This commits extends nullability declarations to the field level, formalizing the interaction between methods and their underlying fields and therefore avoiding any nullability mismatch.
Issue: SPR-15720
Beyond just formally declaring the current behavior, this revision actually enforces non-null behavior in selected signatures now, not tolerating null values anymore when not explicitly documented. It also changes some utility methods with historic null-in/null-out tolerance towards enforced non-null return values, making them a proper citizen in non-null assignments.
Some issues are left as to-do: in particular a thorough revision of spring-test, and a few tests with unclear failures (ignored as "TODO: NULLABLE") to be sorted out in a follow-up commit.
Issue: SPR-15540
This commit introduces 2 new @Nullable and @NonNullApi
annotations that leverage JSR 305 (dormant but available via
Findbugs jsr305 dependency and already used by libraries
like OkHttp) meta-annotations to specify explicitly
null-safety of Spring Framework parameters and return values.
In order to avoid adding too much annotations, the
default is set at package level with @NonNullApi and
@Nullable annotations are added when needed at parameter or
return value level. These annotations are intended to be used
on Spring Framework itself but also by other Spring projects.
@Nullable annotations have been introduced based on Javadoc
and search of patterns like "return null;". It is expected that
nullability of Spring Framework API will be polished with
complementary commits.
In practice, this will make the whole Spring Framework API
null-safe for Kotlin projects (when KT-10942 will be fixed)
since Kotlin will be able to leverage these annotations to
know if a parameter or a return value is nullable or not. But
this is also useful for Java developers as well since IntelliJ
IDEA, for example, also understands these annotations to
generate warnings when unsafe nullable usages are detected.
Issue: SPR-15540
Until this change a single classloader was used to load
all compiled SpEL expressions. This meant in a context where an
expression was repeatedly flipping between compiled and interpreted
mode (which can happen if in MIXED mode compilation and changing
the context around the evaluation) the classloader would continually
load a new compiled version but not orphan the old compiled version.
This eventually uses up all the memory as the number of classes
is ever increasing.
With this change classloaders are used to load 100 compiled
expressions. The 101st will be loaded by a new one. Orphaning the
old classloader means if an expression is ever recompiled there
is more likely to be no anchored references left to the older
compiled form and it can be GC'd. In the MIXED situation above it
should help alleviate the problem of older classes never being
candidates for GC.
Issue: SPR-15460
Without this fix the compiled version of elvis
actual behaved differently to the interpreted version
if the value being queried was an empty string. This
is now fixed. It also now correctly handles the
query value being a primitive and addresses the
findings of SPR-15192 where some type inferencing
logic was trying to be too clever, that code has
been deleted.
Issue: SPR-15192
For SPR-14863 we need to adjust the code generation for OpNE
to use !x.equals(y) rather than x!=y. There are also further
cases in the equalityCheck() code in Operator that were not
being handled in the compilation case (when comparators are
used for example). This latter issue also affects OpEQ.
Rather than add yet more bytecode generation, both OpNE and
OpEQ generateCode() methods have been simplified. The
generated code now delegates to equalityCheck() in Operator
which is exactly what the interpreted case does.
This ensures that the compiled code continues to behave just
like the interpreted case. It ensures changes to the interpreted
case are automatically picked up for the compiled case. It
makes the bytecode generation simpler.
The benefit of compilation of SpEL expressions is to avoid
slow reflective calls - that doesn't apply for a basic
(in)equality test so there is no need to go crazy in bytecode
gen.
Issue: SPR-14863
This commit removes `GuavaCache` and support classes. Caffeine supersedes
the caching support in the Google Guava library with an actively maintained
Java 8+ version in standalone form.
As it is the only Guava feature Spring framework integrates with, this
commit removes effectively any reference to Guava.
Issue: SPR-13797
This commit adds a test runtime dependency on log4j 2 for every project
and migrates all log4j.properties files to log4j2-test.xml files.
Issue: SPR-14431
Prior to this commit the SpEL compiler would generate bad bytecode
if the left hand operand of an instanceof was a primitive or
if the right hand operand was a primitive type reference.
With the fixes primitives on the left hand side are now
correctly boxed and special handling is in place for when the
right hand side is a primitive type reference. Using a primitive
type reference on the right always causes the instanceof
check to return false.
Additionally a guard has been added such that compilation is
not allowed when the right hand side of an expression
is not a type reference. If it is, for example, a variable
reference that evaluates to a type reference then that
cannot be expressed in bytecode so compilation is not performed.
Issue: SPR-14250
Prior to this change SpEL did not have an syntactic
construct enabling easy access to a FactoryBean. With this
change it is now possible to use &foo in an expression when
the factory bean should be returned.
Issue: SPR-9511
Before this change the compilation of a method reference or property/field
access was not properly cleaning up the stack if compilation meant
calling a static method or accessing a static field. In these cases there
is no need for a target object on the stack and it should be removed if
present. For a simple expression it is harmless since the end result of
the expression is the thing on the top of the stack, but for nested
expressions if the inner expression suffered this issue, the outer
expression can find itself operating on the wrong element.
The particular issue covered the case of a static field access but this
fix (and associated tests) cover static method, property and field access.
Issue: SPR-13781
When the plus operator is used between strings in a SpEL
expression and that expression is compiled, it is
important to include a cast if computation of any of
the operands isn't obviously leaving strings on the
stack. Likewise if the stack contents are known to
be strings, a cast should not be included.
Issue: SPR-12426
Before this commit the object that #this would refer to in
nested expressions within projection/selection clauses was always
the root context object. This was incorrect as it should be the
element being projected/selected over. This commit introduces
a scope root context object which is set upon entering a new
scope (like when entering a projection or selection). Any
object. With this change this kind of expression now behaves:
where #this is the element of list1. Unqualified references
are also resolved against this scope root context object.
Issues: SPR-10417, SPR-12035, SPR-13055
This change provides support for map[NEW], map[new], map[T]
and map[t]. Prior to this change the 'new' and 't' had to
be quoted because they were keywords in SpEL for a constructor
reference and type reference respectively.
Issue: SPR-11783
Prior to this change the SpEL compiler would not compile mathematical
expressions where the operands were of differing types (e.g. int
and double). This required the expression writer to do the conversion
in the expression text. For example:
T(Integer).valueOf(someInt).doubleValue()/35d
With this commit the restriction is lifted and it is more like Java
so that you can simply do someInt/35d. The mathematical operators
affected are divide/plus/minus/multiply and modulus. This change
involved removing some guards so that the exitTypeDescriptor (the
trigger for whether compilation is allowed) is set more frequently
and enhancing bytecode generation to perform more sophisticated
conversion/coercion automatically.
Issue: SPR-12789
Without this change when the operands to an == are a mix of a
reference type and a primitive, the failure to box the primitive
would result in a verifyerror when the compiled code is loaded.
Issue: SPR-12557
Without this change the plus operator would fail to
include a CHECKCAST in generated bytecode when it
was compiled in cases where one of the operands
has a runtime type of String but a statically
declared type that was not String (i.e. Object).
Issue: SPR-12426
Before this change isWritable() could return true
for a badly formed expression. That is because the
decision about whether something is writable was made
based on the node type rather than whether the node
represented something that could actually be resolved
to be a real thing. This change ensures a resolution
check is done and isWritable() should only return
true if a subsequent setValue() will succeed.
Issue: SPR-10610
These changes provide more robust handling of function
reference compilation in SpEL expressions. Prior to
this change the isCompilable check was not performing
enough visibility checks on the proposed target
function, causing bytecode to be generated that
would lead to an IllegalAccessError.
The changes also bring the argument handling for
function invocation completely inline with that used
for method invocation allowing some code to be deleted.
Many new tests are also included for function
reference compilation.
Issue: SPR-12359
Juergen Hoeller
Stephane Nicoll
diguage
Sebastien Deleuze
Andy Clement
stonio
Christoph Dreis
Sam Brannen
Johnny Lim
Andy Clement
Eric Bottard
Juergen Hoeller