[ARM] Implement more NEON permutation instructions.

src/arm/assembler-arm.cc

 // Copyright (c) 1994-2006 Sun Microsystems Inc.
 // All Rights Reserved.
 //
 // Redistribution and use in source and binary forms, with or without
 // modification, are permitted provided that the following conditions
 // are met:
 //
 // - Redistributions of source code must retain the above copyright notice,
 // this list of conditions and the following disclaimer.
 //
@@ skipping 3885 matching lines @@
   // Instruction details available in ARM DDI 0406C.b, A8.8.320.
   // 1111(31-28) | 01000(27-23) | D(22) | 10(21-20) | Rn(19-16) |
   // Vd(15-12) | type(11-8) | size(7-6) | align(5-4) | Rm(3-0)
   DCHECK(IsEnabled(NEON));
   int vd, d;
   dst.base().split_code(&vd, &d);
   emit(0xFU*B28 | 4*B24 | d*B22 | 2*B20 | src.rn().code()*B16 | vd*B12 |
        dst.type()*B8 | size*B6 | src.align()*B4 | src.rm().code());
 }
 
-
-void Assembler::vst1(NeonSize size,
-                     const NeonListOperand& src,
+void Assembler::vst1(NeonSize size, const NeonListOperand& src,
                      const NeonMemOperand& dst) {
   // Instruction details available in ARM DDI 0406C.b, A8.8.404.
   // 1111(31-28) | 01000(27-23) | D(22) | 00(21-20) | Rn(19-16) |
   // Vd(15-12) | type(11-8) | size(7-6) | align(5-4) | Rm(3-0)
   DCHECK(IsEnabled(NEON));
   int vd, d;
   src.base().split_code(&vd, &d);
   emit(0xFU*B28 | 4*B24 | d*B22 | dst.rn().code()*B16 | vd*B12 | src.type()*B8 |
        size*B6 | dst.align()*B4 | dst.rm().code());
 }
@@ skipping 58 matching lines @@
   // vmov Arm scalar to core register.
   DCHECK(dt == NeonS32 || dt == NeonU32 || IsEnabled(NEON));
   int vn, n;
   src.split_code(&vn, &n);
   int opc1_opc2 = EncodeScalar(dt, index);
   int u = NeonU(dt);
   emit(0xEEu * B24 | u * B23 | B20 | vn * B16 | dst.code() * B12 | 0xB * B8 |
        n * B7 | B4 | opc1_opc2);
 }
 
-void Assembler::vmov(const QwNeonRegister dst, const QwNeonRegister src) {
+void Assembler::vmov(QwNeonRegister dst, QwNeonRegister src) {
   // Instruction details available in ARM DDI 0406C.b, A8-938.
   // vmov is encoded as vorr.
   vorr(dst, src, src);
 }
 
-void Assembler::vmvn(const QwNeonRegister dst, const QwNeonRegister src) {
-  DCHECK(IsEnabled(NEON));
-  // Instruction details available in ARM DDI 0406C.b, A8-966.
-  DCHECK(VfpRegisterIsAvailable(dst));
-  DCHECK(VfpRegisterIsAvailable(src));
-  int vd, d;
-  dst.split_code(&vd, &d);
-  int vm, m;
-  src.split_code(&vm, &m);
-  emit(0x1E7U * B23 | d * B22 | 3 * B20 | vd * B12 | 0x17 * B6 | m * B5 | vm);
-}
-
-void Assembler::vswp(DwVfpRegister dst, DwVfpRegister src) {
-  // Instruction details available in ARM DDI 0406C.b, A8.8.418.
-  // 1111(31-28) | 00111(27-23) | D(22) | 110010(21-16) |
-  // Vd(15-12) | 000000(11-6) | M(5) | 0(4) | Vm(3-0)
-  DCHECK(IsEnabled(NEON));
-  int vd, d;
-  dst.split_code(&vd, &d);
-  int vm, m;
-  src.split_code(&vm, &m);
-  emit(0xFU * B28 | 7 * B23 | d * B22 | 0x32 * B16 | vd * B12 | m * B5 | vm);
-}
-
-void Assembler::vswp(QwNeonRegister dst, QwNeonRegister src) {
-  // Instruction details available in ARM DDI 0406C.b, A8.8.418.
-  // 1111(31-28) | 00111(27-23) | D(22) | 110010(21-16) |
-  // Vd(15-12) | 000000(11-6) | M(5) | 0(4) | Vm(3-0)
-  DCHECK(IsEnabled(NEON));
-  int vd, d;
-  dst.split_code(&vd, &d);
-  int vm, m;
-  src.split_code(&vm, &m);
-  emit(0xFU * B28 | 7 * B23 | d * B22 | 0x32 * B16 | vd * B12 | B6 | m * B5 |
-       vm);
-}
-
-void Assembler::vdup(NeonSize size, const QwNeonRegister dst,
-                     const Register src) {
+void Assembler::vdup(NeonSize size, QwNeonRegister dst, Register src) {
   DCHECK(IsEnabled(NEON));
   // Instruction details available in ARM DDI 0406C.b, A8-886.
   int B = 0, E = 0;
   switch (size) {
     case Neon8:
       B = 1;
       break;
     case Neon16:
       E = 1;
       break;
     case Neon32:
       break;
     default:
       UNREACHABLE();
       break;
   }
   int vd, d;
   dst.split_code(&vd, &d);
 
   emit(al | 0x1D * B23 | B * B22 | B21 | vd * B16 | src.code() * B12 |
        0xB * B8 | d * B7 | E * B5 | B4);
 }
 
-void Assembler::vdup(const QwNeonRegister dst, const SwVfpRegister src) {
+void Assembler::vdup(QwNeonRegister dst, SwVfpRegister src) {
   DCHECK(IsEnabled(NEON));
   // Instruction details available in ARM DDI 0406C.b, A8-884.
   int index = src.code() & 1;
   int d_reg = src.code() / 2;
   int imm4 = 4 | index << 3;  // esize = 32, index in bit 3.
   int vd, d;
   dst.split_code(&vd, &d);
   int vm, m;
   DwVfpRegister::from_code(d_reg).split_code(&vm, &m);
 
   emit(0x1E7U * B23 | d * B22 | 0x3 * B20 | imm4 * B16 | vd * B12 | 0x18 * B7 |
        B6 | m * B5 | vm);
 }
 
 // Encode NEON vcvt.src_type.dst_type instruction.
-static Instr EncodeNeonVCVT(const VFPType dst_type, const QwNeonRegister dst,
-                            const VFPType src_type, const QwNeonRegister src) {
+static Instr EncodeNeonVCVT(VFPType dst_type, QwNeonRegister dst,
+                            VFPType src_type, QwNeonRegister src) {
   DCHECK(src_type != dst_type);
   DCHECK(src_type == F32 || dst_type == F32);
   // Instruction details available in ARM DDI 0406C.b, A8.8.868.
   int vd, d;
   dst.split_code(&vd, &d);
   int vm, m;
   src.split_code(&vm, &m);
 
   int op = 0;
   if (src_type == F32) {
     DCHECK(dst_type == S32 || dst_type == U32);
     op = dst_type == U32 ? 3 : 2;
   } else {
     DCHECK(src_type == S32 || src_type == U32);
     op = src_type == U32 ? 1 : 0;
   }
 
   return 0x1E7U * B23 | d * B22 | 0x3B * B16 | vd * B12 | 0x3 * B9 | op * B7 |
          B6 | m * B5 | vm;
 }
 
-void Assembler::vcvt_f32_s32(const QwNeonRegister dst,
-                             const QwNeonRegister src) {
+void Assembler::vcvt_f32_s32(QwNeonRegister dst, QwNeonRegister src) {
   DCHECK(IsEnabled(NEON));
   DCHECK(VfpRegisterIsAvailable(dst));
   DCHECK(VfpRegisterIsAvailable(src));
   emit(EncodeNeonVCVT(F32, dst, S32, src));
 }
 
-void Assembler::vcvt_f32_u32(const QwNeonRegister dst,
-                             const QwNeonRegister src) {
+void Assembler::vcvt_f32_u32(QwNeonRegister dst, QwNeonRegister src) {
   DCHECK(IsEnabled(NEON));
   DCHECK(VfpRegisterIsAvailable(dst));
   DCHECK(VfpRegisterIsAvailable(src));
   emit(EncodeNeonVCVT(F32, dst, U32, src));
 }
 
-void Assembler::vcvt_s32_f32(const QwNeonRegister dst,
-                             const QwNeonRegister src) {
+void Assembler::vcvt_s32_f32(QwNeonRegister dst, QwNeonRegister src) {
   DCHECK(IsEnabled(NEON));
   DCHECK(VfpRegisterIsAvailable(dst));
   DCHECK(VfpRegisterIsAvailable(src));
   emit(EncodeNeonVCVT(S32, dst, F32, src));
 }
 
-void Assembler::vcvt_u32_f32(const QwNeonRegister dst,
-                             const QwNeonRegister src) {
+void Assembler::vcvt_u32_f32(QwNeonRegister dst, QwNeonRegister src) {
   DCHECK(IsEnabled(NEON));
   DCHECK(VfpRegisterIsAvailable(dst));
   DCHECK(VfpRegisterIsAvailable(src));
   emit(EncodeNeonVCVT(U32, dst, F32, src));
 }
 
-enum UnaryOp { VABS, VABSF, VNEG, VNEGF };
+enum NeonRegType { NEON_D, NEON_Q };
 
-static Instr EncodeNeonUnaryOp(UnaryOp op, NeonSize size, QwNeonRegister dst,
-                               QwNeonRegister src) {
+enum UnaryOp { VMVN, VSWP, VABS, VABSF, VNEG, VNEGF };
+
+static Instr EncodeNeonUnaryOp(UnaryOp op, NeonRegType reg_type, NeonSize size,
+                               int dst_code, int src_code) {
   int op_encoding = 0;
   switch (op) {
+    case VMVN:
+      DCHECK_EQ(Neon8, size);  // size == 0 for vmvn
+      op_encoding = B10 | 0x3 * B7;
+      break;
+    case VSWP:
+      DCHECK_EQ(Neon8, size);  // size == 0 for vswp
+      op_encoding = B17;
+      break;
     case VABS:
-      op_encoding = 0x6 * B7;
+      op_encoding = B16 | 0x6 * B7;
       break;
     case VABSF:
       DCHECK_EQ(Neon32, size);
-      op_encoding = 0x6 * B7 | B10;
+      op_encoding = B16 | B10 | 0x6 * B7;
       break;
     case VNEG:
-      op_encoding = 0x7 * B7;
+      op_encoding = B16 | 0x7 * B7;
       break;
     case VNEGF:
       DCHECK_EQ(Neon32, size);
-      op_encoding = 0x7 * B7 | B10;
+      op_encoding = B16 | B10 | 0x7 * B7;
       break;
     default:
       UNREACHABLE();
       break;
   }
-  int vd, d;
-  dst.split_code(&vd, &d);
-  int vm, m;
-  src.split_code(&vm, &m);
-  return 0x1E7U * B23 | d * B22 | 0x3 * B20 | size * B18 | B16 | vd * B12 | B6 |
-         m * B5 | vm | op_encoding;
+  if (reg_type == NEON_Q) {
+    op_encoding |= B6;
+    dst_code <<= 1;
+    src_code <<= 1;
+  }
+
+  int vd = dst_code & 0x0F;

[martyn.capewell, 2017/03/09 14:03:36]

Perhaps there should be some static methods to split register codes based on
whether the value represents a S, D or Q register, then use them here, below,
and also refactor QwNeonRegister::split_code() etc to use them.

[bbudge, 2017/03/09 20:51:05]

Good idea. Done.

+  int d = (dst_code & 0x10) >> 4;
+  int vm = src_code & 0x0F;
+  int m = (src_code & 0x10) >> 4;
+  return 0x1E7U * B23 | d * B22 | 0x3 * B20 | size * B18 | vd * B12 | m * B5 |
+         vm | op_encoding;
 }
 
-void Assembler::vabs(const QwNeonRegister dst, const QwNeonRegister src) {
+void Assembler::vmvn(QwNeonRegister dst, QwNeonRegister src) {
+  // Qd = vmvn(Qn, Qm) SIMD bitwise negate.
+  // Instruction details available in ARM DDI 0406C.b, A8-966.
+  DCHECK(IsEnabled(NEON));
+  emit(EncodeNeonUnaryOp(VMVN, NEON_Q, Neon8, dst.code(), src.code()));
+}
+
+void Assembler::vswp(DwVfpRegister dst, DwVfpRegister src) {
+  DCHECK(IsEnabled(NEON));
+  // Dd = vswp(Dn, Dm) SIMD d-register swap.
+  // Instruction details available in ARM DDI 0406C.b, A8.8.418.
+  DCHECK(IsEnabled(NEON));
+  emit(EncodeNeonUnaryOp(VSWP, NEON_D, Neon8, dst.code(), src.code()));
+}
+
+void Assembler::vswp(QwNeonRegister dst, QwNeonRegister src) {
+  // Qd = vswp(Qn, Qm) SIMD q-register swap.
+  // Instruction details available in ARM DDI 0406C.b, A8.8.418.
+  DCHECK(IsEnabled(NEON));
+  emit(EncodeNeonUnaryOp(VSWP, NEON_Q, Neon8, dst.code(), src.code()));
+}
+
+void Assembler::vabs(QwNeonRegister dst, QwNeonRegister src) {
   // Qd = vabs.f<size>(Qn, Qm) SIMD floating point absolute value.
   // Instruction details available in ARM DDI 0406C.b, A8.8.824.
   DCHECK(IsEnabled(NEON));
-  emit(EncodeNeonUnaryOp(VABSF, Neon32, dst, src));
+  emit(EncodeNeonUnaryOp(VABSF, NEON_Q, Neon32, dst.code(), src.code()));
 }
 
-void Assembler::vabs(NeonSize size, const QwNeonRegister dst,
-                     const QwNeonRegister src) {
+void Assembler::vabs(NeonSize size, QwNeonRegister dst, QwNeonRegister src) {
   // Qd = vabs.s<size>(Qn, Qm) SIMD integer absolute value.
   // Instruction details available in ARM DDI 0406C.b, A8.8.824.
   DCHECK(IsEnabled(NEON));
-  emit(EncodeNeonUnaryOp(VABS, size, dst, src));
+  emit(EncodeNeonUnaryOp(VABS, NEON_Q, size, dst.code(), src.code()));
 }
 
-void Assembler::vneg(const QwNeonRegister dst, const QwNeonRegister src) {
+void Assembler::vneg(QwNeonRegister dst, QwNeonRegister src) {
   // Qd = vabs.f<size>(Qn, Qm) SIMD floating point negate.
   // Instruction details available in ARM DDI 0406C.b, A8.8.968.
   DCHECK(IsEnabled(NEON));
-  emit(EncodeNeonUnaryOp(VNEGF, Neon32, dst, src));
+  emit(EncodeNeonUnaryOp(VNEGF, NEON_Q, Neon32, dst.code(), src.code()));
 }
 
-void Assembler::vneg(NeonSize size, const QwNeonRegister dst,
-                     const QwNeonRegister src) {
+void Assembler::vneg(NeonSize size, QwNeonRegister dst, QwNeonRegister src) {
   // Qd = vabs.s<size>(Qn, Qm) SIMD integer negate.
   // Instruction details available in ARM DDI 0406C.b, A8.8.968.
   DCHECK(IsEnabled(NEON));
-  emit(EncodeNeonUnaryOp(VNEG, size, dst, src));
-}
-
-void Assembler::veor(DwVfpRegister dst, DwVfpRegister src1,
-                     DwVfpRegister src2) {
-  // Dd = veor(Dn, Dm) 64 bit integer exclusive OR.
-  // Instruction details available in ARM DDI 0406C.b, A8.8.888.
-  DCHECK(IsEnabled(NEON));
-  int vd, d;
-  dst.split_code(&vd, &d);
-  int vn, n;
-  src1.split_code(&vn, &n);
-  int vm, m;
-  src2.split_code(&vm, &m);
-  emit(0x1E6U * B23 | d * B22 | vn * B16 | vd * B12 | B8 | n * B7 | m * B5 |
-       B4 | vm);
+  emit(EncodeNeonUnaryOp(VNEG, NEON_Q, size, dst.code(), src.code()));
 }
 
 enum BinaryBitwiseOp { VAND, VBIC, VBIF, VBIT, VBSL, VEOR, VORR, VORN };
 
-static Instr EncodeNeonBinaryBitwiseOp(BinaryBitwiseOp op, QwNeonRegister dst,
-                                       QwNeonRegister src1,
-                                       QwNeonRegister src2) {
+static Instr EncodeNeonBinaryBitwiseOp(BinaryBitwiseOp op, NeonRegType reg_type,
+                                       int dst_code, int src_code1,
+                                       int src_code2) {
   int op_encoding = 0;
   switch (op) {
     case VBIC:
       op_encoding = 0x1 * B20;
       break;
     case VBIF:
       op_encoding = B24 | 0x3 * B20;
       break;
     case VBIT:
       op_encoding = B24 | 0x2 * B20;
@@ skipping 10 matching lines @@
     case VORN:
       op_encoding = 0x3 * B20;
       break;
     case VAND:
       // op_encoding is 0.
       break;
     default:
       UNREACHABLE();
       break;
   }
-  int vd, d;
-  dst.split_code(&vd, &d);
-  int vn, n;
-  src1.split_code(&vn, &n);
-  int vm, m;
-  src2.split_code(&vm, &m);
+  if (reg_type == NEON_Q) {
+    op_encoding |= B6;
+    dst_code <<= 1;
+    src_code1 <<= 1;
+    src_code2 <<= 1;
+  }
+
+  int vd = dst_code & 0x0F;
+  int d = (dst_code & 0x10) >> 4;
+  int vn = src_code1 & 0x0F;
+  int n = (src_code1 & 0x10) >> 4;
+  int vm = src_code2 & 0x0F;
+  int m = (src_code2 & 0x10) >> 4;
   return 0x1E4U * B23 | op_encoding | d * B22 | vn * B16 | vd * B12 | B8 |
-         n * B7 | B6 | m * B5 | B4 | vm;
+         n * B7 | m * B5 | B4 | vm;
 }
 
 void Assembler::vand(QwNeonRegister dst, QwNeonRegister src1,
                      QwNeonRegister src2) {
   // Qd = vand(Qn, Qm) SIMD AND.
   // Instruction details available in ARM DDI 0406C.b, A8.8.836.
   DCHECK(IsEnabled(NEON));
-  emit(EncodeNeonBinaryBitwiseOp(VAND, dst, src1, src2));
+  emit(EncodeNeonBinaryBitwiseOp(VAND, NEON_Q, dst.code(), src1.code(),
+                                 src2.code()));
 }
 
-void Assembler::vbsl(QwNeonRegister dst, const QwNeonRegister src1,
-                     const QwNeonRegister src2) {
-  DCHECK(IsEnabled(NEON));
+void Assembler::vbsl(QwNeonRegister dst, QwNeonRegister src1,
+                     QwNeonRegister src2) {
   // Qd = vbsl(Qn, Qm) SIMD bitwise select.
   // Instruction details available in ARM DDI 0406C.b, A8-844.
-  emit(EncodeNeonBinaryBitwiseOp(VBSL, dst, src1, src2));
+  DCHECK(IsEnabled(NEON));
+  emit(EncodeNeonBinaryBitwiseOp(VBSL, NEON_Q, dst.code(), src1.code(),
+                                 src2.code()));
+}
+
+void Assembler::veor(DwVfpRegister dst, DwVfpRegister src1,
+                     DwVfpRegister src2) {
+  // Dd = veor(Dn, Dm) SIMD exclusive OR.
+  // Instruction details available in ARM DDI 0406C.b, A8.8.888.
+  DCHECK(IsEnabled(NEON));
+  emit(EncodeNeonBinaryBitwiseOp(VEOR, NEON_D, dst.code(), src1.code(),
+                                 src2.code()));
 }
 
 void Assembler::veor(QwNeonRegister dst, QwNeonRegister src1,
                      QwNeonRegister src2) {
   // Qd = veor(Qn, Qm) SIMD exclusive OR.
   // Instruction details available in ARM DDI 0406C.b, A8.8.888.
   DCHECK(IsEnabled(NEON));
-  emit(EncodeNeonBinaryBitwiseOp(VEOR, dst, src1, src2));
+  emit(EncodeNeonBinaryBitwiseOp(VEOR, NEON_Q, dst.code(), src1.code(),
+                                 src2.code()));
 }
 
 void Assembler::vorr(QwNeonRegister dst, QwNeonRegister src1,
                      QwNeonRegister src2) {
   // Qd = vorr(Qn, Qm) SIMD OR.
   // Instruction details available in ARM DDI 0406C.b, A8.8.976.
   DCHECK(IsEnabled(NEON));
-  emit(EncodeNeonBinaryBitwiseOp(VORR, dst, src1, src2));
+  emit(EncodeNeonBinaryBitwiseOp(VORR, NEON_Q, dst.code(), src1.code(),
+                                 src2.code()));
 }
 
 enum FPBinOp {
   VADDF,
   VSUBF,
   VMULF,
   VMINF,
   VMAXF,
   VRECPS,
   VRSQRTS,
@@ skipping 174 matching lines @@
 }
 
 void Assembler::vmul(QwNeonRegister dst, QwNeonRegister src1,
                      QwNeonRegister src2) {
   DCHECK(IsEnabled(NEON));
   // Qd = vadd(Qn, Qm) SIMD floating point multiply.
   // Instruction details available in ARM DDI 0406C.b, A8-958.
   emit(EncodeNeonBinOp(VMULF, dst, src1, src2));
 }
 
-void Assembler::vmul(NeonSize size, QwNeonRegister dst,
-                     const QwNeonRegister src1, const QwNeonRegister src2) {
+void Assembler::vmul(NeonSize size, QwNeonRegister dst, QwNeonRegister src1,
+                     QwNeonRegister src2) {
   DCHECK(IsEnabled(NEON));
   // Qd = vadd(Qn, Qm) SIMD integer multiply.
   // Instruction details available in ARM DDI 0406C.b, A8-960.
   emit(EncodeNeonBinOp(VMUL, size, dst, src1, src2));
 }
 
-void Assembler::vmin(const QwNeonRegister dst, const QwNeonRegister src1,
-                     const QwNeonRegister src2) {
+void Assembler::vmin(QwNeonRegister dst, QwNeonRegister src1,
+                     QwNeonRegister src2) {
   DCHECK(IsEnabled(NEON));
   // Qd = vmin(Qn, Qm) SIMD floating point MIN.
   // Instruction details available in ARM DDI 0406C.b, A8-928.
   emit(EncodeNeonBinOp(VMINF, dst, src1, src2));
 }
 
 void Assembler::vmin(NeonDataType dt, QwNeonRegister dst, QwNeonRegister src1,
                      QwNeonRegister src2) {
   DCHECK(IsEnabled(NEON));
   // Qd = vmin(Qn, Qm) SIMD integer MIN.
@@ skipping 93 matching lines @@
 }
 
 void Assembler::vrsqrts(QwNeonRegister dst, QwNeonRegister src1,
                         QwNeonRegister src2) {
   DCHECK(IsEnabled(NEON));
   // Qd = vrsqrts(Qn, Qm) SIMD reciprocal square root refinement step.
   // Instruction details available in ARM DDI 0406C.b, A8-1040.
   emit(EncodeNeonBinOp(VRSQRTS, dst, src1, src2));
 }
 
-enum PairwiseOp { VPMIN, VPMAX };
+enum NeonPairwiseOp { VPMIN, VPMAX };
 
-static Instr EncodeNeonPairwiseOp(PairwiseOp op, NeonDataType dt,
+static Instr EncodeNeonPairwiseOp(NeonPairwiseOp op, NeonDataType dt,
                                   DwVfpRegister dst, DwVfpRegister src1,
                                   DwVfpRegister src2) {
   int op_encoding = 0;
   switch (op) {
     case VPMIN:
       op_encoding = 0xA * B8 | B4;
       break;
     case VPMAX:
       op_encoding = 0xA * B8;
       break;
@@ skipping 78 matching lines @@
 }
 
 void Assembler::vcgt(NeonDataType dt, QwNeonRegister dst, QwNeonRegister src1,
                      QwNeonRegister src2) {
   DCHECK(IsEnabled(NEON));
   // Qd = vcgt(Qn, Qm) SIMD integer compare greater than.
   // Instruction details available in ARM DDI 0406C.b, A8-852.
   emit(EncodeNeonBinOp(VCGT, dt, dst, src1, src2));
 }
 
-void Assembler::vext(QwNeonRegister dst, const QwNeonRegister src1,
-                     const QwNeonRegister src2, int bytes) {
+void Assembler::vext(QwNeonRegister dst, QwNeonRegister src1,
+                     QwNeonRegister src2, int bytes) {
   DCHECK(IsEnabled(NEON));
   // Qd = vext(Qn, Qm) SIMD byte extract.
   // Instruction details available in ARM DDI 0406C.b, A8-890.
   int vd, d;
   dst.split_code(&vd, &d);
   int vn, n;
   src1.split_code(&vn, &n);
   int vm, m;
   src2.split_code(&vm, &m);
   DCHECK_GT(16, bytes);
   emit(0x1E5U * B23 | d * B22 | 0x3 * B20 | vn * B16 | vd * B12 | bytes * B8 |
        n * B7 | B6 | m * B5 | vm);
 }
 
-void Assembler::vzip(NeonSize size, QwNeonRegister dst,
-                     const QwNeonRegister src) {
-  DCHECK(IsEnabled(NEON));
-  // Qd = vzip.<size>(Qn, Qm) SIMD zip (interleave).
-  // Instruction details available in ARM DDI 0406C.b, A8-1102.
+enum NeonSizedOp { VZIP, VUZP, VREV16, VREV32, VREV64, VTRN };
+
+static Instr EncodeNeonSizedOp(NeonSizedOp op, NeonSize size,
+                               QwNeonRegister dst, QwNeonRegister src) {
+  int op_encoding = 0;
+  switch (op) {
+    case VZIP:
+      op_encoding = 0x2 * B16 | B8 | B7;

[martyn.capewell, 2017/03/09 14:03:36]

Some of these seem inconsistently expressed. For example B8 | B7 == 0x3 * B7 and
B8 == 0x2 * B7.

[bbudge, 2017/03/09 20:51:05]

Done.

+      break;
+    case VUZP:
+      op_encoding = 0x2 * B16 | B8;
+      break;
+    case VREV16:
+      op_encoding = 0x2 * B7;
+      break;
+    case VREV32:
+      op_encoding = 0x1 * B7;
+      break;
+    case VREV64:
+      // op_encoding is 0;
+      break;
+    case VTRN:
+      op_encoding = 0x2 * B16 | B7;
+      break;
+    default:
+      UNREACHABLE();
+      break;
+  }
   int vd, d;
   dst.split_code(&vd, &d);
   int vm, m;
   src.split_code(&vm, &m);
   int sz = static_cast<int>(size);
-  emit(0x1E7U * B23 | d * B22 | 0x3 * B20 | sz * B18 | 2 * B16 | vd * B12 |
-       0x3 * B7 | B6 | m * B5 | vm);
+  return 0x1E7U * B23 | d * B22 | 0x3 * B20 | sz * B18 | vd * B12 | B6 |
+         m * B5 | vm | op_encoding;
 }
 
-static Instr EncodeNeonVREV(NeonSize op_size, NeonSize size,
-                            const QwNeonRegister dst,
-                            const QwNeonRegister src) {
+void Assembler::vzip(NeonSize size, QwNeonRegister src1, QwNeonRegister src2) {
+  DCHECK(IsEnabled(NEON));
+  // Qd = vzip.<size>(Qn, Qm) SIMD zip (interleave).
+  // Instruction details available in ARM DDI 0406C.b, A8-1102.
+  emit(EncodeNeonSizedOp(VZIP, size, src1, src2));
+}
+
+void Assembler::vuzp(NeonSize size, QwNeonRegister src1, QwNeonRegister src2) {
+  DCHECK(IsEnabled(NEON));
+  // Qd = vuzp.<size>(Qn, Qm) SIMD un-zip (de-interleave).
+  // Instruction details available in ARM DDI 0406C.b, A8-1100.
+  emit(EncodeNeonSizedOp(VUZP, size, src1, src2));
+}
+
+void Assembler::vrev16(NeonSize size, QwNeonRegister dst, QwNeonRegister src) {
+  DCHECK(IsEnabled(NEON));
   // Qd = vrev<op_size>.<size>(Qn, Qm) SIMD scalar reverse.
   // Instruction details available in ARM DDI 0406C.b, A8-1028.
-  DCHECK_GT(op_size, static_cast<int>(size));
-  int vd, d;
-  dst.split_code(&vd, &d);
-  int vm, m;
-  src.split_code(&vm, &m);
-  int sz = static_cast<int>(size);
-  int op = static_cast<int>(Neon64) - static_cast<int>(op_size);
-  return 0x1E7U * B23 | d * B22 | 0x3 * B20 | sz * B18 | vd * B12 | op * B7 |
-         B6 | m * B5 | vm;
+  emit(EncodeNeonSizedOp(VREV16, size, dst, src));
 }
 
-void Assembler::vrev16(NeonSize size, const QwNeonRegister dst,
-                       const QwNeonRegister src) {
+void Assembler::vrev32(NeonSize size, QwNeonRegister dst, QwNeonRegister src) {
   DCHECK(IsEnabled(NEON));
-  emit(EncodeNeonVREV(Neon16, size, dst, src));
+  // Qd = vrev<op_size>.<size>(Qn, Qm) SIMD scalar reverse.
+  // Instruction details available in ARM DDI 0406C.b, A8-1028.
+  emit(EncodeNeonSizedOp(VREV32, size, dst, src));
 }
 
-void Assembler::vrev32(NeonSize size, const QwNeonRegister dst,
-                       const QwNeonRegister src) {
+void Assembler::vrev64(NeonSize size, QwNeonRegister dst, QwNeonRegister src) {
   DCHECK(IsEnabled(NEON));
-  emit(EncodeNeonVREV(Neon32, size, dst, src));
+  // Qd = vrev<op_size>.<size>(Qn, Qm) SIMD scalar reverse.
+  // Instruction details available in ARM DDI 0406C.b, A8-1028.
+  emit(EncodeNeonSizedOp(VREV64, size, dst, src));
 }
 
-void Assembler::vrev64(NeonSize size, const QwNeonRegister dst,
-                       const QwNeonRegister src) {
+void Assembler::vtrn(NeonSize size, QwNeonRegister src1, QwNeonRegister src2) {
   DCHECK(IsEnabled(NEON));
-  emit(EncodeNeonVREV(Neon64, size, dst, src));
+  // Qd = vrev<op_size>.<size>(Qn, Qm) SIMD scalar reverse.
+  // Instruction details available in ARM DDI 0406C.b, A8-1096.
+  emit(EncodeNeonSizedOp(VTRN, size, src1, src2));
 }
 
 // Encode NEON vtbl / vtbx instruction.
-static Instr EncodeNeonVTB(const DwVfpRegister dst, const NeonListOperand& list,
-                           const DwVfpRegister index, bool vtbx) {
+static Instr EncodeNeonVTB(DwVfpRegister dst, const NeonListOperand& list,
+                           DwVfpRegister index, bool vtbx) {
   // Dd = vtbl(table, Dm) SIMD vector permute, zero at out of range indices.
   // Instruction details available in ARM DDI 0406C.b, A8-1094.
   // Dd = vtbx(table, Dm) SIMD vector permute, skip out of range indices.
   // Instruction details available in ARM DDI 0406C.b, A8-1094.
   int vd, d;
   dst.split_code(&vd, &d);
   int vn, n;
   list.base().split_code(&vn, &n);
   int vm, m;
   index.split_code(&vm, &m);
   int op = vtbx ? 1 : 0;  // vtbl = 0, vtbx = 1.
   return 0x1E7U * B23 | d * B22 | 0x3 * B20 | vn * B16 | vd * B12 | 0x2 * B10 |
          list.length() * B8 | n * B7 | op * B6 | m * B5 | vm;
 }
 
-void Assembler::vtbl(const DwVfpRegister dst, const NeonListOperand& list,
-                     const DwVfpRegister index) {
+void Assembler::vtbl(DwVfpRegister dst, const NeonListOperand& list,
+                     DwVfpRegister index) {
   DCHECK(IsEnabled(NEON));
   emit(EncodeNeonVTB(dst, list, index, false));
 }
 
-void Assembler::vtbx(const DwVfpRegister dst, const NeonListOperand& list,
-                     const DwVfpRegister index) {
+void Assembler::vtbx(DwVfpRegister dst, const NeonListOperand& list,
+                     DwVfpRegister index) {
   DCHECK(IsEnabled(NEON));
   emit(EncodeNeonVTB(dst, list, index, true));
 }
 
 // Pseudo instructions.
 void Assembler::nop(int type) {
   // ARMv6{K/T2} and v7 have an actual NOP instruction but it serializes
   // some of the CPU's pipeline and has to issue. Older ARM chips simply used
   // MOV Rx, Rx as NOP and it performs better even in newer CPUs.
   // We therefore use MOV Rx, Rx, even on newer CPUs, and use Rx to encode
@@ skipping 538 matching lines @@
     DCHECK(is_uint12(offset));
     instr_at_put(pc, SetLdrRegisterImmediateOffset(instr, offset));
   }
 }
 
 
 }  // namespace internal
 }  // namespace v8
 
 #endif  // V8_TARGET_ARCH_ARM