Auto merge of rust-lang#96285 - flip1995:pk-vfe, r=nagisa

Introduce `-Zvirtual-function-elimination` codegen flag

Fixes rust-lang#68262

This PR adds a codegen flag `-Zvirtual-function-elimination` to enable the VFE optimization in LLVM. To make this work, additonal  information has to be added to vtables ([`!vcall_visibility` metadata](https://llvm.org/docs/TypeMetadata.html#vcall-visibility-metadata) and a `typeid` of the trait). Furthermore, instead of just `load`ing functions, the [`llvm.type.checked.load` intrinsic](https://llvm.org/docs/LangRef.html#llvm-type-checked-load-intrinsic) has to be used to map functions to vtables.

For technical details of the changes, see the commit messages.

I also tested this flag on https://github.com/tock/tock on different boards to verify that this fixes the issue tock/tock#2594. This flag is able to improve the size of the resulting binary by about 8k-9k bytes by removing the unused debug print functions.

[Rendered documentation update](https://github.com/flip1995/rust/blob/pk-vfe/src/doc/rustc/src/codegen-options/index.md#virtual-function-elimination)

Author: bors

Cargo.lock

@@ -3695,6 +3695,7 @@ dependencies = [
  "rustc_serialize",
  "rustc_session",
  "rustc_span",
+ "rustc_symbol_mangling",
  "rustc_target",
  "smallvec",
  "tracing",

compiler/rustc_codegen_gcc/src/intrinsic/mod.rs

@@ -356,6 +356,16 @@ impl<'a, 'gcc, 'tcx> IntrinsicCallMethods<'tcx> for Builder<'a, 'gcc, 'tcx> {
         self.context.new_rvalue_from_int(self.int_type, 0)
     }
 
+    fn type_checked_load(
+        &mut self,
+        _llvtable: Self::Value,
+        _vtable_byte_offset: u64,
+        _typeid: Self::Value,
+    ) -> Self::Value {
+        // Unsupported.
+        self.context.new_rvalue_from_int(self.int_type, 0)
+    }
+
     fn va_start(&mut self, _va_list: RValue<'gcc>) -> RValue<'gcc> {
         unimplemented!();
     }

compiler/rustc_codegen_llvm/Cargo.toml

@@ -19,6 +19,7 @@ rustc-demangle = "0.1.21"
 rustc_arena = { path = "../rustc_arena" }
 rustc_attr = { path = "../rustc_attr" }
 rustc_codegen_ssa = { path = "../rustc_codegen_ssa" }
+rustc_symbol_mangling = { path = "../rustc_symbol_mangling" }
 rustc_data_structures = { path = "../rustc_data_structures" }
 rustc_errors = { path = "../rustc_errors" }
 rustc_fs_util = { path = "../rustc_fs_util" }

compiler/rustc_codegen_llvm/src/back/lto.rs

@@ -586,9 +586,21 @@ pub(crate) fn run_pass_manager(
     // LTO-specific optimization passes that LLVM provides.
     //
     // This code is based off the code found in llvm's LTO code generator:
-    //      tools/lto/LTOCodeGenerator.cpp
+    //      llvm/lib/LTO/LTOCodeGenerator.cpp
     debug!("running the pass manager");
     unsafe {
+        if !llvm::LLVMRustHasModuleFlag(
+            module.module_llvm.llmod(),
+            "LTOPostLink".as_ptr().cast(),
+            11,
+        ) {
+            llvm::LLVMRustAddModuleFlag(
+                module.module_llvm.llmod(),
+                llvm::LLVMModFlagBehavior::Error,
+                "LTOPostLink\0".as_ptr().cast(),
+                1,
+            );
+        }
         if llvm_util::should_use_new_llvm_pass_manager(
             &config.new_llvm_pass_manager,
             &cgcx.target_arch,

compiler/rustc_codegen_llvm/src/context.rs

@@ -326,6 +326,15 @@ pub unsafe fn create_module<'ll>(
         )
     }
 
+    if sess.opts.debugging_opts.virtual_function_elimination {
+        llvm::LLVMRustAddModuleFlag(
+            llmod,
+            llvm::LLVMModFlagBehavior::Error,
+            "Virtual Function Elim\0".as_ptr().cast(),
+            1,
+        );
+    }
+
     llmod
 }
 
@@ -656,6 +665,7 @@ impl<'ll> CodegenCx<'ll, '_> {
         let t_isize = self.type_isize();
         let t_f32 = self.type_f32();
         let t_f64 = self.type_f64();
+        let t_metadata = self.type_metadata();
 
         ifn!("llvm.wasm.trunc.unsigned.i32.f32", fn(t_f32) -> t_i32);
         ifn!("llvm.wasm.trunc.unsigned.i32.f64", fn(t_f64) -> t_i32);
@@ -881,11 +891,12 @@ impl<'ll> CodegenCx<'ll, '_> {
             ifn!("llvm.instrprof.increment", fn(i8p, t_i64, t_i32, t_i32) -> void);
         }
 
-        ifn!("llvm.type.test", fn(i8p, self.type_metadata()) -> i1);
+        ifn!("llvm.type.test", fn(i8p, t_metadata) -> i1);
+        ifn!("llvm.type.checked.load", fn(i8p, t_i32, t_metadata) -> mk_struct! {i8p, i1});
 
         if self.sess().opts.debuginfo != DebugInfo::None {
-            ifn!("llvm.dbg.declare", fn(self.type_metadata(), self.type_metadata()) -> void);
-            ifn!("llvm.dbg.value", fn(self.type_metadata(), t_i64, self.type_metadata()) -> void);
+            ifn!("llvm.dbg.declare", fn(t_metadata, t_metadata) -> void);
+            ifn!("llvm.dbg.value", fn(t_metadata, t_i64, t_metadata) -> void);
         }
         None
     }

compiler/rustc_codegen_llvm/src/debuginfo/metadata.rs

@@ -30,20 +30,21 @@ use rustc_hir::def_id::{DefId, LOCAL_CRATE};
 use rustc_index::vec::{Idx, IndexVec};
 use rustc_middle::bug;
 use rustc_middle::mir::{self, GeneratorLayout};
-use rustc_middle::ty::layout::LayoutOf;
-use rustc_middle::ty::layout::TyAndLayout;
+use rustc_middle::ty::layout::{LayoutOf, TyAndLayout};
 use rustc_middle::ty::subst::GenericArgKind;
-use rustc_middle::ty::{self, AdtKind, Instance, ParamEnv, Ty, TyCtxt};
-use rustc_session::config::{self, DebugInfo};
+use rustc_middle::ty::{
+    self, AdtKind, Instance, ParamEnv, PolyExistentialTraitRef, Ty, TyCtxt, Visibility,
+};
+use rustc_session::config::{self, DebugInfo, Lto};
 use rustc_span::symbol::Symbol;
 use rustc_span::FileName;
-use rustc_span::FileNameDisplayPreference;
-use rustc_span::{self, SourceFile};
+use rustc_span::{self, FileNameDisplayPreference, SourceFile};
+use rustc_symbol_mangling::typeid_for_trait_ref;
 use rustc_target::abi::{Align, Size};
 use smallvec::smallvec;
 use tracing::debug;
 
-use libc::{c_longlong, c_uint};
+use libc::{c_char, c_longlong, c_uint};
 use std::borrow::Cow;
 use std::fmt::{self, Write};
 use std::hash::{Hash, Hasher};
@@ -1468,6 +1469,84 @@ fn build_vtable_type_di_node<'ll, 'tcx>(
     .di_node
 }
 
+fn vcall_visibility_metadata<'ll, 'tcx>(
+    cx: &CodegenCx<'ll, 'tcx>,
+    ty: Ty<'tcx>,
+    trait_ref: Option<PolyExistentialTraitRef<'tcx>>,
+    vtable: &'ll Value,
+) {
+    enum VCallVisibility {
+        Public = 0,
+        LinkageUnit = 1,
+        TranslationUnit = 2,
+    }
+
+    let Some(trait_ref) = trait_ref else { return };
+
+    let trait_ref_self = trait_ref.with_self_ty(cx.tcx, ty);
+    let trait_ref_self = cx.tcx.erase_regions(trait_ref_self);
+    let trait_def_id = trait_ref_self.def_id();
+    let trait_vis = cx.tcx.visibility(trait_def_id);
+
+    let cgus = cx.sess().codegen_units();
+    let single_cgu = cgus == 1;
+
+    let lto = cx.sess().lto();
+
+    // Since LLVM requires full LTO for the virtual function elimination optimization to apply,
+    // only the `Lto::Fat` cases are relevant currently.
+    let vcall_visibility = match (lto, trait_vis, single_cgu) {
+        // If there is not LTO and the visibility in public, we have to assume that the vtable can
+        // be seen from anywhere. With multiple CGUs, the vtable is quasi-public.
+        (Lto::No | Lto::ThinLocal, Visibility::Public, _)
+        | (Lto::No, Visibility::Restricted(_) | Visibility::Invisible, false) => {
+            VCallVisibility::Public
+        }
+        // With LTO and a quasi-public visibility, the usages of the functions of the vtable are
+        // all known by the `LinkageUnit`.
+        // FIXME: LLVM only supports this optimization for `Lto::Fat` currently. Once it also
+        // supports `Lto::Thin` the `VCallVisibility` may have to be adjusted for those.
+        (Lto::Fat | Lto::Thin, Visibility::Public, _)
+        | (
+            Lto::ThinLocal | Lto::Thin | Lto::Fat,
+            Visibility::Restricted(_) | Visibility::Invisible,
+            false,
+        ) => VCallVisibility::LinkageUnit,
+        // If there is only one CGU, private vtables can only be seen by that CGU/translation unit
+        // and therefore we know of all usages of functions in the vtable.
+        (_, Visibility::Restricted(_) | Visibility::Invisible, true) => {
+            VCallVisibility::TranslationUnit
+        }
+    };
+
+    let trait_ref_typeid = typeid_for_trait_ref(cx.tcx, trait_ref);
+
+    unsafe {
+        let typeid = llvm::LLVMMDStringInContext(
+            cx.llcx,
+            trait_ref_typeid.as_ptr() as *const c_char,
+            trait_ref_typeid.as_bytes().len() as c_uint,
+        );
+        let v = [cx.const_usize(0), typeid];
+        llvm::LLVMRustGlobalAddMetadata(
+            vtable,
+            llvm::MD_type as c_uint,
+            llvm::LLVMValueAsMetadata(llvm::LLVMMDNodeInContext(
+                cx.llcx,
+                v.as_ptr(),
+                v.len() as c_uint,
+            )),
+        );
+        let vcall_visibility = llvm::LLVMValueAsMetadata(cx.const_u64(vcall_visibility as u64));
+        let vcall_visibility_metadata = llvm::LLVMMDNodeInContext2(cx.llcx, &vcall_visibility, 1);
+        llvm::LLVMGlobalSetMetadata(
+            vtable,
+            llvm::MetadataType::MD_vcall_visibility as c_uint,
+            vcall_visibility_metadata,
+        );
+    }
+}
+
 /// Creates debug information for the given vtable, which is for the
 /// given type.
 ///
@@ -1478,6 +1557,12 @@ pub fn create_vtable_di_node<'ll, 'tcx>(
     poly_trait_ref: Option<ty::PolyExistentialTraitRef<'tcx>>,
     vtable: &'ll Value,
 ) {
+    // FIXME(flip1995): The virtual function elimination optimization only works with full LTO in
+    // LLVM at the moment.
+    if cx.sess().opts.debugging_opts.virtual_function_elimination && cx.sess().lto() == Lto::Fat {
+        vcall_visibility_metadata(cx, ty, poly_trait_ref, vtable);
+    }
+
     if cx.dbg_cx.is_none() {
         return;
     }

compiler/rustc_codegen_llvm/src/intrinsic.rs

@@ -406,6 +406,16 @@ impl<'ll, 'tcx> IntrinsicCallMethods<'tcx> for Builder<'_, 'll, 'tcx> {
         self.call_intrinsic("llvm.type.test", &[bitcast, typeid])
     }
 
+    fn type_checked_load(
+        &mut self,
+        llvtable: &'ll Value,
+        vtable_byte_offset: u64,
+        typeid: &'ll Value,
+    ) -> Self::Value {
+        let vtable_byte_offset = self.const_i32(vtable_byte_offset as i32);
+        self.call_intrinsic("llvm.type.checked.load", &[llvtable, vtable_byte_offset, typeid])
+    }
+
     fn va_start(&mut self, va_list: &'ll Value) -> &'ll Value {
         self.call_intrinsic("llvm.va_start", &[va_list])
     }

compiler/rustc_codegen_llvm/src/llvm/ffi.rs

@@ -442,6 +442,7 @@ pub enum MetadataType {
     MD_nonnull = 11,
     MD_align = 17,
     MD_type = 19,
+    MD_vcall_visibility = 28,
     MD_noundef = 29,
 }
 
@@ -1067,6 +1068,7 @@ extern "C" {
     pub fn LLVMReplaceAllUsesWith<'a>(OldVal: &'a Value, NewVal: &'a Value);
     pub fn LLVMSetMetadata<'a>(Val: &'a Value, KindID: c_uint, Node: &'a Value);
     pub fn LLVMGlobalSetMetadata<'a>(Val: &'a Value, KindID: c_uint, Metadata: &'a Metadata);
+    pub fn LLVMRustGlobalAddMetadata<'a>(Val: &'a Value, KindID: c_uint, Metadata: &'a Metadata);
     pub fn LLVMValueAsMetadata(Node: &Value) -> &Metadata;
 
     // Operations on constants of any type
@@ -1080,6 +1082,11 @@ extern "C" {
         Vals: *const &'a Value,
         Count: c_uint,
     ) -> &'a Value;
+    pub fn LLVMMDNodeInContext2<'a>(
+        C: &'a Context,
+        Vals: *const &'a Metadata,
+        Count: size_t,
+    ) -> &'a Metadata;
     pub fn LLVMAddNamedMetadataOperand<'a>(M: &'a Module, Name: *const c_char, Val: &'a Value);
 
     // Operations on scalar constants
@@ -1936,6 +1943,7 @@ extern "C" {
         name: *const c_char,
         value: u32,
     );
+    pub fn LLVMRustHasModuleFlag(M: &Module, name: *const c_char, len: size_t) -> bool;
 
     pub fn LLVMRustMetadataAsValue<'a>(C: &'a Context, MD: &'a Metadata) -> &'a Value;
 

compiler/rustc_codegen_ssa/src/meth.rs

@@ -1,6 +1,8 @@
 use crate::traits::*;
 
-use rustc_middle::ty::{self, Ty};
+use rustc_middle::ty::{self, subst::GenericArgKind, ExistentialPredicate, Ty, TyCtxt};
+use rustc_session::config::Lto;
+use rustc_symbol_mangling::typeid_for_trait_ref;
 use rustc_target::abi::call::FnAbi;
 
 #[derive(Copy, Clone, Debug)]
@@ -15,20 +17,32 @@ impl<'a, 'tcx> VirtualIndex {
         self,
         bx: &mut Bx,
         llvtable: Bx::Value,
+        ty: Ty<'tcx>,
         fn_abi: &FnAbi<'tcx, Ty<'tcx>>,
     ) -> Bx::Value {
         // Load the data pointer from the object.
-        debug!("get_fn({:?}, {:?})", llvtable, self);
-
+        debug!("get_fn({llvtable:?}, {ty:?}, {self:?})");
         let llty = bx.fn_ptr_backend_type(fn_abi);
         let llvtable = bx.pointercast(llvtable, bx.type_ptr_to(llty));
-        let ptr_align = bx.tcx().data_layout.pointer_align.abi;
-        let gep = bx.inbounds_gep(llty, llvtable, &[bx.const_usize(self.0)]);
-        let ptr = bx.load(llty, gep, ptr_align);
-        bx.nonnull_metadata(ptr);
-        // Vtable loads are invariant.
-        bx.set_invariant_load(ptr);
-        ptr
+
+        if bx.cx().sess().opts.debugging_opts.virtual_function_elimination
+            && bx.cx().sess().lto() == Lto::Fat
+        {
+            let typeid =
+                bx.typeid_metadata(typeid_for_trait_ref(bx.tcx(), get_trait_ref(bx.tcx(), 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);
+            bx.pointercast(func, llty)
+        } else {
+            let ptr_align = bx.tcx().data_layout.pointer_align.abi;
+            let gep = bx.inbounds_gep(llty, llvtable, &[bx.const_usize(self.0)]);
+            let ptr = bx.load(llty, gep, ptr_align);
+            bx.nonnull_metadata(ptr);
+            // Vtable loads are invariant.
+            bx.set_invariant_load(ptr);
+            ptr
+        }
     }
 
     pub fn get_usize<Bx: BuilderMethods<'a, 'tcx>>(
@@ -50,6 +64,24 @@ impl<'a, 'tcx> VirtualIndex {
     }
 }
 
+fn get_trait_ref<'tcx>(tcx: TyCtxt<'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")
+                    }
+                });
+            }
+        }
+    }
+
+    bug!("expected a `dyn Trait` ty, found {ty:?}")
+}
+
 /// Creates a dynamic vtable for the given type and vtable origin.
 /// This is used only for objects.
 ///

compiler/rustc_codegen_ssa/src/mir/block.rs

@@ -401,7 +401,7 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
                 args = &args[..1];
                 (
                     meth::VirtualIndex::from_index(ty::COMMON_VTABLE_ENTRIES_DROPINPLACE)
-                        .get_fn(&mut bx, vtable, &fn_abi),
+                        .get_fn(&mut bx, vtable, ty, &fn_abi),
                     fn_abi,
                 )
             }
@@ -819,17 +819,25 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
                     // 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, &fn_abi),
-                            );
+                            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),
                     }
                 } 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, &fn_abi));
+                    llfn = Some(meth::VirtualIndex::from_index(idx).get_fn(
+                        &mut bx,
+                        meta,
+                        op.layout.ty,
+                        &fn_abi,
+                    ));
                     llargs.push(data_ptr);
                     continue;
                 } else {