[CopyPropagation] Canonicalize all values. by nate-chandler · Pull Request #41943 · swiftlang/swift

nate-chandler · 2022-03-21T22:57:50Z

Previously, the lifetimes of only values which were copied were canonicalized during the non-mandatory CopyPropagation pass. The mandatory pass (i.e. MandatoryCopyPropagation) has been disabled, meaning that non-copied values lifetimes were never canonicalized. So enable canonicalization of all lifetimes during regular CopyPropagation.

rdar://54335055

nate-chandler · 2022-03-21T22:58:14Z

@swift-ci please benchmark

nate-chandler · 2022-03-22T00:25:44Z

------- Performance (x86_64): -O -------

REGRESSION                                     OLD     NEW    DELTA    RATIO
PrefixWhileArrayLazy                           20      26     +30.0%   **0.77x (?)**
DifferentiationIdentity                        247     320    +29.6%   **0.77x (?)**
DifferentiationSquare                          444     537    +20.9%   **0.83x (?)**
PrefixWhileAnySeqCntRange                      114     129    +13.2%   **0.88x (?)**
RemoveWhereMoveInts                            17      19     +11.8%   **0.89x (?)**
Data.hash.Empty                                47      52     +10.6%   **0.90x (?)**

IMPROVEMENT                                    OLD     NEW    DELTA    RATIO
SortArrayInClass                               14976   311    -97.9%   **48.15x**
DropWhileArrayLazy                             49      31     -36.7%   **1.58x**
FlattenListLoop                                1379    939    -31.9%   **1.47x (?)**
FlattenListFlatMap                             4234    3004   -29.1%   **1.41x (?)**
NSArray.nonbridged.mutableCopy.objectAtIndex   578     431    -25.4%   **1.34x (?)**
NSArray.nonbridged.objectAtIndex               560     424    -24.3%   **1.32x (?)**
PrefixWhileSequenceLazy                        26      20     -23.1%   **1.30x (?)**
DropWhileAnyCollectionLazy                     49      38     -22.4%   **1.29x**
NSArray.bridged.objectAtIndex                  278     216    -22.3%   **1.29x**
NSArray.bridged.mutableCopy.objectAtIndex      733     615    -16.1%   **1.19x (?)**
DropWhileAnySeqCRangeIterLazy                  45      38     -15.6%   **1.18x (?)**
DropWhileAnySeqCntRangeLazy                    45      38     -15.6%   **1.18x (?)**
DictionarySubscriptDefaultMutation             172     148    -14.0%   **1.16x (?)**
DropWhileSequenceLazy                          67      62     -7.5%    **1.08x (?)**
PrefixArray                                    14      13     -7.1%    **1.08x (?)**
Diffing.Similar                                373     347    -7.0%    **1.07x (?)**
PrefixWhileAnySequence                         198     185    -6.6%    **1.07x (?)**

------- Code size: -O -------

IMPROVEMENT                    OLD     NEW     DELTA    RATIO
Differentiation.o              7695    6063    -21.2%   **1.27x**
SortArrayInClass.o             2708    2612    -3.5%    **1.04x**
DictionarySubscriptDefault.o   19090   18770   -1.7%    **1.02x**

------- Performance (x86_64): -Osize -------

REGRESSION                                     OLD     NEW    DELTA    RATIO
DropWhileCountableRangeLazy                    49      71     +44.9%   **0.69x (?)**
DifferentiationIdentity                        253     332    +31.2%   **0.76x (?)**
DropWhileAnySeqCRangeIterLazy                  127     166    +30.7%   **0.77x (?)**
DropWhileAnySeqCntRangeLazy                    127     166    +30.7%   **0.77x (?)**
PrefixWhileAnyCollection                       111     144    +29.7%   **0.77x (?)**
Set.filter.Int100.28k                          37      47     +27.0%   **0.79x (?)**
SequenceAlgosRange                             1740    2170   +24.7%   **0.80x (?)**
Set.filter.Int100.16k                          21      26     +23.8%   **0.81x (?)**
Set.filter.Int100.20k                          26      32     +23.1%   **0.81x (?)**
Set.filter.Int100.24k                          31      38     +22.6%   **0.82x (?)**
DifferentiationSquare                          443     541    +22.1%   **0.82x**
SequenceAlgosArray                             1860    2180   +17.2%   **0.85x (?)**
ObjectiveCBridgeStubDateAccess                 130     152    +16.9%   **0.86x (?)**
PrefixWhileAnyCollectionLazy                   94      107    +13.8%   **0.88x (?)**
PrefixWhileAnySequence                         188     213    +13.3%   **0.88x (?)**
RemoveWhereMoveInts                            19      21     +10.5%   **0.90x (?)**
RemoveWhereQuadraticString                     186     201    +8.1%    **0.93x (?)**
LineSink.scalars.complex                       137     148    +8.0%    **0.93x (?)**

IMPROVEMENT                                    OLD     NEW    DELTA    RATIO
SortArrayInClass                               15158   403    -97.3%   **37.61x**
PrefixWhileCountableRangeLazy                  40      21     -47.5%   **1.90x (?)**
PrefixWhileSequenceLazy                        40      26     -35.0%   **1.54x**
DropWhileSequenceLazy                          71      49     -31.0%   **1.45x**
NSArray.nonbridged.objectAtIndex               568     424    -25.4%   **1.34x (?)**
NSArray.nonbridged.mutableCopy.objectAtIndex   577     431    -25.3%   **1.34x (?)**
RecursiveOwnedParameter                        327     251    -23.2%   **1.30x (?)**
PrefixWhileArrayLazy                           26      20     -23.1%   **1.30x (?)**
NSArray.bridged.objectAtIndex                  278     216    -22.3%   **1.29x**
NSStringConversion.UTF8                        718     592    -17.5%   **1.21x (?)**
NSArray.bridged.mutableCopy.objectAtIndex      733     619    -15.6%   **1.18x (?)**
NSStringConversion.Medium                      354     299    -15.5%   **1.18x (?)**
NSStringConversion.InlineBuffer.ASCII          293     257    -12.3%   **1.14x (?)**
SequenceAlgosContiguousArray                   1920    1710   -10.9%   **1.12x (?)**
Set.isSuperset.Seq.Empty.Int                   45      41     -8.9%    **1.10x (?)**
Set.isSubset.Int.Empty                         45      41     -8.9%    **1.10x (?)**
NSStringConversion.Long                        708     649    -8.3%    **1.09x (?)**
StringWalk                                     3520    3240   -8.0%    **1.09x (?)**
PrefixWhileAnySequenceLazy                     1179    1086   -7.9%    **1.09x (?)**
PrefixAnySequenceLazy                          1192    1112   -6.7%    **1.07x (?)**

------- Code size: -Osize -------

REGRESSION           OLD     NEW     DELTA    RATIO
Hash.o               18682   18997   +1.7%    **0.98x**

IMPROVEMENT          OLD     NEW     DELTA    RATIO
Differentiation.o    7225    5881    -18.6%   **1.23x**
SortArrayInClass.o   2956    2876    -2.7%    **1.03x**

------- Performance (x86_64): -Onone -------

REGRESSION                                        OLD     NEW    DELTA    RATIO
NSStringConversion.InlineBuffer.ASCII             5890    6705   +13.8%   **0.88x (?)**
NSStringConversion.InlineBuffer.UTF8              3606    4075   +13.0%   **0.88x (?)**

IMPROVEMENT                                       OLD     NEW    DELTA    RATIO
CharIteration_punctuatedJapanese_unicodeScalars   10360   9680   -6.6%    **1.07x (?)**

------- Code size: -swiftlibs -------

atrick · 2022-03-22T15:02:27Z

We're really starting to see the benefits of ownership representation now
SortArrayInClass 14976 311 -97.9% **48.15x**

atrick

Good. This means the lifetime of owned values will be consistently canonicalized, within the new boundaries of lexical lifetimes.

nate-chandler · 2022-03-23T03:48:26Z

@swift-ci please benchmark

nate-chandler · 2022-03-23T03:48:32Z

@swift-ci please test

nate-chandler · 2022-03-23T14:48:11Z

@swift-ci please benchmark

nate-chandler · 2022-03-23T15:06:29Z

@swift-ci please test macos platform

nate-chandler · 2022-03-23T15:06:37Z

@swift-ci please test linux platform

nate-chandler · 2022-03-23T15:47:33Z

@swift-ci please test linux platform

nate-chandler · 2022-03-23T16:11:03Z

@swift-ci please clean test macos platform

nate-chandler · 2022-03-23T16:11:22Z

@swift-ci please clean test linux platform

atrick · 2022-03-23T16:51:17Z

Thanks for fixing the outliner to handle consuming arguments. It's good to have this pass working well with OSSA!

nate-chandler · 2022-03-23T18:41:51Z

@swift-ci please clean test linux platform

atrick · 2022-04-13T19:32:24Z

@nate-chandler I just realized this never landed. Anything holding it up?

nate-chandler · 2022-04-13T19:33:21Z

There were some docc test failures earlier. Let's see if those were spurious.

nate-chandler · 2022-04-13T19:33:44Z

@swift-ci please test

nate-chandler · 2022-06-01T01:33:02Z

@swift-ci please test

nate-chandler · 2022-06-03T05:27:57Z

@swift-ci please test

The new utility looks through operands whose values are themselves SILPhiArguments and visits those arguments' operands.

nate-chandler · 2022-07-19T03:14:16Z

@swift-ci please test

nate-chandler · 2022-07-20T22:10:15Z

@swift-ci please test

The new utility, given an phi, visits all adjacent phis (i.e. arguments to the same block) which are (potentially iterated) reborrows of a value reaching the given phi.

When canonicalizing lifetime of an owned phi, add adjacent phis which are reborrows of its reaching values to the worklist. Update liveness to consider uses whose operand ownership is specific to guaranteed values (i.e. uses which are derived from these adjacent reborrow phis) as standard, non-lifetime-ending uses. rdar://94346482

Properly handle the different sorts of reborrows when they are visited: - non-branch reborrows (tuples, etc) are non-lifetime-ending uses; such reborrows are relevant defs for liveness - branch reborrows which also consume the owned phi are lifetime-ending uses - branch reborrows which DO NOT consume the owned phi are non-lifetime-ending uses, and the corresponding phis are relevant defs for liveness rdar://97399065

nate-chandler · 2022-07-21T22:52:06Z

@swift-ci please test

Previously, the match failed to find the larger sequence that began with load [copy] and ended with destroy_value because the iterator advanced after finding the load [copy]. Advanced the iterator here. Enables reverting the test outlining test changes introduced in a52b896.

Previously, the lifetimes of only values which were copied were canonicalized during the non-mandatory CopyPropagation pass. The mandatory pass (i.e. MandatoryCopyPropagation) has been disabled, meaning that non-copied values lifetimes were never canonicalized. So enable canonicalization of all lifetimes during regular CopyPropagation. rdar://54335055

Thanks to CopyPropagation canonicalizing all values, not just those for which there is a copy_value instruction, the lifetime of the value that is loaded in the BridgedProperty pattern is shortened and the destroy_value appears right after its use rather than after the full CFG for the bridged property. Updated the BridgedProperty pattern to match that newly hoisted instruction.

Put off calculating the outlined function name until the point at which we will have all the inputs for the name.

Previously, for the BridgedProperty pattern, there were two flavors of outlined function that would be formed, varying on the ownership of the instance whose field is being access: (1) guaranteed_in -- for use by addresses (2) unowned -- for use by values which were not destroyed until some time after the pattern of interest is matched Now that the lifetime of the instance may be shortened to just after the call to the objc method, the latter of these is not appropriate for values. We would have to re-extend the lifetime until after the method returns. Instead, here, we add a new variant (3) owned -- for use by values which were destroyed just after the objc method call Doing so requires new mangling.

nate-chandler · 2022-07-22T15:21:48Z

@swift-ci please test

nate-chandler · 2022-07-22T18:47:39Z

@swift-ci please test source compatibility

nate-chandler requested a review from atrick March 21, 2022 22:57

atrick approved these changes Mar 22, 2022

View reviewed changes

nate-chandler force-pushed the copy_propagation/canonicalize-all branch 5 times, most recently from cd76eb3 to 6b06ce2 Compare March 23, 2022 03:47

nate-chandler force-pushed the copy_propagation/canonicalize-all branch from 6b06ce2 to aaf4750 Compare May 25, 2022 18:46

nate-chandler force-pushed the copy_propagation/canonicalize-all branch from aaf4750 to 899141e Compare June 3, 2022 05:26

[SIL] Add utility to transitively visit operands.

8c44d36

The new utility looks through operands whose values are themselves SILPhiArguments and visits those arguments' operands.

nate-chandler force-pushed the copy_propagation/canonicalize-all branch from 899141e to 6fcd1f6 Compare July 19, 2022 03:14

nate-chandler force-pushed the copy_propagation/canonicalize-all branch from 6fcd1f6 to 228232f Compare July 20, 2022 22:01

nate-chandler added 3 commits July 21, 2022 15:03

[OwnershipUtils] Visit owned phi's reborrows.

fe43531

The new utility, given an phi, visits all adjacent phis (i.e. arguments to the same block) which are (potentially iterated) reborrows of a value reaching the given phi.

nate-chandler force-pushed the copy_propagation/canonicalize-all branch from 228232f to 7feccbd Compare July 21, 2022 22:47

nate-chandler added 6 commits July 22, 2022 08:17

[NFC] [Outliner] Removed unused fields.

b527649

[Outliner] Defer bridged property mangling.

9b290cd

Put off calculating the outlined function name until the point at which we will have all the inputs for the name.

nate-chandler force-pushed the copy_propagation/canonicalize-all branch from 7feccbd to a903805 Compare July 22, 2022 15:21

nate-chandler merged commit dcea147 into swiftlang:main Jul 22, 2022

nate-chandler deleted the copy_propagation/canonicalize-all branch July 22, 2022 23:54

meg-gupta modified the milestone: Swift 5.7 Sep 14, 2022

Conversation

nate-chandler commented Mar 21, 2022

Uh oh!

nate-chandler commented Mar 21, 2022

Uh oh!

nate-chandler commented Mar 22, 2022

Uh oh!

atrick commented Mar 22, 2022

Uh oh!

atrick left a comment

Choose a reason for hiding this comment

Uh oh!

nate-chandler commented Mar 23, 2022

Uh oh!

nate-chandler commented Mar 23, 2022

Uh oh!

nate-chandler commented Mar 23, 2022

Uh oh!

nate-chandler commented Mar 23, 2022

Uh oh!

nate-chandler commented Mar 23, 2022

Uh oh!

nate-chandler commented Mar 23, 2022

Uh oh!

nate-chandler commented Mar 23, 2022

Uh oh!

nate-chandler commented Mar 23, 2022

Uh oh!

atrick commented Mar 23, 2022

Uh oh!

nate-chandler commented Mar 23, 2022

Uh oh!

atrick commented Apr 13, 2022

Uh oh!

nate-chandler commented Apr 13, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

nate-chandler commented Apr 13, 2022

Uh oh!

nate-chandler commented Jun 1, 2022

Uh oh!

nate-chandler commented Jun 3, 2022

Uh oh!

nate-chandler commented Jul 19, 2022

Uh oh!

nate-chandler commented Jul 20, 2022

Uh oh!

nate-chandler commented Jul 21, 2022

Uh oh!

nate-chandler commented Jul 22, 2022

Uh oh!

nate-chandler commented Jul 22, 2022

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

nate-chandler commented Apr 13, 2022 •

edited

Loading