diff options
Diffstat (limited to 'compiler/rustc_codegen_gcc/src')
27 files changed, 1904 insertions, 1127 deletions
diff --git a/compiler/rustc_codegen_gcc/src/abi.rs b/compiler/rustc_codegen_gcc/src/abi.rs index 6fb1cbfad8c..35bb0b6e5f4 100644 --- a/compiler/rustc_codegen_gcc/src/abi.rs +++ b/compiler/rustc_codegen_gcc/src/abi.rs @@ -3,7 +3,9 @@ use rustc_codegen_ssa::traits::{AbiBuilderMethods, BaseTypeMethods}; use rustc_data_structures::fx::FxHashSet; use rustc_middle::bug; use rustc_middle::ty::Ty; -use rustc_target::abi::call::{CastTarget, FnAbi, PassMode, Reg, RegKind}; +#[cfg(feature = "master")] +use rustc_session::config; +use rustc_target::abi::call::{ArgAttributes, CastTarget, FnAbi, PassMode, Reg, RegKind}; use crate::builder::Builder; use crate::context::CodegenCx; @@ -113,37 +115,57 @@ impl<'gcc, 'tcx> FnAbiGccExt<'gcc, 'tcx> for FnAbi<'tcx, Ty<'tcx>> { match self.ret.mode { PassMode::Ignore => cx.type_void(), PassMode::Direct(_) | PassMode::Pair(..) => self.ret.layout.immediate_gcc_type(cx), - PassMode::Cast(ref cast, _) => cast.gcc_type(cx), + PassMode::Cast { ref cast, .. } => cast.gcc_type(cx), PassMode::Indirect { .. } => { argument_tys.push(cx.type_ptr_to(self.ret.memory_ty(cx))); cx.type_void() } }; + #[cfg(feature = "master")] + let apply_attrs = |ty: Type<'gcc>, attrs: &ArgAttributes| { + if cx.sess().opts.optimize != config::OptLevel::No + && attrs.regular.contains(rustc_target::abi::call::ArgAttribute::NoAlias) + { + ty.make_restrict() + } else { + ty + } + }; + #[cfg(not(feature = "master"))] + let apply_attrs = |ty: Type<'gcc>, _attrs: &ArgAttributes| { + ty + }; + for arg in self.args.iter() { let arg_ty = match arg.mode { PassMode::Ignore => continue, - PassMode::Direct(_) => arg.layout.immediate_gcc_type(cx), - PassMode::Pair(..) => { - argument_tys.push(arg.layout.scalar_pair_element_gcc_type(cx, 0, true)); - argument_tys.push(arg.layout.scalar_pair_element_gcc_type(cx, 1, true)); + PassMode::Pair(a, b) => { + argument_tys.push(apply_attrs(arg.layout.scalar_pair_element_gcc_type(cx, 0), &a)); + argument_tys.push(apply_attrs(arg.layout.scalar_pair_element_gcc_type(cx, 1), &b)); continue; } - PassMode::Indirect { extra_attrs: Some(_), .. } => { - unimplemented!(); - } - PassMode::Cast(ref cast, pad_i32) => { + PassMode::Cast { ref cast, pad_i32 } => { // add padding if pad_i32 { argument_tys.push(Reg::i32().gcc_type(cx)); } - cast.gcc_type(cx) + let ty = cast.gcc_type(cx); + apply_attrs(ty, &cast.attrs) } - PassMode::Indirect { extra_attrs: None, on_stack: true, .. } => { + PassMode::Indirect { attrs: _, meta_attrs: None, on_stack: true } => { + // This is a "byval" argument, so we don't apply the `restrict` attribute on it. on_stack_param_indices.insert(argument_tys.len()); arg.memory_ty(cx) }, - PassMode::Indirect { extra_attrs: None, on_stack: false, .. } => cx.type_ptr_to(arg.memory_ty(cx)), + PassMode::Direct(attrs) => apply_attrs(arg.layout.immediate_gcc_type(cx), &attrs), + PassMode::Indirect { attrs, meta_attrs: None, on_stack: false } => { + apply_attrs(cx.type_ptr_to(arg.memory_ty(cx)), &attrs) + } + PassMode::Indirect { attrs, meta_attrs: Some(meta_attrs), on_stack } => { + assert!(!on_stack); + apply_attrs(apply_attrs(cx.type_ptr_to(arg.memory_ty(cx)), &attrs), &meta_attrs) + } }; argument_tys.push(arg_ty); } diff --git a/compiler/rustc_codegen_gcc/src/allocator.rs b/compiler/rustc_codegen_gcc/src/allocator.rs index e90db44ece1..c8c098e2973 100644 --- a/compiler/rustc_codegen_gcc/src/allocator.rs +++ b/compiler/rustc_codegen_gcc/src/allocator.rs @@ -1,14 +1,17 @@ #[cfg(feature="master")] use gccjit::FnAttribute; -use gccjit::{FunctionType, GlobalKind, ToRValue}; -use rustc_ast::expand::allocator::{AllocatorKind, AllocatorTy, ALLOCATOR_METHODS}; +use gccjit::{Context, FunctionType, GlobalKind, ToRValue, Type}; +use rustc_ast::expand::allocator::{ + alloc_error_handler_name, default_fn_name, global_fn_name, AllocatorKind, AllocatorTy, + ALLOCATOR_METHODS, NO_ALLOC_SHIM_IS_UNSTABLE, +}; use rustc_middle::bug; use rustc_middle::ty::TyCtxt; use rustc_session::config::OomStrategy; use crate::GccContext; -pub(crate) unsafe fn codegen(tcx: TyCtxt<'_>, mods: &mut GccContext, _module_name: &str, kind: AllocatorKind) { +pub(crate) unsafe fn codegen(tcx: TyCtxt<'_>, mods: &mut GccContext, _module_name: &str, kind: AllocatorKind, alloc_error_handler_kind: AllocatorKind) { let context = &mods.context; let usize = match tcx.sess.target.pointer_width { @@ -19,76 +22,105 @@ pub(crate) unsafe fn codegen(tcx: TyCtxt<'_>, mods: &mut GccContext, _module_nam }; let i8 = context.new_type::<i8>(); let i8p = i8.make_pointer(); - let void = context.new_type::<()>(); - - for method in ALLOCATOR_METHODS { - let mut types = Vec::with_capacity(method.inputs.len()); - for ty in method.inputs.iter() { - match *ty { - AllocatorTy::Layout => { - types.push(usize); - types.push(usize); - } - AllocatorTy::Ptr => types.push(i8p), - AllocatorTy::Usize => types.push(usize), - AllocatorTy::ResultPtr | AllocatorTy::Unit => panic!("invalid allocator arg"), - } - } - let output = match method.output { - AllocatorTy::ResultPtr => Some(i8p), - AllocatorTy::Unit => None, + if kind == AllocatorKind::Default { + for method in ALLOCATOR_METHODS { + let mut types = Vec::with_capacity(method.inputs.len()); + for input in method.inputs.iter() { + match input.ty { + AllocatorTy::Layout => { + types.push(usize); + types.push(usize); + } + AllocatorTy::Ptr => types.push(i8p), + AllocatorTy::Usize => types.push(usize), - AllocatorTy::Layout | AllocatorTy::Usize | AllocatorTy::Ptr => { - panic!("invalid allocator output") + AllocatorTy::ResultPtr | AllocatorTy::Unit => panic!("invalid allocator arg"), + } } - }; - let name = format!("__rust_{}", method.name); - - let args: Vec<_> = types.iter().enumerate() - .map(|(index, typ)| context.new_parameter(None, *typ, &format!("param{}", index))) - .collect(); - let func = context.new_function(None, FunctionType::Exported, output.unwrap_or(void), &args, name, false); + let output = match method.output { + AllocatorTy::ResultPtr => Some(i8p), + AllocatorTy::Unit => None, - if tcx.sess.target.options.default_hidden_visibility { - #[cfg(feature="master")] - func.add_attribute(FnAttribute::Visibility(gccjit::Visibility::Hidden)); - } - if tcx.sess.must_emit_unwind_tables() { - // TODO(antoyo): emit unwind tables. - } + AllocatorTy::Layout | AllocatorTy::Usize | AllocatorTy::Ptr => { + panic!("invalid allocator output") + } + }; + let from_name = global_fn_name(method.name); + let to_name = default_fn_name(method.name); - let callee = kind.fn_name(method.name); - let args: Vec<_> = types.iter().enumerate() - .map(|(index, typ)| context.new_parameter(None, *typ, &format!("param{}", index))) - .collect(); - let callee = context.new_function(None, FunctionType::Extern, output.unwrap_or(void), &args, callee, false); - #[cfg(feature="master")] - callee.add_attribute(FnAttribute::Visibility(gccjit::Visibility::Hidden)); - - let block = func.new_block("entry"); - - let args = args - .iter() - .enumerate() - .map(|(i, _)| func.get_param(i as i32).to_rvalue()) - .collect::<Vec<_>>(); - let ret = context.new_call(None, callee, &args); - //llvm::LLVMSetTailCall(ret, True); - if output.is_some() { - block.end_with_return(None, ret); + create_wrapper_function(tcx, context, &from_name, &to_name, &types, output); } - else { - block.end_with_void_return(None); - } - - // TODO(@Commeownist): Check if we need to emit some extra debugging info in certain circumstances - // as described in https://github.com/rust-lang/rust/commit/77a96ed5646f7c3ee8897693decc4626fe380643 } + // FIXME(bjorn3): Add noreturn attribute + create_wrapper_function( + tcx, + context, + "__rust_alloc_error_handler", + &alloc_error_handler_name(alloc_error_handler_kind), + &[usize, usize], + None, + ); + let name = OomStrategy::SYMBOL.to_string(); let global = context.new_global(None, GlobalKind::Exported, i8, name); let value = tcx.sess.opts.unstable_opts.oom.should_panic(); let value = context.new_rvalue_from_int(i8, value as i32); global.global_set_initializer_rvalue(value); + + let name = NO_ALLOC_SHIM_IS_UNSTABLE.to_string(); + let global = context.new_global(None, GlobalKind::Exported, i8, name); + let value = context.new_rvalue_from_int(i8, 0); + global.global_set_initializer_rvalue(value); +} + +fn create_wrapper_function( + tcx: TyCtxt<'_>, + context: &Context<'_>, + from_name: &str, + to_name: &str, + types: &[Type<'_>], + output: Option<Type<'_>>, +) { + let void = context.new_type::<()>(); + + let args: Vec<_> = types.iter().enumerate() + .map(|(index, typ)| context.new_parameter(None, *typ, &format!("param{}", index))) + .collect(); + let func = context.new_function(None, FunctionType::Exported, output.unwrap_or(void), &args, from_name, false); + + if tcx.sess.target.options.default_hidden_visibility { + #[cfg(feature="master")] + func.add_attribute(FnAttribute::Visibility(gccjit::Visibility::Hidden)); + } + if tcx.sess.must_emit_unwind_tables() { + // TODO(antoyo): emit unwind tables. + } + + let args: Vec<_> = types.iter().enumerate() + .map(|(index, typ)| context.new_parameter(None, *typ, &format!("param{}", index))) + .collect(); + let callee = context.new_function(None, FunctionType::Extern, output.unwrap_or(void), &args, to_name, false); + #[cfg(feature="master")] + callee.add_attribute(FnAttribute::Visibility(gccjit::Visibility::Hidden)); + + let block = func.new_block("entry"); + + let args = args + .iter() + .enumerate() + .map(|(i, _)| func.get_param(i as i32).to_rvalue()) + .collect::<Vec<_>>(); + let ret = context.new_call(None, callee, &args); + //llvm::LLVMSetTailCall(ret, True); + if output.is_some() { + block.end_with_return(None, ret); + } + else { + block.end_with_void_return(None); + } + + // TODO(@Commeownist): Check if we need to emit some extra debugging info in certain circumstances + // as described in https://github.com/rust-lang/rust/commit/77a96ed5646f7c3ee8897693decc4626fe380643 } diff --git a/compiler/rustc_codegen_gcc/src/asm.rs b/compiler/rustc_codegen_gcc/src/asm.rs index 65de02b3567..f3a9ca77a67 100644 --- a/compiler/rustc_codegen_gcc/src/asm.rs +++ b/compiler/rustc_codegen_gcc/src/asm.rs @@ -107,7 +107,7 @@ enum ConstraintOrRegister { impl<'a, 'gcc, 'tcx> AsmBuilderMethods<'tcx> for Builder<'a, 'gcc, 'tcx> { - fn codegen_inline_asm(&mut self, template: &[InlineAsmTemplatePiece], rust_operands: &[InlineAsmOperandRef<'tcx, Self>], options: InlineAsmOptions, span: &[Span], _instance: Instance<'_>, _dest_catch_funclet: Option<(Self::BasicBlock, Self::BasicBlock, Option<&Self::Funclet>)>) { + fn codegen_inline_asm(&mut self, template: &[InlineAsmTemplatePiece], rust_operands: &[InlineAsmOperandRef<'tcx, Self>], options: InlineAsmOptions, span: &[Span], instance: Instance<'_>, _dest_catch_funclet: Option<(Self::BasicBlock, Self::BasicBlock, Option<&Self::Funclet>)>) { if options.contains(InlineAsmOptions::MAY_UNWIND) { self.sess() .create_err(UnwindingInlineAsm { span: span[0] }) @@ -173,7 +173,7 @@ impl<'a, 'gcc, 'tcx> AsmBuilderMethods<'tcx> for Builder<'a, 'gcc, 'tcx> { let is_target_supported = reg.reg_class().supported_types(asm_arch).iter() .any(|&(_, feature)| { if let Some(feature) = feature { - self.tcx.sess.target_features.contains(&feature) + self.tcx.asm_target_features(instance.def_id()).contains(&feature) } else { true // Register class is unconditionally supported } @@ -452,10 +452,6 @@ impl<'a, 'gcc, 'tcx> AsmBuilderMethods<'tcx> for Builder<'a, 'gcc, 'tcx> { } InlineAsmOperandRef::Const { ref string } => { - // Const operands get injected directly into the template - if att_dialect { - template_str.push('$'); - } template_str.push_str(string); } } @@ -501,7 +497,7 @@ impl<'a, 'gcc, 'tcx> AsmBuilderMethods<'tcx> for Builder<'a, 'gcc, 'tcx> { if options.contains(InlineAsmOptions::NORETURN) { let builtin_unreachable = self.context.get_builtin_function("__builtin_unreachable"); let builtin_unreachable: RValue<'gcc> = unsafe { std::mem::transmute(builtin_unreachable) }; - self.call(self.type_void(), None, builtin_unreachable, &[], None); + self.call(self.type_void(), None, None, builtin_unreachable, &[], None); } // Write results to outputs. @@ -518,7 +514,6 @@ impl<'a, 'gcc, 'tcx> AsmBuilderMethods<'tcx> for Builder<'a, 'gcc, 'tcx> { OperandValue::Immediate(op.tmp_var.to_rvalue()).store(self, place); } } - } } @@ -593,9 +588,13 @@ fn reg_to_gcc(reg: InlineAsmRegOrRegClass) -> ConstraintOrRegister { InlineAsmRegClass::Bpf(BpfInlineAsmRegClass::reg) => "r", InlineAsmRegClass::Bpf(BpfInlineAsmRegClass::wreg) => "w", InlineAsmRegClass::Hexagon(HexagonInlineAsmRegClass::reg) => "r", + InlineAsmRegClass::LoongArch(LoongArchInlineAsmRegClass::reg) => "r", + InlineAsmRegClass::LoongArch(LoongArchInlineAsmRegClass::freg) => "f", InlineAsmRegClass::M68k(M68kInlineAsmRegClass::reg) => "r", InlineAsmRegClass::M68k(M68kInlineAsmRegClass::reg_addr) => "a", InlineAsmRegClass::M68k(M68kInlineAsmRegClass::reg_data) => "d", + InlineAsmRegClass::CSKY(CSKYInlineAsmRegClass::reg) => "r", + InlineAsmRegClass::CSKY(CSKYInlineAsmRegClass::freg) => "f", InlineAsmRegClass::Mips(MipsInlineAsmRegClass::reg) => "d", // more specific than "r" InlineAsmRegClass::Mips(MipsInlineAsmRegClass::freg) => "f", InlineAsmRegClass::Msp430(Msp430InlineAsmRegClass::reg) => "r", @@ -667,9 +666,13 @@ fn dummy_output_type<'gcc, 'tcx>(cx: &CodegenCx<'gcc, 'tcx>, reg: InlineAsmRegCl InlineAsmRegClass::Avr(_) => unimplemented!(), InlineAsmRegClass::Bpf(_) => unimplemented!(), InlineAsmRegClass::Hexagon(HexagonInlineAsmRegClass::reg) => cx.type_i32(), + InlineAsmRegClass::LoongArch(LoongArchInlineAsmRegClass::reg) => cx.type_i32(), + InlineAsmRegClass::LoongArch(LoongArchInlineAsmRegClass::freg) => cx.type_f32(), InlineAsmRegClass::M68k(M68kInlineAsmRegClass::reg) => cx.type_i32(), InlineAsmRegClass::M68k(M68kInlineAsmRegClass::reg_addr) => cx.type_i32(), InlineAsmRegClass::M68k(M68kInlineAsmRegClass::reg_data) => cx.type_i32(), + InlineAsmRegClass::CSKY(CSKYInlineAsmRegClass::reg) => cx.type_i32(), + InlineAsmRegClass::CSKY(CSKYInlineAsmRegClass::freg) => cx.type_f32(), InlineAsmRegClass::Mips(MipsInlineAsmRegClass::reg) => cx.type_i32(), InlineAsmRegClass::Mips(MipsInlineAsmRegClass::freg) => cx.type_f32(), InlineAsmRegClass::Msp430(_) => unimplemented!(), @@ -804,6 +807,7 @@ fn modifier_to_gcc(arch: InlineAsmArch, reg: InlineAsmRegClass, modifier: Option } } InlineAsmRegClass::Hexagon(_) => None, + InlineAsmRegClass::LoongArch(_) => None, InlineAsmRegClass::Mips(_) => None, InlineAsmRegClass::Nvptx(_) => None, InlineAsmRegClass::PowerPC(_) => None, @@ -856,6 +860,7 @@ fn modifier_to_gcc(arch: InlineAsmArch, reg: InlineAsmRegClass, modifier: Option InlineAsmRegClass::S390x(_) => None, InlineAsmRegClass::Msp430(_) => None, InlineAsmRegClass::M68k(_) => None, + InlineAsmRegClass::CSKY(_) => None, InlineAsmRegClass::SpirV(SpirVInlineAsmRegClass::reg) => { bug!("LLVM backend does not support SPIR-V") } diff --git a/compiler/rustc_codegen_gcc/src/attributes.rs b/compiler/rustc_codegen_gcc/src/attributes.rs index db841b1b524..971e019a4f6 100644 --- a/compiler/rustc_codegen_gcc/src/attributes.rs +++ b/compiler/rustc_codegen_gcc/src/attributes.rs @@ -2,68 +2,31 @@ use gccjit::FnAttribute; use gccjit::Function; use rustc_attr::InstructionSetAttr; -use rustc_codegen_ssa::target_features::tied_target_features; -use rustc_data_structures::fx::FxHashMap; +#[cfg(feature="master")] +use rustc_attr::InlineAttr; use rustc_middle::ty; -use rustc_session::Session; +#[cfg(feature="master")] +use rustc_middle::middle::codegen_fn_attrs::CodegenFnAttrFlags; use rustc_span::symbol::sym; -use smallvec::{smallvec, SmallVec}; use crate::{context::CodegenCx, errors::TiedTargetFeatures}; +use crate::gcc_util::{check_tied_features, to_gcc_features}; -// Given a map from target_features to whether they are enabled or disabled, -// ensure only valid combinations are allowed. -pub fn check_tied_features(sess: &Session, features: &FxHashMap<&str, bool>) -> Option<&'static [&'static str]> { - for tied in tied_target_features(sess) { - // Tied features must be set to the same value, or not set at all - let mut tied_iter = tied.iter(); - let enabled = features.get(tied_iter.next().unwrap()); - if tied_iter.any(|feature| enabled != features.get(feature)) { - return Some(tied); +/// Get GCC attribute for the provided inline heuristic. +#[cfg(feature="master")] +#[inline] +fn inline_attr<'gcc, 'tcx>(cx: &CodegenCx<'gcc, 'tcx>, inline: InlineAttr) -> Option<FnAttribute<'gcc>> { + match inline { + InlineAttr::Hint => Some(FnAttribute::Inline), + InlineAttr::Always => Some(FnAttribute::AlwaysInline), + InlineAttr::Never => { + if cx.sess().target.arch != "amdgpu" { + Some(FnAttribute::NoInline) + } else { + None + } } - } - None -} - -// TODO(antoyo): maybe move to a new module gcc_util. -// To find a list of GCC's names, check https://gcc.gnu.org/onlinedocs/gcc/Function-Attributes.html -fn to_gcc_features<'a>(sess: &Session, s: &'a str) -> SmallVec<[&'a str; 2]> { - let arch = if sess.target.arch == "x86_64" { "x86" } else { &*sess.target.arch }; - match (arch, s) { - ("x86", "sse4.2") => smallvec!["sse4.2", "crc32"], - ("x86", "pclmulqdq") => smallvec!["pclmul"], - ("x86", "rdrand") => smallvec!["rdrnd"], - ("x86", "bmi1") => smallvec!["bmi"], - ("x86", "cmpxchg16b") => smallvec!["cx16"], - ("x86", "avx512vaes") => smallvec!["vaes"], - ("x86", "avx512gfni") => smallvec!["gfni"], - ("x86", "avx512vpclmulqdq") => smallvec!["vpclmulqdq"], - // NOTE: seems like GCC requires 'avx512bw' for 'avx512vbmi2'. - ("x86", "avx512vbmi2") => smallvec!["avx512vbmi2", "avx512bw"], - // NOTE: seems like GCC requires 'avx512bw' for 'avx512bitalg'. - ("x86", "avx512bitalg") => smallvec!["avx512bitalg", "avx512bw"], - ("aarch64", "rcpc2") => smallvec!["rcpc-immo"], - ("aarch64", "dpb") => smallvec!["ccpp"], - ("aarch64", "dpb2") => smallvec!["ccdp"], - ("aarch64", "frintts") => smallvec!["fptoint"], - ("aarch64", "fcma") => smallvec!["complxnum"], - ("aarch64", "pmuv3") => smallvec!["perfmon"], - ("aarch64", "paca") => smallvec!["pauth"], - ("aarch64", "pacg") => smallvec!["pauth"], - // Rust ties fp and neon together. In LLVM neon implicitly enables fp, - // but we manually enable neon when a feature only implicitly enables fp - ("aarch64", "f32mm") => smallvec!["f32mm", "neon"], - ("aarch64", "f64mm") => smallvec!["f64mm", "neon"], - ("aarch64", "fhm") => smallvec!["fp16fml", "neon"], - ("aarch64", "fp16") => smallvec!["fullfp16", "neon"], - ("aarch64", "jsconv") => smallvec!["jsconv", "neon"], - ("aarch64", "sve") => smallvec!["sve", "neon"], - ("aarch64", "sve2") => smallvec!["sve2", "neon"], - ("aarch64", "sve2-aes") => smallvec!["sve2-aes", "neon"], - ("aarch64", "sve2-sm4") => smallvec!["sve2-sm4", "neon"], - ("aarch64", "sve2-sha3") => smallvec!["sve2-sha3", "neon"], - ("aarch64", "sve2-bitperm") => smallvec!["sve2-bitperm", "neon"], - (_, s) => smallvec![s], + InlineAttr::None => None, } } @@ -77,6 +40,36 @@ pub fn from_fn_attrs<'gcc, 'tcx>( ) { let codegen_fn_attrs = cx.tcx.codegen_fn_attrs(instance.def_id()); + #[cfg(feature="master")] + { + let inline = + if codegen_fn_attrs.flags.contains(CodegenFnAttrFlags::NAKED) { + InlineAttr::Never + } + else if codegen_fn_attrs.inline == InlineAttr::None && instance.def.requires_inline(cx.tcx) { + InlineAttr::Hint + } + else { + codegen_fn_attrs.inline + }; + if let Some(attr) = inline_attr(cx, inline) { + func.add_attribute(attr); + } + + if codegen_fn_attrs.flags.contains(CodegenFnAttrFlags::COLD) { + func.add_attribute(FnAttribute::Cold); + } + if codegen_fn_attrs.flags.contains(CodegenFnAttrFlags::FFI_RETURNS_TWICE) { + func.add_attribute(FnAttribute::ReturnsTwice); + } + if codegen_fn_attrs.flags.contains(CodegenFnAttrFlags::FFI_PURE) { + func.add_attribute(FnAttribute::Pure); + } + if codegen_fn_attrs.flags.contains(CodegenFnAttrFlags::FFI_CONST) { + func.add_attribute(FnAttribute::Const); + } + } + let function_features = codegen_fn_attrs.target_features.iter().map(|features| features.as_str()).collect::<Vec<&str>>(); @@ -101,11 +94,33 @@ pub fn from_fn_attrs<'gcc, 'tcx>( })) .collect::<Vec<_>>(); - // TODO(antoyo): check if we really need global backend features. (Maybe they could be applied - // globally?) + // TODO(antoyo): cg_llvm adds global features to each function so that LTO keep them. + // Check if GCC requires the same. let mut global_features = cx.tcx.global_backend_features(()).iter().map(|s| s.as_str()); function_features.extend(&mut global_features); - let target_features = function_features.join(","); + let target_features = function_features + .iter() + .filter_map(|feature| { + // FIXME(antoyo): for some reasons, disabling SSE results in the following error when + // compiling Rust for Linux: + // SSE register return with SSE disabled + // TODO(antoyo): support soft-float and retpoline-external-thunk. + if feature.contains("soft-float") || feature.contains("retpoline-external-thunk") || *feature == "-sse" { + return None; + } + + if feature.starts_with('-') { + Some(format!("no{}", feature)) + } + else if feature.starts_with('+') { + Some(feature[1..].to_string()) + } + else { + Some(feature.to_string()) + } + }) + .collect::<Vec<_>>() + .join(","); if !target_features.is_empty() { #[cfg(feature="master")] func.add_attribute(FnAttribute::Target(&target_features)); diff --git a/compiler/rustc_codegen_gcc/src/back/lto.rs b/compiler/rustc_codegen_gcc/src/back/lto.rs new file mode 100644 index 00000000000..529454b119e --- /dev/null +++ b/compiler/rustc_codegen_gcc/src/back/lto.rs @@ -0,0 +1,341 @@ +/// GCC requires to use the same toolchain for the whole compilation when doing LTO. +/// So, we need the same version/commit of the linker (gcc) and lto front-end binaries (lto1, +/// lto-wrapper, liblto_plugin.so). + +// FIXME(antoyo): the executables compiled with LTO are bigger than those compiled without LTO. +// Since it is the opposite for cg_llvm, check if this is normal. +// +// Maybe we embed the bitcode in the final binary? +// It doesn't look like we try to generate fat objects for the final binary. +// Check if the way we combine the object files make it keep the LTO sections on the final link. +// Maybe that's because the combined object files contain the IR (true) and the final link +// does not remove it? +// +// TODO(antoyo): for performance, check which optimizations the C++ frontend enables. +// +// Fix these warnings: +// /usr/bin/ld: warning: type of symbol `_RNvNvNvNtCs5JWOrf9uCus_5rayon11thread_pool19WORKER_THREAD_STATE7___getit5___KEY' changed from 1 to 6 in /tmp/ccKeUSiR.ltrans0.ltrans.o +// /usr/bin/ld: warning: type of symbol `_RNvNvNvNvNtNtNtCsAj5i4SGTR7_3std4sync4mpmc5waker17current_thread_id5DUMMY7___getit5___KEY' changed from 1 to 6 in /tmp/ccKeUSiR.ltrans0.ltrans.o +// /usr/bin/ld: warning: incremental linking of LTO and non-LTO objects; using -flinker-output=nolto-rel which will bypass whole program optimization + +use std::ffi::CString; +use std::fs::{self, File}; +use std::path::{Path, PathBuf}; + +use gccjit::OutputKind; +use object::read::archive::ArchiveFile; +use rustc_codegen_ssa::back::lto::{LtoModuleCodegen, SerializedModule}; +use rustc_codegen_ssa::back::symbol_export; +use rustc_codegen_ssa::back::write::{CodegenContext, FatLtoInput}; +use rustc_codegen_ssa::traits::*; +use rustc_codegen_ssa::{looks_like_rust_object_file, ModuleCodegen, ModuleKind}; +use rustc_data_structures::memmap::Mmap; +use rustc_errors::{FatalError, Handler}; +use rustc_hir::def_id::LOCAL_CRATE; +use rustc_middle::dep_graph::WorkProduct; +use rustc_middle::middle::exported_symbols::{SymbolExportInfo, SymbolExportLevel}; +use rustc_session::config::{CrateType, Lto}; +use tempfile::{TempDir, tempdir}; + +use crate::back::write::save_temp_bitcode; +use crate::errors::{ + DynamicLinkingWithLTO, LtoBitcodeFromRlib, LtoDisallowed, LtoDylib, +}; +use crate::{GccCodegenBackend, GccContext, to_gcc_opt_level}; + +/// We keep track of the computed LTO cache keys from the previous +/// session to determine which CGUs we can reuse. +//pub const THIN_LTO_KEYS_INCR_COMP_FILE_NAME: &str = "thin-lto-past-keys.bin"; + +pub fn crate_type_allows_lto(crate_type: CrateType) -> bool { + match crate_type { + CrateType::Executable | CrateType::Dylib | CrateType::Staticlib | CrateType::Cdylib => true, + CrateType::Rlib | CrateType::ProcMacro => false, + } +} + +struct LtoData { + // TODO(antoyo): use symbols_below_threshold. + //symbols_below_threshold: Vec<CString>, + upstream_modules: Vec<(SerializedModule<ModuleBuffer>, CString)>, + tmp_path: TempDir, +} + +fn prepare_lto(cgcx: &CodegenContext<GccCodegenBackend>, diag_handler: &Handler) -> Result<LtoData, FatalError> { + let export_threshold = match cgcx.lto { + // We're just doing LTO for our one crate + Lto::ThinLocal => SymbolExportLevel::Rust, + + // We're doing LTO for the entire crate graph + Lto::Fat | Lto::Thin => symbol_export::crates_export_threshold(&cgcx.crate_types), + + Lto::No => panic!("didn't request LTO but we're doing LTO"), + }; + + let tmp_path = + match tempdir() { + Ok(tmp_path) => tmp_path, + Err(error) => { + eprintln!("Cannot create temporary directory: {}", error); + return Err(FatalError); + }, + }; + + let symbol_filter = &|&(ref name, info): &(String, SymbolExportInfo)| { + if info.level.is_below_threshold(export_threshold) || info.used { + Some(CString::new(name.as_str()).unwrap()) + } else { + None + } + }; + let exported_symbols = cgcx.exported_symbols.as_ref().expect("needs exported symbols for LTO"); + let mut symbols_below_threshold = { + let _timer = cgcx.prof.generic_activity("GCC_lto_generate_symbols_below_threshold"); + exported_symbols[&LOCAL_CRATE].iter().filter_map(symbol_filter).collect::<Vec<CString>>() + }; + info!("{} symbols to preserve in this crate", symbols_below_threshold.len()); + + // If we're performing LTO for the entire crate graph, then for each of our + // upstream dependencies, find the corresponding rlib and load the bitcode + // from the archive. + // + // We save off all the bytecode and GCC module file path for later processing + // with either fat or thin LTO + let mut upstream_modules = Vec::new(); + if cgcx.lto != Lto::ThinLocal { + // Make sure we actually can run LTO + for crate_type in cgcx.crate_types.iter() { + if !crate_type_allows_lto(*crate_type) { + diag_handler.emit_err(LtoDisallowed); + return Err(FatalError); + } else if *crate_type == CrateType::Dylib { + if !cgcx.opts.unstable_opts.dylib_lto { + diag_handler.emit_err(LtoDylib); + return Err(FatalError); + } + } + } + + if cgcx.opts.cg.prefer_dynamic && !cgcx.opts.unstable_opts.dylib_lto { + diag_handler.emit_err(DynamicLinkingWithLTO); + return Err(FatalError); + } + + for &(cnum, ref path) in cgcx.each_linked_rlib_for_lto.iter() { + let exported_symbols = + cgcx.exported_symbols.as_ref().expect("needs exported symbols for LTO"); + { + let _timer = + cgcx.prof.generic_activity("GCC_lto_generate_symbols_below_threshold"); + symbols_below_threshold + .extend(exported_symbols[&cnum].iter().filter_map(symbol_filter)); + } + + let archive_data = unsafe { + Mmap::map(File::open(&path).expect("couldn't open rlib")) + .expect("couldn't map rlib") + }; + let archive = ArchiveFile::parse(&*archive_data).expect("wanted an rlib"); + let obj_files = archive + .members() + .filter_map(|child| { + child.ok().and_then(|c| { + std::str::from_utf8(c.name()).ok().map(|name| (name.trim(), c)) + }) + }) + .filter(|&(name, _)| looks_like_rust_object_file(name)); + for (name, child) in obj_files { + info!("adding bitcode from {}", name); + let path = tmp_path.path().join(name); + match save_as_file(child.data(&*archive_data).expect("corrupt rlib"), &path) { + Ok(()) => { + let buffer = ModuleBuffer::new(path); + let module = SerializedModule::Local(buffer); + upstream_modules.push((module, CString::new(name).unwrap())); + } + Err(e) => { + diag_handler.emit_err(e); + return Err(FatalError); + } + } + } + } + } + + Ok(LtoData { + //symbols_below_threshold, + upstream_modules, + tmp_path, + }) +} + +fn save_as_file(obj: &[u8], path: &Path) -> Result<(), LtoBitcodeFromRlib> { + fs::write(path, obj) + .map_err(|error| LtoBitcodeFromRlib { + gcc_err: format!("write object file to temp dir: {}", error) + }) +} + +/// Performs fat LTO by merging all modules into a single one and returning it +/// for further optimization. +pub(crate) fn run_fat( + cgcx: &CodegenContext<GccCodegenBackend>, + modules: Vec<FatLtoInput<GccCodegenBackend>>, + cached_modules: Vec<(SerializedModule<ModuleBuffer>, WorkProduct)>, +) -> Result<LtoModuleCodegen<GccCodegenBackend>, FatalError> { + let diag_handler = cgcx.create_diag_handler(); + let lto_data = prepare_lto(cgcx, &diag_handler)?; + /*let symbols_below_threshold = + lto_data.symbols_below_threshold.iter().map(|c| c.as_ptr()).collect::<Vec<_>>();*/ + fat_lto(cgcx, &diag_handler, modules, cached_modules, lto_data.upstream_modules, lto_data.tmp_path, + //&symbols_below_threshold, + ) +} + +fn fat_lto(cgcx: &CodegenContext<GccCodegenBackend>, _diag_handler: &Handler, modules: Vec<FatLtoInput<GccCodegenBackend>>, cached_modules: Vec<(SerializedModule<ModuleBuffer>, WorkProduct)>, mut serialized_modules: Vec<(SerializedModule<ModuleBuffer>, CString)>, tmp_path: TempDir, + //symbols_below_threshold: &[*const libc::c_char], +) -> Result<LtoModuleCodegen<GccCodegenBackend>, FatalError> { + let _timer = cgcx.prof.generic_activity("GCC_fat_lto_build_monolithic_module"); + info!("going for a fat lto"); + + // Sort out all our lists of incoming modules into two lists. + // + // * `serialized_modules` (also and argument to this function) contains all + // modules that are serialized in-memory. + // * `in_memory` contains modules which are already parsed and in-memory, + // such as from multi-CGU builds. + // + // All of `cached_modules` (cached from previous incremental builds) can + // immediately go onto the `serialized_modules` modules list and then we can + // split the `modules` array into these two lists. + let mut in_memory = Vec::new(); + serialized_modules.extend(cached_modules.into_iter().map(|(buffer, wp)| { + info!("pushing cached module {:?}", wp.cgu_name); + (buffer, CString::new(wp.cgu_name).unwrap()) + })); + for module in modules { + match module { + FatLtoInput::InMemory(m) => in_memory.push(m), + FatLtoInput::Serialized { name, buffer } => { + info!("pushing serialized module {:?}", name); + let buffer = SerializedModule::Local(buffer); + serialized_modules.push((buffer, CString::new(name).unwrap())); + } + } + } + + // Find the "costliest" module and merge everything into that codegen unit. + // All the other modules will be serialized and reparsed into the new + // context, so this hopefully avoids serializing and parsing the largest + // codegen unit. + // + // Additionally use a regular module as the base here to ensure that various + // file copy operations in the backend work correctly. The only other kind + // of module here should be an allocator one, and if your crate is smaller + // than the allocator module then the size doesn't really matter anyway. + let costliest_module = in_memory + .iter() + .enumerate() + .filter(|&(_, module)| module.kind == ModuleKind::Regular) + .map(|(i, _module)| { + //let cost = unsafe { llvm::LLVMRustModuleCost(module.module_llvm.llmod()) }; + // TODO(antoyo): compute the cost of a module if GCC allows this. + (0, i) + }) + .max(); + + // If we found a costliest module, we're good to go. Otherwise all our + // inputs were serialized which could happen in the case, for example, that + // all our inputs were incrementally reread from the cache and we're just + // re-executing the LTO passes. If that's the case deserialize the first + // module and create a linker with it. + let mut module: ModuleCodegen<GccContext> = match costliest_module { + Some((_cost, i)) => in_memory.remove(i), + None => { + unimplemented!("Incremental"); + /*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: GccContext::parse(cgcx, &name, buffer.data(), diag_handler)?, + name: name.into_string().unwrap(), + kind: ModuleKind::Regular, + }*/ + } + }; + let mut serialized_bitcode = Vec::new(); + { + info!("using {:?} as a base module", module.name); + + // We cannot load and merge GCC contexts in memory like cg_llvm is doing. + // Instead, we combine the object files into a single object file. + for module in in_memory { + let path = tmp_path.path().to_path_buf().join(&module.name); + let path = path.to_str().expect("path"); + let context = &module.module_llvm.context; + let config = cgcx.config(module.kind); + // NOTE: we need to set the optimization level here in order for LTO to do its job. + context.set_optimization_level(to_gcc_opt_level(config.opt_level)); + context.add_command_line_option("-flto=auto"); + context.add_command_line_option("-flto-partition=one"); + context.compile_to_file(OutputKind::ObjectFile, path); + let buffer = ModuleBuffer::new(PathBuf::from(path)); + let llmod_id = CString::new(&module.name[..]).unwrap(); + serialized_modules.push((SerializedModule::Local(buffer), llmod_id)); + } + // Sort the modules to ensure we produce deterministic results. + serialized_modules.sort_by(|module1, module2| module1.1.cmp(&module2.1)); + + // We add the object files and save in should_combine_object_files that we should combine + // them into a single object file when compiling later. + for (bc_decoded, name) in serialized_modules { + let _timer = cgcx + .prof + .generic_activity_with_arg_recorder("GCC_fat_lto_link_module", |recorder| { + recorder.record_arg(format!("{:?}", name)) + }); + info!("linking {:?}", name); + match bc_decoded { + SerializedModule::Local(ref module_buffer) => { + module.module_llvm.should_combine_object_files = true; + module.module_llvm.context.add_driver_option(module_buffer.0.to_str().expect("path")); + }, + SerializedModule::FromRlib(_) => unimplemented!("from rlib"), + SerializedModule::FromUncompressedFile(_) => unimplemented!("from uncompressed file"), + } + serialized_bitcode.push(bc_decoded); + } + save_temp_bitcode(cgcx, &module, "lto.input"); + + // Internalize everything below threshold to help strip out more modules and such. + /*unsafe { + let ptr = symbols_below_threshold.as_ptr(); + llvm::LLVMRustRunRestrictionPass( + llmod, + ptr as *const *const libc::c_char, + symbols_below_threshold.len() as libc::size_t, + );*/ + save_temp_bitcode(cgcx, &module, "lto.after-restriction"); + //} + } + + // NOTE: save the temporary directory used by LTO so that it gets deleted after linking instead + // of now. + module.module_llvm.temp_dir = Some(tmp_path); + + Ok(LtoModuleCodegen::Fat { module, _serialized_bitcode: serialized_bitcode }) +} + +pub struct ModuleBuffer(PathBuf); + +impl ModuleBuffer { + pub fn new(path: PathBuf) -> ModuleBuffer { + ModuleBuffer(path) + } +} + +impl ModuleBufferMethods for ModuleBuffer { + fn data(&self) -> &[u8] { + unimplemented!("data not needed for GCC codegen"); + } +} diff --git a/compiler/rustc_codegen_gcc/src/back/mod.rs b/compiler/rustc_codegen_gcc/src/back/mod.rs index d692799d764..10187eab0d7 100644 --- a/compiler/rustc_codegen_gcc/src/back/mod.rs +++ b/compiler/rustc_codegen_gcc/src/back/mod.rs @@ -1 +1,2 @@ +pub mod lto; pub mod write; diff --git a/compiler/rustc_codegen_gcc/src/back/write.rs b/compiler/rustc_codegen_gcc/src/back/write.rs index 5f54ac4ebc6..04772d7707a 100644 --- a/compiler/rustc_codegen_gcc/src/back/write.rs +++ b/compiler/rustc_codegen_gcc/src/back/write.rs @@ -2,27 +2,71 @@ use std::{env, fs}; use gccjit::OutputKind; use rustc_codegen_ssa::{CompiledModule, ModuleCodegen}; -use rustc_codegen_ssa::back::write::{CodegenContext, EmitObj, ModuleConfig}; +use rustc_codegen_ssa::back::link::ensure_removed; +use rustc_codegen_ssa::back::write::{BitcodeSection, CodegenContext, EmitObj, ModuleConfig}; use rustc_errors::Handler; +use rustc_fs_util::link_or_copy; use rustc_session::config::OutputType; use rustc_span::fatal_error::FatalError; use rustc_target::spec::SplitDebuginfo; use crate::{GccCodegenBackend, GccContext}; +use crate::errors::CopyBitcode; -pub(crate) unsafe fn codegen(cgcx: &CodegenContext<GccCodegenBackend>, _diag_handler: &Handler, module: ModuleCodegen<GccContext>, config: &ModuleConfig) -> Result<CompiledModule, FatalError> { - let _timer = cgcx.prof.generic_activity_with_arg("LLVM_module_codegen", &*module.name); +pub(crate) unsafe fn codegen(cgcx: &CodegenContext<GccCodegenBackend>, diag_handler: &Handler, module: ModuleCodegen<GccContext>, config: &ModuleConfig) -> Result<CompiledModule, FatalError> { + let _timer = cgcx.prof.generic_activity_with_arg("GCC_module_codegen", &*module.name); { let context = &module.module_llvm.context; let module_name = module.name.clone(); + + let should_combine_object_files = module.module_llvm.should_combine_object_files; + let module_name = Some(&module_name[..]); - let _bc_out = cgcx.output_filenames.temp_path(OutputType::Bitcode, module_name); + // NOTE: Only generate object files with GIMPLE when this environment variable is set for + // now because this requires a particular setup (same gcc/lto1/lto-wrapper commit as libgccjit). + let fat_lto = env::var("EMBED_LTO_BITCODE").as_deref() == Ok("1"); + + let bc_out = cgcx.output_filenames.temp_path(OutputType::Bitcode, module_name); let obj_out = cgcx.output_filenames.temp_path(OutputType::Object, module_name); - if config.bitcode_needed() { + if config.bitcode_needed() && fat_lto { + let _timer = cgcx + .prof + .generic_activity_with_arg("GCC_module_codegen_make_bitcode", &*module.name); + // TODO(antoyo) + /*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_bc || config.emit_obj == EmitObj::Bitcode { + let _timer = cgcx + .prof + .generic_activity_with_arg("GCC_module_codegen_emit_bitcode", &*module.name); + context.add_command_line_option("-flto=auto"); + context.add_command_line_option("-flto-partition=one"); + context.compile_to_file(OutputKind::ObjectFile, bc_out.to_str().expect("path to str")); + } + + if config.emit_obj == EmitObj::ObjectCode(BitcodeSection::Full) { + let _timer = cgcx + .prof + .generic_activity_with_arg("GCC_module_codegen_embed_bitcode", &*module.name); + // TODO(antoyo): maybe we should call embed_bitcode to have the proper iOS fixes? + //embed_bitcode(cgcx, llcx, llmod, &config.bc_cmdline, data); + + context.add_command_line_option("-flto=auto"); + context.add_command_line_option("-flto-partition=one"); + context.add_command_line_option("-ffat-lto-objects"); + // TODO(antoyo): Send -plugin/usr/lib/gcc/x86_64-pc-linux-gnu/11.1.0/liblto_plugin.so to linker (this should be done when specifying the appropriate rustc cli argument). + context.compile_to_file(OutputKind::ObjectFile, bc_out.to_str().expect("path to str")); + } } if config.emit_ir { @@ -32,7 +76,7 @@ pub(crate) unsafe fn codegen(cgcx: &CodegenContext<GccCodegenBackend>, _diag_han if config.emit_asm { let _timer = cgcx .prof - .generic_activity_with_arg("LLVM_module_codegen_emit_asm", &*module.name); + .generic_activity_with_arg("GCC_module_codegen_emit_asm", &*module.name); let path = cgcx.output_filenames.temp_path(OutputType::Assembly, module_name); context.compile_to_file(OutputKind::Assembler, path.to_str().expect("path to str")); } @@ -41,7 +85,7 @@ pub(crate) unsafe fn codegen(cgcx: &CodegenContext<GccCodegenBackend>, _diag_han EmitObj::ObjectCode(_) => { let _timer = cgcx .prof - .generic_activity_with_arg("LLVM_module_codegen_emit_obj", &*module.name); + .generic_activity_with_arg("GCC_module_codegen_emit_obj", &*module.name); if env::var("CG_GCCJIT_DUMP_MODULE_NAMES").as_deref() == Ok("1") { println!("Module {}", module.name); } @@ -60,11 +104,36 @@ pub(crate) unsafe fn codegen(cgcx: &CodegenContext<GccCodegenBackend>, _diag_han context.set_debug_info(true); context.dump_to_file(path, true); } - context.compile_to_file(OutputKind::ObjectFile, obj_out.to_str().expect("path to str")); + if should_combine_object_files && fat_lto { + context.add_command_line_option("-flto=auto"); + context.add_command_line_option("-flto-partition=one"); + + context.add_driver_option("-Wl,-r"); + // NOTE: we need -nostdlib, otherwise, we get the following error: + // /usr/bin/ld: cannot find -lgcc_s: No such file or directory + context.add_driver_option("-nostdlib"); + // NOTE: without -fuse-linker-plugin, we get the following error: + // lto1: internal compiler error: decompressed stream: Destination buffer is too small + context.add_driver_option("-fuse-linker-plugin"); + + // NOTE: this doesn't actually generate an executable. With the above flags, it combines the .o files together in another .o. + context.compile_to_file(OutputKind::Executable, obj_out.to_str().expect("path to str")); + } + else { + context.compile_to_file(OutputKind::ObjectFile, obj_out.to_str().expect("path to str")); + } } EmitObj::Bitcode => { - // TODO(antoyo) + debug!("copying bitcode {:?} to obj {:?}", bc_out, obj_out); + if let Err(err) = link_or_copy(&bc_out, &obj_out) { + diag_handler.emit_err(CopyBitcode { err }); + } + + if !config.emit_bc { + debug!("removing_bitcode {:?}", bc_out); + ensure_removed(diag_handler, &bc_out); + } } EmitObj::None => {} @@ -82,3 +151,18 @@ pub(crate) unsafe fn codegen(cgcx: &CodegenContext<GccCodegenBackend>, _diag_han pub(crate) fn link(_cgcx: &CodegenContext<GccCodegenBackend>, _diag_handler: &Handler, mut _modules: Vec<ModuleCodegen<GccContext>>) -> Result<ModuleCodegen<GccContext>, FatalError> { unimplemented!(); } + +pub(crate) fn save_temp_bitcode(cgcx: &CodegenContext<GccCodegenBackend>, _module: &ModuleCodegen<GccContext>, _name: &str) { + if !cgcx.save_temps { + return; + } + unimplemented!(); + /*unsafe { + let ext = format!("{}.bc", name); + let cgu = Some(&module.name[..]); + let path = cgcx.output_filenames.temp_path_ext(&ext, cgu); + let cstr = path_to_c_string(&path); + let llmod = module.module_llvm.llmod(); + llvm::LLVMWriteBitcodeToFile(llmod, cstr.as_ptr()); + }*/ +} diff --git a/compiler/rustc_codegen_gcc/src/base.rs b/compiler/rustc_codegen_gcc/src/base.rs index dcd560b3dcd..b081e9ff2fd 100644 --- a/compiler/rustc_codegen_gcc/src/base.rs +++ b/compiler/rustc_codegen_gcc/src/base.rs @@ -1,3 +1,4 @@ +use std::collections::HashSet; use std::env; use std::time::Instant; @@ -18,6 +19,7 @@ use rustc_codegen_ssa::traits::DebugInfoMethods; use rustc_session::config::DebugInfo; use rustc_span::Symbol; +use crate::{LockedTargetInfo, gcc_util}; use crate::GccContext; use crate::builder::Builder; use crate::context::CodegenCx; @@ -50,6 +52,7 @@ pub fn global_linkage_to_gcc(linkage: Linkage) -> GlobalKind { pub fn linkage_to_gcc(linkage: Linkage) -> FunctionType { match linkage { Linkage::External => FunctionType::Exported, + // TODO(antoyo): set the attribute externally_visible. Linkage::AvailableExternally => FunctionType::Extern, Linkage::LinkOnceAny => unimplemented!(), Linkage::LinkOnceODR => unimplemented!(), @@ -63,7 +66,7 @@ pub fn linkage_to_gcc(linkage: Linkage) -> FunctionType { } } -pub fn compile_codegen_unit(tcx: TyCtxt<'_>, cgu_name: Symbol, supports_128bit_integers: bool) -> (ModuleCodegen<GccContext>, u64) { +pub fn compile_codegen_unit(tcx: TyCtxt<'_>, cgu_name: Symbol, target_info: LockedTargetInfo) -> (ModuleCodegen<GccContext>, u64) { let prof_timer = tcx.prof.generic_activity("codegen_module"); let start_time = Instant::now(); @@ -71,7 +74,7 @@ pub fn compile_codegen_unit(tcx: TyCtxt<'_>, cgu_name: Symbol, supports_128bit_i let (module, _) = tcx.dep_graph.with_task( dep_node, tcx, - (cgu_name, supports_128bit_integers), + (cgu_name, target_info), module_codegen, Some(dep_graph::hash_result), ); @@ -82,38 +85,28 @@ pub fn compile_codegen_unit(tcx: TyCtxt<'_>, cgu_name: Symbol, supports_128bit_i // the time we needed for codegenning it. let cost = time_to_codegen.as_secs() * 1_000_000_000 + time_to_codegen.subsec_nanos() as u64; - fn module_codegen(tcx: TyCtxt<'_>, (cgu_name, supports_128bit_integers): (Symbol, bool)) -> ModuleCodegen<GccContext> { + fn module_codegen(tcx: TyCtxt<'_>, (cgu_name, target_info): (Symbol, LockedTargetInfo)) -> ModuleCodegen<GccContext> { let cgu = tcx.codegen_unit(cgu_name); // Instantiate monomorphizations without filling out definitions yet... - //let llvm_module = ModuleLlvm::new(tcx, &cgu_name.as_str()); let context = Context::default(); context.add_command_line_option("-fexceptions"); context.add_driver_option("-fexceptions"); + let disabled_features: HashSet<_> = tcx.sess.opts.cg.target_feature.split(',') + .filter(|feature| feature.starts_with('-')) + .map(|string| &string[1..]) + .collect(); + // TODO(antoyo): only set on x86 platforms. context.add_command_line_option("-masm=intel"); - // TODO(antoyo): only add the following cli argument if the feature is supported. - context.add_command_line_option("-msse2"); - context.add_command_line_option("-mavx2"); - // FIXME(antoyo): the following causes an illegal instruction on vmovdqu64 in std_example on my CPU. - // Only add if the CPU supports it. - context.add_command_line_option("-msha"); - context.add_command_line_option("-mpclmul"); - context.add_command_line_option("-mfma"); - context.add_command_line_option("-mfma4"); - context.add_command_line_option("-m64"); - context.add_command_line_option("-mbmi"); - context.add_command_line_option("-mgfni"); - //context.add_command_line_option("-mavxvnni"); // The CI doesn't support this option. - context.add_command_line_option("-mf16c"); - context.add_command_line_option("-maes"); - context.add_command_line_option("-mxsavec"); - context.add_command_line_option("-mbmi2"); - context.add_command_line_option("-mrtm"); - context.add_command_line_option("-mvaes"); - context.add_command_line_option("-mvpclmulqdq"); - context.add_command_line_option("-mavx"); + + if !disabled_features.contains("avx") { + // NOTE: we always enable AVX because the equivalent of llvm.x86.sse2.cmp.pd in GCC for + // SSE2 is multiple builtins, so we use the AVX __builtin_ia32_cmppd instead. + // FIXME(antoyo): use the proper builtins for llvm.x86.sse2.cmp.pd and similar. + context.add_command_line_option("-mavx"); + } for arg in &tcx.sess.opts.cg.llvm_args { context.add_command_line_option(arg); @@ -127,6 +120,16 @@ pub fn compile_codegen_unit(tcx: TyCtxt<'_>, cgu_name: Symbol, supports_128bit_i // NOTE: Rust relies on LLVM doing wrapping on overflow. context.add_command_line_option("-fwrapv"); + if tcx.sess.relocation_model() == rustc_target::spec::RelocModel::Static { + context.add_command_line_option("-mcmodel=kernel"); + context.add_command_line_option("-fno-pie"); + } + + let target_cpu = gcc_util::target_cpu(tcx.sess); + if target_cpu != "generic" { + context.add_command_line_option(&format!("-march={}", target_cpu)); + } + if tcx.sess.opts.unstable_opts.function_sections.unwrap_or(tcx.sess.target.function_sections) { context.add_command_line_option("-ffunction-sections"); context.add_command_line_option("-fdata-sections"); @@ -135,8 +138,14 @@ pub fn compile_codegen_unit(tcx: TyCtxt<'_>, cgu_name: Symbol, supports_128bit_i if env::var("CG_GCCJIT_DUMP_RTL").as_deref() == Ok("1") { context.add_command_line_option("-fdump-rtl-vregs"); } + if env::var("CG_GCCJIT_DUMP_RTL_ALL").as_deref() == Ok("1") { + context.add_command_line_option("-fdump-rtl-all"); + } if env::var("CG_GCCJIT_DUMP_TREE_ALL").as_deref() == Ok("1") { - context.add_command_line_option("-fdump-tree-all"); + context.add_command_line_option("-fdump-tree-all-eh"); + } + if env::var("CG_GCCJIT_DUMP_IPA_ALL").as_deref() == Ok("1") { + context.add_command_line_option("-fdump-ipa-all-eh"); } if env::var("CG_GCCJIT_DUMP_CODE").as_deref() == Ok("1") { context.set_dump_code_on_compile(true); @@ -152,15 +161,19 @@ pub fn compile_codegen_unit(tcx: TyCtxt<'_>, cgu_name: Symbol, supports_128bit_i context.set_keep_intermediates(true); } + if env::var("CG_GCCJIT_VERBOSE").as_deref() == Ok("1") { + context.add_driver_option("-v"); + } + // NOTE: The codegen generates unrechable blocks. context.set_allow_unreachable_blocks(true); { - let cx = CodegenCx::new(&context, cgu, tcx, supports_128bit_integers); + let cx = CodegenCx::new(&context, cgu, tcx, target_info.supports_128bit_int()); let mono_items = cgu.items_in_deterministic_order(tcx); - for &(mono_item, (linkage, visibility)) in &mono_items { - mono_item.predefine::<Builder<'_, '_, '_>>(&cx, linkage, visibility); + 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. @@ -181,7 +194,9 @@ pub fn compile_codegen_unit(tcx: TyCtxt<'_>, cgu_name: Symbol, supports_128bit_i ModuleCodegen { name: cgu_name.to_string(), module_llvm: GccContext { - context + context, + should_combine_object_files: false, + temp_dir: None, }, kind: ModuleKind::Regular, } diff --git a/compiler/rustc_codegen_gcc/src/builder.rs b/compiler/rustc_codegen_gcc/src/builder.rs index a3c8142bea2..b7841808934 100644 --- a/compiler/rustc_codegen_gcc/src/builder.rs +++ b/compiler/rustc_codegen_gcc/src/builder.rs @@ -27,7 +27,6 @@ use rustc_codegen_ssa::traits::{ BaseTypeMethods, BuilderMethods, ConstMethods, - DerivedTypeMethods, LayoutTypeMethods, HasCodegen, OverflowOp, @@ -35,6 +34,7 @@ use rustc_codegen_ssa::traits::{ }; use rustc_data_structures::fx::FxHashSet; use rustc_middle::bug; +use rustc_middle::middle::codegen_fn_attrs::CodegenFnAttrs; use rustc_middle::ty::{ParamEnv, Ty, TyCtxt}; use rustc_middle::ty::layout::{FnAbiError, FnAbiOfHelpers, FnAbiRequest, HasParamEnv, HasTyCtxt, LayoutError, LayoutOfHelpers, TyAndLayout}; use rustc_span::Span; @@ -180,6 +180,8 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> { }) .collect(); + debug_assert_eq!(casted_args.len(), args.len()); + Cow::Owned(casted_args) } @@ -206,7 +208,7 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> { let func_name = format!("{:?}", func_ptr); - let casted_args: Vec<_> = param_types + let mut casted_args: Vec<_> = param_types .into_iter() .zip(args.iter()) .enumerate() @@ -236,20 +238,18 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> { }) .collect(); + // NOTE: to take into account variadic functions. + for i in casted_args.len()..args.len() { + casted_args.push(args[i]); + } + Cow::Owned(casted_args) } fn check_store(&mut self, val: RValue<'gcc>, ptr: RValue<'gcc>) -> RValue<'gcc> { - let dest_ptr_ty = self.cx.val_ty(ptr).make_pointer(); // TODO(antoyo): make sure make_pointer() is okay here. let stored_ty = self.cx.val_ty(val); let stored_ptr_ty = self.cx.type_ptr_to(stored_ty); - - if dest_ptr_ty == stored_ptr_ty { - ptr - } - else { - self.bitcast(ptr, stored_ptr_ty) - } + self.bitcast(ptr, stored_ptr_ty) } pub fn current_func(&self) -> Function<'gcc> { @@ -279,8 +279,17 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> { } } - fn function_ptr_call(&mut self, func_ptr: RValue<'gcc>, args: &[RValue<'gcc>], _funclet: Option<&Funclet>) -> RValue<'gcc> { - let gcc_func = func_ptr.get_type().dyncast_function_ptr_type().expect("function ptr"); + fn function_ptr_call(&mut self, typ: Type<'gcc>, mut func_ptr: RValue<'gcc>, args: &[RValue<'gcc>], _funclet: Option<&Funclet>) -> RValue<'gcc> { + let gcc_func = + match func_ptr.get_type().dyncast_function_ptr_type() { + Some(func) => func, + None => { + // NOTE: due to opaque pointers now being used, we need to cast here. + let new_func_type = typ.dyncast_function_ptr_type().expect("function ptr"); + func_ptr = self.context.new_cast(None, func_ptr, typ); + new_func_type + }, + }; let func_name = format!("{:?}", func_ptr); let previous_arg_count = args.len(); let orig_args = args; @@ -423,16 +432,17 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> { self.llbb().end_with_void_return(None) } - fn ret(&mut self, value: RValue<'gcc>) { - let value = - if self.structs_as_pointer.borrow().contains(&value) { - // NOTE: hack to workaround a limitation of the rustc API: see comment on - // CodegenCx.structs_as_pointer - value.dereference(None).to_rvalue() - } - else { - value - }; + fn ret(&mut self, mut value: RValue<'gcc>) { + if self.structs_as_pointer.borrow().contains(&value) { + // NOTE: hack to workaround a limitation of the rustc API: see comment on + // CodegenCx.structs_as_pointer + value = value.dereference(None).to_rvalue(); + } + let expected_return_type = self.current_func().get_return_type(); + if !expected_return_type.is_compatible_with(value.get_type()) { + // NOTE: due to opaque pointers now being used, we need to cast here. + value = self.context.new_cast(None, value, expected_return_type); + } self.llbb().end_with_return(None, value); } @@ -455,12 +465,12 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> { } #[cfg(feature="master")] - fn invoke(&mut self, typ: Type<'gcc>, _fn_abi: Option<&FnAbi<'tcx, Ty<'tcx>>>, func: RValue<'gcc>, args: &[RValue<'gcc>], then: Block<'gcc>, catch: Block<'gcc>, _funclet: Option<&Funclet>) -> RValue<'gcc> { + fn invoke(&mut self, typ: Type<'gcc>, fn_attrs: Option<&CodegenFnAttrs>, _fn_abi: Option<&FnAbi<'tcx, Ty<'tcx>>>, func: RValue<'gcc>, args: &[RValue<'gcc>], then: Block<'gcc>, catch: Block<'gcc>, _funclet: Option<&Funclet>) -> RValue<'gcc> { let try_block = self.current_func().new_block("try"); let current_block = self.block.clone(); self.block = try_block; - let call = self.call(typ, None, func, args, None); // TODO(antoyo): use funclet here? + let call = self.call(typ, fn_attrs, None, func, args, None); // TODO(antoyo): use funclet here? self.block = current_block; let return_value = self.current_func() @@ -483,8 +493,8 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> { } #[cfg(not(feature="master"))] - fn invoke(&mut self, typ: Type<'gcc>, fn_abi: Option<&FnAbi<'tcx, Ty<'tcx>>>, func: RValue<'gcc>, args: &[RValue<'gcc>], then: Block<'gcc>, catch: Block<'gcc>, _funclet: Option<&Funclet>) -> RValue<'gcc> { - let call_site = self.call(typ, None, func, args, None); + fn invoke(&mut self, typ: Type<'gcc>, fn_attrs: Option<&CodegenFnAttrs>, fn_abi: Option<&FnAbi<'tcx, Ty<'tcx>>>, func: RValue<'gcc>, args: &[RValue<'gcc>], then: Block<'gcc>, catch: Block<'gcc>, _funclet: Option<&Funclet>) -> RValue<'gcc> { + let call_site = self.call(typ, fn_attrs, None, func, args, None); let condition = self.context.new_rvalue_from_int(self.bool_type, 1); self.llbb().end_with_conditional(None, condition, then, catch); if let Some(_fn_abi) = fn_abi { @@ -646,7 +656,7 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> { } fn unchecked_sadd(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { - a + b + self.gcc_add(a, b) } fn unchecked_uadd(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { @@ -654,7 +664,7 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> { } fn unchecked_ssub(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { - a - b + self.gcc_sub(a, b) } fn unchecked_usub(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { @@ -663,11 +673,11 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> { } fn unchecked_smul(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { - a * b + self.gcc_mul(a, b) } fn unchecked_umul(&mut self, a: RValue<'gcc>, b: RValue<'gcc>) -> RValue<'gcc> { - a * b + self.gcc_mul(a, b) } fn fadd_fast(&mut self, lhs: RValue<'gcc>, rhs: RValue<'gcc>) -> RValue<'gcc> { @@ -718,17 +728,25 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> { unimplemented!(); } - fn load(&mut self, pointee_ty: Type<'gcc>, ptr: RValue<'gcc>, _align: Align) -> RValue<'gcc> { + fn load(&mut self, pointee_ty: Type<'gcc>, ptr: RValue<'gcc>, align: Align) -> RValue<'gcc> { let block = self.llbb(); let function = block.get_function(); // NOTE: instead of returning the dereference here, we have to assign it to a variable in // the current basic block. Otherwise, it could be used in another basic block, causing a // dereference after a drop, for instance. - // TODO(antoyo): handle align of the load instruction. - let ptr = self.context.new_cast(None, ptr, pointee_ty.make_pointer()); + // FIXME(antoyo): this check that we don't call get_aligned() a second time on a type. + // Ideally, we shouldn't need to do this check. + let aligned_type = + if pointee_ty == self.cx.u128_type || pointee_ty == self.cx.i128_type { + pointee_ty + } + else { + pointee_ty.get_aligned(align.bytes()) + }; + let ptr = self.context.new_cast(None, ptr, aligned_type.make_pointer()); let deref = ptr.dereference(None).to_rvalue(); unsafe { RETURN_VALUE_COUNT += 1 }; - let loaded_value = function.new_local(None, pointee_ty, &format!("loadedValue{}", unsafe { RETURN_VALUE_COUNT })); + let loaded_value = function.new_local(None, aligned_type, &format!("loadedValue{}", unsafe { RETURN_VALUE_COUNT })); block.add_assignment(None, loaded_value, deref); loaded_value.to_rvalue() } @@ -757,7 +775,7 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> { assert_eq!(place.llextra.is_some(), place.layout.is_unsized()); if place.layout.is_zst() { - return OperandRef::new_zst(self, place.layout); + return OperandRef::zero_sized(place.layout); } fn scalar_load_metadata<'a, 'gcc, 'tcx>(bx: &mut Builder<'a, 'gcc, 'tcx>, load: RValue<'gcc>, scalar: &abi::Scalar) { @@ -796,7 +814,7 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> { let mut load = |i, scalar: &abi::Scalar, align| { let llptr = self.struct_gep(pair_type, place.llval, i as u64); - let llty = place.layout.scalar_pair_element_gcc_type(self, i, false); + let llty = place.layout.scalar_pair_element_gcc_type(self, i); let load = self.load(llty, llptr, align); scalar_load_metadata(self, load, scalar); if scalar.is_bool() { self.trunc(load, self.type_i1()) } else { load } @@ -891,7 +909,9 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> { .add_eval(None, self.context.new_call(None, atomic_store, &[ptr, value, ordering])); } - fn gep(&mut self, _typ: Type<'gcc>, ptr: RValue<'gcc>, indices: &[RValue<'gcc>]) -> RValue<'gcc> { + fn gep(&mut self, typ: Type<'gcc>, ptr: RValue<'gcc>, indices: &[RValue<'gcc>]) -> RValue<'gcc> { + // NOTE: due to opaque pointers now being used, we need to cast here. + let ptr = self.context.new_cast(None, ptr, typ.make_pointer()); let ptr_type = ptr.get_type(); let mut pointee_type = ptr.get_type(); // NOTE: we cannot use array indexing here like in inbounds_gep because array indexing is @@ -902,13 +922,21 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> { // require dereferencing the pointer. for index in indices { pointee_type = pointee_type.get_pointee().expect("pointee type"); + #[cfg(feature="master")] + let pointee_size = { + let size = self.cx.context.new_sizeof(pointee_type); + self.context.new_cast(None, size, index.get_type()) + }; + #[cfg(not(feature="master"))] let pointee_size = self.context.new_rvalue_from_int(index.get_type(), pointee_type.get_size() as i32); result = result + self.gcc_int_cast(*index * pointee_size, self.sizet_type); } self.context.new_bitcast(None, result, ptr_type) } - fn inbounds_gep(&mut self, _typ: Type<'gcc>, ptr: RValue<'gcc>, indices: &[RValue<'gcc>]) -> RValue<'gcc> { + fn inbounds_gep(&mut self, typ: Type<'gcc>, ptr: RValue<'gcc>, indices: &[RValue<'gcc>]) -> RValue<'gcc> { + // NOTE: due to opaque pointers now being used, we need to cast here. + let ptr = self.context.new_cast(None, ptr, typ.make_pointer()); // NOTE: array indexing is always considered in bounds in GCC (TODO(antoyo): to be verified). let mut indices = indices.into_iter(); let index = indices.next().expect("first index in inbounds_gep"); @@ -937,6 +965,8 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> { element.get_address(None) } else if let Some(struct_type) = value_type.is_struct() { + // NOTE: due to opaque pointers now being used, we need to bitcast here. + let ptr = self.bitcast_if_needed(ptr, value_type.make_pointer()); ptr.dereference_field(None, struct_type.get_field(idx as i32)).get_address(None) } else { @@ -1226,6 +1256,11 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> { (value1, value2) } + fn filter_landing_pad(&mut self, pers_fn: RValue<'gcc>) -> (RValue<'gcc>, RValue<'gcc>) { + // TODO(antoyo): generate the correct landing pad + self.cleanup_landing_pad(pers_fn) + } + #[cfg(feature="master")] fn resume(&mut self, exn0: RValue<'gcc>, _exn1: RValue<'gcc>) { let exn_type = exn0.get_type(); @@ -1350,7 +1385,8 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> { fn call( &mut self, - _typ: Type<'gcc>, + typ: Type<'gcc>, + _fn_attrs: Option<&CodegenFnAttrs>, fn_abi: Option<&FnAbi<'tcx, Ty<'tcx>>>, func: RValue<'gcc>, args: &[RValue<'gcc>], @@ -1363,7 +1399,7 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> { } else { // If it's a not function that was defined, it's a function pointer. - self.function_ptr_call(func, args, funclet) + self.function_ptr_call(typ, func, args, funclet) }; if let Some(_fn_abi) = fn_abi { // TODO(bjorn3): Apply function attributes @@ -1385,7 +1421,7 @@ impl<'a, 'gcc, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'gcc, 'tcx> { self.cx } - fn do_not_inline(&mut self, _llret: RValue<'gcc>) { + fn apply_attrs_to_cleanup_callsite(&mut self, _llret: RValue<'gcc>) { // FIXME(bjorn3): implement } @@ -1836,7 +1872,8 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> { #[cfg(feature="master")] let (cond, element_type) = { - let then_val_vector_type = then_val.get_type().dyncast_vector().expect("vector type"); + // TODO(antoyo): dyncast_vector should not require a call to unqualified. + let then_val_vector_type = then_val.get_type().unqualified().dyncast_vector().expect("vector type"); let then_val_element_type = then_val_vector_type.get_element_type(); let then_val_element_size = then_val_element_type.get_size(); diff --git a/compiler/rustc_codegen_gcc/src/callee.rs b/compiler/rustc_codegen_gcc/src/callee.rs index ba1e8656208..9fc77627b1b 100644 --- a/compiler/rustc_codegen_gcc/src/callee.rs +++ b/compiler/rustc_codegen_gcc/src/callee.rs @@ -17,8 +17,8 @@ use crate::context::CodegenCx; pub fn get_fn<'gcc, 'tcx>(cx: &CodegenCx<'gcc, 'tcx>, instance: Instance<'tcx>) -> Function<'gcc> { let tcx = cx.tcx(); - assert!(!instance.substs.needs_infer()); - assert!(!instance.substs.has_escaping_bound_vars()); + assert!(!instance.args.has_infer()); + assert!(!instance.args.has_escaping_bound_vars()); let sym = tcx.symbol_name(instance).name; @@ -100,7 +100,7 @@ pub fn get_fn<'gcc, 'tcx>(cx: &CodegenCx<'gcc, 'tcx>, instance: Instance<'tcx>) // whether we are sharing generics or not. The important thing here is // that the visibility we apply to the declaration is the same one that // has been applied to the definition (wherever that definition may be). - let is_generic = instance.substs.non_erasable_generics().next().is_some(); + let is_generic = instance.args.non_erasable_generics(tcx, instance.def_id()).next().is_some(); if is_generic { // This is a monomorphization. Its expected visibility depends diff --git a/compiler/rustc_codegen_gcc/src/common.rs b/compiler/rustc_codegen_gcc/src/common.rs index ac04b61a306..5f54cb16d8e 100644 --- a/compiler/rustc_codegen_gcc/src/common.rs +++ b/compiler/rustc_codegen_gcc/src/common.rs @@ -1,23 +1,25 @@ use gccjit::LValue; use gccjit::{RValue, Type, ToRValue}; -use rustc_codegen_ssa::mir::place::PlaceRef; use rustc_codegen_ssa::traits::{ BaseTypeMethods, ConstMethods, - DerivedTypeMethods, MiscMethods, StaticMethods, }; use rustc_middle::mir::Mutability; -use rustc_middle::ty::layout::{TyAndLayout, LayoutOf}; +use rustc_middle::ty::layout::{LayoutOf}; use rustc_middle::mir::interpret::{ConstAllocation, GlobalAlloc, Scalar}; -use rustc_target::abi::{self, HasDataLayout, Pointer, Size}; +use rustc_target::abi::{self, HasDataLayout, Pointer}; use crate::consts::const_alloc_to_gcc; use crate::context::CodegenCx; use crate::type_of::LayoutGccExt; impl<'gcc, 'tcx> CodegenCx<'gcc, 'tcx> { + pub fn const_ptrcast(&self, val: RValue<'gcc>, ty: Type<'gcc>) -> RValue<'gcc> { + self.context.new_cast(None, val, ty) + } + pub fn const_bytes(&self, bytes: &[u8]) -> RValue<'gcc> { bytes_in_context(self, bytes) } @@ -110,6 +112,10 @@ impl<'gcc, 'tcx> ConstMethods<'tcx> for CodegenCx<'gcc, 'tcx> { self.const_uint(self.type_u64(), i) } + fn const_u128(&self, i: u128) -> RValue<'gcc> { + self.const_uint_big(self.type_u128(), i) + } + fn const_usize(&self, i: u64) -> RValue<'gcc> { let bit_size = self.data_layout().pointer_size.bits(); if bit_size < 64 { @@ -240,27 +246,21 @@ impl<'gcc, 'tcx> ConstMethods<'tcx> for CodegenCx<'gcc, 'tcx> { const_alloc_to_gcc(self, alloc) } - fn from_const_alloc(&self, layout: TyAndLayout<'tcx>, alloc: ConstAllocation<'tcx>, offset: Size) -> PlaceRef<'tcx, RValue<'gcc>> { - assert_eq!(alloc.inner().align, layout.align.abi); - let ty = self.type_ptr_to(layout.gcc_type(self)); - let value = - if layout.size == Size::ZERO { - let value = self.const_usize(alloc.inner().align.bytes()); - self.const_bitcast(value, ty) + fn const_bitcast(&self, value: RValue<'gcc>, typ: Type<'gcc>) -> RValue<'gcc> { + if value.get_type() == self.bool_type.make_pointer() { + if let Some(pointee) = typ.get_pointee() { + if pointee.dyncast_vector().is_some() { + panic!() + } } - else { - let init = const_alloc_to_gcc(self, alloc); - let base_addr = self.static_addr_of(init, alloc.inner().align, None); - - let array = self.const_bitcast(base_addr, self.type_i8p()); - let value = self.context.new_array_access(None, array, self.const_usize(offset.bytes())).get_address(None); - self.const_bitcast(value, ty) - }; - PlaceRef::new_sized(value, layout) + } + // NOTE: since bitcast makes a value non-constant, don't bitcast if not necessary as some + // SIMD builtins require a constant value. + self.bitcast_if_needed(value, typ) } - fn const_ptrcast(&self, val: RValue<'gcc>, ty: Type<'gcc>) -> RValue<'gcc> { - self.context.new_cast(None, val, ty) + fn const_ptr_byte_offset(&self, base_addr: Self::Value, offset: abi::Size) -> Self::Value { + self.context.new_array_access(None, base_addr, self.const_usize(offset.bytes())).get_address(None) } } diff --git a/compiler/rustc_codegen_gcc/src/consts.rs b/compiler/rustc_codegen_gcc/src/consts.rs index 792ab8f890d..d8a1fd315c0 100644 --- a/compiler/rustc_codegen_gcc/src/consts.rs +++ b/compiler/rustc_codegen_gcc/src/consts.rs @@ -1,6 +1,6 @@ #[cfg(feature = "master")] -use gccjit::FnAttribute; -use gccjit::{Function, GlobalKind, LValue, RValue, ToRValue, Type}; +use gccjit::{FnAttribute, VarAttribute, Visibility}; +use gccjit::{Function, GlobalKind, LValue, RValue, ToRValue}; use rustc_codegen_ssa::traits::{BaseTypeMethods, ConstMethods, DerivedTypeMethods, StaticMethods}; use rustc_middle::span_bug; use rustc_middle::middle::codegen_fn_attrs::{CodegenFnAttrFlags, CodegenFnAttrs}; @@ -16,21 +16,6 @@ use crate::context::CodegenCx; use crate::errors::InvalidMinimumAlignment; use crate::type_of::LayoutGccExt; -impl<'gcc, 'tcx> CodegenCx<'gcc, 'tcx> { - pub fn const_bitcast(&self, value: RValue<'gcc>, typ: Type<'gcc>) -> RValue<'gcc> { - if value.get_type() == self.bool_type.make_pointer() { - if let Some(pointee) = typ.get_pointee() { - if pointee.dyncast_vector().is_some() { - panic!() - } - } - } - // NOTE: since bitcast makes a value non-constant, don't bitcast if not necessary as some - // SIMD builtins require a constant value. - self.bitcast_if_needed(value, typ) - } -} - fn set_global_alignment<'gcc, 'tcx>(cx: &CodegenCx<'gcc, 'tcx>, gv: LValue<'gcc>, mut align: Align) { // The target may require greater alignment for globals than the type does. // Note: GCC and Clang also allow `__attribute__((aligned))` on variables, @@ -39,7 +24,7 @@ fn set_global_alignment<'gcc, 'tcx>(cx: &CodegenCx<'gcc, 'tcx>, gv: LValue<'gcc> match Align::from_bits(min) { Ok(min) => align = align.max(min), Err(err) => { - cx.sess().emit_err(InvalidMinimumAlignment { err }); + cx.sess().emit_err(InvalidMinimumAlignment { err: err.to_string() }); } } } @@ -249,7 +234,8 @@ impl<'gcc, 'tcx> CodegenCx<'gcc, 'tcx> { ); if !self.tcx.is_reachable_non_generic(def_id) { - // TODO(antoyo): set visibility. + #[cfg(feature = "master")] + global.add_attribute(VarAttribute::Visibility(Visibility::Hidden)); } global diff --git a/compiler/rustc_codegen_gcc/src/context.rs b/compiler/rustc_codegen_gcc/src/context.rs index 661681bdb50..dcebd92a61c 100644 --- a/compiler/rustc_codegen_gcc/src/context.rs +++ b/compiler/rustc_codegen_gcc/src/context.rs @@ -7,6 +7,7 @@ use rustc_codegen_ssa::traits::{ BaseTypeMethods, MiscMethods, }; +use rustc_codegen_ssa::errors as ssa_errors; use rustc_data_structures::base_n; use rustc_data_structures::fx::{FxHashMap, FxHashSet}; use rustc_middle::span_bug; @@ -476,10 +477,10 @@ impl<'gcc, 'tcx> LayoutOfHelpers<'tcx> for CodegenCx<'gcc, 'tcx> { #[inline] fn handle_layout_err(&self, err: LayoutError<'tcx>, span: Span, ty: Ty<'tcx>) -> ! { - if let LayoutError::SizeOverflow(_) = err { - self.sess().emit_fatal(respan(span, err)) + if let LayoutError::SizeOverflow(_) | LayoutError::ReferencesError(_) = err { + self.sess().emit_fatal(respan(span, err.into_diagnostic())) } else { - span_bug!(span, "failed to get layout for `{}`: {}", ty, err) + self.tcx.sess.emit_fatal(ssa_errors::FailedToGetLayout { span, ty, err }) } } } @@ -499,21 +500,12 @@ impl<'gcc, 'tcx> FnAbiOfHelpers<'tcx> for CodegenCx<'gcc, 'tcx> { } else { match fn_abi_request { FnAbiRequest::OfFnPtr { sig, extra_args } => { - span_bug!( - span, - "`fn_abi_of_fn_ptr({}, {:?})` failed: {}", - sig, - extra_args, - err - ); + span_bug!(span, "`fn_abi_of_fn_ptr({sig}, {extra_args:?})` failed: {err:?}"); } FnAbiRequest::OfInstance { instance, extra_args } => { span_bug!( span, - "`fn_abi_of_instance({}, {:?})` failed: {}", - instance, - extra_args, - err + "`fn_abi_of_instance({instance}, {extra_args:?})` failed: {err:?}" ); } } diff --git a/compiler/rustc_codegen_gcc/src/coverageinfo.rs b/compiler/rustc_codegen_gcc/src/coverageinfo.rs index 872fc2472e2..849e9886ef3 100644 --- a/compiler/rustc_codegen_gcc/src/coverageinfo.rs +++ b/compiler/rustc_codegen_gcc/src/coverageinfo.rs @@ -1,69 +1,11 @@ -use gccjit::RValue; -use rustc_codegen_ssa::traits::{CoverageInfoBuilderMethods, CoverageInfoMethods}; -use rustc_hir::def_id::DefId; -use rustc_middle::mir::coverage::{ - CodeRegion, - CounterValueReference, - ExpressionOperandId, - InjectedExpressionId, - Op, -}; +use rustc_codegen_ssa::traits::CoverageInfoBuilderMethods; +use rustc_middle::mir::Coverage; use rustc_middle::ty::Instance; use crate::builder::Builder; -use crate::context::CodegenCx; impl<'a, 'gcc, 'tcx> CoverageInfoBuilderMethods<'tcx> for Builder<'a, 'gcc, 'tcx> { - fn set_function_source_hash( - &mut self, - _instance: Instance<'tcx>, - _function_source_hash: u64, - ) -> bool { - unimplemented!(); - } - - fn add_coverage_counter(&mut self, _instance: Instance<'tcx>, _id: CounterValueReference, _region: CodeRegion) -> bool { - // TODO(antoyo) - false - } - - fn add_coverage_counter_expression(&mut self, _instance: Instance<'tcx>, _id: InjectedExpressionId, _lhs: ExpressionOperandId, _op: Op, _rhs: ExpressionOperandId, _region: Option<CodeRegion>) -> bool { - // TODO(antoyo) - false - } - - fn add_coverage_unreachable(&mut self, _instance: Instance<'tcx>, _region: CodeRegion) -> bool { + fn add_coverage(&mut self, _instance: Instance<'tcx>, _coverage: &Coverage) { // TODO(antoyo) - false - } -} - -impl<'gcc, 'tcx> CoverageInfoMethods<'tcx> for CodegenCx<'gcc, 'tcx> { - fn coverageinfo_finalize(&self) { - // TODO(antoyo) - } - - fn get_pgo_func_name_var(&self, _instance: Instance<'tcx>) -> RValue<'gcc> { - unimplemented!(); - } - - /// Functions with MIR-based coverage are normally codegenned _only_ if - /// called. LLVM coverage tools typically expect every function to be - /// defined (even if unused), with at least one call to LLVM intrinsic - /// `instrprof.increment`. - /// - /// Codegen a small function that will never be called, with one counter - /// that will never be incremented. - /// - /// For used/called functions, the coverageinfo was already added to the - /// `function_coverage_map` (keyed by function `Instance`) during codegen. - /// But in this case, since the unused function was _not_ previously - /// codegenned, collect the coverage `CodeRegion`s from the MIR and add - /// them. The first `CodeRegion` is used to add a single counter, with the - /// same counter ID used in the injected `instrprof.increment` intrinsic - /// call. Since the function is never called, all other `CodeRegion`s can be - /// added as `unreachable_region`s. - fn define_unused_fn(&self, _def_id: DefId) { - unimplemented!(); } } diff --git a/compiler/rustc_codegen_gcc/src/debuginfo.rs b/compiler/rustc_codegen_gcc/src/debuginfo.rs index a81585d4128..d1bfd833cd8 100644 --- a/compiler/rustc_codegen_gcc/src/debuginfo.rs +++ b/compiler/rustc_codegen_gcc/src/debuginfo.rs @@ -55,7 +55,7 @@ impl<'gcc, 'tcx> DebugInfoMethods<'tcx> for CodegenCx<'gcc, 'tcx> { _fn_abi: &FnAbi<'tcx, Ty<'tcx>>, _llfn: RValue<'gcc>, _mir: &mir::Body<'tcx>, - ) -> Option<FunctionDebugContext<Self::DIScope, Self::DILocation>> { + ) -> Option<FunctionDebugContext<'tcx, Self::DIScope, Self::DILocation>> { // TODO(antoyo) None } diff --git a/compiler/rustc_codegen_gcc/src/declare.rs b/compiler/rustc_codegen_gcc/src/declare.rs index 4748e7e4be2..e673d0af4c7 100644 --- a/compiler/rustc_codegen_gcc/src/declare.rs +++ b/compiler/rustc_codegen_gcc/src/declare.rs @@ -1,4 +1,6 @@ use gccjit::{Function, FunctionType, GlobalKind, LValue, RValue, Type}; +#[cfg(feature="master")] +use gccjit::{FnAttribute, ToRValue}; use rustc_codegen_ssa::traits::BaseTypeMethods; use rustc_middle::ty::Ty; use rustc_span::Symbol; @@ -114,6 +116,44 @@ fn declare_raw_fn<'gcc>(cx: &CodegenCx<'gcc, '_>, name: &str, _callconv: () /*ll .collect(); let func = cx.context.new_function(None, cx.linkage.get(), return_type, ¶ms, mangle_name(name), variadic); cx.functions.borrow_mut().insert(name.to_string(), func); + + #[cfg(feature="master")] + if name == "rust_eh_personality" { + // NOTE: GCC will sometimes change the personality function set on a function from + // rust_eh_personality to __gcc_personality_v0 as an optimization. + // As such, we need to create a weak alias from __gcc_personality_v0 to + // rust_eh_personality in order to avoid a linker error. + // This needs to be weak in order to still allow using the standard + // __gcc_personality_v0 when the linking to it. + // Since aliases don't work (maybe because of a bug in LTO partitioning?), we + // create a wrapper function that calls rust_eh_personality. + + let params: Vec<_> = param_types.into_iter().enumerate() + .map(|(index, param)| cx.context.new_parameter(None, *param, &format!("param{}", index))) // TODO(antoyo): set name. + .collect(); + let gcc_func = cx.context.new_function(None, FunctionType::Exported, return_type, ¶ms, "__gcc_personality_v0", variadic); + + // We need a normal extern function for the crates that access rust_eh_personality + // without defining it, otherwise we'll get a compiler error. + // + // For the crate defining it, that needs to be a weak alias instead. + gcc_func.add_attribute(FnAttribute::Weak); + + let block = gcc_func.new_block("start"); + let mut args = vec![]; + for param in ¶ms { + args.push(param.to_rvalue()); + } + let call = cx.context.new_call(None, func, &args); + if return_type == cx.type_void() { + block.add_eval(None, call); + block.end_with_void_return(None); + } + else { + block.end_with_return(None, call); + } + } + func }; @@ -132,7 +172,7 @@ fn declare_raw_fn<'gcc>(cx: &CodegenCx<'gcc, '_>, name: &str, _callconv: () /*ll pub fn mangle_name(name: &str) -> String { name.replace(|char: char| { if !char.is_alphanumeric() && char != '_' { - debug_assert!("$.".contains(char), "Unsupported char in function name: {}", char); + debug_assert!("$.*".contains(char), "Unsupported char in function name {}: {}", name, char); true } else { diff --git a/compiler/rustc_codegen_gcc/src/errors.rs b/compiler/rustc_codegen_gcc/src/errors.rs index 9305bd1e043..4bf3b71f503 100644 --- a/compiler/rustc_codegen_gcc/src/errors.rs +++ b/compiler/rustc_codegen_gcc/src/errors.rs @@ -1,238 +1,125 @@ -use rustc_errors::{DiagnosticArgValue, IntoDiagnosticArg}; -use rustc_macros::Diagnostic; -use rustc_middle::ty::Ty; -use rustc_span::{Span, Symbol}; +use rustc_errors::{ + DiagnosticArgValue, DiagnosticBuilder, ErrorGuaranteed, Handler, IntoDiagnostic, IntoDiagnosticArg, +}; +use rustc_macros::{Diagnostic, Subdiagnostic}; +use rustc_span::Span; use std::borrow::Cow; -struct ExitCode(Option<i32>); - -impl IntoDiagnosticArg for ExitCode { - fn into_diagnostic_arg(self) -> DiagnosticArgValue<'static> { - let ExitCode(exit_code) = self; - match exit_code { - Some(t) => t.into_diagnostic_arg(), - None => DiagnosticArgValue::Str(Cow::Borrowed("<signal>")), - } - } -} - -#[derive(Diagnostic)] -#[diag(codegen_gcc_invalid_monomorphization_basic_integer, code = "E0511")] -pub(crate) struct InvalidMonomorphizationBasicInteger<'a> { - #[primary_span] - pub span: Span, - pub name: Symbol, - pub ty: Ty<'a>, -} - -#[derive(Diagnostic)] -#[diag(codegen_gcc_invalid_monomorphization_invalid_float_vector, code = "E0511")] -pub(crate) struct InvalidMonomorphizationInvalidFloatVector<'a> { - #[primary_span] - pub span: Span, - pub name: Symbol, - pub elem_ty: &'a str, - pub vec_ty: Ty<'a>, -} - -#[derive(Diagnostic)] -#[diag(codegen_gcc_invalid_monomorphization_not_float, code = "E0511")] -pub(crate) struct InvalidMonomorphizationNotFloat<'a> { - #[primary_span] - pub span: Span, - pub name: Symbol, - pub ty: Ty<'a>, -} - -#[derive(Diagnostic)] -#[diag(codegen_gcc_invalid_monomorphization_unrecognized, code = "E0511")] -pub(crate) struct InvalidMonomorphizationUnrecognized { - #[primary_span] - pub span: Span, - pub name: Symbol, -} +use crate::fluent_generated as fluent; #[derive(Diagnostic)] -#[diag(codegen_gcc_invalid_monomorphization_expected_signed_unsigned, code = "E0511")] -pub(crate) struct InvalidMonomorphizationExpectedSignedUnsigned<'a> { - #[primary_span] - pub span: Span, - pub name: Symbol, - pub elem_ty: Ty<'a>, - pub vec_ty: Ty<'a>, +#[diag(codegen_gcc_unknown_ctarget_feature_prefix)] +#[note] +pub(crate) struct UnknownCTargetFeaturePrefix<'a> { + pub feature: &'a str, } #[derive(Diagnostic)] -#[diag(codegen_gcc_invalid_monomorphization_unsupported_element, code = "E0511")] -pub(crate) struct InvalidMonomorphizationUnsupportedElement<'a> { - #[primary_span] - pub span: Span, - pub name: Symbol, - pub in_ty: Ty<'a>, - pub elem_ty: Ty<'a>, - pub ret_ty: Ty<'a>, +#[diag(codegen_gcc_unknown_ctarget_feature)] +#[note] +pub(crate) struct UnknownCTargetFeature<'a> { + pub feature: &'a str, + #[subdiagnostic] + pub rust_feature: PossibleFeature<'a>, } -#[derive(Diagnostic)] -#[diag(codegen_gcc_invalid_monomorphization_invalid_bitmask, code = "E0511")] -pub(crate) struct InvalidMonomorphizationInvalidBitmask<'a> { - #[primary_span] - pub span: Span, - pub name: Symbol, - pub ty: Ty<'a>, - pub expected_int_bits: u64, - pub expected_bytes: u64, +#[derive(Subdiagnostic)] +pub(crate) enum PossibleFeature<'a> { + #[help(codegen_gcc_possible_feature)] + Some { rust_feature: &'a str }, + #[help(codegen_gcc_consider_filing_feature_request)] + None, } -#[derive(Diagnostic)] -#[diag(codegen_gcc_invalid_monomorphization_simd_shuffle, code = "E0511")] -pub(crate) struct InvalidMonomorphizationSimdShuffle<'a> { - #[primary_span] - pub span: Span, - pub name: Symbol, - pub ty: Ty<'a>, -} +struct ExitCode(Option<i32>); -#[derive(Diagnostic)] -#[diag(codegen_gcc_invalid_monomorphization_expected_simd, code = "E0511")] -pub(crate) struct InvalidMonomorphizationExpectedSimd<'a> { - #[primary_span] - pub span: Span, - pub name: Symbol, - pub position: &'a str, - pub found_ty: Ty<'a>, +impl IntoDiagnosticArg for ExitCode { + fn into_diagnostic_arg(self) -> DiagnosticArgValue<'static> { + let ExitCode(exit_code) = self; + match exit_code { + Some(t) => t.into_diagnostic_arg(), + None => DiagnosticArgValue::Str(Cow::Borrowed("<signal>")), + } + } } #[derive(Diagnostic)] -#[diag(codegen_gcc_invalid_monomorphization_mask_type, code = "E0511")] -pub(crate) struct InvalidMonomorphizationMaskType<'a> { - #[primary_span] - pub span: Span, - pub name: Symbol, - pub ty: Ty<'a>, -} +#[diag(codegen_gcc_lto_not_supported)] +pub(crate) struct LTONotSupported; #[derive(Diagnostic)] -#[diag(codegen_gcc_invalid_monomorphization_return_length, code = "E0511")] -pub(crate) struct InvalidMonomorphizationReturnLength<'a> { +#[diag(codegen_gcc_unwinding_inline_asm)] +pub(crate) struct UnwindingInlineAsm { #[primary_span] pub span: Span, - pub name: Symbol, - pub in_len: u64, - pub ret_ty: Ty<'a>, - pub out_len: u64, } #[derive(Diagnostic)] -#[diag(codegen_gcc_invalid_monomorphization_return_length_input_type, code = "E0511")] -pub(crate) struct InvalidMonomorphizationReturnLengthInputType<'a> { - #[primary_span] - pub span: Span, - pub name: Symbol, - pub in_len: u64, - pub in_ty: Ty<'a>, - pub ret_ty: Ty<'a>, - pub out_len: u64, +#[diag(codegen_gcc_invalid_minimum_alignment)] +pub(crate) struct InvalidMinimumAlignment { + pub err: String, } #[derive(Diagnostic)] -#[diag(codegen_gcc_invalid_monomorphization_return_element, code = "E0511")] -pub(crate) struct InvalidMonomorphizationReturnElement<'a> { +#[diag(codegen_gcc_tied_target_features)] +#[help] +pub(crate) struct TiedTargetFeatures { #[primary_span] pub span: Span, - pub name: Symbol, - pub in_elem: Ty<'a>, - pub in_ty: Ty<'a>, - pub ret_ty: Ty<'a>, - pub out_ty: Ty<'a>, + pub features: String, } #[derive(Diagnostic)] -#[diag(codegen_gcc_invalid_monomorphization_return_type, code = "E0511")] -pub(crate) struct InvalidMonomorphizationReturnType<'a> { - #[primary_span] - pub span: Span, - pub name: Symbol, - pub in_elem: Ty<'a>, - pub in_ty: Ty<'a>, - pub ret_ty: Ty<'a>, +#[diag(codegen_gcc_copy_bitcode)] +pub(crate) struct CopyBitcode { + pub err: std::io::Error, } #[derive(Diagnostic)] -#[diag(codegen_gcc_invalid_monomorphization_inserted_type, code = "E0511")] -pub(crate) struct InvalidMonomorphizationInsertedType<'a> { - #[primary_span] - pub span: Span, - pub name: Symbol, - pub in_elem: Ty<'a>, - pub in_ty: Ty<'a>, - pub out_ty: Ty<'a>, -} +#[diag(codegen_gcc_dynamic_linking_with_lto)] +#[note] +pub(crate) struct DynamicLinkingWithLTO; #[derive(Diagnostic)] -#[diag(codegen_gcc_invalid_monomorphization_return_integer_type, code = "E0511")] -pub(crate) struct InvalidMonomorphizationReturnIntegerType<'a> { - #[primary_span] - pub span: Span, - pub name: Symbol, - pub ret_ty: Ty<'a>, - pub out_ty: Ty<'a>, +#[diag(codegen_gcc_load_bitcode)] +pub(crate) struct LoadBitcode { + name: String, } #[derive(Diagnostic)] -#[diag(codegen_gcc_invalid_monomorphization_mismatched_lengths, code = "E0511")] -pub(crate) struct InvalidMonomorphizationMismatchedLengths { - #[primary_span] - pub span: Span, - pub name: Symbol, - pub m_len: u64, - pub v_len: u64, -} +#[diag(codegen_gcc_lto_disallowed)] +pub(crate) struct LtoDisallowed; #[derive(Diagnostic)] -#[diag(codegen_gcc_invalid_monomorphization_unsupported_cast, code = "E0511")] -pub(crate) struct InvalidMonomorphizationUnsupportedCast<'a> { - #[primary_span] - pub span: Span, - pub name: Symbol, - pub in_ty: Ty<'a>, - pub in_elem: Ty<'a>, - pub ret_ty: Ty<'a>, - pub out_elem: Ty<'a>, -} +#[diag(codegen_gcc_lto_dylib)] +pub(crate) struct LtoDylib; #[derive(Diagnostic)] -#[diag(codegen_gcc_invalid_monomorphization_unsupported_operation, code = "E0511")] -pub(crate) struct InvalidMonomorphizationUnsupportedOperation<'a> { - #[primary_span] - pub span: Span, - pub name: Symbol, - pub in_ty: Ty<'a>, - pub in_elem: Ty<'a>, +#[diag(codegen_gcc_lto_bitcode_from_rlib)] +pub(crate) struct LtoBitcodeFromRlib { + pub gcc_err: String, } -#[derive(Diagnostic)] -#[diag(codegen_gcc_lto_not_supported)] -pub(crate) struct LTONotSupported; - -#[derive(Diagnostic)] -#[diag(codegen_gcc_unwinding_inline_asm)] -pub(crate) struct UnwindingInlineAsm { - #[primary_span] - pub span: Span, +pub(crate) struct TargetFeatureDisableOrEnable<'a> { + pub features: &'a [&'a str], + pub span: Option<Span>, + pub missing_features: Option<MissingFeatures>, } -#[derive(Diagnostic)] -#[diag(codegen_gcc_invalid_minimum_alignment)] -pub(crate) struct InvalidMinimumAlignment { - pub err: String, -} +#[derive(Subdiagnostic)] +#[help(codegen_gcc_missing_features)] +pub(crate) struct MissingFeatures; -#[derive(Diagnostic)] -#[diag(codegen_gcc_tied_target_features)] -#[help] -pub(crate) struct TiedTargetFeatures { - #[primary_span] - pub span: Span, - pub features: String, +impl IntoDiagnostic<'_, ErrorGuaranteed> for TargetFeatureDisableOrEnable<'_> { + fn into_diagnostic(self, sess: &'_ Handler) -> DiagnosticBuilder<'_, ErrorGuaranteed> { + let mut diag = sess.struct_err(fluent::codegen_gcc_target_feature_disable_or_enable); + if let Some(span) = self.span { + diag.set_span(span); + }; + if let Some(missing_features) = self.missing_features { + diag.subdiagnostic(missing_features); + } + diag.set_arg("features", self.features.join(", ")); + diag + } } diff --git a/compiler/rustc_codegen_gcc/src/gcc_util.rs b/compiler/rustc_codegen_gcc/src/gcc_util.rs new file mode 100644 index 00000000000..0514c9988e0 --- /dev/null +++ b/compiler/rustc_codegen_gcc/src/gcc_util.rs @@ -0,0 +1,223 @@ +#[cfg(feature="master")] +use gccjit::Context; +use smallvec::{smallvec, SmallVec}; + +use rustc_codegen_ssa::target_features::{ + supported_target_features, tied_target_features, RUSTC_SPECIFIC_FEATURES, +}; +use rustc_data_structures::fx::FxHashMap; +use rustc_middle::bug; +use rustc_session::Session; + +use crate::errors::{PossibleFeature, TargetFeatureDisableOrEnable, UnknownCTargetFeature, UnknownCTargetFeaturePrefix}; + +/// The list of GCC features computed from CLI flags (`-Ctarget-cpu`, `-Ctarget-feature`, +/// `--target` and similar). +pub(crate) fn global_gcc_features(sess: &Session, diagnostics: bool) -> Vec<String> { + // Features that come earlier are overridden by conflicting features later in the string. + // Typically we'll want more explicit settings to override the implicit ones, so: + // + // * Features from -Ctarget-cpu=*; are overridden by [^1] + // * Features implied by --target; are overridden by + // * Features from -Ctarget-feature; are overridden by + // * function specific features. + // + // [^1]: target-cpu=native is handled here, other target-cpu values are handled implicitly + // through GCC march implementation. + // + // FIXME(nagisa): it isn't clear what's the best interaction between features implied by + // `-Ctarget-cpu` and `--target` are. On one hand, you'd expect CLI arguments to always + // override anything that's implicit, so e.g. when there's no `--target` flag, features implied + // the host target are overridden by `-Ctarget-cpu=*`. On the other hand, what about when both + // `--target` and `-Ctarget-cpu=*` are specified? Both then imply some target features and both + // flags are specified by the user on the CLI. It isn't as clear-cut which order of precedence + // should be taken in cases like these. + let mut features = vec![]; + + // Features implied by an implicit or explicit `--target`. + features.extend( + sess.target + .features + .split(',') + .filter(|v| !v.is_empty() && backend_feature_name(v).is_some()) + .map(String::from), + ); + + // -Ctarget-features + let supported_features = supported_target_features(sess); + let mut featsmap = FxHashMap::default(); + let feats = sess.opts.cg.target_feature + .split(',') + .filter_map(|s| { + let enable_disable = match s.chars().next() { + None => return None, + Some(c @ ('+' | '-')) => c, + Some(_) => { + if diagnostics { + sess.emit_warning(UnknownCTargetFeaturePrefix { feature: s }); + } + return None; + } + }; + + let feature = backend_feature_name(s)?; + // Warn against use of GCC specific feature names on the CLI. + if diagnostics && !supported_features.iter().any(|&(v, _)| v == feature) { + let rust_feature = supported_features.iter().find_map(|&(rust_feature, _)| { + let gcc_features = to_gcc_features(sess, rust_feature); + if gcc_features.contains(&feature) && !gcc_features.contains(&rust_feature) { + Some(rust_feature) + } else { + None + } + }); + let unknown_feature = + if let Some(rust_feature) = rust_feature { + UnknownCTargetFeature { + feature, + rust_feature: PossibleFeature::Some { rust_feature }, + } + } + else { + UnknownCTargetFeature { feature, rust_feature: PossibleFeature::None } + }; + sess.emit_warning(unknown_feature); + } + + if diagnostics { + // FIXME(nagisa): figure out how to not allocate a full hashset here. + featsmap.insert(feature, enable_disable == '+'); + } + + // rustc-specific features do not get passed down to GCC… + if RUSTC_SPECIFIC_FEATURES.contains(&feature) { + return None; + } + // ... otherwise though we run through `to_gcc_features` when + // passing requests down to GCC. This means that all in-language + // features also work on the command line instead of having two + // different names when the GCC name and the Rust name differ. + Some(to_gcc_features(sess, feature) + .iter() + .flat_map(|feat| to_gcc_features(sess, feat).into_iter()) + .map(|feature| { + if enable_disable == '-' { + format!("-{}", feature) + } + else { + feature.to_string() + } + }) + .collect::<Vec<_>>(), + ) + }) + .flatten(); + features.extend(feats); + + if diagnostics { + if let Some(f) = check_tied_features(sess, &featsmap) { + sess.emit_err(TargetFeatureDisableOrEnable { + features: f, + span: None, + missing_features: None, + }); + } + } + + features +} + +/// Returns a feature name for the given `+feature` or `-feature` string. +/// +/// Only allows features that are backend specific (i.e. not [`RUSTC_SPECIFIC_FEATURES`].) +fn backend_feature_name(s: &str) -> Option<&str> { + // features must start with a `+` or `-`. + let feature = s.strip_prefix(&['+', '-'][..]).unwrap_or_else(|| { + bug!("target feature `{}` must begin with a `+` or `-`", s); + }); + // Rustc-specific feature requests like `+crt-static` or `-crt-static` + // are not passed down to GCC. + if RUSTC_SPECIFIC_FEATURES.contains(&feature) { + return None; + } + Some(feature) +} + +// To find a list of GCC's names, check https://gcc.gnu.org/onlinedocs/gcc/Function-Attributes.html +pub fn to_gcc_features<'a>(sess: &Session, s: &'a str) -> SmallVec<[&'a str; 2]> { + let arch = if sess.target.arch == "x86_64" { "x86" } else { &*sess.target.arch }; + match (arch, s) { + ("x86", "sse4.2") => smallvec!["sse4.2", "crc32"], + ("x86", "pclmulqdq") => smallvec!["pclmul"], + ("x86", "rdrand") => smallvec!["rdrnd"], + ("x86", "bmi1") => smallvec!["bmi"], + ("x86", "cmpxchg16b") => smallvec!["cx16"], + ("x86", "avx512vaes") => smallvec!["vaes"], + ("x86", "avx512gfni") => smallvec!["gfni"], + ("x86", "avx512vpclmulqdq") => smallvec!["vpclmulqdq"], + // NOTE: seems like GCC requires 'avx512bw' for 'avx512vbmi2'. + ("x86", "avx512vbmi2") => smallvec!["avx512vbmi2", "avx512bw"], + // NOTE: seems like GCC requires 'avx512bw' for 'avx512bitalg'. + ("x86", "avx512bitalg") => smallvec!["avx512bitalg", "avx512bw"], + ("aarch64", "rcpc2") => smallvec!["rcpc-immo"], + ("aarch64", "dpb") => smallvec!["ccpp"], + ("aarch64", "dpb2") => smallvec!["ccdp"], + ("aarch64", "frintts") => smallvec!["fptoint"], + ("aarch64", "fcma") => smallvec!["complxnum"], + ("aarch64", "pmuv3") => smallvec!["perfmon"], + ("aarch64", "paca") => smallvec!["pauth"], + ("aarch64", "pacg") => smallvec!["pauth"], + // Rust ties fp and neon together. In GCC neon implicitly enables fp, + // but we manually enable neon when a feature only implicitly enables fp + ("aarch64", "f32mm") => smallvec!["f32mm", "neon"], + ("aarch64", "f64mm") => smallvec!["f64mm", "neon"], + ("aarch64", "fhm") => smallvec!["fp16fml", "neon"], + ("aarch64", "fp16") => smallvec!["fullfp16", "neon"], + ("aarch64", "jsconv") => smallvec!["jsconv", "neon"], + ("aarch64", "sve") => smallvec!["sve", "neon"], + ("aarch64", "sve2") => smallvec!["sve2", "neon"], + ("aarch64", "sve2-aes") => smallvec!["sve2-aes", "neon"], + ("aarch64", "sve2-sm4") => smallvec!["sve2-sm4", "neon"], + ("aarch64", "sve2-sha3") => smallvec!["sve2-sha3", "neon"], + ("aarch64", "sve2-bitperm") => smallvec!["sve2-bitperm", "neon"], + (_, s) => smallvec![s], + } +} + +// Given a map from target_features to whether they are enabled or disabled, +// ensure only valid combinations are allowed. +pub fn check_tied_features(sess: &Session, features: &FxHashMap<&str, bool>) -> Option<&'static [&'static str]> { + for tied in tied_target_features(sess) { + // Tied features must be set to the same value, or not set at all + let mut tied_iter = tied.iter(); + let enabled = features.get(tied_iter.next().unwrap()); + if tied_iter.any(|feature| enabled != features.get(feature)) { + return Some(tied); + } + } + None +} + +fn handle_native(name: &str) -> &str { + if name != "native" { + return name; + } + + #[cfg(feature="master")] + { + // Get the native arch. + let context = Context::default(); + context.get_target_info().arch().unwrap() + .to_str() + .unwrap() + } + #[cfg(not(feature="master"))] + unimplemented!(); +} + +pub fn target_cpu(sess: &Session) -> &str { + match sess.opts.cg.target_cpu { + Some(ref name) => handle_native(name), + None => handle_native(sess.target.cpu.as_ref()), + } +} diff --git a/compiler/rustc_codegen_gcc/src/int.rs b/compiler/rustc_codegen_gcc/src/int.rs index 0cf1204791d..58e0dd56f38 100644 --- a/compiler/rustc_codegen_gcc/src/int.rs +++ b/compiler/rustc_codegen_gcc/src/int.rs @@ -36,7 +36,6 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> { self.cx.context.new_unary_op(None, operation, typ, a) } else { - // TODO(antoyo): use __negdi2 and __negti2 instead? let element_type = typ.dyncast_array().expect("element type"); let values = [ self.cx.context.new_unary_op(None, UnaryOp::BitwiseNegate, element_type, self.low(a)), @@ -52,9 +51,7 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> { self.cx.context.new_unary_op(None, UnaryOp::Minus, a.get_type(), a) } else { - let param_a = self.context.new_parameter(None, a_type, "a"); - let func = self.context.new_function(None, FunctionType::Extern, a_type, &[param_a], "__negti2", false); - self.context.new_call(None, func, &[a]) + self.gcc_add(self.gcc_not(a), self.gcc_int(a_type, 1)) } } @@ -353,23 +350,63 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> { (res.dereference(None).to_rvalue(), overflow) } - pub fn gcc_icmp(&self, op: IntPredicate, mut lhs: RValue<'gcc>, mut rhs: RValue<'gcc>) -> RValue<'gcc> { + pub fn gcc_icmp(&mut self, op: IntPredicate, mut lhs: RValue<'gcc>, mut rhs: RValue<'gcc>) -> RValue<'gcc> { let a_type = lhs.get_type(); let b_type = rhs.get_type(); if self.is_non_native_int_type(a_type) || self.is_non_native_int_type(b_type) { - let signed = a_type.is_compatible_with(self.i128_type); - let sign = - if signed { - "" - } - else { - "u" - }; - let func_name = format!("__{}cmpti2", sign); - let param_a = self.context.new_parameter(None, a_type, "a"); - let param_b = self.context.new_parameter(None, b_type, "b"); - let func = self.context.new_function(None, FunctionType::Extern, self.int_type, &[param_a, param_b], func_name, false); - let cmp = self.context.new_call(None, func, &[lhs, rhs]); + // This algorithm is based on compiler-rt's __cmpti2: + // https://github.com/llvm-mirror/compiler-rt/blob/f0745e8476f069296a7c71accedd061dce4cdf79/lib/builtins/cmpti2.c#L21 + let result = self.current_func().new_local(None, self.int_type, "icmp_result"); + let block1 = self.current_func().new_block("block1"); + let block2 = self.current_func().new_block("block2"); + let block3 = self.current_func().new_block("block3"); + let block4 = self.current_func().new_block("block4"); + let block5 = self.current_func().new_block("block5"); + let block6 = self.current_func().new_block("block6"); + let block7 = self.current_func().new_block("block7"); + let block8 = self.current_func().new_block("block8"); + let after = self.current_func().new_block("after"); + + let native_int_type = a_type.dyncast_array().expect("get element type"); + // NOTE: cast low to its unsigned type in order to perform a comparison correctly (e.g. + // the sign is only on high). + let unsigned_type = native_int_type.to_unsigned(&self.cx); + + let lhs_low = self.context.new_cast(None, self.low(lhs), unsigned_type); + let rhs_low = self.context.new_cast(None, self.low(rhs), unsigned_type); + + let condition = self.context.new_comparison(None, ComparisonOp::LessThan, self.high(lhs), self.high(rhs)); + self.llbb().end_with_conditional(None, condition, block1, block2); + + block1.add_assignment(None, result, self.context.new_rvalue_zero(self.int_type)); + block1.end_with_jump(None, after); + + let condition = self.context.new_comparison(None, ComparisonOp::GreaterThan, self.high(lhs), self.high(rhs)); + block2.end_with_conditional(None, condition, block3, block4); + + block3.add_assignment(None, result, self.context.new_rvalue_from_int(self.int_type, 2)); + block3.end_with_jump(None, after); + + let condition = self.context.new_comparison(None, ComparisonOp::LessThan, lhs_low, rhs_low); + block4.end_with_conditional(None, condition, block5, block6); + + block5.add_assignment(None, result, self.context.new_rvalue_zero(self.int_type)); + block5.end_with_jump(None, after); + + let condition = self.context.new_comparison(None, ComparisonOp::GreaterThan, lhs_low, rhs_low); + block6.end_with_conditional(None, condition, block7, block8); + + block7.add_assignment(None, result, self.context.new_rvalue_from_int(self.int_type, 2)); + block7.end_with_jump(None, after); + + block8.add_assignment(None, result, self.context.new_rvalue_one(self.int_type)); + block8.end_with_jump(None, after); + + // NOTE: since jumps were added in a place rustc does not expect, the current block in the + // state need to be updated. + self.switch_to_block(after); + + let cmp = result.to_rvalue(); let (op, limit) = match op { IntPredicate::IntEQ => { @@ -546,7 +583,12 @@ impl<'gcc, 'tcx> CodegenCx<'gcc, 'tcx> { } pub fn gcc_uint(&self, typ: Type<'gcc>, int: u64) -> RValue<'gcc> { - if self.is_native_int_type_or_bool(typ) { + if typ.is_u128(self) { + // FIXME(antoyo): libgccjit cannot create 128-bit values yet. + let num = self.context.new_rvalue_from_long(self.u64_type, int as i64); + self.gcc_int_cast(num, typ) + } + else if self.is_native_int_type_or_bool(typ) { self.context.new_rvalue_from_long(typ, u64::try_from(int).expect("u64::try_from") as i64) } else { @@ -572,6 +614,7 @@ impl<'gcc, 'tcx> CodegenCx<'gcc, 'tcx> { } } else if typ.is_i128(self) { + // FIXME(antoyo): libgccjit cannot create 128-bit values yet. let num = self.context.new_rvalue_from_long(self.u64_type, num as u64 as i64); self.gcc_int_cast(num, typ) } diff --git a/compiler/rustc_codegen_gcc/src/intrinsic/archs.rs b/compiler/rustc_codegen_gcc/src/intrinsic/archs.rs index 8a4559355ea..e01299d32fd 100644 --- a/compiler/rustc_codegen_gcc/src/intrinsic/archs.rs +++ b/compiler/rustc_codegen_gcc/src/intrinsic/archs.rs @@ -2254,6 +2254,42 @@ match name { "llvm.hexagon.prefetch" => "__builtin_HEXAGON_prefetch", "llvm.hexagon.vmemcpy" => "__builtin_hexagon_vmemcpy", "llvm.hexagon.vmemset" => "__builtin_hexagon_vmemset", + // loongarch + "llvm.loongarch.asrtgt.d" => "__builtin_loongarch_asrtgt_d", + "llvm.loongarch.asrtle.d" => "__builtin_loongarch_asrtle_d", + "llvm.loongarch.break" => "__builtin_loongarch_break", + "llvm.loongarch.cacop.d" => "__builtin_loongarch_cacop_d", + "llvm.loongarch.cacop.w" => "__builtin_loongarch_cacop_w", + "llvm.loongarch.cpucfg" => "__builtin_loongarch_cpucfg", + "llvm.loongarch.crc.w.b.w" => "__builtin_loongarch_crc_w_b_w", + "llvm.loongarch.crc.w.d.w" => "__builtin_loongarch_crc_w_d_w", + "llvm.loongarch.crc.w.h.w" => "__builtin_loongarch_crc_w_h_w", + "llvm.loongarch.crc.w.w.w" => "__builtin_loongarch_crc_w_w_w", + "llvm.loongarch.crcc.w.b.w" => "__builtin_loongarch_crcc_w_b_w", + "llvm.loongarch.crcc.w.d.w" => "__builtin_loongarch_crcc_w_d_w", + "llvm.loongarch.crcc.w.h.w" => "__builtin_loongarch_crcc_w_h_w", + "llvm.loongarch.crcc.w.w.w" => "__builtin_loongarch_crcc_w_w_w", + "llvm.loongarch.csrrd.d" => "__builtin_loongarch_csrrd_d", + "llvm.loongarch.csrrd.w" => "__builtin_loongarch_csrrd_w", + "llvm.loongarch.csrwr.d" => "__builtin_loongarch_csrwr_d", + "llvm.loongarch.csrwr.w" => "__builtin_loongarch_csrwr_w", + "llvm.loongarch.csrxchg.d" => "__builtin_loongarch_csrxchg_d", + "llvm.loongarch.csrxchg.w" => "__builtin_loongarch_csrxchg_w", + "llvm.loongarch.dbar" => "__builtin_loongarch_dbar", + "llvm.loongarch.ibar" => "__builtin_loongarch_ibar", + "llvm.loongarch.iocsrrd.b" => "__builtin_loongarch_iocsrrd_b", + "llvm.loongarch.iocsrrd.d" => "__builtin_loongarch_iocsrrd_d", + "llvm.loongarch.iocsrrd.h" => "__builtin_loongarch_iocsrrd_h", + "llvm.loongarch.iocsrrd.w" => "__builtin_loongarch_iocsrrd_w", + "llvm.loongarch.iocsrwr.b" => "__builtin_loongarch_iocsrwr_b", + "llvm.loongarch.iocsrwr.d" => "__builtin_loongarch_iocsrwr_d", + "llvm.loongarch.iocsrwr.h" => "__builtin_loongarch_iocsrwr_h", + "llvm.loongarch.iocsrwr.w" => "__builtin_loongarch_iocsrwr_w", + "llvm.loongarch.lddir.d" => "__builtin_loongarch_lddir_d", + "llvm.loongarch.ldpte.d" => "__builtin_loongarch_ldpte_d", + "llvm.loongarch.movfcsr2gr" => "__builtin_loongarch_movfcsr2gr", + "llvm.loongarch.movgr2fcsr" => "__builtin_loongarch_movgr2fcsr", + "llvm.loongarch.syscall" => "__builtin_loongarch_syscall", // mips "llvm.mips.absq.s.ph" => "__builtin_mips_absq_s_ph", "llvm.mips.absq.s.qb" => "__builtin_mips_absq_s_qb", @@ -2954,6 +2990,8 @@ match name { "llvm.nvvm.barrier0.and" => "__nvvm_bar0_and", "llvm.nvvm.barrier0.or" => "__nvvm_bar0_or", "llvm.nvvm.barrier0.popc" => "__nvvm_bar0_popc", + "llvm.nvvm.bf2h.rn" => "__nvvm_bf2h_rn", + "llvm.nvvm.bf2h.rn.ftz" => "__nvvm_bf2h_rn_ftz", "llvm.nvvm.bitcast.d2ll" => "__nvvm_bitcast_d2ll", "llvm.nvvm.bitcast.f2i" => "__nvvm_bitcast_f2i", "llvm.nvvm.bitcast.i2f" => "__nvvm_bitcast_i2f", @@ -2967,10 +3005,6 @@ match name { "llvm.nvvm.clz.ll" => "__nvvm_clz_ll", "llvm.nvvm.cos.approx.f" => "__nvvm_cos_approx_f", "llvm.nvvm.cos.approx.ftz.f" => "__nvvm_cos_approx_ftz_f", - "llvm.nvvm.cp.async.ca.shared.global.16" => "__nvvm_cp_async_ca_shared_global_16", - "llvm.nvvm.cp.async.ca.shared.global.4" => "__nvvm_cp_async_ca_shared_global_4", - "llvm.nvvm.cp.async.ca.shared.global.8" => "__nvvm_cp_async_ca_shared_global_8", - "llvm.nvvm.cp.async.cg.shared.global.16" => "__nvvm_cp_async_cg_shared_global_16", "llvm.nvvm.cp.async.commit.group" => "__nvvm_cp_async_commit_group", "llvm.nvvm.cp.async.mbarrier.arrive" => "__nvvm_cp_async_mbarrier_arrive", "llvm.nvvm.cp.async.mbarrier.arrive.noinc" => "__nvvm_cp_async_mbarrier_arrive_noinc", @@ -3020,8 +3054,6 @@ match name { "llvm.nvvm.div.rz.ftz.f" => "__nvvm_div_rz_ftz_f", "llvm.nvvm.ex2.approx.d" => "__nvvm_ex2_approx_d", "llvm.nvvm.ex2.approx.f" => "__nvvm_ex2_approx_f", - "llvm.nvvm.ex2.approx.f16" => "__nvvm_ex2_approx_f16", - "llvm.nvvm.ex2.approx.f16x2" => "__nvvm_ex2_approx_f16x2", "llvm.nvvm.ex2.approx.ftz.f" => "__nvvm_ex2_approx_ftz_f", "llvm.nvvm.f2bf16.rn" => "__nvvm_f2bf16_rn", "llvm.nvvm.f2bf16.rn.relu" => "__nvvm_f2bf16_rn_relu", @@ -3083,21 +3115,17 @@ match name { "llvm.nvvm.fma.rn.bf16x2" => "__nvvm_fma_rn_bf16x2", "llvm.nvvm.fma.rn.d" => "__nvvm_fma_rn_d", "llvm.nvvm.fma.rn.f" => "__nvvm_fma_rn_f", - "llvm.nvvm.fma.rn.f16" => "__nvvm_fma_rn_f16", - "llvm.nvvm.fma.rn.f16x2" => "__nvvm_fma_rn_f16x2", + "llvm.nvvm.fma.rn.ftz.bf16" => "__nvvm_fma_rn_ftz_bf16", + "llvm.nvvm.fma.rn.ftz.bf16x2" => "__nvvm_fma_rn_ftz_bf16x2", "llvm.nvvm.fma.rn.ftz.f" => "__nvvm_fma_rn_ftz_f", - "llvm.nvvm.fma.rn.ftz.f16" => "__nvvm_fma_rn_ftz_f16", - "llvm.nvvm.fma.rn.ftz.f16x2" => "__nvvm_fma_rn_ftz_f16x2", - "llvm.nvvm.fma.rn.ftz.relu.f16" => "__nvvm_fma_rn_ftz_relu_f16", - "llvm.nvvm.fma.rn.ftz.relu.f16x2" => "__nvvm_fma_rn_ftz_relu_f16x2", - "llvm.nvvm.fma.rn.ftz.sat.f16" => "__nvvm_fma_rn_ftz_sat_f16", - "llvm.nvvm.fma.rn.ftz.sat.f16x2" => "__nvvm_fma_rn_ftz_sat_f16x2", + "llvm.nvvm.fma.rn.ftz.relu.bf16" => "__nvvm_fma_rn_ftz_relu_bf16", + "llvm.nvvm.fma.rn.ftz.relu.bf16x2" => "__nvvm_fma_rn_ftz_relu_bf16x2", + "llvm.nvvm.fma.rn.ftz.sat.bf16" => "__nvvm_fma_rn_ftz_sat_bf16", + "llvm.nvvm.fma.rn.ftz.sat.bf16x2" => "__nvvm_fma_rn_ftz_sat_bf16x2", "llvm.nvvm.fma.rn.relu.bf16" => "__nvvm_fma_rn_relu_bf16", "llvm.nvvm.fma.rn.relu.bf16x2" => "__nvvm_fma_rn_relu_bf16x2", - "llvm.nvvm.fma.rn.relu.f16" => "__nvvm_fma_rn_relu_f16", - "llvm.nvvm.fma.rn.relu.f16x2" => "__nvvm_fma_rn_relu_f16x2", - "llvm.nvvm.fma.rn.sat.f16" => "__nvvm_fma_rn_sat_f16", - "llvm.nvvm.fma.rn.sat.f16x2" => "__nvvm_fma_rn_sat_f16x2", + "llvm.nvvm.fma.rn.sat.bf16" => "__nvvm_fma_rn_sat_bf16", + "llvm.nvvm.fma.rn.sat.bf16x2" => "__nvvm_fma_rn_sat_bf16x2", "llvm.nvvm.fma.rp.d" => "__nvvm_fma_rp_d", "llvm.nvvm.fma.rp.f" => "__nvvm_fma_rp_f", "llvm.nvvm.fma.rp.ftz.f" => "__nvvm_fma_rp_ftz_f", @@ -3108,68 +3136,52 @@ match name { "llvm.nvvm.fmax.bf16x2" => "__nvvm_fmax_bf16x2", "llvm.nvvm.fmax.d" => "__nvvm_fmax_d", "llvm.nvvm.fmax.f" => "__nvvm_fmax_f", - "llvm.nvvm.fmax.f16" => "__nvvm_fmax_f16", - "llvm.nvvm.fmax.f16x2" => "__nvvm_fmax_f16x2", + "llvm.nvvm.fmax.ftz.bf16" => "__nvvm_fmax_ftz_bf16", + "llvm.nvvm.fmax.ftz.bf16x2" => "__nvvm_fmax_ftz_bf16x2", "llvm.nvvm.fmax.ftz.f" => "__nvvm_fmax_ftz_f", - "llvm.nvvm.fmax.ftz.f16" => "__nvvm_fmax_ftz_f16", - "llvm.nvvm.fmax.ftz.f16x2" => "__nvvm_fmax_ftz_f16x2", + "llvm.nvvm.fmax.ftz.nan.bf16" => "__nvvm_fmax_ftz_nan_bf16", + "llvm.nvvm.fmax.ftz.nan.bf16x2" => "__nvvm_fmax_ftz_nan_bf16x2", "llvm.nvvm.fmax.ftz.nan.f" => "__nvvm_fmax_ftz_nan_f", - "llvm.nvvm.fmax.ftz.nan.f16" => "__nvvm_fmax_ftz_nan_f16", - "llvm.nvvm.fmax.ftz.nan.f16x2" => "__nvvm_fmax_ftz_nan_f16x2", + "llvm.nvvm.fmax.ftz.nan.xorsign.abs.bf16" => "__nvvm_fmax_ftz_nan_xorsign_abs_bf16", + "llvm.nvvm.fmax.ftz.nan.xorsign.abs.bf16x2" => "__nvvm_fmax_ftz_nan_xorsign_abs_bf16x2", "llvm.nvvm.fmax.ftz.nan.xorsign.abs.f" => "__nvvm_fmax_ftz_nan_xorsign_abs_f", - "llvm.nvvm.fmax.ftz.nan.xorsign.abs.f16" => "__nvvm_fmax_ftz_nan_xorsign_abs_f16", - "llvm.nvvm.fmax.ftz.nan.xorsign.abs.f16x2" => "__nvvm_fmax_ftz_nan_xorsign_abs_f16x2", + "llvm.nvvm.fmax.ftz.xorsign.abs.bf16" => "__nvvm_fmax_ftz_xorsign_abs_bf16", + "llvm.nvvm.fmax.ftz.xorsign.abs.bf16x2" => "__nvvm_fmax_ftz_xorsign_abs_bf16x2", "llvm.nvvm.fmax.ftz.xorsign.abs.f" => "__nvvm_fmax_ftz_xorsign_abs_f", - "llvm.nvvm.fmax.ftz.xorsign.abs.f16" => "__nvvm_fmax_ftz_xorsign_abs_f16", - "llvm.nvvm.fmax.ftz.xorsign.abs.f16x2" => "__nvvm_fmax_ftz_xorsign_abs_f16x2", "llvm.nvvm.fmax.nan.bf16" => "__nvvm_fmax_nan_bf16", "llvm.nvvm.fmax.nan.bf16x2" => "__nvvm_fmax_nan_bf16x2", "llvm.nvvm.fmax.nan.f" => "__nvvm_fmax_nan_f", - "llvm.nvvm.fmax.nan.f16" => "__nvvm_fmax_nan_f16", - "llvm.nvvm.fmax.nan.f16x2" => "__nvvm_fmax_nan_f16x2", "llvm.nvvm.fmax.nan.xorsign.abs.bf16" => "__nvvm_fmax_nan_xorsign_abs_bf16", "llvm.nvvm.fmax.nan.xorsign.abs.bf16x2" => "__nvvm_fmax_nan_xorsign_abs_bf16x2", "llvm.nvvm.fmax.nan.xorsign.abs.f" => "__nvvm_fmax_nan_xorsign_abs_f", - "llvm.nvvm.fmax.nan.xorsign.abs.f16" => "__nvvm_fmax_nan_xorsign_abs_f16", - "llvm.nvvm.fmax.nan.xorsign.abs.f16x2" => "__nvvm_fmax_nan_xorsign_abs_f16x2", "llvm.nvvm.fmax.xorsign.abs.bf16" => "__nvvm_fmax_xorsign_abs_bf16", "llvm.nvvm.fmax.xorsign.abs.bf16x2" => "__nvvm_fmax_xorsign_abs_bf16x2", "llvm.nvvm.fmax.xorsign.abs.f" => "__nvvm_fmax_xorsign_abs_f", - "llvm.nvvm.fmax.xorsign.abs.f16" => "__nvvm_fmax_xorsign_abs_f16", - "llvm.nvvm.fmax.xorsign.abs.f16x2" => "__nvvm_fmax_xorsign_abs_f16x2", "llvm.nvvm.fmin.bf16" => "__nvvm_fmin_bf16", "llvm.nvvm.fmin.bf16x2" => "__nvvm_fmin_bf16x2", "llvm.nvvm.fmin.d" => "__nvvm_fmin_d", "llvm.nvvm.fmin.f" => "__nvvm_fmin_f", - "llvm.nvvm.fmin.f16" => "__nvvm_fmin_f16", - "llvm.nvvm.fmin.f16x2" => "__nvvm_fmin_f16x2", + "llvm.nvvm.fmin.ftz.bf16" => "__nvvm_fmin_ftz_bf16", + "llvm.nvvm.fmin.ftz.bf16x2" => "__nvvm_fmin_ftz_bf16x2", "llvm.nvvm.fmin.ftz.f" => "__nvvm_fmin_ftz_f", - "llvm.nvvm.fmin.ftz.f16" => "__nvvm_fmin_ftz_f16", - "llvm.nvvm.fmin.ftz.f16x2" => "__nvvm_fmin_ftz_f16x2", + "llvm.nvvm.fmin.ftz.nan.bf16" => "__nvvm_fmin_ftz_nan_bf16", + "llvm.nvvm.fmin.ftz.nan.bf16x2" => "__nvvm_fmin_ftz_nan_bf16x2", "llvm.nvvm.fmin.ftz.nan.f" => "__nvvm_fmin_ftz_nan_f", - "llvm.nvvm.fmin.ftz.nan.f16" => "__nvvm_fmin_ftz_nan_f16", - "llvm.nvvm.fmin.ftz.nan.f16x2" => "__nvvm_fmin_ftz_nan_f16x2", + "llvm.nvvm.fmin.ftz.nan.xorsign.abs.bf16" => "__nvvm_fmin_ftz_nan_xorsign_abs_bf16", + "llvm.nvvm.fmin.ftz.nan.xorsign.abs.bf16x2" => "__nvvm_fmin_ftz_nan_xorsign_abs_bf16x2", "llvm.nvvm.fmin.ftz.nan.xorsign.abs.f" => "__nvvm_fmin_ftz_nan_xorsign_abs_f", - "llvm.nvvm.fmin.ftz.nan.xorsign.abs.f16" => "__nvvm_fmin_ftz_nan_xorsign_abs_f16", - "llvm.nvvm.fmin.ftz.nan.xorsign.abs.f16x2" => "__nvvm_fmin_ftz_nan_xorsign_abs_f16x2", + "llvm.nvvm.fmin.ftz.xorsign.abs.bf16" => "__nvvm_fmin_ftz_xorsign_abs_bf16", + "llvm.nvvm.fmin.ftz.xorsign.abs.bf16x2" => "__nvvm_fmin_ftz_xorsign_abs_bf16x2", "llvm.nvvm.fmin.ftz.xorsign.abs.f" => "__nvvm_fmin_ftz_xorsign_abs_f", - "llvm.nvvm.fmin.ftz.xorsign.abs.f16" => "__nvvm_fmin_ftz_xorsign_abs_f16", - "llvm.nvvm.fmin.ftz.xorsign.abs.f16x2" => "__nvvm_fmin_ftz_xorsign_abs_f16x2", "llvm.nvvm.fmin.nan.bf16" => "__nvvm_fmin_nan_bf16", "llvm.nvvm.fmin.nan.bf16x2" => "__nvvm_fmin_nan_bf16x2", "llvm.nvvm.fmin.nan.f" => "__nvvm_fmin_nan_f", - "llvm.nvvm.fmin.nan.f16" => "__nvvm_fmin_nan_f16", - "llvm.nvvm.fmin.nan.f16x2" => "__nvvm_fmin_nan_f16x2", "llvm.nvvm.fmin.nan.xorsign.abs.bf16" => "__nvvm_fmin_nan_xorsign_abs_bf16", "llvm.nvvm.fmin.nan.xorsign.abs.bf16x2" => "__nvvm_fmin_nan_xorsign_abs_bf16x2", "llvm.nvvm.fmin.nan.xorsign.abs.f" => "__nvvm_fmin_nan_xorsign_abs_f", - "llvm.nvvm.fmin.nan.xorsign.abs.f16" => "__nvvm_fmin_nan_xorsign_abs_f16", - "llvm.nvvm.fmin.nan.xorsign.abs.f16x2" => "__nvvm_fmin_nan_xorsign_abs_f16x2", "llvm.nvvm.fmin.xorsign.abs.bf16" => "__nvvm_fmin_xorsign_abs_bf16", "llvm.nvvm.fmin.xorsign.abs.bf16x2" => "__nvvm_fmin_xorsign_abs_bf16x2", "llvm.nvvm.fmin.xorsign.abs.f" => "__nvvm_fmin_xorsign_abs_f", - "llvm.nvvm.fmin.xorsign.abs.f16" => "__nvvm_fmin_xorsign_abs_f16", - "llvm.nvvm.fmin.xorsign.abs.f16x2" => "__nvvm_fmin_xorsign_abs_f16x2", "llvm.nvvm.fns" => "__nvvm_fns", "llvm.nvvm.h2f" => "__nvvm_h2f", "llvm.nvvm.i2d.rm" => "__nvvm_i2d_rm", @@ -4255,6 +4267,28 @@ match name { "llvm.r600.read.tgid.x" => "__builtin_r600_read_tgid_x", "llvm.r600.read.tgid.y" => "__builtin_r600_read_tgid_y", "llvm.r600.read.tgid.z" => "__builtin_r600_read_tgid_z", + // riscv + "llvm.riscv.aes32dsi" => "__builtin_riscv_aes32dsi", + "llvm.riscv.aes32dsmi" => "__builtin_riscv_aes32dsmi", + "llvm.riscv.aes32esi" => "__builtin_riscv_aes32esi", + "llvm.riscv.aes32esmi" => "__builtin_riscv_aes32esmi", + "llvm.riscv.aes64ds" => "__builtin_riscv_aes64ds", + "llvm.riscv.aes64dsm" => "__builtin_riscv_aes64dsm", + "llvm.riscv.aes64es" => "__builtin_riscv_aes64es", + "llvm.riscv.aes64esm" => "__builtin_riscv_aes64esm", + "llvm.riscv.aes64im" => "__builtin_riscv_aes64im", + "llvm.riscv.aes64ks1i" => "__builtin_riscv_aes64ks1i", + "llvm.riscv.aes64ks2" => "__builtin_riscv_aes64ks2", + "llvm.riscv.sha512sig0" => "__builtin_riscv_sha512sig0", + "llvm.riscv.sha512sig0h" => "__builtin_riscv_sha512sig0h", + "llvm.riscv.sha512sig0l" => "__builtin_riscv_sha512sig0l", + "llvm.riscv.sha512sig1" => "__builtin_riscv_sha512sig1", + "llvm.riscv.sha512sig1h" => "__builtin_riscv_sha512sig1h", + "llvm.riscv.sha512sig1l" => "__builtin_riscv_sha512sig1l", + "llvm.riscv.sha512sum0" => "__builtin_riscv_sha512sum0", + "llvm.riscv.sha512sum0r" => "__builtin_riscv_sha512sum0r", + "llvm.riscv.sha512sum1" => "__builtin_riscv_sha512sum1", + "llvm.riscv.sha512sum1r" => "__builtin_riscv_sha512sum1r", // s390 "llvm.s390.efpc" => "__builtin_s390_efpc", "llvm.s390.etnd" => "__builtin_tx_nesting_depth", @@ -5954,6 +5988,18 @@ match name { "llvm.x86.avx2.vpdpbuud.256" => "__builtin_ia32_vpdpbuud256", "llvm.x86.avx2.vpdpbuuds.128" => "__builtin_ia32_vpdpbuuds128", "llvm.x86.avx2.vpdpbuuds.256" => "__builtin_ia32_vpdpbuuds256", + "llvm.x86.avx2.vpdpwsud.128" => "__builtin_ia32_vpdpwsud128", + "llvm.x86.avx2.vpdpwsud.256" => "__builtin_ia32_vpdpwsud256", + "llvm.x86.avx2.vpdpwsuds.128" => "__builtin_ia32_vpdpwsuds128", + "llvm.x86.avx2.vpdpwsuds.256" => "__builtin_ia32_vpdpwsuds256", + "llvm.x86.avx2.vpdpwusd.128" => "__builtin_ia32_vpdpwusd128", + "llvm.x86.avx2.vpdpwusd.256" => "__builtin_ia32_vpdpwusd256", + "llvm.x86.avx2.vpdpwusds.128" => "__builtin_ia32_vpdpwusds128", + "llvm.x86.avx2.vpdpwusds.256" => "__builtin_ia32_vpdpwusds256", + "llvm.x86.avx2.vpdpwuud.128" => "__builtin_ia32_vpdpwuud128", + "llvm.x86.avx2.vpdpwuud.256" => "__builtin_ia32_vpdpwuud256", + "llvm.x86.avx2.vpdpwuuds.128" => "__builtin_ia32_vpdpwuuds128", + "llvm.x86.avx2.vpdpwuuds.256" => "__builtin_ia32_vpdpwuuds256", "llvm.x86.avx2.vperm2i128" => "__builtin_ia32_permti256", "llvm.x86.avx512.add.pd.512" => "__builtin_ia32_addpd512", "llvm.x86.avx512.add.ps.512" => "__builtin_ia32_addps512", @@ -7895,6 +7941,10 @@ match name { "llvm.x86.subborrow.u64" => "__builtin_ia32_subborrow_u64", "llvm.x86.tbm.bextri.u32" => "__builtin_ia32_bextri_u32", "llvm.x86.tbm.bextri.u64" => "__builtin_ia32_bextri_u64", + "llvm.x86.tcmmimfp16ps" => "__builtin_ia32_tcmmimfp16ps", + "llvm.x86.tcmmimfp16ps.internal" => "__builtin_ia32_tcmmimfp16ps_internal", + "llvm.x86.tcmmrlfp16ps" => "__builtin_ia32_tcmmrlfp16ps", + "llvm.x86.tcmmrlfp16ps.internal" => "__builtin_ia32_tcmmrlfp16ps_internal", "llvm.x86.tdpbf16ps" => "__builtin_ia32_tdpbf16ps", "llvm.x86.tdpbf16ps.internal" => "__builtin_ia32_tdpbf16ps_internal", "llvm.x86.tdpbssd" => "__builtin_ia32_tdpbssd", @@ -7947,6 +7997,16 @@ match name { "llvm.x86.vgf2p8mulb.128" => "__builtin_ia32_vgf2p8mulb_v16qi", "llvm.x86.vgf2p8mulb.256" => "__builtin_ia32_vgf2p8mulb_v32qi", "llvm.x86.vgf2p8mulb.512" => "__builtin_ia32_vgf2p8mulb_v64qi", + "llvm.x86.vsha512msg1" => "__builtin_ia32_vsha512msg1", + "llvm.x86.vsha512msg2" => "__builtin_ia32_vsha512msg2", + "llvm.x86.vsha512rnds2" => "__builtin_ia32_vsha512rnds2", + "llvm.x86.vsm3msg1" => "__builtin_ia32_vsm3msg1", + "llvm.x86.vsm3msg2" => "__builtin_ia32_vsm3msg2", + "llvm.x86.vsm3rnds2" => "__builtin_ia32_vsm3rnds2", + "llvm.x86.vsm4key4128" => "__builtin_ia32_vsm4key4128", + "llvm.x86.vsm4key4256" => "__builtin_ia32_vsm4key4256", + "llvm.x86.vsm4rnds4128" => "__builtin_ia32_vsm4rnds4128", + "llvm.x86.vsm4rnds4256" => "__builtin_ia32_vsm4rnds4256", "llvm.x86.wbinvd" => "__builtin_ia32_wbinvd", "llvm.x86.wbnoinvd" => "__builtin_ia32_wbnoinvd", "llvm.x86.wrfsbase.32" => "__builtin_ia32_wrfsbase32", diff --git a/compiler/rustc_codegen_gcc/src/intrinsic/llvm.rs b/compiler/rustc_codegen_gcc/src/intrinsic/llvm.rs index 0edec566be3..5996623bdc5 100644 --- a/compiler/rustc_codegen_gcc/src/intrinsic/llvm.rs +++ b/compiler/rustc_codegen_gcc/src/intrinsic/llvm.rs @@ -236,11 +236,17 @@ pub fn adjust_intrinsic_arguments<'a, 'b, 'gcc, 'tcx>(builder: &Builder<'a, 'gcc let arg2 = builder.context.new_cast(None, arg2, arg2_type); args = vec![new_args[0], arg2].into(); }, + // These builtins are sent one more argument than needed. "__builtin_prefetch" => { let mut new_args = args.to_vec(); new_args.pop(); args = new_args.into(); }, + // The GCC version returns one value of the tuple through a pointer. + "__builtin_ia32_rdrand64_step" => { + let arg = builder.current_func().new_local(None, builder.ulonglong_type, "return_rdrand_arg"); + args = vec![arg.get_address(None)].into(); + }, _ => (), } } @@ -313,6 +319,13 @@ pub fn adjust_intrinsic_arguments<'a, 'b, 'gcc, 'tcx>(builder: &Builder<'a, 'gcc let new_args = args.to_vec(); args = vec![new_args[1], new_args[0], new_args[2], new_args[3], new_args[4]].into(); }, + "__builtin_ia32_vpshrdv_v8di" | "__builtin_ia32_vpshrdv_v4di" | "__builtin_ia32_vpshrdv_v2di" | + "__builtin_ia32_vpshrdv_v16si" | "__builtin_ia32_vpshrdv_v8si" | "__builtin_ia32_vpshrdv_v4si" | + "__builtin_ia32_vpshrdv_v32hi" | "__builtin_ia32_vpshrdv_v16hi" | "__builtin_ia32_vpshrdv_v8hi" => { + // The first two arguments are reversed, compared to LLVM. + let new_args = args.to_vec(); + args = vec![new_args[1], new_args[0], new_args[2]].into(); + }, _ => (), } } @@ -354,6 +367,19 @@ pub fn adjust_intrinsic_return_value<'a, 'gcc, 'tcx>(builder: &Builder<'a, 'gcc, // builtin twice, we overwrite the return value with a dummy value. return_value = builder.context.new_rvalue_zero(builder.int_type); }, + "__builtin_ia32_rdrand64_step" => { + let random_number = args[0].dereference(None).to_rvalue(); + let success_variable = builder.current_func().new_local(None, return_value.get_type(), "success"); + builder.llbb().add_assignment(None, success_variable, return_value); + + let field1 = builder.context.new_field(None, random_number.get_type(), "random_number"); + let field2 = builder.context.new_field(None, return_value.get_type(), "success"); + let struct_type = builder.context.new_struct_type(None, "rdrand_result", &[field1, field2]); + return_value = builder.context.new_struct_constructor(None, struct_type.as_type(), None, &[ + random_number, + success_variable.to_rvalue(), + ]); + }, _ => (), } @@ -606,6 +632,7 @@ pub fn intrinsic<'gcc, 'tcx>(name: &str, cx: &CodegenCx<'gcc, 'tcx>) -> Function "llvm.fshr.v8i16" => "__builtin_ia32_vpshrdv_v8hi", "llvm.x86.fma.vfmadd.sd" => "__builtin_ia32_vfmaddsd3", "llvm.x86.fma.vfmadd.ss" => "__builtin_ia32_vfmaddss3", + "llvm.x86.rdrand.64" => "__builtin_ia32_rdrand64_step", // The above doc points to unknown builtins for the following, so override them: "llvm.x86.avx2.gather.d.d" => "__builtin_ia32_gathersiv4si", diff --git a/compiler/rustc_codegen_gcc/src/intrinsic/mod.rs b/compiler/rustc_codegen_gcc/src/intrinsic/mod.rs index 94dc8c9e93b..9caed459a29 100644 --- a/compiler/rustc_codegen_gcc/src/intrinsic/mod.rs +++ b/compiler/rustc_codegen_gcc/src/intrinsic/mod.rs @@ -10,9 +10,10 @@ use rustc_codegen_ssa::base::wants_msvc_seh; use rustc_codegen_ssa::common::IntPredicate; use rustc_codegen_ssa::mir::operand::{OperandRef, OperandValue}; use rustc_codegen_ssa::mir::place::PlaceRef; -use rustc_codegen_ssa::traits::{ArgAbiMethods, BaseTypeMethods, BuilderMethods, ConstMethods, IntrinsicCallMethods}; +use rustc_codegen_ssa::traits::{ArgAbiMethods, BuilderMethods, ConstMethods, IntrinsicCallMethods}; #[cfg(feature="master")] -use rustc_codegen_ssa::traits::{DerivedTypeMethods, MiscMethods}; +use rustc_codegen_ssa::traits::{BaseTypeMethods, MiscMethods}; +use rustc_codegen_ssa::errors::InvalidMonomorphization; use rustc_middle::bug; use rustc_middle::ty::{self, Instance, Ty}; use rustc_middle::ty::layout::LayoutOf; @@ -31,7 +32,6 @@ use crate::abi::FnAbiGccExt; use crate::builder::Builder; use crate::common::{SignType, TypeReflection}; use crate::context::CodegenCx; -use crate::errors::InvalidMonomorphizationBasicInteger; use crate::type_of::LayoutGccExt; use crate::intrinsic::simd::generic_simd_intrinsic; @@ -92,8 +92,8 @@ impl<'a, 'gcc, 'tcx> IntrinsicCallMethods<'tcx> for Builder<'a, 'gcc, 'tcx> { let tcx = self.tcx; let callee_ty = instance.ty(tcx, ty::ParamEnv::reveal_all()); - let (def_id, substs) = match *callee_ty.kind() { - ty::FnDef(def_id, substs) => (def_id, substs), + let (def_id, fn_args) = match *callee_ty.kind() { + ty::FnDef(def_id, fn_args) => (def_id, fn_args), _ => bug!("expected fn item type, found {}", callee_ty), }; @@ -113,7 +113,7 @@ impl<'a, 'gcc, 'tcx> IntrinsicCallMethods<'tcx> for Builder<'a, 'gcc, 'tcx> { _ if simple.is_some() => { // FIXME(antoyo): remove this cast when the API supports function. let func = unsafe { std::mem::transmute(simple.expect("simple")) }; - self.call(self.type_void(), None, func, &args.iter().map(|arg| arg.immediate()).collect::<Vec<_>>(), None) + self.call(self.type_void(), None, None, func, &args.iter().map(|arg| arg.immediate()).collect::<Vec<_>>(), None) }, sym::likely => { self.expect(args[0].immediate(), true) @@ -142,9 +142,9 @@ impl<'a, 'gcc, 'tcx> IntrinsicCallMethods<'tcx> for Builder<'a, 'gcc, 'tcx> { } sym::volatile_load | sym::unaligned_volatile_load => { - let tp_ty = substs.type_at(0); + let tp_ty = fn_args.type_at(0); let mut ptr = args[0].immediate(); - if let PassMode::Cast(ty, _) = &fn_abi.ret.mode { + if let PassMode::Cast { cast: ty, .. } = &fn_abi.ret.mode { ptr = self.pointercast(ptr, self.type_ptr_to(ty.gcc_type(self))); } let load = self.volatile_load(ptr.get_type(), ptr); @@ -256,7 +256,7 @@ impl<'a, 'gcc, 'tcx> IntrinsicCallMethods<'tcx> for Builder<'a, 'gcc, 'tcx> { _ => bug!(), }, None => { - tcx.sess.emit_err(InvalidMonomorphizationBasicInteger { span, name, ty }); + tcx.sess.emit_err(InvalidMonomorphization::BasicIntegerType { span, name, ty }); return; } } @@ -264,7 +264,7 @@ impl<'a, 'gcc, 'tcx> IntrinsicCallMethods<'tcx> for Builder<'a, 'gcc, 'tcx> { sym::raw_eq => { use rustc_target::abi::Abi::*; - let tp_ty = substs.type_at(0); + let tp_ty = fn_args.type_at(0); let layout = self.layout_of(tp_ty).layout; let _use_integer_compare = match layout.abi() { Scalar(_) | ScalarPair(_, _) => true, @@ -302,6 +302,21 @@ impl<'a, 'gcc, 'tcx> IntrinsicCallMethods<'tcx> for Builder<'a, 'gcc, 'tcx> { } } + sym::compare_bytes => { + let a = args[0].immediate(); + let b = args[1].immediate(); + let n = args[2].immediate(); + + let void_ptr_type = self.context.new_type::<*const ()>(); + let a_ptr = self.bitcast(a, void_ptr_type); + let b_ptr = self.bitcast(b, void_ptr_type); + + // Here we assume that the `memcmp` provided by the target is a NOP for size 0. + let builtin = self.context.get_builtin_function("memcmp"); + let cmp = self.context.new_call(None, builtin, &[a_ptr, b_ptr, n]); + self.sext(cmp, self.type_ix(32)) + } + sym::black_box => { args[0].val.store(self, result); @@ -326,7 +341,7 @@ impl<'a, 'gcc, 'tcx> IntrinsicCallMethods<'tcx> for Builder<'a, 'gcc, 'tcx> { let masked = self.and(addr, mask); self.bitcast(masked, void_ptr_type) }, - + _ if name_str.starts_with("simd_") => { match generic_simd_intrinsic(self, name, callee_ty, args, ret_ty, llret_ty, span) { Ok(llval) => llval, @@ -338,7 +353,7 @@ impl<'a, 'gcc, 'tcx> IntrinsicCallMethods<'tcx> for Builder<'a, 'gcc, 'tcx> { }; if !fn_abi.ret.is_ignore() { - if let PassMode::Cast(ty, _) = &fn_abi.ret.mode { + if let PassMode::Cast { cast: ty, .. } = &fn_abi.ret.mode { let ptr_llty = self.type_ptr_to(ty.gcc_type(self)); let ptr = self.pointercast(result.llval, ptr_llty); self.store(llval, ptr, result.align); @@ -354,7 +369,7 @@ impl<'a, 'gcc, 'tcx> IntrinsicCallMethods<'tcx> for Builder<'a, 'gcc, 'tcx> { fn abort(&mut self) { let func = self.context.get_builtin_function("abort"); let func: RValue<'gcc> = unsafe { std::mem::transmute(func) }; - self.call(self.type_void(), None, func, &[], None); + self.call(self.type_void(), None, None, func, &[], None); } fn assume(&mut self, value: Self::Value) { @@ -434,7 +449,7 @@ impl<'gcc, 'tcx> ArgAbiExt<'gcc, 'tcx> for ArgAbi<'tcx, Ty<'tcx>> { else if self.is_unsized_indirect() { bug!("unsized `ArgAbi` must be handled through `store_fn_arg`"); } - else if let PassMode::Cast(ref cast, _) = self.mode { + else if let PassMode::Cast { ref cast, .. } = self.mode { // FIXME(eddyb): Figure out when the simpler Store is safe, clang // uses it for i16 -> {i8, i8}, but not for i24 -> {i8, i8, i8}. let can_store_through_cast_ptr = false; @@ -496,10 +511,10 @@ impl<'gcc, 'tcx> ArgAbiExt<'gcc, 'tcx> for ArgAbi<'tcx, Ty<'tcx>> { PassMode::Pair(..) => { OperandValue::Pair(next(), next()).store(bx, dst); }, - PassMode::Indirect { extra_attrs: Some(_), .. } => { + PassMode::Indirect { meta_attrs: Some(_), .. } => { OperandValue::Ref(next(), Some(next()), self.layout.align.abi).store(bx, dst); }, - PassMode::Direct(_) | PassMode::Indirect { extra_attrs: None, .. } | PassMode::Cast(..) => { + PassMode::Direct(_) | PassMode::Indirect { meta_attrs: None, .. } | PassMode::Cast { .. } => { let next_arg = next(); self.store(bx, next_arg, dst); }, @@ -551,141 +566,52 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> { let context = &self.cx.context; let result = match width { - 8 => { - // First step. - let left = self.and(value, context.new_rvalue_from_int(typ, 0xF0)); - let left = self.lshr(left, context.new_rvalue_from_int(typ, 4)); - let right = self.and(value, context.new_rvalue_from_int(typ, 0x0F)); - let right = self.shl(right, context.new_rvalue_from_int(typ, 4)); - let step1 = self.or(left, right); - - // Second step. - let left = self.and(step1, context.new_rvalue_from_int(typ, 0xCC)); - let left = self.lshr(left, context.new_rvalue_from_int(typ, 2)); - let right = self.and(step1, context.new_rvalue_from_int(typ, 0x33)); - let right = self.shl(right, context.new_rvalue_from_int(typ, 2)); - let step2 = self.or(left, right); - - // Third step. - let left = self.and(step2, context.new_rvalue_from_int(typ, 0xAA)); - let left = self.lshr(left, context.new_rvalue_from_int(typ, 1)); - let right = self.and(step2, context.new_rvalue_from_int(typ, 0x55)); - let right = self.shl(right, context.new_rvalue_from_int(typ, 1)); - let step3 = self.or(left, right); - - step3 - }, - 16 => { - // First step. - let left = self.and(value, context.new_rvalue_from_int(typ, 0x5555)); - let left = self.shl(left, context.new_rvalue_from_int(typ, 1)); - let right = self.and(value, context.new_rvalue_from_int(typ, 0xAAAA)); - let right = self.lshr(right, context.new_rvalue_from_int(typ, 1)); - let step1 = self.or(left, right); - - // Second step. - let left = self.and(step1, context.new_rvalue_from_int(typ, 0x3333)); - let left = self.shl(left, context.new_rvalue_from_int(typ, 2)); - let right = self.and(step1, context.new_rvalue_from_int(typ, 0xCCCC)); - let right = self.lshr(right, context.new_rvalue_from_int(typ, 2)); - let step2 = self.or(left, right); - - // Third step. - let left = self.and(step2, context.new_rvalue_from_int(typ, 0x0F0F)); - let left = self.shl(left, context.new_rvalue_from_int(typ, 4)); - let right = self.and(step2, context.new_rvalue_from_int(typ, 0xF0F0)); - let right = self.lshr(right, context.new_rvalue_from_int(typ, 4)); - let step3 = self.or(left, right); - - // Fourth step. - let left = self.and(step3, context.new_rvalue_from_int(typ, 0x00FF)); - let left = self.shl(left, context.new_rvalue_from_int(typ, 8)); - let right = self.and(step3, context.new_rvalue_from_int(typ, 0xFF00)); - let right = self.lshr(right, context.new_rvalue_from_int(typ, 8)); - let step4 = self.or(left, right); + 8 | 16 | 32 | 64 => { + let mask = ((1u128 << width) - 1) as u64; + let (m0, m1, m2) = if width > 16 { + ( + context.new_rvalue_from_long(typ, (0x5555555555555555u64 & mask) as i64), + context.new_rvalue_from_long(typ, (0x3333333333333333u64 & mask) as i64), + context.new_rvalue_from_long(typ, (0x0f0f0f0f0f0f0f0fu64 & mask) as i64), + ) + } else { + ( + context.new_rvalue_from_int(typ, (0x5555u64 & mask) as i32), + context.new_rvalue_from_int(typ, (0x3333u64 & mask) as i32), + context.new_rvalue_from_int(typ, (0x0f0fu64 & mask) as i32), + ) + }; + let one = context.new_rvalue_from_int(typ, 1); + let two = context.new_rvalue_from_int(typ, 2); + let four = context.new_rvalue_from_int(typ, 4); - step4 - }, - 32 => { - // TODO(antoyo): Refactor with other implementations. // First step. - let left = self.and(value, context.new_rvalue_from_long(typ, 0x55555555)); - let left = self.shl(left, context.new_rvalue_from_long(typ, 1)); - let right = self.and(value, context.new_rvalue_from_long(typ, 0xAAAAAAAA)); - let right = self.lshr(right, context.new_rvalue_from_long(typ, 1)); + let left = self.lshr(value, one); + let left = self.and(left, m0); + let right = self.and(value, m0); + let right = self.shl(right, one); let step1 = self.or(left, right); // Second step. - let left = self.and(step1, context.new_rvalue_from_long(typ, 0x33333333)); - let left = self.shl(left, context.new_rvalue_from_long(typ, 2)); - let right = self.and(step1, context.new_rvalue_from_long(typ, 0xCCCCCCCC)); - let right = self.lshr(right, context.new_rvalue_from_long(typ, 2)); + let left = self.lshr(step1, two); + let left = self.and(left, m1); + let right = self.and(step1, m1); + let right = self.shl(right, two); let step2 = self.or(left, right); // Third step. - let left = self.and(step2, context.new_rvalue_from_long(typ, 0x0F0F0F0F)); - let left = self.shl(left, context.new_rvalue_from_long(typ, 4)); - let right = self.and(step2, context.new_rvalue_from_long(typ, 0xF0F0F0F0)); - let right = self.lshr(right, context.new_rvalue_from_long(typ, 4)); + let left = self.lshr(step2, four); + let left = self.and(left, m2); + let right = self.and(step2, m2); + let right = self.shl(right, four); let step3 = self.or(left, right); // Fourth step. - let left = self.and(step3, context.new_rvalue_from_long(typ, 0x00FF00FF)); - let left = self.shl(left, context.new_rvalue_from_long(typ, 8)); - let right = self.and(step3, context.new_rvalue_from_long(typ, 0xFF00FF00)); - let right = self.lshr(right, context.new_rvalue_from_long(typ, 8)); - let step4 = self.or(left, right); - - // Fifth step. - let left = self.and(step4, context.new_rvalue_from_long(typ, 0x0000FFFF)); - let left = self.shl(left, context.new_rvalue_from_long(typ, 16)); - let right = self.and(step4, context.new_rvalue_from_long(typ, 0xFFFF0000)); - let right = self.lshr(right, context.new_rvalue_from_long(typ, 16)); - let step5 = self.or(left, right); - - step5 - }, - 64 => { - // First step. - let left = self.shl(value, context.new_rvalue_from_long(typ, 32)); - let right = self.lshr(value, context.new_rvalue_from_long(typ, 32)); - let step1 = self.or(left, right); - - // Second step. - let left = self.and(step1, context.new_rvalue_from_long(typ, 0x0001FFFF0001FFFF)); - let left = self.shl(left, context.new_rvalue_from_long(typ, 15)); - let right = self.and(step1, context.new_rvalue_from_long(typ, 0xFFFE0000FFFE0000u64 as i64)); // TODO(antoyo): transmute the number instead? - let right = self.lshr(right, context.new_rvalue_from_long(typ, 17)); - let step2 = self.or(left, right); - - // Third step. - let left = self.lshr(step2, context.new_rvalue_from_long(typ, 10)); - let left = self.xor(step2, left); - let temp = self.and(left, context.new_rvalue_from_long(typ, 0x003F801F003F801F)); - - let left = self.shl(temp, context.new_rvalue_from_long(typ, 10)); - let left = self.or(temp, left); - let step3 = self.xor(left, step2); - - // Fourth step. - let left = self.lshr(step3, context.new_rvalue_from_long(typ, 4)); - let left = self.xor(step3, left); - let temp = self.and(left, context.new_rvalue_from_long(typ, 0x0E0384210E038421)); - - let left = self.shl(temp, context.new_rvalue_from_long(typ, 4)); - let left = self.or(temp, left); - let step4 = self.xor(left, step3); - - // Fifth step. - let left = self.lshr(step4, context.new_rvalue_from_long(typ, 2)); - let left = self.xor(step4, left); - let temp = self.and(left, context.new_rvalue_from_long(typ, 0x2248884222488842)); - - let left = self.shl(temp, context.new_rvalue_from_long(typ, 2)); - let left = self.or(temp, left); - let step5 = self.xor(left, step4); - - step5 + if width == 8 { + step3 + } else { + self.gcc_bswap(step3, width) + } }, 128 => { // TODO(antoyo): find a more efficient implementation? @@ -1135,7 +1061,7 @@ impl<'a, 'gcc, 'tcx> Builder<'a, 'gcc, 'tcx> { fn try_intrinsic<'a, 'b, 'gcc, 'tcx>(bx: &'b mut Builder<'a, 'gcc, 'tcx>, try_func: RValue<'gcc>, data: RValue<'gcc>, _catch_func: RValue<'gcc>, dest: RValue<'gcc>) { if bx.sess().panic_strategy() == PanicStrategy::Abort { - bx.call(bx.type_void(), None, try_func, &[data], None); + bx.call(bx.type_void(), None, None, try_func, &[data], None); // Return 0 unconditionally from the intrinsic call; // we can never unwind. let ret_align = bx.tcx.data_layout.i32_align.abi; @@ -1204,21 +1130,21 @@ fn codegen_gnu_try<'gcc>(bx: &mut Builder<'_, 'gcc, '_>, try_func: RValue<'gcc>, let zero = bx.cx.context.new_rvalue_zero(bx.int_type); let ptr = bx.cx.context.new_call(None, eh_pointer_builtin, &[zero]); let catch_ty = bx.type_func(&[bx.type_i8p(), bx.type_i8p()], bx.type_void()); - bx.call(catch_ty, None, catch_func, &[data, ptr], None); + bx.call(catch_ty, None, None, catch_func, &[data, ptr], None); bx.ret(bx.const_i32(1)); // NOTE: the blocks must be filled before adding the try/catch, otherwise gcc will not // generate a try/catch. // FIXME(antoyo): add a check in the libgccjit API to prevent this. bx.switch_to_block(current_block); - bx.invoke(try_func_ty, None, try_func, &[data], then, catch, None); + bx.invoke(try_func_ty, None, None, try_func, &[data], then, catch, None); }); let func = unsafe { std::mem::transmute(func) }; // Note that no invoke is used here because by definition this function // can't panic (that's what it's catching). - let ret = bx.call(llty, None, func, &[try_func, data, catch_func], None); + let ret = bx.call(llty, None, None, func, &[try_func, data, catch_func], None); let i32_align = bx.tcx().data_layout.i32_align.abi; bx.store(ret, dest, i32_align); } @@ -1236,19 +1162,19 @@ fn get_rust_try_fn<'a, 'gcc, 'tcx>(cx: &'a CodegenCx<'gcc, 'tcx>, codegen: &mut // Define the type up front for the signature of the rust_try function. let tcx = cx.tcx; - let i8p = tcx.mk_mut_ptr(tcx.types.i8); + let i8p = Ty::new_mut_ptr(tcx,tcx.types.i8); // `unsafe fn(*mut i8) -> ()` - let try_fn_ty = tcx.mk_fn_ptr(ty::Binder::dummy(tcx.mk_fn_sig( + let try_fn_ty = Ty::new_fn_ptr(tcx,ty::Binder::dummy(tcx.mk_fn_sig( iter::once(i8p), - tcx.mk_unit(), + Ty::new_unit(tcx,), false, rustc_hir::Unsafety::Unsafe, Abi::Rust, ))); // `unsafe fn(*mut i8, *mut i8) -> ()` - let catch_fn_ty = tcx.mk_fn_ptr(ty::Binder::dummy(tcx.mk_fn_sig( + let catch_fn_ty = Ty::new_fn_ptr(tcx,ty::Binder::dummy(tcx.mk_fn_sig( [i8p, i8p].iter().cloned(), - tcx.mk_unit(), + Ty::new_unit(tcx,), false, rustc_hir::Unsafety::Unsafe, Abi::Rust, diff --git a/compiler/rustc_codegen_gcc/src/intrinsic/simd.rs b/compiler/rustc_codegen_gcc/src/intrinsic/simd.rs index b59c3a64f57..85d3e7234a0 100644 --- a/compiler/rustc_codegen_gcc/src/intrinsic/simd.rs +++ b/compiler/rustc_codegen_gcc/src/intrinsic/simd.rs @@ -1,11 +1,11 @@ -#[cfg(feature="master")] -use gccjit::{ComparisonOp, UnaryOp}; use gccjit::ToRValue; use gccjit::{BinaryOp, RValue, Type}; +#[cfg(feature = "master")] +use gccjit::{ComparisonOp, UnaryOp}; use rustc_codegen_ssa::base::compare_simd_types; use rustc_codegen_ssa::common::{IntPredicate, TypeKind}; -#[cfg(feature="master")] +#[cfg(feature = "master")] use rustc_codegen_ssa::errors::ExpectedPointerMutability; use rustc_codegen_ssa::errors::InvalidMonomorphization; use rustc_codegen_ssa::mir::operand::OperandRef; @@ -19,21 +19,8 @@ use rustc_span::{sym, Span, Symbol}; use rustc_target::abi::Align; use crate::builder::Builder; -#[cfg(feature="master")] +#[cfg(feature = "master")] use crate::context::CodegenCx; -#[cfg(feature="master")] -use crate::errors::{InvalidMonomorphizationExpectedSignedUnsigned, InvalidMonomorphizationInsertedType}; -use crate::errors::{ - InvalidMonomorphizationExpectedSimd, - InvalidMonomorphizationInvalidBitmask, - InvalidMonomorphizationInvalidFloatVector, InvalidMonomorphizationMaskType, - InvalidMonomorphizationMismatchedLengths, InvalidMonomorphizationNotFloat, - InvalidMonomorphizationReturnElement, InvalidMonomorphizationReturnIntegerType, - InvalidMonomorphizationReturnLength, InvalidMonomorphizationReturnLengthInputType, - InvalidMonomorphizationReturnType, InvalidMonomorphizationSimdShuffle, - InvalidMonomorphizationUnrecognized, InvalidMonomorphizationUnsupportedElement, - InvalidMonomorphizationUnsupportedOperation, -}; pub fn generic_simd_intrinsic<'a, 'gcc, 'tcx>( bx: &mut Builder<'a, 'gcc, 'tcx>, @@ -59,16 +46,8 @@ pub fn generic_simd_intrinsic<'a, 'gcc, 'tcx>( }; } macro_rules! require_simd { - ($ty: expr, $position: expr) => { - require!( - $ty.is_simd(), - InvalidMonomorphizationExpectedSimd { - span, - name, - position: $position, - found_ty: $ty - } - ) + ($ty: expr, $diag: expr) => { + require!($ty.is_simd(), $diag) }; } @@ -78,7 +57,10 @@ pub fn generic_simd_intrinsic<'a, 'gcc, 'tcx>( let arg_tys = sig.inputs(); if name == sym::simd_select_bitmask { - require_simd!(arg_tys[1], "argument"); + require_simd!( + arg_tys[1], + InvalidMonomorphization::SimdArgument { span, name, ty: arg_tys[1] } + ); let (len, _) = arg_tys[1].simd_size_and_type(bx.tcx()); let expected_int_bits = (len.max(8) - 1).next_power_of_two(); @@ -99,10 +81,10 @@ pub fn generic_simd_intrinsic<'a, 'gcc, 'tcx>( let ptr = bx.pointercast(place.llval, bx.cx.type_ptr_to(int_ty)); bx.load(int_ty, ptr, Align::ONE) } - _ => return_error!(InvalidMonomorphizationInvalidBitmask { + _ => return_error!(InvalidMonomorphization::InvalidBitmask { span, name, - ty: mask_ty, + mask_ty, expected_int_bits, expected_bytes }), @@ -116,7 +98,8 @@ pub fn generic_simd_intrinsic<'a, 'gcc, 'tcx>( // NOTE: since the arguments can be vectors of floats, make sure the mask is a vector of // integer. let mask_element_type = bx.type_ix(arg1_element_type.get_size() as u64 * 8); - let vector_mask_type = bx.context.new_vector_type(mask_element_type, arg1_vector_type.get_num_units() as u64); + let vector_mask_type = + bx.context.new_vector_type(mask_element_type, arg1_vector_type.get_num_units() as u64); let mut elements = vec![]; let one = bx.context.new_rvalue_one(mask.get_type()); @@ -131,7 +114,7 @@ pub fn generic_simd_intrinsic<'a, 'gcc, 'tcx>( } // every intrinsic below takes a SIMD vector as its first argument - require_simd!(arg_tys[0], "input"); + require_simd!(arg_tys[0], InvalidMonomorphization::SimdInput { span, name, ty: arg_tys[0] }); let in_ty = arg_tys[0]; let comparison = match name { @@ -146,12 +129,12 @@ pub fn generic_simd_intrinsic<'a, 'gcc, 'tcx>( let (in_len, in_elem) = arg_tys[0].simd_size_and_type(bx.tcx()); if let Some(cmp_op) = comparison { - require_simd!(ret_ty, "return"); + require_simd!(ret_ty, InvalidMonomorphization::SimdReturn { span, name, ty: ret_ty }); let (out_len, out_ty) = ret_ty.simd_size_and_type(bx.tcx()); require!( in_len == out_len, - InvalidMonomorphizationReturnLengthInputType { + InvalidMonomorphization::ReturnLengthInputType { span, name, in_len, @@ -162,51 +145,42 @@ pub fn generic_simd_intrinsic<'a, 'gcc, 'tcx>( ); require!( bx.type_kind(bx.element_type(llret_ty)) == TypeKind::Integer, - InvalidMonomorphizationReturnIntegerType { span, name, ret_ty, out_ty } + InvalidMonomorphization::ReturnIntegerType { span, name, ret_ty, out_ty } ); - return Ok(compare_simd_types( - bx, - args[0].immediate(), - args[1].immediate(), - in_elem, - llret_ty, - cmp_op, - )); + let arg1 = args[0].immediate(); + // NOTE: we get different vector types for the same vector type and libgccjit doesn't + // compare them as equal, so bitcast. + // FIXME(antoyo): allow comparing vector types as equal in libgccjit. + let arg2 = bx.context.new_bitcast(None, args[1].immediate(), arg1.get_type()); + return Ok(compare_simd_types(bx, arg1, arg2, in_elem, llret_ty, cmp_op)); } - if let Some(stripped) = name.as_str().strip_prefix("simd_shuffle") { - let n: u64 = if stripped.is_empty() { - // Make sure this is actually an array, since typeck only checks the length-suffixed - // version of this intrinsic. - match args[2].layout.ty.kind() { - ty::Array(ty, len) if matches!(ty.kind(), ty::Uint(ty::UintTy::U32)) => { - len.try_eval_target_usize(bx.cx.tcx, ty::ParamEnv::reveal_all()).unwrap_or_else( - || span_bug!(span, "could not evaluate shuffle index array length"), - ) - } - _ => return_error!(InvalidMonomorphizationSimdShuffle { - span, - name, - ty: args[2].layout.ty - }), + if name == sym::simd_shuffle { + // Make sure this is actually an array, since typeck only checks the length-suffixed + // version of this intrinsic. + let n: u64 = match args[2].layout.ty.kind() { + ty::Array(ty, len) if matches!(ty.kind(), ty::Uint(ty::UintTy::U32)) => { + len.try_eval_target_usize(bx.cx.tcx, ty::ParamEnv::reveal_all()).unwrap_or_else( + || span_bug!(span, "could not evaluate shuffle index array length"), + ) } - } else { - stripped.parse().unwrap_or_else(|_| { - span_bug!(span, "bad `simd_shuffle` instruction only caught in codegen?") - }) + _ => return_error!(InvalidMonomorphization::SimdShuffle { + span, + name, + ty: args[2].layout.ty + }), }; - - require_simd!(ret_ty, "return"); + require_simd!(ret_ty, InvalidMonomorphization::SimdReturn { span, name, ty: ret_ty }); let (out_len, out_ty) = ret_ty.simd_size_and_type(bx.tcx()); require!( out_len == n, - InvalidMonomorphizationReturnLength { span, name, in_len: n, ret_ty, out_len } + InvalidMonomorphization::ReturnLength { span, name, in_len: n, ret_ty, out_len } ); require!( in_elem == out_ty, - InvalidMonomorphizationReturnElement { span, name, in_elem, in_ty, ret_ty, out_ty } + InvalidMonomorphization::ReturnElement { span, name, in_elem, in_ty, ret_ty, out_ty } ); let vector = args[2].immediate(); @@ -218,7 +192,13 @@ pub fn generic_simd_intrinsic<'a, 'gcc, 'tcx>( if name == sym::simd_insert { require!( in_elem == arg_tys[2], - InvalidMonomorphizationInsertedType { span, name, in_elem, in_ty, out_ty: arg_tys[2] } + InvalidMonomorphization::InsertedType { + span, + name, + in_elem, + in_ty, + out_ty: arg_tys[2] + } ); let vector = args[0].immediate(); let index = args[1].immediate(); @@ -235,7 +215,7 @@ pub fn generic_simd_intrinsic<'a, 'gcc, 'tcx>( if name == sym::simd_extract { require!( ret_ty == in_elem, - InvalidMonomorphizationReturnType { span, name, in_elem, in_ty, ret_ty } + InvalidMonomorphization::ReturnType { span, name, in_elem, in_ty, ret_ty } ); let vector = args[0].immediate(); return Ok(bx.context.new_vector_access(None, vector, args[1].immediate()).to_rvalue()); @@ -244,26 +224,29 @@ pub fn generic_simd_intrinsic<'a, 'gcc, 'tcx>( if name == sym::simd_select { let m_elem_ty = in_elem; let m_len = in_len; - require_simd!(arg_tys[1], "argument"); + require_simd!( + arg_tys[1], + InvalidMonomorphization::SimdArgument { span, name, ty: arg_tys[1] } + ); let (v_len, _) = arg_tys[1].simd_size_and_type(bx.tcx()); require!( m_len == v_len, - InvalidMonomorphizationMismatchedLengths { span, name, m_len, v_len } + InvalidMonomorphization::MismatchedLengths { span, name, m_len, v_len } ); match m_elem_ty.kind() { ty::Int(_) => {} - _ => return_error!(InvalidMonomorphizationMaskType { span, name, ty: m_elem_ty }), + _ => return_error!(InvalidMonomorphization::MaskType { span, name, ty: m_elem_ty }), } return Ok(bx.vector_select(args[0].immediate(), args[1].immediate(), args[2].immediate())); } - #[cfg(feature="master")] + #[cfg(feature = "master")] if name == sym::simd_cast || name == sym::simd_as { - require_simd!(ret_ty, "return"); + require_simd!(ret_ty, InvalidMonomorphization::SimdReturn { span, name, ty: ret_ty }); let (out_len, out_elem) = ret_ty.simd_size_and_type(bx.tcx()); require!( in_len == out_len, - InvalidMonomorphizationReturnLengthInputType { + InvalidMonomorphization::ReturnLengthInputType { span, name, in_len, @@ -283,19 +266,17 @@ pub fn generic_simd_intrinsic<'a, 'gcc, 'tcx>( Unsupported, } - let in_style = - match in_elem.kind() { - ty::Int(_) | ty::Uint(_) => Style::Int, - ty::Float(_) => Style::Float, - _ => Style::Unsupported, - }; + let in_style = match in_elem.kind() { + ty::Int(_) | ty::Uint(_) => Style::Int, + ty::Float(_) => Style::Float, + _ => Style::Unsupported, + }; - let out_style = - match out_elem.kind() { - ty::Int(_) | ty::Uint(_) => Style::Int, - ty::Float(_) => Style::Float, - _ => Style::Unsupported, - }; + let out_style = match out_elem.kind() { + ty::Int(_) | ty::Uint(_) => Style::Int, + ty::Float(_) => Style::Float, + _ => Style::Unsupported, + }; match (in_style, out_style) { (Style::Unsupported, Style::Unsupported) => { @@ -310,7 +291,7 @@ pub fn generic_simd_intrinsic<'a, 'gcc, 'tcx>( out_elem } ); - }, + } _ => return Ok(bx.context.convert_vector(None, args[0].immediate(), llret_ty)), } } @@ -324,7 +305,7 @@ pub fn generic_simd_intrinsic<'a, 'gcc, 'tcx>( })* _ => {}, } - return_error!(InvalidMonomorphizationUnsupportedOperation { span, name, in_ty, in_elem }) + return_error!(InvalidMonomorphization::UnsupportedOperation { span, name, in_ty, in_elem }) })* } } @@ -341,7 +322,8 @@ pub fn generic_simd_intrinsic<'a, 'gcc, 'tcx>( // endian and MSB-first for big endian. let vector = args[0].immediate(); - let vector_type = vector.get_type().dyncast_vector().expect("vector type"); + // TODO(antoyo): dyncast_vector should not require a call to unqualified. + let vector_type = vector.get_type().unqualified().dyncast_vector().expect("vector type"); let elem_type = vector_type.get_element_type(); let expected_int_bits = in_len.max(8); @@ -357,10 +339,13 @@ pub fn generic_simd_intrinsic<'a, 'gcc, 'tcx>( let mut shift = 0; for i in 0..in_len { - let elem = bx.extract_element(vector, bx.context.new_rvalue_from_int(bx.int_type, i as i32)); + let elem = + bx.extract_element(vector, bx.context.new_rvalue_from_int(bx.int_type, i as i32)); let shifted = elem >> sign_shift; let masked = shifted & one; - result = result | (bx.context.new_cast(None, masked, result_type) << bx.context.new_rvalue_from_int(result_type, shift)); + result = result + | (bx.context.new_cast(None, masked, result_type) + << bx.context.new_rvalue_from_int(result_type, shift)); shift += 1; } @@ -409,46 +394,50 @@ pub fn generic_simd_intrinsic<'a, 'gcc, 'tcx>( return Err(()); }}; } - let (elem_ty_str, elem_ty) = - if let ty::Float(f) = in_elem.kind() { - let elem_ty = bx.cx.type_float_from_ty(*f); - match f.bit_width() { - 32 => ("f", elem_ty), - 64 => ("", elem_ty), - _ => { - return_error!(InvalidMonomorphizationInvalidFloatVector { span, name, elem_ty: f.name_str(), vec_ty: in_ty }); - } + let (elem_ty_str, elem_ty) = if let ty::Float(f) = in_elem.kind() { + let elem_ty = bx.cx.type_float_from_ty(*f); + match f.bit_width() { + 32 => ("f", elem_ty), + 64 => ("", elem_ty), + _ => { + return_error!(InvalidMonomorphization::FloatingPointVector { + span, + name, + f_ty: *f, + in_ty + }); } } - else { - return_error!(InvalidMonomorphizationNotFloat { span, name, ty: in_ty }); - }; + } else { + return_error!(InvalidMonomorphization::FloatingPointType { span, name, in_ty }); + }; let vec_ty = bx.cx.type_vector(elem_ty, in_len); - let intr_name = - match name { - sym::simd_ceil => "ceil", - sym::simd_fabs => "fabs", // TODO(antoyo): pand with 170141183420855150465331762880109871103 - sym::simd_fcos => "cos", - sym::simd_fexp2 => "exp2", - sym::simd_fexp => "exp", - sym::simd_flog10 => "log10", - sym::simd_flog2 => "log2", - sym::simd_flog => "log", - sym::simd_floor => "floor", - sym::simd_fma => "fma", - sym::simd_fpowi => "__builtin_powi", - sym::simd_fpow => "pow", - sym::simd_fsin => "sin", - sym::simd_fsqrt => "sqrt", - sym::simd_round => "round", - sym::simd_trunc => "trunc", - _ => return_error!(InvalidMonomorphizationUnrecognized { span, name }) - }; + let intr_name = match name { + sym::simd_ceil => "ceil", + sym::simd_fabs => "fabs", // TODO(antoyo): pand with 170141183420855150465331762880109871103 + sym::simd_fcos => "cos", + sym::simd_fexp2 => "exp2", + sym::simd_fexp => "exp", + sym::simd_flog10 => "log10", + sym::simd_flog2 => "log2", + sym::simd_flog => "log", + sym::simd_floor => "floor", + sym::simd_fma => "fma", + sym::simd_fpowi => "__builtin_powi", + sym::simd_fpow => "pow", + sym::simd_fsin => "sin", + sym::simd_fsqrt => "sqrt", + sym::simd_round => "round", + sym::simd_trunc => "trunc", + _ => return_error!(InvalidMonomorphization::UnrecognizedIntrinsic { span, name }), + }; let builtin_name = format!("{}{}", intr_name, elem_ty_str); let funcs = bx.cx.functions.borrow(); - let function = funcs.get(&builtin_name).unwrap_or_else(|| panic!("unable to find builtin function {}", builtin_name)); + let function = funcs + .get(&builtin_name) + .unwrap_or_else(|| panic!("unable to find builtin function {}", builtin_name)); // TODO(antoyo): add platform-specific behavior here for architectures that have these // intrinsics as instructions (for instance, gpus) @@ -494,8 +483,12 @@ pub fn generic_simd_intrinsic<'a, 'gcc, 'tcx>( return simd_simple_float_intrinsic(name, in_elem, in_ty, in_len, bx, span, args); } - #[cfg(feature="master")] - fn vector_ty<'gcc, 'tcx>(cx: &CodegenCx<'gcc, 'tcx>, elem_ty: Ty<'tcx>, vec_len: u64) -> Type<'gcc> { + #[cfg(feature = "master")] + fn vector_ty<'gcc, 'tcx>( + cx: &CodegenCx<'gcc, 'tcx>, + elem_ty: Ty<'tcx>, + vec_len: u64, + ) -> Type<'gcc> { // FIXME: use cx.layout_of(ty).llvm_type() ? let elem_ty = match *elem_ty.kind() { ty::Int(v) => cx.type_int_from_ty(v), @@ -506,15 +499,22 @@ pub fn generic_simd_intrinsic<'a, 'gcc, 'tcx>( cx.type_vector(elem_ty, vec_len) } - #[cfg(feature="master")] - fn gather<'a, 'gcc, 'tcx>(default: RValue<'gcc>, pointers: RValue<'gcc>, mask: RValue<'gcc>, pointer_count: usize, bx: &mut Builder<'a, 'gcc, 'tcx>, in_len: u64, underlying_ty: Ty<'tcx>, invert: bool) -> RValue<'gcc> { - let vector_type = - if pointer_count > 1 { - bx.context.new_vector_type(bx.usize_type, in_len) - } - else { - vector_ty(bx, underlying_ty, in_len) - }; + #[cfg(feature = "master")] + fn gather<'a, 'gcc, 'tcx>( + default: RValue<'gcc>, + pointers: RValue<'gcc>, + mask: RValue<'gcc>, + pointer_count: usize, + bx: &mut Builder<'a, 'gcc, 'tcx>, + in_len: u64, + underlying_ty: Ty<'tcx>, + invert: bool, + ) -> RValue<'gcc> { + let vector_type = if pointer_count > 1 { + bx.context.new_vector_type(bx.usize_type, in_len) + } else { + vector_ty(bx, underlying_ty, in_len) + }; let elem_type = vector_type.dyncast_vector().expect("vector type").get_element_type(); let mut values = vec![]; @@ -545,13 +545,12 @@ pub fn generic_simd_intrinsic<'a, 'gcc, 'tcx>( if invert { bx.shuffle_vector(vector, default, mask) - } - else { + } else { bx.shuffle_vector(default, vector, mask) } } - #[cfg(feature="master")] + #[cfg(feature = "master")] if name == sym::simd_gather { // simd_gather(values: <N x T>, pointers: <N x *_ T>, // mask: <N x i{M}>) -> <N x T> @@ -560,10 +559,16 @@ pub fn generic_simd_intrinsic<'a, 'gcc, 'tcx>( // * M: any integer width is supported, will be truncated to i1 // All types must be simd vector types - require_simd!(in_ty, "first"); - require_simd!(arg_tys[1], "second"); - require_simd!(arg_tys[2], "third"); - require_simd!(ret_ty, "return"); + require_simd!(in_ty, InvalidMonomorphization::SimdFirst { span, name, ty: in_ty }); + require_simd!( + arg_tys[1], + InvalidMonomorphization::SimdSecond { span, name, ty: arg_tys[1] } + ); + require_simd!( + arg_tys[2], + InvalidMonomorphization::SimdThird { span, name, ty: arg_tys[2] } + ); + require_simd!(ret_ty, InvalidMonomorphization::SimdReturn { span, name, ty: ret_ty }); // Of the same length: let (out_len, _) = arg_tys[1].simd_size_and_type(bx.tcx()); @@ -656,10 +661,19 @@ pub fn generic_simd_intrinsic<'a, 'gcc, 'tcx>( } } - return Ok(gather(args[0].immediate(), args[1].immediate(), args[2].immediate(), pointer_count, bx, in_len, underlying_ty, false)); + return Ok(gather( + args[0].immediate(), + args[1].immediate(), + args[2].immediate(), + pointer_count, + bx, + in_len, + underlying_ty, + false, + )); } - #[cfg(feature="master")] + #[cfg(feature = "master")] if name == sym::simd_scatter { // simd_scatter(values: <N x T>, pointers: <N x *mut T>, // mask: <N x i{M}>) -> () @@ -668,9 +682,15 @@ pub fn generic_simd_intrinsic<'a, 'gcc, 'tcx>( // * M: any integer width is supported, will be truncated to i1 // All types must be simd vector types - require_simd!(in_ty, "first"); - require_simd!(arg_tys[1], "second"); - require_simd!(arg_tys[2], "third"); + require_simd!(in_ty, InvalidMonomorphization::SimdFirst { span, name, ty: in_ty }); + require_simd!( + arg_tys[1], + InvalidMonomorphization::SimdSecond { span, name, ty: arg_tys[1] } + ); + require_simd!( + arg_tys[2], + InvalidMonomorphization::SimdThird { span, name, ty: arg_tys[2] } + ); // Of the same length: let (element_len1, _) = arg_tys[1].simd_size_and_type(bx.tcx()); @@ -759,17 +779,24 @@ pub fn generic_simd_intrinsic<'a, 'gcc, 'tcx>( } } - let result = gather(args[0].immediate(), args[1].immediate(), args[2].immediate(), pointer_count, bx, in_len, underlying_ty, true); + let result = gather( + args[0].immediate(), + args[1].immediate(), + args[2].immediate(), + pointer_count, + bx, + in_len, + underlying_ty, + true, + ); let pointers = args[1].immediate(); - let vector_type = - if pointer_count > 1 { - bx.context.new_vector_type(bx.usize_type, in_len) - } - else { - vector_ty(bx, underlying_ty, in_len) - }; + let vector_type = if pointer_count > 1 { + bx.context.new_vector_type(bx.usize_type, in_len) + } else { + vector_ty(bx, underlying_ty, in_len) + }; let elem_type = vector_type.dyncast_vector().expect("vector type").get_element_type(); for i in 0..in_len { @@ -809,7 +836,7 @@ pub fn generic_simd_intrinsic<'a, 'gcc, 'tcx>( })* _ => {}, } - return_error!(InvalidMonomorphizationUnsupportedOperation { span, name, in_ty, in_elem }) + return_error!(InvalidMonomorphization::UnsupportedOperation { span, name, in_ty, in_elem }) })* } } @@ -824,89 +851,97 @@ pub fn generic_simd_intrinsic<'a, 'gcc, 'tcx>( let rhs = args[1].immediate(); let is_add = name == sym::simd_saturating_add; let ptr_bits = bx.tcx().data_layout.pointer_size.bits() as _; - let (signed, elem_width, elem_ty) = - match *in_elem.kind() { - ty::Int(i) => (true, i.bit_width().unwrap_or(ptr_bits) / 8, bx.cx.type_int_from_ty(i)), - ty::Uint(i) => (false, i.bit_width().unwrap_or(ptr_bits) / 8, bx.cx.type_uint_from_ty(i)), - _ => { - return_error!(InvalidMonomorphizationExpectedSignedUnsigned { + let (signed, elem_width, elem_ty) = match *in_elem.kind() { + ty::Int(i) => (true, i.bit_width().unwrap_or(ptr_bits) / 8, bx.cx.type_int_from_ty(i)), + ty::Uint(i) => { + (false, i.bit_width().unwrap_or(ptr_bits) / 8, bx.cx.type_uint_from_ty(i)) + } + _ => { + return_error!(InvalidMonomorphization::ExpectedVectorElementType { span, name, - elem_ty: arg_tys[0].simd_size_and_type(bx.tcx()).1, - vec_ty: arg_tys[0], + expected_element: arg_tys[0].simd_size_and_type(bx.tcx()).1, + vector_type: arg_tys[0], }); } }; - let result = - match (signed, is_add) { - (false, true) => { - let res = lhs + rhs; - let cmp = bx.context.new_comparison(None, ComparisonOp::LessThan, res, lhs); - res | cmp - }, - (true, true) => { - // Algorithm from: https://codereview.stackexchange.com/questions/115869/saturated-signed-addition - // TODO(antoyo): improve using conditional operators if possible. - let arg_type = lhs.get_type(); - // TODO(antoyo): convert lhs and rhs to unsigned. - let sum = lhs + rhs; - let vector_type = arg_type.dyncast_vector().expect("vector type"); - let unit = vector_type.get_num_units(); - let a = bx.context.new_rvalue_from_int(elem_ty, ((elem_width as i32) << 3) - 1); - let width = bx.context.new_rvalue_from_vector(None, lhs.get_type(), &vec![a; unit]); - - let xor1 = lhs ^ rhs; - let xor2 = lhs ^ sum; - let and = bx.context.new_unary_op(None, UnaryOp::BitwiseNegate, arg_type, xor1) & xor2; - let mask = and >> width; - - let one = bx.context.new_rvalue_one(elem_ty); - let ones = bx.context.new_rvalue_from_vector(None, lhs.get_type(), &vec![one; unit]); - let shift1 = ones << width; - let shift2 = sum >> width; - let mask_min = shift1 ^ shift2; - - let and1 = bx.context.new_unary_op(None, UnaryOp::BitwiseNegate, arg_type, mask) & sum; - let and2 = mask & mask_min; - - and1 + and2 - }, - (false, false) => { - let res = lhs - rhs; - let cmp = bx.context.new_comparison(None, ComparisonOp::LessThanEquals, res, lhs); - res & cmp - }, - (true, false) => { - let arg_type = lhs.get_type(); - // TODO(antoyo): this uses the same algorithm from saturating add, but add the - // negative of the right operand. Find a proper subtraction algorithm. - let rhs = bx.context.new_unary_op(None, UnaryOp::Minus, arg_type, rhs); - - // TODO(antoyo): convert lhs and rhs to unsigned. - let sum = lhs + rhs; - let vector_type = arg_type.dyncast_vector().expect("vector type"); - let unit = vector_type.get_num_units(); - let a = bx.context.new_rvalue_from_int(elem_ty, ((elem_width as i32) << 3) - 1); - let width = bx.context.new_rvalue_from_vector(None, lhs.get_type(), &vec![a; unit]); - - let xor1 = lhs ^ rhs; - let xor2 = lhs ^ sum; - let and = bx.context.new_unary_op(None, UnaryOp::BitwiseNegate, arg_type, xor1) & xor2; - let mask = and >> width; - - let one = bx.context.new_rvalue_one(elem_ty); - let ones = bx.context.new_rvalue_from_vector(None, lhs.get_type(), &vec![one; unit]); - let shift1 = ones << width; - let shift2 = sum >> width; - let mask_min = shift1 ^ shift2; - - let and1 = bx.context.new_unary_op(None, UnaryOp::BitwiseNegate, arg_type, mask) & sum; - let and2 = mask & mask_min; - - and1 + and2 - } - }; + let result = match (signed, is_add) { + (false, true) => { + let res = lhs + rhs; + let cmp = bx.context.new_comparison(None, ComparisonOp::LessThan, res, lhs); + res | cmp + } + (true, true) => { + // Algorithm from: https://codereview.stackexchange.com/questions/115869/saturated-signed-addition + // TODO(antoyo): improve using conditional operators if possible. + // TODO(antoyo): dyncast_vector should not require a call to unqualified. + let arg_type = lhs.get_type().unqualified(); + // TODO(antoyo): convert lhs and rhs to unsigned. + let sum = lhs + rhs; + let vector_type = arg_type.dyncast_vector().expect("vector type"); + let unit = vector_type.get_num_units(); + let a = bx.context.new_rvalue_from_int(elem_ty, ((elem_width as i32) << 3) - 1); + let width = bx.context.new_rvalue_from_vector(None, lhs.get_type(), &vec![a; unit]); + + let xor1 = lhs ^ rhs; + let xor2 = lhs ^ sum; + let and = + bx.context.new_unary_op(None, UnaryOp::BitwiseNegate, arg_type, xor1) & xor2; + let mask = and >> width; + + let one = bx.context.new_rvalue_one(elem_ty); + let ones = + bx.context.new_rvalue_from_vector(None, lhs.get_type(), &vec![one; unit]); + let shift1 = ones << width; + let shift2 = sum >> width; + let mask_min = shift1 ^ shift2; + + let and1 = + bx.context.new_unary_op(None, UnaryOp::BitwiseNegate, arg_type, mask) & sum; + let and2 = mask & mask_min; + + and1 + and2 + } + (false, false) => { + let res = lhs - rhs; + let cmp = bx.context.new_comparison(None, ComparisonOp::LessThanEquals, res, lhs); + res & cmp + } + (true, false) => { + // TODO(antoyo): dyncast_vector should not require a call to unqualified. + let arg_type = lhs.get_type().unqualified(); + // TODO(antoyo): this uses the same algorithm from saturating add, but add the + // negative of the right operand. Find a proper subtraction algorithm. + let rhs = bx.context.new_unary_op(None, UnaryOp::Minus, arg_type, rhs); + + // TODO(antoyo): convert lhs and rhs to unsigned. + let sum = lhs + rhs; + let vector_type = arg_type.dyncast_vector().expect("vector type"); + let unit = vector_type.get_num_units(); + let a = bx.context.new_rvalue_from_int(elem_ty, ((elem_width as i32) << 3) - 1); + let width = bx.context.new_rvalue_from_vector(None, lhs.get_type(), &vec![a; unit]); + + let xor1 = lhs ^ rhs; + let xor2 = lhs ^ sum; + let and = + bx.context.new_unary_op(None, UnaryOp::BitwiseNegate, arg_type, xor1) & xor2; + let mask = and >> width; + + let one = bx.context.new_rvalue_one(elem_ty); + let ones = + bx.context.new_rvalue_from_vector(None, lhs.get_type(), &vec![one; unit]); + let shift1 = ones << width; + let shift2 = sum >> width; + let mask_min = shift1 ^ shift2; + + let and1 = + bx.context.new_unary_op(None, UnaryOp::BitwiseNegate, arg_type, mask) & sum; + let and2 = mask & mask_min; + + and1 + and2 + } + }; return Ok(result); } @@ -917,7 +952,7 @@ pub fn generic_simd_intrinsic<'a, 'gcc, 'tcx>( if name == sym::$name { require!( ret_ty == in_elem, - InvalidMonomorphizationReturnType { span, name, in_elem, in_ty, ret_ty } + InvalidMonomorphization::ReturnType { span, name, in_elem, in_ty, ret_ty } ); return match in_elem.kind() { ty::Int(_) | ty::Uint(_) => { @@ -939,11 +974,12 @@ pub fn generic_simd_intrinsic<'a, 'gcc, 'tcx>( Ok(bx.vector_reduce_op(args[0].immediate(), $vec_op)) } } - _ => return_error!(InvalidMonomorphizationUnsupportedElement { + _ => return_error!(InvalidMonomorphization::UnsupportedSymbol { span, name, + symbol: sym::$name, in_ty, - elem_ty: in_elem, + in_elem, ret_ty }), }; @@ -980,18 +1016,24 @@ pub fn generic_simd_intrinsic<'a, 'gcc, 'tcx>( 1.0 ); - macro_rules! minmax_red { ($name:ident: $int_red:ident, $float_red:ident) => { if name == sym::$name { require!( ret_ty == in_elem, - InvalidMonomorphizationReturnType { span, name, in_elem, in_ty, ret_ty } + InvalidMonomorphization::ReturnType { span, name, in_elem, in_ty, ret_ty } ); return match in_elem.kind() { ty::Int(_) | ty::Uint(_) => Ok(bx.$int_red(args[0].immediate())), ty::Float(_) => Ok(bx.$float_red(args[0].immediate())), - _ => return_error!(InvalidMonomorphizationUnsupportedElement { span, name, in_ty, elem_ty: in_elem, ret_ty }), + _ => return_error!(InvalidMonomorphization::UnsupportedSymbol { + span, + name, + symbol: sym::$name, + in_ty, + in_elem, + ret_ty + }), }; } }; @@ -1009,17 +1051,18 @@ pub fn generic_simd_intrinsic<'a, 'gcc, 'tcx>( let input = if !$boolean { require!( ret_ty == in_elem, - InvalidMonomorphizationReturnType { span, name, in_elem, in_ty, ret_ty } + InvalidMonomorphization::ReturnType { span, name, in_elem, in_ty, ret_ty } ); args[0].immediate() } else { match in_elem.kind() { ty::Int(_) | ty::Uint(_) => {} - _ => return_error!(InvalidMonomorphizationUnsupportedElement { + _ => return_error!(InvalidMonomorphization::UnsupportedSymbol { span, name, + symbol: sym::$name, in_ty, - elem_ty: in_elem, + in_elem, ret_ty }), } @@ -1029,13 +1072,22 @@ pub fn generic_simd_intrinsic<'a, 'gcc, 'tcx>( return match in_elem.kind() { ty::Int(_) | ty::Uint(_) => { let r = bx.vector_reduce_op(input, $op); - Ok(if !$boolean { r } else { bx.icmp(IntPredicate::IntNE, r, bx.context.new_rvalue_zero(r.get_type())) }) + Ok(if !$boolean { + r + } else { + bx.icmp( + IntPredicate::IntNE, + r, + bx.context.new_rvalue_zero(r.get_type()), + ) + }) } - _ => return_error!(InvalidMonomorphizationUnsupportedElement { + _ => return_error!(InvalidMonomorphization::UnsupportedSymbol { span, name, + symbol: sym::$name, in_ty, - elem_ty: in_elem, + in_elem, ret_ty }), }; diff --git a/compiler/rustc_codegen_gcc/src/lib.rs b/compiler/rustc_codegen_gcc/src/lib.rs index 1a20dbcebd4..9c18fc4a0dc 100644 --- a/compiler/rustc_codegen_gcc/src/lib.rs +++ b/compiler/rustc_codegen_gcc/src/lib.rs @@ -2,6 +2,12 @@ * TODO(antoyo): implement equality in libgccjit based on https://zpz.github.io/blog/overloading-equality-operator-in-cpp-class-hierarchy/ (for type equality?) * TODO(antoyo): support #[inline] attributes. * TODO(antoyo): support LTO (gcc's equivalent to Full LTO is -flto -flto-partition=one — https://documentation.suse.com/sbp/all/html/SBP-GCC-10/index.html). + * For Thin LTO, this might be helpful: + * In gcc 4.6 -fwhopr was removed and became default with -flto. The non-whopr path can still be executed via -flto-partition=none. + * + * Maybe some missing optizations enabled by rustc's LTO is in there: https://gcc.gnu.org/onlinedocs/gcc/Optimize-Options.html + * Like -fipa-icf (should be already enabled) and maybe -fdevirtualize-at-ltrans. + * TODO: disable debug info always being emitted. Perhaps this slows down things? * * TODO(antoyo): remove the patches. */ @@ -28,6 +34,7 @@ extern crate rustc_codegen_ssa; extern crate rustc_data_structures; extern crate rustc_errors; extern crate rustc_fluent_macro; +extern crate rustc_fs_util; extern crate rustc_hir; extern crate rustc_macros; extern crate rustc_metadata; @@ -35,7 +42,8 @@ extern crate rustc_middle; extern crate rustc_session; extern crate rustc_span; extern crate rustc_target; -extern crate tempfile; +#[macro_use] +extern crate tracing; // This prevents duplicating functions and statics that are already part of the host rustc process. #[allow(unused_extern_crates)] @@ -57,6 +65,7 @@ mod coverageinfo; mod debuginfo; mod declare; mod errors; +mod gcc_util; mod int; mod intrinsic; mod mono_item; @@ -64,30 +73,44 @@ mod type_; mod type_of; use std::any::Any; -use std::sync::{Arc, Mutex}; - -use crate::errors::LTONotSupported; -use gccjit::{Context, OptimizationLevel, CType}; +use std::fmt::Debug; +use std::sync::Arc; +use std::sync::Mutex; +#[cfg(not(feature="master"))] +use std::sync::atomic::AtomicBool; +#[cfg(not(feature="master"))] +use std::sync::atomic::Ordering; + +use gccjit::{Context, OptimizationLevel}; +#[cfg(feature="master")] +use gccjit::TargetInfo; +#[cfg(not(feature="master"))] +use gccjit::CType; +use errors::LTONotSupported; use rustc_ast::expand::allocator::AllocatorKind; use rustc_codegen_ssa::{CodegenResults, CompiledModule, ModuleCodegen}; use rustc_codegen_ssa::base::codegen_crate; -use rustc_codegen_ssa::back::write::{CodegenContext, FatLTOInput, ModuleConfig, TargetMachineFactoryFn}; +use rustc_codegen_ssa::back::write::{CodegenContext, FatLtoInput, ModuleConfig, TargetMachineFactoryFn}; use rustc_codegen_ssa::back::lto::{LtoModuleCodegen, SerializedModule, ThinModule}; use rustc_codegen_ssa::target_features::supported_target_features; -use rustc_codegen_ssa::traits::{CodegenBackend, ExtraBackendMethods, ModuleBufferMethods, ThinBufferMethods, WriteBackendMethods}; -use rustc_data_structures::fx::FxHashMap; +use rustc_data_structures::fx::FxIndexMap; +use rustc_data_structures::sync::IntoDynSyncSend; +use rustc_codegen_ssa::traits::{CodegenBackend, ExtraBackendMethods, ThinBufferMethods, WriteBackendMethods}; use rustc_errors::{DiagnosticMessage, ErrorGuaranteed, Handler, SubdiagnosticMessage}; use rustc_fluent_macro::fluent_messages; use rustc_metadata::EncodedMetadata; use rustc_middle::dep_graph::{WorkProduct, WorkProductId}; +use rustc_middle::util::Providers; use rustc_middle::ty::TyCtxt; -use rustc_middle::ty::query::Providers; use rustc_session::config::{Lto, OptLevel, OutputFilenames}; use rustc_session::Session; use rustc_span::Symbol; use rustc_span::fatal_error::FatalError; use tempfile::TempDir; +use crate::back::lto::ModuleBuffer; +use crate::gcc_util::target_cpu; + fluent_messages! { "../messages.ftl" } pub struct PrintOnPanic<F: Fn() -> String>(pub F); @@ -100,9 +123,47 @@ impl<F: Fn() -> String> Drop for PrintOnPanic<F> { } } +#[cfg(not(feature="master"))] +#[derive(Debug)] +pub struct TargetInfo { + supports_128bit_integers: AtomicBool, +} + +#[cfg(not(feature="master"))] +impl TargetInfo { + fn cpu_supports(&self, _feature: &str) -> bool { + false + } + + fn supports_128bit_int(&self) -> bool { + self.supports_128bit_integers.load(Ordering::SeqCst) + } +} + +#[derive(Clone)] +pub struct LockedTargetInfo { + info: Arc<Mutex<IntoDynSyncSend<TargetInfo>>>, +} + +impl Debug for LockedTargetInfo { + fn fmt(&self, formatter: &mut std::fmt::Formatter<'_>) -> std::fmt::Result { + self.info.lock().expect("lock").fmt(formatter) + } +} + +impl LockedTargetInfo { + fn cpu_supports(&self, feature: &str) -> bool { + self.info.lock().expect("lock").cpu_supports(feature) + } + + fn supports_128bit_int(&self) -> bool { + self.info.lock().expect("lock").supports_128bit_int() + } +} + #[derive(Clone)] pub struct GccCodegenBackend { - supports_128bit_integers: Arc<Mutex<bool>>, + target_info: LockedTargetInfo, } impl CodegenBackend for GccCodegenBackend { @@ -111,23 +172,41 @@ impl CodegenBackend for GccCodegenBackend { } fn init(&self, sess: &Session) { - if sess.lto() != Lto::No { + #[cfg(feature="master")] + { + let target_cpu = target_cpu(sess); + + // Get the second TargetInfo with the correct CPU features by setting the arch. + let context = Context::default(); + if target_cpu != "generic" { + context.add_command_line_option(&format!("-march={}", target_cpu)); + } + + **self.target_info.info.lock().expect("lock") = context.get_target_info(); + } + + #[cfg(feature="master")] + gccjit::set_global_personality_function_name(b"rust_eh_personality\0"); + if sess.lto() == Lto::Thin { sess.emit_warning(LTONotSupported {}); } - let temp_dir = TempDir::new().expect("cannot create temporary directory"); - let temp_file = temp_dir.into_path().join("result.asm"); - let check_context = Context::default(); - check_context.set_print_errors_to_stderr(false); - let _int128_ty = check_context.new_c_type(CType::UInt128t); - // NOTE: we cannot just call compile() as this would require other files than libgccjit.so. - check_context.compile_to_file(gccjit::OutputKind::Assembler, temp_file.to_str().expect("path to str")); - *self.supports_128bit_integers.lock().expect("lock") = check_context.get_last_error() == Ok(None); + #[cfg(not(feature="master"))] + { + let temp_dir = TempDir::new().expect("cannot create temporary directory"); + let temp_file = temp_dir.into_path().join("result.asm"); + let check_context = Context::default(); + check_context.set_print_errors_to_stderr(false); + let _int128_ty = check_context.new_c_type(CType::UInt128t); + // NOTE: we cannot just call compile() as this would require other files than libgccjit.so. + check_context.compile_to_file(gccjit::OutputKind::Assembler, temp_file.to_str().expect("path to str")); + self.target_info.info.lock().expect("lock").supports_128bit_integers.store(check_context.get_last_error() == Ok(None), Ordering::SeqCst); + } } fn provide(&self, providers: &mut Providers) { - // FIXME(antoyo) compute list of enabled features from cli flags - providers.global_backend_features = |_tcx, ()| vec![]; + providers.global_backend_features = + |tcx, ()| gcc_util::global_gcc_features(tcx.sess, true) } fn codegen_crate<'tcx>(&self, tcx: TyCtxt<'tcx>, metadata: EncodedMetadata, need_metadata_module: bool) -> Box<dyn Any> { @@ -137,7 +216,7 @@ impl CodegenBackend for GccCodegenBackend { Box::new(res) } - fn join_codegen(&self, ongoing_codegen: Box<dyn Any>, sess: &Session, _outputs: &OutputFilenames) -> Result<(CodegenResults, FxHashMap<WorkProductId, WorkProduct>), ErrorGuaranteed> { + fn join_codegen(&self, ongoing_codegen: Box<dyn Any>, sess: &Session, _outputs: &OutputFilenames) -> Result<(CodegenResults, FxIndexMap<WorkProductId, WorkProduct>), ErrorGuaranteed> { let (codegen_results, work_products) = ongoing_codegen .downcast::<rustc_codegen_ssa::back::write::OngoingCodegen<GccCodegenBackend>>() .expect("Expected GccCodegenBackend's OngoingCodegen, found Box<Any>") @@ -158,21 +237,26 @@ impl CodegenBackend for GccCodegenBackend { } fn target_features(&self, sess: &Session, allow_unstable: bool) -> Vec<Symbol> { - target_features(sess, allow_unstable) + target_features(sess, allow_unstable, &self.target_info) } } impl ExtraBackendMethods for GccCodegenBackend { - fn codegen_allocator<'tcx>(&self, tcx: TyCtxt<'tcx>, module_name: &str, kind: AllocatorKind) -> Self::Module { + fn codegen_allocator<'tcx>(&self, tcx: TyCtxt<'tcx>, module_name: &str, kind: AllocatorKind, alloc_error_handler_kind: AllocatorKind) -> Self::Module { let mut mods = GccContext { context: Context::default(), + should_combine_object_files: false, + temp_dir: None, }; - unsafe { allocator::codegen(tcx, &mut mods, module_name, kind); } + + // TODO(antoyo): only set for x86. + mods.context.add_command_line_option("-masm=intel"); + unsafe { allocator::codegen(tcx, &mut mods, module_name, kind, alloc_error_handler_kind); } mods } fn compile_codegen_unit(&self, tcx: TyCtxt<'_>, cgu_name: Symbol) -> (ModuleCodegen<Self::Module>, u64) { - base::compile_codegen_unit(tcx, cgu_name, *self.supports_128bit_integers.lock().expect("lock")) + base::compile_codegen_unit(tcx, cgu_name, self.target_info.clone()) } fn target_machine_factory(&self, _sess: &Session, _opt_level: OptLevel, _features: &[String]) -> TargetMachineFactoryFn<Self> { @@ -183,14 +267,6 @@ impl ExtraBackendMethods for GccCodegenBackend { } } -pub struct ModuleBuffer; - -impl ModuleBufferMethods for ModuleBuffer { - fn data(&self) -> &[u8] { - unimplemented!(); - } -} - pub struct ThinBuffer; impl ThinBufferMethods for ThinBuffer { @@ -201,6 +277,9 @@ impl ThinBufferMethods for ThinBuffer { pub struct GccContext { context: Context<'static>, + should_combine_object_files: bool, + // Temporary directory used by LTO. We keep it here so that it's not removed before linking. + temp_dir: Option<TempDir>, } unsafe impl Send for GccContext {} @@ -215,18 +294,8 @@ impl WriteBackendMethods for GccCodegenBackend { type ThinData = (); type ThinBuffer = ThinBuffer; - fn run_fat_lto(_cgcx: &CodegenContext<Self>, mut modules: Vec<FatLTOInput<Self>>, _cached_modules: Vec<(SerializedModule<Self::ModuleBuffer>, WorkProduct)>) -> Result<LtoModuleCodegen<Self>, FatalError> { - // TODO(antoyo): implement LTO by sending -flto to libgccjit and adding the appropriate gcc linker plugins. - // NOTE: implemented elsewhere. - // TODO(antoyo): what is implemented elsewhere ^ ? - let module = - match modules.remove(0) { - FatLTOInput::InMemory(module) => module, - FatLTOInput::Serialized { .. } => { - unimplemented!(); - } - }; - Ok(LtoModuleCodegen::Fat { module, _serialized_bitcode: vec![] }) + fn run_fat_lto(cgcx: &CodegenContext<Self>, modules: Vec<FatLtoInput<Self>>, cached_modules: Vec<(SerializedModule<Self::ModuleBuffer>, WorkProduct)>) -> Result<LtoModuleCodegen<Self>, FatalError> { + back::lto::run_fat(cgcx, modules, cached_modules) } fn run_thin_lto(_cgcx: &CodegenContext<Self>, _modules: Vec<(String, Self::ThinBuffer)>, _cached_modules: Vec<(SerializedModule<Self::ModuleBuffer>, WorkProduct)>) -> Result<(Vec<LtoModuleCodegen<Self>>, Vec<WorkProduct>), FatalError> { @@ -237,6 +306,10 @@ impl WriteBackendMethods for GccCodegenBackend { unimplemented!(); } + fn print_statistics(&self) { + unimplemented!() + } + unsafe fn optimize(_cgcx: &CodegenContext<Self>, _diag_handler: &Handler, module: &ModuleCodegen<Self::Module>, config: &ModuleConfig) -> Result<(), FatalError> { module.module_llvm.context.set_optimization_level(to_gcc_opt_level(config.opt_level)); Ok(()) @@ -271,8 +344,19 @@ impl WriteBackendMethods for GccCodegenBackend { /// This is the entrypoint for a hot plugged rustc_codegen_gccjit #[no_mangle] pub fn __rustc_codegen_backend() -> Box<dyn CodegenBackend> { + #[cfg(feature="master")] + let info = { + // Check whether the target supports 128-bit integers. + let context = Context::default(); + Arc::new(Mutex::new(IntoDynSyncSend(context.get_target_info()))) + }; + #[cfg(not(feature="master"))] + let info = Arc::new(Mutex::new(IntoDynSyncSend(TargetInfo { + supports_128bit_integers: AtomicBool::new(false), + }))); + Box::new(GccCodegenBackend { - supports_128bit_integers: Arc::new(Mutex::new(false)), + target_info: LockedTargetInfo { info }, }) } @@ -291,22 +375,7 @@ fn to_gcc_opt_level(optlevel: Option<OptLevel>) -> OptimizationLevel { } } -fn handle_native(name: &str) -> &str { - if name != "native" { - return name; - } - - unimplemented!(); -} - -pub fn target_cpu(sess: &Session) -> &str { - match sess.opts.cg.target_cpu { - Some(ref name) => handle_native(name), - None => handle_native(sess.target.cpu.as_ref()), - } -} - -pub fn target_features(sess: &Session, allow_unstable: bool) -> Vec<Symbol> { +pub fn target_features(sess: &Session, allow_unstable: bool, target_info: &LockedTargetInfo) -> Vec<Symbol> { supported_target_features(sess) .iter() .filter_map( @@ -315,26 +384,13 @@ pub fn target_features(sess: &Session, allow_unstable: bool) -> Vec<Symbol> { }, ) .filter(|_feature| { - // TODO(antoyo): implement a way to get enabled feature in libgccjit. - // Probably using the equivalent of __builtin_cpu_supports. - // TODO(antoyo): maybe use whatever outputs the following command: - // gcc -march=native -Q --help=target - #[cfg(feature="master")] - { - // NOTE: the CPU in the CI doesn't support sse4a, so disable it to make the stdarch tests pass in the CI. - (_feature.contains("sse") || _feature.contains("avx")) && !_feature.contains("avx512") && !_feature.contains("sse4a") - } - #[cfg(not(feature="master"))] - { - false - } + target_info.cpu_supports(_feature) /* adx, aes, avx, avx2, avx512bf16, avx512bitalg, avx512bw, avx512cd, avx512dq, avx512er, avx512f, avx512ifma, avx512pf, avx512vbmi, avx512vbmi2, avx512vl, avx512vnni, avx512vp2intersect, avx512vpopcntdq, bmi1, bmi2, cmpxchg16b, ermsb, f16c, fma, fxsr, gfni, lzcnt, movbe, pclmulqdq, popcnt, rdrand, rdseed, rtm, sha, sse, sse2, sse3, sse4.1, sse4.2, sse4a, ssse3, tbm, vaes, vpclmulqdq, xsave, xsavec, xsaveopt, xsaves */ - //false }) .map(|feature| Symbol::intern(feature)) .collect() diff --git a/compiler/rustc_codegen_gcc/src/mono_item.rs b/compiler/rustc_codegen_gcc/src/mono_item.rs index c1f6340866c..3322d56513b 100644 --- a/compiler/rustc_codegen_gcc/src/mono_item.rs +++ b/compiler/rustc_codegen_gcc/src/mono_item.rs @@ -31,7 +31,7 @@ impl<'gcc, 'tcx> PreDefineMethods<'tcx> for CodegenCx<'gcc, 'tcx> { #[cfg_attr(not(feature="master"), allow(unused_variables))] fn predefine_fn(&self, instance: Instance<'tcx>, linkage: Linkage, visibility: Visibility, symbol_name: &str) { - assert!(!instance.substs.needs_infer()); + assert!(!instance.args.has_infer()); let fn_abi = self.fn_abi_of_instance(instance, ty::List::empty()); self.linkage.set(base::linkage_to_gcc(linkage)); diff --git a/compiler/rustc_codegen_gcc/src/type_.rs b/compiler/rustc_codegen_gcc/src/type_.rs index daa661f35c4..31899740514 100644 --- a/compiler/rustc_codegen_gcc/src/type_.rs +++ b/compiler/rustc_codegen_gcc/src/type_.rs @@ -54,6 +54,23 @@ impl<'gcc, 'tcx> CodegenCx<'gcc, 'tcx> { self.u128_type } + pub fn type_ptr_to(&self, ty: Type<'gcc>) -> Type<'gcc> { + ty.make_pointer() + } + + pub fn type_ptr_to_ext(&self, ty: Type<'gcc>, _address_space: AddressSpace) -> Type<'gcc> { + // TODO(antoyo): use address_space, perhaps with TYPE_ADDR_SPACE? + ty.make_pointer() + } + + pub fn type_i8p(&self) -> Type<'gcc> { + self.type_ptr_to(self.type_i8()) + } + + pub fn type_i8p_ext(&self, address_space: AddressSpace) -> Type<'gcc> { + self.type_ptr_to_ext(self.type_i8(), address_space) + } + pub fn type_pointee_for_align(&self, align: Align) -> Type<'gcc> { // FIXME(eddyb) We could find a better approximation if ity.align < align. let ity = Integer::approximate_align(self, align); @@ -149,13 +166,12 @@ impl<'gcc, 'tcx> BaseTypeMethods<'tcx> for CodegenCx<'gcc, 'tcx> { } } - fn type_ptr_to(&self, ty: Type<'gcc>) -> Type<'gcc> { - ty.make_pointer() + fn type_ptr(&self) -> Type<'gcc> { + self.type_ptr_to(self.type_void()) } - fn type_ptr_to_ext(&self, ty: Type<'gcc>, _address_space: AddressSpace) -> Type<'gcc> { - // TODO(antoyo): use address_space, perhaps with TYPE_ADDR_SPACE? - ty.make_pointer() + fn type_ptr_ext(&self, address_space: AddressSpace) -> Type<'gcc> { + self.type_ptr_to_ext(self.type_void(), address_space) } fn element_type(&self, ty: Type<'gcc>) -> Type<'gcc> { @@ -280,16 +296,4 @@ pub fn struct_fields<'gcc, 'tcx>(cx: &CodegenCx<'gcc, 'tcx>, layout: TyAndLayout } impl<'gcc, 'tcx> TypeMembershipMethods<'tcx> for CodegenCx<'gcc, 'tcx> { - fn set_type_metadata(&self, _function: RValue<'gcc>, _typeid: String) { - // Unsupported. - } - - fn typeid_metadata(&self, _typeid: String) -> RValue<'gcc> { - // Unsupported. - self.context.new_rvalue_from_int(self.int_type, 0) - } - - fn set_kcfi_type_metadata(&self, _function: RValue<'gcc>, _kcfi_typeid: u32) { - // Unsupported. - } } diff --git a/compiler/rustc_codegen_gcc/src/type_of.rs b/compiler/rustc_codegen_gcc/src/type_of.rs index 5df8c1a209d..c2eab295acd 100644 --- a/compiler/rustc_codegen_gcc/src/type_of.rs +++ b/compiler/rustc_codegen_gcc/src/type_of.rs @@ -4,7 +4,7 @@ use gccjit::{Struct, Type}; use crate::rustc_codegen_ssa::traits::{BaseTypeMethods, DerivedTypeMethods, LayoutTypeMethods}; use rustc_middle::bug; use rustc_middle::ty::{self, Ty, TypeVisitableExt}; -use rustc_middle::ty::layout::{FnAbiOf, LayoutOf, TyAndLayout}; +use rustc_middle::ty::layout::{LayoutOf, TyAndLayout}; use rustc_middle::ty::print::with_no_trimmed_paths; use rustc_target::abi::{self, Abi, Align, F32, F64, FieldsShape, Int, Integer, Pointer, PointeeInfo, Size, TyAbiInterface, Variants}; use rustc_target::abi::call::{CastTarget, FnAbi, Reg}; @@ -74,8 +74,8 @@ fn uncached_gcc_type<'gcc, 'tcx>(cx: &CodegenCx<'gcc, 'tcx>, layout: TyAndLayout Abi::ScalarPair(..) => { return cx.type_struct( &[ - layout.scalar_pair_element_gcc_type(cx, 0, false), - layout.scalar_pair_element_gcc_type(cx, 1, false), + layout.scalar_pair_element_gcc_type(cx, 0), + layout.scalar_pair_element_gcc_type(cx, 1), ], false, ); @@ -101,7 +101,7 @@ fn uncached_gcc_type<'gcc, 'tcx>(cx: &CodegenCx<'gcc, 'tcx>, layout: TyAndLayout if let (&ty::Generator(_, _, _), &Variants::Single { index }) = (layout.ty.kind(), &layout.variants) { - write!(&mut name, "::{}", ty::GeneratorSubsts::variant_name(index)).unwrap(); + write!(&mut name, "::{}", ty::GeneratorArgs::variant_name(index)).unwrap(); } Some(name) } @@ -150,7 +150,7 @@ pub trait LayoutGccExt<'tcx> { fn gcc_type<'gcc>(&self, cx: &CodegenCx<'gcc, 'tcx>) -> Type<'gcc>; fn immediate_gcc_type<'gcc>(&self, cx: &CodegenCx<'gcc, 'tcx>) -> Type<'gcc>; fn scalar_gcc_type_at<'gcc>(&self, cx: &CodegenCx<'gcc, 'tcx>, scalar: &abi::Scalar, offset: Size) -> Type<'gcc>; - fn scalar_pair_element_gcc_type<'gcc>(&self, cx: &CodegenCx<'gcc, 'tcx>, index: usize, immediate: bool) -> Type<'gcc>; + fn scalar_pair_element_gcc_type<'gcc>(&self, cx: &CodegenCx<'gcc, 'tcx>, index: usize) -> Type<'gcc>; fn gcc_field_index(&self, index: usize) -> u64; fn pointee_info_at<'gcc>(&self, cx: &CodegenCx<'gcc, 'tcx>, offset: Size) -> Option<PointeeInfo>; } @@ -159,8 +159,7 @@ impl<'tcx> LayoutGccExt<'tcx> for TyAndLayout<'tcx> { fn is_gcc_immediate(&self) -> bool { match self.abi { Abi::Scalar(_) | Abi::Vector { .. } => true, - Abi::ScalarPair(..) => false, - Abi::Uninhabited | Abi::Aggregate { .. } => self.is_zst(), + Abi::ScalarPair(..) | Abi::Uninhabited | Abi::Aggregate { .. } => false, } } @@ -183,6 +182,10 @@ impl<'tcx> LayoutGccExt<'tcx> for TyAndLayout<'tcx> { /// of that field's type - this is useful for taking the address of /// that field and ensuring the struct has the right alignment. fn gcc_type<'gcc>(&self, cx: &CodegenCx<'gcc, 'tcx>) -> Type<'gcc> { + use crate::rustc_middle::ty::layout::FnAbiOf; + // This must produce the same result for `repr(transparent)` wrappers as for the inner type! + // In other words, this should generally not look at the type at all, but only at the + // layout. if let Abi::Scalar(ref scalar) = self.abi { // Use a different cache for scalars because pointers to DSTs // can be either fat or thin (data pointers of fat pointers). @@ -191,12 +194,9 @@ impl<'tcx> LayoutGccExt<'tcx> for TyAndLayout<'tcx> { } let ty = match *self.ty.kind() { - ty::Ref(_, ty, _) | ty::RawPtr(ty::TypeAndMut { ty, .. }) => { - cx.type_ptr_to(cx.layout_of(ty).gcc_type(cx)) - } - ty::Adt(def, _) if def.is_box() => { - cx.type_ptr_to(cx.layout_of(self.ty.boxed_ty()).gcc_type(cx)) - } + // NOTE: we cannot remove this match like in the LLVM codegen because the call + // to fn_ptr_backend_type handle the on-stack attribute. + // TODO(antoyo): find a less hackish way to hande the on-stack attribute. ty::FnPtr(sig) => cx.fn_ptr_backend_type(&cx.fn_abi_of_fn_ptr(sig, ty::List::empty())), _ => self.scalar_gcc_type_at(cx, scalar, Size::ZERO), }; @@ -273,23 +273,10 @@ impl<'tcx> LayoutGccExt<'tcx> for TyAndLayout<'tcx> { } } - fn scalar_pair_element_gcc_type<'gcc>(&self, cx: &CodegenCx<'gcc, 'tcx>, index: usize, immediate: bool) -> Type<'gcc> { - // TODO(antoyo): remove llvm hack: - // HACK(eddyb) special-case fat pointers until LLVM removes - // pointee types, to avoid bitcasting every `OperandRef::deref`. - match self.ty.kind() { - ty::Ref(..) | ty::RawPtr(_) => { - return self.field(cx, index).gcc_type(cx); - } - // only wide pointer boxes are handled as pointers - // thin pointer boxes with scalar allocators are handled by the general logic below - ty::Adt(def, substs) if def.is_box() && cx.layout_of(substs.type_at(1)).is_zst() => { - let ptr_ty = cx.tcx.mk_mut_ptr(self.ty.boxed_ty()); - return cx.layout_of(ptr_ty).scalar_pair_element_gcc_type(cx, index, immediate); - } - _ => {} - } - + fn scalar_pair_element_gcc_type<'gcc>(&self, cx: &CodegenCx<'gcc, 'tcx>, index: usize) -> Type<'gcc> { + // This must produce the same result for `repr(transparent)` wrappers as for the inner type! + // In other words, this should generally not look at the type at all, but only at the + // layout. let (a, b) = match self.abi { Abi::ScalarPair(ref a, ref b) => (a, b), _ => bug!("TyAndLayout::scalar_pair_element_llty({:?}): not applicable", self), @@ -368,8 +355,8 @@ impl<'gcc, 'tcx> LayoutTypeMethods<'tcx> for CodegenCx<'gcc, 'tcx> { layout.gcc_field_index(index) } - fn scalar_pair_element_backend_type(&self, layout: TyAndLayout<'tcx>, index: usize, immediate: bool) -> Type<'gcc> { - layout.scalar_pair_element_gcc_type(self, index, immediate) + fn scalar_pair_element_backend_type(&self, layout: TyAndLayout<'tcx>, index: usize, _immediate: bool) -> Type<'gcc> { + layout.scalar_pair_element_gcc_type(self, index) } fn cast_backend_type(&self, ty: &CastTarget) -> Type<'gcc> { @@ -384,8 +371,8 @@ impl<'gcc, 'tcx> LayoutTypeMethods<'tcx> for CodegenCx<'gcc, 'tcx> { unimplemented!(); } - fn fn_decl_backend_type(&self, _fn_abi: &FnAbi<'tcx, Ty<'tcx>>) -> Type<'gcc> { - // FIXME(antoyo): return correct type. - self.type_void() + fn fn_decl_backend_type(&self, fn_abi: &FnAbi<'tcx, Ty<'tcx>>) -> Type<'gcc> { + let (return_type, param_types, variadic, _) = fn_abi.gcc_type(self); + self.context.new_function_pointer_type(None, return_type, ¶m_types, variadic) } } |