Fri Jun 11 16:13:09 BST 1999

Working on nested procedures, two approaches, either:
we work with nested NewLocalVar quadruples or we unravel
the nested NewLocalVars. Unraveling procedures seems attactive
as a mechanism to move quadruples could be later extened to
inline procedures. Working with nested NewLocalVars is a
possibility but may cause several optimization implementations
to break in the process!
----------------------------------------------------------------
Tue Feb  3 15:00:25 GMT 1998

Nearly finished the pipeline code optimization (-Oi586) together
with pentium timings for the apu code generator. Pentium pipeline
optimization improves some loops by up to 33%.
----------------------------------------------------------------
Mon Jul  7 08:40:51 BST 1997

Dynamic code generator now used during stage3 compiler
construction. It also can be used to build the realtime
systems and simulator tools.
A few fixes to the compiler code generator have taken place
since the last log. Better error messages have also been added.
----------------------------------------------------------------
Wed Mar 12 15:53:52 GMT 1997

Compiler now compiles itself, minor inefficienies exist.
Would be good to remove them so thay do not bite later.
----------------------------------------------------------------
Fri Mar  7 10:08:40 GMT 1997

M2GenDyn486 completed and very nearly debugged, one module
M2GenDyn486 remains to execute successfully. Recursive debugging code...
Dynamic code generator, compares favourably with gcc for some modules,
(NameKey gm2 produces better code) for others gcc wins hands down.
Still dynamic code generator produces half the code of the static version.
----------------------------------------------------------------
Tue Nov 19 14:33:03 GMT 1996

Finally completed IndrX and XIndr operators after many alterations
to M2Quads and M2Gen487. In a much better position to attack the gencode
algorithm outlined in the Dragon Book.

----------------------------------------------------------------
Fri Nov  1 08:57:32 GMT 1996

Going to change ModeOfAddr to Values = LValue, RValue
Going to change quadruple meaning as above, only indirection via
Xindir  and indirX  (where X = *)

LeftValue  == ValueAddr
RightValue == OffsetAddr

ParamOp now is mode less.
AddrOp  is only used in two places in M2Quads
need to alter CopyOp  -> indirX and Xindir
     achieve this via implementing a procedure called MakeIndirectX
     and MakeXIndirect
     then remove CopyOp and alter BecomesOp to be mode less.

----------------------------------------------------------------
considering altering the set of quadruples. In particular I would
like to remove the CopyOp and replace it with two quadruples:

   *indir   a   b     meaning   *a = b

   indir*   a   b     meaning    a = *b

and I'd like to remove OffsetAddr & ValueAddr  (ModeOfAddr)
This would make the CSE more efficient and also enable a much
better code generation mechanism (aka GenCode in Dragon Book).
All other quadruples would NOT depend upon the mode of a variable.

Although the WITH, array access, VAR parameters, pointers
need to be throught out very carefully before embarking on
a major alteration. I'm not sure whether I can remove OffsetAddr
and ValueAddr yet, it maybe that these are necessary (LValue and RValue).

BaseOp now computes the constant offset from the ADR(array), this
is much better than computing the ADR(array)+-offset as the constant
calculation can be done independantly of the runtime ADR(array).
Thus CSE can manipulate the constant offset and integrate it with
other constants in the DAG.

Done
----------------------------------------------------------------
