Add a Context::emit_to_memory function.

This function will emit the binary machine code into contiguous raw
memory while sending relocations to a RelocSink.

Add a MemoryCodeSink for generating machine code directly into memory
efficiently. Allow the TargetIsa to provide emit_function
implementations that are specialized to the MemoryCodeSink type to avoid
needless small virtual callbacks to put1() et etc.

lib/cretonne/src/binemit/memorysink.rs

@@ -0,0 +1,108 @@
+//! Code sink that writes binary machine code into contiguous memory.
+//!
+//! The `CodeSink` trait is the most general way of extracting binary machine code from Cretonne,
+//! and it is implemented by things like the `test binemit` file test driver to generate
+//! hexadecimal machine code. The `CodeSink` has some undesirable performance properties because of
+//! the dual abstraction: `TargetIsa` is a trait object implemented by each supported ISA, so it
+//! can't have any generic functions that could be specialized for each `CodeSink` implementation.
+//! This results in many virtual function callbacks (one per `put*` call) when
+//! `TargetIsa::emit_inst()` is used.
+//!
+//! The `MemoryCodeSink` type fixes the performance problem because it is a type known to
+//! `TargetIsa` so it can specialize its machine code generation for the type. The trade-off is
+//! that a `MemoryCodeSink` will always write binary machine code to raw memory. It forwards any
+//! relocations to a `RelocSink` trait object. Relocations are less frequent than the
+//! `CodeSink::put*` methods, so the performance impact of the virtual callbacks is less severe.
+
+use ir::{Ebb, FuncRef, JumpTable};
+use super::{CodeSink, CodeOffset, Reloc};
+use std::ptr::write_unaligned;
+
+/// A `CodeSink` that writes binary machine code directly into memory.
+///
+/// A `MemoryCodeSink` object should be used when emitting a Cretonne IL function into executable
+/// memory. It writes machine code directly to a raw pointer without any bounds checking, so make
+/// sure to allocate enough memory for the whole function. The number of bytes required is returned
+/// by the `Context::compile()` function.
+///
+/// Any relocations in the function are forwarded to the `RelocSink` trait object.
+///
+/// Note that `MemoryCodeSink` writes multi-byte values in the native byte order of the host. This
+/// is not the right thing to do for cross compilation.
+pub struct MemoryCodeSink<'a> {
+    data: *mut u8,
+    offset: isize,
+    relocs: &'a mut RelocSink,
+}
+
+impl<'a> MemoryCodeSink<'a> {
+    /// Create a new memory code sink that writes a function to the memory pointed to by `data`.
+    pub fn new(data: *mut u8, relocs: &mut RelocSink) -> MemoryCodeSink {
+        MemoryCodeSink {
+            data,
+            offset: 0,
+            relocs,
+        }
+    }
+}
+
+/// A trait for receiving relocations for code that is emitted directly into memory.
+pub trait RelocSink {
+    /// Add a relocation referencing an EBB at the current offset.
+    fn reloc_ebb(&mut self, CodeOffset, Reloc, Ebb);
+
+    /// Add a relocation referencing an external function at the current offset.
+    fn reloc_func(&mut self, CodeOffset, Reloc, FuncRef);
+
+    /// Add a relocation referencing a jump table.
+    fn reloc_jt(&mut self, CodeOffset, Reloc, JumpTable);
+}
+
+impl<'a> CodeSink for MemoryCodeSink<'a> {
+    fn offset(&self) -> CodeOffset {
+        self.offset as CodeOffset
+    }
+
+    fn put1(&mut self, x: u8) {
+        unsafe {
+            write_unaligned(self.data.offset(self.offset), x);
+        }
+        self.offset += 1;
+    }
+
+    fn put2(&mut self, x: u16) {
+        unsafe {
+            write_unaligned(self.data.offset(self.offset) as *mut u16, x);
+        }
+        self.offset += 2;
+    }
+
+    fn put4(&mut self, x: u32) {
+        unsafe {
+            write_unaligned(self.data.offset(self.offset) as *mut u32, x);
+        }
+        self.offset += 4;
+    }
+
+    fn put8(&mut self, x: u64) {
+        unsafe {
+            write_unaligned(self.data.offset(self.offset) as *mut u64, x);
+        }
+        self.offset += 8;
+    }
+
+    fn reloc_ebb(&mut self, rel: Reloc, ebb: Ebb) {
+        let ofs = self.offset();
+        self.relocs.reloc_ebb(ofs, rel, ebb);
+    }
+
+    fn reloc_func(&mut self, rel: Reloc, func: FuncRef) {
+        let ofs = self.offset();
+        self.relocs.reloc_func(ofs, rel, func);
+    }
+
+    fn reloc_jt(&mut self, rel: Reloc, jt: JumpTable) {
+        let ofs = self.offset();
+        self.relocs.reloc_jt(ofs, rel, jt);
+    }
+}

lib/cretonne/src/binemit/mod.rs

@@ -4,8 +4,10 @@
 //! binary machine code.
 
 mod relaxation;
+mod memorysink;
 
 pub use self::relaxation::relax_branches;
+pub use self::memorysink::{MemoryCodeSink, RelocSink};
 
 use ir::{Ebb, FuncRef, JumpTable, Function, Inst};
 
@@ -55,3 +57,20 @@ pub fn bad_encoding(func: &Function, inst: Inst) -> ! {
            func.encodings[inst],
            func.dfg.display_inst(inst, None));
 }
+
+/// Emit a function to `sink`, given an instruction emitter function.
+///
+/// This function is called from the `TargetIsa::emit_function()` implementations with the
+/// appropriate instruction emitter.
+pub fn emit_function<CS, EI>(func: &Function, emit_inst: EI, sink: &mut CS)
+    where CS: CodeSink,
+          EI: Fn(&Function, Inst, &mut CS)
+{
+    for ebb in func.layout.ebbs() {
+        assert_eq!(func.offsets[ebb], sink.offset());
+        for inst in func.layout.ebb_insts(ebb) {
+            emit_inst(func, inst, sink);
+        }
+
+    }
+}