mangled name handling ansistring rather than pshortstring based (required
for JVM target; little effect on speed, some extra memory usage)
git-svn-id: branches/jvmbackend@18597 -
o support for JVM arrays in JVM addrnodes and derefnodes (so we
can take the address of var parameters to store them in the
parentfpstruct and later dereference them)
o process loadnode.left also in tjvmloadnode.pass_generate_code() when
handling var-parameters
git-svn-id: branches/jvmbackend@18593 -
o since the JVM target has no stack/framepointer that can be passed
on to nested routines, all local variables and parameters accessed
from nested routines are grouped into a local record whose address
is passed to nested routines. The same technique is also required
for LLVM in the future
git-svn-id: branches/jvmbackend@18588 -
(non-dynamic arrays, records, shortstrings)
- removed the ability to typecast such types directly into related class
types, you have to use the @-operator first now to get a pointer to
the type
o updated the RTL and internal compiler code to properly use this
new convention
o allowed removing several special cases from
tjvmtypeconvnode.target_specific_general_typeconv(), and that
method can probably be removed completely over time
* no longer give compile time errors for pointer-related typecasts that
will fail at run time, because the checking was too complex and could
be worked around via actual pointer typecasts anyway
* removed some unnecessary checkcast operations (for shortstring/
shortstringclass)
git-svn-id: branches/jvmbackend@18574 -
* move the incstack() from a load before the potential "and" to
zero-extend, so that the maximum stack height get calculated
properly
git-svn-id: branches/jvmbackend@18565 -
o support for ansistring constants. It's done via a detour because the
JVM only supports UTF-16 string constants (no array of byte or anything
like that): store every ansicharacter in the lower 8 bits of an
UTF-16 constant string, and at run time copy the characters to an
ansistring. The alternative is to generate code that stores every
character separately to an array.
o the base ansistring support is implemented in a class called
AnsistringClass, and an ansistring is simply an instance of this
class under the hood
o the compiler currently does generate nil pointers as empty
ansistrings unlike for unicodestrings, where we always
explicitly generate an empty string. The reason is that
unicodestrings are the same as JLString and hence common
for Java interoperation, while ansistrings are unlikely to
be used in interaction with external Java code
* fixed indentation
git-svn-id: branches/jvmbackend@18562 -
so it can be intercepted by the JVM backend (it has to create an actual
array)
+ JVM support for the elem_2_open_array hook
git-svn-id: branches/jvmbackend@18561 -
* don't free function result space on stack twice if the result
is not used but a the same time has already been stored in a
temporary funcretnode location
git-svn-id: branches/jvmbackend@18556 -
voidpointerdef to java_jlobject (they're generated by generic code
in nmem; avoids an ifdef for the jvm target there)
git-svn-id: branches/jvmbackend@18554 -
are JVM annotations used by Java's generics support. They cannot be used
for FPC's generics support, but they are useful in other cases
* emit classrefdefs as java.lang.Class, with a signature annotation that
indicates which class they actually refer to
git-svn-id: branches/jvmbackend@18534 -
with java_jlstring in that case, because we have to insert the type
without L prefix and ; suffix for that opcode (which is not done for
objectdefs/recorddefs)
git-svn-id: branches/jvmbackend@18530 -
typeconversion handling
+ support for class reference types in the JVM (although without class virtual
methods, they're not that useful)
git-svn-id: branches/jvmbackend@18516 -
cannot insert typecasting checks on the assignment side, only on the
values that are being assigned
o since valid_for_assignment() is called in tassignmentnode.typecheckpass()
after typecheckpassing left and right, moved the conversion of the
typecheck nodes into JVM-specific constructs from typecheckpass to
pass_1, so that they can use the information whether they are on
the assignment side or not
* forbid casting a child type to a parent type on the assignment side on
managed platforms, because that circumvents the type checking
git-svn-id: branches/jvmbackend@18515 -
o moved several routines from pmodules to ngenutil and overrode them
in njvmutil (for unit initialisation tables, resource strings, ...)
o force the evaluation stack size to at least 1 for the main program,
because the unit initialisation triggers are inserted there afterwards
and they require one stack slot
git-svn-id: branches/jvmbackend@18507 -
only JVM constructs that are already implemented, but also ones that
will be supported in the future but that aren't implemented yet (to
make it easier to already adapt code to the future changes)
git-svn-id: branches/jvmbackend@18498 -
the types as declared in the system unit, since they can also be
used with equivalent but different types (e.g., byte vs shortint)
git-svn-id: branches/jvmbackend@18487 -
assignment-nodes. For global typed constants and typed constants/
local variable initialisers in regular functions/procedurs, the
assignments are performed in the unit initialisation code. For
those in object/record definitions and their methods, it's done
in the class constructor. Since we may not yet have parsed all
method implementations when the class constructor is parsed, part
of these may be initialised in a helper routine called from the
class constructor. The ones known when the class constructor is
parsed are inited there, because the ones marked as "final" and
declared as static class fields must be initialised in the class
constructor for Java
o new set systems_typed_constants_node_init in systems unit that
indicates that a target uses node trees to initialise typed consts
instead of an initialised data section
o mark typed constants in {$j-} mode as "final" for JVM
o mangle the name of staticvarsyms inside localtables a bit to avoid
name clashes (only with procedure names for now, no parameters yet
so can still cause problems with overloaded routines)
o after a routine has been parsed, it is now processed by
cnodeutils.wrap_proc_body(), which can add extra nodes before code
generation (used for injected the typed constant node trees)
git-svn-id: branches/jvmbackend@18475 -
in a single statement, to be added later)
o the unicodestrings are internally simply java.lang.String instances
o at the language level, the unicodestrings are assignment-compatible
with java.lang.String
o constant strings can be implicitly converted to java.lang.String
o since java.lang.String is immutable, in particular changing a
single character in a string is extremely inefficient. This could
be solved by letting unicodestring map to java.lang.StringBuilder,
but that would make integration with plain Java code harder
git-svn-id: branches/jvmbackend@18470 -
o when allocating array temps for the JVM target, use the specified
"forcesize" for the first dimension, since the arraydef size may
be invalid (e.g., in case it's an array constructor)
git-svn-id: branches/jvmbackend@18468 -
asis_target_specific_typecheck, and if so get the real
definition of its pointedtype rather than of the classrefdef
itself (the latter would never be different from the original
one, since there are no formal external classrefdef definitions)
git-svn-id: branches/jvmbackend@18467 -
o initialise class vars that need initialisations (records, arrays) in
class constructors
o treat class constructors as having a "void" resultdef rather than the
class type for JVM (maybe has to be done in general?)
o make it possible to specify pno_noleadingdollar to
tprocdef.customprocname() so it can be used for class constructors
(their name is lower cased because it mustn't conflict with other
identifiers, since their name doesn't matter anyway)
o added tsk_empty synthetic procdef kind which, as the name implies,
generates an empty body (for class generated constructors)
+ auto-generate class constructors in case a class has class vars that
need initialisation
git-svn-id: branches/jvmbackend@18462 -
and initialise global variables that are wrapped (records, arrays)
in those sections
o check whether pd.localst is assigned in dbgjasm, because it's
not for the unit initialisation routine
o moved insertbssdata() from ncgutil to ngenutil and override it
njvmutil (it does nothing in the latter, since global variables
are added as fields to the class representing the unit; the
initialisation is done in gen_initialize_code() in thlcgjvm)
o added force_init() and force_final() methods to ngenutil, so
that targets can force init/final routines separate from the
regular managed types infrastructure (used by JVM for forcing
an init section in case of records/arrays)
git-svn-id: branches/jvmbackend@18460 -
fpcbaserecordtype for the JVM target (intercept in ncnv via new
target_specific_general_typeconv helper, handle in as/is code)
* not only check for related types in htypechk in case they are
objdefs, but always do so (records are related to jlobject/fpcbaserecord
on the JVM target)
git-svn-id: branches/jvmbackend@18458 -
use of the register (can happen if the register is unused afterwards;
remove the alloc/dealloc results in no temp being allocated for the
register, so that instruction becomes invalid)
git-svn-id: branches/jvmbackend@18455 -
happens automatically when required (for constructing arrays/records;
reference counted types don't need special treatment since everything
is garbage collected)
git-svn-id: branches/jvmbackend@18453 -
implemented via classes, all descending from system.FpcBaseRecordType
(records are also considered to be "related" to system.FpcBaseRecordType
on the JVM target)
* several routines are auto-generated for all record-classes: apart
from a default constructor (if there is none), also clone (which
returns a new instance containing a deep copy of the current
instance) and deepCopy (which copies all fields of one instance
into another one)
o added new field "synthetickind" to tprocdef that indicates what
kind of synthetically generated method it is (if any), and
mark such methods also as "synthetic" in the JVM assembler code
o split off the JVM-specific parser code (e.g., to add default
constructors) into pjvm.pas
git-svn-id: branches/jvmbackend@18450 -
(and override for the JVM, making the register type for records
R_ADDRESSREGISTER instead of R_INTREGISTER there)
git-svn-id: branches/jvmbackend@18448 -
svn r17068,17071,17081,17136
* changed all init_paras code in both thlcgobj and ncgutil to use
location_get_data_ref() instead of direct a_load_loc_reg()/
ref.base:=reg so it also works with the JVM target
* changed all init_paras code so it works with targets that do
not pass an implicit high parameter for open array (and a similar
fix in ncgcal)
+ added support for initializing array (both regular and open)
"out" parameters of reference counted types on the JVM target
(the arrays will be initialised with nil rather than an empty
array for implementation reasons, see comments in compproc.inc)
* factored out calling of functions in the system unit directly
from hlcgobj
git-svn-id: branches/jvmbackend@18421 -
of functions at the caller side, since these will already be allocated
at the callee side
* added comment to cpupara explaining why we never allocate the function
result on the caller side and then pass it as an invisible parameter,
but instead always let the callee allocate it
git-svn-id: branches/jvmbackend@18420 -
disabled it under all circumstances on platforms that use garbage
collection for managed types since it's not required there
git-svn-id: branches/jvmbackend@18418 -
in case -sr is used (-sr code cannot be compiled, and is only used for
debugging; with -alr, the stack slot information is printed in the
assembler file)
git-svn-id: branches/jvmbackend@18410 -
nil-pointers typecasted to a class type, strings)
o escape ", \, #10 and #13 in string constants as required by Java
git-svn-id: branches/jvmbackend@18391 -
function, so it can be easily also used for constsym without adding
JVM-specific routines to symtype or duplicating the routine without
inheritance
+ added tconstsym support to jvmdef.jvmmangledbasename()
git-svn-id: branches/jvmbackend@18390 -
o support for copying value parameters at the callee side if they were
passed by reference in hlcg
o JVM g_concatcopy() implementation for arrays
o moved code to get length of an array from njvminl to hlcgcpu so it can
be reused elsewhere as well
o export array copy helpers from system unit for use when assigning one
array to another
o some generic support for types that are normally not implicit pointers,
but which are for the JVM target (such as normal arrays)
* handle assigning nil to a dynamic array by generating a setlength(x,0)
node instead of by hardcoding a call to fpc_dynarray_clear, so
target-specific code can handle it if required
* hook up gethltemp() for JVM ttgjvm so array temps are properly
allocated
git-svn-id: branches/jvmbackend@18388 -
o always create exceptvarsym entry for on-nodes (on all targets) to remove
some special cases when an unnamed exception was caught
o the JVM tryfinally node generates the finally code twice: once for the
case where no exception occurs, and once when it does occur. The reason
is that the JVM's static bytecode verification otherwise cannot prove
that we will only reraise the caught exception when we caught one in
the first place (the old "jsr" opcode to de-duplicate finally code
is no longer used in JDK 1.6 because it suffered from the same problem,
see Sun Java bug
http://webcache.googleusercontent.com/search?q=cache:ZJFtvxuyhfMJ:bugs.sun.com/bugdatabase/view_bug.do%3Fbug_id%3D6491544 )
git-svn-id: branches/jvmbackend@18387 -
o tobjectdef.jvm_full_typename() now gets an extra parameter to determine
whether or not the package name should be prepended, so it can be easily
used to generate the name of the .j file and of the class name inside it
git-svn-id: branches/jvmbackend@18384 -
to tasnode (like for regular Object Pascal classes)
* don't collect WPO information for Java classes in tloadvmtaddrnode.pass_1()
(devirtualization can't work for Java, since classes can always be loaded
at run time, except for final/sealed classes -- but that's not yet
implemented)
+ JVM is-node support, unified JVM type checking codegen for is- and as-nodes
git-svn-id: branches/jvmbackend@18383 -
o since the JVM does not support call-by-reference, setlength() works
by taking an argument pointing to the old array and one to the new
array (the latter is always created in advance on the caller side,
even if not strictly required, because we cannot easily create it
on the callee side in an efficient way). Then we copy parts of the
old array to the new array as necessary
o to represent creating a new dynamic array, the JVM target uses
an in_new_x tinlinenode
+ tasnode support for the JVM. Special: it can also be used to convert
java.lang.Object to dynamic arrays, and dynamic arrays of java.lang.Object
to dynamic arrays with more dimensions (arrays are special JVM objects,
and such support is required for the setlength support)
+ check whether explicit type conversions are valid, and if so, add the
necessary conversion code since we cannot simply reinterpret bit patterns
in most cases in the JVM:
o in case of class and/or dynamic array types, convert to an as-node
o in case of int-to-float or float-to-int, use java.lang.Float/Double
helpers (+ added the definitions of these helpers to the system unit)
git-svn-id: branches/jvmbackend@18378 -
for regular temps. This is required for targets that need special
handling of the temps depending on the type
* converted most gettemp() calls to gethltemp() calls
git-svn-id: branches/jvmbackend@18376 -
do nothing for finalization, assign nil pointer for dynamic arrays
and refcounted strings (not required for JVM per se, but required
because programmer's may use them without initialising them first
and then they should be empty rather than invalid)
git-svn-id: branches/jvmbackend@18368 -
(pushedparasize is already expressed in number of stackslots rather
than in bytes)
o when determining the number of stack slots to pop after calling a function
and not using its function result, round up the function result size to
the nearest multiple of 4 before determining the number os tack slots (to
correctly handle byte/word-sized results)
git-svn-id: branches/jvmbackend@18362 -
o self is encoded as "this" for javac compatibility
+ ait_jvar (for the above) and ait_jcatch (similar, for future try/catch
support) classes
+ support for smallset JVM type encoding (as int)
git-svn-id: branches/jvmbackend@18354 -
and skip comments (so it works again when -ar is used, because
tlhcgjvm.inc/decstack() insert comments in that case)
git-svn-id: branches/jvmbackend@18340 -
that's how these operations also work on other architectures
* fixed tjvmaddnode.second_generic_compare() for this new order
git-svn-id: branches/jvmbackend@18338 -
operation is 32 bit, not 64 bit (both in the compiler and in the
JVM)
* request a 64 rather than a 32 bit shift operation from tjvmshlshrnode
if the result is 64 bit (rather than in case shift value is 64 bit,
since as described above that never happens)
git-svn-id: branches/jvmbackend@18337 -
in case of -ar add to the assembler output the height of the
evaluation stack every time it's increased or decreased (to
more easily track missing/wrong inc/decstack() operations)
git-svn-id: branches/jvmbackend@18335 -
u64bit division/mod, and no overflow checking for divisions
(none yet for add nodes either, but that has to be implemented
in hlcgcpu rather than in the add-nodes themselves)
git-svn-id: branches/jvmbackend@18331 -
o handle them like for regular classes (return a class instance,
although this is technically not true since they don't return
anything; will be changed in the future)
o because of the previous point, make sure that we handle the
"function result" properly and don't pop too many values from
the evaluation stack when calling one constructor from another
o added "extra_pre_call_code" method used by njvmcal to insert
the "new" opcode to create the new class instance before
calling a constructor
o when a constructor does not call any other constructor (inherited
or otherwise), automatically insert a call to the inherited
parameterless constructor as required by the jvm standard)
TODO: check that *if* an inherited or other constructor is called
from another constructor, that it does so as the first statement/
call
git-svn-id: branches/jvmbackend@18328 -
renamed refactored routine to do_release_unused_return_value()
which by default releases temps, and if something else needs to
be done override (and call inherited if temps may have to be
released)
git-svn-id: branches/jvmbackend@18327 -
support), passed through to original ncgutils version in thlcg2ll
+ thlcgobj.location_force_mem() implementation
* order parameters for jvm similar to those for i386 without fixed_stack,
so we don't need temporary paralocations
* converted most of ncgcal to thlcg
* disabled special handling for virtual methods for jvm in ncgcal, as all
invocations are name-based there
+ njvmcal with special jvm callnode support:
o always move the function result into a memory temp
o when freeing an unused function result, use a_pop(2) and adjust
the internal evaluation stack height counter
o after the call instruction, adjust the evaluation stack height counter
by subtracting the number of the pushed parameter slots, adjusted for
the slots taken up by the function result
git-svn-id: branches/jvmbackend@18325 -
as well as insert_regalloc_info_all()
+ "register" allocator for the jvm target, which simply replaces every
virtual register with a temp. This is done for all register types in
one pass, so that the temps can be easily reused without worrying
about conflicts. Small optimisation: alloc/store/dealloc/load
sequences for a single reg are removed (many of these are generated
because most cg nodes return their value in a register which is then
immediately consumed by the parent)
* map addressregisters to integer registers with size R_SUBD, because
they require one stackslot (the subregister type is used by rgcpu
to determine the size of the temp it has to allocate)
git-svn-id: branches/jvmbackend@18316 -
handle vectorfpu (floatdef->MMREG) and softfloat (floatdef->INTREG)
+ thlcg.getregisterfordef(), which uses def2regtyp() to allocate a register
appropriate to hold values of that tdef type
+ generic thlcg.location_force_reg() implementation. Note that for
low-level code generator targets it may be slightly less efficient than
the implementation in hlcg2ll (from ncgutil) because it does not play
any tricks with the register or location size, or with reference offsets,
to truncate values
git-svn-id: branches/jvmbackend@18315 -
+ basic target information for jvm target (assembling/linking
helpers are still dummies for now)
+ basic jasmin assembler writer
+ cpunode and cputarg units to include the target units in the
compiler
git-svn-id: branches/jvmbackend@18309 -
+ added arrayreftype field to treference for the jvm, because
array loads have to be performed using special instructions,
so this information has to be passed on to the code generator.
This information will have to be added in t(cg)vecnode.
* moved concatenating the generated code for a procedure to the
al_procedures list to thlcg.record_generated_code_for_procdef(),
which is overridden by jvm/hlcgcpu to attach that code to the
procdef it belongs with. The reason is that a Java class file
can only contain code for one class, so we have to write out
the assembler grouped per class instead of in the order the
routines appear in the source code
* also committed forgotten changes in psub after moving the
register allocator initialisation to hlcgobj
git-svn-id: branches/jvmbackend@18308 -
on the evaluation stack, while on the callee side they are placed
in local variables 1..n
o the jvm only supports call-by-value. call-by-reference parameters
will have to be emulated at a higher level
o similarly, ret_in_param() always returns false
o all function results are returned on the evaluation stack (both
on the caller and callee side)
git-svn-id: branches/jvmbackend@18306 -
longer aligns the temp offset in case it's odd (not required for
jvm)
* ignore the "def" argument in alloctemp() for the jvm since all stack
slots can be reused for anything there, and reformatted its alloctemp()
git-svn-id: branches/jvmbackend@18297 -
(single, double, string, widestring), basic constructors,
spilling_get_operation_type() implementation
o special: O_MOV_SOURCE/O_MOV_DEST don't make sense for this target,
since it's stack based
git-svn-id: branches/jvmbackend@18294 -
o special things:
* no tasmcond (conditions are currently encoded in opcode names,
may be changed later)
* no actual registers, just a couple of fake ones to indicate
the locals stack and the operands/evaluation stack
* no registers for the return value; they're returned on the
evaluation stack; may be handled like the x87 in the future
git-svn-id: branches/jvmbackend@18291 -
* made ttgobj.create virtual, added a "tgobjclass: class of ttgobj = ttgobj"
variable and use that one to instantiate new temp allocators
* created ttgjvm descendant that only allows allocations of 4 or 8 bytes
(rounding allocations < 4 bytes up to 4 bytes), and that divides the
offsets/sizes by 4 so we get stack slot numbers
git-svn-id: branches/jvmbackend@18282 -
