diff options
Diffstat (limited to 'compiler/rustc_codegen_llvm/src')
41 files changed, 2200 insertions, 1566 deletions
diff --git a/compiler/rustc_codegen_llvm/src/abi.rs b/compiler/rustc_codegen_llvm/src/abi.rs index 8c75125e009..8294e29d07d 100644 --- a/compiler/rustc_codegen_llvm/src/abi.rs +++ b/compiler/rustc_codegen_llvm/src/abi.rs @@ -17,14 +17,13 @@ use rustc_target::callconv::{ use rustc_target::spec::SanitizerSet; use smallvec::SmallVec; -use crate::attributes::llfn_attrs_from_instance; +use crate::attributes::{self, llfn_attrs_from_instance}; use crate::builder::Builder; use crate::context::CodegenCx; use crate::llvm::{self, Attribute, AttributePlace}; use crate::type_::Type; use crate::type_of::LayoutLlvmExt; use crate::value::Value; -use crate::{attributes, llvm_util}; trait ArgAttributesExt { fn apply_attrs_to_llfn(&self, idx: AttributePlace, cx: &CodegenCx<'_, '_>, llfn: &Value); @@ -437,7 +436,6 @@ impl<'ll, 'tcx> FnAbiLlvmExt<'ll, 'tcx> for FnAbi<'tcx, Ty<'tcx>> { let apply_range_attr = |idx: AttributePlace, scalar: rustc_abi::Scalar| { if cx.sess().opts.optimize != config::OptLevel::No - && llvm_util::get_version() >= (19, 0, 0) && matches!(scalar.primitive(), Primitive::Int(..)) // If the value is a boolean, the range is 0..2 and that ultimately // become 0..0 when the type becomes i1, which would be rejected @@ -571,19 +569,6 @@ impl<'ll, 'tcx> FnAbiLlvmExt<'ll, 'tcx> for FnAbi<'tcx, Ty<'tcx>> { } _ => {} } - if bx.cx.sess().opts.optimize != config::OptLevel::No - && llvm_util::get_version() < (19, 0, 0) - && let BackendRepr::Scalar(scalar) = self.ret.layout.backend_repr - && matches!(scalar.primitive(), Primitive::Int(..)) - // If the value is a boolean, the range is 0..2 and that ultimately - // become 0..0 when the type becomes i1, which would be rejected - // by the LLVM verifier. - && !scalar.is_bool() - // LLVM also rejects full range. - && !scalar.is_always_valid(bx) - { - bx.range_metadata(callsite, scalar.valid_range(bx)); - } for arg in self.args.iter() { match &arg.mode { PassMode::Ignore => {} @@ -654,7 +639,7 @@ impl<'ll, 'tcx> FnAbiLlvmExt<'ll, 'tcx> for FnAbi<'tcx, Ty<'tcx>> { } } -impl<'tcx> AbiBuilderMethods<'tcx> for Builder<'_, '_, 'tcx> { +impl AbiBuilderMethods for Builder<'_, '_, '_> { fn get_param(&mut self, index: usize) -> Self::Value { llvm::get_param(self.llfn(), index as c_uint) } diff --git a/compiler/rustc_codegen_llvm/src/allocator.rs b/compiler/rustc_codegen_llvm/src/allocator.rs index 66723cbf882..4a78e694979 100644 --- a/compiler/rustc_codegen_llvm/src/allocator.rs +++ b/compiler/rustc_codegen_llvm/src/allocator.rs @@ -3,33 +3,32 @@ use rustc_ast::expand::allocator::{ ALLOCATOR_METHODS, AllocatorKind, AllocatorTy, NO_ALLOC_SHIM_IS_UNSTABLE, alloc_error_handler_name, default_fn_name, global_fn_name, }; +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::common::AsCCharPtr; -use crate::llvm::{self, Context, False, Module, True, Type}; -use crate::{ModuleLlvm, attributes, debuginfo}; +use crate::builder::SBuilder; +use crate::declare::declare_simple_fn; +use crate::llvm::{self, False, True, Type}; +use crate::{SimpleCx, attributes, debuginfo}; pub(crate) unsafe fn codegen( tcx: TyCtxt<'_>, - module_llvm: &mut ModuleLlvm, + cx: SimpleCx<'_>, module_name: &str, kind: AllocatorKind, alloc_error_handler_kind: AllocatorKind, ) { - let llcx = &*module_llvm.llcx; - let llmod = module_llvm.llmod(); - let usize = unsafe { - match tcx.sess.target.pointer_width { - 16 => llvm::LLVMInt16TypeInContext(llcx), - 32 => llvm::LLVMInt32TypeInContext(llcx), - 64 => llvm::LLVMInt64TypeInContext(llcx), - tws => bug!("Unsupported target word size for int: {}", tws), - } + let usize = match tcx.sess.target.pointer_width { + 16 => cx.type_i16(), + 32 => cx.type_i32(), + 64 => cx.type_i64(), + tws => bug!("Unsupported target word size for int: {}", tws), }; - let i8 = unsafe { llvm::LLVMInt8TypeInContext(llcx) }; - let i8p = unsafe { llvm::LLVMPointerTypeInContext(llcx, 0) }; + let i8 = cx.type_i8(); + let i8p = cx.type_ptr(); if kind == AllocatorKind::Default { for method in ALLOCATOR_METHODS { @@ -55,20 +54,19 @@ 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, llcx, llmod, &from_name, &to_name, &args, output, false); + create_wrapper_function(tcx, &cx, &from_name, &to_name, &args, output, false); } } // rust alloc error handler create_wrapper_function( tcx, - llcx, - llmod, - "__rust_alloc_error_handler", - alloc_error_handler_name(alloc_error_handler_kind), + &cx, + &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, @@ -76,22 +74,22 @@ pub(crate) unsafe fn codegen( unsafe { // __rust_alloc_error_handler_should_panic - let name = OomStrategy::SYMBOL; - let ll_g = llvm::LLVMRustGetOrInsertGlobal(llmod, name.as_c_char_ptr(), name.len(), 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 = llvm::LLVMRustGetOrInsertGlobal(llmod, name.as_c_char_ptr(), name.len(), 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); } if tcx.sess.opts.debuginfo != DebugInfo::None { - let dbg_cx = debuginfo::CodegenUnitDebugContext::new(llmod); + let dbg_cx = debuginfo::CodegenUnitDebugContext::new(cx.llmod); debuginfo::metadata::build_compile_unit_di_node(tcx, module_name, &dbg_cx); dbg_cx.finalize(tcx.sess); } @@ -99,77 +97,64 @@ pub(crate) unsafe fn codegen( fn create_wrapper_function( tcx: TyCtxt<'_>, - llcx: &Context, - llmod: &Module, + cx: &SimpleCx<'_>, from_name: &str, to_name: &str, args: &[&Type], output: Option<&Type>, no_return: bool, ) { - unsafe { - let ty = llvm::LLVMFunctionType( - output.unwrap_or_else(|| llvm::LLVMVoidTypeInContext(llcx)), - args.as_ptr(), - args.len() as c_uint, - False, - ); - let llfn = llvm::LLVMRustGetOrInsertFunction( - llmod, - from_name.as_c_char_ptr(), - from_name.len(), - ty, - ); - let no_return = if no_return { - // -> ! DIFlagNoReturn - let no_return = llvm::AttributeKind::NoReturn.create_attr(llcx); - attributes::apply_to_llfn(llfn, llvm::AttributePlace::Function, &[no_return]); - Some(no_return) - } else { - None - }; - - llvm::set_visibility(llfn, llvm::Visibility::from_generic(tcx.sess.default_visibility())); - - if tcx.sess.must_emit_unwind_tables() { - let uwtable = - attributes::uwtable_attr(llcx, tcx.sess.opts.unstable_opts.use_sync_unwind); - attributes::apply_to_llfn(llfn, llvm::AttributePlace::Function, &[uwtable]); - } + let ty = cx.type_func(args, output.unwrap_or_else(|| cx.type_void())); + let llfn = declare_simple_fn( + &cx, + from_name, + llvm::CallConv::CCallConv, + llvm::UnnamedAddr::Global, + llvm::Visibility::from_generic(tcx.sess.default_visibility()), + ty, + ); + let no_return = if no_return { + // -> ! DIFlagNoReturn + let no_return = llvm::AttributeKind::NoReturn.create_attr(cx.llcx); + attributes::apply_to_llfn(llfn, llvm::AttributePlace::Function, &[no_return]); + Some(no_return) + } else { + None + }; - let callee = - llvm::LLVMRustGetOrInsertFunction(llmod, to_name.as_c_char_ptr(), to_name.len(), ty); - if let Some(no_return) = no_return { - // -> ! DIFlagNoReturn - attributes::apply_to_llfn(callee, llvm::AttributePlace::Function, &[no_return]); - } - llvm::set_visibility(callee, llvm::Visibility::Hidden); - - let llbb = llvm::LLVMAppendBasicBlockInContext(llcx, llfn, c"entry".as_ptr()); - - let llbuilder = llvm::LLVMCreateBuilderInContext(llcx); - llvm::LLVMPositionBuilderAtEnd(llbuilder, llbb); - let args = args - .iter() - .enumerate() - .map(|(i, _)| llvm::LLVMGetParam(llfn, i as c_uint)) - .collect::<Vec<_>>(); - let ret = llvm::LLVMBuildCallWithOperandBundles( - llbuilder, - ty, - callee, - args.as_ptr(), - args.len() as c_uint, - [].as_ptr(), - 0 as c_uint, - c"".as_ptr(), - ); - llvm::LLVMSetTailCall(ret, True); - if output.is_some() { - llvm::LLVMBuildRet(llbuilder, ret); - } else { - llvm::LLVMBuildRetVoid(llbuilder); - } - llvm::LLVMDisposeBuilder(llbuilder); + if tcx.sess.must_emit_unwind_tables() { + let uwtable = + attributes::uwtable_attr(cx.llcx, tcx.sess.opts.unstable_opts.use_sync_unwind); + attributes::apply_to_llfn(llfn, llvm::AttributePlace::Function, &[uwtable]); + } + + let callee = declare_simple_fn( + &cx, + to_name, + llvm::CallConv::CCallConv, + llvm::UnnamedAddr::Global, + llvm::Visibility::Hidden, + ty, + ); + if let Some(no_return) = no_return { + // -> ! DIFlagNoReturn + attributes::apply_to_llfn(callee, llvm::AttributePlace::Function, &[no_return]); + } + llvm::set_visibility(callee, llvm::Visibility::Hidden); + + let llbb = unsafe { llvm::LLVMAppendBasicBlockInContext(cx.llcx, llfn, c"entry".as_ptr()) }; + + let mut bx = SBuilder::build(&cx, llbb); + let args = args + .iter() + .enumerate() + .map(|(i, _)| llvm::get_param(llfn, i as c_uint)) + .collect::<Vec<_>>(); + let ret = bx.call(ty, callee, &args, None); + llvm::LLVMSetTailCall(ret, True); + if output.is_some() { + bx.ret(ret); + } else { + bx.ret_void() } } diff --git a/compiler/rustc_codegen_llvm/src/asm.rs b/compiler/rustc_codegen_llvm/src/asm.rs index be5673eddf9..e481b99afcc 100644 --- a/compiler/rustc_codegen_llvm/src/asm.rs +++ b/compiler/rustc_codegen_llvm/src/asm.rs @@ -1,6 +1,5 @@ use std::assert_matches::assert_matches; -use libc::{c_char, c_uint}; use rustc_abi::{BackendRepr, Float, Integer, Primitive, Scalar}; use rustc_ast::{InlineAsmOptions, InlineAsmTemplatePiece}; use rustc_codegen_ssa::mir::operand::OperandValue; @@ -377,7 +376,7 @@ impl<'ll, 'tcx> AsmBuilderMethods<'tcx> for Builder<'_, 'll, 'tcx> { impl<'tcx> AsmCodegenMethods<'tcx> for CodegenCx<'_, 'tcx> { fn codegen_global_asm( - &self, + &mut self, template: &[InlineAsmTemplatePiece], operands: &[GlobalAsmOperandRef<'tcx>], options: InlineAsmOptions, @@ -483,12 +482,13 @@ pub(crate) fn inline_asm_call<'ll>( debug!("Asm Output Type: {:?}", output); let fty = bx.cx.type_func(&argtys, output); - unsafe { - // Ask LLVM to verify that the constraints are well-formed. - let constraints_ok = llvm::LLVMRustInlineAsmVerify(fty, cons.as_c_char_ptr(), cons.len()); - debug!("constraint verification result: {:?}", constraints_ok); - if constraints_ok { - let v = llvm::LLVMRustInlineAsm( + // Ask LLVM to verify that the constraints are well-formed. + let constraints_ok = + unsafe { llvm::LLVMRustInlineAsmVerify(fty, cons.as_c_char_ptr(), cons.len()) }; + debug!("constraint verification result: {:?}", constraints_ok); + if constraints_ok { + let v = unsafe { + llvm::LLVMRustInlineAsm( fty, asm.as_c_char_ptr(), asm.len(), @@ -498,54 +498,50 @@ pub(crate) fn inline_asm_call<'ll>( alignstack, dia, can_throw, - ); - - let call = if !labels.is_empty() { - assert!(catch_funclet.is_none()); - bx.callbr(fty, None, None, v, inputs, dest.unwrap(), labels, None, None) - } else if let Some((catch, funclet)) = catch_funclet { - bx.invoke(fty, None, None, v, inputs, dest.unwrap(), catch, funclet, None) - } else { - bx.call(fty, None, None, v, inputs, None, None) - }; + ) + }; - // Store mark in a metadata node so we can map LLVM errors - // back to source locations. See #17552. - let key = "srcloc"; - let kind = llvm::LLVMGetMDKindIDInContext( - bx.llcx, - key.as_ptr().cast::<c_char>(), - key.len() as c_uint, - ); + let call = if !labels.is_empty() { + assert!(catch_funclet.is_none()); + bx.callbr(fty, None, None, v, inputs, dest.unwrap(), labels, None, None) + } else if let Some((catch, funclet)) = catch_funclet { + bx.invoke(fty, None, None, v, inputs, dest.unwrap(), catch, funclet, None) + } else { + bx.call(fty, None, None, v, inputs, None, None) + }; - // `srcloc` contains one 64-bit integer for each line of assembly code, - // where the lower 32 bits hold the lo byte position and the upper 32 bits - // hold the hi byte position. - let mut srcloc = vec![]; - if dia == llvm::AsmDialect::Intel && line_spans.len() > 1 { - // LLVM inserts an extra line to add the ".intel_syntax", so add - // a dummy srcloc entry for it. - // - // Don't do this if we only have 1 line span since that may be - // due to the asm template string coming from a macro. LLVM will - // default to the first srcloc for lines that don't have an - // associated srcloc. - srcloc.push(llvm::LLVMValueAsMetadata(bx.const_u64(0))); - } - srcloc.extend(line_spans.iter().map(|span| { - llvm::LLVMValueAsMetadata(bx.const_u64( - u64::from(span.lo().to_u32()) | (u64::from(span.hi().to_u32()) << 32), - )) - })); - let md = llvm::LLVMMDNodeInContext2(bx.llcx, srcloc.as_ptr(), srcloc.len()); - let md = llvm::LLVMMetadataAsValue(&bx.llcx, md); - llvm::LLVMSetMetadata(call, kind, md); + // Store mark in a metadata node so we can map LLVM errors + // back to source locations. See #17552. + let key = "srcloc"; + let kind = bx.get_md_kind_id(key); - Some(call) - } else { - // LLVM has detected an issue with our constraints, bail out - None + // `srcloc` contains one 64-bit integer for each line of assembly code, + // where the lower 32 bits hold the lo byte position and the upper 32 bits + // hold the hi byte position. + let mut srcloc = vec![]; + if dia == llvm::AsmDialect::Intel && line_spans.len() > 1 { + // LLVM inserts an extra line to add the ".intel_syntax", so add + // a dummy srcloc entry for it. + // + // Don't do this if we only have 1 line span since that may be + // due to the asm template string coming from a macro. LLVM will + // default to the first srcloc for lines that don't have an + // associated srcloc. + srcloc.push(llvm::LLVMValueAsMetadata(bx.const_u64(0))); } + srcloc.extend(line_spans.iter().map(|span| { + llvm::LLVMValueAsMetadata( + bx.const_u64(u64::from(span.lo().to_u32()) | (u64::from(span.hi().to_u32()) << 32)), + ) + })); + let md = unsafe { llvm::LLVMMDNodeInContext2(bx.llcx, srcloc.as_ptr(), srcloc.len()) }; + let md = bx.get_metadata_value(md); + llvm::LLVMSetMetadata(call, kind, md); + + Some(call) + } else { + // LLVM has detected an issue with our constraints, bail out + None } } @@ -939,9 +935,10 @@ fn llvm_fixup_input<'ll, 'tcx>( } bx.insert_element(bx.const_undef(vec_ty), value, bx.const_i32(0)) } - (AArch64(AArch64InlineAsmRegClass::vreg_low16), BackendRepr::Vector { element, count }) - if layout.size.bytes() == 8 => - { + ( + AArch64(AArch64InlineAsmRegClass::vreg_low16), + BackendRepr::SimdVector { element, count }, + ) if layout.size.bytes() == 8 => { let elem_ty = llvm_asm_scalar_type(bx.cx, element); let vec_ty = bx.cx.type_vector(elem_ty, count); let indices: Vec<_> = (0..count * 2).map(|x| bx.const_i32(x as i32)).collect(); @@ -954,7 +951,7 @@ fn llvm_fixup_input<'ll, 'tcx>( } ( X86(X86InlineAsmRegClass::xmm_reg | X86InlineAsmRegClass::zmm_reg), - BackendRepr::Vector { .. }, + BackendRepr::SimdVector { .. }, ) if layout.size.bytes() == 64 => bx.bitcast(value, bx.cx.type_vector(bx.cx.type_f64(), 8)), ( X86( @@ -989,7 +986,7 @@ fn llvm_fixup_input<'ll, 'tcx>( | X86InlineAsmRegClass::ymm_reg | X86InlineAsmRegClass::zmm_reg, ), - BackendRepr::Vector { element, count: count @ (8 | 16) }, + BackendRepr::SimdVector { element, count: count @ (8 | 16) }, ) if element.primitive() == Primitive::Float(Float::F16) => { bx.bitcast(value, bx.type_vector(bx.type_i16(), count)) } @@ -1026,7 +1023,7 @@ fn llvm_fixup_input<'ll, 'tcx>( | ArmInlineAsmRegClass::qreg_low4 | ArmInlineAsmRegClass::qreg_low8, ), - BackendRepr::Vector { element, count: count @ (4 | 8) }, + BackendRepr::SimdVector { element, count: count @ (4 | 8) }, ) if element.primitive() == Primitive::Float(Float::F16) => { bx.bitcast(value, bx.type_vector(bx.type_i16(), count)) } @@ -1099,9 +1096,10 @@ fn llvm_fixup_output<'ll, 'tcx>( } value } - (AArch64(AArch64InlineAsmRegClass::vreg_low16), BackendRepr::Vector { element, count }) - if layout.size.bytes() == 8 => - { + ( + AArch64(AArch64InlineAsmRegClass::vreg_low16), + BackendRepr::SimdVector { element, count }, + ) if layout.size.bytes() == 8 => { let elem_ty = llvm_asm_scalar_type(bx.cx, element); let vec_ty = bx.cx.type_vector(elem_ty, count * 2); let indices: Vec<_> = (0..count).map(|x| bx.const_i32(x as i32)).collect(); @@ -1114,7 +1112,7 @@ fn llvm_fixup_output<'ll, 'tcx>( } ( X86(X86InlineAsmRegClass::xmm_reg | X86InlineAsmRegClass::zmm_reg), - BackendRepr::Vector { .. }, + BackendRepr::SimdVector { .. }, ) if layout.size.bytes() == 64 => bx.bitcast(value, layout.llvm_type(bx.cx)), ( X86( @@ -1145,7 +1143,7 @@ fn llvm_fixup_output<'ll, 'tcx>( | X86InlineAsmRegClass::ymm_reg | X86InlineAsmRegClass::zmm_reg, ), - BackendRepr::Vector { element, count: count @ (8 | 16) }, + BackendRepr::SimdVector { element, count: count @ (8 | 16) }, ) if element.primitive() == Primitive::Float(Float::F16) => { bx.bitcast(value, bx.type_vector(bx.type_f16(), count)) } @@ -1182,7 +1180,7 @@ fn llvm_fixup_output<'ll, 'tcx>( | ArmInlineAsmRegClass::qreg_low4 | ArmInlineAsmRegClass::qreg_low8, ), - BackendRepr::Vector { element, count: count @ (4 | 8) }, + BackendRepr::SimdVector { element, count: count @ (4 | 8) }, ) if element.primitive() == Primitive::Float(Float::F16) => { bx.bitcast(value, bx.type_vector(bx.type_f16(), count)) } @@ -1243,9 +1241,10 @@ fn llvm_fixup_output_type<'ll, 'tcx>( let count = 16 / layout.size.bytes(); cx.type_vector(elem_ty, count) } - (AArch64(AArch64InlineAsmRegClass::vreg_low16), BackendRepr::Vector { element, count }) - if layout.size.bytes() == 8 => - { + ( + AArch64(AArch64InlineAsmRegClass::vreg_low16), + BackendRepr::SimdVector { element, count }, + ) if layout.size.bytes() == 8 => { let elem_ty = llvm_asm_scalar_type(cx, element); cx.type_vector(elem_ty, count * 2) } @@ -1256,7 +1255,7 @@ fn llvm_fixup_output_type<'ll, 'tcx>( } ( X86(X86InlineAsmRegClass::xmm_reg | X86InlineAsmRegClass::zmm_reg), - BackendRepr::Vector { .. }, + BackendRepr::SimdVector { .. }, ) if layout.size.bytes() == 64 => cx.type_vector(cx.type_f64(), 8), ( X86( @@ -1284,7 +1283,7 @@ fn llvm_fixup_output_type<'ll, 'tcx>( | X86InlineAsmRegClass::ymm_reg | X86InlineAsmRegClass::zmm_reg, ), - BackendRepr::Vector { element, count: count @ (8 | 16) }, + BackendRepr::SimdVector { element, count: count @ (8 | 16) }, ) if element.primitive() == Primitive::Float(Float::F16) => { cx.type_vector(cx.type_i16(), count) } @@ -1321,7 +1320,7 @@ fn llvm_fixup_output_type<'ll, 'tcx>( | ArmInlineAsmRegClass::qreg_low4 | ArmInlineAsmRegClass::qreg_low8, ), - BackendRepr::Vector { element, count: count @ (4 | 8) }, + BackendRepr::SimdVector { element, count: count @ (4 | 8) }, ) if element.primitive() == Primitive::Float(Float::F16) => { cx.type_vector(cx.type_i16(), count) } diff --git a/compiler/rustc_codegen_llvm/src/attributes.rs b/compiler/rustc_codegen_llvm/src/attributes.rs index 3d7afa17bdf..176fb72dfdc 100644 --- a/compiler/rustc_codegen_llvm/src/attributes.rs +++ b/compiler/rustc_codegen_llvm/src/attributes.rs @@ -1,5 +1,4 @@ //! Set and unset common attributes on LLVM values. - use rustc_attr_parsing::{InlineAttr, InstructionSetAttr, OptimizeAttr}; use rustc_codegen_ssa::traits::*; use rustc_hir::def_id::DefId; @@ -28,6 +27,22 @@ pub(crate) fn apply_to_callsite(callsite: &Value, idx: AttributePlace, attrs: &[ } } +pub(crate) fn has_attr(llfn: &Value, idx: AttributePlace, attr: AttributeKind) -> bool { + llvm::HasAttributeAtIndex(llfn, idx, attr) +} + +pub(crate) fn has_string_attr(llfn: &Value, name: &str) -> bool { + llvm::HasStringAttribute(llfn, name) +} + +pub(crate) fn remove_from_llfn(llfn: &Value, place: AttributePlace, kind: AttributeKind) { + llvm::RemoveRustEnumAttributeAtIndex(llfn, place, kind); +} + +pub(crate) fn remove_string_attr_from_llfn(llfn: &Value, name: &str) { + llvm::RemoveStringAttrFromFn(llfn, name); +} + /// Get LLVM attribute for the provided inline heuristic. #[inline] fn inline_attr<'ll>(cx: &CodegenCx<'ll, '_>, inline: InlineAttr) -> Option<&'ll Attribute> { @@ -407,30 +422,28 @@ pub(crate) fn llfn_attrs_from_instance<'ll, 'tcx>( // Do not set sanitizer attributes for naked functions. to_add.extend(sanitize_attrs(cx, codegen_fn_attrs.no_sanitize)); - if llvm_util::get_version() >= (19, 0, 0) { - // For non-naked functions, set branch protection attributes on aarch64. - if let Some(BranchProtection { bti, pac_ret }) = - cx.sess().opts.unstable_opts.branch_protection - { - assert!(cx.sess().target.arch == "aarch64"); - if bti { - to_add.push(llvm::CreateAttrString(cx.llcx, "branch-target-enforcement")); - } - if let Some(PacRet { leaf, pc, key }) = pac_ret { - if pc { - to_add.push(llvm::CreateAttrString(cx.llcx, "branch-protection-pauth-lr")); - } - to_add.push(llvm::CreateAttrStringValue( - cx.llcx, - "sign-return-address", - if leaf { "all" } else { "non-leaf" }, - )); - to_add.push(llvm::CreateAttrStringValue( - cx.llcx, - "sign-return-address-key", - if key == PAuthKey::A { "a_key" } else { "b_key" }, - )); + // For non-naked functions, set branch protection attributes on aarch64. + if let Some(BranchProtection { bti, pac_ret }) = + cx.sess().opts.unstable_opts.branch_protection + { + assert!(cx.sess().target.arch == "aarch64"); + if bti { + to_add.push(llvm::CreateAttrString(cx.llcx, "branch-target-enforcement")); + } + if let Some(PacRet { leaf, pc, key }) = pac_ret { + if pc { + to_add.push(llvm::CreateAttrString(cx.llcx, "branch-protection-pauth-lr")); } + to_add.push(llvm::CreateAttrStringValue( + cx.llcx, + "sign-return-address", + if leaf { "all" } else { "non-leaf" }, + )); + to_add.push(llvm::CreateAttrStringValue( + cx.llcx, + "sign-return-address-key", + if key == PAuthKey::A { "a_key" } else { "b_key" }, + )); } } } @@ -510,12 +523,6 @@ pub(crate) fn llfn_attrs_from_instance<'ll, 'tcx>( InstructionSetAttr::ArmA32 => "-thumb-mode".to_string(), InstructionSetAttr::ArmT32 => "+thumb-mode".to_string(), })) - // HACK: LLVM versions 19+ do not have the FPMR feature and treat it as always enabled - // It only exists as a feature in LLVM 18, cannot be passed down for any other version - .chain(match &*cx.tcx.sess.target.arch { - "aarch64" if llvm_util::get_version().0 == 18 => vec!["+fpmr".to_string()], - _ => vec![], - }) .collect::<Vec<String>>(); if cx.tcx.sess.target.is_like_wasm { diff --git a/compiler/rustc_codegen_llvm/src/back/archive.rs b/compiler/rustc_codegen_llvm/src/back/archive.rs index 93553f3f364..0a161442933 100644 --- a/compiler/rustc_codegen_llvm/src/back/archive.rs +++ b/compiler/rustc_codegen_llvm/src/back/archive.rs @@ -132,14 +132,33 @@ fn get_llvm_object_symbols( if err.is_null() { return Ok(true); } else { - return Err(unsafe { *Box::from_raw(err as *mut io::Error) }); + let error = unsafe { *Box::from_raw(err as *mut io::Error) }; + // These are the magic constants for LLVM bitcode files: + // https://github.com/llvm/llvm-project/blob/7eadc1960d199676f04add402bb0aa6f65b7b234/llvm/lib/BinaryFormat/Magic.cpp#L90-L97 + if buf.starts_with(&[0xDE, 0xCE, 0x17, 0x0B]) || buf.starts_with(&[b'B', b'C', 0xC0, 0xDE]) + { + // For LLVM bitcode, failure to read the symbols is not fatal. The bitcode may have been + // produced by a newer LLVM version that the one linked to rustc. This is fine provided + // that the linker does use said newer LLVM version. We skip writing the symbols for the + // bitcode to the symbol table of the archive. Traditional linkers don't like this, but + // newer linkers like lld, mold and wild ignore the symbol table anyway, so if they link + // against a new enough LLVM it will work out in the end. + // LLVM's archive writer also has this same behavior of only warning about invalid + // bitcode since https://github.com/llvm/llvm-project/pull/96848 + + // We don't have access to the DiagCtxt here to produce a nice warning in the correct format. + eprintln!("warning: Failed to read symbol table from LLVM bitcode: {}", error); + return Ok(true); + } else { + return Err(error); + } } unsafe extern "C" fn callback(state: *mut c_void, symbol_name: *const c_char) -> *mut c_void { let f = unsafe { &mut *(state as *mut &mut dyn FnMut(&[u8]) -> io::Result<()>) }; match f(unsafe { CStr::from_ptr(symbol_name) }.to_bytes()) { Ok(()) => std::ptr::null_mut(), - Err(err) => Box::into_raw(Box::new(err)) as *mut c_void, + Err(err) => Box::into_raw(Box::new(err) as Box<io::Error>) as *mut c_void, } } @@ -148,7 +167,7 @@ fn get_llvm_object_symbols( Box::into_raw(Box::new(io::Error::new( io::ErrorKind::Other, format!("LLVM error: {}", error.to_string_lossy()), - ))) as *mut c_void + )) as Box<io::Error>) as *mut c_void } } diff --git a/compiler/rustc_codegen_llvm/src/back/lto.rs b/compiler/rustc_codegen_llvm/src/back/lto.rs index 99906ea7bce..39b3a23e0b1 100644 --- a/compiler/rustc_codegen_llvm/src/back/lto.rs +++ b/compiler/rustc_codegen_llvm/src/back/lto.rs @@ -2,6 +2,7 @@ use std::collections::BTreeMap; use std::ffi::{CStr, CString}; use std::fs::File; use std::path::Path; +use std::ptr::NonNull; use std::sync::Arc; use std::{io, iter, slice}; @@ -27,8 +28,9 @@ use crate::back::write::{ use crate::errors::{ DynamicLinkingWithLTO, LlvmError, LtoBitcodeFromRlib, LtoDisallowed, LtoDylib, LtoProcMacro, }; +use crate::llvm::AttributePlace::Function; use crate::llvm::{self, build_string}; -use crate::{LlvmCodegenBackend, ModuleLlvm}; +use crate::{LlvmCodegenBackend, ModuleLlvm, SimpleCx, attributes}; /// We keep track of the computed LTO cache keys from the previous /// session to determine which CGUs we can reuse. @@ -305,11 +307,8 @@ fn fat_lto( assert!(!serialized_modules.is_empty(), "must have at least one serialized module"); let (buffer, name) = serialized_modules.remove(0); info!("no in-memory regular modules to choose from, parsing {:?}", name); - ModuleCodegen { - module_llvm: ModuleLlvm::parse(cgcx, &name, buffer.data(), dcx)?, - name: name.into_string().unwrap(), - kind: ModuleKind::Regular, - } + let llvm_module = ModuleLlvm::parse(cgcx, &name, buffer.data(), dcx)?; + ModuleCodegen::new_regular(name.into_string().unwrap(), llvm_module) } }; { @@ -586,12 +585,10 @@ fn thin_lto( } } -fn enable_autodiff_settings(ad: &[config::AutoDiff], module: &mut ModuleCodegen<ModuleLlvm>) { +fn enable_autodiff_settings(ad: &[config::AutoDiff]) { for &val in ad { + // We intentionally don't use a wildcard, to not forget handling anything new. match val { - config::AutoDiff::PrintModBefore => { - unsafe { llvm::LLVMDumpModule(module.module_llvm.llmod()) }; - } config::AutoDiff::PrintPerf => { llvm::set_print_perf(true); } @@ -605,15 +602,23 @@ fn enable_autodiff_settings(ad: &[config::AutoDiff], module: &mut ModuleCodegen< llvm::set_inline(true); } config::AutoDiff::LooseTypes => { - llvm::set_loose_types(false); + llvm::set_loose_types(true); } config::AutoDiff::PrintSteps => { llvm::set_print(true); } - // We handle this below + // We handle this in the PassWrapper.cpp + config::AutoDiff::PrintPasses => {} + // We handle this in the PassWrapper.cpp + config::AutoDiff::PrintModBefore => {} + // We handle this in the PassWrapper.cpp config::AutoDiff::PrintModAfter => {} + // We handle this in the PassWrapper.cpp + config::AutoDiff::PrintModFinal => {} // This is required and already checked config::AutoDiff::Enable => {} + // We handle this below + config::AutoDiff::NoPostopt => {} } } // This helps with handling enums for now. @@ -647,26 +652,57 @@ pub(crate) fn run_pass_manager( // We then run the llvm_optimize function a second time, to optimize the code which we generated // in the enzyme differentiation pass. let enable_ad = config.autodiff.contains(&config::AutoDiff::Enable); - let stage = - if enable_ad { write::AutodiffStage::DuringAD } else { write::AutodiffStage::PostAD }; + let stage = if thin { + write::AutodiffStage::PreAD + } else { + if enable_ad { write::AutodiffStage::DuringAD } else { write::AutodiffStage::PostAD } + }; if enable_ad { - enable_autodiff_settings(&config.autodiff, module); + enable_autodiff_settings(&config.autodiff); } unsafe { - write::llvm_optimize(cgcx, dcx, module, config, opt_level, opt_stage, stage)?; + write::llvm_optimize(cgcx, dcx, module, None, config, opt_level, opt_stage, stage)?; } - if cfg!(llvm_enzyme) && enable_ad { + if cfg!(llvm_enzyme) && enable_ad && !thin { + let cx = + SimpleCx::new(module.module_llvm.llmod(), &module.module_llvm.llcx, cgcx.pointer_size); + + for function in cx.get_functions() { + let enzyme_marker = "enzyme_marker"; + if attributes::has_string_attr(function, enzyme_marker) { + // Sanity check: Ensure 'noinline' is present before replacing it. + assert!( + !attributes::has_attr(function, Function, llvm::AttributeKind::NoInline), + "Expected __enzyme function to have 'noinline' before adding 'alwaysinline'" + ); + + attributes::remove_from_llfn(function, Function, llvm::AttributeKind::NoInline); + attributes::remove_string_attr_from_llfn(function, enzyme_marker); + + assert!( + !attributes::has_string_attr(function, enzyme_marker), + "Expected function to not have 'enzyme_marker'" + ); + + let always_inline = llvm::AttributeKind::AlwaysInline.create_attr(cx.llcx); + attributes::apply_to_llfn(function, Function, &[always_inline]); + } + } + let opt_stage = llvm::OptStage::FatLTO; let stage = write::AutodiffStage::PostAD; - unsafe { - write::llvm_optimize(cgcx, dcx, module, config, opt_level, opt_stage, stage)?; + if !config.autodiff.contains(&config::AutoDiff::NoPostopt) { + 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()) }; } } @@ -729,6 +765,11 @@ impl ThinBuffer { ThinBuffer(buffer) } } + + pub(crate) unsafe fn from_raw_ptr(ptr: *mut llvm::ThinLTOBuffer) -> ThinBuffer { + let mut ptr = NonNull::new(ptr).unwrap(); + ThinBuffer(unsafe { ptr.as_mut() }) + } } impl ThinBufferMethods for ThinBuffer { @@ -772,11 +813,11 @@ pub(crate) unsafe fn optimize_thin_module( // crates but for locally codegened modules we may be able to reuse // that LLVM Context and Module. let module_llvm = ModuleLlvm::parse(cgcx, module_name, thin_module.data(), dcx)?; - let mut module = ModuleCodegen { - module_llvm, - name: thin_module.name().to_string(), - kind: ModuleKind::Regular, - }; + let mut module = ModuleCodegen::new_regular(thin_module.name(), module_llvm); + // Given that the newly created module lacks a thinlto buffer for embedding, we need to re-add it here. + if cgcx.config(ModuleKind::Regular).embed_bitcode() { + module.thin_lto_buffer = Some(thin_module.data().to_vec()); + } { let target = &*module.module_llvm.tm; let llmod = module.module_llvm.llmod(); @@ -793,7 +834,9 @@ pub(crate) unsafe fn optimize_thin_module( { let _timer = cgcx.prof.generic_activity_with_arg("LLVM_thin_lto_rename", thin_module.name()); - unsafe { llvm::LLVMRustPrepareThinLTORename(thin_module.shared.data.0, llmod, target) }; + unsafe { + llvm::LLVMRustPrepareThinLTORename(thin_module.shared.data.0, llmod, target.raw()) + }; save_temp_bitcode(cgcx, &module, "thin-lto-after-rename"); } @@ -823,7 +866,7 @@ pub(crate) unsafe fn optimize_thin_module( let _timer = cgcx.prof.generic_activity_with_arg("LLVM_thin_lto_import", thin_module.name()); if unsafe { - !llvm::LLVMRustPrepareThinLTOImport(thin_module.shared.data.0, llmod, target) + !llvm::LLVMRustPrepareThinLTOImport(thin_module.shared.data.0, llmod, target.raw()) } { return Err(write::llvm_err(dcx, LlvmError::PrepareThinLtoModule)); } diff --git a/compiler/rustc_codegen_llvm/src/back/owned_target_machine.rs b/compiler/rustc_codegen_llvm/src/back/owned_target_machine.rs index f075f332462..dfde4595590 100644 --- a/compiler/rustc_codegen_llvm/src/back/owned_target_machine.rs +++ b/compiler/rustc_codegen_llvm/src/back/owned_target_machine.rs @@ -1,6 +1,5 @@ use std::ffi::{CStr, c_char}; use std::marker::PhantomData; -use std::ops::Deref; use std::ptr::NonNull; use rustc_data_structures::small_c_str::SmallCStr; @@ -80,12 +79,12 @@ impl OwnedTargetMachine { .map(|tm_unique| Self { tm_unique, phantom: PhantomData }) .ok_or_else(|| LlvmError::CreateTargetMachine { triple: SmallCStr::from(triple) }) } -} - -impl Deref for OwnedTargetMachine { - type Target = llvm::TargetMachine; - fn deref(&self) -> &Self::Target { + /// Returns inner `llvm::TargetMachine` type. + /// + /// This could be a `Deref` implementation, but `llvm::TargetMachine` is an extern type and + /// `Deref::Target: ?Sized`. + pub fn raw(&self) -> &llvm::TargetMachine { // SAFETY: constructing ensures we have a valid pointer created by // llvm::LLVMRustCreateTargetMachine. unsafe { self.tm_unique.as_ref() } diff --git a/compiler/rustc_codegen_llvm/src/back/write.rs b/compiler/rustc_codegen_llvm/src/back/write.rs index b67890c0465..4ac77c8f7f1 100644 --- a/compiler/rustc_codegen_llvm/src/back/write.rs +++ b/compiler/rustc_codegen_llvm/src/back/write.rs @@ -1,6 +1,7 @@ use std::ffi::{CStr, CString}; use std::io::{self, Write}; use std::path::{Path, PathBuf}; +use std::ptr::null_mut; use std::sync::Arc; use std::{fs, slice, str}; @@ -15,7 +16,7 @@ use rustc_codegen_ssa::back::write::{ TargetMachineFactoryFn, }; use rustc_codegen_ssa::traits::*; -use rustc_codegen_ssa::{CompiledModule, ModuleCodegen}; +use rustc_codegen_ssa::{CompiledModule, ModuleCodegen, ModuleKind}; use rustc_data_structures::profiling::SelfProfilerRef; use rustc_data_structures::small_c_str::SmallCStr; use rustc_errors::{DiagCtxtHandle, FatalError, Level}; @@ -118,14 +119,18 @@ pub(crate) fn create_target_machine(tcx: TyCtxt<'_>, mod_name: &str) -> OwnedTar tcx.output_filenames(()).split_dwarf_path( tcx.sess.split_debuginfo(), tcx.sess.opts.unstable_opts.split_dwarf_kind, - Some(mod_name), + mod_name, + tcx.sess.invocation_temp.as_deref(), ) } else { None }; - let output_obj_file = - Some(tcx.output_filenames(()).temp_path(OutputType::Object, Some(mod_name))); + let output_obj_file = Some(tcx.output_filenames(()).temp_path_for_cgu( + OutputType::Object, + mod_name, + tcx.sess.invocation_temp.as_deref(), + )); let config = TargetMachineFactoryConfig { split_dwarf_file, output_obj_file }; target_machine_factory( @@ -329,8 +334,11 @@ pub(crate) fn save_temp_bitcode( return; } let ext = format!("{name}.bc"); - let cgu = Some(&module.name[..]); - let path = cgcx.output_filenames.temp_path_ext(&ext, cgu); + let path = cgcx.output_filenames.temp_path_ext_for_cgu( + &ext, + &module.name, + cgcx.invocation_temp.as_deref(), + ); write_bitcode_to_file(module, &path) } @@ -438,12 +446,9 @@ fn report_inline_asm( let span = if cookie == 0 || matches!(cgcx.lto, Lto::Fat | Lto::Thin) { SpanData::default() } else { - let lo = BytePos::from_u32(cookie as u32); - let hi = BytePos::from_u32((cookie >> 32) as u32); SpanData { - lo, - // LLVM version < 19 silently truncates the cookie to 32 bits in some situations. - hi: if hi.to_u32() != 0 { hi } else { lo }, + lo: BytePos::from_u32(cookie as u32), + hi: BytePos::from_u32((cookie >> 32) as u32), ctxt: SyntaxContext::root(), parent: None, } @@ -551,6 +556,7 @@ pub(crate) unsafe fn llvm_optimize( cgcx: &CodegenContext<LlvmCodegenBackend>, dcx: DiagCtxtHandle<'_>, module: &ModuleCodegen<ModuleLlvm>, + thin_lto_buffer: Option<&mut *mut llvm::ThinLTOBuffer>, config: &ModuleConfig, opt_level: config::OptLevel, opt_stage: llvm::OptStage, @@ -566,6 +572,10 @@ pub(crate) unsafe fn llvm_optimize( let consider_ad = cfg!(llvm_enzyme) && config.autodiff.contains(&config::AutoDiff::Enable); let run_enzyme = autodiff_stage == AutodiffStage::DuringAD; + let print_before_enzyme = config.autodiff.contains(&config::AutoDiff::PrintModBefore); + let print_after_enzyme = config.autodiff.contains(&config::AutoDiff::PrintModAfter); + let print_passes = config.autodiff.contains(&config::AutoDiff::PrintPasses); + let merge_functions; let unroll_loops; let vectorize_slp; let vectorize_loop; @@ -573,18 +583,35 @@ pub(crate) unsafe fn llvm_optimize( // When we build rustc with enzyme/autodiff support, we want to postpone size-increasing // optimizations until after differentiation. Our pipeline is thus: (opt + enzyme), (full opt). // We therefore have two calls to llvm_optimize, if autodiff is used. + // + // We also must disable merge_functions, since autodiff placeholder/dummy bodies tend to be + // identical. We run opts before AD, so there is a chance that LLVM will merge our dummies. + // In that case, we lack some dummy bodies and can't replace them with the real AD code anymore. + // We then would need to abort compilation. This was especially common in test cases. if consider_ad && autodiff_stage != AutodiffStage::PostAD { + merge_functions = false; unroll_loops = false; vectorize_slp = false; vectorize_loop = false; } else { unroll_loops = opt_level != config::OptLevel::Size && opt_level != config::OptLevel::SizeMin; + merge_functions = config.merge_functions; vectorize_slp = config.vectorize_slp; vectorize_loop = config.vectorize_loop; } trace!(?unroll_loops, ?vectorize_slp, ?vectorize_loop, ?run_enzyme); - let using_thin_buffers = opt_stage == llvm::OptStage::PreLinkThinLTO || config.bitcode_needed(); + if thin_lto_buffer.is_some() { + assert!( + matches!( + opt_stage, + llvm::OptStage::PreLinkNoLTO + | llvm::OptStage::PreLinkFatLTO + | llvm::OptStage::PreLinkThinLTO + ), + "the bitcode for LTO can only be obtained at the pre-link stage" + ); + } let pgo_gen_path = get_pgo_gen_path(config); let pgo_use_path = get_pgo_use_path(config); let pgo_sample_use_path = get_pgo_sample_use_path(config); @@ -637,21 +664,26 @@ pub(crate) unsafe fn llvm_optimize( let result = unsafe { llvm::LLVMRustOptimize( module.module_llvm.llmod(), - &*module.module_llvm.tm, + &*module.module_llvm.tm.raw(), to_pass_builder_opt_level(opt_level), opt_stage, cgcx.opts.cg.linker_plugin_lto.enabled(), config.no_prepopulate_passes, config.verify_llvm_ir, config.lint_llvm_ir, - using_thin_buffers, - config.merge_functions, + thin_lto_buffer, + config.emit_thin_lto, + config.emit_thin_lto_summary, + merge_functions, unroll_loops, vectorize_slp, vectorize_loop, config.no_builtins, config.emit_lifetime_markers, run_enzyme, + print_before_enzyme, + print_after_enzyme, + print_passes, sanitizer_options.as_ref(), pgo_gen_path.as_ref().map_or(std::ptr::null(), |s| s.as_ptr()), pgo_use_path.as_ref().map_or(std::ptr::null(), |s| s.as_ptr()), @@ -675,7 +707,7 @@ pub(crate) unsafe fn llvm_optimize( pub(crate) unsafe fn optimize( cgcx: &CodegenContext<LlvmCodegenBackend>, dcx: DiagCtxtHandle<'_>, - module: &ModuleCodegen<ModuleLlvm>, + module: &mut ModuleCodegen<ModuleLlvm>, config: &ModuleConfig, ) -> Result<(), FatalError> { let _timer = cgcx.prof.generic_activity_with_arg("LLVM_module_optimize", &*module.name); @@ -683,11 +715,12 @@ pub(crate) unsafe fn optimize( let llcx = &*module.module_llvm.llcx; let _handlers = DiagnosticHandlers::new(cgcx, dcx, llcx, module, CodegenDiagnosticsStage::Opt); - let module_name = module.name.clone(); - let module_name = Some(&module_name[..]); - if config.emit_no_opt_bc { - let out = cgcx.output_filenames.temp_path_ext("no-opt.bc", module_name); + let out = cgcx.output_filenames.temp_path_ext_for_cgu( + "no-opt.bc", + &module.name, + cgcx.invocation_temp.as_deref(), + ); write_bitcode_to_file(module, &out) } @@ -705,9 +738,56 @@ pub(crate) unsafe fn optimize( // Otherwise we pretend AD is already done and run the normal opt pipeline (=PostAD). let consider_ad = cfg!(llvm_enzyme) && config.autodiff.contains(&config::AutoDiff::Enable); let autodiff_stage = if consider_ad { AutodiffStage::PreAD } else { AutodiffStage::PostAD }; - return unsafe { - llvm_optimize(cgcx, dcx, module, config, opt_level, opt_stage, autodiff_stage) + // The embedded bitcode is used to run LTO/ThinLTO. + // The bitcode obtained during the `codegen` phase is no longer suitable for performing LTO. + // It may have undergone LTO due to ThinLocal, so we need to obtain the embedded bitcode at + // this point. + let mut thin_lto_buffer = if (module.kind == ModuleKind::Regular + && config.emit_obj == EmitObj::ObjectCode(BitcodeSection::Full)) + || config.emit_thin_lto_summary + { + Some(null_mut()) + } else { + None }; + unsafe { + llvm_optimize( + cgcx, + dcx, + module, + thin_lto_buffer.as_mut(), + config, + opt_level, + opt_stage, + autodiff_stage, + ) + }?; + if let Some(thin_lto_buffer) = thin_lto_buffer { + let thin_lto_buffer = unsafe { ThinBuffer::from_raw_ptr(thin_lto_buffer) }; + module.thin_lto_buffer = Some(thin_lto_buffer.data().to_vec()); + let bc_summary_out = cgcx.output_filenames.temp_path_for_cgu( + OutputType::ThinLinkBitcode, + &module.name, + cgcx.invocation_temp.as_deref(), + ); + if config.emit_thin_lto_summary + && let Some(thin_link_bitcode_filename) = bc_summary_out.file_name() + { + let summary_data = thin_lto_buffer.thin_link_data(); + cgcx.prof.artifact_size( + "llvm_bitcode_summary", + thin_link_bitcode_filename.to_string_lossy(), + summary_data.len() as u64, + ); + let _timer = cgcx.prof.generic_activity_with_arg( + "LLVM_module_codegen_emit_bitcode_summary", + &*module.name, + ); + if let Err(err) = fs::write(&bc_summary_out, summary_data) { + dcx.emit_err(WriteBytecode { path: &bc_summary_out, err }); + } + } + } } Ok(()) } @@ -746,8 +826,6 @@ pub(crate) unsafe fn codegen( let llmod = module.module_llvm.llmod(); let llcx = &*module.module_llvm.llcx; let tm = &*module.module_llvm.tm; - let module_name = module.name.clone(); - let module_name = Some(&module_name[..]); let _handlers = DiagnosticHandlers::new(cgcx, dcx, llcx, &module, CodegenDiagnosticsStage::Codegen); @@ -759,60 +837,50 @@ pub(crate) unsafe fn codegen( // 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 = - cgcx.output_filenames.temp_path(OutputType::ThinLinkBitcode, module_name); - let obj_out = cgcx.output_filenames.temp_path(OutputType::Object, module_name); + let bc_out = cgcx.output_filenames.temp_path_for_cgu( + OutputType::Bitcode, + &module.name, + cgcx.invocation_temp.as_deref(), + ); + let obj_out = cgcx.output_filenames.temp_path_for_cgu( + OutputType::Object, + &module.name, + cgcx.invocation_temp.as_deref(), + ); if config.bitcode_needed() { - let _timer = cgcx - .prof - .generic_activity_with_arg("LLVM_module_codegen_make_bitcode", &*module.name); - let thin = ThinBuffer::new(llmod, config.emit_thin_lto, config.emit_thin_lto_summary); - let data = thin.data(); - - if let Some(bitcode_filename) = bc_out.file_name() { - cgcx.prof.artifact_size( - "llvm_bitcode", - bitcode_filename.to_string_lossy(), - data.len() as u64, - ); - } - - if config.emit_thin_lto_summary - && let Some(thin_link_bitcode_filename) = bc_summary_out.file_name() - { - let summary_data = thin.thin_link_data(); - cgcx.prof.artifact_size( - "llvm_bitcode_summary", - thin_link_bitcode_filename.to_string_lossy(), - summary_data.len() as u64, - ); - - let _timer = cgcx.prof.generic_activity_with_arg( - "LLVM_module_codegen_emit_bitcode_summary", - &*module.name, - ); - if let Err(err) = fs::write(&bc_summary_out, summary_data) { - dcx.emit_err(WriteBytecode { path: &bc_summary_out, err }); - } - } - if config.emit_bc || config.emit_obj == EmitObj::Bitcode { + let thin = { + let _timer = cgcx.prof.generic_activity_with_arg( + "LLVM_module_codegen_make_bitcode", + &*module.name, + ); + ThinBuffer::new(llmod, config.emit_thin_lto, false) + }; + let data = thin.data(); let _timer = cgcx .prof .generic_activity_with_arg("LLVM_module_codegen_emit_bitcode", &*module.name); + if let Some(bitcode_filename) = bc_out.file_name() { + cgcx.prof.artifact_size( + "llvm_bitcode", + bitcode_filename.to_string_lossy(), + data.len() as u64, + ); + } if let Err(err) = fs::write(&bc_out, data) { dcx.emit_err(WriteBytecode { path: &bc_out, err }); } } - if config.emit_obj == EmitObj::ObjectCode(BitcodeSection::Full) { + if config.embed_bitcode() && module.kind == ModuleKind::Regular { let _timer = cgcx .prof .generic_activity_with_arg("LLVM_module_codegen_embed_bitcode", &*module.name); + let thin_bc = + module.thin_lto_buffer.as_deref().expect("cannot find embedded bitcode"); unsafe { - embed_bitcode(cgcx, llcx, llmod, &config.bc_cmdline, data); + embed_bitcode(cgcx, llcx, llmod, &config.bc_cmdline, &thin_bc); } } } @@ -820,7 +888,11 @@ pub(crate) unsafe fn codegen( if config.emit_ir { let _timer = cgcx.prof.generic_activity_with_arg("LLVM_module_codegen_emit_ir", &*module.name); - let out = cgcx.output_filenames.temp_path(OutputType::LlvmAssembly, module_name); + let out = cgcx.output_filenames.temp_path_for_cgu( + OutputType::LlvmAssembly, + &module.name, + cgcx.invocation_temp.as_deref(), + ); let out_c = path_to_c_string(&out); extern "C" fn demangle_callback( @@ -862,7 +934,11 @@ pub(crate) unsafe fn codegen( if config.emit_asm { let _timer = cgcx.prof.generic_activity_with_arg("LLVM_module_codegen_emit_asm", &*module.name); - let path = cgcx.output_filenames.temp_path(OutputType::Assembly, module_name); + let path = cgcx.output_filenames.temp_path_for_cgu( + OutputType::Assembly, + &module.name, + cgcx.invocation_temp.as_deref(), + ); // We can't use the same module for asm and object code output, // because that triggers various errors like invalid IR or broken @@ -875,7 +951,7 @@ pub(crate) unsafe fn codegen( }; write_output_file( dcx, - tm, + tm.raw(), config.no_builtins, llmod, &path, @@ -892,7 +968,9 @@ pub(crate) unsafe fn codegen( .prof .generic_activity_with_arg("LLVM_module_codegen_emit_obj", &*module.name); - let dwo_out = cgcx.output_filenames.temp_path_dwo(module_name); + let dwo_out = cgcx + .output_filenames + .temp_path_dwo_for_cgu(&module.name, cgcx.invocation_temp.as_deref()); let dwo_out = match (cgcx.split_debuginfo, cgcx.split_dwarf_kind) { // Don't change how DWARF is emitted when disabled. (SplitDebuginfo::Off, _) => None, @@ -909,7 +987,7 @@ pub(crate) unsafe fn codegen( write_output_file( dcx, - tm, + tm.raw(), config.no_builtins, llmod, &obj_out, @@ -957,6 +1035,7 @@ pub(crate) unsafe fn codegen( config.emit_asm, config.emit_ir, &cgcx.output_filenames, + cgcx.invocation_temp.as_deref(), )) } @@ -984,7 +1063,7 @@ 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" @@ -1037,7 +1116,7 @@ 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" @@ -1056,7 +1135,7 @@ 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/base.rs b/compiler/rustc_codegen_llvm/src/base.rs index d35c7945bae..e4fac35aa44 100644 --- a/compiler/rustc_codegen_llvm/src/base.rs +++ b/compiler/rustc_codegen_llvm/src/base.rs @@ -13,10 +13,10 @@ use std::time::Instant; +use rustc_codegen_ssa::ModuleCodegen; use rustc_codegen_ssa::base::maybe_create_entry_wrapper; use rustc_codegen_ssa::mono_item::MonoItemExt; use rustc_codegen_ssa::traits::*; -use rustc_codegen_ssa::{ModuleCodegen, ModuleKind}; use rustc_data_structures::small_c_str::SmallCStr; use rustc_middle::dep_graph; use rustc_middle::middle::codegen_fn_attrs::CodegenFnAttrs; @@ -83,15 +83,15 @@ pub(crate) fn compile_codegen_unit( // Instantiate monomorphizations without filling out definitions yet... let llvm_module = ModuleLlvm::new(tcx, cgu_name.as_str()); { - let cx = CodegenCx::new(tcx, cgu, &llvm_module); + let mut cx = CodegenCx::new(tcx, cgu, &llvm_module); let mono_items = cx.codegen_unit.items_in_deterministic_order(cx.tcx); for &(mono_item, data) in &mono_items { mono_item.predefine::<Builder<'_, '_, '_>>(&cx, data.linkage, data.visibility); } // ... and now that we have everything pre-defined, fill out those definitions. - for &(mono_item, _) in &mono_items { - mono_item.define::<Builder<'_, '_, '_>>(&cx); + for &(mono_item, item_data) in &mono_items { + mono_item.define::<Builder<'_, '_, '_>>(&mut cx, item_data); } // If this codegen unit contains the main function, also create the @@ -133,11 +133,7 @@ pub(crate) fn compile_codegen_unit( } } - ModuleCodegen { - name: cgu_name.to_string(), - module_llvm: llvm_module, - kind: ModuleKind::Regular, - } + ModuleCodegen::new_regular(cgu_name.to_string(), llvm_module) } (module, cost) diff --git a/compiler/rustc_codegen_llvm/src/builder.rs b/compiler/rustc_codegen_llvm/src/builder.rs index e1609e31c07..04c8118b616 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, @@ -29,24 +30,26 @@ use smallvec::SmallVec; use tracing::{debug, instrument}; use crate::abi::FnAbiLlvmExt; +use crate::attributes; use crate::common::Funclet; -use crate::context::{CodegenCx, SimpleCx}; -use crate::llvm::{self, AtomicOrdering, AtomicRmwBinOp, BasicBlock, False, Metadata, True}; +use crate::context::{CodegenCx, FullCx, GenericCx, SCx}; +use crate::llvm::{ + self, AtomicOrdering, AtomicRmwBinOp, BasicBlock, False, GEPNoWrapFlags, Metadata, True, +}; use crate::type_::Type; use crate::type_of::LayoutLlvmExt; use crate::value::Value; -use crate::{attributes, llvm_util}; #[must_use] -pub(crate) struct GenericBuilder<'a, 'll, CX: Borrow<SimpleCx<'ll>>> { +pub(crate) struct GenericBuilder<'a, 'll, CX: Borrow<SCx<'ll>>> { pub llbuilder: &'ll mut llvm::Builder<'ll>, - pub cx: &'a CX, + pub cx: &'a GenericCx<'ll, CX>, } -pub(crate) type SBuilder<'a, 'll> = GenericBuilder<'a, 'll, SimpleCx<'ll>>; -pub(crate) type Builder<'a, 'll, 'tcx> = GenericBuilder<'a, 'll, CodegenCx<'ll, 'tcx>>; +pub(crate) type SBuilder<'a, 'll> = GenericBuilder<'a, 'll, SCx<'ll>>; +pub(crate) type Builder<'a, 'll, 'tcx> = GenericBuilder<'a, 'll, FullCx<'ll, 'tcx>>; -impl<'a, 'll, CX: Borrow<SimpleCx<'ll>>> Drop for GenericBuilder<'a, 'll, CX> { +impl<'a, 'll, CX: Borrow<SCx<'ll>>> Drop for GenericBuilder<'a, 'll, CX> { fn drop(&mut self) { unsafe { llvm::LLVMDisposeBuilder(&mut *(self.llbuilder as *mut _)); @@ -55,7 +58,7 @@ impl<'a, 'll, CX: Borrow<SimpleCx<'ll>>> Drop for GenericBuilder<'a, 'll, CX> { } impl<'a, 'll> SBuilder<'a, 'll> { - fn call( + pub(crate) fn call( &mut self, llty: &'ll Type, llfn: &'ll Value, @@ -85,97 +88,54 @@ impl<'a, 'll> SBuilder<'a, 'll> { }; call } +} - fn with_scx(scx: &'a SimpleCx<'ll>) -> Self { +impl<'a, 'll, CX: Borrow<SCx<'ll>>> GenericBuilder<'a, 'll, CX> { + fn with_cx(scx: &'a GenericCx<'ll, CX>) -> Self { // Create a fresh builder from the simple context. - let llbuilder = unsafe { llvm::LLVMCreateBuilderInContext(scx.llcx) }; - SBuilder { llbuilder, cx: scx } + let llbuilder = unsafe { llvm::LLVMCreateBuilderInContext(scx.deref().borrow().llcx) }; + GenericBuilder { 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); - } + pub(crate) fn ret_void(&mut self) { + llvm::LLVMBuildRetVoid(self.llbuilder); } - fn ret(&mut self, v: &'ll Value) { + pub(crate) 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); + + pub(crate) fn build(cx: &'a GenericCx<'ll, CX>, llbb: &'ll BasicBlock) -> Self { + let bx = Self::with_cx(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. -const UNNAMED: *const c_char = c"".as_ptr(); - -impl<'ll, 'tcx> BackendTypes for Builder<'_, 'll, 'tcx> { - type Value = <CodegenCx<'ll, 'tcx> as BackendTypes>::Value; - type Metadata = <CodegenCx<'ll, 'tcx> as BackendTypes>::Metadata; - type Function = <CodegenCx<'ll, 'tcx> as BackendTypes>::Function; - type BasicBlock = <CodegenCx<'ll, 'tcx> as BackendTypes>::BasicBlock; - type Type = <CodegenCx<'ll, 'tcx> as BackendTypes>::Type; - type Funclet = <CodegenCx<'ll, 'tcx> as BackendTypes>::Funclet; - - type DIScope = <CodegenCx<'ll, 'tcx> as BackendTypes>::DIScope; - type DILocation = <CodegenCx<'ll, 'tcx> as BackendTypes>::DILocation; - type DIVariable = <CodegenCx<'ll, 'tcx> as BackendTypes>::DIVariable; +pub(crate) const UNNAMED: *const c_char = c"".as_ptr(); + +impl<'ll, CX: Borrow<SCx<'ll>>> BackendTypes for GenericBuilder<'_, 'll, CX> { + type Value = <GenericCx<'ll, CX> as BackendTypes>::Value; + type Metadata = <GenericCx<'ll, CX> as BackendTypes>::Metadata; + type Function = <GenericCx<'ll, CX> as BackendTypes>::Function; + type BasicBlock = <GenericCx<'ll, CX> as BackendTypes>::BasicBlock; + type Type = <GenericCx<'ll, CX> as BackendTypes>::Type; + type Funclet = <GenericCx<'ll, CX> as BackendTypes>::Funclet; + + type DIScope = <GenericCx<'ll, CX> as BackendTypes>::DIScope; + type DILocation = <GenericCx<'ll, CX> as BackendTypes>::DILocation; + type DIVariable = <GenericCx<'ll, CX> as BackendTypes>::DIVariable; } impl abi::HasDataLayout for Builder<'_, '_, '_> { @@ -291,9 +251,7 @@ impl<'a, 'll, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> { } fn ret_void(&mut self) { - unsafe { - llvm::LLVMBuildRetVoid(self.llbuilder); - } + llvm::LLVMBuildRetVoid(self.llbuilder); } fn ret(&mut self, v: &'ll Value) { @@ -354,8 +312,8 @@ impl<'a, 'll, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> { // This function handles switch instructions with more than 2 targets and it needs to // emit branch weights metadata instead of using the intrinsic. // The values 1 and 2000 are the same as the values used by the `llvm.expect` intrinsic. - let cold_weight = unsafe { llvm::LLVMValueAsMetadata(self.cx.const_u32(1)) }; - let hot_weight = unsafe { llvm::LLVMValueAsMetadata(self.cx.const_u32(2000)) }; + let cold_weight = llvm::LLVMValueAsMetadata(self.cx.const_u32(1)); + let hot_weight = llvm::LLVMValueAsMetadata(self.cx.const_u32(2000)); let weight = |is_cold: bool| -> &Metadata { if is_cold { cold_weight } else { hot_weight } }; @@ -403,7 +361,7 @@ impl<'a, 'll, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> { // Emit KCFI operand bundle let kcfi_bundle = self.kcfi_operand_bundle(fn_attrs, fn_abi, instance, llfn); - if let Some(kcfi_bundle) = kcfi_bundle.as_deref() { + if let Some(kcfi_bundle) = kcfi_bundle.as_ref().map(|b| b.raw()) { bundles.push(kcfi_bundle); } @@ -636,6 +594,7 @@ impl<'a, 'll, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> { fn load(&mut self, ty: &'ll Type, ptr: &'ll Value, align: Align) -> &'ll Value { unsafe { let load = llvm::LLVMBuildLoad2(self.llbuilder, ty, ptr, UNNAMED); + let align = align.min(self.cx().tcx.sess.target.max_reliable_alignment()); llvm::LLVMSetAlignment(load, align.bytes() as c_uint); load } @@ -849,6 +808,7 @@ impl<'a, 'll, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> { assert_eq!(self.cx.type_kind(self.cx.val_ty(ptr)), TypeKind::Pointer); unsafe { let store = llvm::LLVMBuildStore(self.llbuilder, val, ptr); + let align = align.min(self.cx().tcx.sess.target.max_reliable_alignment()); let align = if flags.contains(MemFlags::UNALIGNED) { 1 } else { align.bytes() as c_uint }; llvm::LLVMSetAlignment(store, align); @@ -910,13 +870,14 @@ impl<'a, 'll, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> { fn gep(&mut self, ty: &'ll Type, ptr: &'ll Value, indices: &[&'ll Value]) -> &'ll Value { unsafe { - llvm::LLVMBuildGEP2( + llvm::LLVMBuildGEPWithNoWrapFlags( self.llbuilder, ty, ptr, indices.as_ptr(), indices.len() as c_uint, UNNAMED, + GEPNoWrapFlags::default(), ) } } @@ -928,13 +889,33 @@ impl<'a, 'll, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> { indices: &[&'ll Value], ) -> &'ll Value { unsafe { - llvm::LLVMBuildInBoundsGEP2( + llvm::LLVMBuildGEPWithNoWrapFlags( + self.llbuilder, + ty, + ptr, + indices.as_ptr(), + indices.len() as c_uint, + UNNAMED, + GEPNoWrapFlags::InBounds, + ) + } + } + + fn inbounds_nuw_gep( + &mut self, + ty: &'ll Type, + ptr: &'ll Value, + indices: &[&'ll Value], + ) -> &'ll Value { + unsafe { + llvm::LLVMBuildGEPWithNoWrapFlags( self.llbuilder, ty, ptr, indices.as_ptr(), indices.len() as c_uint, UNNAMED, + GEPNoWrapFlags::InBounds | GEPNoWrapFlags::NUW, ) } } @@ -948,11 +929,9 @@ impl<'a, 'll, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> { debug_assert_ne!(self.val_ty(val), dest_ty); let trunc = self.trunc(val, dest_ty); - if llvm_util::get_version() >= (19, 0, 0) { - unsafe { - if llvm::LLVMIsAInstruction(trunc).is_some() { - llvm::LLVMSetNUW(trunc, True); - } + unsafe { + if llvm::LLVMIsAInstruction(trunc).is_some() { + llvm::LLVMSetNUW(trunc, True); } } trunc @@ -962,11 +941,9 @@ impl<'a, 'll, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> { debug_assert_ne!(self.val_ty(val), dest_ty); let trunc = self.trunc(val, dest_ty); - if llvm_util::get_version() >= (19, 0, 0) { - unsafe { - if llvm::LLVMIsAInstruction(trunc).is_some() { - llvm::LLVMSetNSW(trunc, True); - } + unsafe { + if llvm::LLVMIsAInstruction(trunc).is_some() { + llvm::LLVMSetNSW(trunc, True); } } trunc @@ -1096,6 +1073,35 @@ 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, @@ -1410,7 +1416,7 @@ impl<'a, 'll, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> { // Emit KCFI operand bundle let kcfi_bundle = self.kcfi_operand_bundle(fn_attrs, fn_abi, instance, llfn); - if let Some(kcfi_bundle) = kcfi_bundle.as_deref() { + if let Some(kcfi_bundle) = kcfi_bundle.as_ref().map(|b| b.raw()) { bundles.push(kcfi_bundle); } @@ -1453,26 +1459,12 @@ 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) }; - Builder { llbuilder, cx } - } - pub(crate) fn llfn(&self) -> &'ll Value { unsafe { llvm::LLVMGetBasicBlockParent(self.llbb()) } } } -impl<'a, 'll, CX: Borrow<SimpleCx<'ll>>> GenericBuilder<'a, 'll, CX> { +impl<'a, 'll, CX: Borrow<SCx<'ll>>> GenericBuilder<'a, 'll, CX> { fn position_at_start(&mut self, llbb: &'ll BasicBlock) { unsafe { llvm::LLVMRustPositionBuilderAtStart(self.llbuilder, llbb); @@ -1502,7 +1494,7 @@ impl<'a, 'll, 'tcx> Builder<'a, 'll, 'tcx> { } } } -impl<'a, 'll, CX: Borrow<SimpleCx<'ll>>> GenericBuilder<'a, 'll, CX> { +impl<'a, 'll, CX: Borrow<SCx<'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) } } @@ -1603,9 +1595,7 @@ impl<'a, 'll, CX: Borrow<SimpleCx<'ll>>> GenericBuilder<'a, 'll, CX> { 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, @@ -1620,7 +1610,7 @@ impl<'a, 'll, 'tcx> Builder<'a, 'll, 'tcx> { let param_tys = self.cx.func_params_types(fn_ty); - let all_args_match = iter::zip(¶m_tys, args.iter().map(|&v| self.val_ty(v))) + 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 { @@ -1630,7 +1620,7 @@ impl<'a, 'll, 'tcx> Builder<'a, 'll, 'tcx> { let casted_args: Vec<_> = iter::zip(param_tys, args) .enumerate() .map(|(i, (expected_ty, &actual_val))| { - let actual_ty = self.val_ty(actual_val); + let actual_ty = self.cx.val_ty(actual_val); if expected_ty != actual_ty { debug!( "type mismatch in function call of {:?}. \ @@ -1646,12 +1636,12 @@ 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); @@ -1671,7 +1661,7 @@ 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> { +impl<'a, 'll, CX: Borrow<SCx<'ll>>> GenericBuilder<'a, 'll, CX> { pub(crate) fn phi( &mut self, ty: &'ll Type, @@ -1759,7 +1749,7 @@ impl<'a, 'll, 'tcx> Builder<'a, 'll, 'tcx> { // Emit KCFI operand bundle let kcfi_bundle = self.kcfi_operand_bundle(fn_attrs, fn_abi, instance, llfn); - if let Some(kcfi_bundle) = kcfi_bundle.as_deref() { + if let Some(kcfi_bundle) = kcfi_bundle.as_ref().map(|b| b.raw()) { bundles.push(kcfi_bundle); } @@ -1907,10 +1897,6 @@ impl<'a, 'll, 'tcx> Builder<'a, 'll, 'tcx> { hash: &'ll Value, bitmap_bits: &'ll Value, ) { - assert!( - crate::llvm_util::get_version() >= (19, 0, 0), - "MCDC intrinsics require LLVM 19 or later" - ); self.call_intrinsic("llvm.instrprof.mcdc.parameters", &[fn_name, hash, bitmap_bits]); } @@ -1922,10 +1908,6 @@ impl<'a, 'll, 'tcx> Builder<'a, 'll, 'tcx> { bitmap_index: &'ll Value, mcdc_temp: &'ll Value, ) { - assert!( - crate::llvm_util::get_version() >= (19, 0, 0), - "MCDC intrinsics require LLVM 19 or later" - ); let args = &[fn_name, hash, bitmap_index, mcdc_temp]; self.call_intrinsic("llvm.instrprof.mcdc.tvbitmap.update", args); } @@ -1937,10 +1919,6 @@ impl<'a, 'll, 'tcx> Builder<'a, 'll, 'tcx> { #[instrument(level = "debug", skip(self))] pub(crate) fn mcdc_condbitmap_update(&mut self, cond_index: &'ll Value, mcdc_temp: &'ll Value) { - assert!( - crate::llvm_util::get_version() >= (19, 0, 0), - "MCDC intrinsics require LLVM 19 or later" - ); let align = self.tcx.data_layout.i32_align.abi; let current_tv_index = self.load(self.cx.type_i32(), mcdc_temp, align); let new_tv_index = self.add(current_tv_index, cond_index); diff --git a/compiler/rustc_codegen_llvm/src/builder/autodiff.rs b/compiler/rustc_codegen_llvm/src/builder/autodiff.rs index 2c7899975e3..c5c13ac097a 100644 --- a/compiler/rustc_codegen_llvm/src/builder/autodiff.rs +++ b/compiler/rustc_codegen_llvm/src/builder/autodiff.rs @@ -3,11 +3,14 @@ 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::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; -use crate::builder::SBuilder; +use crate::builder::{SBuilder, UNNAMED}; use crate::context::SimpleCx; use crate::declare::declare_simple_fn; use crate::errors::{AutoDiffWithoutEnable, LlvmError}; @@ -17,14 +20,264 @@ 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>, + builder: &SBuilder<'ll, '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_dupv = cx.create_metadata("enzyme_dupv".to_string()).unwrap(); + let enzyme_dupnoneed = cx.create_metadata("enzyme_dupnoneed".to_string()).unwrap(); + let enzyme_dupnoneedv = cx.create_metadata("enzyme_dupnoneedv".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::Dualv => (enzyme_dupv, true), + DiffActivity::DualOnly => (enzyme_dupnoneed, true), + DiffActivity::DualvOnly => (enzyme_dupnoneedv, 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)); + if matches!(diff_activity, DiffActivity::Dualv) { + let next_outer_arg = outer_args[outer_pos + 1]; + let elem_bytes_size: u64 = match inputs[activity_pos + 1] { + DiffActivity::FakeActivitySize(Some(s)) => s.into(), + _ => bug!("incorrect Dualv handling recognized."), + }; + // stride: sizeof(T) * n_elems. + // n_elems is the next integer. + // Now we multiply `4 * next_outer_arg` to get the stride. + let mul = unsafe { + llvm::LLVMBuildMul( + builder.llbuilder, + cx.get_const_i64(elem_bytes_size), + next_outer_arg, + UNNAMED, + ) + }; + args.push(mul); + } + 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. + matches!(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); + + let iterations = + if matches!(diff_activity, DiffActivity::Dualv) { 1 } else { width as usize }; + + for i in 0..iterations { + 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 * iterations; + 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 { + // Enzyme returns a struct of style: + // `{ original_ret(if requested), float, float, ... }` + let mut struct_elements = vec![]; + if attrs.has_primal_ret() { + struct_elements.push(inner_ret_ty); + } + // Next, we push the list of active floats, since they will be lowered to `enzyme_out`, + // and therefore part of the return struct. + let param_tys = cx.func_params_types(fn_ty); + for (act, param_ty) in attrs.input_activity.iter().zip(param_tys) { + if matches!(act, DiffActivity::Active) { + // Now find the float type at position i based on the fn_ty, + // to know what (f16/f32/f64/...) to add to the struct. + struct_elements.push(param_ty); + } + } + ret_ty = cx.type_struct(&struct_elements, false); + } 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 @@ -42,9 +295,6 @@ 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", @@ -92,17 +342,9 @@ 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( @@ -119,6 +361,11 @@ fn generate_enzyme_call<'ll>( let attr = llvm::AttributeKind::NoInline.create_attr(cx.llcx); attributes::apply_to_llfn(ad_fn, Function, &[attr]); + // We add a made-up attribute just such that we can recognize it after AD to update + // (no)-inline attributes. We'll then also remove this attribute. + let enzyme_marker_attr = llvm::CreateAttrString(cx.llcx, "enzyme_marker"); + attributes::apply_to_llfn(outer_fn, Function, &[enzyme_marker_attr]); + // first, remove all calls from fnc let entry = llvm::LLVMGetFirstBasicBlock(outer_fn); let br = llvm::LLVMRustGetTerminator(entry); @@ -131,116 +378,32 @@ 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, + &builder, + 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. @@ -261,7 +424,26 @@ 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); + if attrs.width == 1 { + assert_eq!(cx.type_kind(call_ty), TypeKind::Struct); + } else { + assert_eq!(cx.type_kind(call_ty), TypeKind::Array); + } + llvm::LLVMBuildStore(&builder.llbuilder, call, sret_ptr); + } builder.ret_void(); } else { builder.ret(call); @@ -286,17 +468,17 @@ pub(crate) fn differentiate<'ll>( } let diag_handler = cgcx.create_dcx(); - let cx = SimpleCx { llmod: module.module_llvm.llmod(), llcx: module.module_llvm.llcx }; + + let cx = SimpleCx::new(module.module_llvm.llmod(), module.module_llvm.llcx, cgcx.pointer_size); // First of all, did the user try to use autodiff without using the -Zautodiff=Enable flag? 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. + // Here we replace the placeholder code with the actual autodiff code, which calls Enzyme. for item in diff_items.iter() { let name = item.source.clone(); let fn_def: Option<&llvm::Value> = cx.get_function(&name); diff --git a/compiler/rustc_codegen_llvm/src/common.rs b/compiler/rustc_codegen_llvm/src/common.rs index f17d98fa242..a6f277e4455 100644 --- a/compiler/rustc_codegen_llvm/src/common.rs +++ b/compiler/rustc_codegen_llvm/src/common.rs @@ -1,9 +1,11 @@ //! Code that is useful in various codegen modules. +use std::borrow::Borrow; + use libc::{c_char, c_uint}; use rustc_abi as abi; +use rustc_abi::HasDataLayout; use rustc_abi::Primitive::Pointer; -use rustc_abi::{AddressSpace, HasDataLayout}; use rustc_ast::Mutability; use rustc_codegen_ssa::common::TypeKind; use rustc_codegen_ssa::traits::*; @@ -18,6 +20,7 @@ use tracing::debug; use crate::consts::const_alloc_to_llvm; pub(crate) use crate::context::CodegenCx; +use crate::context::{GenericCx, SCx}; use crate::llvm::{self, BasicBlock, Bool, ConstantInt, False, Metadata, True}; use crate::type_::Type; use crate::value::Value; @@ -77,11 +80,11 @@ impl<'ll> Funclet<'ll> { } pub(crate) fn bundle(&self) -> &llvm::OperandBundle<'ll> { - &self.operand + self.operand.raw() } } -impl<'ll> BackendTypes for CodegenCx<'ll, '_> { +impl<'ll, CX: Borrow<SCx<'ll>>> BackendTypes for GenericCx<'ll, CX> { type Value = &'ll Value; type Metadata = &'ll Metadata; // FIXME(eddyb) replace this with a `Function` "subclass" of `Value`. @@ -118,7 +121,7 @@ impl<'ll> CodegenCx<'ll, '_> { } } -impl<'ll, 'tcx> ConstCodegenMethods<'tcx> for CodegenCx<'ll, 'tcx> { +impl<'ll, 'tcx> ConstCodegenMethods for CodegenCx<'ll, 'tcx> { fn const_null(&self, t: &'ll Type) -> &'ll Value { unsafe { llvm::LLVMConstNull(t) } } @@ -127,10 +130,6 @@ 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) } } @@ -209,28 +208,24 @@ impl<'ll, 'tcx> ConstCodegenMethods<'tcx> for CodegenCx<'ll, 'tcx> { } fn const_str(&self, s: &str) -> (&'ll Value, &'ll Value) { - let str_global = *self - .const_str_cache - .borrow_mut() - .raw_entry_mut() - .from_key(s) - .or_insert_with(|| { - let sc = self.const_bytes(s.as_bytes()); - let sym = self.generate_local_symbol_name("str"); - let g = self.define_global(&sym, self.val_ty(sc)).unwrap_or_else(|| { - bug!("symbol `{}` is already defined", sym); - }); - llvm::set_initializer(g, sc); - unsafe { - llvm::LLVMSetGlobalConstant(g, True); - llvm::LLVMSetUnnamedAddress(g, llvm::UnnamedAddr::Global); - } - llvm::set_linkage(g, llvm::Linkage::InternalLinkage); - // Cast to default address space if globals are in a different addrspace - let g = self.const_pointercast(g, self.type_ptr()); - (s.to_owned(), g) - }) - .1; + let mut const_str_cache = self.const_str_cache.borrow_mut(); + let str_global = const_str_cache.get(s).copied().unwrap_or_else(|| { + let sc = self.const_bytes(s.as_bytes()); + let sym = self.generate_local_symbol_name("str"); + let g = self.define_global(&sym, self.val_ty(sc)).unwrap_or_else(|| { + bug!("symbol `{}` is already defined", sym); + }); + llvm::set_initializer(g, sc); + unsafe { + llvm::LLVMSetGlobalConstant(g, True); + llvm::LLVMSetUnnamedAddress(g, llvm::UnnamedAddr::Global); + } + llvm::set_linkage(g, llvm::Linkage::InternalLinkage); + // Cast to default address space if globals are in a different addrspace + let g = self.const_pointercast(g, self.type_ptr()); + const_str_cache.insert(s.to_owned(), g); + g + }); let len = s.len(); (str_global, self.const_usize(len as u64)) } @@ -274,7 +269,8 @@ impl<'ll, 'tcx> ConstCodegenMethods<'tcx> for CodegenCx<'ll, 'tcx> { } Scalar::Ptr(ptr, _size) => { let (prov, offset) = ptr.into_parts(); - let (base_addr, base_addr_space) = match self.tcx.global_alloc(prov.alloc_id()) { + let global_alloc = self.tcx.global_alloc(prov.alloc_id()); + let base_addr = match global_alloc { GlobalAlloc::Memory(alloc) => { // For ZSTs directly codegen an aligned pointer. // This avoids generating a zero-sized constant value and actually needing a @@ -306,12 +302,10 @@ impl<'ll, 'tcx> ConstCodegenMethods<'tcx> for CodegenCx<'ll, 'tcx> { format!("alloc_{hash:032x}").as_bytes(), ); } - (value, AddressSpace::DATA) + value } } - GlobalAlloc::Function { instance, .. } => { - (self.get_fn_addr(instance), self.data_layout().instruction_address_space) - } + GlobalAlloc::Function { instance, .. } => self.get_fn_addr(instance), GlobalAlloc::VTable(ty, dyn_ty) => { let alloc = self .tcx @@ -324,14 +318,15 @@ impl<'ll, 'tcx> ConstCodegenMethods<'tcx> for CodegenCx<'ll, 'tcx> { .unwrap_memory(); let init = const_alloc_to_llvm(self, alloc, /*static*/ false); let value = self.static_addr_of_impl(init, alloc.inner().align, None); - (value, AddressSpace::DATA) + value } GlobalAlloc::Static(def_id) => { assert!(self.tcx.is_static(def_id)); assert!(!self.tcx.is_thread_local_static(def_id)); - (self.get_static(def_id), AddressSpace::DATA) + self.get_static(def_id) } }; + let base_addr_space = global_alloc.address_space(self); let llval = unsafe { llvm::LLVMConstInBoundsGEP2( self.type_i8(), @@ -350,7 +345,7 @@ impl<'ll, 'tcx> ConstCodegenMethods<'tcx> for CodegenCx<'ll, 'tcx> { } } - fn const_data_from_alloc(&self, alloc: ConstAllocation<'tcx>) -> Self::Value { + fn const_data_from_alloc(&self, alloc: ConstAllocation<'_>) -> Self::Value { const_alloc_to_llvm(self, alloc, /*static*/ false) } diff --git a/compiler/rustc_codegen_llvm/src/consts.rs b/compiler/rustc_codegen_llvm/src/consts.rs index 330e8a8f406..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}; @@ -12,11 +13,10 @@ 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 rustc_session::config::Lto; use tracing::{debug, instrument, trace}; use crate::common::{AsCCharPtr, CodegenCx}; @@ -129,7 +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>( @@ -172,8 +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 @@ -344,11 +368,11 @@ impl<'ll> CodegenCx<'ll, '_> { // Local definitions can never be imported, so we must not apply // the DLLImport annotation. && !dso_local - // ThinLTO can't handle this workaround in all cases, so we don't - // emit the attrs. Instead we make them unnecessary by disallowing - // dynamic linking when linker plugin based LTO is enabled. - && !self.tcx.sess.opts.cg.linker_plugin_lto.enabled() - && self.tcx.sess.lto() != Lto::Thin; + // Linker plugin ThinLTO doesn't create the self-dllimport Rust uses for rlibs + // as the code generation happens out of process. Instead we assume static linkage + // and disallow dynamic linking when linker plugin based LTO is enabled. + // Regular in-process ThinLTO doesn't need this workaround. + && !self.tcx.sess.opts.cg.linker_plugin_lto.enabled(); // If this assertion triggers, there's something wrong with commandline // argument validation. @@ -406,7 +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 @@ -490,7 +514,7 @@ impl<'ll> CodegenCx<'ll, '_> { llvm::LLVMMDStringInContext2(self.llcx, bytes.as_c_char_ptr(), bytes.len()); let data = [section, alloc]; let meta = llvm::LLVMMDNodeInContext2(self.llcx, data.as_ptr(), data.len()); - let val = llvm::LLVMMetadataAsValue(self.llcx, meta); + let val = self.get_metadata_value(meta); llvm::LLVMAddNamedMetadataOperand( self.llmod, c"wasm.custom_sections".as_ptr(), diff --git a/compiler/rustc_codegen_llvm/src/context.rs b/compiler/rustc_codegen_llvm/src/context.rs index e7952bc95e7..ed50515b707 100644 --- a/compiler/rustc_codegen_llvm/src/context.rs +++ b/compiler/rustc_codegen_llvm/src/context.rs @@ -1,10 +1,11 @@ use std::borrow::Borrow; use std::cell::{Cell, RefCell}; use std::ffi::{CStr, c_char, c_uint}; +use std::marker::PhantomData; use std::ops::Deref; use std::str; -use rustc_abi::{HasDataLayout, TargetDataLayout, VariantIdx}; +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; @@ -27,34 +28,36 @@ 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; use crate::back::write::to_llvm_code_model; use crate::callee::get_fn; -use crate::common::{self, AsCCharPtr}; +use crate::common::AsCCharPtr; use crate::debuginfo::metadata::apply_vcall_visibility_metadata; -use crate::llvm::{Metadata, MetadataType}; +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 /// 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(crate) struct SCx<'ll> { pub llmod: &'ll llvm::Module, pub llcx: &'ll llvm::Context, + pub isize_ty: &'ll Type, } -impl<'ll> Borrow<SimpleCx<'ll>> for CodegenCx<'ll, '_> { - fn borrow(&self) -> &SimpleCx<'ll> { +impl<'ll> Borrow<SCx<'ll>> for FullCx<'ll, '_> { + fn borrow(&self) -> &SCx<'ll> { &self.scx } } -impl<'ll, 'tcx> Deref for CodegenCx<'ll, 'tcx> { +impl<'ll, 'tcx> Deref for FullCx<'ll, 'tcx> { type Target = SimpleCx<'ll>; #[inline] @@ -63,10 +66,25 @@ impl<'ll, 'tcx> Deref for CodegenCx<'ll, 'tcx> { } } +pub(crate) struct GenericCx<'ll, T: Borrow<SCx<'ll>>>(T, PhantomData<SCx<'ll>>); + +impl<'ll, T: Borrow<SCx<'ll>>> Deref for GenericCx<'ll, T> { + type Target = T; + + #[inline] + fn deref(&self) -> &Self::Target { + &self.0 + } +} + +pub(crate) type SimpleCx<'ll> = GenericCx<'ll, SCx<'ll>>; + /// 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(crate) type CodegenCx<'ll, 'tcx> = GenericCx<'ll, FullCx<'ll, 'tcx>>; + +pub(crate) struct FullCx<'ll, 'tcx> { pub tcx: TyCtxt<'tcx>, pub scx: SimpleCx<'ll>, pub use_dll_storage_attrs: bool, @@ -104,8 +122,6 @@ pub(crate) struct CodegenCx<'ll, 'tcx> { /// Mapping of scalar types to llvm types. pub scalar_lltypes: RefCell<FxHashMap<Ty<'tcx>, &'ll Type>>, - pub isize_ty: &'ll Type, - /// Extra per-CGU codegen state needed when coverage instrumentation is enabled. pub coverage_cx: Option<coverageinfo::CguCoverageContext<'ll, 'tcx>>, pub dbg_cx: Option<debuginfo::CodegenUnitDebugContext<'ll, 'tcx>>, @@ -148,23 +164,6 @@ pub(crate) unsafe fn create_module<'ll>( let mut target_data_layout = sess.target.data_layout.to_string(); let llvm_version = llvm_util::get_version(); - if llvm_version < (19, 0, 0) { - if sess.target.arch == "aarch64" || sess.target.arch.starts_with("arm64") { - // LLVM 19 sets -Fn32 in its data layout string for 64-bit ARM - // Earlier LLVMs leave this default, so remove it. - // See https://github.com/llvm/llvm-project/pull/90702 - target_data_layout = target_data_layout.replace("-Fn32", ""); - } - } - - if llvm_version < (19, 0, 0) { - if sess.target.arch == "loongarch64" { - // LLVM 19 updates the LoongArch64 data layout. - // See https://github.com/llvm/llvm-project/pull/93814 - target_data_layout = target_data_layout.replace("-n32:64", "-n64"); - } - } - if llvm_version < (20, 0, 0) { if sess.target.arch == "aarch64" || sess.target.arch.starts_with("arm64") { // LLVM 20 defines three additional address spaces for alternate @@ -205,7 +204,7 @@ pub(crate) unsafe fn create_module<'ll>( { let tm = crate::back::write::create_informational_target_machine(tcx.sess, false); unsafe { - llvm::LLVMRustSetDataLayoutFromTargetMachine(llmod, &tm); + llvm::LLVMRustSetDataLayoutFromTargetMachine(llmod, tm.raw()); } let llvm_data_layout = unsafe { llvm::LLVMGetDataLayoutStr(llmod) }; @@ -312,6 +311,22 @@ 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. @@ -579,33 +594,33 @@ impl<'ll, 'tcx> CodegenCx<'ll, 'tcx> { None }; - let isize_ty = Type::ix_llcx(llcx, tcx.data_layout.pointer_size.bits()); - - CodegenCx { - tcx, - scx: SimpleCx { llcx, llmod }, - use_dll_storage_attrs, - tls_model, - codegen_unit, - instances: Default::default(), - vtables: Default::default(), - const_str_cache: Default::default(), - const_globals: Default::default(), - statics_to_rauw: RefCell::new(Vec::new()), - used_statics: RefCell::new(Vec::new()), - compiler_used_statics: RefCell::new(Vec::new()), - type_lowering: Default::default(), - scalar_lltypes: Default::default(), - isize_ty, - coverage_cx, - dbg_cx, - eh_personality: Cell::new(None), - eh_catch_typeinfo: Cell::new(None), - rust_try_fn: Cell::new(None), - intrinsics: Default::default(), - local_gen_sym_counter: Cell::new(0), - renamed_statics: Default::default(), - } + GenericCx( + FullCx { + tcx, + scx: SimpleCx::new(llmod, llcx, tcx.data_layout.pointer_size), + use_dll_storage_attrs, + tls_model, + codegen_unit, + instances: Default::default(), + vtables: Default::default(), + const_str_cache: Default::default(), + const_globals: Default::default(), + statics_to_rauw: RefCell::new(Vec::new()), + used_statics: RefCell::new(Vec::new()), + compiler_used_statics: RefCell::new(Vec::new()), + type_lowering: Default::default(), + scalar_lltypes: Default::default(), + coverage_cx, + dbg_cx, + eh_personality: Cell::new(None), + eh_catch_typeinfo: Cell::new(None), + rust_try_fn: Cell::new(None), + intrinsics: Default::default(), + local_gen_sym_counter: Cell::new(0), + renamed_statics: Default::default(), + }, + PhantomData, + ) } pub(crate) fn statics_to_rauw(&self) -> &RefCell<Vec<(&'ll Value, &'ll Value)>> { @@ -629,23 +644,49 @@ impl<'ll, 'tcx> CodegenCx<'ll, 'tcx> { } } 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, + llcx: &'ll llvm::Context, + pointer_size: Size, + ) -> Self { + let isize_ty = llvm::Type::ix_llcx(llcx, pointer_size.bits()); + Self(SCx { llmod, llcx, isize_ty }, PhantomData) + } +} +impl<'ll, CX: Borrow<SCx<'ll>>> GenericCx<'ll, CX> { pub(crate) fn get_metadata_value(&self, metadata: &'ll Metadata) -> &'ll Value { - unsafe { llvm::LLVMMetadataAsValue(self.llcx, metadata) } + 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.llmod, name.as_ptr()) } + unsafe { llvm::LLVMGetNamedFunction((**self).borrow().llmod, name.as_ptr()) } } - pub(crate) fn get_md_kind_id(&self, name: &str) -> u32 { + pub(crate) fn get_md_kind_id(&self, name: &str) -> llvm::MetadataKindId { unsafe { llvm::LLVMGetMDKindIDInContext( - self.llcx, + self.llcx(), name.as_ptr() as *const c_char, name.len() as c_uint, ) @@ -654,12 +695,18 @@ impl<'ll> SimpleCx<'ll> { pub(crate) fn create_metadata(&self, name: String) -> Option<&'ll Metadata> { Some(unsafe { - llvm::LLVMMDStringInContext2(self.llcx, name.as_ptr() as *const c_char, name.len()) + 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() } + pub(crate) fn get_functions(&self) -> Vec<&'ll Value> { + let mut functions = vec![]; + let mut func = unsafe { llvm::LLVMGetFirstFunction(self.llmod()) }; + while let Some(f) = func { + functions.push(f); + func = unsafe { llvm::LLVMGetNextFunction(f) } + } + functions } } @@ -1108,6 +1155,18 @@ 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); @@ -1152,10 +1211,8 @@ impl<'ll> CodegenCx<'ll, '_> { if self.sess().instrument_coverage() { ifn!("llvm.instrprof.increment", fn(ptr, t_i64, t_i32, t_i32) -> void); - if crate::llvm_util::get_version() >= (19, 0, 0) { - ifn!("llvm.instrprof.mcdc.parameters", fn(ptr, t_i64, t_i32) -> void); - ifn!("llvm.instrprof.mcdc.tvbitmap.update", fn(ptr, t_i64, t_i32, ptr) -> void); - } + ifn!("llvm.instrprof.mcdc.parameters", fn(ptr, t_i64, t_i32) -> void); + ifn!("llvm.instrprof.mcdc.tvbitmap.update", fn(ptr, t_i64, t_i32, ptr) -> void); } ifn!("llvm.type.test", fn(ptr, t_metadata) -> i1); @@ -1181,7 +1238,7 @@ 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)); @@ -1203,27 +1260,18 @@ impl CodegenCx<'_, '_> { name.push_str(&(idx as u64).to_base(ALPHANUMERIC_ONLY)); name } - - /// 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); - } - } } -// 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)] +impl<'ll, CX: Borrow<SCx<'ll>>> GenericCx<'ll, CX> { /// 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); - } + pub(crate) fn set_metadata<'a>( + &self, + val: &'a Value, + kind_id: impl Into<llvm::MetadataKindId>, + md: &'ll Metadata, + ) { + let node = self.get_metadata_value(md); + llvm::LLVMSetMetadata(val, kind_id.into(), node); } } 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.rs b/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen.rs index 9a2473d6cf2..55b1e728b70 100644 --- a/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen.rs +++ b/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen.rs @@ -5,15 +5,11 @@ use rustc_abi::Align; use rustc_codegen_ssa::traits::{ BaseTypeCodegenMethods, ConstCodegenMethods, StaticCodegenMethods, }; -use rustc_data_structures::fx::{FxHashSet, FxIndexMap}; -use rustc_hir::def_id::{DefId, LocalDefId}; +use rustc_data_structures::fx::FxIndexMap; use rustc_index::IndexVec; -use rustc_middle::mir; -use rustc_middle::mir::mono::MonoItemPartitions; -use rustc_middle::ty::{self, TyCtxt}; +use rustc_middle::ty::TyCtxt; use rustc_session::RemapFileNameExt; use rustc_session::config::RemapPathScopeComponents; -use rustc_span::def_id::DefIdSet; use rustc_span::{SourceFile, StableSourceFileId}; use tracing::debug; @@ -24,6 +20,7 @@ use crate::llvm; mod covfun; mod spans; +mod unused; /// Generates and exports the coverage map, which is embedded in special /// linker sections in the final binary. @@ -56,13 +53,6 @@ pub(crate) fn finalize(cx: &CodegenCx<'_, '_>) { None => return, }; - // The order of entries in this global file table needs to be deterministic, - // and ideally should also be independent of the details of stable-hashing, - // because coverage tests snapshots (`.cov-map`) can observe the order and - // would need to be re-blessed if it changes. As long as those requirements - // are satisfied, the order can be arbitrary. - let mut global_file_table = GlobalFileTable::new(); - let mut covfun_records = instances_used .iter() .copied() @@ -70,18 +60,13 @@ pub(crate) fn finalize(cx: &CodegenCx<'_, '_>) { // order that doesn't depend on the stable-hash-based order in which // instances were visited during codegen. .sorted_by_cached_key(|&instance| tcx.symbol_name(instance).name) - .filter_map(|instance| prepare_covfun_record(tcx, &mut global_file_table, instance, true)) + .filter_map(|instance| prepare_covfun_record(tcx, instance, true)) .collect::<Vec<_>>(); // In a single designated CGU, also prepare covfun records for functions // in this crate that were instrumented for coverage, but are unused. if cx.codegen_unit.is_code_coverage_dead_code_cgu() { - let mut unused_instances = gather_unused_function_instances(cx); - // Sort the unused instances by symbol name, for the same reason as the used ones. - unused_instances.sort_by_cached_key(|&instance| tcx.symbol_name(instance).name); - covfun_records.extend(unused_instances.into_iter().filter_map(|instance| { - prepare_covfun_record(tcx, &mut global_file_table, instance, false) - })); + unused::prepare_covfun_records_for_unused_functions(cx, &mut covfun_records); } // If there are no covfun records for this CGU, don't generate a covmap record. @@ -93,91 +78,88 @@ pub(crate) fn finalize(cx: &CodegenCx<'_, '_>) { return; } - // Encode all filenames referenced by coverage mappings in this CGU. - let filenames_buffer = global_file_table.make_filenames_buffer(tcx); - // The `llvm-cov` tool uses this hash to associate each covfun record with - // its corresponding filenames table, since the final binary will typically - // contain multiple covmap records from different compilation units. - let filenames_hash = llvm_cov::hash_bytes(&filenames_buffer); - - let mut unused_function_names = vec![]; + // Prepare the global file table for this CGU, containing all paths needed + // by one or more covfun records. + let global_file_table = + GlobalFileTable::build(tcx, covfun_records.iter().flat_map(|c| c.all_source_files())); for covfun in &covfun_records { - unused_function_names.extend(covfun.mangled_function_name_if_unused()); - - covfun::generate_covfun_record(cx, filenames_hash, covfun) - } - - // For unused functions, we need to take their mangled names and store them - // in a specially-named global array. LLVM's `InstrProfiling` pass will - // detect this global and include those names in its `__llvm_prf_names` - // section. (See `llvm/lib/Transforms/Instrumentation/InstrProfiling.cpp`.) - if !unused_function_names.is_empty() { - assert!(cx.codegen_unit.is_code_coverage_dead_code_cgu()); - - let name_globals = unused_function_names - .into_iter() - .map(|mangled_function_name| cx.const_str(mangled_function_name).0) - .collect::<Vec<_>>(); - let initializer = cx.const_array(cx.type_ptr(), &name_globals); - - let array = llvm::add_global(cx.llmod, cx.val_ty(initializer), c"__llvm_coverage_names"); - llvm::set_global_constant(array, true); - llvm::set_linkage(array, llvm::Linkage::InternalLinkage); - llvm::set_initializer(array, initializer); + covfun::generate_covfun_record(cx, &global_file_table, covfun) } // Generate the coverage map header, which contains the filenames used by // this CGU's coverage mappings, and store it in a well-known global. // (This is skipped if we returned early due to having no covfun records.) - generate_covmap_record(cx, covmap_version, &filenames_buffer); + generate_covmap_record(cx, covmap_version, &global_file_table.filenames_buffer); } -/// Maps "global" (per-CGU) file ID numbers to their underlying source files. +/// Maps "global" (per-CGU) file ID numbers to their underlying source file paths. +#[derive(Debug)] struct GlobalFileTable { /// This "raw" table doesn't include the working dir, so a file's /// global ID is its index in this set **plus one**. - raw_file_table: FxIndexMap<StableSourceFileId, Arc<SourceFile>>, + raw_file_table: FxIndexMap<StableSourceFileId, String>, + + /// The file table in encoded form (possibly compressed), which can be + /// included directly in this CGU's `__llvm_covmap` record. + filenames_buffer: Vec<u8>, + + /// Truncated hash of the bytes in `filenames_buffer`. + /// + /// The `llvm-cov` tool uses this hash to associate each covfun record with + /// its corresponding filenames table, since the final binary will typically + /// contain multiple covmap records from different compilation units. + filenames_hash: u64, } impl GlobalFileTable { - fn new() -> Self { - Self { raw_file_table: FxIndexMap::default() } - } + /// Builds a "global file table" for this CGU, mapping numeric IDs to + /// path strings. + fn build<'a>(tcx: TyCtxt<'_>, all_files: impl Iterator<Item = &'a SourceFile>) -> Self { + let mut raw_file_table = FxIndexMap::default(); + + for file in all_files { + raw_file_table.entry(file.stable_id).or_insert_with(|| { + file.name + .for_scope(tcx.sess, RemapPathScopeComponents::MACRO) + .to_string_lossy() + .into_owned() + }); + } - fn global_file_id_for_file(&mut self, file: &Arc<SourceFile>) -> GlobalFileId { - // Ensure the given file has a table entry, and get its index. - let entry = self.raw_file_table.entry(file.stable_id); - let raw_id = entry.index(); - entry.or_insert_with(|| Arc::clone(file)); + // FIXME(Zalathar): Consider sorting the file table here, but maybe + // only after adding filename support to coverage-dump, so that the + // table order isn't directly visible in `.coverage-map` snapshots. - // The raw file table doesn't include an entry for the working dir - // (which has ID 0), so add 1 to get the correct ID. - GlobalFileId::from_usize(raw_id + 1) - } + let mut table = Vec::with_capacity(raw_file_table.len() + 1); - fn make_filenames_buffer(&self, tcx: TyCtxt<'_>) -> Vec<u8> { - let mut table = Vec::with_capacity(self.raw_file_table.len() + 1); - - // LLVM Coverage Mapping Format version 6 (zero-based encoded as 5) - // requires setting the first filename to the compilation directory. - // Since rustc generates coverage maps with relative paths, the - // compilation directory can be combined with the relative paths - // to get absolute paths, if needed. - table.push( - tcx.sess - .opts - .working_dir - .for_scope(tcx.sess, RemapPathScopeComponents::MACRO) - .to_string_lossy(), - ); + // Since version 6 of the LLVM coverage mapping format, the first entry + // in the global file table is treated as a base directory, used to + // resolve any other entries that are stored as relative paths. + let base_dir = tcx + .sess + .opts + .working_dir + .for_scope(tcx.sess, RemapPathScopeComponents::MACRO) + .to_string_lossy(); + table.push(base_dir.as_ref()); // Add the regular entries after the base directory. - table.extend(self.raw_file_table.values().map(|file| { - file.name.for_scope(tcx.sess, RemapPathScopeComponents::MACRO).to_string_lossy() - })); + table.extend(raw_file_table.values().map(|name| name.as_str())); - llvm_cov::write_filenames_to_buffer(&table) + // Encode the file table into a buffer, and get the hash of its encoded + // bytes, so that we can embed that hash in `__llvm_covfun` records. + let filenames_buffer = llvm_cov::write_filenames_to_buffer(&table); + let filenames_hash = llvm_cov::hash_bytes(&filenames_buffer); + + Self { raw_file_table, filenames_buffer, filenames_hash } + } + + fn get_existing_id(&self, file: &SourceFile) -> Option<GlobalFileId> { + let raw_id = self.raw_file_table.get_index_of(&file.stable_id)?; + // The raw file table doesn't include an entry for the base dir + // (which has ID 0), so add 1 to get the correct ID. + Some(GlobalFileId::from_usize(raw_id + 1)) } } @@ -193,26 +175,31 @@ rustc_index::newtype_index! { struct LocalFileId {} } -/// Holds a mapping from "local" (per-function) file IDs to "global" (per-CGU) -/// file IDs. +/// Holds a mapping from "local" (per-function) file IDs to their corresponding +/// source files. #[derive(Debug, Default)] struct VirtualFileMapping { - local_to_global: IndexVec<LocalFileId, GlobalFileId>, - global_to_local: FxIndexMap<GlobalFileId, LocalFileId>, + local_file_table: IndexVec<LocalFileId, Arc<SourceFile>>, } impl VirtualFileMapping { - fn local_id_for_global(&mut self, global_file_id: GlobalFileId) -> LocalFileId { - *self - .global_to_local - .entry(global_file_id) - .or_insert_with(|| self.local_to_global.push(global_file_id)) + fn push_file(&mut self, source_file: &Arc<SourceFile>) -> LocalFileId { + self.local_file_table.push(Arc::clone(source_file)) } - fn to_vec(&self) -> Vec<u32> { - // This clone could be avoided by transmuting `&[GlobalFileId]` to `&[u32]`, - // but it isn't hot or expensive enough to justify the extra unsafety. - self.local_to_global.iter().map(|&global| GlobalFileId::as_u32(global)).collect() + /// Resolves all of the filenames in this local file mapping to a list of + /// global file IDs in its CGU, for inclusion in this function's + /// `__llvm_covfun` record. + /// + /// The global file IDs are returned as `u32` to make FFI easier. + fn resolve_all(&self, global_file_table: &GlobalFileTable) -> Option<Vec<u32>> { + self.local_file_table + .iter() + .map(|file| try { + let id = global_file_table.get_existing_id(file)?; + GlobalFileId::as_u32(id) + }) + .collect::<Option<Vec<_>>>() } } @@ -249,121 +236,3 @@ fn generate_covmap_record<'ll>(cx: &CodegenCx<'ll, '_>, version: u32, filenames_ cx.add_used_global(covmap_global); } - -/// Each CGU will normally only emit coverage metadata for the functions that it actually generates. -/// But since we don't want unused functions to disappear from coverage reports, we also scan for -/// functions that were instrumented but are not participating in codegen. -/// -/// These unused functions don't need to be codegenned, but we do need to add them to the function -/// coverage map (in a single designated CGU) so that we still emit coverage mappings for them. -/// We also end up adding their symbol names to a special global array that LLVM will include in -/// its embedded coverage data. -fn gather_unused_function_instances<'tcx>(cx: &CodegenCx<'_, 'tcx>) -> Vec<ty::Instance<'tcx>> { - assert!(cx.codegen_unit.is_code_coverage_dead_code_cgu()); - - let tcx = cx.tcx; - let usage = prepare_usage_sets(tcx); - - let is_unused_fn = |def_id: LocalDefId| -> bool { - // Usage sets expect `DefId`, so convert from `LocalDefId`. - let d: DefId = LocalDefId::to_def_id(def_id); - // To be potentially eligible for "unused function" mappings, a definition must: - // - Be eligible for coverage instrumentation - // - Not participate directly in codegen (or have lost all its coverage statements) - // - Not have any coverage statements inlined into codegenned functions - tcx.is_eligible_for_coverage(def_id) - && (!usage.all_mono_items.contains(&d) || usage.missing_own_coverage.contains(&d)) - && !usage.used_via_inlining.contains(&d) - }; - - // FIXME(#79651): Consider trying to filter out dummy instantiations of - // unused generic functions from library crates, because they can produce - // "unused instantiation" in coverage reports even when they are actually - // used by some downstream crate in the same binary. - - tcx.mir_keys(()) - .iter() - .copied() - .filter(|&def_id| is_unused_fn(def_id)) - .map(|def_id| make_dummy_instance(tcx, def_id)) - .collect::<Vec<_>>() -} - -struct UsageSets<'tcx> { - all_mono_items: &'tcx DefIdSet, - used_via_inlining: FxHashSet<DefId>, - missing_own_coverage: FxHashSet<DefId>, -} - -/// 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 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 = codegen_units - .iter() - .flat_map(|cgu| cgu.items().keys()) - .filter_map(|item| match item { - mir::mono::MonoItem::Fn(instance) => Some(instance), - mir::mono::MonoItem::Static(_) | mir::mono::MonoItem::GlobalAsm(_) => None, - }) - // We only need one arbitrary instance per definition. - .filter(move |instance| def_ids_seen.insert(instance.def_id())) - .map(|instance| { - // We don't care about the instance, just its underlying MIR. - let body = tcx.instance_mir(instance.def); - (instance.def_id(), body) - }); - - // Functions whose coverage statements were found inlined into other functions. - let mut used_via_inlining = FxHashSet::default(); - // Functions that were instrumented, but had all of their coverage statements - // removed by later MIR transforms (e.g. UnreachablePropagation). - let mut missing_own_coverage = FxHashSet::default(); - - for (def_id, body) in def_and_mir_for_all_mono_fns { - let mut saw_own_coverage = false; - - // Inspect every coverage statement in the function's MIR. - for stmt in body - .basic_blocks - .iter() - .flat_map(|block| &block.statements) - .filter(|stmt| matches!(stmt.kind, mir::StatementKind::Coverage(_))) - { - if let Some(inlined) = stmt.source_info.scope.inlined_instance(&body.source_scopes) { - // This coverage statement was inlined from another function. - used_via_inlining.insert(inlined.def_id()); - } else { - // Non-inlined coverage statements belong to the enclosing function. - saw_own_coverage = true; - } - } - - if !saw_own_coverage && body.function_coverage_info.is_some() { - missing_own_coverage.insert(def_id); - } - } - - UsageSets { all_mono_items, used_via_inlining, missing_own_coverage } -} - -fn make_dummy_instance<'tcx>(tcx: TyCtxt<'tcx>, local_def_id: LocalDefId) -> ty::Instance<'tcx> { - let def_id = local_def_id.to_def_id(); - - // Make a dummy instance that fills in all generics with placeholders. - ty::Instance::new( - def_id, - ty::GenericArgs::for_item(tcx, def_id, |param, _| { - if let ty::GenericParamDefKind::Lifetime = param.kind { - tcx.lifetimes.re_erased.into() - } else { - tcx.mk_param_from_def(param) - } - }), - ) -} diff --git a/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen/covfun.rs b/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen/covfun.rs index c53ea6d4666..7bdbc685952 100644 --- a/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen/covfun.rs +++ b/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen/covfun.rs @@ -5,17 +5,18 @@ //! [^win]: On Windows the section name is `.lcovfun`. use std::ffi::CString; +use std::sync::Arc; use rustc_abi::Align; use rustc_codegen_ssa::traits::{ - BaseTypeCodegenMethods, ConstCodegenMethods, StaticCodegenMethods, + BaseTypeCodegenMethods as _, ConstCodegenMethods, StaticCodegenMethods, }; use rustc_middle::mir::coverage::{ BasicCoverageBlock, CovTerm, CoverageIdsInfo, Expression, FunctionCoverageInfo, Mapping, MappingKind, Op, }; use rustc_middle::ty::{Instance, TyCtxt}; -use rustc_span::Span; +use rustc_span::{SourceFile, Span}; use rustc_target::spec::HasTargetSpec; use tracing::debug; @@ -38,16 +39,15 @@ pub(crate) struct CovfunRecord<'tcx> { } impl<'tcx> CovfunRecord<'tcx> { - /// FIXME(Zalathar): Make this the responsibility of the code that determines - /// which functions are unused. - pub(crate) fn mangled_function_name_if_unused(&self) -> Option<&'tcx str> { - (!self.is_used).then_some(self.mangled_function_name) + /// Iterator that yields all source files referred to by this function's + /// coverage mappings. Used to build the global file table for the CGU. + pub(crate) fn all_source_files(&self) -> impl Iterator<Item = &SourceFile> { + self.virtual_file_mapping.local_file_table.iter().map(Arc::as_ref) } } pub(crate) fn prepare_covfun_record<'tcx>( tcx: TyCtxt<'tcx>, - global_file_table: &mut GlobalFileTable, instance: Instance<'tcx>, is_used: bool, ) -> Option<CovfunRecord<'tcx>> { @@ -65,7 +65,7 @@ pub(crate) fn prepare_covfun_record<'tcx>( regions: ffi::Regions::default(), }; - fill_region_tables(tcx, global_file_table, fn_cov_info, ids_info, &mut covfun); + fill_region_tables(tcx, fn_cov_info, ids_info, &mut covfun); if covfun.regions.has_no_regions() { debug!(?covfun, "function has no mappings to embed; skipping"); @@ -100,28 +100,37 @@ fn prepare_expressions(ids_info: &CoverageIdsInfo) -> Vec<ffi::CounterExpression /// Populates the mapping region tables in the current function's covfun record. fn fill_region_tables<'tcx>( tcx: TyCtxt<'tcx>, - global_file_table: &mut GlobalFileTable, fn_cov_info: &'tcx FunctionCoverageInfo, ids_info: &'tcx CoverageIdsInfo, covfun: &mut CovfunRecord<'tcx>, ) { - // Currently a function's mappings must all be in the same file as its body span. + // 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()); - - // Look up the global file ID for that file. - let global_file_id = global_file_table.global_file_id_for_file(&source_file); - - // Associate that global file ID with a local file ID for this function. - let local_file_id = covfun.virtual_file_mapping.local_id_for_global(global_file_id); + let 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()); - let ffi::Regions { code_regions, branch_regions, mcdc_branch_regions, mcdc_decision_regions } = - &mut covfun.regions; + let local_file_id = covfun.virtual_file_mapping.push_file(&source_file); - 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 +145,8 @@ 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 } => { @@ -182,7 +182,7 @@ fn fill_region_tables<'tcx>( /// as a global variable in the `__llvm_covfun` section. pub(crate) fn generate_covfun_record<'tcx>( cx: &CodegenCx<'_, 'tcx>, - filenames_hash: u64, + global_file_table: &GlobalFileTable, covfun: &CovfunRecord<'tcx>, ) { let &CovfunRecord { @@ -194,12 +194,19 @@ pub(crate) fn generate_covfun_record<'tcx>( ref regions, } = covfun; + let Some(local_file_table) = virtual_file_mapping.resolve_all(global_file_table) else { + debug_assert!( + false, + "all local files should be present in the global file table: \ + global_file_table = {global_file_table:?}, \ + virtual_file_mapping = {virtual_file_mapping:?}" + ); + return; + }; + // Encode the function's coverage mappings into a buffer. - let coverage_mapping_buffer = llvm_cov::write_function_mappings_to_buffer( - &virtual_file_mapping.to_vec(), - expressions, - regions, - ); + let coverage_mapping_buffer = + llvm_cov::write_function_mappings_to_buffer(&local_file_table, expressions, regions); // A covfun record consists of four target-endian integers, followed by the // encoded mapping data in bytes. Note that the length field is 32 bits. @@ -212,7 +219,7 @@ pub(crate) fn generate_covfun_record<'tcx>( cx.const_u64(func_name_hash), cx.const_u32(coverage_mapping_buffer.len() as u32), cx.const_u64(source_hash), - cx.const_u64(filenames_hash), + cx.const_u64(global_file_table.filenames_hash), cx.const_bytes(&coverage_mapping_buffer), ], // This struct needs to be packed, so that the 32-bit length field 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/coverageinfo/mapgen/unused.rs b/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen/unused.rs new file mode 100644 index 00000000000..68f60f169b5 --- /dev/null +++ b/compiler/rustc_codegen_llvm/src/coverageinfo/mapgen/unused.rs @@ -0,0 +1,170 @@ +use rustc_codegen_ssa::traits::{BaseTypeCodegenMethods, ConstCodegenMethods}; +use rustc_data_structures::fx::FxHashSet; +use rustc_hir::def_id::{DefId, LocalDefId}; +use rustc_middle::mir; +use rustc_middle::mir::mono::MonoItemPartitions; +use rustc_middle::ty::{self, TyCtxt}; +use rustc_span::def_id::DefIdSet; + +use crate::common::CodegenCx; +use crate::coverageinfo::mapgen::covfun::{CovfunRecord, prepare_covfun_record}; +use crate::llvm; + +/// Each CGU will normally only emit coverage metadata for the functions that it actually generates. +/// But since we don't want unused functions to disappear from coverage reports, we also scan for +/// functions that were instrumented but are not participating in codegen. +/// +/// These unused functions don't need to be codegenned, but we do need to add them to the function +/// coverage map (in a single designated CGU) so that we still emit coverage mappings for them. +/// We also end up adding their symbol names to a special global array that LLVM will include in +/// its embedded coverage data. +pub(crate) fn prepare_covfun_records_for_unused_functions<'tcx>( + cx: &CodegenCx<'_, 'tcx>, + covfun_records: &mut Vec<CovfunRecord<'tcx>>, +) { + assert!(cx.codegen_unit.is_code_coverage_dead_code_cgu()); + + let mut unused_instances = gather_unused_function_instances(cx); + // Sort the unused instances by symbol name, so that their order isn't hash-sensitive. + unused_instances.sort_by_key(|instance| instance.symbol_name); + + // Try to create a covfun record for each unused function. + let mut name_globals = Vec::with_capacity(unused_instances.len()); + covfun_records.extend(unused_instances.into_iter().filter_map(|unused| try { + let record = prepare_covfun_record(cx.tcx, unused.instance, false)?; + // If successful, also store its symbol name in a global constant. + name_globals.push(cx.const_str(unused.symbol_name.name).0); + record + })); + + // Store the names of unused functions in a specially-named global array. + // LLVM's `InstrProfilling` pass will detect this array, and include the + // referenced names in its `__llvm_prf_names` section. + // (See `llvm/lib/Transforms/Instrumentation/InstrProfiling.cpp`.) + if !name_globals.is_empty() { + let initializer = cx.const_array(cx.type_ptr(), &name_globals); + + let array = llvm::add_global(cx.llmod, cx.val_ty(initializer), c"__llvm_coverage_names"); + llvm::set_global_constant(array, true); + llvm::set_linkage(array, llvm::Linkage::InternalLinkage); + llvm::set_initializer(array, initializer); + } +} + +/// Holds a dummy function instance along with its symbol name, to avoid having +/// to repeatedly query for the name. +struct UnusedInstance<'tcx> { + instance: ty::Instance<'tcx>, + symbol_name: ty::SymbolName<'tcx>, +} + +fn gather_unused_function_instances<'tcx>(cx: &CodegenCx<'_, 'tcx>) -> Vec<UnusedInstance<'tcx>> { + assert!(cx.codegen_unit.is_code_coverage_dead_code_cgu()); + + let tcx = cx.tcx; + let usage = prepare_usage_sets(tcx); + + let is_unused_fn = |def_id: LocalDefId| -> bool { + // Usage sets expect `DefId`, so convert from `LocalDefId`. + let d: DefId = LocalDefId::to_def_id(def_id); + // To be potentially eligible for "unused function" mappings, a definition must: + // - Be eligible for coverage instrumentation + // - Not participate directly in codegen (or have lost all its coverage statements) + // - Not have any coverage statements inlined into codegenned functions + tcx.is_eligible_for_coverage(def_id) + && (!usage.all_mono_items.contains(&d) || usage.missing_own_coverage.contains(&d)) + && !usage.used_via_inlining.contains(&d) + }; + + // FIXME(#79651): Consider trying to filter out dummy instantiations of + // unused generic functions from library crates, because they can produce + // "unused instantiation" in coverage reports even when they are actually + // used by some downstream crate in the same binary. + + tcx.mir_keys(()) + .iter() + .copied() + .filter(|&def_id| is_unused_fn(def_id)) + .map(|def_id| make_dummy_instance(tcx, def_id)) + .map(|instance| UnusedInstance { instance, symbol_name: tcx.symbol_name(instance) }) + .collect::<Vec<_>>() +} + +struct UsageSets<'tcx> { + all_mono_items: &'tcx DefIdSet, + used_via_inlining: FxHashSet<DefId>, + missing_own_coverage: FxHashSet<DefId>, +} + +/// 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 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 = codegen_units + .iter() + .flat_map(|cgu| cgu.items().keys()) + .filter_map(|item| match item { + mir::mono::MonoItem::Fn(instance) => Some(instance), + mir::mono::MonoItem::Static(_) | mir::mono::MonoItem::GlobalAsm(_) => None, + }) + // We only need one arbitrary instance per definition. + .filter(move |instance| def_ids_seen.insert(instance.def_id())) + .map(|instance| { + // We don't care about the instance, just its underlying MIR. + let body = tcx.instance_mir(instance.def); + (instance.def_id(), body) + }); + + // Functions whose coverage statements were found inlined into other functions. + let mut used_via_inlining = FxHashSet::default(); + // Functions that were instrumented, but had all of their coverage statements + // removed by later MIR transforms (e.g. UnreachablePropagation). + let mut missing_own_coverage = FxHashSet::default(); + + for (def_id, body) in def_and_mir_for_all_mono_fns { + let mut saw_own_coverage = false; + + // Inspect every coverage statement in the function's MIR. + for stmt in body + .basic_blocks + .iter() + .flat_map(|block| &block.statements) + .filter(|stmt| matches!(stmt.kind, mir::StatementKind::Coverage(_))) + { + if let Some(inlined) = stmt.source_info.scope.inlined_instance(&body.source_scopes) { + // This coverage statement was inlined from another function. + used_via_inlining.insert(inlined.def_id()); + } else { + // Non-inlined coverage statements belong to the enclosing function. + saw_own_coverage = true; + } + } + + if !saw_own_coverage && body.function_coverage_info.is_some() { + missing_own_coverage.insert(def_id); + } + } + + UsageSets { all_mono_items, used_via_inlining, missing_own_coverage } +} + +fn make_dummy_instance<'tcx>(tcx: TyCtxt<'tcx>, local_def_id: LocalDefId) -> ty::Instance<'tcx> { + let def_id = local_def_id.to_def_id(); + + // Make a dummy instance that fills in all generics with placeholders. + ty::Instance::new( + def_id, + ty::GenericArgs::for_item(tcx, def_id, |param, _| { + if let ty::GenericParamDefKind::Lifetime = param.kind { + tcx.lifetimes.re_erased.into() + } else { + tcx.mk_param_from_def(param) + } + }), + ) +} 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 f52991b3697..d2591139d6e 100644 --- a/compiler/rustc_codegen_llvm/src/debuginfo/create_scope_map.rs +++ b/compiler/rustc_codegen_llvm/src/debuginfo/create_scope_map.rs @@ -3,7 +3,6 @@ use std::collections::hash_map::Entry; use rustc_codegen_ssa::mir::debuginfo::{DebugScope, FunctionDebugContext}; use rustc_codegen_ssa::traits::*; use rustc_data_structures::fx::FxHashMap; -use rustc_index::Idx; use rustc_index::bit_set::DenseBitSet; use rustc_middle::mir::{Body, SourceScope}; use rustc_middle::ty::layout::{FnAbiOf, HasTypingEnv}; @@ -43,8 +42,7 @@ pub(crate) fn compute_mir_scopes<'ll, 'tcx>( let mut instantiated = DenseBitSet::new_empty(mir.source_scopes.len()); let mut discriminators = FxHashMap::default(); // Instantiate all scopes. - for idx in 0..mir.source_scopes.len() { - let scope = SourceScope::new(idx); + for scope in mir.source_scopes.indices() { make_mir_scope( cx, instance, diff --git a/compiler/rustc_codegen_llvm/src/debuginfo/metadata.rs b/compiler/rustc_codegen_llvm/src/debuginfo/metadata.rs index 98d59f5a8ae..7f3e486ca31 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>, @@ -905,7 +910,8 @@ pub(crate) fn build_compile_unit_di_node<'ll, 'tcx>( && let Some(f) = output_filenames.split_dwarf_path( tcx.sess.split_debuginfo(), tcx.sess.opts.unstable_opts.split_dwarf_kind, - Some(codegen_unit_name), + codegen_unit_name, + tcx.sess.invocation_temp.as_deref(), ) { // We get a path relative to the working directory from split_dwarf_path Some(tcx.sess.source_map().path_mapping().to_real_filename(f)) @@ -929,17 +935,13 @@ 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 +973,7 @@ 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 +985,30 @@ 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 +1017,8 @@ 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 +1102,7 @@ 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 +1152,7 @@ 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 +1197,7 @@ 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 +1295,7 @@ 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), @@ -1287,7 +1313,7 @@ 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()); @@ -1297,16 +1323,7 @@ fn build_generic_type_param_di_nodes<'ll, 'tcx>( 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, - ) - } + Some(cx.create_template_type_parameter(name.as_str(), actual_type_di_node)) }) }) .collect(); @@ -1416,7 +1433,9 @@ 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 +1491,7 @@ 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..e9574108696 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, @@ -721,8 +721,7 @@ fn build_union_fields_for_direct_tag_coroutine<'ll, 'tcx>( _ => unreachable!(), }; - let coroutine_layout = - cx.tcx.coroutine_layout(coroutine_def_id, coroutine_args.kind_ty()).unwrap(); + let coroutine_layout = cx.tcx.coroutine_layout(coroutine_def_id, coroutine_args.args).unwrap(); let common_upvar_names = cx.tcx.closure_saved_names_of_captured_variables(coroutine_def_id); let variant_range = coroutine_args.variant_range(coroutine_def_id, cx.tcx); @@ -820,7 +819,6 @@ 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 +838,23 @@ 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 +868,7 @@ 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 +883,7 @@ 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 +894,7 @@ 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 +905,7 @@ 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..7c701926d2c 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), 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..20a841f2287 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}; @@ -174,10 +174,8 @@ pub(super) fn build_coroutine_di_node<'ll, 'tcx>( DIFlags::FlagZero, ), |cx, coroutine_type_di_node| { - let coroutine_layout = cx - .tcx - .coroutine_layout(coroutine_def_id, coroutine_args.as_coroutine().kind_ty()) - .unwrap(); + let coroutine_layout = + cx.tcx.coroutine_layout(coroutine_def_id, coroutine_args).unwrap(); let Variants::Multiple { tag_encoding: TagEncoding::Direct, ref variants, .. } = coroutine_type_and_layout.variants @@ -363,23 +361,22 @@ 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..56fb12d3c22 100644 --- a/compiler/rustc_codegen_llvm/src/debuginfo/metadata/type_map.rs +++ b/compiler/rustc_codegen_llvm/src/debuginfo/metadata/type_map.rs @@ -247,6 +247,16 @@ pub(super) fn stub<'ll, 'tcx>( StubInfo { metadata, unique_type_id } } +struct AdtStackPopGuard<'ll, 'tcx, 'a> { + cx: &'a CodegenCx<'ll, 'tcx>, +} + +impl<'ll, 'tcx, 'a> Drop for AdtStackPopGuard<'ll, 'tcx, 'a> { + fn drop(&mut self) { + debug_context(self.cx).adt_stack.borrow_mut().pop(); + } +} + /// This function enables creating debuginfo nodes that can recursively refer to themselves. /// It will first insert the given stub into the type map and only then execute the `members` /// and `generics` closures passed in. These closures have access to the stub so they can @@ -257,16 +267,53 @@ 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); + let mut _adt_stack_pop_guard = None; + if let UniqueTypeId::Ty(ty, ..) = stub_info.unique_type_id + && let ty::Adt(adt_def, args) = ty.kind() + { + let def_id = adt_def.did(); + // If any sub type reference the original type definition and the sub type has a type + // parameter that strictly contains the original parameter, the original type is a recursive + // type that can expanding indefinitely. Example, + // ``` + // enum Recursive<T> { + // Recurse(*const Recursive<Wrap<T>>), + // Item(T), + // } + // ``` + let is_expanding_recursive = + debug_context(cx).adt_stack.borrow().iter().any(|(parent_def_id, parent_args)| { + if def_id == *parent_def_id { + args.iter().zip(parent_args.iter()).any(|(arg, parent_arg)| { + if let (Some(arg), Some(parent_arg)) = (arg.as_type(), parent_arg.as_type()) + { + arg != parent_arg && arg.contains(parent_arg) + } else { + false + } + }) + } else { + false + } + }); + if is_expanding_recursive { + // FIXME: indicate that this is an expanding recursive type in stub metadata? + return DINodeCreationResult::new(stub_info.metadata, false); + } else { + debug_context(cx).adt_stack.borrow_mut().push((def_id, args)); + _adt_stack_pop_guard = Some(AdtStackPopGuard { cx }); + } + } + debug_context(cx).type_map.insert(stub_info.unique_type_id, stub_info.metadata); 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..c5085927923 100644 --- a/compiler/rustc_codegen_llvm/src/debuginfo/mod.rs +++ b/compiler/rustc_codegen_llvm/src/debuginfo/mod.rs @@ -6,6 +6,7 @@ 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; @@ -65,6 +66,7 @@ pub(crate) struct CodegenUnitDebugContext<'ll, 'tcx> { created_files: RefCell<UnordMap<Option<(StableSourceFileId, SourceFileHash)>, &'ll DIFile>>, type_map: metadata::TypeMap<'ll, 'tcx>, + adt_stack: RefCell<Vec<(DefId, GenericArgsRef<'tcx>)>>, namespace_map: RefCell<DefIdMap<&'ll DIScope>>, recursion_marker_type: OnceCell<&'ll DIType>, } @@ -79,6 +81,7 @@ impl<'ll, 'tcx> CodegenUnitDebugContext<'ll, 'tcx> { builder, created_files: Default::default(), type_map: Default::default(), + adt_stack: Default::default(), namespace_map: RefCell::new(Default::default()), recursion_marker_type: OnceCell::new(), } @@ -251,7 +254,7 @@ 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 +282,22 @@ 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 +344,8 @@ 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); @@ -483,16 +500,10 @@ impl<'ll, 'tcx> DebugInfoCodegenMethods<'tcx> for CodegenCx<'ll, 'tcx> { 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, - )) - } + Some(cx.create_template_type_parameter( + name.as_str(), + actual_type_metadata, + )) }) }) .collect() diff --git a/compiler/rustc_codegen_llvm/src/declare.rs b/compiler/rustc_codegen_llvm/src/declare.rs index e79662ebc64..2419ec1f888 100644 --- a/compiler/rustc_codegen_llvm/src/declare.rs +++ b/compiler/rustc_codegen_llvm/src/declare.rs @@ -11,6 +11,8 @@ //! * Use define_* family of methods when you might be defining the Value. //! * When in doubt, define. +use std::borrow::Borrow; + use itertools::Itertools; use rustc_codegen_ssa::traits::TypeMembershipCodegenMethods; use rustc_data_structures::fx::FxIndexSet; @@ -22,7 +24,7 @@ use tracing::debug; use crate::abi::FnAbiLlvmExt; use crate::common::AsCCharPtr; -use crate::context::{CodegenCx, SimpleCx}; +use crate::context::{CodegenCx, GenericCx, SCx, SimpleCx}; use crate::llvm::AttributePlace::Function; use crate::llvm::Visibility; use crate::type_::Type; @@ -81,16 +83,25 @@ pub(crate) fn declare_raw_fn<'ll, 'tcx>( llfn } -impl<'ll, 'tcx> CodegenCx<'ll, 'tcx> { +impl<'ll, CX: Borrow<SCx<'ll>>> GenericCx<'ll, CX> { /// Declare a global value. /// /// If there’s a value with the same name already declared, the function will /// return its Value instead. pub(crate) fn declare_global(&self, name: &str, ty: &'ll Type) -> &'ll Value { debug!("declare_global(name={:?})", name); - unsafe { llvm::LLVMRustGetOrInsertGlobal(self.llmod, name.as_c_char_ptr(), name.len(), ty) } + unsafe { + llvm::LLVMRustGetOrInsertGlobal( + (**self).borrow().llmod, + name.as_c_char_ptr(), + name.len(), + ty, + ) + } } +} +impl<'ll, 'tcx> CodegenCx<'ll, 'tcx> { /// Declare a C ABI function. /// /// Only use this for foreign function ABIs and glue. For Rust functions use 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 f68365f6c69..ffeab59b05c 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; @@ -127,15 +128,14 @@ fn get_simple_intrinsic<'ll>( sym::truncf64 => "llvm.trunc.f64", sym::truncf128 => "llvm.trunc.f128", - sym::rintf16 => "llvm.rint.f16", - sym::rintf32 => "llvm.rint.f32", - sym::rintf64 => "llvm.rint.f64", - sym::rintf128 => "llvm.rint.f128", - - sym::nearbyintf16 => "llvm.nearbyint.f16", - sym::nearbyintf32 => "llvm.nearbyint.f32", - sym::nearbyintf64 => "llvm.nearbyint.f64", - sym::nearbyintf128 => "llvm.nearbyint.f128", + // We could use any of `rint`, `nearbyint`, or `roundeven` + // for this -- they are all identical in semantics when + // assuming the default FP environment. + // `rint` is what we used for $forever. + sym::round_ties_even_f16 => "llvm.rint.f16", + sym::round_ties_even_f32 => "llvm.rint.f32", + sym::round_ties_even_f64 => "llvm.rint.f64", + sym::round_ties_even_f128 => "llvm.rint.f128", sym::roundf16 => "llvm.round.f16", sym::roundf32 => "llvm.round.f32", @@ -144,11 +144,6 @@ fn get_simple_intrinsic<'ll>( sym::ptr_mask => "llvm.ptrmask", - sym::roundevenf16 => "llvm.roundeven.f16", - sym::roundevenf32 => "llvm.roundeven.f32", - sym::roundevenf64 => "llvm.roundeven.f64", - sym::roundevenf128 => "llvm.roundeven.f128", - _ => return None, }; Some(cx.get_intrinsic(llvm_name)) @@ -476,7 +471,7 @@ impl<'ll, 'tcx> IntrinsicCallBuilderMethods<'tcx> for Builder<'_, 'll, 'tcx> { let layout = self.layout_of(tp_ty).layout; let use_integer_compare = match layout.backend_repr() { Scalar(_) | ScalarPair(_, _) => true, - Vector { .. } => false, + SimdVector { .. } => false, Memory { .. } => { // For rusty ABIs, small aggregates are actually passed // as `RegKind::Integer` (see `FnAbi::adjust_for_abi`), @@ -655,7 +650,7 @@ impl<'ll, 'tcx> IntrinsicCallBuilderMethods<'tcx> for Builder<'_, 'll, 'tcx> { fn type_test(&mut self, pointer: Self::Value, typeid: Self::Metadata) -> Self::Value { // Test the called operand using llvm.type.test intrinsic. The LowerTypeTests link-time // optimization pass replaces calls to this intrinsic with code to test type membership. - let typeid = unsafe { llvm::LLVMMetadataAsValue(&self.llcx, typeid) }; + let typeid = self.get_metadata_value(typeid); self.call_intrinsic("llvm.type.test", &[pointer, typeid]) } @@ -665,7 +660,7 @@ impl<'ll, 'tcx> IntrinsicCallBuilderMethods<'tcx> for Builder<'_, 'll, 'tcx> { vtable_byte_offset: u64, typeid: &'ll Metadata, ) -> Self::Value { - let typeid = unsafe { llvm::LLVMMetadataAsValue(&self.llcx, typeid) }; + let typeid = self.get_metadata_value(typeid); let vtable_byte_offset = self.const_i32(vtable_byte_offset as i32); let type_checked_load = self.call_intrinsic("llvm.type.checked.load", &[llvtable, vtable_byte_offset, typeid]); @@ -818,7 +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() { @@ -1186,18 +1184,6 @@ 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) => { @@ -1335,7 +1321,7 @@ fn generic_simd_intrinsic<'ll, 'tcx>( )); } - if name == sym::simd_shuffle_generic { + if name == sym::simd_shuffle_const_generic { let idx = fn_args[2].expect_const().to_value().valtree.unwrap_branch(); let n = idx.len() as u64; @@ -1423,7 +1409,7 @@ fn generic_simd_intrinsic<'ll, 'tcx>( return Ok(bx.shuffle_vector(args[0].immediate(), args[1].immediate(), indices)); } - if name == sym::simd_insert { + if name == sym::simd_insert || name == sym::simd_insert_dyn { require!( in_elem == arg_tys[2], InvalidMonomorphization::InsertedType { @@ -1434,40 +1420,49 @@ fn generic_simd_intrinsic<'ll, 'tcx>( out_ty: arg_tys[2] } ); - let idx = bx - .const_to_opt_u128(args[1].immediate(), false) - .expect("typeck should have ensure that this is a const"); - if idx >= in_len.into() { - return_error!(InvalidMonomorphization::SimdIndexOutOfBounds { - span, - name, - arg_idx: 1, - total_len: in_len.into(), - }); - } - return Ok(bx.insert_element( - args[0].immediate(), - args[2].immediate(), - bx.const_i32(idx as i32), - )); + + let index_imm = if name == sym::simd_insert { + let idx = bx + .const_to_opt_u128(args[1].immediate(), false) + .expect("typeck should have ensure that this is a const"); + if idx >= in_len.into() { + return_error!(InvalidMonomorphization::SimdIndexOutOfBounds { + span, + name, + arg_idx: 1, + total_len: in_len.into(), + }); + } + bx.const_i32(idx as i32) + } else { + args[1].immediate() + }; + + return Ok(bx.insert_element(args[0].immediate(), args[2].immediate(), index_imm)); } - if name == sym::simd_extract { + if name == sym::simd_extract || name == sym::simd_extract_dyn { require!( ret_ty == in_elem, InvalidMonomorphization::ReturnType { span, name, in_elem, in_ty, ret_ty } ); - let idx = bx - .const_to_opt_u128(args[1].immediate(), false) - .expect("typeck should have ensure that this is a const"); - if idx >= in_len.into() { - return_error!(InvalidMonomorphization::SimdIndexOutOfBounds { - span, - name, - arg_idx: 1, - total_len: in_len.into(), - }); - } - return Ok(bx.extract_element(args[0].immediate(), bx.const_i32(idx as i32))); + let index_imm = if name == sym::simd_extract { + let idx = bx + .const_to_opt_u128(args[1].immediate(), false) + .expect("typeck should have ensure that this is a const"); + if idx >= in_len.into() { + return_error!(InvalidMonomorphization::SimdIndexOutOfBounds { + span, + name, + arg_idx: 1, + total_len: in_len.into(), + }); + } + bx.const_i32(idx as i32) + } else { + args[1].immediate() + }; + + return Ok(bx.extract_element(args[0].immediate(), index_imm)); } if name == sym::simd_select { @@ -1478,9 +1473,9 @@ fn generic_simd_intrinsic<'ll, 'tcx>( m_len == v_len, InvalidMonomorphization::MismatchedLengths { span, name, m_len, v_len } ); - let in_elem_bitwidth = require_int_ty!( + let in_elem_bitwidth = require_int_or_uint_ty!( m_elem_ty.kind(), - InvalidMonomorphization::MaskType { span, name, ty: m_elem_ty } + InvalidMonomorphization::MaskWrongElementType { 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())); @@ -1501,7 +1496,7 @@ fn generic_simd_intrinsic<'ll, 'tcx>( // Integer vector <i{in_bitwidth} x in_len>: let in_elem_bitwidth = require_int_or_uint_ty!( in_elem.kind(), - InvalidMonomorphization::VectorArgument { span, name, in_ty, in_elem } + InvalidMonomorphization::MaskWrongElementType { span, name, ty: in_elem } ); let i1xn = vector_mask_to_bitmask(bx, args[0].immediate(), in_elem_bitwidth, in_len); @@ -1587,8 +1582,6 @@ fn generic_simd_intrinsic<'ll, 'tcx>( sym::simd_floor => ("floor", bx.type_func(&[vec_ty], vec_ty)), sym::simd_fma => ("fma", bx.type_func(&[vec_ty, vec_ty, vec_ty], vec_ty)), sym::simd_relaxed_fma => ("fmuladd", bx.type_func(&[vec_ty, vec_ty, vec_ty], vec_ty)), - sym::simd_fpowi => ("powi", bx.type_func(&[vec_ty, bx.type_i32()], vec_ty)), - sym::simd_fpow => ("pow", bx.type_func(&[vec_ty, vec_ty], vec_ty)), sym::simd_fsin => ("sin", bx.type_func(&[vec_ty], vec_ty)), sym::simd_fsqrt => ("sqrt", bx.type_func(&[vec_ty], vec_ty)), sym::simd_round => ("round", bx.type_func(&[vec_ty], vec_ty)), @@ -1621,8 +1614,6 @@ fn generic_simd_intrinsic<'ll, 'tcx>( | sym::simd_flog | sym::simd_floor | sym::simd_fma - | sym::simd_fpow - | sym::simd_fpowi | sym::simd_fsin | sym::simd_fsqrt | sym::simd_relaxed_fma @@ -1729,14 +1720,9 @@ fn generic_simd_intrinsic<'ll, 'tcx>( } ); - let mask_elem_bitwidth = require_int_ty!( + let mask_elem_bitwidth = require_int_or_uint_ty!( element_ty2.kind(), - InvalidMonomorphization::ThirdArgElementType { - span, - name, - expected_element: element_ty2, - third_arg: arg_tys[2] - } + InvalidMonomorphization::MaskWrongElementType { span, name, ty: element_ty2 } ); // Alignment of T, must be a constant integer value: @@ -1831,14 +1817,9 @@ fn generic_simd_intrinsic<'ll, 'tcx>( } ); - let m_elem_bitwidth = require_int_ty!( + let m_elem_bitwidth = require_int_or_uint_ty!( mask_elem.kind(), - InvalidMonomorphization::ThirdArgElementType { - span, - name, - expected_element: values_elem, - third_arg: mask_ty, - } + InvalidMonomorphization::MaskWrongElementType { span, name, ty: mask_elem } ); let mask = vector_mask_to_bitmask(bx, args[0].immediate(), m_elem_bitwidth, mask_len); @@ -1921,14 +1902,9 @@ fn generic_simd_intrinsic<'ll, 'tcx>( } ); - let m_elem_bitwidth = require_int_ty!( + let m_elem_bitwidth = require_int_or_uint_ty!( mask_elem.kind(), - InvalidMonomorphization::ThirdArgElementType { - span, - name, - expected_element: values_elem, - third_arg: mask_ty, - } + InvalidMonomorphization::MaskWrongElementType { span, name, ty: mask_elem } ); let mask = vector_mask_to_bitmask(bx, args[0].immediate(), m_elem_bitwidth, mask_len); @@ -2016,15 +1992,10 @@ fn generic_simd_intrinsic<'ll, 'tcx>( } ); - // The element type of the third argument must be a signed integer type of any width: - let mask_elem_bitwidth = require_int_ty!( + // The element type of the third argument must be an integer type of any width: + let mask_elem_bitwidth = require_int_or_uint_ty!( element_ty2.kind(), - InvalidMonomorphization::ThirdArgElementType { - span, - name, - expected_element: element_ty2, - third_arg: arg_tys[2] - } + InvalidMonomorphization::MaskWrongElementType { span, name, ty: element_ty2 } ); // Alignment of T, must be a constant integer value: diff --git a/compiler/rustc_codegen_llvm/src/lib.rs b/compiler/rustc_codegen_llvm/src/lib.rs index e9e1b644f18..e8010ec9fc4 100644 --- a/compiler/rustc_codegen_llvm/src/lib.rs +++ b/compiler/rustc_codegen_llvm/src/lib.rs @@ -6,21 +6,19 @@ // tidy-alphabetical-start #![allow(internal_features)] +#![cfg_attr(bootstrap, feature(let_chains))] #![doc(html_root_url = "https://doc.rust-lang.org/nightly/nightly-rustc/")] #![doc(rust_logo)] #![feature(assert_matches)] #![feature(exact_size_is_empty)] #![feature(extern_types)] #![feature(file_buffered)] -#![feature(hash_raw_entry)] #![feature(if_let_guard)] #![feature(impl_trait_in_assoc_type)] #![feature(iter_intersperse)] -#![feature(let_chains)] #![feature(rustdoc_internals)] #![feature(slice_as_array)] #![feature(try_blocks)] -#![warn(unreachable_pub)] // tidy-alphabetical-end use std::any::Any; @@ -29,8 +27,9 @@ use std::mem::ManuallyDrop; 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_config; use rustc_ast::expand::allocator::AllocatorKind; use rustc_ast::expand::autodiff_attrs::AutoDiffItem; use rustc_codegen_ssa::back::lto::{LtoModuleCodegen, SerializedModule, ThinModule}; @@ -38,7 +37,7 @@ use rustc_codegen_ssa::back::write::{ CodegenContext, FatLtoInput, ModuleConfig, TargetMachineFactoryConfig, TargetMachineFactoryFn, }; use rustc_codegen_ssa::traits::*; -use rustc_codegen_ssa::{CodegenResults, CompiledModule, ModuleCodegen}; +use rustc_codegen_ssa::{CodegenResults, CompiledModule, ModuleCodegen, TargetConfig}; use rustc_data_structures::fx::FxIndexMap; use rustc_errors::{DiagCtxtHandle, FatalError}; use rustc_metadata::EncodedMetadata; @@ -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_; @@ -116,9 +113,11 @@ impl ExtraBackendMethods for LlvmCodegenBackend { kind: AllocatorKind, alloc_error_handler_kind: AllocatorKind, ) -> ModuleLlvm { - let mut module_llvm = ModuleLlvm::new_metadata(tcx, module_name); + let module_llvm = ModuleLlvm::new_metadata(tcx, module_name); + let cx = + SimpleCx::new(module_llvm.llmod(), &module_llvm.llcx, tcx.data_layout.pointer_size); unsafe { - allocator::codegen(tcx, &mut module_llvm, module_name, kind, alloc_error_handler_kind); + allocator::codegen(tcx, cx, module_name, kind, alloc_error_handler_kind); } module_llvm } @@ -194,7 +193,7 @@ impl WriteBackendMethods for LlvmCodegenBackend { unsafe fn optimize( cgcx: &CodegenContext<Self>, dcx: DiagCtxtHandle<'_>, - module: &ModuleCodegen<Self::Module>, + module: &mut ModuleCodegen<Self::Module>, config: &ModuleConfig, ) -> Result<(), FatalError> { unsafe { back::write::optimize(cgcx, dcx, module, config) } @@ -339,8 +338,8 @@ impl CodegenBackend for LlvmCodegenBackend { llvm_util::print_version(); } - fn target_features_cfg(&self, sess: &Session, allow_unstable: bool) -> Vec<Symbol> { - target_features_cfg(sess, allow_unstable) + fn target_config(&self, sess: &Session) -> TargetConfig { + target_config(sess) } fn codegen_crate<'tcx>( 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 3c2c6964a3d..2ad39fc8538 100644 --- a/compiler/rustc_codegen_llvm/src/llvm/enzyme_ffi.rs +++ b/compiler/rustc_codegen_llvm/src/llvm/enzyme_ffi.rs @@ -3,19 +3,35 @@ use libc::{c_char, c_uint}; -use super::ffi::{BasicBlock, Metadata, Module, Type, Value}; +use super::MetadataKindId; +use super::ffi::{AttributeKind, BasicBlock, Metadata, Module, Type, Value}; use crate::llvm::Bool; #[link(name = "llvm-wrapper", kind = "static")] unsafe extern "C" { // Enzyme - pub(crate) fn LLVMRustHasMetadata(I: &Value, KindID: c_uint) -> bool; + pub(crate) safe fn LLVMRustHasMetadata(I: &Value, KindID: MetadataKindId) -> bool; pub(crate) fn LLVMRustEraseInstUntilInclusive(BB: &BasicBlock, I: &Value); pub(crate) fn LLVMRustGetLastInstruction<'a>(BB: &BasicBlock) -> Option<&'a Value>; pub(crate) fn LLVMRustDIGetInstMetadata(I: &Value) -> Option<&Metadata>; 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; + pub(crate) fn LLVMRustHasFnAttribute( + F: &Value, + Name: *const c_char, + NameLen: libc::size_t, + ) -> bool; + pub(crate) fn LLVMRustRemoveFnAttribute(F: &Value, Name: *const c_char, NameLen: libc::size_t); + pub(crate) fn LLVMGetFirstFunction(M: &Module) -> Option<&Value>; + pub(crate) fn LLVMGetNextFunction(Fn: &Value) -> Option<&Value>; + pub(crate) fn LLVMRustRemoveEnumAttributeAtIndex( + Fn: &Value, + index: c_uint, + kind: AttributeKind, + ); } unsafe extern "C" { @@ -30,22 +46,22 @@ 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; - extern "C" { - pub fn EnzymeSetCLBool(arg1: *mut ::std::os::raw::c_void, arg2: u8); + unsafe extern "C" { + pub(crate) fn EnzymeSetCLBool(arg1: *mut ::std::os::raw::c_void, arg2: u8); } - extern "C" { + unsafe extern "C" { static mut EnzymePrintPerf: c_void; static mut EnzymePrintActivity: c_void; static mut EnzymePrintType: c_void; @@ -55,42 +71,42 @@ 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); } @@ -98,34 +114,34 @@ 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 3b0187b9d37..ffb490dcdc2 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; @@ -954,16 +961,38 @@ bitflags! { } } +// These values **must** match with LLVMGEPNoWrapFlags +bitflags! { + #[repr(transparent)] + #[derive(Default)] + pub struct GEPNoWrapFlags : c_uint { + const InBounds = 1 << 0; + const NUSW = 1 << 1; + const NUW = 1 << 2; + } +} + 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(crate) struct MetadataKindId(c_uint); + +impl From<MetadataType> for MetadataKindId { + fn from(value: MetadataType) -> Self { + Self(value as c_uint) + } +} unsafe extern "C" { // Create and destroy contexts. @@ -972,7 +1001,7 @@ unsafe extern "C" { C: &Context, Name: *const c_char, SLen: c_uint, - ) -> c_uint; + ) -> MetadataKindId; // Create modules. pub(crate) fn LLVMModuleCreateWithNameInContext( @@ -1035,14 +1064,13 @@ unsafe extern "C" { pub(crate) fn LLVMMetadataTypeInContext(C: &Context) -> &Type; // Operations on all values - pub(crate) fn LLVMIsUndef(Val: &Value) -> Bool; pub(crate) fn LLVMTypeOf(Val: &Value) -> &Type; pub(crate) fn LLVMGetValueName2(Val: &Value, Length: *mut size_t) -> *const c_char; pub(crate) fn LLVMSetValueName2(Val: &Value, Name: *const c_char, NameLen: size_t); pub(crate) fn LLVMReplaceAllUsesWith<'a>(OldVal: &'a Value, NewVal: &'a Value); - pub(crate) fn LLVMSetMetadata<'a>(Val: &'a Value, KindID: c_uint, Node: &'a Value); + pub(crate) safe fn LLVMSetMetadata<'a>(Val: &'a Value, KindID: MetadataKindId, Node: &'a Value); pub(crate) fn LLVMGlobalSetMetadata<'a>(Val: &'a Value, KindID: c_uint, Metadata: &'a Metadata); - pub(crate) fn LLVMValueAsMetadata(Node: &Value) -> &Metadata; + pub(crate) safe fn LLVMValueAsMetadata(Node: &Value) -> &Metadata; // Operations on constants of any type pub(crate) fn LLVMConstNull(Ty: &Type) -> &Value; @@ -1136,7 +1164,7 @@ unsafe extern "C" { pub(crate) fn LLVMSetThreadLocalMode(GlobalVar: &Value, Mode: ThreadLocalMode); pub(crate) fn LLVMIsGlobalConstant(GlobalVar: &Value) -> Bool; pub(crate) fn LLVMSetGlobalConstant(GlobalVar: &Value, IsConstant: Bool); - pub(crate) fn LLVMSetTailCall(CallInst: &Value, IsTailCall: Bool); + pub(crate) safe fn LLVMSetTailCall(CallInst: &Value, IsTailCall: Bool); // Operations on attributes pub(crate) fn LLVMCreateStringAttribute( @@ -1152,7 +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 @@ -1193,7 +1221,7 @@ unsafe extern "C" { pub(crate) fn LLVMGetCurrentDebugLocation2<'a>(Builder: &Builder<'a>) -> Option<&'a Metadata>; // Terminators - pub(crate) fn LLVMBuildRetVoid<'a>(B: &Builder<'a>) -> &'a Value; + pub(crate) safe fn LLVMBuildRetVoid<'a>(B: &Builder<'a>) -> &'a Value; pub(crate) fn LLVMBuildRet<'a>(B: &Builder<'a>, V: &'a Value) -> &'a Value; pub(crate) fn LLVMBuildBr<'a>(B: &Builder<'a>, Dest: &'a BasicBlock) -> &'a Value; pub(crate) fn LLVMBuildCondBr<'a>( @@ -1454,21 +1482,14 @@ unsafe extern "C" { pub(crate) fn LLVMBuildStore<'a>(B: &Builder<'a>, Val: &'a Value, Ptr: &'a Value) -> &'a Value; - pub(crate) fn LLVMBuildGEP2<'a>( - B: &Builder<'a>, - Ty: &'a Type, - Pointer: &'a Value, - Indices: *const &'a Value, - NumIndices: c_uint, - Name: *const c_char, - ) -> &'a Value; - pub(crate) fn LLVMBuildInBoundsGEP2<'a>( + pub(crate) fn LLVMBuildGEPWithNoWrapFlags<'a>( B: &Builder<'a>, Ty: &'a Type, Pointer: &'a Value, Indices: *const &'a Value, NumIndices: c_uint, Name: *const c_char, + Flags: GEPNoWrapFlags, ) -> &'a Value; // Casts @@ -1676,7 +1697,7 @@ unsafe extern "C" { Packed: Bool, ); - pub(crate) fn LLVMMetadataAsValue<'a>(C: &'a Context, MD: &'a Metadata) -> &'a Value; + pub(crate) safe fn LLVMMetadataAsValue<'a>(C: &'a Context, MD: &'a Metadata) -> &'a Value; pub(crate) fn LLVMSetUnnamedAddress(Global: &Value, UnnamedAddr: UnnamedAddr); @@ -2006,6 +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, @@ -2421,7 +2444,9 @@ unsafe extern "C" { NoPrepopulatePasses: bool, VerifyIR: bool, LintIR: bool, - UseThinLTOBuffers: bool, + ThinLTOBuffer: Option<&mut *mut ThinLTOBuffer>, + EmitThinLTO: bool, + EmitThinLTOSummary: bool, MergeFunctions: bool, UnrollLoops: bool, SLPVectorize: bool, @@ -2429,6 +2454,9 @@ unsafe extern "C" { DisableSimplifyLibCalls: bool, EmitLifetimeMarkers: bool, RunEnzyme: bool, + PrintBeforeEnzyme: bool, + PrintAfterEnzyme: bool, + PrintPasses: bool, SanitizerOptions: Option<&SanitizerOptions>, PGOGenPath: *const c_char, PGOUsePath: *const c_char, diff --git a/compiler/rustc_codegen_llvm/src/llvm/mod.rs b/compiler/rustc_codegen_llvm/src/llvm/mod.rs index 5ec93424131..d14aab06073 100644 --- a/compiler/rustc_codegen_llvm/src/llvm/mod.rs +++ b/compiler/rustc_codegen_llvm/src/llvm/mod.rs @@ -1,7 +1,6 @@ #![allow(non_snake_case)] use std::ffi::{CStr, CString}; -use std::ops::Deref; use std::ptr; use std::str::FromStr; use std::string::FromUtf8Error; @@ -10,18 +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<(), ()> { @@ -42,6 +41,32 @@ pub(crate) fn AddFunctionAttributes<'ll>( } } +pub(crate) fn HasAttributeAtIndex<'ll>( + llfn: &'ll Value, + idx: AttributePlace, + kind: AttributeKind, +) -> bool { + unsafe { LLVMRustHasAttributeAtIndex(llfn, idx.as_uint(), kind) } +} + +pub(crate) fn HasStringAttribute<'ll>(llfn: &'ll Value, name: &str) -> bool { + unsafe { LLVMRustHasFnAttribute(llfn, name.as_c_char_ptr(), name.len()) } +} + +pub(crate) fn RemoveStringAttrFromFn<'ll>(llfn: &'ll Value, name: &str) { + unsafe { LLVMRustRemoveFnAttribute(llfn, name.as_c_char_ptr(), name.len()) } +} + +pub(crate) fn RemoveRustEnumAttributeAtIndex( + llfn: &Value, + place: AttributePlace, + kind: AttributeKind, +) { + unsafe { + LLVMRustRemoveEnumAttributeAtIndex(llfn, place.as_uint(), kind); + } +} + pub(crate) fn AddCallSiteAttributes<'ll>( callsite: &'ll Value, idx: AttributePlace, @@ -128,7 +153,7 @@ pub(crate) fn CreateRangeAttr(llcx: &Context, size: Size, range: WrappingRange) } #[derive(Copy, Clone)] -pub enum AttributePlace { +pub(crate) enum AttributePlace { ReturnValue, Argument(u32), Function, @@ -146,7 +171,7 @@ impl AttributePlace { #[derive(Copy, Clone, PartialEq)] #[repr(C)] -pub enum CodeGenOptSize { +pub(crate) enum CodeGenOptSize { CodeGenOptSizeNone = 0, CodeGenOptSizeDefault = 1, CodeGenOptSizeAggressive = 2, @@ -355,6 +380,16 @@ impl<'a> OperandBundleOwned<'a> { }; OperandBundleOwned { raw: ptr::NonNull::new(raw).unwrap() } } + + /// Returns inner `OperandBundle` type. + /// + /// This could be a `Deref` implementation, but `OperandBundle` contains an extern type and + /// `Deref::Target: ?Sized`. + pub(crate) fn raw(&self) -> &OperandBundle<'a> { + // SAFETY: The returned reference is opaque and can only used for FFI. + // It is valid for as long as `&self` is. + unsafe { self.raw.as_ref() } + } } impl Drop for OperandBundleOwned<'_> { @@ -365,16 +400,6 @@ impl Drop for OperandBundleOwned<'_> { } } -impl<'a> Deref for OperandBundleOwned<'a> { - type Target = OperandBundle<'a>; - - fn deref(&self) -> &Self::Target { - // SAFETY: The returned reference is opaque and can only used for FFI. - // It is valid for as long as `&self` is. - unsafe { self.raw.as_ref() } - } -} - pub(crate) fn add_module_flag_u32( module: &Module, merge_behavior: ModuleFlagMergeBehavior, diff --git a/compiler/rustc_codegen_llvm/src/llvm_util.rs b/compiler/rustc_codegen_llvm/src/llvm_util.rs index b4b5d6a5b19..507cbf20d89 100644 --- a/compiler/rustc_codegen_llvm/src/llvm_util.rs +++ b/compiler/rustc_codegen_llvm/src/llvm_util.rs @@ -6,6 +6,7 @@ use std::sync::Once; use std::{ptr, slice, str}; use libc::c_int; +use rustc_codegen_ssa::TargetConfig; use rustc_codegen_ssa::base::wants_wasm_eh; use rustc_codegen_ssa::codegen_attrs::check_tied_features; use rustc_data_structures::fx::{FxHashMap, FxHashSet}; @@ -256,7 +257,6 @@ pub(crate) fn to_llvm_features<'a>(sess: &Session, s: &'a str) -> Option<LLVMFea ("aarch64", "pmuv3") => Some(LLVMFeature::new("perfmon")), ("aarch64", "paca") => Some(LLVMFeature::new("pauth")), ("aarch64", "pacg") => Some(LLVMFeature::new("pauth")), - ("aarch64", "pauth-lr") if get_version().0 < 19 => None, // Before LLVM 20 those two features were packaged together as b16b16 ("aarch64", "sve-b16b16") if get_version().0 < 20 => Some(LLVMFeature::new("b16b16")), ("aarch64", "sme-b16b16") if get_version().0 < 20 => Some(LLVMFeature::new("b16b16")), @@ -270,17 +270,16 @@ pub(crate) fn to_llvm_features<'a>(sess: &Session, s: &'a str) -> Option<LLVMFea ("aarch64", "fhm") => Some(LLVMFeature::new("fp16fml")), ("aarch64", "fp16") => Some(LLVMFeature::new("fullfp16")), // Filter out features that are not supported by the current LLVM version - ("aarch64", "fpmr") if get_version().0 != 18 => None, + ("aarch64", "fpmr") => None, // only existed in 18 ("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 => { - Some(LLVMFeature::new("fast-unaligned-access")) + // Filter out features that are not supported by the current LLVM version + ("loongarch64", "div32" | "lam-bh" | "lamcas" | "ld-seq-sa" | "scq") + if get_version().0 < 20 => + { + None } // Filter out features that are not supported by the current LLVM version - ("riscv32" | "riscv64", "zaamo") if get_version().0 < 19 => None, - ("riscv32" | "riscv64", "zabha") if get_version().0 < 19 => None, - ("riscv32" | "riscv64", "zalrsc") if get_version().0 < 19 => None, + ("riscv32" | "riscv64", "zacas") if get_version().0 < 20 => None, // Enable the evex512 target feature if an avx512 target feature is enabled. ("x86", s) if s.starts_with("avx512") => { Some(LLVMFeature::with_dependency(s, TargetFeatureFoldStrength::EnableOnly("evex512"))) @@ -291,11 +290,17 @@ pub(crate) fn to_llvm_features<'a>(sess: &Session, s: &'a str) -> Option<LLVMFea ("sparc", "leoncasa") => Some(LLVMFeature::new("hasleoncasa")), // In LLVM 19, there is no `v8plus` feature and `v9` means "SPARC-V9 instruction available and SPARC-V8+ ABI used". // https://github.com/llvm/llvm-project/blob/llvmorg-19.1.0/llvm/lib/Target/Sparc/MCTargetDesc/SparcELFObjectWriter.cpp#L27-L28 - // Before LLVM 19, there is no `v8plus` feature and `v9` means "SPARC-V9 instruction available". + // Before LLVM 19, there was no `v8plus` feature and `v9` means "SPARC-V9 instruction available". // https://github.com/llvm/llvm-project/blob/llvmorg-18.1.0/llvm/lib/Target/Sparc/MCTargetDesc/SparcELFObjectWriter.cpp#L26 ("sparc", "v8plus") if get_version().0 == 19 => Some(LLVMFeature::new("v9")), - ("sparc", "v8plus") if get_version().0 < 19 => None, ("powerpc", "power8-crypto") => Some(LLVMFeature::new("crypto")), + // 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)), } } @@ -304,44 +309,44 @@ pub(crate) fn to_llvm_features<'a>(sess: &Session, s: &'a str) -> Option<LLVMFea /// Must express features in the way Rust understands them. /// /// We do not have to worry about RUSTC_SPECIFIC_FEATURES here, those are handled outside codegen. -pub(crate) fn target_features_cfg(sess: &Session, allow_unstable: bool) -> Vec<Symbol> { - let mut features: FxHashSet<Symbol> = Default::default(); - +pub(crate) fn target_config(sess: &Session) -> TargetConfig { // Add base features for the target. // We do *not* add the -Ctarget-features there, and instead duplicate the logic for that below. // The reason is that if LLVM considers a feature implied but we do not, we don't want that to // show up in `cfg`. That way, `cfg` is entirely under our control -- except for the handling of - // the target CPU, that is still expanded to target features (with all their implied features) by - // LLVM. + // the target CPU, that is still expanded to target features (with all their implied features) + // by LLVM. let target_machine = create_informational_target_machine(sess, true); - // Compute which of the known target features are enabled in the 'base' target machine. - // We only consider "supported" features; "forbidden" features are not reflected in `cfg` as of now. - features.extend( - sess.target - .rust_target_features() - .iter() - .filter(|(feature, _, _)| { - // skip checking special features, as LLVM may not understand them - if RUSTC_SPECIAL_FEATURES.contains(feature) { - return true; - } - // check that all features in a given smallvec are enabled - if let Some(feat) = to_llvm_features(sess, feature) { - for llvm_feature in feat { - let cstr = SmallCStr::new(llvm_feature); - if !unsafe { llvm::LLVMRustHasFeature(&target_machine, cstr.as_ptr()) } { - return false; - } + // Compute which of the known target features are enabled in the 'base' target machine. We only + // consider "supported" features; "forbidden" features are not reflected in `cfg` as of now. + let mut features: FxHashSet<Symbol> = sess + .target + .rust_target_features() + .iter() + .filter(|(feature, _, _)| { + // skip checking special features, as LLVM may not understand them + if RUSTC_SPECIAL_FEATURES.contains(feature) { + return true; + } + if let Some(feat) = to_llvm_features(sess, feature) { + for llvm_feature in feat { + let cstr = SmallCStr::new(llvm_feature); + // `LLVMRustHasFeature` is moderately expensive. On targets with many + // features (e.g. x86) these calls take a non-trivial fraction of runtime + // when compiling very small programs. + if !unsafe { llvm::LLVMRustHasFeature(target_machine.raw(), cstr.as_ptr()) } { + return false; } - true - } else { - false } - }) - .map(|(feature, _, _)| Symbol::intern(feature)), - ); + true + } else { + false + } + }) + .map(|(feature, _, _)| Symbol::intern(feature)) + .collect(); - // Add enabled features + // Add enabled and remove disabled features. for (enabled, feature) in sess.opts.cg.target_feature.split(',').filter_map(|s| match s.chars().next() { Some('+') => Some((true, Symbol::intern(&s[1..]))), @@ -357,7 +362,7 @@ pub(crate) fn target_features_cfg(sess: &Session, allow_unstable: bool) -> Vec<S #[allow(rustc::potential_query_instability)] features.extend( sess.target - .implied_target_features(std::iter::once(feature.as_str())) + .implied_target_features(feature.as_str()) .iter() .map(|s| Symbol::intern(s)), ); @@ -368,11 +373,7 @@ pub(crate) fn target_features_cfg(sess: &Session, allow_unstable: bool) -> Vec<S // `features.contains` below. #[allow(rustc::potential_query_instability)] features.retain(|f| { - if sess - .target - .implied_target_features(std::iter::once(f.as_str())) - .contains(&feature.as_str()) - { + if sess.target.implied_target_features(f.as_str()).contains(&feature.as_str()) { // If `f` if implies `feature`, then `!feature` implies `!f`, so we have to // remove `f`. (This is the standard logical contraposition principle.) false @@ -384,25 +385,109 @@ pub(crate) fn target_features_cfg(sess: &Session, allow_unstable: bool) -> Vec<S } } - // Filter enabled features based on feature gates - sess.target - .rust_target_features() - .iter() - .filter_map(|(feature, gate, _)| { - // The `allow_unstable` set is used by rustc internally to determined which target - // features are truly available, so we want to return even perma-unstable "forbidden" - // features. - if allow_unstable - || (gate.in_cfg() && (sess.is_nightly_build() || gate.requires_nightly().is_none())) - { - Some(*feature) - } else { - None - } - }) - .filter(|feature| features.contains(&Symbol::intern(feature))) - .map(|feature| Symbol::intern(feature)) - .collect() + // Filter enabled features based on feature gates. + let f = |allow_unstable| { + sess.target + .rust_target_features() + .iter() + .filter_map(|(feature, gate, _)| { + // The `allow_unstable` set is used by rustc internally to determined which target + // features are truly available, so we want to return even perma-unstable + // "forbidden" features. + if allow_unstable + || (gate.in_cfg() + && (sess.is_nightly_build() || gate.requires_nightly().is_none())) + { + Some(Symbol::intern(feature)) + } else { + None + } + }) + .filter(|feature| features.contains(&feature)) + .collect() + }; + + let target_features = f(false); + let unstable_target_features = f(true); + let mut cfg = TargetConfig { + target_features, + unstable_target_features, + has_reliable_f16: true, + has_reliable_f16_math: true, + has_reliable_f128: true, + has_reliable_f128_math: true, + }; + + update_target_reliable_float_cfg(sess, &mut cfg); + cfg +} + +/// Determine whether or not experimental float types are reliable based on known bugs. +fn update_target_reliable_float_cfg(sess: &Session, cfg: &mut TargetConfig) { + let target_arch = sess.target.arch.as_ref(); + let target_os = sess.target.options.os.as_ref(); + let target_env = sess.target.options.env.as_ref(); + let target_abi = sess.target.options.abi.as_ref(); + let target_pointer_width = sess.target.pointer_width; + + cfg.has_reliable_f16 = match (target_arch, target_os) { + // Selection failure <https://github.com/llvm/llvm-project/issues/50374> + ("s390x", _) => false, + // Unsupported <https://github.com/llvm/llvm-project/issues/94434> + ("arm64ec", _) => false, + // MinGW ABI bugs <https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115054> + ("x86_64", "windows") if target_env == "gnu" && target_abi != "llvm" => false, + // Infinite recursion <https://github.com/llvm/llvm-project/issues/97981> + ("csky", _) => false, + ("hexagon", _) => false, + ("powerpc" | "powerpc64", _) => false, + ("sparc" | "sparc64", _) => false, + ("wasm32" | "wasm64", _) => false, + // `f16` support only requires that symbols converting to and from `f32` are available. We + // provide these in `compiler-builtins`, so `f16` should be available on all platforms that + // do not have other ABI issues or LLVM crashes. + _ => true, + }; + + cfg.has_reliable_f128 = match (target_arch, target_os) { + // Unsupported <https://github.com/llvm/llvm-project/issues/94434> + ("arm64ec", _) => false, + // Selection bug <https://github.com/llvm/llvm-project/issues/96432> + ("mips64" | "mips64r6", _) => false, + // Selection bug <https://github.com/llvm/llvm-project/issues/95471> + ("nvptx64", _) => false, + // ABI bugs <https://github.com/rust-lang/rust/issues/125109> et al. (full + // list at <https://github.com/rust-lang/rust/issues/116909>) + ("powerpc" | "powerpc64", _) => false, + // ABI unsupported <https://github.com/llvm/llvm-project/issues/41838> + ("sparc", _) => false, + // Stack alignment bug <https://github.com/llvm/llvm-project/issues/77401>. NB: tests may + // not fail if our compiler-builtins is linked. + ("x86", _) => false, + // MinGW ABI bugs <https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115054> + ("x86_64", "windows") if target_env == "gnu" && target_abi != "llvm" => false, + // There are no known problems on other platforms, so the only requirement is that symbols + // are available. `compiler-builtins` provides all symbols required for core `f128` + // support, so this should work for everything else. + _ => true, + }; + + // Assume that working `f16` means working `f16` math for most platforms, since + // operations just go through `f32`. + cfg.has_reliable_f16_math = cfg.has_reliable_f16; + + cfg.has_reliable_f128_math = match (target_arch, target_os) { + // LLVM lowers `fp128` math to `long double` symbols even on platforms where + // `long double` is not IEEE binary128. See + // <https://github.com/llvm/llvm-project/issues/44744>. + // + // This rules out anything that doesn't have `long double` = `binary128`; <= 32 bits + // (ld is `f64`), anything other than Linux (Windows and MacOS use `f64`), and `x86` + // (ld is 80-bit extended precision). + ("x86_64", _) => false, + (_, "linux") if target_pointer_width == 64 => true, + _ => false, + } && cfg.has_reliable_f128; } pub(crate) fn print_version() { @@ -451,8 +536,8 @@ pub(crate) fn print(req: &PrintRequest, out: &mut String, sess: &Session) { require_inited(); let tm = create_informational_target_machine(sess, false); match req.kind { - PrintKind::TargetCPUs => print_target_cpus(sess, &tm, out), - PrintKind::TargetFeatures => print_target_features(sess, &tm, out), + PrintKind::TargetCPUs => print_target_cpus(sess, tm.raw(), out), + PrintKind::TargetFeatures => print_target_features(sess, tm.raw(), out), _ => bug!("rustc_codegen_llvm can't handle print request: {:?}", req), } } @@ -679,14 +764,14 @@ pub(crate) fn global_llvm_features( for feature in sess.opts.cg.target_feature.split(',') { if let Some(feature) = feature.strip_prefix('+') { all_rust_features.extend( - UnordSet::from(sess.target.implied_target_features(std::iter::once(feature))) + UnordSet::from(sess.target.implied_target_features(feature)) .to_sorted_stable_ord() .iter() .map(|&&s| (true, s)), ) } else if let Some(feature) = feature.strip_prefix('-') { // FIXME: Why do we not remove implied features on "-" here? - // We do the equivalent above in `target_features_cfg`. + // We do the equivalent above in `target_config`. // See <https://github.com/rust-lang/rust/issues/134792>. all_rust_features.push((false, feature)); } else if !feature.is_empty() { 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/type_.rs b/compiler/rustc_codegen_llvm/src/type_.rs index d61ce417562..169036f5152 100644 --- a/compiler/rustc_codegen_llvm/src/type_.rs +++ b/compiler/rustc_codegen_llvm/src/type_.rs @@ -1,3 +1,4 @@ +use std::borrow::Borrow; use std::{fmt, ptr}; use libc::{c_char, c_uint}; @@ -11,7 +12,7 @@ use rustc_middle::ty::{self, Ty}; use rustc_target::callconv::{CastTarget, FnAbi}; use crate::abi::{FnAbiLlvmExt, LlvmType}; -use crate::context::{CodegenCx, SimpleCx}; +use crate::context::{CodegenCx, GenericCx, SCx}; pub(crate) use crate::llvm::Type; use crate::llvm::{Bool, False, Metadata, True}; use crate::type_of::LayoutLlvmExt; @@ -36,29 +37,29 @@ impl fmt::Debug for Type { } impl<'ll> CodegenCx<'ll, '_> {} -impl<'ll> SimpleCx<'ll> { +impl<'ll, CX: Borrow<SCx<'ll>>> GenericCx<'ll, CX> { pub(crate) fn type_named_struct(&self, name: &str) -> &'ll Type { let name = SmallCStr::new(name); - unsafe { llvm::LLVMStructCreateNamed(self.llcx, name.as_ptr()) } + unsafe { llvm::LLVMStructCreateNamed(self.llcx(), name.as_ptr()) } } 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) } + unsafe { llvm::LLVMVoidTypeInContext(self.llcx()) } } pub(crate) fn type_token(&self) -> &'ll Type { - unsafe { llvm::LLVMTokenTypeInContext(self.llcx) } + unsafe { llvm::LLVMTokenTypeInContext(self.llcx()) } } pub(crate) fn type_metadata(&self) -> &'ll Type { - unsafe { llvm::LLVMMetadataTypeInContext(self.llcx) } + unsafe { llvm::LLVMMetadataTypeInContext(self.llcx()) } } ///x Creates an integer type with the given number of bits, e.g., i24 pub(crate) fn type_ix(&self, num_bits: u64) -> &'ll Type { - unsafe { llvm::LLVMIntTypeInContext(self.llcx, num_bits as c_uint) } + unsafe { llvm::LLVMIntTypeInContext(self.llcx(), num_bits as c_uint) } } pub(crate) fn type_vector(&self, ty: &'ll Type, len: u64) -> &'ll Type { @@ -121,19 +122,32 @@ impl<'ll, 'tcx> CodegenCx<'ll, 'tcx> { self.type_array(self.type_from_integer(unit), size / unit_size) } } -impl<'ll> SimpleCx<'ll> { + +impl<'ll, CX: Borrow<SCx<'ll>>> GenericCx<'ll, CX> { + pub(crate) fn llcx(&self) -> &'ll llvm::Context { + (**self).borrow().llcx + } + + pub(crate) fn llmod(&self) -> &'ll llvm::Module { + (**self).borrow().llmod + } + + pub(crate) fn isize_ty(&self) -> &'ll Type { + (**self).borrow().isize_ty + } + 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) } } pub(crate) fn type_i1(&self) -> &'ll Type { - unsafe { llvm::LLVMInt1TypeInContext(self.llcx) } + unsafe { llvm::LLVMInt1TypeInContext(self.llcx()) } } pub(crate) fn type_struct(&self, els: &[&'ll Type], packed: bool) -> &'ll Type { unsafe { llvm::LLVMStructTypeInContext( - self.llcx, + self.llcx(), els.as_ptr(), els.len() as c_uint, packed as Bool, @@ -142,45 +156,45 @@ impl<'ll> SimpleCx<'ll> { } } -impl<'ll, 'tcx> BaseTypeCodegenMethods<'tcx> for CodegenCx<'ll, 'tcx> { +impl<'ll, CX: Borrow<SCx<'ll>>> BaseTypeCodegenMethods for GenericCx<'ll, CX> { fn type_i8(&self) -> &'ll Type { - unsafe { llvm::LLVMInt8TypeInContext(self.llcx) } + unsafe { llvm::LLVMInt8TypeInContext(self.llcx()) } } fn type_i16(&self) -> &'ll Type { - unsafe { llvm::LLVMInt16TypeInContext(self.llcx) } + unsafe { llvm::LLVMInt16TypeInContext(self.llcx()) } } fn type_i32(&self) -> &'ll Type { - unsafe { llvm::LLVMInt32TypeInContext(self.llcx) } + unsafe { llvm::LLVMInt32TypeInContext(self.llcx()) } } fn type_i64(&self) -> &'ll Type { - unsafe { llvm::LLVMInt64TypeInContext(self.llcx) } + unsafe { llvm::LLVMInt64TypeInContext(self.llcx()) } } fn type_i128(&self) -> &'ll Type { - unsafe { llvm::LLVMIntTypeInContext(self.llcx, 128) } + unsafe { llvm::LLVMIntTypeInContext(self.llcx(), 128) } } fn type_isize(&self) -> &'ll Type { - self.isize_ty + self.isize_ty() } fn type_f16(&self) -> &'ll Type { - unsafe { llvm::LLVMHalfTypeInContext(self.llcx) } + unsafe { llvm::LLVMHalfTypeInContext(self.llcx()) } } fn type_f32(&self) -> &'ll Type { - unsafe { llvm::LLVMFloatTypeInContext(self.llcx) } + unsafe { llvm::LLVMFloatTypeInContext(self.llcx()) } } fn type_f64(&self) -> &'ll Type { - unsafe { llvm::LLVMDoubleTypeInContext(self.llcx) } + unsafe { llvm::LLVMDoubleTypeInContext(self.llcx()) } } fn type_f128(&self) -> &'ll Type { - unsafe { llvm::LLVMFP128TypeInContext(self.llcx) } + unsafe { llvm::LLVMFP128TypeInContext(self.llcx()) } } fn type_func(&self, args: &[&'ll Type], ret: &'ll Type) -> &'ll Type { @@ -196,7 +210,7 @@ impl<'ll, 'tcx> BaseTypeCodegenMethods<'tcx> for CodegenCx<'ll, 'tcx> { } fn type_ptr_ext(&self, address_space: AddressSpace) -> &'ll Type { - unsafe { llvm::LLVMPointerTypeInContext(self.llcx, address_space.0) } + unsafe { llvm::LLVMPointerTypeInContext(self.llcx(), address_space.0) } } fn element_type(&self, ty: &'ll Type) -> &'ll Type { diff --git a/compiler/rustc_codegen_llvm/src/type_of.rs b/compiler/rustc_codegen_llvm/src/type_of.rs index ba01fbff385..4e7096da502 100644 --- a/compiler/rustc_codegen_llvm/src/type_of.rs +++ b/compiler/rustc_codegen_llvm/src/type_of.rs @@ -19,7 +19,7 @@ fn uncached_llvm_type<'a, 'tcx>( ) -> &'a Type { match layout.backend_repr { BackendRepr::Scalar(_) => bug!("handled elsewhere"), - BackendRepr::Vector { element, count } => { + BackendRepr::SimdVector { element, count } => { let element = layout.scalar_llvm_type_at(cx, element); return cx.type_vector(element, count); } @@ -171,7 +171,7 @@ pub(crate) trait LayoutLlvmExt<'tcx> { impl<'tcx> LayoutLlvmExt<'tcx> for TyAndLayout<'tcx> { fn is_llvm_immediate(&self) -> bool { match self.backend_repr { - BackendRepr::Scalar(_) | BackendRepr::Vector { .. } => true, + BackendRepr::Scalar(_) | BackendRepr::SimdVector { .. } => true, BackendRepr::ScalarPair(..) | BackendRepr::Memory { .. } => false, } } @@ -179,9 +179,9 @@ impl<'tcx> LayoutLlvmExt<'tcx> for TyAndLayout<'tcx> { fn is_llvm_scalar_pair(&self) -> bool { match self.backend_repr { BackendRepr::ScalarPair(..) => true, - BackendRepr::Scalar(_) | BackendRepr::Vector { .. } | BackendRepr::Memory { .. } => { - false - } + BackendRepr::Scalar(_) + | BackendRepr::SimdVector { .. } + | BackendRepr::Memory { .. } => 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), |