Add support for passing and returning vectors in accordance with the x86 calling convention.

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 229 matching lines @@
   if (Func->hasError())
     return;
   T_deletePhis.printElapsedUs(Context, "deletePhis()");
   Func->dump("After Phi lowering");
 
   // Address mode optimization.
   Timer T_doAddressOpt;
   Func->doAddressOpt();
   T_doAddressOpt.printElapsedUs(Context, "doAddressOpt()");
 
+  // Argument lowering
+  Timer T_argLowering;
+  Func->doArgLowering();
+  T_argLowering.printElapsedUs(Context, "lowerArguments()");
+
   // Target lowering.  This requires liveness analysis for some parts
   // of the lowering decisions, such as compare/branch fusing.  If
   // non-lightweight liveness analysis is used, the instructions need
   // to be renumbered first.  TODO: This renumbering should only be
   // necessary if we're actually calculating live intervals, which we
   // only do for register allocation.
   Timer T_renumber1;
   Func->renumberInstructions();
   if (Func->hasError())
     return;
   T_renumber1.printElapsedUs(Context, "renumberInstructions()");
+
   // TODO: It should be sufficient to use the fastest liveness
   // calculation, i.e. livenessLightweight().  However, for some
   // reason that slows down the rest of the translation.  Investigate.
   Timer T_liveness1;
   Func->liveness(Liveness_Basic);
   if (Func->hasError())
     return;
   T_liveness1.printElapsedUs(Context, "liveness()");
   Func->dump("After x86 address mode opt");
+
   Timer T_genCode;
   Func->genCode();
   if (Func->hasError())
     return;
   T_genCode.printElapsedUs(Context, "genCode()");
 
   // Register allocation.  This requires instruction renumbering and
   // full liveness analysis.
   Timer T_renumber2;
   Func->renumberInstructions();
@@ skipping 42 matching lines @@
   if (Func->hasError())
     return;
   T_placePhiStores.printElapsedUs(Context, "placePhiStores()");
   Timer T_deletePhis;
   Func->deletePhis();
   if (Func->hasError())
     return;
   T_deletePhis.printElapsedUs(Context, "deletePhis()");
   Func->dump("After Phi lowering");
 
+  Timer T_argLowering;
+  Func->doArgLowering();
+  T_argLowering.printElapsedUs(Context, "lowerArguments()");
+
   Timer T_genCode;
   Func->genCode();
   if (Func->hasError())
     return;
   T_genCode.printElapsedUs(Context, "genCode()");
   Func->dump("After initial x8632 codegen");
 
   Timer T_genFrame;
   Func->genFrame();
   if (Func->hasError())
@@ skipping 63 matching lines @@
   if (!hasFramePointer())
     Offset += getStackAdjustment();
   if (Offset) {
     if (Offset > 0)
       Str << "+";
     Str << Offset;
   }
   Str << "]";
 }
 
-// Helper function for addProlog().  Sets the frame offset for Arg,
-// updates InArgsSizeBytes according to Arg's width, and generates an
-// instruction to copy Arg into its assigned register if applicable.
-// For an I64 arg that has been split into Lo and Hi components, it
-// calls itself recursively on the components, taking care to handle
-// Lo first because of the little-endian architecture.
-void TargetX8632::setArgOffsetAndCopy(Variable *Arg, Variable *FramePtr,
-                                      size_t BasicFrameOffset,
-                                      size_t &InArgsSizeBytes) {
+void TargetX8632::lowerArguments() {
+  VarList &Args = Func->getArgs();
+  // The first four arguments of vector type, regardless of their
+  // position relative to the other arguments in the argument list, are
+  // passed in registers xmm0 - xmm3.
+  unsigned NumXmmArgs = 0;
+
+  Context.init(Func->getEntryNode());
+  Context.setInsertPoint(Context.getCur());
+
+  for (SizeT I = 0, E = Args.size(); I < E && NumXmmArgs < 4; ++I) {

[jvoung (off chromium), 2014/07/09 04:26:53]

Might be worth making the "4" a symbolic constant too. Then the file where the
constant is defined can also document the calling convention. At least for LLVM
it's kind of nice to have a file like "X86CallingConv.td" that also serves as a
reference. On the other hand, we only support one calling convention, so it's
not that big of a reference =)

[Jim Stichnoth, 2014/07/09 16:56:13]

I agree, except that between x86-32 and x86-64 lowering (not to mention raw
versus sandboxed), there will likely be significant refactoring/reuse that makes
this more relevant.

[wala, 2014/07/09 19:05:12]

Done.

+    Variable *Arg = Args[I];
+    Type Ty = Arg->getType();
+    if (!isVectorType(Ty))
+      continue;
+    // Replace Arg in the argument list with the home register.  Then
+    // generate an instruction in the prolog to copy the home register
+    // to the assigned location of Arg.
+    int32_t RegNum = Reg_xmm0 + NumXmmArgs;
+    ++NumXmmArgs;
+    IceString Name = "home_reg:" + Arg->getName();
+    const CfgNode *DefNode = NULL;
+    Variable *RegisterArg = Func->makeVariable(Ty, DefNode, Name);
+    RegisterArg->setRegNum(RegNum);
+    RegisterArg->setIsArg(Func);
+    Arg->setIsArg(Func, false);
+
+    Args[I] = RegisterArg;
+    Context.insert(InstAssign::create(Func, Arg, RegisterArg));
+  }
+}
+
+// Helper function for addProlog().
+//
+// This assumes Arg is an argument passed on the stack.  This sets the
+// frame offset for Arg and updates InArgsSizeBytes according to Arg's
+// width.  For an I64 arg that has been split into Lo and Hi components,
+// it calls itself recursively on the components, taking care to handle
+// Lo first because of the little-endian architecture.  Lastly, this
+// function generates an instruction to copy Arg into its assigned
+// register if applicable.
+void TargetX8632::finishArgumentLowering(Variable *Arg, Variable *FramePtr,
+                                         size_t BasicFrameOffset,
+                                         size_t &InArgsSizeBytes) {
   Variable *Lo = Arg->getLo();
   Variable *Hi = Arg->getHi();
   Type Ty = Arg->getType();
   if (Lo && Hi && Ty == IceType_i64) {
     assert(Lo->getType() != IceType_i64); // don't want infinite recursion
     assert(Hi->getType() != IceType_i64); // don't want infinite recursion
-    setArgOffsetAndCopy(Lo, FramePtr, BasicFrameOffset, InArgsSizeBytes);
-    setArgOffsetAndCopy(Hi, FramePtr, BasicFrameOffset, InArgsSizeBytes);
+    finishArgumentLowering(Lo, FramePtr, BasicFrameOffset, InArgsSizeBytes);
+    finishArgumentLowering(Hi, FramePtr, BasicFrameOffset, InArgsSizeBytes);
     return;
   }
   Arg->setStackOffset(BasicFrameOffset + InArgsSizeBytes);
+  InArgsSizeBytes += typeWidthInBytesOnStack(Ty);
   if (Arg->hasReg()) {
     assert(Ty != IceType_i64);
     OperandX8632Mem *Mem = OperandX8632Mem::create(
         Func, Ty, FramePtr,
         Ctx->getConstantInt(IceType_i32, Arg->getStackOffset()));
-    _mov(Arg, Mem);
+    if (isVectorType(Arg->getType())) {
+      _movp(Arg, Mem);
+    } else {
+      _mov(Arg, Mem);
+    }
   }
-  InArgsSizeBytes += typeWidthInBytesOnStack(Ty);
 }
 
 Type TargetX8632::stackSlotType() { return IceType_i32; }
 
 void TargetX8632::addProlog(CfgNode *Node) {
   // If SimpleCoalescing is false, each variable without a register
   // gets its own unique stack slot, which leads to large stack
   // frames.  If SimpleCoalescing is true, then each "global" variable
   // without a register gets its own slot, but "local" variable slots
   // are reused across basic blocks.  E.g., if A and B are local to
@@ skipping 29 matching lines @@
   RegsUsed = llvm::SmallBitVector(CalleeSaves.size());
   const VarList &Variables = Func->getVariables();
   const VarList &Args = Func->getArgs();
   for (VarList::const_iterator I = Variables.begin(), E = Variables.end();
        I != E; ++I) {
     Variable *Var = *I;
     if (Var->hasReg()) {
       RegsUsed[Var->getRegNum()] = true;
       continue;
     }
-    // An argument passed on the stack already has a stack slot.
+    // An argument either does not need a stack slot (if passed in a
+    // register) or already has one (if passed on the stack).
     if (Var->getIsArg())
       continue;
     // An unreferenced variable doesn't need a stack slot.
     if (ComputedLiveRanges && Var->getLiveRange().isEmpty())
       continue;
     // A spill slot linked to a variable with a stack slot should reuse
     // that stack slot.
     if (Var->getWeight() == RegWeight::Zero && Var->getRegisterOverlap()) {
       if (Variable *Linked = Var->getPreferredRegister()) {
         if (!Linked->hasReg())
@@ skipping 37 matching lines @@
     _mov(ebp, esp);
   }
 
   // Generate "sub esp, LocalsSizeBytes"
   if (LocalsSizeBytes)
     _sub(getPhysicalRegister(Reg_esp),
          Ctx->getConstantInt(IceType_i32, LocalsSizeBytes));
 
   resetStackAdjustment();
 
-  // Fill in stack offsets for args, and copy args into registers for
-  // those that were register-allocated.  Args are pushed right to
+  // Fill in stack offsets for stack args, and copy args into registers
+  // for those that were register-allocated.  Args are pushed right to
   // left, so Arg[0] is closest to the stack/frame pointer.
-  //
-  // TODO: Make this right for different width args, calling
-  // conventions, etc.  For one thing, args passed in registers will
-  // need to be copied/shuffled to their home registers (the
-  // RegManager code may have some permutation logic to leverage),
-  // and if they have no home register, home space will need to be
-  // allocated on the stack to copy into.
   Variable *FramePtr = getPhysicalRegister(getFrameOrStackReg());
   size_t BasicFrameOffset = PreservedRegsSizeBytes + RetIpSizeBytes;
   if (!IsEbpBasedFrame)
     BasicFrameOffset += LocalsSizeBytes;
+
+  unsigned NumXmmArgs = 0;
   for (SizeT i = 0; i < Args.size(); ++i) {
     Variable *Arg = Args[i];
-    setArgOffsetAndCopy(Arg, FramePtr, BasicFrameOffset, InArgsSizeBytes);
+    // Skip arguments passed in registers.
+    if (isVectorType(Arg->getType()) && NumXmmArgs < 4) {
+      ++NumXmmArgs;
+      continue;
+    }
+    finishArgumentLowering(Arg, FramePtr, BasicFrameOffset, InArgsSizeBytes);
   }
 
   // Fill in stack offsets for locals.
   size_t TotalGlobalsSize = GlobalsSize;
   GlobalsSize = 0;
   LocalsSize.assign(LocalsSize.size(), 0);
   size_t NextStackOffset = 0;
   for (VarList::const_iterator I = Variables.begin(), E = Variables.end();
        I != E; ++I) {
     Variable *Var = *I;
@@ skipping 666 matching lines @@
     Operand *Src0Lo = loOperand(Src0);
     Operand *Src0Hi = hiOperand(Src0);
     Variable *DestLo = llvm::cast<Variable>(loOperand(Dest));
     Variable *DestHi = llvm::cast<Variable>(hiOperand(Dest));
     Variable *T_Lo = NULL, *T_Hi = NULL;
     _mov(T_Lo, Src0Lo);
     _mov(DestLo, T_Lo);
     _mov(T_Hi, Src0Hi);
     _mov(DestHi, T_Hi);
   } else {
-    const bool AllowOverlap = true;
+    const bool AllowOverlap = false;
     // RI is either a physical register or an immediate.
     Operand *RI = legalize(Src0, Legal_Reg | Legal_Imm, AllowOverlap);
-    _mov(Dest, RI);
+    if (isVectorType(Dest->getType()))
+      _movp(Dest, RI);
+    else
+      _mov(Dest, RI);
   }
 }
 
 void TargetX8632::lowerBr(const InstBr *Inst) {
   if (Inst->isUnconditional()) {
     _br(Inst->getTargetUnconditional());
   } else {
     Operand *Src0 = legalize(Inst->getCondition());
     Constant *Zero = Ctx->getConstantZero(IceType_i32);
     _cmp(Src0, Zero);
     _br(InstX8632Br::Br_ne, Inst->getTargetTrue(), Inst->getTargetFalse());
   }
 }
 
 void TargetX8632::lowerCall(const InstCall *Instr) {
-  // Generate a sequence of push instructions, pushing right to left,
-  // keeping track of stack offsets in case a push involves a stack
-  // operand and we are using an esp-based frame.
+  // For stack arguments, generate a sequence of push instructions,
+  // pushing right to left, keeping track of stack offsets in case a
+  // push involves a stack operand and we are using an esp-based frame.
   uint32_t StackOffset = 0;
+  // Keep track of the number of xmm registers that get used to pass
+  // arguments.
+  unsigned NumXmmArgs = 0;
+  VarList RegisterArgs;
   // TODO: If for some reason the call instruction gets dead-code
   // eliminated after lowering, we would need to ensure that the
   // pre-call push instructions and the post-call esp adjustment get
   // eliminated as well.
   for (SizeT NumArgs = Instr->getNumArgs(), i = 0; i < NumArgs; ++i) {
     Operand *Arg = legalize(Instr->getArg(NumArgs - i - 1));
+    bool ArgInRegister = false;
     if (Arg->getType() == IceType_i64) {
       _push(hiOperand(Arg));
       _push(loOperand(Arg));
     } else if (Arg->getType() == IceType_f64) {
       // If the Arg turns out to be a memory operand, we need to push
       // 8 bytes, which requires two push instructions.  This ends up
       // being somewhat clumsy in the current IR, so we use a
       // workaround.  Force the operand into a (xmm) register, and
       // then push the register.  An xmm register push is actually not
       // possible in x86, but the Push instruction emitter handles
       // this by decrementing the stack pointer and directly writing
       // the xmm register value.
       Variable *T = NULL;
       _mov(T, Arg);
       _push(T);
+    } else if (isVectorType(Arg->getType())) {
+      if (NumXmmArgs < 4) {
+        Variable *Reg = legalizeToVar(Arg, false, Reg_xmm0 + NumXmmArgs);
+        ++NumXmmArgs;
+        ArgInRegister = true;
+        RegisterArgs.push_back(Reg);
+      } else {
+        // sub    esp,   16
+        // movups [esp], legalize_to_reg(Arg)
+        Variable *esp = getPhysicalRegister(Reg_esp);
+        size_t Width = typeWidthInBytesOnStack(Arg->getType());
+        _sub(esp, Ctx->getConstantInt(IceType_i8, Width));
+        Constant *Zero = Ctx->getConstantZero(IceType_i8);
+        OperandX8632Mem *Dest =
+            OperandX8632Mem::create(Func, Arg->getType(), esp, Zero);
+        _storep(legalize(Arg, Legal_Reg), Dest);
+      }
     } else {
       // Otherwise PNaCl requires parameter types to be at least 32-bits.
       assert(Arg->getType() == IceType_f32 || Arg->getType() == IceType_i32);
       _push(Arg);
     }
-    StackOffset += typeWidthInBytesOnStack(Arg->getType());
+    if (!ArgInRegister) {
+      StackOffset += typeWidthInBytesOnStack(Arg->getType());
+    }
+  }
+  // Generate a FakeUse of all register arguments so that they do not
+  // get dead code eliminated as a result of the FakeKill of scratch
+  // registers after the call.

[Jim Stichnoth, 2014/07/09 16:56:13]

I think it wouldn't hurt to assert(RegisterArgs.size()==NumXmmArgs) here.

[wala, 2014/07/09 19:05:12]

Unnecessary in the revised code.

+  for (VarList::const_iterator I = RegisterArgs.begin(), E = RegisterArgs.end();
+       I != E; ++I) {
+    Context.insert(InstFakeUse::create(Func, *I));
   }
   // Generate the call instruction.  Assign its result to a temporary
   // with high register allocation weight.
   Variable *Dest = Instr->getDest();
-  Variable *eax = NULL; // doubles as RegLo as necessary
-  Variable *edx = NULL;
+  // ReturnReg doubles as ReturnRegLo as necessary.
+  Variable *ReturnReg = NULL;
+  Variable *ReturnRegHi = NULL;
   if (Dest) {
     switch (Dest->getType()) {
     case IceType_NUM:
       llvm_unreachable("Invalid Call dest type");
       break;
     case IceType_void:
       break;
     case IceType_i1:
     case IceType_i8:
     case IceType_i16:
     case IceType_i32:
-      eax = makeReg(Dest->getType(), Reg_eax);
+      ReturnReg = makeReg(Dest->getType(), Reg_eax);
       break;
     case IceType_i64:
-      eax = makeReg(IceType_i32, Reg_eax);
-      edx = makeReg(IceType_i32, Reg_edx);
+      ReturnReg = makeReg(IceType_i32, Reg_eax);
+      ReturnRegHi = makeReg(IceType_i32, Reg_edx);
       break;
     case IceType_f32:
     case IceType_f64:
-      // Leave eax==edx==NULL, and capture the result with the fstp
-      // instruction.
+      // Leave ReturnReg==ReturnRegHi==NULL, and capture the result with
+      // the fstp instruction.
       break;
     case IceType_v4i1:
     case IceType_v8i1:
     case IceType_v16i1:
     case IceType_v16i8:
     case IceType_v8i16:
     case IceType_v4i32:
-    case IceType_v4f32: {
-      // TODO(wala): Handle return values of vector type in the caller.
-      IceString Ty;
-      llvm::raw_string_ostream BaseOS(Ty);
-      Ostream OS(&BaseOS);
-      OS << Dest->getType();
-      Func->setError("Unhandled dest type: " + BaseOS.str());
-      return;
-    }
+    case IceType_v4f32:
+      ReturnReg = makeReg(Dest->getType(), Reg_xmm0);
+      break;
     }
   }
   // TODO(stichnot): LEAHACK: remove Legal_All (and use default) once
   // a proper emitter is used.
   Operand *CallTarget = legalize(Instr->getCallTarget(), Legal_All);
-  Inst *NewCall = InstX8632Call::create(Func, eax, CallTarget);
+  Inst *NewCall = InstX8632Call::create(Func, ReturnReg, CallTarget);
   Context.insert(NewCall);
-  if (edx)
-    Context.insert(InstFakeDef::create(Func, edx));
+  if (ReturnRegHi)
+    Context.insert(InstFakeDef::create(Func, ReturnRegHi));
 
   // Add the appropriate offset to esp.
   if (StackOffset) {
     Variable *esp = Func->getTarget()->getPhysicalRegister(Reg_esp);
     _add(esp, Ctx->getConstantInt(IceType_i32, StackOffset));
   }
 
   // Insert a register-kill pseudo instruction.
   VarList KilledRegs;
   for (SizeT i = 0; i < ScratchRegs.size(); ++i) {
     if (ScratchRegs[i])
       KilledRegs.push_back(Func->getTarget()->getPhysicalRegister(i));
   }
   Context.insert(InstFakeKill::create(Func, KilledRegs, NewCall));
 
   // Generate a FakeUse to keep the call live if necessary.
-  if (Instr->hasSideEffects() && eax) {
-    Inst *FakeUse = InstFakeUse::create(Func, eax);
+  if (Instr->hasSideEffects() && ReturnReg) {
+    Inst *FakeUse = InstFakeUse::create(Func, ReturnReg);
     Context.insert(FakeUse);
   }
 
-  // Generate Dest=eax assignment.
-  if (Dest && eax) {
-    if (edx) {
+  // Assign the result of the call to Dest.
+  if (!Dest) {
+    return;
+  } else if (ReturnReg) {

[Jim Stichnoth, 2014/07/09 16:56:13]

Don't use "else if" here.
http://llvm.org/docs/CodingStandards.html#don-t-use-else-after-a-return

[wala, 2014/07/09 19:05:12]

Done.

+    if (ReturnRegHi) {
+      assert(Dest->getType() == IceType_i64);
       split64(Dest);
       Variable *DestLo = Dest->getLo();
       Variable *DestHi = Dest->getHi();
-      DestLo->setPreferredRegister(eax, false);
-      DestHi->setPreferredRegister(edx, false);
-      _mov(DestLo, eax);
-      _mov(DestHi, edx);
+      DestLo->setPreferredRegister(ReturnReg, false);
+      DestHi->setPreferredRegister(ReturnRegHi, false);
+      _mov(DestLo, ReturnReg);
+      _mov(DestHi, ReturnRegHi);
     } else {
-      Dest->setPreferredRegister(eax, false);
-      _mov(Dest, eax);
+      assert(Dest->getType() == IceType_i32 || Dest->getType() == IceType_i16 ||
+             Dest->getType() == IceType_i8 || Dest->getType() == IceType_i1 ||
+             isVectorType(Dest->getType()));
+      Dest->setPreferredRegister(ReturnReg, false);
+      if (isVectorType(Dest->getType())) {
+        _movp(Dest, ReturnReg);
+      } else {
+        _mov(Dest, ReturnReg);
+      }
     }
-  }
-
-  // Special treatment for an FP function which returns its result in
-  // st(0).
-  if (Dest &&
-      (Dest->getType() == IceType_f32 || Dest->getType() == IceType_f64)) {
+  } else if (Dest->getType() == IceType_f32 || Dest->getType() == IceType_f64) {
+    // Special treatment for an FP function which returns its result in
+    // st(0).
     _fstp(Dest);
-    // If Dest ends up being a physical xmm register, the fstp emit
-    // code will route st(0) through a temporary stack slot.
+    // If Dest ends up being a physical xmm register, the fstp emit code
+    // will route st(0) through a temporary stack slot.
   }
 }
 
 void TargetX8632::lowerCast(const InstCast *Inst) {
   // a = cast(b) ==> t=cast(b); a=t; (link t->b, link a->t, no overlap)
   InstCast::OpKind CastKind = Inst->getCastKind();
   Variable *Dest = Inst->getDest();
   switch (CastKind) {
   default:
     Func->setError("Cast type not supported");
@@ skipping 1325 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