| Commit message (Collapse) | Author | Age | Files | Lines |
|---|
| |
|
|
|
|
|
|
| |
Native pascal strings shall be just treated as normal Java strings on the Java side.
Hence drop the length parameter across generated API, i.e.
- C Function bindings
- Java Callbacks
|
| |
|
|
| |
compound attribute
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
GlueGen Revised Struct Mapping (esp pointer to array or single element), Struct String Charset, .. and Documentation
- Documentation:
- Added README.md
Let's have a proper face for the git repo
- Added doc/GlueGen_Mapping.md (and its html conversion doc/GlueGen_Mapping.html)
Created a new document covering application and implementation details suitable for users/devs.
- Added doc/JogAmpMacOSVersions.md conversion to doc/JogAmpMacOSVersions.html
- Updated www/index.html
- Use *CodeUnit instead of PrintWriter, representing a Java or C code unit covering a set of functions and structs.
The CCodeUnit also handles common code shared by its unit across functions etc.
- Dropping 'static initializer', as its no more required
due to simplified `JVMUtil_NewDirectByteBufferCopy()` variant.
- Revised Struct Mapping:
- Pure Java implementation to map primitive and struct fields within a struct
by utilizing ElementBuffer.
Only 'Function Pointer' fields within a struct require native code.
Exposes `static boolean usesNativeCode()` to query whether native code is used/required.
- Transparent native memory address API
Expose `long getDirectBufferAddress()` and `static TK_Struct derefPointer(long addr)`,
allowing to
- pass the native struct-pointer with native code
- reconstruct the struct from a native struct-pointer
- have a fully functional `TK_Struct.derefPointer(struct.getDirectBufferAddress())` cycle.
- Add 'boolean is<Val>Null() to query whether a pointer (array) is NULL
- *Changed* array get/set method for more flexibility alike `System.arraycopy(src, srcPos, dest, destPos, len)`,
where 'src' is being dropped for the getter and 'dest' is being dropped for the setter
as both objects are reflected by the struct instance.
- *Changed* `get<Val>ArrayLength()` -> `get<Val>ElemCount()` for clarity
- Considering all ConstElemCount values with config 'ReturnedArrayLength <int>'
to be owned by native code -> NativeOwnership -> Not changing the underlying memory region!
JavaOwnership is considered for all pointer-arrays not of NativeOwnership.
Hence any setter on a NativeOwnership pointer-array will fail with non-matching elem-count.
- Add 'release<Val>()' for JavaOwnership pointer-arrays,
allowing to release the Java owned native memory incl. null-ing pointer and set<Val>ElemCount(0).
- Support setter for 'const <type>*' w/ JavaOwnership, i.e. pointer to const value of a primitive or struct,
setter and getter using pointer to array or single element in general.
- Added Config `ImmutableAccess symbol` to disable all setter for whole struct or a field
- Added Config `MaxOneElement symbol` to restrict a pointer to maximum one element and unset
initial value (zero elements)
- Added Config `ReturnsStringOnly symbol` to restrict mapping only to a Java String,
dropping the ByteBuffer variant for 'char'
- String mapping default is UTF-8 and can be read and set via [get|set]Charset(..) per class.
- Dynamic string length retrieval in case no `ReturnedArrayLength` has been configured
has changed from `strlen()` to `strnlen(aptr, max_len)` to be on the safe site.
The maximum length default is 8192 bytes and can be read and set via [get|set]MaxStrnlen(..) per class.
FIXME: strnlen(..) using EOS byte non-functional for non 8-bit codecs like UTF-8, US-ASCII.
This is due to e.g. UTF-16 doesn't use an EOS byte, but interprets it as part of a code point.
- TODO: Perhaps a few more unit tests
- TODO: Allow plain 'int' to be mapped in structs IFF their size is same for all MachineDescriptions used.
Currently this is the case -> 4 bytes like int32_t.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
renamed original
'DelegateImplementation' is a variation of 'ManuallyImplement'.
'ManuallyImplement' emits the interface method, but suppresses
the Java and native-code implementation.
The latter shall be implemented manually by the user.
'DelegateImplementation' emits the interface method,
and the _private_ renamed Java and native-code implementation.
Both can be called from the manual user implementation,
hence delegation.
Configuration:
DelegateImplementation <ORIG-SYMBOL> <RENAMED-IMPL-SYMBOL>
I.e. delegation model shall apply to <ORIG-SYMBOL>
and the Java and native-code implementation renamed to <RENAMED-IMPL-SYMBOL>.
The user manual implementation of <ORIG-SYMBOL>
may delegate to <RENAMED-IMPL-SYMBOL>.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
- Code regarding 'aliased typedef struct' is tagged in JavaEmitter and HeaderParser:
'NOTE: Struct Name Resolution (JavaEmitter, HeaderParser)'
Prefers containing cstruct typedef pointer
if available _and_ if cstruct is _not_ a typedef!
- Removed: 'HeaderParser.resolveAnonCompound(..)' no more required,
since CompoundType always sets its name!
Commit cf9f28cf249393f42d7d2835775521dfadee6b92
- JavaEmitter.emitStruct:
- Regard above 'aliased typedef struct' NOTE
- JavaEmitter.typeToJavaType:
- Regard above 'aliased typedef struct' NOTE
- ReferencedStructs
- Drop duplicate CompoundType instances of same name.
This can happen due to const/volatile and ASTLocusTag variants.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
- We shall be able to import 'most' vanilla GL header,
i.e. only change the typedef part using our GlueGen types
- Type Safety:
- GlueGen now detects '#define' and 'enum' redefines
and throw an exception in this case.
This helps detecting wrongly renamed GL extensions into core!
- GlueGen now detects function redefines (overloading)
and throw an exception in this case.
Hence the semantics of duplicate functions has to be equal!
This helps detecting wrongly renamed GL extensions into core!
- Semantic equality for all types is provided
via interface TypeComparator.SemanticEqualityOp, i.e. 'boolean equalSemantics(..)'
implemented by com.jogamp.gluegen.cgram.types.Type.
Semantic equality can be relaxed via config "RelaxedEqualSemanticsTest true",
i.e. ignoring integer size, and const / volatile qualifiers.
- All equality/hash methods of 'com.jogamp.gluegen.cgram.types.*'
are restructured.
- Track and simplify renamed 'symbol', i.e. use a common
sub-interface for all renamed symbols (ConstantDefinition, FunctionSymbol, ..)
- This is provided in a unified manner
via interface com.jogamp.gluegen.cgram.types.AliasedSymbol
and its common implementation AliasedSymbolImpl
- All JavaConfiguration.shouldIgnore* methods operate w/
'AliasedSymbol' trying to match all aliases.
- Support 'struct NAME [ { ... } ]' w/o typedef's
- New GL / CL headers do not use typedef's for anonymous opaque types
- Opaque Type handling
- JavaConfiguration.typeInfo(..), identifying opaque types,
no more back references from target-type -> typedef.
Hence the following is possible now:
typedef void * Opaque01; // Opaque
typedef void * APointerBuffer; // A Buffer
- All Logger instances are no more static
and derive their warning level from the package's root Logger
via Logging.getLogger(..).
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
and string (code generation)
Enhance compound access as delivered by Bug 1022,
to also generate accessors (getter and setter) for
array, pointer and string types.
Allow configuration of array length either via
their internal size (c-header) or config 'ReturnedArrayLength'.
'ReturnedArrayLength' allows specifying a java expression.
Canonical field names of compounds are _now_ specified as
follows for configuration entries:
COMPOUND.FIELD
e.g.
StructA.fieldB
Also allow configuration of pointer fields to be treated as
referenced arrays via 'ReturnedArrayLength'.
Further, allow specifying 'pointer fields' as String values
via 'ReturnsString' configuration.
++++
Implementation details:
- handle above described accessor features
- enhance JavaDoc for generated accessors
- generate native JNI compound and string accessor on demand
- encapsule accessor code generation in their own methods
- enhance exception messages
- enhance type verbosity in debug mode
- verbose debug output via GlueGen.debug()
Tests:
- Features covered by test1.[ch]
and Test1p1JavaEmitter and Test1p2ProcAddressEmitter
- Validated compilation and unit tests for modules:
- joal
- jogl (minor config changes req.)
- jocl (minor config changes req.)
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
C-Enum values for array length
- Parser (HeaderParser.g): Support using C-Enum values for array length specification
- Will throw an exception if enum identifier is unknown or exceeds int-size
- Add StructEmitter supports for compound-arrays
- Add Debug stderr verbose info:
- Struct Emitter prefix 'SE.' - to analyze emitting struct fields (offset+size and accessors)
- Struct Layout prefix 'SL.' - to analyze memory layout (based on MachineDescription.StaticConfig.X86_64_UNIX)
Tested via test1.[ch] BaseClass ..
|
| |
|
|
| |
call-by-value' JavaMethodBindingEmitter, array may still require NIO handling, also consider NIO array length.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Completing commit c3054a01990e55ab35756ea23ab7d7c05f24dd37
by allowing passing compound arrays via 'call-by-value.
- Creating linear temp heap, copying NIO values into it
and passing to C function.
Copy-back if not 'const', see below.
- Respect 'const' qualifier to skip write-back of
temp heap passed to C function
- See tag: // FIXME: Compound and Compound-Arrays
for code changes and validation of completeness
- triggers for compound arrays are:
- javaType.isArrayOfCompoundTypeWrappers()
- type.isArray()
- simplified const query by c-type: FunctionEmitter.isBaseTypeConst(ctype)
+++
Tests: Added call-by-value to test1.[ch] binding test!
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
returning struct instance
'compound call-by-value' is not efficient.
However, to allow mapping APIs utilizing passing small structs
as arguments and return values - this feature has been added.
+++
To return the struct value, native code needs to
allocate a NIO ByteBuffer and copy the data.
The stack return value can be dismissed afterwards and the
NIO buffer is returned.
We include this functionality for all generated [impl] classes,
native method:
'static jobject JVMUtil_NewDirectByteBufferCopy(JNIEnv *env, void * source_address, jlong capacity)'
(See: 'JavaEmitter.initClassAccessCode')
Since this code requires knowledge of java classes and methods,
for which a reference needs to be acquired, a static initialization method
has been introduced for all generated [impl] classes:
'private static native boolean initializeImpl();'
Per default the this method will be called in the new
static initializer block of the class,
which can be supressed via the configuration element:
'ManualStaticInit <class-name>'
'ManualStaticInit' can be used to issue the 'initializeImpl()'
call in a custom static initializer written by the user.
However, at the time 'initializeImpl()' gets called
the JNI native library must have been loaded, of course!
+++
- See tag: // FIXME: Compound call-by-value
for code changes and validation of completeness
Trigger for compond call-by-value in CMethodBindingEmitter is:
!cArgType.isPointer() && javaArgType.isCompoundTypeWrapper()
Trigger for compond call-by-value in JavaEmitter is:
t.isCompound()
+++
Further more we do tolerate 'javaType.isCPrimitivePointerType()',
i.e. adding comments for offset/size and field entries,
which are all NOP.
This allows to utilize the remaining fields of the native structure.
+++
Tests: Added call-by-value to test1.[ch] binding test!
|
| | |
|
| |
|
|
|
|
|
|
| |
Fix 1: Only emit "int * _offsetHandle = NULL" if it will be used, to
avoid unused variable warning. Fix 2: Add "unsigned" to typecasts in C
function calls when needed to avoid implicit typecast warning. This
commit also adds a unit test for a method that uses an "unsigned char
**" parameter, to mimic the JOCL clCreateProgramWithBinary() function
that caused the typecast warnings.
|
| |
|
|
|
|
|
|
|
|
|
|
| |
test array offset working correct.
The 'carray' pointer returned from GetPrimitiveArrayCritical(..) is moved about the array offset
and used in ReleasePrimitiveArrayCritical(..) to release the pinpointed memory.
Even though this 'is' a bug by violating the _sparse_ specification, Hotspot impl. doesn't use the value at all (NOP)
and hence this code didn't produce an error since .. (Same w/ Dalvik).
A followup commit will fix this issue.
|
| |
|
|
|
|
| |
PointerBuffer, since referenced memory-size is arch dependent
Added extensive PointerBuffer unit tests w/ new mapping in generated test class.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
- Fix 'NIOOnly' impl
- Use 'final' in gen. Java stubs ;
- No null check for Buffers.isDirect(arrayArg) [required]
- Clarify Buffer arg API doc (NIOOnly or NIODirectOnly)
- cleaned up loop / branch (CMethodBindingEmitter)
- remove unsued vars/code
- Tests:
- covers normal/NIOOnly/NIODirectOnly
- covers passing null for array and NIO
|
| |
|
|
|
|
|
|
|
| |
conversion, ie. 32/64 bits)
An opaque type still needs to be converted to the right size (32/64 bit).
In case of a conversion, respect the pointer type.
This fixes bug 536 .
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
const, ie no write-back
We shall consider the C header declaration as being correct
and no modification shall happen on const arrays.
Tested w/ unit tests and JOGL
+++
Cleanup JavaType: final immutable fields, proper CVoidPointer name
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
- Fin
MachineDesction == MD
MD.StaticConfig:
- enum for all supported static configs (ID -> MD)
- verified at runtime: test runtime queried-MD versus static-MD,
hard fail if not compatible (size/alignment)
SizeThunk primitive sizes:
- Add notion of fixed native size (eg. int64_t) and otherwise (eg. long)
java struct 'wrappers' code generation:
- single class using size/offset arrays of all MachineDescription configurations
- at runtime the array idx is queried in static block
- type aligment for not fixed-native-size types (SizeThunk, undef long/int) via StructAccessor
junit test:
- add float test
- fix native code
- add java (create, write) -> native (verify) test
works (tested) on: linux 32/64 and windows 32/64
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
com.jogamp.gluegen.cgram.types -> com.jogamp.gluegen.runtime.types
This is required for desired runtime memory layout.
- Split CompoundType to StructType + UnionType
- StructLayout:
- Utilizing SizeThunk alignment
- Alignment
1) Natural type alignment
2) Add Size
3) Trailing padding w/ largest element alignment
- Only perform memory layout once for type.
Status:
- Unit test passes w/ static MachineDescriptor64Bit
- FIXME static 32bit is faulty, uses 64bit size/alignment
- TODO runtime struct layout to please all platforms w/o worrying
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
- Preparation.
Currently GlueGen fails for type long (size) and some alignments (see package.html).
- The size and alignment values shall be queried at runtime.
- Compound alignment needs to follow the described natural alignment (also @runtime).
-
- Build
- add Linux Arm7 (EABI)
- junit test
- added compound/struct tests, pointing out the shortcomings of current impl.
- package.html
- Added alignment documentation
- remove intptr.cfg
- add GluGen types int8_t, int16_t, uint8_t, uint16_t
- move MachineDescription* into runtime
- Platform
- has runtime MachineDescription
- moved size, .. to MachineDescription
- use enums for OSType, CPUArch and CPUType defined by os.name/os.arch,
triggering exception if os/arch is not supported.
This avoids Java String comparison and conscious os/arch detection.
- MachineDescription:
- compile time instances MachineDescription32Bits, MachineDescription64Bits
- runtime queried instance MachineDescriptionRuntime
- correct size, alignment, page size, ..
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
alloc referenced data map if used only)
This patch doesn't impact GlueGen's code generation,
but enhance and fix PointerBuffer usage only.
remove explicit backup array
As suggested by Michael Bien with a proposed patch,
PointerBuffer's backup array is not only redundant in case it's not used,
but also erroneous - due to possible sliced buffers.
Removes the explicit backup array implementation
leaving it up to the user, ie how PointerBuffer is created (alloc/allocDirect)
and use the underlying nio's buffer backup array, if available.
This also fixes the (never tested) case of indirect w/ backup array usage
on 32bit platform size. In this case the array shall be of type int[],
holding 32bit pointer - on 64bit long[].
Previous to this patch, it was always long[].
Added more thorough tests of PointerBuffer, notably indirect w/ backup array
and native deep copy and filling of a pointer array.
alloc referenced data map if used only
As suggested by Michael Bien with a proposed patch,
the allocation of the dataMap hash map is redundant in case it's not used.
The hash map will be initialized lazy, if needed only.
|
| |
|
|
|
|
|
| |
'toString()' ; Added type tests.
For size_t and uintptr_t this actually generated invalid C code.
Added type tests for those.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Fixed junit test: test1
- Create seperate native libraries to reflect a real world example:
test1 - the library to bind to (no more declaring __stdcall @nn functions)
BindingTest1p1 - the dynamic fixed linkage binding test1p1,
references dynamic library test1 at linktime.
BindingTest1p2 - the dynamic runtime linkage binding test1p2,
loads dynamic library test1 at runtime.
Generic:
- gluegen-cpptasks-base.xml
- target 'gluegen.cpptasks.detect.os'
Set new property 'system.env.library.path'
DYLD_LIBRARY_PATH (macosx)
LD_LIBRARY_PATH (unix)
PATH (windows)
- target 'gluegen.cpptasks.striplibs'
Strips the symbols out of the native libraries
in case c.compiler.debug is false.
Maybe configured with the properties:
c.strip.tool, c.strip.args
- Using system.env.library.path in junit call
to find the test1 library in case of runtime linkage and lookup (test1p2).
- Use gluegen.cpptasks.striplibs for all native libs ..
- Added macosx32 in analogy to macosx64, both defaults to true now
- com.jogamp.common.os.WindowsDynamicLinkerImpl:lookupSymbol()
- Added lookup for __stdcall @nn (stepping alignment 4, max-args: 12)
in case no undecorated __cdecl symbol is found.
Fixed Windows platform:
- Use proper path.seperator on Windows.
- test1.dll needs proper soname inside for fixed linkage (test1p1)
hence the output name must be test1.dll, not libtest1.so
+++
http://jogamp.org/bugzilla/show_bug.cgi?id=394
Fix MacOsX platform:
The commit of cpptasks.jar, git hash 129e783741d91e9ee5cd7da5d5c962c32ec96b0b,
broke the universal binary build on MacOSX.
The above change used cpptasks-1.05b with a few patches in regards to crosscompilation,
but missed one, which accepts the '-arch' argument for GccLinker undecorated.
The new cpptasks.jar is vanilla 1.05b + cpptasks-1.0b5-darwin-patch.diff,
the latter a more refined one.
This version accepts the '-arch' argument undecorated on the darwin platform.
+++
|
| |
|
