diff options
Diffstat (limited to 'compiler')
32 files changed, 926 insertions, 750 deletions
diff --git a/compiler/rustc_codegen_llvm/src/back/command_line_args.rs b/compiler/rustc_codegen_llvm/src/back/command_line_args.rs new file mode 100644 index 00000000000..b14713969b3 --- /dev/null +++ b/compiler/rustc_codegen_llvm/src/back/command_line_args.rs @@ -0,0 +1,37 @@ +#[cfg(test)] +mod tests; + +/// Joins command-line arguments into a single space-separated string, quoting +/// and escaping individual arguments as necessary. +/// +/// The result is intended to be informational, for embedding in debug metadata, +/// and might not be properly quoted/escaped for actual command-line use. +pub(crate) fn quote_command_line_args(args: &[String]) -> String { + // Start with a decent-sized buffer, since rustc invocations tend to be long. + let mut buf = String::with_capacity(128); + + for arg in args { + if !buf.is_empty() { + buf.push(' '); + } + + print_arg_quoted(&mut buf, arg); + } + + buf +} + +/// Equivalent to LLVM's `sys::printArg` with quoting always enabled +/// (see llvm/lib/Support/Program.cpp). +fn print_arg_quoted(buf: &mut String, arg: &str) { + buf.reserve(arg.len() + 2); + + buf.push('"'); + for ch in arg.chars() { + if matches!(ch, '"' | '\\' | '$') { + buf.push('\\'); + } + buf.push(ch); + } + buf.push('"'); +} diff --git a/compiler/rustc_codegen_llvm/src/back/command_line_args/tests.rs b/compiler/rustc_codegen_llvm/src/back/command_line_args/tests.rs new file mode 100644 index 00000000000..69641fed3bc --- /dev/null +++ b/compiler/rustc_codegen_llvm/src/back/command_line_args/tests.rs @@ -0,0 +1,25 @@ +#[test] +fn quote_command_line_args() { + use super::quote_command_line_args; + + struct Case<'a> { + args: &'a [&'a str], + expected: &'a str, + } + + let cases = &[ + Case { args: &[], expected: "" }, + Case { args: &["--hello", "world"], expected: r#""--hello" "world""# }, + Case { args: &["--hello world"], expected: r#""--hello world""# }, + Case { + args: &["plain", "$dollar", "spa ce", r"back\slash", r#""quote""#, "plain"], + expected: r#""plain" "\$dollar" "spa ce" "back\\slash" "\"quote\"" "plain""#, + }, + ]; + + for &Case { args, expected } in cases { + let args = args.iter().copied().map(str::to_owned).collect::<Vec<_>>(); + let actual = quote_command_line_args(&args); + assert_eq!(actual, expected, "args {args:?}"); + } +} diff --git a/compiler/rustc_codegen_llvm/src/back/mod.rs b/compiler/rustc_codegen_llvm/src/back/mod.rs index 6cb89f80ab8..fe3883e8c73 100644 --- a/compiler/rustc_codegen_llvm/src/back/mod.rs +++ b/compiler/rustc_codegen_llvm/src/back/mod.rs @@ -1,4 +1,5 @@ pub(crate) mod archive; +mod command_line_args; pub(crate) mod lto; pub(crate) mod owned_target_machine; mod profiling; diff --git a/compiler/rustc_codegen_llvm/src/back/owned_target_machine.rs b/compiler/rustc_codegen_llvm/src/back/owned_target_machine.rs index 6d8178320fe..d5228f0e0de 100644 --- a/compiler/rustc_codegen_llvm/src/back/owned_target_machine.rs +++ b/compiler/rustc_codegen_llvm/src/back/owned_target_machine.rs @@ -1,4 +1,3 @@ -use std::assert_matches::assert_matches; use std::ffi::CStr; use std::marker::PhantomData; use std::ptr::NonNull; @@ -39,13 +38,10 @@ impl OwnedTargetMachine { output_obj_file: &CStr, debug_info_compression: &CStr, use_emulated_tls: bool, - args_cstr_buff: &[u8], + argv0: &str, + command_line_args: &str, use_wasm_eh: bool, ) -> Result<Self, LlvmError<'static>> { - // The argument list is passed as the concatenation of one or more C strings. - // This implies that there must be a last byte, and it must be 0. - assert_matches!(args_cstr_buff, [.., b'\0'], "the last byte must be a NUL terminator"); - // SAFETY: llvm::LLVMRustCreateTargetMachine copies pointed to data let tm_ptr = unsafe { llvm::LLVMRustCreateTargetMachine( @@ -70,8 +66,10 @@ impl OwnedTargetMachine { output_obj_file.as_ptr(), debug_info_compression.as_ptr(), use_emulated_tls, - args_cstr_buff.as_ptr(), - args_cstr_buff.len(), + argv0.as_ptr(), + argv0.len(), + command_line_args.as_ptr(), + command_line_args.len(), use_wasm_eh, ) }; diff --git a/compiler/rustc_codegen_llvm/src/back/write.rs b/compiler/rustc_codegen_llvm/src/back/write.rs index bda81fbd19e..c4881f0aafc 100644 --- a/compiler/rustc_codegen_llvm/src/back/write.rs +++ b/compiler/rustc_codegen_llvm/src/back/write.rs @@ -31,6 +31,7 @@ use rustc_span::{BytePos, InnerSpan, Pos, SpanData, SyntaxContext, sym}; use rustc_target::spec::{CodeModel, FloatAbi, RelocModel, SanitizerSet, SplitDebuginfo, TlsModel}; use tracing::{debug, trace}; +use crate::back::command_line_args::quote_command_line_args; use crate::back::lto::ThinBuffer; use crate::back::owned_target_machine::OwnedTargetMachine; use crate::back::profiling::{ @@ -249,23 +250,15 @@ pub(crate) fn target_machine_factory( let use_emulated_tls = matches!(sess.tls_model(), TlsModel::Emulated); - // copy the exe path, followed by path all into one buffer - // null terminating them so we can use them as null terminated strings - let args_cstr_buff = { - let mut args_cstr_buff: Vec<u8> = Vec::new(); - let exe_path = std::env::current_exe().unwrap_or_default(); - let exe_path_str = exe_path.into_os_string().into_string().unwrap_or_default(); - - args_cstr_buff.extend_from_slice(exe_path_str.as_bytes()); - args_cstr_buff.push(0); - - for arg in sess.expanded_args.iter() { - args_cstr_buff.extend_from_slice(arg.as_bytes()); - args_cstr_buff.push(0); - } - - args_cstr_buff - }; + // Command-line information to be included in the target machine. + // This seems to only be used for embedding in PDB debuginfo files. + // FIXME(Zalathar): Maybe skip this for non-PDB targets? + let argv0 = std::env::current_exe() + .unwrap_or_default() + .into_os_string() + .into_string() + .unwrap_or_default(); + let command_line_args = quote_command_line_args(&sess.expanded_args); let debuginfo_compression = sess.opts.debuginfo_compression.to_string(); match sess.opts.debuginfo_compression { @@ -323,7 +316,8 @@ pub(crate) fn target_machine_factory( &output_obj_file, &debuginfo_compression, use_emulated_tls, - &args_cstr_buff, + &argv0, + &command_line_args, use_wasm_eh, ) }) diff --git a/compiler/rustc_codegen_llvm/src/lib.rs b/compiler/rustc_codegen_llvm/src/lib.rs index 6fb23d09843..13bdb7cb1a2 100644 --- a/compiler/rustc_codegen_llvm/src/lib.rs +++ b/compiler/rustc_codegen_llvm/src/lib.rs @@ -45,6 +45,7 @@ use rustc_middle::util::Providers; use rustc_session::Session; use rustc_session::config::{OptLevel, OutputFilenames, PrintKind, PrintRequest}; use rustc_span::Symbol; +use rustc_target::spec::{RelocModel, TlsModel}; mod abi; mod allocator; @@ -244,16 +245,7 @@ impl CodegenBackend for LlvmCodegenBackend { match req.kind { PrintKind::RelocationModels => { writeln!(out, "Available relocation models:").unwrap(); - for name in &[ - "static", - "pic", - "pie", - "dynamic-no-pic", - "ropi", - "rwpi", - "ropi-rwpi", - "default", - ] { + for name in RelocModel::ALL.iter().map(RelocModel::desc).chain(["default"]) { writeln!(out, " {name}").unwrap(); } writeln!(out).unwrap(); @@ -267,9 +259,7 @@ impl CodegenBackend for LlvmCodegenBackend { } PrintKind::TlsModels => { writeln!(out, "Available TLS models:").unwrap(); - for name in - &["global-dynamic", "local-dynamic", "initial-exec", "local-exec", "emulated"] - { + for name in TlsModel::ALL.iter().map(TlsModel::desc) { writeln!(out, " {name}").unwrap(); } writeln!(out).unwrap(); diff --git a/compiler/rustc_codegen_llvm/src/llvm/ffi.rs b/compiler/rustc_codegen_llvm/src/llvm/ffi.rs index 1124ebc3d44..9a86e4373d8 100644 --- a/compiler/rustc_codegen_llvm/src/llvm/ffi.rs +++ b/compiler/rustc_codegen_llvm/src/llvm/ffi.rs @@ -2491,8 +2491,10 @@ unsafe extern "C" { OutputObjFile: *const c_char, DebugInfoCompression: *const c_char, UseEmulatedTls: bool, - ArgsCstrBuff: *const c_uchar, // See "PTR_LEN_STR". - ArgsCstrBuffLen: usize, + Argv0: *const c_uchar, // See "PTR_LEN_STR". + Argv0Len: size_t, + CommandLineArgs: *const c_uchar, // See "PTR_LEN_STR". + CommandLineArgsLen: size_t, UseWasmEH: bool, ) -> *mut TargetMachine; diff --git a/compiler/rustc_codegen_llvm/src/llvm_util.rs b/compiler/rustc_codegen_llvm/src/llvm_util.rs index 8461c8b03d5..45c5c9aa551 100644 --- a/compiler/rustc_codegen_llvm/src/llvm_util.rs +++ b/compiler/rustc_codegen_llvm/src/llvm_util.rs @@ -217,27 +217,16 @@ impl<'a> IntoIterator for LLVMFeature<'a> { /// Rust can also be build with an external precompiled version of LLVM which might lead to failures /// if the oldest tested / supported LLVM version doesn't yet support the relevant intrinsics. pub(crate) fn to_llvm_features<'a>(sess: &Session, s: &'a str) -> Option<LLVMFeature<'a>> { - let arch = if sess.target.arch == "x86_64" { - "x86" - } else if sess.target.arch == "arm64ec" { - "aarch64" - } else if sess.target.arch == "sparc64" { - "sparc" - } else if sess.target.arch == "powerpc64" { - "powerpc" - } else { - &*sess.target.arch + let raw_arch = &*sess.target.arch; + let arch = match raw_arch { + "x86_64" => "x86", + "arm64ec" => "aarch64", + "sparc64" => "sparc", + "powerpc64" => "powerpc", + _ => raw_arch, }; + let (major, _, _) = get_version(); match (arch, s) { - ("x86", "sse4.2") => Some(LLVMFeature::with_dependencies( - "sse4.2", - smallvec![TargetFeatureFoldStrength::EnableOnly("crc32")], - )), - ("x86", "pclmulqdq") => Some(LLVMFeature::new("pclmul")), - ("x86", "rdrand") => Some(LLVMFeature::new("rdrnd")), - ("x86", "bmi1") => Some(LLVMFeature::new("bmi")), - ("x86", "cmpxchg16b") => Some(LLVMFeature::new("cx16")), - ("x86", "lahfsahf") => Some(LLVMFeature::new("sahf")), ("aarch64", "rcpc2") => Some(LLVMFeature::new("rcpc-immo")), ("aarch64", "dpb") => Some(LLVMFeature::new("ccpp")), ("aarch64", "dpb2") => Some(LLVMFeature::new("ccdp")), @@ -260,14 +249,23 @@ pub(crate) fn to_llvm_features<'a>(sess: &Session, s: &'a str) -> Option<LLVMFea ("aarch64", "fpmr") => None, // only existed in 18 ("arm", "fp16") => Some(LLVMFeature::new("fullfp16")), // Filter out features that are not supported by the current LLVM version - ("loongarch32" | "loongarch64", "32s") if get_version().0 < 21 => None, + ("loongarch32" | "loongarch64", "32s") if major < 21 => None, + ("powerpc", "power8-crypto") => Some(LLVMFeature::new("crypto")), + ("sparc", "leoncasa") => Some(LLVMFeature::new("hasleoncasa")), + ("x86", "sse4.2") => Some(LLVMFeature::with_dependencies( + "sse4.2", + smallvec![TargetFeatureFoldStrength::EnableOnly("crc32")], + )), + ("x86", "pclmulqdq") => Some(LLVMFeature::new("pclmul")), + ("x86", "rdrand") => Some(LLVMFeature::new("rdrnd")), + ("x86", "bmi1") => Some(LLVMFeature::new("bmi")), + ("x86", "cmpxchg16b") => Some(LLVMFeature::new("cx16")), + ("x86", "lahfsahf") => Some(LLVMFeature::new("sahf")), // Enable the evex512 target feature if an avx512 target feature is enabled. ("x86", s) if s.starts_with("avx512") => Some(LLVMFeature::with_dependencies( s, smallvec![TargetFeatureFoldStrength::EnableOnly("evex512")], )), - ("sparc", "leoncasa") => Some(LLVMFeature::new("hasleoncasa")), - ("powerpc", "power8-crypto") => Some(LLVMFeature::new("crypto")), ("x86", "avx10.1") => Some(LLVMFeature::new("avx10.1-512")), ("x86", "avx10.2") => Some(LLVMFeature::new("avx10.2-512")), ("x86", "apxf") => Some(LLVMFeature::with_dependencies( diff --git a/compiler/rustc_codegen_ssa/messages.ftl b/compiler/rustc_codegen_ssa/messages.ftl index 1dd65d38a2b..91c3806df4c 100644 --- a/compiler/rustc_codegen_ssa/messages.ftl +++ b/compiler/rustc_codegen_ssa/messages.ftl @@ -8,8 +8,6 @@ codegen_ssa_aix_strip_not_used = using host's `strip` binary to cross-compile to codegen_ssa_archive_build_failure = failed to build archive at `{$path}`: {$error} -codegen_ssa_autodiff_without_lto = using the autodiff feature requires using fat-lto - codegen_ssa_bare_instruction_set = `#[instruction_set]` requires an argument codegen_ssa_binary_output_to_tty = option `-o` or `--emit` is used to write binary output type `{$shorthand}` to stdout, but stdout is a tty diff --git a/compiler/rustc_codegen_ssa/src/errors.rs b/compiler/rustc_codegen_ssa/src/errors.rs index fb5a8205140..d5c30c5c7a6 100644 --- a/compiler/rustc_codegen_ssa/src/errors.rs +++ b/compiler/rustc_codegen_ssa/src/errors.rs @@ -38,10 +38,6 @@ pub(crate) struct CguNotRecorded<'a> { } #[derive(Diagnostic)] -#[diag(codegen_ssa_autodiff_without_lto)] -pub struct AutodiffWithoutLto; - -#[derive(Diagnostic)] #[diag(codegen_ssa_unknown_reuse_kind)] pub(crate) struct UnknownReuseKind { #[primary_span] diff --git a/compiler/rustc_expand/src/mbe/macro_rules.rs b/compiler/rustc_expand/src/mbe/macro_rules.rs index 946265eba8b..1d147a0385c 100644 --- a/compiler/rustc_expand/src/mbe/macro_rules.rs +++ b/compiler/rustc_expand/src/mbe/macro_rules.rs @@ -702,7 +702,7 @@ pub fn compile_declarative_macro( kinds |= MacroKinds::DERIVE; let derive_keyword_span = p.prev_token.span; if !features.macro_derive() { - feature_err(sess, sym::macro_attr, span, "`macro_rules!` derives are unstable") + feature_err(sess, sym::macro_derive, span, "`macro_rules!` derives are unstable") .emit(); } if let Some(guar) = check_no_eof(sess, &p, "expected `()` after `derive`") { diff --git a/compiler/rustc_hir_typeck/src/method/suggest.rs b/compiler/rustc_hir_typeck/src/method/suggest.rs index a39ac0fcb6e..024b9ee08c2 100644 --- a/compiler/rustc_hir_typeck/src/method/suggest.rs +++ b/compiler/rustc_hir_typeck/src/method/suggest.rs @@ -2194,7 +2194,7 @@ impl<'a, 'tcx> FnCtxt<'a, 'tcx> { fn suggest_associated_call_syntax( &self, err: &mut Diag<'_>, - static_candidates: &Vec<CandidateSource>, + static_candidates: &[CandidateSource], rcvr_ty: Ty<'tcx>, source: SelfSource<'tcx>, item_name: Ident, @@ -2422,7 +2422,7 @@ impl<'a, 'tcx> FnCtxt<'a, 'tcx> { let span_included = match parent_expr.kind { hir::ExprKind::Struct(_, eps, _) => { - eps.len() > 0 && eps.last().is_some_and(|ep| ep.span.contains(span)) + eps.last().is_some_and(|ep| ep.span.contains(span)) } // `..=` desugars into `::std::ops::RangeInclusive::new(...)`. hir::ExprKind::Call(func, ..) => func.span.contains(span), @@ -2484,7 +2484,7 @@ impl<'a, 'tcx> FnCtxt<'a, 'tcx> { simplify_type(tcx, ty, TreatParams::InstantiateWithInfer) .and_then(|simp| { tcx.incoherent_impls(simp) - .into_iter() + .iter() .find_map(|&id| self.associated_value(id, item_name)) }) .is_some() @@ -2617,7 +2617,7 @@ impl<'a, 'tcx> FnCtxt<'a, 'tcx> { if let Node::Expr(call_expr) = self.tcx.parent_hir_node(seg1.hir_id) && let ControlFlow::Break(Some(expr)) = - (LetVisitor { ident_name: seg1.ident.name }).visit_body(&body) + (LetVisitor { ident_name: seg1.ident.name }).visit_body(body) && let Some(self_ty) = self.node_ty_opt(expr.hir_id) { let probe = self.lookup_probe_for_diagnostic( @@ -2960,14 +2960,11 @@ impl<'a, 'tcx> FnCtxt<'a, 'tcx> { .collect::<Vec<_>>() .into(); for pred in &local_preds { - match pred.self_ty().kind() { - ty::Adt(def, _) => { - local_spans.push_span_label( - self.tcx.def_span(def.did()), - format!("must implement `{}`", pred.trait_ref.print_trait_sugared()), - ); - } - _ => {} + if let ty::Adt(def, _) = pred.self_ty().kind() { + local_spans.push_span_label( + self.tcx.def_span(def.did()), + format!("must implement `{}`", pred.trait_ref.print_trait_sugared()), + ); } } if local_spans.primary_span().is_some() { @@ -3006,14 +3003,11 @@ impl<'a, 'tcx> FnCtxt<'a, 'tcx> { .collect::<Vec<_>>() .into(); for pred in &foreign_preds { - match pred.self_ty().kind() { - ty::Adt(def, _) => { - foreign_spans.push_span_label( - self.tcx.def_span(def.did()), - format!("not implement `{}`", pred.trait_ref.print_trait_sugared()), - ); - } - _ => {} + if let ty::Adt(def, _) = pred.self_ty().kind() { + foreign_spans.push_span_label( + self.tcx.def_span(def.did()), + format!("not implement `{}`", pred.trait_ref.print_trait_sugared()), + ); } } if foreign_spans.primary_span().is_some() { @@ -3595,7 +3589,7 @@ impl<'a, 'tcx> FnCtxt<'a, 'tcx> { // would take care of them. && !skippable.contains(&Some(pick.item.container_id(self.tcx))) // Do not suggest pinning when the method is directly on `Pin`. - && pick.item.impl_container(self.tcx).map_or(true, |did| { + && pick.item.impl_container(self.tcx).is_none_or(|did| { match self.tcx.type_of(did).skip_binder().kind() { ty::Adt(def, _) => Some(def.did()) != self.tcx.lang_items().pin_type(), _ => true, @@ -3653,7 +3647,7 @@ impl<'a, 'tcx> FnCtxt<'a, 'tcx> { vec![ ( rcvr.span.shrink_to_lo(), - format!("let mut pinned = std::pin::pin!("), + "let mut pinned = std::pin::pin!(".to_string(), ), ( rcvr.span.shrink_to_hi(), @@ -4128,7 +4122,7 @@ impl<'a, 'tcx> FnCtxt<'a, 'tcx> { ); let trait_span = self.tcx.def_span(trait_def_id); let mut multi_span: MultiSpan = trait_span.into(); - multi_span.push_span_label(trait_span, format!("this is the trait that is needed")); + multi_span.push_span_label(trait_span, "this is the trait that is needed".to_string()); let descr = self.tcx.associated_item(item_def_id).descr(); let rcvr_ty = rcvr_ty.map(|t| format!("`{t}`")).unwrap_or_else(|| "the receiver".to_string()); @@ -4146,7 +4140,7 @@ impl<'a, 'tcx> FnCtxt<'a, 'tcx> { } multi_span.push_span_label( self.tcx.def_span(def_id), - format!("this is the trait that was imported"), + "this is the trait that was imported".to_string(), ); } err.span_note(multi_span, msg); diff --git a/compiler/rustc_hir_typeck/src/op.rs b/compiler/rustc_hir_typeck/src/op.rs index 05443537943..a8e8582c51c 100644 --- a/compiler/rustc_hir_typeck/src/op.rs +++ b/compiler/rustc_hir_typeck/src/op.rs @@ -566,7 +566,7 @@ impl<'a, 'tcx> FnCtxt<'a, 'tcx> { rhs_ty, lhs_expr, lhs_ty, - |lhs_ty, rhs_ty| is_compatible_after_call(lhs_ty, rhs_ty), + is_compatible_after_call, ) { // Cool } @@ -1170,8 +1170,8 @@ fn deref_ty_if_possible(ty: Ty<'_>) -> Ty<'_> { } } -/// Returns `true` if this is a built-in arithmetic operation (e.g., u32 -/// + u32, i16x4 == i16x4) and false if these types would have to be +/// Returns `true` if this is a built-in arithmetic operation (e.g., +/// u32 + u32, i16x4 == i16x4) and false if these types would have to be /// overloaded to be legal. There are two reasons that we distinguish /// builtin operations from overloaded ones (vs trying to drive /// everything uniformly through the trait system and intrinsics or @@ -1191,7 +1191,7 @@ fn is_builtin_binop<'tcx>(lhs: Ty<'tcx>, rhs: Ty<'tcx>, category: BinOpCategory) // (See https://github.com/rust-lang/rust/issues/57447.) let (lhs, rhs) = (deref_ty_if_possible(lhs), deref_ty_if_possible(rhs)); - match category.into() { + match category { BinOpCategory::Shortcircuit => true, BinOpCategory::Shift => { lhs.references_error() diff --git a/compiler/rustc_llvm/llvm-wrapper/PassWrapper.cpp b/compiler/rustc_llvm/llvm-wrapper/PassWrapper.cpp index ab5d5c03e81..7518b40799b 100644 --- a/compiler/rustc_llvm/llvm-wrapper/PassWrapper.cpp +++ b/compiler/rustc_llvm/llvm-wrapper/PassWrapper.cpp @@ -271,8 +271,9 @@ extern "C" LLVMTargetMachineRef LLVMRustCreateTargetMachine( bool TrapUnreachable, bool Singlethread, bool VerboseAsm, bool EmitStackSizeSection, bool RelaxELFRelocations, bool UseInitArray, const char *SplitDwarfFile, const char *OutputObjFile, - const char *DebugInfoCompression, bool UseEmulatedTls, - const char *ArgsCstrBuff, size_t ArgsCstrBuffLen, bool UseWasmEH) { + const char *DebugInfoCompression, bool UseEmulatedTls, const char *Argv0, + size_t Argv0Len, const char *CommandLineArgs, size_t CommandLineArgsLen, + bool UseWasmEH) { auto OptLevel = fromRust(RustOptLevel); auto RM = fromRust(RustReloc); @@ -343,25 +344,10 @@ extern "C" LLVMTargetMachineRef LLVMRustCreateTargetMachine( Options.EmitStackSizeSection = EmitStackSizeSection; - if (ArgsCstrBuff != nullptr) { - size_t buffer_offset = 0; - assert(ArgsCstrBuff[ArgsCstrBuffLen - 1] == '\0'); - auto Arg0 = std::string(ArgsCstrBuff); - buffer_offset = Arg0.size() + 1; - - std::string CommandlineArgs; - raw_string_ostream OS(CommandlineArgs); - ListSeparator LS(" "); - for (StringRef Arg : split(StringRef(ArgsCstrBuff + buffer_offset, - ArgsCstrBuffLen - buffer_offset), - '\0')) { - OS << LS; - sys::printArg(OS, Arg, /*Quote=*/true); - } - OS.flush(); - Options.MCOptions.Argv0 = Arg0; - Options.MCOptions.CommandlineArgs = CommandlineArgs; - } + if (Argv0 != nullptr) + Options.MCOptions.Argv0 = {Argv0, Argv0Len}; + if (CommandLineArgs != nullptr) + Options.MCOptions.CommandlineArgs = {CommandLineArgs, CommandLineArgsLen}; #if LLVM_VERSION_GE(21, 0) TargetMachine *TM = TheTarget->createTargetMachine(Trip, CPU, Feature, diff --git a/compiler/rustc_mir_transform/Cargo.toml b/compiler/rustc_mir_transform/Cargo.toml index 08c43a4648c..511c1960e40 100644 --- a/compiler/rustc_mir_transform/Cargo.toml +++ b/compiler/rustc_mir_transform/Cargo.toml @@ -6,6 +6,7 @@ edition = "2024" [dependencies] # tidy-alphabetical-start either = "1" +hashbrown = "0.15" itertools = "0.12" rustc_abi = { path = "../rustc_abi" } rustc_arena = { path = "../rustc_arena" } diff --git a/compiler/rustc_mir_transform/src/gvn.rs b/compiler/rustc_mir_transform/src/gvn.rs index 30e68a3f650..ebec3d12500 100644 --- a/compiler/rustc_mir_transform/src/gvn.rs +++ b/compiler/rustc_mir_transform/src/gvn.rs @@ -85,8 +85,10 @@ //! that contain `AllocId`s. use std::borrow::Cow; +use std::hash::{Hash, Hasher}; use either::Either; +use hashbrown::hash_table::{Entry, HashTable}; use itertools::Itertools as _; use rustc_abi::{self as abi, BackendRepr, FIRST_VARIANT, FieldIdx, Primitive, Size, VariantIdx}; use rustc_const_eval::const_eval::DummyMachine; @@ -94,7 +96,7 @@ use rustc_const_eval::interpret::{ ImmTy, Immediate, InterpCx, MemPlaceMeta, MemoryKind, OpTy, Projectable, Scalar, intern_const_alloc_for_constprop, }; -use rustc_data_structures::fx::{FxIndexSet, MutableValues}; +use rustc_data_structures::fx::FxHasher; use rustc_data_structures::graph::dominators::Dominators; use rustc_hir::def::DefKind; use rustc_index::bit_set::DenseBitSet; @@ -152,9 +154,29 @@ impl<'tcx> crate::MirPass<'tcx> for GVN { } newtype_index! { + /// This represents a `Value` in the symbolic execution. + #[debug_format = "_v{}"] struct VnIndex {} } +/// Marker type to forbid hashing and comparing opaque values. +/// This struct should only be constructed by `ValueSet::insert_unique` to ensure we use that +/// method to create non-unifiable values. It will ICE if used in `ValueSet::insert`. +#[derive(Copy, Clone, Debug, Eq)] +struct VnOpaque; +impl PartialEq for VnOpaque { + fn eq(&self, _: &VnOpaque) -> bool { + // ICE if we try to compare unique values + unreachable!() + } +} +impl Hash for VnOpaque { + fn hash<T: Hasher>(&self, _: &mut T) { + // ICE if we try to hash unique values + unreachable!() + } +} + #[derive(Copy, Clone, Debug, PartialEq, Eq, Hash)] enum AddressKind { Ref(BorrowKind), @@ -166,15 +188,17 @@ enum Value<'tcx> { // Root values. /// Used to represent values we know nothing about. /// The `usize` is a counter incremented by `new_opaque`. - Opaque(usize), + Opaque(VnOpaque), /// Evaluated or unevaluated constant value. Constant { value: Const<'tcx>, /// Some constants do not have a deterministic value. To avoid merging two instances of the /// same `Const`, we assign them an additional integer index. - // `disambiguator` is 0 iff the constant is deterministic. - disambiguator: usize, + // `disambiguator` is `None` iff the constant is deterministic. + disambiguator: Option<VnOpaque>, }, + + // Aggregates. /// An aggregate value, either tuple/closure/struct/enum. /// This does not contain unions, as we cannot reason with the value. Aggregate(VariantIdx, Vec<VnIndex>), @@ -192,7 +216,7 @@ enum Value<'tcx> { place: Place<'tcx>, kind: AddressKind, /// Give each borrow and pointer a different provenance, so we don't merge them. - provenance: usize, + provenance: VnOpaque, }, // Extractions. @@ -211,6 +235,107 @@ enum Value<'tcx> { }, } +/// Stores and deduplicates pairs of `(Value, Ty)` into in `VnIndex` numbered values. +/// +/// This data structure is mostly a partial reimplementation of `FxIndexMap<VnIndex, (Value, Ty)>`. +/// We do not use a regular `FxIndexMap` to skip hashing values that are unique by construction, +/// like opaque values, address with provenance and non-deterministic constants. +struct ValueSet<'tcx> { + indices: HashTable<VnIndex>, + hashes: IndexVec<VnIndex, u64>, + values: IndexVec<VnIndex, Value<'tcx>>, + types: IndexVec<VnIndex, Ty<'tcx>>, +} + +impl<'tcx> ValueSet<'tcx> { + fn new(num_values: usize) -> ValueSet<'tcx> { + ValueSet { + indices: HashTable::with_capacity(num_values), + hashes: IndexVec::with_capacity(num_values), + values: IndexVec::with_capacity(num_values), + types: IndexVec::with_capacity(num_values), + } + } + + /// Insert a `(Value, Ty)` pair without hashing or deduplication. + /// This always creates a new `VnIndex`. + #[inline] + fn insert_unique( + &mut self, + ty: Ty<'tcx>, + value: impl FnOnce(VnOpaque) -> Value<'tcx>, + ) -> VnIndex { + let value = value(VnOpaque); + + debug_assert!(match value { + Value::Opaque(_) | Value::Address { .. } => true, + Value::Constant { disambiguator, .. } => disambiguator.is_some(), + _ => false, + }); + + let index = self.hashes.push(0); + let _index = self.types.push(ty); + debug_assert_eq!(index, _index); + let _index = self.values.push(value); + debug_assert_eq!(index, _index); + index + } + + /// Insert a `(Value, Ty)` pair to be deduplicated. + /// Returns `true` as second tuple field if this value did not exist previously. + #[allow(rustc::pass_by_value)] // closures take `&VnIndex` + fn insert(&mut self, ty: Ty<'tcx>, value: Value<'tcx>) -> (VnIndex, bool) { + debug_assert!(match value { + Value::Opaque(_) | Value::Address { .. } => false, + Value::Constant { disambiguator, .. } => disambiguator.is_none(), + _ => true, + }); + + let hash: u64 = { + let mut h = FxHasher::default(); + value.hash(&mut h); + ty.hash(&mut h); + h.finish() + }; + + let eq = |index: &VnIndex| self.values[*index] == value && self.types[*index] == ty; + let hasher = |index: &VnIndex| self.hashes[*index]; + match self.indices.entry(hash, eq, hasher) { + Entry::Occupied(entry) => { + let index = *entry.get(); + (index, false) + } + Entry::Vacant(entry) => { + let index = self.hashes.push(hash); + entry.insert(index); + let _index = self.values.push(value); + debug_assert_eq!(index, _index); + let _index = self.types.push(ty); + debug_assert_eq!(index, _index); + (index, true) + } + } + } + + /// Return the `Value` associated with the given `VnIndex`. + #[inline] + fn value(&self, index: VnIndex) -> &Value<'tcx> { + &self.values[index] + } + + /// Return the type associated with the given `VnIndex`. + #[inline] + fn ty(&self, index: VnIndex) -> Ty<'tcx> { + self.types[index] + } + + /// Replace the value associated with `index` with an opaque value. + #[inline] + fn forget(&mut self, index: VnIndex) { + self.values[index] = Value::Opaque(VnOpaque); + } +} + struct VnState<'body, 'tcx> { tcx: TyCtxt<'tcx>, ecx: InterpCx<'tcx, DummyMachine>, @@ -221,11 +346,9 @@ struct VnState<'body, 'tcx> { /// Locals that are assigned that value. // This vector does not hold all the values of `VnIndex` that we create. rev_locals: IndexVec<VnIndex, SmallVec<[Local; 1]>>, - values: FxIndexSet<(Value<'tcx>, Ty<'tcx>)>, + values: ValueSet<'tcx>, /// Values evaluated as constants if possible. evaluated: IndexVec<VnIndex, Option<OpTy<'tcx>>>, - /// Counter to generate different values. - next_opaque: usize, /// Cache the deref values. derefs: Vec<VnIndex>, ssa: &'body SsaLocals, @@ -256,9 +379,8 @@ impl<'body, 'tcx> VnState<'body, 'tcx> { is_coroutine: body.coroutine.is_some(), locals: IndexVec::from_elem(None, local_decls), rev_locals: IndexVec::with_capacity(num_values), - values: FxIndexSet::with_capacity_and_hasher(num_values, Default::default()), + values: ValueSet::new(num_values), evaluated: IndexVec::with_capacity(num_values), - next_opaque: 1, derefs: Vec::new(), ssa, dominators, @@ -272,8 +394,7 @@ impl<'body, 'tcx> VnState<'body, 'tcx> { #[instrument(level = "trace", skip(self), ret)] fn insert(&mut self, ty: Ty<'tcx>, value: Value<'tcx>) -> VnIndex { - let (index, new) = self.values.insert_full((value, ty)); - let index = VnIndex::from_usize(index); + let (index, new) = self.values.insert(ty, value); if new { // Grow `evaluated` and `rev_locals` here to amortize the allocations. let evaluated = self.eval_to_const(index); @@ -285,18 +406,16 @@ impl<'body, 'tcx> VnState<'body, 'tcx> { index } - fn next_opaque(&mut self) -> usize { - let next_opaque = self.next_opaque; - self.next_opaque += 1; - next_opaque - } - /// Create a new `Value` for which we have no information at all, except that it is distinct /// from all the others. #[instrument(level = "trace", skip(self), ret)] fn new_opaque(&mut self, ty: Ty<'tcx>) -> VnIndex { - let value = Value::Opaque(self.next_opaque()); - self.insert(ty, value) + let index = self.values.insert_unique(ty, Value::Opaque); + let _index = self.evaluated.push(None); + debug_assert_eq!(index, _index); + let _index = self.rev_locals.push(SmallVec::new()); + debug_assert_eq!(index, _index); + index } /// Create a new `Value::Address` distinct from all the others. @@ -309,18 +428,49 @@ impl<'body, 'tcx> VnState<'body, 'tcx> { } AddressKind::Address(mutbl) => Ty::new_ptr(self.tcx, pty, mutbl.to_mutbl_lossy()), }; - let value = Value::Address { place, kind, provenance: self.next_opaque() }; - self.insert(ty, value) + let index = + self.values.insert_unique(ty, |provenance| Value::Address { place, kind, provenance }); + let evaluated = self.eval_to_const(index); + let _index = self.evaluated.push(evaluated); + debug_assert_eq!(index, _index); + let _index = self.rev_locals.push(SmallVec::new()); + debug_assert_eq!(index, _index); + index + } + + #[instrument(level = "trace", skip(self), ret)] + fn insert_constant(&mut self, value: Const<'tcx>) -> VnIndex { + let (index, new) = if value.is_deterministic() { + // The constant is deterministic, no need to disambiguate. + let constant = Value::Constant { value, disambiguator: None }; + self.values.insert(value.ty(), constant) + } else { + // Multiple mentions of this constant will yield different values, + // so assign a different `disambiguator` to ensure they do not get the same `VnIndex`. + let index = self.values.insert_unique(value.ty(), |disambiguator| Value::Constant { + value, + disambiguator: Some(disambiguator), + }); + (index, true) + }; + if new { + let evaluated = self.eval_to_const(index); + let _index = self.evaluated.push(evaluated); + debug_assert_eq!(index, _index); + let _index = self.rev_locals.push(SmallVec::new()); + debug_assert_eq!(index, _index); + } + index } #[inline] fn get(&self, index: VnIndex) -> &Value<'tcx> { - &self.values.get_index(index.as_usize()).unwrap().0 + self.values.value(index) } #[inline] fn ty(&self, index: VnIndex) -> Ty<'tcx> { - self.values.get_index(index.as_usize()).unwrap().1 + self.values.ty(index) } /// Record that `local` is assigned `value`. `local` must be SSA. @@ -331,33 +481,18 @@ impl<'body, 'tcx> VnState<'body, 'tcx> { self.rev_locals[value].push(local); } - fn insert_constant(&mut self, value: Const<'tcx>) -> VnIndex { - let disambiguator = if value.is_deterministic() { - // The constant is deterministic, no need to disambiguate. - 0 - } else { - // Multiple mentions of this constant will yield different values, - // so assign a different `disambiguator` to ensure they do not get the same `VnIndex`. - let disambiguator = self.next_opaque(); - // `disambiguator: 0` means deterministic. - debug_assert_ne!(disambiguator, 0); - disambiguator - }; - self.insert(value.ty(), Value::Constant { value, disambiguator }) - } - fn insert_bool(&mut self, flag: bool) -> VnIndex { // Booleans are deterministic. let value = Const::from_bool(self.tcx, flag); debug_assert!(value.is_deterministic()); - self.insert(self.tcx.types.bool, Value::Constant { value, disambiguator: 0 }) + self.insert(self.tcx.types.bool, Value::Constant { value, disambiguator: None }) } fn insert_scalar(&mut self, ty: Ty<'tcx>, scalar: Scalar) -> VnIndex { // Scalars are deterministic. let value = Const::from_scalar(self.tcx, scalar, ty); debug_assert!(value.is_deterministic()); - self.insert(ty, Value::Constant { value, disambiguator: 0 }) + self.insert(ty, Value::Constant { value, disambiguator: None }) } fn insert_tuple(&mut self, ty: Ty<'tcx>, values: Vec<VnIndex>) -> VnIndex { @@ -372,8 +507,7 @@ impl<'body, 'tcx> VnState<'body, 'tcx> { fn invalidate_derefs(&mut self) { for deref in std::mem::take(&mut self.derefs) { - let opaque = self.next_opaque(); - self.values.get_index_mut2(deref.index()).unwrap().0 = Value::Opaque(opaque); + self.values.forget(deref); } } @@ -1572,7 +1706,7 @@ impl<'tcx> VnState<'_, 'tcx> { // This was already constant in MIR, do not change it. If the constant is not // deterministic, adding an additional mention of it in MIR will not give the same value as // the former mention. - if let Value::Constant { value, disambiguator: 0 } = *self.get(index) { + if let Value::Constant { value, disambiguator: None } = *self.get(index) { debug_assert!(value.is_deterministic()); return Some(ConstOperand { span: DUMMY_SP, user_ty: None, const_: value }); } diff --git a/compiler/rustc_next_trait_solver/src/canonicalizer.rs b/compiler/rustc_next_trait_solver/src/canonical/canonicalizer.rs index 4b4ec4956eb..b25671d676b 100644 --- a/compiler/rustc_next_trait_solver/src/canonicalizer.rs +++ b/compiler/rustc_next_trait_solver/src/canonical/canonicalizer.rs @@ -57,7 +57,7 @@ enum CanonicalizeMode { }, } -pub struct Canonicalizer<'a, D: SolverDelegate<Interner = I>, I: Interner> { +pub(super) struct Canonicalizer<'a, D: SolverDelegate<Interner = I>, I: Interner> { delegate: &'a D, // Immutable field. @@ -83,7 +83,7 @@ pub struct Canonicalizer<'a, D: SolverDelegate<Interner = I>, I: Interner> { } impl<'a, D: SolverDelegate<Interner = I>, I: Interner> Canonicalizer<'a, D, I> { - pub fn canonicalize_response<T: TypeFoldable<I>>( + pub(super) fn canonicalize_response<T: TypeFoldable<I>>( delegate: &'a D, max_input_universe: ty::UniverseIndex, variables: &'a mut Vec<I::GenericArg>, @@ -112,7 +112,6 @@ impl<'a, D: SolverDelegate<Interner = I>, I: Interner> Canonicalizer<'a, D, I> { let (max_universe, variables) = canonicalizer.finalize(); Canonical { max_universe, variables, value } } - fn canonicalize_param_env( delegate: &'a D, variables: &'a mut Vec<I::GenericArg>, @@ -195,7 +194,7 @@ impl<'a, D: SolverDelegate<Interner = I>, I: Interner> Canonicalizer<'a, D, I> { /// /// We want to keep the option of canonicalizing `'static` to an existential /// variable in the future by changing the way we detect global where-bounds. - pub fn canonicalize_input<P: TypeFoldable<I>>( + pub(super) fn canonicalize_input<P: TypeFoldable<I>>( delegate: &'a D, variables: &'a mut Vec<I::GenericArg>, input: QueryInput<I, P>, diff --git a/compiler/rustc_next_trait_solver/src/canonical/mod.rs b/compiler/rustc_next_trait_solver/src/canonical/mod.rs new file mode 100644 index 00000000000..e3520e238ed --- /dev/null +++ b/compiler/rustc_next_trait_solver/src/canonical/mod.rs @@ -0,0 +1,364 @@ +//! Canonicalization is used to separate some goal from its context, +//! throwing away unnecessary information in the process. +//! +//! This is necessary to cache goals containing inference variables +//! and placeholders without restricting them to the current `InferCtxt`. +//! +//! Canonicalization is fairly involved, for more details see the relevant +//! section of the [rustc-dev-guide][c]. +//! +//! [c]: https://rustc-dev-guide.rust-lang.org/solve/canonicalization.html + +use std::iter; + +use canonicalizer::Canonicalizer; +use rustc_index::IndexVec; +use rustc_type_ir::inherent::*; +use rustc_type_ir::relate::solver_relating::RelateExt; +use rustc_type_ir::{ + self as ty, Canonical, CanonicalVarKind, CanonicalVarValues, InferCtxtLike, Interner, + TypeFoldable, +}; +use tracing::instrument; + +use crate::delegate::SolverDelegate; +use crate::resolve::eager_resolve_vars; +use crate::solve::{ + CanonicalInput, CanonicalResponse, Certainty, ExternalConstraintsData, Goal, + NestedNormalizationGoals, PredefinedOpaquesData, QueryInput, Response, inspect, +}; + +pub mod canonicalizer; + +trait ResponseT<I: Interner> { + fn var_values(&self) -> CanonicalVarValues<I>; +} + +impl<I: Interner> ResponseT<I> for Response<I> { + fn var_values(&self) -> CanonicalVarValues<I> { + self.var_values + } +} + +impl<I: Interner, T> ResponseT<I> for inspect::State<I, T> { + fn var_values(&self) -> CanonicalVarValues<I> { + self.var_values + } +} + +/// Canonicalizes the goal remembering the original values +/// for each bound variable. +/// +/// This expects `goal` and `opaque_types` to be eager resolved. +pub(super) fn canonicalize_goal<D, I>( + delegate: &D, + goal: Goal<I, I::Predicate>, + opaque_types: Vec<(ty::OpaqueTypeKey<I>, I::Ty)>, +) -> (Vec<I::GenericArg>, CanonicalInput<I, I::Predicate>) +where + D: SolverDelegate<Interner = I>, + I: Interner, +{ + let mut orig_values = Default::default(); + let canonical = Canonicalizer::canonicalize_input( + delegate, + &mut orig_values, + QueryInput { + goal, + predefined_opaques_in_body: delegate + .cx() + .mk_predefined_opaques_in_body(PredefinedOpaquesData { opaque_types }), + }, + ); + let query_input = ty::CanonicalQueryInput { canonical, typing_mode: delegate.typing_mode() }; + (orig_values, query_input) +} + +pub(super) fn canonicalize_response<D, I, T>( + delegate: &D, + max_input_universe: ty::UniverseIndex, + value: T, +) -> ty::Canonical<I, T> +where + D: SolverDelegate<Interner = I>, + I: Interner, + T: TypeFoldable<I>, +{ + let mut orig_values = Default::default(); + let canonical = + Canonicalizer::canonicalize_response(delegate, max_input_universe, &mut orig_values, value); + canonical +} + +/// After calling a canonical query, we apply the constraints returned +/// by the query using this function. +/// +/// This happens in three steps: +/// - we instantiate the bound variables of the query response +/// - we unify the `var_values` of the response with the `original_values` +/// - we apply the `external_constraints` returned by the query, returning +/// the `normalization_nested_goals` +pub(super) fn instantiate_and_apply_query_response<D, I>( + delegate: &D, + param_env: I::ParamEnv, + original_values: &[I::GenericArg], + response: CanonicalResponse<I>, + span: I::Span, +) -> (NestedNormalizationGoals<I>, Certainty) +where + D: SolverDelegate<Interner = I>, + I: Interner, +{ + let instantiation = + compute_query_response_instantiation_values(delegate, &original_values, &response, span); + + let Response { var_values, external_constraints, certainty } = + delegate.instantiate_canonical(response, instantiation); + + unify_query_var_values(delegate, param_env, &original_values, var_values, span); + + let ExternalConstraintsData { region_constraints, opaque_types, normalization_nested_goals } = + &*external_constraints; + + register_region_constraints(delegate, region_constraints, span); + register_new_opaque_types(delegate, opaque_types, span); + + (normalization_nested_goals.clone(), certainty) +} + +/// This returns the canonical variable values to instantiate the bound variables of +/// the canonical response. This depends on the `original_values` for the +/// bound variables. +fn compute_query_response_instantiation_values<D, I, T>( + delegate: &D, + original_values: &[I::GenericArg], + response: &Canonical<I, T>, + span: I::Span, +) -> CanonicalVarValues<I> +where + D: SolverDelegate<Interner = I>, + I: Interner, + T: ResponseT<I>, +{ + // FIXME: Longterm canonical queries should deal with all placeholders + // created inside of the query directly instead of returning them to the + // caller. + let prev_universe = delegate.universe(); + let universes_created_in_query = response.max_universe.index(); + for _ in 0..universes_created_in_query { + delegate.create_next_universe(); + } + + let var_values = response.value.var_values(); + assert_eq!(original_values.len(), var_values.len()); + + // If the query did not make progress with constraining inference variables, + // we would normally create a new inference variables for bound existential variables + // only then unify this new inference variable with the inference variable from + // the input. + // + // We therefore instantiate the existential variable in the canonical response with the + // inference variable of the input right away, which is more performant. + let mut opt_values = IndexVec::from_elem_n(None, response.variables.len()); + for (original_value, result_value) in iter::zip(original_values, var_values.var_values.iter()) { + match result_value.kind() { + ty::GenericArgKind::Type(t) => { + // We disable the instantiation guess for inference variables + // and only use it for placeholders. We need to handle the + // `sub_root` of type inference variables which would make this + // more involved. They are also a lot rarer than region variables. + if let ty::Bound(debruijn, b) = t.kind() + && !matches!( + response.variables.get(b.var().as_usize()).unwrap(), + CanonicalVarKind::Ty { .. } + ) + { + assert_eq!(debruijn, ty::INNERMOST); + opt_values[b.var()] = Some(*original_value); + } + } + ty::GenericArgKind::Lifetime(r) => { + if let ty::ReBound(debruijn, br) = r.kind() { + assert_eq!(debruijn, ty::INNERMOST); + opt_values[br.var()] = Some(*original_value); + } + } + ty::GenericArgKind::Const(c) => { + if let ty::ConstKind::Bound(debruijn, bv) = c.kind() { + assert_eq!(debruijn, ty::INNERMOST); + opt_values[bv.var()] = Some(*original_value); + } + } + } + } + CanonicalVarValues::instantiate(delegate.cx(), response.variables, |var_values, kind| { + if kind.universe() != ty::UniverseIndex::ROOT { + // A variable from inside a binder of the query. While ideally these shouldn't + // exist at all (see the FIXME at the start of this method), we have to deal with + // them for now. + delegate.instantiate_canonical_var(kind, span, &var_values, |idx| { + prev_universe + idx.index() + }) + } else if kind.is_existential() { + // As an optimization we sometimes avoid creating a new inference variable here. + // + // All new inference variables we create start out in the current universe of the caller. + // This is conceptually wrong as these inference variables would be able to name + // more placeholders then they should be able to. However the inference variables have + // to "come from somewhere", so by equating them with the original values of the caller + // later on, we pull them down into their correct universe again. + if let Some(v) = opt_values[ty::BoundVar::from_usize(var_values.len())] { + v + } else { + delegate.instantiate_canonical_var(kind, span, &var_values, |_| prev_universe) + } + } else { + // For placeholders which were already part of the input, we simply map this + // universal bound variable back the placeholder of the input. + original_values[kind.expect_placeholder_index()] + } + }) +} + +/// Unify the `original_values` with the `var_values` returned by the canonical query.. +/// +/// This assumes that this unification will always succeed. This is the case when +/// applying a query response right away. However, calling a canonical query, doing any +/// other kind of trait solving, and only then instantiating the result of the query +/// can cause the instantiation to fail. This is not supported and we ICE in this case. +/// +/// We always structurally instantiate aliases. Relating aliases needs to be different +/// depending on whether the alias is *rigid* or not. We're only really able to tell +/// whether an alias is rigid by using the trait solver. When instantiating a response +/// from the solver we assume that the solver correctly handled aliases and therefore +/// always relate them structurally here. +#[instrument(level = "trace", skip(delegate))] +fn unify_query_var_values<D, I>( + delegate: &D, + param_env: I::ParamEnv, + original_values: &[I::GenericArg], + var_values: CanonicalVarValues<I>, + span: I::Span, +) where + D: SolverDelegate<Interner = I>, + I: Interner, +{ + assert_eq!(original_values.len(), var_values.len()); + + for (&orig, response) in iter::zip(original_values, var_values.var_values.iter()) { + let goals = + delegate.eq_structurally_relating_aliases(param_env, orig, response, span).unwrap(); + assert!(goals.is_empty()); + } +} + +fn register_region_constraints<D, I>( + delegate: &D, + outlives: &[ty::OutlivesPredicate<I, I::GenericArg>], + span: I::Span, +) where + D: SolverDelegate<Interner = I>, + I: Interner, +{ + for &ty::OutlivesPredicate(lhs, rhs) in outlives { + match lhs.kind() { + ty::GenericArgKind::Lifetime(lhs) => delegate.sub_regions(rhs, lhs, span), + ty::GenericArgKind::Type(lhs) => delegate.register_ty_outlives(lhs, rhs, span), + ty::GenericArgKind::Const(_) => panic!("const outlives: {lhs:?}: {rhs:?}"), + } + } +} + +fn register_new_opaque_types<D, I>( + delegate: &D, + opaque_types: &[(ty::OpaqueTypeKey<I>, I::Ty)], + span: I::Span, +) where + D: SolverDelegate<Interner = I>, + I: Interner, +{ + for &(key, ty) in opaque_types { + let prev = delegate.register_hidden_type_in_storage(key, ty, span); + // We eagerly resolve inference variables when computing the query response. + // This can cause previously distinct opaque type keys to now be structurally equal. + // + // To handle this, we store any duplicate entries in a separate list to check them + // at the end of typeck/borrowck. We could alternatively eagerly equate the hidden + // types here. However, doing so is difficult as it may result in nested goals and + // any errors may make it harder to track the control flow for diagnostics. + if let Some(prev) = prev { + delegate.add_duplicate_opaque_type(key, prev, span); + } + } +} + +/// Used by proof trees to be able to recompute intermediate actions while +/// evaluating a goal. The `var_values` not only include the bound variables +/// of the query input, but also contain all unconstrained inference vars +/// created while evaluating this goal. +pub fn make_canonical_state<D, I, T>( + delegate: &D, + var_values: &[I::GenericArg], + max_input_universe: ty::UniverseIndex, + data: T, +) -> inspect::CanonicalState<I, T> +where + D: SolverDelegate<Interner = I>, + I: Interner, + T: TypeFoldable<I>, +{ + let var_values = CanonicalVarValues { var_values: delegate.cx().mk_args(var_values) }; + let state = inspect::State { var_values, data }; + let state = eager_resolve_vars(delegate, state); + Canonicalizer::canonicalize_response(delegate, max_input_universe, &mut vec![], state) +} + +// FIXME: needs to be pub to be accessed by downstream +// `rustc_trait_selection::solve::inspect::analyse`. +pub fn instantiate_canonical_state<D, I, T>( + delegate: &D, + span: I::Span, + param_env: I::ParamEnv, + orig_values: &mut Vec<I::GenericArg>, + state: inspect::CanonicalState<I, T>, +) -> T +where + D: SolverDelegate<Interner = I>, + I: Interner, + T: TypeFoldable<I>, +{ + // In case any fresh inference variables have been created between `state` + // and the previous instantiation, extend `orig_values` for it. + orig_values.extend( + state.value.var_values.var_values.as_slice()[orig_values.len()..] + .iter() + .map(|&arg| delegate.fresh_var_for_kind_with_span(arg, span)), + ); + + let instantiation = + compute_query_response_instantiation_values(delegate, orig_values, &state, span); + + let inspect::State { var_values, data } = delegate.instantiate_canonical(state, instantiation); + + unify_query_var_values(delegate, param_env, orig_values, var_values, span); + data +} + +pub fn response_no_constraints_raw<I: Interner>( + cx: I, + max_universe: ty::UniverseIndex, + variables: I::CanonicalVarKinds, + certainty: Certainty, +) -> CanonicalResponse<I> { + ty::Canonical { + max_universe, + variables, + value: Response { + var_values: ty::CanonicalVarValues::make_identity(cx, variables), + // FIXME: maybe we should store the "no response" version in cx, like + // we do for cx.types and stuff. + external_constraints: cx.mk_external_constraints(ExternalConstraintsData::default()), + certainty, + }, + } +} diff --git a/compiler/rustc_next_trait_solver/src/lib.rs b/compiler/rustc_next_trait_solver/src/lib.rs index d3965e14c68..5fa29b7d9f8 100644 --- a/compiler/rustc_next_trait_solver/src/lib.rs +++ b/compiler/rustc_next_trait_solver/src/lib.rs @@ -10,7 +10,7 @@ #![allow(rustc::usage_of_type_ir_traits)] // tidy-alphabetical-end -pub mod canonicalizer; +pub mod canonical; pub mod coherence; pub mod delegate; pub mod placeholder; diff --git a/compiler/rustc_next_trait_solver/src/solve/eval_ctxt/canonical.rs b/compiler/rustc_next_trait_solver/src/solve/eval_ctxt/canonical.rs deleted file mode 100644 index 889588afe61..00000000000 --- a/compiler/rustc_next_trait_solver/src/solve/eval_ctxt/canonical.rs +++ /dev/null @@ -1,517 +0,0 @@ -//! Canonicalization is used to separate some goal from its context, -//! throwing away unnecessary information in the process. -//! -//! This is necessary to cache goals containing inference variables -//! and placeholders without restricting them to the current `InferCtxt`. -//! -//! Canonicalization is fairly involved, for more details see the relevant -//! section of the [rustc-dev-guide][c]. -//! -//! [c]: https://rustc-dev-guide.rust-lang.org/solve/canonicalization.html - -use std::iter; - -use rustc_index::IndexVec; -use rustc_type_ir::data_structures::HashSet; -use rustc_type_ir::inherent::*; -use rustc_type_ir::relate::solver_relating::RelateExt; -use rustc_type_ir::solve::OpaqueTypesJank; -use rustc_type_ir::{ - self as ty, Canonical, CanonicalVarKind, CanonicalVarValues, InferCtxtLike, Interner, - TypeFoldable, -}; -use tracing::{debug, instrument, trace}; - -use crate::canonicalizer::Canonicalizer; -use crate::delegate::SolverDelegate; -use crate::resolve::eager_resolve_vars; -use crate::solve::eval_ctxt::CurrentGoalKind; -use crate::solve::{ - CanonicalInput, CanonicalResponse, Certainty, EvalCtxt, ExternalConstraintsData, Goal, - MaybeCause, NestedNormalizationGoals, NoSolution, PredefinedOpaquesData, QueryInput, - QueryResult, Response, inspect, response_no_constraints_raw, -}; - -trait ResponseT<I: Interner> { - fn var_values(&self) -> CanonicalVarValues<I>; -} - -impl<I: Interner> ResponseT<I> for Response<I> { - fn var_values(&self) -> CanonicalVarValues<I> { - self.var_values - } -} - -impl<I: Interner, T> ResponseT<I> for inspect::State<I, T> { - fn var_values(&self) -> CanonicalVarValues<I> { - self.var_values - } -} - -impl<D, I> EvalCtxt<'_, D> -where - D: SolverDelegate<Interner = I>, - I: Interner, -{ - /// Canonicalizes the goal remembering the original values - /// for each bound variable. - /// - /// This expects `goal` and `opaque_types` to be eager resolved. - pub(super) fn canonicalize_goal( - delegate: &D, - goal: Goal<I, I::Predicate>, - opaque_types: Vec<(ty::OpaqueTypeKey<I>, I::Ty)>, - ) -> (Vec<I::GenericArg>, CanonicalInput<I, I::Predicate>) { - let mut orig_values = Default::default(); - let canonical = Canonicalizer::canonicalize_input( - delegate, - &mut orig_values, - QueryInput { - goal, - predefined_opaques_in_body: delegate - .cx() - .mk_predefined_opaques_in_body(PredefinedOpaquesData { opaque_types }), - }, - ); - let query_input = - ty::CanonicalQueryInput { canonical, typing_mode: delegate.typing_mode() }; - (orig_values, query_input) - } - - /// To return the constraints of a canonical query to the caller, we canonicalize: - /// - /// - `var_values`: a map from bound variables in the canonical goal to - /// the values inferred while solving the instantiated goal. - /// - `external_constraints`: additional constraints which aren't expressible - /// using simple unification of inference variables. - /// - /// This takes the `shallow_certainty` which represents whether we're confident - /// that the final result of the current goal only depends on the nested goals. - /// - /// In case this is `Certainty::Maybe`, there may still be additional nested goals - /// or inference constraints required for this candidate to be hold. The candidate - /// always requires all already added constraints and nested goals. - #[instrument(level = "trace", skip(self), ret)] - pub(in crate::solve) fn evaluate_added_goals_and_make_canonical_response( - &mut self, - shallow_certainty: Certainty, - ) -> QueryResult<I> { - self.inspect.make_canonical_response(shallow_certainty); - - let goals_certainty = self.try_evaluate_added_goals()?; - assert_eq!( - self.tainted, - Ok(()), - "EvalCtxt is tainted -- nested goals may have been dropped in a \ - previous call to `try_evaluate_added_goals!`" - ); - - // We only check for leaks from universes which were entered inside - // of the query. - self.delegate.leak_check(self.max_input_universe).map_err(|NoSolution| { - trace!("failed the leak check"); - NoSolution - })?; - - let (certainty, normalization_nested_goals) = - match (self.current_goal_kind, shallow_certainty) { - // When normalizing, we've replaced the expected term with an unconstrained - // inference variable. This means that we dropped information which could - // have been important. We handle this by instead returning the nested goals - // to the caller, where they are then handled. We only do so if we do not - // need to recompute the `NormalizesTo` goal afterwards to avoid repeatedly - // uplifting its nested goals. This is the case if the `shallow_certainty` is - // `Certainty::Yes`. - (CurrentGoalKind::NormalizesTo, Certainty::Yes) => { - let goals = std::mem::take(&mut self.nested_goals); - // As we return all ambiguous nested goals, we can ignore the certainty - // returned by `self.try_evaluate_added_goals()`. - if goals.is_empty() { - assert!(matches!(goals_certainty, Certainty::Yes)); - } - ( - Certainty::Yes, - NestedNormalizationGoals( - goals.into_iter().map(|(s, g, _)| (s, g)).collect(), - ), - ) - } - _ => { - let certainty = shallow_certainty.and(goals_certainty); - (certainty, NestedNormalizationGoals::empty()) - } - }; - - if let Certainty::Maybe { - cause: cause @ MaybeCause::Overflow { keep_constraints: false, .. }, - opaque_types_jank, - } = certainty - { - // If we have overflow, it's probable that we're substituting a type - // into itself infinitely and any partial substitutions in the query - // response are probably not useful anyways, so just return an empty - // query response. - // - // This may prevent us from potentially useful inference, e.g. - // 2 candidates, one ambiguous and one overflow, which both - // have the same inference constraints. - // - // Changing this to retain some constraints in the future - // won't be a breaking change, so this is good enough for now. - return Ok(self.make_ambiguous_response_no_constraints(cause, opaque_types_jank)); - } - - let external_constraints = - self.compute_external_query_constraints(certainty, normalization_nested_goals); - let (var_values, mut external_constraints) = - eager_resolve_vars(self.delegate, (self.var_values, external_constraints)); - - // Remove any trivial or duplicated region constraints once we've resolved regions - let mut unique = HashSet::default(); - external_constraints.region_constraints.retain(|outlives| { - outlives.0.as_region().is_none_or(|re| re != outlives.1) && unique.insert(*outlives) - }); - - let canonical = Canonicalizer::canonicalize_response( - self.delegate, - self.max_input_universe, - &mut Default::default(), - Response { - var_values, - certainty, - external_constraints: self.cx().mk_external_constraints(external_constraints), - }, - ); - - // HACK: We bail with overflow if the response would have too many non-region - // inference variables. This tends to only happen if we encounter a lot of - // ambiguous alias types which get replaced with fresh inference variables - // during generalization. This prevents hangs caused by an exponential blowup, - // see tests/ui/traits/next-solver/coherence-alias-hang.rs. - match self.current_goal_kind { - // We don't do so for `NormalizesTo` goals as we erased the expected term and - // bailing with overflow here would prevent us from detecting a type-mismatch, - // causing a coherence error in diesel, see #131969. We still bail with overflow - // when later returning from the parent AliasRelate goal. - CurrentGoalKind::NormalizesTo => {} - CurrentGoalKind::Misc | CurrentGoalKind::CoinductiveTrait => { - let num_non_region_vars = canonical - .variables - .iter() - .filter(|c| !c.is_region() && c.is_existential()) - .count(); - if num_non_region_vars > self.cx().recursion_limit() { - debug!(?num_non_region_vars, "too many inference variables -> overflow"); - return Ok(self.make_ambiguous_response_no_constraints( - MaybeCause::Overflow { - suggest_increasing_limit: true, - keep_constraints: false, - }, - OpaqueTypesJank::AllGood, - )); - } - } - } - - Ok(canonical) - } - - /// Constructs a totally unconstrained, ambiguous response to a goal. - /// - /// Take care when using this, since often it's useful to respond with - /// ambiguity but return constrained variables to guide inference. - pub(in crate::solve) fn make_ambiguous_response_no_constraints( - &self, - cause: MaybeCause, - opaque_types_jank: OpaqueTypesJank, - ) -> CanonicalResponse<I> { - response_no_constraints_raw( - self.cx(), - self.max_input_universe, - self.variables, - Certainty::Maybe { cause, opaque_types_jank }, - ) - } - - /// Computes the region constraints and *new* opaque types registered when - /// proving a goal. - /// - /// If an opaque was already constrained before proving this goal, then the - /// external constraints do not need to record that opaque, since if it is - /// further constrained by inference, that will be passed back in the var - /// values. - #[instrument(level = "trace", skip(self), ret)] - fn compute_external_query_constraints( - &self, - certainty: Certainty, - normalization_nested_goals: NestedNormalizationGoals<I>, - ) -> ExternalConstraintsData<I> { - // We only return region constraints once the certainty is `Yes`. This - // is necessary as we may drop nested goals on ambiguity, which may result - // in unconstrained inference variables in the region constraints. It also - // prevents us from emitting duplicate region constraints, avoiding some - // unnecessary work. This slightly weakens the leak check in case it uses - // region constraints from an ambiguous nested goal. This is tested in both - // `tests/ui/higher-ranked/leak-check/leak-check-in-selection-5-ambig.rs` and - // `tests/ui/higher-ranked/leak-check/leak-check-in-selection-6-ambig-unify.rs`. - let region_constraints = if certainty == Certainty::Yes { - self.delegate.make_deduplicated_outlives_constraints() - } else { - Default::default() - }; - - // We only return *newly defined* opaque types from canonical queries. - // - // Constraints for any existing opaque types are already tracked by changes - // to the `var_values`. - let opaque_types = self - .delegate - .clone_opaque_types_added_since(self.initial_opaque_types_storage_num_entries); - - ExternalConstraintsData { region_constraints, opaque_types, normalization_nested_goals } - } - - /// After calling a canonical query, we apply the constraints returned - /// by the query using this function. - /// - /// This happens in three steps: - /// - we instantiate the bound variables of the query response - /// - we unify the `var_values` of the response with the `original_values` - /// - we apply the `external_constraints` returned by the query, returning - /// the `normalization_nested_goals` - pub(super) fn instantiate_and_apply_query_response( - delegate: &D, - param_env: I::ParamEnv, - original_values: &[I::GenericArg], - response: CanonicalResponse<I>, - span: I::Span, - ) -> (NestedNormalizationGoals<I>, Certainty) { - let instantiation = Self::compute_query_response_instantiation_values( - delegate, - &original_values, - &response, - span, - ); - - let Response { var_values, external_constraints, certainty } = - delegate.instantiate_canonical(response, instantiation); - - Self::unify_query_var_values(delegate, param_env, &original_values, var_values, span); - - let ExternalConstraintsData { - region_constraints, - opaque_types, - normalization_nested_goals, - } = &*external_constraints; - - Self::register_region_constraints(delegate, region_constraints, span); - Self::register_new_opaque_types(delegate, opaque_types, span); - - (normalization_nested_goals.clone(), certainty) - } - - /// This returns the canonical variable values to instantiate the bound variables of - /// the canonical response. This depends on the `original_values` for the - /// bound variables. - fn compute_query_response_instantiation_values<T: ResponseT<I>>( - delegate: &D, - original_values: &[I::GenericArg], - response: &Canonical<I, T>, - span: I::Span, - ) -> CanonicalVarValues<I> { - // FIXME: Longterm canonical queries should deal with all placeholders - // created inside of the query directly instead of returning them to the - // caller. - let prev_universe = delegate.universe(); - let universes_created_in_query = response.max_universe.index(); - for _ in 0..universes_created_in_query { - delegate.create_next_universe(); - } - - let var_values = response.value.var_values(); - assert_eq!(original_values.len(), var_values.len()); - - // If the query did not make progress with constraining inference variables, - // we would normally create a new inference variables for bound existential variables - // only then unify this new inference variable with the inference variable from - // the input. - // - // We therefore instantiate the existential variable in the canonical response with the - // inference variable of the input right away, which is more performant. - let mut opt_values = IndexVec::from_elem_n(None, response.variables.len()); - for (original_value, result_value) in - iter::zip(original_values, var_values.var_values.iter()) - { - match result_value.kind() { - ty::GenericArgKind::Type(t) => { - // We disable the instantiation guess for inference variables - // and only use it for placeholders. We need to handle the - // `sub_root` of type inference variables which would make this - // more involved. They are also a lot rarer than region variables. - if let ty::Bound(debruijn, b) = t.kind() - && !matches!( - response.variables.get(b.var().as_usize()).unwrap(), - CanonicalVarKind::Ty { .. } - ) - { - assert_eq!(debruijn, ty::INNERMOST); - opt_values[b.var()] = Some(*original_value); - } - } - ty::GenericArgKind::Lifetime(r) => { - if let ty::ReBound(debruijn, br) = r.kind() { - assert_eq!(debruijn, ty::INNERMOST); - opt_values[br.var()] = Some(*original_value); - } - } - ty::GenericArgKind::Const(c) => { - if let ty::ConstKind::Bound(debruijn, bv) = c.kind() { - assert_eq!(debruijn, ty::INNERMOST); - opt_values[bv.var()] = Some(*original_value); - } - } - } - } - CanonicalVarValues::instantiate(delegate.cx(), response.variables, |var_values, kind| { - if kind.universe() != ty::UniverseIndex::ROOT { - // A variable from inside a binder of the query. While ideally these shouldn't - // exist at all (see the FIXME at the start of this method), we have to deal with - // them for now. - delegate.instantiate_canonical_var(kind, span, &var_values, |idx| { - prev_universe + idx.index() - }) - } else if kind.is_existential() { - // As an optimization we sometimes avoid creating a new inference variable here. - // - // All new inference variables we create start out in the current universe of the caller. - // This is conceptually wrong as these inference variables would be able to name - // more placeholders then they should be able to. However the inference variables have - // to "come from somewhere", so by equating them with the original values of the caller - // later on, we pull them down into their correct universe again. - if let Some(v) = opt_values[ty::BoundVar::from_usize(var_values.len())] { - v - } else { - delegate.instantiate_canonical_var(kind, span, &var_values, |_| prev_universe) - } - } else { - // For placeholders which were already part of the input, we simply map this - // universal bound variable back the placeholder of the input. - original_values[kind.expect_placeholder_index()] - } - }) - } - - /// Unify the `original_values` with the `var_values` returned by the canonical query.. - /// - /// This assumes that this unification will always succeed. This is the case when - /// applying a query response right away. However, calling a canonical query, doing any - /// other kind of trait solving, and only then instantiating the result of the query - /// can cause the instantiation to fail. This is not supported and we ICE in this case. - /// - /// We always structurally instantiate aliases. Relating aliases needs to be different - /// depending on whether the alias is *rigid* or not. We're only really able to tell - /// whether an alias is rigid by using the trait solver. When instantiating a response - /// from the solver we assume that the solver correctly handled aliases and therefore - /// always relate them structurally here. - #[instrument(level = "trace", skip(delegate))] - fn unify_query_var_values( - delegate: &D, - param_env: I::ParamEnv, - original_values: &[I::GenericArg], - var_values: CanonicalVarValues<I>, - span: I::Span, - ) { - assert_eq!(original_values.len(), var_values.len()); - - for (&orig, response) in iter::zip(original_values, var_values.var_values.iter()) { - let goals = - delegate.eq_structurally_relating_aliases(param_env, orig, response, span).unwrap(); - assert!(goals.is_empty()); - } - } - - fn register_region_constraints( - delegate: &D, - outlives: &[ty::OutlivesPredicate<I, I::GenericArg>], - span: I::Span, - ) { - for &ty::OutlivesPredicate(lhs, rhs) in outlives { - match lhs.kind() { - ty::GenericArgKind::Lifetime(lhs) => delegate.sub_regions(rhs, lhs, span), - ty::GenericArgKind::Type(lhs) => delegate.register_ty_outlives(lhs, rhs, span), - ty::GenericArgKind::Const(_) => panic!("const outlives: {lhs:?}: {rhs:?}"), - } - } - } - - fn register_new_opaque_types( - delegate: &D, - opaque_types: &[(ty::OpaqueTypeKey<I>, I::Ty)], - span: I::Span, - ) { - for &(key, ty) in opaque_types { - let prev = delegate.register_hidden_type_in_storage(key, ty, span); - // We eagerly resolve inference variables when computing the query response. - // This can cause previously distinct opaque type keys to now be structurally equal. - // - // To handle this, we store any duplicate entries in a separate list to check them - // at the end of typeck/borrowck. We could alternatively eagerly equate the hidden - // types here. However, doing so is difficult as it may result in nested goals and - // any errors may make it harder to track the control flow for diagnostics. - if let Some(prev) = prev { - delegate.add_duplicate_opaque_type(key, prev, span); - } - } - } -} - -/// Used by proof trees to be able to recompute intermediate actions while -/// evaluating a goal. The `var_values` not only include the bound variables -/// of the query input, but also contain all unconstrained inference vars -/// created while evaluating this goal. -pub(in crate::solve) fn make_canonical_state<D, T, I>( - delegate: &D, - var_values: &[I::GenericArg], - max_input_universe: ty::UniverseIndex, - data: T, -) -> inspect::CanonicalState<I, T> -where - D: SolverDelegate<Interner = I>, - I: Interner, - T: TypeFoldable<I>, -{ - let var_values = CanonicalVarValues { var_values: delegate.cx().mk_args(var_values) }; - let state = inspect::State { var_values, data }; - let state = eager_resolve_vars(delegate, state); - Canonicalizer::canonicalize_response(delegate, max_input_universe, &mut vec![], state) -} - -// FIXME: needs to be pub to be accessed by downstream -// `rustc_trait_selection::solve::inspect::analyse`. -pub fn instantiate_canonical_state<D, I, T: TypeFoldable<I>>( - delegate: &D, - span: I::Span, - param_env: I::ParamEnv, - orig_values: &mut Vec<I::GenericArg>, - state: inspect::CanonicalState<I, T>, -) -> T -where - D: SolverDelegate<Interner = I>, - I: Interner, -{ - // In case any fresh inference variables have been created between `state` - // and the previous instantiation, extend `orig_values` for it. - orig_values.extend( - state.value.var_values.var_values.as_slice()[orig_values.len()..] - .iter() - .map(|&arg| delegate.fresh_var_for_kind_with_span(arg, span)), - ); - - let instantiation = - EvalCtxt::compute_query_response_instantiation_values(delegate, orig_values, &state, span); - - let inspect::State { var_values, data } = delegate.instantiate_canonical(state, instantiation); - - EvalCtxt::unify_query_var_values(delegate, param_env, orig_values, var_values, span); - data -} diff --git a/compiler/rustc_next_trait_solver/src/solve/eval_ctxt/mod.rs b/compiler/rustc_next_trait_solver/src/solve/eval_ctxt/mod.rs index 5df7c92d881..bb86357a85f 100644 --- a/compiler/rustc_next_trait_solver/src/solve/eval_ctxt/mod.rs +++ b/compiler/rustc_next_trait_solver/src/solve/eval_ctxt/mod.rs @@ -17,6 +17,10 @@ use rustc_type_ir::{ use tracing::{debug, instrument, trace}; use super::has_only_region_constraints; +use crate::canonical::{ + canonicalize_goal, canonicalize_response, instantiate_and_apply_query_response, + response_no_constraints_raw, +}; use crate::coherence; use crate::delegate::SolverDelegate; use crate::placeholder::BoundVarReplacer; @@ -24,12 +28,11 @@ use crate::resolve::eager_resolve_vars; use crate::solve::search_graph::SearchGraph; use crate::solve::ty::may_use_unstable_feature; use crate::solve::{ - CanonicalInput, Certainty, FIXPOINT_STEP_LIMIT, Goal, GoalEvaluation, GoalSource, - GoalStalledOn, HasChanged, NestedNormalizationGoals, NoSolution, QueryInput, QueryResult, - inspect, + CanonicalInput, CanonicalResponse, Certainty, ExternalConstraintsData, FIXPOINT_STEP_LIMIT, + Goal, GoalEvaluation, GoalSource, GoalStalledOn, HasChanged, MaybeCause, + NestedNormalizationGoals, NoSolution, QueryInput, QueryResult, Response, inspect, }; -pub(super) mod canonical; mod probe; /// The kind of goal we're currently proving. @@ -464,8 +467,7 @@ where let opaque_types = self.delegate.clone_opaque_types_lookup_table(); let (goal, opaque_types) = eager_resolve_vars(self.delegate, (goal, opaque_types)); - let (orig_values, canonical_goal) = - Self::canonicalize_goal(self.delegate, goal, opaque_types); + let (orig_values, canonical_goal) = canonicalize_goal(self.delegate, goal, opaque_types); let canonical_result = self.search_graph.evaluate_goal( self.cx(), canonical_goal, @@ -480,7 +482,7 @@ where let has_changed = if !has_only_region_constraints(response) { HasChanged::Yes } else { HasChanged::No }; - let (normalization_nested_goals, certainty) = Self::instantiate_and_apply_query_response( + let (normalization_nested_goals, certainty) = instantiate_and_apply_query_response( self.delegate, goal.param_env, &orig_values, @@ -1223,6 +1225,198 @@ where vec![] } } + + /// To return the constraints of a canonical query to the caller, we canonicalize: + /// + /// - `var_values`: a map from bound variables in the canonical goal to + /// the values inferred while solving the instantiated goal. + /// - `external_constraints`: additional constraints which aren't expressible + /// using simple unification of inference variables. + /// + /// This takes the `shallow_certainty` which represents whether we're confident + /// that the final result of the current goal only depends on the nested goals. + /// + /// In case this is `Certainty::Maybe`, there may still be additional nested goals + /// or inference constraints required for this candidate to be hold. The candidate + /// always requires all already added constraints and nested goals. + #[instrument(level = "trace", skip(self), ret)] + pub(in crate::solve) fn evaluate_added_goals_and_make_canonical_response( + &mut self, + shallow_certainty: Certainty, + ) -> QueryResult<I> { + self.inspect.make_canonical_response(shallow_certainty); + + let goals_certainty = self.try_evaluate_added_goals()?; + assert_eq!( + self.tainted, + Ok(()), + "EvalCtxt is tainted -- nested goals may have been dropped in a \ + previous call to `try_evaluate_added_goals!`" + ); + + // We only check for leaks from universes which were entered inside + // of the query. + self.delegate.leak_check(self.max_input_universe).map_err(|NoSolution| { + trace!("failed the leak check"); + NoSolution + })?; + + let (certainty, normalization_nested_goals) = + match (self.current_goal_kind, shallow_certainty) { + // When normalizing, we've replaced the expected term with an unconstrained + // inference variable. This means that we dropped information which could + // have been important. We handle this by instead returning the nested goals + // to the caller, where they are then handled. We only do so if we do not + // need to recompute the `NormalizesTo` goal afterwards to avoid repeatedly + // uplifting its nested goals. This is the case if the `shallow_certainty` is + // `Certainty::Yes`. + (CurrentGoalKind::NormalizesTo, Certainty::Yes) => { + let goals = std::mem::take(&mut self.nested_goals); + // As we return all ambiguous nested goals, we can ignore the certainty + // returned by `self.try_evaluate_added_goals()`. + if goals.is_empty() { + assert!(matches!(goals_certainty, Certainty::Yes)); + } + ( + Certainty::Yes, + NestedNormalizationGoals( + goals.into_iter().map(|(s, g, _)| (s, g)).collect(), + ), + ) + } + _ => { + let certainty = shallow_certainty.and(goals_certainty); + (certainty, NestedNormalizationGoals::empty()) + } + }; + + if let Certainty::Maybe { + cause: cause @ MaybeCause::Overflow { keep_constraints: false, .. }, + opaque_types_jank, + } = certainty + { + // If we have overflow, it's probable that we're substituting a type + // into itself infinitely and any partial substitutions in the query + // response are probably not useful anyways, so just return an empty + // query response. + // + // This may prevent us from potentially useful inference, e.g. + // 2 candidates, one ambiguous and one overflow, which both + // have the same inference constraints. + // + // Changing this to retain some constraints in the future + // won't be a breaking change, so this is good enough for now. + return Ok(self.make_ambiguous_response_no_constraints(cause, opaque_types_jank)); + } + + let external_constraints = + self.compute_external_query_constraints(certainty, normalization_nested_goals); + let (var_values, mut external_constraints) = + eager_resolve_vars(self.delegate, (self.var_values, external_constraints)); + + // Remove any trivial or duplicated region constraints once we've resolved regions + let mut unique = HashSet::default(); + external_constraints.region_constraints.retain(|outlives| { + outlives.0.as_region().is_none_or(|re| re != outlives.1) && unique.insert(*outlives) + }); + + let canonical = canonicalize_response( + self.delegate, + self.max_input_universe, + Response { + var_values, + certainty, + external_constraints: self.cx().mk_external_constraints(external_constraints), + }, + ); + + // HACK: We bail with overflow if the response would have too many non-region + // inference variables. This tends to only happen if we encounter a lot of + // ambiguous alias types which get replaced with fresh inference variables + // during generalization. This prevents hangs caused by an exponential blowup, + // see tests/ui/traits/next-solver/coherence-alias-hang.rs. + match self.current_goal_kind { + // We don't do so for `NormalizesTo` goals as we erased the expected term and + // bailing with overflow here would prevent us from detecting a type-mismatch, + // causing a coherence error in diesel, see #131969. We still bail with overflow + // when later returning from the parent AliasRelate goal. + CurrentGoalKind::NormalizesTo => {} + CurrentGoalKind::Misc | CurrentGoalKind::CoinductiveTrait => { + let num_non_region_vars = canonical + .variables + .iter() + .filter(|c| !c.is_region() && c.is_existential()) + .count(); + if num_non_region_vars > self.cx().recursion_limit() { + debug!(?num_non_region_vars, "too many inference variables -> overflow"); + return Ok(self.make_ambiguous_response_no_constraints( + MaybeCause::Overflow { + suggest_increasing_limit: true, + keep_constraints: false, + }, + OpaqueTypesJank::AllGood, + )); + } + } + } + + Ok(canonical) + } + + /// Constructs a totally unconstrained, ambiguous response to a goal. + /// + /// Take care when using this, since often it's useful to respond with + /// ambiguity but return constrained variables to guide inference. + pub(in crate::solve) fn make_ambiguous_response_no_constraints( + &self, + cause: MaybeCause, + opaque_types_jank: OpaqueTypesJank, + ) -> CanonicalResponse<I> { + response_no_constraints_raw( + self.cx(), + self.max_input_universe, + self.variables, + Certainty::Maybe { cause, opaque_types_jank }, + ) + } + + /// Computes the region constraints and *new* opaque types registered when + /// proving a goal. + /// + /// If an opaque was already constrained before proving this goal, then the + /// external constraints do not need to record that opaque, since if it is + /// further constrained by inference, that will be passed back in the var + /// values. + #[instrument(level = "trace", skip(self), ret)] + fn compute_external_query_constraints( + &self, + certainty: Certainty, + normalization_nested_goals: NestedNormalizationGoals<I>, + ) -> ExternalConstraintsData<I> { + // We only return region constraints once the certainty is `Yes`. This + // is necessary as we may drop nested goals on ambiguity, which may result + // in unconstrained inference variables in the region constraints. It also + // prevents us from emitting duplicate region constraints, avoiding some + // unnecessary work. This slightly weakens the leak check in case it uses + // region constraints from an ambiguous nested goal. This is tested in both + // `tests/ui/higher-ranked/leak-check/leak-check-in-selection-5-ambig.rs` and + // `tests/ui/higher-ranked/leak-check/leak-check-in-selection-6-ambig-unify.rs`. + let region_constraints = if certainty == Certainty::Yes { + self.delegate.make_deduplicated_outlives_constraints() + } else { + Default::default() + }; + + // We only return *newly defined* opaque types from canonical queries. + // + // Constraints for any existing opaque types are already tracked by changes + // to the `var_values`. + let opaque_types = self + .delegate + .clone_opaque_types_added_since(self.initial_opaque_types_storage_num_entries); + + ExternalConstraintsData { region_constraints, opaque_types, normalization_nested_goals } + } } /// Eagerly replace aliases with inference variables, emitting `AliasRelate` @@ -1363,7 +1557,7 @@ pub(super) fn evaluate_root_goal_for_proof_tree<D: SolverDelegate<Interner = I>, let opaque_types = delegate.clone_opaque_types_lookup_table(); let (goal, opaque_types) = eager_resolve_vars(delegate, (goal, opaque_types)); - let (orig_values, canonical_goal) = EvalCtxt::canonicalize_goal(delegate, goal, opaque_types); + let (orig_values, canonical_goal) = canonicalize_goal(delegate, goal, opaque_types); let (canonical_result, final_revision) = delegate.cx().evaluate_root_goal_for_proof_tree_raw(canonical_goal); @@ -1380,7 +1574,7 @@ pub(super) fn evaluate_root_goal_for_proof_tree<D: SolverDelegate<Interner = I>, Ok(response) => response, }; - let (normalization_nested_goals, _certainty) = EvalCtxt::instantiate_and_apply_query_response( + let (normalization_nested_goals, _certainty) = instantiate_and_apply_query_response( delegate, goal.param_env, &proof_tree.orig_values, diff --git a/compiler/rustc_next_trait_solver/src/solve/inspect/build.rs b/compiler/rustc_next_trait_solver/src/solve/inspect/build.rs index 2675ed0d0da..4369148baf9 100644 --- a/compiler/rustc_next_trait_solver/src/solve/inspect/build.rs +++ b/compiler/rustc_next_trait_solver/src/solve/inspect/build.rs @@ -10,8 +10,8 @@ use derive_where::derive_where; use rustc_type_ir::inherent::*; use rustc_type_ir::{self as ty, Interner}; +use crate::canonical; use crate::delegate::SolverDelegate; -use crate::solve::eval_ctxt::canonical; use crate::solve::{Certainty, Goal, GoalSource, QueryResult, inspect}; /// We need to know whether to build a prove tree while evaluating. We diff --git a/compiler/rustc_next_trait_solver/src/solve/inspect/mod.rs b/compiler/rustc_next_trait_solver/src/solve/inspect/mod.rs index 0d8c0060126..65f32f1947f 100644 --- a/compiler/rustc_next_trait_solver/src/solve/inspect/mod.rs +++ b/compiler/rustc_next_trait_solver/src/solve/inspect/mod.rs @@ -2,5 +2,3 @@ pub use rustc_type_ir::solve::inspect::*; mod build; pub(in crate::solve) use build::*; - -pub use crate::solve::eval_ctxt::canonical::instantiate_canonical_state; diff --git a/compiler/rustc_next_trait_solver/src/solve/mod.rs b/compiler/rustc_next_trait_solver/src/solve/mod.rs index fb900b592d1..afb86aaf8ab 100644 --- a/compiler/rustc_next_trait_solver/src/solve/mod.rs +++ b/compiler/rustc_next_trait_solver/src/solve/mod.rs @@ -380,25 +380,6 @@ where } } -fn response_no_constraints_raw<I: Interner>( - cx: I, - max_universe: ty::UniverseIndex, - variables: I::CanonicalVarKinds, - certainty: Certainty, -) -> CanonicalResponse<I> { - ty::Canonical { - max_universe, - variables, - value: Response { - var_values: ty::CanonicalVarValues::make_identity(cx, variables), - // FIXME: maybe we should store the "no response" version in cx, like - // we do for cx.types and stuff. - external_constraints: cx.mk_external_constraints(ExternalConstraintsData::default()), - certainty, - }, - } -} - /// The result of evaluating a goal. pub struct GoalEvaluation<I: Interner> { /// The goal we've evaluated. This is the input goal, but potentially with its diff --git a/compiler/rustc_next_trait_solver/src/solve/search_graph.rs b/compiler/rustc_next_trait_solver/src/solve/search_graph.rs index 289325d7055..aa9dfc9a9a2 100644 --- a/compiler/rustc_next_trait_solver/src/solve/search_graph.rs +++ b/compiler/rustc_next_trait_solver/src/solve/search_graph.rs @@ -6,6 +6,7 @@ use rustc_type_ir::search_graph::{self, PathKind}; use rustc_type_ir::solve::{CanonicalInput, Certainty, NoSolution, QueryResult}; use rustc_type_ir::{Interner, TypingMode}; +use crate::canonical::response_no_constraints_raw; use crate::delegate::SolverDelegate; use crate::solve::{ EvalCtxt, FIXPOINT_STEP_LIMIT, has_no_inference_or_external_constraints, inspect, @@ -127,7 +128,7 @@ fn response_no_constraints<I: Interner>( input: CanonicalInput<I>, certainty: Certainty, ) -> QueryResult<I> { - Ok(super::response_no_constraints_raw( + Ok(response_no_constraints_raw( cx, input.canonical.max_universe, input.canonical.variables, diff --git a/compiler/rustc_session/src/config.rs b/compiler/rustc_session/src/config.rs index 297df7c2c97..795cb2b2cfe 100644 --- a/compiler/rustc_session/src/config.rs +++ b/compiler/rustc_session/src/config.rs @@ -1509,6 +1509,11 @@ impl Options { pub fn get_symbol_mangling_version(&self) -> SymbolManglingVersion { self.cg.symbol_mangling_version.unwrap_or(SymbolManglingVersion::Legacy) } + + #[inline] + pub fn autodiff_enabled(&self) -> bool { + self.unstable_opts.autodiff.contains(&AutoDiff::Enable) + } } impl UnstableOptions { diff --git a/compiler/rustc_session/src/config/cfg.rs b/compiler/rustc_session/src/config/cfg.rs index 8f63ce6f0ae..7e970461ab7 100644 --- a/compiler/rustc_session/src/config/cfg.rs +++ b/compiler/rustc_session/src/config/cfg.rs @@ -374,11 +374,13 @@ impl CheckCfg { ins!(sym::overflow_checks, no_values); - ins!(sym::panic, empty_values).extend(&PanicStrategy::all()); + ins!(sym::panic, empty_values) + .extend(PanicStrategy::ALL.iter().map(PanicStrategy::desc_symbol)); ins!(sym::proc_macro, no_values); - ins!(sym::relocation_model, empty_values).extend(RelocModel::all()); + ins!(sym::relocation_model, empty_values) + .extend(RelocModel::ALL.iter().map(RelocModel::desc_symbol)); let sanitize_values = SanitizerSet::all() .into_iter() diff --git a/compiler/rustc_session/src/session.rs b/compiler/rustc_session/src/session.rs index 3525c7c1d1a..d0dd2cdac0c 100644 --- a/compiler/rustc_session/src/session.rs +++ b/compiler/rustc_session/src/session.rs @@ -600,6 +600,13 @@ impl Session { /// Calculates the flavor of LTO to use for this compilation. pub fn lto(&self) -> config::Lto { + // Autodiff currently requires fat-lto to have access to the llvm-ir of all (indirectly) used functions and types. + // fat-lto is the easiest solution to this requirement, but quite expensive. + // FIXME(autodiff): Make autodiff also work with embed-bc instead of fat-lto. + if self.opts.autodiff_enabled() { + return config::Lto::Fat; + } + // If our target has codegen requirements ignore the command line if self.target.requires_lto { return config::Lto::Fat; diff --git a/compiler/rustc_target/src/lib.rs b/compiler/rustc_target/src/lib.rs index b3fe1fffcce..8c6a77cba8b 100644 --- a/compiler/rustc_target/src/lib.rs +++ b/compiler/rustc_target/src/lib.rs @@ -76,10 +76,10 @@ fn find_relative_libdir(sysroot: &Path) -> std::borrow::Cow<'static, str> { macro_rules! target_spec_enum { ( $( #[$attr:meta] )* - pub enum $name:ident { + pub enum $Name:ident { $( $( #[$variant_attr:meta] )* - $variant:ident = $string:literal, + $Variant:ident = $string:literal, )* } parse_error_type = $parse_error_type:literal; @@ -87,20 +87,20 @@ macro_rules! target_spec_enum { $( #[$attr] )* #[derive(Clone, Copy, PartialEq, Eq, Hash, Debug, PartialOrd, Ord)] #[derive(schemars::JsonSchema)] - pub enum $name { + pub enum $Name { $( $( #[$variant_attr] )* #[serde(rename = $string)] // for JSON schema generation only - $variant, + $Variant, )* } - impl FromStr for $name { + impl FromStr for $Name { type Err = String; fn from_str(s: &str) -> Result<Self, Self::Err> { Ok(match s { - $( $string => Self::$variant, )* + $( $string => Self::$Variant, )* _ => { let all = [$( concat!("'", $string, "'") ),*].join(", "); return Err(format!("invalid {}: '{s}'. allowed values: {all}", $parse_error_type)); @@ -109,24 +109,25 @@ macro_rules! target_spec_enum { } } - impl $name { + impl $Name { + pub const ALL: &'static [$Name] = &[ $( $Name::$Variant, )* ]; pub fn desc(&self) -> &'static str { match self { - $( Self::$variant => $string, )* + $( Self::$Variant => $string, )* } } } - impl crate::json::ToJson for $name { + impl crate::json::ToJson for $Name { fn to_json(&self) -> crate::json::Json { self.desc().to_json() } } - crate::json::serde_deserialize_from_str!($name); + crate::json::serde_deserialize_from_str!($Name); - impl std::fmt::Display for $name { + impl std::fmt::Display for $Name { fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result { f.write_str(self.desc()) } diff --git a/compiler/rustc_target/src/spec/mod.rs b/compiler/rustc_target/src/spec/mod.rs index 07fb1ce63f7..f705af52bd8 100644 --- a/compiler/rustc_target/src/spec/mod.rs +++ b/compiler/rustc_target/src/spec/mod.rs @@ -854,10 +854,6 @@ impl PanicStrategy { PanicStrategy::Abort => sym::abort, } } - - pub const fn all() -> [Symbol; 2] { - [Self::Abort.desc_symbol(), Self::Unwind.desc_symbol()] - } } crate::target_spec_enum! { @@ -974,18 +970,6 @@ impl RelocModel { RelocModel::RopiRwpi => sym::ropi_rwpi, } } - - pub const fn all() -> [Symbol; 7] { - [ - RelocModel::Static.desc_symbol(), - RelocModel::Pic.desc_symbol(), - RelocModel::Pie.desc_symbol(), - RelocModel::DynamicNoPic.desc_symbol(), - RelocModel::Ropi.desc_symbol(), - RelocModel::Rwpi.desc_symbol(), - RelocModel::RopiRwpi.desc_symbol(), - ] - } } crate::target_spec_enum! { diff --git a/compiler/rustc_trait_selection/src/solve/inspect/analyse.rs b/compiler/rustc_trait_selection/src/solve/inspect/analyse.rs index 086a7a44786..c010add0fc5 100644 --- a/compiler/rustc_trait_selection/src/solve/inspect/analyse.rs +++ b/compiler/rustc_trait_selection/src/solve/inspect/analyse.rs @@ -18,9 +18,9 @@ use rustc_middle::traits::ObligationCause; use rustc_middle::traits::solve::{Certainty, Goal, GoalSource, NoSolution, QueryResult}; use rustc_middle::ty::{TyCtxt, VisitorResult, try_visit}; use rustc_middle::{bug, ty}; +use rustc_next_trait_solver::canonical::instantiate_canonical_state; use rustc_next_trait_solver::resolve::eager_resolve_vars; -use rustc_next_trait_solver::solve::inspect::{self, instantiate_canonical_state}; -use rustc_next_trait_solver::solve::{MaybeCause, SolverDelegateEvalExt as _}; +use rustc_next_trait_solver::solve::{MaybeCause, SolverDelegateEvalExt as _, inspect}; use rustc_span::Span; use tracing::instrument; diff --git a/compiler/rustc_windows_rc/src/lib.rs b/compiler/rustc_windows_rc/src/lib.rs index caa5e5ef276..5e95557501e 100644 --- a/compiler/rustc_windows_rc/src/lib.rs +++ b/compiler/rustc_windows_rc/src/lib.rs @@ -35,8 +35,11 @@ pub fn compile_windows_resource_file( resources_dir.push("resources"); fs::create_dir_all(&resources_dir).unwrap(); - let resource_compiler = - find_resource_compiler(&env::var("CARGO_CFG_TARGET_ARCH").unwrap()).expect("found rc.exe"); + let resource_compiler = if let Ok(path) = env::var("RUSTC_WINDOWS_RC") { + path.into() + } else { + find_resource_compiler(&env::var("CARGO_CFG_TARGET_ARCH").unwrap()).expect("found rc.exe") + }; let rc_path = resources_dir.join(file_stem.with_extension("rc")); |