diff options
Diffstat (limited to 'compiler/rustc_codegen_llvm/src')
18 files changed, 434 insertions, 214 deletions
diff --git a/compiler/rustc_codegen_llvm/src/abi.rs b/compiler/rustc_codegen_llvm/src/abi.rs index 1d35138b013..31ee0eeca11 100644 --- a/compiler/rustc_codegen_llvm/src/abi.rs +++ b/compiler/rustc_codegen_llvm/src/abi.rs @@ -1,3 +1,4 @@ +use std::borrow::Borrow; use std::cmp; use libc::c_uint; @@ -312,7 +313,7 @@ impl<'ll, 'tcx> ArgAbiBuilderMethods<'tcx> for Builder<'_, 'll, 'tcx> { pub(crate) trait FnAbiLlvmExt<'ll, 'tcx> { fn llvm_type(&self, cx: &CodegenCx<'ll, 'tcx>) -> &'ll Type; fn ptr_to_llvm_type(&self, cx: &CodegenCx<'ll, 'tcx>) -> &'ll Type; - fn llvm_cconv(&self) -> llvm::CallConv; + fn llvm_cconv(&self, cx: &CodegenCx<'ll, 'tcx>) -> llvm::CallConv; /// Apply attributes to a function declaration/definition. fn apply_attrs_llfn( @@ -404,8 +405,8 @@ impl<'ll, 'tcx> FnAbiLlvmExt<'ll, 'tcx> for FnAbi<'tcx, Ty<'tcx>> { cx.type_ptr_ext(cx.data_layout().instruction_address_space) } - fn llvm_cconv(&self) -> llvm::CallConv { - self.conv.into() + fn llvm_cconv(&self, cx: &CodegenCx<'ll, 'tcx>) -> llvm::CallConv { + llvm::CallConv::from_conv(self.conv, cx.tcx.sess.target.arch.borrow()) } fn apply_attrs_llfn( @@ -617,7 +618,7 @@ impl<'ll, 'tcx> FnAbiLlvmExt<'ll, 'tcx> for FnAbi<'tcx, Ty<'tcx>> { } } - let cconv = self.llvm_cconv(); + let cconv = self.llvm_cconv(&bx.cx); if cconv != llvm::CCallConv { llvm::SetInstructionCallConv(callsite, cconv); } @@ -655,8 +656,8 @@ impl<'tcx> AbiBuilderMethods<'tcx> for Builder<'_, '_, 'tcx> { } } -impl From<Conv> for llvm::CallConv { - fn from(conv: Conv) -> Self { +impl llvm::CallConv { + pub fn from_conv(conv: Conv, arch: &str) -> Self { match conv { Conv::C | Conv::Rust @@ -666,6 +667,15 @@ impl From<Conv> for llvm::CallConv { Conv::Cold => llvm::ColdCallConv, Conv::PreserveMost => llvm::PreserveMost, Conv::PreserveAll => llvm::PreserveAll, + Conv::GpuKernel => { + if arch == "amdgpu" { + llvm::AmdgpuKernel + } else if arch == "nvptx64" { + llvm::PtxKernel + } else { + panic!("Architecture {arch} does not support GpuKernel calling convention"); + } + } Conv::AvrInterrupt => llvm::AvrInterrupt, Conv::AvrNonBlockingInterrupt => llvm::AvrNonBlockingInterrupt, Conv::ArmAapcs => llvm::ArmAapcsCallConv, diff --git a/compiler/rustc_codegen_llvm/src/attributes.rs b/compiler/rustc_codegen_llvm/src/attributes.rs index 95e0481b035..3d7afa17bdf 100644 --- a/compiler/rustc_codegen_llvm/src/attributes.rs +++ b/compiler/rustc_codegen_llvm/src/attributes.rs @@ -333,9 +333,12 @@ pub(crate) fn llfn_attrs_from_instance<'ll, 'tcx>( let mut to_add = SmallVec::<[_; 16]>::new(); match codegen_fn_attrs.optimize { - OptimizeAttr::None => { + OptimizeAttr::Default => { to_add.extend(default_optimisation_attrs(cx)); } + OptimizeAttr::DoNotOptimize => { + to_add.push(llvm::AttributeKind::OptimizeNone.create_attr(cx.llcx)); + } OptimizeAttr::Size => { to_add.push(llvm::AttributeKind::MinSize.create_attr(cx.llcx)); to_add.push(llvm::AttributeKind::OptimizeForSize.create_attr(cx.llcx)); @@ -343,12 +346,12 @@ pub(crate) fn llfn_attrs_from_instance<'ll, 'tcx>( OptimizeAttr::Speed => {} } - let inline = - if codegen_fn_attrs.inline == InlineAttr::None && instance.def.requires_inline(cx.tcx) { - InlineAttr::Hint - } else { - codegen_fn_attrs.inline - }; + // `optnone` requires `noinline` + let inline = match (codegen_fn_attrs.inline, &codegen_fn_attrs.optimize) { + (_, OptimizeAttr::DoNotOptimize) => InlineAttr::Never, + (InlineAttr::None, _) if instance.def.requires_inline(cx.tcx) => InlineAttr::Hint, + (inline, _) => inline, + }; to_add.extend(inline_attr(cx, inline)); // The `uwtable` attribute according to LLVM is: diff --git a/compiler/rustc_codegen_llvm/src/back/write.rs b/compiler/rustc_codegen_llvm/src/back/write.rs index 509b24dd703..4706744f353 100644 --- a/compiler/rustc_codegen_llvm/src/back/write.rs +++ b/compiler/rustc_codegen_llvm/src/back/write.rs @@ -769,12 +769,9 @@ pub(crate) unsafe fn codegen( } } - // Two things to note: - // - If object files are just LLVM bitcode we write bitcode, copy it to - // the .o file, and delete the bitcode if it wasn't otherwise - // requested. - // - If we don't have the integrated assembler then we need to emit - // asm from LLVM and use `gcc` to create the object file. + // Note that if object files are just LLVM bitcode we write bitcode, + // copy it to the .o file, and delete the bitcode if it wasn't + // otherwise requested. let bc_out = cgcx.output_filenames.temp_path(OutputType::Bitcode, module_name); let bc_summary_out = diff --git a/compiler/rustc_codegen_llvm/src/builder.rs b/compiler/rustc_codegen_llvm/src/builder.rs index 5a34b52e6ef..d2de62b17f0 100644 --- a/compiler/rustc_codegen_llvm/src/builder.rs +++ b/compiler/rustc_codegen_llvm/src/builder.rs @@ -1,4 +1,4 @@ -use std::borrow::Cow; +use std::borrow::{Borrow, Cow}; use std::ops::Deref; use std::{iter, ptr}; @@ -31,20 +31,22 @@ use tracing::{debug, instrument}; use crate::abi::FnAbiLlvmExt; use crate::attributes; use crate::common::Funclet; -use crate::context::CodegenCx; +use crate::context::{CodegenCx, SimpleCx}; use crate::llvm::{self, AtomicOrdering, AtomicRmwBinOp, BasicBlock, False, True}; use crate::type_::Type; use crate::type_of::LayoutLlvmExt; use crate::value::Value; -// All Builders must have an llfn associated with them #[must_use] -pub(crate) struct Builder<'a, 'll, 'tcx> { +pub(crate) struct GenericBuilder<'a, 'll, CX: Borrow<SimpleCx<'ll>>> { pub llbuilder: &'ll mut llvm::Builder<'ll>, - pub cx: &'a CodegenCx<'ll, 'tcx>, + pub cx: &'a CX, } -impl Drop for Builder<'_, '_, '_> { +pub(crate) type SBuilder<'a, 'll> = GenericBuilder<'a, 'll, SimpleCx<'ll>>; +pub(crate) type Builder<'a, 'll, 'tcx> = GenericBuilder<'a, 'll, CodegenCx<'ll, 'tcx>>; + +impl<'a, 'll, CX: Borrow<SimpleCx<'ll>>> Drop for GenericBuilder<'a, 'll, CX> { fn drop(&mut self) { unsafe { llvm::LLVMDisposeBuilder(&mut *(self.llbuilder as *mut _)); @@ -52,6 +54,112 @@ impl Drop for Builder<'_, '_, '_> { } } +impl<'a, 'll> SBuilder<'a, 'll> { + fn call( + &mut self, + llty: &'ll Type, + llfn: &'ll Value, + args: &[&'ll Value], + funclet: Option<&Funclet<'ll>>, + ) -> &'ll Value { + debug!("call {:?} with args ({:?})", llfn, args); + + let args = self.check_call("call", llty, llfn, args); + let funclet_bundle = funclet.map(|funclet| funclet.bundle()); + let mut bundles: SmallVec<[_; 2]> = SmallVec::new(); + if let Some(funclet_bundle) = funclet_bundle { + bundles.push(funclet_bundle); + } + + let call = unsafe { + llvm::LLVMBuildCallWithOperandBundles( + self.llbuilder, + llty, + llfn, + args.as_ptr() as *const &llvm::Value, + args.len() as c_uint, + bundles.as_ptr(), + bundles.len() as c_uint, + c"".as_ptr(), + ) + }; + call + } + + fn with_scx(scx: &'a SimpleCx<'ll>) -> Self { + // Create a fresh builder from the simple context. + let llbuilder = unsafe { llvm::LLVMCreateBuilderInContext(scx.llcx) }; + SBuilder { llbuilder, cx: scx } + } +} +impl<'a, 'll, CX: Borrow<SimpleCx<'ll>>> GenericBuilder<'a, 'll, CX> { + pub(crate) fn bitcast(&mut self, val: &'ll Value, dest_ty: &'ll Type) -> &'ll Value { + unsafe { llvm::LLVMBuildBitCast(self.llbuilder, val, dest_ty, UNNAMED) } + } + + fn ret_void(&mut self) { + unsafe { + llvm::LLVMBuildRetVoid(self.llbuilder); + } + } + + fn ret(&mut self, v: &'ll Value) { + unsafe { + llvm::LLVMBuildRet(self.llbuilder, v); + } + } +} +impl<'a, 'll> SBuilder<'a, 'll> { + fn build(cx: &'a SimpleCx<'ll>, llbb: &'ll BasicBlock) -> SBuilder<'a, 'll> { + let bx = SBuilder::with_scx(cx); + unsafe { + llvm::LLVMPositionBuilderAtEnd(bx.llbuilder, llbb); + } + bx + } + + fn check_call<'b>( + &mut self, + typ: &str, + fn_ty: &'ll Type, + llfn: &'ll Value, + args: &'b [&'ll Value], + ) -> Cow<'b, [&'ll Value]> { + assert!( + self.cx.type_kind(fn_ty) == TypeKind::Function, + "builder::{typ} not passed a function, but {fn_ty:?}" + ); + + let param_tys = self.cx.func_params_types(fn_ty); + + let all_args_match = iter::zip(¶m_tys, args.iter().map(|&v| self.cx.val_ty(v))) + .all(|(expected_ty, actual_ty)| *expected_ty == actual_ty); + + if all_args_match { + return Cow::Borrowed(args); + } + + let casted_args: Vec<_> = iter::zip(param_tys, args) + .enumerate() + .map(|(i, (expected_ty, &actual_val))| { + let actual_ty = self.cx.val_ty(actual_val); + if expected_ty != actual_ty { + debug!( + "type mismatch in function call of {:?}. \ + Expected {:?} for param {}, got {:?}; injecting bitcast", + llfn, expected_ty, i, actual_ty + ); + self.bitcast(actual_val, expected_ty) + } else { + actual_val + } + }) + .collect(); + + Cow::Owned(casted_args) + } +} + /// Empty string, to be used where LLVM expects an instruction name, indicating /// that the instruction is to be left unnamed (i.e. numbered, in textual IR). // FIXME(eddyb) pass `&CStr` directly to FFI once it's a thin pointer. @@ -1222,6 +1330,14 @@ impl<'ll> StaticBuilderMethods for Builder<'_, 'll, '_> { } impl<'a, 'll, 'tcx> Builder<'a, 'll, 'tcx> { + fn build(cx: &'a CodegenCx<'ll, 'tcx>, llbb: &'ll BasicBlock) -> Builder<'a, 'll, 'tcx> { + let bx = Builder::with_cx(cx); + unsafe { + llvm::LLVMPositionBuilderAtEnd(bx.llbuilder, llbb); + } + bx + } + fn with_cx(cx: &'a CodegenCx<'ll, 'tcx>) -> Self { // Create a fresh builder from the crate context. let llbuilder = unsafe { llvm::LLVMCreateBuilderInContext(cx.llcx) }; @@ -1231,13 +1347,16 @@ impl<'a, 'll, 'tcx> Builder<'a, 'll, 'tcx> { pub(crate) fn llfn(&self) -> &'ll Value { unsafe { llvm::LLVMGetBasicBlockParent(self.llbb()) } } +} +impl<'a, 'll, CX: Borrow<SimpleCx<'ll>>> GenericBuilder<'a, 'll, CX> { fn position_at_start(&mut self, llbb: &'ll BasicBlock) { unsafe { llvm::LLVMRustPositionBuilderAtStart(self.llbuilder, llbb); } } - +} +impl<'a, 'll, 'tcx> Builder<'a, 'll, 'tcx> { fn align_metadata(&mut self, load: &'ll Value, align: Align) { unsafe { let md = [llvm::LLVMValueAsMetadata(self.cx.const_u64(align.bytes()))]; @@ -1259,7 +1378,8 @@ impl<'a, 'll, 'tcx> Builder<'a, 'll, 'tcx> { self.set_metadata(inst, llvm::MD_unpredictable, md); } } - +} +impl<'a, 'll, CX: Borrow<SimpleCx<'ll>>> GenericBuilder<'a, 'll, CX> { pub(crate) fn minnum(&mut self, lhs: &'ll Value, rhs: &'ll Value) -> &'ll Value { unsafe { llvm::LLVMRustBuildMinNum(self.llbuilder, lhs, rhs) } } @@ -1360,7 +1480,9 @@ impl<'a, 'll, 'tcx> Builder<'a, 'll, 'tcx> { let ret = unsafe { llvm::LLVMBuildCatchRet(self.llbuilder, funclet.cleanuppad(), unwind) }; ret.expect("LLVM does not have support for catchret") } +} +impl<'a, 'll, 'tcx> Builder<'a, 'll, 'tcx> { fn check_call<'b>( &mut self, typ: &str, @@ -1401,11 +1523,13 @@ impl<'a, 'll, 'tcx> Builder<'a, 'll, 'tcx> { Cow::Owned(casted_args) } - +} +impl<'a, 'll, CX: Borrow<SimpleCx<'ll>>> GenericBuilder<'a, 'll, CX> { pub(crate) fn va_arg(&mut self, list: &'ll Value, ty: &'ll Type) -> &'ll Value { unsafe { llvm::LLVMBuildVAArg(self.llbuilder, list, ty, UNNAMED) } } - +} +impl<'a, 'll, 'tcx> Builder<'a, 'll, 'tcx> { pub(crate) fn call_intrinsic(&mut self, intrinsic: &str, args: &[&'ll Value]) -> &'ll Value { let (ty, f) = self.cx.get_intrinsic(intrinsic); self.call(ty, None, None, f, args, None, None) @@ -1423,7 +1547,8 @@ impl<'a, 'll, 'tcx> Builder<'a, 'll, 'tcx> { self.call_intrinsic(intrinsic, &[self.cx.const_u64(size), ptr]); } - +} +impl<'a, 'll, CX: Borrow<SimpleCx<'ll>>> GenericBuilder<'a, 'll, CX> { pub(crate) fn phi( &mut self, ty: &'ll Type, @@ -1443,7 +1568,8 @@ impl<'a, 'll, 'tcx> Builder<'a, 'll, 'tcx> { llvm::LLVMAddIncoming(phi, &val, &bb, 1 as c_uint); } } - +} +impl<'a, 'll, 'tcx> Builder<'a, 'll, 'tcx> { fn fptoint_sat(&mut self, signed: bool, val: &'ll Value, dest_ty: &'ll Type) -> &'ll Value { let src_ty = self.cx.val_ty(val); let (float_ty, int_ty, vector_length) = if self.cx.type_kind(src_ty) == TypeKind::Vector { diff --git a/compiler/rustc_codegen_llvm/src/builder/autodiff.rs b/compiler/rustc_codegen_llvm/src/builder/autodiff.rs index 38f7eaa090f..6b17b5f6989 100644 --- a/compiler/rustc_codegen_llvm/src/builder/autodiff.rs +++ b/compiler/rustc_codegen_llvm/src/builder/autodiff.rs @@ -3,20 +3,19 @@ use std::ptr; use rustc_ast::expand::autodiff_attrs::{AutoDiffAttrs, AutoDiffItem, DiffActivity, DiffMode}; use rustc_codegen_ssa::ModuleCodegen; use rustc_codegen_ssa::back::write::ModuleConfig; -use rustc_codegen_ssa::traits::{BaseTypeCodegenMethods, BuilderMethods}; use rustc_errors::FatalError; -use rustc_middle::ty::TyCtxt; use rustc_session::config::Lto; use tracing::{debug, trace}; use crate::back::write::{llvm_err, llvm_optimize}; -use crate::builder::Builder; -use crate::declare::declare_raw_fn; +use crate::builder::SBuilder; +use crate::context::SimpleCx; +use crate::declare::declare_simple_fn; use crate::errors::LlvmError; use crate::llvm::AttributePlace::Function; use crate::llvm::{Metadata, True}; use crate::value::Value; -use crate::{CodegenContext, LlvmCodegenBackend, ModuleLlvm, attributes, context, llvm}; +use crate::{CodegenContext, LlvmCodegenBackend, ModuleLlvm, attributes, llvm}; fn get_params(fnc: &Value) -> Vec<&Value> { unsafe { @@ -38,8 +37,8 @@ fn get_params(fnc: &Value) -> Vec<&Value> { /// [^1]: <https://enzyme.mit.edu/getting_started/CallingConvention/> // FIXME(ZuseZ4): `outer_fn` should include upstream safety checks to // cover some assumptions of enzyme/autodiff, which could lead to UB otherwise. -fn generate_enzyme_call<'ll, 'tcx>( - cx: &context::CodegenCx<'ll, 'tcx>, +fn generate_enzyme_call<'ll>( + cx: &SimpleCx<'ll>, fn_to_diff: &'ll Value, outer_fn: &'ll Value, attrs: AutoDiffAttrs, @@ -112,7 +111,7 @@ fn generate_enzyme_call<'ll, 'tcx>( //FIXME(ZuseZ4): the CC/Addr/Vis values are best effort guesses, we should look at tests and // think a bit more about what should go here. let cc = llvm::LLVMGetFunctionCallConv(outer_fn); - let ad_fn = declare_raw_fn( + let ad_fn = declare_simple_fn( cx, &ad_name, llvm::CallConv::try_from(cc).expect("invalid callconv"), @@ -132,7 +131,7 @@ fn generate_enzyme_call<'ll, 'tcx>( llvm::LLVMRustEraseInstFromParent(br); let last_inst = llvm::LLVMRustGetLastInstruction(entry).unwrap(); - let mut builder = Builder::build(cx, entry); + let mut builder = SBuilder::build(cx, entry); let num_args = llvm::LLVMCountParams(&fn_to_diff); let mut args = Vec::with_capacity(num_args as usize + 1); @@ -236,7 +235,7 @@ fn generate_enzyme_call<'ll, 'tcx>( } } - let call = builder.call(enzyme_ty, None, None, ad_fn, &args, None, None); + 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 @@ -274,10 +273,9 @@ fn generate_enzyme_call<'ll, 'tcx>( } } -pub(crate) fn differentiate<'ll, 'tcx>( +pub(crate) fn differentiate<'ll>( module: &'ll ModuleCodegen<ModuleLlvm>, cgcx: &CodegenContext<LlvmCodegenBackend>, - tcx: TyCtxt<'tcx>, diff_items: Vec<AutoDiffItem>, config: &ModuleConfig, ) -> Result<(), FatalError> { @@ -286,8 +284,7 @@ pub(crate) fn differentiate<'ll, 'tcx>( } let diag_handler = cgcx.create_dcx(); - let (_, cgus) = tcx.collect_and_partition_mono_items(()); - let cx = context::CodegenCx::new(tcx, &cgus.first().unwrap(), &module.module_llvm); + let cx = SimpleCx { llmod: module.module_llvm.llmod(), llcx: module.module_llvm.llcx }; // Before dumping the module, we want all the TypeTrees to become part of the module. for item in diff_items.iter() { diff --git a/compiler/rustc_codegen_llvm/src/common.rs b/compiler/rustc_codegen_llvm/src/common.rs index adfe8aeb5c5..b4e9b9f44f4 100644 --- a/compiler/rustc_codegen_llvm/src/common.rs +++ b/compiler/rustc_codegen_llvm/src/common.rs @@ -126,6 +126,10 @@ impl<'ll, 'tcx> ConstCodegenMethods<'tcx> for CodegenCx<'ll, 'tcx> { unsafe { llvm::LLVMGetUndef(t) } } + fn is_undef(&self, v: &'ll Value) -> bool { + unsafe { llvm::LLVMIsUndef(v) == True } + } + fn const_poison(&self, t: &'ll Type) -> &'ll Value { unsafe { llvm::LLVMGetPoison(t) } } diff --git a/compiler/rustc_codegen_llvm/src/context.rs b/compiler/rustc_codegen_llvm/src/context.rs index d8fbe51b975..79381f35a3c 100644 --- a/compiler/rustc_codegen_llvm/src/context.rs +++ b/compiler/rustc_codegen_llvm/src/context.rs @@ -1,11 +1,13 @@ use std::borrow::Borrow; use std::cell::{Cell, RefCell}; use std::ffi::{CStr, c_char, c_uint}; +use std::ops::Deref; use std::str; use rustc_abi::{HasDataLayout, 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}; @@ -30,23 +32,46 @@ use smallvec::SmallVec; use crate::back::write::to_llvm_code_model; use crate::callee::get_fn; -use crate::common::AsCCharPtr; +use crate::common::{self, AsCCharPtr}; use crate::debuginfo::metadata::apply_vcall_visibility_metadata; use crate::llvm::{Metadata, MetadataType}; use crate::type_::Type; use crate::value::Value; use crate::{attributes, 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 +/// other compiler pieces. Here we define a small subset which has enough information and can be +/// moved around more freely. +pub(crate) struct SimpleCx<'ll> { + pub llmod: &'ll llvm::Module, + pub llcx: &'ll llvm::Context, +} + +impl<'ll> Borrow<SimpleCx<'ll>> for CodegenCx<'ll, '_> { + fn borrow(&self) -> &SimpleCx<'ll> { + &self.scx + } +} + +impl<'ll, 'tcx> Deref for CodegenCx<'ll, 'tcx> { + type Target = SimpleCx<'ll>; + + #[inline] + fn deref(&self) -> &Self::Target { + &self.scx + } +} + /// There is one `CodegenCx` per codegen unit. Each one has its own LLVM /// `llvm::Context` so that several codegen units may be processed in parallel. /// All other LLVM data structures in the `CodegenCx` are tied to that `llvm::Context`. pub(crate) struct CodegenCx<'ll, 'tcx> { pub tcx: TyCtxt<'tcx>, + pub scx: SimpleCx<'ll>, pub use_dll_storage_attrs: bool, pub tls_model: llvm::ThreadLocalMode, - pub llmod: &'ll llvm::Module, - pub llcx: &'ll llvm::Context, pub codegen_unit: &'tcx CodegenUnit<'tcx>, /// Cache instances of monomorphic and polymorphic items @@ -553,10 +578,9 @@ impl<'ll, 'tcx> CodegenCx<'ll, 'tcx> { CodegenCx { tcx, + scx: SimpleCx { llcx, llmod }, use_dll_storage_attrs, tls_model, - llmod, - llcx, codegen_unit, instances: Default::default(), vtables: Default::default(), @@ -600,6 +624,11 @@ impl<'ll, 'tcx> CodegenCx<'ll, 'tcx> { llvm::set_section(g, c"llvm.metadata"); } } +} +impl<'ll> SimpleCx<'ll> { + pub(crate) fn val_ty(&self, v: &'ll Value) -> &'ll Type { + common::val_ty(v) + } pub(crate) fn get_metadata_value(&self, metadata: &'ll Metadata) -> &'ll Value { unsafe { llvm::LLVMMetadataAsValue(self.llcx, metadata) } @@ -625,6 +654,10 @@ impl<'ll, 'tcx> CodegenCx<'ll, 'tcx> { llvm::LLVMMDStringInContext2(self.llcx, name.as_ptr() as *const c_char, name.len()) }) } + + pub(crate) fn type_kind(&self, ty: &'ll Type) -> TypeKind { + unsafe { llvm::LLVMRustGetTypeKind(ty).to_generic() } + } } impl<'ll, 'tcx> MiscCodegenMethods<'tcx> for CodegenCx<'ll, 'tcx> { @@ -741,14 +774,16 @@ impl<'ll, 'tcx> MiscCodegenMethods<'tcx> for CodegenCx<'ll, 'tcx> { if self.get_declared_value(entry_name).is_none() { Some(self.declare_entry_fn( entry_name, - self.sess().target.entry_abi.into(), + llvm::CallConv::from_conv( + self.sess().target.entry_abi, + self.sess().target.arch.borrow(), + ), llvm::UnnamedAddr::Global, fn_type, )) } else { // If the symbol already exists, it is an error: for example, the user wrote // #[no_mangle] extern "C" fn main(..) {..} - // instead of #[start] None } } @@ -1176,6 +1211,20 @@ impl CodegenCx<'_, '_> { } } +// This is a duplication of the set_metadata function above. However, so far it's the only one +// shared between both contexts, so it doesn't seem worth it to make the Cx generic like we did it +// for the Builder. +impl SimpleCx<'_> { + #[allow(unused)] + /// A wrapper for [`llvm::LLVMSetMetadata`], but it takes `Metadata` as a parameter instead of `Value`. + pub(crate) fn set_metadata<'a>(&self, val: &'a Value, kind_id: MetadataType, md: &'a Metadata) { + unsafe { + let node = llvm::LLVMMetadataAsValue(&self.llcx, md); + llvm::LLVMSetMetadata(val, kind_id as c_uint, node); + } + } +} + impl HasDataLayout for CodegenCx<'_, '_> { #[inline] fn data_layout(&self) -> &TargetDataLayout { diff --git a/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen.rs b/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen.rs index b3ad2a0e409..fd22421c7fc 100644 --- a/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen.rs +++ b/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen.rs @@ -9,6 +9,7 @@ use rustc_data_structures::fx::{FxHashSet, FxIndexMap}; use rustc_hir::def_id::{DefId, LocalDefId}; use rustc_index::IndexVec; use rustc_middle::mir; +use rustc_middle::mir::mono::MonoItemPartitions; use rustc_middle::ty::{self, TyCtxt}; use rustc_session::RemapFileNameExt; use rustc_session::config::RemapPathScopeComponents; @@ -297,12 +298,13 @@ struct UsageSets<'tcx> { /// Prepare sets of definitions that are relevant to deciding whether something /// is an "unused function" for coverage purposes. fn prepare_usage_sets<'tcx>(tcx: TyCtxt<'tcx>) -> UsageSets<'tcx> { - let (all_mono_items, cgus) = tcx.collect_and_partition_mono_items(()); + let MonoItemPartitions { all_mono_items, codegen_units } = + tcx.collect_and_partition_mono_items(()); // Obtain a MIR body for each function participating in codegen, via an // arbitrary instance. let mut def_ids_seen = FxHashSet::default(); - let def_and_mir_for_all_mono_fns = cgus + let def_and_mir_for_all_mono_fns = codegen_units .iter() .flat_map(|cgu| cgu.items().keys()) .filter_map(|item| match item { diff --git a/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen/covfun.rs b/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen/covfun.rs index 5428d776f41..460a4664615 100644 --- a/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen/covfun.rs +++ b/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen/covfun.rs @@ -51,7 +51,7 @@ pub(crate) fn prepare_covfun_record<'tcx>( is_used: bool, ) -> Option<CovfunRecord<'tcx>> { let fn_cov_info = tcx.instance_mir(instance.def).function_coverage_info.as_deref()?; - let ids_info = tcx.coverage_ids_info(instance.def); + let ids_info = tcx.coverage_ids_info(instance.def)?; let expressions = prepare_expressions(fn_cov_info, ids_info, is_used); diff --git a/compiler/rustc_codegen_llvm/src/coverageinfo/mod.rs b/compiler/rustc_codegen_llvm/src/coverageinfo/mod.rs index 7311cd9d230..021108cd51c 100644 --- a/compiler/rustc_codegen_llvm/src/coverageinfo/mod.rs +++ b/compiler/rustc_codegen_llvm/src/coverageinfo/mod.rs @@ -8,7 +8,6 @@ use rustc_codegen_ssa::traits::{ use rustc_data_structures::fx::{FxHashMap, FxIndexSet}; use rustc_middle::mir::coverage::CoverageKind; use rustc_middle::ty::Instance; -use rustc_middle::ty::layout::HasTyCtxt; use tracing::{debug, instrument}; use crate::builder::Builder; @@ -147,6 +146,10 @@ impl<'tcx> CoverageInfoBuilderMethods<'tcx> for Builder<'_, '_, 'tcx> { debug!("function has a coverage statement but no coverage info"); return; }; + let Some(ids_info) = bx.tcx.coverage_ids_info(instance.def) else { + debug!("function has a coverage statement but no IDs info"); + return; + }; // Mark the instance as used in this CGU, for coverage purposes. // This includes functions that were not partitioned into this CGU, @@ -162,8 +165,7 @@ impl<'tcx> CoverageInfoBuilderMethods<'tcx> for Builder<'_, '_, 'tcx> { // be smaller than the number originally inserted by the instrumentor, // if some high-numbered counters were removed by MIR optimizations. // If so, LLVM's profiler runtime will use fewer physical counters. - let num_counters = - bx.tcx().coverage_ids_info(instance.def).num_counters_after_mir_opts(); + let num_counters = ids_info.num_counters_after_mir_opts(); assert!( num_counters as usize <= function_coverage_info.num_counters, "num_counters disagreement: query says {num_counters} but function info only has {}", diff --git a/compiler/rustc_codegen_llvm/src/debuginfo/create_scope_map.rs b/compiler/rustc_codegen_llvm/src/debuginfo/create_scope_map.rs index e545ce386ed..11eb9651af6 100644 --- a/compiler/rustc_codegen_llvm/src/debuginfo/create_scope_map.rs +++ b/compiler/rustc_codegen_llvm/src/debuginfo/create_scope_map.rs @@ -9,7 +9,7 @@ use rustc_middle::mir::{Body, SourceScope}; use rustc_middle::ty::layout::{FnAbiOf, HasTypingEnv}; use rustc_middle::ty::{self, Instance}; use rustc_session::config::DebugInfo; -use rustc_span::{BytePos, hygiene}; +use rustc_span::{BytePos, DUMMY_SP, hygiene}; use super::metadata::file_metadata; use super::utils::DIB; @@ -85,23 +85,15 @@ fn make_mir_scope<'ll, 'tcx>( discriminators, parent, ); - if let Some(parent_scope) = debug_context.scopes[parent] { - parent_scope - } else { - // If the parent scope could not be represented then no children - // can be either. - debug_context.scopes[scope] = None; - instantiated.insert(scope); - return; - } + debug_context.scopes[parent] } else { // The root is the function itself. let file = cx.sess().source_map().lookup_source_file(mir.span.lo()); - debug_context.scopes[scope] = Some(DebugScope { + debug_context.scopes[scope] = DebugScope { file_start_pos: file.start_pos, file_end_pos: file.end_position(), - ..debug_context.scopes[scope].unwrap() - }); + ..debug_context.scopes[scope] + }; instantiated.insert(scope); return; }; @@ -112,7 +104,7 @@ fn make_mir_scope<'ll, 'tcx>( { // Do not create a DIScope if there are no variables defined in this // MIR `SourceScope`, and it's not `inlined`, to avoid debuginfo bloat. - debug_context.scopes[scope] = Some(parent_scope); + debug_context.scopes[scope] = parent_scope; instantiated.insert(scope); return; } @@ -145,14 +137,7 @@ fn make_mir_scope<'ll, 'tcx>( }, }; - let mut debug_scope = Some(DebugScope { - dbg_scope, - inlined_at: parent_scope.inlined_at, - file_start_pos: loc.file.start_pos, - file_end_pos: loc.file.end_position(), - }); - - if let Some((_, callsite_span)) = scope_data.inlined { + let inlined_at = scope_data.inlined.map(|(_, callsite_span)| { let callsite_span = hygiene::walk_chain_collapsed(callsite_span, mir.span); let callsite_scope = parent_scope.adjust_dbg_scope_for_span(cx, callsite_span); let loc = cx.dbg_loc(callsite_scope, parent_scope.inlined_at, callsite_span); @@ -175,29 +160,29 @@ fn make_mir_scope<'ll, 'tcx>( // Note further that we can't key this hashtable on the span itself, // because these spans could have distinct SyntaxContexts. We have // to key on exactly what we're giving to LLVM. - let inlined_at = match discriminators.entry(callsite_span.lo()) { + match discriminators.entry(callsite_span.lo()) { Entry::Occupied(mut o) => { *o.get_mut() += 1; + // NB: We have to emit *something* here or we'll fail LLVM IR verification + // in at least some circumstances (see issue #135322) so if the required + // discriminant cannot be encoded fall back to the dummy location. unsafe { llvm::LLVMRustDILocationCloneWithBaseDiscriminator(loc, *o.get()) } + .unwrap_or_else(|| { + cx.dbg_loc(callsite_scope, parent_scope.inlined_at, DUMMY_SP) + }) } Entry::Vacant(v) => { v.insert(0); - Some(loc) - } - }; - match inlined_at { - Some(inlined_at) => { - debug_scope.as_mut().unwrap().inlined_at = Some(inlined_at); - } - None => { - // LLVM has a maximum discriminator that it can encode (currently - // it uses 12 bits for 4096 possible values). If we exceed that - // there is little we can do but drop the debug info. - debug_scope = None; + loc } } - } + }); - debug_context.scopes[scope] = debug_scope; + debug_context.scopes[scope] = DebugScope { + dbg_scope, + inlined_at: inlined_at.or(parent_scope.inlined_at), + file_start_pos: loc.file.start_pos, + file_end_pos: loc.file.end_position(), + }; instantiated.insert(scope); } diff --git a/compiler/rustc_codegen_llvm/src/debuginfo/metadata.rs b/compiler/rustc_codegen_llvm/src/debuginfo/metadata.rs index 88e43e1c678..8d782a618fc 100644 --- a/compiler/rustc_codegen_llvm/src/debuginfo/metadata.rs +++ b/compiler/rustc_codegen_llvm/src/debuginfo/metadata.rs @@ -442,7 +442,7 @@ pub(crate) fn type_di_node<'ll, 'tcx>(cx: &CodegenCx<'ll, 'tcx>, t: Ty<'tcx>) -> // (or if there is no allocator argument). ty::Adt(def, args) if def.is_box() - && args.get(1).map_or(true, |arg| cx.layout_of(arg.expect_ty()).is_1zst()) => + && args.get(1).is_none_or(|arg| cx.layout_of(arg.expect_ty()).is_1zst()) => { build_pointer_or_reference_di_node(cx, t, t.expect_boxed_ty(), unique_type_id) } diff --git a/compiler/rustc_codegen_llvm/src/debuginfo/mod.rs b/compiler/rustc_codegen_llvm/src/debuginfo/mod.rs index 755f4816acf..e6778411365 100644 --- a/compiler/rustc_codegen_llvm/src/debuginfo/mod.rs +++ b/compiler/rustc_codegen_llvm/src/debuginfo/mod.rs @@ -295,12 +295,12 @@ impl<'ll, 'tcx> DebugInfoCodegenMethods<'tcx> for CodegenCx<'ll, 'tcx> { } // Initialize fn debug context (including scopes). - let empty_scope = Some(DebugScope { + let empty_scope = DebugScope { dbg_scope: self.dbg_scope_fn(instance, fn_abi, Some(llfn)), inlined_at: None, file_start_pos: BytePos(0), file_end_pos: BytePos(0), - }); + }; let mut fn_debug_context = FunctionDebugContext { scopes: IndexVec::from_elem(empty_scope, &mir.source_scopes), inlined_function_scopes: Default::default(), diff --git a/compiler/rustc_codegen_llvm/src/declare.rs b/compiler/rustc_codegen_llvm/src/declare.rs index 3ec386f6b07..bdc83267cca 100644 --- a/compiler/rustc_codegen_llvm/src/declare.rs +++ b/compiler/rustc_codegen_llvm/src/declare.rs @@ -21,26 +21,26 @@ use tracing::debug; use crate::abi::{FnAbi, FnAbiLlvmExt}; use crate::common::AsCCharPtr; -use crate::context::CodegenCx; +use crate::context::{CodegenCx, SimpleCx}; use crate::llvm::AttributePlace::Function; use crate::llvm::Visibility; use crate::type_::Type; use crate::value::Value; use crate::{attributes, llvm}; -/// Declare a function. +/// Declare a function with a SimpleCx. /// /// If there’s a value with the same name already declared, the function will /// update the declaration and return existing Value instead. -pub(crate) fn declare_raw_fn<'ll>( - cx: &CodegenCx<'ll, '_>, +pub(crate) fn declare_simple_fn<'ll>( + cx: &SimpleCx<'ll>, name: &str, callconv: llvm::CallConv, unnamed: llvm::UnnamedAddr, visibility: llvm::Visibility, ty: &'ll Type, ) -> &'ll Value { - debug!("declare_raw_fn(name={:?}, ty={:?})", name, ty); + debug!("declare_simple_fn(name={:?}, ty={:?})", name, ty); let llfn = unsafe { llvm::LLVMRustGetOrInsertFunction(cx.llmod, name.as_c_char_ptr(), name.len(), ty) }; @@ -49,6 +49,24 @@ pub(crate) fn declare_raw_fn<'ll>( llvm::SetUnnamedAddress(llfn, unnamed); llvm::set_visibility(llfn, visibility); + llfn +} + +/// Declare a function. +/// +/// If there’s a value with the same name already declared, the function will +/// update the declaration and return existing Value instead. +pub(crate) fn declare_raw_fn<'ll, 'tcx>( + cx: &CodegenCx<'ll, 'tcx>, + name: &str, + callconv: llvm::CallConv, + unnamed: llvm::UnnamedAddr, + visibility: llvm::Visibility, + ty: &'ll Type, +) -> &'ll Value { + debug!("declare_raw_fn(name={:?}, ty={:?})", name, ty); + let llfn = declare_simple_fn(cx, name, callconv, unnamed, visibility, ty); + let mut attrs = SmallVec::<[_; 4]>::new(); if cx.tcx.sess.opts.cg.no_redzone.unwrap_or(cx.tcx.sess.target.disable_redzone) { @@ -125,7 +143,7 @@ impl<'ll, 'tcx> CodegenCx<'ll, 'tcx> { let llfn = declare_raw_fn( self, name, - fn_abi.llvm_cconv(), + fn_abi.llvm_cconv(self), llvm::UnnamedAddr::Global, llvm::Visibility::Default, fn_abi.llvm_type(self), diff --git a/compiler/rustc_codegen_llvm/src/intrinsic.rs b/compiler/rustc_codegen_llvm/src/intrinsic.rs index cabcfc9b42b..eab4a9f30c9 100644 --- a/compiler/rustc_codegen_llvm/src/intrinsic.rs +++ b/compiler/rustc_codegen_llvm/src/intrinsic.rs @@ -1081,11 +1081,11 @@ fn codegen_emcc_try<'ll>( // Helper function to give a Block to a closure to codegen a shim function. // This is currently primarily used for the `try` intrinsic functions above. -fn gen_fn<'ll, 'tcx>( - cx: &CodegenCx<'ll, 'tcx>, +fn gen_fn<'a, 'll, 'tcx>( + cx: &'a CodegenCx<'ll, 'tcx>, name: &str, rust_fn_sig: ty::PolyFnSig<'tcx>, - codegen: &mut dyn FnMut(Builder<'_, 'll, 'tcx>), + codegen: &mut dyn FnMut(Builder<'a, 'll, 'tcx>), ) -> (&'ll Type, &'ll Value) { let fn_abi = cx.fn_abi_of_fn_ptr(rust_fn_sig, ty::List::empty()); let llty = fn_abi.llvm_type(cx); @@ -1104,9 +1104,9 @@ fn gen_fn<'ll, 'tcx>( // catch exceptions. // // This function is only generated once and is then cached. -fn get_rust_try_fn<'ll, 'tcx>( - cx: &CodegenCx<'ll, 'tcx>, - codegen: &mut dyn FnMut(Builder<'_, 'll, 'tcx>), +fn get_rust_try_fn<'a, 'll, 'tcx>( + cx: &'a CodegenCx<'ll, 'tcx>, + codegen: &mut dyn FnMut(Builder<'a, 'll, 'tcx>), ) -> (&'ll Type, &'ll Value) { if let Some(llfn) = cx.rust_try_fn.get() { return llfn; @@ -1182,6 +1182,60 @@ fn generic_simd_intrinsic<'ll, 'tcx>( }}; } + /// Returns the bitwidth of the `$ty` argument if it is an `Int` type. + macro_rules! require_int_ty { + ($ty: expr, $diag: expr) => { + match $ty { + ty::Int(i) => i.bit_width().unwrap_or_else(|| bx.data_layout().pointer_size.bits()), + _ => { + return_error!($diag); + } + } + }; + } + + /// Returns the bitwidth of the `$ty` argument if it is an `Int` or `Uint` type. + macro_rules! require_int_or_uint_ty { + ($ty: expr, $diag: expr) => { + match $ty { + ty::Int(i) => i.bit_width().unwrap_or_else(|| bx.data_layout().pointer_size.bits()), + ty::Uint(i) => { + i.bit_width().unwrap_or_else(|| bx.data_layout().pointer_size.bits()) + } + _ => { + return_error!($diag); + } + } + }; + } + + /// Converts a vector mask, where each element has a bit width equal to the data elements it is used with, + /// down to an i1 based mask that can be used by llvm intrinsics. + /// + /// The rust simd semantics are that each element should either consist of all ones or all zeroes, + /// but this information is not available to llvm. Truncating the vector effectively uses the lowest bit, + /// but codegen for several targets is better if we consider the highest bit by shifting. + /// + /// For x86 SSE/AVX targets this is beneficial since most instructions with mask parameters only consider the highest bit. + /// So even though on llvm level we have an additional shift, in the final assembly there is no shift or truncate and + /// instead the mask can be used as is. + /// + /// For aarch64 and other targets there is a benefit because a mask from the sign bit can be more + /// efficiently converted to an all ones / all zeroes mask by comparing whether each element is negative. + fn vector_mask_to_bitmask<'a, 'll, 'tcx>( + bx: &mut Builder<'a, 'll, 'tcx>, + i_xn: &'ll Value, + in_elem_bitwidth: u64, + in_len: u64, + ) -> &'ll Value { + // Shift the MSB to the right by "in_elem_bitwidth - 1" into the first bit position. + let shift_idx = bx.cx.const_int(bx.type_ix(in_elem_bitwidth), (in_elem_bitwidth - 1) as _); + let shift_indices = vec![shift_idx; in_len as _]; + let i_xn_msb = bx.lshr(i_xn, bx.const_vector(shift_indices.as_slice())); + // Truncate vector to an <i1 x N> + bx.trunc(i_xn_msb, bx.type_vector(bx.type_i1(), in_len)) + } + let tcx = bx.tcx(); let sig = tcx.normalize_erasing_late_bound_regions(bx.typing_env(), callee_ty.fn_sig(tcx)); let arg_tys = sig.inputs(); @@ -1433,14 +1487,13 @@ fn generic_simd_intrinsic<'ll, 'tcx>( m_len, v_len }); - match m_elem_ty.kind() { - ty::Int(_) => {} - _ => return_error!(InvalidMonomorphization::MaskType { span, name, ty: m_elem_ty }), - } - // truncate the mask to a vector of i1s - let i1 = bx.type_i1(); - let i1xn = bx.type_vector(i1, m_len as u64); - let m_i1s = bx.trunc(args[0].immediate(), i1xn); + let in_elem_bitwidth = + require_int_ty!(m_elem_ty.kind(), InvalidMonomorphization::MaskType { + span, + name, + ty: m_elem_ty + }); + let m_i1s = vector_mask_to_bitmask(bx, args[0].immediate(), in_elem_bitwidth, m_len); return Ok(bx.select(m_i1s, args[1].immediate(), args[2].immediate())); } @@ -1457,33 +1510,15 @@ fn generic_simd_intrinsic<'ll, 'tcx>( let expected_bytes = in_len.div_ceil(8); // Integer vector <i{in_bitwidth} x in_len>: - let (i_xn, in_elem_bitwidth) = match in_elem.kind() { - ty::Int(i) => ( - args[0].immediate(), - i.bit_width().unwrap_or_else(|| bx.data_layout().pointer_size.bits()), - ), - ty::Uint(i) => ( - args[0].immediate(), - i.bit_width().unwrap_or_else(|| bx.data_layout().pointer_size.bits()), - ), - _ => return_error!(InvalidMonomorphization::VectorArgument { + let in_elem_bitwidth = + require_int_or_uint_ty!(in_elem.kind(), InvalidMonomorphization::VectorArgument { span, name, in_ty, in_elem - }), - }; + }); - // LLVM doesn't always know the inputs are `0` or `!0`, so we shift here so it optimizes to - // `pmovmskb` and similar on x86. - let shift_indices = - vec![ - bx.cx.const_int(bx.type_ix(in_elem_bitwidth), (in_elem_bitwidth - 1) as _); - in_len as _ - ]; - let i_xn_msb = bx.lshr(i_xn, bx.const_vector(shift_indices.as_slice())); - // Truncate vector to an <i1 x N> - let i1xn = bx.trunc(i_xn_msb, bx.type_vector(bx.type_i1(), in_len)); + let i1xn = vector_mask_to_bitmask(bx, args[0].immediate(), in_elem_bitwidth, in_len); // Bitcast <i1 x N> to iN: let i_ = bx.bitcast(i1xn, bx.type_ix(in_len)); @@ -1704,28 +1739,21 @@ fn generic_simd_intrinsic<'ll, 'tcx>( } ); - match element_ty2.kind() { - ty::Int(_) => (), - _ => { - return_error!(InvalidMonomorphization::ThirdArgElementType { - span, - name, - expected_element: element_ty2, - third_arg: arg_tys[2] - }); - } - } + let mask_elem_bitwidth = + require_int_ty!(element_ty2.kind(), InvalidMonomorphization::ThirdArgElementType { + span, + name, + expected_element: element_ty2, + third_arg: arg_tys[2] + }); // Alignment of T, must be a constant integer value: let alignment_ty = bx.type_i32(); let alignment = bx.const_i32(bx.align_of(in_elem).bytes() as i32); // Truncate the mask vector to a vector of i1s: - let (mask, mask_ty) = { - let i1 = bx.type_i1(); - let i1xn = bx.type_vector(i1, in_len); - (bx.trunc(args[2].immediate(), i1xn), i1xn) - }; + let mask = vector_mask_to_bitmask(bx, args[2].immediate(), mask_elem_bitwidth, in_len); + let mask_ty = bx.type_vector(bx.type_i1(), in_len); // Type of the vector of pointers: let llvm_pointer_vec_ty = llvm_vector_ty(bx, element_ty1, in_len); @@ -1810,27 +1838,21 @@ fn generic_simd_intrinsic<'ll, 'tcx>( } ); - require!( - matches!(mask_elem.kind(), ty::Int(_)), - InvalidMonomorphization::ThirdArgElementType { + let m_elem_bitwidth = + require_int_ty!(mask_elem.kind(), InvalidMonomorphization::ThirdArgElementType { span, name, expected_element: values_elem, third_arg: mask_ty, - } - ); + }); + + let mask = vector_mask_to_bitmask(bx, args[0].immediate(), m_elem_bitwidth, mask_len); + let mask_ty = bx.type_vector(bx.type_i1(), mask_len); // Alignment of T, must be a constant integer value: let alignment_ty = bx.type_i32(); let alignment = bx.const_i32(bx.align_of(values_elem).bytes() as i32); - // Truncate the mask vector to a vector of i1s: - let (mask, mask_ty) = { - let i1 = bx.type_i1(); - let i1xn = bx.type_vector(i1, mask_len); - (bx.trunc(args[0].immediate(), i1xn), i1xn) - }; - let llvm_pointer = bx.type_ptr(); // Type of the vector of elements: @@ -1901,27 +1923,21 @@ fn generic_simd_intrinsic<'ll, 'tcx>( } ); - require!( - matches!(mask_elem.kind(), ty::Int(_)), - InvalidMonomorphization::ThirdArgElementType { + let m_elem_bitwidth = + require_int_ty!(mask_elem.kind(), InvalidMonomorphization::ThirdArgElementType { span, name, expected_element: values_elem, third_arg: mask_ty, - } - ); + }); + + let mask = vector_mask_to_bitmask(bx, args[0].immediate(), m_elem_bitwidth, mask_len); + let mask_ty = bx.type_vector(bx.type_i1(), mask_len); // Alignment of T, must be a constant integer value: let alignment_ty = bx.type_i32(); let alignment = bx.const_i32(bx.align_of(values_elem).bytes() as i32); - // Truncate the mask vector to a vector of i1s: - let (mask, mask_ty) = { - let i1 = bx.type_i1(); - let i1xn = bx.type_vector(i1, in_len); - (bx.trunc(args[0].immediate(), i1xn), i1xn) - }; - let ret_t = bx.type_void(); let llvm_pointer = bx.type_ptr(); @@ -1995,28 +2011,21 @@ fn generic_simd_intrinsic<'ll, 'tcx>( ); // The element type of the third argument must be a signed integer type of any width: - match element_ty2.kind() { - ty::Int(_) => (), - _ => { - return_error!(InvalidMonomorphization::ThirdArgElementType { - span, - name, - expected_element: element_ty2, - third_arg: arg_tys[2] - }); - } - } + let mask_elem_bitwidth = + require_int_ty!(element_ty2.kind(), InvalidMonomorphization::ThirdArgElementType { + span, + name, + expected_element: element_ty2, + third_arg: arg_tys[2] + }); // Alignment of T, must be a constant integer value: let alignment_ty = bx.type_i32(); let alignment = bx.const_i32(bx.align_of(in_elem).bytes() as i32); // Truncate the mask vector to a vector of i1s: - let (mask, mask_ty) = { - let i1 = bx.type_i1(); - let i1xn = bx.type_vector(i1, in_len); - (bx.trunc(args[2].immediate(), i1xn), i1xn) - }; + let mask = vector_mask_to_bitmask(bx, args[2].immediate(), mask_elem_bitwidth, in_len); + let mask_ty = bx.type_vector(bx.type_i1(), in_len); let ret_t = bx.type_void(); @@ -2164,8 +2173,13 @@ fn generic_simd_intrinsic<'ll, 'tcx>( }); args[0].immediate() } else { - match in_elem.kind() { - ty::Int(_) | ty::Uint(_) => {} + let bitwidth = match in_elem.kind() { + ty::Int(i) => { + i.bit_width().unwrap_or_else(|| bx.data_layout().pointer_size.bits()) + } + ty::Uint(i) => { + i.bit_width().unwrap_or_else(|| bx.data_layout().pointer_size.bits()) + } _ => return_error!(InvalidMonomorphization::UnsupportedSymbol { span, name, @@ -2174,12 +2188,9 @@ fn generic_simd_intrinsic<'ll, 'tcx>( in_elem, ret_ty }), - } + }; - // boolean reductions operate on vectors of i1s: - let i1 = bx.type_i1(); - let i1xn = bx.type_vector(i1, in_len as u64); - bx.trunc(args[0].immediate(), i1xn) + vector_mask_to_bitmask(bx, args[0].immediate(), bitwidth, in_len as _) }; return match in_elem.kind() { ty::Int(_) | ty::Uint(_) => { diff --git a/compiler/rustc_codegen_llvm/src/lib.rs b/compiler/rustc_codegen_llvm/src/lib.rs index 06afe8bb3ad..4a84fd29e44 100644 --- a/compiler/rustc_codegen_llvm/src/lib.rs +++ b/compiler/rustc_codegen_llvm/src/lib.rs @@ -237,7 +237,6 @@ impl WriteBackendMethods for LlvmCodegenBackend { /// Generate autodiff rules fn autodiff( cgcx: &CodegenContext<Self>, - tcx: TyCtxt<'_>, module: &ModuleCodegen<Self::Module>, diff_fncs: Vec<AutoDiffItem>, config: &ModuleConfig, @@ -246,7 +245,7 @@ impl WriteBackendMethods for LlvmCodegenBackend { let dcx = cgcx.create_dcx(); return Err(dcx.handle().emit_almost_fatal(AutoDiffWithoutLTO)); } - builder::autodiff::differentiate(module, cgcx, tcx, diff_fncs, config) + builder::autodiff::differentiate(module, cgcx, diff_fncs, config) } } diff --git a/compiler/rustc_codegen_llvm/src/llvm/ffi.rs b/compiler/rustc_codegen_llvm/src/llvm/ffi.rs index cb4a8c9a5f2..009d15a932f 100644 --- a/compiler/rustc_codegen_llvm/src/llvm/ffi.rs +++ b/compiler/rustc_codegen_llvm/src/llvm/ffi.rs @@ -120,6 +120,7 @@ pub enum CallConv { X86_Intr = 83, AvrNonBlockingInterrupt = 84, AvrInterrupt = 85, + AmdgpuKernel = 91, } /// Must match the layout of `LLVMLinkage`. @@ -740,8 +741,11 @@ pub mod debuginfo { pub type DIEnumerator = DIDescriptor; pub type DITemplateTypeParameter = DIDescriptor; - // These values **must** match with LLVMRustDIFlags!! bitflags! { + /// Must match the layout of `LLVMDIFlags` in the LLVM-C API. + /// + /// Each value declared here must also be covered by the static + /// assertions in `RustWrapper.cpp` used by `fromRust(LLVMDIFlags)`. #[repr(transparent)] #[derive(Clone, Copy, Default)] pub struct DIFlags: u32 { @@ -751,7 +755,7 @@ pub mod debuginfo { const FlagPublic = 3; const FlagFwdDecl = (1 << 2); const FlagAppleBlock = (1 << 3); - const FlagBlockByrefStruct = (1 << 4); + const FlagReservedBit4 = (1 << 4); const FlagVirtual = (1 << 5); const FlagArtificial = (1 << 6); const FlagExplicit = (1 << 7); @@ -762,10 +766,21 @@ pub mod debuginfo { const FlagStaticMember = (1 << 12); const FlagLValueReference = (1 << 13); const FlagRValueReference = (1 << 14); - const FlagExternalTypeRef = (1 << 15); + const FlagReserved = (1 << 15); + const FlagSingleInheritance = (1 << 16); + const FlagMultipleInheritance = (2 << 16); + const FlagVirtualInheritance = (3 << 16); const FlagIntroducedVirtual = (1 << 18); const FlagBitField = (1 << 19); const FlagNoReturn = (1 << 20); + // The bit at (1 << 21) is unused, but was `LLVMDIFlagMainSubprogram`. + const FlagTypePassByValue = (1 << 22); + const FlagTypePassByReference = (1 << 23); + const FlagEnumClass = (1 << 24); + const FlagThunk = (1 << 25); + const FlagNonTrivial = (1 << 26); + const FlagBigEndian = (1 << 27); + const FlagLittleEndian = (1 << 28); } } @@ -917,6 +932,7 @@ unsafe extern "C" { pub fn LLVMMetadataTypeInContext(C: &Context) -> &Type; // Operations on all values + pub fn LLVMIsUndef(Val: &Value) -> Bool; pub fn LLVMTypeOf(Val: &Value) -> &Type; pub fn LLVMGetValueName2(Val: &Value, Length: *mut size_t) -> *const c_char; pub fn LLVMSetValueName2(Val: &Value, Name: *const c_char, NameLen: size_t); diff --git a/compiler/rustc_codegen_llvm/src/type_.rs b/compiler/rustc_codegen_llvm/src/type_.rs index 6aec078e0de..c56ad886120 100644 --- a/compiler/rustc_codegen_llvm/src/type_.rs +++ b/compiler/rustc_codegen_llvm/src/type_.rs @@ -11,7 +11,7 @@ use rustc_middle::ty::{self, Ty}; use rustc_target::callconv::{CastTarget, FnAbi, Reg}; use crate::abi::{FnAbiLlvmExt, LlvmType}; -use crate::context::CodegenCx; +use crate::context::{CodegenCx, SimpleCx}; pub(crate) use crate::llvm::Type; use crate::llvm::{Bool, False, Metadata, True}; use crate::type_of::LayoutLlvmExt; @@ -35,7 +35,8 @@ impl fmt::Debug for Type { } } -impl<'ll> CodegenCx<'ll, '_> { +impl<'ll> CodegenCx<'ll, '_> {} +impl<'ll> SimpleCx<'ll> { pub(crate) fn type_named_struct(&self, name: &str) -> &'ll Type { let name = SmallCStr::new(name); unsafe { llvm::LLVMStructCreateNamed(self.llcx, name.as_ptr()) } @@ -44,11 +45,9 @@ impl<'ll> CodegenCx<'ll, '_> { pub(crate) fn set_struct_body(&self, ty: &'ll Type, els: &[&'ll Type], packed: bool) { unsafe { llvm::LLVMStructSetBody(ty, els.as_ptr(), els.len() as c_uint, packed as Bool) } } - pub(crate) fn type_void(&self) -> &'ll Type { unsafe { llvm::LLVMVoidTypeInContext(self.llcx) } } - pub(crate) fn type_token(&self) -> &'ll Type { unsafe { llvm::LLVMTokenTypeInContext(self.llcx) } } @@ -75,7 +74,8 @@ impl<'ll> CodegenCx<'ll, '_> { args } } - +} +impl<'ll, 'tcx> CodegenCx<'ll, 'tcx> { pub(crate) fn type_bool(&self) -> &'ll Type { self.type_i8() } @@ -120,7 +120,8 @@ impl<'ll> CodegenCx<'ll, '_> { assert_eq!(size % unit_size, 0); self.type_array(self.type_from_integer(unit), size / unit_size) } - +} +impl<'ll> SimpleCx<'ll> { pub(crate) fn type_variadic_func(&self, args: &[&'ll Type], ret: &'ll Type) -> &'ll Type { unsafe { llvm::LLVMFunctionType(ret, args.as_ptr(), args.len() as c_uint, True) } } |