[compiler][wasm] Align Frame slots to value size

- Adds an AlignedSlotAllocator class and tests, to unify slot
  allocation. This attempts to use alignment holes for smaller
  values.
- Reworks Frame to use the new allocator for stack slots.
- Reworks LinkageAllocator to use the new allocator for stack
  slots and for ARMv7 FP register aliasing.
- Fixes the RegisterAllocator to align spill slots.
- Fixes InstructionSelector to align spill slots.

Bug: v8:9198

Change-Id: Ida148db428be89ef95de748ec5fc0e7b0358f523
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2512840
Commit-Queue: Bill Budge <bbudge@chromium.org>
Reviewed-by: Georg Neis <neis@chromium.org>
Reviewed-by: Andreas Haas <ahaas@chromium.org>
Cr-Commit-Position: refs/heads/master@{#71644}

BUILD.gn

@@ -2427,6 +2427,8 @@ v8_source_set("v8_base_without_compiler") {
     "src/builtins/constants-table-builder.cc",
     "src/builtins/constants-table-builder.h",
     "src/builtins/profile-data-reader.h",
+    "src/codegen/aligned-slot-allocator.cc",
+    "src/codegen/aligned-slot-allocator.h",
     "src/codegen/assembler-arch.h",
     "src/codegen/assembler-inl.h",
     "src/codegen/assembler.cc",

src/codegen/aligned-slot-allocator.cc

+// Copyright 2020 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 "src/codegen/aligned-slot-allocator.h"
+
+#include "src/base/bits.h"
+#include "src/base/logging.h"
+
+namespace v8 {
+namespace internal {
+
+int AlignedSlotAllocator::NextSlot(int n) const {
+  DCHECK(n == 1 || n == 2 || n == 4);
+  if (n <= 1 && IsValid(next1_)) return next1_;
+  if (n <= 2 && IsValid(next2_)) return next2_;
+  DCHECK(IsValid(next4_));
+  return next4_;
+}
+
+int AlignedSlotAllocator::Allocate(int n) {
+  DCHECK(n == 1 || n == 2 || n == 4);
+  // Check invariants.
+  DCHECK_EQ(0, next4_ & 3);
+  DCHECK_IMPLIES(IsValid(next2_), (next2_ & 1) == 0);
+
+  // The sentinel value kInvalidSlot is used to indicate no slot.
+  // next1_ is the index of the 1 slot fragment, or kInvalidSlot.
+  // next2_ is the 2-aligned index of the 2 slot fragment, or kInvalidSlot.
+  // next4_ is the 4-aligned index of the next 4 slot group. It is always valid.
+  // In order to ensure we only have a single 1- or 2-slot fragment, we greedily
+  // use any fragment that satisfies the request.
+  int result = kInvalidSlot;
+  switch (n) {
+    case 1: {
+      if (IsValid(next1_)) {
+        result = next1_;
+        next1_ = kInvalidSlot;
+      } else if (IsValid(next2_)) {
+        result = next2_;
+        next1_ = result + 1;
+        next2_ = kInvalidSlot;
+      } else {
+        result = next4_;
+        next1_ = result + 1;
+        next2_ = result + 2;
+        next4_ += 4;
+      }
+      break;
+    }
+    case 2: {
+      if (IsValid(next2_)) {
+        result = next2_;
+        next2_ = kInvalidSlot;
+      } else {
+        result = next4_;
+        next2_ = result + 2;
+        next4_ += 4;
+      }
+      break;
+    }
+    case 4: {
+      result = next4_;
+      next4_ += 4;
+      break;
+    }
+    default:
+      UNREACHABLE();
+      break;
+  }
+  DCHECK(IsValid(result));
+  size_ = std::max(size_, result + n);
+  return result;
+}
+
+int AlignedSlotAllocator::AllocateUnaligned(int n) {
+  DCHECK_GE(n, 0);
+  // Check invariants.
+  DCHECK_EQ(0, next4_ & 3);
+  DCHECK_IMPLIES(IsValid(next2_), (next2_ & 1) == 0);
+
+  // Reserve |n| slots at |size_|, invalidate fragments below the new |size_|,
+  // and add any new fragments beyond the new |size_|.
+  int result = size_;
+  size_ += n;
+  switch (size_ & 3) {
+    case 0: {
+      next1_ = next2_ = kInvalidSlot;
+      next4_ = size_;
+      break;
+    }
+    case 1: {
+      next1_ = size_;
+      next2_ = size_ + 1;
+      next4_ = size_ + 3;
+      break;
+    }
+    case 2: {
+      next1_ = kInvalidSlot;
+      next2_ = size_;
+      next4_ = size_ + 2;
+      break;
+    }
+    case 3: {
+      next1_ = size_;
+      next2_ = kInvalidSlot;
+      next4_ = size_ + 1;
+      break;
+    }
+  }
+  return result;
+}
+
+int AlignedSlotAllocator::Align(int n) {
+  DCHECK(base::bits::IsPowerOfTwo(n));
+  DCHECK_LE(n, 4);
+  int mask = n - 1;
+  int misalignment = size_ & mask;
+  int padding = (n - misalignment) & mask;
+  AllocateUnaligned(padding);
+  return padding;
+}
+
+}  // namespace internal
+}  // namespace v8

src/codegen/aligned-slot-allocator.h

+// Copyright 2020 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.
+
+#ifndef V8_CODEGEN_ALIGNED_SLOT_ALLOCATOR_H_
+#define V8_CODEGEN_ALIGNED_SLOT_ALLOCATOR_H_
+
+#include "src/base/macros.h"
+#include "src/base/platform/platform.h"
+#include "src/common/globals.h"
+
+namespace v8 {
+namespace internal {
+
+// An aligned slot allocator. Allocates groups of 1, 2, or 4 slots such that the
+// first slot of the group is aligned to the group size. The allocator can also
+// allocate unaligned groups of arbitrary size, and an align the number of slots
+// to 1, 2, or 4 slots. The allocator tries to be as thrifty as possible by
+// reusing alignment padding slots in subsequent smaller slot allocations.
+class V8_EXPORT_PRIVATE AlignedSlotAllocator {
+ public:
+  // Slots are always multiples of pointer-sized units.
+  static constexpr int kSlotSize = kSystemPointerSize;
+
+  static int NumSlotsForWidth(int bytes) {
+    return (bytes + kSlotSize - 1) / kSlotSize;
+  }
+
+  AlignedSlotAllocator() = default;
+
+  // Allocates |n| slots, where |n| must be 1, 2, or 4. Padding slots may be
+  // inserted for alignment.
+  // Returns the starting index of the slots, which is evenly divisible by |n|.
+  int Allocate(int n);
+
+  // Gets the starting index of the slots that would be returned by Allocate(n).
+  int NextSlot(int n) const;
+
+  // Allocates the given number of slots at the current end of the slot area,
+  // and returns the starting index of the slots. This resets any fragment
+  // slots, so subsequent allocations will be after the end of this one.
+  // AllocateUnaligned(0) can be used to partition the slot area, for example
+  // to make sure tagged values follow untagged values on a Frame.
+  int AllocateUnaligned(int n);
+
+  // Aligns the slot area so that future allocations begin at the alignment.
+  // Returns the number of slots needed to align the slot area.
+  int Align(int n);
+
+  // Returns the size of the slot area, in slots. This will be greater than any
+  // already allocated slot index.
+  int Size() const { return size_; }
+
+ private:
+  static constexpr int kInvalidSlot = -1;
+
+  static bool IsValid(int slot) { return slot > kInvalidSlot; }
+
+  int next1_ = kInvalidSlot;
+  int next2_ = kInvalidSlot;
+  int next4_ = 0;
+  int size_ = 0;
+
+  DISALLOW_NEW_AND_DELETE()
+};
+
+}  // namespace internal
+}  // namespace v8
+
+#endif  // V8_CODEGEN_ALIGNED_SLOT_ALLOCATOR_H_

src/compiler/backend/arm/instruction-selector-arm.cc

@@ -490,7 +490,7 @@ void VisitPairAtomicBinOp(InstructionSelector* selector, Node* node,
 
 void InstructionSelector::VisitStackSlot(Node* node) {
   StackSlotRepresentation rep = StackSlotRepresentationOf(node->op());
-  int slot = frame_->AllocateSpillSlot(rep.size());
+  int slot = frame_->AllocateSpillSlot(rep.size(), rep.alignment());
   OperandGenerator g(this);
 
   Emit(kArchStackSlot, g.DefineAsRegister(node),

src/compiler/backend/arm64/instruction-selector-arm64.cc

@@ -557,7 +557,7 @@ int32_t LeftShiftForReducedMultiply(Matcher* m) {
 
 void InstructionSelector::VisitStackSlot(Node* node) {
   StackSlotRepresentation rep = StackSlotRepresentationOf(node->op());
-  int slot = frame_->AllocateSpillSlot(rep.size());
+  int slot = frame_->AllocateSpillSlot(rep.size(), rep.alignment());
   OperandGenerator g(this);
 
   Emit(kArchStackSlot, g.DefineAsRegister(node),

src/compiler/backend/ia32/instruction-selector-ia32.cc

@@ -352,7 +352,7 @@ void VisitRROI8x16SimdShift(InstructionSelector* selector, Node* node,
 
 void InstructionSelector::VisitStackSlot(Node* node) {
   StackSlotRepresentation rep = StackSlotRepresentationOf(node->op());
-  int slot = frame_->AllocateSpillSlot(rep.size());
+  int slot = frame_->AllocateSpillSlot(rep.size(), rep.alignment());
   OperandGenerator g(this);
 
   Emit(kArchStackSlot, g.DefineAsRegister(node),

src/compiler/backend/ppc/instruction-selector-ppc.cc

@@ -155,7 +155,7 @@ void VisitBinop(InstructionSelector* selector, Node* node,
 
 void InstructionSelector::VisitStackSlot(Node* node) {
   StackSlotRepresentation rep = StackSlotRepresentationOf(node->op());
-  int slot = frame_->AllocateSpillSlot(rep.size());
+  int slot = frame_->AllocateSpillSlot(rep.size(), rep.alignment());
   OperandGenerator g(this);
 
   Emit(kArchStackSlot, g.DefineAsRegister(node),

src/compiler/backend/register-allocator.cc

@@ -4502,9 +4502,11 @@ void OperandAssigner::AssignSpillSlots() {
   for (SpillRange* range : spill_ranges) {
     data()->tick_counter()->TickAndMaybeEnterSafepoint();
     if (range == nullptr || range->IsEmpty()) continue;
-    // Allocate a new operand referring to the spill slot.
     if (!range->HasSlot()) {
-      int index = data()->frame()->AllocateSpillSlot(range->byte_width());
+      // Allocate a new operand referring to the spill slot, aligned to the
+      // operand size.
+      int width = range->byte_width();
+      int index = data()->frame()->AllocateSpillSlot(width, width);
       range->set_assigned_slot(index);
     }
   }

src/compiler/backend/s390/instruction-selector-s390.cc

@@ -680,7 +680,7 @@ void VisitBinOp(InstructionSelector* selector, Node* node,
 
 void InstructionSelector::VisitStackSlot(Node* node) {
   StackSlotRepresentation rep = StackSlotRepresentationOf(node->op());
-  int slot = frame_->AllocateSpillSlot(rep.size());
+  int slot = frame_->AllocateSpillSlot(rep.size(), rep.alignment());
   OperandGenerator g(this);
 
   Emit(kArchStackSlot, g.DefineAsRegister(node),

src/compiler/backend/x64/instruction-selector-x64.cc

@@ -335,7 +335,7 @@ ArchOpcode GetStoreOpcode(StoreRepresentation store_rep) {
 
 void InstructionSelector::VisitStackSlot(Node* node) {
   StackSlotRepresentation rep = StackSlotRepresentationOf(node->op());
-  int slot = frame_->AllocateSpillSlot(rep.size());
+  int slot = frame_->AllocateSpillSlot(rep.size(), rep.alignment());
   OperandGenerator g(this);
 
   Emit(kArchStackSlot, g.DefineAsRegister(node),

src/compiler/frame.cc

@@ -12,14 +12,15 @@ namespace compiler {
 
 Frame::Frame(int fixed_frame_size_in_slots)
     : fixed_slot_count_(fixed_frame_size_in_slots),
-      frame_slot_count_(fixed_frame_size_in_slots),
       spill_slot_count_(0),
       return_slot_count_(0),
       allocated_registers_(nullptr),
-      allocated_double_registers_(nullptr) {}
+      allocated_double_registers_(nullptr) {
+  slot_allocator_.AllocateUnaligned(fixed_frame_size_in_slots);
+}
 
 void Frame::AlignFrame(int alignment) {
-  int alignment_slots = alignment / kSystemPointerSize;
+  int alignment_slots = AlignedSlotAllocator::NumSlotsForWidth(alignment);
   // In the calculations below we assume that alignment_slots is a power of 2.
   DCHECK(base::bits::IsPowerOfTwo(alignment_slots));
 
@@ -28,11 +29,12 @@ void Frame::AlignFrame(int alignment) {
   int return_delta =
       alignment_slots - (return_slot_count_ & (alignment_slots - 1));
   if (return_delta != alignment_slots) {
-    frame_slot_count_ += return_delta;
+    slot_allocator_.Align(alignment_slots);
   }
-  int delta = alignment_slots - (frame_slot_count_ & (alignment_slots - 1));
+  int delta =
+      alignment_slots - (slot_allocator_.Size() & (alignment_slots - 1));
   if (delta != alignment_slots) {
-    frame_slot_count_ += delta;
+    slot_allocator_.Align(alignment_slots);
     if (spill_slot_count_ != 0) {
       spill_slot_count_ += delta;
     }

src/compiler/frame.h

@@ -5,6 +5,7 @@
 #ifndef V8_COMPILER_FRAME_H_
 #define V8_COMPILER_FRAME_H_
 
+#include "src/codegen/aligned-slot-allocator.h"
 #include "src/execution/frame-constants.h"
 #include "src/utils/bit-vector.h"
 
@@ -92,7 +93,7 @@ class V8_EXPORT_PRIVATE Frame : public ZoneObject {
   Frame(const Frame&) = delete;
   Frame& operator=(const Frame&) = delete;
 
-  inline int GetTotalFrameSlotCount() const { return frame_slot_count_; }
+  inline int GetTotalFrameSlotCount() const { return slot_allocator_.Size(); }
   inline int GetFixedSlotCount() const { return fixed_slot_count_; }
   inline int GetSpillSlotCount() const { return spill_slot_count_; }
   inline int GetReturnSlotCount() const { return return_slot_count_; }
@@ -112,36 +113,39 @@ class V8_EXPORT_PRIVATE Frame : public ZoneObject {
   }
 
   void AlignSavedCalleeRegisterSlots(int alignment = kDoubleSize) {
-    int alignment_slots = alignment / kSystemPointerSize;
-    int delta = alignment_slots - (frame_slot_count_ & (alignment_slots - 1));
-    if (delta != alignment_slots) {
-      frame_slot_count_ += delta;
-    }
-    spill_slot_count_ += delta;
+    int alignment_slots = AlignedSlotAllocator::NumSlotsForWidth(alignment);
+    int padding = slot_allocator_.Align(alignment_slots);
+    spill_slot_count_ += padding;
   }
 
   void AllocateSavedCalleeRegisterSlots(int count) {
-    frame_slot_count_ += count;
+    slot_allocator_.AllocateUnaligned(count);
   }
 
   int AllocateSpillSlot(int width, int alignment = 0) {
-    DCHECK_EQ(frame_slot_count_,
+    DCHECK_EQ(GetTotalFrameSlotCount(),
               fixed_slot_count_ + spill_slot_count_ + return_slot_count_);
-    int frame_slot_count_before = frame_slot_count_;
-    if (alignment > kSystemPointerSize) {
-      // Slots are pointer sized, so alignment greater than a pointer size
-      // requires allocating additional slots.
-      width += alignment - kSystemPointerSize;
+    int slots = AlignedSlotAllocator::NumSlotsForWidth(width);
+    int alignment_slots =
+        std::max(1, AlignedSlotAllocator::NumSlotsForWidth(alignment));
+    int old_end = slot_allocator_.Size();
+    int slot;
+    if (slots == alignment_slots) {
+      slot = slot_allocator_.Allocate(slots);
+    } else {
+      slot_allocator_.Align(alignment_slots);
+      slot = slot_allocator_.AllocateUnaligned(slots);
     }
-    AllocateAlignedFrameSlots(width);
-    spill_slot_count_ += frame_slot_count_ - frame_slot_count_before;
-    return frame_slot_count_ - return_slot_count_ - 1;
+    int end = slot_allocator_.Size();
+
+    spill_slot_count_ += end - old_end;
+    return slot + slots - return_slot_count_ - 1;
   }
 
   void EnsureReturnSlots(int count) {
     if (count > return_slot_count_) {
       count -= return_slot_count_;
-      frame_slot_count_ += count;
+      slot_allocator_.AllocateUnaligned(count);
       return_slot_count_ += count;
     }
   }
@@ -151,28 +155,15 @@ class V8_EXPORT_PRIVATE Frame : public ZoneObject {
   int ReserveSpillSlots(size_t slot_count) {
     DCHECK_EQ(0, spill_slot_count_);
     spill_slot_count_ += static_cast<int>(slot_count);
-    frame_slot_count_ += static_cast<int>(slot_count);
-    return frame_slot_count_ - 1;
-  }
-
- private:
-  void AllocateAlignedFrameSlots(int width) {
-    DCHECK_LT(0, width);
-    int new_frame_slots = (width + kSystemPointerSize - 1) / kSystemPointerSize;
-    // Align to 8 bytes if width is a multiple of 8 bytes, and to 16 bytes if
-    // multiple of 16.
-    int align_to =
-        (width & 15) == 0 ? 16 : (width & 7) == 0 ? 8 : kSystemPointerSize;
-    frame_slot_count_ = RoundUp(frame_slot_count_ + new_frame_slots,
-                                align_to / kSystemPointerSize);
-    DCHECK_LT(0, frame_slot_count_);
+    slot_allocator_.AllocateUnaligned(static_cast<int>(slot_count));
+    return slot_allocator_.Size() - 1;
   }
 
  private:
   int fixed_slot_count_;
-  int frame_slot_count_;
   int spill_slot_count_;
   int return_slot_count_;
+  AlignedSlotAllocator slot_allocator_;
   BitVector* allocated_registers_;
   BitVector* allocated_double_registers_;
 };

src/compiler/wasm-compiler.cc

@@ -7952,6 +7952,7 @@ class LinkageLocationAllocator {
 
   void SetStackOffset(int offset) { allocator_.SetStackOffset(offset); }
   int NumStackSlots() const { return allocator_.NumStackSlots(); }
+  void EndSlotArea() { allocator_.EndSlotArea(); }
 
  private:
   wasm::LinkageAllocator allocator_;
@@ -7989,6 +7990,10 @@ CallDescriptor* GetWasmCallDescriptor(
     auto l = params.Next(param);
     locations.AddParamAt(i + param_offset, l);
   }
+
+  // End the untagged area, so tagged slots come after.
+  params.EndSlotArea();
+
   for (size_t i = 0; i < parameter_count; i++) {
     MachineRepresentation param = fsig->GetParam(i).machine_representation();
     // Skip untagged parameters.

src/wasm/wasm-linkage.h

@@ -5,6 +5,7 @@
 #ifndef V8_WASM_WASM_LINKAGE_H_
 #define V8_WASM_WASM_LINKAGE_H_
 
+#include "src/codegen/aligned-slot-allocator.h"
 #include "src/codegen/assembler-arch.h"
 #include "src/codegen/machine-type.h"
 #include "src/codegen/signature.h"
@@ -44,7 +45,7 @@ constexpr DoubleRegister kFpReturnRegisters[] = {xmm1, xmm2};
 // ===========================================================================
 constexpr Register kGpParamRegisters[] = {r3, r0, r2, r6};
 constexpr Register kGpReturnRegisters[] = {r0, r1};
-// ARM d-registers must be in ascending order for correct allocation.
+// ARM d-registers must be in even/odd D-register pairs for correct allocation.
 constexpr DoubleRegister kFpParamRegisters[] = {d0, d1, d2, d3, d4, d5, d6, d7};
 constexpr DoubleRegister kFpReturnRegisters[] = {d0, d1};
 
@@ -133,18 +134,27 @@ class LinkageAllocator {
   bool CanAllocateFP(MachineRepresentation rep) const {
 #if V8_TARGET_ARCH_ARM
     switch (rep) {
-      case MachineRepresentation::kFloat32:
-        return fp_offset_ < fp_count_ && fp_regs_[fp_offset_].code() < 16;
-      case MachineRepresentation::kFloat64:
-        return extra_double_reg_ >= 0 || fp_offset_ < fp_count_;
-      case MachineRepresentation::kSimd128:
-        return ((fp_offset_ + 1) & ~1) + 1 < fp_count_;
+      case MachineRepresentation::kFloat32: {
+        // Get the next D-register (Liftoff only uses the even S-registers).
+        int next = fp_allocator_.NextSlot(2) / 2;
+        // Only the lower 16 D-registers alias S-registers.
+        return next < fp_count_ && fp_regs_[next].code() < 16;
+      }
+      case MachineRepresentation::kFloat64: {
+        int next = fp_allocator_.NextSlot(2) / 2;
+        return next < fp_count_;
+      }
+      case MachineRepresentation::kSimd128: {
+        int next = fp_allocator_.NextSlot(4) / 2;
+        return next < fp_count_ - 1;  // 2 D-registers are required.
+      }
       default:
         UNREACHABLE();
         return false;
     }
-#endif
+#else
     return fp_offset_ < fp_count_;
+#endif
   }
 
   int NextGpReg() {
@@ -153,80 +163,58 @@ class LinkageAllocator {
   }
 
   int NextFpReg(MachineRepresentation rep) {
+    DCHECK(CanAllocateFP(rep));
 #if V8_TARGET_ARCH_ARM
     switch (rep) {
       case MachineRepresentation::kFloat32: {
-        // Liftoff uses only even-numbered f32 registers, and encodes them using
-        // the code of the corresponding f64 register. This limits the calling
-        // interface to only using the even-numbered f32 registers.
+        // Liftoff uses only even-numbered S-registers, and encodes them using
+        // the code of the corresponding D-register. This limits the calling
+        // interface to only using the even-numbered S-registers.
         int d_reg_code = NextFpReg(MachineRepresentation::kFloat64);
-        DCHECK_GT(16, d_reg_code);  // D-registers 16 - 31 can't split.
+        DCHECK_GT(16, d_reg_code);  // D16 - D31 don't alias S-registers.
         return d_reg_code * 2;
       }
       case MachineRepresentation::kFloat64: {
-        // Use the extra D-register if there is one.
-        if (extra_double_reg_ >= 0) {
-          int reg_code = extra_double_reg_;
-          extra_double_reg_ = -1;
-          return reg_code;
-        }
-        DCHECK_LT(fp_offset_, fp_count_);
-        return fp_regs_[fp_offset_++].code();
+        int next = fp_allocator_.Allocate(2) / 2;
+        return fp_regs_[next].code();
       }
       case MachineRepresentation::kSimd128: {
-        // Q-register must be an even-odd pair, so we must try to allocate at
-        // the end, not using extra_double_reg_. If we are at an odd D-register,
-        // skip past it (saving it to extra_double_reg_).
-        DCHECK_LT(((fp_offset_ + 1) & ~1) + 1, fp_count_);
-        int d_reg1_code = fp_regs_[fp_offset_++].code();
-        if (d_reg1_code % 2 != 0) {
-          // If we're misaligned then extra_double_reg_ must have been consumed.
-          DCHECK_EQ(-1, extra_double_reg_);
-          int odd_double_reg = d_reg1_code;
-          d_reg1_code = fp_regs_[fp_offset_++].code();
-          extra_double_reg_ = odd_double_reg;
-        }
-        // Combine the current D-register with the next to form a Q-register.
-        int d_reg2_code = fp_regs_[fp_offset_++].code();
-        DCHECK_EQ(0, d_reg1_code % 2);
-        DCHECK_EQ(d_reg1_code + 1, d_reg2_code);
-        USE(d_reg2_code);
-        return d_reg1_code / 2;
+        int next = fp_allocator_.Allocate(4) / 2;
+        int d_reg_code = fp_regs_[next].code();
+        // Check that result and the next D-register pair.
+        DCHECK_EQ(0, d_reg_code % 2);
+        DCHECK_EQ(d_reg_code + 1, fp_regs_[next + 1].code());
+        return d_reg_code / 2;
       }
       default:
         UNREACHABLE();
     }
 #else
-    DCHECK_LT(fp_offset_, fp_count_);
     return fp_regs_[fp_offset_++].code();
 #endif
   }
 
   // Stackslots are counted upwards starting from 0 (or the offset set by
-  // {SetStackOffset}.
-  int NumStackSlots(MachineRepresentation type) {
-    return std::max(1, ElementSizeInBytes(type) / kSystemPointerSize);
-  }
-
-  // Stackslots are counted upwards starting from 0 (or the offset set by
-  // {SetStackOffset}. If {type} needs more than
-  // one stack slot, the lowest used stack slot is returned.
+  // {SetStackOffset}. If {type} needs more than one stack slot, the lowest
+  // used stack slot is returned.
   int NextStackSlot(MachineRepresentation type) {
-    int num_stack_slots = NumStackSlots(type);
-    int offset = stack_offset_;
-    stack_offset_ += num_stack_slots;
-    return offset;
+    int num_slots =
+        AlignedSlotAllocator::NumSlotsForWidth(ElementSizeInBytes(type));
+    int slot = slot_allocator_.Allocate(num_slots);
+    return slot;
   }
 
   // Set an offset for the stack slots returned by {NextStackSlot} and
   // {NumStackSlots}. Can only be called before any call to {NextStackSlot}.
-  void SetStackOffset(int num) {
-    DCHECK_LE(0, num);
-    DCHECK_EQ(0, stack_offset_);
-    stack_offset_ = num;
+  void SetStackOffset(int offset) {
+    DCHECK_LE(0, offset);
+    DCHECK_EQ(0, slot_allocator_.Size());
+    slot_allocator_.AllocateUnaligned(offset);
   }
 
-  int NumStackSlots() const { return stack_offset_; }
+  int NumStackSlots() const { return slot_allocator_.Size(); }
+
+  void EndSlotArea() { slot_allocator_.AllocateUnaligned(0); }
 
  private:
   const int gp_count_;
@@ -234,16 +222,16 @@ class LinkageAllocator {
   const Register* const gp_regs_;
 
   const int fp_count_;
-  int fp_offset_ = 0;
-  const DoubleRegister* const fp_regs_;
-
 #if V8_TARGET_ARCH_ARM
-  // Track fragments of registers below fp_offset_ here. There can only be one
-  // extra double register.
-  int extra_double_reg_ = -1;
+  // Use an aligned slot allocator to model ARM FP register aliasing. The slots
+  // are 32 bits, so 2 slots are required for a D-register, 4 for a Q-register.
+  AlignedSlotAllocator fp_allocator_;
+#else
+  int fp_offset_ = 0;
 #endif
+  const DoubleRegister* const fp_regs_;
 
-  int stack_offset_ = 0;
+  AlignedSlotAllocator slot_allocator_;
 };
 
 }  // namespace wasm

test/unittests/BUILD.gn

@@ -215,6 +215,7 @@ v8_source_set("unittests_sources") {
     "base/threaded-list-unittest.cc",
     "base/utils/random-number-generator-unittest.cc",
     "base/vlq-base64-unittest.cc",
+    "codegen/aligned-slot-allocator-unittest.cc",
     "codegen/code-stub-assembler-unittest.cc",
     "codegen/code-stub-assembler-unittest.h",
     "codegen/register-configuration-unittest.cc",

test/unittests/codegen/aligned-slot-allocator-unittest.cc

+// Copyright 2020 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 "src/codegen/aligned-slot-allocator.h"
+
+#include "src/base/bits.h"
+#include "testing/gtest-support.h"
+
+namespace v8 {
+namespace internal {
+
+class AlignedSlotAllocatorUnitTest : public ::testing::Test {
+ public:
+  AlignedSlotAllocatorUnitTest() = default;
+  ~AlignedSlotAllocatorUnitTest() override = default;
+
+  // Helper method to test AlignedSlotAllocator::Allocate.
+  void Allocate(int size, int expected) {
+    int next = allocator_.NextSlot(size);
+    int result = allocator_.Allocate(size);
+    EXPECT_EQ(next, result);  // NextSlot/Allocate are consistent.
+    EXPECT_EQ(expected, result);
+    EXPECT_EQ(0, result & (size - 1));  // result is aligned to size.
+    int slot_end = result + static_cast<int>(base::bits::RoundUpToPowerOfTwo32(
+                                static_cast<uint32_t>(size)));
+    EXPECT_LE(slot_end, allocator_.Size());  // allocator Size is beyond slot.
+  }
+
+  // Helper method to test AlignedSlotAllocator::AllocateUnaligned.
+  void AllocateUnaligned(int size, int expected, int expected1, int expected2,
+                         int expected4) {
+    int size_before = allocator_.Size();
+    int result = allocator_.AllocateUnaligned(size);
+    EXPECT_EQ(size_before, result);  // AllocateUnaligned/Size are consistent.
+    EXPECT_EQ(expected, result);
+    EXPECT_EQ(result + size, allocator_.Size());
+    EXPECT_EQ(expected1, allocator_.NextSlot(1));
+    EXPECT_EQ(expected2, allocator_.NextSlot(2));
+    EXPECT_EQ(expected4, allocator_.NextSlot(4));
+  }
+
+  AlignedSlotAllocator allocator_;
+};
+
+TEST_F(AlignedSlotAllocatorUnitTest, Allocate1) {
+  Allocate(1, 0);
+  EXPECT_EQ(2, allocator_.NextSlot(2));
+  EXPECT_EQ(4, allocator_.NextSlot(4));
+
+  Allocate(1, 1);
+  EXPECT_EQ(2, allocator_.NextSlot(2));
+  EXPECT_EQ(4, allocator_.NextSlot(4));
+
+  Allocate(1, 2);
+  EXPECT_EQ(4, allocator_.NextSlot(2));
+  EXPECT_EQ(4, allocator_.NextSlot(4));
+
+  Allocate(1, 3);
+  EXPECT_EQ(4, allocator_.NextSlot(2));
+  EXPECT_EQ(4, allocator_.NextSlot(4));
+
+  // Make sure we use 1-fragments.
+  Allocate(1, 4);
+  Allocate(2, 6);
+  Allocate(1, 5);
+
+  // Make sure we use 2-fragments.
+  Allocate(2, 8);
+  Allocate(1, 10);
+  Allocate(1, 11);
+}
+
+TEST_F(AlignedSlotAllocatorUnitTest, Allocate2) {
+  Allocate(2, 0);
+  EXPECT_EQ(2, allocator_.NextSlot(1));
+  EXPECT_EQ(4, allocator_.NextSlot(4));
+
+  Allocate(2, 2);
+  EXPECT_EQ(4, allocator_.NextSlot(1));
+  EXPECT_EQ(4, allocator_.NextSlot(4));
+
+  // Make sure we use 2-fragments.
+  Allocate(1, 4);
+  Allocate(2, 6);
+  Allocate(2, 8);
+}
+
+TEST_F(AlignedSlotAllocatorUnitTest, Allocate4) {
+  Allocate(4, 0);
+  EXPECT_EQ(4, allocator_.NextSlot(1));
+  EXPECT_EQ(4, allocator_.NextSlot(2));
+
+  Allocate(1, 4);
+  Allocate(4, 8);
+
+  Allocate(2, 6);
+  Allocate(4, 12);
+}
+
+TEST_F(AlignedSlotAllocatorUnitTest, AllocateUnaligned) {
+  AllocateUnaligned(1, 0, 1, 2, 4);
+  AllocateUnaligned(1, 1, 2, 2, 4);
+
+  Allocate(1, 2);
+
+  AllocateUnaligned(2, 3, 5, 6, 8);
+
+  // Advance to leave 1- and 2- fragments below Size.
+  Allocate(4, 8);
+
+  // AllocateUnaligned should allocate at the end, and clear fragments.
+  AllocateUnaligned(0, 12, 12, 12, 12);
+}
+
+TEST_F(AlignedSlotAllocatorUnitTest, LargeAllocateUnaligned) {
+  AllocateUnaligned(11, 0, 11, 12, 12);
+  AllocateUnaligned(11, 11, 22, 22, 24);
+  AllocateUnaligned(13, 22, 35, 36, 36);
+}
+
+TEST_F(AlignedSlotAllocatorUnitTest, Size) {
+  allocator_.Allocate(1);
+  EXPECT_EQ(1, allocator_.Size());
+  // Allocate 2, leaving a fragment at 1. Size should be at 4.
+  allocator_.Allocate(2);
+  EXPECT_EQ(4, allocator_.Size());
+  // Allocate should consume fragment.
+  EXPECT_EQ(1, allocator_.Allocate(1));
+  // Size should still be 4.
+  EXPECT_EQ(4, allocator_.Size());
+}
+
+TEST_F(AlignedSlotAllocatorUnitTest, Align) {
+  EXPECT_EQ(0, allocator_.Align(1));
+  EXPECT_EQ(0, allocator_.Size());
+
+  // Allocate 1 to become misaligned.
+  Allocate(1, 0);
+
+  // 4-align.
+  EXPECT_EQ(3, allocator_.Align(4));
+  EXPECT_EQ(4, allocator_.NextSlot(1));
+  EXPECT_EQ(4, allocator_.NextSlot(2));
+  EXPECT_EQ(4, allocator_.NextSlot(4));
+  EXPECT_EQ(4, allocator_.Size());
+
+  // Allocate 2 to become misaligned.
+  Allocate(2, 4);
+
+  // 4-align.
+  EXPECT_EQ(2, allocator_.Align(4));
+  EXPECT_EQ(8, allocator_.NextSlot(1));
+  EXPECT_EQ(8, allocator_.NextSlot(2));
+  EXPECT_EQ(8, allocator_.NextSlot(4));
+  EXPECT_EQ(8, allocator_.Size());
+
+  // No change when we're already aligned.
+  EXPECT_EQ(0, allocator_.Align(2));
+  EXPECT_EQ(8, allocator_.NextSlot(1));
+  EXPECT_EQ(8, allocator_.NextSlot(2));
+  EXPECT_EQ(8, allocator_.NextSlot(4));
+  EXPECT_EQ(8, allocator_.Size());
+}
+
+}  // namespace internal
+}  // namespace v8