diff options
Diffstat (limited to 'compiler/rustc_codegen_llvm/src')
28 files changed, 854 insertions, 553 deletions
diff --git a/compiler/rustc_codegen_llvm/src/allocator.rs b/compiler/rustc_codegen_llvm/src/allocator.rs index e614115f64b..4a78e694979 100644 --- a/compiler/rustc_codegen_llvm/src/allocator.rs +++ b/compiler/rustc_codegen_llvm/src/allocator.rs @@ -7,6 +7,7 @@ use rustc_codegen_ssa::traits::BaseTypeCodegenMethods as _; use rustc_middle::bug; use rustc_middle::ty::TyCtxt; use rustc_session::config::{DebugInfo, OomStrategy}; +use rustc_symbol_mangling::mangle_internal_symbol; use crate::builder::SBuilder; use crate::declare::declare_simple_fn; @@ -53,8 +54,8 @@ pub(crate) unsafe fn codegen( } }; - let from_name = global_fn_name(method.name); - let to_name = default_fn_name(method.name); + let from_name = mangle_internal_symbol(tcx, &global_fn_name(method.name)); + let to_name = mangle_internal_symbol(tcx, &default_fn_name(method.name)); create_wrapper_function(tcx, &cx, &from_name, &to_name, &args, output, false); } @@ -64,8 +65,8 @@ pub(crate) unsafe fn codegen( create_wrapper_function( tcx, &cx, - "__rust_alloc_error_handler", - alloc_error_handler_name(alloc_error_handler_kind), + &mangle_internal_symbol(tcx, "__rust_alloc_error_handler"), + &mangle_internal_symbol(tcx, alloc_error_handler_name(alloc_error_handler_kind)), &[usize, usize], // size, align None, true, @@ -73,15 +74,15 @@ pub(crate) unsafe fn codegen( unsafe { // __rust_alloc_error_handler_should_panic - let name = OomStrategy::SYMBOL; - let ll_g = cx.declare_global(name, i8); + let name = mangle_internal_symbol(tcx, OomStrategy::SYMBOL); + let ll_g = cx.declare_global(&name, i8); 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::set_initializer(ll_g, llval); - let name = NO_ALLOC_SHIM_IS_UNSTABLE; - let ll_g = cx.declare_global(name, i8); + let name = mangle_internal_symbol(tcx, NO_ALLOC_SHIM_IS_UNSTABLE); + let ll_g = cx.declare_global(&name, i8); llvm::set_visibility(ll_g, llvm::Visibility::from_generic(tcx.sess.default_visibility())); let llval = llvm::LLVMConstInt(i8, 0, False); llvm::set_initializer(ll_g, llval); diff --git a/compiler/rustc_codegen_llvm/src/back/lto.rs b/compiler/rustc_codegen_llvm/src/back/lto.rs index 668795191a2..a8b49e9552c 100644 --- a/compiler/rustc_codegen_llvm/src/back/lto.rs +++ b/compiler/rustc_codegen_llvm/src/back/lto.rs @@ -610,6 +610,8 @@ fn enable_autodiff_settings(ad: &[config::AutoDiff], module: &mut ModuleCodegen< } // We handle this below config::AutoDiff::PrintModAfter => {} + // We handle this below + config::AutoDiff::PrintModFinal => {} // This is required and already checked config::AutoDiff::Enable => {} } @@ -657,14 +659,20 @@ pub(crate) fn run_pass_manager( } if cfg!(llvm_enzyme) && enable_ad { + // This is the post-autodiff IR, mainly used for testing and educational purposes. + if config.autodiff.contains(&config::AutoDiff::PrintModAfter) { + unsafe { llvm::LLVMDumpModule(module.module_llvm.llmod()) }; + } + let opt_stage = llvm::OptStage::FatLTO; let stage = write::AutodiffStage::PostAD; unsafe { write::llvm_optimize(cgcx, dcx, module, None, 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) { + // This is the final IR, so people should be able to inspect the optimized autodiff output, + // for manual inspection. + if config.autodiff.contains(&config::AutoDiff::PrintModFinal) { unsafe { llvm::LLVMDumpModule(module.module_llvm.llmod()) }; } } @@ -728,7 +736,7 @@ impl ThinBuffer { } } - pub unsafe fn from_raw_ptr(ptr: *mut llvm::ThinLTOBuffer) -> ThinBuffer { + pub(crate) unsafe fn from_raw_ptr(ptr: *mut llvm::ThinLTOBuffer) -> ThinBuffer { let mut ptr = NonNull::new(ptr).unwrap(); ThinBuffer(unsafe { ptr.as_mut() }) } diff --git a/compiler/rustc_codegen_llvm/src/back/write.rs b/compiler/rustc_codegen_llvm/src/back/write.rs index bead4c82a81..bf6138142b6 100644 --- a/compiler/rustc_codegen_llvm/src/back/write.rs +++ b/compiler/rustc_codegen_llvm/src/back/write.rs @@ -1024,7 +1024,7 @@ fn create_section_with_flags_asm(section_name: &str, section_flags: &str, data: } pub(crate) fn bitcode_section_name(cgcx: &CodegenContext<LlvmCodegenBackend>) -> &'static CStr { - if cgcx.target_is_like_osx { + if cgcx.target_is_like_darwin { c"__LLVM,__bitcode" } else if cgcx.target_is_like_aix { c".ipa" @@ -1077,7 +1077,7 @@ unsafe fn embed_bitcode( // and COFF we emit the sections using module level inline assembly for that // reason (see issue #90326 for historical background). unsafe { - if cgcx.target_is_like_osx + if cgcx.target_is_like_darwin || cgcx.target_is_like_aix || cgcx.target_arch == "wasm32" || cgcx.target_arch == "wasm64" @@ -1096,7 +1096,7 @@ unsafe fn embed_bitcode( 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 { + let section = if cgcx.target_is_like_darwin { c"__LLVM,__cmdline" } else if cgcx.target_is_like_aix { c".info" diff --git a/compiler/rustc_codegen_llvm/src/builder.rs b/compiler/rustc_codegen_llvm/src/builder.rs index 55d34f5f2ef..297f104d124 100644 --- a/compiler/rustc_codegen_llvm/src/builder.rs +++ b/compiler/rustc_codegen_llvm/src/builder.rs @@ -14,6 +14,7 @@ use rustc_codegen_ssa::mir::place::PlaceRef; use rustc_codegen_ssa::traits::*; use rustc_data_structures::small_c_str::SmallCStr; use rustc_hir::def_id::DefId; +use rustc_middle::bug; use rustc_middle::middle::codegen_fn_attrs::CodegenFnAttrs; use rustc_middle::ty::layout::{ FnAbiError, FnAbiOfHelpers, FnAbiRequest, HasTypingEnv, LayoutError, LayoutOfHelpers, @@ -1074,6 +1075,35 @@ impl<'a, 'll, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> { unsafe { llvm::LLVMBuildFCmp(self.llbuilder, op as c_uint, lhs, rhs, UNNAMED) } } + fn three_way_compare( + &mut self, + ty: Ty<'tcx>, + lhs: Self::Value, + rhs: Self::Value, + ) -> Option<Self::Value> { + // FIXME: See comment on the definition of `three_way_compare`. + if crate::llvm_util::get_version() < (20, 0, 0) { + return None; + } + + let name = match (ty.is_signed(), ty.primitive_size(self.tcx).bits()) { + (true, 8) => "llvm.scmp.i8.i8", + (true, 16) => "llvm.scmp.i8.i16", + (true, 32) => "llvm.scmp.i8.i32", + (true, 64) => "llvm.scmp.i8.i64", + (true, 128) => "llvm.scmp.i8.i128", + + (false, 8) => "llvm.ucmp.i8.i8", + (false, 16) => "llvm.ucmp.i8.i16", + (false, 32) => "llvm.ucmp.i8.i32", + (false, 64) => "llvm.ucmp.i8.i64", + (false, 128) => "llvm.ucmp.i8.i128", + + _ => bug!("three-way compare unsupported for type {ty:?}"), + }; + Some(self.call_intrinsic(name, &[lhs, rhs])) + } + /* Miscellaneous instructions */ fn memcpy( &mut self, diff --git a/compiler/rustc_codegen_llvm/src/builder/autodiff.rs b/compiler/rustc_codegen_llvm/src/builder/autodiff.rs index 71705ecb4d0..7d264ba4d00 100644 --- a/compiler/rustc_codegen_llvm/src/builder/autodiff.rs +++ b/compiler/rustc_codegen_llvm/src/builder/autodiff.rs @@ -3,8 +3,10 @@ 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 as _; +use rustc_codegen_ssa::common::TypeKind; +use rustc_codegen_ssa::traits::BaseTypeCodegenMethods; use rustc_errors::FatalError; +use rustc_middle::bug; use tracing::{debug, trace}; use crate::back::write::llvm_err; @@ -18,14 +20,221 @@ use crate::value::Value; use crate::{CodegenContext, LlvmCodegenBackend, ModuleLlvm, attributes, llvm}; fn get_params(fnc: &Value) -> Vec<&Value> { + let param_num = llvm::LLVMCountParams(fnc) as usize; + let mut fnc_args: Vec<&Value> = vec![]; + fnc_args.reserve(param_num); 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 } + fnc_args +} + +fn has_sret(fnc: &Value) -> bool { + let num_args = llvm::LLVMCountParams(fnc) as usize; + if num_args == 0 { + false + } else { + unsafe { llvm::LLVMRustHasAttributeAtIndex(fnc, 0, llvm::AttributeKind::StructRet) } + } +} + +// When we call the `__enzyme_autodiff` or `__enzyme_fwddiff` function, we need to pass all the +// original inputs, as well as metadata and the additional shadow arguments. +// This function matches the arguments from the outer function to the inner enzyme call. +// +// This function also considers 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()? +fn match_args_from_caller_to_enzyme<'ll>( + cx: &SimpleCx<'ll>, + width: u32, + args: &mut Vec<&'ll llvm::Value>, + inputs: &[DiffActivity], + outer_args: &[&'ll llvm::Value], + has_sret: bool, +) { + 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 + // 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; + + if has_sret { + // Then the first outer arg is the sret pointer. Enzyme doesn't know about sret, so the + // inner function will still return something. We increase our outer_pos by one, + // and once we're done with all other args we will take the return of the inner call and + // update the sret pointer with it + outer_pos = 1; + } + + 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(); + + while activity_pos < inputs.len() { + 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 diff_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_eq!(cx.type_kind(next_outer_ty), TypeKind::Integer); + + for i in 0..(width as usize) { + let next_outer_arg2 = outer_args[outer_pos + 2 * (i + 1)]; + let next_outer_ty2 = cx.val_ty(next_outer_arg2); + assert_eq!(cx.type_kind(next_outer_ty2), TypeKind::Pointer); + let next_outer_arg3 = outer_args[outer_pos + 2 * (i + 1) + 1]; + let next_outer_ty3 = cx.val_ty(next_outer_arg3); + assert_eq!(cx.type_kind(next_outer_ty3), TypeKind::Integer); + args.push(next_outer_arg2); + } + args.push(cx.get_metadata_value(enzyme_const)); + args.push(next_outer_arg); + outer_pos += 2 + 2 * width as usize; + 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, ...). + if matches!(diff_activity, DiffActivity::Duplicated | DiffActivity::DuplicatedOnly) + { + assert_eq!(cx.type_kind(next_outer_ty), 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; + + // Now, if width > 1, we need to account for that + for _ in 1..width { + let next_outer_arg = outer_args[outer_pos]; + args.push(next_outer_arg); + outer_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; + } + } +} + +// 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. This is not always trivial, e.g. because sret. +// Beyond sret, this article describes our challenges nicely: +// <https://yorickpeterse.com/articles/the-mess-that-is-handling-structure-arguments-and-returns-in-llvm/> +// I.e. (i32, f32) will get merged into i64, but we don't handle that yet. +fn compute_enzyme_fn_ty<'ll>( + cx: &SimpleCx<'ll>, + attrs: &AutoDiffAttrs, + fn_to_diff: &'ll Value, + outer_fn: &'ll Value, +) -> &'ll llvm::Type { + let fn_ty = cx.get_type_of_global(outer_fn); + let mut ret_ty = cx.get_return_type(fn_ty); + + let has_sret = has_sret(outer_fn); + + if has_sret { + // Now we don't just forward the return type, so we have to figure it out based on the + // primal return type, in combination with the autodiff settings. + let fn_ty = cx.get_type_of_global(fn_to_diff); + let inner_ret_ty = cx.get_return_type(fn_ty); + + let void_ty = unsafe { llvm::LLVMVoidTypeInContext(cx.llcx) }; + if inner_ret_ty == void_ty { + // This indicates that even the inner function has an sret. + // Right now I only look for an sret in the outer function. + // This *probably* needs some extra handling, but I never ran + // into such a case. So I'll wait for user reports to have a test case. + bug!("sret in inner function"); + } + + if attrs.width == 1 { + todo!("Handle sret for scalar ad"); + } else { + // First we check if we also have to deal with the primal return. + match attrs.mode { + DiffMode::Forward => match attrs.ret_activity { + DiffActivity::Dual => { + let arr_ty = + unsafe { llvm::LLVMArrayType2(inner_ret_ty, attrs.width as u64 + 1) }; + ret_ty = arr_ty; + } + DiffActivity::DualOnly => { + let arr_ty = + unsafe { llvm::LLVMArrayType2(inner_ret_ty, attrs.width as u64) }; + ret_ty = arr_ty; + } + DiffActivity::Const => { + todo!("Not sure, do we need to do something here?"); + } + _ => { + bug!("unreachable"); + } + }, + DiffMode::Reverse => { + todo!("Handle sret for reverse mode"); + } + _ => { + bug!("unreachable"); + } + } + } + } + + // LLVM can figure out the input types on it's own, so we take a shortcut here. + unsafe { llvm::LLVMFunctionType(ret_ty, ptr::null(), 0, True) } } /// When differentiating `fn_to_diff`, take a `outer_fn` and generate another @@ -43,9 +252,6 @@ fn generate_enzyme_call<'ll>( 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. let mut ad_name: String = match attrs.mode { DiffMode::Forward => "__enzyme_fwddiff", @@ -93,17 +299,9 @@ fn generate_enzyme_call<'ll>( // } // ``` 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); + let enzyme_ty = compute_enzyme_fn_ty(cx, &attrs, fn_to_diff, outer_fn); - // 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 + // 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_simple_fn( @@ -132,116 +330,31 @@ fn generate_enzyme_call<'ll>( 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)); - } - _ => {} + if matches!(attrs.ret_activity, DiffActivity::Dual | DiffActivity::Active) { + args.push(cx.get_metadata_value(enzyme_primal_ret)); + } + if attrs.width > 1 { + let enzyme_width = cx.create_metadata("enzyme_width".to_string()).unwrap(); + args.push(cx.get_metadata_value(enzyme_width)); + args.push(cx.get_const_i64(attrs.width as u64)); } - 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 - // 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 has_sret = has_sret(outer_fn); let outer_args: Vec<&llvm::Value> = get_params(outer_fn); - while activity_pos < inputs.len() { - 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 diff_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, ...). - 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; - } - } 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; - } - } + match_args_from_caller_to_enzyme( + &cx, + attrs.width, + &mut args, + &attrs.input_activity, + &outer_args, + has_sret, + ); let call = builder.call(enzyme_ty, ad_fn, &args, 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 + // metadata attached 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. @@ -262,7 +375,22 @@ fn generate_enzyme_call<'ll>( // Now that we copied the metadata, get rid of dummy code. llvm::LLVMRustEraseInstUntilInclusive(entry, last_inst); - if cx.val_ty(call) == cx.type_void() { + if cx.val_ty(call) == cx.type_void() || has_sret { + if has_sret { + // This is what we already have in our outer_fn (shortened): + // define void @_foo(ptr <..> sret([32 x i8]) initializes((0, 32)) %0, <...>) { + // %7 = call [4 x double] (...) @__enzyme_fwddiff_foo(ptr @square, metadata !"enzyme_width", i64 4, <...>) + // <Here we are, we want to add the following two lines> + // store [4 x double] %7, ptr %0, align 8 + // ret void + // } + + // now store the result of the enzyme call into the sret pointer. + let sret_ptr = outer_args[0]; + let call_ty = cx.val_ty(call); + assert_eq!(cx.type_kind(call_ty), TypeKind::Array); + llvm::LLVMBuildStore(&builder.llbuilder, call, sret_ptr); + } builder.ret_void(); } else { builder.ret(call); @@ -294,8 +422,7 @@ pub(crate) fn differentiate<'ll>( if !diff_items.is_empty() && !cgcx.opts.unstable_opts.autodiff.contains(&rustc_session::config::AutoDiff::Enable) { - let dcx = cgcx.create_dcx(); - return Err(dcx.handle().emit_almost_fatal(AutoDiffWithoutEnable)); + return Err(diag_handler.handle().emit_almost_fatal(AutoDiffWithoutEnable)); } // Before dumping the module, we want all the TypeTrees to become part of the module. diff --git a/compiler/rustc_codegen_llvm/src/consts.rs b/compiler/rustc_codegen_llvm/src/consts.rs index a4e5749b3ac..bf81eb648f8 100644 --- a/compiler/rustc_codegen_llvm/src/consts.rs +++ b/compiler/rustc_codegen_llvm/src/consts.rs @@ -5,6 +5,7 @@ use rustc_abi::{ }; use rustc_codegen_ssa::common; use rustc_codegen_ssa::traits::*; +use rustc_hir::LangItem; use rustc_hir::def::DefKind; use rustc_hir::def_id::DefId; use rustc_middle::middle::codegen_fn_attrs::{CodegenFnAttrFlags, CodegenFnAttrs}; @@ -12,9 +13,9 @@ use rustc_middle::mir::interpret::{ Allocation, ConstAllocation, ErrorHandled, InitChunk, Pointer, Scalar as InterpScalar, read_target_uint, }; -use rustc_middle::mir::mono::MonoItem; -use rustc_middle::ty::Instance; +use rustc_middle::mir::mono::{Linkage, MonoItem}; use rustc_middle::ty::layout::{HasTypingEnv, LayoutOf}; +use rustc_middle::ty::{self, Instance}; use rustc_middle::{bug, span_bug}; use tracing::{debug, instrument, trace}; @@ -128,7 +129,12 @@ pub(crate) fn const_alloc_to_llvm<'ll>( append_chunks_of_init_and_uninit_bytes(&mut llvals, cx, alloc, range); } - cx.const_struct(&llvals, true) + // Avoid wrapping in a struct if there is only a single value. This ensures + // that LLVM is able to perform the string merging optimization if the constant + // is a valid C string. LLVM only considers bare arrays for this optimization, + // not arrays wrapped in a struct. LLVM handles this at: + // https://github.com/rust-lang/llvm-project/blob/acaea3d2bb8f351b740db7ebce7d7a40b9e21488/llvm/lib/Target/TargetLoweringObjectFile.cpp#L249-L280 + if let &[data] = &*llvals { data } else { cx.const_struct(&llvals, true) } } fn codegen_static_initializer<'ll, 'tcx>( @@ -171,8 +177,27 @@ fn check_and_apply_linkage<'ll, 'tcx>( if let Some(linkage) = attrs.import_linkage { debug!("get_static: sym={} linkage={:?}", sym, linkage); - // Declare a symbol `foo` with the desired linkage. - let g1 = cx.declare_global(sym, cx.type_i8()); + // Declare a symbol `foo`. If `foo` is an extern_weak symbol, we declare + // an extern_weak function, otherwise a global with the desired linkage. + let g1 = if matches!(attrs.import_linkage, Some(Linkage::ExternalWeak)) { + // An `extern_weak` function is represented as an `Option<unsafe extern ...>`, + // we extract the function signature and declare it as an extern_weak function + // instead of an extern_weak i8. + let instance = Instance::mono(cx.tcx, def_id); + if let ty::Adt(struct_def, args) = instance.ty(cx.tcx, cx.typing_env()).kind() + && cx.tcx.is_lang_item(struct_def.did(), LangItem::Option) + && let ty::FnPtr(sig, header) = args.type_at(0).kind() + { + let fn_sig = sig.with(*header); + + let fn_abi = cx.fn_abi_of_fn_ptr(fn_sig, ty::List::empty()); + cx.declare_fn(sym, &fn_abi, None) + } else { + cx.declare_global(sym, cx.type_i8()) + } + } else { + cx.declare_global(sym, cx.type_i8()) + }; llvm::set_linkage(g1, base::linkage_to_llvm(linkage)); // Declare an internal global `extern_with_linkage_foo` which @@ -405,7 +430,7 @@ impl<'ll> CodegenCx<'ll, '_> { let val_llty = self.val_ty(v); let g = self.get_static_inner(def_id, val_llty); - let llty = llvm::LLVMGlobalGetValueType(g); + let llty = self.get_type_of_global(g); let g = if val_llty == llty { g diff --git a/compiler/rustc_codegen_llvm/src/context.rs b/compiler/rustc_codegen_llvm/src/context.rs index 9f8ec0ea526..1e8861116e2 100644 --- a/compiler/rustc_codegen_llvm/src/context.rs +++ b/compiler/rustc_codegen_llvm/src/context.rs @@ -8,6 +8,7 @@ use std::str; use rustc_abi::{HasDataLayout, Size, TargetDataLayout, VariantIdx}; use rustc_codegen_ssa::back::versioned_llvm_target; use rustc_codegen_ssa::base::{wants_msvc_seh, wants_wasm_eh}; +use rustc_codegen_ssa::common::TypeKind; use rustc_codegen_ssa::errors as ssa_errors; use rustc_codegen_ssa::traits::*; use rustc_data_structures::base_n::{ALPHANUMERIC_ONLY, ToBaseN}; @@ -27,6 +28,7 @@ use rustc_session::config::{ }; use rustc_span::source_map::Spanned; use rustc_span::{DUMMY_SP, Span}; +use rustc_symbol_mangling::mangle_internal_symbol; use rustc_target::spec::{HasTargetSpec, RelocModel, SmallDataThresholdSupport, Target, TlsModel}; use smallvec::SmallVec; @@ -37,7 +39,7 @@ use crate::debuginfo::metadata::apply_vcall_visibility_metadata; use crate::llvm::Metadata; use crate::type_::Type; use crate::value::Value; -use crate::{attributes, coverageinfo, debuginfo, llvm, llvm_util}; +use crate::{attributes, common, coverageinfo, debuginfo, llvm, llvm_util}; /// `TyCtxt` (and related cache datastructures) can't be move between threads. /// However, there are various cx related functions which we want to be available to the builder and @@ -326,6 +328,22 @@ pub(crate) unsafe fn create_module<'ll>( pfe.prefix().into(), ); } + + // Add "kcfi-arity" module flag if KCFI arity indicator is enabled. (See + // https://github.com/llvm/llvm-project/pull/117121.) + if sess.is_sanitizer_kcfi_arity_enabled() { + // KCFI arity indicator requires LLVM 21.0.0 or later. + if llvm_version < (21, 0, 0) { + tcx.dcx().emit_err(crate::errors::SanitizerKcfiArityRequiresLLVM2100); + } + + llvm::add_module_flag_u32( + llmod, + llvm::ModuleFlagMergeBehavior::Override, + "kcfi-arity", + 1, + ); + } } // Control Flow Guard is currently only supported by MSVC and LLVM on Windows. @@ -642,7 +660,18 @@ impl<'ll, 'tcx> CodegenCx<'ll, 'tcx> { llvm::set_section(g, c"llvm.metadata"); } } - +impl<'ll> SimpleCx<'ll> { + pub(crate) fn get_return_type(&self, ty: &'ll Type) -> &'ll Type { + assert_eq!(self.type_kind(ty), TypeKind::Function); + unsafe { llvm::LLVMGetReturnType(ty) } + } + pub(crate) fn get_type_of_global(&self, val: &'ll Value) -> &'ll Type { + unsafe { llvm::LLVMGlobalGetValueType(val) } + } + pub(crate) fn val_ty(&self, v: &'ll Value) -> &'ll Type { + common::val_ty(v) + } +} impl<'ll> SimpleCx<'ll> { pub(crate) fn new( llmod: &'ll llvm::Module, @@ -659,6 +688,13 @@ impl<'ll, CX: Borrow<SCx<'ll>>> GenericCx<'ll, CX> { llvm::LLVMMetadataAsValue(self.llcx(), metadata) } + // FIXME(autodiff): We should split `ConstCodegenMethods` to pull the reusable parts + // onto a trait that is also implemented for GenericCx. + pub(crate) fn get_const_i64(&self, n: u64) -> &'ll Value { + let ty = unsafe { llvm::LLVMInt64TypeInContext(self.llcx()) }; + unsafe { llvm::LLVMConstInt(ty, n, llvm::False) } + } + pub(crate) fn get_function(&self, name: &str) -> Option<&'ll Value> { let name = SmallCStr::new(name); unsafe { llvm::LLVMGetNamedFunction((**self).borrow().llmod, name.as_ptr()) } @@ -1126,6 +1162,18 @@ impl<'ll> CodegenCx<'ll, '_> { ifn!("llvm.usub.sat.i64", fn(t_i64, t_i64) -> t_i64); ifn!("llvm.usub.sat.i128", fn(t_i128, t_i128) -> t_i128); + ifn!("llvm.scmp.i8.i8", fn(t_i8, t_i8) -> t_i8); + ifn!("llvm.scmp.i8.i16", fn(t_i16, t_i16) -> t_i8); + ifn!("llvm.scmp.i8.i32", fn(t_i32, t_i32) -> t_i8); + ifn!("llvm.scmp.i8.i64", fn(t_i64, t_i64) -> t_i8); + ifn!("llvm.scmp.i8.i128", fn(t_i128, t_i128) -> t_i8); + + ifn!("llvm.ucmp.i8.i8", fn(t_i8, t_i8) -> t_i8); + ifn!("llvm.ucmp.i8.i16", fn(t_i16, t_i16) -> t_i8); + ifn!("llvm.ucmp.i8.i32", fn(t_i32, t_i32) -> t_i8); + ifn!("llvm.ucmp.i8.i64", fn(t_i64, t_i64) -> t_i8); + ifn!("llvm.ucmp.i8.i128", fn(t_i128, t_i128) -> t_i8); + ifn!("llvm.lifetime.start.p0i8", fn(t_i64, ptr) -> void); ifn!("llvm.lifetime.end.p0i8", fn(t_i64, ptr) -> void); @@ -1199,7 +1247,7 @@ impl<'ll> CodegenCx<'ll, '_> { Some(def_id) => self.get_static(def_id), _ => { let ty = self.type_struct(&[self.type_ptr(), self.type_ptr()], false); - self.declare_global("rust_eh_catch_typeinfo", ty) + self.declare_global(&mangle_internal_symbol(self.tcx, "rust_eh_catch_typeinfo"), ty) } }; self.eh_catch_typeinfo.set(Some(eh_catch_typeinfo)); diff --git a/compiler/rustc_codegen_llvm/src/coverageinfo/ffi.rs b/compiler/rustc_codegen_llvm/src/coverageinfo/ffi.rs index b617f4d37f5..f6000e72840 100644 --- a/compiler/rustc_codegen_llvm/src/coverageinfo/ffi.rs +++ b/compiler/rustc_codegen_llvm/src/coverageinfo/ffi.rs @@ -146,6 +146,7 @@ pub(crate) struct CoverageSpan { #[derive(Clone, Debug, Default)] pub(crate) struct Regions { pub(crate) code_regions: Vec<CodeRegion>, + pub(crate) expansion_regions: Vec<ExpansionRegion>, pub(crate) branch_regions: Vec<BranchRegion>, pub(crate) mcdc_branch_regions: Vec<MCDCBranchRegion>, pub(crate) mcdc_decision_regions: Vec<MCDCDecisionRegion>, @@ -154,10 +155,16 @@ pub(crate) struct Regions { 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; + let Self { + code_regions, + expansion_regions, + branch_regions, + mcdc_branch_regions, + mcdc_decision_regions, + } = self; code_regions.is_empty() + && expansion_regions.is_empty() && branch_regions.is_empty() && mcdc_branch_regions.is_empty() && mcdc_decision_regions.is_empty() @@ -172,6 +179,14 @@ pub(crate) struct CodeRegion { pub(crate) counter: Counter, } +/// Must match the layout of `LLVMRustCoverageExpansionRegion`. +#[derive(Clone, Debug)] +#[repr(C)] +pub(crate) struct ExpansionRegion { + pub(crate) cov_span: CoverageSpan, + pub(crate) expanded_file_id: u32, +} + /// Must match the layout of `LLVMRustCoverageBranchRegion`. #[derive(Clone, Debug)] #[repr(C)] diff --git a/compiler/rustc_codegen_llvm/src/coverageinfo/llvm_cov.rs b/compiler/rustc_codegen_llvm/src/coverageinfo/llvm_cov.rs index 2cd7fa3225a..907d6d41a1f 100644 --- a/compiler/rustc_codegen_llvm/src/coverageinfo/llvm_cov.rs +++ b/compiler/rustc_codegen_llvm/src/coverageinfo/llvm_cov.rs @@ -63,8 +63,18 @@ pub(crate) fn write_function_mappings_to_buffer( expressions: &[ffi::CounterExpression], regions: &ffi::Regions, ) -> Vec<u8> { - let ffi::Regions { code_regions, branch_regions, mcdc_branch_regions, mcdc_decision_regions } = - regions; + let ffi::Regions { + code_regions, + expansion_regions, + branch_regions, + mcdc_branch_regions, + mcdc_decision_regions, + } = regions; + + // SAFETY: + // - All types are FFI-compatible and have matching representations in Rust/C++. + // - For pointer/length pairs, the pointer and length come from the same vector or slice. + // - C++ code does not retain any pointers after the call returns. llvm::build_byte_buffer(|buffer| unsafe { llvm::LLVMRustCoverageWriteFunctionMappingsToBuffer( virtual_file_mapping.as_ptr(), @@ -73,6 +83,8 @@ pub(crate) fn write_function_mappings_to_buffer( expressions.len(), code_regions.as_ptr(), code_regions.len(), + expansion_regions.as_ptr(), + expansion_regions.len(), branch_regions.as_ptr(), branch_regions.len(), mcdc_branch_regions.as_ptr(), diff --git a/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen/covfun.rs b/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen/covfun.rs index 80e54bf045e..048e1988c32 100644 --- a/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen/covfun.rs +++ b/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen/covfun.rs @@ -105,9 +105,14 @@ fn fill_region_tables<'tcx>( ids_info: &'tcx CoverageIdsInfo, covfun: &mut CovfunRecord<'tcx>, ) { - // Currently a function's mappings must all be in the same file as its body span. + // Currently a function's mappings must all be in the same file, so use the + // first mapping's span to determine the file. let source_map = tcx.sess.source_map(); - let source_file = source_map.lookup_source_file(fn_cov_info.body_span.lo()); + let Some(first_span) = (try { fn_cov_info.mappings.first()?.span }) else { + debug_assert!(false, "function has no mappings: {:?}", covfun.mangled_function_name); + return; + }; + let source_file = source_map.lookup_source_file(first_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); @@ -115,13 +120,22 @@ fn fill_region_tables<'tcx>( // 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) - }; + // In rare cases, _all_ of a function's spans are discarded, and coverage + // codegen needs to handle that gracefully to avoid #133606. + // It's hard for tests to trigger this organically, so instead we set + // `-Zcoverage-options=discard-all-spans-in-codegen` to force it to occur. let discard_all = tcx.sess.coverage_discard_all_spans_in_codegen(); + let make_coords = |span: Span| { + if discard_all { None } else { spans::make_coords(source_map, &source_file, span) } + }; + + let ffi::Regions { + code_regions, + expansion_regions: _, // FIXME(Zalathar): Fill out support for expansion regions + branch_regions, + mcdc_branch_regions, + mcdc_decision_regions, + } = &mut covfun.regions; // For each counter/region pair in this function+file, convert it to a // form suitable for FFI. @@ -136,17 +150,8 @@ fn fill_region_tables<'tcx>( ffi::Counter::from_term(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; - } + let Some(coords) = make_coords(span) else { continue }; + let cov_span = coords.make_coverage_span(local_file_id); match *kind { MappingKind::Code { bcb } => { diff --git a/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen/spans.rs b/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen/spans.rs index 6d1d91340c2..39a59560c9d 100644 --- a/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen/spans.rs +++ b/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen/spans.rs @@ -1,4 +1,3 @@ -use rustc_middle::mir::coverage::FunctionCoverageInfo; use rustc_span::source_map::SourceMap; use rustc_span::{BytePos, Pos, SourceFile, Span}; use tracing::debug; @@ -6,24 +5,41 @@ use tracing::debug; use crate::coverageinfo::ffi; use crate::coverageinfo::mapgen::LocalFileId; +/// Line and byte-column coordinates of a source code span within some file. +/// The file itself must be tracked separately. +#[derive(Clone, Copy, Debug)] +pub(crate) struct Coords { + /// 1-based starting line of the source code span. + pub(crate) start_line: u32, + /// 1-based starting column (in bytes) of the source code span. + pub(crate) start_col: u32, + /// 1-based ending line of the source code span. + pub(crate) end_line: u32, + /// 1-based ending column (in bytes) of the source code span. High bit must be unset. + pub(crate) end_col: u32, +} + +impl Coords { + /// Attaches a local file ID to these coordinates to produce an `ffi::CoverageSpan`. + pub(crate) fn make_coverage_span(&self, local_file_id: LocalFileId) -> ffi::CoverageSpan { + let &Self { start_line, start_col, end_line, end_col } = self; + let file_id = local_file_id.as_u32(); + ffi::CoverageSpan { file_id, start_line, start_col, end_line, end_col } + } +} + /// 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, +/// Returns `None` if the conversion failed for some reason. This should be uncommon, /// 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)?; +pub(crate) fn make_coords(source_map: &SourceMap, file: &SourceFile, span: Span) -> Option<Coords> { + let span = ensure_non_empty_span(source_map, span)?; let lo = span.lo(); let hi = span.hi(); @@ -46,8 +62,7 @@ pub(crate) fn make_coverage_span( 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(), + check_coords(Coords { start_line: start_line as u32, start_col: start_col as u32, end_line: end_line as u32, @@ -55,36 +70,22 @@ pub(crate) fn make_coverage_span( }) } -fn ensure_non_empty_span( - source_map: &SourceMap, - fn_cov_info: &FunctionCoverageInfo, - span: Span, -) -> Option<Span> { +fn ensure_non_empty_span(source_map: &SourceMap, 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; - } - + // The span is empty, so try to enlarge it to cover an adjacent '{' or '}'. 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))) + if src.as_bytes().get(end).copied() == Some(b'{') { + Some(span.with_hi(span.hi() + BytePos(1))) + } else if start > 0 && src.as_bytes()[start - 1] == b'}' { + Some(span.with_lo(span.lo() - BytePos(1))) } else { None } @@ -96,8 +97,8 @@ fn ensure_non_empty_span( /// 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; +fn check_coords(coords: Coords) -> Option<Coords> { + let Coords { start_line, start_col, end_line, end_col } = coords; // Line/column coordinates are supposed to be 1-based. If we ever emit // coordinates of 0, `llvm-cov` might misinterpret them. @@ -110,17 +111,17 @@ fn check_coverage_span(cov_span: ffi::CoverageSpan) -> Option<ffi::CoverageSpan> 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) + Some(coords) } else { debug!( - ?cov_span, + ?coords, ?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:?}"); + debug_assert!(false, "Improper source region: {coords:?}"); None } } diff --git a/compiler/rustc_codegen_llvm/src/debuginfo/metadata.rs b/compiler/rustc_codegen_llvm/src/debuginfo/metadata.rs index 98d59f5a8ae..2eaaf127e41 100644 --- a/compiler/rustc_codegen_llvm/src/debuginfo/metadata.rs +++ b/compiler/rustc_codegen_llvm/src/debuginfo/metadata.rs @@ -2,6 +2,7 @@ use std::borrow::Cow; use std::fmt::{self, Write}; use std::hash::{Hash, Hasher}; use std::path::{Path, PathBuf}; +use std::sync::Arc; use std::{iter, ptr}; use libc::{c_char, c_longlong, c_uint}; @@ -38,8 +39,8 @@ use crate::debuginfo::metadata::type_map::build_type_with_children; use crate::debuginfo::utils::{WidePtrKind, wide_pointer_kind}; use crate::llvm; use crate::llvm::debuginfo::{ - DIDescriptor, DIFile, DIFlags, DILexicalBlock, DIScope, DIType, DebugEmissionKind, - DebugNameTableKind, + DIBasicType, DIBuilder, DICompositeType, DIDescriptor, DIFile, DIFlags, DILexicalBlock, + DIScope, DIType, DebugEmissionKind, DebugNameTableKind, }; use crate::value::Value; @@ -68,7 +69,8 @@ pub(super) const UNKNOWN_COLUMN_NUMBER: c_uint = 0; const NO_SCOPE_METADATA: Option<&DIScope> = None; /// A function that returns an empty list of generic parameter debuginfo nodes. -const NO_GENERICS: for<'ll> fn(&CodegenCx<'ll, '_>) -> SmallVec<&'ll DIType> = |_| SmallVec::new(); +const NO_GENERICS: for<'ll> fn(&CodegenCx<'ll, '_>) -> SmallVec<Option<&'ll DIType>> = + |_| SmallVec::new(); // SmallVec is used quite a bit in this module, so create a shorthand. // The actual number of elements is not so important. @@ -243,7 +245,7 @@ fn build_pointer_or_reference_di_node<'ll, 'tcx>( cx, owner, addr_field_name, - (addr_field.size, addr_field.align.abi), + addr_field, layout.fields.offset(WIDE_PTR_ADDR), DIFlags::FlagZero, data_ptr_type_di_node, @@ -253,7 +255,7 @@ fn build_pointer_or_reference_di_node<'ll, 'tcx>( cx, owner, extra_field_name, - (extra_field.size, extra_field.align.abi), + extra_field, layout.fields.offset(WIDE_PTR_EXTRA), DIFlags::FlagZero, type_di_node(cx, extra_field.ty), @@ -311,12 +313,7 @@ fn build_subroutine_type_di_node<'ll, 'tcx>( debug_context(cx).type_map.unique_id_to_di_node.borrow_mut().remove(&unique_type_id); - let fn_di_node = unsafe { - llvm::LLVMRustDIBuilderCreateSubroutineType( - DIB(cx), - create_DIArray(DIB(cx), &signature_di_nodes[..]), - ) - }; + let fn_di_node = create_subroutine_type(cx, create_DIArray(DIB(cx), &signature_di_nodes[..])); // This is actually a function pointer, so wrap it in pointer DI. let name = compute_debuginfo_type_name(cx.tcx, fn_ty, false); @@ -340,6 +337,13 @@ fn build_subroutine_type_di_node<'ll, 'tcx>( DINodeCreationResult::new(di_node, false) } +pub(super) fn create_subroutine_type<'ll>( + cx: &CodegenCx<'ll, '_>, + signature: &'ll DICompositeType, +) -> &'ll DICompositeType { + unsafe { llvm::LLVMRustDIBuilderCreateSubroutineType(DIB(cx), signature) } +} + /// Create debuginfo for `dyn SomeTrait` types. Currently these are empty structs /// we with the correct type name (e.g. "dyn SomeTrait<Foo, Item=u32> + Sync"). fn build_dyn_type_di_node<'ll, 'tcx>( @@ -487,26 +491,22 @@ pub(crate) fn type_di_node<'ll, 'tcx>(cx: &CodegenCx<'ll, 'tcx>, t: Ty<'tcx>) -> // FIXME(mw): Cache this via a regular UniqueTypeId instead of an extra field in the debug context. fn recursion_marker_type_di_node<'ll, 'tcx>(cx: &CodegenCx<'ll, 'tcx>) -> &'ll DIType { *debug_context(cx).recursion_marker_type.get_or_init(move || { - unsafe { - // The choice of type here is pretty arbitrary - - // anything reading the debuginfo for a recursive - // type is going to see *something* weird - the only - // question is what exactly it will see. - // - // FIXME: the name `<recur_type>` does not fit the naming scheme - // of other types. - // - // FIXME: it might make sense to use an actual pointer type here - // so that debuggers can show the address. - let name = "<recur_type>"; - llvm::LLVMRustDIBuilderCreateBasicType( - DIB(cx), - name.as_c_char_ptr(), - name.len(), - cx.tcx.data_layout.pointer_size.bits(), - dwarf_const::DW_ATE_unsigned, - ) - } + // The choice of type here is pretty arbitrary - + // anything reading the debuginfo for a recursive + // type is going to see *something* weird - the only + // question is what exactly it will see. + // + // FIXME: the name `<recur_type>` does not fit the naming scheme + // of other types. + // + // FIXME: it might make sense to use an actual pointer type here + // so that debuggers can show the address. + create_basic_type( + cx, + "<recur_type>", + cx.tcx.data_layout.pointer_size, + dwarf_const::DW_ATE_unsigned, + ) }) } @@ -620,42 +620,38 @@ pub(crate) fn file_metadata<'ll>(cx: &CodegenCx<'ll, '_>, source_file: &SourceFi let source = cx.sess().opts.unstable_opts.embed_source.then_some(()).and(source_file.src.as_ref()); - unsafe { - llvm::LLVMRustDIBuilderCreateFile( - DIB(cx), - file_name.as_c_char_ptr(), - file_name.len(), - directory.as_c_char_ptr(), - directory.len(), - hash_kind, - hash_value.as_c_char_ptr(), - hash_value.len(), - source.map_or(ptr::null(), |x| x.as_c_char_ptr()), - source.map_or(0, |x| x.len()), - ) - } + create_file(DIB(cx), &file_name, &directory, &hash_value, hash_kind, source) } } fn unknown_file_metadata<'ll>(cx: &CodegenCx<'ll, '_>) -> &'ll DIFile { - debug_context(cx).created_files.borrow_mut().entry(None).or_insert_with(|| unsafe { - let file_name = "<unknown>"; - let directory = ""; - let hash_value = ""; + debug_context(cx).created_files.borrow_mut().entry(None).or_insert_with(|| { + create_file(DIB(cx), "<unknown>", "", "", llvm::ChecksumKind::None, None) + }) +} +fn create_file<'ll>( + builder: &DIBuilder<'ll>, + file_name: &str, + directory: &str, + hash_value: &str, + hash_kind: llvm::ChecksumKind, + source: Option<&Arc<String>>, +) -> &'ll DIFile { + unsafe { llvm::LLVMRustDIBuilderCreateFile( - DIB(cx), + builder, file_name.as_c_char_ptr(), file_name.len(), directory.as_c_char_ptr(), directory.len(), - llvm::ChecksumKind::None, + hash_kind, hash_value.as_c_char_ptr(), hash_value.len(), - ptr::null(), - 0, + source.map_or(ptr::null(), |x| x.as_c_char_ptr()), + source.map_or(0, |x| x.len()), ) - }) + } } trait MsvcBasicName { @@ -742,7 +738,7 @@ fn build_cpp_f16_di_node<'ll, 'tcx>(cx: &CodegenCx<'ll, 'tcx>) -> DINodeCreation cx, float_di_node, "bits", - cx.size_and_align_of(bits_ty), + cx.layout_of(bits_ty), Size::ZERO, DIFlags::FlagZero, type_di_node(cx, bits_ty), @@ -788,15 +784,7 @@ fn build_basic_type_di_node<'ll, 'tcx>( _ => bug!("debuginfo::build_basic_type_di_node - `t` is invalid type"), }; - let ty_di_node = unsafe { - llvm::LLVMRustDIBuilderCreateBasicType( - DIB(cx), - name.as_c_char_ptr(), - name.len(), - cx.size_of(t).bits(), - encoding, - ) - }; + let ty_di_node = create_basic_type(cx, name, cx.size_of(t), encoding); if !cpp_like_debuginfo { return DINodeCreationResult::new(ty_di_node, false); @@ -824,6 +812,23 @@ fn build_basic_type_di_node<'ll, 'tcx>( DINodeCreationResult::new(typedef_di_node, false) } +fn create_basic_type<'ll, 'tcx>( + cx: &CodegenCx<'ll, 'tcx>, + name: &str, + size: Size, + encoding: u32, +) -> &'ll DIBasicType { + unsafe { + llvm::LLVMRustDIBuilderCreateBasicType( + DIB(cx), + name.as_c_char_ptr(), + name.len(), + size.bits(), + encoding, + ) + } +} + fn build_foreign_type_di_node<'ll, 'tcx>( cx: &CodegenCx<'ll, 'tcx>, t: Ty<'tcx>, @@ -929,17 +934,13 @@ pub(crate) fn build_compile_unit_di_node<'ll, 'tcx>( }; unsafe { - let compile_unit_file = llvm::LLVMRustDIBuilderCreateFile( + let compile_unit_file = create_file( debug_context.builder.as_ref(), - name_in_debuginfo.as_c_char_ptr(), - name_in_debuginfo.len(), - work_dir.as_c_char_ptr(), - work_dir.len(), + &name_in_debuginfo, + &work_dir, + "", llvm::ChecksumKind::None, - ptr::null(), - 0, - ptr::null(), - 0, + None, ); let unit_metadata = llvm::LLVMRustDIBuilderCreateCompileUnit( @@ -971,7 +972,7 @@ fn build_field_di_node<'ll, 'tcx>( cx: &CodegenCx<'ll, 'tcx>, owner: &'ll DIScope, name: &str, - size_and_align: (Size, Align), + layout: TyAndLayout<'tcx>, offset: Size, flags: DIFlags, type_di_node: &'ll DIType, @@ -983,6 +984,30 @@ fn build_field_di_node<'ll, 'tcx>( } else { (unknown_file_metadata(cx), UNKNOWN_LINE_NUMBER) }; + create_member_type( + cx, + owner, + name, + file_metadata, + line_number, + layout, + offset, + flags, + type_di_node, + ) +} + +fn create_member_type<'ll, 'tcx>( + cx: &CodegenCx<'ll, 'tcx>, + owner: &'ll DIScope, + name: &str, + file_metadata: &'ll DIType, + line_number: u32, + layout: TyAndLayout<'tcx>, + offset: Size, + flags: DIFlags, + type_di_node: &'ll DIType, +) -> &'ll DIType { unsafe { llvm::LLVMRustDIBuilderCreateMemberType( DIB(cx), @@ -991,8 +1016,8 @@ fn build_field_di_node<'ll, 'tcx>( name.len(), file_metadata, line_number, - size_and_align.0.bits(), - size_and_align.1.bits() as u32, + layout.size.bits(), + layout.align.abi.bits() as u32, offset.bits(), flags, type_di_node, @@ -1076,7 +1101,7 @@ fn build_struct_type_di_node<'ll, 'tcx>( cx, owner, &field_name[..], - (field_layout.size, field_layout.align.abi), + field_layout, struct_type_and_layout.fields.offset(i), visibility_di_flags(cx, f.did, adt_def.did()), type_di_node(cx, field_layout.ty), @@ -1126,7 +1151,7 @@ fn build_upvar_field_di_nodes<'ll, 'tcx>( cx, closure_or_coroutine_di_node, capture_name.as_str(), - cx.size_and_align_of(up_var_ty), + cx.layout_of(up_var_ty), layout.fields.offset(index), DIFlags::FlagZero, type_di_node(cx, up_var_ty), @@ -1171,7 +1196,7 @@ fn build_tuple_type_di_node<'ll, 'tcx>( cx, tuple_di_node, &tuple_field_name(index), - cx.size_and_align_of(component_type), + cx.layout_of(component_type), tuple_type_and_layout.fields.offset(index), DIFlags::FlagZero, type_di_node(cx, component_type), @@ -1269,7 +1294,7 @@ fn build_union_type_di_node<'ll, 'tcx>( cx, owner, f.name.as_str(), - size_and_align_of(field_layout), + field_layout, Size::ZERO, DIFlags::FlagZero, type_di_node(cx, field_layout.ty), @@ -1287,32 +1312,33 @@ fn build_union_type_di_node<'ll, 'tcx>( fn build_generic_type_param_di_nodes<'ll, 'tcx>( cx: &CodegenCx<'ll, 'tcx>, ty: Ty<'tcx>, -) -> SmallVec<&'ll DIType> { +) -> SmallVec<Option<&'ll DIType>> { if let ty::Adt(def, args) = *ty.kind() { - if args.types().next().is_some() { - let generics = cx.tcx.generics_of(def.did()); - let names = get_parameter_names(cx, generics); - let template_params: SmallVec<_> = iter::zip(args, names) - .filter_map(|(kind, name)| { - kind.as_type().map(|ty| { - let actual_type = cx.tcx.normalize_erasing_regions(cx.typing_env(), ty); - let actual_type_di_node = type_di_node(cx, actual_type); - let name = name.as_str(); - unsafe { - llvm::LLVMRustDIBuilderCreateTemplateTypeParameter( - DIB(cx), - None, - name.as_c_char_ptr(), - name.len(), - actual_type_di_node, - ) - } - }) + let generics = cx.tcx.generics_of(def.did()); + return get_template_parameters(cx, generics, args); + } + + return smallvec![]; +} + +pub(super) fn get_template_parameters<'ll, 'tcx>( + cx: &CodegenCx<'ll, 'tcx>, + generics: &ty::Generics, + args: ty::GenericArgsRef<'tcx>, +) -> SmallVec<Option<&'ll DIType>> { + if args.types().next().is_some() { + let names = get_parameter_names(cx, generics); + let template_params: SmallVec<_> = iter::zip(args, names) + .filter_map(|(kind, name)| { + kind.as_type().map(|ty| { + let actual_type = cx.tcx.normalize_erasing_regions(cx.typing_env(), ty); + let actual_type_di_node = type_di_node(cx, actual_type); + Some(cx.create_template_type_parameter(name.as_str(), actual_type_di_node)) }) - .collect(); + }) + .collect(); - return template_params; - } + return template_params; } return smallvec![]; @@ -1416,7 +1442,9 @@ fn build_vtable_type_di_node<'ll, 'tcx>( let void_pointer_ty = Ty::new_imm_ptr(tcx, tcx.types.unit); let void_pointer_type_di_node = type_di_node(cx, void_pointer_ty); let usize_di_node = type_di_node(cx, tcx.types.usize); - let (pointer_size, pointer_align) = cx.size_and_align_of(void_pointer_ty); + let pointer_layout = cx.layout_of(void_pointer_ty); + let pointer_size = pointer_layout.size; + let pointer_align = pointer_layout.align.abi; // If `usize` is not pointer-sized and -aligned then the size and alignment computations // for the vtable as a whole would be wrong. Let's make sure this holds even on weird // platforms. @@ -1472,7 +1500,7 @@ fn build_vtable_type_di_node<'ll, 'tcx>( cx, vtable_type_di_node, &field_name, - (pointer_size, pointer_align), + pointer_layout, field_offset, DIFlags::FlagZero, field_type_di_node, 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 a72e205c9b2..07075be55fa 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 @@ -17,8 +17,8 @@ use crate::debuginfo::metadata::enums::DiscrResult; use crate::debuginfo::metadata::type_map::{self, Stub, UniqueTypeId}; use crate::debuginfo::metadata::{ DINodeCreationResult, NO_GENERICS, NO_SCOPE_METADATA, SmallVec, UNKNOWN_LINE_NUMBER, - build_field_di_node, file_metadata, file_metadata_from_def_id, size_and_align_of, type_di_node, - unknown_file_metadata, visibility_di_flags, + build_field_di_node, create_member_type, file_metadata, file_metadata_from_def_id, + size_and_align_of, type_di_node, unknown_file_metadata, visibility_di_flags, }; use crate::debuginfo::utils::DIB; use crate::llvm::debuginfo::{DIFile, DIFlags, DIType}; @@ -370,9 +370,9 @@ fn build_single_variant_union_fields<'ll, 'tcx>( cx, enum_type_di_node, &variant_union_field_name(variant_index), - // NOTE: We use the size and align of the entire type, not from variant_layout + // NOTE: We use the layout of the entire type, not from variant_layout // since the later is sometimes smaller (if it has fewer fields). - size_and_align_of(enum_type_and_layout), + enum_type_and_layout, Size::ZERO, visibility_flags, variant_struct_type_wrapper_di_node, @@ -560,7 +560,7 @@ fn build_variant_struct_wrapper_type_di_node<'ll, 'tcx>( cx, wrapper_struct_type_di_node, "value", - size_and_align_of(enum_or_coroutine_type_and_layout), + enum_or_coroutine_type_and_layout, Size::ZERO, DIFlags::FlagZero, variant_struct_type_di_node, @@ -820,7 +820,6 @@ fn build_union_fields_for_direct_tag_enum_or_coroutine<'ll, 'tcx>( .unwrap_or_else(|| (unknown_file_metadata(cx), UNKNOWN_LINE_NUMBER)); let field_name = variant_union_field_name(variant_member_info.variant_index); - let (size, align) = size_and_align_of(enum_type_and_layout); let variant_struct_type_wrapper = build_variant_struct_wrapper_type_di_node( cx, @@ -840,27 +839,23 @@ fn build_union_fields_for_direct_tag_enum_or_coroutine<'ll, 'tcx>( }, ); - // We use LLVMRustDIBuilderCreateMemberType() member type directly because + // We use create_member_type() member type directly because // the build_field_di_node() function does not support specifying a source location, // which is something that we don't do anywhere else. - unsafe { - llvm::LLVMRustDIBuilderCreateMemberType( - DIB(cx), - enum_type_di_node, - field_name.as_c_char_ptr(), - field_name.len(), - file_di_node, - line_number, - // NOTE: We use the size and align of the entire type, not from variant_layout - // since the later is sometimes smaller (if it has fewer fields). - size.bits(), - align.bits() as u32, - // Union fields are always at offset zero - Size::ZERO.bits(), - di_flags, - variant_struct_type_wrapper, - ) - } + create_member_type( + cx, + enum_type_di_node, + &field_name, + file_di_node, + line_number, + // NOTE: We use the layout of the entire type, not from variant_layout + // since the later is sometimes smaller (if it has fewer fields). + enum_type_and_layout, + // Union fields are always at offset zero + Size::ZERO, + di_flags, + variant_struct_type_wrapper, + ) })); assert_eq!( @@ -874,7 +869,7 @@ fn build_union_fields_for_direct_tag_enum_or_coroutine<'ll, 'tcx>( if is_128_bits { let type_di_node = type_di_node(cx, cx.tcx.types.u64); - let size_and_align = cx.size_and_align_of(cx.tcx.types.u64); + let u64_layout = cx.layout_of(cx.tcx.types.u64); let (lo_offset, hi_offset) = match cx.tcx.data_layout.endian { Endian::Little => (0, 8), @@ -889,7 +884,7 @@ fn build_union_fields_for_direct_tag_enum_or_coroutine<'ll, 'tcx>( cx, enum_type_di_node, TAG_FIELD_NAME_128_LO, - size_and_align, + u64_layout, lo_offset, di_flags, type_di_node, @@ -900,7 +895,7 @@ fn build_union_fields_for_direct_tag_enum_or_coroutine<'ll, 'tcx>( cx, enum_type_di_node, TAG_FIELD_NAME_128_HI, - size_and_align, + u64_layout, hi_offset, DIFlags::FlagZero, type_di_node, @@ -911,7 +906,7 @@ fn build_union_fields_for_direct_tag_enum_or_coroutine<'ll, 'tcx>( cx, enum_type_di_node, TAG_FIELD_NAME, - cx.size_and_align_of(enum_type_and_layout.field(cx, tag_field).ty), + enum_type_and_layout.field(cx, tag_field), enum_type_and_layout.fields.offset(tag_field), di_flags, tag_base_type_di_node, 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 9f6a5cc89e0..6792c307fdc 100644 --- a/compiler/rustc_codegen_llvm/src/debuginfo/metadata/enums/mod.rs +++ b/compiler/rustc_codegen_llvm/src/debuginfo/metadata/enums/mod.rs @@ -249,7 +249,7 @@ fn build_enum_variant_struct_type_di_node<'ll, 'tcx>( cx, struct_type_di_node, &field_name, - (field_layout.size, field_layout.align.abi), + field_layout, variant_layout.fields.offset(field_index), di_flags, type_di_node(cx, field_layout.ty), @@ -332,7 +332,7 @@ fn build_coroutine_variant_struct_type_di_node<'ll, 'tcx>( cx, variant_struct_type_di_node, &field_name, - cx.size_and_align_of(field_type), + cx.layout_of(field_type), variant_layout.fields.offset(field_index), DIFlags::FlagZero, type_di_node(cx, field_type), @@ -352,7 +352,7 @@ fn build_coroutine_variant_struct_type_di_node<'ll, 'tcx>( cx, variant_struct_type_di_node, upvar_name.as_str(), - cx.size_and_align_of(upvar_ty), + cx.layout_of(upvar_ty), coroutine_type_and_layout.fields.offset(index), DIFlags::FlagZero, type_di_node(cx, upvar_ty), @@ -363,6 +363,7 @@ fn build_coroutine_variant_struct_type_di_node<'ll, 'tcx>( state_specific_fields.into_iter().chain(common_fields).collect() }, + // FIXME: this is a no-op. `build_generic_type_param_di_nodes` only works for Adts. |cx| build_generic_type_param_di_nodes(cx, coroutine_type_and_layout.ty), ) .di_node 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 187d97c54c8..bfd131cfd3d 100644 --- a/compiler/rustc_codegen_llvm/src/debuginfo/metadata/enums/native.rs +++ b/compiler/rustc_codegen_llvm/src/debuginfo/metadata/enums/native.rs @@ -13,9 +13,9 @@ use smallvec::smallvec; use crate::common::{AsCCharPtr, CodegenCx}; use crate::debuginfo::metadata::type_map::{self, Stub, StubInfo, UniqueTypeId}; use crate::debuginfo::metadata::{ - DINodeCreationResult, NO_GENERICS, SmallVec, UNKNOWN_LINE_NUMBER, file_metadata, - file_metadata_from_def_id, size_and_align_of, type_di_node, unknown_file_metadata, - visibility_di_flags, + DINodeCreationResult, NO_GENERICS, SmallVec, UNKNOWN_LINE_NUMBER, create_member_type, + file_metadata, file_metadata_from_def_id, size_and_align_of, type_di_node, + unknown_file_metadata, visibility_di_flags, }; use crate::debuginfo::utils::{DIB, create_DIArray, get_namespace_for_item}; use crate::llvm::debuginfo::{DIFile, DIFlags, DIType}; @@ -363,23 +363,22 @@ fn build_discr_member_di_node<'ll, 'tcx>( &Variants::Multiple { tag_field, .. } => { let tag_base_type = tag_base_type(cx.tcx, enum_or_coroutine_type_and_layout); - let (size, align) = cx.size_and_align_of(tag_base_type); - - unsafe { - Some(llvm::LLVMRustDIBuilderCreateMemberType( - DIB(cx), - containing_scope, - tag_name.as_c_char_ptr(), - tag_name.len(), - unknown_file_metadata(cx), - UNKNOWN_LINE_NUMBER, - size.bits(), - align.bits() as u32, - enum_or_coroutine_type_and_layout.fields.offset(tag_field).bits(), - DIFlags::FlagArtificial, - type_di_node(cx, tag_base_type), - )) - } + let ty = type_di_node(cx, tag_base_type); + let file = unknown_file_metadata(cx); + + let layout = cx.layout_of(tag_base_type); + + Some(create_member_type( + cx, + containing_scope, + &tag_name, + file, + UNKNOWN_LINE_NUMBER, + layout, + enum_or_coroutine_type_and_layout.fields.offset(tag_field), + DIFlags::FlagArtificial, + ty, + )) } } } diff --git a/compiler/rustc_codegen_llvm/src/debuginfo/metadata/type_map.rs b/compiler/rustc_codegen_llvm/src/debuginfo/metadata/type_map.rs index af1d503ad6a..ae2ab32ef53 100644 --- a/compiler/rustc_codegen_llvm/src/debuginfo/metadata/type_map.rs +++ b/compiler/rustc_codegen_llvm/src/debuginfo/metadata/type_map.rs @@ -257,7 +257,7 @@ pub(super) fn build_type_with_children<'ll, 'tcx>( cx: &CodegenCx<'ll, 'tcx>, stub_info: StubInfo<'ll, 'tcx>, members: impl FnOnce(&CodegenCx<'ll, 'tcx>, &'ll DIType) -> SmallVec<&'ll DIType>, - generics: impl FnOnce(&CodegenCx<'ll, 'tcx>) -> SmallVec<&'ll DIType>, + generics: impl FnOnce(&CodegenCx<'ll, 'tcx>) -> SmallVec<Option<&'ll DIType>>, ) -> DINodeCreationResult<'ll> { assert_eq!(debug_context(cx).type_map.di_node_for_unique_id(stub_info.unique_type_id), None); @@ -265,8 +265,7 @@ pub(super) fn build_type_with_children<'ll, 'tcx>( let members: SmallVec<_> = members(cx, stub_info.metadata).into_iter().map(|node| Some(node)).collect(); - let generics: SmallVec<Option<&'ll DIType>> = - generics(cx).into_iter().map(|node| Some(node)).collect(); + let generics = generics(cx); if !(members.is_empty() && generics.is_empty()) { unsafe { diff --git a/compiler/rustc_codegen_llvm/src/debuginfo/mod.rs b/compiler/rustc_codegen_llvm/src/debuginfo/mod.rs index 10819a53b1d..ae7d080db66 100644 --- a/compiler/rustc_codegen_llvm/src/debuginfo/mod.rs +++ b/compiler/rustc_codegen_llvm/src/debuginfo/mod.rs @@ -2,10 +2,11 @@ use std::cell::{OnceCell, RefCell}; use std::ops::Range; +use std::ptr; use std::sync::Arc; -use std::{iter, ptr}; use libc::c_uint; +use metadata::create_subroutine_type; use rustc_abi::Size; use rustc_codegen_ssa::debuginfo::type_names; use rustc_codegen_ssa::mir::debuginfo::VariableKind::*; @@ -34,8 +35,8 @@ use crate::builder::Builder; use crate::common::{AsCCharPtr, CodegenCx}; use crate::llvm; use crate::llvm::debuginfo::{ - DIArray, DIBuilderBox, DIFile, DIFlags, DILexicalBlock, DILocation, DISPFlags, DIScope, DIType, - DIVariable, + DIArray, DIBuilderBox, DIFile, DIFlags, DILexicalBlock, DILocation, DISPFlags, DIScope, + DITemplateTypeParameter, DIType, DIVariable, }; use crate::value::Value; @@ -251,7 +252,7 @@ struct DebugLoc { col: u32, } -impl CodegenCx<'_, '_> { +impl<'ll> CodegenCx<'ll, '_> { /// Looks up debug source information about a `BytePos`. // FIXME(eddyb) rename this to better indicate it's a duplicate of // `lookup_char_pos` rather than `dbg_loc`, perhaps by making @@ -279,6 +280,22 @@ impl CodegenCx<'_, '_> { DebugLoc { file, line, col } } } + + fn create_template_type_parameter( + &self, + name: &str, + actual_type_metadata: &'ll DIType, + ) -> &'ll DITemplateTypeParameter { + unsafe { + llvm::LLVMRustDIBuilderCreateTemplateTypeParameter( + DIB(self), + None, + name.as_c_char_ptr(), + name.len(), + actual_type_metadata, + ) + } + } } impl<'ll, 'tcx> DebugInfoCodegenMethods<'tcx> for CodegenCx<'ll, 'tcx> { @@ -325,10 +342,8 @@ impl<'ll, 'tcx> DebugInfoCodegenMethods<'tcx> for CodegenCx<'ll, 'tcx> { let loc = self.lookup_debug_loc(span.lo()); let file_metadata = file_metadata(self, &loc.file); - let function_type_metadata = unsafe { - let fn_signature = get_function_signature(self, fn_abi); - llvm::LLVMRustDIBuilderCreateSubroutineType(DIB(self), fn_signature) - }; + let function_type_metadata = + create_subroutine_type(self, get_function_signature(self, fn_abi)); let mut name = String::with_capacity(64); type_names::push_item_name(tcx, def_id, false, &mut name); @@ -471,46 +486,10 @@ impl<'ll, 'tcx> DebugInfoCodegenMethods<'tcx> for CodegenCx<'ll, 'tcx> { generics: &ty::Generics, args: GenericArgsRef<'tcx>, ) -> &'ll DIArray { - if args.types().next().is_none() { - return create_DIArray(DIB(cx), &[]); - } - - // Again, only create type information if full debuginfo is enabled - let template_params: Vec<_> = if cx.sess().opts.debuginfo == DebugInfo::Full { - let names = get_parameter_names(cx, generics); - iter::zip(args, names) - .filter_map(|(kind, name)| { - kind.as_type().map(|ty| { - let actual_type = cx.tcx.normalize_erasing_regions(cx.typing_env(), ty); - let actual_type_metadata = type_di_node(cx, actual_type); - let name = name.as_str(); - unsafe { - Some(llvm::LLVMRustDIBuilderCreateTemplateTypeParameter( - DIB(cx), - None, - name.as_c_char_ptr(), - name.len(), - actual_type_metadata, - )) - } - }) - }) - .collect() - } else { - vec![] - }; - + let template_params = metadata::get_template_parameters(cx, generics, args); create_DIArray(DIB(cx), &template_params) } - fn get_parameter_names(cx: &CodegenCx<'_, '_>, generics: &ty::Generics) -> Vec<Symbol> { - let mut names = generics.parent.map_or_else(Vec::new, |def_id| { - get_parameter_names(cx, cx.tcx.generics_of(def_id)) - }); - names.extend(generics.own_params.iter().map(|param| param.name)); - names - } - /// Returns a scope, plus `true` if that's a type scope for "class" methods, /// otherwise `false` for plain namespace scopes. fn get_containing_scope<'ll, 'tcx>( diff --git a/compiler/rustc_codegen_llvm/src/errors.rs b/compiler/rustc_codegen_llvm/src/errors.rs index 4c5a78ca74f..ecf108f988f 100644 --- a/compiler/rustc_codegen_llvm/src/errors.rs +++ b/compiler/rustc_codegen_llvm/src/errors.rs @@ -217,3 +217,7 @@ pub(crate) struct MismatchedDataLayout<'a> { pub(crate) struct FixedX18InvalidArch<'a> { pub arch: &'a str, } + +#[derive(Diagnostic)] +#[diag(codegen_llvm_sanitizer_kcfi_arity_requires_llvm_21_0_0)] +pub(crate) struct SanitizerKcfiArityRequiresLLVM2100; diff --git a/compiler/rustc_codegen_llvm/src/intrinsic.rs b/compiler/rustc_codegen_llvm/src/intrinsic.rs index 660fc7ec4c4..67135fcc308 100644 --- a/compiler/rustc_codegen_llvm/src/intrinsic.rs +++ b/compiler/rustc_codegen_llvm/src/intrinsic.rs @@ -14,6 +14,7 @@ use rustc_middle::ty::layout::{FnAbiOf, HasTyCtxt, HasTypingEnv, LayoutOf}; use rustc_middle::ty::{self, GenericArgsRef, Ty}; use rustc_middle::{bug, span_bug}; use rustc_span::{Span, Symbol, sym}; +use rustc_symbol_mangling::mangle_internal_symbol; use rustc_target::callconv::{FnAbi, PassMode}; use rustc_target::spec::{HasTargetSpec, PanicStrategy}; use tracing::debug; @@ -812,7 +813,10 @@ fn codegen_msvc_try<'ll>( let type_name = bx.const_bytes(b"rust_panic\0"); let type_info = bx.const_struct(&[type_info_vtable, bx.const_null(bx.type_ptr()), type_name], false); - let tydesc = bx.declare_global("__rust_panic_type_info", bx.val_ty(type_info)); + let tydesc = bx.declare_global( + &mangle_internal_symbol(bx.tcx, "__rust_panic_type_info"), + bx.val_ty(type_info), + ); llvm::set_linkage(tydesc, llvm::Linkage::LinkOnceODRLinkage); if bx.cx.tcx.sess.target.supports_comdat() { diff --git a/compiler/rustc_codegen_llvm/src/lib.rs b/compiler/rustc_codegen_llvm/src/lib.rs index dec93bbccc3..425381b0ffa 100644 --- a/compiler/rustc_codegen_llvm/src/lib.rs +++ b/compiler/rustc_codegen_llvm/src/lib.rs @@ -19,7 +19,6 @@ #![feature(rustdoc_internals)] #![feature(slice_as_array)] #![feature(try_blocks)] -#![warn(unreachable_pub)] // tidy-alphabetical-end use std::any::Any; @@ -30,7 +29,7 @@ use back::owned_target_machine::OwnedTargetMachine; use back::write::{create_informational_target_machine, create_target_machine}; use context::SimpleCx; use errors::{AutoDiffWithoutLTO, ParseTargetMachineConfig}; -pub(crate) use llvm_util::target_features_cfg; +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}; @@ -72,9 +71,7 @@ mod debuginfo; mod declare; mod errors; mod intrinsic; -// FIXME(Zalathar): Fix all the unreachable-pub warnings that would occur if -// this isn't pub, then make it not pub. -pub mod llvm; +mod llvm; mod llvm_util; mod mono_item; mod type_; diff --git a/compiler/rustc_codegen_llvm/src/llvm/archive_ro.rs b/compiler/rustc_codegen_llvm/src/llvm/archive_ro.rs index 63b2b15c514..51bcc4d123d 100644 --- a/compiler/rustc_codegen_llvm/src/llvm/archive_ro.rs +++ b/compiler/rustc_codegen_llvm/src/llvm/archive_ro.rs @@ -5,17 +5,17 @@ use std::{slice, str}; use rustc_fs_util::path_to_c_string; -pub struct ArchiveRO { +pub(crate) struct ArchiveRO { pub raw: &'static mut super::Archive, } unsafe impl Send for ArchiveRO {} -pub struct Iter<'a> { +pub(crate) struct Iter<'a> { raw: &'a mut super::ArchiveIterator<'a>, } -pub struct Child<'a> { +pub(crate) struct Child<'a> { pub raw: &'a mut super::ArchiveChild<'a>, } diff --git a/compiler/rustc_codegen_llvm/src/llvm/diagnostic.rs b/compiler/rustc_codegen_llvm/src/llvm/diagnostic.rs index 11043b664f5..0e0f2b0eab0 100644 --- a/compiler/rustc_codegen_llvm/src/llvm/diagnostic.rs +++ b/compiler/rustc_codegen_llvm/src/llvm/diagnostic.rs @@ -3,13 +3,13 @@ use libc::c_uint; use rustc_span::InnerSpan; -pub use self::Diagnostic::*; -pub use self::OptimizationDiagnosticKind::*; +pub(crate) use self::Diagnostic::*; +use self::OptimizationDiagnosticKind::*; use super::{DiagnosticInfo, SMDiagnostic}; use crate::value::Value; #[derive(Copy, Clone, Debug)] -pub enum OptimizationDiagnosticKind { +pub(crate) enum OptimizationDiagnosticKind { OptimizationRemark, OptimizationMissed, OptimizationAnalysis, @@ -19,9 +19,10 @@ pub enum OptimizationDiagnosticKind { OptimizationRemarkOther, } -pub struct OptimizationDiagnostic<'ll> { +pub(crate) struct OptimizationDiagnostic<'ll> { pub kind: OptimizationDiagnosticKind, pub pass_name: String, + #[expect(dead_code)] pub function: &'ll Value, pub line: c_uint, pub column: c_uint, @@ -73,14 +74,14 @@ impl<'ll> OptimizationDiagnostic<'ll> { } } -pub struct SrcMgrDiagnostic { +pub(crate) struct SrcMgrDiagnostic { pub level: super::DiagnosticLevel, pub message: String, pub source: Option<(String, Vec<InnerSpan>)>, } impl SrcMgrDiagnostic { - pub unsafe fn unpack(diag: &SMDiagnostic) -> SrcMgrDiagnostic { + pub(crate) unsafe fn unpack(diag: &SMDiagnostic) -> SrcMgrDiagnostic { // Recover the post-substitution assembly code from LLVM for better // diagnostics. let mut have_source = false; @@ -120,7 +121,7 @@ impl SrcMgrDiagnostic { } #[derive(Clone)] -pub struct InlineAsmDiagnostic { +pub(crate) struct InlineAsmDiagnostic { pub level: super::DiagnosticLevel, pub cookie: u64, pub message: String, @@ -158,7 +159,7 @@ impl InlineAsmDiagnostic { } } -pub enum Diagnostic<'ll> { +pub(crate) enum Diagnostic<'ll> { Optimization(OptimizationDiagnostic<'ll>), InlineAsm(InlineAsmDiagnostic), PGO(&'ll DiagnosticInfo), @@ -166,11 +167,12 @@ pub enum Diagnostic<'ll> { Unsupported(&'ll DiagnosticInfo), /// LLVM has other types that we do not wrap here. + #[expect(dead_code)] UnknownDiagnostic(&'ll DiagnosticInfo), } impl<'ll> Diagnostic<'ll> { - pub unsafe fn unpack(di: &'ll DiagnosticInfo) -> Self { + pub(crate) unsafe fn unpack(di: &'ll DiagnosticInfo) -> Self { use super::DiagnosticKind as Dk; unsafe { diff --git a/compiler/rustc_codegen_llvm/src/llvm/enzyme_ffi.rs b/compiler/rustc_codegen_llvm/src/llvm/enzyme_ffi.rs index f6b23862907..a9b3bdf7344 100644 --- a/compiler/rustc_codegen_llvm/src/llvm/enzyme_ffi.rs +++ b/compiler/rustc_codegen_llvm/src/llvm/enzyme_ffi.rs @@ -4,7 +4,7 @@ use libc::{c_char, c_uint}; use super::MetadataKindId; -use super::ffi::{BasicBlock, Metadata, Module, Type, Value}; +use super::ffi::{AttributeKind, BasicBlock, Metadata, Module, Type, Value}; use crate::llvm::Bool; #[link(name = "llvm-wrapper", kind = "static")] @@ -17,6 +17,8 @@ unsafe extern "C" { pub(crate) fn LLVMRustEraseInstFromParent(V: &Value); pub(crate) fn LLVMRustGetTerminator<'a>(B: &BasicBlock) -> &'a Value; pub(crate) fn LLVMRustVerifyFunction(V: &Value, action: LLVMRustVerifierFailureAction) -> Bool; + pub(crate) fn LLVMRustHasAttributeAtIndex(V: &Value, i: c_uint, Kind: AttributeKind) -> bool; + pub(crate) fn LLVMRustGetArrayNumElements(Ty: &Type) -> u64; } unsafe extern "C" { @@ -31,20 +33,20 @@ unsafe extern "C" { #[repr(C)] #[derive(Copy, Clone, PartialEq)] -pub enum LLVMRustVerifierFailureAction { +pub(crate) enum LLVMRustVerifierFailureAction { LLVMAbortProcessAction = 0, LLVMPrintMessageAction = 1, LLVMReturnStatusAction = 2, } #[cfg(llvm_enzyme)] -pub use self::Enzyme_AD::*; +pub(crate) use self::Enzyme_AD::*; #[cfg(llvm_enzyme)] -pub mod Enzyme_AD { +pub(crate) mod Enzyme_AD { use libc::c_void; unsafe extern "C" { - pub fn EnzymeSetCLBool(arg1: *mut ::std::os::raw::c_void, arg2: u8); + pub(crate) fn EnzymeSetCLBool(arg1: *mut ::std::os::raw::c_void, arg2: u8); } unsafe extern "C" { static mut EnzymePrintPerf: c_void; @@ -56,42 +58,42 @@ pub mod Enzyme_AD { static mut EnzymeInline: c_void; static mut RustTypeRules: c_void; } - pub fn set_print_perf(print: bool) { + pub(crate) fn set_print_perf(print: bool) { unsafe { EnzymeSetCLBool(std::ptr::addr_of_mut!(EnzymePrintPerf), print as u8); } } - pub fn set_print_activity(print: bool) { + pub(crate) fn set_print_activity(print: bool) { unsafe { EnzymeSetCLBool(std::ptr::addr_of_mut!(EnzymePrintActivity), print as u8); } } - pub fn set_print_type(print: bool) { + pub(crate) fn set_print_type(print: bool) { unsafe { EnzymeSetCLBool(std::ptr::addr_of_mut!(EnzymePrintType), print as u8); } } - pub fn set_print(print: bool) { + pub(crate) fn set_print(print: bool) { unsafe { EnzymeSetCLBool(std::ptr::addr_of_mut!(EnzymePrint), print as u8); } } - pub fn set_strict_aliasing(strict: bool) { + pub(crate) fn set_strict_aliasing(strict: bool) { unsafe { EnzymeSetCLBool(std::ptr::addr_of_mut!(EnzymeStrictAliasing), strict as u8); } } - pub fn set_loose_types(loose: bool) { + pub(crate) fn set_loose_types(loose: bool) { unsafe { EnzymeSetCLBool(std::ptr::addr_of_mut!(looseTypeAnalysis), loose as u8); } } - pub fn set_inline(val: bool) { + pub(crate) fn set_inline(val: bool) { unsafe { EnzymeSetCLBool(std::ptr::addr_of_mut!(EnzymeInline), val as u8); } } - pub fn set_rust_rules(val: bool) { + pub(crate) fn set_rust_rules(val: bool) { unsafe { EnzymeSetCLBool(std::ptr::addr_of_mut!(RustTypeRules), val as u8); } @@ -99,34 +101,34 @@ pub mod Enzyme_AD { } #[cfg(not(llvm_enzyme))] -pub use self::Fallback_AD::*; +pub(crate) use self::Fallback_AD::*; #[cfg(not(llvm_enzyme))] -pub mod Fallback_AD { +pub(crate) mod Fallback_AD { #![allow(unused_variables)] - pub fn set_inline(val: bool) { + pub(crate) fn set_inline(val: bool) { unimplemented!() } - pub fn set_print_perf(print: bool) { + pub(crate) fn set_print_perf(print: bool) { unimplemented!() } - pub fn set_print_activity(print: bool) { + pub(crate) fn set_print_activity(print: bool) { unimplemented!() } - pub fn set_print_type(print: bool) { + pub(crate) fn set_print_type(print: bool) { unimplemented!() } - pub fn set_print(print: bool) { + pub(crate) fn set_print(print: bool) { unimplemented!() } - pub fn set_strict_aliasing(strict: bool) { + pub(crate) fn set_strict_aliasing(strict: bool) { unimplemented!() } - pub fn set_loose_types(loose: bool) { + pub(crate) fn set_loose_types(loose: bool) { unimplemented!() } - pub fn set_rust_rules(val: bool) { + pub(crate) fn set_rust_rules(val: bool) { unimplemented!() } } diff --git a/compiler/rustc_codegen_llvm/src/llvm/ffi.rs b/compiler/rustc_codegen_llvm/src/llvm/ffi.rs index 39087a4d6f4..9ff04f72903 100644 --- a/compiler/rustc_codegen_llvm/src/llvm/ffi.rs +++ b/compiler/rustc_codegen_llvm/src/llvm/ffi.rs @@ -32,10 +32,10 @@ use crate::llvm; /// 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(crate) type Bool = c_int; -pub const True: Bool = 1 as Bool; -pub const False: Bool = 0 as Bool; +pub(crate) const True: Bool = 1 as Bool; +pub(crate) const False: Bool = 0 as Bool; /// Wrapper for a raw enum value returned from LLVM's C APIs. /// @@ -44,7 +44,7 @@ pub const False: Bool = 0 as Bool; /// value and returns it. Instead, return this raw wrapper, then convert to the /// Rust-side enum explicitly. #[repr(transparent)] -pub struct RawEnum<T> { +pub(crate) struct RawEnum<T> { value: u32, /// We don't own or consume a `T`, but we can produce one. _rust_side_type: PhantomData<fn() -> T>, @@ -64,7 +64,7 @@ impl<T: TryFrom<u32>> RawEnum<T> { #[derive(Copy, Clone, PartialEq)] #[repr(C)] #[allow(dead_code)] // Variants constructed by C++. -pub enum LLVMRustResult { +pub(crate) enum LLVMRustResult { Success, Failure, } @@ -83,7 +83,7 @@ pub enum LLVMRustResult { /// C++ API. #[derive(Copy, Clone, PartialEq)] #[repr(C)] -pub enum ModuleFlagMergeBehavior { +pub(crate) enum ModuleFlagMergeBehavior { Error = 1, Warning = 2, Require = 3, @@ -101,7 +101,7 @@ pub enum ModuleFlagMergeBehavior { /// See <https://github.com/llvm/llvm-project/blob/main/llvm/include/llvm/IR/CallingConv.h> #[derive(Copy, Clone, PartialEq, Debug, TryFromU32)] #[repr(C)] -pub enum CallConv { +pub(crate) enum CallConv { CCallConv = 0, FastCallConv = 8, ColdCallConv = 9, @@ -126,7 +126,7 @@ pub enum CallConv { /// Must match the layout of `LLVMLinkage`. #[derive(Copy, Clone, PartialEq, TryFromU32)] #[repr(C)] -pub enum Linkage { +pub(crate) enum Linkage { ExternalLinkage = 0, AvailableExternallyLinkage = 1, LinkOnceAnyLinkage = 2, @@ -153,7 +153,7 @@ pub enum Linkage { /// Must match the layout of `LLVMVisibility`. #[repr(C)] #[derive(Copy, Clone, PartialEq, TryFromU32)] -pub enum Visibility { +pub(crate) enum Visibility { Default = 0, Hidden = 1, Protected = 2, @@ -171,8 +171,9 @@ impl Visibility { /// LLVMUnnamedAddr #[repr(C)] -pub enum UnnamedAddr { +pub(crate) enum UnnamedAddr { No, + #[expect(dead_code)] Local, Global, } @@ -180,7 +181,7 @@ pub enum UnnamedAddr { /// LLVMDLLStorageClass #[derive(Copy, Clone)] #[repr(C)] -pub enum DLLStorageClass { +pub(crate) enum DLLStorageClass { #[allow(dead_code)] Default = 0, DllImport = 1, // Function to be imported from DLL. @@ -193,7 +194,8 @@ pub enum DLLStorageClass { /// though it is not ABI compatible (since it's a C++ enum) #[repr(C)] #[derive(Copy, Clone, Debug)] -pub enum AttributeKind { +#[expect(dead_code, reason = "Some variants are unused, but are kept to match the C++")] +pub(crate) enum AttributeKind { AlwaysInline = 0, ByVal = 1, Cold = 2, @@ -241,7 +243,7 @@ pub enum AttributeKind { /// LLVMIntPredicate #[derive(Copy, Clone)] #[repr(C)] -pub enum IntPredicate { +pub(crate) enum IntPredicate { IntEQ = 32, IntNE = 33, IntUGT = 34, @@ -275,7 +277,7 @@ impl IntPredicate { /// LLVMRealPredicate #[derive(Copy, Clone)] #[repr(C)] -pub enum RealPredicate { +pub(crate) enum RealPredicate { RealPredicateFalse = 0, RealOEQ = 1, RealOGT = 2, @@ -321,7 +323,8 @@ impl RealPredicate { /// LLVMTypeKind #[derive(Copy, Clone, PartialEq, Debug)] #[repr(C)] -pub enum TypeKind { +#[expect(dead_code, reason = "Some variants are unused, but are kept to match LLVM-C")] +pub(crate) enum TypeKind { Void = 0, Half = 1, Float = 2, @@ -373,7 +376,7 @@ impl TypeKind { /// LLVMAtomicRmwBinOp #[derive(Copy, Clone)] #[repr(C)] -pub enum AtomicRmwBinOp { +pub(crate) enum AtomicRmwBinOp { AtomicXchg = 0, AtomicAdd = 1, AtomicSub = 2, @@ -409,7 +412,7 @@ impl AtomicRmwBinOp { /// LLVMAtomicOrdering #[derive(Copy, Clone)] #[repr(C)] -pub enum AtomicOrdering { +pub(crate) enum AtomicOrdering { #[allow(dead_code)] NotAtomic = 0, Unordered = 1, @@ -438,7 +441,7 @@ impl AtomicOrdering { /// LLVMRustFileType #[derive(Copy, Clone)] #[repr(C)] -pub enum FileType { +pub(crate) enum FileType { AssemblyFile, ObjectFile, } @@ -446,7 +449,8 @@ pub enum FileType { /// LLVMMetadataType #[derive(Copy, Clone)] #[repr(C)] -pub enum MetadataType { +#[expect(dead_code, reason = "Some variants are unused, but are kept to match LLVM-C")] +pub(crate) enum MetadataType { MD_dbg = 0, MD_tbaa = 1, MD_prof = 2, @@ -470,7 +474,7 @@ pub enum MetadataType { /// LLVMRustAsmDialect #[derive(Copy, Clone, PartialEq)] #[repr(C)] -pub enum AsmDialect { +pub(crate) enum AsmDialect { Att, Intel, } @@ -478,7 +482,7 @@ pub enum AsmDialect { /// LLVMRustCodeGenOptLevel #[derive(Copy, Clone, PartialEq)] #[repr(C)] -pub enum CodeGenOptLevel { +pub(crate) enum CodeGenOptLevel { None, Less, Default, @@ -487,7 +491,7 @@ pub enum CodeGenOptLevel { /// LLVMRustPassBuilderOptLevel #[repr(C)] -pub enum PassBuilderOptLevel { +pub(crate) enum PassBuilderOptLevel { O0, O1, O2, @@ -499,7 +503,7 @@ pub enum PassBuilderOptLevel { /// LLVMRustOptStage #[derive(PartialEq)] #[repr(C)] -pub enum OptStage { +pub(crate) enum OptStage { PreLinkNoLTO, PreLinkThinLTO, PreLinkFatLTO, @@ -509,7 +513,7 @@ pub enum OptStage { /// LLVMRustSanitizerOptions #[repr(C)] -pub struct SanitizerOptions { +pub(crate) struct SanitizerOptions { pub sanitize_address: bool, pub sanitize_address_recover: bool, pub sanitize_cfi: bool, @@ -530,7 +534,7 @@ pub struct SanitizerOptions { /// LLVMRustRelocModel #[derive(Copy, Clone, PartialEq)] #[repr(C)] -pub enum RelocModel { +pub(crate) enum RelocModel { Static, PIC, DynamicNoPic, @@ -542,7 +546,7 @@ pub enum RelocModel { /// LLVMRustFloatABI #[derive(Copy, Clone, PartialEq)] #[repr(C)] -pub enum FloatAbi { +pub(crate) enum FloatAbi { Default, Soft, Hard, @@ -551,7 +555,7 @@ pub enum FloatAbi { /// LLVMRustCodeModel #[derive(Copy, Clone)] #[repr(C)] -pub enum CodeModel { +pub(crate) enum CodeModel { Tiny, Small, Kernel, @@ -564,7 +568,7 @@ pub enum CodeModel { #[derive(Copy, Clone)] #[repr(C)] #[allow(dead_code)] // Variants constructed by C++. -pub enum DiagnosticKind { +pub(crate) enum DiagnosticKind { Other, InlineAsm, StackSize, @@ -587,7 +591,7 @@ pub enum DiagnosticKind { #[derive(Copy, Clone)] #[repr(C)] #[allow(dead_code)] // Variants constructed by C++. -pub enum DiagnosticLevel { +pub(crate) enum DiagnosticLevel { Error, Warning, Note, @@ -597,7 +601,7 @@ pub enum DiagnosticLevel { /// LLVMRustArchiveKind #[derive(Copy, Clone)] #[repr(C)] -pub enum ArchiveKind { +pub(crate) enum ArchiveKind { K_GNU, K_BSD, K_DARWIN, @@ -607,15 +611,15 @@ pub enum ArchiveKind { unsafe extern "C" { // LLVMRustThinLTOData - pub type ThinLTOData; + pub(crate) type ThinLTOData; // LLVMRustThinLTOBuffer - pub type ThinLTOBuffer; + pub(crate) type ThinLTOBuffer; } /// LLVMRustThinLTOModule #[repr(C)] -pub struct ThinLTOModule { +pub(crate) struct ThinLTOModule { pub identifier: *const c_char, pub data: *const u8, pub len: usize, @@ -624,7 +628,8 @@ pub struct ThinLTOModule { /// LLVMThreadLocalMode #[derive(Copy, Clone)] #[repr(C)] -pub enum ThreadLocalMode { +pub(crate) enum ThreadLocalMode { + #[expect(dead_code)] NotThreadLocal, GeneralDynamic, LocalDynamic, @@ -635,7 +640,7 @@ pub enum ThreadLocalMode { /// LLVMRustChecksumKind #[derive(Copy, Clone)] #[repr(C)] -pub enum ChecksumKind { +pub(crate) enum ChecksumKind { None, MD5, SHA1, @@ -645,7 +650,7 @@ pub enum ChecksumKind { /// LLVMRustMemoryEffects #[derive(Copy, Clone)] #[repr(C)] -pub enum MemoryEffects { +pub(crate) enum MemoryEffects { None, ReadOnly, InaccessibleMemOnly, @@ -654,7 +659,8 @@ pub enum MemoryEffects { /// LLVMOpcode #[derive(Copy, Clone, PartialEq, Eq)] #[repr(C)] -pub enum Opcode { +#[expect(dead_code, reason = "Some variants are unused, but are kept to match LLVM-C")] +pub(crate) enum Opcode { Ret = 1, Br = 2, Switch = 3, @@ -735,48 +741,48 @@ struct InvariantOpaque<'a> { // Opaque pointer types unsafe extern "C" { - pub type Module; - pub type Context; - pub type Type; - pub type Value; - pub type ConstantInt; - pub type Attribute; - pub type Metadata; - pub type BasicBlock; - pub type Comdat; + pub(crate) type Module; + pub(crate) type Context; + pub(crate) type Type; + pub(crate) type Value; + pub(crate) type ConstantInt; + pub(crate) type Attribute; + pub(crate) type Metadata; + pub(crate) type BasicBlock; + pub(crate) type Comdat; } #[repr(C)] -pub struct Builder<'a>(InvariantOpaque<'a>); +pub(crate) struct Builder<'a>(InvariantOpaque<'a>); #[repr(C)] -pub struct PassManager<'a>(InvariantOpaque<'a>); +pub(crate) struct PassManager<'a>(InvariantOpaque<'a>); unsafe extern "C" { pub type TargetMachine; - pub type Archive; + pub(crate) type Archive; } #[repr(C)] -pub struct ArchiveIterator<'a>(InvariantOpaque<'a>); +pub(crate) struct ArchiveIterator<'a>(InvariantOpaque<'a>); #[repr(C)] -pub struct ArchiveChild<'a>(InvariantOpaque<'a>); +pub(crate) struct ArchiveChild<'a>(InvariantOpaque<'a>); unsafe extern "C" { - pub type Twine; - pub type DiagnosticInfo; - pub type SMDiagnostic; + pub(crate) type Twine; + pub(crate) type DiagnosticInfo; + pub(crate) type SMDiagnostic; } #[repr(C)] -pub struct RustArchiveMember<'a>(InvariantOpaque<'a>); +pub(crate) struct RustArchiveMember<'a>(InvariantOpaque<'a>); /// Opaque pointee of `LLVMOperandBundleRef`. #[repr(C)] pub(crate) struct OperandBundle<'a>(InvariantOpaque<'a>); #[repr(C)] -pub struct Linker<'a>(InvariantOpaque<'a>); +pub(crate) struct Linker<'a>(InvariantOpaque<'a>); unsafe extern "C" { - pub type DiagnosticHandler; + pub(crate) type DiagnosticHandler; } -pub type DiagnosticHandlerTy = unsafe extern "C" fn(&DiagnosticInfo, *mut c_void); +pub(crate) type DiagnosticHandlerTy = unsafe extern "C" fn(&DiagnosticInfo, *mut c_void); -pub mod debuginfo { +pub(crate) mod debuginfo { use std::ptr; use bitflags::bitflags; @@ -793,7 +799,7 @@ pub mod debuginfo { /// builder reference typically has a shorter lifetime than the LLVM /// session (`'ll`) that it participates in. #[repr(C)] - pub struct DIBuilder<'ll>(InvariantOpaque<'ll>); + pub(crate) struct DIBuilder<'ll>(InvariantOpaque<'ll>); /// Owning pointer to a `DIBuilder<'ll>` that will dispose of the builder /// when dropped. Use `.as_ref()` to get the underlying `&DIBuilder` @@ -822,22 +828,22 @@ pub mod debuginfo { } } - pub type DIDescriptor = Metadata; - pub type DILocation = Metadata; - pub type DIScope = DIDescriptor; - pub type DIFile = DIScope; - pub type DILexicalBlock = DIScope; - pub type DISubprogram = DIScope; - pub type DIType = DIDescriptor; - pub type DIBasicType = DIType; - pub type DIDerivedType = DIType; - pub type DICompositeType = DIDerivedType; - pub type DIVariable = DIDescriptor; - pub type DIGlobalVariableExpression = DIDescriptor; - pub type DIArray = DIDescriptor; - pub type DISubrange = DIDescriptor; - pub type DIEnumerator = DIDescriptor; - pub type DITemplateTypeParameter = DIDescriptor; + pub(crate) type DIDescriptor = Metadata; + pub(crate) type DILocation = Metadata; + pub(crate) type DIScope = DIDescriptor; + pub(crate) type DIFile = DIScope; + pub(crate) type DILexicalBlock = DIScope; + pub(crate) type DISubprogram = DIScope; + pub(crate) type DIType = DIDescriptor; + pub(crate) type DIBasicType = DIType; + pub(crate) type DIDerivedType = DIType; + pub(crate) type DICompositeType = DIDerivedType; + pub(crate) type DIVariable = DIDescriptor; + pub(crate) type DIGlobalVariableExpression = DIDescriptor; + pub(crate) type DIArray = DIDescriptor; + pub(crate) type DISubrange = DIDescriptor; + pub(crate) type DIEnumerator = DIDescriptor; + pub(crate) type DITemplateTypeParameter = DIDescriptor; bitflags! { /// Must match the layout of `LLVMDIFlags` in the LLVM-C API. @@ -846,7 +852,7 @@ pub mod debuginfo { /// assertions in `RustWrapper.cpp` used by `fromRust(LLVMDIFlags)`. #[repr(transparent)] #[derive(Clone, Copy, Default)] - pub struct DIFlags: u32 { + pub(crate) struct DIFlags: u32 { const FlagZero = 0; const FlagPrivate = 1; const FlagProtected = 2; @@ -886,7 +892,7 @@ pub mod debuginfo { bitflags! { #[repr(transparent)] #[derive(Clone, Copy, Default)] - pub struct DISPFlags: u32 { + pub(crate) struct DISPFlags: u32 { const SPFlagZero = 0; const SPFlagVirtual = 1; const SPFlagPureVirtual = 2; @@ -900,7 +906,7 @@ pub mod debuginfo { /// LLVMRustDebugEmissionKind #[derive(Copy, Clone)] #[repr(C)] - pub enum DebugEmissionKind { + pub(crate) enum DebugEmissionKind { NoDebug, FullDebug, LineTablesOnly, @@ -932,8 +938,9 @@ pub mod debuginfo { /// LLVMRustDebugNameTableKind #[derive(Clone, Copy)] #[repr(C)] - pub enum DebugNameTableKind { + pub(crate) enum DebugNameTableKind { Default, + #[expect(dead_code)] Gnu, None, } @@ -943,7 +950,7 @@ pub mod debuginfo { bitflags! { #[repr(transparent)] #[derive(Default)] - pub struct AllocKindFlags : u64 { + pub(crate) struct AllocKindFlags : u64 { const Unknown = 0; const Alloc = 1; const Realloc = 1 << 1; @@ -966,19 +973,20 @@ bitflags! { } unsafe extern "C" { - pub type ModuleBuffer; + pub(crate) type ModuleBuffer; } -pub type SelfProfileBeforePassCallback = +pub(crate) type SelfProfileBeforePassCallback = unsafe extern "C" fn(*mut c_void, *const c_char, *const c_char); -pub type SelfProfileAfterPassCallback = unsafe extern "C" fn(*mut c_void); +pub(crate) type SelfProfileAfterPassCallback = unsafe extern "C" fn(*mut c_void); -pub type GetSymbolsCallback = unsafe extern "C" fn(*mut c_void, *const c_char) -> *mut c_void; -pub type GetSymbolsErrorCallback = unsafe extern "C" fn(*const c_char) -> *mut c_void; +pub(crate) type GetSymbolsCallback = + unsafe extern "C" fn(*mut c_void, *const c_char) -> *mut c_void; +pub(crate) type GetSymbolsErrorCallback = unsafe extern "C" fn(*const c_char) -> *mut c_void; #[derive(Copy, Clone)] #[repr(transparent)] -pub struct MetadataKindId(c_uint); +pub(crate) struct MetadataKindId(c_uint); impl From<MetadataType> for MetadataKindId { fn from(value: MetadataType) -> Self { @@ -1172,7 +1180,7 @@ unsafe extern "C" { // Operations on parameters pub(crate) fn LLVMIsAArgument(Val: &Value) -> Option<&Value>; - pub(crate) fn LLVMCountParams(Fn: &Value) -> c_uint; + pub(crate) safe fn LLVMCountParams(Fn: &Value) -> c_uint; pub(crate) fn LLVMGetParam(Fn: &Value, Index: c_uint) -> &Value; // Operations on basic blocks @@ -2019,6 +2027,8 @@ unsafe extern "C" { NumExpressions: size_t, CodeRegions: *const crate::coverageinfo::ffi::CodeRegion, NumCodeRegions: size_t, + ExpansionRegions: *const crate::coverageinfo::ffi::ExpansionRegion, + NumExpansionRegions: size_t, BranchRegions: *const crate::coverageinfo::ffi::BranchRegion, NumBranchRegions: size_t, MCDCBranchRegions: *const crate::coverageinfo::ffi::MCDCBranchRegion, diff --git a/compiler/rustc_codegen_llvm/src/llvm/mod.rs b/compiler/rustc_codegen_llvm/src/llvm/mod.rs index a36226b25a2..6ca81c651ed 100644 --- a/compiler/rustc_codegen_llvm/src/llvm/mod.rs +++ b/compiler/rustc_codegen_llvm/src/llvm/mod.rs @@ -9,18 +9,18 @@ use libc::c_uint; use rustc_abi::{Align, Size, WrappingRange}; use rustc_llvm::RustString; -pub use self::CallConv::*; -pub use self::CodeGenOptSize::*; -pub use self::MetadataType::*; -pub use self::ffi::*; +pub(crate) use self::CallConv::*; +pub(crate) use self::CodeGenOptSize::*; +pub(crate) use self::MetadataType::*; +pub(crate) use self::ffi::*; use crate::common::AsCCharPtr; -pub mod archive_ro; -pub mod diagnostic; -pub mod enzyme_ffi; +pub(crate) mod archive_ro; +pub(crate) mod diagnostic; +pub(crate) mod enzyme_ffi; mod ffi; -pub use self::enzyme_ffi::*; +pub(crate) use self::enzyme_ffi::*; impl LLVMRustResult { pub(crate) fn into_result(self) -> Result<(), ()> { @@ -127,7 +127,7 @@ pub(crate) fn CreateRangeAttr(llcx: &Context, size: Size, range: WrappingRange) } #[derive(Copy, Clone)] -pub enum AttributePlace { +pub(crate) enum AttributePlace { ReturnValue, Argument(u32), Function, @@ -145,7 +145,7 @@ impl AttributePlace { #[derive(Copy, Clone, PartialEq)] #[repr(C)] -pub enum CodeGenOptSize { +pub(crate) enum CodeGenOptSize { CodeGenOptSizeNone = 0, CodeGenOptSizeDefault = 1, CodeGenOptSizeAggressive = 2, diff --git a/compiler/rustc_codegen_llvm/src/llvm_util.rs b/compiler/rustc_codegen_llvm/src/llvm_util.rs index 4a166b0872d..4e85286ed55 100644 --- a/compiler/rustc_codegen_llvm/src/llvm_util.rs +++ b/compiler/rustc_codegen_llvm/src/llvm_util.rs @@ -274,7 +274,9 @@ pub(crate) fn to_llvm_features<'a>(sess: &Session, s: &'a str) -> Option<LLVMFea ("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 => { + ("riscv32" | "riscv64", "unaligned-scalar-mem" | "unaligned-vector-mem") + if get_version().0 == 18 => + { Some(LLVMFeature::new("fast-unaligned-access")) } // Filter out features that are not supported by the current LLVM version @@ -298,6 +300,13 @@ pub(crate) fn to_llvm_features<'a>(sess: &Session, s: &'a str) -> Option<LLVMFea ("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")), + // These new `amx` variants and `movrs` were introduced in LLVM20 + ("x86", "amx-avx512" | "amx-fp8" | "amx-movrs" | "amx-tf32" | "amx-transpose") + if get_version().0 < 20 => + { + None + } + ("x86", "movrs") if get_version().0 < 20 => None, (_, s) => Some(LLVMFeature::new(s)), } } diff --git a/compiler/rustc_codegen_llvm/src/mono_item.rs b/compiler/rustc_codegen_llvm/src/mono_item.rs index a64627eaf59..fdf62a08065 100644 --- a/compiler/rustc_codegen_llvm/src/mono_item.rs +++ b/compiler/rustc_codegen_llvm/src/mono_item.rs @@ -120,7 +120,7 @@ impl CodegenCx<'_, '_> { } // Match clang by only supporting COFF and ELF for now. - if self.tcx.sess.target.is_like_osx { + if self.tcx.sess.target.is_like_darwin { return false; } diff --git a/compiler/rustc_codegen_llvm/src/va_arg.rs b/compiler/rustc_codegen_llvm/src/va_arg.rs index 8baa69cefe1..c216f0f4a09 100644 --- a/compiler/rustc_codegen_llvm/src/va_arg.rs +++ b/compiler/rustc_codegen_llvm/src/va_arg.rs @@ -399,7 +399,7 @@ pub(super) fn emit_va_arg<'ll, 'tcx>( emit_ptr_va_arg(bx, addr, target_ty, false, Align::from_bytes(8).unwrap(), false) } // macOS / iOS AArch64 - "aarch64" if target.is_like_osx => { + "aarch64" if target.is_like_darwin => { emit_ptr_va_arg(bx, addr, target_ty, false, Align::from_bytes(8).unwrap(), true) } "aarch64" => emit_aapcs_va_arg(bx, addr, target_ty), |