Skip to content

Commit

Permalink
Add support for ARM
Browse files Browse the repository at this point in the history
  • Loading branch information
Amanieu committed Nov 2, 2016
1 parent ba5bb38 commit 81fb8ec
Show file tree
Hide file tree
Showing 6 changed files with 320 additions and 9 deletions.
2 changes: 1 addition & 1 deletion Cargo.toml
Original file line number Diff line number Diff line change
Expand Up @@ -15,7 +15,7 @@ libc = "0.2.14"
optional = true
version = "1.0.0"

[dev-dependencies]
[target.'cfg(any(target_arch = "x86", target_arch = "x86_64", target_arch = "aarch64"))'.dev-dependencies]
simd = "0.1"

[features]
Expand Down
14 changes: 8 additions & 6 deletions README.md
Original file line number Diff line number Diff line change
Expand Up @@ -118,7 +118,7 @@ there should be at least 8 KiB of free stack space, or panicking will result in

## Limitations

The architectures currently supported are: x86, x86_64, aarch64, or1k.
The architectures currently supported are: x86, x86_64, aarch64, arm, or1k.

The platforms currently supported are: bare metal, Linux (any libc),
FreeBSD, DragonFly BSD, macOS.
Expand Down Expand Up @@ -176,13 +176,15 @@ of callee-saved registers.

### Call stack splicing

Non-Windows platforms use [DWARF][] for both stack unwinding and debugging. DWARF call frame
information is very generic to be ABI-agnostic—it defines a bytecode that describes the actions
that need to be performed to simulate returning from a function. libfringe uses this bytecode
to specify that, after the generator function has returned, execution continues at the point
where the generator function was resumed the last time.
Non-Windows platforms use [DWARF][] (or the highly similar [ARM EHABI][ehabi]) for both stack
unwinding and debugging. DWARF call frame information is very generic to be ABI-agnostic—
it defines a bytecode that describes the actions that need to be performed to simulate
returning from a function. libfringe uses this bytecode to specify that, after the generator
function has returned, execution continues at the point where the generator function was
resumed the last time.

[dwarf]: http://dwarfstd.org
[ehabi]: http://infocenter.arm.com/help/topic/com.arm.doc.ihi0038b/IHI0038B_ehabi.pdf

## Windows compatibility

Expand Down
304 changes: 304 additions & 0 deletions src/arch/arm.rs
Original file line number Diff line number Diff line change
@@ -0,0 +1,304 @@
// This file is part of libfringe, a low-level green threading library.
// Copyright (c) Nathan Zadoks <nathan@nathan7.eu>,
// whitequark <whitequark@whitequark.org>
// Amanieu d'Antras <amanieu@gmail.com>
// Licensed under the Apache License, Version 2.0, <LICENSE-APACHE or
// http://apache.org/licenses/LICENSE-2.0> or the MIT license <LICENSE-MIT or
// http://opensource.org/licenses/MIT>, at your option. This file may not be
// copied, modified, or distributed except according to those terms.

// To understand the machine code in this file, keep in mind these facts:
// * ARM AAPCS ABI passes the first argument in r0. We also use r0 to pass a value
// while swapping context; this is an arbitrary choice
// (we clobber all registers and could use any of them) but this allows us
// to reuse the swap function to perform the initial call.
//
// To understand the ARM EHABI CFI code in this file, keep in mind these facts:
// * CFI is "call frame information"; a set of instructions to a debugger or
// an unwinder that allow it to simulate returning from functions. This implies
// restoring every register to its pre-call state, as well as the stack pointer.
// * CFA is "call frame address"; the value of stack pointer right before the call
// instruction in the caller. Everything strictly below CFA (and inclusive until
// the next CFA) is the call frame of the callee. This implies that the return
// address is the part of callee's call frame.
// * Logically, ARM EHABI CFI is a table where rows are instruction pointer values and
// columns describe where registers are spilled (mostly using expressions that
// compute a memory location as CFA+n). A .save pseudoinstruction changes
// the state of a column for all IP numerically larger than the one it's placed
// after. A .pad or .setfp pseudoinstructions change the CFA value similarly.
// * Simulating return is as easy as restoring register values from the CFI table
// and then setting stack pointer to CFA.
//
// A high-level overview of the function of the trampolines is:
// * The 2nd init trampoline puts a controlled value (written in swap to `new_cfa`)
// into r11. This is then used as the CFA for the 1st trampoline.
// * This controlled value points to the bottom of the stack of the parent context,
// which holds the saved r11 and lr from the call to swap().
// * The 1st init trampoline tells the unwinder to restore r11 and lr
// from the stack frame at r11 (in the parent stack), thus continuing
// unwinding at the swap call site instead of falling off the end of context stack.
use core::mem;
use arch::StackPointer;
use unwind;

pub const STACK_ALIGNMENT: usize = 8;

pub unsafe fn init(stack_base: *mut u8, f: unsafe fn(usize, StackPointer)) -> StackPointer {
#[cfg(not(target_vendor = "apple"))]
#[naked]
unsafe extern "C" fn trampoline_1() {
asm!(
r#"
# gdb has a hardcoded check that rejects backtraces where frame addresses
# do not monotonically decrease. It is turned off if the function is called
# "__morestack" and that is hardcoded. So, to make gdb backtraces match
# the actual unwinder behavior, we call ourselves "__morestack" and mark
# the symbol as local; it shouldn't interfere with anything.
__morestack:
.local __morestack
# Set up the first part of our ARM EHABI CFI linking stacks together. When
# we reach this function from unwinding, r11 will be pointing at the bottom
# of the parent linked stack. This link is set each time swap() is called.
# When unwinding the frame corresponding to this function, a ARM EHABI unwinder
# will use r11+16 as the next call frame address, restore return address (lr)
# from CFA-8 and restore r11 from CFA-16. This mirrors what the second half
# of `swap_trampoline` does.
# .setfp fp, sp
# .save {fp, lr}
.cfi_def_cfa fp, 8
.cfi_offset lr, -4
.cfi_offset fp, -8
# This nop is here so that the initial swap doesn't return to the start
# of the trampoline, which confuses the unwinder since it will look for
# frame information in the previous symbol rather than this one. It is
# never actually executed.
nop
.Lend:
.size __morestack, .Lend-__morestack
"#
: : : : "volatile")
}

#[cfg(target_vendor = "apple")]
#[naked]
unsafe extern "C" fn trampoline_1() {
asm!(
r#"
# Identical to the above, except avoids .local/.size that aren't available on Mach-O.
__morestack:
.private_extern __morestack
# .setfp fp, sp
# .save {fp, lr}
.cfi_def_cfa fp, 8
.cfi_offset lr, -4
.cfi_offset fp, -8
nop
"#
: : : : "volatile")
}

#[naked]
unsafe extern "C" fn trampoline_2() {
asm!(
r#"
# Set up the second part of our ARM EHABI CFI.
# When unwinding the frame corresponding to this function, a DWARF unwinder
# will restore r11 (and thus CFA of the first trampoline) from the stack slot.
# This stack slot is updated every time swap() is called to point to the bottom
# of the stack of the context switch just switched from.
# .setfp fp, sp
# .save {fp, lr}
.cfi_def_cfa fp, 8
.cfi_offset lr, -4
.cfi_offset fp, -8
# This nop is here so that the return address of the swap trampoline
# doesn't point to the start of the symbol. This confuses gdb's backtraces,
# causing them to think the parent function is trampoline_1 instead of
# trampoline_2.
nop
# Call unwind_wrapper with the provided function and the stack base address.
add r2, sp, #16
ldr r3, [sp, #8]
bl ${0}
# Restore the stack pointer of the parent context. No CFI adjustments
# are needed since we have the same stack frame as trampoline_1.
ldr sp, [sp]
# Load frame and instruction pointers of the parent context.
pop {fp, lr}
.cfi_adjust_cfa_offset -8
.cfi_restore fp
.cfi_restore lr
# If the returned value is nonzero, trigger an unwind in the parent
# context with the given exception object.
cmp r0, #0
bne ${1}
# Clear the stack pointer. We can't call into this context any more once
# the function has returned.
mov r1, #0
# Return into the new context. Use `r12` instead of `lr` to avoid
# return address mispredictions.
mov r12, lr
bx r12
"#
:
: "s" (unwind::unwind_wrapper as usize)
"s" (unwind::start_unwind as usize)
: : "volatile")
}

// We set up the stack in a somewhat special way so that to the unwinder it
// looks like trampoline_1 has called trampoline_2, which has in turn called
// swap::trampoline.
//
// There are 2 call frames in this setup, each containing the return address
// followed by the r11 value for that frame. This setup supports unwinding
// using DWARF CFI as well as the frame pointer-based unwinding used by tools
// such as perf or dtrace.
let mut sp = StackPointer::new(stack_base);

sp.push(0 as usize); // Padding to ensure the stack is properly aligned
sp.push(f as usize); // Function that trampoline_2 should call

// Call frame for trampoline_2. The CFA slot is updated by swap::trampoline
// each time a context switch is performed.
sp.push(trampoline_1 as usize + 4); // Return after the nop
sp.push(0xdead0cfa); // CFA slot

// Call frame for swap::trampoline. We set up the r11 value to point to the
// parent call frame.
let frame = sp.offset(0);
sp.push(trampoline_2 as usize + 4); // Entry point, skip initial nop
sp.push(frame as usize); // Pointer to parent call frame

sp
}

#[inline(always)]
pub unsafe fn swap_link(arg: usize, new_sp: StackPointer,
new_stack_base: *mut u8) -> (usize, Option<StackPointer>) {
let ret: usize;
let ret_sp: usize;
asm!(
r#"
# Set up the link register
adr lr, 0f
# Save the frame pointer and link register; the unwinder uses them to find
# the CFA of the caller, and so they have to have the correct value immediately
# after the call instruction that invoked the trampoline.
push {fp, lr}
# Pass the stack pointer of the old context to the new one.
mov r1, sp
# Link the call stacks together by writing the current stack bottom
# address to the CFA slot in the new stack.
str sp, [r3, #-16]
# Load stack pointer of the new context.
mov sp, r2
# Load frame and instruction pointers of the new context.
pop {fp, r12}
# Return into the new context. Use `r12` instead of `lr` to avoid
# return address mispredictions.
bx r12
0:
"#
: "={r0}" (ret)
"={r1}" (ret_sp)
: "{r0}" (arg)
"{r2}" (new_sp.0)
"{r3}" (new_stack_base)
:/*r0, r1,*/ "r2", "r3", "r4", "r5", "r6", "r7",
"r8", "r9", "r10",/*r11,*/"r12",/*sp,*/ "lr", /*pc,*/
"d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7",
"d8", "d9", "d10", "d11", "d12", "d13", "d14", "d15",
"d16", "d17", "d18", "d19", "d20", "d21", "d22", "d23",
"d24", "d25", "d26", "d27", "d28", "d29", "d30", "d31",
"cc", "memory"
: "volatile");
(ret, mem::transmute(ret_sp))
}

#[inline(always)]
pub unsafe fn swap(arg: usize, new_sp: StackPointer) -> (usize, StackPointer) {
// This is identical to swap_link, but without the write to the CFA slot.
let ret: usize;
let ret_sp: usize;
asm!(
r#"
adr lr, 0f
push {fp, lr}
mov r1, sp
mov sp, r2
pop {fp, r12}
bx r12
0:
"#
: "={r0}" (ret)
"={r1}" (ret_sp)
: "{r0}" (arg)
"{r2}" (new_sp.0)
:/*r0, r1,*/ "r2", "r3", "r4", "r5", "r6", "r7",
"r8", "r9", "r10",/*r11,*/"r12",/*sp,*/ "lr", /*pc,*/
"d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7",
"d8", "d9", "d10", "d11", "d12", "d13", "d14", "d15",
"d16", "d17", "d18", "d19", "d20", "d21", "d22", "d23",
"d24", "d25", "d26", "d27", "d28", "d29", "d30", "d31",
"cc", "memory"
// We need the "alignstack" attribute here to ensure that the stack is
// properly aligned if a call to start_unwind needs to be injected into
// our stack context.
: "volatile", "alignstack");
(ret, mem::transmute(ret_sp))
}

#[inline(always)]
pub unsafe fn unwind(new_sp: StackPointer, new_stack_base: *mut u8) {
// Argument to pass to start_unwind, based on the stack base address.
let arg = unwind::unwind_arg(new_stack_base);

// This is identical to swap_link, except that it performs a tail call to
// start_unwind instead of returning into the target context.
asm!(
r#"
adr lr, 0f
push {fp, lr}
str sp, [r3, #-16]
mov sp, r2
pop {fp, r12}
# Jump to the start_unwind function, which will force a stack unwind in
# the target context. This will eventually return to us through the
# stack link.
b ${0}
0:
"#
:
: "s" (unwind::start_unwind as usize)
"{r0}" (arg)
"{r2}" (new_sp.0)
"{r3}" (new_stack_base)
: "r0", "r1", "r2", "r3", "r4", "r5", "r6", "r7",
"r8", "r9", "r10",/*r11,*/"r12",/*sp,*/ "lr", /*pc,*/
"d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7",
"d8", "d9", "d10", "d11", "d12", "d13", "d14", "d15",
"d16", "d17", "d18", "d19", "d20", "d21", "d22", "d23",
"d24", "d25", "d26", "d27", "d28", "d29", "d30", "d31",
"cc", "memory"
: "volatile");
}
3 changes: 3 additions & 0 deletions src/arch/mod.rs
Original file line number Diff line number Diff line change
Expand Up @@ -13,6 +13,7 @@ use core::nonzero::NonZero;
#[cfg_attr(target_arch = "x86", path = "x86.rs")]
#[cfg_attr(target_arch = "x86_64", path = "x86_64.rs")]
#[cfg_attr(target_arch = "aarch64", path = "aarch64.rs")]
#[cfg_attr(target_arch = "arm", path = "arm.rs")]
#[cfg_attr(target_arch = "or1k", path = "or1k.rs")]
mod imp;

Expand Down Expand Up @@ -40,6 +41,7 @@ impl StackPointer {
#[cfg(test)]
mod tests {
extern crate test;
#[cfg(any(target_arch = "x86", target_arch = "x86_64", target_arch = "aarch64"))]
extern crate simd;

use arch::{self, StackPointer};
Expand All @@ -66,6 +68,7 @@ mod tests {
}
}

#[cfg(any(target_arch = "x86", target_arch = "x86_64", target_arch = "aarch64"))]
#[test]
fn context_simd() {
unsafe fn permuter(arg: usize, stack_ptr: StackPointer) {
Expand Down
2 changes: 1 addition & 1 deletion src/arch/x86.rs
Original file line number Diff line number Diff line change
Expand Up @@ -343,7 +343,7 @@ pub unsafe fn unwind(new_sp: StackPointer, new_stack_base: *mut u8) {

asm!(
r#"
call ${0:c}@plt
call ${0:c}
"#
:
: "s" (trampoline as usize)
Expand Down
4 changes: 3 additions & 1 deletion src/unwind.rs
Original file line number Diff line number Diff line change
Expand Up @@ -28,7 +28,9 @@ fn have_cross_stack_unwind() -> bool {
// for now.
// - iOS on ARM uses setjmp/longjmp instead of DWARF-2 unwinding, which needs
// to be explicitly saved/restored when switching contexts.
!(cfg!(windows) || cfg!(all(target_os = "ios", target_arch = "arm")))
// - LLVM doesn't currently support ARM EHABI directives in inline assembly so
// we instead need to propagate exceptions manually across contexts.
!(cfg!(windows) || cfg!(target_arch = "arm"))
}

// Wrapper around the root function of a generator which handles unwinding.
Expand Down

0 comments on commit 81fb8ec

Please sign in to comment.