2 Added list of member entites to union. unioned_types and delimiters
3 are superfluous now -- removed.
4 Rationale: we want to express selection of a union member as a Sel
5 node. for this we need to have entities. The entities contain the
6 unioned type and have as name the delimiters...
9 Changed Makefile/MakeTargets to install headers and libraries
13 Changed Makefile and Configure scripts to produce a proper tar file
14 for distribution (libfirm-VERSION.tar)
15 Use 'make distrib' instead of 'make dist'
18 Changed implementation of tr module. With this I also changed the
19 interface!! Type casts are unnecessary now, but some access funcions
23 ======================================================================
24 ==================== New version: 0.3.0 ==============================
25 ======================================================================
31 added creation of a tar archive for distribution
32 replaced ir/.dummy.in by ir/Makefile.in
35 Changed equivalent_node() and copy_preds() so that the end block
36 is not optimized away if it has no real predecessors. This fixes
37 the problem with the while examples.
38 Added while_example in Makefile
41 Fooled around with if_while_example: Why does cse make it abort?
42 Cse removes path to End node, Bad remains as pred of Endblock.
43 The block walker can't deal with this.
44 Made new example with only loop: while_example.
47 added auto documentation in Makefiles, configure, etc
48 example in ir/ir/ircons.c - much must be added
49 creation of documentation by 'make autodoc' in architecture directory
52 added lots of #ifdef HAVE_CONFIG_G #include "config.h" #endif
53 replaced some C++ style comments (//) by standard C comments /**/
56 - configure (with some features, see --help for more)
57 - make [depend | tags | install]
58 - make [run] in testprograms
60 - compilation on SunOS
62 - automatic call of autoconf if required
65 More colours in vcg output.
66 Fixed bug: cse found that start block equals it's
67 first successor as both have the Start block as predecessor.
68 This only happens for local_optimize_graph() as during construction
69 the Start block is not entered in the value table.
70 Changed implementation: irgmod: local_optimize_graph() did not
71 empty the hash table with the irnodes for cse. This might cause that
72 cse turns up with a node that seems better but was removed from the
73 graph in an earlier optimization phase.
74 Removed old #if 1 from ircons.c
75 Commented in call of optimization for Block nodes in irgopt.c, copy_preds():
76 Optimizes blocks that had bad predecessors.
81 Moved struct from irprog.h to irprog_t.h, same for irmode.h.
82 Added a module deb_info in debug.h. Adapted makefile. Added example
84 Removed use of debug.h in ident.c. Now debug.c|h can be moved to
86 Removed inabled code for dead node elimination in irgopt.h.
87 Renamed some missnamed access routines (_of_). Macros with old names
88 are in old_fctnames.h.
89 Edited makefiles to remove #* and *.flc.
92 Some minor bugfixes...
95 After compacting of in arrays in dead_node_elimination
96 merge blocks and remove Phis with one pred.
99 Freeing of obstack in dead_node_elimination commented in.
100 Compacting in arrays of Block and Phi nodes in dead_node_
105 added set/get_SymConst_type_or_id in irnode.h
106 added field offset to entity in entity.h
107 added field size to type_class in type.h
108 include "gmp.h" in tv.h with "" instead <>
111 Some changes to make the lib better portable
114 Initialize max_node_nr in irprog.
117 Now outputs warning if faulty type in irdump, dump_type_info.
118 Explicit string termination in irdump, vcg_open.
121 Added external to declaration of irp.
122 Added typedef before pns and pnc_number.
123 irmode.c, init_mode(): We need to memset all new modes as
124 tarval_vrfy reads fields before they are initialized: It
125 compares to min/max when tarvals for min/max are allocated!
128 Changed parameter of Methods in type.h and type.c from
129 class to clss. Class is a C++ keyword.
132 Changed some testprograms to accomodate Uebprakt.
135 Finished implementation of dump_all_types.
136 Added new testprogram inheritance_example.
139 Added seven access routines in type.[hc]:
140 get_class_n_member, get_class_n_subtype, get_class_n_supertype,
141 get_strct_n_member, add_strct_member, get_strct_member and set_strct_member.
142 Fixed the type_walk_2 for those seven routines.
145 Added irdump routine dump_all_types.
146 Did changes to irprog, type, typewalk to achieve this.
147 typewalk needs to walk along members, sub and supertypes, this
148 is still missing! irdump then needs to output these edges.
150 ======================================================================
151 ==================== New version: 0.2.5 ==============================
152 ======================================================================
155 Made new Version: 0.2.5
157 15.8.2000 Goetz until 3.9.00
158 Separated irnode.h and many others into two files: one public
160 Added funcionality to extract headers for lib.
163 type_dump also dumps irg->ent
164 if_example: corrected nr of locals
167 irgopt: Correct copy of self pointer in start block.
168 irgopt: Correct treatment of loops: pre function in walker
169 breaks all possible loops in Phi and Block nodes.
170 irr_loop_example: wrong Return statement
171 Now I fixed all but one error: in memory_example.
172 This error showed a conceptual problem, so I wrote
173 a slightly different algorithm. This now runs for
175 dead_node_example: cse is buggy!! merges start and the
179 looking at bug in dead_node_elimination.
181 irnode: all get_..._arr functions changed
182 ircons: new_r_Block: set matured flag.
183 ircons: new_r_SymConst: needs mode_p for linkage_ptr_info.
184 adapted call_str_example to new concepts.
187 Implemented simpler Phi construction algorithm that constructs
188 correct code for three_cfpred_example in ircons. This algorithm
189 generates more Id nodes than the original algorithm and therefore
191 Also made some changes to optimizations of Not.
194 Fixed a bug, caused by the recursive use in type.h and entity.h.
195 Deleted the _TYPE_TYPEDEF_ and inserted the _ENTITY_TYPEDEF_ instead.
196 Fixed a problem in if_while testprogram, so it now works.
199 Fixed bug in optimize(). Returned Null instead n if
200 get_optimize() == false.
203 Added some access routines in type.[hc]. Solved a recursive definition
204 of type_class struct, entity struct and type union.
207 Separated entity header into pure functional header and header
211 ... a bunch small changes not documented
214 Changed output of oo_prog_example: now uses the internal list
215 of all irgs. Had to implement some of the access routines for
216 that. Also fixed a bug in irdump. Main routines did not use
217 current_ir_graph properly.
220 Implemented some missing parts in irgopt.
221 Added field irg to entity. Only valid if entity is a method.
222 Had to resove a cyclic de[endency between irgraph.h and entity.h
224 Added fields sub and super to type_class for inheritance.
225 Added field main_irg to irprog.
228 Fixed some bugs in irgopt.c
229 Added some missing 'get_new_node' methods, so the new obstack contains
233 Renamed ir_vrfy to irn_vrfy and vrfy_graphg to irg_vrfy.
234 Solved some copy'n'paste problems. Deleted most debug-code!
235 Added comments to several files.
238 Fixed bug in _r_ constructors for all four Div operations.
239 Fixed bug in turn_into_tuple: didn't remember the block.
240 Tested transform_node.
243 Fixed some testprograms to new semantics of firm.
244 Added testprogram global_var_example.
245 Fixed a bug in irop: Size of Sel node was (entity *) instead of
246 sel_attr. Therefor array_heap_example aborted.
247 Also found a missing break in typewalk, case iro_Alloc.
248 Added a construct that unifies all things constructed by the lib
249 and needed to represent a program: irprog.ch
250 This contains a list of all types as well as a list of all ir graphs.
251 I added calls to initialization and constructors so that this is constructed
252 automatically. I changed firm.c, type.c, irgraph.c and ev. more.
253 Removed strange block struct and constructor from ircons.c.
254 Moved routines for automatic Phi construction from irgmod to ircons.
255 Sorted functions in ircons to reflect the different interfaces.
256 Renamed dataseg to globals to be conformant with documentation.
257 Added conservation of current_ir_graph to irgopt.c.
258 Corrected bug in Makefiles: wrong path to directory ./inlude.
261 Commented a whole bunch of stuff, e.g. in ircons.h (Procedure to construct)
262 We should change the naming of the Block constructor.
263 Removed acces routines to attr. "closed" of Block.
264 Removed second implementation of new_ir_node from ircons and some comments
265 concerned with the now resolved cyclic dependency.
266 Fixed some bugs in irgopt so that the compilation goes through.
269 Changed the irnode attribut 'visit' into 'visited'.
270 Added some access functions, like get/set_irn_visit etc.
273 Renamed some access funtion in irgraph.*.
274 Deleted two global flags (ir_visited and block_visited) from irgraph.*.
275 Added two attributes to the irgraph struct. The visited flags belongs
276 to a ir graph, not to a whole program.
279 A new optimization has been finished. The dead-node-elimination copies
280 the existing ir graph from one opstack to another. While walking the
281 graph, only the reachable nodes will be copied, so the "dead-nodes"
282 (unreachable in the controlflow) are removed.
285 Solved some problems in the 'copy_node' function. The in array was not
286 as easy as it seamed.
289 Wrote a new 'post' funtion, for future use in 'dead-node-elimination'
290 optimization, which is needed in the paramter list of 'irg_walk'.
293 Moved 'optimze_in_place_wrapper' from iropt.* tothe new files irgopt.*.
294 New functions 'local_optimze_graph' and 'dead_node_elimination' are added,
298 Removed some files from the archive, after dependencies and usage are
300 - 'ir/common/strerror.c'
302 - 'ir/ident/xx_ident.h'
303 an empty file, only included by 'ident.c', but never used;
304 - 'ir/tv/label.c' and
306 contained a numeric counter;
307 The files were included in 'tv.h' and implemented in 'tarval.c'. The two
308 functions 'tarval_label' and 'tarval_forall_labeled' were also never used.
309 In the old fiasco compiler, the label thing is used in the backend.
310 Search for 'babil' to see the usage.
311 Also removed the two attributes 'lab' and 'used' in the tarval struct.
314 Moved the 'new_ir_node' constructor from ircons.* to irnode.* and fixed
315 afterwards some recursive includes, so libfirm works again.
318 Did a lot of changes, which I never commented until now (4.7.00).
320 * Added new result to Start node: Pointer to global data segment.
321 * Extended Semantics of SymConst node to represent information for the
323 * Added arithmeitc nodes (Shrs, Minus ...)
325 Rearranged the directory structure and adjusted the makefiles.
326 The directories contain:
327 ir: everything for the intermediate representation (better: src?)
328 /ir: the ir itself, and standard optimizations.
329 /tv: the target value module
330 /tr: the type and entity representation
331 /adt: abstract data types
332 /common: stuff needed by all other dirs
333 /debug: debugging Unterstuetzung
335 include: external files needen as includes
336 testprograms: examples to test the lib.
338 The makefiles generate files with extension .d that contain the dependencies
342 Added access routine to attribute link of irnode in irnode.ch.
343 Added get_negated_pnc to irnode.ch.
346 Added an iterator in iropt that calls the optimization for all nodes.
347 Copied optimize's code to optimize_in_place and removed deallocations.
348 Made set_irn_n in irnode.c public.
349 Added Bad as possible cfg predecessor in is_cfg_node in irnode. Bads
350 are allowded as cfgpredecessors to represent dead "incoming" control
352 Added call to optimize_in_place in mature_block. Is this useful?
355 Changed tests from comparing enums to comparing pointers. This is more
356 efficient (is it?) and reads better. e.g., instead get_irn_opcode == irm_And
357 now get_irn_op == op_And
360 Compared equivalent_value to the one in the original fiasco.
361 Added a test in case iro_And.
362 Changed implementation of iro_Div. Added routines turn_into_tuple
363 and set_irn_op() in irgmod/irnode.
364 Changed implementation of irn_Minus (although commented out).
365 Now all features are ported or documented for this function.
368 Compared computed_value to the one in the original fiasco.
369 Added one further optimization in case iro_Proj.
370 Now all features are ported for this function.
373 Updated comments in ircons.h.
376 Changed possible attribute of SymConst from type_class to type.
377 SymConst size could be the size of a union, array, ...
380 Implemented two dumpers that dump type information and a graph with
384 Implemented a walker that walks over the type informatin that can
385 be reached from an ir node.
386 Added routine get_kind(firm_thing) in common.h, Added file common.c.
387 Added files typewalk.ch and routines to irdump.
388 Added visited flags and initialization to entity and all types.
389 Implemented some of the missing access routines in type.c, entity.c
392 Implemented constructor for ShrA.
394 Added routine that iterates the graph and verifies all nodes.
395 Added Phi_in stack to irgraph. This allows to build several
396 graphs at once with respect to the Phi building algorithm.
399 Implemented dumper for control flow graphs. Changed irr_* and dead_*
400 examples to dump cfg.
403 Implemeted explicit stack for the Phi node generation. See ircons.c.
406 added include string.h in xoprintf.c to get rid of warning:
407 implicit declaration of function `bcopy'
411 Added more cases to equivalent_node.
412 Portet gigo, identify, identify_remember.
413 The set/pset library can not handle several hash tables. At least,
414 compiling several procedures causes a segmentation fault during access
415 to the tables (oo_program_example). Need to work on this.
416 transform_node still needs to be ported.
417 some cases in equivalent_node are still missing.
419 To get it running I had to remove the call to optimize in the
420 new_Block constructor. It can only be called when the Block is
424 Introduced a walker that walks only over blocks. The asserts
425 might be too strict for general graphs, but so far it works.
426 For the walker I introduced an additional flag in the attribute
427 of Blocks, an a global block_visited variable (in irgraph.h).
428 Furhter I modularized dump_ir_node, implemented dump_ir_graph
429 with the existing walker, and added new routines to output a
431 Changed the examples to call the new dump routine.
434 Added auxiliary routines to ir_node:
435 skip_Proj, skip_Nop, is_Bad ...
436 Changed access routines to use get_irn_n instead of explicit array
437 accesses. Changed get_irn_n to call skip_Id, so that the access routines
438 never will return Id nodes.
441 Repaired Phi computation, changed new_r_Phi_in. Added test program to
442 test irregular controlflow.
445 Updated optimizations computed_value and equivalent_node to new
446 Syntax of FIRM nodes and to use access routines.
447 Added routine skip_Proj in irnode.
449 2.2.2000 Boris Boesler
450 Added files irgwalk.c and irgwalk.h. They include a simple function to
451 traverse an ir graph and execute pre and post functions to current node.
454 Added arrays with parameters and result types to the type informtion
455 of a procedure. This is needed for the code generation. Also added
456 the corresponding access routines.
457 Adapted the testprograms.
460 Found a bug / complication: the call order
461 get_value (makes Phi0, put's it into graph_arr)
462 set_value (overwrites Phi0 in graph_arr)
463 mature_block (upgrades Phi0, puts it again into graph_arr, overwriting
465 fails. Added test wether graph_arr is already set in phi_merge.
467 Added tons of comments to explain Phi generation -- the stack hack.
471 irdump now outputs the pnc_number for projs of Cmp nodes.
472 Added routine to access the pnc_string.
475 Added result "dataseg" to Start node, with all it's consequences.
476 We need this for the compiler lab to translate global variables.
477 Now they can be selected from the data segment, and the start
478 node supplies the pointer to this segment.
479 These changes are guarded by preprocessor variable UEBPRAKT in
481 Changed files: common.h, irgraph.ch, irnode.h
485 removed array_example, added array-heap_example and
486 array-stack_example and adapted makefile.
488 Entity now also allows type_method as owner. This is to model
489 the stack, Sel nodes that select a variable (array) from the
490 stack need an entity that describes what they select. This
491 entity gets the type of the enclosing procedure as owner.
494 Added future access routines in `irmode.[ch]' as comments,
495 that all attributes in the ir_mode struct can be user-defined.
496 Added access routines in `irgraph.[ch]', for all attributes.
499 Moved verify routines from `irgraph.[ch]' to own files `irvrfy.[ch]'.
500 Added `irvrfy.[ch]' to MAKEFILE
501 Added "inline" before access routines in `irnode.h', that no warnings
502 will appear furthermore.
503 Added three more access routines in `irmode.[ch]':
504 struct tarval *get_min_of_mode (ir_mode *mode);
505 struct tarval *get_max_of_mode (ir_mode *mode);
506 struct tarval *get_null_of_mode (ir_mode *mode);
509 Adaption of Alloc, Sel and Entity to new types:
510 Changed array_example to use type instead of type_class.
511 Changed entity to allow all types as owner instead of
512 only type_class. Added routine "assert_legal_owner_of_ent(type* type)".
513 Changed alloc to allow allocating variables of any type instead of
515 Adapted all testprograms to avoid warnings.
517 removed dtest ... from Makefile, replaced by test.
520 Added "run" to makefile, added array_example to makefile of testprograms.
521 Corrected testprogram array
522 Changed assert in irvrfy -- Conv.