Fix vector load/stores in the ARM assembler.

src/IceAssemblerARM32.cpp

 //===- subzero/src/IceAssemblerARM32.cpp - Assembler for ARM32 --*- C++ -*-===//
 //
 // Copyright (c) 2013, the Dart project authors.  Please see the AUTHORS file
 // for details. All rights reserved. Use of this source code is governed by a
 // BSD-style license that can be found in the LICENSE file.
 //
 // Modified by the Subzero authors.
 //
 //===----------------------------------------------------------------------===//
 //
@@ skipping 2526 matching lines @@
       encodeAddress(OpAddress, Address, TInfo, RotatedImm8Div4Address);
   (void)AddressEncoding;
   assert(AddressEncoding == EncodedAsImmRegOffset);
   IValueT Encoding = B27 | B26 | B24 | B20 | B11 | B9 |
                      (encodeCondition(Cond) << kConditionShift) |
                      (getYInRegXXXXY(Sd) << 22) |
                      (getXXXXInRegXXXXY(Sd) << 12) | Address;
   emitInst(Encoding);
 }
 
+void AssemblerARM32::emitVMem1Op(IValueT Opcode, IValueT Dd, IValueT Rn,
+                                 IValueT Rm, DRegListSize NumDRegs,
+                                 size_t ElmtSize, IValueT Align,
+                                 const char *InstName) {
+  assert(Utils::IsAbsoluteUint(2, Align));
+  IValueT EncodedElmtSize;
+  switch (ElmtSize) {
+  default: {
+    std::string Buffer;
+    llvm::raw_string_ostream StrBuf(Buffer);
+    StrBuf << InstName << ": found invalid vector element size " << ElmtSize;
+    llvm::report_fatal_error(StrBuf.str());
+  }
+  case 8:
+    EncodedElmtSize = 0;
+    break;
+  case 16:
+    EncodedElmtSize = 1;
+    break;
+  case 32:
+    EncodedElmtSize = 2;
+    break;
+  case 64:
+    EncodedElmtSize = 3;
+  }
+  const IValueT Encoding =
+      Opcode | (encodeCondition(CondARM32::kNone) << kConditionShift) |
+      (getYInRegYXXXX(Dd) << 22) | (Rn << kRnShift) |
+      (getXXXXInRegYXXXX(Dd) << kRdShift) | (NumDRegs << 8) |
+      (EncodedElmtSize << 6) | (Align << 4) | Rm;
+  emitInst(Encoding);
+}
+
+void AssemblerARM32::vld1qr(size_t ElmtSize, const Operand *OpQd,
+                            const Operand *OpAddress, const TargetInfo &TInfo) {
+  // VLD1 (multiple single elements) - ARM section A8.8.320, encoding A1:
+  //   vld1.<size> <Qd>, [<Rn>]
+  //
+  // 111101000D10nnnnddd0ttttssaammmm where tttt=DRegListSize2, Dddd=Qd,
+  // nnnn=Rn, aa=0 (use default alignment), size=ElmtSize, and ss is the
+  // encoding of ElmtSize.
+  constexpr const char *Vld1qr = "vld1qr";
+  const IValueT Qd = encodeQRegister(OpQd, "Qd", Vld1qr);
+  const IValueT Dd = mapQRegToDReg(Qd);
+  IValueT Address;
+  if (encodeAddress(OpAddress, Address, TInfo, NoImmOffsetAddress) !=
+      EncodedAsImmRegOffset)
+    llvm::report_fatal_error(std::string(Vld1qr) + ": malform memory address");
+  const IValueT Rn = mask(Address, kRnShift, 4);
+  constexpr IValueT Rm = RegARM32::Reg_pc;
+  constexpr IValueT Opcode = B26 | B21;
+  constexpr IValueT Align = 0; // use default alignment.
+  emitVMem1Op(Opcode, Dd, Rn, Rm, DRegListSize2, ElmtSize, Align, Vld1qr);
+}
+
 void AssemblerARM32::vmovd(const Operand *OpDd,
                            const OperandARM32FlexFpImm *OpFpImm,
                            CondARM32::Cond Cond) {
   // VMOV (immediate) - ARM section A8.8.339, encoding A2:
   //   vmov<c>.f64 <Dd>, #<imm>
   //
   // cccc11101D11xxxxdddd10110000yyyy where cccc=Cond, ddddD=Sn, xxxxyyyy=imm.
   constexpr const char *Vmovd = "vmovd";
   IValueT Dd = encodeSRegister(OpDd, "Dd", Vmovd);
   IValueT Imm8 = OpFpImm->getModifiedImm();
@@ skipping 294 matching lines @@
   IValueT AddressEncoding =
       encodeAddress(OpAddress, Address, TInfo, RotatedImm8Div4Address);
   (void)AddressEncoding;
   assert(AddressEncoding == EncodedAsImmRegOffset);
   IValueT Encoding =
       B27 | B26 | B24 | B11 | B9 | (encodeCondition(Cond) << kConditionShift) |
       (getYInRegXXXXY(Sd) << 22) | (getXXXXInRegXXXXY(Sd) << 12) | Address;
   emitInst(Encoding);
 }
 
+void AssemblerARM32::vst1qr(size_t ElmtSize, const Operand *OpQd,
+                            const Operand *OpAddress, const TargetInfo &TInfo) {
+  // VST1 (multiple single elements) - ARM section A8.8.404, encoding A1:
+  //   vst1.<size> <Qd>, [<Rn>]
+  //
+  // 111101000D00nnnnddd0ttttssaammmm where tttt=DRegListSize2, Dddd=Qd,
+  // nnnn=Rn, aa=0 (use default alignment), size=ElmtSize, and ss is the
+  // encoding of ElmtSize.
+  constexpr const char *Vst1qr = "vst1qr";
+  const IValueT Qd = encodeQRegister(OpQd, "Qd", Vst1qr);
+  const IValueT Dd = mapQRegToDReg(Qd);
+  IValueT Address;
+  if (encodeAddress(OpAddress, Address, TInfo, NoImmOffsetAddress) !=
+      EncodedAsImmRegOffset)
+    llvm::report_fatal_error(std::string(Vst1qr) + ": malform memory address");
+  const IValueT Rn = mask(Address, kRnShift, 4);
+  constexpr IValueT Rm = RegARM32::Reg_pc;
+  constexpr IValueT Opcode = B26;
+  constexpr IValueT Align = 0; // use default alignment.
+  emitVMem1Op(Opcode, Dd, Rn, Rm, DRegListSize2, ElmtSize, Align, Vst1qr);
+}
+
 void AssemblerARM32::vsubs(const Operand *OpSd, const Operand *OpSn,
                            const Operand *OpSm, CondARM32::Cond Cond) {
   // VSUB (floating-point) - ARM section A8.8.415, encoding A2:
   //   vsub<c>.f32 <Sd>, <Sn>, <Sm>
   //
   // cccc11100D11nnnndddd101sN1M0mmmm where cccc=Cond, s=0, ddddD=Rd, nnnnN=Rn,
   // and mmmmM=Rm.
   constexpr const char *Vsubs = "vsubs";
   constexpr IValueT VsubsOpcode = B21 | B20 | B6;
   emitVFPsss(Cond, VsubsOpcode, OpSd, OpSn, OpSm, Vsubs);
@@ skipping 103 matching lines @@
   constexpr const char *Vsqrts = "vsqrts";
   IValueT Sd = encodeSRegister(OpSd, "Sd", Vsqrts);
   IValueT Sm = encodeSRegister(OpSm, "Sm", Vsqrts);
   constexpr IValueT VsqrtsOpcode = B23 | B21 | B20 | B16 | B7 | B6;
   constexpr IValueT S0 = 0;
   emitVFPsss(Cond, VsqrtsOpcode, Sd, S0, Sm);
 }
 
 } // end of namespace ARM32
 } // end of namespace Ice