fix(ext/ffi): trampoline for fast calls (#15139)

.dprint.json

@@ -38,6 +38,7 @@
     "cli/tsc/*typescript.js",
     "gh-pages",
     "target",
+    "test_ffi/tests/test.js",
     "test_util/std",
     "test_util/wpt",
     "third_party",

.gitmodules

@@ -9,3 +9,6 @@
 [submodule "test_util/wpt"]
 	path = test_util/wpt
 	url = https://github.com/web-platform-tests/wpt.git
+[submodule "ext/ffi/tinycc"]
+	path = ext/ffi/tinycc
+	url = https://github.com/TinyCC/tinycc

ext/ffi/README.md

@@ -1,3 +1,25 @@
 # deno_ffi
 
 This crate implements dynamic library ffi.
+
+## Performance
+
+Deno FFI calls have extremely low overhead (~1ns on M1 16GB RAM) and perform on
+par with native code. Deno leverages V8 fast api calls and JIT compiled bindings
+to achieve these high speeds.
+
+`Deno.dlopen` generates an optimized and a fallback path. Optimized paths are
+triggered when V8 decides to optimize the function, hence call through the Fast
+API. Fallback paths handle types like function callbacks and implement proper
+error handling for unexpected types, that is not supported in Fast calls.
+
+Optimized calls enter a JIT compiled function "trampoline" that translates Fast
+API values directly for symbol calls. JIT compilation itself is super fast,
+thanks to `tinycc`. Currently, the optimized path is only supported on Linux and
+MacOS.
+
+To run benchmarks:
+
+```bash
+target/release/deno bench --allow-ffi --allow-read --unstable ./test_ffi/tests/bench.js
+```

ext/ffi/build.rs

@@ -0,0 +1,63 @@
+// Copyright 2018-2022 the Deno authors. All rights reserved. MIT license.
+
+use std::env;
+
+fn build_tcc() {
+  {
+    // TODO(@littledivy): Windows support for fast call.
+    // let tcc_path = root
+    //   .parent()
+    //   .unwrap()
+    //   .to_path_buf()
+    //   .parent()
+    //   .unwrap()
+    //   .to_path_buf()
+    //   .join("third_party")
+    //   .join("prebuilt")
+    //   .join("win");
+    // println!("cargo:rustc-link-search=native={}", tcc_path.display());
+  }
+  #[cfg(not(target_os = "windows"))]
+  {
+    use std::path::PathBuf;
+    use std::process::exit;
+    use std::process::Command;
+
+    let root = PathBuf::from(concat!(env!("CARGO_MANIFEST_DIR")));
+    let tcc_src = root.join("tinycc");
+    dbg!(&tcc_src);
+    let out_dir = PathBuf::from(env::var("OUT_DIR").unwrap());
+    let mut configure = Command::new(tcc_src.join("configure"));
+    configure.current_dir(&out_dir);
+    configure.args(&["--enable-static", "--extra-cflags=-fPIC -O3 -g -static"]);
+    let status = configure.status().unwrap();
+    if !status.success() {
+      eprintln!("Fail to configure: {:?}", status);
+      exit(1);
+    }
+
+    let mut make = Command::new("make");
+    make.current_dir(&out_dir).arg(format!(
+      "-j{}",
+      env::var("NUM_JOBS").unwrap_or_else(|_| String::from("1"))
+    ));
+    make.args(&["libtcc.a"]);
+    let status = make.status().unwrap();
+
+    if !status.success() {
+      eprintln!("Fail to make: {:?}", status);
+      exit(1);
+    }
+    println!("cargo:rustc-link-search=native={}", out_dir.display());
+    println!("cargo:rerun-if-changed={}", tcc_src.display());
+  }
+}
+
+#[cfg(target_os = "windows")]
+fn main() {}
+
+#[cfg(not(target_os = "windows"))]
+fn main() {
+  build_tcc();
+  println!("cargo:rustc-link-lib=static=tcc");
+}

ext/ffi/jit_trampoline.rs

@@ -0,0 +1,153 @@
+// Copyright 2018-2022 the Deno authors. All rights reserved. MIT license.
+
+use crate::NativeType;
+use crate::{tcc::Compiler, Symbol};
+use std::ffi::c_void;
+use std::ffi::CString;
+use std::fmt::Write as _;
+
+pub(crate) struct Allocation {
+  pub addr: *mut c_void,
+  _ctx: Compiler,
+  _sym: Box<Symbol>,
+}
+
+macro_rules! cstr {
+  ($st:expr) => {
+    &CString::new($st).unwrap()
+  };
+}
+
+fn native_arg_to_c(ty: &NativeType) -> &'static str {
+  match ty {
+    NativeType::U8 | NativeType::U16 | NativeType::U32 => "uint32_t",
+    NativeType::I8 | NativeType::I16 | NativeType::I32 => "int32_t",
+    NativeType::Void => "void",
+    NativeType::F32 => "float",
+    NativeType::F64 => "double",
+    _ => unimplemented!(),
+  }
+}
+
+fn native_to_c(ty: &NativeType) -> &'static str {
+  match ty {
+    NativeType::U8 => "uint8_t",
+    NativeType::U16 => "uint16_t",
+    NativeType::U32 => "uint32_t",
+    NativeType::I8 => "int8_t",
+    NativeType::I16 => "uint16_t",
+    NativeType::I32 => "int32_t",
+    NativeType::Void => "void",
+    NativeType::F32 => "float",
+    NativeType::F64 => "double",
+    _ => unimplemented!(),
+  }
+}
+
+pub(crate) fn codegen(sym: &crate::Symbol) -> String {
+  let mut c = String::from("#include <stdint.h>\n");
+  let ret = native_to_c(&sym.result_type);
+
+  // extern <return_type> func(
+  c += "\nextern ";
+  c += ret;
+  c += " func(";
+  // <param_type> p0, <param_type> p1, ...);
+  for (i, ty) in sym.parameter_types.iter().enumerate() {
+    if i > 0 {
+      c += ", ";
+    }
+    c += native_to_c(ty);
+    let _ = write!(c, " p{i}");
+  }
+  c += ");\n\n";
+
+  // void* recv, <param_type> p0, <param_type> p1, ...);
+  c += ret;
+  c += " func_trampoline(";
+  c += "void* recv";
+  for (i, ty) in sym.parameter_types.iter().enumerate() {
+    c += ", ";
+    c += native_arg_to_c(ty);
+    let _ = write!(c, " p{i}");
+  }
+  c += ") {\n";
+  // return func(p0, p1, ...);
+  c += "  return func(";
+  for (i, _) in sym.parameter_types.iter().enumerate() {
+    if i > 0 {
+      c += ", ";
+    }
+    let _ = write!(c, "p{i}");
+  }
+  c += ");\n}\n\n";
+  c
+}
+
+pub(crate) fn gen_trampoline(
+  sym: Box<crate::Symbol>,
+) -> Result<Box<Allocation>, ()> {
+  let mut ctx = Compiler::new()?;
+  ctx.set_options(cstr!("-nostdlib"));
+  // SAFETY: symbol satisfies ABI requirement.
+  unsafe { ctx.add_symbol(cstr!("func"), sym.ptr.0 as *const c_void) };
+  let c = codegen(&sym);
+
+  ctx.compile_string(cstr!(c))?;
+  let alloc = Allocation {
+    addr: ctx.relocate_and_get_symbol(cstr!("func_trampoline"))?,
+    _ctx: ctx,
+    _sym: sym,
+  };
+  Ok(Box::new(alloc))
+}
+
+#[cfg(test)]
+mod tests {
+  use super::*;
+  use libffi::middle::Type;
+  use std::ptr::null_mut;
+
+  fn codegen(parameters: Vec<NativeType>, ret: NativeType) -> String {
+    let sym = Box::new(crate::Symbol {
+      cif: libffi::middle::Cif::new(vec![], Type::void()),
+      ptr: libffi::middle::CodePtr(null_mut()),
+      parameter_types: parameters,
+      result_type: ret,
+      can_callback: false,
+    });
+    super::codegen(&sym)
+  }
+
+  #[test]
+  fn test_gen_trampoline() {
+    assert_eq!(
+      codegen(vec![], NativeType::Void),
+      "#include <stdint.h>\n\nextern void func();\n\nvoid func_trampoline(void* recv) {\n  return func();\n}\n\n"
+    );
+    assert_eq!(
+      codegen(vec![NativeType::U32, NativeType::U32], NativeType::U32),
+      "#include <stdint.h>\n\nextern uint32_t func(uint32_t p0, uint32_t p1);\n\nuint32_t func_trampoline(void* recv, uint32_t p0, uint32_t p1) {\n  return func(p0, p1);\n}\n\n"
+    );
+    assert_eq!(
+      codegen(vec![NativeType::I32, NativeType::I32], NativeType::I32),
+      "#include <stdint.h>\n\nextern int32_t func(int32_t p0, int32_t p1);\n\nint32_t func_trampoline(void* recv, int32_t p0, int32_t p1) {\n  return func(p0, p1);\n}\n\n"
+    );
+    assert_eq!(
+      codegen(vec![NativeType::F32, NativeType::F32], NativeType::F32),
+      "#include <stdint.h>\n\nextern float func(float p0, float p1);\n\nfloat func_trampoline(void* recv, float p0, float p1) {\n  return func(p0, p1);\n}\n\n"
+    );
+    assert_eq!(
+      codegen(vec![NativeType::F64, NativeType::F64], NativeType::F64),
+      "#include <stdint.h>\n\nextern double func(double p0, double p1);\n\ndouble func_trampoline(void* recv, double p0, double p1) {\n  return func(p0, p1);\n}\n\n"
+    );
+  }
+
+  #[test]
+  fn test_gen_trampoline_implicit_cast() {
+    assert_eq!(
+      codegen(vec![NativeType::I8, NativeType::U8], NativeType::I8),
+      "#include <stdint.h>\n\nextern int8_t func(int8_t p0, uint8_t p1);\n\nint8_t func_trampoline(void* recv, int32_t p0, uint32_t p1) {\n  return func(p0, p1);\n}\n\n"
+    )
+  }
+}

ext/ffi/lib.rs

@@ -39,6 +39,11 @@ use std::path::PathBuf;
 use std::ptr;
 use std::rc::Rc;
 
+#[cfg(not(target_os = "windows"))]
+mod jit_trampoline;
+#[cfg(not(target_os = "windows"))]
+mod tcc;
+
 thread_local! {
   static LOCAL_ISOLATE_POINTER: RefCell<*const v8::Isolate> = RefCell::new(ptr::null());
 }
@@ -72,6 +77,8 @@ struct Symbol {
   ptr: libffi::middle::CodePtr,
   parameter_types: Vec<NativeType>,
   result_type: NativeType,
+  // This is dead code only on Windows
+  #[allow(dead_code)]
   can_callback: bool,
 }
 
@@ -678,6 +685,7 @@ impl From<&NativeType> for fast_api::Type {
   }
 }
 
+#[cfg(not(target_os = "windows"))]
 fn is_fast_api(rv: NativeType) -> bool {
   !matches!(
     rv,
@@ -696,25 +704,36 @@ fn make_sync_fn<'s>(
   scope: &mut v8::HandleScope<'s>,
   sym: Box<Symbol>,
 ) -> v8::Local<'s, v8::Function> {
-  let mut fast_ffi_templ = None;
+  #[cfg(not(target_os = "windows"))]
+  let mut fast_ffi_templ: Option<FfiFastCallTemplate> = None;
 
+  #[cfg(target_os = "windows")]
+  let fast_ffi_templ: Option<FfiFastCallTemplate> = None;
+
+  #[cfg(not(target_os = "windows"))]
+  let mut fast_allocations: Option<*mut ()> = None;
+  #[cfg(not(target_os = "windows"))]
   if !sym.can_callback
     && !sym.parameter_types.iter().any(|t| !is_fast_api(*t))
     && is_fast_api(sym.result_type)
   {
+    let ret = fast_api::Type::from(&sym.result_type);
+
     let mut args = sym
       .parameter_types
       .iter()
       .map(|t| t.into())
       .collect::<Vec<_>>();
-    if args.is_empty() {
-      args.push(fast_api::Type::V8Value);
-    }
+    // recv
+    args.insert(0, fast_api::Type::V8Value);
+    let symbol_trampoline =
+      jit_trampoline::gen_trampoline(sym.clone()).expect("gen_trampoline");
     fast_ffi_templ = Some(FfiFastCallTemplate {
       args: args.into_boxed_slice(),
-      ret: (&fast_api::Type::from(&sym.result_type)).into(),
-      symbol_ptr: sym.ptr.as_ptr() as *const c_void,
+      ret: (&ret).into(),
+      symbol_ptr: symbol_trampoline.addr,
     });
+    fast_allocations = Some(Box::into_raw(symbol_trampoline) as *mut ());
   }
 
   let sym = Box::leak(sym);
@@ -754,7 +773,13 @@ fn make_sync_fn<'s>(
     Box::new(move |_| {
       // SAFETY: This is never called twice. pointer obtained
       // from Box::into_raw, hence, satisfies memory layout requirements.
-      unsafe { Box::from_raw(sym) };
+      unsafe {
+        Box::from_raw(sym);
+        #[cfg(not(target_os = "windows"))]
+        if let Some(fast_allocations) = fast_allocations {
+          Box::from_raw(fast_allocations as *mut jit_trampoline::Allocation);
+        }
+      }
     }),
   );
 

ext/ffi/tcc.rs

@@ -0,0 +1,143 @@
+// Copyright 2018-2022 the Deno authors. All rights reserved. MIT license.
+
+use std::{
+  ffi::CStr,
+  marker::PhantomData,
+  os::raw::{c_char, c_int, c_void},
+  ptr::null_mut,
+};
+
+#[repr(C)]
+#[derive(Debug)]
+pub struct TCCState {
+  _unused: [u8; 0],
+}
+pub const TCC_OUTPUT_MEMORY: i32 = 1;
+
+extern "C" {
+  pub fn tcc_new() -> *mut TCCState;
+  pub fn tcc_delete(s: *mut TCCState);
+  pub fn tcc_set_options(s: *mut TCCState, str: *const c_char);
+  pub fn tcc_compile_string(s: *mut TCCState, buf: *const c_char) -> c_int;
+  pub fn tcc_add_symbol(
+    s: *mut TCCState,
+    name: *const c_char,
+    val: *const c_void,
+  ) -> c_int;
+  pub fn tcc_set_output_type(s: *mut TCCState, output_type: c_int) -> c_int;
+  pub fn tcc_relocate(s1: *mut TCCState, ptr: *mut c_void) -> c_int;
+  pub fn tcc_get_symbol(s: *mut TCCState, name: *const c_char) -> *mut c_void;
+}
+
+/// Compilation context.
+pub struct Compiler {
+  inner: *mut TCCState,
+  _phantom: PhantomData<TCCState>,
+  pub bin: Option<Vec<u8>>,
+}
+
+impl Compiler {
+  pub fn new() -> Result<Self, ()> {
+    // SAFETY: There is one context per thread.
+    let inner = unsafe { tcc_new() };
+    if inner.is_null() {
+      Err(())
+    } else {
+      let ret =
+        // SAFETY: set output to memory.
+        unsafe { tcc_set_output_type(inner, TCC_OUTPUT_MEMORY as c_int) };
+      assert_eq!(ret, 0);
+      Ok(Self {
+        inner,
+        _phantom: PhantomData,
+        bin: None,
+      })
+    }
+  }
+
+  pub fn set_options(&mut self, option: &CStr) -> &mut Self {
+    // SAFETY: option is a null-terminated C string.
+    unsafe {
+      tcc_set_options(self.inner, option.as_ptr());
+    }
+    self
+  }
+
+  pub fn compile_string(&mut self, p: &CStr) -> Result<(), ()> {
+    // SAFETY: p is a null-terminated C string.
+    let ret = unsafe { tcc_compile_string(self.inner, p.as_ptr()) };
+    if ret == 0 {
+      Ok(())
+    } else {
+      Err(())
+    }
+  }
+
+  /// # Safety
+  /// Symbol need satisfy ABI requirement.
+  pub unsafe fn add_symbol(&mut self, sym: &CStr, val: *const c_void) {
+    // SAFETY: sym is a null-terminated C string.
+    let ret = tcc_add_symbol(self.inner, sym.as_ptr(), val);
+    assert_eq!(ret, 0);
+  }
+
+  pub fn relocate_and_get_symbol(
+    &mut self,
+    sym: &CStr,
+  ) -> Result<*mut c_void, ()> {
+    // SAFETY: pass null ptr to get required length
+    let len = unsafe { tcc_relocate(self.inner, null_mut()) };
+    if len == -1 {
+      return Err(());
+    };
+    let mut bin = Vec::with_capacity(len as usize);
+    let ret =
+      // SAFETY: bin is allocated up to len.
+      unsafe { tcc_relocate(self.inner, bin.as_mut_ptr() as *mut c_void) };
+    if ret != 0 {
+      return Err(());
+    }
+    // SAFETY: if ret == 0, bin is initialized.
+    unsafe {
+      bin.set_len(len as usize);
+    }
+    self.bin = Some(bin);
+    // SAFETY: sym is a null-terminated C string.
+    let addr = unsafe { tcc_get_symbol(self.inner, sym.as_ptr()) };
+    Ok(addr)
+  }
+}
+
+impl Drop for Compiler {
+  fn drop(&mut self) {
+    // SAFETY: delete state from tcc_new()
+    unsafe { tcc_delete(self.inner) };
+  }
+}
+
+#[cfg(test)]
+mod test {
+  use super::*;
+  use std::ffi::CString;
+
+  #[test]
+  fn test_compiler_jit() {
+    let p = CString::new(
+      r#"
+      #include <stdint.h>
+      int32_t add(int32_t a, int32_t b) {
+          return a + b;
+      }
+      "#
+      .as_bytes(),
+    )
+    .unwrap();
+    let sym = CString::new("add".as_bytes()).unwrap();
+
+    let mut ctx = Compiler::new().unwrap();
+    let ops = CString::new("-nostdlib").unwrap();
+    ctx.set_options(&ops);
+    assert!(ctx.compile_string(&p).is_ok());
+    ctx.relocate_and_get_symbol(&sym).unwrap();
+  }
+}

ext/ffi/tinycc

@@ -0,0 +1 @@
+Subproject commit afc136262e93ae85fb3643005b36dbfc30d99c42

test_ffi/tests/integration_tests.rs

@@ -30,6 +30,7 @@ fn basic() {
     .arg("--allow-read")
     .arg("--unstable")
     .arg("--quiet")
+    .arg(r#"--v8-flags=--allow-natives-syntax"#)
     .arg("tests/test.js")
     .env("NO_COLOR", "1")
     .output()
@@ -62,6 +63,7 @@ fn basic() {
     true\n\
     579\n\
     579\n\
+    579\n\
     8589934590n\n\
     -8589934590n\n\
     8589934590n\n\

test_ffi/tests/test.js

@@ -1,6 +1,8 @@
 // Copyright 2018-2022 the Deno authors. All rights reserved. MIT license.
 // deno-lint-ignore-file
 
+// Run using cargo test or `--v8-options=--allow-natives-syntax`
+
 import { assertThrows } from "../../test_util/std/testing/asserts.ts";
 
 const targetDir = Deno.execPath().replace(/[^\/\\]+$/, "");
@@ -182,8 +184,9 @@ const dylib = Deno.dlopen(libPath, {
     type: "pointer",
   },
 });
+const { symbols } = dylib;
 
-dylib.symbols.printSomething();
+symbols.printSomething();
 const buffer = new Uint8Array([1, 2, 3, 4, 5, 6, 7, 8]);
 const buffer2 = new Uint8Array([9, 10]);
 dylib.symbols.print_buffer(buffer, buffer.length);
@@ -238,7 +241,15 @@ const before = performance.now();
 await sleepNonBlocking.call(100);
 console.log(performance.now() - before >= 100);
 
-console.log(dylib.symbols.add_u32(123, 456));
+const { add_u32 } = symbols;
+function addU32Fast(a, b) {
+  return add_u32(a, b);
+};
+
+%PrepareFunctionForOptimization(addU32Fast);
+console.log(addU32Fast(123, 456));
+%OptimizeFunctionOnNextCall(addU32Fast);
+console.log(addU32Fast(123, 456));
 
 console.log(dylib.symbols.add_i32(123, 456));
 console.log(dylib.symbols.add_u64(0xffffffffn, 0xffffffffn));

third_party

@@ -1 +1 @@
-Subproject commit 4fd74a381b2a9f357ea7be80c12c24863596841f
+Subproject commit 9f314cefb507e3b9de08edc6046353e4012279fc