Subzero: ARM32: lowering of vector insert and extract.

src/IceTargetLoweringARM32.h

 //===- subzero/src/IceTargetLoweringARM32.h - ARM32 lowering ----*- C++ -*-===//
 //
 //                        The Subzero Code Generator
 //
 // This file is distributed under the University of Illinois Open Source
 // License. See LICENSE.TXT for details.
 //
 //===----------------------------------------------------------------------===//
 ///
 /// \file
@@ skipping 67 matching lines @@
   bool doBranchOpt(Inst *I, const CfgNode *NextNode) override;
 
   SizeT getNumRegisters() const override { return RegARM32::Reg_NUM; }
   Variable *getPhysicalRegister(SizeT RegNum, Type Ty = IceType_void) override;
   IceString getRegName(SizeT RegNum, Type Ty) const override;
   llvm::SmallBitVector getRegisterSet(RegSetMask Include,
                                       RegSetMask Exclude) const override;
   const llvm::SmallBitVector &
   getRegistersForVariable(const Variable *Var) const override {
     RegClass RC = Var->getRegClass();
-    assert(RC < RC_Target);
-    return TypeToRegisterSet[RC];
+    switch (RC) {
+    default:
+      assert(RC < RC_Target);
+      return TypeToRegisterSet[RC];
+    case RegARM32::RCARM32_QtoS:
+      assert(RegARM32::RCARM32_QtoS < RegARM32::RCARM32_NUM);

[Jim Stichnoth, 2016/02/03 15:28:38]

I think this assert is pretty much guaranteed to be true and you can remove it. 
(Another possibility is to use static_assert, but it's trivially true enough
that I just wouldn't bother.)

[Eric Holk, 2016/02/03 21:02:22]

Done.

+      return TypeToRegisterSet[RC];
+    }
   }
   const llvm::SmallBitVector &
   getAllRegistersForVariable(const Variable *Var) const override {
     RegClass RC = Var->getRegClass();
-    assert(RC < RC_Target);
+    assert((RegARM32::RegClassARM32)RC < RegARM32::RCARM32_NUM);
     return TypeToRegisterSetUnfiltered[RC];
   }
   const llvm::SmallBitVector &getAliasesForRegister(SizeT Reg) const override {
     return RegisterAliases[Reg];
   }
   bool hasFramePointer() const override { return UsesFramePointer; }
   void setHasFramePointer() override { UsesFramePointer = true; }
   SizeT getStackReg() const override { return RegARM32::Reg_sp; }
   SizeT getFrameReg() const override { return RegARM32::Reg_fp; }
   SizeT getFrameOrStackReg() const override {
@@ skipping 301 matching lines @@
     auto *Instr = Context.insert<InstARM32Mov>(Dest, Src0, Pred);
     Instr->setDestRedefined();
     if (Instr->isMultiDest()) {
       // If Instr is multi-dest, then Dest must be a Variable64On32. We add a
       // fake-def for Instr.DestHi here.
       assert(llvm::isa<Variable64On32>(Dest));
       Context.insert<InstFakeDef>(Instr->getDestHi());
     }
   }
 
+  // Generates a vmov instruction to extract the given index from a vector
+  // register.
+  void _extract(Variable *Dest, Variable *Src, uint32_t Index,

[Jim Stichnoth, 2016/02/03 15:28:38]

I would probably rename Src to Src0 for consistency.

[John, 2016/02/03 16:06:52]

It's been common practice (at least in the ARM32 backend) to name something Src
if it is the only Src. Unfortunately, this is not always followed, but I think
it makes moar sense -- if I see Src0 in the middle of a function I am tempted to
think there might be a Src1.

regardless of parameter naming, I would name this method _extract_element (and
_insert_element) simply because _insert and _extract are too generic.

[Eric Holk, 2016/02/03 21:02:22]

Done.

[Eric Holk, 2016/02/03 21:02:22]

Done.

+                CondARM32::Cond Pred = CondARM32::AL) {
+    Context.insert<InstARM32Extract>(Dest, Src, Index, Pred);
+  }
+
+  // Generates a vmov instruction to insert a value into the given index of a
+  // vector register.
+  void _insert(Variable *Dest, Variable *Val, uint32_t Index,

[Jim Stichnoth, 2016/02/03 15:28:38]

I would probably rename Val to Src0 for consistency.

[Eric Holk, 2016/02/03 21:02:22]

Done.

+               CondARM32::Cond Pred = CondARM32::AL) {
+    Context.insert<InstARM32Insert>(Dest, Val, Index, Pred);
+  }
+
   // --------------------------------------------------------------------------
   // Begin bool folding machinery.
   //
   // There are three types of boolean lowerings handled by this target:
   //
   // 1) Boolean expressions leading to a boolean Variable definition
   // ---------------------------------------------------------------
   //
   // Whenever a i1 Variable is live out (i.e., its live range extends beyond
   // the defining basic block) we do not fold the operation. We instead
@@ skipping 790 matching lines @@
 private:
   ~TargetHeaderARM32() = default;
 
   TargetARM32Features CPUFeatures;
 };
 
 } // end of namespace ARM32
 } // end of namespace Ice
 
 #endif // SUBZERO_SRC_ICETARGETLOWERINGARM32_H