S390: Optimize allocate sequence

src/s390/macro-assembler-s390.cc

 // Copyright 2014 the V8 project authors. All rights reserved.
 // Use of this source code is governed by a BSD-style license that can be
 // found in the LICENSE file.
 
 #include <assert.h>  // For assert
 #include <limits.h>  // For LONG_MIN, LONG_MAX.
 
 #if V8_TARGET_ARCH_S390
 
 #include "src/base/bits.h"
@@ skipping 1561 matching lines @@
       AllocationUtils::GetAllocationTopReference(isolate(), flags);
   ExternalReference allocation_limit =
       AllocationUtils::GetAllocationLimitReference(isolate(), flags);
 
   intptr_t top = reinterpret_cast<intptr_t>(allocation_top.address());
   intptr_t limit = reinterpret_cast<intptr_t>(allocation_limit.address());
   DCHECK((limit - top) == kPointerSize);
 
   // Set up allocation top address register.
   Register top_address = scratch1;
-  // This code stores a temporary value in ip. This is OK, as the code below
-  // does not need ip for implicit literal generation.
-  Register alloc_limit = ip;
   Register result_end = scratch2;
   mov(top_address, Operand(allocation_top));
 
   if ((flags & RESULT_CONTAINS_TOP) == 0) {
     // Load allocation top into result and allocation limit into ip.
     LoadP(result, MemOperand(top_address));
-    LoadP(alloc_limit, MemOperand(top_address, kPointerSize));
   } else {
     if (emit_debug_code()) {
       // Assert that result actually contains top on entry.
-      LoadP(alloc_limit, MemOperand(top_address));
-      CmpP(result, alloc_limit);
+      CmpP(result, MemOperand(top_address));
       Check(eq, kUnexpectedAllocationTop);
     }
-    // Load allocation limit. Result already contains allocation top.
-    LoadP(alloc_limit, MemOperand(top_address, limit - top));
   }
 
   if ((flags & DOUBLE_ALIGNMENT) != 0) {
 // Align the next allocation. Storing the filler map without checking top is
 // safe in new-space because the limit of the heap is aligned there.
 #if V8_TARGET_ARCH_S390X
     STATIC_ASSERT(kPointerAlignment == kDoubleAlignment);
 #else
     STATIC_ASSERT(kPointerAlignment * 2 == kDoubleAlignment);
     AndP(result_end, result, Operand(kDoubleAlignmentMask));
     Label aligned;
     beq(&aligned, Label::kNear);
     if ((flags & PRETENURE) != 0) {
-      CmpLogicalP(result, alloc_limit);
+      CmpLogicalP(result, MemOperand(top_address, limit - top));
       bge(gc_required);
     }
     mov(result_end, Operand(isolate()->factory()->one_pointer_filler_map()));
     StoreW(result_end, MemOperand(result));
     AddP(result, result, Operand(kDoubleSize / 2));
     bind(&aligned);
 #endif
   }
 
-  // Calculate new top and bail out if new space is exhausted. Use result
-  // to calculate the new top.
-  SubP(r0, alloc_limit, result);
-  if (is_int16(object_size)) {
-    CmpP(r0, Operand(object_size));
-    blt(gc_required);
-    AddP(result_end, result, Operand(object_size));
-  } else {
-    mov(result_end, Operand(object_size));
-    CmpP(r0, result_end);
-    blt(gc_required);
-    AddP(result_end, result, result_end);
-  }
+  AddP(result_end, result, Operand(object_size));
+
+  // Compare with allocation limit.
+  CmpLogicalP(result_end, MemOperand(top_address, limit - top));
+  bge(gc_required);
 
   if ((flags & ALLOCATION_FOLDING_DOMINATOR) == 0) {
     // The top pointer is not updated for allocation folding dominators.
     StoreP(result_end, MemOperand(top_address));
   }
 
+  // Prefetch the allocation_top's next cache line in advance to
+  // help alleviate potential cache misses.
+  // Mode 2 - Prefetch the data into a cache line for store access.
+  pfd(r2, MemOperand(result, 256));
+
   // Tag object.
-  AddP(result, result, Operand(kHeapObjectTag));
+  la(result, MemOperand(result, kHeapObjectTag));
 }
 
 void MacroAssembler::Allocate(Register object_size, Register result,
                               Register result_end, Register scratch,
                               Label* gc_required, AllocationFlags flags) {
   DCHECK((flags & ALLOCATION_FOLDED) == 0);
   if (!FLAG_inline_new) {
     if (emit_debug_code()) {
       // Trash the registers to simulate an allocation failure.
       LoadImmP(result, Operand(0x7091));
@@ skipping 14 matching lines @@
   ExternalReference allocation_top =
       AllocationUtils::GetAllocationTopReference(isolate(), flags);
   ExternalReference allocation_limit =
       AllocationUtils::GetAllocationLimitReference(isolate(), flags);
   intptr_t top = reinterpret_cast<intptr_t>(allocation_top.address());
   intptr_t limit = reinterpret_cast<intptr_t>(allocation_limit.address());
   DCHECK((limit - top) == kPointerSize);
 
   // Set up allocation top address and allocation limit registers.
   Register top_address = scratch;
-  // This code stores a temporary value in ip. This is OK, as the code below
-  // does not need ip for implicit literal generation.
-  Register alloc_limit = ip;
   mov(top_address, Operand(allocation_top));
 
   if ((flags & RESULT_CONTAINS_TOP) == 0) {
-    // Load allocation top into result and allocation limit into alloc_limit..
+    // Load allocation top into result
     LoadP(result, MemOperand(top_address));
-    LoadP(alloc_limit, MemOperand(top_address, kPointerSize));
   } else {
     if (emit_debug_code()) {
       // Assert that result actually contains top on entry.
-      LoadP(alloc_limit, MemOperand(top_address));
-      CmpP(result, alloc_limit);
+      CmpP(result, MemOperand(top_address));
       Check(eq, kUnexpectedAllocationTop);
     }
-    // Load allocation limit. Result already contains allocation top.
-    LoadP(alloc_limit, MemOperand(top_address, limit - top));
   }
 
   if ((flags & DOUBLE_ALIGNMENT) != 0) {
 // Align the next allocation. Storing the filler map without checking top is
 // safe in new-space because the limit of the heap is aligned there.
 #if V8_TARGET_ARCH_S390X
     STATIC_ASSERT(kPointerAlignment == kDoubleAlignment);
 #else
     STATIC_ASSERT(kPointerAlignment * 2 == kDoubleAlignment);
     AndP(result_end, result, Operand(kDoubleAlignmentMask));
     Label aligned;
     beq(&aligned, Label::kNear);
     if ((flags & PRETENURE) != 0) {
-      CmpLogicalP(result, alloc_limit);
+      CmpLogicalP(result, MemOperand(top_address, limit - top));
       bge(gc_required);
     }
     mov(result_end, Operand(isolate()->factory()->one_pointer_filler_map()));
     StoreW(result_end, MemOperand(result));
     AddP(result, result, Operand(kDoubleSize / 2));
     bind(&aligned);
 #endif
   }
 
   // Calculate new top and bail out if new space is exhausted. Use result
   // to calculate the new top. Object size may be in words so a shift is
   // required to get the number of bytes.
-  SubP(r0, alloc_limit, result);
   if ((flags & SIZE_IN_WORDS) != 0) {
     ShiftLeftP(result_end, object_size, Operand(kPointerSizeLog2));
-    CmpP(r0, result_end);
-    blt(gc_required);
     AddP(result_end, result, result_end);
   } else {
-    CmpP(r0, object_size);
-    blt(gc_required);
     AddP(result_end, result, object_size);
   }
+  CmpLogicalP(result_end, MemOperand(top_address, limit - top));
+  bge(gc_required);
 
   // Update allocation top. result temporarily holds the new top.
   if (emit_debug_code()) {
     AndP(r0, result_end, Operand(kObjectAlignmentMask));
     Check(eq, kUnalignedAllocationInNewSpace, cr0);
   }
   if ((flags & ALLOCATION_FOLDING_DOMINATOR) == 0) {
     // The top pointer is not updated for allocation folding dominators.
     StoreP(result_end, MemOperand(top_address));
   }
 
+  // Prefetch the allocation_top's next cache line in advance to
+  // help alleviate potential cache misses.
+  // Mode 2 - Prefetch the data into a cache line for store access.
+  pfd(r2, MemOperand(result, 256));
+
   // Tag object.
-  AddP(result, result, Operand(kHeapObjectTag));
+  la(result, MemOperand(result, kHeapObjectTag));
 }
 
 void MacroAssembler::FastAllocate(Register object_size, Register result,
                                   Register result_end, Register scratch,
                                   AllocationFlags flags) {
   // |object_size| and |result_end| may overlap if the DOUBLE_ALIGNMENT flag
   // is not specified. Other registers must not overlap.
   DCHECK(!AreAliased(object_size, result, scratch, ip));
   DCHECK(!AreAliased(result_end, result, scratch, ip));
   DCHECK((flags & DOUBLE_ALIGNMENT) == 0 || !object_size.is(result_end));
@@ skipping 31 matching lines @@
     AddP(result_end, result, object_size);
   }
 
   // Update allocation top. result temporarily holds the new top.
   if (emit_debug_code()) {
     AndP(r0, result_end, Operand(kObjectAlignmentMask));
     Check(eq, kUnalignedAllocationInNewSpace, cr0);
   }
   StoreP(result_end, MemOperand(top_address));
 
+  // Prefetch the allocation_top's next cache line in advance to
+  // help alleviate potential cache misses.
+  // Mode 2 - Prefetch the data into a cache line for store access.
+  pfd(r2, MemOperand(result, 256));
+
   // Tag object.
-  AddP(result, result, Operand(kHeapObjectTag));
+  la(result, MemOperand(result, kHeapObjectTag));
 }
 
 void MacroAssembler::FastAllocate(int object_size, Register result,
                                   Register scratch1, Register scratch2,
                                   AllocationFlags flags) {
   DCHECK(object_size <= kMaxRegularHeapObjectSize);
   DCHECK(!AreAliased(result, scratch1, scratch2, ip));
 
   // Make object size into bytes.
   if ((flags & SIZE_IN_WORDS) != 0) {
@@ skipping 20 matching lines @@
     AndP(result_end, result, Operand(kDoubleAlignmentMask));
     Label aligned;
     beq(&aligned, Label::kNear);
     mov(result_end, Operand(isolate()->factory()->one_pointer_filler_map()));
     StoreW(result_end, MemOperand(result));
     AddP(result, result, Operand(kDoubleSize / 2));
     bind(&aligned);
 #endif
   }
 
-  // Calculate new top using result.
-  AddP(result_end, result, Operand(object_size));
+  if (CpuFeatures::IsSupported(GENERAL_INSTR_EXT) && is_int8(object_size)) {
+    // Update allocation top.
+    AddP(MemOperand(top_address), Operand(object_size));
+  } else {
+    // Calculate new top using result.
+    AddP(result_end, result, Operand(object_size));
 
-  // The top pointer is not updated for allocation folding dominators.
-  StoreP(result_end, MemOperand(top_address));
+    // Update allocation top.
+    StoreP(result_end, MemOperand(top_address));
+  }
+
+  // Prefetch the allocation_top's next cache line in advance to
+  // help alleviate potential cache misses.
+  // Mode 2 - Prefetch the data into a cache line for store access.
+  pfd(r2, MemOperand(result, 256));
 
   // Tag object.
-  AddP(result, result, Operand(kHeapObjectTag));
+  la(result, MemOperand(result, kHeapObjectTag));
 }
 
 void MacroAssembler::CompareObjectType(Register object, Register map,
                                        Register type_reg, InstanceType type) {
   const Register temp = type_reg.is(no_reg) ? r0 : type_reg;
 
   LoadP(map, FieldMemOperand(object, HeapObject::kMapOffset));
   CompareInstanceType(map, temp, type);
 }
 
@@ skipping 3253 matching lines @@
   }
   if (mag.shift > 0) ShiftRightArith(result, result, Operand(mag.shift));
   ExtractBit(r0, dividend, 31);
   AddP(result, r0);
 }
 
 }  // namespace internal
 }  // namespace v8
 
 #endif  // V8_TARGET_ARCH_S390