[sdcc-devel] [ sdcc-Bugs-1086921 ] Several bugs concerning the stack

[SourceForge.net <no...@so...>]

Bugs item #1086921, was opened at 2004-12-17 08:04
Message generated for change (Comment added) made by tecodev
You can respond by visiting: 
https://sourceforge.net/tracker/?func=detail&atid=100599&aid=1086921&group_id=599

Category: pic16 target
Group: fixed
Status: Closed
Resolution: Fixed
Priority: 5
Submitted By: ccsporters (tecodev)
Assigned to: Vangelis Rokas (vrokas)
Summary: Several bugs concerning the stack

Initial Comment:
PIC16's gen.c contains several bugs concerning values
on the stack.

(1) values are accessed via FSR2
    The current implementation is off by one (FSR2
points to highest
    address available for local variables on the stack
but offset 0
    seems never to be used; instead we overwrite the
first saved
    register...)
    fixed in aopForSym() and pic16_freeAsmop()

(2) stack space for locals on the stack is allocated
incorrectly
    In genFunction() fsr1h in decremented unless fsr0l
has underflowed
    (Carry bit is set iff SUBWF returns a value >= 0).

(3) result values in WREG aer overwritten
    In genEndFunction() functions using the stack for
locals and
    returning a value via WREG the result value in WREG is
    overwritten by the MOVLW (sym->stack) instruction used
    to pop the locals from the stack.
    My fix is rather a workaround as it ignores
functions not
    returning a value (producing unneccessary slow (but
still
    correct) code.

(4) addresses on symbols on the stack
    In genAddrOf() the addresses are calculated using
ADDWF and
    a conditional INCF on FSR2. As FSR2 contains the
largest
    address used for locals this has to be a
conditional decrement
    on FSR2H. For "clarity" I changed it to SUBWF,
conditional DECF.

I also meddled in stdbool.h telling the system to use
chars for booleans on the PIC16 port (as bits aer not
yet too wel supported).
A last change (rather cosmetic) has been applied to
frexpf.c using the constant EXCESS instead of the
literal value 0x7e for exponent calculations (this
would impact no all ports so double check!).

Although floating point still doesn't really work
(sqrtf, sinf etc. still return rather random values)
these fixes help on getting them done.
Most bugs from above can be pinpointed by compiling
sqrtf.c and frexpf.c from device/lib for pic16, so I do
not attach another demo file.
I do upload another implementation on sqrt using the
"slow" fixpoint iteration, which actually works in the
curent state of the port!

Regards,
  Raphael Neider

----------------------------------------------------------------------

>Comment By: ccsporters (tecodev)
Date: 2004-12-26 19:31

Message:
Logged In: YES 
user_id=1115835

I have to correct my patch: in (1) and (4) we have to
consider two cases:
(a) arguments are refenreced via FSR2 + sym->stack
    with sym->stack >= 2 (this was broken by my patch)
(b) local variable on the stack are accessed via
FSR2+sym->stack+1
    with sym->stack < 0 (sym->stack == 0 is never used thus
I add 1)

Fixed in aopForSym(), freeAsmOp() and genAddrOf().

genEndFunction() has another problem: on entry we save FSR2,
create the local frame and save registers (in this order).
On exit we first removed the local frame (thus throwing away
saved registers), restored registers and FSR2. This lead to
uncontrolled behaviour after function calls (as the caller
variables were overwritten).
Fixed in genEndFunction().

The attached patch is against the current CVS version!

Regards,
    Raphael

----------------------------------------------------------------------

Comment By: ccsporters (tecodev)
Date: 2004-12-17 08:09

Message:
Logged In: YES 
user_id=1115835

Being a it, here's my device/lib/Makefile which (hopefully)
builds the pic16.lib library containing nearly all functions
from the standard SDCC lib! Yes, it works!

Regards,
    Raphael Neider

----------------------------------------------------------------------

Comment By: ccsporters (tecodev)
Date: 2004-12-17 08:07

Message:
Logged In: YES 
user_id=1115835

float.c proposes the slow but working iterative sqrt
calculation.
Compile with -DIN_VALUE=xxx and link against the neccessary
.rel files from device/lib/build/pic16 (after having build
them, which requires modifications to device/lib/Makefile).


----------------------------------------------------------------------

You can respond by visiting: 
https://sourceforge.net/tracker/?func=detail&atid=100599&aid=1086921&group_id=599