This is a big change: apart from removing LLVM 14 it also removes typed
pointer support (which was only fully supported in LLVM up to version
14). This removes about 200 lines of code, but more importantly removes
a ton of special cases for LLVM 14.
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.
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.
This removes the parentHandle argument from the internal calling convention.
It was formerly used to implment coroutines.
Now that coroutines have been removed, it is no longer necessary.
This is fake debug info. It doesn't point to a source location because
there is no source location. However, it helps to correctly attribute
code size usage to particular packages.
I've also updated builder/sizes.go with some debugging helpers.
This commit has a few related changes:
* It sets the optsize attribute immediately in the compiler instead of
adding it to each function afterwards in a loop. This seems to me
like the more appropriate way to do it.
* It centralizes setting the optsize attribute in the transform
package, to make later changes easier.
* It sets the optsize in a few more places: to runtime.initAll and to
WebAssembly i64 wrappers.
This commit does not affect the binary size of any of the smoke tests,
so should be risk-free.
This commit simplifies the IR a little bit: instead of calling
pseudo-functions runtime.interfaceImplements and
runtime.interfaceMethod, real declared functions are being called that
are then defined in the interface lowering pass. This should simplify
the interaction between various transformation passes. It also reduces
the number of lines of code, which is generally a good thing.
This commit includes two changes:
* It makes unexported interface methods package-private, so that it's
not possible to type-assert on an unexported method in a different
package.
* It makes the globals used to identify interface methods defined
globals, so that they can (eventually) be left in the program for an
eventual non-LTO build mode.
These two passes are related, but can definitely work independently.
Which is what this change does: it splits the two passes. This should
make it easier to change these two new passes in the future.
This change now also enables slightly better testing by testing these
two passes independently. In particular, the reflect lowering pass got
some actual tests: it was barely unit-tested before.
I have verified that this doesn't really change code size, at least not
on the microbit target. Two tests do change, but in a very minor way
(and in opposite direction).
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 simplifies future changes. While the move itself is very simple, it
required some other changes to a few transforms that create new
functions to add the optsize attribute manually. It also required
abstracting away the optimization level flags (based on the -opt flag)
so that it can easily be retrieved from the config object.
This commit does not impact binary size on baremetal and WebAssembly.
I've seen a few tests on linux/amd64 grow slightly in size, but I'm not
too worried about those.
In rare cases the signature might change as a result of LLVM renaming
some named struct types when multiple LLVM modules are merged. The
easiest workaround is to detect such mismatched signatures and adding a
bitcast: this should be safe as the underlying data is effectively of
the same type.
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.
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.
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.
I ran into an issue where I did a method call on a nil interface and it
resulted in a HardFault. Luckily I quickly realized what was going on so
I could fix it, but I think undefined behavior is definitely the wrong
behavior in this case. This commit therefore changes such calls to cause
a nil panic instead of introducing undefined behavior.
This does have a code size impact. It's relatively minor, much lower
than I expected. When comparing the before and after of the drivers
smoke tests (probably the most representative sample available), I found
that most did not change at all and those that did change, normally not
more than 100 bytes (16 or 32 byte changes are typical).
Right now the pattern is the following:
switch typecode {
case 1:
call method 1
case 2:
call method 2
default:
nil panic
}
I also tried the following (in the hope that it would be easier to
optimize), but it didn't really result in a code size reduction:
switch typecode {
case 1:
call method 1
case 2:
call method 2
case 0:
nil panic
default:
unreachable
}
Some code got smaller, while other code (the majority) got bigger. Maybe
this can be improved once range[1] is finally allowed[2] on function
parameters, but it'll probably take a while before that is implemented.
[1]: https://llvm.org/docs/LangRef.html#range-metadata
[2]: https://github.com/rust-lang/rust/issues/50156
This is a common case, but it also complicates the code. Removing this
special case does have a negative effect on code size in rare cases, but
I don't think it's worth keeping around (and possibly causing bugs) for
such uncommon cases.
This should not result in functional changes, although the output (as
stated above) sometimes changes a little bit.
It appears that LLVM can sometimes recognize that multiple calls to
runtime.interfaceMethod can be merged into one. When that happens, the
interface lowering pass shows an error as it didn't expect that
situation.
Luckily the fix is very easy.
Panics are bad for usability: whenever something breaks, the user is
shown a (not very informative) backtrace. Replace it with real error
messages instead, that even try to display the Go source location.
When using reflect, arbitrary types can be synthesized. This invalidates
a few assumptions in the interface-lowering pass, that think they can
see all types that are in use in a program and optimize accordingly.
The file size impact depends on the specific program. Sometimes it's
nonexistent, sometimes it's rather hefty (up to 30% bigger). Especially
the samd21 targets seem to be affected, with a 2000-6000 bytes increase
in code size. A moderately large case (the stdlib test) increases by
4%/6%/15% depending on the target.
I hope that this increase could be mitigated, but I don't see an obvious
way to do that.
A bitcast was inserted when the receiver of the call wasn't a *i8. This
is a pretty common case, and did not play well with goroutines.
Avoid this bitcast by changing each call to a direct call, after
unpacking the receiver type from the *i8 parameter. This might also fix
some undefined behavior in the resulting program, as it is technically
not allowed to call a function with a different signature (even if the
signature is compatible).
In LLVM 8, the AVR backend has moved all function pointers to address
space 1 by default. Much of the code still assumes function pointers
live in address space 0, leading to assertion failures.
This commit fixes this problem by autodetecting function pointers and
avoiding them in interface pseudo-calls.
This makes sure the most commonly used types have the lowest type codes.
This was intended to be the case, but apparently I forgot to sort them
the right way.
This commit changes many things:
* Most interface-related operations are moved into an optimization
pass for more modularity. IR construction creates pseudo-calls which
are lowered in this pass.
* Type codes are assigned in this interface lowering pass, after DCE.
* Type codes are sorted by usage: types more often used in type
asserts are assigned lower numbers to ease jump table construction
during machine code generation.
* Interface assertions are optimized: they are replaced by constant
false, comparison against a constant, or a typeswitch with only
concrete types in the general case.
* Interface calls are replaced with unreachable, direct calls, or a
concrete type switch with direct calls depending on the number of
implementing types. This hopefully makes some interface patterns
zero-cost.
These changes lead to a ~0.5K reduction in code size on Cortex-M for
testdata/interface.go. It appears that a major cause for this is the
replacement of function pointers with direct calls, which are far more
susceptible to optimization. Also, not having a fixed global array of
function pointers greatly helps dead code elimination.
This change also makes future optimizations easier, like optimizations
on interface value comparisons.
Ayke van Laethem
Damian Gryski
Nia Waldvogel
Jaden Weiss
Jaden Weiss