This commit optimizes string literals and globals by setting the
appropriate alignment and using a nil pointer in zero-length strings.
- Setting the alignment for string values has a surprisingly large
effect, up to around 2% in binary size. I suspect that LLVM will
pick some default alignment for larger byte arrays if no alignment
has been specified and forcing an alignment of 1 will pack all
strings closer together.
- Using nil for zero-length strings also has a positive effect, but
I'm not sure why. Perhaps it makes some optimizations more trivial.
- Always setting the alignment on globals improves code size slightly,
probably for the same reasons setting the alignment of string
literals improves code size. The effect is much smaller, however.
This commit might have an effect on performance, but if it does this
should be tested separately and such a large win in binary size should
definitely not be ignored for small embedded systems.
Sometimes, LLVM may rename named structs when merging modules.
Therefore, we can't rely on typecodeID structs to retain their struct
names.
This commit changes the interface lowering pass to not rely on these
names. The interp package does however still rely on this name, but I
hope to fix that in the future.
This commit adds a new transform that converts reflect Implements()
calls to runtime.interfaceImplements. At the moment, the Implements()
method is not yet implemented (how ironic) but if the value passed to
Implements is known at compile time the method call can be optimized to
runtime.interfaceImplements to make it a regular interface assert.
This commit is the last change necessary to add basic support for the
encoding/json package. The json package is certainly not yet fully
supported, but some trivial objects can be converted to JSON.
This is important as golden test output and to verify that the output is
correct. Later improvements and bug fixes are clearly visible in the IR,
and unintentional changes will also be immediately spotted.
This patch fixes a use of the global context. I've seen a few instances
of crashes in the llvm.ConstInt function when called from
makeStructTypeFields, which I believe are caused by this bug.
Previously there was code to avoid impossible type asserts but it wasn't
great and in fact was too aggressive when combined with reflection.
This commit improves this by checking all types that exist in the
program that may appear in an interface (even struct fields and the
like) but without creating runtime.typecodeID objects with the type
assert. This has two advantages:
* As mentioned, it optimizes impossible type asserts away.
* It allows methods on types that were only asserted on (in
runtime.typeAssert) but never used in an interface to be optimized
away using GlobalDCE. This may have a cascading effect so that other
parts of the code can be further optimized.
This sometimes massively improves code size and mostly negates the code
size regression of the previous commit.
This distinction was useful before when reflect wasn't properly
supported. Back then it made sense to only include method sets that were
actually used in an interface. But now that it is possible to get to
other values (for example, by extracting fields from structs) and it is
possible to turn them back into interfaces, it is necessary to preserve
all method sets that can possibly be used in the program in a type
assert, interface assert or interface method call.
In the future, this logic will need to be revisited again when
reflect.New or reflect.Zero gets implemented.
Code size increases a bit in some cases, but usually in a very limited
way (except for one outlier in the drivers smoke tests). The next commit
will improve the situation significantly.
Previously we used the --export-all linker flag to export most
functions. However, this is not needed and possibly increases binary
size. Instead, we should be exporting the specific functions to be
exported.
An allocated object is never nil, so there is no need for a nil check.
This probably does not result in any better optimization (the nil check
is easily optimized away by LLVM because the result of runtime.alloc is
marked nonnull) but it makes the slice tests a bit cleaner.
It's difficult to create clean test cases while remaining compatible
with multiple LLVM versions. Most test outputs are much more readable
after an instcombine pass but instcombine rules change between LLVM
versions, leading to different (but semantically equivalent) test
outputs.
This reduces the test coverage a little bit (because old LLVM versions
aren't tested as well), but it als makes it easier to add more complex
tests.
In the future it might be a good idea to make the compiler output a bit
less messy so these workarounds are not needed.
This commit switches from the previous behavior of compiling the whole
program at once, to compiling every package in parallel and linking the
LLVM bitcode files together for further whole-program optimization.
This is a small performance win, but it has several advantages in the
future:
- There are many more things that can be done per package in parallel,
avoiding the bottleneck at the end of the compiler phase. This
should speed up the compiler futher.
- This change is a necessary step towards a non-LTO build mode for
fast incremental builds that only rebuild the changed package, when
compiler speed is more important than binary size.
- This change refactors the compiler in such a way that it will be
easier to inspect the IR for one package only. Inspecting this IR
will be very helpful for compiler developers.
The SimpleDCE pass was previously used to only compile the parts of the
program that were in use. However, lately the only real purpose has been
to speed up the compiler a bit by only compiling the necessary
functions.
This pass however is a problem for compiling (and caching) packages in
parallel. Therefore, this commit removes it as a preparatory step
towards that goal.
This is a leftover from a long time ago, when everything was still in
the global context. The fact that this uses the global context is most
certainly a bug.
I have seen occasional crashes in the build-packages-indepedently branch
(and PRs based on it) which I suspect are caused by this bug. I think
this is a long-dormant bug that only surfaced when doing the compilation
steps in parallel.
This doesn't yet add support for actually making use of variadic
functions, but at least allows (unintended) variadic functions like the
following to work:
void foo();
Because of a bug in the ARM backend of LLVM, the cmpxchg instruction is
lowered using ldrexd/strexd instructions which don't exist on Cortex-M
cores. This leads to an "undefined instruction" exception at runtime.
Therefore, this patch works around this by lowering directly to a call
to the __sync_val_compare_and_swap_8 function, which is what the backend
should be doing.
For details, see: https://reviews.llvm.org/D95891
To test this patch, you can run the code on a Cortex-M3 or higher
microcontroller, for example:
tinygo flash -target=pca10040 ./testdata/atomic.go
Before this patch, this would trigger an error. With this patch, the
behavior is correct. The error (without this patch) could look like
this:
fatal error: undefined instruction with sp=0x200007cc pc=nil
Moving settings to a separate config struct has two benefits:
- It decouples the compiler a bit from other packages, most
importantly the compileopts package. Decoupling is generally a good
thing.
- Perhaps more importantly, it precisely specifies which settings are
used while compiling and affect the resulting LLVM module. This will
be necessary for caching the LLVM module.
While it would have been possible to cache without this refactor, it
would have been very easy to miss a setting and thus let the
compiler work with invalid/stale data.
This fixes a longstanding TODO comment and similar to
https://github.com/tinygo-org/tinygo/pull/1593 it removes some code out
of the compiler.CompileProgram function that doesn't need to be there.
This is a small refactor to move code away from compiler.CompilePackage,
with the goal that compiler.CompilePackage will eventually be removed
entirely in favor of compiler.CompilePackage.
Since https://github.com/tinygo-org/tinygo/pull/1571 (in particular, the first
commit that sets the main package path), the main package is always named
"main". This makes the callMain() workaround in the runtime unnecessary and
allows directly calling the main.main function with a //go:linkname pragma.
This package was long making the design of the compiler more complicated
than it needs to be. Previously this package implemented several
optimization passes, but those passes have since moved to work directly
with LLVM IR instead of Go SSA. The only remaining pass is the SimpleDCE
pass.
This commit removes the *ir.Function type that permeated the whole
compiler and instead switches to use *ssa.Function directly. The
SimpleDCE pass is kept but is far less tightly coupled to the rest of
the compiler so that it can easily be removed once the switch to
building and caching packages individually happens.
This change extends defer support to all supported builitin functions.
Not all of them make sense (such as len, cap, real, imag, etc) but this
change for example adds support for `defer(delete(m, key))` which is
used in the Go 1.15 encoding/json package.
This works around some UB in LLVM, where an out-of-bounds conversion would produce a poison value.
The selected behavior is saturating, except that NaN is mapped to the minimum value.
This commit finally introduces unit tests for the compiler, to check
whether input Go code is converted to the expected output IR.
To make this necessary, a few refactors were needed. Hopefully these
refactors (to compile a program package by package instead of all at
once) will eventually become standard, so that packages can all be
compiled separate from each other and be cached between compiles.
Before this change, the compiler could panic with the following message:
panic: 20 not an Int
That of course doesn't make much sense. But it apparently is expected
behavior, see https://github.com/golang/go/issues/43165 for details.
This commit fixes this issue by converting the constant to an integer if
needed.
Because the parentHandle parameter wasn't always set to the right value,
the coroutine lowering pass would sometimes panic with "trying to make
exported function async" even though there was no exported function
involved. Therefore, it should unconditionally be set to avoid this.
The parent function doesn't always have the parentHandle function
parameter set because it can only be set after defining a function, not
when it is only declared.
Previously, EmitPointerPack would generate an out-of-bounds read from an
alloca. This commit fixes that by creating an alloca of the appropriate
size instead of using the size of the to-be-packed data (which might be
smaller than a pointer).
I discovered this error while working on a rewrite of the interp
package, which checks for out-of-bounds reads and writes. There I
discovered this issue when the image package was compiled.
Test binaries must be run in the source directory of the package to be
tested. This wasn't done, leading to a few "file not found" errors.
This commit implements this. Unfortunately, it does not allow more
packages to be tested as both affected packages (debug/macho and
debug/plan9obj) will still fail with this patch even though the "file
not found" errors are gone.
There were a few problems with the go/packages package. While it is more
or less designed for our purpose, it didn't work quite well as it didn't
provide access to indirectly imported packages (most importantly the
runtime package). This led to a workaround that sometimes broke
`tinygo test`.
This PR contains a number of related changes:
* It uses `go list` directly to retrieve the list of packages/files to
compile, instead of relying on the go/packages package.
* It replaces our custom TestMain replace code with the standard code
for running tests (generated by `go list`).
* It adds a dummy runtime/pprof package and modifies the testing
package, to get tests to run again with the code generated by
`go list`.
This is a big change that will determine the stack size for many
goroutines automatically. Functions that aren't recursive and don't call
function pointers can in many cases have an automatically determined
worst case stack size. This is useful, as the stack size is usually much
lower than the previous hardcoded default of 1024 bytes: somewhere
around 200-500 bytes is common.
A side effect of this change is that the default stack sizes (including
the stack size for other architectures such as AVR) can now be changed
in the config JSON file, making it tunable per application.
This is required by the coroutines pass, otherwise it will panic. It
checks for the proper parameter names to make sure the function is not
exported. In this case, the runtime.initAll function wasn't exported but
simply didn't have the correct parameter names so the check triggered
even though it shouldn't.
This is necessary to avoid a circular dependency between the device/avr
and runtime/interrupts package in the next commit.
It may be worth replacing existing calls like device/arm.Asm to
device.Asm, to have a single place where these are defined.
Previously, chansend and chanrecv allocated a heap object before blocking on a channel.
This object was used to implement a linked list of goroutines blocked on the channel.
The chansend and chanrecv now instead accept a buffer to store this object in as an argument.
The compiler now creates a stack allocation for this object and passes it in.
There were a few cases left where a named type would cause a crash in
the compiler. While going through enough code would have found them
eventually, I specifically looked for the `Type().(` pattern: a Type()
call that is then used in a type assert. Most of those were indeed bugs,
although for some I couldn't come up with a reproducer so I left them
as-is.
