diff options
Diffstat (limited to 'compiler/rustc_codegen_llvm/src')
20 files changed, 190 insertions, 208 deletions
diff --git a/compiler/rustc_codegen_llvm/src/abi.rs b/compiler/rustc_codegen_llvm/src/abi.rs index 685b2f37c9c..28f423efc21 100644 --- a/compiler/rustc_codegen_llvm/src/abi.rs +++ b/compiler/rustc_codegen_llvm/src/abi.rs @@ -664,7 +664,7 @@ impl<'tcx> AbiBuilderMethods<'tcx> for Builder<'_, '_, 'tcx> { } impl llvm::CallConv { - pub fn from_conv(conv: Conv, arch: &str) -> Self { + pub(crate) fn from_conv(conv: Conv, arch: &str) -> Self { match conv { Conv::C | Conv::Rust @@ -687,7 +687,6 @@ impl llvm::CallConv { Conv::AvrNonBlockingInterrupt => llvm::AvrNonBlockingInterrupt, Conv::ArmAapcs => llvm::ArmAapcsCallConv, Conv::Msp430Intr => llvm::Msp430Intr, - Conv::PtxKernel => llvm::PtxKernel, Conv::X86Fastcall => llvm::X86FastcallCallConv, Conv::X86Intr => llvm::X86_Intr, Conv::X86Stdcall => llvm::X86StdcallCallConv, diff --git a/compiler/rustc_codegen_llvm/src/allocator.rs b/compiler/rustc_codegen_llvm/src/allocator.rs index 149ded28356..66723cbf882 100644 --- a/compiler/rustc_codegen_llvm/src/allocator.rs +++ b/compiler/rustc_codegen_llvm/src/allocator.rs @@ -81,13 +81,13 @@ pub(crate) unsafe fn codegen( llvm::set_visibility(ll_g, llvm::Visibility::from_generic(tcx.sess.default_visibility())); let val = tcx.sess.opts.unstable_opts.oom.should_panic(); let llval = llvm::LLVMConstInt(i8, val as u64, False); - llvm::LLVMSetInitializer(ll_g, llval); + llvm::set_initializer(ll_g, llval); let name = NO_ALLOC_SHIM_IS_UNSTABLE; let ll_g = llvm::LLVMRustGetOrInsertGlobal(llmod, name.as_c_char_ptr(), name.len(), i8); llvm::set_visibility(ll_g, llvm::Visibility::from_generic(tcx.sess.default_visibility())); let llval = llvm::LLVMConstInt(i8, 0, False); - llvm::LLVMSetInitializer(ll_g, llval); + llvm::set_initializer(ll_g, llval); } if tcx.sess.opts.debuginfo != DebugInfo::None { diff --git a/compiler/rustc_codegen_llvm/src/asm.rs b/compiler/rustc_codegen_llvm/src/asm.rs index 3722d4350a2..be5673eddf9 100644 --- a/compiler/rustc_codegen_llvm/src/asm.rs +++ b/compiler/rustc_codegen_llvm/src/asm.rs @@ -286,7 +286,9 @@ impl<'ll, 'tcx> AsmBuilderMethods<'tcx> for Builder<'_, 'll, 'tcx> { InlineAsmArch::M68k => { constraints.push("~{ccr}".to_string()); } - InlineAsmArch::CSKY => {} + InlineAsmArch::CSKY => { + constraints.push("~{psr}".to_string()); + } } } if !options.contains(InlineAsmOptions::NOMEM) { 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..7262fce4911 100644 --- a/compiler/rustc_codegen_llvm/src/back/lto.rs +++ b/compiler/rustc_codegen_llvm/src/back/lto.rs @@ -606,10 +606,31 @@ pub(crate) fn run_pass_manager( // 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)?; + write::llvm_optimize( + cgcx, + dcx, + module, + config, + opt_level, + opt_stage, + write::AutodiffStage::DuringAD, + )?; + } + // FIXME(ZuseZ4): Make this more granular + if cfg!(llvm_enzyme) && !thin { + unsafe { + write::llvm_optimize( + cgcx, + dcx, + module, + config, + opt_level, + llvm::OptStage::FatLTO, + write::AutodiffStage::PostAD, + )?; + } } debug!("lto done"); Ok(()) @@ -621,7 +642,7 @@ 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 +684,7 @@ 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..58933a77e53 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( @@ -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 @@ -550,12 +567,16 @@ pub(crate) unsafe fn llvm_optimize( let unroll_loops; let vectorize_slp; let vectorize_loop; + let run_enzyme = cfg!(llvm_enzyme) && autodiff_stage == AutodiffStage::DuringAD; // When we build rustc with enzyme/autodiff support, we want to postpone size-increasing - // optimizations until after differentiation. FIXME(ZuseZ4): Before shipping on nightly, + // optimizations until after differentiation. Our pipeline is thus: (opt + enzyme), (full opt). + // We therefore have two calls to llvm_optimize, if autodiff is used. + // + // 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) { + if cfg!(llvm_enzyme) && autodiff_stage != AutodiffStage::PostAD { unroll_loops = false; vectorize_slp = false; vectorize_loop = false; @@ -565,7 +586,7 @@ 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 +654,7 @@ 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 +683,6 @@ 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 +691,7 @@ 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 +704,14 @@ 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). + // FIXME(ZuseZ4): Make this more granular, only set PreAD if we actually have autodiff + // usages, not just if we build rustc with autodiff support. + let autodiff_stage = + if cfg!(llvm_enzyme) { 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 +760,6 @@ 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 +878,17 @@ 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 +912,17 @@ 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 +1049,18 @@ 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 +1100,29 @@ 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. diff --git a/compiler/rustc_codegen_llvm/src/builder.rs b/compiler/rustc_codegen_llvm/src/builder.rs index 264d43c6d46..acae0b444c0 100644 --- a/compiler/rustc_codegen_llvm/src/builder.rs +++ b/compiler/rustc_codegen_llvm/src/builder.rs @@ -543,7 +543,8 @@ impl<'a, 'll, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> { unsafe { let alloca = llvm::LLVMBuildAlloca(bx.llbuilder, ty, UNNAMED); llvm::LLVMSetAlignment(alloca, align.bytes() as c_uint); - alloca + // Cast to default addrspace if necessary + llvm::LLVMBuildPointerCast(bx.llbuilder, alloca, self.cx().type_ptr(), UNNAMED) } } @@ -552,7 +553,8 @@ impl<'a, 'll, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> { let alloca = llvm::LLVMBuildArrayAlloca(self.llbuilder, self.cx().type_i8(), size, UNNAMED); llvm::LLVMSetAlignment(alloca, align.bytes() as c_uint); - alloca + // Cast to default addrspace if necessary + llvm::LLVMBuildPointerCast(self.llbuilder, alloca, self.cx().type_ptr(), UNNAMED) } } diff --git a/compiler/rustc_codegen_llvm/src/builder/autodiff.rs b/compiler/rustc_codegen_llvm/src/builder/autodiff.rs index dd5e726160d..b2c1088e3fc 100644 --- a/compiler/rustc_codegen_llvm/src/builder/autodiff.rs +++ b/compiler/rustc_codegen_llvm/src/builder/autodiff.rs @@ -4,10 +4,9 @@ use rustc_ast::expand::autodiff_attrs::{AutoDiffAttrs, AutoDiffItem, DiffActivit use rustc_codegen_ssa::ModuleCodegen; use rustc_codegen_ssa::back::write::ModuleConfig; use rustc_errors::FatalError; -use rustc_session::config::Lto; use tracing::{debug, trace}; -use crate::back::write::{llvm_err, llvm_optimize}; +use crate::back::write::llvm_err; use crate::builder::SBuilder; use crate::context::SimpleCx; use crate::declare::declare_simple_fn; @@ -53,8 +52,6 @@ fn generate_enzyme_call<'ll>( 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(); @@ -153,7 +150,7 @@ fn generate_enzyme_call<'ll>( _ => {} } - trace!("matching autodiff arguments"); + debug!("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 @@ -164,10 +161,10 @@ fn generate_enzyme_call<'ll>( 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]; + let diff_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 { + let (activity, duplicated): (&Metadata, bool) = match diff_activity { DiffActivity::None => panic!("not a valid input activity"), DiffActivity::Const => (enzyme_const, false), DiffActivity::Active => (enzyme_out, false), @@ -222,7 +219,15 @@ fn generate_enzyme_call<'ll>( // 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); + if matches!( + diff_activity, + DiffActivity::Duplicated | DiffActivity::DuplicatedOnly + ) { + assert!( + llvm::LLVMRustGetTypeKind(next_outer_ty) == llvm::TypeKind::Pointer + ); + } + // In the case of Dual we don't have assumptions, e.g. f32 would be valid. args.push(next_outer_arg); outer_pos += 2; activity_pos += 1; @@ -277,7 +282,7 @@ pub(crate) fn differentiate<'ll>( module: &'ll ModuleCodegen<ModuleLlvm>, cgcx: &CodegenContext<LlvmCodegenBackend>, diff_items: Vec<AutoDiffItem>, - config: &ModuleConfig, + _config: &ModuleConfig, ) -> Result<(), FatalError> { for item in &diff_items { trace!("{}", item); @@ -318,29 +323,6 @@ pub(crate) fn differentiate<'ll>( // 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/common.rs b/compiler/rustc_codegen_llvm/src/common.rs index 8c94a46ebf3..78b3a7f8541 100644 --- a/compiler/rustc_codegen_llvm/src/common.rs +++ b/compiler/rustc_codegen_llvm/src/common.rs @@ -219,8 +219,8 @@ impl<'ll, 'tcx> ConstCodegenMethods<'tcx> for CodegenCx<'ll, 'tcx> { let g = self.define_global(&sym, self.val_ty(sc)).unwrap_or_else(|| { bug!("symbol `{}` is already defined", sym); }); + llvm::set_initializer(g, sc); unsafe { - llvm::LLVMSetInitializer(g, sc); llvm::LLVMSetGlobalConstant(g, True); llvm::LLVMSetUnnamedAddress(g, llvm::UnnamedAddr::Global); } diff --git a/compiler/rustc_codegen_llvm/src/consts.rs b/compiler/rustc_codegen_llvm/src/consts.rs index c6855dd42e5..4a5491ec7a1 100644 --- a/compiler/rustc_codegen_llvm/src/consts.rs +++ b/compiler/rustc_codegen_llvm/src/consts.rs @@ -191,7 +191,7 @@ fn check_and_apply_linkage<'ll, 'tcx>( }) }); llvm::set_linkage(g2, llvm::Linkage::InternalLinkage); - unsafe { llvm::LLVMSetInitializer(g2, g1) }; + llvm::set_initializer(g2, g1); g2 } else if cx.tcx.sess.target.arch == "x86" && common::is_mingw_gnu_toolchain(&cx.tcx.sess.target) @@ -235,7 +235,7 @@ impl<'ll> CodegenCx<'ll, '_> { } _ => self.define_private_global(self.val_ty(cv)), }; - unsafe { llvm::LLVMSetInitializer(gv, cv) }; + llvm::set_initializer(gv, cv); set_global_alignment(self, gv, align); llvm::SetUnnamedAddress(gv, llvm::UnnamedAddr::Global); gv @@ -458,7 +458,7 @@ impl<'ll> CodegenCx<'ll, '_> { new_g }; set_global_alignment(self, g, alloc.align); - llvm::LLVMSetInitializer(g, v); + llvm::set_initializer(g, v); if self.should_assume_dso_local(g, true) { llvm::LLVMRustSetDSOLocal(g, true); diff --git a/compiler/rustc_codegen_llvm/src/context.rs b/compiler/rustc_codegen_llvm/src/context.rs index ba4fd75fb94..7fe527a4c07 100644 --- a/compiler/rustc_codegen_llvm/src/context.rs +++ b/compiler/rustc_codegen_llvm/src/context.rs @@ -616,12 +616,10 @@ impl<'ll, 'tcx> CodegenCx<'ll, 'tcx> { pub(crate) fn create_used_variable_impl(&self, name: &'static CStr, values: &[&'ll Value]) { let array = self.const_array(self.type_ptr(), values); - unsafe { - let g = llvm::LLVMAddGlobal(self.llmod, self.val_ty(array), name.as_ptr()); - llvm::LLVMSetInitializer(g, array); - llvm::set_linkage(g, llvm::Linkage::AppendingLinkage); - llvm::set_section(g, c"llvm.metadata"); - } + let g = llvm::add_global(self.llmod, self.val_ty(array), name); + llvm::set_initializer(g, array); + llvm::set_linkage(g, llvm::Linkage::AppendingLinkage); + llvm::set_section(g, c"llvm.metadata"); } } impl<'ll> SimpleCx<'ll> { diff --git a/compiler/rustc_codegen_llvm/src/debuginfo/gdb.rs b/compiler/rustc_codegen_llvm/src/debuginfo/gdb.rs index 2c9f1cda13a..54c5d445f66 100644 --- a/compiler/rustc_codegen_llvm/src/debuginfo/gdb.rs +++ b/compiler/rustc_codegen_llvm/src/debuginfo/gdb.rs @@ -73,7 +73,7 @@ pub(crate) fn get_or_insert_gdb_debug_scripts_section_global<'ll>( .define_global(section_var_name, llvm_type) .unwrap_or_else(|| bug!("symbol `{}` is already defined", section_var_name)); llvm::set_section(section_var, c".debug_gdb_scripts"); - llvm::LLVMSetInitializer(section_var, cx.const_bytes(section_contents)); + llvm::set_initializer(section_var, cx.const_bytes(section_contents)); llvm::LLVMSetGlobalConstant(section_var, llvm::True); llvm::LLVMSetUnnamedAddress(section_var, llvm::UnnamedAddr::Global); llvm::set_linkage(section_var, llvm::Linkage::LinkOnceODRLinkage); diff --git a/compiler/rustc_codegen_llvm/src/declare.rs b/compiler/rustc_codegen_llvm/src/declare.rs index bdc83267cca..cebceef1b93 100644 --- a/compiler/rustc_codegen_llvm/src/declare.rs +++ b/compiler/rustc_codegen_llvm/src/declare.rs @@ -235,7 +235,7 @@ impl<'ll, 'tcx> CodegenCx<'ll, 'tcx> { /// name. pub(crate) fn get_defined_value(&self, name: &str) -> Option<&'ll Value> { self.get_declared_value(name).and_then(|val| { - let declaration = unsafe { llvm::LLVMIsDeclaration(val) != 0 }; + let declaration = llvm::is_declaration(val); if !declaration { Some(val) } else { None } }) } diff --git a/compiler/rustc_codegen_llvm/src/intrinsic.rs b/compiler/rustc_codegen_llvm/src/intrinsic.rs index 3200c94d977..8b976885904 100644 --- a/compiler/rustc_codegen_llvm/src/intrinsic.rs +++ b/compiler/rustc_codegen_llvm/src/intrinsic.rs @@ -824,7 +824,7 @@ fn codegen_msvc_try<'ll>( if bx.cx.tcx.sess.target.supports_comdat() { llvm::SetUniqueComdat(bx.llmod, tydesc); } - unsafe { llvm::LLVMSetInitializer(tydesc, type_info) }; + llvm::set_initializer(tydesc, type_info); // The flag value of 8 indicates that we are catching the exception by // reference instead of by value. We can't use catch by value because diff --git a/compiler/rustc_codegen_llvm/src/lib.rs b/compiler/rustc_codegen_llvm/src/lib.rs index f0d04b2b644..bcb1cc72b56 100644 --- a/compiler/rustc_codegen_llvm/src/lib.rs +++ b/compiler/rustc_codegen_llvm/src/lib.rs @@ -30,7 +30,7 @@ use std::mem::ManuallyDrop; use back::owned_target_machine::OwnedTargetMachine; use back::write::{create_informational_target_machine, create_target_machine}; use errors::{AutoDiffWithoutLTO, ParseTargetMachineConfig}; -pub use llvm_util::target_features_cfg; +pub(crate) 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}; diff --git a/compiler/rustc_codegen_llvm/src/llvm/ffi.rs b/compiler/rustc_codegen_llvm/src/llvm/ffi.rs index 50a40c9c309..441d144ce50 100644 --- a/compiler/rustc_codegen_llvm/src/llvm/ffi.rs +++ b/compiler/rustc_codegen_llvm/src/llvm/ffi.rs @@ -2359,7 +2359,7 @@ unsafe extern "C" { ); pub fn LLVMRustWriteOutputFile<'a>( T: &'a TargetMachine, - PM: &PassManager<'a>, + PM: *mut PassManager<'a>, M: &'a Module, Output: *const c_char, DwoOutput: *const c_char, @@ -2382,6 +2382,7 @@ unsafe extern "C" { LoopVectorize: bool, DisableSimplifyLibCalls: bool, EmitLifetimeMarkers: bool, + RunEnzyme: bool, SanitizerOptions: Option<&SanitizerOptions>, PGOGenPath: *const c_char, PGOUsePath: *const c_char, diff --git a/compiler/rustc_codegen_llvm/src/llvm/mod.rs b/compiler/rustc_codegen_llvm/src/llvm/mod.rs index 707aeba22cc..7becba4ccd4 100644 --- a/compiler/rustc_codegen_llvm/src/llvm/mod.rs +++ b/compiler/rustc_codegen_llvm/src/llvm/mod.rs @@ -241,6 +241,10 @@ pub fn set_linkage(llglobal: &Value, linkage: Linkage) { } } +pub fn is_declaration(llglobal: &Value) -> bool { + unsafe { LLVMIsDeclaration(llglobal) == ffi::True } +} + pub fn get_visibility(llglobal: &Value) -> Visibility { unsafe { LLVMGetVisibility(llglobal) }.to_rust() } diff --git a/compiler/rustc_codegen_llvm/src/llvm_util.rs b/compiler/rustc_codegen_llvm/src/llvm_util.rs index 53611c746a7..b4b5d6a5b19 100644 --- a/compiler/rustc_codegen_llvm/src/llvm_util.rs +++ b/compiler/rustc_codegen_llvm/src/llvm_util.rs @@ -271,6 +271,7 @@ pub(crate) fn to_llvm_features<'a>(sess: &Session, s: &'a str) -> Option<LLVMFea ("aarch64", "fp16") => Some(LLVMFeature::new("fullfp16")), // Filter out features that are not supported by the current LLVM version ("aarch64", "fpmr") if get_version().0 != 18 => None, + ("arm", "fp16") => Some(LLVMFeature::new("fullfp16")), // In LLVM 18, `unaligned-scalar-mem` was merged with `unaligned-vector-mem` into a single // feature called `fast-unaligned-access`. In LLVM 19, it was split back out. ("riscv32" | "riscv64", "unaligned-scalar-mem") if get_version().0 == 18 => { @@ -303,7 +304,7 @@ 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_cfg(sess: &Session, allow_unstable: bool) -> Vec<Symbol> { +pub(crate) fn target_features_cfg(sess: &Session, allow_unstable: bool) -> Vec<Symbol> { let mut features: FxHashSet<Symbol> = Default::default(); // Add base features for the target. diff --git a/compiler/rustc_codegen_llvm/src/type_.rs b/compiler/rustc_codegen_llvm/src/type_.rs index 34cede54aaf..d61ce417562 100644 --- a/compiler/rustc_codegen_llvm/src/type_.rs +++ b/compiler/rustc_codegen_llvm/src/type_.rs @@ -237,11 +237,11 @@ impl<'ll, 'tcx> BaseTypeCodegenMethods<'tcx> for CodegenCx<'ll, 'tcx> { impl Type { /// Creates an integer type with the given number of bits, e.g., i24 - pub fn ix_llcx(llcx: &llvm::Context, num_bits: u64) -> &Type { + pub(crate) fn ix_llcx(llcx: &llvm::Context, num_bits: u64) -> &Type { unsafe { llvm::LLVMIntTypeInContext(llcx, num_bits as c_uint) } } - pub fn ptr_llcx(llcx: &llvm::Context) -> &Type { + pub(crate) fn ptr_llcx(llcx: &llvm::Context) -> &Type { unsafe { llvm::LLVMPointerTypeInContext(llcx, AddressSpace::DATA.0) } } } |