Subzero: Use scalar arithmetic when no vector instruction exists.

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 1278 matching lines @@
         Variable *Src1R = LEGAL_HACK(Src1);
         _pshufd(T2, Src0R, Mask1030);
         _pshufd(T3, Src1R, Mask1030);
         _pmuludq(T1, Src1R);
         _pmuludq(T2, T3);
         _shufps(T1, T2, Ctx->getConstantInt(IceType_i8, Mask0202));
         _pshufd(T4, T1, Ctx->getConstantInt(IceType_i8, Mask0213));
         _movp(Dest, T4);
       } else {
         assert(Dest->getType() == IceType_v16i8);
-        // Sz_mul_v16i8
-        const IceString Helper = "Sz_mul_v16i8";
-        const SizeT MaxSrcs = 2;
-        InstCall *Call = makeHelperCall(Helper, Dest, MaxSrcs);
-        Call->addArg(Src0);
-        Call->addArg(Src1);
-        lowerCall(Call);
+        scalarizeArithmetic(InstArithmetic::Mul, Dest, Src0, Src1);

[Jim Stichnoth, 2014/08/07 19:58:10]

Consider using Inst->getOp() instead of ::Mul?

[wala, 2014/08/07 20:30:35]

Done.

       }
     } break;
-    case InstArithmetic::Shl: {
-      // Sz_shl_v4i32, Sz_shl_v8i16, Sz_shl_v16i8
-      const IceString Helper = "Sz_shl_" + typeIdentString(Dest->getType());
-      const SizeT MaxSrcs = 2;
-      InstCall *Call = makeHelperCall(Helper, Dest, MaxSrcs);
-      Call->addArg(Src0);
-      Call->addArg(Src1);
-      lowerCall(Call);
-    } break;
-    case InstArithmetic::Lshr: {
-      // Sz_lshr_v4i32, Sz_lshr_v8i16, Sz_lshr_v16i8
-      const IceString Helper = "Sz_lshr_" + typeIdentString(Dest->getType());
-      const SizeT MaxSrcs = 2;
-      InstCall *Call = makeHelperCall(Helper, Dest, MaxSrcs);
-      Call->addArg(Src0);
-      Call->addArg(Src1);
-      lowerCall(Call);
-    } break;
-    case InstArithmetic::Ashr: {
-      // Sz_ashr_v4i32, Sz_ashr_v8i16, Sz_ashr_v16i8
-      const IceString Helper = "Sz_ashr_" + typeIdentString(Dest->getType());
-      const SizeT MaxSrcs = 2;
-      InstCall *Call = makeHelperCall(Helper, Dest, MaxSrcs);
-      Call->addArg(Src0);
-      Call->addArg(Src1);
-      lowerCall(Call);
-    } break;
-    case InstArithmetic::Udiv: {
-      // Sz_udiv_v4i32, Sz_udiv_v8i16, Sz_udiv_v16i8
-      const IceString Helper = "Sz_udiv_" + typeIdentString(Dest->getType());
-      const SizeT MaxSrcs = 2;
-      InstCall *Call = makeHelperCall(Helper, Dest, MaxSrcs);
-      Call->addArg(Src0);
-      Call->addArg(Src1);
-      lowerCall(Call);
-    } break;
-    case InstArithmetic::Sdiv: {
-      // Sz_sdiv_v4i32, Sz_sdiv_v8i16, Sz_sdiv_v16i8
-      const IceString Helper = "Sz_sdiv_" + typeIdentString(Dest->getType());
-      const SizeT MaxSrcs = 2;
-      InstCall *Call = makeHelperCall(Helper, Dest, MaxSrcs);
-      Call->addArg(Src0);
-      Call->addArg(Src1);
-      lowerCall(Call);
-    } break;
-    case InstArithmetic::Urem: {
-      // Sz_urem_v4i32, Sz_urem_v8i16, Sz_urem_v16i8
-      const IceString Helper = "Sz_urem_" + typeIdentString(Dest->getType());
-      const SizeT MaxSrcs = 2;
-      InstCall *Call = makeHelperCall(Helper, Dest, MaxSrcs);
-      Call->addArg(Src0);
-      Call->addArg(Src1);
-      lowerCall(Call);
-    } break;
-    case InstArithmetic::Srem: {
-      // Sz_srem_v4i32, Sz_srem_v8i16, Sz_srem_v16i8
-      const IceString Helper = "Sz_srem_" + typeIdentString(Dest->getType());
-      const SizeT MaxSrcs = 2;
-      InstCall *Call = makeHelperCall(Helper, Dest, MaxSrcs);
-      Call->addArg(Src0);
-      Call->addArg(Src1);
-      lowerCall(Call);
-    } break;
+    case InstArithmetic::Shl:
+    case InstArithmetic::Lshr:
+    case InstArithmetic::Ashr:
+    case InstArithmetic::Udiv:
+    case InstArithmetic::Urem:
+    case InstArithmetic::Sdiv:
+    case InstArithmetic::Srem:
+      scalarizeArithmetic(Inst->getOp(), Dest, Src0, Src1);
+      break;
     case InstArithmetic::Fadd: {
       Variable *T = makeReg(Dest->getType());
       _movp(T, Src0);
       _addps(T, LEGAL_HACK(Src1));
       _movp(Dest, T);
     } break;
     case InstArithmetic::Fsub: {
       Variable *T = makeReg(Dest->getType());
       _movp(T, Src0);
       _subps(T, LEGAL_HACK(Src1));
       _movp(Dest, T);
     } break;
     case InstArithmetic::Fmul: {
       Variable *T = makeReg(Dest->getType());
       _movp(T, Src0);
       _mulps(T, LEGAL_HACK(Src1));
       _movp(Dest, T);
     } break;
     case InstArithmetic::Fdiv: {
       Variable *T = makeReg(Dest->getType());
       _movp(T, Src0);
       _divps(T, LEGAL_HACK(Src1));
       _movp(Dest, T);
     } break;
-    case InstArithmetic::Frem: {
-      const SizeT MaxSrcs = 2;
-      InstCall *Call = makeHelperCall("Sz_frem_v4f32", Dest, MaxSrcs);
-      Call->addArg(Src0);
-      Call->addArg(Src1);
-      lowerCall(Call);
-    } break;
+    case InstArithmetic::Frem:
+      scalarizeArithmetic(InstArithmetic::Frem, Dest, Src0, Src1);

[Jim Stichnoth, 2014/08/07 19:58:10]

Consider using Inst->getOp() instead of ::Frem?

[wala, 2014/08/07 20:30:35]

Done.

+      break;
     }
 #undef LEGAL_HACK
   } else { // Dest->getType() is non-i64 scalar
     Variable *T_edx = NULL;
     Variable *T = NULL;
     switch (Inst->getOp()) {
     case InstArithmetic::_num:
       llvm_unreachable("Unknown arithmetic operator");
       break;
     case InstArithmetic::Add:
@@ skipping 71 matching lines @@
       } else {
         Constant *Zero = Ctx->getConstantZero(IceType_i32);
         _mov(T, Src0, Reg_eax);
         _mov(T_edx, Zero, Reg_edx);
         _div(T, Src1, T_edx);
         _mov(Dest, T);
       }
       break;
     case InstArithmetic::Sdiv:
       Src1 = legalize(Src1, Legal_Reg | Legal_Mem);
-      T_edx = makeReg(IceType_i32, Reg_edx);
-      _mov(T, Src0, Reg_eax);
-      _cdq(T_edx, T);
-      _idiv(T, Src1, T_edx);
-      _mov(Dest, T);
+      if (Dest->getType() == IceType_i8) {
+        _mov(T, Src0, Reg_eax);
+        _cbwdq(T, T);
+        _idiv(T, Src1, T);
+        _mov(Dest, T);
+      } else {
+        T_edx = makeReg(IceType_i32, Reg_edx);
+        _mov(T, Src0, Reg_eax);
+        _cbwdq(T_edx, T);
+        _idiv(T, Src1, T_edx);
+        _mov(Dest, T);
+      }
       break;
     case InstArithmetic::Urem:
       Src1 = legalize(Src1, Legal_Reg | Legal_Mem);
       if (Dest->getType() == IceType_i8) {
         Variable *T_ah = NULL;
         Constant *Zero = Ctx->getConstantZero(IceType_i8);
         _mov(T, Src0, Reg_eax);
         _mov(T_ah, Zero, Reg_ah);
         _div(T_ah, Src1, T);
         _mov(Dest, T_ah);
       } else {
         Constant *Zero = Ctx->getConstantZero(IceType_i32);
         _mov(T_edx, Zero, Reg_edx);
         _mov(T, Src0, Reg_eax);
         _div(T_edx, Src1, T);
         _mov(Dest, T_edx);
       }
       break;
     case InstArithmetic::Srem:
       Src1 = legalize(Src1, Legal_Reg | Legal_Mem);
-      T_edx = makeReg(IceType_i32, Reg_edx);
-      _mov(T, Src0, Reg_eax);
-      _cdq(T_edx, T);
-      _idiv(T_edx, Src1, T);
-      _mov(Dest, T_edx);
+      if (Dest->getType() == IceType_i8) {
+        Variable *T_ah = makeReg(IceType_i8, Reg_ah);
+        _mov(T, Src0, Reg_eax);
+        _cbwdq(T, T);
+        Context.insert(InstFakeDef::create(Func, T_ah));
+        _idiv(T_ah, Src1, T);
+        _mov(Dest, T_ah);
+      } else {
+        T_edx = makeReg(IceType_i32, Reg_edx);
+        _mov(T, Src0, Reg_eax);
+        _cbwdq(T_edx, T);
+        _idiv(T_edx, Src1, T);
+        _mov(Dest, T_edx);
+      }
       break;
     case InstArithmetic::Fadd:
       _mov(T, Src0);
       _addss(T, Src1);
       _mov(Dest, T);
       break;
     case InstArithmetic::Fsub:
       _mov(T, Src0);
       _subss(T, Src1);
       _mov(Dest, T);
@@ skipping 2204 matching lines @@
     Src0 = legalize(Src0, Legal_All, true);
   for (SizeT I = 0; I < NumCases; ++I) {
     Operand *Value = Ctx->getConstantInt(IceType_i32, Inst->getValue(I));
     _cmp(Src0, Value);
     _br(InstX8632Br::Br_e, Inst->getLabel(I));
   }
 
   _br(Inst->getLabelDefault());
 }
 
+void TargetX8632::scalarizeArithmetic(InstArithmetic::OpKind K, Variable *Dest,

[Jim Stichnoth, 2014/08/07 19:58:10]

Name the arg "Kind", or something else more descriptive than "K"?

[wala, 2014/08/07 20:30:35]

Done.

+                                      Operand *Src0, Operand *Src1) {
+  assert(isVectorType(Dest->getType()));
+  Type Ty = Dest->getType();
+  Type ElementTy = typeElementType(Ty);
+  SizeT NumElements = typeNumElements(Ty);
+
+  Operand *T = Ctx->getConstantUndef(Ty);
+  for (SizeT I = 0; I < NumElements; ++I) {
+    Constant *Index = Ctx->getConstantInt(IceType_i32, I);
+
+    // Extract the next two inputs.
+    Variable *Op0 = Func->makeVariable(ElementTy, Context.getNode());
+    lowerExtractElement(InstExtractElement::create(Func, Op0, Src0, Index));
+    Variable *Op1 = Func->makeVariable(ElementTy, Context.getNode());
+    lowerExtractElement(InstExtractElement::create(Func, Op1, Src1, Index));
+
+    // Perform the arithmetic as a scalar operation.
+    Variable *Res = Func->makeVariable(ElementTy, Context.getNode());
+    lowerArithmetic(InstArithmetic::create(Func, K, Res, Op0, Op1));
+
+    // Insert the result into position.
+    Variable *DestT = Func->makeVariable(Ty, Context.getNode());
+    lowerInsertElement(InstInsertElement::create(Func, DestT, T, Res, Index));
+    T = DestT;
+    // TODO: Use postLower() in -Om1 mode to avoid buildup of infinite
+    // weight temporaries.
+  }
+
+  lowerAssign(InstAssign::create(Func, Dest, T));
+}
+
 // The following pattern occurs often in lowered C and C++ code:
 //
 //   %cmp     = fcmp/icmp pred <n x ty> %src0, %src1
 //   %cmp.ext = sext <n x i1> %cmp to <n x ty>
 //
 // We can eliminate the sext operation by copying the result of pcmpeqd,
 // pcmpgtd, or cmpps (which produce sign extended results) to the result
 // of the sext operation.
 void
 TargetX8632::eliminateNextVectorSextInstruction(Variable *SignExtendedResult) {
@@ skipping 406 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