[LV][ARM] Improve reduction costmodel for mismatching extension types.

[dmgreen]

Summary

Given a MLA reduction from two different types (say i8 and i16), we were previously failing to find the reduction pattern, often making us chose the lower vector factor. This improves that by using the largest of the two extension types, allowing us to use the larger VF as the type of the reduction.

As per https://godbolt.org/z/KP549EEYM the backend handles this valiantly, leading to better performance.

[sdesmalen]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

7086

Is there any rationale for using the largest instead of the smallest type here? And would changing it to the smallest change the result of the test?

[dmgreen]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

7086

Yeah - the largest type is the correct one to use. We end up with, for example, add(mul(ext(A), ext(ext(B))), where add(mul(ext, ext) is the reduction pattern. There is then an extra ext - I haven't added the cost of that, but we might as well. It will often be free or cheap under MVE in reality, but I don't think it should be a problem to add it, so long as the correct vector factor is chosen.

[dmgreen]

Update with extra ext cost.

[sdesmalen]

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

7106

From what I can see, these two are mutually exclusive, i.e. (Op0Ty != LargestOpTy && Op1Ty != LargestOpTy) <=> false. Can you rewrite this as:

Value *ExtraExtOp = (Op0Ty != LargestOpTy) ? Op0 : Op1;
InstructionCost ExtraExtCost =
    TTI.getCastInstrCost(ExtraExtOp->getOpcode(), ExtType,
                         VectorType::get(ExtraExtOp->getType(), VectorTy),
                         TTI::CastContextHint::None, CostKind, ExtraExtOp);

[dmgreen]

Thanks for the suggestion. Updated to have a single getCastInstrCost.

[sdesmalen]

Thanks, LGTM!

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

InstructionCost ExtCost =
         TTI.getCastInstrCost(RedOp->getOpcode(), VectorTy, ExtType,
                              TTI::CastContextHint::None, CostKind, RedOp);
     if (RedCost.isValid() && RedCost < BaseCost + ExtCost)
       return I == RetI ? RedCost : 0;
   } else if (RedOp &&
              match(RedOp, m_Mul(m_Instruction(Op0), m_Instruction(Op1)))) {
     if (match(Op0, m_ZExtOrSExt(m_Value())) &&
         Op0->getOpcode() == Op1->getOpcode() &&
-        Op0->getOperand(0)->getType() == Op1->getOperand(0)->getType() &&
         !TheLoop->isLoopInvariant(Op0) && !TheLoop->isLoopInvariant(Op1)) {
       bool IsUnsigned = isa<ZExtInst>(Op0);
-      auto *ExtType = VectorType::get(Op0->getOperand(0)->getType(), VectorTy);
+      Type *Op0Ty = Op0->getOperand(0)->getType();
+      Type *Op1Ty = Op1->getOperand(0)->getType();
+      Type *LargestOpTy =
+          Op0Ty->getIntegerBitWidth() < Op1Ty->getIntegerBitWidth() ? Op1Ty
+                                                                    : Op0Ty;
+      auto *ExtType = VectorType::get(LargestOpTy, VectorTy);
       // Matched reduce(mul(ext, ext))
       InstructionCost ExtCost =
           TTI.getCastInstrCost(Op0->getOpcode(), VectorTy, ExtType,
                                TTI::CastContextHint::None, CostKind, Op0);
       InstructionCost MulCost =
           TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy, CostKind);
 
       InstructionCost RedCost = TTI.getExtendedAddReductionCost(
           /*IsMLA=*/true, IsUnsigned, RdxDesc.getRecurrenceType(), ExtType,
           CostKind);
 
       if (RedCost.isValid() && RedCost < ExtCost * 2 + MulCost + BaseCost)
         return I == RetI ? RedCost : 0;
     } else if (!match(I, m_ZExtOrSExt(m_Value()))) {
       // Matched reduce(mul())
       InstructionCost MulCost =
           TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy, CostKind);
 
       InstructionCost RedCost = TTI.getExtendedAddReductionCost(
           /*IsMLA=*/true, true, RdxDesc.getRecurrenceType(), VectorTy,
           CostKind);
 
       if (RedCost.isValid() && RedCost < MulCost + BaseCost)
         return I == RetI ? RedCost : 0;
     }

llvm/test/Transforms/LoopVectorize/ARM/mve-reductions.ll

for.body:                                         ; preds = %entry, %for.body
   %exitcond = icmp eq i32 %inc, %n
   br i1 %exitcond, label %for.cond.cleanup, label %for.body
 
 for.cond.cleanup:                                 ; preds = %for.body, %entry
   %r.0.lcssa = phi i8 [ 0, %entry ], [ %add, %for.body ]
   ret i8 %r.0.lcssa
 }
 
-; 4x or 8x as different types
+; 8x as different types
 define i32 @red_mla_ext_s8_s16_s32(i8* noalias nocapture readonly %A, i16* noalias nocapture readonly %B, i32 %n) #0 {
 ; CHECK-LABEL: @red_mla_ext_s8_s16_s32(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    [[CMP9_NOT:%.*]] = icmp eq i32 [[N:%.*]], 0
 ; CHECK-NEXT:    br i1 [[CMP9_NOT]], label [[FOR_COND_CLEANUP:%.*]], label [[VECTOR_PH:%.*]]
 ; CHECK:       vector.ph:
-; CHECK-NEXT:    [[N_RND_UP:%.*]] = add i32 [[N]], 3
-; CHECK-NEXT:    [[N_VEC:%.*]] = and i32 [[N_RND_UP]], -4
+; CHECK-NEXT:    [[N_RND_UP:%.*]] = add i32 [[N]], 7
+; CHECK-NEXT:    [[N_VEC:%.*]] = and i32 [[N_RND_UP]], -8
 ; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
 ; CHECK:       vector.body:
 ; CHECK-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
 ; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[TMP9:%.*]], [[VECTOR_BODY]] ]
-; CHECK-NEXT:    [[ACTIVE_LANE_MASK:%.*]] = call <4 x i1> @llvm.get.active.lane.mask.v4i1.i32(i32 [[INDEX]], i32 [[N]])
+; CHECK-NEXT:    [[ACTIVE_LANE_MASK:%.*]] = call <8 x i1> @llvm.get.active.lane.mask.v8i1.i32(i32 [[INDEX]], i32 [[N]])
 ; CHECK-NEXT:    [[TMP0:%.*]] = getelementptr inbounds i8, i8* [[A:%.*]], i32 [[INDEX]]
-; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i8* [[TMP0]] to <4 x i8>*
-; CHECK-NEXT:    [[WIDE_MASKED_LOAD:%.*]] = call <4 x i8> @llvm.masked.load.v4i8.p0v4i8(<4 x i8>* [[TMP1]], i32 1, <4 x i1> [[ACTIVE_LANE_MASK]], <4 x i8> poison)
-; CHECK-NEXT:    [[TMP2:%.*]] = sext <4 x i8> [[WIDE_MASKED_LOAD]] to <4 x i32>
+; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i8* [[TMP0]] to <8 x i8>*
+; CHECK-NEXT:    [[WIDE_MASKED_LOAD:%.*]] = call <8 x i8> @llvm.masked.load.v8i8.p0v8i8(<8 x i8>* [[TMP1]], i32 1, <8 x i1> [[ACTIVE_LANE_MASK]], <8 x i8> poison)
+; CHECK-NEXT:    [[TMP2:%.*]] = sext <8 x i8> [[WIDE_MASKED_LOAD]] to <8 x i32>
 ; CHECK-NEXT:    [[TMP3:%.*]] = getelementptr inbounds i16, i16* [[B:%.*]], i32 [[INDEX]]
-; CHECK-NEXT:    [[TMP4:%.*]] = bitcast i16* [[TMP3]] to <4 x i16>*
-; CHECK-NEXT:    [[WIDE_MASKED_LOAD1:%.*]] = call <4 x i16> @llvm.masked.load.v4i16.p0v4i16(<4 x i16>* [[TMP4]], i32 2, <4 x i1> [[ACTIVE_LANE_MASK]], <4 x i16> poison)
-; CHECK-NEXT:    [[TMP5:%.*]] = sext <4 x i16> [[WIDE_MASKED_LOAD1]] to <4 x i32>
-; CHECK-NEXT:    [[TMP6:%.*]] = mul nsw <4 x i32> [[TMP5]], [[TMP2]]
-; CHECK-NEXT:    [[TMP7:%.*]] = select <4 x i1> [[ACTIVE_LANE_MASK]], <4 x i32> [[TMP6]], <4 x i32> zeroinitializer
-; CHECK-NEXT:    [[TMP8:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[TMP7]])
+; CHECK-NEXT:    [[TMP4:%.*]] = bitcast i16* [[TMP3]] to <8 x i16>*
+; CHECK-NEXT:    [[WIDE_MASKED_LOAD1:%.*]] = call <8 x i16> @llvm.masked.load.v8i16.p0v8i16(<8 x i16>* [[TMP4]], i32 2, <8 x i1> [[ACTIVE_LANE_MASK]], <8 x i16> poison)
+; CHECK-NEXT:    [[TMP5:%.*]] = sext <8 x i16> [[WIDE_MASKED_LOAD1]] to <8 x i32>
+; CHECK-NEXT:    [[TMP6:%.*]] = mul nsw <8 x i32> [[TMP5]], [[TMP2]]
+; CHECK-NEXT:    [[TMP7:%.*]] = select <8 x i1> [[ACTIVE_LANE_MASK]], <8 x i32> [[TMP6]], <8 x i32> zeroinitializer
+; CHECK-NEXT:    [[TMP8:%.*]] = call i32 @llvm.vector.reduce.add.v8i32(<8 x i32> [[TMP7]])
 ; CHECK-NEXT:    [[TMP9]] = add i32 [[TMP8]], [[VEC_PHI]]
-; CHECK-NEXT:    [[INDEX_NEXT]] = add i32 [[INDEX]], 4
+; CHECK-NEXT:    [[INDEX_NEXT]] = add i32 [[INDEX]], 8
 ; CHECK-NEXT:    [[TMP10:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]
 ; CHECK-NEXT:    br i1 [[TMP10]], label [[FOR_COND_CLEANUP]], label [[VECTOR_BODY]], !llvm.loop [[LOOP26:![0-9]+]]
 ; CHECK:       for.cond.cleanup:
 ; CHECK-NEXT:    [[S_0_LCSSA:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[TMP9]], [[VECTOR_BODY]] ]
 ; CHECK-NEXT:    ret i32 [[S_0_LCSSA]]
 ;
 entry:
   %cmp9.not = icmp eq i32 %n, 0