[SLP][RISCV] Implement unsigned getMinVectorRegisterBitWidth() for RISCV

[luke957]

Summary

Implement method unsigned getMinVectorRegisterBitWidth() for RISCV

[luke957]

Should we implement method unsigned getMinVectorRegisterBitWidth() to avoid potential errors?

[craig.topper]

Tests?

[craig.topper]

Is this function just picking the minimum size of vector SLP will use? Do we want that to follow getRegisterBitWidth() or should it be allowed to be lower?

[luke957]

In D108973#2974682, @craig.topper wrote:

Tests?

Em, I haven't got a proper test case for this. Without this methed being implemented, the mothed in super class will be called and got a return value of 128.

In D108973#2974808, @craig.topper wrote:

Is this function just picking the minimum size of vector SLP will use? Do we want that to follow getRegisterBitWidth() or should it be allowed to be lower?

Oh, I get your point. For RVV, we can use a vector registor fractionally, so minimum size of vector could be less than MinVectorRegisterBitWidth. But I'm not sure whether it is appropriate for a methed naming getMinVectorRegisterBitWidth.

[craig.topper]

In D108973#2976093, @luke957 wrote:

In D108973#2974682, @craig.topper wrote:

Tests?

Em, I haven't got a proper test case for this. Without this methed being implemented, the mothed in super class will be called and got a return value of 128.

In D108973#2974808, @craig.topper wrote:

Is this function just picking the minimum size of vector SLP will use? Do we want that to follow getRegisterBitWidth() or should it be allowed to be lower?

Oh, I get your point. For RVV, we can use a vector registor fractionally, so minimum size of vector could be less than MinVectorRegisterBitWidth. But I'm not sure whether it is appropriate for a methed naming getMinVectorRegisterBitWidth.

There was a vague suggestion about possibly adding a new getMinVectorizationBitWidth mentioned here https://reviews.llvm.org/D103925

Anyway this patch is probably good as a starting point. LGTM

[luke957]

In D108973#2976094, @craig.topper wrote:

In D108973#2976093, @luke957 wrote:

In D108973#2974682, @craig.topper wrote:

Tests?

Em, I haven't got a proper test case for this. Without this methed being implemented, the mothed in super class will be called and got a return value of 128.

In D108973#2974808, @craig.topper wrote:

Is this function just picking the minimum size of vector SLP will use? Do we want that to follow getRegisterBitWidth() or should it be allowed to be lower?

Oh, I get your point. For RVV, we can use a vector registor fractionally, so minimum size of vector could be less than MinVectorRegisterBitWidth. But I'm not sure whether it is appropriate for a methed naming getMinVectorRegisterBitWidth.

There was a vague suggestion about possibly adding a new getMinVectorizationBitWidth mentioned here https://reviews.llvm.org/D103925

Anyway this patch is probably good as a starting point. LGTM

Thanks. I'll look into this patch.

llvm/lib/Target/RISCV/RISCVTargetTransformInfo.h

TypeSize getRegisterBitWidth(TargetTransformInfo::RegisterKind K) const {
     case TargetTransformInfo::RGK_ScalableVector:
       return TypeSize::getScalable(
           ST->hasStdExtV() ? RISCV::RVVBitsPerBlock : 0);
     }
 
     llvm_unreachable("Unsupported register kind");
   }
 
+  unsigned getMinVectorRegisterBitWidth() const {
+    return ST->hasStdExtV() ? ST->getMinRVVVectorSizeInBits() : 0;
+  }
+
   InstructionCost getGatherScatterOpCost(unsigned Opcode, Type *DataTy,
                                          const Value *Ptr, bool VariableMask,
                                          Align Alignment,
                                          TTI::TargetCostKind CostKind,
                                          const Instruction *I);
 
   bool isLegalElementTypeForRVV(Type *ScalarTy) const {
     if (ScalarTy->isPointerTy())