Conversation

Notices

1. Adrian Cochrane (alcinnz@floss.social)'s status on Friday, 24-Feb-2023 10:28:30 JST Adrian Cochrane

   Let me revisit my hypothetical about designing hardware to reimplement a browser (like the ones I actually am implementing) & it's underlying OS upon.

   https://adrian.geek.nz/docs/web-cpu-hypothetical.html

   What sort of programming language would be ideal for implementing webpage layout in?

   The language for parsing input is straightforward: I'd want something resembling YACC. May need some software optimization, but mostly I'd have cause to mostly do that in hardware as a "decoder" to bitmask-addressed microcode.

   1/?

   • Adrian Cochrane (alcinnz@floss.social)'s status on Friday, 24-Feb-2023 10:50:48 JST Adrian Cochrane

     in reply to

     My ideal language for implementing CSS layout would let me write formulas for each CSS unit, keyword, or function emitting number(s). From there the compiler could compute a partial ordering controlflow. Letting me specify fallback values to break cycles.

     I'd need the ability to wrap these formulas in pattern-matched conditions upon other properties.

     All data would be fixed-length arrays of numbers.

     I'd be tempted to lower this language using the tree-SIMT it'd target, should be capable.

     2/?

   • Adrian Cochrane (alcinnz@floss.social)'s status on Friday, 24-Feb-2023 11:11:32 JST Adrian Cochrane

     in reply to

     Tree-traversals would be triggered in that language by referencing parent or child data. I'd probably have this language lower arbitrarily-structured trees into binary trees that easier to implement in hardware.

     ---

     However CSS layout isn't the only thing I'd want to run on this tree-SIMT, & the nicer language would likely yield code which would heavily benefit from further optimization.

     Text layout can be heavily involved. And I'd involve it in rasterizing vector graphics like fonts.

     3/4?

   • Adrian Cochrane (alcinnz@floss.social)'s status on Friday, 24-Feb-2023 11:31:59 JST Adrian Cochrane

     in reply to

     This lower-level language would still operate in terms of numeric-arrays for a datamodel. But now I'd add explicit controlflow, includes loops over arrays & (binary-tree) children.

     If I incorporate immutability that'd simplify most standard optimizations LLVM or GCC implements.

     It'd especially help to eliminate common-subexpressions between branches (since branching is inefficient in SIMTs), propagate constants, & generating formulas for array lookups. As for register allocation...

     4/4.1!!

   • Adrian Cochrane (alcinnz@floss.social)'s status on Friday, 24-Feb-2023 11:40:23 JST Adrian Cochrane

     in reply to

     Since I have shift registers it'd be trivial for me to incorporate cheap stack(s) operations into the machinecode, which could reduce the demands on a register allocator. Without removing the need.

     An optimizing compiler would be very useful for implementing shaping, & to allow for the nicer syntax abstracting away controlflow!

     4.5/4.5