src/cpu/sparc/vm/vm_version_sparc.cpp

*** 42,65 ****
    _features = determine_features();
    PrefetchCopyIntervalInBytes = prefetch_copy_interval_in_bytes();
    PrefetchScanIntervalInBytes = prefetch_scan_interval_in_bytes();
    PrefetchFieldsAhead         = prefetch_fields_ahead();
  
    // Allocation prefetch settings
!   intx cache_line_size = L1_data_cache_line_size();
    if( cache_line_size > AllocatePrefetchStepSize )
      AllocatePrefetchStepSize = cache_line_size;
!   if( FLAG_IS_DEFAULT(AllocatePrefetchLines) )
!     AllocatePrefetchLines = 3; // Optimistic value
!   assert( AllocatePrefetchLines > 0, "invalid value");
    if( AllocatePrefetchLines < 1 ) // set valid value in product VM
!     AllocatePrefetchLines = 1; // Conservative value
  
    AllocatePrefetchDistance = allocate_prefetch_distance();
    AllocatePrefetchStyle    = allocate_prefetch_style();
  
!   assert(AllocatePrefetchDistance % AllocatePrefetchStepSize == 0, "invalid value");
  
    if (AllocatePrefetchStyle == 3 && !has_blk_init()) {
      warning("BIS instructions are not available on this CPU");
      FLAG_SET_DEFAULT(AllocatePrefetchStyle, 1);
    }
--- 42,76 ----
    _features = determine_features();
    PrefetchCopyIntervalInBytes = prefetch_copy_interval_in_bytes();
    PrefetchScanIntervalInBytes = prefetch_scan_interval_in_bytes();
    PrefetchFieldsAhead         = prefetch_fields_ahead();
  
+   assert(0 <= AllocatePrefetchInstr && AllocatePrefetchInstr <= 1, "invalid value");
+   if( AllocatePrefetchInstr < 0 ) AllocatePrefetchInstr = 0;
+   if( AllocatePrefetchInstr > 1 ) AllocatePrefetchInstr = 0;
+ 
    // Allocation prefetch settings
!   intx cache_line_size = prefetch_data_size();
    if( cache_line_size > AllocatePrefetchStepSize )
      AllocatePrefetchStepSize = cache_line_size;
! 
!   assert(AllocatePrefetchLines > 0, "invalid value");
    if( AllocatePrefetchLines < 1 )     // set valid value in product VM
!     AllocatePrefetchLines = 3;
!   assert(AllocateInstancePrefetchLines > 0, "invalid value");
!   if( AllocateInstancePrefetchLines < 1 ) // set valid value in product VM
!     AllocateInstancePrefetchLines = 1;
  
    AllocatePrefetchDistance = allocate_prefetch_distance();
    AllocatePrefetchStyle    = allocate_prefetch_style();
  
!   assert((AllocatePrefetchDistance % AllocatePrefetchStepSize) == 0 &&
!          (AllocatePrefetchDistance > 0), "invalid value");
!   if ((AllocatePrefetchDistance % AllocatePrefetchStepSize) != 0 ||
!       (AllocatePrefetchDistance <= 0)) {
!     AllocatePrefetchDistance = AllocatePrefetchStepSize;
!   }
  
    if (AllocatePrefetchStyle == 3 && !has_blk_init()) {
      warning("BIS instructions are not available on this CPU");
      FLAG_SET_DEFAULT(AllocatePrefetchStyle, 1);
    }
*** 97,120 ****
      // aligned on a single instruction boundary
      if (FLAG_IS_DEFAULT(InteriorEntryAlignment)) {
        FLAG_SET_DEFAULT(InteriorEntryAlignment, 4);
      }
      if (is_niagara_plus()) {
!       if (has_blk_init() && AllocatePrefetchStyle > 0 &&
!           FLAG_IS_DEFAULT(AllocatePrefetchStyle)) {
!         // Use BIS instruction for allocation prefetch.
!         FLAG_SET_DEFAULT(AllocatePrefetchStyle, 3);
          if (FLAG_IS_DEFAULT(AllocatePrefetchDistance)) {
!           // Use smaller prefetch distance on N2 with BIS
!           FLAG_SET_DEFAULT(AllocatePrefetchDistance, 64);
          }
        }
        if (AllocatePrefetchStyle != 3 && FLAG_IS_DEFAULT(AllocatePrefetchDistance)) {
          // Use different prefetch distance without BIS
!         FLAG_SET_DEFAULT(AllocatePrefetchDistance, 256);
        }
      }
  #endif
    }
  
    // Use hardware population count instruction if available.
    if (has_hardware_popc()) {
--- 108,154 ----
      // aligned on a single instruction boundary
      if (FLAG_IS_DEFAULT(InteriorEntryAlignment)) {
        FLAG_SET_DEFAULT(InteriorEntryAlignment, 4);
      }
      if (is_niagara_plus()) {
!       if (has_blk_init() && UseTLAB &&
!           FLAG_IS_DEFAULT(AllocatePrefetchInstr)) {
!         // Use BIS instruction for TLAB allocation prefetch.
!         FLAG_SET_ERGO(intx, AllocatePrefetchInstr, 1);
!         if (FLAG_IS_DEFAULT(AllocatePrefetchStyle)) {
!           FLAG_SET_ERGO(intx, AllocatePrefetchStyle, 3);
!         }
          if (FLAG_IS_DEFAULT(AllocatePrefetchDistance)) {
!           // Use smaller prefetch distance with BIS
!           FLAG_SET_ERGO(intx, AllocatePrefetchDistance, 64);
          }
        }
+       if (is_T4()) {
+         // Double number of prefetched cache lines on T4
+         // since L2 cache line size is smaller (32 bytes).
+         if (FLAG_IS_DEFAULT(AllocatePrefetchLines)) {
+           FLAG_SET_ERGO(intx, AllocatePrefetchLines, AllocatePrefetchLines*2);
+         }
+         if (FLAG_IS_DEFAULT(AllocateInstancePrefetchLines)) {
+           FLAG_SET_ERGO(intx, AllocateInstancePrefetchLines, AllocateInstancePrefetchLines*2);
+         }
+       }
        if (AllocatePrefetchStyle != 3 && FLAG_IS_DEFAULT(AllocatePrefetchDistance)) {
          // Use different prefetch distance without BIS
!         FLAG_SET_ERGO(intx, AllocatePrefetchDistance, 256);
        }
+       if (AllocatePrefetchInstr == 1) {
+         // Need a space at the end of TLAB for BIS since it
+         // will fault when accessing memory outside of heap.
+ 
+         // +1 for rounding up to next cache line, +1 to be safe
+         int lines = AllocatePrefetchLines + 2;
+         int step_size = AllocatePrefetchStepSize;
+         int distance = AllocatePrefetchDistance;
+         _reserve_for_allocation_prefetch = (distance + step_size*lines)/(int)HeapWordSize;
        }
+     }
  #endif
    }
  
    // Use hardware population count instruction if available.
    if (has_hardware_popc()) {
*** 183,200 ****
    if (!has_vis1()) // Drop to 0 if no VIS1 support
      UseVIS = 0;
  
  #ifndef PRODUCT
    if (PrintMiscellaneous && Verbose) {
!     tty->print("Allocation: ");
      if (AllocatePrefetchStyle <= 0) {
!       tty->print_cr("no prefetching");
      } else {
        if (AllocatePrefetchLines > 1) {
!         tty->print_cr("PREFETCH %d, %d lines of size %d bytes", AllocatePrefetchDistance, AllocatePrefetchLines, AllocatePrefetchStepSize);
        } else {
!         tty->print_cr("PREFETCH %d, one line", AllocatePrefetchDistance);
        }
      }
      if (PrefetchCopyIntervalInBytes > 0) {
        tty->print_cr("PrefetchCopyIntervalInBytes %d", PrefetchCopyIntervalInBytes);
      }
--- 217,240 ----
    if (!has_vis1()) // Drop to 0 if no VIS1 support
      UseVIS = 0;
  
  #ifndef PRODUCT
    if (PrintMiscellaneous && Verbose) {
!     tty->print("Allocation");
      if (AllocatePrefetchStyle <= 0) {
!       tty->print_cr(": no prefetching");
      } else {
+       tty->print(" prefetching: ");
+       if (AllocatePrefetchInstr == 0) {
+           tty->print("PREFETCH");
+       } else if (AllocatePrefetchInstr == 1) {
+           tty->print("BIS");
+       }
        if (AllocatePrefetchLines > 1) {
!         tty->print_cr(" at distance %d, %d lines of %d bytes", AllocatePrefetchDistance, AllocatePrefetchLines, AllocatePrefetchStepSize);
        } else {
!         tty->print_cr(" at distance %d, one line of %d bytes", AllocatePrefetchDistance, AllocatePrefetchStepSize);
        }
      }
      if (PrefetchCopyIntervalInBytes > 0) {
        tty->print_cr("PrefetchCopyIntervalInBytes %d", PrefetchCopyIntervalInBytes);
      }