I also had to fix up one test and eliminated a test that did not really make
sense in the first place to test with ossa. (Specifically,
copy_value_destroy_value is a test I added when adding copy_value,
destroy_value, it really doesn't make sense since we will just lower it to
retain, release before we hit IRGen).
The XXOptUtils.h convention is already established and parallels
the SIL/XXUtils convention.
New:
- InstOptUtils.h
- CFGOptUtils.h
- BasicBlockOptUtils.h
- ValueLifetime.h
Removed:
- Local.h
- Two conflicting CFG.h files
This reorganization is helpful before I introduce more
utilities for block cloning similar to SinkAddressProjections.
Move the control flow utilies out of Local.h, which was an
unreadable, unprincipled mess. Rename it to InstOptUtils.h, and
confine it to small APIs for working with individual instructions.
These are the optimizer's additions to /SIL/InstUtils.h.
Rename CFG.h to CFGOptUtils.h and remove the one in /Analysis. Now
there is only SIL/CFG.h, resolving the naming conflict within the
swift project (this has always been a problem for source tools). Limit
this header to low-level APIs for working with branches and CFG edges.
Add BasicBlockOptUtils.h for block level transforms (it makes me sad
that I can't use BBOptUtils.h, but SIL already has
BasicBlockUtils.h). These are larger APIs for cloning or removing
whole blocks.
This flag, currently staged in as `-experimental-skip-non-inlinable-function-bodies`, will cause the typechecker to skip typechecking bodies of functions that will not be serialized in the resulting `.swiftmodule`. This patch also includes a SIL verifier that ensures that we don’t accidentally include a body that we should have skipped.
There is still some work left to make sure the emitted .swiftmodule is exactly the same as what’s emitted without the flag, which is what’s causing the benchmark noise above. I’ll be committing follow-up patches to address those, but for now I’m going to land the implementation behind a flag.
DestroyHoisting moves destroys of memory locations up the control flow as far as possible.
Beside destroy_addr, also "store [assign]" is considered a destroy, because is is equivalent to an destroy_addr + a "store [init]".
The main purpose of this optimization is to minimize copy-on-write operations for arrays, etc. Especially if such COW containers are used as enum payloads and modified in-place. E.g.
switch e {
case .A(var arr):
arr.append(x)
self = .A(arr)
...
In such a case DestroyHoisting can move the destroy of the self-assignment up before the switch and thus let the array buffer only be single-referenced at the time of the append.
When we have ownership SIL throughout the pass pipeline this optimization will replace the current destroy hoisting optimization in CopyForwarding.
For now, this optimization only runs in the mandatory pipeline (but not for -Onone) where we already have ownership SIL.
SR-10605
rdar://problem/50463362
the builtin.globalStringTablePointer to the new OSLog overlay.
Modify the new OSLog implementation to use this SPI instead of
`withCString` to pass the (compiler-generated) format string to
the C os_log_impl ABI.
Move the OSLogOptimization pass before constant propagation in
the pass pipeline so that the SPI and the builtin it uses can be
folded to a string_literal instruction.
Update OSLogTests to work with the changes in the implementation.
I am going to use this to ensure some end-to-end tests that do not inline from
the stdlib will work after flipping the switch and stripping ownership after
serialization.
The new pass is based on existing asserts in DiagnoseStaticExclusivity.
They were compiled out in release builds and only checked for captures of
inout parameters. This patch converts the assertions into diagnostics and
adds checks for captures of non-escaping function values.
Unlike the Sema-based checks that this replaces, the new code handles
transitive captures from recursive local functions, which means certain
invalid code that used to compile will now be rejected with an error.
The new analysis also looks at the ultimate usages of a local function
instead of just assuming all local functions are escaping, which fixes
issues where the compiler would reject valid code.
Fixes a bunch of related issues, including:
- <rdar://problem/29403178>
- <https://bugs.swift.org/browse/SR-8546> / <rdar://problem/43355341>
- <https://bugs.swift.org/browse/SR-9043> / <rdar://problem/45511834>
yields in generalized accessors: _read and _modify, which are
yield-once corountines. This pass is based on the existing SIL verifier
checks but diagnoses only those errors that can be introduced by programmers
when using yields.
<rdar://43578476>
Once the flag is flipped, ownership stripping will no longer be done in the
diagnostics pipeline. Instead what will happen is that:
* Onone: At -Onone, we run the entire diagnostics pipeline with ownership
enabled and do not strip ownership until after we serialize in the Onone
"optimization" pass pipeline plan.
* -O: At -O, to temporarily work around serialization issues with transparent
functions, we strip ownership from all but transparent functions at the
beginning of the performance pipeline, serialize, and then strip ownership from
transparent functions. For this to work, I needed to make sure that the
performance pipeline passes that do not support ownership SIL, just skip such
functions. So the transparent functions will arrive (mostly) untouched in
serialized SIL and the rest of the pipeline will optimize non-transparent
functions as they should.
The key thing about the -O change is that it /should/ be performance neutral
since after we serialize we re-run the entire pipeline so we can optimize
semantic functions that we only can inline after we serialize.
This normalizes the creation of pass pipelines by ensuring that all pass
pipelines take a SILOption instead of only some. It also makes it so that we do
not need to propagate options through various pipeline creation helpers.
I discovered while updating PMO for ownership that for ~5 years there has been a
bug where we were treating copy_addr of trivial values like an "Assign" (in PMO
terminology) of a non-trivial value and thus stopping allocation
elimination. When I fixed this I discovered that this caused us to no longer
emit diagnostics in a predictable way. Specifically, consider the following
swift snippet:
var _: UInt = (-1) >> 0
Today, we emit a diagnostic that -1 can not be put into a UInt. This occurs
since even though the underlying allocation is only stored into, the copy_addr
assign keeps it alive, causing the diagnostics pass to see the conversion. With
my fix though, we see that we are only storing into the allocation, causing the
allocation to be eliminated before the constant propagation diagnostic pass
runs, causing the diagnostic to no longer be emitted.
We should truly not be performing this type of DCE before we emit such
diagnostics. So in this commit, I split the pass into two parts:
1. A load promotion pass that performs the SSA formation needed for SSA based
diagnostics to actually work.
2. An allocation elimination passes that run /after/ SSA based diagnostics.
This should be NFC since the constant propagation SSA based diagnostics do not
create memory operations so the output should be the same.
NOTE: This is not in the mandatory passes (which run before this). This will
enable me to strip out ownership after we serialize without touching frontend
code. It also makes Onone and O use the same code paths for serialization
instead of one happening in the driver (Onone today) and the other in a SIL pass
(-O, -Osize).
The reason that I updated the sil-func-extractor test is that I found a bug in
how we emit sib files, namely if you try to emit a sib file to stdout, the
llvm-bcanalyzer flags it as malformed. If I output the .sib into a file rather
than trying to use stdout, everything works.
The idea is that eventually down the line we will split this pass into a
mandatory/performance parts once ownership SSA is eliminated after the
guaranteed passes run. For now though, just run this when optimization is
enabled.
This pass only runs on the stdlib/overlays so far so there shouldn't have been
any slow downs in non-stdlib -Onone builds.
We've been running doxygen with the autobrief option for a couple of
years now. This makes the \brief markers into our comments
redundant. Since they are a visual distraction and we don't want to
encourage more \brief markers in new code either, this patch removes
them all.
Patch produced by
for i in $(git grep -l '\\brief'); do perl -pi -e 's/\\brief //g' $i & done
To be clear this only runs on the stdlibs/overlays since it is gated behind a
flag that is set in our cmake.
I can not move it past closure lifetime fixup since the transformation as
written does not express all of the necessary ownership constraints explicitly
in the IR by using addresses.
General case:
begin_access A
...
strong_release / release_value / destroy
end_access
The release instruction can be sunk below the end_access instruction,
This extends the lifetime of the released value, but, might allow us to
Mark the access scope as no nested conflict.
General case:
—
begin_access A (may or may not have no_nested_conflict)
load/store
end_access
apply // may have a scoped access that conflicts with A
begin_access A [no_nested_conflict]
load/store
end_access A
—
The second access scope does not need to be emitted.
NOTE: KeyPath access must be identified at the top-level, non-inlinable stdlib entry point.
As such, The sodlib entry pointed is annotated by a new @_semantics that is equivalent to inline(never)
Introduce an "early redundant load elimination", which does not optimize loads from arrays.
Later array optimizations, like ABCOpt, get confused if an array load in a loop is converted to a pattern with a phi argument.
This problem was introduced with accessors.
rdar://problem/44184763
Major refactoring + tuning of LICM. Includes:
Support for hosting more array semantic calls
Remove restrictions for sinking instructions
Add support for hoisting and sinking instruction pairs (begin and end accesses)
Testing with Exclusivity enabled on a couple of benchmarks shows:
ReversedArray 7x improvement
StringWalk 2.6x improvement
I am doing this so I can start writing DI tests without this lowering occuring.
There never was a real reason for this code to be in DI beyond convenience. Now
it just makes writing tests more difficult. To prevent any test delta, I changed
all current DI tests to run this pass after DI.
There are ~100 significant benchmark regressions (of ~350) with -O
-enforce-exclusivity=checked.
This optimization roughly cuts the overhead in half for almost all of those
regressions. These are the top 30 improvements with the optimization enabled.
XorShift....................................................2.83x
ReversedArray...............................................2.76x
RangeIterationSigned........................................2.67x
ExclusivityGlobal...........................................2.57x
Random......................................................2.44x
ReversedDictionary..........................................2.41x
GeekbenchGEMM...............................................2.35x
ArrayInClass................................................2.31x
StringWalk..................................................2.29x
Ary.........................................................2.25x
Ary3........................................................2.25x
Ary2........................................................2.21x
MultiFileTogether...........................................2.17x
MultiFileSeparate...........................................2.17x
RecursiveOwnedParameter.....................................2.14x
LevenshteinDistance.........................................2.04x
HashTest....................................................1.97x
Voronoi.....................................................1.94x
NopDeinit...................................................1.92x
Life........................................................1.89x
Richards....................................................1.84x
Rectangles..................................................1.74x
MatMul......................................................1.71x
LinkedList..................................................1.51x
GeekbenchFFT................................................1.47x
Xcbuild_OutputByteStreamPerfTests...........................1.39x
ObjectAllocation............................................1.33x
MapReduceLazyCollection.....................................1.30x
Prims.......................................................1.28x
CharIndexing_tweet_unicodeScalars_Backwards.................1.28x
I am going to be adding logic here to enable apple/swift#1550 to be completed.
The rename makes sense due to precedent from LLVM's codegen prepare and also
since I am going to be expanding what the pass is doing beyond just "cleaning
up". It is really a grab bag pass for performing simple transformations that we
do not want to pollute IRGen's logic with.
https://github.com/apple/swift/pull/15502
rdar://39335800
As a first step to getting mandatory inlining out of the business
of 'linking' (walking the function graph and deserializing all
referenced functions), add a new optimizer pass which links
everything in the mandatory pipeline.
For now this is mostly NFC, except it regresses an optimization
I made recently by linking in bodies of methods of deserialized
vtables eagerly. This will be addressed in upcoming patches.
As a first step to getting mandatory inlining out of the business
of 'linking' (walking the function graph and deserializing all
referenced functions), add a new optimizer pass which links
everything in the mandatory pipeline.
For now this is mostly NFC, except it regresses an optimization
I made recently by linking in bodies of methods of deserialized
vtables eagerly. This will be addressed in upcoming patches.
