diff options
Diffstat (limited to 'compiler/rustc_codegen_llvm/src/back')
| -rw-r--r-- | compiler/rustc_codegen_llvm/src/back/archive.rs | 11 | ||||
| -rw-r--r-- | compiler/rustc_codegen_llvm/src/back/lto.rs | 76 | ||||
| -rw-r--r-- | compiler/rustc_codegen_llvm/src/back/owned_target_machine.rs | 2 | ||||
| -rw-r--r-- | compiler/rustc_codegen_llvm/src/back/write.rs | 256 |
4 files changed, 186 insertions, 159 deletions
diff --git a/compiler/rustc_codegen_llvm/src/back/archive.rs b/compiler/rustc_codegen_llvm/src/back/archive.rs index 33a956e552f..93553f3f364 100644 --- a/compiler/rustc_codegen_llvm/src/back/archive.rs +++ b/compiler/rustc_codegen_llvm/src/back/archive.rs @@ -11,7 +11,7 @@ use rustc_codegen_ssa::back::archive::{ use rustc_session::Session; use crate::llvm::archive_ro::{ArchiveRO, Child}; -use crate::llvm::{self, ArchiveKind}; +use crate::llvm::{self, ArchiveKind, last_error}; /// Helper for adding many files to an archive. #[must_use = "must call build() to finish building the archive"] @@ -169,6 +169,8 @@ impl<'a> LlvmArchiveBuilder<'a> { .unwrap_or_else(|kind| self.sess.dcx().emit_fatal(UnknownArchiveKind { kind })); let mut additions = mem::take(&mut self.additions); + // Values in the `members` list below will contain pointers to the strings allocated here. + // So they need to get dropped after all elements of `members` get freed. let mut strings = Vec::new(); let mut members = Vec::new(); @@ -229,12 +231,7 @@ impl<'a> LlvmArchiveBuilder<'a> { self.sess.target.arch == "arm64ec", ); let ret = if r.into_result().is_err() { - let err = llvm::LLVMRustGetLastError(); - let msg = if err.is_null() { - "failed to write archive".into() - } else { - String::from_utf8_lossy(CStr::from_ptr(err).to_bytes()) - }; + let msg = last_error().unwrap_or_else(|| "failed to write archive".into()); Err(io::Error::new(io::ErrorKind::Other, msg)) } else { Ok(!members.is_empty()) diff --git a/compiler/rustc_codegen_llvm/src/back/lto.rs b/compiler/rustc_codegen_llvm/src/back/lto.rs index 78c759bbe8c..99906ea7bce 100644 --- a/compiler/rustc_codegen_llvm/src/back/lto.rs +++ b/compiler/rustc_codegen_llvm/src/back/lto.rs @@ -362,8 +362,8 @@ fn fat_lto( ptr as *const *const libc::c_char, symbols_below_threshold.len() as libc::size_t, ); - save_temp_bitcode(cgcx, &module, "lto.after-restriction"); } + save_temp_bitcode(cgcx, &module, "lto.after-restriction"); } Ok(LtoModuleCodegen::Fat(module)) @@ -586,6 +586,42 @@ fn thin_lto( } } +fn enable_autodiff_settings(ad: &[config::AutoDiff], module: &mut ModuleCodegen<ModuleLlvm>) { + for &val in ad { + match val { + config::AutoDiff::PrintModBefore => { + unsafe { llvm::LLVMDumpModule(module.module_llvm.llmod()) }; + } + config::AutoDiff::PrintPerf => { + llvm::set_print_perf(true); + } + config::AutoDiff::PrintAA => { + llvm::set_print_activity(true); + } + config::AutoDiff::PrintTA => { + llvm::set_print_type(true); + } + config::AutoDiff::Inline => { + llvm::set_inline(true); + } + config::AutoDiff::LooseTypes => { + llvm::set_loose_types(false); + } + config::AutoDiff::PrintSteps => { + llvm::set_print(true); + } + // We handle this below + config::AutoDiff::PrintModAfter => {} + // This is required and already checked + config::AutoDiff::Enable => {} + } + } + // This helps with handling enums for now. + llvm::set_strict_aliasing(false); + // FIXME(ZuseZ4): Test this, since it was added a long time ago. + llvm::set_rust_rules(true); +} + pub(crate) fn run_pass_manager( cgcx: &CodegenContext<LlvmCodegenBackend>, dcx: DiagCtxtHandle<'_>, @@ -604,13 +640,37 @@ pub(crate) fn run_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); - // 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"); + // The PostAD behavior is the same that we would have if no autodiff was used. + // It will run the default optimization pipeline. If AD is enabled we select + // the DuringAD stage, which will disable vectorization and loop unrolling, and + // schedule two autodiff optimization + differentiation passes. + // We then run the llvm_optimize function a second time, to optimize the code which we generated + // in the enzyme differentiation pass. + let enable_ad = config.autodiff.contains(&config::AutoDiff::Enable); + let stage = + if enable_ad { write::AutodiffStage::DuringAD } else { write::AutodiffStage::PostAD }; + + if enable_ad { + enable_autodiff_settings(&config.autodiff, module); + } + unsafe { - write::llvm_optimize(cgcx, dcx, module, config, opt_level, opt_stage, first_run)?; + write::llvm_optimize(cgcx, dcx, module, config, opt_level, opt_stage, stage)?; + } + + if cfg!(llvm_enzyme) && enable_ad { + let opt_stage = llvm::OptStage::FatLTO; + let stage = write::AutodiffStage::PostAD; + unsafe { + write::llvm_optimize(cgcx, dcx, module, config, opt_level, opt_stage, stage)?; + } + + // This is the final IR, so people should be able to inspect the optimized autodiff output. + if config.autodiff.contains(&config::AutoDiff::PrintModAfter) { + unsafe { llvm::LLVMDumpModule(module.module_llvm.llmod()) }; + } } + debug!("lto done"); Ok(()) } @@ -621,7 +681,7 @@ unsafe impl Send for ModuleBuffer {} unsafe impl Sync for ModuleBuffer {} impl ModuleBuffer { - pub fn new(m: &llvm::Module) -> ModuleBuffer { + pub(crate) fn new(m: &llvm::Module) -> ModuleBuffer { ModuleBuffer(unsafe { llvm::LLVMRustModuleBufferCreate(m) }) } } @@ -663,7 +723,7 @@ unsafe impl Send for ThinBuffer {} unsafe impl Sync for ThinBuffer {} impl ThinBuffer { - pub fn new(m: &llvm::Module, is_thin: bool, emit_summary: bool) -> ThinBuffer { + pub(crate) fn new(m: &llvm::Module, is_thin: bool, emit_summary: bool) -> ThinBuffer { unsafe { let buffer = llvm::LLVMRustThinLTOBufferCreate(m, is_thin, emit_summary); ThinBuffer(buffer) 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 4cbd49aa44d..f075f332462 100644 --- a/compiler/rustc_codegen_llvm/src/back/owned_target_machine.rs +++ b/compiler/rustc_codegen_llvm/src/back/owned_target_machine.rs @@ -17,7 +17,7 @@ pub struct OwnedTargetMachine { } impl OwnedTargetMachine { - pub fn new( + pub(crate) fn new( triple: &CStr, cpu: &CStr, features: &CStr, diff --git a/compiler/rustc_codegen_llvm/src/back/write.rs b/compiler/rustc_codegen_llvm/src/back/write.rs index 4706744f353..b67890c0465 100644 --- a/compiler/rustc_codegen_llvm/src/back/write.rs +++ b/compiler/rustc_codegen_llvm/src/back/write.rs @@ -40,7 +40,7 @@ use crate::errors::{ WithLlvmError, WriteBytecode, }; use crate::llvm::diagnostic::OptimizationDiagnosticKind::*; -use crate::llvm::{self, DiagnosticInfo, PassManager}; +use crate::llvm::{self, DiagnosticInfo}; use crate::type_::Type; use crate::{LlvmCodegenBackend, ModuleLlvm, base, common, llvm_util}; @@ -54,7 +54,7 @@ pub(crate) fn llvm_err<'a>(dcx: DiagCtxtHandle<'_>, err: LlvmError<'a>) -> Fatal fn write_output_file<'ll>( dcx: DiagCtxtHandle<'_>, target: &'ll llvm::TargetMachine, - pm: &llvm::PassManager<'ll>, + no_builtins: bool, m: &'ll llvm::Module, output: &Path, dwo_output: Option<&Path>, @@ -63,16 +63,19 @@ fn write_output_file<'ll>( verify_llvm_ir: bool, ) -> Result<(), FatalError> { debug!("write_output_file output={:?} dwo_output={:?}", output, dwo_output); - unsafe { - let output_c = path_to_c_string(output); - let dwo_output_c; - let dwo_output_ptr = if let Some(dwo_output) = dwo_output { - dwo_output_c = path_to_c_string(dwo_output); - dwo_output_c.as_ptr() - } else { - std::ptr::null() - }; - let result = llvm::LLVMRustWriteOutputFile( + let output_c = path_to_c_string(output); + let dwo_output_c; + let dwo_output_ptr = if let Some(dwo_output) = dwo_output { + dwo_output_c = path_to_c_string(dwo_output); + dwo_output_c.as_ptr() + } else { + std::ptr::null() + }; + let result = unsafe { + let pm = llvm::LLVMCreatePassManager(); + llvm::LLVMAddAnalysisPasses(target, pm); + llvm::LLVMRustAddLibraryInfo(pm, m, no_builtins); + llvm::LLVMRustWriteOutputFile( target, pm, m, @@ -80,22 +83,22 @@ fn write_output_file<'ll>( dwo_output_ptr, file_type, verify_llvm_ir, - ); + ) + }; - // Record artifact sizes for self-profiling - if result == llvm::LLVMRustResult::Success { - let artifact_kind = match file_type { - llvm::FileType::ObjectFile => "object_file", - llvm::FileType::AssemblyFile => "assembly_file", - }; - record_artifact_size(self_profiler_ref, artifact_kind, output); - if let Some(dwo_file) = dwo_output { - record_artifact_size(self_profiler_ref, "dwo_file", dwo_file); - } + // Record artifact sizes for self-profiling + if result == llvm::LLVMRustResult::Success { + let artifact_kind = match file_type { + llvm::FileType::ObjectFile => "object_file", + llvm::FileType::AssemblyFile => "assembly_file", + }; + record_artifact_size(self_profiler_ref, artifact_kind, output); + if let Some(dwo_file) = dwo_output { + record_artifact_size(self_profiler_ref, "dwo_file", dwo_file); } - - result.into_result().map_err(|()| llvm_err(dcx, LlvmError::WriteOutput { path: output })) } + + result.into_result().map_err(|()| llvm_err(dcx, LlvmError::WriteOutput { path: output })) } pub(crate) fn create_informational_target_machine( @@ -138,7 +141,7 @@ fn to_llvm_opt_settings(cfg: config::OptLevel) -> (llvm::CodeGenOptLevel, llvm:: match cfg { No => (llvm::CodeGenOptLevel::None, llvm::CodeGenOptSizeNone), Less => (llvm::CodeGenOptLevel::Less, llvm::CodeGenOptSizeNone), - Default => (llvm::CodeGenOptLevel::Default, llvm::CodeGenOptSizeNone), + More => (llvm::CodeGenOptLevel::Default, llvm::CodeGenOptSizeNone), Aggressive => (llvm::CodeGenOptLevel::Aggressive, llvm::CodeGenOptSizeNone), Size => (llvm::CodeGenOptLevel::Default, llvm::CodeGenOptSizeDefault), SizeMin => (llvm::CodeGenOptLevel::Default, llvm::CodeGenOptSizeAggressive), @@ -150,7 +153,7 @@ fn to_pass_builder_opt_level(cfg: config::OptLevel) -> llvm::PassBuilderOptLevel match cfg { No => llvm::PassBuilderOptLevel::O0, Less => llvm::PassBuilderOptLevel::O1, - Default => llvm::PassBuilderOptLevel::O2, + More => llvm::PassBuilderOptLevel::O2, Aggressive => llvm::PassBuilderOptLevel::O3, Size => llvm::PassBuilderOptLevel::Os, SizeMin => llvm::PassBuilderOptLevel::Oz, @@ -325,13 +328,17 @@ pub(crate) fn save_temp_bitcode( if !cgcx.save_temps { return; } + let ext = format!("{name}.bc"); + let cgu = Some(&module.name[..]); + let path = cgcx.output_filenames.temp_path_ext(&ext, cgu); + write_bitcode_to_file(module, &path) +} + +fn write_bitcode_to_file(module: &ModuleCodegen<ModuleLlvm>, path: &Path) { unsafe { - let ext = format!("{name}.bc"); - let cgu = Some(&module.name[..]); - let path = cgcx.output_filenames.temp_path_ext(&ext, cgu); - let cstr = path_to_c_string(&path); + let path = path_to_c_string(&path); let llmod = module.module_llvm.llmod(); - llvm::LLVMWriteBitcodeToFile(llmod, cstr.as_ptr()); + llvm::LLVMWriteBitcodeToFile(llmod, path.as_ptr()); } } @@ -530,6 +537,16 @@ fn get_instr_profile_output_path(config: &ModuleConfig) -> Option<CString> { config.instrument_coverage.then(|| c"default_%m_%p.profraw".to_owned()) } +// PreAD will run llvm opts but disable size increasing opts (vectorization, loop unrolling) +// DuringAD is the same as above, but also runs the enzyme opt and autodiff passes. +// PostAD will run all opts, including size increasing opts. +#[derive(Debug, Eq, PartialEq)] +pub(crate) enum AutodiffStage { + PreAD, + DuringAD, + PostAD, +} + pub(crate) unsafe fn llvm_optimize( cgcx: &CodegenContext<LlvmCodegenBackend>, dcx: DiagCtxtHandle<'_>, @@ -537,7 +554,7 @@ pub(crate) unsafe fn llvm_optimize( config: &ModuleConfig, opt_level: config::OptLevel, opt_stage: llvm::OptStage, - skip_size_increasing_opts: bool, + autodiff_stage: AutodiffStage, ) -> Result<(), FatalError> { // Enzyme: // The whole point of compiler based AD is to differentiate optimized IR instead of unoptimized @@ -547,15 +564,16 @@ pub(crate) unsafe fn llvm_optimize( // FIXME(ZuseZ4): In a future update we could figure out how to only optimize individual functions getting // differentiated. + let consider_ad = cfg!(llvm_enzyme) && config.autodiff.contains(&config::AutoDiff::Enable); + let run_enzyme = autodiff_stage == AutodiffStage::DuringAD; 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) { + // optimizations until after differentiation. Our pipeline is thus: (opt + enzyme), (full opt). + // We therefore have two calls to llvm_optimize, if autodiff is used. + if consider_ad && autodiff_stage != AutodiffStage::PostAD { unroll_loops = false; vectorize_slp = false; vectorize_loop = false; @@ -565,7 +583,7 @@ pub(crate) unsafe fn llvm_optimize( vectorize_slp = config.vectorize_slp; vectorize_loop = config.vectorize_loop; } - trace!(?unroll_loops, ?vectorize_slp, ?vectorize_loop); + trace!(?unroll_loops, ?vectorize_slp, ?vectorize_loop, ?run_enzyme); 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); @@ -633,6 +651,7 @@ pub(crate) unsafe fn llvm_optimize( vectorize_loop, config.no_builtins, config.emit_lifetime_markers, + run_enzyme, sanitizer_options.as_ref(), pgo_gen_path.as_ref().map_or(std::ptr::null(), |s| s.as_ptr()), pgo_use_path.as_ref().map_or(std::ptr::null(), |s| s.as_ptr()), @@ -661,7 +680,6 @@ pub(crate) unsafe fn optimize( ) -> Result<(), FatalError> { let _timer = cgcx.prof.generic_activity_with_arg("LLVM_module_optimize", &*module.name); - let llmod = module.module_llvm.llmod(); let llcx = &*module.module_llvm.llcx; let _handlers = DiagnosticHandlers::new(cgcx, dcx, llcx, module, CodegenDiagnosticsStage::Opt); @@ -670,8 +688,7 @@ pub(crate) unsafe fn optimize( if config.emit_no_opt_bc { let out = cgcx.output_filenames.temp_path_ext("no-opt.bc", module_name); - let out = path_to_c_string(&out); - unsafe { llvm::LLVMWriteBitcodeToFile(llmod, out.as_ptr()) }; + write_bitcode_to_file(module, &out) } // FIXME(ZuseZ4): support SanitizeHWAddress and prevent illegal/unsupported opts @@ -684,18 +701,12 @@ pub(crate) unsafe fn optimize( _ => llvm::OptStage::PreLinkNoLTO, }; - // If we know that we will later run AD, then we disable vectorization and loop unrolling - let skip_size_increasing_opts = cfg!(llvm_enzyme); + // If we know that we will later run AD, then we disable vectorization and loop unrolling. + // Otherwise we pretend AD is already done and run the normal opt pipeline (=PostAD). + let consider_ad = cfg!(llvm_enzyme) && config.autodiff.contains(&config::AutoDiff::Enable); + let autodiff_stage = if consider_ad { AutodiffStage::PreAD } else { AutodiffStage::PostAD }; return unsafe { - llvm_optimize( - cgcx, - dcx, - module, - config, - opt_level, - opt_stage, - skip_size_increasing_opts, - ) + llvm_optimize(cgcx, dcx, module, config, opt_level, opt_stage, autodiff_stage) }; } Ok(()) @@ -744,31 +755,6 @@ pub(crate) unsafe fn codegen( create_msvc_imps(cgcx, llcx, llmod); } - // A codegen-specific pass manager is used to generate object - // files for an LLVM module. - // - // Apparently each of these pass managers is a one-shot kind of - // thing, so we create a new one for each type of output. The - // pass manager passed to the closure should be ensured to not - // escape the closure itself, and the manager should only be - // used once. - unsafe fn with_codegen<'ll, F, R>( - tm: &'ll llvm::TargetMachine, - llmod: &'ll llvm::Module, - no_builtins: bool, - f: F, - ) -> R - where - F: FnOnce(&'ll mut PassManager<'ll>) -> R, - { - unsafe { - let cpm = llvm::LLVMCreatePassManager(); - llvm::LLVMAddAnalysisPasses(tm, cpm); - llvm::LLVMRustAddLibraryInfo(cpm, llmod, no_builtins); - f(cpm) - } - } - // Note that if object files are just LLVM bitcode we write bitcode, // copy it to the .o file, and delete the bitcode if it wasn't // otherwise requested. @@ -887,21 +873,17 @@ pub(crate) unsafe fn codegen( } else { llmod }; - unsafe { - with_codegen(tm, llmod, config.no_builtins, |cpm| { - write_output_file( - dcx, - tm, - cpm, - llmod, - &path, - None, - llvm::FileType::AssemblyFile, - &cgcx.prof, - config.verify_llvm_ir, - ) - })?; - } + write_output_file( + dcx, + tm, + config.no_builtins, + llmod, + &path, + None, + llvm::FileType::AssemblyFile, + &cgcx.prof, + config.verify_llvm_ir, + )?; } match config.emit_obj { @@ -925,21 +907,17 @@ pub(crate) unsafe fn codegen( (_, SplitDwarfKind::Split) => Some(dwo_out.as_path()), }; - unsafe { - with_codegen(tm, llmod, config.no_builtins, |cpm| { - write_output_file( - dcx, - tm, - cpm, - llmod, - &obj_out, - dwo_out, - llvm::FileType::ObjectFile, - &cgcx.prof, - config.verify_llvm_ir, - ) - })?; - } + write_output_file( + dcx, + tm, + config.no_builtins, + llmod, + &obj_out, + dwo_out, + llvm::FileType::ObjectFile, + &cgcx.prof, + config.verify_llvm_ir, + )?; } EmitObj::Bitcode => { @@ -1066,24 +1044,18 @@ unsafe fn embed_bitcode( { // We don't need custom section flags, create LLVM globals. let llconst = common::bytes_in_context(llcx, bitcode); - let llglobal = llvm::LLVMAddGlobal( - llmod, - common::val_ty(llconst), - c"rustc.embedded.module".as_ptr(), - ); - llvm::LLVMSetInitializer(llglobal, llconst); + let llglobal = + llvm::add_global(llmod, common::val_ty(llconst), c"rustc.embedded.module"); + llvm::set_initializer(llglobal, llconst); llvm::set_section(llglobal, bitcode_section_name(cgcx)); llvm::set_linkage(llglobal, llvm::Linkage::PrivateLinkage); llvm::LLVMSetGlobalConstant(llglobal, llvm::True); let llconst = common::bytes_in_context(llcx, cmdline.as_bytes()); - let llglobal = llvm::LLVMAddGlobal( - llmod, - common::val_ty(llconst), - c"rustc.embedded.cmdline".as_ptr(), - ); - llvm::LLVMSetInitializer(llglobal, llconst); + let llglobal = + llvm::add_global(llmod, common::val_ty(llconst), c"rustc.embedded.cmdline"); + llvm::set_initializer(llglobal, llconst); let section = if cgcx.target_is_like_osx { c"__LLVM,__cmdline" } else if cgcx.target_is_like_aix { @@ -1123,31 +1095,29 @@ fn create_msvc_imps( // underscores added in front). let prefix = if cgcx.target_arch == "x86" { "\x01__imp__" } else { "\x01__imp_" }; - unsafe { - let ptr_ty = Type::ptr_llcx(llcx); - let globals = base::iter_globals(llmod) - .filter(|&val| { - llvm::get_linkage(val) == llvm::Linkage::ExternalLinkage - && llvm::LLVMIsDeclaration(val) == 0 - }) - .filter_map(|val| { - // Exclude some symbols that we know are not Rust symbols. - let name = llvm::get_value_name(val); - if ignored(name) { None } else { Some((val, name)) } - }) - .map(move |(val, name)| { - let mut imp_name = prefix.as_bytes().to_vec(); - imp_name.extend(name); - let imp_name = CString::new(imp_name).unwrap(); - (imp_name, val) - }) - .collect::<Vec<_>>(); + let ptr_ty = Type::ptr_llcx(llcx); + let globals = base::iter_globals(llmod) + .filter(|&val| { + llvm::get_linkage(val) == llvm::Linkage::ExternalLinkage && !llvm::is_declaration(val) + }) + .filter_map(|val| { + // Exclude some symbols that we know are not Rust symbols. + let name = llvm::get_value_name(val); + if ignored(name) { None } else { Some((val, name)) } + }) + .map(move |(val, name)| { + let mut imp_name = prefix.as_bytes().to_vec(); + imp_name.extend(name); + let imp_name = CString::new(imp_name).unwrap(); + (imp_name, val) + }) + .collect::<Vec<_>>(); - for (imp_name, val) in globals { - let imp = llvm::LLVMAddGlobal(llmod, ptr_ty, imp_name.as_ptr()); - llvm::LLVMSetInitializer(imp, val); - llvm::set_linkage(imp, llvm::Linkage::ExternalLinkage); - } + for (imp_name, val) in globals { + let imp = llvm::add_global(llmod, ptr_ty, &imp_name); + + llvm::set_initializer(imp, val); + llvm::set_linkage(imp, llvm::Linkage::ExternalLinkage); } // Use this function to exclude certain symbols from `__imp` generation. |