Pack200: Packed Class Archive Specification - JDK 5 Documentation v6.0, Java 2 SDK 英文文档

5.5.6. Metadata Layouts

  [NH[(1)]]
  [RSH NH[RUH(1)]]
  [TB
    (66,67,73,83,90) [KIH]
    (68)  [KDH]
    (70)  [KFH]
    (74)  [KJH]
    (99)  [RSH]
    (101) [RSH RUH]
    (115) [RUH]
    (91)  [NH[(0)]]
    (64)  [RSH [RUH(0)]]
    ()    []
    ]
 

  [NB[(1)]]
  [NH[(1)]]
  [RSH NH[RUH(1)]]
  [TB...]
 

5.5.7. Unusual Layout Usages

5.6. Source File Abbreviation

• Every character up to and including the last slash or dot ('/' or '.') is removed. (This removes any package prefix.)
• If there is a character of code 0x2D or lower in the remaining string, the first such and all following characters are removed. (This removes any nested class name mangling.)
• The suffix characters ".java" are appended.

This rule matches the historical behavior of many Java compilers, and allows the compressor to avoid allocating Utf8 strings for "obvious" derived SourceFile names. (If a compressor needs to transmit an irregular SourceFile attribute with a true null reference, it must use a non-standard layout.) Here is a table of examples of class names and the corresponding derived SourceFile names.

Class	SourceFile
foo	foo.java
foo/bar	bar.java
foo/bar$baz	bar.java
foo/bar#baz#1	bar.java
foo.bar.baz#1	baz.java

5.7. Nested Classes

  ic_bands:
        *ic_this_class :UDELTA5 [#ic_count] (cp_Class)
        *ic_flags :UNSIGNED5 [#ic_count]
        *ic_outer_class :DELTA5 [COUNT(1<<16,...)]  (null or cp_Class)
        *ic_name :DELTA5 [LENGTH(*ic_outer_class)] (null or cp_Utf8)
 

The ic_this_class and ic_flags bands are both of length #ic_count, and the corresponding elements of these bands specify, for each tuple, the nested class identity (represented as a cp_Class reference) and the flags bitmask.

The nested class flag bit at position 16 (as set in the mask 0x00010000) is inside the archive file to indicate if there are corresponding entries for the tuple in the ic_outer_class and ic_name bands. Thus, the length of both of these bands is the sum of all flag bits at position 16. Typically, only a few percent of nested classes need to set this bit and specify outer and name fields explicitly.

If a tuple has an entry in the ic_outer_class and ic_name bands, these specify its outer class and simple name. (They are represented respectively as a possibly null cp_Class reference and a possibly null cp_Utf8 reference.) Otherwise, the tuple's outer class and name are said to be predicted. In this case, they must be correctly predictable from the name of the nested class itself, by parsing its spelling.

A nested class has a bytecode name which names the class within class files. The spelling of this name, sometimes called a "mangled name", has extra punctuation signs and perhaps digits as well as the name of a containing class. If the bytecode name can be parsed into an outer class and a class name, and this class and name are identical with the true outer class and class name of the nested class, then we say the outer class and class name are predictable.

The extraction of the predictable outer class and class name must follow the following grammar for class bytecode names, as applied to the spelling of the nested class name. (This grammar is independent of the grammar governing band structure, or any other grammar appearing in other parts of this specification.) The terminal DOLLAR refers to any character (such as '$' or '#') whose code is 0x2D or lower. The terminals SLASH refers to the slash or dot characters '/' or '.', which have the codes 0x2E and 0x2F. The terminal DIGIT refers to an ASCII decimal digit, one of the ten character codes from 0x30 to 0x39 inclusive. The terminal LETTER refers to any other character. That is, it refers to any character whose code is 0x3A or higher.

  bcn:
        (bcnCase1 | bcnCase2 | bcnCase3 | bcnCase4)
  bcnCase1:
        packageQual (namePart)? DOLLAR number
  bcnCase2:
        packageQual (namePart)? DOLLAR number DOLLAR predictableICName
  bcnCase3:
        predictableOuter DOLLAR predictableICName
  bcnCase4:
        packageQual (namePart)?

  predictableOuter:
        packageQual namePart
  predictableICName:
        LETTER (LETTER | DIGIT)*

  namePart:
        (LETTER | DIGIT | DOLLAR)+
  number:
        (DIGIT)+
  packageQual:
        (namePart SLASH)*
 

This grammar ambiguously divides an arbitrary class name into several parts, which may include an optional predictedOuter prefix, an optional predictedICName suffix, and a optional numeric suffix. Any ambiguity must be resolved by preferring the alternative cases for bcn nonterminal in the order given. E.g., if bcnCase1 matches, it is used, even though either or both other cases may match also.

The predictable nested class name is the string corresponding to the predictableICName nonterminal, if it was parsed. Otherwise the predictable nested class name is taken to be null.

If a predictableOuter outer nonterminal is parsed, the corresponding string is the predictable outer class name. Otherwise the predictable outer class reference is taken to be null. Note that if a number nonterminal is parsed, no predictableOuter can be parsed. Here are some examples of prediction, nonprediction, and misprediction:

nested class:	member named Entry of java/util/Map
mangled name:	java/util/Map$Entry
outer, name:	java/util/Map, Entry
predictable?	yes (since Map.Entry is a member of Map)
nested class:	anonymous
mangled name:	java/util/AbstractList$1
outer, name:	(none), (none)
predictable?	yes (since the ic_name is null)
nested class:	non-member, named Local
mangled name:	java/util/AbstractList$2$Local
outer, name:	(none), Local
predictable?	yes (since the ic_name is Local)
nested class:	non-member, named Local
mangled name:	java/util/AbstractList#2#Local
outer, name:	(none), Local
predictable?	yes (since the ic_name is Local)
nested class:	member named $2$Local of Foo
mangled name:	Foo$$2$Local
outer, name:	(none), Local
predictable?	no (outer name missing, ic_name mispredicted)
nested class:	class named Red$Herring
mangled name:	Red$Herring
outer, name:	Red, Herring
predictable?	no (the predicted ic_name Herring)
nested class:	member named Q of X$1
mangled name:	X$1$Q
outer, name:	(none), Q
predictable?	no (since Q is a member of X$1)
nested class:	member named Z of X$Y
mangled name:	X$Y$Z
outer, name:	X$Y, Z
predictable?	yes (since X$Y.Z is a member of X$Y)
nested class:	member named Y$Z of X
mangled name:	X$Y$Z
outer, name:	X$Y, Z
predictable?	no (outer name and ic_name are mispredicted)

When a decompressor is processing a nested class name marked "predictable", it must parse the nested class's bytecode name into an enclosing class, an optional number, and an optional name. (The compressor might choose to always specify outer classes and names explicitly, in which case the decompressor would not be required to parse the nested class names at all. However, decompressors must always be prepared to perform this parsing.)

If the compressor did not transmit an entry in cp_Utf8 or cp_Class for a predicted name or outer class, the decompressor must create such a constant internally. It will be referred to as a cp_Utf8 or cp_Class entry, even though it is not in the constant transmission sequence cp_All. However, if the compressor did transmit a constant with the same spelling as the predicted name or outer class, the decompressor must use that constant rather than creating a new one. The creation of such internal constants within the decompressor is detectable in an output file because they are inserted in a special order in the output file's constant pool. (See discussion of the output ordering rules below.)

Please see the Appendix section for pseudo-code explaining this concept.

When the decompressor receives the ic_this_class, ic_flags, ic_name, and ic_outer_class bands, and performs any required name parsing, it creates the set ic_All of four-tuples. This set is the primary source of InnerClasses attributes synthesized for individual class files by the decompressor.

5.8. Class Schema

  class_bands:
        *class_this :DELTA5 [#class_count] (cp_Class)
        *class_super :DELTA5 [#class_count] (cp_Class)
        *class_interface_count [#class_count] :DELTA5
        *class_interface :DELTA5 [SUM(*class_interface_count)] (cp_Class)
        *class_field_count :DELTA5 [#class_count]
        *class_method_count :DELTA5 [#class_count]

        *field_descr :DELTA5 [SUM(*class_field_count)] (cp_Descr)
        field_attr_bands

        *method_descr :MDELTA5 [SUM(*class_method_count)] (cp_Descr)
        method_attr_bands

        class_attr_bands
        code_bands
 

The class_this, class_super, class_flags_lo, class_flags_hi (if present), class_interface_count, class_field_count, and class_method_count bands are all of length #class_count, and the corresponding elements of these bands transmit in turn each class's name and super-class, and the number of implemented interfaces, declared fields, and declared methods. (The access modifier bits in each class's flags word are mixed with attribute indicators, and are transmitted in class_flags_lo, as defined above.)

The class_interface band contains runs of class interface declarations, one run for each element of class_interface_count, and applying to the corresponding class.

Each element of class_this, class_super, and class_interface is a reference (in fact, a non-null reference) to the constant pool cp_Class.

In the unique case of java/lang/Object, the stored super-class must be a null reference. Rather than disturbing the transmission of all other elements of the class_super band, we use the convention that a compressor, when it encounters a null super-class reference, must transmit in its place a duplicate of the current class's reference. (Since it is illegal for a class to inherit from itself, this renumbering of null references is unambiguous.)

The field_descr bands contain one run of elements for each element of class_field_count, and apply to successive fields in the corresponding class. The method_descr bands contain one run of elements for each element of class_method_count, and apply to successive methods in the corresponding class.

(Unlike the class file format, field and method descriptors are stored as single references to a "name-and-type" constant pool, rather than as pairs of name references and type references.)

The method_descr band has a primary encoding which performs best if the method references are mostly sorted. Compressors may elect to take advantage of this fact by sorting the methods in each class. (Note: It is generally acceptable for compressors to reorder methods for best compression, but fields must not be reordered, since their order is apparent through reflection and significant to some facilities, such as serialization.)

Attribute bands are placed in the positions indicated by the grammar. They are described below. The attribute bands include flags bands, which carry both access modifiers and attribute indicators for the corresponding classes, fields, and methods.

The Pack200 archive ends with bands devoted to method code blocks and the bytecodes themselves. If a method has a code attribute, the compressor must mention it in the flags bit (the 6th LSB) to which it is assigned, or as an overflow attribute (with an index of 6). The fields of a Code attribute are transmitted in the bands organized under the code_bands nonterminal.

The specification of each Code attribute begins with three key parameters, which declare the number of stack and local slots, and the number of handlers. #Stack is the number of stack slots used by the code. #NALocal is the number of non-argument locals used by the code. (The actual number of locals declared in the class file will be #NALocal plus the number of locals required by method parameters, as determined by the method's signature.) #Handler is the number of exception handlers.

. The code_headers band contains a series of bytes, each of which tersely encodes the first three key parameters of a code attribute: Each of the 255 non-zero byte values encodes a unique triple of #NALocal, #Stack, and #Handler. There is (of course) one code header for each Code attribute transmitted.

The special code header byte zero (0x00) indicates that the code attribute's three parameters are to be found, instead, in the code_max_stack, code_max_na_locals, and code_handler_count bands. If the code header byte is non-zero, these three bands do not transmit entries for the corresponding Code attribute.

The code_flags_lo band transmits an entry for a corresponding Code attribute if and only if one or both of the following is true: the Code attribute's code header byte is zero, or the have_all_code_flags bit is set in the #archive_options word. If a Code attribute does not have a transmitted code_flags_lo value, its flags word is taken to be zero, and it has no sub-attributes. The code_flags_hi band transmits a corresponding entry if and only if the code_flags_lo value was transmitted as just described, and also #have_code_flags_hi is set.

Non-zero code header bytes denote code attribute parameters according to the following scheme:

Header Range	#Stack	#NALocal	#Handler
1 <= x <= 144	(x-1) % 12	(x-1) / 12	0
145 <= x <= 208	(x-145) % 8	(x-145) / 8	1
209 <= x <= 255	(x-209) % 7	(x-209) / 7	2
x = 0	code_max_stack	code_max_na_locals	code_handler_count

This scheme allows a single byte to describe a Code attribute in about 95% of all cases. (Note: If debugging attributes are not stripped, the compressor should probably set the have_all_code_flags bit, so that the presence of these attributes will not force the use of the special zero code header byte.)

Regardless of whether the exception handler count is encoded in the one-byte code header, or whether it is an explicit element of code_handler_count, for each code attribute there is a run of values in code_handler_start_P, code_handler_end_PO, code_handler_catch_PO, and code_handler_class_RCN, one for each handler, as if those bands were governed by an attribute layout 'NV[PHPOHPOHRCNH]' (there is no band governed by the 'NV' layout element itself; it is the handler count).

That is, each triplet of Handler.start, Handler.end, and Handler.catch values are transmitted as renumbered (by renumber_bci). Also, Handler.end is transmitted as a difference of its renumbering with that of Handler.start in the same triplet, and Handler.catch is transmitted as a difference of its renumbering with that of Handler.end in the same triplet. Finally, code_handler_class_RCN is transmitted as an incremented reference into cp_Class, or zero if the handler class reference is null.

  code_bands:
        *code_headers :BYTE1 [COUNT(Code,...)]

        *code_max_stack :UNSIGNED5 [COUNT(0,*code_headers)]
        *code_max_na_locals :UNSIGNED5 [COUNT(0,*code_headers)]

        *code_handler_count :UNSIGNED5 [COUNT(0,*code_headers)]
        *code_handler_start_P :BCI5 [SUM(*code_handler_count)]
        *code_handler_end_PO :BRANCH5 [SUM(*code_handler_count)]
        *code_handler_catch_PO :BRANCH5 [SUM(*code_handler_count)]
        *code_handler_class_RCN :UNSIGNED5 [SUM(*code_handler_count)] (null or cp_Class)

        code_attr_bands
 

5.9. Attribute Bands

If the compressor marks a class (resp. field, method, or code) as having overflow attributes, it must transmit a corresponding element in the class_attr_count band (resp. field_attr_count, method_attr_count, or code_attr_count band). This count, in turn, specifies the size of a run in class_attr_indexes (resp. field_attr_indexes, method_attr_indexes, or code_attr_indexes) of indexes of attribute layouts governing attributes of the class (resp. field, method, or code). The compressor must transmit a attribute layout definition index for each overflow attribute.

For each attribute layout selected by a bit in an element of class_flags or by an element of class_attr_indexes, the compressor must retrieve data governed by the layout elements from the class attribute and transmit elements encoding this data in the bands governed by the layout. Similar conditions apply for field, method, and code attributes.

For each attribute layout that transmits data and contains backward calls, the compressor must transmit the number of times each backward callable will be the subject of a backward call as the decompressor works its way through attribute layouts. These call counts are transmitted in the class_attr_calls, field_attr_calls, method_attr_calls, and code_attr_calls bands, according to the context type of the layouts they apply to. The call counts are transmitted, one per backward callable, in the definition order of the callables. (See above.) Call counts are only transmitted for backward callables which occur within layouts that are used at least once. Call counts do not count entries to any callable because of a forward call, nor do they count the initial call to a callable which initiates attribute processing. A call count might be zero if a backward callable occurs in a layout that is used, but which does not happen to reach the backward callable. These call counts are needed to break a circularity inherent in the sizing of bands for mutually-recursive layouts. They provide the minimum information necessary for the decompressor to locate all the bands in the archive, prior to distributing the band values to the various output classes. For this reason, the call counts are supplied one per layout, summed over all attribute occurrences of that each layout.

The decompressor is responsible for processing any explicit attribute layout definitions transmitted by the compressor. It must prepare to receive the additional bands governed by those layouts. When it reads layout definition indexes, it must prepare to read in the correct number of values transmitted in each of those additional bands.

The grammar of bands defined in this specification includes bands governed by all predefined attribute layouts. For clarity, some band names mention a part of the layout element that created them. Note that the Deprecated attributes have no bands, because their layouts are empty.

Attributes named "Synthetic", although part of the standard class file format in earlier versions of Java, are not directly supported in this specification, because that attribute has been replaced by a new flag bit (ACC_SYNTHETIC, 0x1000) in more recent versions of Java. However, compressors are encouraged to process Synthetic attributes, and any other zero-length attributes not predefined here, by assigning them unused flag bits, and emitting explicit zero-length layout definitions for decompressors to follow.

If the compressor elects to define new layouts, the bands governed by those layouts are added immediately after the bands for the predefined layouts. (The order and structure of the predefined attribute bands reflects the predefined layout definitions in a regular way, as if the compressor had in fact defined them explicitly.)

  class_attr_bands:
        *class_flags_hi :UNSIGNED5 [#class_count*#have_class_flags_hi]
        *class_flags_lo :UNSIGNED5 [#class_count]
        *class_attr_count :UNSIGNED5 [COUNT(1<<16,...)]
        *class_attr_indexes :UNSIGNED5 [SUM(*class_attr_count)]
        *class_attr_calls :UNSIGNED5 [...]
        *class_SourceFile_RUN :UNSIGNED5 [COUNT(SourceFile,...)] (null or cp_Utf8)
        *class_EnclosingMethod_RC :UNSIGNED5 [COUNT(EnclosingMethod,...)] (cp_Class)
        *class_EnclosingMethod_RDN :UNSIGNED5 [COUNT(EnclosingMethod,...)] (null or cp_Descr)
        *class_Signature_RS :UNSIGNED5 [COUNT(Signature,..)] (cp_Signature)
        class_metadata_bands
        ic_local_bands
        *class_file_version_minor_H :UNSIGNED5 [COUNT(version,...)]
        *class_file_version_major_H :UNSIGNED5 [COUNT(version,...)]

  field_attr_bands:
        *field_flags_hi :UNSIGNED5 [SUM(*class_field_count)*#have_field_flags_hi]
        *field_flags_lo :UNSIGNED5 [SUM(*class_field_count)]
        *field_attr_count :UNSIGNED5 [COUNT(1<<16,...)]
        *field_attr_indexes :UNSIGNED5 [SUM(*field_attr_count)]
        *field_attr_calls :UNSIGNED5 [...]
        *field_ConstantValue_KQ :UNSIGNED5 [COUNT(ConstantValue,...)] (cp_Int, etc.; see note)
        *field_Signature_RS :UNSIGNED5 [COUNT(Signature,...)] (cp_Signature)
        field_metadata_bands

  method_attr_bands:
        *method_flags_hi :UNSIGNED5 [SUM(*class_method_count)*#have_method_flags_hi]
        *method_flags_lo :UNSIGNED5 [SUM(*class_method_count)]
        *method_attr_count :UNSIGNED5 [COUNT(1<<16,...)]
        *method_attr_indexes :UNSIGNED5 [SUM(*method_attr_count)]
        *method_attr_calls :UNSIGNED5 [...]
        *method_Exceptions_N :UNSIGNED5 [COUNT(Exceptions,...)]
        *method_Exceptions_RC :UNSIGNED5 [SUM(*method_Exceptions_N)] (cp_Class)
        *method_Signature_RS :UNSIGNED5 [COUNT(Signature,...)] (cp_Signature)
        method_metadata_bands

  code_attr_bands:
        *code_flags_hi :UNSIGNED5 [...*#have_code_flags_hi]
        *code_flags_lo :UNSIGNED5 [...]
        *code_attr_count :UNSIGNED5 [COUNT(1<<16,...)]
        *code_attr_indexes :UNSIGNED5 [SUM(*code_attr_count)]
        *code_attr_calls :UNSIGNED5 [...]
        *code_StackMapTable_N :UNSIGNED5 [COUNT(StackMapTable,...)]
        *code_StackMapTable_frame_T :BYTE1 [SUM(*code_StackMapTable_N)]
        *code_StackMapTable_local_N :UNSIGNED5 [COUNT(255,*code_StackMapTable_frame_T)]
        *code_StackMapTable_stack_N :UNSIGNED5 [COUNT(255,*code_StackMapTable_frame_T)]
        *code_StackMapTable_offset :UNSIGNED5 [...]
        *code_StackMapTable_T :BYTE1 [...]
        *code_StackMapTable_RC :UNSIGNED5 [COUNT(7,*code_StackMapTable_T)]
        *code_StackMapTable_P :BCI5 [COUNT(8,*code_StackMapTable_T)]
        *code_LineNumberTable_N :UNSIGNED5 [...]
        *code_LineNumberTable_bci_P :BCI5 [...]
        *code_LineNumberTable_line :UNSIGNED5 [...]
        *code_LocalVariableTable_N :UNSIGNED5 [...]
        *code_LocalVariableTable_bci_P :BCI5 [...]
        *code_LocalVariableTable_span_O :BRANCH5 [...]
        *code_LocalVariableTable_name_RU :UNSIGNED5 [...] (cp_Utf8)
        *code_LocalVariableTable_type_RS :UNSIGNED5 [...] (cp_Signature)
        *code_LocalVariableTable_slot :UNSIGNED5 [...]
        *code_LocalVariableTypeTable_N :UNSIGNED5 [...]
        *code_LocalVariableTypeTable_bci_P :BCI5 [...]
        *code_LocalVariableTypeTable_span_O :BRANCH5 [...]
        *code_LocalVariableTypeTable_name_RU :UNSIGNED5 [...] (cp_Utf8)
        *code_LocalVariableTypeTable_type_RS :UNSIGNED5 [...] (cp_Signature)
        *code_LocalVariableTypeTable_slot :UNSIGNED5 [...]
 

A class possesses a "class-file version" pseudo-attribute if its class file's minor or major version differs from the #default_class_minver or #default_class_majver, respectively. This pseudo-attribute is a pair of 16-bit integers giving the major and minor version numbers of the class's file. These integers are stored in the header of the class's file, rather than in an attribute record. As is the convention with classfiles, the minor version number comes first. Thus minor version numbers are transmitted in the band class_file_version_minor_H, and major version numbers in class_file_version_major_H. Decompressors are not required to process archives with greater minor or major version numbers than those of the specification they were engineered to process. Decompressors are required to to process archives with the same major and smaller or equal minor version numbers.

Local InnerClasses Attributes

The four-tuples of all local InnerClasses attributes are transmitted in five bands:

  ic_local_bands:
        *class_InnerClasses_N :UNSIGNED5 [COUNT(InnerClasses,...)]
        *class_InnerClasses_RC :UNSIGNED5 [SUM(*class_InnerClasses_N)] (cp_Class)
        *class_InnerClasses_F :UNSIGNED5 [SUM(*class_InnerClasses_N)]
        *class_InnerClasses_outer_RCN :UNSIGNED5 [COUNT(!=0,*class_InnerClasses_F)]  (null or cp_Class)
        *class_InnerClasses_name_RUN :UNSIGNED5 [COUNT(!=0,*class_InnerClasses_F)] (null or cp_Utf8)
 

Each four-tuple <C,F,C2,N> transmitted in these attribute bands is called a local IC tuple. (Note that the class file format stores the elements of these four-tuples in a different order, (C,C2,N,F).) For a given class file X, the sequence of local IC tuples is called ic_Local(X).

For each local IC tuple, the compressor transmits, at minimum, a C value and an F value. If a transmitted flags value F is not zero, there are also corresponding transmitted constant pool references C2 and N. As a whole, the transmitted local tuple may or may not be equivalent to a global tuple from ic_All.

As an abbreviation, the compressor may transmit only a C and a flags value of zero, if the local IC tuple to be transmitted is equivalent to a member of ic_All, and no other member of ic_All defines the same class C. In this case the decompressor must behave exactly as if all four tuple components had been explicitly transmitted.

(If all four components of a local tuple are transmitted, and the flags value to be transmitted is in fact zero, the value 0x00010000 must be transmitted instead of zero for the flags.)

Frequently, the compressor will not need to transmit any local IC tuples at all, since the set of four-tuples to be stored in the class's InnerClasses attribute will be exactly ic_Relevant(X), with no adjustment required. If extraneous four-tuples (beyond those required by a minimized constant pool) are found in the input class file, then some local IC tuples will be needed (with zero flags), to provide local "roots" for the extra InnerClasses entries. This can occur if a compiler requires an InnerClasses entry for a class mentioned only in a signature. Compressors are required to predict which classes require such extra local roots, and transmit only the unexpected four-tuples.

5.9.1. Metadata Transmission

  class_metadata_bands:
        class_RVA_bands
        class_RIA_bands

  field_metadata_bands:
        field_RVA_bands
        field_RIA_bands

  method_metadata_bands:
        method_RVA_bands
        method_RIA_bands
        method_RVPA_bands
        method_RIPA_bands
        method_AD_bands

  class_RVA_bands:
        *class_RVA_anno_N :UNSIGNED5 [...] 
        *class_RVA_type_RS :UNSIGNED5 [...] (cp_Signature)
        *class_RVA_pair_N :UNSIGNED5 [...] 
        *class_RVA_name_RU :UNSIGNED5 [...] (cp_Utf8)
        *class_RVA_T :BYTE1 [...] 
        *class_RVA_caseI_KI :UNSIGNED5 [...] (cp_Int)
        *class_RVA_caseD_KD :UNSIGNED5 [...] (cp_Double)
        *class_RVA_caseF_KF :UNSIGNED5 [...] (cp_Float)
        *class_RVA_caseJ_KJ :UNSIGNED5 [...] (cp_Long)
        *class_RVA_casec_RS :UNSIGNED5 [...] (cp_Signature)
        *class_RVA_caseet_RS :UNSIGNED5 [...] (cp_Signature)
        *class_RVA_caseec_RU :UNSIGNED5 [...] (cp_Utf8)
        *class_RVA_cases_RU :UNSIGNED5 [...] (cp_Utf8)
        *class_RVA_casearray_N :UNSIGNED5 [...] 
        *class_RVA_nesttype_RS :UNSIGNED5 [...] (cp_Signature)
        *class_RVA_nestpair_N :UNSIGNED5 [...] 
        *class_RVA_nestname_RU :UNSIGNED5 [...] (cp_Utf8)

  class_RIA_bands:
        *class_RIA_anno_N :UNSIGNED5 [...] 
        (analogous to class_RVA_bands)

  field_RVA_bands:
        *field_RVA_anno_N :UNSIGNED5 [...] 
        (analogous to class_RVA_bands)

  field_RIA_bands:
        *field_RIA_anno_N :UNSIGNED5 [...] 
        (analogous to field_RVA_bands)

  method_RVA_bands:
        *method_RVA_anno_N :UNSIGNED5 [...] 
        (analogous to class_RVA_bands)

  method_RIA_bands:
        *method_RIA_anno_N :UNSIGNED5 [...] 
        (analogous to method_RIA_bands)

  method_RVPA_bands:
        *method_RVPA_param_NB :BYTE1 [...] 
        *method_RVPA_anno_N :UNSIGNED5 [...] 
        (analogous to method_RVA_bands)

  method_RIPA_bands:
        *method_RIPA_param_NB :BYTE1 [...] 
        *method_RIPA_anno_N :UNSIGNED5 [...] 
        (analogous to method_RVPA_bands)

  method_AD_bands
        *method_AD_T :BYTE1 [...] 
        *method_AD_caseI_KI :UNSIGNED5 [...] (cp_Int)
        *method_AD_caseD_KD :UNSIGNED5 [...] (cp_Double)
        *method_AD_caseF_KF :UNSIGNED5 [...] (cp_Float)
        *method_AD_caseJ_KJ :UNSIGNED5 [...] (cp_Long)
        *method_AD_casec_RS :UNSIGNED5 [...] (cp_Signature)
        *method_AD_caseet_RS :UNSIGNED5 [...] (cp_Signature)
        *method_AD_caseec_RU :UNSIGNED5 [...] (cp_Utf8)
        *method_AD_cases_RU :UNSIGNED5 [...] (cp_Utf8)
        *method_AD_casearray_N :UNSIGNED5 [...] 
        *method_AD_nesttype_RS :UNSIGNED5 [...] (cp_Signature)
        *method_AD_nestpair_N :UNSIGNED5 [...] 
        *method_AD_nestname_RU :UNSIGNED5 [...] (cp_Utf8)

As an aid to understanding the metadata layout, here is a brief description of each of the method metadata bands, as used by classfiles which comply with JSR 175. The class and field metadata bands function in a similar way.

Note that JSR 175 does not allow embedded references to CONSTANT_Class entries, even where a reference to a class is required. Instead, JSR 175 requires that references to classes be encoded as field signatures. (Their names are bracketed by 'L' and ';'.) The Pack200 archive transmits all such values as references into cp_Signature, not cp_Class.

For each use of a parameter annotation attribute, the method_RVPA_param_NB band or method_RIPA_param_NB band (for invisible annotations) transmits an unsigned byte indicating the parameter count. For each use of a non-parameter annotation attribute, the method_RVA_anno_N band or method_RIA_anno_N band (for invisible annotations) transmits an annotation count. Likewise, for each annotated parameter, the method_RVPA_anno_N band or method_RIPA_anno_N band (for invisible annotations) transmits an annotation count.

For each visible non-parameter annotation, the method_RVA_type_RS band transmits an annotation's type, and the method_RVA_pair_N band transmits the number of that annotation's member-value pairs. Analogous values for invisible annotations and parameter annotations are transmitted in the bands method_RIA_type_RS, method_RIA_pair_N, method_RVPA_type_RS, method_RVPA_pair_N, method_RIPA_type_RS, and method_RIPA_pair_N. For each member-value pair transmitted as a direct part of a visible non-parameter annotation, the method_RVA_name_RU band transmits the member name. Analogous values for invisible annotations and parameter annotations are transmitted in the bands method_RIA_name_RU, method_RVPA_name_RU, and method_RIPA_name_RU.

For each value transmitted with a visible non-parameter annotation, whether directly, or indirectly via a nested value or annotation, the method_RVA_T band transmits a byte which selects the format of the annotation value, and analogous values are transmitted in the bands method_RIA_T, method_RVPA_T, method_RIPA_T, and (for annotation defaults) method_AD_T, Direct uses of these value tag bands are counted by summing the values of the corresponding preceding pair-count band (method_RVA_pair_N, etc.). This count also includes the sum of the corresponding following nested pair-count band (method_RVA_nestpair_N, etc.) and nested array length bands (method_RVA_casearray_N, etc.).

Since those latter sums come from backward calls in the attribute layout, they cannot be directly computed by the decompressor, but their sum is reported by the compressor as an element of method_attr_calls. (Note that the last callable in each metadata layout is the target of two backward calls from inside itself.) That band contains these backward call counts for method_RVA_T, method_RIA_T method_RVPA_T, method_RIPA_T method_AD_T, in that order. If there are no occurrences of a method metadata attribute, then the corresponding backward call is omitted. (Thus, there are up to five backward call counts transmitted for method metadata.) If there are compressor-defined recursive layouts used in the archive, their backward call counts follow the counts for metadata in method_attr_calls.

For each value tag of 'B', 'C', 'I', 'S', or 'Z', there is a corresponding element transmitted in method_RVA_caseI_KI which supplies the value as a cp_Int reference. Analogous integer references within invisible and parameter annotations and annotation defaults are transmitted in the bands method_RIA_caseI_KI, method_RVPA_caseI_KI, method_RIPA_caseI_KI, and method_AD_caseI_KI. For each value tag of 'e', the bands method_RVA_caseet_RS and method_RVA_caseec_RU transmit the signature of the class and name of the member of an enumeration constant, as references into cp_Signature and cp_Utf8. For each value tag of '[', the band method_RVA_casearray_N transmits the length of a nested value array. For each value tag of '@', the bands method_RVA_nesttype_RS and method_RVA_nestpair_N transmit the class signtaure of a nested annotation and the number of its pairs. For each pair in such a nested annotation, the band method_RVA_nestname_RU transmits the name of the pair.

The bands in the following table transmit appropriately-typed constant pool references for each occurence of specific tag characters.

Tag(s)	Band	Reference
'B','C','I','S','Z'	method_RVA_caseI_KI	cp_Int
'D'	method_RVA_caseD_KD	cp_Double
'F'	method_RVA_caseF_KF	cp_Float
'J'	method_RVA_caseJ_KJ	cp_Long
'c'	method_RVA_casec_RS	cp_Signature
'e'	method_RVA_caseet_RS method_RVA_caseec_RU	cp_Signature cp_Utf8
's'	method_RVA_cases_RU	cp_Utf8

Analogous groups of bands transmit values within the other four method annotation types, and within the visible and invisible annotations of classes and fields.

5.10. Bytecode Instructions

Here are the bands which transmit bytecode instructions:

  bc_bands:
        *bc_codes :BYTE1 [...]
        *bc_case_count :UNSIGNED5 [COUNT(switch,*bc_codes)]
        *bc_case_value :DELTA5 [...]
        *bc_byte :BYTE1 [...]
        *bc_short :DELTA5 [...]
        *bc_local :UNSIGNED5 [...]
        *bc_label :BRANCH5 [...]
        *bc_intref :DELTA5 [...] (cp_Int)
        *bc_floatref :DELTA5 [...] (cp_Float)
        *bc_longref :DELTA5 [...] (cp_Long)
        *bc_doubleref :DELTA5 [...] (cp_Double)
        *bc_stringref :DELTA5 [...] (cp_String)
        *bc_classref :UNSIGNED5 [...] (current class or cp_Class)
        *bc_fieldref :DELTA5 [...] (cp_Field)
        *bc_methodref :UNSIGNED5 [...] (cp_Method)
        *bc_imethodref :DELTA5 [...] (cp_Imethod)
        *bc_thisfield :UNSIGNED5 [...] (cp_Field, only for current class)
        *bc_superfield :UNSIGNED5 [...] (cp_Field, only for current super)
        *bc_thismethod :UNSIGNED5 [...] (cp_Method, only for current class)
        *bc_supermethod :UNSIGNED5 [...] (cp_Method, only for current super)
        *bc_initref :UNSIGNED5 [...] (cp_Field, only for most recent new)
        *bc_escref :UNSIGNED5 [...] (cp_All)
        *bc_escrefsize :UNSIGNED5 [...]
        *bc_escsize :UNSIGNED5 [...]
        *bc_escbyte :BYTE1 [...]
 

In order to transmit bytecode instructions more efficiently, some instructions may be rewritten into a transmission form. The ldc, ldc_w, and ldc2_w bytecodes must be rewritten by the compressor into strongly-typed operations; see below. Some getstatic, putstatic, getfield, putfield, invokevirtual, invokespecial, and invokestatic bytecodes may (at the compressor's option) be rewritten into more specialized and compact forms which, because they are context sensitive, select their operands from a smaller set of possibilities.

In all cases, the first byte of each instruction (perhaps rewritten) is determined by a byte transmitted in bc_codes. All operand bytes are decoded into operands and transmitted in separate bands according to the type of operand encoded. The padding bytes of switch instructions and the last two bytes of invokeinterface instructions are discarded, because they can be reconstructed by the decompressor.

A "wide" prefix bytecode is transmitted in bc_codes. The instruction following the prefix is decoded in its wide format, but (except in the case of iinc) it is transmitted in the same way, and using the same bands, as the normal (non-wide) format. The wide format of the iinc instruction uses bc_short instead of bc_byte to transmit its second operand.

Every branch instruction transmits its target (or targets, in the case of the switches) in the bc_label band, encoded as the difference between the renumbered BCI of the branch target and the renumbered BCI of the branch itself. (The renumbering, as described above, numbers the first instruction as zero, the second as one, and so forth.) Differences between renumbered BCIs are very compact; the primary encoding of BRANCH5 also takes advantage of the fact that most such differences are positive, because most branches are forward branches.

The bc_classref band carries incremented indexes into the cp_Class constant pool. The incrementing (by unity) reserves the code zero, which always refers to the current class. (This provides a a compact form for the most common class reference, which is a self-reference.)

The bc_byte and bc_short bands carry fixed-sized operands. These operands are not transmitted as general 32-bit integers because their instructions, as originally defined, already serve as specially compressed transmission formats. Integer values of these bands are treated as unsigned, even when they are semantically signed as operands. The bc_byte band is used for the last operands of multianewarray and non-wide iinc instructions, and for the operands of bipush and newarray instructions. The bc_short band is used for the last operands of wide iinc instructions, and for the operands of sipush instructions.

The bands bc_intref, bc_floatref, bc_longref, bc_doubleref, bc_stringref, bc_fieldref, bc_methodref, and bc_imethodref transmit ordinary indexes into the constant pools cp_Int, cp_Float, cp_Long, cp_Double, cp_String, cp_Field, cp_Method, and cp_Imethod, respectively.

For every lookupswitch or tableswitch instruction, the number of cases is transmitted in bc_case_count, and the default label is transmitted in bc_label. Each case value and case target of a lookupswitch is transmitted in bc_case_value and bc_label, respectively. The initial case value of a tableswitch is transmitted in bc_case_value, and each of its case targets is transmitted in bc_label.

If the the first byte of an instruction has a code in the range [202..255], or if an instruction's operands cannot be parsed according to the requirements of this specification, it is a non-standard instruction. Every byte of a non-standard instruction must be transmitted in a special envelope in order to warn the decompressor to accept it literally. The envelope consists of a series of "byte_escape" and "ref_escape" opcodes in the bc_code band, corresponding sizes in bc_escsize and bc_escrefsize, bytes in bc_escbyte, and constant pool references in bc_escref. Every constant pool reference in a non-standard instruction must be escaped and transmitted in the bc_escref band. Each ref_escape wraps one such constant pool reference, of size up to four bytes. The transmitted reference is an index into cp_All, such that zero refers to the first CONSTANT_Utf8 constant, and so on.

Escape Opcode	Size Operand	Operand Size	Size Band	Data Band	Opcode Value
byte_escape	N<=255	N bytes	bc_escsize	bc_escbyte	254
ref_escape	N<=2	N bytes	bc_escrefsize	bc_escref	253

There is no need for special transmission of any other operand type within non-standard instructions, because the Pack200 format exactly preserves all instruction bytes except for constant pool references. This specification does not address the means by which compressors adapt to the presence and format of non-standard instructions. It simply requires decompressors to decode them correctly.

Here is a table of the specially rewritten transmission forms for instructions:

Original Instruction	Operand	Transmission Instruction	Rewrite Required?	Opcode
ldc	cp_String[i]	aldc	yes	18 (=ldc)
ldc	cp_Class[i]	cldc	yes	233
ldc	cp_Int[i]	ildc	yes	234
ldc	cp_Float[i]	fldc	yes	235
ldc_w	cp_String[i]	aldc_w	yes	19 (=ldc_w)
ldc_w	cp_Class[i]	cldc_w	yes	236
ldc_w	cp_Int[i]	ildc_w	yes	237
ldc_w	cp_Float[i]	fldc_w	yes	238
ldc2_w	cp_Long[i]	lldc2_w	yes	20 (=ldc2_w)
ldc2_w	cp_Double[i]	dldc2_w	yes	239
getstatic	(this class member)	getstatic_this	no	202
putstatic	(this class member)	putstatic_this	no	203
getfield	(this class member)	getfield_this	no	204
putfield	(this class member)	putfield_this	no	205
invokevirtual	(this class member)	invokevirtual_this	no	206
invokespecial	(this class member)	invokespecial_this	no	207
invokestatic	(this class member)	invokestatic_this	no	208
aload_0; getstatic	(this class member)	aload_0_getstatic_this	no	209
aload_0; putstatic	(this class member)	aload_0_putstatic_this	no	210
aload_0; getfield	(this class member)	aload_0_getfield_this	no	211
aload_0; putfield	(this class member)	aload_0_putfield_this	no	212
aload_0; invokevirtual	(this class member)	aload_0_invokevirtual_this	no	213
aload_0; invokespecial	(this class member)	aload_0_invokespecial_this	no	214
aload_0; invokestatic	(this class member)	aload_0_invokestatic_this	no	215
getstatic	(super class member)	getstatic_super	no	216
putstatic	(super class member)	putstatic_super	no	217
getfield	(super class member)	getfield_super	no	218
putfield	(super class member)	putfield_super	no	219
invokevirtual	(super class member)	invokevirtual_super	no	220
invokespecial	(super class member)	invokespecial_super	no	221
invokestatic	(super class member)	invokestatic_super	no	222
aload_0; getstatic	(super class member)	aload_0_getstatic_super	no	223
aload_0; putstatic	(super class member)	aload_0_putstatic_super	no	224
aload_0; getfield	(super class member)	aload_0_getfield_super	no	225
aload_0; putfield	(super class member)	aload_0_putfield_super	no	226
aload_0; invokevirtual	(super class member)	aload_0_invokevirtual_super	no	227
aload_0; invokespecial	(super class member)	aload_0_invokespecial_super	no	228
aload_0; invokestatic	(super class member)	aload_0_invokestatic_super	no	229
invokespecial	(this class <init>)	invokespecial_this_init	no	230
invokespecial	(super class <init>)	invokespecial_super_init	no	231
invokespecial	(new class <init>)	invokespecial_new_init	no	232

Every bytecode instruction is contained by a class, called the current class. The superclass (if any) of the current class is the current super class. The operand of the textually most recent "new" instruction (in the same method) is called the current new class.

If an instruction refers to a field or method in the current class, it may (at the compressor's option) be rewritten for transmission as a corresponding opcode spelled with "_this". Likewise, an instruction referring to a field or method in the current super class may be rewritten as a corresponding opcode spelled with "_super". In either case, if the immediately preceding instruction is aload_0 (opcode 42), the transmission of that instruction may be suppressed by the compressor, and the corresponding opcode spelled with "aload_0_" selected instead; otherwise the "aload_0_" variant may not be selected.

If an invokespecial instruction refers to a method named <init> in the current class, the current super class, or the current new class, the compressor may choose to rewrite it as an invokespecial_this_init, invokespecial_super_init, or invokespecial_new_init, respectively.

Field (resp. method) operands of rewritten instructions spelled with "_this" (but not "_init") are transmitted in the special band bc_thisfield (resp. bc_thismethod). The numbering of these operands is defined by taking the sequence of symbols in cp_Field (resp. cp_Method) and selecting only members of the current class. The resulting subset, without changing its order, is renumbered starting with zero. This provides a compact mapping of small integers to members of the current class.

Likewise, operands of rewritten instructions spelled with "_super" (but not "_init") are transmitted in the bc_superfield or bc_supermethod band, and renumbered as a subset of cp_Field or cp_Method, selecting only members of the current super class.

Finally, operands of the rewritten instructions spelled with "_init" are transmitted in the band bc_initref, and renumbered as a subset of cp_Method, selected according to the appropriate class (current, current super, or current new), and also selected to have the name <init>. (The index transmitted is usually very small; in effect it selects only the signature of the invoked method, and most classes possess only a few constructors.)

Here is a table summarizing the transmission of instructions of more than one byte.

Instruction	Operand	Transmitted Value	Band
bipush	(byte) x	x & 0xFF	bc_byte
sipush	(short) x	x & 0xFFFF	bc_short
ildc	cp_Int[i]	i	bc_intref
fldc	cp_Float[i]	i	bc_floatref
aldc	cp_String[i]	i	bc_stringref
cldc	current class	0	bc_classref
cldc	cp_Class[i]	i+1	bc_classref
ildc_w	cp_Int[i]	i	bc_intref
fldc_w	cp_Float[i]	i	bc_floatref
aldc_w	cp_String[i]	i	bc_stringref
cldc_w	current class	0	bc_classref
cldc_w	cp_Class[i]	i+1	bc_classref
lldc2_w	cp_Long[i]	i	bc_long
dldc2_w	cp_Double[i]	i	bc_double
*load	locals[i]	i	bc_local
*store	locals[i]	i	bc_local
ret	locals[i]	i	bc_local
iinc	locals[i]	i	bc_local
iinc (not wide)	(byte) x	x & 0xFF	bc_byte
iinc (wide)	(short) x	x & 0xFFFF	bc_short
if**	pc	(delta/renumbered)	bc_label
if_**	pc	(delta/renumbered)	bc_label
goto	pc	(delta/renumbered)	bc_label
jsr	pc	(delta/renumbered)	bc_label
goto_w	pc	(delta/renumbered)	bc_label
jsr_w	pc	(delta/renumbered)	bc_label
tableswitch	case count	count	bc_case_count
tableswitch	default pc	(delta/renumbered)	bc_label
tableswitch	first case value	value	bc_case_value
tableswitch	each case pc	(delta/renumbered)	bc_label
lookupswitch	case count	count	bc_case_count
lookupswitch	default pc	(delta/renumbered)	bc_label
lookupswitch	each case value	value	bc_case_value
lookupswitch	each case pc	(delta/renumbered)	bc_label
new	current class	0	bc_classref
new	cp_Class[i]	1+i	bc_classref
newarray	type code	value	bc_byte
anewarray	current class	0	bc_classref
anewarray	cp_Class[i]	1+i	bc_classref
checkcast	current class	0	bc_classref
checkcast	cp_Class[i]	1+i	bc_classref
instanceof	current class	0	bc_classref
instanceof	cp_Class[i]	1+i	bc_classref
multianewarray	cp_Class[i]	1+i	bc_classref
multianewarray	rank	rank & 0xFF	bc_byte
getstatic	cp_Field[i]	i	bc_fieldref
putstatic	cp_Field[i]	i	bc_fieldref
getfield	cp_Field[i]	i	bc_fieldref
putfield	cp_Field[i]	i	bc_fieldref
invokevirtual	cp_Method[i]	i	bc_methodref
invokespecial	cp_Method[i]	i	bc_methodref
invokestatic	cp_Method[i]	i	bc_methodref
invokeinterface	cp_Imethod[i]	i	bc_imethodref
**_this	this_fields[i]	i	bc_thisfield
**_this	this_methods[i]	i	bc_thismethod
**_super	super_fields[i]	i	bc_superfield
**_super	super_methods[i]	i	bc_supermethod
invokespecial_this_init	this_constructors[i]	i	bc_initref
invokespecial_super_init	super_constructors[i]	i	bc_initref
invokespecial_new_init	new_constructors[i]	i	bc_initref

6. Specification of Band Coding

Each integer is coded for transmission as a byte sequence, using between one and five bytes, according to a previously determined encoding. The encoding may by a primary encoding for the current band as specified as part of this Pack200 archive format specification, or the encoding may depend on information transmitted before the occurrence of the encoded integer.

(Note: In this account of encodings, bytes are indivisible octets which express unsigned values in [0,255].)

6.1. Encoding of Small Whole Numbers

6.1.1. Scheme of Multiple Codings

Given any two values B and H, there is a coding (B,H) which establishes a one-to-one correspondence between an initial sequence of non-negative integers, and an "encoding sets" of short sequences of bytes.

Moreover, no byte sequence in the encoding set is a proper prefix of another byte sequence in the same set. prefix. This means, informally, that encodings are self-sizing, or "parseable".

Also, the encoding set is as full as possible, so that every long-enough sequence of bytes begins with a unique prefix of bytes drawn from the encoding set. In particular, every sequence of B bytes has a (unique) prefix in the encoding set for the (B,H) coding.

6.1.2. Definition of Encoding Byte Sequences

Given a (B,H) coding, a sequence of bytes is in its encoding set if and only if all of the following are true:

• the sequence contains at least one and no more than B bytes
• all bytes except the last are "high"
• either the last byte is "low", or else the sequence has B bytes

As a consequence of this definition, the encoding set be regarded as satisfying this regular expression, in terms of high and low bytes:

  encoding_set = (high* low) | (high)^B
 

(Note: In this specification the up-caret sign '^' denotes exponentiation, either of regular expressions as above, or of numbers.)

If n is less than B then the number of (B,H) encodings of length exactly n is (H^(n-1) * L). The number of (B,H) encodings of length exactly B is (H^(B-1) * (L+H)). The sum of these values for all n in [1..B] is the total size of the encoding set for a (B,H) coding. It is called Card(B,H) and is therefore defined as:

  Card(B,H) = (L * (1-H^B)/(1-H)) + H^B
    if H>1, or else
  Card(B,1) = B*255+1
 

If H is 256, there are no low bytes, and the encoding set consists of all possible sequences of exactly B bytes, and Card(B,H) is 256^B.

Otherwise, the encoding set consists of sequences of various lengths, from 1 to B inclusive. In particular, there are L one-byte sequences. We will use these elsewhere to encode the "favored" integers of the highest frequency. Note that the encoding rules specified elsewhere in this document are designed to produce smaller integers more frequently than larger ones.

Note that L can be viewed as a "sharpness" parameter, expressing the sharpness of the distribution about zero of the expected values to be encoded. The larger L is, the more the encoding favors distributions of values which are close to zero, and rarely far from it. Larger H values favor "flatter" distributions, up to H=256, L=0, which favorably encodes completely random data within the range of the coding.

6.1.3. Definition of Decoded Whole Number Values

Given the sequence of N byte values b[0] .. b[N-1], the sorting-based definition given above implies that the decoded integer value of this byte sequence is the little-endian base-H scaled sum of the bytes, and is called Decode(B,H; b):

  Decode(B,H; b[*]) = Sum[0<=i<N]( b[i] * H^i )
 

As a consequence of this definition, each byte sequence in the encoding set corresponds uniquely to an arithmetic integer value in that range. This correspondence may be observed by ordering the byte sequences first by length, and second according to their antilexicographic (little-endian) order. Each element in resulting sequence decodes into its own zero-based index within the sequence.

Note that the number zero is always encoded by a sequence of B zero bytes (if H is 256) or else by single zero byte.

The largest encoded arithmetic value is always encoded by a sequence of B bytes of value 255.

Note that the (B,H) encodings (1,256), (2,256), and (4,256) are identical with the conventional little-endian representations of unsigned bytes, unsigned 16-bit integers, and unsigned 32-bit integers.

Note that some may represent arithmetic values greater than 2^31-1 or even 2^32-1. This is a feature of the (B,H) coding scheme. (But sometimes intermediate 64-bit values are required when multiple sign bits are being used, as described below. The rules given below for signed values require the compressor to emit the simplest encoding that will suffice to transmit the required 32-bit integer, even if more complex sequences would produce the same 32-bit value, after truncation.)

Note that implementations are not always free to truncate intermediate values when computing (B,H) coding values. However, the final decoded values (after restoration of signs) will always fit in 32 bits, as described in the next section.

6.2. Encoding of Signed Integers

The whole number U obtained from a (B,H) coding is converted to a signed 32-bit value X in a manner which depends on the parameter S. This process is called sign conversion.

Informally, sign conversion is performed by 32-bit truncation if S=0. Otherwise, the low-order bits of U are collectively treated as a sign bit, so that X will be negative only if all sign bits are set. The two conversions from U to positive X and U to negative X are independently dense and monotonic in opposite directions, as defined below.

Unlike the (B,H) coding, a (B,H,S) coding represents only numbers in a specifically defined range, which is never larger than [-2^31..2^31-1].

In what follows, for each (B,H,S) coding, we define Range(B,H,S), and the method used to represent 32-bit signed values with that coding.

A (B,H) encoding byte sequence is not a legal input to a corresponding (B,H,S) coding if it represents a whole number U which does not convert into the range of the (B,H,S) coding. Also, if there are two byte sequences encoding values U1 and U2 which both map into the same point X in the signed range, only the byte sequence for the smaller of U1 and U2 is legal.

Thus, a (B,H,S) coding may have a smaller cardinality than the corresponding (B,H) coding, due to illegal encoding byte sequences.

The range of a (B,H,S) coding contains all 32-bit signed integers [-2^31..2^31-1] if Card(B,H) is 2^32 or more. (Negative values are represented by 31-bit unsigned overflow.) If the coding encodes fewer unsigned values, the top of its range is clipped to 2^31-1, and the bottom of its range is clipped to -2^31.

More precisely, the range of a (B,H,S) coding is called Range(B,H,S) and is defined (partially, for S=0) as:

  Range(B,H,S) = [-2^31..2^31-1]
    if S=0, Card(B,H) >= 2^32, or
  Range(B,H,S) = [0..2^31-1]
    if S=0, 2^31 < Card(B,H) < 2^32, or else
  Range(B,H,S) = [0..Card(B,H)-1]
    if S=0, Card(B,H) <= 2^31
 

(Note that decompressors can produce arbitrary answers for out-of-range codings, since compressors are not allowed to emit them. This allows decompressors to use 32-bit arithmetic for most calculations, ignoring the possibility of undesired wraparond.)

If S>0, the decoding of a byte sequence is performed first by decoding it as a (B,H) coding, and then by converting the decoded whole number into a signed 32-bit number. This sign conversion depends only on the arithmetic value U obtained from the (B,H) coding, and on the value of S. The arithmetic value must be preserved beyond 32 bits of precision, if the sign conversion will produce a value within the range of the coding. If the cardinality of the (B,H) coding is 2^32 or more, the (B,H,S) coding produces all possible signed 32-bit negative values, by 32-bit truncation of the intermediate unsigned whole number. The intermediate unsigned arithmetic can be carried out (with some care) using 32-bit unsigned numbers, or it can be carried out on 64-bit signed or unsigned numbers.

More precisely, the sign conversion operation SignConvert(S;U) is defined as:

  SignConvert(S; U) = Cast32(U)
    if S==0; or
  SignConvert(S; U) = U - Floor(U / 2^S)
    if S>0, (U % 2^S) < 2^S-1, Card(B,H) < 2^32
  SignConvert(S; U) = Cast32(U - Floor(U / 2^S))
    if S>0, (U % 2^S) < 2^S-1, Card(B,H) >= 2^32
  SignConvert(S; U) = -Floor(U / 2^S)-1
    if S>0, (U % 2^S) == 2^S-1
 

  Cast32(U) = ((U + 2^31) mod 2^32) - 2^31
 

6.2.1. Further Discussion of Sign Conversion

As a consequence of the definition of SignConvert(S;U), the S low-order bits of U function as a sign bit. In effect, U is partitioned into a sign field U0 (consisting of the S low-order bits) and a significand U1 (consisting of all other bits of U).

In this alternative view of the definition of SignConvert, if all S bits of U0 are set, then the decoded signed value is the ones complement of U1. (Since this includes the indefinite number of zero high-order bits of U, the resulting value is arithmetically negative.)

Otherwise, if some of the S bits of U0 are clear, then the significand U1 is scaled by the number of possible values of U0 (i.e., (2^2)-1 if S is 2) and U0 is added back in.

This provides a one-to-one correspondence between some portion of the arithmetic interval [0..Card(B,H)-1] and Range(B,H,S).

The integers "favored" by this encoding are those of small magnitude, but they can have either sign. If S is one, the encoding has a balanced number of positive and negative favored values. If S is two (or more), the encoding is skewed toward positive numbers, but also favors a few negative numbers. We use such codings elsewhere to encode values, such as branch displacements or deltas of mostly-sorted data, which are usually positive but sometimes negative.

Note that if S=1, this definition of sign conversion is equivalent to an unsigned right shift followed by an exclusive-or of the sign bit with all remaining bits:

  SignConvert(1; U) = (U >>> 1) ^ -(U & 1)
 

For S>0, Range(B,H,S) is simply defined as the intersection of the 32-bit signed range [-2^31..2^31-1] with the set SignConvert(S; [0..Card(B,H)-1]), obtained by element-wise application of sign conversion to every arithmetically representable value of (B,H). Therefore, we can complete the definition of Range as follows:

  Range(B,H,S) = [-2^31..2^31-1]
    if Card(B,H) >= 2^32, or
  Range(B,H,S) = [0..2^31-1]
    if S=0, 2^31 < Card(B,H) < 2^32, or else
  Range(B,H,S) = [0..Card(B,H)-1]
    if S=0, Card(B,H) <= 2^31

  Range(B,H,S) = [max(  -2^31, min SignConvert(S; Range(B,H,0)) )
               .. min( 2^31-1, max SignConvert(S; Range(B,H,0)) )]
 

(Note: When computing the bounds of a signed encoding (B,H,S), it is useful to start with the largest unsigned value M of (B,H,0) and perform sign-conversion on M, M-1, M-2, etc., noting the first positive and the first negative value to be encountered. These will be the inclusive bounds of the signed encoding's range.)

6.3. Attributes of Codings

  Card(B,H) = (L * (1-H^B)/(1-H)) + H^B
 

The cardinality of a (B,H,S) coding is determined by that of its range:

  Card(B,H,S) = Card Range(B,H,S) <= Card(B,H)
 

As defined, the range of any coding (B,H,S) is dense about zero and can therefore be expressed as a closed interval:

  Range(B,H,S) = [Min(B,H,S)..Max(B,H,S)]
  -2^31 <= Min(B,H,S) <= 0
  0     <  Max(B,H,S) <= 2^31-1
 

Certain additional attributes can be predicated of codings. A coding can either be signed or not, and it can be a full-range coding, or a sub-range coding (or neither).

A (B,H,S) coding is signed if it can encode at least one negative value. This is true if and only if S is non-zero or S is zero and Card(B,H) is 2^32 or more.

A (B,H,S) coding is full-range if Range(B,H,S) is [-2^31..2^31-1].

A sub-range coding delivers less than 2^31 distinct values. More precisely, a coding (B,H,S) is sub-range if Card(B,H,S) <= 2^31-1. (Some codings are neither full-range nor sub-range codings.)

As a consequence of these definitions, any coding for which B<=3 is a sub-range coding. Longer codings can also be sub-ranges if they are "sharp" enough. Some examples are (4,192,0), (5,32,0), and (5,32,1).

Also, the unsigned coding (4,256,0) is full-range, as is the signed version (4,256,1), but not the doubly-signed version (4,256,2).

The coding (4,255,0) is neither a sub-range nor a full-range coding, since its cardinality is 2^31. It represents only non-negative 32-bit integers, but its cardinality cannot be so represented.

6.4. Encoding of Correlated Sequences

If D is zero, no differencing is performed, and the (B,H,S,0) coding is in all ways identical to the corresponding (B,H,S) coding. If D is one, the values are encoded in terms of their successive differences.

(Note: Sequences which are mostly monotonically ascending will often be favorably encoded using unsigned delta codings of the form (5,H,0,1).)

Suppose a sequence of values X[i] is given, for some range of i in [0..N-1]. This sequence will be representable by a (B,H,S,1) coding only if the (B,H,S) coding is a sub-range or a full-range coding. (If it is neither, there is no legal (B,H,S,1) coding.)

The sequence X[i] will be representable, only if there is a series of "delta values" D[i] whose partial sums Sum[j<=i](D[j]) can represent the values X[i].

Each (B,H,S,1) coding has a range, in which all X[i] values must lie. This range is defined as Range(B,H,S,1) = Range(B,H,0). The range of a delta coding (B,H,S,1) contains negative numbers only if it is based on a full-range coding (B,H,S). Otherwise, the delta coding can only represent non-negative numbers. In practice, this is not a severe restriction, since in practice very few band elements are negative.

The partial sums Sum[j<=i](D[j]) are allowed to leave the range, but the final X[i] values are always brought back into range by adding or subtracting a multiple of Card Range(B,H,0).

More precisely:

  X[i] = Sum[j<=i](D[j]) mod Card(B,H,0)
    if (B,H,S) is sub-range
  X[i] = (int32) Sum[j<=i](D[j])
    if (B,H,S) is full-range
 

Note that implementations can simply ignore 32-bit wraparound when computing partial sums, if the coding is full-range. Otherwise, more care must be taken to bring partial sums back into range by subtracting or adding multiples of the range cardinality. Also, 32-bit wraparound may be a hazard for ranges of size 2^30 or more. It is possible to implement this using only 32-bit signed arithmetic.

6.5. Encodings of Uncorrelated Values

The compressor may elect to use a population-based coding transformation in such cases. This transformation takes a sequence S of N arithmetic values and converts it into three value sequences:

• F a table (arbitrary length K+1 << N) of favored values
• T a sequence (length N) of value tokens
• U a sequence (length << N, depends on T) of unfavored values

Each of these sequences (F, T, and U) is in turn treated as a sub-band, and transmitted with an independent encoding. Each non-zero value in T indexes a value from F, while each zero value in T selects (in order) a value from U:

  S[i] = F[T[i]-1]
    if T[i] != 0
  U = { S[i] such that T[i] == 0 }
 

6.5.1. Table of Favorites

The "central value" X of F is defined as that element of F which is arithmetically closest to zero. If F contains both a positive X and a negative -X of minimal absolute value, then -X is defined as the central value. (I.e., a negative sign is the tie-breaker. The central value is chosen so as to have a favorable encoding.)

(Note that implementations may compare values for centrality by using the 32-bit signed int expression (X>>31)^(X<<1) as an unsigned comparison key.)

A valid sentinel value is either the central value of F, or the last value of F. Thus, while parsing F, the decompressor must look for a repetition of either the central value (so far) or of the immediately previous value.

Let K be the number of values in F, ignoring the repetition of the sentinel value. In the next band, values in [1..K] will refer (as 1-origin indexes) to corresponding elements of F.

6.5.2. Sequence of Tokens

Each element of T is a token for either a favored or an unfavored value, encoded in the arithmetic range [0..K]. Each non-zero value corresponds in order to an element of F, and produces that favored value. Each zero value in T is a placeholder for an "unfavored" value, which is still unknown.

As mentioned above, each element of F must be referred to by at least one element of T. That is, the following two sets must be the same:

  { F[T[i]-1] such that T[i] != 0 }
  { F[j] }
 

The statistics of the sequence T are quite favorable for compact encoding in a suitable (B,H) scheme, assuming that the coder made good choices for the contents of F. Generally speaking, the most commonly occurring input values should be given early positions in F.

A greedy algorithm could sort the input values by number of occurrences and place the most common values at the front of F. This would be likely to produce an improved encoding for the input. Such techniques are neither mandated nor discouraged by this specification.

6.5.3. Sequence of Unfavored Values

Taken together, the elements of F and U selected by the elements of T must be identical in value and order to the input of the population transformation.

The decompressor can then read F into memory, read T into memory, and then make a second pass over T, translating token values, either referring to F or reading from U as necessary.

6.6. Adaptive Encodings

An adaptive coding method handles this situation by allowing value sequences to be partitioned (effectively into sub-bands), with an independent coding used locally in each part.

An adaptive coding method is specified by a count K, a coding method A which will be used to encode K values, and another coding method B which will handle the rest of the values.

Since bands are sized by context, when an adaptive coding method is applied to the bytes of an encoded band, the method will be provided in advance with the count N of values to decode. Thus, after the method A has decoded K values, the method B will be used to decode the rest of the (N-K) values in the band.

6.7. Meta-Coding

Any compressor may elect not to provide a band coding specifier, in which case the band's primary encoding governs the transmission of elements. If the encoding of the first band element under the primary encoding accidentally appears to be a band coding specifier, the compressor is obligated to transmit an explicit band coding specifier that reaffirms the band's primary encoding. This is rare, because (as noted below) band coding specifiers present themselves as negative numbers in the primary encoding, and negative numbers are rare in most bands.

6.7.1. Coding Specifier Structure

  BandCodingSpecifier:
        (Default | BHSDCode | RunCode | PopCode)
  BHSDCode:
        CanonicalBHSDCode | ArbitraryBHSDCode
  CanonicalBHSDCode:
        ( '(1,256,0)' | '(1,256,1)' | ... )
  ArbitraryBHSDCode:
        'arb' ( B H S D )
  B, H, S, D: 
        Integer
  RunCode:
        'run' K ACode BCode
  K:
        Integer
  ACode:
        (Default | BHSDCode | PopCode)
  BCode:
        (Default | BHSDCode | RunCode | PopCode)
  PopCode:
        'pop' ( FCode TCode UCode )
  FCode:
        (Default | BHSDCode | RunCode)
  TCode:
        (Default | BHSDCode | RunCode)
  UCode:
        (Default | BHSDCode | RunCode)
  Integer:
        ( '0' | '1' | ... )
  Default:
        'default'
 

Also, population coding methods can contain adaptive coding methods, and vice versa. However, although the grammar does not show this fact, population coding methods must not be nested, even indirectly.

Not all possible integer values of K and L are equally well expressible. The intended values of K are commensurate with compressor windows, and the intended values of L are sharpness parameters for BHS encodings.

6.7.2. Coding Specifier Semantics

• N, an expected number of band values to decode, or "infinity"
• D, a default BHSD coding
• S, a coding specifier

We denote the application of S in the context of N and D as "S(N,D)". This application decodes up to N elements of transmitted band data.

Just before a band is received, the decompressor knows an expected band length N and a default coding method D, specified statically as the band's primary encoding. The decompressor then reads zero or more bytes which constitute a band encoding specifier, and decodes N elements of band data based on the band encoding specifier.

When the coding specifier is 'default', N values are decoded using the default coding D, which is the band's primary encoding. (This rule is necessary if the band's first element appears to introduce a non-empty coding specifier. Such a default coding specifier is the only kind that a compressor is obligated to emit; the rest are all optional.)

When the coding specifier is another BHSDCode, N values are decoded using that coding. The ambient default D is ignored.

When the coding specifier is a 'run' of K, ACode, BCode, two steps are taken. First, the specifier ACode is given control, as ACode(K,D). That is, N is temporarily set to K. Second, the specifier BCode is given control as BCode(N-K,D). (Note: If N is infinity, N-K will also be infinity.) In both steps, D remains unchanged, so that occurrences of 'default' for ACode or BCode cause D to be used.

It is illegal for a 'run' coding specifier to specify a K of zero, or for it to be used to decode a run of K or fewer values.

When the coding specifier is a 'pop' of FCode, TCode, UCode, three steps are taken. First, FCode is applied as FCode(infinity, D). The decoding of the F values stops when the first duplicate value is encountered. (As specified in the section above on population coding, that duplicate value acts as a sentinel, and must be either the most "central" value read, or else the value immediately previous to the sentinel value.) Let K be the number of unique decoded values.

During the decoding of the F values by FCode, every 'run' specifier in FCode must exhaust its count 'K' without decoding the sentinel value. That is, the sentinel must be decoded by the last simple BHSD coding in FCode.

In the second step, a different coding TCode is used, since the T values are expected to be a dense encoding. TCode is applied as TCode(N,D), and the tokens are read which determine the band values to be delivered. (Note that a 'pop' coding method can only be applied if the N value supplied is finite. This will be true, since the only bands which are read with an infinite N are byte bands such as bc_codes.) Let Z be the number of zeroes decoded using TCode.

The third step is to use UCode to decode Z values, using the original default encoding D. UCode is applied as U(Z,D). The resulting band is produced from the T sequence, as documented above, by replacing zeroes in the T sequence by successive U values, and replacing other elements of the T sequence by those indexed in the F sequence.

It is illegal for a 'pop' coding specifier to be used to decode a run of no values, or of an infinite number of values. If Z (the number of unfavored values) is zero, it is illegal for UCode to be anything other than 'default'.

6.7.3. Coding Specifier Meta-Encoding

The common band coding specifier of 'default' is encoded in a way that often requires no extra bytes. If the band has no elements, no parsing at all is done. Otherwise, if the band's default encoding is of variable length, the decompressor is obligated to decode one value X from the initial bytes of the band using the band's default coding D. If possible, the value X is converted into an unsigned byte value XB, which is taken to be the first byte of a band coding specifier, and X will be discarded. Otherwise, the band coding specifier is taken to be 'default', and so D is applied to the entire band, including the bytes which produced the initial value X.

D must be of the form (B,H,S) or (B,H,S,D), where B>1 and H<256. The value is D is irrelevant to the decoding of the first value, and the decoding of X is done using (B,H,S,0). If S is not zero, and if the value X is in the range [-256..-1], the coding specifier byte XB is defined as XB = (-1-X), and X is discarded. Otherwise, if S is zero, and if the value of X is in the range [(256-H)..(511-H)], the coding specifier byte XB is defined as XB = (X-(256-H)), and X is discarded. Otherwise, there is not a coding specifier byte XB, and X is the first band value, as described above.

The value XB, if produced, is taken to be the initial byte of a coding specifier preceding the actual band data. This specifier is parsed, starting with XB, and continuing (if necessary) with bytes bytes taken from band_headers. The parsed specifier is then used to control decoding, starting at the next byte, of all band elements.

If the band really does start with an element X1 in a range ([-256..-1] or [L,L+255]) that requires the production of a band specififer byte XB, but the compressor wishes to use the default encoding, the compressor is required to specify a band header, lest X1 mislead the decompressor. If the compressor wishes to keep the default coding, it must in such cases specify an explicit coding specifier of 'default', using an XB value of zero (as specified below). This zero XB is transmitted, depending on the default encoding, as an X value of -1 (usually the byte 1) or of L (usually the bytes 192,0).

(An immediate consequence of this design is that byte bands can never have non-default encodings, but all other bands can, since all the other bands have default encodigns which are variable length and have a large dynamic range. The "escape" values chosen for X are rare in practice, so they are not usually confused with real band data. Adding an extra zero value to reaffirm the default will always cost one extra byte before the actual band data, and no extra bytes in the band_headers band. In general, the encoding of explicit X values will always require two bytes if S is zero, and at most two bytes if S is not zero.)

The band band_headers transmits the non-initial band header bytes (if any) of each band that has a band header. The order of transmission is consistent with the overall order of all bands in the archive, as given in the present specification's band grammar. The length of band_headers is independently specified (as #band_headers_size) in the archive header. (It should be clear that the decompressor needs to find the end of band_headers before it can read any further bands.)

Thus, the encoding of a coding specifier consists of a sequence of bytes: an initial byte XB, and zero or more non-initial bytes taken from band_headers. The encodings of the little language specified in the previous section are as follows. The encoding 'default' is a zero byte. Of all the possible BHSDCode encodings, a sequence of 115 canonical encodings (defined below) is selected to cover the expected range of uses. A BHSDCode is represented as a single byte whose value is the index (1-based) of the selected canonical coding.

  Enc{default}  = (0)
  Enc{CanonicalBHSDCode} = value in [1..115]
 

The special value 116 introduces an arbitrary, possibly non-canonical BHSD code, described in the following byte.

  Enc{ arb ( B H S D ) } =
      116
    & (D:[0..1] + 2*S[0..2] + 8*(B:[1..5]-1))
    & (H[1..256]-1):
 

The 'run' construct encodings begin with a byte in [117..140], and may be followed by an optional byte KB and then by one or two encoding specifiers ACode and BCode. The offset from 117 represents bitwise the following data:

• a two-bit field KX
• a bit KBFlag
• a bit ADef (ABDef == 1)
• a bit BDef (ABDef == 2)

  Enc{ run ( K ACode BCode ) } =
      (117 + (KX:[0..3]) + 4*(KBFlag:[0..1]) + 8*(ABDef:[0..2]))
    & KB: one of [0..255] if KBFlag=1
    & Enc{ ACode } if ADef=0  (ABDef != 1)
    & Enc{ BCode } if BDef=0  (ABDef != 2)
 

After the leading byte is parsed, a byte KB is expected if the KBFlag is set, otherwise KB is implicitly given the value 3.

The K value for the 'run' construct can take a range of values, and is determined as follows:

  K = (KB+1) * 16^KX
 

The ACode coding is understood to be 'default' if the ADef bit is set. Otherwise, the representation of ACode is parsed next. Finally, the BCode coding is understood to be 'default' if the BDef bit is set. Otherwise, the representation of BCode is parsed last. Both ADef and BDef may be clear, but both may not be set.

The 'pop' construct encodings begin with a byte in [141..188], and may be followed by one, two, or three encoding specifiers, FCode, UCode, and TCode. The offset from 141 bitwise encodes the following data:

• a bit FDef
• a bit UDef
• a bit TDef (TDefL > 0)
• a value L in [1..11] if TDef=1

  Enc{ pop ( FCode TCode UCode ) }
    = (141 + (FDef:[0..1]) + 2*UDef:[0..1] + 4*(TDefL:[0..11]))
    & Enc{ FCode } if FDef=0
    & Enc{ TCode } if TDef=0  (TDefL==0)
    & Enc{ UCode } if UDef=0
 

If TDef is zero, an explicit encoding specifier determines TCode, and the L parameter is not present. If TDef is one, the L parameter is present and is derived from TDefL, according to the following table:

TDefL	L
1	4
2	8
3	16
4	32
5	64
6	128
7	192
8	224
9	240
10	248
11	252

If the L parameter is present, the TCode is derived from the K and L values. If K < 256, then TCode is BYTE1 (1,255,0), and L is ignored. Otherwise, TCode is the (B,H,S) coding where S=0, H=(256-L), and B is the smallest value such that [0..K] is contained in Range(B,H,S). (It is illegal to specify an L so large that Range(5,256-L,0) does not contain K.)

The FCode coding is understood to be 'default' if the FDef bit is set. Otherwise, the representation of FCode is parsed immediately after the initial byte. The TCode coding is understood to be derived from K and L if the TDef bit is set. Otherwise, the representation of TCode is parsed next. Finally, the UCode coding is understood to be 'default' if the UDef bit is set. Otherwise, the representation of UCode is parsed last.

6.7.4. Canonical BHSD Codings

7. Stability of Decompressor Output

However, for any given Pack200 archive, every decompressor is required to produce a particular byte-wise image for each class file transmitted. This requirement is placed on decompressors in order to make it possible for compressors to transmit information, such as message digests, which relates to the eventual byte-wise contents of transmitted class files. This section describes the restrictions placed on every decompressor that makes the byte-wise contents of its output files a well-defined function of its input.

In general, the order of elements in a decompressed class file must be consistent with the transmission order in the Pack200 archive. For example, the order of a class's fields declared in the class file must correspond to the order in which that class's field descriptors were transmitted in the field_descr band. (This also happens to correspond to the order in the field_flags bands.) This table gives all of the required correspondences of class file order with archive transmission order:

element in class file	band which determines order
implemented interfaces	class_interface
declared fields	field_descr
declared methods	method_descr
code handler list	code_handler_start_P
class attribute list	class_flags, class_attr_indexes
field attribute list	field_flags, field_attr_indexes
method attribute list	method_flags, method_attr_indexes
code attribute list	code_attr_indexes
constant pool entries	cp_Utf8, etc. (see below)

Together, these required correspondences of order determine the exact contents of the decompressed class file. The ordering of interfaces, fields, methods, and exception handlers is directly determined by the ordering of the bands that transmit them.

7.1. Ordering of Attribute Lists

If an InnerClasses attribute must be added to a class, under the rules given in the next section, that attribute must come last in the class's attribute list.

7.2. Ordering of Constant Pools

• cp(X) is not yet defined,
• there is no InnerClasses attribute in X (even if was an ic_Local(X) transmitted), and
• the locations of constant references within X are noted but not initialized with numbers.

Then cp(X) is defined as if by the following steps, which populate it from the global constant pool cp_All, and also potentially produce an InnerClasses attribute for X.

• cp(X) is initially empty.
• Every constant referenced from X is added to cp(X).
• In the next step, any constant from cp_Signature must be treated as making no additional references, despite its transmitted structure.
• Every constant referenced from cp(X) is added to cp(X), if it is not already there. (For example, a constant from cp_Method refers to constants from cp_Class and cp_Descr.)
• The previous step is iterated until no further changes occur. (There will be at most two more of these steps.)

At this point the constant pool cp(X) is well enough defined to allow the decompressor to compute the set of relevant nested classes, called ic_Relevant(X).

• ic_Relevant(X) is initially empty.
• Every four-tuple from ic_All which mentions the current class (defined by X) as an outer class is added to ic_Relevant(X).
• For every class constant K referenced from both ic_this_class and cp(X), the corresponding four-tuple from ic_All is selected, and added to ic_Relevant(X), if not already there.
• The previous step is repeated until closure. (That is, until ic_Relevant(X) no longer changes.)
• The set ic_Relevant(X) is ordered into a sequence, so that its order is consistent with ic_All. (I.e., it is a subsequence of ic_All.)

With ic_Relevant(X) and an optionally transmitted ic_Local(X) in hand, the decompressor must next decide whether to store an InnerClasses attribute ic_Stored(X) for X, using these steps.

• If ic_Local(X) is present and empty, no InnerClasses attribute will be stored for X, and none of the following steps will be taken.
• Also, if ic_Local(X) is not present and ic_Relevant(X) is empty, no InnerClasses attribute will be stored for X.
• Otherwise, ic_Stored(X) is set to be the elements in ic_Local(X) followed by the elements of ic_Relevant(X) (both in transmission order).
• Next, any elements in both ic_Local(X) and ic_Relevant(X) are removed from ic_Stored(X), thus forming a symmetric difference.
• The resulting sequence ic_Stored(X), even if empty, is added to the decompressor's output as an InnerClasses attribute.
• All cp_Class and cp_Utf8 entries directly or indirectly required by ic_Stored(X), as well as the fixed cp_Utf8 entry spelled "InnerClasses", are added to cp(X) if not already there.

With the last contributions from nested class records, the decompressor finishes the constant pool using these steps:

• Every cp_Signature entry in cp(X) is "flattened" into a cp_Utf8 entry with the same spelling.
• If any cp_Utf8 entry E in cp(X) was not transmitted in cp_All but had an equivalently spelled signature S transmitted, then E is replaced by S in cp(X).
• The elements of cp(X) which are also in cp_All are sorted into the same order as their occurrences in cp_All.
• All cp_Utf8 constants in cp(X) not also in cp_All are moved to the end of cp(X), and are sorted in the order defined by String.compareTo.
• Next, any cp_Class constants in cp(X) not also in cp_All are moved to the end of cp(X), in the order defined by String.compareTo on their class names.
• (At this point, the decompressor has uniquely determined a position for every element of cp(X).)
• All elements of cp(X) which are referred to via a one-byte index (i.e., a ldc bytecode) are moved in mass to the beginning of cp(X), without changing their relative order.
• All elements of cp(X) are transformed into their equivalent class file constant types, with signatures being "flattened" into equivalently spelled CONSTANT_Utf8 constants.
• The resulting sequence of class file constants, cp(X), is stored as the constant pool of X.

After this process, the stored constant pool of X has referential integrity, and all constant references can be coded as indexes into cp(X), which has a definite order derived from cp_All. In particular, if a constant in cp(X) needs to refer to another constant, that constant must already have been inserted into cp(X), during the closure steps.

Note that the ordering of cp(X) is consistent with that of cp_All, except that some signature references are redirected to equivalent pre-existing elements of cp_Utf8, and all operands of ldc bytecodes are forced to the front of the constant pool.

When assigning local indexes to elements of cp(X), the decompressor must of course respect the reservation of empty constant pool slots for index zero, and for CONSTANT_Long and CONSTANT_Double constants, as required by the class file format. Together with these rules, the construction and ordering of cp(X) completely determines the assignment of concrete indexes to constant references within X.

8. Appendixes

8.1. Appendix: List of Bands

8.2. Appendix: Pseudo-Code Illustrations

8.2.1. Representation of cp_Utf8 Constant Pool

  assert(cp_Utf8[0].equals(""));
  int cursor = 0;
  int big_cursor = 0;
  for (int i = 1; i < cp_Utf8_count; i++) {
    String thisString = cp_Utf8[i];

    int prefix = (i == 1)? 0: cp_Utf8_prefix[i-2];
    int suffix = thisString.length() - prefix;
    String prevString = cp_Utf8[i-1];
    String prevPrefix = prevString.substring(0, prefix);
    String thisPrefix = thisString.substring(0, prefix);
    assert(prevPrefix.equals(thisPrefix));

    int small_suffix = cp_Utf8_suffix[i-1];
    char[] suffix_chars;
    int offset;
    if (small_suffix != 0) {
      assert(suffix == small_suffix);
      suffix_chars = cp_Utf8_chars;
      offset = cursor;
      cursor += suffix;
    } else {
      assert(suffix == cp_Utf8_big_suffix[big_cursor]);
      suffix_chars = cp_Utf8_big_chars[big_cursor];
      offset = 0;
      assert(suffix == suffix_chars.length);
      big_cursor += 1;
    }
    String thisSuffix = thisString.substring(prefix);
    String theseChars = new String(suffix_chars, offset, suffix);
    assert(thisSuffix.equals(theseChars));
  }
  assert(cp_Utf8_prefix.length == Math.max(0, cp_Utf8_count-2));
  assert(cp_Utf8_suffix.length == Math.max(0, cp_Utf8_count-1));
  assert(cp_Utf8_chars.length == cursor);
  assert(cp_Utf8_big_suffix.length == big_cursor);
  assert(cp_Utf8_big_chars.length == big_cursor);
 

8.2.2. Representation of cp_Signature Constant Pool

  int cursor = 0;
  for (int i = 0; i < cp_Signature_count; i++) {
    String sign = cp_Signature[i];
    String form = cp_Signature_form[i];
    int form_ptr = 0;
    int sign_ptr = 0;
    for (; form_ptr < form.length(); form_ptr++) {
      assert(form.charAt(form_ptr) == sign.charAt(sign_ptr));
      sign_ptr += 1;
      if (form.charAt(form_ptr) == 'L') {
        String cls = cp_Class[cursor];
        assert(sign.startsWith(cls, sign_ptr));
        cursor += 1;
        sign_ptr += cls.length();
      }
    }
    assert(sign_ptr == sign.length());
  }
  assert(cp_Signature_form.length == cp_Signature_count);
  assert(cp_Signature_classes.length == cursor);

8.2.3. Representation of Byte Offsets

  // ins_pos is a display of all instruction boundaries
  int[] ins_pos;
  ...
  Arrays.sort(ins_pos);
  assert(ins_pos[0] == 0);
  assert(ins_pos[ins_pos.length-1] == bytecodes.length);
  int regulars = ins_pos.length; //  # instruction boundaries
  ...
  int renumber_bci(int bci) {
    int i = Arrays.binarySearch(ins_pos, bci);
    return (i >= 0) ? i : (i == -1) ? bci : ins_pos.length + bci - (-i-1);
  }
  ...
  for (int bci = -100; bci <= bytecodes.length+100; bci++) {
    int bci_numbering_for_band = renumber_bci(bci);
    int i = Arrays.binarySearch(ins_pos, bci);
    if (i >= 0) {
      assert(ins_pos[i] == bci);
      assert(bci_numbering_for_band < regulars);
      assert(bci_numbering_for_band == i);
    } else if (0 < bci && bci < bytecodes.length) {
      int nexti = (-i-1);  // index of next instruction
      assert(ins_pos[nexti-1] < bci && bci < ins_pos[nexti]);
      int prevRegulars = nexti;
      int prevAll = bci;
      int prevIrregulars = prevAll - prevRegulars;
      assert(bci_numbering_for_band >= regulars);
      assert(bci_numbering_for_band == regulars + prevIrregulars);
    } else {
      // other (random) numbers are unchanged by renumbering
      assert(bci_numbering_for_band >= bytecodes.length ||
	     bci_numbering_for_band < 0);
      assert(bci_numbering_for_band == bci);
    }
  }

8.2.4. Representation of Predictable Nested Class Names

  String bcn, predictableOuter, predictableICName;
  ...
  String name = bcn.substring(bcn.lastIndexOf('/')+1);
  int dollar2 = name.lastIndexOf('$');
  assert(predictableICName == null ||
         (predictableICName.charAt(0) > '9' &&
          predictableICName.indexOf('$') < 0));
  if (predictableICName == null) {
    // bcnCase1 or bcnCase4
    assert(predictableOuter == null);
    assert(dollar2 == -1 ||
           dollar2+1 == name.length() ||
           isDigit(name.charAt(dollar2+1)));
  } else if (predictableOuter == null) {
    // bcnCase2
    assert(name.endsWith("$"+predictableICName));
    int dollar1 = name.substring(0, dollar2).lastIndexOf('$');
    assert(dollar1 >= 0 && dollar1+1 < dollar2);
    assert(isDigitString(name.substring(dollar1+1, dollar2)));
  } else {
    // bcnCase3
    assert(bcn.equals(predictableOuter+"$"+predictableICName));
  }
 

8.3. Appendix: Design FAQ

8.3.1. General Questions

1. Why not use an existing generic compression mechanism, such as .jar, .zip, or .tar.gz files? First, it is better to know the structure of a thing compressed. Second, a zip or jar archive is compressed element-wise not globally. This means that any symbol shared by several classes must be mentioned independently once in each class file. Third, the structure of an individual class file is really an interleaving of many different kind of data, each with their own statistics, which limits the effectiveness of a single-stream compressor such as gzip. The specific benefit of reordering the data into bands is about a factor of two, independently of the quality of the off-the-shelf compressor used as a post-pass. There are also significant marginal benefits from the various type-specific recoding techniques. The net result is to improve a compression factor from 2-4X to 7-9X.
2. Why is this specification so complex? The techniques described here are the result of much experimentation over several years. Each feature of this specification is thought to contribute measurably to the overall compression ratio achieved by this technology. The authors would be happy to be shown, by benchmarks, that some given feature can be omitted without significant loss of compression performance.

   (A compression performance multiplier of 1.002 or more for some particular technique is significant, because such multipliers accumulate readily into performance that end-users notice. A multiplier of less than 1.0005 is insignificant.)

3. The Pack200 transmission format is closely tied to the current class file format. Won't it go out of date as soon as the class file format evolves further? Further developments are likely to take the form of new attributes or perhaps new bytecodes. The layout language defined by Pack200 supports a very wide range of attribute formats, including arbitrary mixes of bytes and constant pool references. Likewise, the bytecode representation includes escape operators which can represent arbitrary mixes of data and constant pool references. Although such constructs will not compress as optimally as the features for which this specification is directly designed, it appears that reasonable future extensions will continue to be transmittable, without disturbing the compression of current features, and without requiring decompressors to be updated.
4. Since Pack200 is a lossy compression algorithm, won't it break signed JAR files? Signed JAR files contain secure hashes over the bytewise contents of individual class files. Any change to the bits of a class file will change its hash code, making the innocuous changes of Pack200 indistinguishable from attacks on the application code.

   Pack200, like many other compression algorithms, allows the compressor many degrees of freedom in choosing the contents of the compressed archive, but no degrees of freedom in choosing the contents of the decompressed JAR file. Given a compressed archive, all compliant Pack200 decompressors must produce the same class file bytes, for each transmitted class file. This stability of output persists for class files even if a resource file (such as the manifest) changes its contents. Thus, for any given class file, compression may be lossy, but the decompression of each class file must be exact, in a useful way defined by the Pack200 specification.

   This means that a compressor with the right stability properties can be used to produce a signed, packed JAR using these steps:

   1. Pack the original signed JAR file.
   2. Unpack it, producing perturbed class files.
   3. Re-sign it, ensuring that only the manifest changes.
   4. Re-pack it, using the updated manifest.

   (Note: If this specification allows two compliant decompressors to produce different JAR archive elements for the same compressed archive input, it is a bug in the specification. The specification is thought to be free of such bugs, but if you find one please report it.)

5. What is the relationship between file names in Pack200 archives and file names in JAR (or ZIP) archives or disk files? Pack200 specifies that file names are transmitted as Utf8 strings. Since these strings are intended to function as names of JAR elements in functioning Java applications, the strings must also conform to Java usage, which also represents pathnames as Utf8 strings in JAR files and 16-bit Unicode in memory. Also, in JAR files, pathname components are separated by the forward slash character ('/'), and not by any system-specific character. This allows class loaders to easily convert the internal representation of the class names (e.g., "java/lang/Object") to pathnames (e.g., "java/lang/Object.class").

   The Pack200 specification does not dictate the interpretation of file name strings with respect to any other tool or operating system. However, the natural mapping would be as coherent as possible with Java's usages.

6. What is the relationship between file dates in JAR (or ZIP) archives or disk files? Pack200 specifies that file dates are transmitted as 32-bit counts of seconds relative to Java's time base (i.e., System.currentTimeMillis divided by one thousand). This provides absolute times relative to UTC, at a one-second granularity. If an operating system provides more accurate file times, they must be adjusted to a nearby whole second before transmission in a Pack200 archive.

   JAR and ZIP store dates in local format (i.e., "YYYY/MM/DD HH:MM:SS") with no timezone specification. This means that conversion to and from Java's UTC-based times requires a guess at the timezone under which the compressor was operating. (This problem is not unique to Pack200. It is a problem with all uses of ZIP and JAR archives.) For many purposes, the standard guess is that the decompressor and compressor were operating in the same timezone and during the same yearly daylight savings time regime. However, in order to provide more stability in the times transmitted in Pack200 archives, it is expected that most compressors and decompressors will agree to use UTC as the timezone when interpreting ZIP-style local times.

7. Doesn't the JEFF^tm file format also oprovide a class-specific compression algorithm? Yes, the JEFF Working Group has defined a standard file format (ISO/IEC 20970) which allows class files to be compressed about 50% and also loaded directly into memory for interpretive execution. As compression performance, it is comparable with the DEFLATE algorithm used within JAR archives. Pack200 provides much greater compression. However, Pack200 archives are not designed for direct execution by any virtual machine. The greater compression level of Pack200 is won at a cost of complexity in the unpacker, a complexity incompatible with any requirement of direct loading or execution.
8. Why doesn't Pack200 use Huffman encoding? (Same question for LZW, or BWT, or move-to-front, or any other standard compression technique.) As a matter of simplicity and division of labor, Pack200 intentionally focuses on finding and removing large-scale redundancies specific to class files. Pack200 produces a byte-oriented output, hopefully with clear patterns and a simple alphabet statistics. It relies on on a post-pass compressor to encode these bytes in some more efficient string-sharing, bit-sliced representation.

   The following considerations support this design:

   • Repeating similar compression tactics in two places in a pipeline of transformations can actually degrade overall compression. For example, doing move-to-front in Pack200 band transmission would make it harder for the DEFLATE algorithm to detect repeated patterns, because they would have varying spellings in the tokens DEFLATE would observe. (Delta encoding also has this risk, which is why it's optional in Pack200. But, off-the-shelf back ends don't generally do delta encoding.)
   • Relying on an off-the-shelf backend for final compression simplifies the Pack200 design.
   • It also lets Pack200 leverage the latest byte-oriented compression tools.
   • Such factoring also simplifies engineering of packer heuristics.
   • The job of Pack200 is to find redundancies and patterns at class and package scales. It seems cleaner to avoid messing with byte-level scales.
   • Some specific standard compression techniques may still be encumbered by patents, even on simple, self-evident techniques. Avoiding such patents is a significant cost in engineering a compressor. We would rather avoid this cost by using a standard tool.
9. How does this specification cope with archive format changes? The major and minor version numbers of a Pack200 archive advertise which version of this standard has been generated by a packer. As of 2005 there are two versions, 160.1 for Java 6 and 150.7 for Java 5. Because of the flexibility of attribute layouts, packers can sometimes represent new classfile formats in old archive formats, but newer archive formats may be required for reasons of functionality or performance.

   Unpacker implementations are strongly encouraged to support each standard version of the archive format, since there is not always a strong alignment between packer and unpacker versions at either end of a deployment channel. Packer implementations are encouraged to preserve the ability to emit older archive formats, to maintain maximum compatibility with unpackers.

   The reference implementation chooses to maintain backward compatibility by producing a 1.5 pack format if the input JAR archive contains 1.5 (or lower) classfiles. Otherwise it will produce a 1.6 pack file. The reference implementation unpacker is compatible with all previous standard versions.