Please create a bug ticket in http://sourceforge.net/p/sdcc/bugs/, and attach
your patch.
Thanks.
在2016年08月30 08时12分, "Raphael Neider"<rnei...@web.de>写道:
Hi,
I believe I have a fix for all issues you reported:
- useless MOVFF x,x removed (in this particular situation -- there may be more
situations like this)
- missing BANKSEL added (incorrect removal was caused by the useless MOVFF,
because it used an operand that -- if we had a BANKSEL for it -- would have
made the missing one obsolete. For MOVFF operands, we have no BANKSEL, so this
was a bug)
- MOVF a, W; MOVWF a turned to MOVFF (this can be better, if a is not used
before/afterwards and if a is not known to be already in W; it can be worse if
a is used again (copied somewhere else) or if it's already in W)
Patch attached. If anyone wants to review it -- you are welcome. Esp. the
SPIL_LOC && IS_TEMP is something I do not really understand -- it works for our
situation, but I am not sure how correct / generally applicable that is :-(.
I'll try to dig into this again tonight and check it in around 23:00 CEST today.
Good night,
Raphael
2016-08-29 15:33 GMT+02:00 Kustaa Nyholm <kustaa.nyh...@planmeca.com>:
Here is a minimal self contained test case that produces the problem:
//------------------------------------------
typedef struct {
char member;
} test_type;
extern test_type g_test1[];
extern test_type g_test2;
void foobar() {
g_test2.member = g_test1[0].member;
g_test2.member = g_test1[0].member;
}
//------------------------------------------
generated code:
//------------------------------------------
01008 S_main__foobar code
000000 01009 _foobar:
01010 ; .line 66; main.c g_test2.member =
g_test1[0].member;
000000 CFD9 FFE5 01011 MOVFF FSR2L, POSTDEC1
000004 CFE1 FFD9 01012 MOVFF FSR1L, FSR2L
01013 ; .line 67; main.c }
000008 C??? F??? 01014 MOVFF _g_test1, _g_test1
00000C ???? 01015 BANKSEL _g_test1
00000E 51?? 01016 MOVF _g_test1, W, B
000010 ???? 01017 BANKSEL _g_test2
000012 6F?? 01018 MOVWF _g_test2, B
01019 ; .line 68; main.c
000014 C??? F??? 01020 MOVFF _g_test1, _g_test1
000018 51?? 01021 MOVF _g_test1, W, B
01022 ; removed redundant BANKSEL
00001A 6F?? 01023 MOVWF _g_test2, B
00001C CFE4 FFD9 01024 MOVFF PREINC1, FSR2L
000020 0012 01025 RETURN
//------------------------------------------
compiled with:
//------------------------------------------
/Users/nyholku/sdcc-3.6.0/bin/sdcc --no-crt --ivt-loc=0x800 -V -L
/Users/nyholku/sdcc-3.4.0/share/sdcc/non-free/lib/pic16 -Wa,-S,0
-Wl,-m,-s18f45k50.lkr -mpic16 -p18f45k50 --disable-warning 85 --std-sdcc99
--obanksel=3 --use-non-free main.c
//------------------------------------------
Thinking this through while the code would work with without the BANKSEL
optimisation,
it would of course be preferable if the assignment would be optimised to a
single MOVFF instruction.
And I think (without knowing the details of code generation or peephole
optimisation)
that this should be easy to do because the compiler at that point knows both
the source
and destination absolute (even if symbolic) addresses.
wbr Kusti
This e-mail may contain confidential or privileged information. If you are not
the intended recipient (or have received this e-mail in error) please notify
the sender immediately and destroy this e-mail. Any unauthorized copying,
disclosure or distribution of the material in this e-mail is strictly
forbidden. We will not be liable for direct, indirect, special or consequential
damages arising from alteration of the contents of this message by a third
party or as a result of any virus being passed on or as of transmission of this
e-mail in general.
------------------------------------------------------------------------------
_______________________________________________
Sdcc-user mailing list
Sdcc-user@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/sdcc-user
------------------------------------------------------------------------------
_______________________________________________
Sdcc-user mailing list
Sdcc-user@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/sdcc-user
