Reland of "MIPS: Optimize load/store with large offset"

Currently, we are using the following sequence for load/store
with large offset (offset > 16b):
    lui at, 0x1234
    ori at, at, 0x5678
    add at, s0, at
    lw a0, 0(at)

This sequence can be optimized in the following way:
    lui at, 0x1234
    add at, s0, at
    lw a0, 0x5678(at)

BUG=

Review-Url: https://codereview.chromium.org/2503493002
Cr-Commit-Position: refs/heads/master@{#40988}

src/compiler/mips/instruction-selector-mips.cc

@@ -80,12 +80,33 @@ class MipsOperandGenerator final : public OperandGenerator {
       case kMipsSub:
       case kMipsXor:
         return is_uint16(value);
+      case kMipsLb:
+      case kMipsLbu:
+      case kMipsSb:
+      case kMipsLh:
+      case kMipsLhu:
+      case kMipsSh:
+      case kMipsLw:
+      case kMipsSw:
+      case kMipsLwc1:
+      case kMipsSwc1:
       case kMipsLdc1:
       case kMipsSdc1:
+      case kCheckedLoadInt8:
+      case kCheckedLoadUint8:
+      case kCheckedLoadInt16:
+      case kCheckedLoadUint16:
+      case kCheckedLoadWord32:
+      case kCheckedStoreWord8:
+      case kCheckedStoreWord16:
+      case kCheckedStoreWord32:
+      case kCheckedLoadFloat32:
       case kCheckedLoadFloat64:
+      case kCheckedStoreFloat32:
       case kCheckedStoreFloat64:
-        return std::numeric_limits<int16_t>::min() <= (value + kIntSize) &&
-               std::numeric_limits<int16_t>::max() >= (value + kIntSize);
+        // true even for 32b values, offsets > 16b
+        // are handled in assembler-mips.cc
+        return is_int32(value);
       default:
         return is_int16(value);
     }
@@ -1801,6 +1822,7 @@ void InstructionSelector::VisitAtomicLoad(Node* node) {
       UNREACHABLE();
       return;
   }
+
   if (g.CanBeImmediate(index, opcode)) {
     Emit(opcode | AddressingModeField::encode(kMode_MRI),
          g.DefineAsRegister(node), g.UseRegister(base), g.UseImmediate(index));

src/mips/assembler-mips.cc

@@ -1784,13 +1784,44 @@ void Assembler::LoadRegPlusOffsetToAt(const MemOperand& src) {
   addu(at, at, src.rm());  // Add base register.
 }
 
+// Helper for base-reg + upper part of offset, when offset is larger than int16.
+// Loads higher part of the offset to AT register.
+// Returns lower part of the offset to be used as offset
+// in Load/Store instructions
+int32_t Assembler::LoadRegPlusUpperOffsetPartToAt(const MemOperand& src) {
+  DCHECK(!src.rm().is(at));
+  int32_t hi = (src.offset_ >> kLuiShift) & kImm16Mask;
+  // If the highest bit of the lower part of the offset is 1, this would make
+  // the offset in the load/store instruction negative. We need to compensate
+  // for this by adding 1 to the upper part of the offset.
+  if (src.offset_ & kNegOffset) {
+    hi += 1;
+  }
+  lui(at, hi);
+  addu(at, at, src.rm());
+  return (src.offset_ & kImm16Mask);
+}
+
+// Helper for loading base-reg + upper offset's part to AT reg when we are using
+// two 32-bit loads/stores instead of one 64-bit
+int32_t Assembler::LoadUpperOffsetForTwoMemoryAccesses(const MemOperand& src) {
+  DCHECK(!src.rm().is(at));
+  if (is_int16((src.offset_ & kImm16Mask) + kIntSize)) {
+    // Only if lower part of offset + kIntSize fits in 16bits
+    return LoadRegPlusUpperOffsetPartToAt(src);
+  }
+  // In case offset's lower part + kIntSize doesn't fit in 16bits,
+  // load reg + hole offset to AT
+  LoadRegPlusOffsetToAt(src);
+  return 0;
+}
 
 void Assembler::lb(Register rd, const MemOperand& rs) {
   if (is_int16(rs.offset_)) {
     GenInstrImmediate(LB, rs.rm(), rd, rs.offset_);
   } else {  // Offset > 16 bits, use multiple instructions to load.
-    LoadRegPlusOffsetToAt(rs);
-    GenInstrImmediate(LB, at, rd, 0);  // Equiv to lb(rd, MemOperand(at, 0));
+    int32_t off16 = LoadRegPlusUpperOffsetPartToAt(rs);
+    GenInstrImmediate(LB, at, rd, off16);
   }
 }
 
@@ -1799,8 +1830,8 @@ void Assembler::lbu(Register rd, const MemOperand& rs) {
   if (is_int16(rs.offset_)) {
     GenInstrImmediate(LBU, rs.rm(), rd, rs.offset_);
   } else {  // Offset > 16 bits, use multiple instructions to load.
-    LoadRegPlusOffsetToAt(rs);
-    GenInstrImmediate(LBU, at, rd, 0);  // Equiv to lbu(rd, MemOperand(at, 0));
+    int32_t off16 = LoadRegPlusUpperOffsetPartToAt(rs);
+    GenInstrImmediate(LBU, at, rd, off16);
   }
 }
 
@@ -1809,8 +1840,8 @@ void Assembler::lh(Register rd, const MemOperand& rs) {
   if (is_int16(rs.offset_)) {
     GenInstrImmediate(LH, rs.rm(), rd, rs.offset_);
   } else {  // Offset > 16 bits, use multiple instructions to load.
-    LoadRegPlusOffsetToAt(rs);
-    GenInstrImmediate(LH, at, rd, 0);  // Equiv to lh(rd, MemOperand(at, 0));
+    int32_t off16 = LoadRegPlusUpperOffsetPartToAt(rs);
+    GenInstrImmediate(LH, at, rd, off16);
   }
 }
 
@@ -1819,8 +1850,8 @@ void Assembler::lhu(Register rd, const MemOperand& rs) {
   if (is_int16(rs.offset_)) {
     GenInstrImmediate(LHU, rs.rm(), rd, rs.offset_);
   } else {  // Offset > 16 bits, use multiple instructions to load.
-    LoadRegPlusOffsetToAt(rs);
-    GenInstrImmediate(LHU, at, rd, 0);  // Equiv to lhu(rd, MemOperand(at, 0));
+    int32_t off16 = LoadRegPlusUpperOffsetPartToAt(rs);
+    GenInstrImmediate(LHU, at, rd, off16);
   }
 }
 
@@ -1829,8 +1860,8 @@ void Assembler::lw(Register rd, const MemOperand& rs) {
   if (is_int16(rs.offset_)) {
     GenInstrImmediate(LW, rs.rm(), rd, rs.offset_);
   } else {  // Offset > 16 bits, use multiple instructions to load.
-    LoadRegPlusOffsetToAt(rs);
-    GenInstrImmediate(LW, at, rd, 0);  // Equiv to lw(rd, MemOperand(at, 0));
+    int32_t off16 = LoadRegPlusUpperOffsetPartToAt(rs);
+    GenInstrImmediate(LW, at, rd, off16);
   }
 }
 
@@ -1855,8 +1886,8 @@ void Assembler::sb(Register rd, const MemOperand& rs) {
   if (is_int16(rs.offset_)) {
     GenInstrImmediate(SB, rs.rm(), rd, rs.offset_);
   } else {  // Offset > 16 bits, use multiple instructions to store.
-    LoadRegPlusOffsetToAt(rs);
-    GenInstrImmediate(SB, at, rd, 0);  // Equiv to sb(rd, MemOperand(at, 0));
+    int32_t off16 = LoadRegPlusUpperOffsetPartToAt(rs);
+    GenInstrImmediate(SB, at, rd, off16);
   }
 }
 
@@ -1865,8 +1896,8 @@ void Assembler::sh(Register rd, const MemOperand& rs) {
   if (is_int16(rs.offset_)) {
     GenInstrImmediate(SH, rs.rm(), rd, rs.offset_);
   } else {  // Offset > 16 bits, use multiple instructions to store.
-    LoadRegPlusOffsetToAt(rs);
-    GenInstrImmediate(SH, at, rd, 0);  // Equiv to sh(rd, MemOperand(at, 0));
+    int32_t off16 = LoadRegPlusUpperOffsetPartToAt(rs);
+    GenInstrImmediate(SH, at, rd, off16);
   }
 }
 
@@ -1875,8 +1906,8 @@ void Assembler::sw(Register rd, const MemOperand& rs) {
   if (is_int16(rs.offset_)) {
     GenInstrImmediate(SW, rs.rm(), rd, rs.offset_);
   } else {  // Offset > 16 bits, use multiple instructions to store.
-    LoadRegPlusOffsetToAt(rs);
-    GenInstrImmediate(SW, at, rd, 0);  // Equiv to sw(rd, MemOperand(at, 0));
+    int32_t off16 = LoadRegPlusUpperOffsetPartToAt(rs);
+    GenInstrImmediate(SW, at, rd, off16);
   }
 }
 
@@ -2172,8 +2203,8 @@ void Assembler::lwc1(FPURegister fd, const MemOperand& src) {
   if (is_int16(src.offset_)) {
     GenInstrImmediate(LWC1, src.rm(), fd, src.offset_);
   } else {  // Offset > 16 bits, use multiple instructions to load.
-    LoadRegPlusOffsetToAt(src);
-    GenInstrImmediate(LWC1, at, fd, 0);
+    int32_t off16 = LoadRegPlusUpperOffsetPartToAt(src);
+    GenInstrImmediate(LWC1, at, fd, off16);
   }
 }
 
@@ -2190,11 +2221,11 @@ void Assembler::ldc1(FPURegister fd, const MemOperand& src) {
       GenInstrImmediate(LWC1, src.rm(), nextfpreg,
                         src.offset_ + Register::kExponentOffset);
     } else {  // Offset > 16 bits, use multiple instructions to load.
-      LoadRegPlusOffsetToAt(src);
-      GenInstrImmediate(LWC1, at, fd, Register::kMantissaOffset);
+      int32_t off16 = LoadUpperOffsetForTwoMemoryAccesses(src);
+      GenInstrImmediate(LWC1, at, fd, off16 + Register::kMantissaOffset);
       FPURegister nextfpreg;
       nextfpreg.setcode(fd.code() + 1);
-      GenInstrImmediate(LWC1, at, nextfpreg, Register::kExponentOffset);
+      GenInstrImmediate(LWC1, at, nextfpreg, off16 + Register::kExponentOffset);
     }
   } else {
     DCHECK(IsFp64Mode() || IsFpxxMode());
@@ -2207,9 +2238,9 @@ void Assembler::ldc1(FPURegister fd, const MemOperand& src) {
                         src.offset_ + Register::kExponentOffset);
       mthc1(at, fd);
     } else {  // Offset > 16 bits, use multiple instructions to load.
-      LoadRegPlusOffsetToAt(src);
-      GenInstrImmediate(LWC1, at, fd, Register::kMantissaOffset);
-      GenInstrImmediate(LW, at, at, Register::kExponentOffset);
+      int32_t off16 = LoadUpperOffsetForTwoMemoryAccesses(src);
+      GenInstrImmediate(LWC1, at, fd, off16 + Register::kMantissaOffset);
+      GenInstrImmediate(LW, at, at, off16 + Register::kExponentOffset);
       mthc1(at, fd);
     }
   }
@@ -2220,8 +2251,8 @@ void Assembler::swc1(FPURegister fd, const MemOperand& src) {
   if (is_int16(src.offset_)) {
     GenInstrImmediate(SWC1, src.rm(), fd, src.offset_);
   } else {  // Offset > 16 bits, use multiple instructions to load.
-    LoadRegPlusOffsetToAt(src);
-    GenInstrImmediate(SWC1, at, fd, 0);
+    int32_t off16 = LoadRegPlusUpperOffsetPartToAt(src);
+    GenInstrImmediate(SWC1, at, fd, off16);
   }
 }
 
@@ -2240,11 +2271,11 @@ void Assembler::sdc1(FPURegister fd, const MemOperand& src) {
       GenInstrImmediate(SWC1, src.rm(), nextfpreg,
                         src.offset_ + Register::kExponentOffset);
     } else {  // Offset > 16 bits, use multiple instructions to load.
-      LoadRegPlusOffsetToAt(src);
-      GenInstrImmediate(SWC1, at, fd, Register::kMantissaOffset);
+      int32_t off16 = LoadUpperOffsetForTwoMemoryAccesses(src);
+      GenInstrImmediate(SWC1, at, fd, off16 + Register::kMantissaOffset);
       FPURegister nextfpreg;
       nextfpreg.setcode(fd.code() + 1);
-      GenInstrImmediate(SWC1, at, nextfpreg, Register::kExponentOffset);
+      GenInstrImmediate(SWC1, at, nextfpreg, off16 + Register::kExponentOffset);
     }
   } else {
     DCHECK(IsFp64Mode() || IsFpxxMode());
@@ -2257,10 +2288,10 @@ void Assembler::sdc1(FPURegister fd, const MemOperand& src) {
       GenInstrImmediate(SW, src.rm(), at,
                         src.offset_ + Register::kExponentOffset);
     } else {  // Offset > 16 bits, use multiple instructions to load.
-      LoadRegPlusOffsetToAt(src);
-      GenInstrImmediate(SWC1, at, fd, Register::kMantissaOffset);
+      int32_t off16 = LoadUpperOffsetForTwoMemoryAccesses(src);
+      GenInstrImmediate(SWC1, at, fd, off16 + Register::kMantissaOffset);
       mfhc1(t8, fd);
-      GenInstrImmediate(SW, at, t8, Register::kExponentOffset);
+      GenInstrImmediate(SW, at, t8, off16 + Register::kExponentOffset);
     }
   }
 }

src/mips/assembler-mips.h

@@ -1178,6 +1178,8 @@ class Assembler : public AssemblerBase {
 
   // Helpers.
   void LoadRegPlusOffsetToAt(const MemOperand& src);
+  int32_t LoadRegPlusUpperOffsetPartToAt(const MemOperand& src);
+  int32_t LoadUpperOffsetForTwoMemoryAccesses(const MemOperand& src);
 
   // Relocation for a type-recording IC has the AST id added to it.  This
   // member variable is a way to pass the information from the call site to

test/unittests/compiler/mips/instruction-selector-mips-unittest.cc

@@ -928,14 +928,13 @@ const MemoryAccessImm kMemoryAccessesImm[] = {
       -87, -86, -82, -44, -23, -3, 0, 7, 10, 39, 52, 69, 71, 91, 92, 107, 109,
       115, 124, 286, 655, 1362, 1569, 2587, 3067, 3096, 3462, 3510, 4095}}};
 
-
 const MemoryAccessImm1 kMemoryAccessImmMoreThan16bit[] = {
     {MachineType::Int8(),
      kMipsLb,
      kMipsSb,
      &InstructionSelectorTest::Stream::IsInteger,
      {-65000, -55000, 32777, 55000, 65000}},
-    {MachineType::Int8(),
+    {MachineType::Uint8(),
      kMipsLbu,
      kMipsSb,
      &InstructionSelectorTest::Stream::IsInteger,
@@ -945,7 +944,7 @@ const MemoryAccessImm1 kMemoryAccessImmMoreThan16bit[] = {
      kMipsSh,
      &InstructionSelectorTest::Stream::IsInteger,
      {-65000, -55000, 32777, 55000, 65000}},
-    {MachineType::Int16(),
+    {MachineType::Uint16(),
      kMipsLhu,
      kMipsSh,
      &InstructionSelectorTest::Stream::IsInteger,
@@ -1160,11 +1159,9 @@ TEST_P(InstructionSelectorMemoryAccessImmMoreThan16bitTest,
     StreamBuilder m(this, memacc.type, MachineType::Pointer());
     m.Return(m.Load(memacc.type, m.Parameter(0), m.Int32Constant(index)));
     Stream s = m.Build();
-    ASSERT_EQ(2U, s.size());
-    // kMipsAdd is expected opcode.
-    // size more than 16 bits wide.
-    EXPECT_EQ(kMipsAdd, s[0]->arch_opcode());
-    EXPECT_EQ(kMode_None, s[0]->addressing_mode());
+    ASSERT_EQ(1U, s.size());
+    EXPECT_EQ(memacc.load_opcode, s[0]->arch_opcode());
+    EXPECT_EQ(kMode_MRI, s[0]->addressing_mode());
     EXPECT_EQ(2U, s[0]->InputCount());
     EXPECT_EQ(1U, s[0]->OutputCount());
   }
@@ -1181,13 +1178,11 @@ TEST_P(InstructionSelectorMemoryAccessImmMoreThan16bitTest,
             m.Int32Constant(index), m.Parameter(1), kNoWriteBarrier);
     m.Return(m.Int32Constant(0));
     Stream s = m.Build();
-    ASSERT_EQ(2U, s.size());
-    // kMipsAdd is expected opcode
-    // size more than 16 bits wide
-    EXPECT_EQ(kMipsAdd, s[0]->arch_opcode());
-    EXPECT_EQ(kMode_None, s[0]->addressing_mode());
-    EXPECT_EQ(2U, s[0]->InputCount());
-    EXPECT_EQ(1U, s[0]->OutputCount());
+    ASSERT_EQ(1U, s.size());
+    EXPECT_EQ(memacc.store_opcode, s[0]->arch_opcode());
+    EXPECT_EQ(kMode_MRI, s[0]->addressing_mode());
+    EXPECT_EQ(3U, s[0]->InputCount());
+    EXPECT_EQ(0U, s[0]->OutputCount());
   }
 }