RFR: Move Shenandoah stubs generation into ShenandoahBarrierSetAssembler

[Aleksey Shipilev]

On 06/13/2018 09:57 PM, Roman Kennke wrote:
> This moves another bunch of Shenandoah code out of shared into
> Shenandoah specific places. It removes a very significant amount of
> upstream mess :-) BTW, ShenandoahBarrierSetAssembler is approaching 1000
> LOC :-)
> 
> It builds on top of my previous patch to move runtime entries into
> ShenandoahRuntime. This needs to go in first.
> 
> Testing: tier3_gc_shenandoah (ongoing, expected clean before push)
> 
> http://cr.openjdk.java.net/~rkennke/shstubgen/webrev.00/

Good change!

*) I think you can put the assert right inside the SBSA::shenandoah_wb() call in both
macroAssembler_aarch64.cpp and macroAssembler_x86.cpp:

4099   assert(ShenandoahBarrierSetAssembler::shenandoah_wb() != NULL, "need write barrier stub");
4100   far_call(RuntimeAddress(CAST_FROM_FN_PTR(address,
ShenandoahBarrierSetAssembler::shenandoah_wb())));


*) This revert, you sure aarch64 still runs fine after it? We can take in on faith, and see what
testing shows, but at least we should have basic idea about its safety:

  44   code_size2 = 22000           // simply increase if too small (assembler will crash if too small)


*) In lcm.cpp, you need to #include "utilities/macros.hpp" to gain access to INCLUDE_SHENANDOAHGC macro

*) In lcm.cpp, please use braces for "else":

 899 #if INCLUDE_SHENANDOAHGC
 900   if (UseShenandoahGC && mcall->entry_point() ==
ShenandoahBarrierSetAssembler::shenandoah_wb_C()) {
 901     assert(op == Op_CallLeafNoFP, "shenandoah_wb_C should be called with Op_CallLeafNoFP");
 902     add_call_kills(proj, regs, save_policy, exclude_soe, true);
 903   } else
 904 #endif
       {  <--- here
 905     add_call_kills(proj, regs, save_policy, exclude_soe, false);
       }  <--- and here
 906
 907   return node_cnt;


Thanks,
-Aleksey