[ARM] Add NEON instructions for implementing SIMD.

src/arm/macro-assembler-arm.cc

Index: src/arm/macro-assembler-arm.cc
diff --git a/src/arm/macro-assembler-arm.cc b/src/arm/macro-assembler-arm.cc
index 8363d5e44c4e347a2a9c8f1f0bb4007339624d7d..b34d218fdff3e9f75474af4e8ed7bbea96ba56e4 100644
--- a/src/arm/macro-assembler-arm.cc
+++ b/src/arm/macro-assembler-arm.cc
@@ -1185,6 +1185,64 @@ void MacroAssembler::ReplaceLane(QwNeonRegister dst, QwNeonRegister src,
   VmovExtended(s_code, src_lane.code(), scratch);
 }
 
+void MacroAssembler::Swizzle(QwNeonRegister dst, QwNeonRegister src,
+                             Register scratch, NeonSize size, uint32_t lanes) {
+  // TODO(bbudge) Handle Int16x8, Int8x16 vectors.
+  DCHECK_EQ(Neon32, size);
+  DCHECK_IMPLIES(size == Neon32, lanes < 0xFFFFu);
+  if (size == Neon32) {
+    switch (lanes) {
+      // TODO(bbudge) Handle more special cases.
+      case 0x3210:  // Identity.
+        Move(dst, src);
+        return;
+      case 0x1032:  // Swap top and bottom.

[Rodolph Perfetta (ARM), 2016/12/16 20:50:11]

vext(dst, src, src, 8);

[bbudge, 2016/12/17 01:08:32]

Nice, done.

+        Move(dst, src);
+        vswp(dst.low(), dst.high());
+        return;
+      case 0x2103:  // Rotation.
+        vext(dst, src, src, 12);
+        return;
+      case 0x0321:  // Rotation.
+        vext(dst, src, src, 4);
+        return;
+      case 0x0000:  // Equivalent to vdup.
+      case 0x1111:
+      case 0x2222:
+      case 0x3333: {
+        int lane_code = src.code() * 4 + (lanes & 0xF);
+        if (lane_code >= SwVfpRegister::kMaxNumRegisters) {

[Rodolph Perfetta (ARM), 2016/12/16 20:50:11]

TODO: use vdup (vdup.32 dst, Dsrc[lane]) once implemented.

[bbudge, 2016/12/17 01:08:32]

Done.

+          int temp_code = kScratchDoubleReg.code() * 2;
+          VmovExtended(temp_code, lane_code, scratch);
+          lane_code = temp_code;
+        }
+        vdup(dst, SwVfpRegister::from_code(lane_code));
+        return;
+      }
+      case 0x2301:  // Swap lanes 0, 1 and lanes 2, 3.
+        vrev(Neon64, Neon32, dst, src);
+        return;
+      default:  // Handle all other cases with vmovs.
+        int src_code = src.code() * 4;
+        int dst_code = dst.code() * 4;
+        bool in_place = src.is(dst);
+        if (in_place) {
+          vmov(kScratchQuadReg, src);
+          src_code = kScratchQuadReg.code() * 4;
+        }
+        for (int i = 0; i < 4; i++) {
+          int lane = (lanes >> (i * 4) & 0xF);
+          VmovExtended(dst_code + i, src_code + lane, scratch);
+        }
+        if (in_place) {
+          // Restore zero reg.
+          veor(kDoubleRegZero, kDoubleRegZero, kDoubleRegZero);
+        }
+        return;
+    }
+  }
+}
+
 void MacroAssembler::LslPair(Register dst_low, Register dst_high,
                              Register src_low, Register src_high,
                              Register scratch, Register shift) {