Subzero. ARM32. Vector lowering. Add.

src/IceTargetLoweringARM32.cpp

 //===- subzero/src/IceTargetLoweringARM32.cpp - ARM32 lowering ------------===//
 //
 //                        The Subzero Code Generator
 //
 // This file is distributed under the University of Illinois Open Source
 // License. See LICENSE.TXT for details.
 //
 //===----------------------------------------------------------------------===//
 ///
 /// \file
@@ skipping 2780 matching lines @@
   }
 
   Operand *Src0 = legalizeUndef(Instr->getSrc(0));
   Operand *Src1 = legalizeUndef(Instr->getSrc(1));
   if (DestTy == IceType_i64) {
     lowerInt64Arithmetic(Instr->getOp(), Instr->getDest(), Src0, Src1);
     return;
   }
 
   if (isVectorType(DestTy)) {
-    UnimplementedLoweringError(this, Instr);
-    return;
+    switch(Instr->getOp()) {
+      default:
+        UnimplementedLoweringError(this, Instr);
+        return;
+      // this explicitly whitelists vector instructions we have some reason to
+      // believe works.
+      case InstArithmetic::Fadd:
+      case InstArithmetic::Add:
+        break;

[Jim Stichnoth, 2016/01/25 21:35:31]

Is it generally true that vector lowering will be basically identical to scalar
lowering, except for minor differences like _vadd instead of _add?   And it
applies as well to the more complicated stuff like mla?  If so, that's really
cool and refreshingly unlike x86.

[Eric Holk, 2016/01/25 23:56:58]

So far this seems true. I was pleasantly surprised. It looks like there might be
subtleties for things like element-wise multiply vs multiply by scalar.

+    }
   }
 
   // DestTy is a non-i64 scalar.

[Jim Stichnoth, 2016/01/25 21:35:31]

This comment is no longer true.

[Eric Holk, 2016/01/25 23:56:58]

I'll remove the comment.

   Variable *T = makeReg(DestTy);
 
   // * Handle div/rem separately. They require a non-legalized Src1 to inspect
   // whether or not Src1 is a non-zero constant. Once legalized it is more
   // difficult to determine (constant may be moved to a register).
   // * Handle floating point arithmetic separately: they require Src1 to be
   // legalized to a register.
   switch (Instr->getOp()) {
   default:
     break;
@@ skipping 77 matching lines @@
   }
 
   // Handle everything else here.
   Int32Operands Srcs(Src0, Src1);
   switch (Instr->getOp()) {
   case InstArithmetic::_num:
     llvm::report_fatal_error("Unknown arithmetic operator");
     return;
   case InstArithmetic::Add: {
     if (const Inst *Src1Producer = Computations.getProducerOf(Src1)) {
+      assert(!isVectorType(DestTy));
       Variable *Src0R = legalizeToReg(Src0);
       Variable *Src1R = legalizeToReg(Src1Producer->getSrc(0));
       Variable *Src2R = legalizeToReg(Src1Producer->getSrc(1));
       _mla(T, Src1R, Src2R, Src0R);
       _mov(Dest, T);
       return;
     }
 
     if (Srcs.hasConstOperand()) {
       if (!Srcs.immediateIsFlexEncodable() &&
           Srcs.negatedImmediateIsFlexEncodable()) {
+        assert(!isVectorType(DestTy));
         Variable *Src0R = Srcs.src0R(this);
         Operand *Src1F = Srcs.negatedSrc1F(this);
         if (!Srcs.swappedOperands()) {
           _sub(T, Src0R, Src1F);
         } else {
           _rsb(T, Src0R, Src1F);
         }
         _mov(Dest, T);
         return;
       }
     }
     Variable *Src0R = Srcs.src0R(this);
-    Operand *Src1RF = Srcs.src1RF(this);
-    _add(T, Src0R, Src1RF);
+    if (isVectorType(DestTy)) {
+      Variable *Src1R = legalizeToReg(Src1);
+      _vadd(T, Src0R, Src1R);
+    }
+    else {
+      Operand *Src1RF = Srcs.src1RF(this);
+      _add(T, Src0R, Src1RF);
+    }
     _mov(Dest, T);
     return;
   }
   case InstArithmetic::And: {
     if (Srcs.hasConstOperand()) {
       if (!Srcs.immediateIsFlexEncodable() &&
           Srcs.invertedImmediateIsFlexEncodable()) {
         Variable *Src0R = Srcs.src0R(this);
         Operand *Src1F = Srcs.invertedSrc1F(this);
         _bic(T, Src0R, Src1F);
@@ skipping 3528 matching lines @@
   // Technically R9 is used for TLS with Sandboxing, and we reserve it.
   // However, for compatibility with current NaCl LLVM, don't claim that.
   Str << ".eabi_attribute 14, 3   @ Tag_ABI_PCS_R9_use: Not used\n";
 }
 
 llvm::SmallBitVector TargetARM32::TypeToRegisterSet[RegARM32::RCARM32_NUM];
 llvm::SmallBitVector TargetARM32::RegisterAliases[RegARM32::Reg_NUM];
 
 } // end of namespace ARM32
 } // end of namespace Ice