On 10/29/2014 04:31 AM, Andi Kleen wrote:
> 2014-10-28 Andi Kleen <a...@linux.intel.com>
>
> PR target/63672
> * config/i386/i386.c (ix86_expand_builtin): Generate memory
> barrier after abort.
> * config/i386/i386.md (xbegin): Add memory barrier.
> (xend): Rename to ...
> (xend_1): New. Generate memory barrier and emit xend.
Richi's comment is spot on.
The insns themselves should hold the barrier, not being separate like
>
> -(define_insn "xend"
> +(define_insn "xend_1"
> [(unspec_volatile [(const_int 0)] UNSPECV_XEND)]
> "TARGET_RTM"
> "xend"
> [(set_attr "type" "other")
> (set_attr "length" "3")])
>
> +(define_expand "xend"
> + [(set (match_dup 0)
> + (unspec:BLK [(const_int 0)] UNSPECV_XEND))] /* or match_dup 0 ? */
> + "TARGET_RTM"
> +{
> + emit_insn (gen_xend_1 ());
> +
> + operands[0] = gen_rtx_MEM (BLKmode, gen_rtx_SCRATCH (Pmode));
> + MEM_VOLATILE_P (operands[0]) = 1;
> +
> + DONE;
> +})
this, which generates two separate insns. C.f. sse2_lfence.
r~
