At 18:07 08/11/2005, you wrote:
Jump by offset label number? I don't understand it.
_______________________________________________
fpc-pascal maillist - fpc-pascal@lists.freepascal.org
http://lists.freepascal.org/mailman/listinfo/fpc-pascal
Sorry Eduardo, am not knowledgeable of compiler internals.
There is (?) an optimization in FPC that uses a jump table(?)
instead of scanning each label value of a case statement. Hoping
that this is the case, as it seems to be with i386. This would mean
faster execution of the case. Yes?
Don't know how fpc ppc works internally. Surely this optimization is
not done, it's uncommon, i think.
For now, the only optimization to the compiler output maybe use other
crs for comparations and separate the branch from the comparation to
prevent stalls. Here an example.
Original code
# [735] case l3o3.FormType of
// Select 3o3 processor
lwz r2,160(r1)
cmplwi cr0,r2,0
beq cr0,L1051
cmplwi cr0,r2,1
beq cr0,L1052
cmplwi cr0,r2,2
beq cr0,L1053
cmplwi cr0,r2,3
beq cr0,L1054
Modified code
# [735] case l3o3.FormType of
// Select 3o3 processor
lwz r3,160(r1) // r3, has the value to check
cmplwi cr0,r3,0
cmplwi cr2,r3,1 // cr1 is used for fpu condition and
cannot be used for integer
cmplwi cr3,r3,2
cmplwi cr4,r3,3
beq cr0,L1051 // separate branch from comp (4 lines
up) can make the branch be pre-calculated and executed in zero cycles
cmplwi cr5,r3,4
beq cr2,L1052
cmplwi cr0,r3,5
beq cr3,L1053
cmplwi cr2,r3,6
beq cr4,L1054
and so on...
_______________________________________________
fpc-pascal maillist - fpc-pascal@lists.freepascal.org
http://lists.freepascal.org/mailman/listinfo/fpc-pascal
