Trailing-Edge
-
PDP-10 Archives
-
SRI_NIC_PERM_FS_1_19910112
-
c/kcc5/notes.txt
There are 5 other files named notes.txt in the archive. Click here to see a list.
NOTES.TXT
This file is a collection of various notes and thoughts about possible
improvements to the KCC C compiler, retained for posterity or future
consideration. Remember that existence of a item in this file does
NOT necessarily mean it would be a Good Thing to implement; moreover,
some of the things noted have already been done or fixed.
Unless otherwise specified, most of this stuff is the product of KLH's
imagination. Pre-KLH notes (from KRONJ, etc) take up the last portion.
KCC bugs:
Still loses track of proper line count (for err msgs) when
#asm code occurs in the file. Not serious, but annoying. Fix
when tmpfile() is ready.
- Flush %CHRBP from CCOUT generation and from all library source.
(PERROR, OPEN, READ, UNLINK, WRITE)
LIBC stuff:
- lseek uses SIZEF which may bomb if file is newly created (GTFDB
screws up) or I/O bytesize is different from FDB bytesize.
Invent a new internal file-size function that both fstat() and
lseek() can use.
- Add LCTLECH to the local mode word for ttys (ioctl). Make this
the default at runtime startup. See BUGT20TTY.
- STDIO is still not well described. Need state diagram, especially
to make sense of update streams with ungetc, and
improve fseek/ftell.
- read() is inefficient. Prime candidate for using PMAP.
- getenv() needs to be re-designed. Ideally could invent some scheme
to reliably pass both envp and argp. May need to use
temp disk files.
Todo:
Make CCOUT complain if it tries to expand a skipped immediate op into
a two-word XMOVEI+op.
Try to preclude expansion earlier by changing CAIx R,addr into
CAMx R,[$$SECT,,addr] (for extended machines only). Fix oneinstr(),
perhaps move it to CCOUT.
Fix up CCGEN for MIDAS output. Remove outstrs and printfs,
create "outbset" and "outbend" to bracket "outbyte". Also
create "outblk(size)".
MIDAS barfs on "-$$extsym", specifically ASH R,-$$BSHF. Need to
invent a symbol with negative value.
Fix up users of jsys.h macros (esp C lib) to use new monsym.h.
What about avoiding error msgs in _KCCsymfnd constructs similar
to #if (defined(FOO) && FOO)? Worry about it later.
Replace all #include ""s in LIB with <>s, add -H to -make-.cmd.
Perhaps force CPU and CRT from LIBC to be loaded first, so that
LINK doesn't run out of space trying to store all the fixups awaiting
the symbols they define.
Ideas:
- Beebe would like args of the form "foo.fai" to cause KCC to call
the assembler on those files, just like Un*x does for foo.s.
- More work on PSECT support so can load directly into sect N,
get info from PDV, etc.
- Permit compile-time initializations of the form (int)(ptr) and
(ptr)(int) whereby a pointer is turned into an integer and vice
versa. Since this is a no-op, may as well allow it.
- Have KCC check for doing unsigned compares vs 0, as in
foo < 0, and grumble if so? Also complain if find tests that
will never succeed, such as comparing a masked # vs a larger #.
- Errors pertaining to semantics (not syntax) such as type,
definedness, etc should check for N_UNDEF or N_ERROR and try to avoid
further complaints.
- Check for malloc/realloc/calloc (at end of compilation?) and print
warning if types are not "char *". Nonstandard, but cause of many screws
on PDP-10.
- Find some way to combine both UNIX and OS error numbers into
"errno".
- Add special _exitcontinue(&argc, &argv) function which exits in
such a way that it can return from the call with a new command invocation
stored into argc/argv. All I/O will be reset, etc. This will help
some programs set up lots of initialization and then suspend themselves
for dumping. _exitreenter? _exitsuspend? _exitrestart?
The USYS code needs a special routine like _usreset() which will re-init
all USYS variable stuff -- for example, sigvec(), getpid(), time(), and
the like can all have static data areas that need re-initialization.
Perhaps it would be best to simply put this in the CRT.C code and allow
starting at $INIT to take care of CRT-level mapping/initialization, whereupon
user can dump image; normal start vector then bypasses that stuff.
- Include UNV.C from KCC dir for kicks?
- Change extended-addr startup so that section 3 isn't
automatically allocated; let sbrk() do it if necessary? Or start
allocating in section 2, use some room from top of stack. Idea is to
minimize overhead for usual programs. Provide routines that main() can
call at runtime (or a patch location) so that programs which really need
gigantic amounts of stack space can get it. Note ITS pgms also need a way
to specify stack size somehow.
- Allow malloc to use portions of stack space, to squeeze out
max memory usage?
KCC: C compiler itself
- Array refs ought to be checked where possible, so that arr[2] would
complain if the size was only 2 (but not if the expression was
&arr[2], since ANSI explicitly permits p[size+1] to be a valid pointer)
Hard to do this at parse time since it requires expression
evaluation to get at the constants. Must work with *(p+1) too.
- for #if COMMENT / #endif COMMENT make the non-whitespace in the
#endif do something -- print warning depending on -P level?
- Given #if 0 / ... /#endif, if there is unterminated comment inside
the conditional, complaint at EOF is about unterm condit, not about
unterm comment! Fix this.
- Decl problem: doing "static char foo[];" and then later
a "static char foo[] = { "str", "str" }"
causes an error about multiple definition!
ANSI allows multiple static declarations, CARM mentions
problem on p.56.
- H&S: 8.4 last parag - make KCC complain about inits within compnd
stmts that are switch bodies. Do general check and complain
if any unlabelled code.
- H&S: Clean up generation of declaration initializers.
- H&S: Thoroughly go over all type checking and type conversion stuff.
Ensure all H&S type conversions are applied to all expr ops.
- Handle parameter conversions (sec 9.4).
- H&S: Fix KCC so section 7.10 (constant expressions) is followed.
Currently still too liberal. Invent new N_PCONST type?
(Pointer constant)
- Implement direct .REL file generation. In preparation for this,
must fix up #asm and convert all outxxx() calls into
single-word output calls or equivalent.
- Perhaps allow auto array initialization.
- Perhaps allow union initialization.
- Introduce code for checking funct arg types with formal param types,
will be useful when ANSI comes along.
- Fix compound assignment bug the right way (as opposed to current
kludge using SETM).
Requires BIG revisions to pseudo-code and peepholer.
- Add a "constant" flag to NODE flags? Each operator propagates this
flag if its operands have the flag set. Faster folding and
checking.
- Flush unused error messages from ccerr.h - maybe revamp all
error messages and standardize again.
- Helpful check - if function is "main", ensure return type is
either void or int, and issue warning (not error) otherwise.
- Add option/switch to generate non-extended code only.
Use for KA/KI/KS and some KL programs.
Counterpart switch for extended code only?
- Error handling is really poor. One idea for improving is to
not scan for an expected char, but instead just assume
it is seen, and continue from where we are.
- General code cleanup
Apply "static" to all non-external functions.
Make function definitions consistent -- decide whether
to have type on same line, or on previous line.
"void foo()" vs "void\nfoo()".
Isolate all target-mach dependencies. Add t-m char map table.
Move more defs from cc.h to right places.
Check uses of immedop, directop in CCGEN2
Clean up CCOUT, check for p-code string and do simops diffntly
- Flush "entry" after a while.
- Introduce "$asm();" hack to allow instrs within function body.
Only valid where a statement is legal (or at top level).
$asm("string"); just passes on string.
$asm(OP,REG,ADDR); can interpret/integrate instruction.
OP = actual op code? or string? or KCC token value?
REG = string, or reg value 0-017
ADDR = constant addr expression?
Use semicolon in $asm text to separate instructions, then
can ignore EOL. Problem: multi-word literals. Maybe could
then have #asm turn enclosed stuff into $asm text.
- Find way to speed up struct assignments, or to push on stack faster.
Pass pointer to temporary struct (rather than doing
push/return on stack).
Calling SPUSH/SPOP is slow. Invent a P_PUSHN op? Could
expand into series of PUSH instructions (either to stack
or directly to target) Set cutoff point for calling general
SPUSH. Could even code directly for BLT if both addrs
known to be in local section (simple data). Can do table
entries this way...
- Extend "register" declaration flag to VREG struct?
- Add edit history comments? Ho ho.
-Random idea for assembler collision avoidance: if identifier
symbol will conflict with an asm pseudo, change to prepend
a "." on it (or something).
-Random idea: remember knowledge of vreg or local contents,
ie anything not "volatile", so can use knowledge for
optimization. 8 possibilities plus a constant val:
Unknown (never set), Unknown (but set), Known to be one
of < <= == >= > != relative to constant (often 0).
-Random idea: have way for parser to know if token came from macro
arg, and if it is part of a side-effect expression,
give warning. Tough but handy.
-Not-so-random idea: have way to halt program (specifically KCC)
in such a way it can be dumped as an EXE which when
invoked can proceed from where it left off. Very handy
for generating a version with symtab already initialized!
-Give warning for case of "if (x = 3)". That is, if an assignment
of a constant is being used as a conditional expression.
-Give warning or error for symbols which will cause external
conflicts. If two static symbols conflict in this way, gensym new
unique names? Requires storing 6-char representation, but will
need to do this anyway for .REL file output.
- Note that KCC-extension of `sym` is stored with initial ` as
SPC_char, so that the symbols FOO and `FOO` are not considered
the same. But on output, barf if any 2 idents become the same
as result of transformation to external symbol.
KRONJ et al:
(1) Propagate pseudo-label for return in endlabels
- Change JRST to ret label into a POPJ
- Careful not to try to skip over this
- This can be done by checking in gboolean() and jumping over jump
- Or just always jumping over jump and letting it get optimized
- Maybe have $RETZ, $RETO, $RETT for 0, -1, 1 returns?
(2) Re-install automatic indirection
- Make sure all optimizations safe
- Perhaps undo when offsets other than zero used from same cell
- Don't do it for LSH
(3) Allow assembly code within C sources
- Make this work inside functions
- Define pseudo-macros for args and locals
- Parse and re-emit (binary search on dec20op[] so comment alpha order)
(4) Improve tail recursion for more expressions in places of variables
(5) Fold ops on adjacent bit/byte field together.
- Turn byte field ops into bit field ones where possible.
- Merge adjacent deposits to or loads from literal byte pointers
- More halfword ops
- Change MOVE + ANDI into HRRZ (this actually happens in runtm)
(8) Vectorize named array initialization loops into BLTs.
- Combine BLT loops into bigger BLTs.
- Use register loops for unnamed arrays...
(9) Fix calling conventions for pseudo-instructions ($DMOVE etc)
- Use AC1 and AC2 rather than AC15 and AC16
- Push more of it into code generation rather than code emission
- Use getret/getrpair etc
- This allows more folds, simpler behavior for CSE, etc
(10) Clean up peephole optimizer
- Make sure pushneg() not called in foldboth() when reg later used as index
- Put monument in pushneg() about it not checking for future uses
- Move more optimizations from code gen to optimizer modules (esp code0, 4)
- Break up large hairy messes into simple functions
- Pull comparison operand in code0 across multi-instr code for other op
(11) Improve anti-aliasing
- Byte pointer changing ops can't affect cells used as word pointers
- Indexed expr with offset can't affect labeled cell with smaller offset
(12) More creative uses of the PDP-10 instruction set
- Pull AOS R,S across in foldplus
- Push back AOS over IMULI by doing IDIVI / ASH afterwards
- Use DMOVE, DMOVEM, and EXCH to replace combinations of MOVE and MOVEM
- MOVE R,x + MOVEM R,1(S) => PUSH S,x (choke)
- MOVE R,x + XMOVEI 16,y + CAMx R,16 => XMOVEI R,y + CAMy R,x
- MOVN + MOVEM becomes MOVNM + MOVN (in foldboth; flush MOVN in genrelease)
- CAIE+CAIN+TRNA => CAIL+CAILE even when skipped as long as CAILE reskips
- ADDB + AOS => ADDI + ADDB
- Flush duplicate assignments to same place
- Use ROT for LSH/AND/OR combos
- Fold NEG into COMPL and vice versa when useful
- Fold multiple AND/OR into one; fold binary+unary into appropriate binary
- PUSHJ can be ADJSP 17,1 + JRST or TRNA (both useless)... better tricks?
(13) Clean up runtimes
- Either remove call to jsys() in runtm or use it more regularly
- wait() in runtm can just be a jsys, move rest of mess elsewhere
- Make pipe() know how to use PTYs if PIP: doesn't work
- Teach _gtjfn() about ~username syntax
- Implement berkeley dir stuff (see man 3 directory), ...
- Work out some PRARG% convention to supercede RSCAN% passing if used
(14) Language features
- Structure-structure assignment ($BLT?), vectorizing loops?
- Long variable names
- Warn about local variable shadowing function arg, built-in typedef
- Warn about statements without side effects
(15) Programs
- More Unix programs and filters (as runtimes work or source found)
- Make AS68, LD68, CC68, C2 work (already have CCOM and CPP working)
- Get a shell running
(16) Compiler execution
- REL instead of FAIL (remain able to make FAIL) (long externs optional)
- Code generation uses table of PDP-10 opcode structure by actual op
- Make fixprev() back up to save peephole buffer space
- Allocate various data structures dynamically
(17) Simplified instruction set for non-emission use
- Flush immedop() and directop() in favor of using IMMED forms
Found in CCGEN1: Should be:
CAIG 11,1 CAIL 4,77
CAIN 4,100 CAILE 4,102
JRST $21 CAILE 11,1
CAIE 4,77 JRST $21
CAIN 4,102
JRST $21
CAIN 4,101
JRST $21
CAIE 5,126 CAIL 5,125
CAIN 5,125 CAILE 5,127
JRST $83 JRST $82
CAIE 5,127
JRST $82
$83::
TRNA CAIN 1,2
$155:: JRST $158
MOVEI 1,1 CAIE 1,1
CAIN 1,2 JRST $159
JRST $158 $155::
CAIE 1,1
JRST $159
Found in hock:
$186::
MOVEI 3,1
MOVE 5,-3(17)
MOVE 4,2(5)
MOVEM 3,0(4)
CAML 3,-7(17)
JRST $177 ...
MOVEI 6,60 MOVE 3,-5(17)
ADD 3,-5(17) LDB 10,1(3)
LDB 10,0(3) CAIL 10,60
CAMLE 6,10 CAILE 10,71
JRST $177 JRST $177
MOVE 11,-5(17)
LDB 12,1(11)
CAILE 12,71
JRST $177
PUSH 17,-1(17) SKIPN 4,-2(17)
MOVE 4,0(17) POPJ 17,
JUMPN 4,$356 PUSH 17,4
ADJSP 17,-1
POPJ 17,
$356::
IBP -4(17) CAIN 4,0
JUMPN 4,$923 AOSA 5,-1(17)
SETO 3, TRNA
ADJBP 3,-4(17) JRST $919
MOVEM 3,-4(17) IBP -4(17)
AOS 5,-1(17) JRST $923
JRST $919
MOVNI 4,2 MOVNI 4,2
ADJBP 4,-5(17) ADJBP 4,-5(17)
DPB 3,4 IDPB 3,4
SETZ 5, SETZ 5,
SETO 6, DPB 5,4
ADJBP 6,-5(17)
DPB 5,6
MOVE 3,210(10) MOVE 3,210(10)
MOVE 6,-4(17) MOVE 6,-4(17)
ADD 6,period SKIPE 7,period
ADDB 3,31(6) ADDM 3,31(6)
SKIPN 7,period ADD 6,7
JRST $1334 ADDB 3,31(6)
MOVE 4,210(10)
MOVE 11,-4(17)
ADDB 4,31(11)
$1334::
gettok:
ADJSP 17,2 SETZB 4,ntoks
SETZB 4,ntoks MOVEM 4,nctoks
MOVEM 4,nctoks MOVEM 4,tokind
MOVEM 4,tokind LDB 5,-3(17)
XMOVEI 3,tokbuf CAIN 5,0
IOR 3,$BYTE POPJ 17,
MOVEM 3,0(17) ADJSP 17,2
LDB 5,-3(17) XMOVEI 3,0(17)
CAME 5,4 IOR 3,$BYTE
JRST $198 MOVEM 3,0(17)
ADJSP 17,-2
POPJ 17,
$198::
JUMPE 5,$200 CAIN 5,40
CAIE 5,40 AOSA 3,tokind
CAIN 5,11 TRNA
TRNA JRST $199
JRST $200 CAIE 5,11
CAIE 5,11 JRST $200
AOSA 3,tokind MOVN 3,tokind
SKIPA 4,[10] IDIVI 3,10
JRST $199 ADDI 4,10
MOVE 5,tokind ADDB 4,tokind
IDIVI 5,10
SUB 4,6
ADDB 4,tokind
MOVE 6,-1(17) XMOVEI 4,linbuf
XMOVEI 4,linbuf IOR 4,$BYTE
IOR 4,$BYTE CAME 4,-1(17)
CAME 6,4 JRST $222
JRST $222
doincl:
ADJSP 17,4 PUSH 17,nctoks
PUSH 17,nctoks PUSH 17,-2(17)
PUSH 17,-6(17) PUSHJ 17,maktnc
PUSHJ 17,maktnc ADJSP 17,-2
ADJSP 17,-2 CAIN 1,0
MOVEM 1,-3(17) POPJ 17,
JUMPN 1,$225 PUSH 17,1
ADJSP 17,-4 ADJSP 17,3
POPJ 17, MOVEM 1,-1(17)
$225::
MOVE 3,-3(17)
MOVEM 3,-1(17)
MOVE 6,0(17) POP 17,6
LDB 3,[331106,,147] LDB 1,[331106,,147]
JUMPN 3,$289 CAIN 1,0
LDB 4,[331106,,206] LDB 1,[331106,,206]
JUMPE 4,$288 CAIE 1,0
$289:: LDB 1,[331106,,1]
MOVE 1,0(17) CAIE 1,0
LDB 3,[331101,,1] MOVE 1,6
JUMPE 3,$288 POPJ 17,
ADJSP 17,-1
POPJ 17,
$288::
SETZ 1,
ADJSP 17,-1
POPJ 17,
pctfra:
ADJSP 17,1 PUSH 17,-3(17)
PUSH 17,-4(17) PUSH 17,-3(17)
PUSH 17,-4(17) PUSH 17,-3(17)
PUSH 17,-4(17) PUSHJ 17,pldfra
PUSHJ 17,pldfra ADJSP 17,-3
ADJSP 17,-3 LDB 3,1
MOVEM 1,0(17) CAIN 3,60
LDB 3,1 IBP 1
CAIN 3,60 POPJ 17,
IBP 0(17)
POP 17,1
POPJ 17,
SKIPL 7,-4(17) MOVE 7,-4(17)
SKIPGE 5,-5(17) XOR 7,-5(17)
TRNA JUMPGE 7,$1635
JRST $1635
SKIPGE 4,-5(17)
JUMPL 7,$1635
LDB 3,-4(17) LDB 1,-4(17)
JUMPE 3,$1654 CAIE 1,0
SETO 1, SETO 1,
JRST $1655
$1654::
SETZ 1,
upcase:
MOVEI 3,141 MOVE 1,-1(17)
CAMLE 3,-1(17) CAILE 1,141
JRST $1661 CAILE 1,172
MOVE 1,-1(17) TRNA
CAILE 1,172 SUBI 1,40
JRST $1661 POPJ 17,
SUBI 1,40
POPJ 17,
$1661::
MOVE 1,-1(17)
POPJ 17,
10-Dec-85 08:48:23-PST,1880;000000000005
Return-Path: <EPPSTEIN@CS.COLUMBIA.EDU>
Received: from CS.COLUMBIA.EDU by SRI-NIC.ARPA with TCP; Tue 10 Dec 85 08:48:14-PST
Date: 10 Dec 1985 11:47 EST (Tue)
Message-ID: <EPPSTEIN.12166024376.BABYL@CS.COLUMBIA.EDU>
From: David Eppstein <Eppstein@CS.COLUMBIA.EDU>
To: Ian Macky <Ian@SRI-NIC.ARPA>, KLH@SRI-NIC.ARPA
Subject: floating-point operations
In-reply-to: Msg of Mon 9 Dec 1985 17:01 PST from Ian Macky <Ian@SRI-NIC>
I've been thinking lately about how to avoid the flushcode() in double
compares. I came up with a sequence using DSUB instead of DFSB, and
another with an ADDI, but these all share a problem with the DFSB
suggestion: they don't work right if the addition overflows.
The following might be a way to do it: Switch the position and sense
of the first two compares. You will get something like
CAML(E) R+1,S+1
CAMGE R,S
CAMG R,S
(the middle op can also be CAME but that will take three instead of
two instructions in more cases). The R+1/S+1 compare is safe to
optimize because R+1 and S+1 are not used anytime after that. The
only time we have to worry is when we are generating the middle
compare, so that we don't lose the values for the final compare. So
if we make sure code0() does not do any optimization if the preceding
op is a skip we will be safe. Since you'd be adding code for this
case there anyway you might as well make code0() do a setskip() and
take out that code from double generation. You would also want to fix
invskip() to track the changed sequence, and might want to insert some
fixup for compares against zero...