cow3.txt

Dos6.0

Introduction

This document describes implementation details of modules of CW.
Applications should not depend upon anything contained herein.


SWAPPING
--------

The CODE SWAPPING/LRU scheme used by CW is significantly different than
Windows.
The concepts as used in CW are described below and notably how they differ
from Windows.
This is not an attempt to document the Windows system.

THUNKS
------

The thunks used for CW are shorter (6 vs 10 bytes) than Windows,
do not require the INT 3FH handler
and implement a more efficient form of LRU.
Noteably the SAR CS:[xxxx],1 instruction is not present in CW

Thunks may take one of three forms:

FORM 1 (ENTMOVE) : Resident code segment.
	JMPF	seg:off
	DB	segno
  -- action of jumping to thunk:
	jump to resident program

FORM 2 (ENTMOVE1) : Resident 
	CALLN	relative to JMPF RelruSegment
	DW	offset
	DB	segno
  -- action of jumping to thunk:
	update LRU for segment, change ENTMOVE1 -> ENTMOVE
	then jump to resident program

FORM 3 (ENTMOVE2) : Not-resident
	CALLN	relative to JMPF RelruSegment
	DW	offset
	DB	segno
  -- action of jumping to thunk:
	load in segment, change ENTMOVE1 -> ENTMOVE
	then jump to resident program



LRU
---

LRU for discardable (i.e. moveable) code segments is defined completely
by the segment reference bytes (1 per code segment).

The following values for these bytes are used:
	0	- Most recently used
	1..0x7f	- less recently used
	0x80	- non-resident (i.e. discarded) code segment
	0xFF	- FIXED segment

The "ReLru" process only occurs for resident segments (i.e. segrefbyte <= 0x7f).
This consists of setting the proper segrefbyte to 0 indicating that this
segment is the Most Recently Used segment. The "Relru" process is initiated
by the program calling a thunk of FORM2.

The "Sweep" process occurs periodically (currently about every 1/4 of a second)
and asynchronously to the main program.
The "Sweep" process consists of bumping the segref bytes in the range 0 .. 0x7e
(this keeps 0x7f .. 0xff from being changed), and going through the entry
table (i.e. the list of thunks) and changing ENTMOVE (FORM1) to ENTMOVE (FORM2)
thunks.

Question : Why are we doing this ?
Answer : the Window's LRU system puts a SAR CS:[xxxx],1 instruction at the
start of each thunk (requiring 31 cycles on an 8088 - more than tripling the
time it takes for a thunk).

So far all the changes in the LRU scheme have been improvements over the
Windows LRU. The thing involved with LRU is the determination of which
segments are the least recently used. This is where the CW system suffers
(i.e. is less efficient than Windows). In windows (due to the fact Windows
updates a linked LRU list) it is just a matter of removing the head item
from the lru linked list. For CW the extra work of scanning the array
of segref bytes is needed.
This should happens relatively rarely (i.e. over a long session,
each discard will correspond to re-reading in the code -
making this an already slow process).


RETURN THUNKS
-------------

Return thunks are also dissimilar from Windows.
First, they are of a different form (ENTRET) :
	CALLN relative to JMPF ReturnThunk
	DB segno
	DW offset

Note : these are kept in a table (pretthunks). There is one per code segment
(even for FIXED segments, in which case it is not used).

The segno is the segment number (and is used since dividing by 5 is slow on
the 8086).
The offset is either -1 (0xffff) or a valid offset:
	if offset == 0xffff then this means that either:
		a) the segment is resident
	   or   b) the segment is not resident, and there are no returns
			to that segment on the stack.
	if offset != 0xffff then this means that the segment IS NOT resident
		AND the