Hi, Can the issue be resolved in a target independent manner as suggested below? Or is it better to deal with this in the target code?
Best regards, Oleg Endo On Fri, 2024-09-27 at 00:26 -0400, Pietro Monteiro wrote: > The prefetch instruction that is emitted by __builtin_prefetch is > re-ordered on GCC, but not on clang[0]. GCC's behavior is surprising > because when using the builtin you want the instruction to be placed at > the exact point where you put it. Moving it around, specially across > load/stores, may end up being a pessimization. Adding a blockage > instruction before the prefetch prevents the scheduler from moving it. > > [0] https://godbolt.org/z/Ycjr7Tq8b > > > -- 8< -- > > > diff --git a/gcc/builtins.cc b/gcc/builtins.cc > index 37c7c98e5c..fec751e0d6 100644 > --- a/gcc/builtins.cc > +++ b/gcc/builtins.cc > @@ -1329,7 +1329,12 @@ expand_builtin_prefetch (tree exp) > create_integer_operand (&ops[1], INTVAL (op1)); > create_integer_operand (&ops[2], INTVAL (op2)); > if (maybe_expand_insn (targetm.code_for_prefetch, 3, ops)) > - return; > + { > + /* Prevent the prefetch from being moved. */ > + rtx_insn *last = get_last_insn (); > + emit_insn_before (gen_blockage (), last); > + return; > + } > } > > /* Don't do anything with direct references to volatile memory, but
