add a "raw" function definition interface to cranelift-module (#1400)

* move trap site definitions into cranelift-module

`cranelift-faerie` and `cranelift-object` already have identical
definitions of structures to represent trap sites.  We might as well
merge them ahead of work to define functions via a raw slice of bytes
with associated traps, which will need some kind of common structure for
representing traps anyway.

* cranelift-module: add `define_function_bytes` interface

This interface is useful when the client needs to precisely specify the
ordering of bytes in a particular function.

* add comment about saving files for `perf`

cranelift/faerie/src/backend.rs

@@ -10,9 +10,10 @@ use cranelift_codegen::isa::TargetIsa;
 use cranelift_codegen::{self, binemit, ir};
 use cranelift_module::{
     Backend, DataContext, DataDescription, DataId, FuncId, Init, Linkage, ModuleError,
-    ModuleNamespace, ModuleResult,
+    ModuleNamespace, ModuleResult, TrapSite,
 };
 use faerie;
+use std::convert::TryInto;
 use std::fs::File;
 use target_lexicon::Triple;
 
@@ -200,6 +201,31 @@ impl Backend for FaerieBackend {
         Ok(FaerieCompiledFunction { code_length })
     }
 
+    fn define_function_bytes(
+        &mut self,
+        _id: FuncId,
+        name: &str,
+        bytes: &[u8],
+        _namespace: &ModuleNamespace<Self>,
+        traps: Vec<TrapSite>,
+    ) -> ModuleResult<FaerieCompiledFunction> {
+        let code_length: u32 = match bytes.len().try_into() {
+            Ok(code_length) => code_length,
+            _ => Err(ModuleError::FunctionTooLarge(name.to_string()))?,
+        };
+
+        if let Some(ref mut trap_manifest) = self.trap_manifest {
+            let trap_sink = FaerieTrapSink::new_with_sites(name, code_length, traps);
+            trap_manifest.add_sink(trap_sink);
+        }
+
+        self.artifact
+            .define(name, bytes.to_vec())
+            .expect("inconsistent declaration");
+
+        Ok(FaerieCompiledFunction { code_length })
+    }
+
     fn define_data(
         &mut self,
         _id: DataId,

cranelift/faerie/src/traps.rs

@@ -2,17 +2,7 @@
 //! for every function in the module. This data may be useful at runtime.
 
 use cranelift_codegen::{binemit, ir};
-
+use cranelift_module::TrapSite;
-/// Record of the arguments cranelift passes to `TrapSink::trap`
-#[derive(Debug)]
-pub struct FaerieTrapSite {
-    /// Offset into function
-    pub offset: binemit::CodeOffset,
-    /// Source location given to cranelift
-    pub srcloc: ir::SourceLoc,
-    /// Trap code, as determined by cranelift
-    pub code: ir::TrapCode,
-}
 
 /// Record of the trap sites for a given function
 #[derive(Debug)]
@@ -22,7 +12,7 @@ pub struct FaerieTrapSink {
     /// Total code size of function
     pub code_size: u32,
     /// All trap sites collected in function
-    pub sites: Vec<FaerieTrapSite>,
+    pub sites: Vec<TrapSite>,
 }
 
 impl FaerieTrapSink {
@@ -34,11 +24,20 @@ impl FaerieTrapSink {
             code_size,
         }
     }
+
+    /// Create a `FaerieTrapSink` pre-populated with `traps`
+    pub fn new_with_sites(name: &str, code_size: u32, traps: Vec<TrapSite>) -> Self {
+        Self {
+            sites: traps,
+            name: name.to_owned(),
+            code_size,
+        }
+    }
 }
 
 impl binemit::TrapSink for FaerieTrapSink {
     fn trap(&mut self, offset: binemit::CodeOffset, srcloc: ir::SourceLoc, code: ir::TrapCode) {
-        self.sites.push(FaerieTrapSite {
+        self.sites.push(TrapSite {
             offset,
             srcloc,
             code,

cranelift/module/src/backend.rs

@@ -6,6 +6,7 @@ use crate::FuncId;
 use crate::Linkage;
 use crate::ModuleNamespace;
 use crate::ModuleResult;
+use crate::TrapSite;
 use core::marker;
 use cranelift_codegen::isa::TargetIsa;
 use cranelift_codegen::Context;
@@ -14,6 +15,7 @@ use cranelift_codegen::{binemit, ir};
 use std::borrow::ToOwned;
 use std::boxed::Box;
 use std::string::String;
+use std::vec::Vec;
 
 /// A `Backend` implements the functionality needed to support a `Module`.
 ///
@@ -85,6 +87,18 @@ where
         code_size: u32,
     ) -> ModuleResult<Self::CompiledFunction>;
 
+    /// Define a function, taking the function body from the given `bytes`.
+    ///
+    /// Functions must be declared before being defined.
+    fn define_function_bytes(
+        &mut self,
+        id: FuncId,
+        name: &str,
+        bytes: &[u8],
+        namespace: &ModuleNamespace<Self>,
+        traps: Vec<TrapSite>,
+    ) -> ModuleResult<Self::CompiledFunction>;
+
     /// Define a zero-initialized data object of the given size.
     ///
     /// Data objects must be declared before being defined.

cranelift/module/src/lib.rs

@@ -35,6 +35,7 @@ use std::collections::{hash_map, HashMap};
 mod backend;
 mod data_context;
 mod module;
+mod traps;
 
 pub use crate::backend::{default_libcall_names, Backend};
 pub use crate::data_context::{DataContext, DataDescription, Init};
@@ -42,6 +43,7 @@ pub use crate::module::{
     DataId, FuncId, FuncOrDataId, Linkage, Module, ModuleError, ModuleFunction, ModuleNamespace,
     ModuleResult,
 };
+pub use crate::traps::TrapSite;
 
 /// Version number of this crate.
 pub const VERSION: &str = env!("CARGO_PKG_VERSION");

cranelift/module/src/module.rs

@@ -7,12 +7,14 @@
 
 use super::HashMap;
 use crate::data_context::DataContext;
+use crate::traps::TrapSite;
 use crate::Backend;
 use cranelift_codegen::binemit::{self, CodeInfo};
 use cranelift_codegen::entity::{entity_impl, PrimaryMap};
 use cranelift_codegen::{ir, isa, CodegenError, Context};
 use log::info;
 use std::borrow::ToOwned;
+use std::convert::TryInto;
 use std::string::String;
 use std::vec::Vec;
 use thiserror::Error;
@@ -139,6 +141,9 @@ pub enum ModuleError {
     /// Indicates an identifier was defined, but was declared as an import
     #[error("Invalid to define identifier declared as an import: {0}")]
     InvalidImportDefinition(String),
+    /// Indicates a too-long function was defined
+    #[error("Function {0} exceeds the maximum function size")]
+    FunctionTooLarge(String),
     /// Wraps a `cranelift-codegen` error
     #[error("Compilation error: {0}")]
     Compilation(#[from] CodegenError),
@@ -573,6 +578,48 @@ where
         Ok(total_size)
     }
 
+    /// Define a function, taking the function body from the given `bytes`.
+    ///
+    /// This function is generally only useful if you need to precisely specify
+    /// the emitted instructions for some reason; otherwise, you should use
+    /// `define_function`.
+    ///
+    /// Returns the size of the function's code.
+    pub fn define_function_bytes(
+        &mut self,
+        func: FuncId,
+        bytes: &[u8],
+        traps: Vec<TrapSite>,
+    ) -> ModuleResult<binemit::CodeOffset> {
+        info!("defining function {} with bytes", func);
+        let info = &self.contents.functions[func];
+        if info.compiled.is_some() {
+            return Err(ModuleError::DuplicateDefinition(info.decl.name.clone()));
+        }
+        if !info.decl.linkage.is_definable() {
+            return Err(ModuleError::InvalidImportDefinition(info.decl.name.clone()));
+        }
+
+        let total_size: u32 = match bytes.len().try_into() {
+            Ok(total_size) => total_size,
+            _ => Err(ModuleError::FunctionTooLarge(info.decl.name.clone()))?,
+        };
+
+        let compiled = Some(self.backend.define_function_bytes(
+            func,
+            &info.decl.name,
+            bytes,
+            &ModuleNamespace::<B> {
+                contents: &self.contents,
+            },
+            traps,
+        )?);
+
+        self.contents.functions[func].compiled = compiled;
+        self.functions_to_finalize.push(func);
+        Ok(total_size)
+    }
+
     /// Define a data object, producing the data contents from the given `DataContext`.
     pub fn define_data(&mut self, data: DataId, data_ctx: &DataContext) -> ModuleResult<()> {
         let compiled = {

cranelift/module/src/traps.rs

@@ -0,0 +1,14 @@
+//! Defines `TrapSite`.
+
+use cranelift_codegen::{binemit, ir};
+
+/// Record of the arguments cranelift passes to `TrapSink::trap`.
+#[derive(Clone, Debug)]
+pub struct TrapSite {
+    /// Offset into function.
+    pub offset: binemit::CodeOffset,
+    /// Source location given to cranelift.
+    pub srcloc: ir::SourceLoc,
+    /// Trap code, as determined by cranelift.
+    pub code: ir::TrapCode,
+}

cranelift/object/src/backend.rs

@@ -1,6 +1,6 @@
 //! Defines `ObjectBackend`.
 
-use crate::traps::{ObjectTrapSink, ObjectTrapSite};
+use crate::traps::ObjectTrapSink;
 use cranelift_codegen::binemit::{
     Addend, CodeOffset, NullStackmapSink, NullTrapSink, Reloc, RelocSink,
 };
@@ -9,7 +9,7 @@ use cranelift_codegen::isa::TargetIsa;
 use cranelift_codegen::{self, binemit, ir};
 use cranelift_module::{
     Backend, DataContext, DataDescription, DataId, FuncId, Init, Linkage, ModuleNamespace,
-    ModuleResult,
+    ModuleResult, TrapSite,
 };
 use object::write::{
     Object, Relocation, SectionId, StandardSection, Symbol, SymbolId, SymbolSection,
@@ -79,7 +79,7 @@ pub struct ObjectBackend {
     object: Object,
     functions: SecondaryMap<FuncId, Option<SymbolId>>,
     data_objects: SecondaryMap<DataId, Option<SymbolId>>,
-    traps: SecondaryMap<FuncId, Vec<ObjectTrapSite>>,
+    traps: SecondaryMap<FuncId, Vec<TrapSite>>,
     relocs: Vec<SymbolRelocs>,
     libcalls: HashMap<ir::LibCall, SymbolId>,
     libcall_names: Box<dyn Fn(ir::LibCall) -> String>,
@@ -226,6 +226,23 @@ impl Backend for ObjectBackend {
         Ok(ObjectCompiledFunction)
     }
 
+    fn define_function_bytes(
+        &mut self,
+        func_id: FuncId,
+        _name: &str,
+        bytes: &[u8],
+        _namespace: &ModuleNamespace<Self>,
+        traps: Vec<TrapSite>,
+    ) -> ModuleResult<ObjectCompiledFunction> {
+        let symbol = self.functions[func_id].unwrap();
+        let section = self.object.section_id(StandardSection::Text);
+        let _offset = self
+            .object
+            .add_symbol_data(symbol, section, bytes, self.function_alignment);
+        self.traps[func_id] = traps;
+        Ok(ObjectCompiledFunction)
+    }
+
     fn define_data(
         &mut self,
         data_id: DataId,
@@ -462,7 +479,7 @@ pub struct ObjectProduct {
     /// Symbol IDs for data objects (both declared and defined).
     pub data_objects: SecondaryMap<DataId, Option<SymbolId>>,
     /// Trap sites for defined functions.
-    pub traps: SecondaryMap<FuncId, Vec<ObjectTrapSite>>,
+    pub traps: SecondaryMap<FuncId, Vec<TrapSite>>,
 }
 
 impl ObjectProduct {

cranelift/object/src/lib.rs

@@ -29,7 +29,7 @@ mod backend;
 mod traps;
 
 pub use crate::backend::{ObjectBackend, ObjectBuilder, ObjectProduct, ObjectTrapCollection};
-pub use crate::traps::{ObjectTrapSink, ObjectTrapSite};
+pub use crate::traps::ObjectTrapSink;
 
 /// Version number of this crate.
 pub const VERSION: &str = env!("CARGO_PKG_VERSION");

cranelift/object/src/traps.rs

@@ -2,28 +2,18 @@
 //! for every function in the module. This data may be useful at runtime.
 
 use cranelift_codegen::{binemit, ir};
-
+use cranelift_module::TrapSite;
-/// Record of the arguments cranelift passes to `TrapSink::trap`
-#[derive(Clone)]
-pub struct ObjectTrapSite {
-    /// Offset into function
-    pub offset: binemit::CodeOffset,
-    /// Source location given to cranelift
-    pub srcloc: ir::SourceLoc,
-    /// Trap code, as determined by cranelift
-    pub code: ir::TrapCode,
-}
 
 /// Record of the trap sites for a given function
 #[derive(Default, Clone)]
 pub struct ObjectTrapSink {
     /// All trap sites collected in function
-    pub sites: Vec<ObjectTrapSite>,
+    pub sites: Vec<TrapSite>,
 }
 
 impl binemit::TrapSink for ObjectTrapSink {
     fn trap(&mut self, offset: binemit::CodeOffset, srcloc: ir::SourceLoc, code: ir::TrapCode) {
-        self.sites.push(ObjectTrapSite {
+        self.sites.push(TrapSite {
             offset,
             srcloc,
             code,

cranelift/simplejit/src/backend.rs

@@ -8,7 +8,7 @@ use cranelift_codegen::isa::TargetIsa;
 use cranelift_codegen::{self, ir, settings};
 use cranelift_module::{
     Backend, DataContext, DataDescription, DataId, FuncId, Init, Linkage, ModuleNamespace,
-    ModuleResult,
+    ModuleResult, TrapSite,
 };
 use cranelift_native;
 #[cfg(not(windows))]
@@ -191,6 +191,23 @@ impl SimpleJITBackend {
             _ => panic!("invalid ExternalName {}", name),
         }
     }
+
+    fn record_function_for_perf(&self, ptr: *mut u8, size: usize, name: &str) {
+        // The Linux perf tool supports JIT code via a /tmp/perf-$PID.map file,
+        // which contains memory regions and their associated names.  If we
+        // are profiling with perf and saving binaries to PERF_BUILDID_DIR
+        // for post-profile analysis, write information about each function
+        // we define.
+        if cfg!(target_os = "linux") && ::std::env::var_os("PERF_BUILDID_DIR").is_some() {
+            let mut map_file = ::std::fs::OpenOptions::new()
+                .create(true)
+                .append(true)
+                .open(format!("/tmp/perf-{}.map", ::std::process::id()))
+                .unwrap();
+
+            let _ = writeln!(map_file, "{:x} {:x} {}", ptr as usize, size, name);
+        }
+    }
 }
 
 impl<'simple_jit_backend> Backend for SimpleJITBackend {
@@ -267,15 +284,7 @@ impl<'simple_jit_backend> Backend for SimpleJITBackend {
             .allocate(size, EXECUTABLE_DATA_ALIGNMENT)
             .expect("TODO: handle OOM etc.");
 
-        if cfg!(target_os = "linux") && ::std::env::var_os("PERF_BUILDID_DIR").is_some() {
+        self.record_function_for_perf(ptr, size, name);
-            let mut map_file = ::std::fs::OpenOptions::new()
-                .create(true)
-                .append(true)
-                .open(format!("/tmp/perf-{}.map", ::std::process::id()))
-                .unwrap();
-
-            let _ = writeln!(map_file, "{:x} {:x} {}", ptr as usize, code_size, name);
-        }
 
         let mut reloc_sink = SimpleJITRelocSink::new();
         // Ignore traps for now. For now, frontends should just avoid generating code
@@ -299,6 +308,34 @@ impl<'simple_jit_backend> Backend for SimpleJITBackend {
         })
     }
 
+    fn define_function_bytes(
+        &mut self,
+        _id: FuncId,
+        name: &str,
+        bytes: &[u8],
+        _namespace: &ModuleNamespace<Self>,
+        _traps: Vec<TrapSite>,
+    ) -> ModuleResult<Self::CompiledFunction> {
+        let size = bytes.len();
+        let ptr = self
+            .memory
+            .code
+            .allocate(size, EXECUTABLE_DATA_ALIGNMENT)
+            .expect("TODO: handle OOM etc.");
+
+        self.record_function_for_perf(ptr, size, name);
+
+        unsafe {
+            ptr::copy_nonoverlapping(bytes.as_ptr(), ptr, size);
+        }
+
+        Ok(Self::CompiledFunction {
+            code: ptr,
+            size,
+            relocs: vec![],
+        })
+    }
+
     fn define_data(
         &mut self,
         _id: DataId,