Implement SetStallEPIn, SetStallEPOut, ClearStallEPIn, and ClearStallEPOut methods on the nRF52840 USBDevice struct using the hardware EPSTALL register to support MSC driver requirements. Also, correct the handleUSBIRQ loops to iterate over physical endpoint numbers rather than dynamic configuration entries to prevent missed or incorrectly routed endpoint interrupts.
Replace static endpoint variables with EndpointIN and EndpointOUT
functions. Allows flexible endpoint remapping across USB configs.
Map CDC, HID, MIDI and MSC USB devices to bidirectional endpoints consistently across different configurations.
Emit fault checkpoints around compiler-generated runtime assertions so
runtimePanicAt can unwind through the existing defer/recover machinery
instead of aborting. This lets panics from bounds checks, type checks,
and other compiler-inserted runtime checks be recovered by deferred
functions.
Mark functions that call recover as noinline. Inlining such a function
into a deferred closure can make recover observe the wrong call context
and report success when it should return nil.
Addresses tinygo-org/tinygo issues 2759 and 3510.
This has the following effects:
* It fixes issue #5414.
* It bumps the minimum Go version to Go 1.24 (from 1.23).
I looked, and there doesn't seem to be a version of golang.org/x/tools
that supports the improved `new` builtin without also bumping the
minimum Go version.
* interp: bail out of loops that iterate too many times
The existing loop guard (errLoopUnrolled) only fires when a loop body
emits runtime instructions. Loops that are fully evaluable at compile
time, such as inserting thousands of entries into a map, were not
caught and could hang the compiler.
Add a per-basic-block iteration counter that triggers a recoverable
error (errLoopTooLong) when any block is entered more than 1000 times
in a single function call. This defers the init function to runtime,
which is the same behavior as other interp bailouts.
Profiling showed that 83% of CPU time was spent in GC, caused by
allocation pressure from the interp memory cloning on each map
mutation. The iteration limit avoids this entirely by bailing out
before the quadratic cost becomes significant.
Performance on the reproducer from #2090 (map init with strconv.Itoa):
entries before after
5,000 7.4s 2.1s
10,000 17.5s 2.1s
20,000 48.0s 2.8s
65,536 >180s (OOM) 3.2s
* Add -interp-loop-limit
Several places used the (*[1 << N]T)(ptr)[:len:len] pattern to convert
C pointers to Go slices. This has a hardcoded size limit that can panic
if exceeded; RunTool hit this with >1024 linker arguments when many
files are embedded.
Replace all instances with unsafe.Slice, which handles any size.
When a string variable already had a source-level default value, the -ldflags
"-X" flag was silently ignored and the variable remained the default value
at runtime.
Fix by stripping the InitOrder entry for each -X variable before
LoadSSA() is called. With no entry in InitOrder, go/ssa emits no init
store, so the global stays zero-valued in the IR. makeGlobalsModule
still injects the actual -X values at final link time.
The -X values remain out of the per-package build cache with only the
variable names appearing in the cache key.
Signed-off-by: deadprogram <ron@hybridgroup.com>
The -march flag for compiling C libraries (compiler-rt, picolibc) was
hardcoded to rv32imac for all riscv32 targets. This is incorrect for
targets that lack the atomic extension, such as ESP32-C3 (rv32imc) and
TKey (rv32iczmmul).
Extract the -march value from the target's cflags when available,
falling back to the previous defaults (rv32imac, rv64gc) for targets
that don't override it.
Fixes#5114
Signed-off-by: deadprogram <ron@hybridgroup.com>
The previous approach of erasing basic blocks one-by-one from
duplicate function definitions could crash with a segfault. When
a function has complex control flow (branches, switches, PHI nodes),
deleting a basic block that is still referenced by instructions in
other blocks of the same function causes use-after-free in LLVM.
LLVM's C++ Function::deleteBody() avoids this by calling
dropAllReferences() on all blocks first, but that API is not
exposed in the Go LLVM bindings.
Fix this by using a completely different approach: instead of
manually deleting the function body, set the duplicate function's
linkage to LinkOnceODRLinkage. This tells the LLVM linker that the
definition can be merged with another copy, causing it to prefer
the already-linked ExternalLinkage definition in the destination
module. The result is the same (the runtime's definition wins) but
without any unsafe block manipulation.
Go 1.26 changed syscall.loadlibrary, syscall.loadsystemlibrary, and
syscall.getprocaddress from declarations to definitions in
syscall/dll_windows.go. TinyGo's runtime also defines these via
//go:linkname, causing "symbol multiply defined!" during LLVM module
linking.
Resolve this by detecting duplicate function definitions before calling
llvm.LinkModules and turning the incoming duplicate into a declaration,
so the runtime's version wins. TinyGo's implementations must take
precedence because Go 1.26's versions depend on //go:cgo_import_dynamic,
which TinyGo does not support.
Also fix the signature of syscall_loadsystemlibrary to match the
standard library (remove unused absoluteFilepath parameter).
Signed-off-by: deadprogram <ron@hybridgroup.com>
This fixes an "index out of range" panic that occurs when calculating
binary sizes. The strings.SplitN operation in findPackagePath can sometimes
return a slice with only one element, so we now verify the length of the
slice before attempting to access the second element.
Signed-off-by: deadprogram <ron@hybridgroup.com>
* begin adding ring512 implementation
* refactor to make operation driven fuzz test
* refactor USBCDC.Read to use ring512
* try txhandler separate to flush
* working USBCDC with large packets
* fix binary size
* remove comment
* documentation improvements
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.
Remove the sendUSBPacket maxLen param because this greatly confused the compiler.
It also fixes a bug where the length provided to the hardware may not match the length of the packet.
sendUSBPacket now panics if the sent packet is too big.
I also fixed some of the string descriptor logic where we could create a packet without fully populating it.
RP2* systems might require some more work since they are implemented very differently?
I don't have any of those to test with yet, so maybe someone can deal with them in a seperate PR?
Instead of looping over each block, we can use bit hacks to operate on an entire state byte.
I deinterleaved the state bits in order to enable these tricks.
Sweep used to count free/freed allocations/blocks.
I managed to move/remove all of these counters:
- The free space is now calculated in buildFreeRanges by adding the range lengths.
- ReadMemStats counts freed objects by subtracting live objects from allocated objects.
- gcFreedBlocks was never necessary because MemStats.HeapAlloc is the same as MemStats.HeapInUse.
The allocator originally just looped through the blocks until it found a sufficiently-long range.
This is simple, but it fragments very easily and can degrade to a full heap scan for long requests.
Instead, we now maintain a sorted nested list of free ranges by size.
The allocator will select the shortest sufficient-length range, generally reducing fragmentation.
This data structure can find a range in time directly proportional to the requested length.
Loop over valid pointer locations in heap objects instead of checking if each location is valid.
The conservative scanning code is now shared between markRoots and the heap scan.
This also removes the ending alignment requirement from markRoots, since the new scan* functions do not require an aligned length.
This requirement was occasionally violated by the linux global marking code.
This saves some code space and has negligible impact on performance.
The blocks GC originally used a fixed-size stack to hold objects to scan.
When this stack overflowed, the GC would fully rescan all marked objects.
This could cause the GC to degrade to O(n^2) when scanning large linked data structures.
Instead of using a fixed-size stack, we now add a pointer field to the start of each object.
This pointer field is used to implement an unbounded linked stack.
This also consolidates the heap object scanning into one place, which simplifies the process.
This comes at the cost of introducing a pointer field to the start of the object, plus the cost of aligning the result.
This translates to:
- 16 bytes of overhead on x86/arm64 with the conservative collector
- 0 bytes of overhead on x86/arm64 with the precise collector (the layout field cost gets aligned up to 16 bytes anyway)
- 8 bytes of overhead on other 64-bit systems
- 4 bytes of overhead on 32-bit systems
- 2 bytes of overhead on AVR
This fixes a bug where runtime.alloc would not clear the padding from rounding the allocation up to a multiple of the block size.
The GC still has to scan this, and this could result in permanent false positives (and therefore memory leaks).
It also handles overflow in the size calculation.
Long sleep durations weren't implemented correctly, by simply casting to
a smaller number of bits. What we want is a saturating downcast, which
is what I've used here instead.
This should fix https://github.com/tinygo-org/tinygo/issues/3356.
By disabling some configuration options (and updating the library), the
library becomes a lot smaller. With `-no-debug`, binaries become ~18kB
smaller on Linux, ~4kB smaller on MacOS, ~9kB smaller on Windows, and
~12kB smaller on WebAssembly.
This function is called when a hard fault occurs. Hard faults happen
when something really bad happens - like writing to unwritable memory or
an unaligned memory access on Cortex-M0. It is not generally possible to
recover from these.
This commit optimizes the code size overhead of hard fault handling:
* It removes the stack overflow checking code.
This may seem like a bad thing, but the only thing this could check
were stack overflows outside goroutines. In practice, this could
only really happen on a stack overflow in the scheduler (unlikely),
or in interrupt code (possible, but interrupts are small so still
unlikely). Most stack overflows happen in regular goroutines, and
weren't caught in the HardFault.
* It makes the panic message similar to a regular panic. This has two
advantages:
* It reduces code size, because the string can be reused between
the HardFault handler and the runtime panic function.
* Using the same pattern automatically makes `-monitor` print the
source address for the hard fault. Not a big benefit as we could
trivially add any other pattern but a nice benefit nonetheless.
Result:
$ tinygo flash -target=microbit -size=short -programmer=openocd -monitor examples/serial
code data bss | flash ram
3036 8 2256 | 3044 2264
[...snip]
Connected to /dev/ttyACM0. Press Ctrl-C to exit.
panic: runtime error at 0x00000344: HardFault with sp=0x200007d0
[tinygo: panic at /home/ayke/src/tinygo/tinygo/src/internal/task/task_stack_cortexm.go:48:4]
(This is with https://github.com/tinygo-org/tinygo/pull/3680 not yet
fixed and some local changes to configure the UART so I can actually see
the panic).
For atsamd21/nrf51 chips this results in a binary size reduction of
around 100 bytes. For other Cortex-M chips it's around 24 bytes but I
hope to change this in the future because a lot of the fault decoding in
runtime_cortexm_hardfault_debug.go should IMHO be done by the TinyGo
monitor instead (I estimate that this would save around 800 bytes on
these chips).
This switches the Espressif fork from LLVM 19 to LLVM 20, so we can use
the improvements made between those LLVM versions. It also better aligns
with the system-LLVM build method, which currently also defaults to LLVM
20.
Note that this disables the machine outliner for RISC-V. It appears
there's a bug in there somewhere, with the machine outliner enabled the
crypto/elliptic package tests fail with -target=riscv-qemu.
This should ideally be investigated and reported upstream.
Account for the sleep queue base time in the computation of the wakeup
time.
Tested with the following program on pico2.
func main() {
go func() {
for i := range 60 {
const delay = 20 * time.Millisecond
before := time.Now()
time.Sleep(delay)
if d := time.Since(before); true || d < delay {
log.Println(i, "actual", d, "delay", delay)
}
}
}()
time.Sleep(500 * time.Millisecond)
log.Println("******** done sleeping ********")
select {}
}
Without this change, the program would print lines such as:
17 actual 15.494ms delay 20ms
18 actual 15.49ms delay 20ms
19 actual 15.585ms delay 20ms
20 actual 15.493ms delay 20ms
21 actual 15.494ms delay 20ms
22 actual 15.487ms delay 20ms
23 actual 15.498ms delay 20ms
******** done sleeping ********
24 actual 15.548ms delay 20ms
25 actual 20.011ms delay 20ms
26 actual 20.01ms delay 20ms
27 actual 20.011ms delay 20ms
28 actual 20.015ms delay 20ms
Note that while more than one sleeping goroutine is in the timer queue,
the sleep duration is 5ms short.
This adds support for `-gc=boehm` on `-target=wasip1` and `-target=wasm`
(in a browser or NodeJS). Notably it does *not* add Boehm GC support for
`-target=wasip2`, since that target doesn't have a real libc.
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.
This adds support for the 'go install' case, where we link against the
system LLVM (Homebrew, Linux distro version, etc). This does not have an
effect on `make`, which still uses LLVM 19 for now.
The main reason for doing this is because newer distros like Fedora 42
and Homebrew have switched to LLVM 20, so switching to this newer
version helps those people. (I'm one of those people, I'm on Fedora 42).
There are two small changes for WebAssembly included:
* nontrapping-fptoint was enabled in the wasmbuiltin library used for
wasm-unknown. This matches wasm-unknown which already had this
enabled, so it doesn't add any new instructions.
* bulk-memory was enabled in wasi-libc. This was previously enabled in
all the other WebAssembly targets (wasm, wasip1, wasip2) so this
does't add any new instructions. I think it was also enabled in the
wasi-libc build before
https://github.com/tinygo-org/tinygo/pull/4820 but I don't know for
sure.
Without this change, a pending interrupt would spuriously trigger
immediately after enabling. This happens if an interrupt is triggered
during flashing (e.g. by DMA), which survives the subsequent reset.
This behaviour matches e.g. `machine.irqSet` in machine_rp2_rp2350.go.
See 7f970a45, whose symptoms were likely caused by spurious interrupts.
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 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.
Konstantin Sharlaimov
あーるどん
sago35
Jake Bailey
Ayke van Laethem
deadprogram
Damian Gryski
Pat Whittingslow
Elias Naur
Nia Waldvogel
cjpeterson
sivchari
luoliwoshang