[Turbofan] Add ARM support for simd128 moves and swaps.

src/compiler/arm/code-generator-arm.cc

Index: src/compiler/arm/code-generator-arm.cc
diff --git a/src/compiler/arm/code-generator-arm.cc b/src/compiler/arm/code-generator-arm.cc
index c473b9b6aa5d04923149a5c41f4b1f5d27beb5f6..e15e8c91156a7711b976c6d7a88a3e444b8b7fd1 100644
--- a/src/compiler/arm/code-generator-arm.cc
+++ b/src/compiler/arm/code-generator-arm.cc
@@ -1891,8 +1891,7 @@ void CodeGenerator::AssembleMove(InstructionOperand* source,
         DCHECK(destination->IsDoubleStackSlot());
         __ vstr(src, g.ToMemOperand(destination));
       }
-    } else {
-      DCHECK_EQ(MachineRepresentation::kFloat32, rep);
+    } else if (rep == MachineRepresentation::kFloat32) {
       // GapResolver may give us reg codes that don't map to actual s-registers.
       // Generate code to work around those cases.
       int src_code = LocationOperand::cast(source)->register_code();
@@ -1903,6 +1902,23 @@ void CodeGenerator::AssembleMove(InstructionOperand* source,
         DCHECK(destination->IsFloatStackSlot());
         __ VmovExtended(g.ToMemOperand(destination), src_code, kScratchReg);
       }
+    } else {
+      DCHECK_EQ(MachineRepresentation::kSimd128, rep);
+      QwNeonRegister src = g.ToSimd128Register(source);
+      if (destination->IsSimd128Register()) {
+        QwNeonRegister dst = g.ToSimd128Register(destination);
+        __ Move(dst, src);
+      } else {
+        DCHECK(destination->IsSimd128StackSlot());
+        MemOperand dst = g.ToMemOperand(destination);
+        __ add(kScratchReg, dst.rn(), Operand(dst.offset()));
+        __ vst1(Neon8, NeonListOperand(src.low(), 2),
+                NeonMemOperand(kScratchReg));
+        // // TODO(bbudge) Use vst1 when it's available.

[martyn.capewell, 2016/11/24 11:40:58]

I hadn't realised vld1/vst1 had already been implemented. I guess these comments
can go now.

As these are NEON instructions, this will make NEON a requirement for
Simd128Register support. Is that the expectation?

I notice that in other parts of this patch, there is non-NEON fallback code for
Q-sized operations, eg. in the macro assembler for Move and Swap Q. If you can
assume Simd128 implies Q registers, then Q registers implies NEON, and the
fallback code for Q in the macro assembler can be removed.

Otherwise, you may need to make a macro Vld1 which selects between vld1 and a
vldr sequence on D registers, depending on NEON availability, and similarly for
Vst1.

[bbudge, 2016/11/24 13:48:05]

I forgot to remove these commented out sections.

It's true that we should not be emitting q-register moves/swaps here if NEON
isn't supported (in which case we do SIMD scalar lowering) so I think we should
remove the fallback code for q-register move/swap in the macro assembler. 

+        // __ vstr(src.low(), dst);
+        // dst.set_offset(dst.offset() + kDoubleSize);
+        // __ vstr(src.high(), dst);
+      }
     }
   } else if (source->IsFPStackSlot()) {
     MemOperand src = g.ToMemOperand(source);
@@ -1911,24 +1927,49 @@ void CodeGenerator::AssembleMove(InstructionOperand* source,
     if (destination->IsFPRegister()) {
       if (rep == MachineRepresentation::kFloat64) {
         __ vldr(g.ToDoubleRegister(destination), src);
-      } else {
-        DCHECK_EQ(MachineRepresentation::kFloat32, rep);
+      } else if (rep == MachineRepresentation::kFloat32) {
         // GapResolver may give us reg codes that don't map to actual
         // s-registers. Generate code to work around those cases.
         int dst_code = LocationOperand::cast(destination)->register_code();
         __ VmovExtended(dst_code, src, kScratchReg);
+      } else {
+        DCHECK_EQ(MachineRepresentation::kSimd128, rep);
+        QwNeonRegister dst = g.ToSimd128Register(destination);
+        __ add(kScratchReg, src.rn(), Operand(src.offset()));
+        __ vld1(Neon8, NeonListOperand(dst.low(), 2),
+                NeonMemOperand(kScratchReg));
+        // // TODO(bbudge) Use vld1 when it's available.
+        // __ vldr(dst.low(), src);
+        // src.set_offset(src.offset() + kDoubleSize);
+        // __ vldr(dst.high(), src);
       }
-    } else {
+    } else if (rep == MachineRepresentation::kFloat64) {
       DCHECK(destination->IsFPStackSlot());
       if (rep == MachineRepresentation::kFloat64) {
         DwVfpRegister temp = kScratchDoubleReg;
         __ vldr(temp, src);
         __ vstr(temp, g.ToMemOperand(destination));
-      } else {
-        DCHECK_EQ(MachineRepresentation::kFloat32, rep);
+      } else if (rep == MachineRepresentation::kFloat32) {
         SwVfpRegister temp = kScratchDoubleReg.low();
         __ vldr(temp, src);
         __ vstr(temp, g.ToMemOperand(destination));
+      } else {
+        DCHECK_EQ(MachineRepresentation::kSimd128, rep);
+        MemOperand dst = g.ToMemOperand(destination);
+        __ add(kScratchReg, src.rn(), Operand(src.offset()));
+        __ vld1(Neon8, NeonListOperand(kScratchQuadReg.low(), 2),
+                NeonMemOperand(kScratchReg));
+        __ add(kScratchReg, dst.rn(), Operand(dst.offset()));
+        __ vst1(Neon8, NeonListOperand(kScratchQuadReg.low(), 2),
+                NeonMemOperand(kScratchReg));
+        __ veor(kDoubleRegZero, kDoubleRegZero, kDoubleRegZero);
+        // // TODO(bbudge) Use vld1/vst1 when they're available.
+        // __ vldr(kScratchDoubleReg, src);
+        // __ vstr(kScratchDoubleReg, dst);
+        // src.set_offset(src.offset() + kDoubleSize);
+        // dst.set_offset(dst.offset() + kDoubleSize);
+        // __ vldr(kScratchDoubleReg, src);
+        // __ vstr(kScratchDoubleReg, dst);
       }
     }
   } else {
@@ -1936,7 +1977,6 @@ void CodeGenerator::AssembleMove(InstructionOperand* source,
   }
 }
 
-
 void CodeGenerator::AssembleSwap(InstructionOperand* source,
                                  InstructionOperand* destination) {
   ArmOperandConverter g(this, nullptr);
@@ -1975,7 +2015,7 @@ void CodeGenerator::AssembleSwap(InstructionOperand* source,
       DwVfpRegister src = g.ToDoubleRegister(source);
       if (destination->IsFPRegister()) {
         DwVfpRegister dst = g.ToDoubleRegister(destination);
-        __ vswp(src, dst);
+        __ Swap(src, dst);
       } else {
         DCHECK(destination->IsFPStackSlot());
         MemOperand dst = g.ToMemOperand(destination);
@@ -1983,8 +2023,7 @@ void CodeGenerator::AssembleSwap(InstructionOperand* source,
         __ vldr(src, dst);
         __ vstr(temp, dst);
       }
-    } else {
-      DCHECK_EQ(MachineRepresentation::kFloat32, rep);
+    } else if (rep == MachineRepresentation::kFloat32) {
       int src_code = LocationOperand::cast(source)->register_code();
       if (destination->IsFPRegister()) {
         int dst_code = LocationOperand::cast(destination)->register_code();
@@ -1998,29 +2037,65 @@ void CodeGenerator::AssembleSwap(InstructionOperand* source,
         __ VmovExtended(src_code, dst, kScratchReg);
         __ vstr(temp.low(), dst);
       }
+    } else {
+      DCHECK_EQ(MachineRepresentation::kSimd128, rep);
+      QwNeonRegister src = g.ToSimd128Register(source);
+      if (destination->IsFPRegister()) {
+        QwNeonRegister dst = g.ToSimd128Register(destination);
+        __ Swap(src, dst);
+      } else {
+        DCHECK(destination->IsFPStackSlot());
+        MemOperand dst = g.ToMemOperand(destination);
+        __ Move(kScratchQuadReg, src);
+        __ add(kScratchReg, dst.rn(), Operand(dst.offset()));
+        __ vld1(Neon8, NeonListOperand(src.low(), 2),
+                NeonMemOperand(kScratchReg));
+        __ vst1(Neon8, NeonListOperand(kScratchQuadReg.low(), 2),
+                NeonMemOperand(kScratchReg));
+        __ veor(kDoubleRegZero, kDoubleRegZero, kDoubleRegZero);
+        // QwNeonRegister temp = kScratchQuadReg;
+        // MemOperand dst2(dst.rn(), dst.offset() + kDoubleSize);
+        // // TODO(bbudge) Use vld1 / vst1 when they're available.
+        // __ Move(temp, src);
+        // __ vldr(src.low(), dst);
+        // __ vldr(src.high(), dst2);
+        // __ vstr(temp.low(), dst);
+        // __ vstr(temp.high(), dst2);
+        // // Restore the 0 register.
+        // __ veor(kDoubleRegZero, kDoubleRegZero, kDoubleRegZero);
+      }
     }
   } else if (source->IsFPStackSlot()) {
     DCHECK(destination->IsFPStackSlot());
-    Register temp_0 = kScratchReg;
-    LowDwVfpRegister temp_1 = kScratchDoubleReg;
-    MemOperand src0 = g.ToMemOperand(source);
-    MemOperand dst0 = g.ToMemOperand(destination);
+    MemOperand src = g.ToMemOperand(source);
+    MemOperand dst = g.ToMemOperand(destination);
     MachineRepresentation rep = LocationOperand::cast(source)->representation();
     if (rep == MachineRepresentation::kFloat64) {
-      MemOperand src1(src0.rn(), src0.offset() + kPointerSize);
-      MemOperand dst1(dst0.rn(), dst0.offset() + kPointerSize);
-      __ vldr(temp_1, dst0);  // Save destination in temp_1.
-      __ ldr(temp_0, src0);   // Then use temp_0 to copy source to destination.
-      __ str(temp_0, dst0);
-      __ ldr(temp_0, src1);
-      __ str(temp_0, dst1);
-      __ vstr(temp_1, src0);
+      __ vldr(kScratchDoubleReg, dst);
+      __ vldr(kDoubleRegZero, src);
+      __ vstr(kScratchDoubleReg, src);
+      __ vstr(kDoubleRegZero, dst);
+      // Restore the 0 register.
+      __ veor(kDoubleRegZero, kDoubleRegZero, kDoubleRegZero);
+    } else if (rep == MachineRepresentation::kFloat32) {
+      __ vldr(kScratchDoubleReg.low(), dst);
+      __ vldr(kScratchDoubleReg.high(), src);
+      __ vstr(kScratchDoubleReg.low(), src);
+      __ vstr(kScratchDoubleReg.high(), dst);
     } else {
-      DCHECK_EQ(MachineRepresentation::kFloat32, rep);
-      __ vldr(temp_1.low(), dst0);  // Save destination in temp_1.
-      __ ldr(temp_0, src0);  // Then use temp_0 to copy source to destination.
-      __ str(temp_0, dst0);
-      __ vstr(temp_1.low(), src0);
+      DCHECK_EQ(MachineRepresentation::kSimd128, rep);
+      __ vldr(kScratchDoubleReg, dst);
+      __ vldr(kDoubleRegZero, src);
+      __ vstr(kScratchDoubleReg, src);
+      __ vstr(kDoubleRegZero, dst);
+      src.set_offset(src.offset() + kDoubleSize);
+      dst.set_offset(dst.offset() + kDoubleSize);
+      __ vldr(kScratchDoubleReg, dst);
+      __ vldr(kDoubleRegZero, src);
+      __ vstr(kScratchDoubleReg, src);
+      __ vstr(kDoubleRegZero, dst);
+      // Restore the 0 register.
+      __ veor(kDoubleRegZero, kDoubleRegZero, kDoubleRegZero);
     }
   } else {
     // No other combinations are possible.