Lower insertelement and extractelement.

src/IceTargetLoweringX8632.cpp

 //===- subzero/src/IceTargetLoweringX8632.cpp - x86-32 lowering -----------===//
 //
 //                        The Subzero Code Generator
 //
 // This file is distributed under the University of Illinois Open Source
 // License. See LICENSE.TXT for details.
 //
 //===----------------------------------------------------------------------===//
 //
 // This file implements the TargetLoweringX8632 class, which
@@ skipping 67 matching lines @@
 #undef X
   };
 const size_t TableIcmp64Size = llvm::array_lengthof(TableIcmp64);
 
 InstX8632::BrCond getIcmp32Mapping(InstIcmp::ICond Cond) {
   size_t Index = static_cast<size_t>(Cond);
   assert(Index < TableIcmp32Size);
   return TableIcmp32[Index].Mapping;
 }
 
+// Return the type which the elements of the vector have in the X86
+// representation.
+Type getLoweredElementType(Type Ty) {
+  assert(isVectorType(Ty));
+  switch (Ty) {
+  default:

[Jim Stichnoth, 2014/07/17 19:49:01]

Can this be encoded in ICETYPEX8632_TABLE?

[wala, 2014/07/17 22:14:12]

Done.

I've changed the name of this function to getInVectorElementType in the latest
commit, because I think that's a clearer name.

+    return typeElementType(Ty);
+  case IceType_v4i1:
+    return IceType_i32;
+  case IceType_v8i1:
+    return IceType_i16;
+  case IceType_v16i1:
+    return IceType_i8;
+  }
+}
+
 // The maximum number of arguments to pass in XMM registers
 const unsigned X86_MAX_XMM_ARGS = 4;
 // The number of bits in a byte
 const unsigned X86_CHAR_BIT = 8;
 
 // In some cases, there are x-macros tables for both high-level and
 // low-level instructions/operands that use the same enum key value.
 // The tables are kept separate to maintain a proper separation
 // between abstraction layers.  There is a risk that the tables
 // could get out of sync if enum values are reordered or if entries
@@ skipping 1840 matching lines @@
     case IceType_v4i32:
     case IceType_v4f32: {
       _movp(Dest, legalizeToVar(Src0));
     } break;
     }
     break;
   }
   }
 }
 
+void TargetX8632::lowerExtractElement(const InstExtractElement *Inst) {
+  Operand *SourceVectOperand = Inst->getSrc(0);
+  ConstantInteger *ElementIndex =
+      llvm::dyn_cast<ConstantInteger>(Inst->getSrc(1));
+  // Only constant indices are allowed in PNaCl IR.
+  assert(ElementIndex);
+
+  unsigned Index = ElementIndex->getValue();
+  Type Ty = SourceVectOperand->getType();
+  Type ElementTy = typeElementType(Ty);
+  Type LoweredElementTy = getLoweredElementType(Ty);
+  Variable *ExtractedElement = makeReg(LoweredElementTy);
+
+  // TODO(wala): Determine the best lowering sequences for each type.
+  if (Ty == IceType_v4i32 || Ty == IceType_v4f32 || Ty == IceType_v4i1) {
+    // Lower extractelement operations where the element is 32 bits
+    // wide with pshufd.
+    // TODO(wala): SSE4.1 has extractps and pextrd
+//
+// ALIGNHACK: Force vector operands to registers in instructions that
+// require aligned memory operands until support for stack alignment
+// is implemented.
+#define ALIGN_HACK(Vect) legalizeToVar(Vect)

[Jim Stichnoth, 2014/07/17 19:49:01]

To be absolutely macro-safe, I think this should be legalizeToVar((Vect)).  Here
and below.

[wala, 2014/07/17 22:14:12]

Done.

+    Operand *T = NULL;
+    if (Index) {
+      // The shuffle only needs to occur if the element to be extracted
+      // is not at the lowest index.
+      Constant *Mask = Ctx->getConstantInt(IceType_i8, Index);
+      T = makeReg(Ty);
+      _pshufd(llvm::cast<Variable>(T), ALIGN_HACK(SourceVectOperand), Mask);
+    } else {
+      // TODO(wala): If SourceVectOperand is in memory, express it as
+      // mem32 so that the call to legalizeToVar() is made unnecessary.
+      // _movd and _movss only take mem32 memory operands.
+      T = legalizeToVar(SourceVectOperand);
+    }
+
+    if (LoweredElementTy == IceType_i32) {
+      _movd(ExtractedElement, T);
+    } else { // LoweredElementTy == IceType_f32
+      _movss(ExtractedElement, T);
+    }
+#undef ALIGN_HACK
+  } else if (Ty == IceType_v8i16 || Ty == IceType_v8i1) {
+    Constant *Mask = Ctx->getConstantInt(IceType_i8, Index);
+    _pextrw(ExtractedElement, legalizeToVar(SourceVectOperand), Mask);
+  } else {
+    assert(Ty == IceType_v16i8 || Ty == IceType_v16i1);
+    // Spill the value to a stack slot and do the extraction in memory.
+    // TODO(wala): SSE4.1 has pextrb.
+    //
+    // TODO(wala): use legalize(SourceVectOperand, Legal_Mem) when
+    // support for legalizing to mem is implemented.
+    Variable *Slot = Func->makeVariable(Ty, Context.getNode());
+    Slot->setWeight(RegWeight::Zero);
+    _movp(Slot, legalizeToVar(SourceVectOperand));
+
+    // Compute the location of the element in memory.
+    unsigned Offset = Index * typeWidthInBytes(LoweredElementTy);
+    OperandX8632Mem *Loc =
+        getMemoryOperandForStackSlot(LoweredElementTy, Slot, Offset);
+    _mov(ExtractedElement, Loc);
+  }
+
+  if (ElementTy == IceType_i1) {
+    // Truncate extracted integers to i1s if necessary.
+    Variable *T = makeReg(IceType_i1);
+    InstCast *Cast =
+        InstCast::create(Func, InstCast::Trunc, T, ExtractedElement);
+    lowerCast(Cast);
+    ExtractedElement = T;
+  }
+
+  // Copy the element to the destination.
+  Variable *Dest = Inst->getDest();
+  _mov(Dest, ExtractedElement);
+}
+
 void TargetX8632::lowerFcmp(const InstFcmp *Inst) {
   Operand *Src0 = Inst->getSrc(0);
   Operand *Src1 = Inst->getSrc(1);
   Variable *Dest = Inst->getDest();
   // Lowering a = fcmp cond, b, c
   //   ucomiss b, c       /* only if C1 != Br_None */
   //                      /* but swap b,c order if SwapOperands==true */
   //   mov a, <default>
   //   j<C1> label        /* only if C1 != Br_None */
   //   j<C2> label        /* only if C2 != Br_None */
@@ skipping 111 matching lines @@
       legalize(Src0, IsSrc1ImmOrReg ? Legal_All : Legal_Reg, true);
   InstX8632Label *Label = InstX8632Label::create(Func, this);
   _cmp(Src0New, Src1);
   _mov(Dest, One);
   _br(getIcmp32Mapping(Inst->getCondition()), Label);
   Context.insert(InstFakeUse::create(Func, Dest));
   _mov(Dest, Zero);
   Context.insert(Label);
 }
 
+void TargetX8632::lowerInsertElement(const InstInsertElement *Inst) {
+  Operand *SourceVectOperand = Inst->getSrc(0);
+  Operand *ElementToInsert = Inst->getSrc(1);
+  ConstantInteger *ElementIndex =
+      llvm::dyn_cast<ConstantInteger>(Inst->getSrc(2));
+  // Only constant indices are allowed in PNaCl IR.
+  assert(ElementIndex);
+  unsigned Index = ElementIndex->getValue();
+
+  Type Ty = SourceVectOperand->getType();
+  Type ElementTy = typeElementType(Ty);
+  Type LoweredElementTy = getLoweredElementType(Ty);
+
+  if (ElementTy == IceType_i1) {
+    // Expand the element to the appropriate size for it to be inserted
+    // in the vector.
+    Variable *Expanded =
+        Func->makeVariable(LoweredElementTy, Context.getNode());
+    InstCast *Cast =
+        InstCast::create(Func, InstCast::Zext, Expanded, ElementToInsert);
+    lowerCast(Cast);
+    ElementToInsert = Expanded;
+  }
+
+  if (Ty == IceType_v4i32 || Ty == IceType_v4f32 || Ty == IceType_v4i1) {
+    // Lower insertelement with 32-bit wide elements using shufps.
+    // TODO(wala): SSE4.1 has pinsrd and insertps.
+//
+// ALIGNHACK: Force vector operands to registers in instructions that
+// require aligned memory operands until support for stack alignment
+// is implemented.
+#define ALIGN_HACK(Vect) legalizeToVar(Vect)

[jvoung (off chromium), 2014/07/17 19:36:38]

Could move ALIGN_HACK down (near the Index < 2), if it's not actually used up
here.

[wala, 2014/07/17 22:14:12]

Done.

+
+    Variable *Element = NULL;
+    if (LoweredElementTy == IceType_f32) {
+      // Element will be in an XMM register since it is floating point.
+      Element = legalizeToVar(ElementToInsert);
+    } else {
+      // Copy an integer to an XMM register.
+      Operand *T = legalize(ElementToInsert, Legal_Reg | Legal_Mem);
+      Element = makeReg(Ty);
+      _movd(Element, T);
+    }
+
+    // shufps treats the source and desination operands as vectors of
+    // four doublewords.  The destination's two high doublewords are
+    // selected from the source operand and the two low doublewords are
+    // selected from the (original value of) the destination operand.
+    // An insertelement operation can be effected with a sequence of two
+    // shufps operations with appropriate masks.  In all cases below,
+    // Element[0] is being inserted into SourceVectOperand.  Indices are
+    // ordered from left to right.
+    //
+    // insertelement into index 0 (result is stored in Element):
+    //   Element := Element[0, 0] SourceVectOperand[0, 1]
+    //   Element := Element[0, 3] SourceVectOperand[2, 3]
+    //
+    // insertelement into index 1 (result is stored in Element):
+    //   Element := Element[0, 0] SourceVectOperand[0, 0]
+    //   Element := Element[3, 0] SourceVectOperand[2, 3]
+    //
+    // insertelement into index 2 (result is stored in T):
+    //   T := SourceVectOperand
+    //   Element := Element[0, 0] T[0, 3]
+    //   T := T[0, 1] Element[0, 3]
+    //
+    // insertelement into index 3 (result is stored in T):
+    //   T := SourceVectOperand
+    //   Element := Element[0, 0] T[0, 2]
+    //   T := T[0, 1] Element[3, 0]
+    const unsigned char Mask1[4] = {64, 0, 192, 128};
+    const unsigned char Mask2[4] = {236, 227, 196, 52};
+
+    Constant *Mask1Constant = Ctx->getConstantInt(IceType_i8, Mask1[Index]);
+    Constant *Mask2Constant = Ctx->getConstantInt(IceType_i8, Mask2[Index]);
+
+    if (Index < 2) {
+      SourceVectOperand = ALIGN_HACK(SourceVectOperand);
+      _shufps(Element, SourceVectOperand, Mask1Constant);
+      _shufps(Element, SourceVectOperand, Mask2Constant);
+      _movp(Inst->getDest(), Element);
+    } else {
+      Variable *T = makeReg(Ty);
+      _movp(T, SourceVectOperand);
+      _shufps(Element, T, Mask1Constant);
+      _shufps(T, Element, Mask2Constant);
+      _movp(Inst->getDest(), T);
+    }
+#undef ALIGN_HACK
+  } else if (Ty == IceType_v8i16 || Ty == IceType_v8i1) {
+    Operand *Element = legalize(ElementToInsert, Legal_Mem | Legal_Reg);
+    Variable *T = makeReg(Ty);
+    _movp(T, SourceVectOperand);
+    _pinsrw(T, Element, Ctx->getConstantInt(IceType_i8, Index));
+    _movp(Inst->getDest(), T);
+  } else {
+    assert(Ty == IceType_v16i8 || Ty == IceType_v16i1);
+    // Spill the value to a stack slot and perform the insertion in
+    // memory.
+    // TODO(wala): SSE4.1 has pinsrb.
+    //
+    // TODO(wala): use legalize(SourceVectOperand, Legal_Mem) when
+    // support for legalizing to mem is implemented.
+    Variable *Slot = Func->makeVariable(Ty, Context.getNode());
+    Slot->setWeight(RegWeight::Zero);
+    _movp(Slot, legalizeToVar(SourceVectOperand));
+
+    // Compute the location of the position to insert in memory.
+    unsigned Offset = Index * typeWidthInBytes(LoweredElementTy);
+    OperandX8632Mem *Loc =
+        getMemoryOperandForStackSlot(LoweredElementTy, Slot, Offset);
+    _store(legalizeToVar(ElementToInsert), Loc);
+
+    Variable *T = makeReg(Ty);
+    _movp(T, Slot);
+    _movp(Inst->getDest(), T);
+  }
+}
+
 void TargetX8632::lowerIntrinsicCall(const InstIntrinsicCall *Instr) {
   switch (Instr->getIntrinsicInfo().ID) {
   case Intrinsics::AtomicCmpxchg: {
     if (!Intrinsics::VerifyMemoryOrder(
              llvm::cast<ConstantInteger>(Instr->getArg(3))->getValue())) {
       Func->setError("Unexpected memory ordering (success) for AtomicCmpxchg");
       return;
     }
     if (!Intrinsics::VerifyMemoryOrder(
              llvm::cast<ConstantInteger>(Instr->getArg(4))->getValue())) {
@@ skipping 911 matching lines @@
   // this value is initialized using register operations.
   Variable *Dest = makeVectorOfZeros(Ty, RegNum);
   Variable *MinusOne = makeReg(Ty);
   // Insert a FakeDef so the live range of MinusOne is not overestimated.
   Context.insert(InstFakeDef::create(Func, MinusOne));
   _pcmpeq(MinusOne, MinusOne);
   _psub(Dest, MinusOne);
   return Dest;
 }
 
+OperandX8632Mem *TargetX8632::getMemoryOperandForStackSlot(Type Ty,
+                                                           Variable *Slot,
+                                                           uint32_t Offset) {
+  // Ensure that Loc is a stack slot.
+  assert(Slot->getWeight() == RegWeight::Zero);

[Jim Stichnoth, 2014/07/17 19:49:01]

May also want to assert NoRegister.

[wala, 2014/07/17 22:14:12]

Done.

+  // Compute the location of Loc in memory.
+  // TODO(wala,stichnot): lea should not be required.  The address of
+  // the stack slot is known at compile time (although not until after
+  // addProlog()).
+  const Type PointerType = IceType_i32;
+  Variable *Loc = makeReg(PointerType);
+  _lea(Loc, Slot);
+  Constant *ConstantOffset = Ctx->getConstantInt(IceType_i32, Offset);
+  return OperandX8632Mem::create(Func, Ty, Loc, ConstantOffset);
+}
+
 // Helper for legalize() to emit the right code to lower an operand to a
 // register of the appropriate type.
 Variable *TargetX8632::copyToReg(Operand *Src, int32_t RegNum) {
   Type Ty = Src->getType();
   Variable *Reg = makeReg(Ty, RegNum);
   if (isVectorType(Ty)) {
     _movp(Reg, Src);
   } else {
     _mov(Reg, Src);
   }
@@ skipping 301 matching lines @@
     for (SizeT i = 0; i < Size; ++i) {
       Str << "\t.byte\t" << (((unsigned)Data[i]) & 0xff) << "\n";
     }
     Str << "\t.size\t" << MangledName << ", " << Size << "\n";
   }
   Str << "\t" << (IsInternal ? ".local" : ".global") << "\t" << MangledName
       << "\n";
 }
 
 } // end of namespace Ice