diff options
| author | Matthew Kelly <matthew.kelly2@gmail.com> | 2022-08-31 19:39:39 -0400 |
|---|---|---|
| committer | Matthew Kelly <matthew.kelly2@gmail.com> | 2022-08-31 19:39:39 -0400 |
| commit | eda2a401457ba645a32bdc5b9e7e90214e3e4e24 (patch) | |
| tree | 76c4a12cb26666f03aa37a81abe27782def16f1d /compiler/rustc_codegen_ssa/src | |
| parent | 4a443dfb8227d407ff3f0542cb6e688833708ba9 (diff) | |
| parent | 9243168fa5615ec8ebe9164c6bc2fdcccffd08b6 (diff) | |
| download | rust-eda2a401457ba645a32bdc5b9e7e90214e3e4e24.tar.gz rust-eda2a401457ba645a32bdc5b9e7e90214e3e4e24.zip | |
Merge remote-tracking branch 'origin/master' into mpk/add-long-error-message-for-E0311
Diffstat (limited to 'compiler/rustc_codegen_ssa/src')
| -rw-r--r-- | compiler/rustc_codegen_ssa/src/back/link.rs | 24 | ||||
| -rw-r--r-- | compiler/rustc_codegen_ssa/src/back/metadata.rs | 12 | ||||
| -rw-r--r-- | compiler/rustc_codegen_ssa/src/back/symbol_export.rs | 18 | ||||
| -rw-r--r-- | compiler/rustc_codegen_ssa/src/back/write.rs | 2 | ||||
| -rw-r--r-- | compiler/rustc_codegen_ssa/src/base.rs | 1 | ||||
| -rw-r--r-- | compiler/rustc_codegen_ssa/src/lib.rs | 25 | ||||
| -rw-r--r-- | compiler/rustc_codegen_ssa/src/meth.rs | 18 | ||||
| -rw-r--r-- | compiler/rustc_codegen_ssa/src/mir/analyze.rs | 6 | ||||
| -rw-r--r-- | compiler/rustc_codegen_ssa/src/mir/block.rs | 133 | ||||
| -rw-r--r-- | compiler/rustc_codegen_ssa/src/mir/intrinsic.rs | 4 | ||||
| -rw-r--r-- | compiler/rustc_codegen_ssa/src/mir/mod.rs | 12 | ||||
| -rw-r--r-- | compiler/rustc_codegen_ssa/src/traits/builder.rs | 157 |
12 files changed, 129 insertions, 283 deletions
diff --git a/compiler/rustc_codegen_ssa/src/back/link.rs b/compiler/rustc_codegen_ssa/src/back/link.rs index 7f6947e3c79..d2f2c7bf798 100644 --- a/compiler/rustc_codegen_ssa/src/back/link.rs +++ b/compiler/rustc_codegen_ssa/src/back/link.rs @@ -1958,7 +1958,6 @@ fn linker_with_args<'a>( // Upstream rust libraries are not supposed to depend on our local native // libraries as that would violate the structure of the DAG, in that // scenario they are required to link to them as well in a shared fashion. - // (The current implementation still doesn't prevent it though, see the FIXME below.) // // Note that upstream rust libraries may contain native dependencies as // well, but they also can't depend on what we just started to add to the @@ -1979,15 +1978,16 @@ fn linker_with_args<'a>( // and move this option back to the top. cmd.add_as_needed(); - // FIXME: Move this below to other native libraries - // (or alternatively link all native libraries after their respective crates). - // This change is somewhat breaking in practice due to local static libraries being linked - // as whole-archive (#85144), so removing whole-archive may be a pre-requisite. + // Local native libraries of all kinds. + // + // If `-Zlink-native-libraries=false` is set, then the assumption is that an + // external build system already has the native dependencies defined, and it + // will provide them to the linker itself. if sess.opts.unstable_opts.link_native_libraries { add_local_native_libraries(cmd, sess, codegen_results); } - // Upstream rust libraries and their non-bundled static libraries + // Upstream rust libraries and their (possibly bundled) static native libraries. add_upstream_rust_crates( cmd, sess, @@ -1997,11 +1997,11 @@ fn linker_with_args<'a>( tmpdir, ); - // Upstream dynamic native libraries linked with `#[link]` attributes at and `-l` - // command line options. - // If -Zlink-native-libraries=false is set, then the assumption is that an - // external build system already has the native dependencies defined, and it - // will provide them to the linker itself. + // Dynamic native libraries from upstream crates. + // + // FIXME: Merge this to `add_upstream_rust_crates` so that all native libraries are linked + // together with their respective upstream crates, and in their originally specified order. + // This may be slightly breaking due to our use of `--as-needed` and needs a crater run. if sess.opts.unstable_opts.link_native_libraries { add_upstream_native_libraries(cmd, sess, codegen_results); } @@ -2415,7 +2415,7 @@ fn add_upstream_rust_crates<'a>( // or is an rlib already included via some other dylib crate, the symbols from // native libs will have already been included in that dylib. // - // If -Zlink-native-libraries=false is set, then the assumption is that an + // If `-Zlink-native-libraries=false` is set, then the assumption is that an // external build system already has the native dependencies defined, and it // will provide them to the linker itself. if sess.opts.unstable_opts.link_native_libraries { diff --git a/compiler/rustc_codegen_ssa/src/back/metadata.rs b/compiler/rustc_codegen_ssa/src/back/metadata.rs index 0302c28815a..b92e146bee2 100644 --- a/compiler/rustc_codegen_ssa/src/back/metadata.rs +++ b/compiler/rustc_codegen_ssa/src/back/metadata.rs @@ -187,12 +187,12 @@ pub enum MetadataPosition { Last, } -// For rlibs we "pack" rustc metadata into a dummy object file. When rustc -// creates a dylib crate type it will pass `--whole-archive` (or the -// platform equivalent) to include all object files from an rlib into the -// final dylib itself. This causes linkers to iterate and try to include all -// files located in an archive, so if metadata is stored in an archive then -// it needs to be of a form that the linker will be able to process. +// For rlibs we "pack" rustc metadata into a dummy object file. +// +// Historically it was needed because rustc linked rlibs as whole-archive in some cases. +// In that case linkers try to include all files located in an archive, so if metadata is stored +// in an archive then it needs to be of a form that the linker is able to process. +// Now it's not clear whether metadata still needs to be wrapped into an object file or not. // // Note, though, that we don't actually want this metadata to show up in any // final output of the compiler. Instead this is purely for rustc's own diff --git a/compiler/rustc_codegen_ssa/src/back/symbol_export.rs b/compiler/rustc_codegen_ssa/src/back/symbol_export.rs index e6b6055759d..32b340832ce 100644 --- a/compiler/rustc_codegen_ssa/src/back/symbol_export.rs +++ b/compiler/rustc_codegen_ssa/src/back/symbol_export.rs @@ -103,18 +103,14 @@ fn reachable_non_generics_provider(tcx: TyCtxt<'_>, cnum: CrateNum) -> DefIdMap< } }) .map(|def_id| { - let (export_level, used) = if special_runtime_crate { - let name = tcx.symbol_name(Instance::mono(tcx, def_id.to_def_id())).name; - // We won't link right if these symbols are stripped during LTO. - let used = match name { - "rust_eh_personality" - | "rust_eh_register_frames" - | "rust_eh_unregister_frames" => true, - _ => false, - }; - (SymbolExportLevel::Rust, used) + // We won't link right if this symbol is stripped during LTO. + let name = tcx.symbol_name(Instance::mono(tcx, def_id.to_def_id())).name; + let used = name == "rust_eh_personality"; + + let export_level = if special_runtime_crate { + SymbolExportLevel::Rust } else { - (symbol_export_level(tcx, def_id.to_def_id()), false) + symbol_export_level(tcx, def_id.to_def_id()) }; let codegen_attrs = tcx.codegen_fn_attrs(def_id.to_def_id()); debug!( diff --git a/compiler/rustc_codegen_ssa/src/back/write.rs b/compiler/rustc_codegen_ssa/src/back/write.rs index 2930d09d71f..68f3b19b715 100644 --- a/compiler/rustc_codegen_ssa/src/back/write.rs +++ b/compiler/rustc_codegen_ssa/src/back/write.rs @@ -1892,7 +1892,7 @@ impl<B: ExtraBackendMethods> OngoingCodegen<B> { } }); - sess.cgu_reuse_tracker.check_expected_reuse(sess.diagnostic()); + sess.cgu_reuse_tracker.check_expected_reuse(sess); sess.abort_if_errors(); diff --git a/compiler/rustc_codegen_ssa/src/base.rs b/compiler/rustc_codegen_ssa/src/base.rs index a840b270974..4c6be3f9108 100644 --- a/compiler/rustc_codegen_ssa/src/base.rs +++ b/compiler/rustc_codegen_ssa/src/base.rs @@ -151,6 +151,7 @@ pub fn unsized_info<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>( let old_info = old_info.expect("unsized_info: missing old info for trait upcasting coercion"); if data_a.principal_def_id() == data_b.principal_def_id() { + // A NOP cast that doesn't actually change anything, should be allowed even with invalid vtables. return old_info; } diff --git a/compiler/rustc_codegen_ssa/src/lib.rs b/compiler/rustc_codegen_ssa/src/lib.rs index 1802eedf193..0faf51b062b 100644 --- a/compiler/rustc_codegen_ssa/src/lib.rs +++ b/compiler/rustc_codegen_ssa/src/lib.rs @@ -168,6 +168,13 @@ pub struct CodegenResults { pub crate_info: CrateInfo, } +pub enum CodegenErrors<'a> { + WrongFileType, + EmptyVersionNumber, + EncodingVersionMismatch { version_array: String, rlink_version: u32 }, + RustcVersionMismatch { rustc_version: String, current_version: &'a str }, +} + pub fn provide(providers: &mut Providers) { crate::back::symbol_export::provide(providers); crate::base::provide(providers); @@ -212,30 +219,34 @@ impl CodegenResults { encoder.finish() } - pub fn deserialize_rlink(data: Vec<u8>) -> Result<Self, String> { + pub fn deserialize_rlink<'a>(data: Vec<u8>) -> Result<Self, CodegenErrors<'a>> { // The Decodable machinery is not used here because it panics if the input data is invalid // and because its internal representation may change. if !data.starts_with(RLINK_MAGIC) { - return Err("The input does not look like a .rlink file".to_string()); + return Err(CodegenErrors::WrongFileType); } let data = &data[RLINK_MAGIC.len()..]; if data.len() < 4 { - return Err("The input does not contain version number".to_string()); + return Err(CodegenErrors::EmptyVersionNumber); } let mut version_array: [u8; 4] = Default::default(); version_array.copy_from_slice(&data[..4]); if u32::from_be_bytes(version_array) != RLINK_VERSION { - return Err(".rlink file was produced with encoding version {version_array}, but the current version is {RLINK_VERSION}".to_string()); + return Err(CodegenErrors::EncodingVersionMismatch { + version_array: String::from_utf8_lossy(&version_array).to_string(), + rlink_version: RLINK_VERSION, + }); } let mut decoder = MemDecoder::new(&data[4..], 0); let rustc_version = decoder.read_str(); let current_version = RUSTC_VERSION.unwrap(); if rustc_version != current_version { - return Err(format!( - ".rlink file was produced by rustc version {rustc_version}, but the current version is {current_version}." - )); + return Err(CodegenErrors::RustcVersionMismatch { + rustc_version: rustc_version.to_string(), + current_version, + }); } let codegen_results = CodegenResults::decode(&mut decoder); diff --git a/compiler/rustc_codegen_ssa/src/meth.rs b/compiler/rustc_codegen_ssa/src/meth.rs index 27d791d90a5..f8e982b7751 100644 --- a/compiler/rustc_codegen_ssa/src/meth.rs +++ b/compiler/rustc_codegen_ssa/src/meth.rs @@ -1,6 +1,6 @@ use crate::traits::*; -use rustc_middle::ty::{self, subst::GenericArgKind, ExistentialPredicate, Ty, TyCtxt}; +use rustc_middle::ty::{self, subst::GenericArgKind, Ty}; use rustc_session::config::Lto; use rustc_symbol_mangling::typeid_for_trait_ref; use rustc_target::abi::call::FnAbi; @@ -29,7 +29,7 @@ impl<'a, 'tcx> VirtualIndex { && bx.cx().sess().lto() == Lto::Fat { let typeid = - bx.typeid_metadata(typeid_for_trait_ref(bx.tcx(), get_trait_ref(bx.tcx(), ty))); + bx.typeid_metadata(typeid_for_trait_ref(bx.tcx(), expect_dyn_trait_in_self(ty))); let vtable_byte_offset = self.0 * bx.data_layout().pointer_size.bytes(); let type_checked_load = bx.type_checked_load(llvtable, vtable_byte_offset, typeid); let func = bx.extract_value(type_checked_load, 0); @@ -64,17 +64,13 @@ impl<'a, 'tcx> VirtualIndex { } } -fn get_trait_ref<'tcx>(tcx: TyCtxt<'tcx>, ty: Ty<'tcx>) -> ty::PolyExistentialTraitRef<'tcx> { +/// This takes a valid `self` receiver type and extracts the principal trait +/// ref of the type. +fn expect_dyn_trait_in_self<'tcx>(ty: Ty<'tcx>) -> ty::PolyExistentialTraitRef<'tcx> { for arg in ty.peel_refs().walk() { if let GenericArgKind::Type(ty) = arg.unpack() { - if let ty::Dynamic(trait_refs, _) = ty.kind() { - return trait_refs[0].map_bound(|trait_ref| match trait_ref { - ExistentialPredicate::Trait(tr) => tr, - ExistentialPredicate::Projection(proj) => proj.trait_ref(tcx), - ExistentialPredicate::AutoTrait(_) => { - bug!("auto traits don't have functions") - } - }); + if let ty::Dynamic(data, _) = ty.kind() { + return data.principal().expect("expected principal trait object"); } } } diff --git a/compiler/rustc_codegen_ssa/src/mir/analyze.rs b/compiler/rustc_codegen_ssa/src/mir/analyze.rs index 24da48ead63..c7617d2e464 100644 --- a/compiler/rustc_codegen_ssa/src/mir/analyze.rs +++ b/compiler/rustc_codegen_ssa/src/mir/analyze.rs @@ -266,7 +266,7 @@ pub fn cleanup_kinds(mir: &mir::Body<'_>) -> IndexVec<mir::BasicBlock, CleanupKi result: &mut IndexVec<mir::BasicBlock, CleanupKind>, mir: &mir::Body<'tcx>, ) { - for (bb, data) in mir.basic_blocks().iter_enumerated() { + for (bb, data) in mir.basic_blocks.iter_enumerated() { match data.terminator().kind { TerminatorKind::Goto { .. } | TerminatorKind::Resume @@ -296,7 +296,7 @@ pub fn cleanup_kinds(mir: &mir::Body<'_>) -> IndexVec<mir::BasicBlock, CleanupKi } fn propagate<'tcx>(result: &mut IndexVec<mir::BasicBlock, CleanupKind>, mir: &mir::Body<'tcx>) { - let mut funclet_succs = IndexVec::from_elem(None, mir.basic_blocks()); + let mut funclet_succs = IndexVec::from_elem(None, &mir.basic_blocks); let mut set_successor = |funclet: mir::BasicBlock, succ| match funclet_succs[funclet] { ref mut s @ None => { @@ -359,7 +359,7 @@ pub fn cleanup_kinds(mir: &mir::Body<'_>) -> IndexVec<mir::BasicBlock, CleanupKi } } - let mut result = IndexVec::from_elem(CleanupKind::NotCleanup, mir.basic_blocks()); + let mut result = IndexVec::from_elem(CleanupKind::NotCleanup, &mir.basic_blocks); discover_masters(&mut result, mir); propagate(&mut result, mir); diff --git a/compiler/rustc_codegen_ssa/src/mir/block.rs b/compiler/rustc_codegen_ssa/src/mir/block.rs index 3eee58d9d1c..5c67d3b6431 100644 --- a/compiler/rustc_codegen_ssa/src/mir/block.rs +++ b/compiler/rustc_codegen_ssa/src/mir/block.rs @@ -21,7 +21,7 @@ use rustc_middle::ty::{self, Instance, Ty, TypeVisitable}; use rustc_span::source_map::Span; use rustc_span::{sym, Symbol}; use rustc_symbol_mangling::typeid::typeid_for_fnabi; -use rustc_target::abi::call::{ArgAbi, FnAbi, PassMode}; +use rustc_target::abi::call::{ArgAbi, FnAbi, PassMode, Reg}; use rustc_target::abi::{self, HasDataLayout, WrappingRange}; use rustc_target::spec::abi::Abi; @@ -324,7 +324,7 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> { bx.unreachable(); return; } - let llval = match self.fn_abi.ret.mode { + let llval = match &self.fn_abi.ret.mode { PassMode::Ignore | PassMode::Indirect { .. } => { bx.ret_void(); return; @@ -339,7 +339,7 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> { } } - PassMode::Cast(cast_ty) => { + PassMode::Cast(cast_ty, _) => { let op = match self.locals[mir::RETURN_PLACE] { LocalRef::Operand(Some(op)) => op, LocalRef::Operand(None) => bug!("use of return before def"), @@ -360,7 +360,7 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> { llval } }; - let ty = bx.cast_backend_type(&cast_ty); + let ty = bx.cast_backend_type(cast_ty); let addr = bx.pointercast(llslot, bx.type_ptr_to(ty)); bx.load(ty, addr, self.fn_abi.ret.layout.align.abi) } @@ -798,58 +798,55 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> { let mut op = self.codegen_operand(&mut bx, arg); if let (0, Some(ty::InstanceDef::Virtual(_, idx))) = (i, def) { - if let Pair(..) = op.val { - // In the case of Rc<Self>, we need to explicitly pass a - // *mut RcBox<Self> with a Scalar (not ScalarPair) ABI. This is a hack - // that is understood elsewhere in the compiler as a method on - // `dyn Trait`. - // To get a `*mut RcBox<Self>`, we just keep unwrapping newtypes until - // we get a value of a built-in pointer type - 'descend_newtypes: while !op.layout.ty.is_unsafe_ptr() - && !op.layout.ty.is_region_ptr() - { - for i in 0..op.layout.fields.count() { - let field = op.extract_field(&mut bx, i); - if !field.layout.is_zst() { - // we found the one non-zero-sized field that is allowed - // now find *its* non-zero-sized field, or stop if it's a - // pointer - op = field; - continue 'descend_newtypes; + match op.val { + Pair(data_ptr, meta) => { + // In the case of Rc<Self>, we need to explicitly pass a + // *mut RcBox<Self> with a Scalar (not ScalarPair) ABI. This is a hack + // that is understood elsewhere in the compiler as a method on + // `dyn Trait`. + // To get a `*mut RcBox<Self>`, we just keep unwrapping newtypes until + // we get a value of a built-in pointer type + 'descend_newtypes: while !op.layout.ty.is_unsafe_ptr() + && !op.layout.ty.is_region_ptr() + { + for i in 0..op.layout.fields.count() { + let field = op.extract_field(&mut bx, i); + if !field.layout.is_zst() { + // we found the one non-zero-sized field that is allowed + // now find *its* non-zero-sized field, or stop if it's a + // pointer + op = field; + continue 'descend_newtypes; + } } + + span_bug!(span, "receiver has no non-zero-sized fields {:?}", op); } - span_bug!(span, "receiver has no non-zero-sized fields {:?}", op); + // now that we have `*dyn Trait` or `&dyn Trait`, split it up into its + // data pointer and vtable. Look up the method in the vtable, and pass + // the data pointer as the first argument + llfn = Some(meth::VirtualIndex::from_index(idx).get_fn( + &mut bx, + meta, + op.layout.ty, + &fn_abi, + )); + llargs.push(data_ptr); + continue 'make_args; } - - // now that we have `*dyn Trait` or `&dyn Trait`, split it up into its - // data pointer and vtable. Look up the method in the vtable, and pass - // the data pointer as the first argument - match op.val { - Pair(data_ptr, meta) => { - llfn = Some(meth::VirtualIndex::from_index(idx).get_fn( - &mut bx, - meta, - op.layout.ty, - &fn_abi, - )); - llargs.push(data_ptr); - continue 'make_args; - } - other => bug!("expected a Pair, got {:?}", other), + Ref(data_ptr, Some(meta), _) => { + // by-value dynamic dispatch + llfn = Some(meth::VirtualIndex::from_index(idx).get_fn( + &mut bx, + meta, + op.layout.ty, + &fn_abi, + )); + llargs.push(data_ptr); + continue; } - } else if let Ref(data_ptr, Some(meta), _) = op.val { - // by-value dynamic dispatch - llfn = Some(meth::VirtualIndex::from_index(idx).get_fn( - &mut bx, - meta, - op.layout.ty, - &fn_abi, - )); - llargs.push(data_ptr); - continue; - } else { - span_bug!(span, "can't codegen a virtual call on {:?}", op); + _ => span_bug!(span, "can't codegen a virtual call on {:?}", op), } } @@ -1161,39 +1158,35 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> { llargs: &mut Vec<Bx::Value>, arg: &ArgAbi<'tcx, Ty<'tcx>>, ) { - // Fill padding with undef value, where applicable. - if let Some(ty) = arg.pad { - llargs.push(bx.const_undef(bx.reg_backend_type(&ty))) - } - - if arg.is_ignore() { - return; - } - - if let PassMode::Pair(..) = arg.mode { - match op.val { + match arg.mode { + PassMode::Ignore => return, + PassMode::Cast(_, true) => { + // Fill padding with undef value, where applicable. + llargs.push(bx.const_undef(bx.reg_backend_type(&Reg::i32()))); + } + PassMode::Pair(..) => match op.val { Pair(a, b) => { llargs.push(a); llargs.push(b); return; } _ => bug!("codegen_argument: {:?} invalid for pair argument", op), - } - } else if arg.is_unsized_indirect() { - match op.val { + }, + PassMode::Indirect { attrs: _, extra_attrs: Some(_), on_stack: _ } => match op.val { Ref(a, Some(b), _) => { llargs.push(a); llargs.push(b); return; } _ => bug!("codegen_argument: {:?} invalid for unsized indirect argument", op), - } + }, + _ => {} } // Force by-ref if we have to load through a cast pointer. let (mut llval, align, by_ref) = match op.val { Immediate(_) | Pair(..) => match arg.mode { - PassMode::Indirect { .. } | PassMode::Cast(_) => { + PassMode::Indirect { .. } | PassMode::Cast(..) => { let scratch = PlaceRef::alloca(bx, arg.layout); op.val.store(bx, scratch); (scratch.llval, scratch.align, true) @@ -1225,8 +1218,8 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> { if by_ref && !arg.is_indirect() { // Have to load the argument, maybe while casting it. - if let PassMode::Cast(ty) = arg.mode { - let llty = bx.cast_backend_type(&ty); + if let PassMode::Cast(ty, _) = &arg.mode { + let llty = bx.cast_backend_type(ty); let addr = bx.pointercast(llval, bx.type_ptr_to(llty)); llval = bx.load(llty, addr, align.min(arg.layout.align.abi)); } else { @@ -1625,7 +1618,7 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> { } DirectOperand(index) => { // If there is a cast, we have to store and reload. - let op = if let PassMode::Cast(_) = ret_abi.mode { + let op = if let PassMode::Cast(..) = ret_abi.mode { let tmp = PlaceRef::alloca(bx, ret_abi.layout); tmp.storage_live(bx); bx.store_arg(&ret_abi, llval, tmp); diff --git a/compiler/rustc_codegen_ssa/src/mir/intrinsic.rs b/compiler/rustc_codegen_ssa/src/mir/intrinsic.rs index 94ac71a4dd2..16aad07194d 100644 --- a/compiler/rustc_codegen_ssa/src/mir/intrinsic.rs +++ b/compiler/rustc_codegen_ssa/src/mir/intrinsic.rs @@ -597,8 +597,8 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> { }; if !fn_abi.ret.is_ignore() { - if let PassMode::Cast(ty) = fn_abi.ret.mode { - let ptr_llty = bx.type_ptr_to(bx.cast_backend_type(&ty)); + if let PassMode::Cast(ty, _) = &fn_abi.ret.mode { + let ptr_llty = bx.type_ptr_to(bx.cast_backend_type(ty)); let ptr = bx.pointercast(result.llval, ptr_llty); bx.store(llval, ptr, result.align); } else { diff --git a/compiler/rustc_codegen_ssa/src/mir/mod.rs b/compiler/rustc_codegen_ssa/src/mir/mod.rs index 8ee375fa9e3..d6bbcd99234 100644 --- a/compiler/rustc_codegen_ssa/src/mir/mod.rs +++ b/compiler/rustc_codegen_ssa/src/mir/mod.rs @@ -150,13 +150,13 @@ pub fn codegen_mir<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>( let start_llbb = Bx::append_block(cx, llfn, "start"); let mut bx = Bx::build(cx, start_llbb); - if mir.basic_blocks().iter().any(|bb| bb.is_cleanup) { + if mir.basic_blocks.iter().any(|bb| bb.is_cleanup) { bx.set_personality_fn(cx.eh_personality()); } let cleanup_kinds = analyze::cleanup_kinds(&mir); let cached_llbbs: IndexVec<mir::BasicBlock, Option<Bx::BasicBlock>> = mir - .basic_blocks() + .basic_blocks .indices() .map(|bb| if bb == mir::START_BLOCK { Some(start_llbb) } else { None }) .collect(); @@ -172,8 +172,8 @@ pub fn codegen_mir<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>( unreachable_block: None, double_unwind_guard: None, cleanup_kinds, - landing_pads: IndexVec::from_elem(None, mir.basic_blocks()), - funclets: IndexVec::from_fn_n(|_| None, mir.basic_blocks().len()), + landing_pads: IndexVec::from_elem(None, &mir.basic_blocks), + funclets: IndexVec::from_fn_n(|_| None, mir.basic_blocks.len()), locals: IndexVec::new(), debug_context, per_local_var_debug_info: None, @@ -283,7 +283,7 @@ fn arg_local_refs<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>( for i in 0..tupled_arg_tys.len() { let arg = &fx.fn_abi.args[idx]; idx += 1; - if arg.pad.is_some() { + if let PassMode::Cast(_, true) = arg.mode { llarg_idx += 1; } let pr_field = place.project_field(bx, i); @@ -309,7 +309,7 @@ fn arg_local_refs<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>( let arg = &fx.fn_abi.args[idx]; idx += 1; - if arg.pad.is_some() { + if let PassMode::Cast(_, true) = arg.mode { llarg_idx += 1; } diff --git a/compiler/rustc_codegen_ssa/src/traits/builder.rs b/compiler/rustc_codegen_ssa/src/traits/builder.rs index 9f49749bb41..10cf8948b5a 100644 --- a/compiler/rustc_codegen_ssa/src/traits/builder.rs +++ b/compiler/rustc_codegen_ssa/src/traits/builder.rs @@ -1,6 +1,5 @@ use super::abi::AbiBuilderMethods; use super::asm::AsmBuilderMethods; -use super::consts::ConstMethods; use super::coverageinfo::CoverageInfoBuilderMethods; use super::debuginfo::DebugInfoBuilderMethods; use super::intrinsic::IntrinsicCallMethods; @@ -15,7 +14,6 @@ use crate::mir::operand::OperandRef; use crate::mir::place::PlaceRef; use crate::MemFlags; -use rustc_apfloat::{ieee, Float, Round, Status}; use rustc_middle::ty::layout::{HasParamEnv, TyAndLayout}; use rustc_middle::ty::Ty; use rustc_span::Span; @@ -188,8 +186,8 @@ pub trait BuilderMethods<'a, 'tcx>: fn trunc(&mut self, val: Self::Value, dest_ty: Self::Type) -> Self::Value; fn sext(&mut self, val: Self::Value, dest_ty: Self::Type) -> Self::Value; - fn fptoui_sat(&mut self, val: Self::Value, dest_ty: Self::Type) -> Option<Self::Value>; - fn fptosi_sat(&mut self, val: Self::Value, dest_ty: Self::Type) -> Option<Self::Value>; + fn fptoui_sat(&mut self, val: Self::Value, dest_ty: Self::Type) -> Self::Value; + fn fptosi_sat(&mut self, val: Self::Value, dest_ty: Self::Type) -> Self::Value; fn fptoui(&mut self, val: Self::Value, dest_ty: Self::Type) -> Self::Value; fn fptosi(&mut self, val: Self::Value, dest_ty: Self::Type) -> Self::Value; fn uitofp(&mut self, val: Self::Value, dest_ty: Self::Type) -> Self::Value; @@ -223,156 +221,7 @@ pub trait BuilderMethods<'a, 'tcx>: return if signed { self.fptosi(x, dest_ty) } else { self.fptoui(x, dest_ty) }; } - let try_sat_result = - if signed { self.fptosi_sat(x, dest_ty) } else { self.fptoui_sat(x, dest_ty) }; - if let Some(try_sat_result) = try_sat_result { - return try_sat_result; - } - - let int_width = self.cx().int_width(int_ty); - let float_width = self.cx().float_width(float_ty); - // LLVM's fpto[su]i returns undef when the input x is infinite, NaN, or does not fit into the - // destination integer type after rounding towards zero. This `undef` value can cause UB in - // safe code (see issue #10184), so we implement a saturating conversion on top of it: - // Semantically, the mathematical value of the input is rounded towards zero to the next - // mathematical integer, and then the result is clamped into the range of the destination - // integer type. Positive and negative infinity are mapped to the maximum and minimum value of - // the destination integer type. NaN is mapped to 0. - // - // Define f_min and f_max as the largest and smallest (finite) floats that are exactly equal to - // a value representable in int_ty. - // They are exactly equal to int_ty::{MIN,MAX} if float_ty has enough significand bits. - // Otherwise, int_ty::MAX must be rounded towards zero, as it is one less than a power of two. - // int_ty::MIN, however, is either zero or a negative power of two and is thus exactly - // representable. Note that this only works if float_ty's exponent range is sufficiently large. - // f16 or 256 bit integers would break this property. Right now the smallest float type is f32 - // with exponents ranging up to 127, which is barely enough for i128::MIN = -2^127. - // On the other hand, f_max works even if int_ty::MAX is greater than float_ty::MAX. Because - // we're rounding towards zero, we just get float_ty::MAX (which is always an integer). - // This already happens today with u128::MAX = 2^128 - 1 > f32::MAX. - let int_max = |signed: bool, int_width: u64| -> u128 { - let shift_amount = 128 - int_width; - if signed { i128::MAX as u128 >> shift_amount } else { u128::MAX >> shift_amount } - }; - let int_min = |signed: bool, int_width: u64| -> i128 { - if signed { i128::MIN >> (128 - int_width) } else { 0 } - }; - - let compute_clamp_bounds_single = |signed: bool, int_width: u64| -> (u128, u128) { - let rounded_min = - ieee::Single::from_i128_r(int_min(signed, int_width), Round::TowardZero); - assert_eq!(rounded_min.status, Status::OK); - let rounded_max = - ieee::Single::from_u128_r(int_max(signed, int_width), Round::TowardZero); - assert!(rounded_max.value.is_finite()); - (rounded_min.value.to_bits(), rounded_max.value.to_bits()) - }; - let compute_clamp_bounds_double = |signed: bool, int_width: u64| -> (u128, u128) { - let rounded_min = - ieee::Double::from_i128_r(int_min(signed, int_width), Round::TowardZero); - assert_eq!(rounded_min.status, Status::OK); - let rounded_max = - ieee::Double::from_u128_r(int_max(signed, int_width), Round::TowardZero); - assert!(rounded_max.value.is_finite()); - (rounded_min.value.to_bits(), rounded_max.value.to_bits()) - }; - // To implement saturation, we perform the following steps: - // - // 1. Cast x to an integer with fpto[su]i. This may result in undef. - // 2. Compare x to f_min and f_max, and use the comparison results to select: - // a) int_ty::MIN if x < f_min or x is NaN - // b) int_ty::MAX if x > f_max - // c) the result of fpto[su]i otherwise - // 3. If x is NaN, return 0.0, otherwise return the result of step 2. - // - // This avoids resulting undef because values in range [f_min, f_max] by definition fit into the - // destination type. It creates an undef temporary, but *producing* undef is not UB. Our use of - // undef does not introduce any non-determinism either. - // More importantly, the above procedure correctly implements saturating conversion. - // Proof (sketch): - // If x is NaN, 0 is returned by definition. - // Otherwise, x is finite or infinite and thus can be compared with f_min and f_max. - // This yields three cases to consider: - // (1) if x in [f_min, f_max], the result of fpto[su]i is returned, which agrees with - // saturating conversion for inputs in that range. - // (2) if x > f_max, then x is larger than int_ty::MAX. This holds even if f_max is rounded - // (i.e., if f_max < int_ty::MAX) because in those cases, nextUp(f_max) is already larger - // than int_ty::MAX. Because x is larger than int_ty::MAX, the return value of int_ty::MAX - // is correct. - // (3) if x < f_min, then x is smaller than int_ty::MIN. As shown earlier, f_min exactly equals - // int_ty::MIN and therefore the return value of int_ty::MIN is correct. - // QED. - - let float_bits_to_llval = |bx: &mut Self, bits| { - let bits_llval = match float_width { - 32 => bx.cx().const_u32(bits as u32), - 64 => bx.cx().const_u64(bits as u64), - n => bug!("unsupported float width {}", n), - }; - bx.bitcast(bits_llval, float_ty) - }; - let (f_min, f_max) = match float_width { - 32 => compute_clamp_bounds_single(signed, int_width), - 64 => compute_clamp_bounds_double(signed, int_width), - n => bug!("unsupported float width {}", n), - }; - let f_min = float_bits_to_llval(self, f_min); - let f_max = float_bits_to_llval(self, f_max); - let int_max = self.cx().const_uint_big(int_ty, int_max(signed, int_width)); - let int_min = self.cx().const_uint_big(int_ty, int_min(signed, int_width) as u128); - let zero = self.cx().const_uint(int_ty, 0); - - // If we're working with vectors, constants must be "splatted": the constant is duplicated - // into each lane of the vector. The algorithm stays the same, we are just using the - // same constant across all lanes. - let maybe_splat = |bx: &mut Self, val| { - if bx.cx().type_kind(dest_ty) == TypeKind::Vector { - bx.vector_splat(bx.vector_length(dest_ty), val) - } else { - val - } - }; - let f_min = maybe_splat(self, f_min); - let f_max = maybe_splat(self, f_max); - let int_max = maybe_splat(self, int_max); - let int_min = maybe_splat(self, int_min); - let zero = maybe_splat(self, zero); - - // Step 1 ... - let fptosui_result = if signed { self.fptosi(x, dest_ty) } else { self.fptoui(x, dest_ty) }; - let less_or_nan = self.fcmp(RealPredicate::RealULT, x, f_min); - let greater = self.fcmp(RealPredicate::RealOGT, x, f_max); - - // Step 2: We use two comparisons and two selects, with %s1 being the - // result: - // %less_or_nan = fcmp ult %x, %f_min - // %greater = fcmp olt %x, %f_max - // %s0 = select %less_or_nan, int_ty::MIN, %fptosi_result - // %s1 = select %greater, int_ty::MAX, %s0 - // Note that %less_or_nan uses an *unordered* comparison. This - // comparison is true if the operands are not comparable (i.e., if x is - // NaN). The unordered comparison ensures that s1 becomes int_ty::MIN if - // x is NaN. - // - // Performance note: Unordered comparison can be lowered to a "flipped" - // comparison and a negation, and the negation can be merged into the - // select. Therefore, it not necessarily any more expensive than an - // ordered ("normal") comparison. Whether these optimizations will be - // performed is ultimately up to the backend, but at least x86 does - // perform them. - let s0 = self.select(less_or_nan, int_min, fptosui_result); - let s1 = self.select(greater, int_max, s0); - - // Step 3: NaN replacement. - // For unsigned types, the above step already yielded int_ty::MIN == 0 if x is NaN. - // Therefore we only need to execute this step for signed integer types. - if signed { - // LLVM has no isNaN predicate, so we use (x == x) instead - let cmp = self.fcmp(RealPredicate::RealOEQ, x, x); - self.select(cmp, s1, zero) - } else { - s1 - } + if signed { self.fptosi_sat(x, dest_ty) } else { self.fptoui_sat(x, dest_ty) } } fn icmp(&mut self, op: IntPredicate, lhs: Self::Value, rhs: Self::Value) -> Self::Value; |
