1 2 ijar: A tool for generating interface .jars from normal .jars 3 ============================================================= 4 5 Alan Donovan, 26 May 2007. 6 7 Rationale: 8 9 In order to improve the speed of compilation of Java programs in 10 Bazel, the output of build steps is cached. 11 12 This works very nicely for C++ compilation: a compilation unit 13 includes a .cc source file and typically dozens of header files. 14 Header files change relatively infrequently, so the need for a 15 rebuild is usually driven by a change in the .cc file. Even after 16 syncing a slightly newer version of the tree and doing a rebuild, 17 many hits in the cache are still observed. 18 19 In Java, by contrast, a compilation unit involves a set of .java 20 source files, plus a set of .jar files containing already-compiled 21 JVM .class files. Class files serve a dual purpose: from the JVM's 22 perspective, they are containers of executable code, but from the 23 compiler's perspective, they are interface definitions. The problem 24 here is that .jar files are very much more sensitive to change than 25 C++ header files, so even a change that is insignificant to the 26 compiler (such as the addition of a print statement to a method in a 27 prerequisite class) will cause the jar to change, and any code that 28 depends on this jar's interface will be recompiled unnecessarily. 29 30 The purpose of ijar is to produce, from a .jar file, a much smaller, 31 simpler .jar file containing only the parts that are significant for 32 the purposes of compilation. In other words, an interface .jar 33 file. By changing ones compilation dependencies to be the interface 34 jar files, unnecessary recompilation is avoided when upstream 35 changes don't affect the interface. 36 37 Details: 38 39 ijar is a tool that reads a .jar file and emits a .jar file 40 containing only the parts that are relevant to Java compilation. 41 For example, it throws away: 42 43 - Files whose name does not end in ".class". 44 - All executable method code. 45 - All private methods and fields. 46 - All constants and attributes except the minimal set necessary to 47 describe the class interface. 48 - All debugging information 49 (LineNumberTable, SourceFile, LocalVariableTables attributes). 50 51 It also sets to zero the file modification times in the index of the 52 .jar file. 53 54 Implementation: 55 56 ijar is implemented in C++, and runs very quickly. For example 57 (when optimized) it takes only 530ms to process a 42MB 58 .jar file containing 5878 classe, resulting in an interface .jar 59 file of only 11.4MB in size. For more usual .jar sizes of a few 60 megabytes, a runtime of 50ms is typical. 61 62 The implementation strategy is to mmap both the input jar and the 63 newly-created _interface.jar, and to scan through the former and 64 emit the latter in a single pass. There are a couple of locations 65 where some kind of "backpatching" is required: 66 67 - in the .zip file format, for each file, the size field precedes 68 the data. We emit a zero but note its location, generate and emit 69 the stripped classfile, then poke the correct size into the 70 location. 71 72 - for JVM .class files, the header (including the constant table) 73 precedes the body, but cannot be emitted before it because it's 74 not until we emit the body that we know which constants are 75 referenced and which are garbage. So we emit the body into a 76 temporary buffer, then emit the header to the output jar, followed 77 by the contents of the temp buffer. 78 79 Also note that the zip file format has unnecessary duplication of 80 the index metadata: it has header+data for each file, then another 81 set of (similar) headers at the end. Rather than save the metadata 82 explicitly in some datastructure, we just record the addresses of 83 the already-emitted zip metadata entries in the output file, and 84 then read from there as necessary. 85 86 Notes: 87 88 This code has no dependency except on the STL and on zlib. 89 90 Almost all of the getX/putX/ReadX/WriteX functions in the code 91 advance their first argument pointer, which is passed by reference. 92 93 It's tempting to discard package-private classes and class members. 94 However, this would be incorrect because they are a necessary part 95 of the package interface, as a Java package is often compiled in 96 multiple stages. For example: in Bazel, both java tests and java 97 code inhabit the same Java package but are compiled separately. 98 99 Assumptions: 100 101 We assume that jar files are uncompressed v1.0 zip files (created 102 with 'jar c0f') with a zero general_purpose_bit_flag. 103 104 We assume that javap/javac don't need the correct CRC checksums in 105 the .jar file. 106 107 We assume that it's better simply to abort in the face of unknown 108 input than to risk leaving out something important from the output 109 (although in the case of annotations, it should be safe to ignore 110 ones we don't understand). 111 112 TODO: 113 Maybe: ensure a canonical sort order is used for every list (jar 114 entries, class members, attributes, etc.) This isn't essential 115 because we can assume the compiler is deterministic and the order in 116 the source files changes little. Also, it would require two passes. :( 117 118 Maybe: delete dynamically-allocated memory. 119 120 Add (a lot) more tests. Include a test of idempotency. 121
