Digital-Research-Source-Code/CPM OPERATING SYSTEMS/CPM 68K/1.0X SOURCES/v103/cgen/log

.he "C68 Code Generator"Change Log"Page %"
.de bg
.sp
.in +5
..
.de eg
.sp
.in -5
..
1. 7/20/82 - Post fix operators
.bg
Post fix operators were being handled a la PDP-11, i.e. if( x++ ) ...
would be true if the result of the post-increment were true rather than
the original value of x.
Also, if( --x ) ... would generate an unnecessary test instruction.
Added scodegen (delay???) routine to pick off post-fix operators,
call codegen to generate the code for the pruned tree, then generate code
for the post-fix operators.
This will generate horrendous code for something like: if( x[i]++ );
but t.s.
.eg
2. 7/21/82 - divide by long constant
.bg
The construct: x / 10L did not generate a ldiv call because long constants
Sethy-Ullman numbers are being calculated normally.
Added DCLONG operator and check in canon to check for this case.
This prevents sucomp from incorrectly computing the Sethy-Ullman number.
.eg
3. 7/23/82 - shifting long or byte address by 1
.bg
The construct: x =<< 1 generated an invalid assembly instruction if x
was an automatic or static byte or long.
This is because you may only shift words in memory (The 68000
clone strikes again!!!).
Changed code skeletons for eqshft to only match for addressable words.
.eg
4. 8/7/82 - complex long divide/modulus expressions
.bg
Complex expressions involving long over short divides or modulus would
not work, not consistently sign extending word result.
Changed test in expand to check for DIV or MOD operators in long tree
and set extf flag.
.eg
5. 8/7/82 - add of quick negative did not generate subq
.bg
The expression: x = y - 1; with x & y registers generated an add #-1,Rx
rather than suq #1,Rx.  This was because rcgen was turning ADD into EQADD.
Added test in ucodegen for both ADD and EQADD.
.eg
6. 8/12/82 - multiply(divide) by -1
.bg
The expression: x =* -1; was being "optimized" to negate x, however it
did not store the result in x.
This is because of multiply/divide by -1 optimization in canon not
checking for =* or =/.
Added checks.
.eg
7. 8/27/82 - character array assigned function value
.bg
The expression: x[i] = f(i); did not generate correct code, the address
register for x[i] was computed, but not saved on stack, then f was called.
This resulted in all further function calls using the first argument
push rather than a move.
In ucodegen, when optimizing for address register shifted left by 1 or 2
did not do fixresult call which would have saved register on stack.
Added this call.
Also added code skeleton to compile the function expression to the stack,
then move that directly to the computed address.
.eg
8. 9/1/82 - address-indexed addressing mode used .b index
.bg
If the index register used in an address-indexed addressing mode
operand was byte, the index register was given a .b extension even
though it had correctly extended the register to a word.
In outaexpr for INDEXED addressing mode an outatype of the index
type should have been used rather than outtype.
.eg
9. 9/1/82 - x =* -1, x =/ -1, x=^ -1 optimizations
.bg
Added EQNEG and EQNOT operators and added checks in canon so that
x =* -1 and x =/ -1 are changed to neg operations and x =^ -1 is
changed to not operation.
.eg
10. 9/15/82 - function redeclaration bug
.bg
If an external function was redeclared in a function body, a
redeclaration error occured.
This was fixed in dodecl by checking for a function of AUTO type
and changing it to a function of EXTERNAL type.
.eg
11. 9/15/82 - Hard long to register fix
.bg
The syntax: b =* x; where b was a long register would give an error
"hard long to register".
This was because the lamul function expected to do the assignment to
a memory location.
Changed canon so that hard =*, =/ and =% would be changed to use
the normal long multiply and divide routines, and then do an assignment.
.eg
12. 9/15/82 - Casting Problems (Not fixed)
.bg
The syntax: i = (char) i; where i is not a register would result in
i being assigned the high-order byte of i.
This also is a problem for any cast where the cast type is smaller
than the type being cast.
Generally this should be solved by moving the expression into a data
register, and then obtaining the result from there.
Added DOCAST operator to handle these problems.
However, this is all tightly interwined with conversions in the parser.
.eg
13. 11/24/82 - More than two constants couldn't be combined
.bg
Constant optimization failed to mathematically compute constant
arithmetic fully (the results were lost) thus yielding the result
as the computation of the last two constants in the expression
only.
.eg
14. 12/13/82 - structure assignments
.bg
One structure may be assigned to another provided that the one on
the right is smaller than or equal to the left one in size.
.eg
15. 1/10/83 - floating point
.bg
Floating point handling including constants, conversions and operations
added.
.eg
16.	1/14/83 - pop stack instruction
.bg
The popstack instruction of 1..8 was turned into an addq.w which only
affected half of the sp register.  The "add #d,sp" has been replaced
by either "addq.l #d,sp" or "adda.l #d,sp".
.eg
17.	1/14/83 - optimization eqmult, eqdiv, eqmod
.bg
Generated a new code skeleton entry to deal with register integer mults,
divs, and mods (eg. register int i; i =* 34;).
.eg
18.	1/17/83 - return of 0L
.bg
Code skeletons generated a clr R0 regardless of operand type.
.eg
19. 1/21/83 - prefix ++, -- problems with bytes
.bg
The code skeletons for fr_eqop were a total mess, particularly for
the left hand side being a character.
Rewrote all fr_eqop and fr_eqmult code skeletons to handle these cases.
Unfortunately rcgen also gets into the act turning all =+, =-, =&, =| and
pre++, pre-- compiled for register into compiled for effect then returning
the left hand side symbol.
This is probably OK for word and long LHS but for byte it results in
lost precision.
Changed rcgen to check for CHAR LHS and not to compile for effect if so.
.eg
20. 2/2/83 - long eq_shift int
.bg
Bad code was generated for long's left shifted by integer values.
.eg
21. 2/23/83 - longs divided by integers
.bg
Ext.l was not always done.
.eg
22.  3/1/83 - casting to char
.bg
Added the docast operator, forces the variable to be put into a register
and an ext.w to be acted upon it.
.eg
23.  3/2/83 - casting register char * to int
.bg
Bad code was generated for casting a character to an integer if the 
character was pointed to by a register character pointer.
.eg
24. 3/3/83 - Compiling complex indirect into A Reg.
.bg
The construct: *p->q->r generated an intermediate load into a D reg
then move to A reg, rather than compiling directly into A reg.
This was because of some hysterical PDP-11 code in expand checking
for whether an A register is appropriate.
.eg
25. 3/3/83 - dbra instruction
.bg
The construct:  if( --i != -1 ) goto x; now generates a dbra i,x
instruction.  This combined with the new for and while loops easily
enables you to generate dbra instructions for loops.  This affected
code in condbr to check for the specific expression tree above.
.eg
26. 3/3/83 - Bit field assigned to bit field
.bg
The construct:  x.bf1 = x.bf2 failed due to fixbfield not fixing up
the RHS of the expression.  Added recursive call to fixbfield to
fix up RHS's of equal operators.
.eg
27.  3/4/83 - Post increment with ?: op
.bg
Post increment and post decrement operations can not be delayed in the
case of a ?: operation.  The post operation was taking place regardless
of whether the case in question was executed.
.eg
28.  3/7/83 - long div/mod/mul/sub/add const 1L or const 0L
.bg
Added optimization handling for long constant zero and one handling.
An error is now generated is a mod or div 0 is found.  Add/Sub of
a consant 0 does not generate any code.  Mult/Div 1 does not generate
any code
.eg
29.  3/8/83 - eqmod 1 and eqmult 0
.bg
Incorrect storage of simple expressions eqmod 1 and eqmult 0 which
result in a zero store.  Also added for the long case.
.eg
30.  3/18/83 - reg ptr assign and character compare
.bg
The statement if (*(regptr = str)); did not do the test for the character
pointed to but rather the result of the assignment.  This was caused
by a codegen optimization which should not include a-registers.
.eg
31. 3/21/83 - Complex assignment to character array
.bg
The expression: ca[i] = f(); would move i to the stack as a word and pop
it as a long.
This was caused by the code skeleton macro (S_FORSTACK + S_INDR) assuming
what was going on the stack will always be a long.
Did quick and dirty fix in expand to extend an int sub-tree to a long
if compiling subtree FORSTACK+INDR.
.eg
32.  4/4/83 - unary operations on address registers			[vlh] 4.1
.bg
Operations involving unary operations on addresses was not forcing the
use of a temporary data register to do the actual operation.  This
was causing bad code to be generated for certain expressions, eg.
array[-i], or array[~i].
.eg
33.  5/23/83 - long constant optimizations						[vlh] 4.1
.bg
Changed all the constant optimizations to handle longs as well as
ints.  This will make life more difficult on the pdp-11 but makes 
for more optimal code (eg. l *= 2...).
.eg
34.  5/25/83 - pre-operators over &&, ||, ?						[vlh] 4.1
.bg
Preincrement/decrement were not working correctly over &&, || and ?
operators.
This was due to rcgen generating code for preinc/dec even if && or ||
was seen in tree.
Added test in rcodegen to stop trying for register optimizations over &&, ||
and ?.
.eg
35. 5/25/83 - post-operators over &&, ||.						[vlh] 4.1
.bg
Postincrement/decrement operators were not working correctly over && and ||.
This was due to scodegen/addptree pulling out all post-operators over && and
|| operators.  Added check in addptree to stop at && and ||.
Condbr then does an scodegen over the && and || to do the post-operators.
.eg
36. 5/25/83 - post-increments over assigns and autoinc/dec		[mc] 4.1
.bg
Postincrement/decrement over expressions involving assignment operators and
autoinc/autodec did not work correctly. For example: c[i=+1]++ and c[*p++]++;
This was because the sub-expression tree which is copied did not adjust for
things which are being altered in the sub-expression.
Fixed tcopy, addptree and fixbfield to fix this problem.
This is not guaranteed to work in extremely complicated code involving
assignment and post-op operators, for example, something like:
c[c[i=+1]++]++;  And is especially not guaranteed to work with bit-fields
in something like: (c[i=+1]++)->bitfield.
.eg
37. 6/1/83 - long multiplication/division as array subscripts	[vlh] 4.1
.bg
Array subscripts which involved long operations which required calls to function
routines expected the results to be left in the first address register rather
than the first data register.  This caused bad code to be generated.
.eg
38. 6/1/83 - allow long array indices							[vlh] 4.1
.bg
In order to generate arrays which are as big as memory the arrays must be larger
than an int in side we need to also allow array indices to be longs.  This was
tied in with the long constant evaluation.
.eg
39. 6/21/83 - link file                                         [vlh] 4.2
.bg
Added the link file as an input file, this is necessary for auto inits
and block variables.
Added the symbol "%" as a readicode symbol to signify the need to pull
an entry out of the link file which is being created by the parser.
.eg
40. 7/6/83 - dbra instruction                                         4.2
.bg
Fixed the dbra conditional test to not be used in the case of a complex
expression (eg. while(--regi != -1 && *l != '\n'); )..	
.eg
41. 7/11/83 - if test float 									[vlh] 4.2
.bg
The testing for float true was not being handled.  The construction
if (fp); now does a test for the floating point value being a non zero
32 bit value.  This is taking advantage of the fact that both floating
point constructs denote the value zero as a zero.
.eg
42. 7/13/83 - short vs. long vs. int							[vlh] 4.2
.bg
For the sake of portability and the likelihood of generating the same
code on the vax which supports a 16 bit short, a 32 bit int and a 32
bit long versus Regulus which supports a 16 bit short, a 16 bit int and
a 32 bit long the code generator has been changed to use only longs
and shorts on the theory that most machines support 16 bit shorts and
32 bit longs.
.eg
43. 7/20/83 - symbolic debugger flag							[vlh] 4.2
.bg
Added the '-g' flag.  Required a change to the outline routine to outline
cdb type line number identifying lables rather than the standard sort.
Also altered readicode to output a line number identifier on a null
expression (allows a line number to be placed previous to asm() code
as well as loops).
.eg
44. 7/22/83 - Suppress optimization	which gens bad code			[vlh] 4.2
.bg
Changed rcodegen to not do the rcgen call if the operator is less-than,
or greater-than.  This keeps certain expressions from being evaluated
in the wrong order.  In particular ((regi=f(3)) > f(1)).
.eg
45. 7/25/83 - unsigned char/long								[vlh] 4.2
.bg
Initial work required to generate code for unsigned char and unsigned
long types.
.eg
46. 7/26/83 - preinc/dec, postinc/dec floats					[vlh] 4.2
.bg
Added OPCALL entries for preinc/predec and postinc/postdec.
.eg
47. 8/2/83 - bad optimizations									[vlh] 4.2
.bg
An expression anded (&) with a 0 or an expression multiplied (*) by a
zero in which the expression has side effects (eg. function call which
changes global variables) would cause bad code to be generated.  This
has been a problem in a number of other compilers and was noted on the
usenet by DMR himself.
.eg
48. 8/3/83 - 68010 changes										[vlh] 4.2
.bg
Added the 68010 flag (called m68010).  Code now generates a "move cc"
instead of "move sr" in scodegen if the 68010 flag is specified.  An
alternative routine fakecmpm has been added because the 68010 does not 
properly handle the cmpm instruction.
.eg
49. 8/9/83 - source file										[vlh] 4.2
.bg
Fixed readicode to pick up a source filename as well as the current line
number, for use in error reporting.
.eg
50. 8/10/83 - int cast pushed on stack as long					[vlh] 4.2
.bg
Fixed match() so that it would not optimize out conversion operators on 
items destined to be put on the stack (eg. f(((int)l)&0xff);).  Required
adding a test for cookie != FORSP.
.eg
51. 8/23/83 - long indexes										[vlh] 4.2
.bg
Fixed optim to not optimize long constant offsets into arrays into 16 bit
displacements.
.eg
52. 8/26/83 - constant int/long assigned to char				[vlh] 4.2
.bg
Added code to optim to truncate Constant integers and longs which were
being assigned to character values via ops: equal, eqadd, eqsub, eqand, 
eqor, eqxor.
.eg
53. 8/30/83 - if (main)...										[vlh] 4.2
.bg
Altered outaexpr to not generate the '#' preceding a symbol expression
which does not have an offset and is and external or static variable
being generated for condition codes.
.eg