I've found the cause of the omitted type casts. It turns out that I missed a branch in the instruction combination step which specifically checks for extension of values stored in memory. The exiting code did the usual job of assuming that typecasts from byte to word can be removed without consequence. Of course this is a tragic mistake for the TMS9900.
I need to find a way to allocate a scratch register during split to convert this:
and_mem_mem
movb @memb_a, r1
inv r1
movb @memb_b, r2
szcb r1, r2
movb r2, r1
sra r1, 8
b *r11
to this:
and_mem_mem
movb @memb_a, r2
inv r2
movb @memb_b, r1
szcb r2, r1 <--- eliminates MOV instruction
sra r1, 8
b *r11
Other stuff to do:
Check PC as fixed function again, might be different now.
Check R11 used as data register in non-leaf functions
Subscribe to:
Post Comments (Atom)
No comments:
Post a Comment