Login
You're viewing the functional.cafe public feed.
  • Jan 4, 2026, 1:20 AM

    Emitting a little section with one jmp to the start of every BB from after the last real instruction with cheeky little mov rax, (index of BB) every 16 steps for readability doesn't penalize the generated code as much as putting a jmp to X, nop, mov rax, (index of BB); nop, X in between BBs, but both the regular view in perf report's annotate view and objdump's --visualize-jumps aren't very practical for the "all jmps in one place" approach ...

    JITDUMP doesn't seem to allow me to have overlapping ranges, so I can't do something like have one range for the whole function and another symbol for each BB inside of it. If I do just that but don't have them overlap, then every BB will appear as its own whole function in the overview and not be disassembled together I think ...

    I could spit out a fake source file that just has "BB 1" through "BB 100000" on each line and generate line number information records but uggghhhh I don't like that either :|

    advice?

    #pldev #assembly #x86_64 #rakulang #moarvm

    💬 1🔄 1⭐ 1

Replies

  • Jan 4, 2026, 1:28 AM

    Here's the (very organised) mess that --visualize-jumps makes out of the lotsa-jumps section ... It scrolls at least one more screenful to the left ...

    Screenshot of "objdump" disassembler. On the far right is a column of `jmp` instructions, to the left are lines drawn in ASCII going to the left and then bending upwards, all crossing over each other
    💬 0🔄 0⭐ 0