Previously we only supported recursive types in structs. But there can
be other kinds of recursive types, like slices:
type RecursiveSlice []RecursiveSlice
This doesn't involve structs, so it led to infinite recursion in the
compiler. This fix avoids recursion at the proper level: at the place
where the named type is defined.
Previously we were using a really weird calculation to determine the
alignment in bits - written by me (no idea what I was thinking at the
time, it's obviously incorrect). Just replace it with the much more
obviously correct multiplication by 8 to get bits from bytes.
Found while working on properly dealing with alignment in `-size=full`.
There was a mostly benign race condition in the compiler. The issue was
that there is a check for type aliases (which can alias types in another
function), but this check was _after_ accessing a property of the
function that might not have been completed.
I don't think this can have any serious effects, as the function is
skipped anyway, but such bugs should certainly be fixed.
If there is no source location, at least use the file (but not line) for
debug information.
This attributes almost 1kB of string data for ./testdata/stdlib.go when
compiling for WASI.
This is helpful for WebAssembly: it makes it possible to attribute many
more data to locations in the source code, which makes `-size=full` more
useful.
I have checked this conversion is not needed anymore after the previous
commit, by running various smoke tests of which none triggered this
optimization. The only case where the optimization would have kicked in
is in syscall/syscall_windows.go:76 of the Go standard library.
Therefore, I prefer to remove it to reduce code complexity.
This metadata is emitted by Clang and I found it is important for source
level debugging on MacOS. This patch does not get source level debugging
to work yet (for that, it seems like packages need to be built
separately), but it is a step in the right direction.
This is a big commit that changes the way runtime type information is stored in
the binary. Instead of compressing it and storing it in a number of sidetables,
it is stored similar to how the Go compiler toolchain stores it (but still more
compactly).
This has a number of advantages:
* It is much easier to add new features to reflect support. They can simply
be added to these structs without requiring massive changes (especially in
the reflect lowering pass).
* It removes the reflect lowering pass, which was a large amount of hard to
understand and debug code.
* The reflect lowering pass also required merging all LLVM IR into one
module, which is terrible for performance especially when compiling large
amounts of code. See issue 2870 for details.
* It is (probably!) easier to reason about for the compiler.
The downside is that it increases code size a bit, especially when reflect is
involved. I hope to fix some of that in later patches.
This was actually surprising once I got TinyGo to build on Windows 11
ARM64. All the changes are exactly what you'd expect for a new
architecture, there was no special weirdness just for arm64.
Actually getting TinyGo to build was kind of involved though. The very
short summary is: install arm64 versions of some pieces of software
(like golang, cmake) instead of installing them though choco. In
particular, use the llvm-mingw[1] toolchain instead of using standard
mingw.
[1]: https://github.com/mstorsjo/llvm-mingw/releases
Using ThinLTO manages to optimize binaries quite significantly. The
exact amount varies a lot by program but it's about 10-15% usually.
Don't remove non-ThinLTO support yet. It would not surprise me if this
triggered some unintended side effect. Eventually, non-ThinLTO support
should be removed though.
I found that when I enable ThinLTO, a miscompilation triggers that had
been hidden all the time previously. The bug appears to happen as
follows:
1. TinyGo generates a function with a runtime.trackPointer call, but
without an alloca (or the alloca gets optimized away).
2. LLVM sees that no alloca needs to be kept alive across the
runtime.trackPointer call, and therefore it adds the 'tail' flag.
One of the effects of this flag is that it makes it undefined
behavior to keep allocas alive across the call (which is still safe
at that point).
3. The GC lowering pass adds a stack slot alloca and converts
runtime.trackPointer calls into alloca stores.
The last step triggers the bug: the compiler inserts an alloca where
there was none before but that's not valid as long as the 'tail' flag is
present.
This patch fixes the bug in a somewhat dirty way, by always creating a
dummy alloca so that LLVM won't do the optimization in step 2 (and
possibly other optimizations that rely on there being no alloca
instruction).
This is purely a refactor, to make the next change simpler.
The wordpack functionality used to be necessary in transform passes, but
now it isn't anymore so let's not complicate this more than necessary.
This implements the block-based GC as a partially precise GC. This means
that for most heap allocations it is known which words contain a pointer
and which don't. This should in theory make the GC faster (because it
can skip non-pointer object) and have fewer false positives in a GC
cycle. It does however use a bit more RAM to store the layout of each
object.
Right now this GC seems to be slower than the conservative GC, but
should be less likely to run out of memory as a result of false
positives.
This makes it much easier to read the value at runtime, as pointer
indices are naturally little endian. It should not affect anything else
in the program.
There were two types that could result in a compiler stack overflow.
This is difficult to fix in LLVM 14, so I won't even bother. However,
this is trivial to fix with opaque pointers in LLVM 15. Therefore, this
fix is for LLVM 15 only.
Fixes: https://github.com/tinygo-org/tinygo/issues/3341
ThinLTO results in a small code size reduction, which is nice
(especially on these very small chips). It also brings us one step
closer to using ThinLTO everywhere.
This flag is necessary in LLVM 15 because it appears that LLVM 15 has
changed the default target ABI from lp64 to lp64d. This results in a
linker failure. Setting the "target-abi" forces the RISC-V backend to
use the intended target ABI.
A number of llvm.Const* functions (in particular extractvalue and
insertvalue) were removed in LLVM 15, so we have to use a builder
instead. This builder will create the same constant values, it simply
uses a different API.
--allow-undefined can be a problem: it allows compiling code that will
fail when loaded. This change makes sure that if some symbols are
undefined, they are reported as an error by the linker.
Previously, people could get away with importing a function that was not
defined, like this:
func add(int a, int b) int
func test() {
println(add(3, 5))
}
This was always unintended but mostly worked. With this change, it isn't
possible anymore. Now every function needs to be marked with //export
explicitly:
//export add
func add(int a, int b) int
func test() {
println(add(3, 5))
}
As before, functions will be placed in the `env` module with the name
set from the `//export` tag. This can be overridden with
`//go:import-module`:
//go:import-module math
//export add
func add(int a, int b) int
func test() {
println(add(3, 5))
}
For the syscall/js package, I needed to give a list of symbols that are
undefined. This list is based on the JavaScript functions defined in
targets/wasm_exec.js.
This is really a few more-or-less separate changes:
* Remove all math aliases that were used in Go 1.16 and below (the
math.[A-Z] aliases).
* Replace math aliases with an assembly implementation (the math.arch*
aliases) with a LLVM intrinsic, where one is available.
* Include missing math functions in picolibc build.
This leaves just four math aliases:
* math.archHypot and math.archModf do not have a LLVM builtin
equivalent. They could be replaced with calls to libm, and I think
that would be a good idea in the long term.
* math.archMax and math.archMin do have a LLVM builtin equivalent
(llvm.maximum.f64, llvm.minimum.f64), but unfortunately they crash
when used. Apparently these exact operations are not yet widely
supported in hardware and they don't have a libm equivalent either.
There are more LLVM builtins that we could use for the math package
(such as FMA), but I will leave that to a future change. It could
potentially speed up some math operations.
Instead of changing the calls, replace the function bodies themselves.
This is useful for a number of reasons, see
https://github.com/tinygo-org/tinygo/pull/2920 for more information.
I have removed the math intrinsics tests because they are no longer
useful. Instead, I think `tinygo test math` should suffice.
This gives some more optimization opportunities to LLVM, because it
understands these intrinsics. For example, it might convert
llvm.sqrt.f64 to llvm.sqrt.f32 if possible.
For some reason, the type of a function parameter can sometimes be of
interface type, while it should be the underlying type. This might be a
bug in the x/tools/go/ssa package but this is a simple workaround.
