Hi all,
I noticed that the PDK port is lacking a basic optimisation for
conditional testing of single bits in IO registers (e.g. "if(PA & (1 <<
4))"), where it uses a sequence of "mov.io", "and", "ceqsn"/"cneqsn"
instructions, when instead it could be using "t0sn.io"/"t1sn.io"
instructions.
I am trying to implement a couple of peephole rules to do this, but they
aren't being applied everywhere I would expect them to. It seems to be
that the "notUsed('a')" condition I have is misbehaving and thinks the A
reg is used when it isn't. If I omit the notUsed('a') but leave
everything else the same, the rules apply everywhere I expect them to.
My rules are as follows:
replace restart {
mov.io a, %1
and a, #%3
ceqsn a, #0x00
goto %2
} by {
t0sn.io %1, #%4
goto %2
; peephole replaced IO reg single bit test-and-branch with 't0sn'
(pdk13).
} if isPort('pdk13'), notUsed('a'), operandsLiteral(%3), immdInRange(0 7
'singleSetBit' %3 8 %4)
replace restart {
mov.io a, %1
and a, #%3
cneqsn a, #0x00
goto %2
} by {
t1sn.io %1, #%4
goto %2
; peephole replaced IO reg single bit test-and-branch with 't1sn'
(pdk14, pdk15).
} if isPort('pdk14' 'pdk15'), notUsed('a'), operandsLiteral(%3),
immdInRange(0 7 'singleSetBit' %3 8 %4)
One example of a situation where the rules aren't being applied is:
void interrupt_handler(void) __interrupt(0) {
if(INTRQ & INTRQ_T16) {
timeout_expired = true;
INTEN &= ~INTEN_T16;
INTRQ &= ~INTRQ_T16;
}
}
The assembly for that is:
_interrupt_handler:
push af
mov a, p
push af
mov.io a, __intrq
and a, #0x04
ceqsn a, #0x00
goto 00111$
00112$:
goto 00103$
00111$:
mov a, #0x01
mov _timeout_expired+0, a
set0.io __inten, #2
set0.io __intrq, #2
00103$:
pop af
mov p, a
pop af
reti
You can see that for the lines following those that would match the
first rule, no matter which branch is taken, the first use of A reg is
to have something assigned to it - either "mov a, #0x01" or "pop af".
The notUsed('a') condition should be returning true here.
Is notUsed() on PDK buggy, or am I missing something?
This is with RC2 of 4.4.0 (r14579) on Windows x64.
Regards,
Basil Hussain
_______________________________________________
Sdcc-user mailing list
Sdcc-user@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/sdcc-user
