, pub(crate) false_counter: Counter, pub(crate) mcdc_branch_params: mcdc::BranchParameters, @@ -183,6 +195,6 @@ pub(crate) struct MCDCBranchRegion { #[derive(Clone, Debug)] #[repr(C)] pub(crate) struct MCDCDecisionRegion { - pub(crate) span: CoverageSpan, + pub(crate) cov_span: CoverageSpan, pub(crate) mcdc_decision_params: mcdc::DecisionParameters, } diff --git a/compiler/rustc_codegen_llvm/src/coverageinfo/llvm_cov.rs b/compiler/rustc_codegen_llvm/src/coverageinfo/llvm_cov.rs index 99c2d12b261..2cd7fa3225a 100644 --- a/compiler/rustc_codegen_llvm/src/coverageinfo/llvm_cov.rs +++ b/compiler/rustc_codegen_llvm/src/coverageinfo/llvm_cov.rs @@ -40,11 +40,10 @@ pub(crate) fn create_pgo_func_name_var<'ll>( } } -pub(crate) fn write_filenames_to_buffer<'a>( - filenames: impl IntoIterator<Item = &'a str>, -) -> Vec<u8> { +pub(crate) fn write_filenames_to_buffer(filenames: &[impl AsRef<str>]) -> Vec<u8> { let (pointers, lengths) = filenames .into_iter() + .map(AsRef::as_ref) .map(|s: &str| (s.as_c_char_ptr(), s.len())) .unzip::<_, _, Vec<_>, Vec<_>>(); @@ -62,11 +61,10 @@ pub(crate) fn write_filenames_to_buffer<'a>( pub(crate) fn write_function_mappings_to_buffer( virtual_file_mapping: &[u32], expressions: &[ffi::CounterExpression], - code_regions: &[ffi::CodeRegion], - branch_regions: &[ffi::BranchRegion], - mcdc_branch_regions: &[ffi::MCDCBranchRegion], - mcdc_decision_regions: &[ffi::MCDCDecisionRegion], + regions: &ffi::Regions, ) -> Vec<u8> { + let ffi::Regions { code_regions, branch_regions, mcdc_branch_regions, mcdc_decision_regions } = + regions; llvm::build_byte_buffer(|buffer| unsafe { llvm::LLVMRustCoverageWriteFunctionMappingsToBuffer( virtual_file_mapping.as_ptr(), diff --git a/compiler/rustc_codegen_llvm/src/coverageinfo/map_data.rs b/compiler/rustc_codegen_llvm/src/coverageinfo/map_data.rs deleted file mode 100644 index 0752c718c70..00000000000 --- a/compiler/rustc_codegen_llvm/src/coverageinfo/map_data.rs +++ /dev/null @@ -1,232 +0,0 @@ -use rustc_data_structures::captures::Captures; -use rustc_data_structures::fx::FxIndexSet; -use rustc_index::bit_set::BitSet; -use rustc_middle::mir::CoverageIdsInfo; -use rustc_middle::mir::coverage::{ - CounterId, CovTerm, Expression, ExpressionId, FunctionCoverageInfo, Mapping, MappingKind, Op, - SourceRegion, -}; -use rustc_middle::ty::Instance; -use tracing::debug; - -use crate::coverageinfo::ffi::{Counter, CounterExpression, ExprKind}; - -/// Holds all of the coverage mapping data associated with a function instance, -/// collected during traversal of `Coverage` statements in the function's MIR. -#[derive(Debug)] -pub(crate) struct FunctionCoverageCollector<'tcx> { - /// Coverage info that was attached to this function by the instrumentor. - function_coverage_info: &'tcx FunctionCoverageInfo, - ids_info: &'tcx CoverageIdsInfo, - is_used: bool, -} - -impl<'tcx> FunctionCoverageCollector<'tcx> { - /// Creates a new set of coverage data for a used (called) function. - pub(crate) fn new( - instance: Instance<'tcx>, - function_coverage_info: &'tcx FunctionCoverageInfo, - ids_info: &'tcx CoverageIdsInfo, - ) -> Self { - Self::create(instance, function_coverage_info, ids_info, true) - } - - /// Creates a new set of coverage data for an unused (never called) function. - pub(crate) fn unused( - instance: Instance<'tcx>, - function_coverage_info: &'tcx FunctionCoverageInfo, - ids_info: &'tcx CoverageIdsInfo, - ) -> Self { - Self::create(instance, function_coverage_info, ids_info, false) - } - - fn create( - instance: Instance<'tcx>, - function_coverage_info: &'tcx FunctionCoverageInfo, - ids_info: &'tcx CoverageIdsInfo, - is_used: bool, - ) -> Self { - let num_counters = function_coverage_info.num_counters; - let num_expressions = function_coverage_info.expressions.len(); - debug!( - "FunctionCoverage::create(instance={instance:?}) has \ - num_counters={num_counters}, num_expressions={num_expressions}, is_used={is_used}" - ); - - Self { function_coverage_info, ids_info, is_used } - } - - /// Identify expressions that will always have a value of zero, and note - /// their IDs in [`ZeroExpressions`]. Mappings that refer to a zero expression - /// can instead become mappings to a constant zero value. - /// - /// This method mainly exists to preserve the simplifications that were - /// already being performed by the Rust-side expression renumbering, so that - /// the resulting coverage mappings don't get worse. - fn identify_zero_expressions(&self) -> ZeroExpressions { - // The set of expressions that either were optimized out entirely, or - // have zero as both of their operands, and will therefore always have - // a value of zero. Other expressions that refer to these as operands - // can have those operands replaced with `CovTerm::Zero`. - let mut zero_expressions = ZeroExpressions::default(); - - // Simplify a copy of each expression based on lower-numbered expressions, - // and then update the set of always-zero expressions if necessary. - // (By construction, expressions can only refer to other expressions - // that have lower IDs, so one pass is sufficient.) - for (id, expression) in self.function_coverage_info.expressions.iter_enumerated() { - if !self.is_used || !self.ids_info.expressions_seen.contains(id) { - // If an expression was not seen, it must have been optimized away, - // so any operand that refers to it can be replaced with zero. - zero_expressions.insert(id); - continue; - } - - // We don't need to simplify the actual expression data in the - // expressions list; we can just simplify a temporary copy and then - // use that to update the set of always-zero expressions. - let Expression { mut lhs, op, mut rhs } = *expression; - - // If an expression has an operand that is also an expression, the - // operand's ID must be strictly lower. This is what lets us find - // all zero expressions in one pass. - let assert_operand_expression_is_lower = |operand_id: ExpressionId| { - assert!( - operand_id < id, - "Operand {operand_id:?} should be less than {id:?} in {expression:?}", - ) - }; - - // If an operand refers to a counter or expression that is always - // zero, then that operand can be replaced with `CovTerm::Zero`. - let maybe_set_operand_to_zero = |operand: &mut CovTerm| { - if let CovTerm::Expression(id) = *operand { - assert_operand_expression_is_lower(id); - } - - if is_zero_term(&self.ids_info.counters_seen, &zero_expressions, *operand) { - *operand = CovTerm::Zero; - } - }; - maybe_set_operand_to_zero(&mut lhs); - maybe_set_operand_to_zero(&mut rhs); - - // Coverage counter values cannot be negative, so if an expression - // involves subtraction from zero, assume that its RHS must also be zero. - // (Do this after simplifications that could set the LHS to zero.) - if lhs == CovTerm::Zero && op == Op::Subtract { - rhs = CovTerm::Zero; - } - - // After the above simplifications, if both operands are zero, then - // we know that this expression is always zero too. - if lhs == CovTerm::Zero && rhs == CovTerm::Zero { - zero_expressions.insert(id); - } - } - - zero_expressions - } - - pub(crate) fn into_finished(self) -> FunctionCoverage<'tcx> { - let zero_expressions = self.identify_zero_expressions(); - let FunctionCoverageCollector { function_coverage_info, ids_info, is_used, .. } = self; - - FunctionCoverage { function_coverage_info, ids_info, is_used, zero_expressions } - } -} - -pub(crate) struct FunctionCoverage<'tcx> { - pub(crate) function_coverage_info: &'tcx FunctionCoverageInfo, - ids_info: &'tcx CoverageIdsInfo, - is_used: bool, - - zero_expressions: ZeroExpressions, -} - -impl<'tcx> FunctionCoverage<'tcx> { - /// Returns true for a used (called) function, and false for an unused function. - pub(crate) fn is_used(&self) -> bool { - self.is_used - } - - /// Return the source hash, generated from the HIR node structure, and used to indicate whether - /// or not the source code structure changed between different compilations. - pub(crate) fn source_hash(&self) -> u64 { - if self.is_used { self.function_coverage_info.function_source_hash } else { 0 } - } - - /// Convert this function's coverage expression data into a form that can be - /// passed through FFI to LLVM. - pub(crate) fn counter_expressions( - &self, - ) -> impl Iterator<Item = CounterExpression> + ExactSizeIterator + Captures<'_> { - // We know that LLVM will optimize out any unused expressions before - // producing the final coverage map, so there's no need to do the same - // thing on the Rust side unless we're confident we can do much better. - // (See `CounterExpressionsMinimizer` in `CoverageMappingWriter.cpp`.) - - self.function_coverage_info.expressions.iter().map(move |&Expression { lhs, op, rhs }| { - CounterExpression { - lhs: self.counter_for_term(lhs), - kind: match op { - Op::Add => ExprKind::Add, - Op::Subtract => ExprKind::Subtract, - }, - rhs: self.counter_for_term(rhs), - } - }) - } - - /// Converts this function's coverage mappings into an intermediate form - /// that will be used by `mapgen` when preparing for FFI. - pub(crate) fn counter_regions( - &self, - ) -> impl Iterator<Item = (MappingKind, &SourceRegion)> + ExactSizeIterator { - self.function_coverage_info.mappings.iter().map(move |mapping| { - let Mapping { kind, source_region } = mapping; - let kind = - kind.map_terms(|term| if self.is_zero_term(term) { CovTerm::Zero } else { term }); - (kind, source_region) - }) - } - - fn counter_for_term(&self, term: CovTerm) -> Counter { - if self.is_zero_term(term) { Counter::ZERO } else { Counter::from_term(term) } - } - - fn is_zero_term(&self, term: CovTerm) -> bool { - !self.is_used || is_zero_term(&self.ids_info.counters_seen, &self.zero_expressions, term) - } -} - -/// Set of expression IDs that are known to always evaluate to zero. -/// Any mapping or expression operand that refers to these expressions can have -/// that reference replaced with a constant zero value. -#[derive(Default)] -struct ZeroExpressions(FxIndexSet<ExpressionId>); - -impl ZeroExpressions { - fn insert(&mut self, id: ExpressionId) { - self.0.insert(id); - } - - fn contains(&self, id: ExpressionId) -> bool { - self.0.contains(&id) - } -} - -/// Returns `true` if the given term is known to have a value of zero, taking -/// into account knowledge of which counters are unused and which expressions -/// are always zero. -fn is_zero_term( - counters_seen: &BitSet<CounterId>, - zero_expressions: &ZeroExpressions, - term: CovTerm, -) -> bool { - match term { - CovTerm::Zero => true, - CovTerm::Counter(id) => !counters_seen.contains(id), - CovTerm::Expression(id) => zero_expressions.contains(id), - } -} diff --git a/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen.rs b/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen.rs index 8c24579fa7c..b3ad2a0e409 100644 --- a/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen.rs +++ b/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen.rs @@ -1,29 +1,29 @@ -use std::ffi::CString; -use std::iter; +use std::sync::Arc; -use itertools::Itertools as _; +use itertools::Itertools; use rustc_abi::Align; use rustc_codegen_ssa::traits::{ BaseTypeCodegenMethods, ConstCodegenMethods, StaticCodegenMethods, }; -use rustc_data_structures::fx::{FxHashSet, FxIndexMap, FxIndexSet}; +use rustc_data_structures::fx::{FxHashSet, FxIndexMap}; use rustc_hir::def_id::{DefId, LocalDefId}; use rustc_index::IndexVec; -use rustc_middle::mir::coverage::MappingKind; +use rustc_middle::mir; use rustc_middle::ty::{self, TyCtxt}; -use rustc_middle::{bug, mir}; use rustc_session::RemapFileNameExt; use rustc_session::config::RemapPathScopeComponents; use rustc_span::def_id::DefIdSet; -use rustc_span::{Span, Symbol}; -use rustc_target::spec::HasTargetSpec; +use rustc_span::{SourceFile, StableSourceFileId}; use tracing::debug; use crate::common::CodegenCx; -use crate::coverageinfo::map_data::{FunctionCoverage, FunctionCoverageCollector}; -use crate::coverageinfo::{ffi, llvm_cov}; +use crate::coverageinfo::llvm_cov; +use crate::coverageinfo::mapgen::covfun::prepare_covfun_record; use crate::llvm; +mod covfun; +mod spans; + /// Generates and exports the coverage map, which is embedded in special /// linker sections in the final binary. /// @@ -49,83 +49,62 @@ pub(crate) fn finalize(cx: &CodegenCx<'_, '_>) { debug!("Generating coverage map for CodegenUnit: `{}`", cx.codegen_unit.name()); - // In order to show that unused functions have coverage counts of zero (0), LLVM requires the - // functions exist. Generate synthetic functions with a (required) single counter, and add the - // MIR `Coverage` code regions to the `function_coverage_map`, before calling - // `ctx.take_function_coverage_map()`. - if cx.codegen_unit.is_code_coverage_dead_code_cgu() { - add_unused_functions(cx); - } - // FIXME(#132395): Can this be none even when coverage is enabled? - let function_coverage_map = match cx.coverage_cx { - Some(ref cx) => cx.take_function_coverage_map(), + let instances_used = match cx.coverage_cx { + Some(ref cx) => cx.instances_used.borrow(), None => return, }; - if function_coverage_map.is_empty() { - // This module has no functions with coverage instrumentation - return; - } - let function_coverage_entries = function_coverage_map - .into_iter() - .map(|(instance, function_coverage)| (instance, function_coverage.into_finished())) - .collect::<Vec<_>>(); + // The order of entries in this global file table needs to be deterministic, + // and ideally should also be independent of the details of stable-hashing, + // because coverage tests snapshots (`.cov-map`) can observe the order and + // would need to be re-blessed if it changes. As long as those requirements + // are satisfied, the order can be arbitrary. + let mut global_file_table = GlobalFileTable::new(); - let all_file_names = function_coverage_entries + let mut covfun_records = instances_used .iter() - .map(|(_, fn_cov)| fn_cov.function_coverage_info.body_span) - .map(|span| span_file_name(tcx, span)); - let global_file_table = GlobalFileTable::new(all_file_names); - - // Encode all filenames referenced by coverage mappings in this CGU. - let filenames_buffer = global_file_table.make_filenames_buffer(tcx); - - let filenames_size = filenames_buffer.len(); - let filenames_val = cx.const_bytes(&filenames_buffer); - let filenames_ref = llvm_cov::hash_bytes(&filenames_buffer); - - // Generate the coverage map header, which contains the filenames used by - // this CGU's coverage mappings, and store it in a well-known global. - generate_covmap_record(cx, covmap_version, filenames_size, filenames_val); - - let mut unused_function_names = Vec::new(); + .copied() + // 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, &mut global_file_table, instance, true)) + .collect::<Vec<_>>(); - // Encode coverage mappings and generate function records - for (instance, function_coverage) in function_coverage_entries { - debug!("Generate function coverage for {}, {:?}", cx.codegen_unit.name(), instance); + // In a single designated CGU, also prepare covfun records for functions + // in this crate that were instrumented for coverage, but are unused. + if cx.codegen_unit.is_code_coverage_dead_code_cgu() { + let mut unused_instances = gather_unused_function_instances(cx); + // Sort the unused instances by symbol name, for the same reason as the used ones. + unused_instances.sort_by_cached_key(|&instance| tcx.symbol_name(instance).name); + covfun_records.extend(unused_instances.into_iter().filter_map(|instance| { + prepare_covfun_record(tcx, &mut global_file_table, instance, false) + })); + } - let mangled_function_name = tcx.symbol_name(instance).name; - let source_hash = function_coverage.source_hash(); - let is_used = function_coverage.is_used(); + // If there are no covfun records for this CGU, don't generate a covmap record. + // Emitting a covmap record without any covfun records causes `llvm-cov` to + // fail when generating coverage reports, and if there are no covfun records + // then the covmap record isn't useful anyway. + // This should prevent a repeat of <https://github.com/rust-lang/rust/issues/133606>. + if covfun_records.is_empty() { + return; + } - let coverage_mapping_buffer = - encode_mappings_for_function(tcx, &global_file_table, &function_coverage); + // Encode all filenames referenced by coverage mappings in this CGU. + let filenames_buffer = global_file_table.make_filenames_buffer(tcx); + // The `llvm-cov` tool uses this hash to associate each covfun record with + // its corresponding filenames table, since the final binary will typically + // contain multiple covmap records from different compilation units. + let filenames_hash = llvm_cov::hash_bytes(&filenames_buffer); - if coverage_mapping_buffer.is_empty() { - if function_coverage.is_used() { - bug!( - "A used function should have had coverage mapping data but did not: {}", - mangled_function_name - ); - } else { - debug!("unused function had no coverage mapping data: {}", mangled_function_name); - continue; - } - } + let mut unused_function_names = vec![]; - if !is_used { - unused_function_names.push(mangled_function_name); - } + for covfun in &covfun_records { + unused_function_names.extend(covfun.mangled_function_name_if_unused()); - generate_covfun_record( - cx, - mangled_function_name, - source_hash, - filenames_ref, - coverage_mapping_buffer, - is_used, - ); + covfun::generate_covfun_record(cx, filenames_hash, covfun) } // For unused functions, we need to take their mangled names and store them @@ -146,56 +125,58 @@ pub(crate) fn finalize(cx: &CodegenCx<'_, '_>) { llvm::set_linkage(array, llvm::Linkage::InternalLinkage); llvm::set_initializer(array, initializer); } + + // Generate the coverage map header, which contains the filenames used by + // this CGU's coverage mappings, and store it in a well-known global. + // (This is skipped if we returned early due to having no covfun records.) + generate_covmap_record(cx, covmap_version, &filenames_buffer); } -/// Maps "global" (per-CGU) file ID numbers to their underlying filenames. +/// Maps "global" (per-CGU) file ID numbers to their underlying source files. struct GlobalFileTable { - /// This "raw" table doesn't include the working dir, so a filename's + /// This "raw" table doesn't include the working dir, so a file's /// global ID is its index in this set **plus one**. - raw_file_table: FxIndexSet<Symbol>, + raw_file_table: FxIndexMap<StableSourceFileId, Arc<SourceFile>>, } impl GlobalFileTable { - fn new(all_file_names: impl IntoIterator<Item = Symbol>) -> Self { - // Collect all of the filenames into a set. Filenames usually come in - // contiguous runs, so we can dedup adjacent ones to save work. - let mut raw_file_table = all_file_names.into_iter().dedup().collect::<FxIndexSet<Symbol>>(); - - // Sort the file table by its actual string values, not the arbitrary - // ordering of its symbols. - raw_file_table.sort_unstable_by(|a, b| a.as_str().cmp(b.as_str())); - - Self { raw_file_table } + fn new() -> Self { + Self { raw_file_table: FxIndexMap::default() } } - fn global_file_id_for_file_name(&self, file_name: Symbol) -> GlobalFileId { - let raw_id = self.raw_file_table.get_index_of(&file_name).unwrap_or_else(|| { - bug!("file name not found in prepared global file table: {file_name}"); - }); + fn global_file_id_for_file(&mut self, file: &Arc<SourceFile>) -> GlobalFileId { + // Ensure the given file has a table entry, and get its index. + let entry = self.raw_file_table.entry(file.stable_id); + let raw_id = entry.index(); + entry.or_insert_with(|| Arc::clone(file)); + // The raw file table doesn't include an entry for the working dir // (which has ID 0), so add 1 to get the correct ID. GlobalFileId::from_usize(raw_id + 1) } fn make_filenames_buffer(&self, tcx: TyCtxt<'_>) -> Vec<u8> { + let mut table = Vec::with_capacity(self.raw_file_table.len() + 1); + // LLVM Coverage Mapping Format version 6 (zero-based encoded as 5) // requires setting the first filename to the compilation directory. // Since rustc generates coverage maps with relative paths, the // compilation directory can be combined with the relative paths // to get absolute paths, if needed. - use rustc_session::RemapFileNameExt; - use rustc_session::config::RemapPathScopeComponents; - let working_dir: &str = &tcx - .sess - .opts - .working_dir - .for_scope(tcx.sess, RemapPathScopeComponents::MACRO) - .to_string_lossy(); - - // Insert the working dir at index 0, before the other filenames. - let filenames = - iter::once(working_dir).chain(self.raw_file_table.iter().map(Symbol::as_str)); - llvm_cov::write_filenames_to_buffer(filenames) + table.push( + tcx.sess + .opts + .working_dir + .for_scope(tcx.sess, RemapPathScopeComponents::MACRO) + .to_string_lossy(), + ); + + // Add the regular entries after the base directory. + table.extend(self.raw_file_table.values().map(|file| { + file.name.for_scope(tcx.sess, RemapPathScopeComponents::MACRO).to_string_lossy() + })); + + llvm_cov::write_filenames_to_buffer(&table) } } @@ -208,12 +189,12 @@ rustc_index::newtype_index! { /// An index into a function's list of global file IDs. That underlying list /// of local-to-global mappings will be embedded in the function's record in /// the `__llvm_covfun` linker section. - pub(crate) struct LocalFileId {} + struct LocalFileId {} } /// Holds a mapping from "local" (per-function) file IDs to "global" (per-CGU) /// file IDs. -#[derive(Default)] +#[derive(Debug, Default)] struct VirtualFileMapping { local_to_global: IndexVec<LocalFileId, GlobalFileId>, global_to_local: FxIndexMap<GlobalFileId, LocalFileId>, @@ -227,187 +208,45 @@ impl VirtualFileMapping { .or_insert_with(|| self.local_to_global.push(global_file_id)) } - fn into_vec(self) -> Vec<u32> { - // This conversion should be optimized away to ~zero overhead. - // In any case, it's probably not hot enough to worry about. - self.local_to_global.into_iter().map(|global| global.as_u32()).collect() - } -} - -fn span_file_name(tcx: TyCtxt<'_>, span: Span) -> Symbol { - let source_file = tcx.sess.source_map().lookup_source_file(span.lo()); - let name = - source_file.name.for_scope(tcx.sess, RemapPathScopeComponents::MACRO).to_string_lossy(); - Symbol::intern(&name) -} - -/// Using the expressions and counter regions collected for a single function, -/// generate the variable-sized payload of its corresponding `__llvm_covfun` -/// entry. The payload is returned as a vector of bytes. -/// -/// Newly-encountered filenames will be added to the global file table. -fn encode_mappings_for_function( - tcx: TyCtxt<'_>, - global_file_table: &GlobalFileTable, - function_coverage: &FunctionCoverage<'_>, -) -> Vec<u8> { - let counter_regions = function_coverage.counter_regions(); - if counter_regions.is_empty() { - return Vec::new(); - } - - let expressions = function_coverage.counter_expressions().collect::<Vec<_>>(); - - let mut virtual_file_mapping = VirtualFileMapping::default(); - let mut code_regions = vec![]; - let mut branch_regions = vec![]; - let mut mcdc_branch_regions = vec![]; - let mut mcdc_decision_regions = vec![]; - - // Currently a function's mappings must all be in the same file as its body span. - let file_name = span_file_name(tcx, function_coverage.function_coverage_info.body_span); - - // Look up the global file ID for that filename. - let global_file_id = global_file_table.global_file_id_for_file_name(file_name); - - // Associate that global file ID with a local file ID for this function. - let local_file_id = virtual_file_mapping.local_id_for_global(global_file_id); - debug!(" file id: {local_file_id:?} => {global_file_id:?} = '{file_name:?}'"); - - // For each counter/region pair in this function+file, convert it to a - // form suitable for FFI. - for (mapping_kind, region) in counter_regions { - debug!("Adding counter {mapping_kind:?} to map for {region:?}"); - let span = ffi::CoverageSpan::from_source_region(local_file_id, region); - match mapping_kind { - MappingKind::Code(term) => { - code_regions.push(ffi::CodeRegion { span, counter: ffi::Counter::from_term(term) }); - } - MappingKind::Branch { true_term, false_term } => { - branch_regions.push(ffi::BranchRegion { - span, - true_counter: ffi::Counter::from_term(true_term), - false_counter: ffi::Counter::from_term(false_term), - }); - } - MappingKind::MCDCBranch { true_term, false_term, mcdc_params } => { - mcdc_branch_regions.push(ffi::MCDCBranchRegion { - span, - true_counter: ffi::Counter::from_term(true_term), - false_counter: ffi::Counter::from_term(false_term), - mcdc_branch_params: ffi::mcdc::BranchParameters::from(mcdc_params), - }); - } - MappingKind::MCDCDecision(mcdc_decision_params) => { - mcdc_decision_regions.push(ffi::MCDCDecisionRegion { - span, - mcdc_decision_params: ffi::mcdc::DecisionParameters::from(mcdc_decision_params), - }); - } - } + fn to_vec(&self) -> Vec<u32> { + // This clone could be avoided by transmuting `&[GlobalFileId]` to `&[u32]`, + // but it isn't hot or expensive enough to justify the extra unsafety. + self.local_to_global.iter().map(|&global| GlobalFileId::as_u32(global)).collect() } - - // Encode the function's coverage mappings into a buffer. - llvm_cov::write_function_mappings_to_buffer( - &virtual_file_mapping.into_vec(), - &expressions, - &code_regions, - &branch_regions, - &mcdc_branch_regions, - &mcdc_decision_regions, - ) } /// Generates the contents of the covmap record for this CGU, which mostly /// consists of a header and a list of filenames. The record is then stored /// as a global variable in the `__llvm_covmap` section. -fn generate_covmap_record<'ll>( - cx: &CodegenCx<'ll, '_>, - version: u32, - filenames_size: usize, - filenames_val: &'ll llvm::Value, -) { - debug!("cov map: filenames_size = {}, 0-based version = {}", filenames_size, version); - - // Create the coverage data header (Note, fields 0 and 2 are now always zero, - // as of `llvm::coverage::CovMapVersion::Version4`.) - let zero_was_n_records_val = cx.const_u32(0); - let filenames_size_val = cx.const_u32(filenames_size as u32); - let zero_was_coverage_size_val = cx.const_u32(0); - let version_val = cx.const_u32(version); - let cov_data_header_val = cx.const_struct( - &[zero_was_n_records_val, filenames_size_val, zero_was_coverage_size_val, version_val], - /*packed=*/ false, - ); - - // Create the complete LLVM coverage data value to add to the LLVM IR - let covmap_data = - cx.const_struct(&[cov_data_header_val, filenames_val], /*packed=*/ false); - - let llglobal = llvm::add_global(cx.llmod, cx.val_ty(covmap_data), &llvm_cov::covmap_var_name()); - llvm::set_initializer(llglobal, covmap_data); - llvm::set_global_constant(llglobal, true); - llvm::set_linkage(llglobal, llvm::Linkage::PrivateLinkage); - llvm::set_section(llglobal, &llvm_cov::covmap_section_name(cx.llmod)); - // LLVM's coverage mapping format specifies 8-byte alignment for items in this section. - // <https://llvm.org/docs/CoverageMappingFormat.html> - llvm::set_alignment(llglobal, Align::EIGHT); - cx.add_used_global(llglobal); -} - -/// Generates the contents of the covfun record for this function, which -/// contains the function's coverage mapping data. The record is then stored -/// as a global variable in the `__llvm_covfun` section. -fn generate_covfun_record( - cx: &CodegenCx<'_, '_>, - mangled_function_name: &str, - source_hash: u64, - filenames_ref: u64, - coverage_mapping_buffer: Vec<u8>, - is_used: bool, -) { - // Concatenate the encoded coverage mappings - let coverage_mapping_size = coverage_mapping_buffer.len(); - let coverage_mapping_val = cx.const_bytes(&coverage_mapping_buffer); - - let func_name_hash = llvm_cov::hash_bytes(mangled_function_name.as_bytes()); - let func_name_hash_val = cx.const_u64(func_name_hash); - let coverage_mapping_size_val = cx.const_u32(coverage_mapping_size as u32); - let source_hash_val = cx.const_u64(source_hash); - let filenames_ref_val = cx.const_u64(filenames_ref); - let func_record_val = cx.const_struct( +fn generate_covmap_record<'ll>(cx: &CodegenCx<'ll, '_>, version: u32, filenames_buffer: &[u8]) { + // A covmap record consists of four target-endian u32 values, followed by + // the encoded filenames table. Two of the header fields are unused in + // modern versions of the LLVM coverage mapping format, and are always 0. + // <https://llvm.org/docs/CoverageMappingFormat.html#llvm-ir-representation> + // See also `src/llvm-project/clang/lib/CodeGen/CoverageMappingGen.cpp`. + let covmap_header = cx.const_struct( &[ - func_name_hash_val, - coverage_mapping_size_val, - source_hash_val, - filenames_ref_val, - coverage_mapping_val, + cx.const_u32(0), // (unused) + cx.const_u32(filenames_buffer.len() as u32), + cx.const_u32(0), // (unused) + cx.const_u32(version), ], - /*packed=*/ true, + /* packed */ false, ); - - // Choose a variable name to hold this function's covfun data. - // Functions that are used have a suffix ("u") to distinguish them from - // unused copies of the same function (from different CGUs), so that if a - // linker sees both it won't discard the used copy's data. - let func_record_var_name = - CString::new(format!("__covrec_{:X}{}", func_name_hash, if is_used { "u" } else { "" })) - .unwrap(); - debug!("function record var name: {:?}", func_record_var_name); - - let llglobal = llvm::add_global(cx.llmod, cx.val_ty(func_record_val), &func_record_var_name); - llvm::set_initializer(llglobal, func_record_val); - llvm::set_global_constant(llglobal, true); - llvm::set_linkage(llglobal, llvm::Linkage::LinkOnceODRLinkage); - llvm::set_visibility(llglobal, llvm::Visibility::Hidden); - llvm::set_section(llglobal, cx.covfun_section_name()); + let covmap_record = cx + .const_struct(&[covmap_header, cx.const_bytes(filenames_buffer)], /* packed */ false); + + let covmap_global = + llvm::add_global(cx.llmod, cx.val_ty(covmap_record), &llvm_cov::covmap_var_name()); + llvm::set_initializer(covmap_global, covmap_record); + llvm::set_global_constant(covmap_global, true); + llvm::set_linkage(covmap_global, llvm::Linkage::PrivateLinkage); + llvm::set_section(covmap_global, &llvm_cov::covmap_section_name(cx.llmod)); // LLVM's coverage mapping format specifies 8-byte alignment for items in this section. // <https://llvm.org/docs/CoverageMappingFormat.html> - llvm::set_alignment(llglobal, Align::EIGHT); - if cx.target_spec().supports_comdat() { - llvm::set_comdat(cx.llmod, llglobal, &func_record_var_name); - } - cx.add_used_global(llglobal); + llvm::set_alignment(covmap_global, Align::EIGHT); + + cx.add_used_global(covmap_global); } /// Each CGU will normally only emit coverage metadata for the functions that it actually generates. @@ -418,39 +257,35 @@ fn generate_covfun_record( /// coverage map (in a single designated CGU) so that we still emit coverage mappings for them. /// We also end up adding their symbol names to a special global array that LLVM will include in /// its embedded coverage data. -fn add_unused_functions(cx: &CodegenCx<'_, '_>) { +fn gather_unused_function_instances<'tcx>(cx: &CodegenCx<'_, 'tcx>) -> Vec<ty::Instance<'tcx>> { assert!(cx.codegen_unit.is_code_coverage_dead_code_cgu()); let tcx = cx.tcx; let usage = prepare_usage_sets(tcx); let is_unused_fn = |def_id: LocalDefId| -> bool { - let def_id = def_id.to_def_id(); - - // To be eligible for "unused function" mappings, a definition must: - // - Be function-like + // Usage sets expect `DefId`, so convert from `LocalDefId`. + let d: DefId = LocalDefId::to_def_id(def_id); + // To be potentially eligible for "unused function" mappings, a definition must: + // - Be eligible for coverage instrumentation // - Not participate directly in codegen (or have lost all its coverage statements) // - Not have any coverage statements inlined into codegenned functions - tcx.def_kind(def_id).is_fn_like() - && (!usage.all_mono_items.contains(&def_id) - || usage.missing_own_coverage.contains(&def_id)) - && !usage.used_via_inlining.contains(&def_id) + tcx.is_eligible_for_coverage(def_id) + && (!usage.all_mono_items.contains(&d) || usage.missing_own_coverage.contains(&d)) + && !usage.used_via_inlining.contains(&d) }; - // Scan for unused functions that were instrumented for coverage. - for def_id in tcx.mir_keys(()).iter().copied().filter(|&def_id| is_unused_fn(def_id)) { - // Get the coverage info from MIR, skipping functions that were never instrumented. - let body = tcx.optimized_mir(def_id); - let Some(function_coverage_info) = body.function_coverage_info.as_deref() else { continue }; + // FIXME(#79651): Consider trying to filter out dummy instantiations of + // unused generic functions from library crates, because they can produce + // "unused instantiation" in coverage reports even when they are actually + // used by some downstream crate in the same binary. - // FIXME(79651): Consider trying to filter out dummy instantiations of - // unused generic functions from library crates, because they can produce - // "unused instantiation" in coverage reports even when they are actually - // used by some downstream crate in the same binary. - - debug!("generating unused fn: {def_id:?}"); - add_unused_function_coverage(cx, def_id, function_coverage_info); - } + tcx.mir_keys(()) + .iter() + .copied() + .filter(|&def_id| is_unused_fn(def_id)) + .map(|def_id| make_dummy_instance(tcx, def_id)) + .collect::<Vec<_>>() } struct UsageSets<'tcx> { @@ -515,16 +350,11 @@ fn prepare_usage_sets<'tcx>(tcx: TyCtxt<'tcx>) -> UsageSets<'tcx> { UsageSets { all_mono_items, used_via_inlining, missing_own_coverage } } -fn add_unused_function_coverage<'tcx>( - cx: &CodegenCx<'_, 'tcx>, - def_id: LocalDefId, - function_coverage_info: &'tcx mir::coverage::FunctionCoverageInfo, -) { - let tcx = cx.tcx; - let def_id = def_id.to_def_id(); +fn make_dummy_instance<'tcx>(tcx: TyCtxt<'tcx>, local_def_id: LocalDefId) -> ty::Instance<'tcx> { + let def_id = local_def_id.to_def_id(); // Make a dummy instance that fills in all generics with placeholders. - let instance = ty::Instance::new( + ty::Instance::new( def_id, ty::GenericArgs::for_item(tcx, def_id, |param, _| { if let ty::GenericParamDefKind::Lifetime = param.kind { @@ -533,14 +363,5 @@ fn add_unused_function_coverage<'tcx>( tcx.mk_param_from_def(param) } }), - ); - - // An unused function's mappings will all be rewritten to map to zero. - let function_coverage = FunctionCoverageCollector::unused( - instance, - function_coverage_info, - tcx.coverage_ids_info(instance.def), - ); - - cx.coverage_cx().function_coverage_map.borrow_mut().insert(instance, function_coverage); + ) } diff --git a/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen/covfun.rs b/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen/covfun.rs new file mode 100644 index 00000000000..5428d776f41 --- /dev/null +++ b/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen/covfun.rs @@ -0,0 +1,252 @@ +//! For each function that was instrumented for coverage, we need to embed its +//! corresponding coverage mapping metadata inside the `__llvm_covfun`[^win] +//! linker section of the final binary. +//! +//! [^win]: On Windows the section name is `.lcovfun`. + +use std::ffi::CString; + +use rustc_abi::Align; +use rustc_codegen_ssa::traits::{ + BaseTypeCodegenMethods, ConstCodegenMethods, StaticCodegenMethods, +}; +use rustc_middle::mir::coverage::{ + CovTerm, CoverageIdsInfo, Expression, FunctionCoverageInfo, Mapping, MappingKind, Op, +}; +use rustc_middle::ty::{Instance, TyCtxt}; +use rustc_span::Span; +use rustc_target::spec::HasTargetSpec; +use tracing::debug; + +use crate::common::CodegenCx; +use crate::coverageinfo::mapgen::{GlobalFileTable, VirtualFileMapping, spans}; +use crate::coverageinfo::{ffi, llvm_cov}; +use crate::llvm; + +/// Intermediate coverage metadata for a single function, used to help build +/// the final record that will be embedded in the `__llvm_covfun` section. +#[derive(Debug)] +pub(crate) struct CovfunRecord<'tcx> { + mangled_function_name: &'tcx str, + source_hash: u64, + is_used: bool, + + virtual_file_mapping: VirtualFileMapping, + expressions: Vec<ffi::CounterExpression>, + regions: ffi::Regions, +} + +impl<'tcx> CovfunRecord<'tcx> { + /// FIXME(Zalathar): Make this the responsibility of the code that determines + /// which functions are unused. + pub(crate) fn mangled_function_name_if_unused(&self) -> Option<&'tcx str> { + (!self.is_used).then_some(self.mangled_function_name) + } +} + +pub(crate) fn prepare_covfun_record<'tcx>( + tcx: TyCtxt<'tcx>, + global_file_table: &mut GlobalFileTable, + instance: Instance<'tcx>, + is_used: bool, +) -> Option<CovfunRecord<'tcx>> { + let fn_cov_info = tcx.instance_mir(instance.def).function_coverage_info.as_deref()?; + let ids_info = tcx.coverage_ids_info(instance.def); + + let expressions = prepare_expressions(fn_cov_info, ids_info, is_used); + + let mut covfun = CovfunRecord { + mangled_function_name: tcx.symbol_name(instance).name, + source_hash: if is_used { fn_cov_info.function_source_hash } else { 0 }, + is_used, + virtual_file_mapping: VirtualFileMapping::default(), + expressions, + regions: ffi::Regions::default(), + }; + + fill_region_tables(tcx, global_file_table, fn_cov_info, ids_info, &mut covfun); + + if covfun.regions.has_no_regions() { + debug!(?covfun, "function has no mappings to embed; skipping"); + return None; + } + + Some(covfun) +} + +/// Convert the function's coverage-counter expressions into a form suitable for FFI. +fn prepare_expressions( + fn_cov_info: &FunctionCoverageInfo, + ids_info: &CoverageIdsInfo, + is_used: bool, +) -> Vec<ffi::CounterExpression> { + // If any counters or expressions were removed by MIR opts, replace their + // terms with zero. + let counter_for_term = |term| { + if !is_used || ids_info.is_zero_term(term) { + ffi::Counter::ZERO + } else { + ffi::Counter::from_term(term) + } + }; + + // We know that LLVM will optimize out any unused expressions before + // producing the final coverage map, so there's no need to do the same + // thing on the Rust side unless we're confident we can do much better. + // (See `CounterExpressionsMinimizer` in `CoverageMappingWriter.cpp`.) + fn_cov_info + .expressions + .iter() + .map(move |&Expression { lhs, op, rhs }| ffi::CounterExpression { + lhs: counter_for_term(lhs), + kind: match op { + Op::Add => ffi::ExprKind::Add, + Op::Subtract => ffi::ExprKind::Subtract, + }, + rhs: counter_for_term(rhs), + }) + .collect::<Vec<_>>() +} + +/// Populates the mapping region tables in the current function's covfun record. +fn fill_region_tables<'tcx>( + tcx: TyCtxt<'tcx>, + global_file_table: &mut GlobalFileTable, + fn_cov_info: &'tcx FunctionCoverageInfo, + ids_info: &'tcx CoverageIdsInfo, + covfun: &mut CovfunRecord<'tcx>, +) { + // Currently a function's mappings must all be in the same file as its body span. + let source_map = tcx.sess.source_map(); + let source_file = source_map.lookup_source_file(fn_cov_info.body_span.lo()); + + // Look up the global file ID for that file. + let global_file_id = global_file_table.global_file_id_for_file(&source_file); + + // Associate that global file ID with a local file ID for this function. + let local_file_id = covfun.virtual_file_mapping.local_id_for_global(global_file_id); + + let ffi::Regions { code_regions, branch_regions, mcdc_branch_regions, mcdc_decision_regions } = + &mut covfun.regions; + + let make_cov_span = |span: Span| { + spans::make_coverage_span(local_file_id, source_map, fn_cov_info, &source_file, span) + }; + let discard_all = tcx.sess.coverage_discard_all_spans_in_codegen(); + + // For each counter/region pair in this function+file, convert it to a + // form suitable for FFI. + let is_zero_term = |term| !covfun.is_used || ids_info.is_zero_term(term); + for &Mapping { ref kind, span } in &fn_cov_info.mappings { + // If the mapping refers to counters/expressions that were removed by + // MIR opts, replace those occurrences with zero. + let kind = kind.map_terms(|term| if is_zero_term(term) { CovTerm::Zero } else { term }); + + // Convert the `Span` into coordinates that we can pass to LLVM, or + // discard the span if conversion fails. In rare, cases _all_ of a + // function's spans are discarded, and the rest of coverage codegen + // needs to handle that gracefully to avoid a repeat of #133606. + // We don't have a good test case for triggering that organically, so + // instead we set `-Zcoverage-options=discard-all-spans-in-codegen` + // to force it to occur. + let Some(cov_span) = make_cov_span(span) else { continue }; + if discard_all { + continue; + } + + match kind { + MappingKind::Code(term) => { + code_regions + .push(ffi::CodeRegion { cov_span, counter: ffi::Counter::from_term(term) }); + } + MappingKind::Branch { true_term, false_term } => { + branch_regions.push(ffi::BranchRegion { + cov_span, + true_counter: ffi::Counter::from_term(true_term), + false_counter: ffi::Counter::from_term(false_term), + }); + } + MappingKind::MCDCBranch { true_term, false_term, mcdc_params } => { + mcdc_branch_regions.push(ffi::MCDCBranchRegion { + cov_span, + true_counter: ffi::Counter::from_term(true_term), + false_counter: ffi::Counter::from_term(false_term), + mcdc_branch_params: ffi::mcdc::BranchParameters::from(mcdc_params), + }); + } + MappingKind::MCDCDecision(mcdc_decision_params) => { + mcdc_decision_regions.push(ffi::MCDCDecisionRegion { + cov_span, + mcdc_decision_params: ffi::mcdc::DecisionParameters::from(mcdc_decision_params), + }); + } + } + } +} + +/// Generates the contents of the covfun record for this function, which +/// contains the function's coverage mapping data. The record is then stored +/// as a global variable in the `__llvm_covfun` section. +pub(crate) fn generate_covfun_record<'tcx>( + cx: &CodegenCx<'_, 'tcx>, + filenames_hash: u64, + covfun: &CovfunRecord<'tcx>, +) { + let &CovfunRecord { + mangled_function_name, + source_hash, + is_used, + ref virtual_file_mapping, + ref expressions, + ref regions, + } = covfun; + + // Encode the function's coverage mappings into a buffer. + let coverage_mapping_buffer = llvm_cov::write_function_mappings_to_buffer( + &virtual_file_mapping.to_vec(), + expressions, + regions, + ); + + // A covfun record consists of four target-endian integers, followed by the + // encoded mapping data in bytes. Note that the length field is 32 bits. + // <https://llvm.org/docs/CoverageMappingFormat.html#llvm-ir-representation> + // See also `src/llvm-project/clang/lib/CodeGen/CoverageMappingGen.cpp` and + // `COVMAP_V3` in `src/llvm-project/llvm/include/llvm/ProfileData/InstrProfData.inc`. + let func_name_hash = llvm_cov::hash_bytes(mangled_function_name.as_bytes()); + let covfun_record = cx.const_struct( + &[ + cx.const_u64(func_name_hash), + cx.const_u32(coverage_mapping_buffer.len() as u32), + cx.const_u64(source_hash), + cx.const_u64(filenames_hash), + cx.const_bytes(&coverage_mapping_buffer), + ], + // This struct needs to be packed, so that the 32-bit length field + // doesn't have unexpected padding. + true, + ); + + // Choose a variable name to hold this function's covfun data. + // Functions that are used have a suffix ("u") to distinguish them from + // unused copies of the same function (from different CGUs), so that if a + // linker sees both it won't discard the used copy's data. + let u = if is_used { "u" } else { "" }; + let covfun_var_name = CString::new(format!("__covrec_{func_name_hash:X}{u}")).unwrap(); + debug!("function record var name: {covfun_var_name:?}"); + + let covfun_global = llvm::add_global(cx.llmod, cx.val_ty(covfun_record), &covfun_var_name); + llvm::set_initializer(covfun_global, covfun_record); + llvm::set_global_constant(covfun_global, true); + llvm::set_linkage(covfun_global, llvm::Linkage::LinkOnceODRLinkage); + llvm::set_visibility(covfun_global, llvm::Visibility::Hidden); + llvm::set_section(covfun_global, cx.covfun_section_name()); + // LLVM's coverage mapping format specifies 8-byte alignment for items in this section. + // <https://llvm.org/docs/CoverageMappingFormat.html> + llvm::set_alignment(covfun_global, Align::EIGHT); + if cx.target_spec().supports_comdat() { + llvm::set_comdat(cx.llmod, covfun_global, &covfun_var_name); + } + + cx.add_used_global(covfun_global); +} diff --git a/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen/spans.rs b/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen/spans.rs new file mode 100644 index 00000000000..6d1d91340c2 --- /dev/null +++ b/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen/spans.rs @@ -0,0 +1,126 @@ +use rustc_middle::mir::coverage::FunctionCoverageInfo; +use rustc_span::source_map::SourceMap; +use rustc_span::{BytePos, Pos, SourceFile, Span}; +use tracing::debug; + +use crate::coverageinfo::ffi; +use crate::coverageinfo::mapgen::LocalFileId; + +/// Converts the span into its start line and column, and end line and column. +/// +/// Line numbers and column numbers are 1-based. Unlike most column numbers emitted by +/// the compiler, these column numbers are denoted in **bytes**, because that's what +/// LLVM's `llvm-cov` tool expects to see in coverage maps. +/// +/// Returns `None` if the conversion failed for some reason. This shouldn't happen, +/// but it's hard to rule out entirely (especially in the presence of complex macros +/// or other expansions), and if it does happen then skipping a span or function is +/// better than an ICE or `llvm-cov` failure that the user might have no way to avoid. +pub(crate) fn make_coverage_span( + file_id: LocalFileId, + source_map: &SourceMap, + fn_cov_info: &FunctionCoverageInfo, + file: &SourceFile, + span: Span, +) -> Option<ffi::CoverageSpan> { + let span = ensure_non_empty_span(source_map, fn_cov_info, span)?; + + let lo = span.lo(); + let hi = span.hi(); + + // Column numbers need to be in bytes, so we can't use the more convenient + // `SourceMap` methods for looking up file coordinates. + let line_and_byte_column = |pos: BytePos| -> Option<(usize, usize)> { + let rpos = file.relative_position(pos); + let line_index = file.lookup_line(rpos)?; + let line_start = file.lines()[line_index]; + // Line numbers and column numbers are 1-based, so add 1 to each. + Some((line_index + 1, (rpos - line_start).to_usize() + 1)) + }; + + let (mut start_line, start_col) = line_and_byte_column(lo)?; + let (mut end_line, end_col) = line_and_byte_column(hi)?; + + // Apply an offset so that code in doctests has correct line numbers. + // FIXME(#79417): Currently we have no way to offset doctest _columns_. + start_line = source_map.doctest_offset_line(&file.name, start_line); + end_line = source_map.doctest_offset_line(&file.name, end_line); + + check_coverage_span(ffi::CoverageSpan { + file_id: file_id.as_u32(), + start_line: start_line as u32, + start_col: start_col as u32, + end_line: end_line as u32, + end_col: end_col as u32, + }) +} + +fn ensure_non_empty_span( + source_map: &SourceMap, + fn_cov_info: &FunctionCoverageInfo, + span: Span, +) -> Option<Span> { + if !span.is_empty() { + return Some(span); + } + + let lo = span.lo(); + let hi = span.hi(); + + // The span is empty, so try to expand it to cover an adjacent '{' or '}', + // but only within the bounds of the body span. + let try_next = hi < fn_cov_info.body_span.hi(); + let try_prev = fn_cov_info.body_span.lo() < lo; + if !(try_next || try_prev) { + return None; + } + + source_map + .span_to_source(span, |src, start, end| try { + // Adjusting span endpoints by `BytePos(1)` is normally a bug, + // but in this case we have specifically checked that the character + // we're skipping over is one of two specific ASCII characters, so + // adjusting by exactly 1 byte is correct. + if try_next && src.as_bytes()[end] == b'{' { + Some(span.with_hi(hi + BytePos(1))) + } else if try_prev && src.as_bytes()[start - 1] == b'}' { + Some(span.with_lo(lo - BytePos(1))) + } else { + None + } + }) + .ok()? +} + +/// If `llvm-cov` sees a source region that is improperly ordered (end < start), +/// it will immediately exit with a fatal error. To prevent that from happening, +/// discard regions that are improperly ordered, or might be interpreted in a +/// way that makes them improperly ordered. +fn check_coverage_span(cov_span: ffi::CoverageSpan) -> Option<ffi::CoverageSpan> { + let ffi::CoverageSpan { file_id: _, start_line, start_col, end_line, end_col } = cov_span; + + // Line/column coordinates are supposed to be 1-based. If we ever emit + // coordinates of 0, `llvm-cov` might misinterpret them. + let all_nonzero = [start_line, start_col, end_line, end_col].into_iter().all(|x| x != 0); + // Coverage mappings use the high bit of `end_col` to indicate that a + // region is actually a "gap" region, so make sure it's unset. + let end_col_has_high_bit_unset = (end_col & (1 << 31)) == 0; + // If a region is improperly ordered (end < start), `llvm-cov` will exit + // with a fatal error, which is inconvenient for users and hard to debug. + let is_ordered = (start_line, start_col) <= (end_line, end_col); + + if all_nonzero && end_col_has_high_bit_unset && is_ordered { + Some(cov_span) + } else { + debug!( + ?cov_span, + ?all_nonzero, + ?end_col_has_high_bit_unset, + ?is_ordered, + "Skipping source region that would be misinterpreted or rejected by LLVM" + ); + // If this happens in a debug build, ICE to make it easier to notice. + debug_assert!(false, "Improper source region: {cov_span:?}"); + None + } +} diff --git a/compiler/rustc_codegen_llvm/src/coverageinfo/mod.rs b/compiler/rustc_codegen_llvm/src/coverageinfo/mod.rs index c2fcb33f98b..7311cd9d230 100644 --- a/compiler/rustc_codegen_llvm/src/coverageinfo/mod.rs +++ b/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, FxIndexMap}; +use rustc_data_structures::fx::{FxHashMap, FxIndexSet}; use rustc_middle::mir::coverage::CoverageKind; use rustc_middle::ty::Instance; use rustc_middle::ty::layout::HasTyCtxt; @@ -13,19 +13,16 @@ use tracing::{debug, instrument}; use crate::builder::Builder; use crate::common::CodegenCx; -use crate::coverageinfo::map_data::FunctionCoverageCollector; use crate::llvm; pub(crate) mod ffi; mod llvm_cov; -pub(crate) mod map_data; 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) function_coverage_map: - RefCell<FxIndexMap<Instance<'tcx>, FunctionCoverageCollector<'tcx>>>, + pub(crate) instances_used: RefCell<FxIndexSet<Instance<'tcx>>>, pub(crate) pgo_func_name_var_map: RefCell<FxHashMap<Instance<'tcx>, &'ll llvm::Value>>, pub(crate) mcdc_condition_bitmap_map: RefCell<FxHashMap<Instance<'tcx>, Vec<&'ll llvm::Value>>>, @@ -35,19 +32,13 @@ pub(crate) struct CguCoverageContext<'ll, 'tcx> { impl<'ll, 'tcx> CguCoverageContext<'ll, 'tcx> { pub(crate) fn new() -> Self { Self { - function_coverage_map: Default::default(), + instances_used: RefCell::<FxIndexSet<_>>::default(), pgo_func_name_var_map: Default::default(), mcdc_condition_bitmap_map: Default::default(), covfun_section_name: Default::default(), } } - fn take_function_coverage_map( - &self, - ) -> FxIndexMap<Instance<'tcx>, FunctionCoverageCollector<'tcx>> { - self.function_coverage_map.replace(FxIndexMap::default()) - } - /// LLVM use a temp value to record evaluated mcdc test vector of each decision, which is /// called condition bitmap. In order to handle nested decisions, several condition bitmaps can /// be allocated for a function body. These values are named `mcdc.addr.{i}` and are a 32-bit @@ -160,13 +151,7 @@ impl<'tcx> CoverageInfoBuilderMethods<'tcx> for Builder<'_, '_, 'tcx> { // 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.function_coverage_map.borrow_mut().entry(instance).or_insert_with(|| { - FunctionCoverageCollector::new( - instance, - function_coverage_info, - bx.tcx.coverage_ids_info(instance.def), - ) - }); + coverage_cx.instances_used.borrow_mut().insert(instance); match *kind { CoverageKind::SpanMarker | CoverageKind::BlockMarker { .. } => unreachable!( diff --git a/compiler/rustc_codegen_llvm/src/debuginfo/gdb.rs b/compiler/rustc_codegen_llvm/src/debuginfo/gdb.rs index aef8642f199..2c9f1cda13a 100644 --- a/compiler/rustc_codegen_llvm/src/debuginfo/gdb.rs +++ b/compiler/rustc_codegen_llvm/src/debuginfo/gdb.rs @@ -7,7 +7,7 @@ use rustc_hir::def_id::LOCAL_CRATE; use rustc_middle::bug; use rustc_middle::middle::debugger_visualizer::DebuggerVisualizerType; use rustc_session::config::{CrateType, DebugInfo}; -use rustc_span::symbol::sym; +use rustc_span::sym; use crate::builder::Builder; use crate::common::CodegenCx; diff --git a/compiler/rustc_codegen_llvm/src/debuginfo/metadata.rs b/compiler/rustc_codegen_llvm/src/debuginfo/metadata.rs index dee67baaee9..40248a9009a 100644 --- a/compiler/rustc_codegen_llvm/src/debuginfo/metadata.rs +++ b/compiler/rustc_codegen_llvm/src/debuginfo/metadata.rs @@ -16,8 +16,7 @@ use rustc_middle::ty::{ self, AdtKind, CoroutineArgsExt, Instance, PolyExistentialTraitRef, Ty, TyCtxt, Visibility, }; use rustc_session::config::{self, DebugInfo, Lto}; -use rustc_span::symbol::Symbol; -use rustc_span::{DUMMY_SP, FileName, FileNameDisplayPreference, SourceFile, hygiene}; +use rustc_span::{DUMMY_SP, FileName, FileNameDisplayPreference, SourceFile, Symbol, hygiene}; use rustc_symbol_mangling::typeid_for_trait_ref; use rustc_target::spec::DebuginfoKind; use smallvec::smallvec; @@ -471,8 +470,6 @@ pub(crate) fn type_di_node<'ll, 'tcx>(cx: &CodegenCx<'ll, 'tcx>, t: Ty<'tcx>) -> AdtKind::Enum => enums::build_enum_type_di_node(cx, unique_type_id), }, ty::Tuple(_) => build_tuple_type_di_node(cx, unique_type_id), - // Type parameters from polymorphized functions. - ty::Param(_) => build_param_type_di_node(cx, t), _ => bug!("debuginfo: unexpected type in type_di_node(): {:?}", t), }; @@ -871,26 +868,6 @@ fn build_foreign_type_di_node<'ll, 'tcx>( ) } -fn build_param_type_di_node<'ll, 'tcx>( - cx: &CodegenCx<'ll, 'tcx>, - t: Ty<'tcx>, -) -> DINodeCreationResult<'ll> { - debug!("build_param_type_di_node: {:?}", t); - let name = format!("{t:?}"); - DINodeCreationResult { - di_node: unsafe { - llvm::LLVMRustDIBuilderCreateBasicType( - DIB(cx), - name.as_c_char_ptr(), - name.len(), - Size::ZERO.bits(), - DW_ATE_unsigned, - ) - }, - already_stored_in_typemap: false, - } -} - pub(crate) fn build_compile_unit_di_node<'ll, 'tcx>( tcx: TyCtxt<'tcx>, codegen_unit_name: &str, diff --git a/compiler/rustc_codegen_llvm/src/debuginfo/metadata/enums/cpp_like.rs b/compiler/rustc_codegen_llvm/src/debuginfo/metadata/enums/cpp_like.rs index d374767f187..23e11748e52 100644 --- a/compiler/rustc_codegen_llvm/src/debuginfo/metadata/enums/cpp_like.rs +++ b/compiler/rustc_codegen_llvm/src/debuginfo/metadata/enums/cpp_like.rs @@ -212,21 +212,17 @@ pub(super) fn build_enum_type_di_node<'ll, 'tcx>( ), |cx, enum_type_di_node| { match enum_type_and_layout.variants { - Variants::Single { index: variant_index } => { - if enum_adt_def.variants().is_empty() { - // Uninhabited enums have Variants::Single. We don't generate - // any members for them. - return smallvec![]; - } - - build_single_variant_union_fields( - cx, - enum_adt_def, - enum_type_and_layout, - enum_type_di_node, - variant_index, - ) + Variants::Empty => { + // We don't generate any members for uninhabited types. + return smallvec![]; } + Variants::Single { index: variant_index } => build_single_variant_union_fields( + cx, + enum_adt_def, + enum_type_and_layout, + enum_type_di_node, + variant_index, + ), Variants::Multiple { tag_encoding: TagEncoding::Direct, ref variants, @@ -303,6 +299,7 @@ pub(super) fn build_coroutine_di_node<'ll, 'tcx>( ) } Variants::Single { .. } + | Variants::Empty | Variants::Multiple { tag_encoding: TagEncoding::Niche { .. }, .. } => { bug!( "Encountered coroutine with non-direct-tag layout: {:?}", diff --git a/compiler/rustc_codegen_llvm/src/debuginfo/metadata/enums/mod.rs b/compiler/rustc_codegen_llvm/src/debuginfo/metadata/enums/mod.rs index 65ab22ad89e..9f6a5cc89e0 100644 --- a/compiler/rustc_codegen_llvm/src/debuginfo/metadata/enums/mod.rs +++ b/compiler/rustc_codegen_llvm/src/debuginfo/metadata/enums/mod.rs @@ -392,7 +392,7 @@ fn compute_discriminant_value<'ll, 'tcx>( variant_index: VariantIdx, ) -> DiscrResult { match enum_type_and_layout.layout.variants() { - &Variants::Single { .. } => DiscrResult::NoDiscriminant, + &Variants::Single { .. } | &Variants::Empty => DiscrResult::NoDiscriminant, &Variants::Multiple { tag_encoding: TagEncoding::Direct, .. } => DiscrResult::Value( enum_type_and_layout.ty.discriminant_for_variant(cx.tcx, variant_index).unwrap().val, ), diff --git a/compiler/rustc_codegen_llvm/src/debuginfo/metadata/enums/native.rs b/compiler/rustc_codegen_llvm/src/debuginfo/metadata/enums/native.rs index 241bf167a81..11824398f24 100644 --- a/compiler/rustc_codegen_llvm/src/debuginfo/metadata/enums/native.rs +++ b/compiler/rustc_codegen_llvm/src/debuginfo/metadata/enums/native.rs @@ -358,8 +358,8 @@ fn build_discr_member_di_node<'ll, 'tcx>( let containing_scope = enum_or_coroutine_type_di_node; match enum_or_coroutine_type_and_layout.layout.variants() { - // A single-variant enum has no discriminant. - &Variants::Single { .. } => None, + // A single-variant or no-variant enum has no discriminant. + &Variants::Single { .. } | &Variants::Empty => None, &Variants::Multiple { tag_field, .. } => { let tag_base_type = tag_base_type(cx.tcx, enum_or_coroutine_type_and_layout); diff --git a/compiler/rustc_codegen_llvm/src/debuginfo/mod.rs b/compiler/rustc_codegen_llvm/src/debuginfo/mod.rs index a8fdfbed592..fae698bea2a 100644 --- a/compiler/rustc_codegen_llvm/src/debuginfo/mod.rs +++ b/compiler/rustc_codegen_llvm/src/debuginfo/mod.rs @@ -19,9 +19,8 @@ use rustc_middle::ty::layout::{HasTypingEnv, LayoutOf}; use rustc_middle::ty::{self, GenericArgsRef, Instance, Ty, TypeVisitableExt}; use rustc_session::Session; use rustc_session::config::{self, DebugInfo}; -use rustc_span::symbol::Symbol; use rustc_span::{ - BytePos, Pos, SourceFile, SourceFileAndLine, SourceFileHash, Span, StableSourceFileId, + BytePos, Pos, SourceFile, SourceFileAndLine, SourceFileHash, Span, StableSourceFileId, Symbol, }; use smallvec::SmallVec; use tracing::debug; diff --git a/compiler/rustc_codegen_llvm/src/declare.rs b/compiler/rustc_codegen_llvm/src/declare.rs index d338c848754..3ec386f6b07 100644 --- a/compiler/rustc_codegen_llvm/src/declare.rs +++ b/compiler/rustc_codegen_llvm/src/declare.rs @@ -32,7 +32,7 @@ use crate::{attributes, llvm}; /// /// If there’s a value with the same name already declared, the function will /// update the declaration and return existing Value instead. -fn declare_raw_fn<'ll>( +pub(crate) fn declare_raw_fn<'ll>( cx: &CodegenCx<'ll, '_>, name: &str, callconv: llvm::CallConv, diff --git a/compiler/rustc_codegen_llvm/src/errors.rs b/compiler/rustc_codegen_llvm/src/errors.rs index 3cdb5b971d9..f340b06e876 100644 --- a/compiler/rustc_codegen_llvm/src/errors.rs +++ b/compiler/rustc_codegen_llvm/src/errors.rs @@ -90,6 +90,11 @@ impl<G: EmissionGuarantee> Diagnostic<'_, G> for ParseTargetMachineConfig<'_> { } #[derive(Diagnostic)] +#[diag(codegen_llvm_autodiff_without_lto)] +#[note] +pub(crate) struct AutoDiffWithoutLTO; + +#[derive(Diagnostic)] #[diag(codegen_llvm_lto_disallowed)] pub(crate) struct LtoDisallowed; @@ -131,6 +136,8 @@ pub enum LlvmError<'a> { PrepareThinLtoModule, #[diag(codegen_llvm_parse_bitcode)] ParseBitcode, + #[diag(codegen_llvm_prepare_autodiff)] + PrepareAutoDiff { src: String, target: String, error: String }, } pub(crate) struct WithLlvmError<'a>(pub LlvmError<'a>, pub String); @@ -152,6 +159,7 @@ impl<G: EmissionGuarantee> Diagnostic<'_, G> for WithLlvmError<'_> { } PrepareThinLtoModule => fluent::codegen_llvm_prepare_thin_lto_module_with_llvm_err, ParseBitcode => fluent::codegen_llvm_parse_bitcode_with_llvm_err, + PrepareAutoDiff { .. } => fluent::codegen_llvm_prepare_autodiff_with_llvm_err, }; self.0 .into_diag(dcx, level) diff --git a/compiler/rustc_codegen_llvm/src/intrinsic.rs b/compiler/rustc_codegen_llvm/src/intrinsic.rs index d8b055137b3..cabcfc9b42b 100644 --- a/compiler/rustc_codegen_llvm/src/intrinsic.rs +++ b/compiler/rustc_codegen_llvm/src/intrinsic.rs @@ -340,6 +340,37 @@ impl<'ll, 'tcx> IntrinsicCallBuilderMethods<'tcx> for Builder<'_, 'll, 'tcx> { self.const_i32(cache_type), ]) } + sym::carrying_mul_add => { + let (size, signed) = fn_args.type_at(0).int_size_and_signed(self.tcx); + + let wide_llty = self.type_ix(size.bits() * 2); + let args = args.as_array().unwrap(); + let [a, b, c, d] = args.map(|a| self.intcast(a.immediate(), wide_llty, signed)); + + let wide = if signed { + let prod = self.unchecked_smul(a, b); + let acc = self.unchecked_sadd(prod, c); + self.unchecked_sadd(acc, d) + } else { + let prod = self.unchecked_umul(a, b); + let acc = self.unchecked_uadd(prod, c); + self.unchecked_uadd(acc, d) + }; + + let narrow_llty = self.type_ix(size.bits()); + let low = self.trunc(wide, narrow_llty); + let bits_const = self.const_uint(wide_llty, size.bits()); + // No need for ashr when signed; LLVM changes it to lshr anyway. + let high = self.lshr(wide, bits_const); + // FIXME: could be `trunc nuw`, even for signed. + let high = self.trunc(high, narrow_llty); + + let pair_llty = self.type_struct(&[narrow_llty, narrow_llty], false); + let pair = self.const_poison(pair_llty); + let pair = self.insert_value(pair, low, 0); + let pair = self.insert_value(pair, high, 1); + pair + } sym::ctlz | sym::ctlz_nonzero | sym::cttz @@ -352,84 +383,84 @@ impl<'ll, 'tcx> IntrinsicCallBuilderMethods<'tcx> for Builder<'_, 'll, 'tcx> { | sym::saturating_add | sym::saturating_sub => { let ty = arg_tys[0]; - match int_type_width_signed(ty, self) { - Some((width, signed)) => match name { - sym::ctlz | sym::cttz => { - let y = self.const_bool(false); - let ret = self.call_intrinsic(&format!("llvm.{name}.i{width}"), &[ - args[0].immediate(), - y, - ]); - - self.intcast(ret, llret_ty, false) - } - sym::ctlz_nonzero => { - let y = self.const_bool(true); - let llvm_name = &format!("llvm.ctlz.i{width}"); - let ret = self.call_intrinsic(llvm_name, &[args[0].immediate(), y]); - self.intcast(ret, llret_ty, false) - } - sym::cttz_nonzero => { - let y = self.const_bool(true); - let llvm_name = &format!("llvm.cttz.i{width}"); - let ret = self.call_intrinsic(llvm_name, &[args[0].immediate(), y]); - self.intcast(ret, llret_ty, false) - } - sym::ctpop => { - let ret = self.call_intrinsic(&format!("llvm.ctpop.i{width}"), &[args - [0] - .immediate()]); - self.intcast(ret, llret_ty, false) - } - sym::bswap => { - if width == 8 { - args[0].immediate() // byte swap a u8/i8 is just a no-op - } else { - self.call_intrinsic(&format!("llvm.bswap.i{width}"), &[ - args[0].immediate() - ]) - } - } - sym::bitreverse => self - .call_intrinsic(&format!("llvm.bitreverse.i{width}"), &[ + if !ty.is_integral() { + tcx.dcx().emit_err(InvalidMonomorphization::BasicIntegerType { + span, + name, + ty, + }); + return Ok(()); + } + let (size, signed) = ty.int_size_and_signed(self.tcx); + let width = size.bits(); + match name { + sym::ctlz | sym::cttz => { + let y = self.const_bool(false); + let ret = self.call_intrinsic(&format!("llvm.{name}.i{width}"), &[ + args[0].immediate(), + y, + ]); + + self.intcast(ret, llret_ty, false) + } + sym::ctlz_nonzero => { + let y = self.const_bool(true); + let llvm_name = &format!("llvm.ctlz.i{width}"); + let ret = self.call_intrinsic(llvm_name, &[args[0].immediate(), y]); + self.intcast(ret, llret_ty, false) + } + sym::cttz_nonzero => { + let y = self.const_bool(true); + let llvm_name = &format!("llvm.cttz.i{width}"); + let ret = self.call_intrinsic(llvm_name, &[args[0].immediate(), y]); + self.intcast(ret, llret_ty, false) + } + sym::ctpop => { + let ret = self.call_intrinsic(&format!("llvm.ctpop.i{width}"), &[ + args[0].immediate() + ]); + self.intcast(ret, llret_ty, false) + } + sym::bswap => { + if width == 8 { + args[0].immediate() // byte swap a u8/i8 is just a no-op + } else { + self.call_intrinsic(&format!("llvm.bswap.i{width}"), &[ args[0].immediate() - ]), - sym::rotate_left | sym::rotate_right => { - let is_left = name == sym::rotate_left; - let val = args[0].immediate(); - let raw_shift = args[1].immediate(); - // rotate = funnel shift with first two args the same - let llvm_name = - &format!("llvm.fsh{}.i{}", if is_left { 'l' } else { 'r' }, width); - - // llvm expects shift to be the same type as the values, but rust - // always uses `u32`. - let raw_shift = self.intcast(raw_shift, self.val_ty(val), false); - - self.call_intrinsic(llvm_name, &[val, val, raw_shift]) + ]) } - sym::saturating_add | sym::saturating_sub => { - let is_add = name == sym::saturating_add; - let lhs = args[0].immediate(); - let rhs = args[1].immediate(); - let llvm_name = &format!( - "llvm.{}{}.sat.i{}", - if signed { 's' } else { 'u' }, - if is_add { "add" } else { "sub" }, - width - ); - self.call_intrinsic(llvm_name, &[lhs, rhs]) - } - _ => bug!(), - }, - None => { - tcx.dcx().emit_err(InvalidMonomorphization::BasicIntegerType { - span, - name, - ty, - }); - return Ok(()); } + sym::bitreverse => self + .call_intrinsic(&format!("llvm.bitreverse.i{width}"), &[ + args[0].immediate() + ]), + sym::rotate_left | sym::rotate_right => { + let is_left = name == sym::rotate_left; + let val = args[0].immediate(); + let raw_shift = args[1].immediate(); + // rotate = funnel shift with first two args the same + let llvm_name = + &format!("llvm.fsh{}.i{}", if is_left { 'l' } else { 'r' }, width); + + // llvm expects shift to be the same type as the values, but rust + // always uses `u32`. + let raw_shift = self.intcast(raw_shift, self.val_ty(val), false); + + self.call_intrinsic(llvm_name, &[val, val, raw_shift]) + } + sym::saturating_add | sym::saturating_sub => { + let is_add = name == sym::saturating_add; + let lhs = args[0].immediate(); + let rhs = args[1].immediate(); + let llvm_name = &format!( + "llvm.{}{}.sat.i{}", + if signed { 's' } else { 'u' }, + if is_add { "add" } else { "sub" }, + width + ); + self.call_intrinsic(llvm_name, &[lhs, rhs]) + } + _ => bug!(), } } @@ -2531,19 +2562,3 @@ fn generic_simd_intrinsic<'ll, 'tcx>( span_bug!(span, "unknown SIMD intrinsic"); } - -// Returns the width of an int Ty, and if it's signed or not -// Returns None if the type is not an integer -// FIXME: there’s multiple of this functions, investigate using some of the already existing -// stuffs. -fn int_type_width_signed(ty: Ty<'_>, cx: &CodegenCx<'_, '_>) -> Option<(u64, bool)> { - match ty.kind() { - ty::Int(t) => { - Some((t.bit_width().unwrap_or(u64::from(cx.tcx.sess.target.pointer_width)), true)) - } - ty::Uint(t) => { - Some((t.bit_width().unwrap_or(u64::from(cx.tcx.sess.target.pointer_width)), false)) - } - _ => None, - } -} diff --git a/compiler/rustc_codegen_llvm/src/lib.rs b/compiler/rustc_codegen_llvm/src/lib.rs index 3dfb86d422d..06afe8bb3ad 100644 --- a/compiler/rustc_codegen_llvm/src/lib.rs +++ b/compiler/rustc_codegen_llvm/src/lib.rs @@ -17,6 +17,8 @@ #![feature(iter_intersperse)] #![feature(let_chains)] #![feature(rustdoc_internals)] +#![feature(slice_as_array)] +#![feature(try_blocks)] #![warn(unreachable_pub)] // tidy-alphabetical-end @@ -26,9 +28,10 @@ use std::mem::ManuallyDrop; use back::owned_target_machine::OwnedTargetMachine; use back::write::{create_informational_target_machine, create_target_machine}; -use errors::ParseTargetMachineConfig; -pub use llvm_util::target_features; +use errors::{AutoDiffWithoutLTO, ParseTargetMachineConfig}; +pub use llvm_util::target_features_cfg; use rustc_ast::expand::allocator::AllocatorKind; +use rustc_ast::expand::autodiff_attrs::AutoDiffItem; use rustc_codegen_ssa::back::lto::{LtoModuleCodegen, SerializedModule, ThinModule}; use rustc_codegen_ssa::back::write::{ CodegenContext, FatLtoInput, ModuleConfig, TargetMachineFactoryConfig, TargetMachineFactoryFn, @@ -36,14 +39,14 @@ use rustc_codegen_ssa::back::write::{ use rustc_codegen_ssa::traits::*; use rustc_codegen_ssa::{CodegenResults, CompiledModule, ModuleCodegen}; use rustc_data_structures::fx::FxIndexMap; -use rustc_errors::{DiagCtxtHandle, ErrorGuaranteed, FatalError}; +use rustc_errors::{DiagCtxtHandle, FatalError}; use rustc_metadata::EncodedMetadata; use rustc_middle::dep_graph::{WorkProduct, WorkProductId}; use rustc_middle::ty::TyCtxt; use rustc_middle::util::Providers; use rustc_session::Session; -use rustc_session::config::{OptLevel, OutputFilenames, PrintKind, PrintRequest}; -use rustc_span::symbol::Symbol; +use rustc_session::config::{Lto, OptLevel, OutputFilenames, PrintKind, PrintRequest}; +use rustc_span::Symbol; mod back { pub(crate) mod archive; @@ -231,6 +234,20 @@ impl WriteBackendMethods for LlvmCodegenBackend { fn serialize_module(module: ModuleCodegen<Self::Module>) -> (String, Self::ModuleBuffer) { (module.name, back::lto::ModuleBuffer::new(module.module_llvm.llmod())) } + /// Generate autodiff rules + fn autodiff( + cgcx: &CodegenContext<Self>, + tcx: TyCtxt<'_>, + module: &ModuleCodegen<Self::Module>, + diff_fncs: Vec<AutoDiffItem>, + config: &ModuleConfig, + ) -> Result<(), FatalError> { + if cgcx.lto != Lto::Fat { + let dcx = cgcx.create_dcx(); + return Err(dcx.handle().emit_almost_fatal(AutoDiffWithoutLTO)); + } + builder::autodiff::differentiate(module, cgcx, tcx, diff_fncs, config) + } } unsafe impl Send for LlvmCodegenBackend {} // Llvm is on a per-thread basis @@ -330,8 +347,8 @@ impl CodegenBackend for LlvmCodegenBackend { llvm_util::print_version(); } - fn target_features(&self, sess: &Session, allow_unstable: bool) -> Vec<Symbol> { - target_features(sess, allow_unstable) + fn target_features_cfg(&self, sess: &Session, allow_unstable: bool) -> Vec<Symbol> { + target_features_cfg(sess, allow_unstable) } fn codegen_crate<'tcx>( @@ -370,19 +387,14 @@ impl CodegenBackend for LlvmCodegenBackend { (codegen_results, work_products) } - fn link( - &self, - sess: &Session, - codegen_results: CodegenResults, - outputs: &OutputFilenames, - ) -> Result<(), ErrorGuaranteed> { + fn link(&self, sess: &Session, codegen_results: CodegenResults, outputs: &OutputFilenames) { use rustc_codegen_ssa::back::link::link_binary; use crate::back::archive::LlvmArchiveBuilderBuilder; // Run the linker on any artifacts that resulted from the LLVM run. // This should produce either a finished executable or library. - link_binary(sess, &LlvmArchiveBuilderBuilder, codegen_results, outputs) + link_binary(sess, &LlvmArchiveBuilderBuilder, codegen_results, outputs); } } diff --git a/compiler/rustc_codegen_llvm/src/llvm/diagnostic.rs b/compiler/rustc_codegen_llvm/src/llvm/diagnostic.rs index a4cb5a25d1b..11043b664f5 100644 --- a/compiler/rustc_codegen_llvm/src/llvm/diagnostic.rs +++ b/compiler/rustc_codegen_llvm/src/llvm/diagnostic.rs @@ -151,7 +151,7 @@ impl InlineAsmDiagnostic { unsafe { SrcMgrDiagnostic::unpack(super::LLVMRustGetSMDiagnostic(di, &mut cookie)) }; InlineAsmDiagnostic { level: smdiag.level, - cookie: cookie.into(), + cookie, message: smdiag.message, source: smdiag.source, } diff --git a/compiler/rustc_codegen_llvm/src/llvm/enzyme_ffi.rs b/compiler/rustc_codegen_llvm/src/llvm/enzyme_ffi.rs new file mode 100644 index 00000000000..729d6f62e24 --- /dev/null +++ b/compiler/rustc_codegen_llvm/src/llvm/enzyme_ffi.rs @@ -0,0 +1,29 @@ +#![allow(non_camel_case_types)] + +use libc::{c_char, c_uint}; + +use super::ffi::{BasicBlock, Metadata, Module, Type, Value}; +use crate::llvm::Bool; +extern "C" { + // Enzyme + pub fn LLVMRustHasMetadata(I: &Value, KindID: c_uint) -> bool; + pub fn LLVMRustEraseInstBefore(BB: &BasicBlock, I: &Value); + pub fn LLVMRustGetLastInstruction<'a>(BB: &BasicBlock) -> Option<&'a Value>; + pub fn LLVMRustDIGetInstMetadata(I: &Value) -> Option<&Metadata>; + pub fn LLVMRustEraseInstFromParent(V: &Value); + pub fn LLVMRustGetTerminator<'a>(B: &BasicBlock) -> &'a Value; + pub fn LLVMRustVerifyFunction(V: &Value, action: LLVMRustVerifierFailureAction) -> Bool; + + pub fn LLVMGetFunctionCallConv(F: &Value) -> c_uint; + pub fn LLVMGetReturnType(T: &Type) -> &Type; + pub fn LLVMGetParams(Fnc: &Value, parms: *mut &Value); + pub fn LLVMGetNamedFunction(M: &Module, Name: *const c_char) -> Option<&Value>; +} + +#[repr(C)] +#[derive(Copy, Clone, PartialEq)] +pub enum LLVMRustVerifierFailureAction { + LLVMAbortProcessAction = 0, + LLVMPrintMessageAction = 1, + LLVMReturnStatusAction = 2, +} diff --git a/compiler/rustc_codegen_llvm/src/llvm/ffi.rs b/compiler/rustc_codegen_llvm/src/llvm/ffi.rs index b1ace0033ba..472d4a3a72b 100644 --- a/compiler/rustc_codegen_llvm/src/llvm/ffi.rs +++ b/compiler/rustc_codegen_llvm/src/llvm/ffi.rs @@ -17,7 +17,9 @@ use super::debuginfo::{ DebugEmissionKind, DebugNameTableKind, }; -pub type Bool = c_uint; +/// In the LLVM-C API, boolean values are passed as `typedef int LLVMBool`, +/// which has a different ABI from Rust or C++ `bool`. +pub type Bool = c_int; pub const True: Bool = 1 as Bool; pub const False: Bool = 0 as Bool; @@ -54,25 +56,6 @@ pub enum LLVMRustResult { Failure, } -// Rust version of the C struct with the same name in rustc_llvm/llvm-wrapper/RustWrapper.cpp. -#[repr(C)] -pub struct LLVMRustCOFFShortExport { - pub name: *const c_char, - pub ordinal_present: bool, - /// value of `ordinal` only important when `ordinal_present` is true - pub ordinal: u16, -} - -impl LLVMRustCOFFShortExport { - pub fn new(name: *const c_char, ordinal: Option<u16>) -> LLVMRustCOFFShortExport { - LLVMRustCOFFShortExport { - name, - ordinal_present: ordinal.is_some(), - ordinal: ordinal.unwrap_or(0), - } - } -} - /// Translation of LLVM's MachineTypes enum, defined in llvm\include\llvm\BinaryFormat\COFF.h. /// /// We include only architectures supported on Windows. @@ -116,7 +99,7 @@ pub enum ModuleFlagMergeBehavior { /// LLVM CallingConv::ID. Should we wrap this? /// /// See <https://github.com/llvm/llvm-project/blob/main/llvm/include/llvm/IR/CallingConv.h> -#[derive(Copy, Clone, PartialEq, Debug)] +#[derive(Copy, Clone, PartialEq, Debug, TryFromU32)] #[repr(C)] pub enum CallConv { CCallConv = 0, @@ -543,7 +526,7 @@ pub struct SanitizerOptions { pub sanitize_kernel_address_recover: bool, } -/// LLVMRelocMode +/// LLVMRustRelocModel #[derive(Copy, Clone, PartialEq)] #[repr(C)] pub enum RelocModel { @@ -555,6 +538,15 @@ pub enum RelocModel { ROPI_RWPI, } +/// LLVMRustFloatABI +#[derive(Copy, Clone, PartialEq)] +#[repr(C)] +pub enum FloatAbi { + Default, + Soft, + Hard, +} + /// LLVMRustCodeModel #[derive(Copy, Clone)] #[repr(C)] @@ -2209,7 +2201,7 @@ unsafe extern "C" { Model: CodeModel, Reloc: RelocModel, Level: CodeGenOptLevel, - UseSoftFP: bool, + FloatABIType: FloatAbi, FunctionSections: bool, DataSections: bool, UniqueSectionNames: bool, @@ -2317,7 +2309,7 @@ unsafe extern "C" { pub fn LLVMRustGetSMDiagnostic<'a>( DI: &'a DiagnosticInfo, - cookie_out: &mut c_uint, + cookie_out: &mut u64, ) -> &'a SMDiagnostic; pub fn LLVMRustUnpackSMDiagnostic( @@ -2345,15 +2337,6 @@ unsafe extern "C" { ) -> &'a mut RustArchiveMember<'a>; pub fn LLVMRustArchiveMemberFree<'a>(Member: &'a mut RustArchiveMember<'a>); - pub fn LLVMRustWriteImportLibrary( - ImportName: *const c_char, - Path: *const c_char, - Exports: *const LLVMRustCOFFShortExport, - NumExports: usize, - Machine: u16, - MinGW: bool, - ) -> LLVMRustResult; - pub fn LLVMRustSetDataLayoutFromTargetMachine<'a>(M: &'a Module, TM: &'a TargetMachine); pub fn LLVMRustPositionBuilderAtStart<'a>(B: &Builder<'a>, BB: &'a BasicBlock); diff --git a/compiler/rustc_codegen_llvm/src/llvm/mod.rs b/compiler/rustc_codegen_llvm/src/llvm/mod.rs index 909afe35a17..2592a7df95c 100644 --- a/compiler/rustc_codegen_llvm/src/llvm/mod.rs +++ b/compiler/rustc_codegen_llvm/src/llvm/mod.rs @@ -22,8 +22,11 @@ use crate::common::AsCCharPtr; pub mod archive_ro; pub mod diagnostic; +pub mod enzyme_ffi; mod ffi; +pub use self::enzyme_ffi::*; + impl LLVMRustResult { pub fn into_result(self) -> Result<(), ()> { match self { diff --git a/compiler/rustc_codegen_llvm/src/llvm_util.rs b/compiler/rustc_codegen_llvm/src/llvm_util.rs index db2b03d9aed..628c0b1c29c 100644 --- a/compiler/rustc_codegen_llvm/src/llvm_util.rs +++ b/compiler/rustc_codegen_llvm/src/llvm_util.rs @@ -15,9 +15,9 @@ use rustc_fs_util::path_to_c_string; use rustc_middle::bug; use rustc_session::Session; use rustc_session::config::{PrintKind, PrintRequest}; -use rustc_span::symbol::Symbol; +use rustc_span::Symbol; use rustc_target::spec::{MergeFunctions, PanicStrategy, SmallDataThresholdSupport}; -use rustc_target::target_features::{RUSTC_SPECIAL_FEATURES, RUSTC_SPECIFIC_FEATURES, Stability}; +use rustc_target::target_features::{RUSTC_SPECIAL_FEATURES, RUSTC_SPECIFIC_FEATURES}; use crate::back::write::create_informational_target_machine; use crate::errors::{ @@ -230,6 +230,8 @@ pub(crate) fn to_llvm_features<'a>(sess: &Session, s: &'a str) -> Option<LLVMFea "aarch64" } else if sess.target.arch == "sparc64" { "sparc" + } else if sess.target.arch == "powerpc64" { + "powerpc" } else { &*sess.target.arch }; @@ -289,6 +291,7 @@ pub(crate) fn to_llvm_features<'a>(sess: &Session, s: &'a str) -> Option<LLVMFea // https://github.com/llvm/llvm-project/blob/llvmorg-18.1.0/llvm/lib/Target/Sparc/MCTargetDesc/SparcELFObjectWriter.cpp#L26 ("sparc", "v8plus") if get_version().0 == 19 => Some(LLVMFeature::new("v9")), ("sparc", "v8plus") if get_version().0 < 19 => None, + ("powerpc", "power8-crypto") => Some(LLVMFeature::new("crypto")), (_, s) => Some(LLVMFeature::new(s)), } } @@ -297,7 +300,7 @@ pub(crate) fn to_llvm_features<'a>(sess: &Session, s: &'a str) -> Option<LLVMFea /// Must express features in the way Rust understands them. /// /// We do not have to worry about RUSTC_SPECIFIC_FEATURES here, those are handled outside codegen. -pub fn target_features(sess: &Session, allow_unstable: bool) -> Vec<Symbol> { +pub fn target_features_cfg(sess: &Session, allow_unstable: bool) -> Vec<Symbol> { let mut features: FxHashSet<Symbol> = Default::default(); // Add base features for the target. @@ -313,7 +316,7 @@ pub fn target_features(sess: &Session, allow_unstable: bool) -> Vec<Symbol> { sess.target .rust_target_features() .iter() - .filter(|(_, gate, _)| gate.is_supported()) + .filter(|(_, gate, _)| gate.in_cfg()) .filter(|(feature, _, _)| { // skip checking special features, as LLVM may not understand them if RUSTC_SPECIAL_FEATURES.contains(feature) { @@ -369,10 +372,10 @@ pub fn target_features(sess: &Session, allow_unstable: bool) -> Vec<Symbol> { sess.target .rust_target_features() .iter() - .filter(|(_, gate, _)| gate.is_supported()) - .filter_map(|&(feature, gate, _)| { - if sess.is_nightly_build() || allow_unstable || gate.is_stable() { - Some(feature) + .filter(|(_, gate, _)| gate.in_cfg()) + .filter_map(|(feature, gate, _)| { + if sess.is_nightly_build() || allow_unstable || gate.requires_nightly().is_none() { + Some(*feature) } else { None } @@ -490,7 +493,7 @@ fn print_target_features(sess: &Session, tm: &llvm::TargetMachine, out: &mut Str .rust_target_features() .iter() .filter_map(|(feature, gate, _implied)| { - if !gate.is_supported() { + if !gate.in_cfg() { // Only list (experimentally) supported features. return None; } @@ -713,13 +716,17 @@ pub(crate) fn global_llvm_features( }; sess.dcx().emit_warn(unknown_feature); } - Some((_, Stability::Stable, _)) => {} - Some((_, Stability::Unstable(_), _)) => { - // An unstable feature. Warn about using it. - sess.dcx().emit_warn(UnstableCTargetFeature { feature }); - } - Some((_, Stability::Forbidden { reason }, _)) => { - sess.dcx().emit_warn(ForbiddenCTargetFeature { feature, reason }); + Some((_, stability, _)) => { + if let Err(reason) = + stability.toggle_allowed(&sess.target, enable_disable == '+') + { + sess.dcx().emit_warn(ForbiddenCTargetFeature { feature, reason }); + } else if stability.requires_nightly().is_some() { + // An unstable feature. Warn about using it. It makes little sense + // to hard-error here since we just warn about fully unknown + // features above. + sess.dcx().emit_warn(UnstableCTargetFeature { feature }); + } } } diff --git a/compiler/rustc_codegen_llvm/src/type_of.rs b/compiler/rustc_codegen_llvm/src/type_of.rs index 2b05e24a7ba..b0b6da869da 100644 --- a/compiler/rustc_codegen_llvm/src/type_of.rs +++ b/compiler/rustc_codegen_llvm/src/type_of.rs @@ -38,7 +38,7 @@ fn uncached_llvm_type<'a, 'tcx>( if let (&ty::Adt(def, _), &Variants::Single { index }) = (layout.ty.kind(), &layout.variants) { - if def.is_enum() && !def.variants().is_empty() { + if def.is_enum() { write!(&mut name, "::{}", def.variant(index).name).unwrap(); } } |