Yep, eventually it will come down to experiments. Or I might end up implementing an optimizer before the Go compiler. Thanks!
On Sun, Nov 12, 2017 at 4:01 PM Jesper Louis Andersen < jesper.louis.ander...@gmail.com> wrote: > On Sun, Nov 12, 2017 at 9:15 PM Petar Maymounkov <pet...@gmail.com> wrote: > >> This is an interesting suggestion. But it makes me wonder how this >> compares against generating directly a bag of LLVM statically-typed >> functions >> and trying the intelligence of the LLVM SSA optimizer. Do you know if >> this one resorts to global analysis? >> >> > My major concern with LLVM is the same as the Go compiler. Does it do > global optimizations well? A whole-program compiler, on the other hand, is > forced by construction to be efficient at this problem: either there is no > need for whole-program compilation, or you run out of memory, whichever > comes first :) > > Another concern, if you are not doing tail-calls, but rather static calls > that really do have a valid return, is that the inliner almost have to be > polyvariant (it needs to determine inlining at each call site and not at a > global enable/disable level). Some compilers simply disable inlining for > any function which is called too often. Some times this is the correct > *optimization* as it alleviates Instruction cache pressure. Other times, > however, the inlining triggers vast simplifications if it is done, so the > overhead in instruction expansion (and thus insn cache pressure) is worth > the faster execution. > > OTOH: modern hardware is excellent at making funcalls fast. Hardware > engineers tend to optimize for C-like programs and fix obvious mistakes > compiler engineers make :) > > In short: Make some small experiments. Try to scale them up. Who is > fastest on a carry-lookahead adder? > -- You received this message because you are subscribed to the Google Groups "golang-nuts" group. To unsubscribe from this group and stop receiving emails from it, send an email to golang-nuts+unsubscr...@googlegroups.com. For more options, visit https://groups.google.com/d/optout.
