[NVPTX] declare no vector registers

[jingyue]

Summary

Without this patch, LoopVectorizer in certain cases (see loop-vectorize.ll)
produces code with complex control flow which hurts later optimizations. Since
NVPTX doesn't have vector registers in LLVM's sense
(NVPTXTTI::getRegisterBitWidth(true) == 32), we for now declare no vector
registers to effectively disable loop vectorization.

[jingyue]

Justin, I wonder why NVPTX doesn't leverage vector instructions (such as
vadd) at all. llc on

fadd <2 x float> %a, %b

gives me two add.f32 instead of vadd.f32 or like.

Jingyue

[jholewinski]

LGTM

[jholewinski]

The short answer is that ptxas doesn't handle vector registers very well. It may be good to revisit this, but ptxas currently prefers scalar ops.

Last time I looked into this, the implementation cost greatly outweighed any potential benefits. At the SASS level, we don't have vector fp ops anyway.

[jingyue]

Makes sense. Thank you.

llvm/trunk/lib/Target/NVPTX/NVPTXTargetTransformInfo.h

public:
   bool isSourceOfDivergence(const Value *V);
 
   unsigned getArithmeticInstrCost(
       unsigned Opcode, Type *Ty,
       TTI::OperandValueKind Opd1Info = TTI::OK_AnyValue,
       TTI::OperandValueKind Opd2Info = TTI::OK_AnyValue,
       TTI::OperandValueProperties Opd1PropInfo = TTI::OP_None,
       TTI::OperandValueProperties Opd2PropInfo = TTI::OP_None);
+
+  unsigned getNumberOfRegisters(bool Vector);
 };
 
 } // end namespace llvm
 
 #endif

llvm/trunk/lib/Target/NVPTX/NVPTXTargetTransformInfo.cpp

case ISD::AND:
     // those on types that can fit into one machine register.
     if (LT.second.SimpleTy == MVT::i64)
       return 2 * LT.first;
     // Delegate other cases to the basic TTI.
     return BaseT::getArithmeticInstrCost(Opcode, Ty, Opd1Info, Opd2Info,
                                          Opd1PropInfo, Opd2PropInfo);
   }
 }
+
+unsigned NVPTXTTIImpl::getNumberOfRegisters(bool Vector) {
+  if (Vector)
+    return 0;
+  return BaseT::getNumberOfRegisters(Vector);
+}

llvm/trunk/test/CodeGen/NVPTX/loop-vectorize.ll

+; RUN: opt < %s -O3 -S | FileCheck %s
+
+target datalayout = "e-i64:64-v16:16-v32:32-n16:32:64"
+target triple = "nvptx64-nvidia-cuda"
+
+define void @no_vectorization(i32 %n, i32 %a, i32 %b) {
+; CHECK-LABEL: no_vectorization(
+; CHECK-NOT: <4 x i32>
+; CHECK-NOT: <4 x i1>
+entry:
+  %cmp.5 = icmp sgt i32 %n, 0
+  br i1 %cmp.5, label %for.body.preheader, label %for.cond.cleanup
+
+for.body.preheader:                               ; preds = %entry
+  br label %for.body
+
+for.cond.cleanup.loopexit:                        ; preds = %for.body
+  br label %for.cond.cleanup
+
+for.cond.cleanup:                                 ; preds = %for.cond.cleanup.loopexit, %entry
+  ret void
+
+for.body:                                         ; preds = %for.body.preheader, %for.body
+  %i.06 = phi i32 [ %inc, %for.body ], [ 0, %for.body.preheader ]
+  %add = add nsw i32 %i.06, %a
+  %mul = mul nsw i32 %add, %b
+  %cmp1 = icmp sgt i32 %mul, -1
+  tail call void @llvm.assume(i1 %cmp1)
+  %inc = add nuw nsw i32 %i.06, 1
+  %exitcond = icmp eq i32 %inc, %n
+  br i1 %exitcond, label %for.cond.cleanup.loopexit, label %for.body
+}
+
+declare void @llvm.assume(i1) #0
+
+attributes #0 = { nounwind }
+
+!nvvm.annotations = !{!0}
+!0 = !{void (i32, i32, i32)* @no_vectorization, !"kernel", i32 1}