CHANGES

##############################################################################
##
## BRASS source file
##
## SCORE TDF compiler:  Change log
## $Revision: 1.160 $
##
##############################################################################
##
## Version discipline:  Version "1.XXX" gets CVS tag "v1_XXX".
##                      After "cvs commit",
##                         do "cvs tag v1_XXX" in tdfc top directory.
##
##############################################################################

1.160  (9/1/05, eylon)
- Verilog emission "-everilog" got
    * New option "-syntech" specifies Synplify technology: tech-part-pkg-speed
        (e.g. VIRTEX2P-XC2VP70-FF1704-7, VIRTEX4-XC4VLX100-FF1148-11)
    * Improved stream queues
        + "Q_pipe.v" family now explicitly uses regs, not SRL
        + "Q_srl.v" + "Q_srl_reserve.v" output regs are now non-retimable
        + New family "Q_lowqli.v" has both input-side and output-side
            logic pipelining, plus wire pipelining
    * New options to distinguish between input-side and output-side
      logic pipelining depths  (use with "Q_lowqli.v")
        + "-qili" "-qilo"  (replace "-qil")
        + "-qoli" "-qolo"  (replace "-qol")
        + "-qlli" "-qllo"  (replace "-qll")
    * Distinguish 2 kinds of compositional ops
        + (1) netlist = composition of compositions,   queues are wires
        + (2) page    = composition of behavioral ops, queues from "-qim" etc.
        + top composition is netlist is page partitioning, page if not
    * Many new scripts for mapping to FPGA  (tools directory)
- Environment variable "CXX" now overrides compiler used for pre-processing

1.159  (12/8/04, eylon)
- Improved type-checking of calls to built-ins (e.g. cat)
- Added final type-checking after instance generation, constant folding,
    and page transformations

1.158  (10/13/04, eylon)
- Verilog emission "-everilog" got
    * ROM emission for read-only, inline arrays  (not in "-ecc", use "-xc")
    * Faster stream queues  ("Q_srl" + "Q_srl_reserve" with data output reg)
    * New module "<page>_dpq" containing data-path + output queues
        (implemented as "<page>_noin" with black-box FSM, may include segments)
        (intended to test datapath pipelining with "-qom Q_lwpipe")
- C back-end "-ecc" now treats all segment data as "long long"
    to match runtime C implementation of segments
    (previously inferred type from data bit width)

1.157  (8/10/04, eylon)
- Page netlist emission ("-epn") now emits files in
    AT&T GraphVis format (opname.dot) in addition to
    Randy Huang's format (opname.pagenetlist)
    (note: "-epn" is additive to and happens after other -eXXX emission)
- New options for choosing logic pipelining depth and interconnect/wire
  pipelining depth of Verilog streams:  (use with "Q_srl_reserve_lwpipe.v")
    -qil <l>   : use logic pipe depth <l> for page input  queues
    -qiw <w>   : use wire  pipe depth <w> for page input  queues
    -qol <l>   : use logic pipe depth <l> for page output queues
    -qow <w>   : use wire  pipe depth <w> for page output queues
    -qll <l>   : use logic pipe depth <l> for page local  queues
    -qlw <w>   : use wire  pipe depth <w> for page local  queues
- Verilog emission ("-everilog") got
    * New stream queues  (e.g. pipelined stream "Q_srl_reserve_lwpipe.v")
    * New implementation "rev_4__200mhz" targetting 200MHz for "-synplify"
    * Code cleanup + bug fixes

1.156  (11/18/03, eylon)
- Verilog emission ("-everilog", "-synplify") now emits pages
  induced by compositional operators, w/queues for input/output/local streams
    * Verilog module hierarchy:
        op_page = { op_qin, op_noin }
        op_noin = { op_q,   behavop1, behavop2, ..., segop1, segop2, ...}
        behavop = { behavop_fsm, behavop_dp }
    * Segment ops emit as Verilog black-boxes
    * Page partitioning modes:
        "single page"  mode:  no partitioning, all ops and queues in 1 page
        "op per page"  mode:  "-pp1" emits pages containing 1 op and I/O queues
        "manual partn" mode:  user writes TDF 2 level compositional hierarchy,
			      top is page graph, 2nd is pages (UNIMPLEMENTED)
        "auto partn"   mode:  automatically generate TDF 2 level compositional
			      hierarchy (UNIMPLEMENTED)
- New option "-pp1" for "op per page" dummy partitioning mode,
    creates pages each containing 1 behavioral op plus input/output queues
    (flattens and wraps each behavioral op in its own compositional op)
- New options for choosing module and depth of Verilog queues:
    -qim <m>  :  use Verilog module <m> for page input  queues
    -qid <d>  :  use         depth  <d> for page input  queues
    -qom <m>  :  use Verilog module <m> for page output queues
    -qod <d>  :  use         depth  <d> for page output queues
    -qlm <m>  :  use Verilog module <m> for page local  queues
    -qld <d>  :  use         depth  <d> for page local  queues
- Fixed C back-end bug where firing should stall on full output
    but control falls through to "safe EOS" termination
    (split firing condition into 2 levels: 1 for inputs, 1 for outputs,
     with "safe EOS" being else-case for level 1)
    (note, bug still exists where do not match any of multiple input
     signatures/state-cases but control falls through to "safe EOS")
- Added lots of scripts for Synplify and Xilinx compilation

1.155  (7/7/03, eylon)
- constant folding now propagates the value of read-only reg/local variables
- TDF transformations now respect the order of variable declarations
- TDF syntax now admits empty statements ";"
         and empty variable declarations ";"
    (except under "if", where we still need a non-empty stmt or stmt block).

1.154  (6/13/03, eylon)
- constant folding now masks high bits of result to limit bit width
    * this is more strict w.r.t. TDF typing rules, e.g. 1<<2 is 0 not 0b100
    * this may change previously erroneous CC emission
- to avoid conflict with gcc naming convention,
    * copy operators     are now generated with name "_copy" not "__copy"
    * pipe regs in "-tt" are now generated with name "_pipe" not "__pipe"

1.153  (6/10/03, eylon)
- new option "-everilog" emits Verilog
    * for each page p, emit 3 modules in separate files:
        datapath "p_dp.v",  FSM "p_fsm.v",  composition "p.v"
- new option "-synplify" emits Synplify project files
    * create dir "proj"
    * for each page p, create project dirs "proj/p_dp" "proj/p_fsm" "proj/p"
    * for each project create project files and 3 implementation directories
    * presently hard-coded to map for XC240 speed grade -6)
- fixed "-pt", "-esd" to admit a single behavioral operator,
    so no longer need a dummy top-level compositional operator

1.152  (3/27/03, eylon)
- fixed bug in timing transform (blockdfg.cc) to recreate partition DFGs
    using correct edge adjacency list order
    (depending on edge order is brittle, should probably change it)
- Linux build now uses LEDA 3.7.1 (recompiled) instead of fakeLEDA

1.151  (9/24/02, eylon)
- updated to use RedHat 7.3, Linux kernel 2.4.18, g++ 2.96
- Linux build now uses Yury's "fakeLEDA" (LEDA API with STL under the hood);
    good for debugging, but expands tdfc size and runtime about 3x
- fixed bug in "-esd" option to set start node when copying automata

1.150  (5/15/02, eylon)
- new option "-esd" Emits Stream Depth bound analysis
    (try composing automata for pairs of connected behavioral operators
     with increasingly larger queues, as per EE290N project 5/15/02)
- new option "-SD <d>" to set max depth tried by "-esd" option (default 2)

1.149  (5/7/02, eylon)
- new option "-ppmetis" emits page partitioning graph for each op
  in extended Metis format, "<instancename>.metis"
    * vertex line format:  a v1 e1s e1b v2 e2s e2b ...
        (a=area, vi=vertex, eis=state xfer count, eib=bit width)
    * page partn graph has state nodes, state xfer edges, data-flow edges
        (multi-graph, i.e. multiple edges connecting same vertices)
    * page partn graph is after dismantling shared stream/reg access
        and pre-clustering states that access same stream/reg

1.148  (5/3/02, eylon)
- new option "-pt1" for page timing transformation, dummy timing model
    * converts any state-case with a multi-cycle action
        into a sequence of single-cycle states
    * dummy timing model:  all primitives take 1 cycle (const, reg write = 0)
    * does pseudo-ssa conversion
        (multiple asst to a var in same state induces new, single-asst vars)
    * does dead-code elimination within a state action
- new option "-pt" for page timing transformation, HSRA-based timing model
    * cascade 2 LUTs or 4 carries per cycle
    * not fully debugged
- new option "-T <n>" to set cycle length to n
    * for "-pt1" option:  base cycle length n=1
    * for "-pt"  option:  base cycle length n=1000

1.147  (1/23/02, eylon)
- page partitioning is now done by "-pp" option,
    before and independent of any "-eXXX" emission option  ("-ecl" is defunct)

1.146  (1/14/02, eylon)
- migrated to CVS
    * version numbering discipline:  version 1.XXX gets CVS tag "v1_XXX";
                                     use "cvs tag v1_XXX" after "cvs commit"
    * RCS history was retained;
        old RCS version 1.XXX is accessible using appropriate CVS tag "v1_XXX"
    * Makefile now deduces TDFC_VERSION directly from CHANGES file

1.145  (1/8/02, eylon)
- new option "-epn" to Emit a Page-level Netlist for each instance (page);
    creates netlist file "opname.pagenetlist" formatted for rhuang's router;
    option "-epn" is additive to and happens after other -eXXX emission.    
- casting from signed int to a narrower signed int is now allowed
    (I am not happy about allowing destructive casting,
     but it seems to be the only reasonable syntax we have
     to discard high-order bits from a signed int)
- bug fix in C back-end to avoid incompletely specified types
    (we are still using TYPE_ANY to represent unknown / un-type-checked types;
     in the future, we should use NULL types instead,
     then use getType()/typeCheck() to automatically generate on demand)

1.144  (12/26/01, eylon)
- fixed C backend emission of boolean params + initial values (Joe's bug)

1.143  (11/30/01, eylon)
- fixed C backend emission of cat() to properly widen C types for shift+mask
- fixed C backend emission of casting to properly widen + sign-upgrade types
    (note: casting from unsigned to narrower unsigned does NOT mask high bits)

1.142  (11/14/01, eylon)
- casting from unsigned int to a narrower unsigned int is now allowed
    (this is a special case of bit selection, since have no inline bit sel op)
    (user is discouraged from using this in TDF source--
     it was added as a hack for internal optimizations + back-ends)
- minor fix to build in Mac OS X 10.1

1.141  (11/2/01, eylon)
- fixed automatic fanout generation (copy operators)
    to distinguish between copy(signed) and copy(unsigned)

1.140  (10/30/01, eylon)
- page synthesis:  extended "-ecl" (state clustering) to
    * synthesize TDF code for pages of partitioned operators
    * modify top-level compositional op to expose pages
    * print TDF of new composition and pages  ("*** RESULTS OF PARTITIONING")
- fixed "-ecl" bug that sometimes lost first cluster (partition)
- fixed constant folding of bitsof() in instance generation
    (andre always stripped bitsof; new code may keep it)
- added basic bitsof() support to Laura's IR
    (strip bitsof(), add DF edge to its arg)
- "-xc" now bypasses/removes "dumb" start state (contains goto, no inputs)

1.139  (10/2/01, eylon)
- option "-xa":  dismantle reg/stream access
    * extends operator decomposition;
        is simple model to guarantee sequentially-consistent access
        to shared regs/streams across all pages of decomposed operator
    * implementation:
        (1) front-end xform s.t. each reg/stream access gets own state
        (2) in state clustering, pre-cluster states that access same reg/stream
    * "-xa" must be used with both "-ecl" (state cluster) and "-as" profiling
- fixed bug in "-ecl" state clustering, in choosing of nodes to hoist;
    was accepting nodes that, together with source node,
    violated partition area even before min-cut;  now optional
- DIV area model now assumes NxM array (like mult; was 0 area const shift)
- linking of 'goto' stmt now "relinks" target state,
    i.e. finds target by name in present scope + fixes target ptr
    (this should fix laurap's complaint about bad gotos in IR code)
- now builds in Mac OS X with LEDA 3.7.1

1.138  (8/15/01, eylon)
- made -ecl option more quiet (most output is now in warnings, can use -q)

1.137  (8/13/01, eylon)
- fixed -as bug that lead to execution errors and inflated firing counts

1.136  (7/23/01, eylon)
- reduced work for -ecl state clustering:
    when evaluating nodes to hoist across cut,
    skip "equivalent" nodes that would lead to a previously-seen min-cut

1.135  (7/4/01, eylon)
- revised version of -ecl state clustering
    * Wong algorithm now evaluates all neighbor nodes
        when choosing a node to hoist across the min-cut
    * previous clusters are collapsed into a common node that serves
        as sink for next cluster's min-cut

1.134  (6/28/01, eylon)
- bug fixes for -ecl state clustering
    * node merging now sums flows/residuals for merged edges
    * simplified backwards edge model-- every edge now gets a bwd counterpart

1.133  (6/26/01, eylon)
- added support for fanout using copy()
    * new builtin operator copy(i,o1,...oN) for fanout, available in compos ops
    * automatic inferencing of fanout in compositional operators;
        when stream has multiple readers, generate corresponding copy() calls
    * automatic generation of behavioral code for copy operators
        (creates parameterized ops "__copy_XN" where X={b,i,f}, N=fanout)
- no longer need g++ 2.95 to compile clusterstates.cc,
    g++ 2.91 now compiles everything

1.132  (6/7/01, eylon)
- various bug fixes for "-ecl" state clustering:
    * skip single-state operators, assuming they are pipelines
    * disallow merging src+dst nodes during Wong min-cut
    * choose dst node as largest state (not largest, previously unused state)
    * statistics

1.131  (5/29/01, eylon)
- "-ecl" state clustering now exposes register data-flow edges between states
    for min-cut state clustering
- experimental option "-mix" sets data-flow/state-flow mixing coeff for "-ecl"
    ("-mix 0" for pure DF cut, "-mix 1" for pure SF cut, frac [0..1] to mix).
- Special note:  "-ecl" data-flow/state-flow mixing does NOT yield cut-sizes
    monotonic in the mixing coefficient!  This is theoretical, not bug.
- "-ecl" min-cut now chooses src/dst states as largest nodes, not random
    (chooses largest nodes not used as src/dst since last graph modification);
    this heuristic seems to improve mix behavior.
- "-ecl" now treats zero-frequency states/transitions from "-as"
    state profiling (i.e. from ".ffsm" files) as frequency=1,
    so they are not ignored altogether.
- "-ecl" bug fixes, including hiding zero-frequency edges before min-cut.

1.130  (5/17/01, eylon)
- TDF printf() now prints to stderr to avoid conflict w/ScoreRuntime output
- new flag -ads adds debug code that prints state name+inputs at state entry
- fixed bug in -xc basic block decomposition where upgrading a
    local variable to a register mistakenly moved its initialization point
- added pass to -xc to bypass+remove "dumb" states (containing only a goto)
    and unreferenced states
- incorporated 1.120ab changes to SCORE API calls for faster simulation:
    * Array-side     NEW_*_STREAM*()              becomes *_ARRAY()
    * Array-side     STREAM_READ/WRITE/DATA/EOS() become  *_ARRAY()
    * Processor-side STREAM_READ/WRITE()          become  *_NOACC()

1.129  (4/11/01, eylon)
- new counter "doneCount" in pagestep() code, counts stall cycles after done
- fixes to support g++ 2.95.2 and LEDA 4.2  (not used in Makefile yet)

1.128  (3/9/01, laurap)
- made the IR code part more user friendly. now a message shows which are the
    files that contain data flow area and fsm information
- taken many comments out
- modified some constructors in ir_graph.h/cc

1.127  (2/9/01, laurap)
- fixed a bug on FSM kiss2 format emission.
    (modified:  ir_graph.h ir_create_dfg.cc ir_create_sfg.cc)

1.126  (2/5/01, eylon)
- state clustering (-ecl) now uses SFG from IR (for state flow, state DF area)

1.125  (2/4/01, eylon)
- reading FSM feedback files (-ecl) now checks if state exists in AST

1.124  (2/4/01, eylon)
- option -eIR no longer turns on -ecc
- changed names of many command-line options for consistency
- state clustering:
    * Implements Wong balanced min-cut on state transition probabilities
    * Presently no IO constraints
    * Presently ignores IR, builds own SF graph from FSM profiling results,
        pretends all states are size 1
    * option -as:   add code in CC back-end (-ecc) to profile FSMs at run-time
    * option -ecl:  emit state clustering results  (uses -as profiling results)

1.123  (2/4/01, eylon)
- cleaned up the passing of instance pages to IR generation
    (reverted instance.cc/.h +  bindvalues.cc/.h to version 1.121)

1.122  (2/1/01, laurap)
- new intermediate representation (IR),
  including pipeline extraction, FSM synthesis, etc. (new files: ir_*)
    * option -eIR:  use IR, emit synthesis info
    * option -evcg: emit .vcg files to graph DFGs, SFGs

1.121  (1/20/01, eylon)
- fixed bug supporting 64 arguments per operator

1.120  (11/13/00, eylon)
- fixed bug where compositional op disallowed initialization of local array

1.119  (9/11/00, eylon)
- support multiple variable decls. per line  (e.g.: "unsigned[8] a,b,c;")
- support stream depth hint  (e.g.: "unsigned[8] a(500)" on local stream decl.)
    * depth hint specifies minimum recommended buffer space for stream
    * depth hint is element of SymbolVar, which is used to represent
      local stream declarations inside a compositional operator

1.118  (8/8/00, yurym)
- modified Makefile to force gmake to use _only_ local compilers
  this should make it easier to distribute tdfc to any linux box

1.117  (8/8/00, eylon)
- added "printf" to TDF
    usage: printf("format", exprs...)
    all args except format string are cast to (long long), so use "%Ld"

1.116  (8/7/00, eylon)
- fixed bit selection bugs for data widths 33-to-64  (Antoine Hue's bug report)

1.115  (5/30/00, yurym)
- unified bitmask type for operator arguments among tdfc and runtime

1.114  (5/25/00, eylon)
- operators now support up to 64 arguments
- feedback files now handle long lines (up to 4096 chars)
- fixed numerous type errors to satisfy the latest g++ (2.95.2)

1.113  (2/3/00, mmchu)
- added a check to input_free[N] before STREAM_FREE in eos state cases.

1.112  (1/6/00, mmchu)
- added a "fictional" done state to the C++ instance output so that a 
  page that has signalled done can be safely rescheduled into the
  array and guarantee to "do no harm".

1.111  (1/4/00, mmchu)
- fixed a bug in cceval.cc for simplify_select. changed -1 to +1.

1.110  (1/4/00, eylon)
- fixed segment initialization in CC backend

1.109  (1/3/00, mmchu)
- added STREAM_MARKREADSTREAM and STREAM_MARKWRITESTREAM to bindInput and
  bindOutput for proc_run.

1.108  (1/3/00, eylon)
- fixed parser bug which disallowed local variable names to override names
    from enclosing blocks
- fixed bug whereby operator instances ignored the "-xc" flag

1.107  (12/30/99, eylon)
- "-E" option now exits cleanly
- fixed error reporting for bad arguments to segment operators

1.106  (12/28/99, mmchu)
- changed STREAM_READ and STREAM_WRITE => STREAM_READ_NOACC and 
  STREAM_WRITE_NOACC.

1.105  (12/8/99, eylon)
- "-xc" and "-ecc" can now be used together
    ("-ecc" was seg-faulting on NULL tokens created by "-xc")
- fixed CC emission of segments (garbage width)
- another CC bug fix to admit uninitialized local arrays (scratch segments)

1.104  (12/7/99, mmchu)
- fixed precendence problem when CC outputting conditionals.
  (a+b?c:d) is wrong! (a+(b?c:d)) is right!

1.103  (12/6/99, eylon)
- front-end accepts hex (0x...), octal (0...), and binary (0b...) numbers
    (bug: bad octal numbers, e.g."08", "09", are read as 0 without error msg)

1.102  (12/5/99, mmchu)
- Fixed "else-bracket" nesting bug for proc_run code as well.

1.101  (12/4/99, eylon)
- CC back-end now admits uninitialized local arrays (scratch segments)
- fixed CC back-end bug which printed a pointer for a shift amount
- front-end now allows arrays of booleans
- fixed front-end bug which created untyped segment operators

1.100  (12/4/99, mmchu)
- Added C linkage support to C++ operator output. Now, if C/C++ driver code
  always instantiate operators via NEW_* and refer to operator types as
  OPERATOR_* then SCORE tdf operators may be freely used in either C or
  C++ code!

1.99  (11/11/99, mmchu)
- Removed outputting of local state variable (so that state changes can
  be reflected in the state variable defined in the ScorePage class).

1.98  (11/10/99, mmchu)
- Fixed "else-bracket" nesting bug where, in ccpage for the pagestep code,
  if a state had multiple cases, the else did not correctly generate
  the nesting bracket.

1.97  (11/10/99, mmchu)
- fixed ccpage so that the global produces/consumes arrays are appropriately
  typed.

1.96  (11/9/99, mmchu)
- moved the calculation and generation of produces/consumes arrays in ccpage
  down after all the other setAnnotes() to CC_STREAM_ID. this fixes a bug
  where incorrect values were being written out for produces/consumes when
  a compositional operator was used.
- fixed ccpage to actually output STREAM_FULL checks before doing a
  STREAM_WRITE.

1.95  (11/4/99, eylon)
- "-xc" now exlines
    (lifts inlined calls and array-mem references from behavioral operators)
    (still makes "-ecc" segfault)

1.94  (11/3/99, mmchu)
- changed the C++ output so that generated C++ code only includes the
    Score.h file instead of individual .h files (i.e. ScoreStream.h, etc.).

1.93  (11/3/99, amd)
- fix eylon state name truncation bug
    (though eylon should check StmtGoto code and see if I should
    have had to do what I did...or if this is just a kludgy work
    around a bug in StmtGoto)

1.91  (10/26/99, mmchu)
- changed C++ output code so that segment obj->id conversion is done through
    the macro SEGMENT_OBJ_TO_ID instead of directly through the method call.
    this guarantees that we will always call noAccess() on the segment during
    the obj->id conversion of a segment.
- made it so that ScoreStream.h was added to the #include list of operator
    .h files.

1.90  (10/26/99, mmchu)
- changed ccbody.cc and ccinstance.cc so that the outputted C++ code would
    include the .h files it needed from ScoreRuntime instead of assuming they
    were included by some other include file.
    In particular, generated operators
    now include ScoreUserSegmentOperators.h (a compositional .h file) for the
    ScoreSegmentOperator* classes and instance files include ScoreSegment.h
    ScoreStream.h and ScoreUserSegments.h (a compositional .h file of
    ScoreSegment*).

1.89  (10/25/99, mmchu)
- changed tdfc so that its outputted C++ code would interface correctly to
    ScoreSegment* and ScoreSegmentOperator*. This is after the discussion to
    clear up ScoreSegmentOperator*. Also, in compositional operators,
    ScoreSegmentOperator* are simply forked off (all special pthread code
    was removed).

1.88  (10/21/99, amd)
- defang bogus complaint about overspecified segment info in feedback
- (remove comment left from debug below)

1.87  (10/21/99, amd)
- annotes: added a few (ccannote now just includes annotes for backward compat)
- ccinstance/ccprep: changed to count (and name) instances better
    (rather than unique operators as was before)

1.86  (10/20/99, eylon)
- "-xc" option now flattens states into basic-blocks  (exline still not done)
- bug fixes for array type-checking and symtab scoping

1.85  (10/14/99, amd)
- checked-in ccinstance, ccbody, ccoperator, cccopy
- CC body fix: segment initiation and maybe naming
- added operator "terminate" (also "copy")
- instance support for copy/terminate, and fixed segment numbering
    (problem remains with segment counting)
    (that's ccinstance, ccbody, and operator not ccoperator)
- added cccopy to makefile

1.84  (10/14/99, eylon)
- "if" now checks type of condition (must be boolean)
- issue warnings for stmt calls which ignore operator's return value

1.83  (10/14/99, eylon)
- new "-xc" option enables canonical-form xforms (exlining, etc.),
    presently implemented only lifting inlined-calls/array-refs to stmt level
- "-etdf" option now prints type predicates as TDF comments
- CC backend: added missing segment cast

1.82  (9/8/99, eylon)
- reduced memory footprint and runtime by 100x!
    (changed LEDA implementation of Tree annotations from "map" to "d_array")

1.81  (9/2/99, eylon)
- fixed type-checking bug for cat() builtin  (amd spotted early exit)

1.80  (9/2/99, eylon)
- type regeneration now catches internal type-checking errors (fatal)
- fixed bug in mapping over memory-array refs, where gc corrupted the refs
- fixed bug in constant-folding type predicates
- cosmetic changes to reduce g++ warnings when compiling tdfc

1.79  (9/1/99, eylon)
- garbage collector is now more careful to not delete built-in
  operators, exprs, types, symbols

1.78  (8/31/99, amd)
- added support to delete types, even at stmt level at several points

1.77  (8/31/99, amd)
- added support to frontend (operator, expr) for a builtin copy
  (wanted to go ahead and check this back in now so we stay in-sync
     ...also think eylon may want to look at typechecking for cat)

1.76  (8/30/99, amd)
- ccbody/define segments in proc_run/compose case
- some tweaks to instance in preparation for overhaul (again)

1.75  (8/27/99, amd)
- add -dps option
- ccbody/create result

1.74  (8/27/99, eylon)
- new "-gc" option for AST garbage-collection
    (we presently gc only after type-checking & instance generation,
     could be more frequent to save more memory)
- Type info may now be manually deleted, regenerated on-demand
    (we presently do not actually delete types anywhere)
- relaxed initial-value assignment of too-wide a word from fatal to warning

1.73  (8/26/99, amd)
- cosmetic typkind->int casts 
- add id->obj cast to stream types
- work on typing and constructors for segments
- start tracking return/output value assignment problems in instance

1.72  (8/26/99, rhuang)
- initial fix of typed Segments in cctypes.h

1.71  (8/24/99, amd)
- support typed ScoreSegments
- fixed minor spacing problem w/ segments
- reverted code added in 1.69 which screwed up stream creation for
   instances
- fixed result stream creation for proc_run
- removed local typedecleration in Page code which shadowed 
   object state (thanks Michael)

1.70  (8/20/99, amd)
- Makefile update for linux (3.8a) LEDA
- ccinstance generated lingering reference to group, changed to page_group

1.69  (8/19/99, amd)
- better support segments in cc backend
  + deal with EXPR_ARRAY and TYPE_ARRAY
  + deal with passing/typing/creating/initialization
  + support to distinguish params which don't cause instances
- support hex constants in frontend

1.68  (8/5/99, amd)
- make read values persistent (always stash in retime array and read from there)
- cleanup some constant folding
- cleanup/fold bit select code
- add segment start (well, start adding, it's problematic as it is now)

1.67  (8/5/99, eylon)
- allow more expressions in compositional operators
    (warning, did not implement binding-time analysis to allow only
     expressions with bound arguments, i.e. consts/params)

1.66  (8/5/99, eylon)
- fixed prototypes of builtin segment operators (R,W ports were reversed)

1.65  (8/5/99, eylon)
- fixed AST node-sharing bugs (in Expr type-checking)
- fixed inlined-call type-checking to regenerate cleared types
- fixed disallowing of input-streams as inlined-call args in compositional ops

1.64  (8/5/99, eylon)
- "-pm" memory-profiling option now reads process memory, not system memory

1.63  (8/4/99, eylon)
- "-dshared" debug option finds shared AST nodes
- fixed tons of AST node-sharing bugs (more pending)
- fixed various expression type-checking bugs
- fixed generation of type-predicates (now at AST leaves only, not propagated)
- changed handling of type-widths to improve type-checking, sharing, memory use
    * Type with constant width does not keep widthExpr
       (Type::getWidth()>=0 --> Type::getWidthExpr()==NULL)
    * Type::makeWidthExpr() generates unique copy of widthExpr on demand
    * similarly for TypeFixed::makeIntWidthExpr,
                    TypeFixed::makeFracWidthExpr,
                    TypeArray::makeNelemsExpr
- added non-recursive AST destructors (prep for discarding Types, GC)

1.62  (8/3/99, amd)
- comparison test for builtin types in instance flatten expr was faulty
- minor change, but once added compiles stuff in cctest w/out crashing

1.61  (8/3/99, amd)
- more work on instance resolution
  (dealing with null environments, getting args evaluate in right
     environment)
- initial attempt at segment handling code in instance generation
- fixed bug where param assignments all going to param[0]
- feedback file reading deal with 0x specification for values
- fix constant folding bug which led to count+1 being reduced to count
- minor fix to expr typing on builtins to not trash type

1.60  (8/2/99, eylon)
- clearType() AST method clears ptrs to memoized types
    (presently for mem profiling; must GC unreferenced types elsewhere)

1.59  (7/30/99, eylon)
- support built-in segment operators (CMBs) in compositional operators

1.58  (7/29/99, amd)
- added linux specific memory usage monitor
- added tree node option to timestamp
- revamped bindvalues to use map2
- fixed a couple of problems in ccinstance pagestep code generation
  + mismatch in assignments of numbers to done_ eos_
  + using wrong return value to create bogus return stream

1.57  (7/28/99, eylon/amd)
- "-pm" for memory profiling, "-pt" for time profiling
- variable duplication bug (eylon)
- "type vs. symtab" bug in instance.cc (amd)

1.56  (7/28/99, amd)
- include operator name in proc_run verbose (-dpr) state print

1.55  (7/28/99, amd)
- generate .instances file
- hoist flattening out of instance generation loop
- continue to work on / revamp resolution in instance.cc
  fixed many problems with things being evaluated in wrong environment

1.54  (7/26/99, eylon)
- fixed more type-sharing bugs

1.53  (7/22/99, amd)
- added flag (-dpr) to generate verbose information in proc_run state
    translation (prints out state name and input values on each transition)
- worked on builtin_cat in bindvalues (looks like there's still some
    duplicate/relink problems)
- instance.cc needed some builtin_cat/widthof/bitsof support

1.52  (7/22/99, amd)
- proper handling of result value in ccbody (processor side compose)
  create result stream when necessary
  rename functional largs to result when necessary
- with this done, everything currently in cctest compiles to a .o file
  (master file)

1.51  (7/21/99, amd)
- revamp instance generation to avoid symbol proliferation
  (do own interpretation and assignment as much as possible)
  uses map2
- add support for builtin_done and builtin_cat
- feedback deal more robustly when fgets not return data
- with this done, everything produces .cc files w/out crashing

1.50  (7/21/99, eylon)
- fixed duplication of builtin calls (ExprBuiltin)

1.49  (7/21/99, eylon)
- AST threading is now done in node constructors, obviating thread() method
- fixed more type-sharing bugs
- documented map, map2

1.48  (7/21/99, eylon)
- bug-fixes for map2

1.47  (7/21/99, eylon)
- overhauled map2  (now allows changing AST node to different kind)
- version string is now "unknown" if Makefile cannot figure it out

1.46  (7/20/99, eylon)
- support builtin done()
- support implicit 0-width sync tokens for inlined calls with no arguments
- fixed more duplication & link bugs

1.45  (7/20/99, eylon)
- version string is now in macro TDFC_VERSION,
    not dynamic file "version.h", so make recompiles fewer files
    (version computed by Makefile, default "unknown" in new file version.h)

1.44  (7/19/99, amd)
- minor fix in ccinstance to guard retime structure consistently between 
  declaration and use (pretty minor change, but is orthogonal to what
  I expect to be toucing next)

1.43  (7/16/99, jyeh)
- added basic code for rate identification
- should be stable in the sense of no compile time errors
- or run-time core dumps but still a lot a work to do

1.42  (7/15/99, eylon)
- fixed cases of shared Type nodes (see 1.36 comments)
- added beta-renaming (replace formals by actuals) in return-type
    of inlined calls  (no beta-renaming of streams during type-checking yet)
- new Tree::map2() allows walking + modifying AST in-place
- "-ecc" now renames vars of all operators before any further processing

1.41  (7/14/99, amd)
- added code (mostly in ccinstance) to handle retiming for page step code
- fixed some symbol finding in bindvalue
- added lots of noise/debug to instance generation (for now)
- comment out debug noise in cctype

1.40  (7/13/99, amd)
- added code (mostly in ccbody) to handle retiming for proc_run case
- also split out case header common code into cccase

1.39  (7/12/99, amd)
- add code in C++ generation to rename variables to avoid conflicts with
  variables added during compilation and C++ keywords
- added timestamps code

1.38  (7/6/99, amd)
- fixed up stuff so output instance c++ code will compile
   (mostly moving around initialization into constructor)
- fixed fencepost error on default eos
- fixed produce value initialization problem
- made ccstmt take a state_prefix argument

1.37  (7/5/99, amd)
- fix check eos default exit at end of explicit state signatures
- fix (put in scheme to fix, anyway) problem with return value/call arg
  types being resolved in wrong environment ... try to rename to 
  syms to acount for formal/actual assignment
- bit of cosmetic work on alignment
- added ifdefs for cygwin, tweaked makefile for curret linux/scratch setup 
 
1.36  (7/2/99, eylon)
- Type nodes in AST are no longer shared  (was problematic for duplication)

1.35  (7/1/99, amd)
- cleanup spacing in ccinstance (pagestep code)
- force more things through EvalExpr before passing to ccEvalExpr
- broaden instance compose exprs (allow values), more useful error message
- reunited return sym and function name

1.34  (7/1/99, eylon/amd)
- more fixes for duplicate+link of symbols, symtabs  (obviated amd's rebind.cc)
- compute produces/consumes for page generation [untested] (amd)
- NULL-else bug fix (procrun) (amd)
- compose exprs allow math (assume typechecked elsewhere) (amd)
- cast parens balance (amd)

1.33  (7/1/99, eylon)
- fixed duplicate+link for child symbol tables

1.32  (6/30/99, amd)
- checked in changes w/ pagestep code addition
- code compiles, but no run tests
- bindvalue work
- before integrating eylon's 1.31 changes

1.31  (6/30/99, eylon)
- fixed AST duplicate+link to support operator instantiation ("unique dup")

1.30  (6/29/99, amd)
- [some 1.30 files checked in as 1.29]
- first pass at bindvalues
- some code in ccinstance to declare top level variables in page

1.29  (6/29/99, eylon)
- fixed bug where binary operator '%' did not print in TDF output
- fixed bug which reversed high/low bit-range-positions in arrays (x[1][2:3])
- TDF output now has smarter (fewer) parentheses
- more type-checking for arrays
- [see 1.30]

1.28  (6/28/99, amd)
- catchup getValues->getValue (for scalars, still not handle memories)
- some work dealing with cygwin32 (most of it still comments of what will
     need to change)
- work cleaning up flattening/duplication in instance generation

1.27  (6/24/99, eylon)
- added support for memory arrays
- fixed initial values for variables

1.26  (6/17/99, eylon)
- fixed bugs in AST duplication (forgot to set parent pointers)

1.25  (6/17/99, eylon)
- fixed bugs in operator duplication

1.24  (6/16/99, eylon)
- fixed handling of "stay"  (stored internally as "goto own-state")
- fixed seg-faulting bugs due to copying NULL tokens  (tokens need not copy)
- made file locking more portable  (no flock() on Solaris; use lockf())

1.23  (6/16/99, amd)
- added first pass at flattenning code (needs debug)
- started backend ScoreOperatorInstance/ScorePage generation 

1.22  (6/15/99, amd)
- added frontend stuff to read feedback files and sort out
     instances that need to be generated
- probably some cleaning up of C++ code generation

1.21  (6/11/99, eylon)
- g++ version used for preprocessing is now macro "CXX" set by Makefile
    (still looked up thru $PATH, default "g++")
- added option '-I<dir>' for preprocessing

1.20  (6/11/99, eylon)
- now invokes C-preprocessor via g++
- added options '-E', '-D...' for preprocessing
- added options '-etdf' '-ecc' to emit target language ('-ecc' engine by amd)
- examples:  modified Makefile, renamed source files '.d', '.pd' to '.tdf'

1.19  (6/10/99, amd)
- internal changes in '-ecc' engine

1.18  (6/10/99, eylon)
- fixed type-checking of operator '?'
- fixed creation and printing of variable symbol-tables 

1.17  (6/9/99, eylon/amd)
- behavioral operators now remember their start-state (eylon)
- support for renaming symbols (vars,streams,params,operators,states) (eylon)
- separated build directories:  src, architecture-build, examples, etc. (amd)
- modified Makefile for multiple architectures (working: sparc, linux) (amd)

1.16  (6/3/99, eylon)
- restructured #includes
    (moved inclusion of all AST stuff from "parse.h" to new "treeincl.h")
- fixed minor bug which threw away constants

1.15  (6/2/99, amd)
- modified Makefile to build on Solaris 2.5 (previously only 2.6, 2.7)

1.14  (4/21/99, eylon)
- changed compiler name to "tdfc"  ("dfrtl"-->"tdf" in text & file names)
- new "eos()" syntax/semantics  (now a token in state signatures)
- added mechanism to map a funct over a tree (map())
- fixed version info ("-v") to get version number from src files, not RCS files
- added U.C.Regents disclaimer and RCS revision tag to all source files

1.13  (4/18/99, eylon)
- added more type-checking code to detect illegal stream usage (e.g. wrong dir)
- fixed compositional operator syntax to accept pass-through assignment "a=b"
- version info ("-v") now reflects all source files
    (scans RCS files, stored in automatically-created file "version.h")

1.12  (4/16/99, eylon)
- added type-checking
- added "-q" quiet option to supress warnings
- removed lots of old, commented-out type-checking code (from v1.5)

1.11  (4/11/99, eylon)
- added boolean constants "true", "false" to lexer
- fixed parser actions which generated reversed lists (e.g. arg lists)
- removed syntax for typeless arg-decls (e.g. "param n")
    to avoid confusion regarding implicit type
- activated+debugged some unactivated linking code
- added code to print AST
- added code to compare exprs, types, syms
- tokens now point into a set of file names instead of duplicating strings

1.10  (4/7/99, eylon)
- added "-v" option to front end to show revision info
    (only shows revision for dfrtl.cc, not across all files)

1.9  (4/7/99, eylon)
- created superclass "Tree" for all parts of AST,
    providing homogeneous access to token, parent, symtab scope, etc.
- added wildcard type for unlinked & polymorphic types (in builtins),
    also nil type for non-typed tree nodes
- added linking code with minimal semantic checks
    (e.g. disallow certain stmts/expr in compositional operators)
- dfrtl_init() now handles all initialization of pre-defined objects

1.8  (4/2/99, eylon)
- global "gSuite" of new class "Suite" contains all known operators
- converted several list<> implementations to set<> for performance

1.7  (4/1/99, eylon)
- parser now constructs abstract syntax tree (AST),
    with minimal semantic checks during parse
- lexer/parser returns real tokens
- introduced dfrtl_init(), dfrtl_exit(),
    for now just to initialize builtin operators

1.6  (3/24/99, eylon)
- fixed expression grammar to be more C-like.
    now allows mixing of boolean and numeric expressions.
    note:  unlike C, TDF statements (e.g. assignment) do NOT have values.

1.5  (3/24/99, eylon)
- added lots of preliminary, commented-out linking/type-checking code
- filled in some interface details of parser classes
- added dependencies Makefile

1.4  (3/15/99, eylon)
- overhauled grammar
- first try at parser classes in separate operator/state/stmt/expr files

1.1-1.3  (amd)
- Andre's initial work