The GC bitmap length is measured in multiples of the pointer alignment.
This is equal to the pointer size on all architectures except AVR.
Replace the hardcoded lengths with lengths that are computed naturally.
The compiler now implements the copy builtin directly instead of calling sliceCopy.
The length is calculated with the llvm.umax.* intrinsics, and the move is performed by llvm.memmove.*.
Both of these operations are easily understood by LLVM's optimization passes.
The type's alignment is also provided to llvm.memmove.*, which is useful when rewriting the move.
Interp no longer needs to reimplement sliceCopy.
Some edge case handling was implemented by sliceCopy but not llvm.memmove.*/llvm.memcpy.*.
I copied this over, so copies of external slices should work now.
Volatile moves/copies are now run at runtime by interp.
There is a 4-byte size increase due to some confusing length logic in sendUSBPacket.
I will look at sendUSBPacket in a future PR.
In order to scan stacks, the GC preempts all other threads and has them scan their own stack.
This is somewhat expensive since all of these threads have to fight over a single lock.
Instead, save the stack bounds and let the GC thread perform the scan.
This also fixes a few other bugs I ran into:
1. The GC starts scanning before the world stops. This can cause it to miss some objects (and mistakenly free them) if memory is modified while stopping.
2. The GC does not wait for threads to resume. This can cause notifications to be misinterpreted due to signal nesting if the GC is re-run before all threads wake.
Previously, we created joinable threads.
As a result, all threads would be kept alive after completion so that pthread_join could wait for them.
We never call pthread_join, so threads would never get cleaned up.
Additionally, the thread creation error check was performed after waiting for the thread to start.
If pthread_create failed, the caller would get stuck waiting for a thread that was never created.
Found this bug while trying to use the upstream testing package instead
of our own. The io/fs package wasn't passing, because the test was run
in a separate goroutine (and therefore a separate thread, with its own
stack) instead of all in the same thread with our own stack
creation/switching implementation.
This makes sure system calls like read don't return EINTR but instead
restart the call on an interrupt. This is by far the more sensible
option, the default POSIX behavior of returning EINTR is extremely
error-prone.
Found this bug while trying to use the upstream testing package instead
of our own.
The previous versions calculated at init() prevented `interp` from running
in many cases, increasing compile times due to the increased need to revert
the partially interpreted results and also increasing binary runtime because
fewer optimizations had happened during interp.
This commit adds support for a scheduler that runs a scheduler on all
available cores. It is meant to be used on baremetal systems with a
fixed number of cores, such as the RP2040.
The initial implementation adds support for multicore scheduling to the
riscv-qemu target as a convenient testing target. This means that this
new multicore scheduler is tested in CI, including a bunch of standard
library tests (`make tinygo-test-baremetal`). This should ensure the new
scheduler is reasonably well tested before trying to use it on
harder-to-debug targets like the RP2040.
The system stack is only needed when we're not on it. So we can directly
call task.SystemStack() without problems.
This also saves a tiny bit of binary size.
For the threads scheduler, it makes sense to have NumCPU available.
For all other schedulers, the number of available CPUs is practically
limited to one by the scheduler (even though the system might have more
CPUs).
This is more descriptive: the call is to exit a task, not to pause it.
This also makes it more obvious that there's an optimization
opportunity: to free the stack explicitly after the goroutine returns
(or to keep it as a cache for the next stack allocation).
This is not a scheduler in the runtime, instead every goroutine is
mapped to a single OS thread - meaning 1:1 scheduling.
While this may not perform well (or at all) for large numbers of
threads, it greatly simplifies many things in the runtime. For example,
blocking syscalls can be called directly instead of having to use epoll
or similar. Also, we don't need to do anything special to call C code -
the default stack is all we need.
This directive caused the code to be put in a non-executable area on
Windows which caused a segmentation fault. This patch fixes the issue by
removing `.section` directives, fixing windows/arm64 support.
The upstream one assumes it's running on a Unix system (which makes
sense), but this package is also used on baremetal. So replace it on
systems that need a replaced syscall package.
The mutex implementation needs a different implementation once support
for threading lands. This implementation just moves code to the
internal/task package to centralize these algorithms.
See the code comments for details. But in short, this implements a futex
for the cooperative scheduler (that is single threaded and
non-reentrant). Similar implementations can be made for basically every
other operating system, and even WebAssembly with the threading
(actually: atomics) proposal.
Using this basic futex implementation means we can use the same
implementation for synchronisation primitives on cooperative and
multicore systems.
For more information on futex across operating systems:
https://outerproduct.net/futex-dictionary.html
This adds some non-atomic types that have the same interface as the ones
in sync/atomic.
We currently don't need these to be atomic (because the scheduler is
entirely cooperative), but once we add support for a scheduler with
multiple threads and/or preemptive scheduling we can trivially add some
type aliases under a different build tag in the future when real
atomicity is needed for threading.
Previously the assembler was reordering this code:
jal tinygo_scanstack
move $a0, $sp
Into this:
jal tinygo_scanstack
nop
move $a0, $sp
So it was "helpfully" inserting a branch delay slot, even though this
was already being taken care of.
Somehow this didn't break, but it does break in the WIP threading branch
(https://github.com/tinygo-org/tinygo/pull/4559) where this bug leads to
a crash.
This uses uint32 instead of uint64. The reason for this is that uint64
atomic operations aren't universally available (especially on 32-bit
architectures). We could also use uintptr, but that seems needlessly
complicated: it's unlikely real-world programs will use more than a
billion select states (2^30).
This requires an API introduced in MacOS 11. I think that's fine, since
the version before that (MacOS 10.15) is EOL since 2022. Though if
needed, we could certainly work around it by using an older and slightly
less nice API.
Damian Gryski
Dima Jolkin
Jake Bailey
Nia Waldvogel
cjpeterson
Ayke van Laethem
Randy Reddig
Elliott Sales de Andrade