search

SamCoVT / TaliForth2

A Subroutine Threaded Code (STC) ANSI-like Forth for the 65c02
Other
29 stars 7 forks source link

Avoid generating JMP in conditionals since it forces NN #135

Open patricksurry opened 2 months ago

patricksurry commented 2 months ago

Most user words with conditionals end up as NN because they involve absolute jmp either literally or via zero_branch_runtime. which uses an absolute target rather than a 16-bit offset.

We could replace these with bra in many cases, and the others using a jsr farbra / .word offset construct so that we never generate an absolute jump target. I believe all code generation for jmp is routed through cmpl_jump, cmpl_jump_later, cmpl_jump_tos, cmpl_0branch_tos and cmpl_0branch_later so it shouldn't be too difficult to rewrite them.

Performance/code-size would improve a little for jmp -> bra but would be a bit worse for jsr farbra. On the other hand, small conditional words would potentially benefit from inlining.

A 16-bit offset would only allow jump up to +/- 32Kb so it would fail if you wrote a word that contained a conditional jump across say a 32K+ alloc'd block. It's hard to see that being a significant restriction in practice so could generate an error message?

(feel free to assign this to me)

SamCoVT commented 2 months ago

Doesn't the 16-bit offset wrap? Doesn't that make the entire 64K address space available as a jump target? I don't believe this needs any special handling if you are doing dest_addr - current_addr to compute the offset, and current_addr + offset to calculate the destination of the branch.

patricksurry commented 2 months ago

doh! yes that sounds right.