complete rewrite!

featuring 7ns inlineable context switches, no more separately-built
assembly objects, and a vastly nicer interface.
incontext/outcontext are no more, context switch calls now take a single
context structure, which functions as both.
everything now also functions without any heap allocations -- for the
context setup, only an FnOnce() value is necessary.

Cargo.toml

@@ -3,6 +3,3 @@ name = "lwkt"
 version = "0.0.1"
 authors = ["edef <edef@edef.eu>"]
 build = "build.rs"
-
-[dependencies]
-fn_box = "^1.0.1"

benches/swap.rs

@@ -1,36 +1,23 @@
-#![feature(unboxed_closures, default_type_params, box_syntax)]
 extern crate test;
-extern crate libc;
 extern crate lwkt;
-extern crate fn_box;
-use test::Bencher;
 use lwkt::Context;
-use fn_box::FnBox;
-use std::ptr::null_mut;
-use std::mem::{transmute, forget};
+
+static mut ctx_slot: *mut Context = 0 as *mut Context;
 
 #[bench]
-fn swap(b: &mut Bencher) {
-  let mut native = unsafe { Context::native() };
-  let f: Box<FnBox() + Send + 'static> = unsafe { transmute((1u, 1u)) };
-
-  let mut ctx = box { (&mut native as *mut Context, null_mut()) };
-  let mut green = Context::new(init, &mut *ctx as *mut _, f);
-  ctx.1 = &mut green as *mut Context;
-
-  fn init(ctx: *mut (*mut Context, *mut Context), f: Box<FnBox()>) -> ! {
+fn swap(b: &mut test::Bencher) {
   unsafe {
-      let (native, green) = *ctx;
-      forget(f);
-      loop { Context::swap(&mut *green, &mut *native); }
-    }
+    let mut ctx = Context::new(move |:| {
+      let ctx_ptr = ctx_slot;
+      loop {
+        (*ctx_ptr).swap()
       }
+    });
 
-  unsafe {
-    Context::swap(&mut native, &mut green);
-  }
+    ctx_slot = &mut ctx;
 
-  b.iter(|| unsafe {
-    Context::swap(&mut native, &mut green);
-  })
+    ctx.swap();
+
+    b.iter(|| ctx.swap());
+  }
 }

build.rs

@@ -22,13 +22,6 @@ fn main() {
                           .arg(outpath.clone())
                           .status().unwrap();
       }
-      else if let Some(basename) = eat_extension(filename, ".s") {
-        outpath = format!("{}/{}.o", out_dir, basename);
-
-        Command::new("nasm").args(&[filepath.as_slice(), "-felf64", "-o"])
-                            .arg(outpath.clone())
-                            .status().unwrap();
-      }
       else { continue }
 
       objects.push(outpath);

src/arch.rs

@@ -1,102 +1,69 @@
 use core::prelude::*;
-use core::simd::u64x2;
-use core::mem::{size_of, zeroed};
+use core::mem::{size_of, align_of};
+use core::cmp::max;
+use core::ptr;
 
 use stack::Stack;
 
-extern "C" {
-  #[link_name = "lwt_bootstrap"]
-  pub fn bootstrap();
-  #[link_name = "lwt_swapcontext"]
-  pub fn swapcontext(save: *mut Registers, restore: *mut Registers);
-  #[link_name = "lwt_abort"]
-  pub fn abort() -> !;
-}
-
 #[allow(non_camel_case_types)]
 pub type uintptr_t = u64;
 
-#[repr(C)]
-#[allow(dead_code)]
 pub struct Registers {
-  rbx: u64,
-  rsp: u64,
-  rbp: u64,
-  rdi: u64,
-  r12: u64,
-  r13: u64,
-  r14: u64,
-  r15: u64,
-  ip:  u64,
-  xmm0: u64x2,
-  xmm1: u64x2,
-  xmm2: u64x2,
-  xmm3: u64x2,
-  xmm4: u64x2,
-  xmm5: u64x2,
+  rsp: *mut uintptr_t
 }
 
-impl Registers {
-  pub fn new() -> Registers {
-    unsafe {
-      Registers {
-        ip: abort as uintptr_t,
-        .. zeroed()
-      }
-    }
-  }
-}
-
-pub fn initialise_call_frame(stack: &mut Stack, init: uintptr_t, args: &[uintptr_t]) -> Registers {
-  let sp = stack.top() as *mut uintptr_t;
-  let sp = align_down_mut(sp, 16);
-  let sp = offset_mut(sp, -1);
-  unsafe {
-    *sp = 0;
-  }
-
-  let mut regs = Registers {
-    rbp: 0,
-    rsp: sp as uintptr_t,
-    ip: bootstrap as uintptr_t,
-    rbx: init,
-    .. Registers::new()
-  };
-
-  match into_fields!(regs { rdi, r12, r13, r14, r15 } <- args.iter().cloned()) {
-    Some(mut args) => if args.next().is_some() {
-      panic!("too many arguments")
-    },
-    None => {}
-  }
-
-  regs
-}
-
-// Rust stores a stack limit at [fs:0x70]. These two functions set and retrieve
-// the limit. They're marked as #[inline(always)] so that they can be used in
-// situations where the stack limit is invalid.
+impl Copy for Registers {}
 
 #[inline(always)]
-pub unsafe fn get_sp_limit() -> *const u8 {
-  let limit;
-  asm!("movq %fs:0x70, $0" : "=r"(limit) ::: "volatile");
-  limit
-}
-
-#[inline(always)]
-pub unsafe fn set_sp_limit(limit: *const u8) {
-  asm!("movq $0, %fs:0x70" :: "r"(limit) :: "volatile");
+pub unsafe fn swap(regs: &mut Registers) {
+  asm!(include_str!("swap.s")
+        :
+        : "{rdi}" (&mut regs.rsp)
+        : "rax", "rbx", "rcx", "rdx", "rsi", "rdi",
+          "r8", "r9", "r10", "r11", "r12", "r13", "r14", "r15",
+          "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7",
+          "xmm8", "xmm9", "xmm10", "xmm11", "xmm12", "xmm13", "xmm14", "xmm15",
+          "cc"
+        : "volatile");
 }
 
 #[inline]
+pub unsafe fn initialize_call_frame<S, F>(stack: &mut S, f: F) -> Registers where S: Stack, F: FnOnce() {
+  let sp_limit = stack.limit();
+  let mut sp = stack.top() as *mut uintptr_t;
+  let f_ptr = push(&mut sp, f);
+
+  asm!(include_str!("init.s")
+        : "={rdi}"(sp)
+        : "{rdi}" (sp),
+          "{rsi}" (rust_trampoline::<F>),
+          "{rdx}" (f_ptr),
+          "{rcx}" (sp_limit)
+        :
+        : "volatile");
+
+  Registers { rsp: sp }
+}
+
+unsafe extern "C" fn rust_trampoline<F: FnOnce()>(f: *const F) {
+  ptr::read(f)()
+}
+
+unsafe fn push<T>(spp: &mut *mut uintptr_t, value: T) -> *mut T {
+  let mut sp = *spp as *mut T;
+  sp = offset_mut(sp, -1);
+  sp = align_down_mut(sp, max(align_of::<T>(), 16));
+  *sp = value;
+  *spp = sp as *mut uintptr_t;
+  sp
+}
+
 fn align_down_mut<T>(sp: *mut T, n: usize) -> *mut T {
   let sp = (sp as usize) & !(n - 1);
   sp as *mut T
 }
 
 // ptr::offset_mut is positive ints only
-#[inline]
 pub fn offset_mut<T>(ptr: *mut T, count: isize) -> *mut T {
   (ptr as isize + count * (size_of::<T>() as isize)) as *mut T
 }

src/arch.s

@@ -1,99 +0,0 @@
-; vim: ft=nasm
-BITS 64
-
-;; the structure containing every register that is saved on context switches.
-;; this needs to match the struct in arch.rs, or shit will break badly.
-struc context
-  ctx_rbx resq 1
-  ctx_rsp resq 1
-  ctx_rbp resq 1
-  ctx_rdi resq 1
-  ctx_r12 resq 1
-  ctx_r13 resq 1
-  ctx_r14 resq 1
-  ctx_r15 resq 1
-  ctx_ip:
-    resq 1
-    alignb 16
-  ctx_xmm0 resq 2
-  ctx_xmm1 resq 2
-  ctx_xmm2 resq 2
-  ctx_xmm3 resq 2
-  ctx_xmm4 resq 2
-  ctx_xmm5 resq 2
-endstruc
-
-global lwt_swapcontext
-lwt_swapcontext:
-;; this is where the actual context switching takes place. first, save every
-;; register in the current context into the leaving context, pointed at by rdi,
-;; making sure the return address ends up in the IP slot. then, restore every
-;; register from the entering context, pointed at by rsi, and jump to the
-;; instruction pointer.
-  pop rax
-
-  ; save instruction pointer
-  mov [rdi+ctx_ip], rax
-
-  ; save non-volatile integer registers (including rsp)
-  mov [rdi+ctx_rbx], rbx
-  mov [rdi+ctx_rsp], rsp
-  mov [rdi+ctx_rbp], rbp
-  mov [rdi+ctx_r12], r12
-  mov [rdi+ctx_r13], r13
-  mov [rdi+ctx_r14], r14
-  mov [rdi+ctx_r15], r15
-
-  ; save 0th argument register
-  mov [rdi+ctx_rdi], rdi
-
-  ; save non-volatile XMM registers
-  movapd [rdi+ctx_xmm0], xmm0
-  movapd [rdi+ctx_xmm1], xmm1
-  movapd [rdi+ctx_xmm2], xmm2
-  movapd [rdi+ctx_xmm3], xmm3
-  movapd [rdi+ctx_xmm4], xmm4
-  movapd [rdi+ctx_xmm5], xmm5
-
-  ; restore non-volatile integer registers
-  mov rbx, [rsi+ctx_rbx]
-  mov rsp, [rsi+ctx_rsp]
-  mov rbp, [rsi+ctx_rbp]
-  mov r12, [rsi+ctx_r12]
-  mov r13, [rsi+ctx_r13]
-  mov r14, [rsi+ctx_r14]
-  mov r15, [rsi+ctx_r15]
-
-  ; restore 0th argument register
-  mov rdi, [rsi+ctx_rdi]
-
-  ; restore non-volatile XMM registers
-  movapd xmm0, [rsi+ctx_xmm0]
-  movapd xmm1, [rsi+ctx_xmm1]
-  movapd xmm2, [rsi+ctx_xmm2]
-  movapd xmm3, [rsi+ctx_xmm3]
-  movapd xmm4, [rsi+ctx_xmm4]
-  movapd xmm5, [rsi+ctx_xmm5]
-
-  jmp [rsi+ctx_ip]
-
-global lwt_bootstrap
-lwt_bootstrap:
-;; some of the parameter registers aren't saved on context switch, and thus
-;; can't be set into the struct directly. thus, initialisation from Rust-land
-;; places the parameters in unrelated registers, and we frob them into place
-;; out here, in assembly-land. below are the parameter registers in order,
-;; along with the alternative register used in parentheses, if there is one.
-;; rdi, rsi (r12), rdx (r13), rcx (r14), r8(r15), r9
-  mov rsi, r12
-  mov rdx, r13
-  mov rcx, r14
-  mov r8, r15
-  jmp rbx
-
-global lwt_abort
-lwt_abort:
-;; when a context is created for a native thread, it should only be switched
-;; out of. if it's accidentally switched into, it'll hit this, because that's
-;; what we set the initial IP to.
-  ud2

src/context.rs

@@ -1,61 +1,26 @@
 use core::prelude::*;
-
-use core::mem::transmute;
-use core::raw;
-use alloc::boxed::Box;
-use fn_box::FnBox;
-
-use stack::Stack;
+use platform::Stack;
 use arch::{self, Registers};
+use platform;
 
 pub struct Context {
   regs: Registers,
-  stack: Stack
+  _stack: platform::Stack
 }
 
-pub type BoxedFn<Args, Result> = Box<FnBox<Args, Result> + Send + 'static>;
-pub type StartFn<T, Args, Result> = fn(data: *mut T, f: BoxedFn<Args, Result>) -> !;
-
 impl Context {
-  pub fn new<T, Args, Result>(init: StartFn<T, Args, Result>, data: *mut T,
-                              f: BoxedFn<Args, Result>) -> Context {
+  #[inline]
+  pub unsafe fn new<F>(f: F) -> Context where F: FnOnce() + Send + 'static {
     let mut stack = Stack::new(4 << 20);
-    let f: raw::TraitObject = unsafe { transmute(f) };
-
+    let regs = arch::initialize_call_frame(&mut stack, f);
     Context {
-      regs: arch::initialise_call_frame(&mut stack,
-        init_ctx::<T, Args, Result> as arch::uintptr_t,
-        &[init as arch::uintptr_t,
-          data as arch::uintptr_t,
-          f.data as arch::uintptr_t,
-          f.vtable as arch::uintptr_t]),
-      stack: stack
+      regs: regs,
+      _stack: stack
     }
   }
-}
-
-unsafe extern "C" fn init_ctx<T, A, R>(start: StartFn<T, A, R>, data: *mut T,
-                                       f_data: *mut (), f_vtable: *mut ()) -> ! {
-  let f: BoxedFn<A, R> = transmute(raw::TraitObject {
-    data: f_data,
-    vtable: f_vtable
-  });
-
-  start(data, f)
-}
-
-impl Context {
-  pub unsafe fn native() -> Context {
-    Context {
-      regs: Registers::new(),
-      stack: Stack::native(arch::get_sp_limit())
-    }
-  }
-
 
   #[inline(always)]
-  pub unsafe fn swap(out_context: &mut Context, in_context: &mut Context) {
-    arch::set_sp_limit(in_context.stack.limit());
-    arch::swapcontext(&mut out_context.regs, &mut in_context.regs);
+  pub unsafe fn swap(&mut self) {
+    arch::swap(&mut self.regs)
   }
 }

src/init.s

@@ -0,0 +1,17 @@
+xchg %rsp, %rdi
+
+pushq %rsi
+pushq %rdx
+pushq %rcx
+call 1f
+
+popq %fs:0x70
+popq %rdi
+popq %rax
+
+movq $$0, %rbp
+call *%rax
+ud2
+
+1:
+  xchg %rsp, %rdi

src/lib.rs

@@ -2,16 +2,12 @@
 #![no_std]
 
 #[macro_use]
+#[allow(unstable)]
 extern crate core;
-extern crate alloc;
-extern crate fn_box;
 
 pub use context::Context;
 
-mod std { pub use core::fmt; }
-
-#[macro_use]
-mod macros;
+mod std { pub use core::*; }
 
 mod context;
 mod stack;

src/macros.rs

@@ -1,29 +0,0 @@
-macro_rules! __into_fields {
-  ($x:ident { $field:ident } <- $iter:ident) => {
-    match $iter.next() {
-      Some(value) => {
-        $x.$field = value;
-        Some($iter)
-      }
-      None => None
-    }
-  };
-  ($x:ident { $field:ident, $($fields_rest:ident),* } <- $iter:ident) => {
-    match $iter.next() {
-      Some(value) => {
-        $x.$field = value;
-        __into_fields!($x { $($fields_rest),* } <- $iter)
-      }
-      None => None
-    }
-  };
-}
-
-macro_rules! into_fields {
-  ($x:ident { $($fields_rest:ident),* } <- $iter:expr) => {
-    {
-      let mut iter = $iter;
-      __into_fields!($x { $($fields_rest),* } <- iter)
-    }
-  }
-}

src/main.rs

@@ -1,36 +1,19 @@
+#![feature(thread_local)]
 extern crate lwkt;
-extern crate fn_box;
-
-use std::ptr::null_mut;
-use std::intrinsics::abort;
 use lwkt::Context;
-use fn_box::FnBox;
+
+#[thread_local]
+static mut ctx_slot: *mut Context = 0 as *mut Context;
 
 fn main() {
-  let f = Box::new(move |:| {
-    println!("Hello, world!")
+  unsafe {
+    let mut ctx = Context::new(move |:| {
+      println!("it's alive!");
+      (*ctx_slot).swap();
     });
 
-  let mut native = unsafe { Context::native() };
+    ctx_slot = &mut ctx;
 
-  fn init(ctx: *mut (*mut Context, *mut Context), f: Box<FnBox<(), ()>>) -> ! {
-    unsafe {
-      let (native, green) = *ctx;
-
-      f();
-
-      Context::swap(&mut *green, &mut *native);
-      abort();
+    (*ctx_slot).swap();
   }
 }
-
-  let mut ctx = Box::new((&mut native as *mut Context, null_mut()));
-  let mut green = Context::new(init, &mut *ctx as *mut _, f);
-  ctx.1 = &mut green as *mut Context;
-
-  unsafe {
-    Context::swap(&mut native, &mut green);
-  }
-
-  println!("size_of::<Context>() == {}", std::mem::size_of::<Context>());
-}

src/platform.rs

@@ -1,8 +1,10 @@
+#![allow(unstable)]
 extern crate libc;
 extern crate std;
 use self::std::prelude::v1::*;
 use self::std::os::{errno, page_size, MemoryMap};
 use self::std::os::MapOption::{MapReadable, MapWritable, MapNonStandardFlags};
+use stack;
 
 extern "C" {
   #[link_name = "lwt_stack_register"]
@@ -62,14 +64,14 @@ impl Drop for Stack {
   }
 }
 
-impl Stack {
-  pub fn top(&mut self) -> *mut u8 {
+impl stack::Stack for Stack {
+  fn top(&mut self) -> *mut u8 {
     unsafe {
       self.buf.data().offset(self.buf.len() as isize)
     }
   }
 
-  pub fn limit(&self) -> *const u8 {
+  fn limit(&self) -> *const u8 {
     unsafe {
       self.buf.data().offset(page_size() as isize) as *const _
     }

src/stack.rs

@@ -1,35 +1,4 @@
-use platform;
-use core::ptr;
-
-pub enum Stack {
-  Native {
-    sp_limit: *const u8
-  },
-  Managed(platform::Stack)
-}
-
-impl Stack {
-  pub fn new(size: usize) -> Stack {
-    Stack::Managed(platform::Stack::new(size))
-  }
-
-  pub unsafe fn native(limit: *const u8) -> Stack {
-    Stack::Native {
-      sp_limit: limit
-    }
-  }
-
-  pub fn top(&mut self) -> *mut u8 {
-    match *self {
-      Stack::Native { .. } => ptr::null_mut(),
-      Stack::Managed(ref mut stack) => stack.top()
-    }
-  }
-
-  pub fn limit(&self) -> *const u8 {
-    match *self {
-      Stack::Native { sp_limit, .. } => sp_limit,
-      Stack::Managed(ref stack) => stack.limit()
-    }
-  }
+pub trait Stack {
+  fn top(&mut self) -> *mut u8;
+  fn limit(&self) -> *const u8;
 }

src/swap.s

@@ -0,0 +1,17 @@
+sub $$128, %rsp
+pushq %fs:0x70
+pushq %rbp
+call 1f
+
+popq %rbp
+popq %fs:0x70
+add $$128, %rsp
+jmp 2f
+
+1:
+  movq (%rdi), %rax
+  movq %rsp, (%rdi)
+  movq %rax, %rsp
+  popq %rax
+  jmpq *%rax
+2: