The fact.pasm and substr.pasm examples can be changed to use more
intelligent branching. (Can you say hand-optimized assembler?)
If you apply this patch, you might have to write test cases again
for the missing instructions. :-)
Now to my real problem:
Some pasm examples have problems with the jit compiler on i386 wrt
branching. I suspect some problem to generate "near" jumps??
Please do not apply the change for jit/i386/core.jit. The current code
looks bogus to make a "jmp()". Funny thing is leaving that code away
does not work, but that "cmp/jne" code seems to make some examples
work. Probably "JUMP" is completely wrong here??
I am running on a current Red Hat Linux machine here.
thanks,
Florian La Roche
--- examples/assembly/fact.pasm
+++ examples/assembly/fact.pasm
@@ -22,14 +22,12 @@ $loop:
print I0
print "\n"
inc I1,1
- eq I1,7,$done
- branch $loop
-$done:
+ ne I1,7,$loop
end
# I0 is the number to compute
fact:
- clonei
+ pushi
lt I0,2,$is_one
set I1,I0
dec I0,1
--- examples/assembly/substr.pasm
+++ examples/assembly/substr.pasm
@@ -18,12 +18,10 @@ WAX: substr S2, S1, I3, I4
print S2
print "\n"
add I4, I4, I2
- eq I4, I5, WANE
- branch WAX
+ ne I4, I5, WAX
WANE: length I1, S1
print S1
print "\n"
chopn S1, 1
- eq I1, I3, DONE
- branch WANE
-DONE: end
+ ne I1, I3, WANE
+ end
--- jit/i386/core.jit
+++ jit/i386/core.jit
@@ -83,7 +83,9 @@ Parrot_if_i_ic {
}
Parrot_branch_ic {
- jmp ($0x0)
+ cmp %eax,%eax
+ jne ($0x0)
+; jmp ($0x0)
JUMP(INT_CONST[1])
}
