Avoid null checks when emitting "".== and SomeModule.== #2954

retronym · 2013-09-17T14:14:55Z

Although the same the code would be later optimized by -Yconst-opt,
we can offer the same lean byte code to those compiling without that
option by being more discerning when translating ==.

This helps people using bytecode based code coverage tools
such as jacoco that would emit "branch not covered" warnings
for the impossible null check.

Before:

scala> class C { (null: Any) match { case Nil => case "" => case _ => } }
defined class C

scala> :javap C

public C();
  Code:
   Stack=2, Locals=5, Args_size=1
   0:   aload_0
   1:   invokespecial   #9; //Method java/lang/Object."<init>":()V
   4:   aconst_null
   5:   astore_1
   6:   getstatic   #15; //Field scala/collection/immutable/Nil$.MODULE$:Lscala/collection/immutable/Nil$;
   9:   aload_1
   10:  astore_2
   11:  dup
   12:  ifnonnull   23
   15:  pop
   16:  aload_2
   17:  ifnull  30
   20:  goto    37
   23:  aload_2
   24:  invokevirtual   #19; //Method java/lang/Object.equals:(Ljava/lang/Object;)Z
   27:  ifeq    37
   30:  getstatic   #25; //Field scala/runtime/BoxedUnit.UNIT:Lscala/runtime/BoxedUnit;
   33:  astore_3
   34:  goto    74
   37:  ldc #27; //String
   39:  aload_1
   40:  astore  4
   42:  dup
   43:  ifnonnull   55
   46:  pop
   47:  aload   4
   49:  ifnull  63
   52:  goto    70
   55:  aload   4
   57:  invokevirtual   #19; //Method java/lang/Object.equals:(Ljava/lang/Object;)Z
   60:  ifeq    70
   63:  getstatic   #25; //Field scala/runtime/BoxedUnit.UNIT:Lscala/runtime/BoxedUnit;
   66:  astore_3
   67:  goto    74
   70:  getstatic   #25; //Field scala/runtime/BoxedUnit.UNIT:Lscala/runtime/BoxedUnit;
   73:  astore_3
   74:  return
  LocalVariableTable:
   Start  Length  Slot  Name   Signature

After:

public C();
  Code:
   Stack=2, Locals=3, Args_size=1
   0:   aload_0
   1:   invokespecial   #9; //Method java/lang/Object."<init>":()V
   4:   aconst_null
   5:   astore_1
   6:   getstatic   #15; //Field scala/collection/immutable/Nil$.MODULE$:Lscala/collection/immutable/Nil$;
   9:   aload_1
   10:  invokevirtual   #19; //Method java/lang/Object.equals:(Ljava/lang/Object;)Z
   13:  ifeq    23
   16:  getstatic   #25; //Field scala/runtime/BoxedUnit.UNIT:Lscala/runtime/BoxedUnit;
   19:  astore_2
   20:  goto    43
   23:  ldc #27; //String
   25:  aload_1
   26:  invokevirtual   #19; //Method java/lang/Object.equals:(Ljava/lang/Object;)Z
   29:  ifeq    39
   32:  getstatic   #25; //Field scala/runtime/BoxedUnit.UNIT:Lscala/runtime/BoxedUnit;
   35:  astore_2
   36:  goto    43
   39:  getstatic   #25; //Field scala/runtime/BoxedUnit.UNIT:Lscala/runtime/BoxedUnit;
   42:  astore_2
   43:  return

Review by @gkossakowski or @JamesIry, /cc @dragos

Will be somewhat leaner after the next commit.

Although the same the code would be later optimized by -Yconst-opt, we can offer the same lean byte code to those compiling without that option by being more discerning when translating ==. This helps people using bytecode based code coverage tools such as jacoco that would emit "branch not covered" warnings for the impossible null check.

For which we can also elide the null check, and will in the next commit.

retronym · 2013-09-17T14:25:41Z

I'm happy to rework the tests a little, it will be a bit less brittle if we assert the absence of null check instructions.

gkossakowski · 2013-09-17T23:25:23Z

@retronym: rewriting those tests to test the specific property we care about would be very welcome.

Also, please reopen once test failures are addressed.

For the same reasons outlined in the previous commits.

- Directly count null checks, rather than diffing the textual bytecode of the entire method - Add a test to show that the LHS needs to be a direct module reference, not just a tree with a module type, to elide the null check.

retronym · 2013-09-18T15:16:57Z

@gkossakowski The latest push does both of those things :) I left the bytcode-diffing tests in the history as they might be easier to review, the the last commit switches to counting null check instructions.

dragos · 2013-09-20T07:58:31Z

LGTM! I like the tests.

gkossakowski · 2013-09-22T23:30:37Z

LGTM as well!

I'll hold the merge until build bot comes back and can verify this branch.

retronym · 2013-09-23T07:59:02Z

Here are green builds for each commit

2bb960 https://scala-webapps.epfl.ch/jenkins/job/pr-scala/2632/
4faaa8 https://scala-webapps.epfl.ch/jenkins/job/pr-scala/2627/
9fee7b https://scala-webapps.epfl.ch/jenkins/job/pr-scala/2757/
354f42 https://scala-webapps.epfl.ch/jenkins/job/pr-scala/2755/
355eff https://scala-webapps.epfl.ch/jenkins/job/pr-scala/2742/

Avoid null checks when emitting "".== and SomeModule.==

Avoid null checks for "someLiteral".== and SomeModule.==. This has been implemented in GenICode in scala#2954. Introduces a trait to share code between GenICode and GenBCode. This is just a start, more such refactorings will come quite certainly.

lrytz · 2016-02-12T10:54:36Z

The optimization is not correct for modules, https://issues.scala-lang.org/browse/SI-9655

sjrd · 2016-02-12T11:09:05Z

The optimization is not correct for modules, https://issues.scala-lang.org/browse/SI-9655

Do we care? I mean, code where a module load returns null is already broken, right?

FTR Scala.js considers this UB.

lrytz · 2016-02-12T12:15:08Z

well, define broken :) the example prints null, and behaves as specified

class C {
  println(O)
}
object O extends C

i'd love to make the assumption that module loads are always non-null. maybe we can, but it definitely changes program semantics.

retronym added 3 commits September 17, 2013 15:39

SI-7852 Test to show the status quo bytecode for "" == ...

2bb960c

Will be somewhat leaner after the next commit.

SI-7852 Test to show status quo of for SomeModule.==

4faaa82

For which we can also elide the null check, and will in the next commit.

gkossakowski closed this Sep 17, 2013

retronym added 2 commits September 18, 2013 17:00

SI-7852 Omit null check for SomeModule.==

354f428

For the same reasons outlined in the previous commits.

SI-7852 Refactor null-elision tests to be more focussed

355eff4

- Directly count null checks, rather than diffing the textual bytecode of the entire method - Add a test to show that the LHS needs to be a direct module reference, not just a tree with a module type, to elide the null check.

retronym reopened this Sep 18, 2013

ghost assigned gkossakowski Sep 18, 2013

retronym added a commit that referenced this pull request Sep 23, 2013

Merge pull request #2954 from retronym/ticket/7852

469d303

Avoid null checks when emitting "".== and SomeModule.==

retronym merged commit 469d303 into scala:master Sep 23, 2013

lrytz mentioned this pull request May 7, 2014

SI-7852 for GenBCode #3721

Merged

lrytz mentioned this pull request Mar 30, 2016

Clarify / unify delaying of module / class initializers by the compiler / optimizer scala/scala-dev#112

Open

This was referenced Apr 7, 2017

Generate leaner code for "" = expr without -Yconst-opt scala/bug#7852

Closed

Omitting the null check for modules in == is not correct scala/bug#9655

Open

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Avoid null checks when emitting "".== and SomeModule.== #2954

Avoid null checks when emitting "".== and SomeModule.== #2954

Uh oh!

retronym commented Sep 17, 2013

Uh oh!

retronym commented Sep 17, 2013

Uh oh!

gkossakowski commented Sep 17, 2013

Uh oh!

retronym commented Sep 18, 2013

Uh oh!

dragos commented Sep 20, 2013

Uh oh!

gkossakowski commented Sep 22, 2013

Uh oh!

retronym commented Sep 23, 2013

Uh oh!

lrytz commented Feb 12, 2016

Uh oh!

sjrd commented Feb 12, 2016

Uh oh!

lrytz commented Feb 12, 2016

Uh oh!

Uh oh!

Avoid null checks when emitting "".== and SomeModule.== #2954

Avoid null checks when emitting "".== and SomeModule.== #2954

Uh oh!

Conversation

retronym commented Sep 17, 2013

Uh oh!

retronym commented Sep 17, 2013

Uh oh!

gkossakowski commented Sep 17, 2013

Uh oh!

retronym commented Sep 18, 2013

Uh oh!

dragos commented Sep 20, 2013

Uh oh!

gkossakowski commented Sep 22, 2013

Uh oh!

retronym commented Sep 23, 2013

Uh oh!

lrytz commented Feb 12, 2016

Uh oh!

sjrd commented Feb 12, 2016

Uh oh!

lrytz commented Feb 12, 2016

Uh oh!

Uh oh!