S390: [crankshaft] Fragmentation-free allocation folding.

Port 61f5fbbb
Port 6e15433d

Original commit message:

      The new allocation folding implementation avoids fragmentation between folded allocation.
      As a consequence, our heap will always be iterable i.e. we do not have to perform a
      garbage collection before iterating the heap.

R=hpayer@chromium.org, joransiu@ca.ibm.com, bjaideep@ca.ibm.com, michael_dawson@ca.ibm.com, mbrandy@us.ibm.com
BUG=

Review-Url: https://codereview.chromium.org/1973883003
Cr-Commit-Position: refs/heads/master@{#36243}

src/crankshaft/s390/lithium-codegen-s390.cc

@@ -5257,6 +5257,12 @@ void LCodeGen::DoAllocate(LAllocate* instr) {
     flags = static_cast<AllocationFlags>(flags | PRETENURE);
   }
 
+  if (instr->hydrogen()->IsAllocationFoldingDominator()) {
+    flags = static_cast<AllocationFlags>(flags | ALLOCATION_FOLDING_DOMINATOR);
+  }
+
+  DCHECK(!instr->hydrogen()->IsAllocationFolded());
+
   if (instr->size()->IsConstantOperand()) {
     int32_t size = ToInteger32(LConstantOperand::cast(instr->size()));
     CHECK(size <= Page::kMaxRegularHeapObjectSize);
@@ -5335,6 +5341,49 @@ void LCodeGen::DoDeferredAllocate(LAllocate* instr) {
   CallRuntimeFromDeferred(Runtime::kAllocateInTargetSpace, 2, instr,
                           instr->context());
   __ StoreToSafepointRegisterSlot(r2, result);
+
+  if (instr->hydrogen()->IsAllocationFoldingDominator()) {
+    AllocationFlags allocation_flags = NO_ALLOCATION_FLAGS;
+    if (instr->hydrogen()->IsOldSpaceAllocation()) {
+      DCHECK(!instr->hydrogen()->IsNewSpaceAllocation());
+      allocation_flags = static_cast<AllocationFlags>(flags | PRETENURE);
+    }
+    // If the allocation folding dominator allocate triggered a GC, allocation
+    // happend in the runtime. We have to reset the top pointer to virtually
+    // undo the allocation.
+    ExternalReference allocation_top =
+        AllocationUtils::GetAllocationTopReference(isolate(), allocation_flags);
+    Register top_address = scratch0();
+    __ SubP(r2, r2, Operand(kHeapObjectTag));
+    __ mov(top_address, Operand(allocation_top));
+    __ StoreP(r2, MemOperand(top_address));
+    __ AddP(r2, r2, Operand(kHeapObjectTag));
+  }
+}
+
+void LCodeGen::DoFastAllocate(LFastAllocate* instr) {
+  DCHECK(instr->hydrogen()->IsAllocationFolded());
+  DCHECK(!instr->hydrogen()->IsAllocationFoldingDominator());
+  Register result = ToRegister(instr->result());
+  Register scratch1 = ToRegister(instr->temp1());
+  Register scratch2 = ToRegister(instr->temp2());
+
+  AllocationFlags flags = ALLOCATION_FOLDED;
+  if (instr->hydrogen()->MustAllocateDoubleAligned()) {
+    flags = static_cast<AllocationFlags>(flags | DOUBLE_ALIGNMENT);
+  }
+  if (instr->hydrogen()->IsOldSpaceAllocation()) {
+    DCHECK(!instr->hydrogen()->IsNewSpaceAllocation());
+    flags = static_cast<AllocationFlags>(flags | PRETENURE);
+  }
+  if (instr->size()->IsConstantOperand()) {
+    int32_t size = ToInteger32(LConstantOperand::cast(instr->size()));
+    CHECK(size <= Page::kMaxRegularHeapObjectSize);
+    __ FastAllocate(size, result, scratch1, scratch2, flags);
+  } else {
+    Register size = ToRegister(instr->size());
+    __ FastAllocate(size, result, scratch1, scratch2, flags);
+  }
 }
 
 void LCodeGen::DoTypeof(LTypeof* instr) {

src/crankshaft/s390/lithium-s390.cc

@@ -2124,13 +2124,18 @@ LInstruction* LChunkBuilder::DoStringCharFromCode(HStringCharFromCode* instr) {
 }
 
 LInstruction* LChunkBuilder::DoAllocate(HAllocate* instr) {
-  info()->MarkAsDeferredCalling();
-  LOperand* context = UseAny(instr->context());
   LOperand* size = UseRegisterOrConstant(instr->size());
   LOperand* temp1 = TempRegister();
   LOperand* temp2 = TempRegister();
-  LAllocate* result = new (zone()) LAllocate(context, size, temp1, temp2);
-  return AssignPointerMap(DefineAsRegister(result));
+  if (instr->IsAllocationFolded()) {
+    LFastAllocate* result = new (zone()) LFastAllocate(size, temp1, temp2);
+    return DefineAsRegister(result);
+  } else {
+    info()->MarkAsDeferredCalling();
+    LOperand* context = UseAny(instr->context());
+    LAllocate* result = new (zone()) LAllocate(context, size, temp1, temp2);
+    return AssignPointerMap(DefineAsRegister(result));
+  }
 }
 
 LInstruction* LChunkBuilder::DoOsrEntry(HOsrEntry* instr) {

src/crankshaft/s390/lithium-s390.h

@@ -67,6 +67,7 @@ class LCodeGen;
   V(Drop)                                    \
   V(Dummy)                                   \
   V(DummyUse)                                \
+  V(FastAllocate)                            \
   V(FlooringDivByConstI)                     \
   V(FlooringDivByPowerOf2I)                  \
   V(FlooringDivI)                            \
@@ -2165,6 +2166,22 @@ class LAllocate final : public LTemplateInstruction<1, 2, 2> {
   DECLARE_HYDROGEN_ACCESSOR(Allocate)
 };
 
+class LFastAllocate final : public LTemplateInstruction<1, 1, 2> {
+ public:
+  LFastAllocate(LOperand* size, LOperand* temp1, LOperand* temp2) {
+    inputs_[0] = size;
+    temps_[0] = temp1;
+    temps_[1] = temp2;
+  }
+
+  LOperand* size() { return inputs_[0]; }
+  LOperand* temp1() { return temps_[0]; }
+  LOperand* temp2() { return temps_[1]; }
+
+  DECLARE_CONCRETE_INSTRUCTION(FastAllocate, "fast-allocate")
+  DECLARE_HYDROGEN_ACCESSOR(Allocate)
+};
+
 class LTypeof final : public LTemplateInstruction<1, 2, 0> {
  public:
   LTypeof(LOperand* context, LOperand* value) {

src/s390/macro-assembler-s390.cc

@@ -1706,6 +1706,7 @@ void MacroAssembler::Allocate(int object_size, Register result,
                               Register scratch1, Register scratch2,
                               Label* gc_required, AllocationFlags flags) {
   DCHECK(object_size <= Page::kMaxRegularHeapObjectSize);
+  DCHECK((flags & ALLOCATION_FOLDED) == 0);
   if (!FLAG_inline_new) {
     if (emit_debug_code()) {
       // Trash the registers to simulate an allocation failure.
@@ -1767,7 +1768,7 @@ void MacroAssembler::Allocate(int object_size, Register result,
     STATIC_ASSERT(kPointerAlignment * 2 == kDoubleAlignment);
     AndP(result_end, result, Operand(kDoubleAlignmentMask));
     Label aligned;
-    beq(&aligned);
+    beq(&aligned, Label::kNear);
     if ((flags & PRETENURE) != 0) {
       CmpLogicalP(result, alloc_limit);
       bge(gc_required);
@@ -1792,7 +1793,11 @@ void MacroAssembler::Allocate(int object_size, Register result,
     blt(gc_required);
     AddP(result_end, result, result_end);
   }
-  StoreP(result_end, MemOperand(top_address));
+
+  if ((flags & ALLOCATION_FOLDING_DOMINATOR) == 0) {
+    // The top pointer is not updated for allocation folding dominators.
+    StoreP(result_end, MemOperand(top_address));
+  }
 
   // Tag object.
   AddP(result, result, Operand(kHeapObjectTag));
@@ -1801,6 +1806,7 @@ void MacroAssembler::Allocate(int object_size, Register result,
 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.
@@ -1858,7 +1864,7 @@ void MacroAssembler::Allocate(Register object_size, Register result,
     STATIC_ASSERT(kPointerAlignment * 2 == kDoubleAlignment);
     AndP(result_end, result, Operand(kDoubleAlignmentMask));
     Label aligned;
-    beq(&aligned);
+    beq(&aligned, Label::kNear);
     if ((flags & PRETENURE) != 0) {
       CmpLogicalP(result, alloc_limit);
       bge(gc_required);
@@ -1890,6 +1896,110 @@ void MacroAssembler::Allocate(Register object_size, Register result,
     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));
+  }
+
+  // Tag object.
+  AddP(result, result, Operand(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));
+
+  ExternalReference allocation_top =
+      AllocationUtils::GetAllocationTopReference(isolate(), flags);
+
+  Register top_address = scratch;
+  mov(top_address, Operand(allocation_top));
+  LoadP(result, MemOperand(top_address));
+
+  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
+    DCHECK(kPointerAlignment * 2 == kDoubleAlignment);
+    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. Object size may be in words so a shift is
+  // required to get the number of bytes.
+  if ((flags & SIZE_IN_WORDS) != 0) {
+    ShiftLeftP(result_end, object_size, Operand(kPointerSizeLog2));
+    AddP(result_end, result, result_end);
+  } else {
+    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));
+
+  // Tag object.
+  AddP(result, result, Operand(kHeapObjectTag));
+}
+
+void MacroAssembler::FastAllocate(int object_size, Register result,
+                                  Register scratch1, Register scratch2,
+                                  AllocationFlags flags) {
+  DCHECK(object_size <= Page::kMaxRegularHeapObjectSize);
+  DCHECK(!AreAliased(result, scratch1, scratch2, ip));
+
+  // Make object size into bytes.
+  if ((flags & SIZE_IN_WORDS) != 0) {
+    object_size *= kPointerSize;
+  }
+  DCHECK_EQ(0, object_size & kObjectAlignmentMask);
+
+  ExternalReference allocation_top =
+      AllocationUtils::GetAllocationTopReference(isolate(), flags);
+
+  // Set up allocation top address register.
+  Register top_address = scratch1;
+  Register result_end = scratch2;
+  mov(top_address, Operand(allocation_top));
+  LoadP(result, MemOperand(top_address));
+
+  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
+    DCHECK(kPointerAlignment * 2 == kDoubleAlignment);
+    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));
+
+  // The top pointer is not updated for allocation folding dominators.
   StoreP(result_end, MemOperand(top_address));
 
   // Tag object.

src/s390/macro-assembler-s390.h

@@ -965,6 +965,15 @@ class MacroAssembler : public Assembler {
   void Allocate(Register object_size, Register result, Register result_end,
                 Register scratch, Label* gc_required, AllocationFlags flags);
 
+  // FastAllocate is right now only used for folded allocations. It just
+  // increments the top pointer without checking against limit. This can only
+  // be done if it was proved earlier that the allocation will succeed.
+  void FastAllocate(int object_size, Register result, Register scratch1,
+                    Register scratch2, AllocationFlags flags);
+
+  void FastAllocate(Register object_size, Register result, Register result_end,
+                    Register scratch, AllocationFlags flags);
+
   void AllocateTwoByteString(Register result, Register length,
                              Register scratch1, Register scratch2,
                              Register scratch3, Label* gc_required);