[TTI][X86] Add SSE2 sub-128bit vXi16/32 and v2i64 stride 2 interleaved load costs

[RKSimon]

Summary

These cases uses the same codegen as AVX2 (pshuflw/pshufd) for the sub-128bit vector deinterleaving, and unpcklqdq for v2i64.

It's going to take a while to add full interleaved cost coverage, but since these are the same for SSE2 -> AVX2 it should be an easy win.

Fixes PR47437

[lebedev.ri]

llvm/lib/Target/X86/X86TargetTransformInfo.cpp

5224

Looking at llvm-project/llvm/test/CodeGen/X86/vector-interleaved-load-i16-stride-2.ll,
VF4 codegen is really different between SSE2 and AVX2.

5230

Looking at llvm-project/llvm/test/CodeGen/X86/vector-interleaved-load-i32-stride-2.ll,
@load_i32_stride2_vf4 also seems to match.

[RKSimon]

llvm/lib/Target/X86/X86TargetTransformInfo.cpp

5224

nice catch!

5230

every little helps :)

[RKSimon]

Address review comment

[lebedev.ri]

lg

llvm/lib/Target/X86/X86TargetTransformInfo.cpp

InstructionCost X86TTIImpl::getInterleavedMemoryOpCost(
   //
   static const CostTblEntry AVX2InterleavedLoadTbl[] = {
       {2, MVT::v2i8, 2},  // (load 4i8 and) deinterleave into 2 x 2i8
       {2, MVT::v4i8, 2},  // (load 8i8 and) deinterleave into 2 x 4i8
       {2, MVT::v8i8, 2},  // (load 16i8 and) deinterleave into 2 x 8i8
       {2, MVT::v16i8, 4}, // (load 32i8 and) deinterleave into 2 x 16i8
       {2, MVT::v32i8, 6}, // (load 64i8 and) deinterleave into 2 x 32i8
 
-      {2, MVT::v2i16, 2},   // (load 4i16 and) deinterleave into 2 x 2i16
       {2, MVT::v4i16, 2},   // (load 8i16 and) deinterleave into 2 x 4i16
       {2, MVT::v8i16, 6},   // (load 16i16 and) deinterleave into 2 x 8i16
       {2, MVT::v16i16, 9},  // (load 32i16 and) deinterleave into 2 x 16i16
       {2, MVT::v32i16, 18}, // (load 64i16 and) deinterleave into 2 x 32i16
 
-      {2, MVT::v2i32, 2},   // (load 4i32 and) deinterleave into 2 x 2i32
-      {2, MVT::v4i32, 2},   // (load 8i32 and) deinterleave into 2 x 4i32
       {2, MVT::v8i32, 4},   // (load 16i32 and) deinterleave into 2 x 8i32
       {2, MVT::v16i32, 8},  // (load 32i32 and) deinterleave into 2 x 16i32
       {2, MVT::v32i32, 16}, // (load 64i32 and) deinterleave into 2 x 32i32
 
-      {2, MVT::v2i64, 2},   // (load 4i64 and) deinterleave into 2 x 2i64
       {2, MVT::v4i64, 4},   // (load 8i64 and) deinterleave into 2 x 4i64
       {2, MVT::v8i64, 8},   // (load 16i64 and) deinterleave into 2 x 8i64
       {2, MVT::v16i64, 16}, // (load 32i64 and) deinterleave into 2 x 16i64
 
       {3, MVT::v2i8, 3},   // (load 6i8 and) deinterleave into 3 x 2i8
       {3, MVT::v4i8, 3},   // (load 12i8 and) deinterleave into 3 x 4i8
       {3, MVT::v8i8, 6},   // (load 24i8 and) deinterleave into 3 x 8i8
       {3, MVT::v16i8, 11}, // (load 48i8 and) deinterleave into 3 x 16i8
static const CostTblEntry AVX2InterleavedLoadTbl[] = {
 
       {6, MVT::v2i64, 6},  // (load 12i64 and) deinterleave into 6 x 2i64
       {6, MVT::v4i64, 18}, // (load 24i64 and) deinterleave into 6 x 4i64
       {6, MVT::v8i64, 36}, // (load 48i64 and) deinterleave into 6 x 8i64
 
       {8, MVT::v8i32, 40} // (load 64i32 and) deinterleave into 8 x 8i32
   };
 
+  static const CostTblEntry SSE2InterleavedLoadTbl[] = {
+      {2, MVT::v2i16, 2},   // (load 4i16 and) deinterleave into 2 x 2i16
+
+      {2, MVT::v2i32, 2},   // (load 4i32 and) deinterleave into 2 x 2i32
+      {2, MVT::v4i32, 2},   // (load 8i32 and) deinterleave into 2 x 4i32
+
+      {2, MVT::v2i64, 2},   // (load 4i64 and) deinterleave into 2 x 2i64
+  };
+
   static const CostTblEntry AVX2InterleavedStoreTbl[] = {
       {2, MVT::v2i8, 1},  // interleave 2 x 2i8 into 4i8 (and store)
       {2, MVT::v4i8, 1},  // interleave 2 x 4i8 into 8i8 (and store)
       {2, MVT::v8i8, 1},  // interleave 2 x 8i8 into 16i8 (and store)
       {2, MVT::v16i8, 3}, // interleave 2 x 16i8 into 32i8 (and store)
       {2, MVT::v32i8, 4}, // interleave 2 x 32i8 into 64i8 (and store)
 
       {2, MVT::v2i16, 1},  // interleave 2 x 2i16 into 4i16 (and store)
InstructionCost X86TTIImpl::getInterleavedMemoryOpCost(
 
   if (Opcode == Instruction::Load) {
     // FIXME: if we have a partially-interleaved groups, with gaps,
     //        should we discount the not-demanded indicies?
     if (ST->hasAVX2())
       if (const auto *Entry = CostTableLookup(AVX2InterleavedLoadTbl, Factor,
                                               ETy.getSimpleVT()))
         return MemOpCosts + Entry->Cost;
+
+    if (ST->hasSSE2())
+      if (const auto *Entry = CostTableLookup(SSE2InterleavedLoadTbl, Factor,
+                                              ETy.getSimpleVT()))
+        return MemOpCosts + Entry->Cost;
   } else {
     assert(Opcode == Instruction::Store &&
            "Expected Store Instruction at this point");
     assert((!Indices.size() || Indices.size() == Factor) &&
            "Interleaved store only supports fully-interleaved groups.");
     if (ST->hasAVX2())
       if (const auto *Entry = CostTableLookup(AVX2InterleavedStoreTbl, Factor,
                                               ETy.getSimpleVT()))
         return MemOpCosts + Entry->Cost;
   }
 
   return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
                                            Alignment, AddressSpace, CostKind,
                                            UseMaskForCond, UseMaskForGaps);
 }

llvm/test/Analysis/CostModel/X86/interleaved-load-f32-stride-2.ll

 ; RUN: opt -loop-vectorize -vectorizer-maximize-bandwidth -S -mattr=+sse2 --debug-only=loop-vectorize < %s 2>&1 | FileCheck %s --check-prefixes=CHECK,SSE2
 ; RUN: opt -loop-vectorize -vectorizer-maximize-bandwidth -S -mattr=+avx  --debug-only=loop-vectorize < %s 2>&1 | FileCheck %s --check-prefixes=CHECK,AVX1
 ; RUN: opt -loop-vectorize -vectorizer-maximize-bandwidth -S -mattr=+avx2 --debug-only=loop-vectorize < %s 2>&1 | FileCheck %s --check-prefixes=CHECK,AVX2
 ; RUN: opt -loop-vectorize -vectorizer-maximize-bandwidth -S -mattr=+avx512bw,+avx512vl --debug-only=loop-vectorize < %s 2>&1 | FileCheck %s --check-prefixes=CHECK,AVX512
 ; REQUIRES: asserts
 
 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
 target triple = "x86_64-unknown-linux-gnu"
 
 @A = global [1024 x float] zeroinitializer, align 128
 @B = global [1024 x i8] zeroinitializer, align 128
 
 ; CHECK: LV: Checking a loop in "test"
 ;
 ; SSE2: LV: Found an estimated cost of 1 for VF 1 For instruction:   %v0 = load float, float* %in0, align 4
-; SSE2: LV: Found an estimated cost of 6 for VF 2 For instruction:   %v0 = load float, float* %in0, align 4
-; SSE2: LV: Found an estimated cost of 14 for VF 4 For instruction:   %v0 = load float, float* %in0, align 4
+; SSE2: LV: Found an estimated cost of 3 for VF 2 For instruction:   %v0 = load float, float* %in0, align 4
+; SSE2: LV: Found an estimated cost of 4 for VF 4 For instruction:   %v0 = load float, float* %in0, align 4
 ; SSE2: LV: Found an estimated cost of 28 for VF 8 For instruction:   %v0 = load float, float* %in0, align 4
 ; SSE2: LV: Found an estimated cost of 56 for VF 16 For instruction:   %v0 = load float, float* %in0, align 4
 ;
 ; AVX1: LV: Found an estimated cost of 1 for VF 1 For instruction:   %v0 = load float, float* %in0, align 4
-; AVX1: LV: Found an estimated cost of 6 for VF 2 For instruction:   %v0 = load float, float* %in0, align 4
-; AVX1: LV: Found an estimated cost of 17 for VF 4 For instruction:   %v0 = load float, float* %in0, align 4
+; AVX1: LV: Found an estimated cost of 3 for VF 2 For instruction:   %v0 = load float, float* %in0, align 4
+; AVX1: LV: Found an estimated cost of 3 for VF 4 For instruction:   %v0 = load float, float* %in0, align 4
 ; AVX1: LV: Found an estimated cost of 38 for VF 8 For instruction:   %v0 = load float, float* %in0, align 4
 ; AVX1: LV: Found an estimated cost of 76 for VF 16 For instruction:   %v0 = load float, float* %in0, align 4
 ; AVX1: LV: Found an estimated cost of 152 for VF 32 For instruction:   %v0 = load float, float* %in0, align 4
 ;;
 ; AVX2: LV: Found an estimated cost of 1 for VF 1 For instruction:   %v0 = load float, float* %in0, align 4
 ; AVX2: LV: Found an estimated cost of 3 for VF 2 For instruction:   %v0 = load float, float* %in0, align 4
 ; AVX2: LV: Found an estimated cost of 3 for VF 4 For instruction:   %v0 = load float, float* %in0, align 4
 ; AVX2: LV: Found an estimated cost of 6 for VF 8 For instruction:   %v0 = load float, float* %in0, align 4

llvm/test/Analysis/CostModel/X86/interleaved-load-f64-stride-2.ll

 ; RUN: opt -loop-vectorize -vectorizer-maximize-bandwidth -S -mattr=+sse2 --debug-only=loop-vectorize < %s 2>&1 | FileCheck %s --check-prefixes=CHECK,SSE2
 ; RUN: opt -loop-vectorize -vectorizer-maximize-bandwidth -S -mattr=+avx  --debug-only=loop-vectorize < %s 2>&1 | FileCheck %s --check-prefixes=CHECK,AVX1
 ; RUN: opt -loop-vectorize -vectorizer-maximize-bandwidth -S -mattr=+avx2 --debug-only=loop-vectorize < %s 2>&1 | FileCheck %s --check-prefixes=CHECK,AVX2
 ; RUN: opt -loop-vectorize -vectorizer-maximize-bandwidth -S -mattr=+avx512bw,+avx512vl --debug-only=loop-vectorize < %s 2>&1 | FileCheck %s --check-prefixes=CHECK,AVX512
 ; REQUIRES: asserts
 
 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
 target triple = "x86_64-unknown-linux-gnu"
 
 @A = global [1024 x double] zeroinitializer, align 128
 @B = global [1024 x i8] zeroinitializer, align 128
 
 ; CHECK: LV: Checking a loop in "test"
 ;
 ; SSE2: LV: Found an estimated cost of 1 for VF 1 For instruction:   %v0 = load double, double* %in0, align 8
-; SSE2: LV: Found an estimated cost of 6 for VF 2 For instruction:   %v0 = load double, double* %in0, align 8
+; SSE2: LV: Found an estimated cost of 4 for VF 2 For instruction:   %v0 = load double, double* %in0, align 8
 ; SSE2: LV: Found an estimated cost of 12 for VF 4 For instruction:   %v0 = load double, double* %in0, align 8
 ; SSE2: LV: Found an estimated cost of 24 for VF 8 For instruction:   %v0 = load double, double* %in0, align 8
 ; SSE2: LV: Found an estimated cost of 48 for VF 16 For instruction:   %v0 = load double, double* %in0, align 8
 ;
 ; AVX1: LV: Found an estimated cost of 1 for VF 1 For instruction:   %v0 = load double, double* %in0, align 8
-; AVX1: LV: Found an estimated cost of 7 for VF 2 For instruction:   %v0 = load double, double* %in0, align 8
+; AVX1: LV: Found an estimated cost of 3 for VF 2 For instruction:   %v0 = load double, double* %in0, align 8
 ; AVX1: LV: Found an estimated cost of 16 for VF 4 For instruction:   %v0 = load double, double* %in0, align 8
 ; AVX1: LV: Found an estimated cost of 32 for VF 8 For instruction:   %v0 = load double, double* %in0, align 8
 ; AVX1: LV: Found an estimated cost of 64 for VF 16 For instruction:   %v0 = load double, double* %in0, align 8
 ; AVX1: LV: Found an estimated cost of 128 for VF 32 For instruction:   %v0 = load double, double* %in0, align 8
 ;;
 ; AVX2: LV: Found an estimated cost of 1 for VF 1 For instruction:   %v0 = load double, double* %in0, align 8
 ; AVX2: LV: Found an estimated cost of 3 for VF 2 For instruction:   %v0 = load double, double* %in0, align 8
 ; AVX2: LV: Found an estimated cost of 6 for VF 4 For instruction:   %v0 = load double, double* %in0, align 8

llvm/test/Analysis/CostModel/X86/interleaved-load-i16-stride-2.ll

 ; RUN: opt -loop-vectorize -vectorizer-maximize-bandwidth -S -mattr=+sse2 --debug-only=loop-vectorize < %s 2>&1 | FileCheck %s --check-prefixes=CHECK,SSE2
 ; RUN: opt -loop-vectorize -vectorizer-maximize-bandwidth -S -mattr=+avx  --debug-only=loop-vectorize < %s 2>&1 | FileCheck %s --check-prefixes=CHECK,AVX1
 ; RUN: opt -loop-vectorize -vectorizer-maximize-bandwidth -S -mattr=+avx2 --debug-only=loop-vectorize < %s 2>&1 | FileCheck %s --check-prefixes=CHECK,AVX2
 ; RUN: opt -loop-vectorize -vectorizer-maximize-bandwidth -S -mattr=+avx512bw,+avx512vl --debug-only=loop-vectorize < %s 2>&1 | FileCheck %s --check-prefixes=CHECK,AVX512
 ; REQUIRES: asserts
 
 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
 target triple = "x86_64-unknown-linux-gnu"
 
 @A = global [1024 x i16] zeroinitializer, align 128
 @B = global [1024 x i8] zeroinitializer, align 128
 
 ; CHECK: LV: Checking a loop in "test"
 ;
 ; SSE2: LV: Found an estimated cost of 1 for VF 1 For instruction:   %v0 = load i16, i16* %in0, align 2
-; SSE2: LV: Found an estimated cost of 9 for VF 2 For instruction:   %v0 = load i16, i16* %in0, align 2
+; SSE2: LV: Found an estimated cost of 3 for VF 2 For instruction:   %v0 = load i16, i16* %in0, align 2
 ; SSE2: LV: Found an estimated cost of 17 for VF 4 For instruction:   %v0 = load i16, i16* %in0, align 2
 ; SSE2: LV: Found an estimated cost of 34 for VF 8 For instruction:   %v0 = load i16, i16* %in0, align 2
 ; SSE2: LV: Found an estimated cost of 68 for VF 16 For instruction:   %v0 = load i16, i16* %in0, align 2
 ;
 ; AVX1: LV: Found an estimated cost of 1 for VF 1 For instruction:   %v0 = load i16, i16* %in0, align 2
-; AVX1: LV: Found an estimated cost of 9 for VF 2 For instruction:   %v0 = load i16, i16* %in0, align 2
+; AVX1: LV: Found an estimated cost of 3 for VF 2 For instruction:   %v0 = load i16, i16* %in0, align 2
 ; AVX1: LV: Found an estimated cost of 17 for VF 4 For instruction:   %v0 = load i16, i16* %in0, align 2
 ; AVX1: LV: Found an estimated cost of 41 for VF 8 For instruction:   %v0 = load i16, i16* %in0, align 2
 ; AVX1: LV: Found an estimated cost of 86 for VF 16 For instruction:   %v0 = load i16, i16* %in0, align 2
 ; AVX1: LV: Found an estimated cost of 172 for VF 32 For instruction:   %v0 = load i16, i16* %in0, align 2
 ;
 ; AVX2: LV: Found an estimated cost of 1 for VF 1 For instruction:   %v0 = load i16, i16* %in0, align 2
 ; AVX2: LV: Found an estimated cost of 3 for VF 2 For instruction:   %v0 = load i16, i16* %in0, align 2
 ; AVX2: LV: Found an estimated cost of 3 for VF 4 For instruction:   %v0 = load i16, i16* %in0, align 2

llvm/test/Analysis/CostModel/X86/interleaved-load-i32-stride-2-indices-0u.ll

 ; RUN: opt -loop-vectorize -vectorizer-maximize-bandwidth -S -mattr=+sse2 --debug-only=loop-vectorize < %s 2>&1 | FileCheck %s --check-prefixes=CHECK,SSE2
 ; RUN: opt -loop-vectorize -vectorizer-maximize-bandwidth -S -mattr=+avx  --debug-only=loop-vectorize < %s 2>&1 | FileCheck %s --check-prefixes=CHECK,AVX1
 ; RUN: opt -loop-vectorize -vectorizer-maximize-bandwidth -S -mattr=+avx2 --debug-only=loop-vectorize < %s 2>&1 | FileCheck %s --check-prefixes=CHECK,AVX2
 ; RUN: opt -loop-vectorize -vectorizer-maximize-bandwidth -S -mattr=+avx512bw,+avx512vl --debug-only=loop-vectorize < %s 2>&1 | FileCheck %s --check-prefixes=CHECK,AVX512
 ; REQUIRES: asserts
 
 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
 target triple = "x86_64-unknown-linux-gnu"
 
 @A = global [1024 x i32] zeroinitializer, align 128
 @B = global [1024 x i8] zeroinitializer, align 128
 
 ; CHECK: LV: Checking a loop in "test"
 ;
 ; SSE2: LV: Found an estimated cost of 1 for VF 1 For instruction:   %v0 = load i32, i32* %in0, align 4
-; SSE2: LV: Found an estimated cost of 7 for VF 2 For instruction:   %v0 = load i32, i32* %in0, align 4
-; SSE2: LV: Found an estimated cost of 15 for VF 4 For instruction:   %v0 = load i32, i32* %in0, align 4
+; SSE2: LV: Found an estimated cost of 3 for VF 2 For instruction:   %v0 = load i32, i32* %in0, align 4
+; SSE2: LV: Found an estimated cost of 4 for VF 4 For instruction:   %v0 = load i32, i32* %in0, align 4
 ; SSE2: LV: Found an estimated cost of 30 for VF 8 For instruction:   %v0 = load i32, i32* %in0, align 4
 ; SSE2: LV: Found an estimated cost of 60 for VF 16 For instruction:   %v0 = load i32, i32* %in0, align 4
 ;
 ; AVX1: LV: Found an estimated cost of 1 for VF 1 For instruction:   %v0 = load i32, i32* %in0, align 4
-; AVX1: LV: Found an estimated cost of 5 for VF 2 For instruction:   %v0 = load i32, i32* %in0, align 4
-; AVX1: LV: Found an estimated cost of 11 for VF 4 For instruction:   %v0 = load i32, i32* %in0, align 4
+; AVX1: LV: Found an estimated cost of 3 for VF 2 For instruction:   %v0 = load i32, i32* %in0, align 4
+; AVX1: LV: Found an estimated cost of 3 for VF 4 For instruction:   %v0 = load i32, i32* %in0, align 4
 ; AVX1: LV: Found an estimated cost of 24 for VF 8 For instruction:   %v0 = load i32, i32* %in0, align 4
 ; AVX1: LV: Found an estimated cost of 48 for VF 16 For instruction:   %v0 = load i32, i32* %in0, align 4
 ; AVX1: LV: Found an estimated cost of 96 for VF 32 For instruction:   %v0 = load i32, i32* %in0, align 4
 ;
 ; AVX2: LV: Found an estimated cost of 1 for VF 1 For instruction:   %v0 = load i32, i32* %in0, align 4
 ; AVX2: LV: Found an estimated cost of 3 for VF 2 For instruction:   %v0 = load i32, i32* %in0, align 4
 ; AVX2: LV: Found an estimated cost of 3 for VF 4 For instruction:   %v0 = load i32, i32* %in0, align 4
 ; AVX2: LV: Found an estimated cost of 6 for VF 8 For instruction:   %v0 = load i32, i32* %in0, align 4

llvm/test/Analysis/CostModel/X86/interleaved-load-i32-stride-2.ll

 ; RUN: opt -loop-vectorize -vectorizer-maximize-bandwidth -S -mattr=+sse2 --debug-only=loop-vectorize < %s 2>&1 | FileCheck %s --check-prefixes=CHECK,SSE2
 ; RUN: opt -loop-vectorize -vectorizer-maximize-bandwidth -S -mattr=+avx  --debug-only=loop-vectorize < %s 2>&1 | FileCheck %s --check-prefixes=CHECK,AVX1
 ; RUN: opt -loop-vectorize -vectorizer-maximize-bandwidth -S -mattr=+avx2 --debug-only=loop-vectorize < %s 2>&1 | FileCheck %s --check-prefixes=CHECK,AVX2
 ; RUN: opt -loop-vectorize -vectorizer-maximize-bandwidth -S -mattr=+avx512bw,+avx512vl --debug-only=loop-vectorize < %s 2>&1 | FileCheck %s --check-prefixes=CHECK,AVX512
 ; REQUIRES: asserts
 
 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
 target triple = "x86_64-unknown-linux-gnu"
 
 @A = global [1024 x i32] zeroinitializer, align 128
 @B = global [1024 x i8] zeroinitializer, align 128
 
 ; CHECK: LV: Checking a loop in "test"
 ;
 ; SSE2: LV: Found an estimated cost of 1 for VF 1 For instruction:   %v0 = load i32, i32* %in0, align 4
-; SSE2: LV: Found an estimated cost of 14 for VF 2 For instruction:   %v0 = load i32, i32* %in0, align 4
-; SSE2: LV: Found an estimated cost of 30 for VF 4 For instruction:   %v0 = load i32, i32* %in0, align 4
+; SSE2: LV: Found an estimated cost of 3 for VF 2 For instruction:   %v0 = load i32, i32* %in0, align 4
+; SSE2: LV: Found an estimated cost of 4 for VF 4 For instruction:   %v0 = load i32, i32* %in0, align 4
 ; SSE2: LV: Found an estimated cost of 60 for VF 8 For instruction:   %v0 = load i32, i32* %in0, align 4
 ; SSE2: LV: Found an estimated cost of 120 for VF 16 For instruction:   %v0 = load i32, i32* %in0, align 4
 ;
 ; AVX1: LV: Found an estimated cost of 1 for VF 1 For instruction:   %v0 = load i32, i32* %in0, align 4
-; AVX1: LV: Found an estimated cost of 9 for VF 2 For instruction:   %v0 = load i32, i32* %in0, align 4
-; AVX1: LV: Found an estimated cost of 21 for VF 4 For instruction:   %v0 = load i32, i32* %in0, align 4
+; AVX1: LV: Found an estimated cost of 3 for VF 2 For instruction:   %v0 = load i32, i32* %in0, align 4
+; AVX1: LV: Found an estimated cost of 3 for VF 4 For instruction:   %v0 = load i32, i32* %in0, align 4
 ; AVX1: LV: Found an estimated cost of 46 for VF 8 For instruction:   %v0 = load i32, i32* %in0, align 4
 ; AVX1: LV: Found an estimated cost of 92 for VF 16 For instruction:   %v0 = load i32, i32* %in0, align 4
 ; AVX1: LV: Found an estimated cost of 184 for VF 32 For instruction:   %v0 = load i32, i32* %in0, align 4
 ;
 ; AVX2: LV: Found an estimated cost of 1 for VF 1 For instruction:   %v0 = load i32, i32* %in0, align 4
 ; AVX2: LV: Found an estimated cost of 3 for VF 2 For instruction:   %v0 = load i32, i32* %in0, align 4
 ; AVX2: LV: Found an estimated cost of 3 for VF 4 For instruction:   %v0 = load i32, i32* %in0, align 4
 ; AVX2: LV: Found an estimated cost of 6 for VF 8 For instruction:   %v0 = load i32, i32* %in0, align 4

llvm/test/Analysis/CostModel/X86/interleaved-load-i64-stride-2.ll

 ; RUN: opt -loop-vectorize -vectorizer-maximize-bandwidth -S -mattr=+sse2 --debug-only=loop-vectorize < %s 2>&1 | FileCheck %s --check-prefixes=CHECK,SSE2
 ; RUN: opt -loop-vectorize -vectorizer-maximize-bandwidth -S -mattr=+avx  --debug-only=loop-vectorize < %s 2>&1 | FileCheck %s --check-prefixes=CHECK,AVX1
 ; RUN: opt -loop-vectorize -vectorizer-maximize-bandwidth -S -mattr=+avx2 --debug-only=loop-vectorize < %s 2>&1 | FileCheck %s --check-prefixes=CHECK,AVX2
 ; RUN: opt -loop-vectorize -vectorizer-maximize-bandwidth -S -mattr=+avx512bw,+avx512vl --debug-only=loop-vectorize < %s 2>&1 | FileCheck %s --check-prefixes=CHECK,AVX512
 ; REQUIRES: asserts
 
 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
 target triple = "x86_64-unknown-linux-gnu"
 
 @A = global [1024 x i64] zeroinitializer, align 128
 @B = global [1024 x i8] zeroinitializer, align 128
 
 ; CHECK: LV: Checking a loop in "test"
 ;
 ; SSE2: LV: Found an estimated cost of 1 for VF 1 For instruction:   %v0 = load i64, i64* %in0, align 8
-; SSE2: LV: Found an estimated cost of 14 for VF 2 For instruction:   %v0 = load i64, i64* %in0, align 8
+; SSE2: LV: Found an estimated cost of 4 for VF 2 For instruction:   %v0 = load i64, i64* %in0, align 8
 ; SSE2: LV: Found an estimated cost of 28 for VF 4 For instruction:   %v0 = load i64, i64* %in0, align 8
 ; SSE2: LV: Found an estimated cost of 56 for VF 8 For instruction:   %v0 = load i64, i64* %in0, align 8
 ; SSE2: LV: Found an estimated cost of 112 for VF 16 For instruction:   %v0 = load i64, i64* %in0, align 8
 ;
 ; AVX1: LV: Found an estimated cost of 1 for VF 1 For instruction:   %v0 = load i64, i64* %in0, align 8
-; AVX1: LV: Found an estimated cost of 11 for VF 2 For instruction:   %v0 = load i64, i64* %in0, align 8
+; AVX1: LV: Found an estimated cost of 3 for VF 2 For instruction:   %v0 = load i64, i64* %in0, align 8
 ; AVX1: LV: Found an estimated cost of 26 for VF 4 For instruction:   %v0 = load i64, i64* %in0, align 8
 ; AVX1: LV: Found an estimated cost of 52 for VF 8 For instruction:   %v0 = load i64, i64* %in0, align 8
 ; AVX1: LV: Found an estimated cost of 104 for VF 16 For instruction:   %v0 = load i64, i64* %in0, align 8
 ; AVX1: LV: Found an estimated cost of 208 for VF 32 For instruction:   %v0 = load i64, i64* %in0, align 8
 ;
 ; AVX2: LV: Found an estimated cost of 1 for VF 1 For instruction:   %v0 = load i64, i64* %in0, align 8
 ; AVX2: LV: Found an estimated cost of 3 for VF 2 For instruction:   %v0 = load i64, i64* %in0, align 8
 ; AVX2: LV: Found an estimated cost of 6 for VF 4 For instruction:   %v0 = load i64, i64* %in0, align 8

llvm/test/Transforms/LoopVectorize/X86/interleaving.ll

 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
 ; RUN: opt -S -mtriple=x86_64-pc_linux -loop-vectorize -instcombine < %s | FileCheck %s --check-prefix=SSE
 ; RUN: opt -S -mtriple=x86_64-pc_linux -loop-vectorize -instcombine -mcpu=sandybridge < %s | FileCheck %s --check-prefix=AVX1
 ; RUN: opt -S -mtriple=x86_64-pc_linux -loop-vectorize -instcombine -mcpu=haswell < %s | FileCheck %s --check-prefix=AVX2
 ; RUN: opt -S -mtriple=x86_64-pc_linux -loop-vectorize -instcombine -mcpu=slm < %s | FileCheck %s --check-prefix=SSE
-; RUN: opt -S -mtriple=x86_64-pc_linux -loop-vectorize -instcombine -mcpu=atom < %s | FileCheck %s --check-prefix=SSE
+; RUN: opt -S -mtriple=x86_64-pc_linux -loop-vectorize -instcombine -mcpu=atom < %s | FileCheck %s --check-prefix=ATOM
 
 define void @foo(i32* noalias nocapture %a, i32* noalias nocapture readonly %b) {
 ; SSE-LABEL: @foo(
 ; SSE-NEXT:  entry:
+; SSE-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
+; SSE:       vector.ph:
+; SSE-NEXT:    br label [[VECTOR_BODY:%.*]]
+; SSE:       vector.body:
+; SSE-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
+; SSE-NEXT:    [[TMP0:%.*]] = shl nsw i64 [[INDEX]], 1
+; SSE-NEXT:    [[TMP1:%.*]] = shl i64 [[INDEX]], 1
+; SSE-NEXT:    [[TMP2:%.*]] = or i64 [[TMP1]], 8
+; SSE-NEXT:    [[TMP3:%.*]] = getelementptr inbounds i32, i32* [[B:%.*]], i64 [[TMP0]]
+; SSE-NEXT:    [[TMP4:%.*]] = getelementptr inbounds i32, i32* [[B]], i64 [[TMP2]]
+; SSE-NEXT:    [[TMP5:%.*]] = bitcast i32* [[TMP3]] to <8 x i32>*
+; SSE-NEXT:    [[TMP6:%.*]] = bitcast i32* [[TMP4]] to <8 x i32>*
+; SSE-NEXT:    [[WIDE_VEC:%.*]] = load <8 x i32>, <8 x i32>* [[TMP5]], align 4
+; SSE-NEXT:    [[WIDE_VEC1:%.*]] = load <8 x i32>, <8 x i32>* [[TMP6]], align 4
+; SSE-NEXT:    [[STRIDED_VEC:%.*]] = shufflevector <8 x i32> [[WIDE_VEC]], <8 x i32> poison, <4 x i32> <i32 0, i32 2, i32 4, i32 6>
+; SSE-NEXT:    [[STRIDED_VEC2:%.*]] = shufflevector <8 x i32> [[WIDE_VEC1]], <8 x i32> poison, <4 x i32> <i32 0, i32 2, i32 4, i32 6>
+; SSE-NEXT:    [[STRIDED_VEC3:%.*]] = shufflevector <8 x i32> [[WIDE_VEC]], <8 x i32> poison, <4 x i32> <i32 1, i32 3, i32 5, i32 7>
+; SSE-NEXT:    [[STRIDED_VEC4:%.*]] = shufflevector <8 x i32> [[WIDE_VEC1]], <8 x i32> poison, <4 x i32> <i32 1, i32 3, i32 5, i32 7>
+; SSE-NEXT:    [[TMP7:%.*]] = add nsw <4 x i32> [[STRIDED_VEC3]], [[STRIDED_VEC]]
+; SSE-NEXT:    [[TMP8:%.*]] = add nsw <4 x i32> [[STRIDED_VEC4]], [[STRIDED_VEC2]]
+; SSE-NEXT:    [[TMP9:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i64 [[INDEX]]
+; SSE-NEXT:    [[TMP10:%.*]] = bitcast i32* [[TMP9]] to <4 x i32>*
+; SSE-NEXT:    store <4 x i32> [[TMP7]], <4 x i32>* [[TMP10]], align 4
+; SSE-NEXT:    [[TMP11:%.*]] = getelementptr inbounds i32, i32* [[TMP9]], i64 4
+; SSE-NEXT:    [[TMP12:%.*]] = bitcast i32* [[TMP11]] to <4 x i32>*
+; SSE-NEXT:    store <4 x i32> [[TMP8]], <4 x i32>* [[TMP12]], align 4
+; SSE-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 8
+; SSE-NEXT:    [[TMP13:%.*]] = icmp eq i64 [[INDEX_NEXT]], 1024
+; SSE-NEXT:    br i1 [[TMP13]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
+; SSE:       middle.block:
+; SSE-NEXT:    br i1 true, label [[FOR_COND_CLEANUP:%.*]], label [[SCALAR_PH]]
+; SSE:       scalar.ph:
 ; SSE-NEXT:    br label [[FOR_BODY:%.*]]
 ; SSE:       for.cond.cleanup:
 ; SSE-NEXT:    ret void
 ; SSE:       for.body:
-; SSE-NEXT:    [[INDVARS_IV:%.*]] = phi i64 [ 0, [[ENTRY:%.*]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]
-; SSE-NEXT:    [[TMP0:%.*]] = shl nuw nsw i64 [[INDVARS_IV]], 1
-; SSE-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i32, i32* [[B:%.*]], i64 [[TMP0]]
-; SSE-NEXT:    [[TMP1:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
-; SSE-NEXT:    [[TMP2:%.*]] = or i64 [[TMP0]], 1
-; SSE-NEXT:    [[ARRAYIDX3:%.*]] = getelementptr inbounds i32, i32* [[B]], i64 [[TMP2]]
-; SSE-NEXT:    [[TMP3:%.*]] = load i32, i32* [[ARRAYIDX3]], align 4
-; SSE-NEXT:    [[ADD4:%.*]] = add nsw i32 [[TMP3]], [[TMP1]]
-; SSE-NEXT:    [[ARRAYIDX6:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i64 [[INDVARS_IV]]
-; SSE-NEXT:    store i32 [[ADD4]], i32* [[ARRAYIDX6]], align 4
-; SSE-NEXT:    [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
-; SSE-NEXT:    [[EXITCOND:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], 1024
-; SSE-NEXT:    br i1 [[EXITCOND]], label [[FOR_COND_CLEANUP:%.*]], label [[FOR_BODY]]
+; SSE-NEXT:    br i1 undef, label [[FOR_COND_CLEANUP]], label [[FOR_BODY]], !llvm.loop [[LOOP2:![0-9]+]]
 ;
 ; AVX1-LABEL: @foo(
 ; AVX1-NEXT:  entry:
 ; AVX1-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
 ; AVX1:       vector.ph:
 ; AVX1-NEXT:    br label [[VECTOR_BODY:%.*]]
 ; AVX1:       vector.body:
 ; AVX1-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
 ; AVX1-NEXT:    [[TMP0:%.*]] = shl nsw i64 [[INDEX]], 1
-; AVX1-NEXT:    [[TMP1:%.*]] = getelementptr inbounds i32, i32* [[B:%.*]], i64 [[TMP0]]
-; AVX1-NEXT:    [[TMP2:%.*]] = bitcast i32* [[TMP1]] to <8 x i32>*
-; AVX1-NEXT:    [[WIDE_VEC:%.*]] = load <8 x i32>, <8 x i32>* [[TMP2]], align 4
+; AVX1-NEXT:    [[TMP1:%.*]] = shl i64 [[INDEX]], 1
+; AVX1-NEXT:    [[TMP2:%.*]] = or i64 [[TMP1]], 8
+; AVX1-NEXT:    [[TMP3:%.*]] = shl i64 [[INDEX]], 1
+; AVX1-NEXT:    [[TMP4:%.*]] = or i64 [[TMP3]], 16
+; AVX1-NEXT:    [[TMP5:%.*]] = shl i64 [[INDEX]], 1
+; AVX1-NEXT:    [[TMP6:%.*]] = or i64 [[TMP5]], 24
+; AVX1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds i32, i32* [[B:%.*]], i64 [[TMP0]]
+; AVX1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds i32, i32* [[B]], i64 [[TMP2]]
+; AVX1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds i32, i32* [[B]], i64 [[TMP4]]
+; AVX1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds i32, i32* [[B]], i64 [[TMP6]]
+; AVX1-NEXT:    [[TMP11:%.*]] = bitcast i32* [[TMP7]] to <8 x i32>*
+; AVX1-NEXT:    [[TMP12:%.*]] = bitcast i32* [[TMP8]] to <8 x i32>*
+; AVX1-NEXT:    [[TMP13:%.*]] = bitcast i32* [[TMP9]] to <8 x i32>*
+; AVX1-NEXT:    [[TMP14:%.*]] = bitcast i32* [[TMP10]] to <8 x i32>*
+; AVX1-NEXT:    [[WIDE_VEC:%.*]] = load <8 x i32>, <8 x i32>* [[TMP11]], align 4
+; AVX1-NEXT:    [[WIDE_VEC1:%.*]] = load <8 x i32>, <8 x i32>* [[TMP12]], align 4
+; AVX1-NEXT:    [[WIDE_VEC2:%.*]] = load <8 x i32>, <8 x i32>* [[TMP13]], align 4
+; AVX1-NEXT:    [[WIDE_VEC3:%.*]] = load <8 x i32>, <8 x i32>* [[TMP14]], align 4
 ; AVX1-NEXT:    [[STRIDED_VEC:%.*]] = shufflevector <8 x i32> [[WIDE_VEC]], <8 x i32> poison, <4 x i32> <i32 0, i32 2, i32 4, i32 6>
-; AVX1-NEXT:    [[STRIDED_VEC1:%.*]] = shufflevector <8 x i32> [[WIDE_VEC]], <8 x i32> poison, <4 x i32> <i32 1, i32 3, i32 5, i32 7>
-; AVX1-NEXT:    [[TMP3:%.*]] = add nsw <4 x i32> [[STRIDED_VEC1]], [[STRIDED_VEC]]
-; AVX1-NEXT:    [[TMP4:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i64 [[INDEX]]
-; AVX1-NEXT:    [[TMP5:%.*]] = bitcast i32* [[TMP4]] to <4 x i32>*
-; AVX1-NEXT:    store <4 x i32> [[TMP3]], <4 x i32>* [[TMP5]], align 4
-; AVX1-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
-; AVX1-NEXT:    [[TMP6:%.*]] = icmp eq i64 [[INDEX_NEXT]], 1024
-; AVX1-NEXT:    br i1 [[TMP6]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
+; AVX1-NEXT:    [[STRIDED_VEC4:%.*]] = shufflevector <8 x i32> [[WIDE_VEC1]], <8 x i32> poison, <4 x i32> <i32 0, i32 2, i32 4, i32 6>
+; AVX1-NEXT:    [[STRIDED_VEC5:%.*]] = shufflevector <8 x i32> [[WIDE_VEC2]], <8 x i32> poison, <4 x i32> <i32 0, i32 2, i32 4, i32 6>
+; AVX1-NEXT:    [[STRIDED_VEC6:%.*]] = shufflevector <8 x i32> [[WIDE_VEC3]], <8 x i32> poison, <4 x i32> <i32 0, i32 2, i32 4, i32 6>
+; AVX1-NEXT:    [[STRIDED_VEC7:%.*]] = shufflevector <8 x i32> [[WIDE_VEC]], <8 x i32> poison, <4 x i32> <i32 1, i32 3, i32 5, i32 7>
+; AVX1-NEXT:    [[STRIDED_VEC8:%.*]] = shufflevector <8 x i32> [[WIDE_VEC1]], <8 x i32> poison, <4 x i32> <i32 1, i32 3, i32 5, i32 7>
+; AVX1-NEXT:    [[STRIDED_VEC9:%.*]] = shufflevector <8 x i32> [[WIDE_VEC2]], <8 x i32> poison, <4 x i32> <i32 1, i32 3, i32 5, i32 7>
+; AVX1-NEXT:    [[STRIDED_VEC10:%.*]] = shufflevector <8 x i32> [[WIDE_VEC3]], <8 x i32> poison, <4 x i32> <i32 1, i32 3, i32 5, i32 7>
+; AVX1-NEXT:    [[TMP15:%.*]] = add nsw <4 x i32> [[STRIDED_VEC7]], [[STRIDED_VEC]]
+; AVX1-NEXT:    [[TMP16:%.*]] = add nsw <4 x i32> [[STRIDED_VEC8]], [[STRIDED_VEC4]]
+; AVX1-NEXT:    [[TMP17:%.*]] = add nsw <4 x i32> [[STRIDED_VEC9]], [[STRIDED_VEC5]]
+; AVX1-NEXT:    [[TMP18:%.*]] = add nsw <4 x i32> [[STRIDED_VEC10]], [[STRIDED_VEC6]]
+; AVX1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i64 [[INDEX]]
+; AVX1-NEXT:    [[TMP20:%.*]] = bitcast i32* [[TMP19]] to <4 x i32>*
+; AVX1-NEXT:    store <4 x i32> [[TMP15]], <4 x i32>* [[TMP20]], align 4
+; AVX1-NEXT:    [[TMP21:%.*]] = getelementptr inbounds i32, i32* [[TMP19]], i64 4
+; AVX1-NEXT:    [[TMP22:%.*]] = bitcast i32* [[TMP21]] to <4 x i32>*
+; AVX1-NEXT:    store <4 x i32> [[TMP16]], <4 x i32>* [[TMP22]], align 4
+; AVX1-NEXT:    [[TMP23:%.*]] = getelementptr inbounds i32, i32* [[TMP19]], i64 8
+; AVX1-NEXT:    [[TMP24:%.*]] = bitcast i32* [[TMP23]] to <4 x i32>*
+; AVX1-NEXT:    store <4 x i32> [[TMP17]], <4 x i32>* [[TMP24]], align 4
+; AVX1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds i32, i32* [[TMP19]], i64 12
+; AVX1-NEXT:    [[TMP26:%.*]] = bitcast i32* [[TMP25]] to <4 x i32>*
+; AVX1-NEXT:    store <4 x i32> [[TMP18]], <4 x i32>* [[TMP26]], align 4
+; AVX1-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 16
+; AVX1-NEXT:    [[TMP27:%.*]] = icmp eq i64 [[INDEX_NEXT]], 1024
+; AVX1-NEXT:    br i1 [[TMP27]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
 ; AVX1:       middle.block:
 ; AVX1-NEXT:    br i1 true, label [[FOR_COND_CLEANUP:%.*]], label [[SCALAR_PH]]
 ; AVX1:       scalar.ph:
 ; AVX1-NEXT:    br label [[FOR_BODY:%.*]]
 ; AVX1:       for.cond.cleanup:
 ; AVX1-NEXT:    ret void
 ; AVX1:       for.body:
 ; AVX1-NEXT:    br i1 undef, label [[FOR_COND_CLEANUP]], label [[FOR_BODY]], !llvm.loop [[LOOP2:![0-9]+]]

 ; AVX2-NEXT:    br i1 true, label [[FOR_COND_CLEANUP:%.*]], label [[SCALAR_PH]]
 ; AVX2:       scalar.ph:
 ; AVX2-NEXT:    br label [[FOR_BODY:%.*]]
 ; AVX2:       for.cond.cleanup:
 ; AVX2-NEXT:    ret void
 ; AVX2:       for.body:
 ; AVX2-NEXT:    br i1 undef, label [[FOR_COND_CLEANUP]], label [[FOR_BODY]], !llvm.loop [[LOOP2:![0-9]+]]
 ;
+; ATOM-LABEL: @foo(
+; ATOM-NEXT:  entry:
+; ATOM-NEXT:    br label [[FOR_BODY:%.*]]
+; ATOM:       for.cond.cleanup:
+; ATOM-NEXT:    ret void
+; ATOM:       for.body:
+; ATOM-NEXT:    [[INDVARS_IV:%.*]] = phi i64 [ 0, [[ENTRY:%.*]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]
+; ATOM-NEXT:    [[TMP0:%.*]] = shl nuw nsw i64 [[INDVARS_IV]], 1
+; ATOM-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i32, i32* [[B:%.*]], i64 [[TMP0]]
+; ATOM-NEXT:    [[TMP1:%.*]] = load i32, i32* [[ARRAYIDX]], align 4
+; ATOM-NEXT:    [[TMP2:%.*]] = or i64 [[TMP0]], 1
+; ATOM-NEXT:    [[ARRAYIDX3:%.*]] = getelementptr inbounds i32, i32* [[B]], i64 [[TMP2]]
+; ATOM-NEXT:    [[TMP3:%.*]] = load i32, i32* [[ARRAYIDX3]], align 4
+; ATOM-NEXT:    [[ADD4:%.*]] = add nsw i32 [[TMP3]], [[TMP1]]
+; ATOM-NEXT:    [[ARRAYIDX6:%.*]] = getelementptr inbounds i32, i32* [[A:%.*]], i64 [[INDVARS_IV]]
+; ATOM-NEXT:    store i32 [[ADD4]], i32* [[ARRAYIDX6]], align 4
+; ATOM-NEXT:    [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
+; ATOM-NEXT:    [[EXITCOND:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], 1024
+; ATOM-NEXT:    br i1 [[EXITCOND]], label [[FOR_COND_CLEANUP:%.*]], label [[FOR_BODY]]
+;
 entry:
   br label %for.body
 
 for.cond.cleanup:                                 ; preds = %for.body
   ret void
 
 for.body:                                         ; preds = %for.body, %entry
   %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]

llvm/test/Transforms/LoopVectorize/X86/pr47437.ll

 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
 ; RUN: opt < %s -S -loop-vectorize -mtriple=x86_64-- -mattr=+sse2   | FileCheck %s --check-prefix=SSE2
 ; RUN: opt < %s -S -loop-vectorize -mtriple=x86_64-- -mattr=+sse4.1 | FileCheck %s --check-prefix=SSE41
 ; RUN: opt < %s -S -loop-vectorize -mtriple=x86_64-- -mattr=+avx    | FileCheck %s --check-prefix=AVX1
 ; RUN: opt < %s -S -loop-vectorize -mtriple=x86_64-- -mattr=+avx2   | FileCheck %s --check-prefix=AVX2
 ; RUN: opt < %s -S -loop-vectorize -mtriple=x86_64-- -mcpu=slm      | FileCheck %s --check-prefix=SSE2
 
 define void @test_muladd(i32* noalias nocapture %d1, i16* noalias nocapture readonly %s1, i16* noalias nocapture readonly %s2, i32 %n) {
 ; SSE2-LABEL: @test_muladd(
 ; SSE2-NEXT:  entry:
 ; SSE2-NEXT:    [[CMP30:%.*]] = icmp sgt i32 [[N:%.*]], 0
 ; SSE2-NEXT:    br i1 [[CMP30]], label [[FOR_BODY_PREHEADER:%.*]], label [[FOR_END:%.*]]
 ; SSE2:       for.body.preheader:
 ; SSE2-NEXT:    [[WIDE_TRIP_COUNT:%.*]] = zext i32 [[N]] to i64
+; SSE2-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[WIDE_TRIP_COUNT]], 2
+; SSE2-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
+; SSE2:       vector.ph:
+; SSE2-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[WIDE_TRIP_COUNT]], 2
+; SSE2-NEXT:    [[N_VEC:%.*]] = sub i64 [[WIDE_TRIP_COUNT]], [[N_MOD_VF]]
+; SSE2-NEXT:    br label [[VECTOR_BODY:%.*]]
+; SSE2:       vector.body:
+; SSE2-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
+; SSE2-NEXT:    [[TMP0:%.*]] = add i64 [[INDEX]], 0
+; SSE2-NEXT:    [[TMP1:%.*]] = shl nuw nsw i64 [[TMP0]], 1
+; SSE2-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i16, i16* [[S1:%.*]], i64 [[TMP1]]
+; SSE2-NEXT:    [[TMP3:%.*]] = getelementptr inbounds i16, i16* [[TMP2]], i32 0
+; SSE2-NEXT:    [[TMP4:%.*]] = bitcast i16* [[TMP3]] to <4 x i16>*
+; SSE2-NEXT:    [[WIDE_VEC:%.*]] = load <4 x i16>, <4 x i16>* [[TMP4]], align 2
+; SSE2-NEXT:    [[STRIDED_VEC:%.*]] = shufflevector <4 x i16> [[WIDE_VEC]], <4 x i16> poison, <2 x i32> <i32 0, i32 2>
+; SSE2-NEXT:    [[STRIDED_VEC1:%.*]] = shufflevector <4 x i16> [[WIDE_VEC]], <4 x i16> poison, <2 x i32> <i32 1, i32 3>
+; SSE2-NEXT:    [[TMP5:%.*]] = sext <2 x i16> [[STRIDED_VEC]] to <2 x i32>
+; SSE2-NEXT:    [[TMP6:%.*]] = getelementptr inbounds i16, i16* [[S2:%.*]], i64 [[TMP1]]
+; SSE2-NEXT:    [[TMP7:%.*]] = getelementptr inbounds i16, i16* [[TMP6]], i32 0
+; SSE2-NEXT:    [[TMP8:%.*]] = bitcast i16* [[TMP7]] to <4 x i16>*
+; SSE2-NEXT:    [[WIDE_VEC2:%.*]] = load <4 x i16>, <4 x i16>* [[TMP8]], align 2
+; SSE2-NEXT:    [[STRIDED_VEC3:%.*]] = shufflevector <4 x i16> [[WIDE_VEC2]], <4 x i16> poison, <2 x i32> <i32 0, i32 2>
+; SSE2-NEXT:    [[STRIDED_VEC4:%.*]] = shufflevector <4 x i16> [[WIDE_VEC2]], <4 x i16> poison, <2 x i32> <i32 1, i32 3>
+; SSE2-NEXT:    [[TMP9:%.*]] = sext <2 x i16> [[STRIDED_VEC3]] to <2 x i32>
+; SSE2-NEXT:    [[TMP10:%.*]] = mul nsw <2 x i32> [[TMP9]], [[TMP5]]
+; SSE2-NEXT:    [[TMP11:%.*]] = or i64 [[TMP1]], 1
+; SSE2-NEXT:    [[TMP12:%.*]] = getelementptr inbounds i16, i16* [[S1]], i64 [[TMP11]]
+; SSE2-NEXT:    [[TMP13:%.*]] = sext <2 x i16> [[STRIDED_VEC1]] to <2 x i32>
+; SSE2-NEXT:    [[TMP14:%.*]] = getelementptr inbounds i16, i16* [[S2]], i64 [[TMP11]]
+; SSE2-NEXT:    [[TMP15:%.*]] = sext <2 x i16> [[STRIDED_VEC4]] to <2 x i32>
+; SSE2-NEXT:    [[TMP16:%.*]] = mul nsw <2 x i32> [[TMP15]], [[TMP13]]
+; SSE2-NEXT:    [[TMP17:%.*]] = add nsw <2 x i32> [[TMP16]], [[TMP10]]
+; SSE2-NEXT:    [[TMP18:%.*]] = getelementptr inbounds i32, i32* [[D1:%.*]], i64 [[TMP0]]
+; SSE2-NEXT:    [[TMP19:%.*]] = getelementptr inbounds i32, i32* [[TMP18]], i32 0
+; SSE2-NEXT:    [[TMP20:%.*]] = bitcast i32* [[TMP19]] to <2 x i32>*
+; SSE2-NEXT:    store <2 x i32> [[TMP17]], <2 x i32>* [[TMP20]], align 4
+; SSE2-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 2
+; SSE2-NEXT:    [[TMP21:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
+; SSE2-NEXT:    br i1 [[TMP21]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
+; SSE2:       middle.block:
+; SSE2-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[WIDE_TRIP_COUNT]], [[N_VEC]]
+; SSE2-NEXT:    br i1 [[CMP_N]], label [[FOR_END_LOOPEXIT:%.*]], label [[SCALAR_PH]]
+; SSE2:       scalar.ph:
+; SSE2-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[FOR_BODY_PREHEADER]] ]
 ; SSE2-NEXT:    br label [[FOR_BODY:%.*]]
 ; SSE2:       for.body:
-; SSE2-NEXT:    [[INDVARS_IV:%.*]] = phi i64 [ 0, [[FOR_BODY_PREHEADER]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]
-; SSE2-NEXT:    [[TMP0:%.*]] = shl nuw nsw i64 [[INDVARS_IV]], 1
-; SSE2-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[S1:%.*]], i64 [[TMP0]]
-; SSE2-NEXT:    [[TMP1:%.*]] = load i16, i16* [[ARRAYIDX]], align 2
-; SSE2-NEXT:    [[CONV:%.*]] = sext i16 [[TMP1]] to i32
-; SSE2-NEXT:    [[ARRAYIDX4:%.*]] = getelementptr inbounds i16, i16* [[S2:%.*]], i64 [[TMP0]]
-; SSE2-NEXT:    [[TMP2:%.*]] = load i16, i16* [[ARRAYIDX4]], align 2
-; SSE2-NEXT:    [[CONV5:%.*]] = sext i16 [[TMP2]] to i32
+; SSE2-NEXT:    [[INDVARS_IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]
+; SSE2-NEXT:    [[TMP22:%.*]] = shl nuw nsw i64 [[INDVARS_IV]], 1
+; SSE2-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[S1]], i64 [[TMP22]]
+; SSE2-NEXT:    [[TMP23:%.*]] = load i16, i16* [[ARRAYIDX]], align 2
+; SSE2-NEXT:    [[CONV:%.*]] = sext i16 [[TMP23]] to i32
+; SSE2-NEXT:    [[ARRAYIDX4:%.*]] = getelementptr inbounds i16, i16* [[S2]], i64 [[TMP22]]
+; SSE2-NEXT:    [[TMP24:%.*]] = load i16, i16* [[ARRAYIDX4]], align 2
+; SSE2-NEXT:    [[CONV5:%.*]] = sext i16 [[TMP24]] to i32
 ; SSE2-NEXT:    [[MUL6:%.*]] = mul nsw i32 [[CONV5]], [[CONV]]
-; SSE2-NEXT:    [[TMP3:%.*]] = or i64 [[TMP0]], 1
-; SSE2-NEXT:    [[ARRAYIDX10:%.*]] = getelementptr inbounds i16, i16* [[S1]], i64 [[TMP3]]
-; SSE2-NEXT:    [[TMP4:%.*]] = load i16, i16* [[ARRAYIDX10]], align 2
-; SSE2-NEXT:    [[CONV11:%.*]] = sext i16 [[TMP4]] to i32
-; SSE2-NEXT:    [[ARRAYIDX15:%.*]] = getelementptr inbounds i16, i16* [[S2]], i64 [[TMP3]]
-; SSE2-NEXT:    [[TMP5:%.*]] = load i16, i16* [[ARRAYIDX15]], align 2
-; SSE2-NEXT:    [[CONV16:%.*]] = sext i16 [[TMP5]] to i32
+; SSE2-NEXT:    [[TMP25:%.*]] = or i64 [[TMP22]], 1
+; SSE2-NEXT:    [[ARRAYIDX10:%.*]] = getelementptr inbounds i16, i16* [[S1]], i64 [[TMP25]]
+; SSE2-NEXT:    [[TMP26:%.*]] = load i16, i16* [[ARRAYIDX10]], align 2
+; SSE2-NEXT:    [[CONV11:%.*]] = sext i16 [[TMP26]] to i32
+; SSE2-NEXT:    [[ARRAYIDX15:%.*]] = getelementptr inbounds i16, i16* [[S2]], i64 [[TMP25]]
+; SSE2-NEXT:    [[TMP27:%.*]] = load i16, i16* [[ARRAYIDX15]], align 2
+; SSE2-NEXT:    [[CONV16:%.*]] = sext i16 [[TMP27]] to i32
 ; SSE2-NEXT:    [[MUL17:%.*]] = mul nsw i32 [[CONV16]], [[CONV11]]
 ; SSE2-NEXT:    [[ADD18:%.*]] = add nsw i32 [[MUL17]], [[MUL6]]
-; SSE2-NEXT:    [[ARRAYIDX20:%.*]] = getelementptr inbounds i32, i32* [[D1:%.*]], i64 [[INDVARS_IV]]
+; SSE2-NEXT:    [[ARRAYIDX20:%.*]] = getelementptr inbounds i32, i32* [[D1]], i64 [[INDVARS_IV]]
 ; SSE2-NEXT:    store i32 [[ADD18]], i32* [[ARRAYIDX20]], align 4
 ; SSE2-NEXT:    [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
 ; SSE2-NEXT:    [[EXITCOND_NOT:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], [[WIDE_TRIP_COUNT]]
-; SSE2-NEXT:    br i1 [[EXITCOND_NOT]], label [[FOR_END_LOOPEXIT:%.*]], label [[FOR_BODY]]
+; SSE2-NEXT:    br i1 [[EXITCOND_NOT]], label [[FOR_END_LOOPEXIT]], label [[FOR_BODY]], !llvm.loop [[LOOP2:![0-9]+]]
 ; SSE2:       for.end.loopexit:
 ; SSE2-NEXT:    br label [[FOR_END]]
 ; SSE2:       for.end:
 ; SSE2-NEXT:    ret void
 ;
 ; SSE41-LABEL: @test_muladd(
 ; SSE41-NEXT:  entry:
 ; SSE41-NEXT:    [[CMP30:%.*]] = icmp sgt i32 [[N:%.*]], 0
 ; SSE41-NEXT:    br i1 [[CMP30]], label [[FOR_BODY_PREHEADER:%.*]], label [[FOR_END:%.*]]
 ; SSE41:       for.body.preheader:
 ; SSE41-NEXT:    [[WIDE_TRIP_COUNT:%.*]] = zext i32 [[N]] to i64
 ; SSE41-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[WIDE_TRIP_COUNT]], 4
 ; SSE41-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
 ; SSE41:       vector.ph:
 ; SSE41-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[WIDE_TRIP_COUNT]], 4
 ; SSE41-NEXT:    [[N_VEC:%.*]] = sub i64 [[WIDE_TRIP_COUNT]], [[N_MOD_VF]]
 ; SSE41-NEXT:    br label [[VECTOR_BODY:%.*]]
 ; SSE41:       vector.body:
 ; SSE41-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
 ; SSE41-NEXT:    [[TMP0:%.*]] = add i64 [[INDEX]], 0
-; SSE41-NEXT:    [[TMP1:%.*]] = shl nuw nsw i64 [[TMP0]], 1
-; SSE41-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i16, i16* [[S1:%.*]], i64 [[TMP1]]
-; SSE41-NEXT:    [[TMP3:%.*]] = getelementptr inbounds i16, i16* [[TMP2]], i32 0
-; SSE41-NEXT:    [[TMP4:%.*]] = bitcast i16* [[TMP3]] to <8 x i16>*
-; SSE41-NEXT:    [[WIDE_VEC:%.*]] = load <8 x i16>, <8 x i16>* [[TMP4]], align 2
-; SSE41-NEXT:    [[STRIDED_VEC:%.*]] = shufflevector <8 x i16> [[WIDE_VEC]], <8 x i16> poison, <4 x i32> <i32 0, i32 2, i32 4, i32 6>
-; SSE41-NEXT:    [[STRIDED_VEC1:%.*]] = shufflevector <8 x i16> [[WIDE_VEC]], <8 x i16> poison, <4 x i32> <i32 1, i32 3, i32 5, i32 7>
-; SSE41-NEXT:    [[TMP5:%.*]] = sext <4 x i16> [[STRIDED_VEC]] to <4 x i32>
-; SSE41-NEXT:    [[TMP6:%.*]] = getelementptr inbounds i16, i16* [[S2:%.*]], i64 [[TMP1]]
-; SSE41-NEXT:    [[TMP7:%.*]] = getelementptr inbounds i16, i16* [[TMP6]], i32 0
-; SSE41-NEXT:    [[TMP8:%.*]] = bitcast i16* [[TMP7]] to <8 x i16>*
-; SSE41-NEXT:    [[WIDE_VEC2:%.*]] = load <8 x i16>, <8 x i16>* [[TMP8]], align 2
-; SSE41-NEXT:    [[STRIDED_VEC3:%.*]] = shufflevector <8 x i16> [[WIDE_VEC2]], <8 x i16> poison, <4 x i32> <i32 0, i32 2, i32 4, i32 6>
-; SSE41-NEXT:    [[STRIDED_VEC4:%.*]] = shufflevector <8 x i16> [[WIDE_VEC2]], <8 x i16> poison, <4 x i32> <i32 1, i32 3, i32 5, i32 7>
-; SSE41-NEXT:    [[TMP9:%.*]] = sext <4 x i16> [[STRIDED_VEC3]] to <4 x i32>
-; SSE41-NEXT:    [[TMP10:%.*]] = mul nsw <4 x i32> [[TMP9]], [[TMP5]]
-; SSE41-NEXT:    [[TMP11:%.*]] = or i64 [[TMP1]], 1
-; SSE41-NEXT:    [[TMP12:%.*]] = getelementptr inbounds i16, i16* [[S1]], i64 [[TMP11]]
-; SSE41-NEXT:    [[TMP13:%.*]] = sext <4 x i16> [[STRIDED_VEC1]] to <4 x i32>
-; SSE41-NEXT:    [[TMP14:%.*]] = getelementptr inbounds i16, i16* [[S2]], i64 [[TMP11]]
-; SSE41-NEXT:    [[TMP15:%.*]] = sext <4 x i16> [[STRIDED_VEC4]] to <4 x i32>
-; SSE41-NEXT:    [[TMP16:%.*]] = mul nsw <4 x i32> [[TMP15]], [[TMP13]]
-; SSE41-NEXT:    [[TMP17:%.*]] = add nsw <4 x i32> [[TMP16]], [[TMP10]]
-; SSE41-NEXT:    [[TMP18:%.*]] = getelementptr inbounds i32, i32* [[D1:%.*]], i64 [[TMP0]]
-; SSE41-NEXT:    [[TMP19:%.*]] = getelementptr inbounds i32, i32* [[TMP18]], i32 0
-; SSE41-NEXT:    [[TMP20:%.*]] = bitcast i32* [[TMP19]] to <4 x i32>*
-; SSE41-NEXT:    store <4 x i32> [[TMP17]], <4 x i32>* [[TMP20]], align 4
+; SSE41-NEXT:    [[TMP1:%.*]] = add i64 [[INDEX]], 2
+; SSE41-NEXT:    [[TMP2:%.*]] = shl nuw nsw i64 [[TMP0]], 1
+; SSE41-NEXT:    [[TMP3:%.*]] = shl nuw nsw i64 [[TMP1]], 1
+; SSE41-NEXT:    [[TMP4:%.*]] = getelementptr inbounds i16, i16* [[S1:%.*]], i64 [[TMP2]]
+; SSE41-NEXT:    [[TMP5:%.*]] = getelementptr inbounds i16, i16* [[S1]], i64 [[TMP3]]
+; SSE41-NEXT:    [[TMP6:%.*]] = getelementptr inbounds i16, i16* [[TMP4]], i32 0
+; SSE41-NEXT:    [[TMP7:%.*]] = bitcast i16* [[TMP6]] to <4 x i16>*
+; SSE41-NEXT:    [[TMP8:%.*]] = getelementptr inbounds i16, i16* [[TMP5]], i32 0
+; SSE41-NEXT:    [[TMP9:%.*]] = bitcast i16* [[TMP8]] to <4 x i16>*
+; SSE41-NEXT:    [[WIDE_VEC:%.*]] = load <4 x i16>, <4 x i16>* [[TMP7]], align 2
+; SSE41-NEXT:    [[WIDE_VEC1:%.*]] = load <4 x i16>, <4 x i16>* [[TMP9]], align 2
+; SSE41-NEXT:    [[STRIDED_VEC:%.*]] = shufflevector <4 x i16> [[WIDE_VEC]], <4 x i16> poison, <2 x i32> <i32 0, i32 2>
+; SSE41-NEXT:    [[STRIDED_VEC2:%.*]] = shufflevector <4 x i16> [[WIDE_VEC1]], <4 x i16> poison, <2 x i32> <i32 0, i32 2>
+; SSE41-NEXT:    [[STRIDED_VEC3:%.*]] = shufflevector <4 x i16> [[WIDE_VEC]], <4 x i16> poison, <2 x i32> <i32 1, i32 3>
+; SSE41-NEXT:    [[STRIDED_VEC4:%.*]] = shufflevector <4 x i16> [[WIDE_VEC1]], <4 x i16> poison, <2 x i32> <i32 1, i32 3>
+; SSE41-NEXT:    [[TMP10:%.*]] = sext <2 x i16> [[STRIDED_VEC]] to <2 x i32>
+; SSE41-NEXT:    [[TMP11:%.*]] = sext <2 x i16> [[STRIDED_VEC2]] to <2 x i32>
+; SSE41-NEXT:    [[TMP12:%.*]] = getelementptr inbounds i16, i16* [[S2:%.*]], i64 [[TMP2]]
+; SSE41-NEXT:    [[TMP13:%.*]] = getelementptr inbounds i16, i16* [[S2]], i64 [[TMP3]]
+; SSE41-NEXT:    [[TMP14:%.*]] = getelementptr inbounds i16, i16* [[TMP12]], i32 0
+; SSE41-NEXT:    [[TMP15:%.*]] = bitcast i16* [[TMP14]] to <4 x i16>*
+; SSE41-NEXT:    [[TMP16:%.*]] = getelementptr inbounds i16, i16* [[TMP13]], i32 0
+; SSE41-NEXT:    [[TMP17:%.*]] = bitcast i16* [[TMP16]] to <4 x i16>*
+; SSE41-NEXT:    [[WIDE_VEC5:%.*]] = load <4 x i16>, <4 x i16>* [[TMP15]], align 2
+; SSE41-NEXT:    [[WIDE_VEC6:%.*]] = load <4 x i16>, <4 x i16>* [[TMP17]], align 2
+; SSE41-NEXT:    [[STRIDED_VEC7:%.*]] = shufflevector <4 x i16> [[WIDE_VEC5]], <4 x i16> poison, <2 x i32> <i32 0, i32 2>
+; SSE41-NEXT:    [[STRIDED_VEC8:%.*]] = shufflevector <4 x i16> [[WIDE_VEC6]], <4 x i16> poison, <2 x i32> <i32 0, i32 2>
+; SSE41-NEXT:    [[STRIDED_VEC9:%.*]] = shufflevector <4 x i16> [[WIDE_VEC5]], <4 x i16> poison, <2 x i32> <i32 1, i32 3>
+; SSE41-NEXT:    [[STRIDED_VEC10:%.*]] = shufflevector <4 x i16> [[WIDE_VEC6]], <4 x i16> poison, <2 x i32> <i32 1, i32 3>
+; SSE41-NEXT:    [[TMP18:%.*]] = sext <2 x i16> [[STRIDED_VEC7]] to <2 x i32>
+; SSE41-NEXT:    [[TMP19:%.*]] = sext <2 x i16> [[STRIDED_VEC8]] to <2 x i32>
+; SSE41-NEXT:    [[TMP20:%.*]] = mul nsw <2 x i32> [[TMP18]], [[TMP10]]
+; SSE41-NEXT:    [[TMP21:%.*]] = mul nsw <2 x i32> [[TMP19]], [[TMP11]]
+; SSE41-NEXT:    [[TMP22:%.*]] = or i64 [[TMP2]], 1
+; SSE41-NEXT:    [[TMP23:%.*]] = or i64 [[TMP3]], 1
+; SSE41-NEXT:    [[TMP24:%.*]] = getelementptr inbounds i16, i16* [[S1]], i64 [[TMP22]]
+; SSE41-NEXT:    [[TMP25:%.*]] = getelementptr inbounds i16, i16* [[S1]], i64 [[TMP23]]
+; SSE41-NEXT:    [[TMP26:%.*]] = sext <2 x i16> [[STRIDED_VEC3]] to <2 x i32>
+; SSE41-NEXT:    [[TMP27:%.*]] = sext <2 x i16> [[STRIDED_VEC4]] to <2 x i32>
+; SSE41-NEXT:    [[TMP28:%.*]] = getelementptr inbounds i16, i16* [[S2]], i64 [[TMP22]]
+; SSE41-NEXT:    [[TMP29:%.*]] = getelementptr inbounds i16, i16* [[S2]], i64 [[TMP23]]
+; SSE41-NEXT:    [[TMP30:%.*]] = sext <2 x i16> [[STRIDED_VEC9]] to <2 x i32>
+; SSE41-NEXT:    [[TMP31:%.*]] = sext <2 x i16> [[STRIDED_VEC10]] to <2 x i32>
+; SSE41-NEXT:    [[TMP32:%.*]] = mul nsw <2 x i32> [[TMP30]], [[TMP26]]
+; SSE41-NEXT:    [[TMP33:%.*]] = mul nsw <2 x i32> [[TMP31]], [[TMP27]]
+; SSE41-NEXT:    [[TMP34:%.*]] = add nsw <2 x i32> [[TMP32]], [[TMP20]]
+; SSE41-NEXT:    [[TMP35:%.*]] = add nsw <2 x i32> [[TMP33]], [[TMP21]]
+; SSE41-NEXT:    [[TMP36:%.*]] = getelementptr inbounds i32, i32* [[D1:%.*]], i64 [[TMP0]]
+; SSE41-NEXT:    [[TMP37:%.*]] = getelementptr inbounds i32, i32* [[D1]], i64 [[TMP1]]
+; SSE41-NEXT:    [[TMP38:%.*]] = getelementptr inbounds i32, i32* [[TMP36]], i32 0
+; SSE41-NEXT:    [[TMP39:%.*]] = bitcast i32* [[TMP38]] to <2 x i32>*
+; SSE41-NEXT:    store <2 x i32> [[TMP34]], <2 x i32>* [[TMP39]], align 4
+; SSE41-NEXT:    [[TMP40:%.*]] = getelementptr inbounds i32, i32* [[TMP36]], i32 2
+; SSE41-NEXT:    [[TMP41:%.*]] = bitcast i32* [[TMP40]] to <2 x i32>*
+; SSE41-NEXT:    store <2 x i32> [[TMP35]], <2 x i32>* [[TMP41]], align 4
 ; SSE41-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
-; SSE41-NEXT:    [[TMP21:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
-; SSE41-NEXT:    br i1 [[TMP21]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
+; SSE41-NEXT:    [[TMP42:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
+; SSE41-NEXT:    br i1 [[TMP42]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
 ; SSE41:       middle.block:
 ; SSE41-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[WIDE_TRIP_COUNT]], [[N_VEC]]
 ; SSE41-NEXT:    br i1 [[CMP_N]], label [[FOR_END_LOOPEXIT:%.*]], label [[SCALAR_PH]]
 ; SSE41:       scalar.ph:
 ; SSE41-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[FOR_BODY_PREHEADER]] ]
 ; SSE41-NEXT:    br label [[FOR_BODY:%.*]]
 ; SSE41:       for.body:
 ; SSE41-NEXT:    [[INDVARS_IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]
-; SSE41-NEXT:    [[TMP22:%.*]] = shl nuw nsw i64 [[INDVARS_IV]], 1
-; SSE41-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[S1]], i64 [[TMP22]]
-; SSE41-NEXT:    [[TMP23:%.*]] = load i16, i16* [[ARRAYIDX]], align 2
-; SSE41-NEXT:    [[CONV:%.*]] = sext i16 [[TMP23]] to i32
-; SSE41-NEXT:    [[ARRAYIDX4:%.*]] = getelementptr inbounds i16, i16* [[S2]], i64 [[TMP22]]
-; SSE41-NEXT:    [[TMP24:%.*]] = load i16, i16* [[ARRAYIDX4]], align 2
-; SSE41-NEXT:    [[CONV5:%.*]] = sext i16 [[TMP24]] to i32
+; SSE41-NEXT:    [[TMP43:%.*]] = shl nuw nsw i64 [[INDVARS_IV]], 1
+; SSE41-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[S1]], i64 [[TMP43]]
+; SSE41-NEXT:    [[TMP44:%.*]] = load i16, i16* [[ARRAYIDX]], align 2
+; SSE41-NEXT:    [[CONV:%.*]] = sext i16 [[TMP44]] to i32
+; SSE41-NEXT:    [[ARRAYIDX4:%.*]] = getelementptr inbounds i16, i16* [[S2]], i64 [[TMP43]]
+; SSE41-NEXT:    [[TMP45:%.*]] = load i16, i16* [[ARRAYIDX4]], align 2
+; SSE41-NEXT:    [[CONV5:%.*]] = sext i16 [[TMP45]] to i32
 ; SSE41-NEXT:    [[MUL6:%.*]] = mul nsw i32 [[CONV5]], [[CONV]]
-; SSE41-NEXT:    [[TMP25:%.*]] = or i64 [[TMP22]], 1
-; SSE41-NEXT:    [[ARRAYIDX10:%.*]] = getelementptr inbounds i16, i16* [[S1]], i64 [[TMP25]]
-; SSE41-NEXT:    [[TMP26:%.*]] = load i16, i16* [[ARRAYIDX10]], align 2
-; SSE41-NEXT:    [[CONV11:%.*]] = sext i16 [[TMP26]] to i32
-; SSE41-NEXT:    [[ARRAYIDX15:%.*]] = getelementptr inbounds i16, i16* [[S2]], i64 [[TMP25]]
-; SSE41-NEXT:    [[TMP27:%.*]] = load i16, i16* [[ARRAYIDX15]], align 2
-; SSE41-NEXT:    [[CONV16:%.*]] = sext i16 [[TMP27]] to i32
+; SSE41-NEXT:    [[TMP46:%.*]] = or i64 [[TMP43]], 1
+; SSE41-NEXT:    [[ARRAYIDX10:%.*]] = getelementptr inbounds i16, i16* [[S1]], i64 [[TMP46]]
+; SSE41-NEXT:    [[TMP47:%.*]] = load i16, i16* [[ARRAYIDX10]], align 2
+; SSE41-NEXT:    [[CONV11:%.*]] = sext i16 [[TMP47]] to i32
+; SSE41-NEXT:    [[ARRAYIDX15:%.*]] = getelementptr inbounds i16, i16* [[S2]], i64 [[TMP46]]
+; SSE41-NEXT:    [[TMP48:%.*]] = load i16, i16* [[ARRAYIDX15]], align 2
+; SSE41-NEXT:    [[CONV16:%.*]] = sext i16 [[TMP48]] to i32
 ; SSE41-NEXT:    [[MUL17:%.*]] = mul nsw i32 [[CONV16]], [[CONV11]]
 ; SSE41-NEXT:    [[ADD18:%.*]] = add nsw i32 [[MUL17]], [[MUL6]]
 ; SSE41-NEXT:    [[ARRAYIDX20:%.*]] = getelementptr inbounds i32, i32* [[D1]], i64 [[INDVARS_IV]]
 ; SSE41-NEXT:    store i32 [[ADD18]], i32* [[ARRAYIDX20]], align 4
 ; SSE41-NEXT:    [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
 ; SSE41-NEXT:    [[EXITCOND_NOT:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], [[WIDE_TRIP_COUNT]]
 ; SSE41-NEXT:    br i1 [[EXITCOND_NOT]], label [[FOR_END_LOOPEXIT]], label [[FOR_BODY]], !llvm.loop [[LOOP2:![0-9]+]]
 ; SSE41:       for.end.loopexit:
 ; SSE41-NEXT:    br label [[FOR_END]]
 ; SSE41:       for.end:
 ; SSE41-NEXT:    ret void
 ;
 ; AVX1-LABEL: @test_muladd(
 ; AVX1-NEXT:  entry:
 ; AVX1-NEXT:    [[CMP30:%.*]] = icmp sgt i32 [[N:%.*]], 0
 ; AVX1-NEXT:    br i1 [[CMP30]], label [[FOR_BODY_PREHEADER:%.*]], label [[FOR_END:%.*]]
 ; AVX1:       for.body.preheader:
 ; AVX1-NEXT:    [[WIDE_TRIP_COUNT:%.*]] = zext i32 [[N]] to i64
-; AVX1-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[WIDE_TRIP_COUNT]], 4
+; AVX1-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[WIDE_TRIP_COUNT]], 8
 ; AVX1-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
 ; AVX1:       vector.ph:
-; AVX1-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[WIDE_TRIP_COUNT]], 4
+; AVX1-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[WIDE_TRIP_COUNT]], 8
 ; AVX1-NEXT:    [[N_VEC:%.*]] = sub i64 [[WIDE_TRIP_COUNT]], [[N_MOD_VF]]
 ; AVX1-NEXT:    br label [[VECTOR_BODY:%.*]]
 ; AVX1:       vector.body:
 ; AVX1-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
 ; AVX1-NEXT:    [[TMP0:%.*]] = add i64 [[INDEX]], 0
-; AVX1-NEXT:    [[TMP1:%.*]] = shl nuw nsw i64 [[TMP0]], 1
-; AVX1-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i16, i16* [[S1:%.*]], i64 [[TMP1]]
-; AVX1-NEXT:    [[TMP3:%.*]] = getelementptr inbounds i16, i16* [[TMP2]], i32 0
-; AVX1-NEXT:    [[TMP4:%.*]] = bitcast i16* [[TMP3]] to <8 x i16>*
-; AVX1-NEXT:    [[WIDE_VEC:%.*]] = load <8 x i16>, <8 x i16>* [[TMP4]], align 2
-; AVX1-NEXT:    [[STRIDED_VEC:%.*]] = shufflevector <8 x i16> [[WIDE_VEC]], <8 x i16> poison, <4 x i32> <i32 0, i32 2, i32 4, i32 6>
-; AVX1-NEXT:    [[STRIDED_VEC1:%.*]] = shufflevector <8 x i16> [[WIDE_VEC]], <8 x i16> poison, <4 x i32> <i32 1, i32 3, i32 5, i32 7>
-; AVX1-NEXT:    [[TMP5:%.*]] = sext <4 x i16> [[STRIDED_VEC]] to <4 x i32>
-; AVX1-NEXT:    [[TMP6:%.*]] = getelementptr inbounds i16, i16* [[S2:%.*]], i64 [[TMP1]]
-; AVX1-NEXT:    [[TMP7:%.*]] = getelementptr inbounds i16, i16* [[TMP6]], i32 0
-; AVX1-NEXT:    [[TMP8:%.*]] = bitcast i16* [[TMP7]] to <8 x i16>*
-; AVX1-NEXT:    [[WIDE_VEC2:%.*]] = load <8 x i16>, <8 x i16>* [[TMP8]], align 2
-; AVX1-NEXT:    [[STRIDED_VEC3:%.*]] = shufflevector <8 x i16> [[WIDE_VEC2]], <8 x i16> poison, <4 x i32> <i32 0, i32 2, i32 4, i32 6>
-; AVX1-NEXT:    [[STRIDED_VEC4:%.*]] = shufflevector <8 x i16> [[WIDE_VEC2]], <8 x i16> poison, <4 x i32> <i32 1, i32 3, i32 5, i32 7>
-; AVX1-NEXT:    [[TMP9:%.*]] = sext <4 x i16> [[STRIDED_VEC3]] to <4 x i32>
-; AVX1-NEXT:    [[TMP10:%.*]] = mul nsw <4 x i32> [[TMP9]], [[TMP5]]
-; AVX1-NEXT:    [[TMP11:%.*]] = or i64 [[TMP1]], 1
-; AVX1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds i16, i16* [[S1]], i64 [[TMP11]]
-; AVX1-NEXT:    [[TMP13:%.*]] = sext <4 x i16> [[STRIDED_VEC1]] to <4 x i32>
-; AVX1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds i16, i16* [[S2]], i64 [[TMP11]]
-; AVX1-NEXT:    [[TMP15:%.*]] = sext <4 x i16> [[STRIDED_VEC4]] to <4 x i32>
-; AVX1-NEXT:    [[TMP16:%.*]] = mul nsw <4 x i32> [[TMP15]], [[TMP13]]
-; AVX1-NEXT:    [[TMP17:%.*]] = add nsw <4 x i32> [[TMP16]], [[TMP10]]
-; AVX1-NEXT:    [[TMP18:%.*]] = getelementptr inbounds i32, i32* [[D1:%.*]], i64 [[TMP0]]
-; AVX1-NEXT:    [[TMP19:%.*]] = getelementptr inbounds i32, i32* [[TMP18]], i32 0
-; AVX1-NEXT:    [[TMP20:%.*]] = bitcast i32* [[TMP19]] to <4 x i32>*
-; AVX1-NEXT:    store <4 x i32> [[TMP17]], <4 x i32>* [[TMP20]], align 4
-; AVX1-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
-; AVX1-NEXT:    [[TMP21:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
-; AVX1-NEXT:    br i1 [[TMP21]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
+; AVX1-NEXT:    [[TMP1:%.*]] = add i64 [[INDEX]], 2
+; AVX1-NEXT:    [[TMP2:%.*]] = add i64 [[INDEX]], 4
+; AVX1-NEXT:    [[TMP3:%.*]] = add i64 [[INDEX]], 6
+; AVX1-NEXT:    [[TMP4:%.*]] = shl nuw nsw i64 [[TMP0]], 1
+; AVX1-NEXT:    [[TMP5:%.*]] = shl nuw nsw i64 [[TMP1]], 1
+; AVX1-NEXT:    [[TMP6:%.*]] = shl nuw nsw i64 [[TMP2]], 1
+; AVX1-NEXT:    [[TMP7:%.*]] = shl nuw nsw i64 [[TMP3]], 1
+; AVX1-NEXT:    [[TMP8:%.*]] = getelementptr inbounds i16, i16* [[S1:%.*]], i64 [[TMP4]]
+; AVX1-NEXT:    [[TMP9:%.*]] = getelementptr inbounds i16, i16* [[S1]], i64 [[TMP5]]
+; AVX1-NEXT:    [[TMP10:%.*]] = getelementptr inbounds i16, i16* [[S1]], i64 [[TMP6]]
+; AVX1-NEXT:    [[TMP11:%.*]] = getelementptr inbounds i16, i16* [[S1]], i64 [[TMP7]]
+; AVX1-NEXT:    [[TMP12:%.*]] = getelementptr inbounds i16, i16* [[TMP8]], i32 0
+; AVX1-NEXT:    [[TMP13:%.*]] = bitcast i16* [[TMP12]] to <4 x i16>*
+; AVX1-NEXT:    [[TMP14:%.*]] = getelementptr inbounds i16, i16* [[TMP9]], i32 0
+; AVX1-NEXT:    [[TMP15:%.*]] = bitcast i16* [[TMP14]] to <4 x i16>*
+; AVX1-NEXT:    [[TMP16:%.*]] = getelementptr inbounds i16, i16* [[TMP10]], i32 0
+; AVX1-NEXT:    [[TMP17:%.*]] = bitcast i16* [[TMP16]] to <4 x i16>*
+; AVX1-NEXT:    [[TMP18:%.*]] = getelementptr inbounds i16, i16* [[TMP11]], i32 0
+; AVX1-NEXT:    [[TMP19:%.*]] = bitcast i16* [[TMP18]] to <4 x i16>*
+; AVX1-NEXT:    [[WIDE_VEC:%.*]] = load <4 x i16>, <4 x i16>* [[TMP13]], align 2
+; AVX1-NEXT:    [[WIDE_VEC1:%.*]] = load <4 x i16>, <4 x i16>* [[TMP15]], align 2
+; AVX1-NEXT:    [[WIDE_VEC2:%.*]] = load <4 x i16>, <4 x i16>* [[TMP17]], align 2
+; AVX1-NEXT:    [[WIDE_VEC3:%.*]] = load <4 x i16>, <4 x i16>* [[TMP19]], align 2
+; AVX1-NEXT:    [[STRIDED_VEC:%.*]] = shufflevector <4 x i16> [[WIDE_VEC]], <4 x i16> poison, <2 x i32> <i32 0, i32 2>
+; AVX1-NEXT:    [[STRIDED_VEC4:%.*]] = shufflevector <4 x i16> [[WIDE_VEC1]], <4 x i16> poison, <2 x i32> <i32 0, i32 2>
+; AVX1-NEXT:    [[STRIDED_VEC5:%.*]] = shufflevector <4 x i16> [[WIDE_VEC2]], <4 x i16> poison, <2 x i32> <i32 0, i32 2>
+; AVX1-NEXT:    [[STRIDED_VEC6:%.*]] = shufflevector <4 x i16> [[WIDE_VEC3]], <4 x i16> poison, <2 x i32> <i32 0, i32 2>
+; AVX1-NEXT:    [[STRIDED_VEC7:%.*]] = shufflevector <4 x i16> [[WIDE_VEC]], <4 x i16> poison, <2 x i32> <i32 1, i32 3>
+; AVX1-NEXT:    [[STRIDED_VEC8:%.*]] = shufflevector <4 x i16> [[WIDE_VEC1]], <4 x i16> poison, <2 x i32> <i32 1, i32 3>
+; AVX1-NEXT:    [[STRIDED_VEC9:%.*]] = shufflevector <4 x i16> [[WIDE_VEC2]], <4 x i16> poison, <2 x i32> <i32 1, i32 3>
+; AVX1-NEXT:    [[STRIDED_VEC10:%.*]] = shufflevector <4 x i16> [[WIDE_VEC3]], <4 x i16> poison, <2 x i32> <i32 1, i32 3>
+; AVX1-NEXT:    [[TMP20:%.*]] = sext <2 x i16> [[STRIDED_VEC]] to <2 x i32>
+; AVX1-NEXT:    [[TMP21:%.*]] = sext <2 x i16> [[STRIDED_VEC4]] to <2 x i32>
+; AVX1-NEXT:    [[TMP22:%.*]] = sext <2 x i16> [[STRIDED_VEC5]] to <2 x i32>
+; AVX1-NEXT:    [[TMP23:%.*]] = sext <2 x i16> [[STRIDED_VEC6]] to <2 x i32>
+; AVX1-NEXT:    [[TMP24:%.*]] = getelementptr inbounds i16, i16* [[S2:%.*]], i64 [[TMP4]]
+; AVX1-NEXT:    [[TMP25:%.*]] = getelementptr inbounds i16, i16* [[S2]], i64 [[TMP5]]
+; AVX1-NEXT:    [[TMP26:%.*]] = getelementptr inbounds i16, i16* [[S2]], i64 [[TMP6]]
+; AVX1-NEXT:    [[TMP27:%.*]] = getelementptr inbounds i16, i16* [[S2]], i64 [[TMP7]]
+; AVX1-NEXT:    [[TMP28:%.*]] = getelementptr inbounds i16, i16* [[TMP24]], i32 0
+; AVX1-NEXT:    [[TMP29:%.*]] = bitcast i16* [[TMP28]] to <4 x i16>*
+; AVX1-NEXT:    [[TMP30:%.*]] = getelementptr inbounds i16, i16* [[TMP25]], i32 0
+; AVX1-NEXT:    [[TMP31:%.*]] = bitcast i16* [[TMP30]] to <4 x i16>*
+; AVX1-NEXT:    [[TMP32:%.*]] = getelementptr inbounds i16, i16* [[TMP26]], i32 0
+; AVX1-NEXT:    [[TMP33:%.*]] = bitcast i16* [[TMP32]] to <4 x i16>*
+; AVX1-NEXT:    [[TMP34:%.*]] = getelementptr inbounds i16, i16* [[TMP27]], i32 0
+; AVX1-NEXT:    [[TMP35:%.*]] = bitcast i16* [[TMP34]] to <4 x i16>*
+; AVX1-NEXT:    [[WIDE_VEC11:%.*]] = load <4 x i16>, <4 x i16>* [[TMP29]], align 2
+; AVX1-NEXT:    [[WIDE_VEC12:%.*]] = load <4 x i16>, <4 x i16>* [[TMP31]], align 2
+; AVX1-NEXT:    [[WIDE_VEC13:%.*]] = load <4 x i16>, <4 x i16>* [[TMP33]], align 2
+; AVX1-NEXT:    [[WIDE_VEC14:%.*]] = load <4 x i16>, <4 x i16>* [[TMP35]], align 2
+; AVX1-NEXT:    [[STRIDED_VEC15:%.*]] = shufflevector <4 x i16> [[WIDE_VEC11]], <4 x i16> poison, <2 x i32> <i32 0, i32 2>
+; AVX1-NEXT:    [[STRIDED_VEC16:%.*]] = shufflevector <4 x i16> [[WIDE_VEC12]], <4 x i16> poison, <2 x i32> <i32 0, i32 2>
+; AVX1-NEXT:    [[STRIDED_VEC17:%.*]] = shufflevector <4 x i16> [[WIDE_VEC13]], <4 x i16> poison, <2 x i32> <i32 0, i32 2>
+; AVX1-NEXT:    [[STRIDED_VEC18:%.*]] = shufflevector <4 x i16> [[WIDE_VEC14]], <4 x i16> poison, <2 x i32> <i32 0, i32 2>
+; AVX1-NEXT:    [[STRIDED_VEC19:%.*]] = shufflevector <4 x i16> [[WIDE_VEC11]], <4 x i16> poison, <2 x i32> <i32 1, i32 3>
+; AVX1-NEXT:    [[STRIDED_VEC20:%.*]] = shufflevector <4 x i16> [[WIDE_VEC12]], <4 x i16> poison, <2 x i32> <i32 1, i32 3>
+; AVX1-NEXT:    [[STRIDED_VEC21:%.*]] = shufflevector <4 x i16> [[WIDE_VEC13]], <4 x i16> poison, <2 x i32> <i32 1, i32 3>
+; AVX1-NEXT:    [[STRIDED_VEC22:%.*]] = shufflevector <4 x i16> [[WIDE_VEC14]], <4 x i16> poison, <2 x i32> <i32 1, i32 3>
+; AVX1-NEXT:    [[TMP36:%.*]] = sext <2 x i16> [[STRIDED_VEC15]] to <2 x i32>
+; AVX1-NEXT:    [[TMP37:%.*]] = sext <2 x i16> [[STRIDED_VEC16]] to <2 x i32>
+; AVX1-NEXT:    [[TMP38:%.*]] = sext <2 x i16> [[STRIDED_VEC17]] to <2 x i32>
+; AVX1-NEXT:    [[TMP39:%.*]] = sext <2 x i16> [[STRIDED_VEC18]] to <2 x i32>
+; AVX1-NEXT:    [[TMP40:%.*]] = mul nsw <2 x i32> [[TMP36]], [[TMP20]]
+; AVX1-NEXT:    [[TMP41:%.*]] = mul nsw <2 x i32> [[TMP37]], [[TMP21]]
+; AVX1-NEXT:    [[TMP42:%.*]] = mul nsw <2 x i32> [[TMP38]], [[TMP22]]
+; AVX1-NEXT:    [[TMP43:%.*]] = mul nsw <2 x i32> [[TMP39]], [[TMP23]]
+; AVX1-NEXT:    [[TMP44:%.*]] = or i64 [[TMP4]], 1
+; AVX1-NEXT:    [[TMP45:%.*]] = or i64 [[TMP5]], 1
+; AVX1-NEXT:    [[TMP46:%.*]] = or i64 [[TMP6]], 1
+; AVX1-NEXT:    [[TMP47:%.*]] = or i64 [[TMP7]], 1
+; AVX1-NEXT:    [[TMP48:%.*]] = getelementptr inbounds i16, i16* [[S1]], i64 [[TMP44]]
+; AVX1-NEXT:    [[TMP49:%.*]] = getelementptr inbounds i16, i16* [[S1]], i64 [[TMP45]]
+; AVX1-NEXT:    [[TMP50:%.*]] = getelementptr inbounds i16, i16* [[S1]], i64 [[TMP46]]
+; AVX1-NEXT:    [[TMP51:%.*]] = getelementptr inbounds i16, i16* [[S1]], i64 [[TMP47]]
+; AVX1-NEXT:    [[TMP52:%.*]] = sext <2 x i16> [[STRIDED_VEC7]] to <2 x i32>
+; AVX1-NEXT:    [[TMP53:%.*]] = sext <2 x i16> [[STRIDED_VEC8]] to <2 x i32>
+; AVX1-NEXT:    [[TMP54:%.*]] = sext <2 x i16> [[STRIDED_VEC9]] to <2 x i32>
+; AVX1-NEXT:    [[TMP55:%.*]] = sext <2 x i16> [[STRIDED_VEC10]] to <2 x i32>
+; AVX1-NEXT:    [[TMP56:%.*]] = getelementptr inbounds i16, i16* [[S2]], i64 [[TMP44]]
+; AVX1-NEXT:    [[TMP57:%.*]] = getelementptr inbounds i16, i16* [[S2]], i64 [[TMP45]]
+; AVX1-NEXT:    [[TMP58:%.*]] = getelementptr inbounds i16, i16* [[S2]], i64 [[TMP46]]
+; AVX1-NEXT:    [[TMP59:%.*]] = getelementptr inbounds i16, i16* [[S2]], i64 [[TMP47]]
+; AVX1-NEXT:    [[TMP60:%.*]] = sext <2 x i16> [[STRIDED_VEC19]] to <2 x i32>
+; AVX1-NEXT:    [[TMP61:%.*]] = sext <2 x i16> [[STRIDED_VEC20]] to <2 x i32>
+; AVX1-NEXT:    [[TMP62:%.*]] = sext <2 x i16> [[STRIDED_VEC21]] to <2 x i32>
+; AVX1-NEXT:    [[TMP63:%.*]] = sext <2 x i16> [[STRIDED_VEC22]] to <2 x i32>
+; AVX1-NEXT:    [[TMP64:%.*]] = mul nsw <2 x i32> [[TMP60]], [[TMP52]]
+; AVX1-NEXT:    [[TMP65:%.*]] = mul nsw <2 x i32> [[TMP61]], [[TMP53]]
+; AVX1-NEXT:    [[TMP66:%.*]] = mul nsw <2 x i32> [[TMP62]], [[TMP54]]
+; AVX1-NEXT:    [[TMP67:%.*]] = mul nsw <2 x i32> [[TMP63]], [[TMP55]]
+; AVX1-NEXT:    [[TMP68:%.*]] = add nsw <2 x i32> [[TMP64]], [[TMP40]]
+; AVX1-NEXT:    [[TMP69:%.*]] = add nsw <2 x i32> [[TMP65]], [[TMP41]]
+; AVX1-NEXT:    [[TMP70:%.*]] = add nsw <2 x i32> [[TMP66]], [[TMP42]]
+; AVX1-NEXT:    [[TMP71:%.*]] = add nsw <2 x i32> [[TMP67]], [[TMP43]]
+; AVX1-NEXT:    [[TMP72:%.*]] = getelementptr inbounds i32, i32* [[D1:%.*]], i64 [[TMP0]]
+; AVX1-NEXT:    [[TMP73:%.*]] = getelementptr inbounds i32, i32* [[D1]], i64 [[TMP1]]
+; AVX1-NEXT:    [[TMP74:%.*]] = getelementptr inbounds i32, i32* [[D1]], i64 [[TMP2]]
+; AVX1-NEXT:    [[TMP75:%.*]] = getelementptr inbounds i32, i32* [[D1]], i64 [[TMP3]]
+; AVX1-NEXT:    [[TMP76:%.*]] = getelementptr inbounds i32, i32* [[TMP72]], i32 0
+; AVX1-NEXT:    [[TMP77:%.*]] = bitcast i32* [[TMP76]] to <2 x i32>*
+; AVX1-NEXT:    store <2 x i32> [[TMP68]], <2 x i32>* [[TMP77]], align 4
+; AVX1-NEXT:    [[TMP78:%.*]] = getelementptr inbounds i32, i32* [[TMP72]], i32 2
+; AVX1-NEXT:    [[TMP79:%.*]] = bitcast i32* [[TMP78]] to <2 x i32>*
+; AVX1-NEXT:    store <2 x i32> [[TMP69]], <2 x i32>* [[TMP79]], align 4
+; AVX1-NEXT:    [[TMP80:%.*]] = getelementptr inbounds i32, i32* [[TMP72]], i32 4
+; AVX1-NEXT:    [[TMP81:%.*]] = bitcast i32* [[TMP80]] to <2 x i32>*
+; AVX1-NEXT:    store <2 x i32> [[TMP70]], <2 x i32>* [[TMP81]], align 4
+; AVX1-NEXT:    [[TMP82:%.*]] = getelementptr inbounds i32, i32* [[TMP72]], i32 6
+; AVX1-NEXT:    [[TMP83:%.*]] = bitcast i32* [[TMP82]] to <2 x i32>*
+; AVX1-NEXT:    store <2 x i32> [[TMP71]], <2 x i32>* [[TMP83]], align 4
+; AVX1-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 8
+; AVX1-NEXT:    [[TMP84:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
+; AVX1-NEXT:    br i1 [[TMP84]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
 ; AVX1:       middle.block:
 ; AVX1-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[WIDE_TRIP_COUNT]], [[N_VEC]]
 ; AVX1-NEXT:    br i1 [[CMP_N]], label [[FOR_END_LOOPEXIT:%.*]], label [[SCALAR_PH]]
 ; AVX1:       scalar.ph:
 ; AVX1-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[FOR_BODY_PREHEADER]] ]
 ; AVX1-NEXT:    br label [[FOR_BODY:%.*]]
 ; AVX1:       for.body:
 ; AVX1-NEXT:    [[INDVARS_IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]
-; AVX1-NEXT:    [[TMP22:%.*]] = shl nuw nsw i64 [[INDVARS_IV]], 1
-; AVX1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[S1]], i64 [[TMP22]]
-; AVX1-NEXT:    [[TMP23:%.*]] = load i16, i16* [[ARRAYIDX]], align 2
-; AVX1-NEXT:    [[CONV:%.*]] = sext i16 [[TMP23]] to i32
-; AVX1-NEXT:    [[ARRAYIDX4:%.*]] = getelementptr inbounds i16, i16* [[S2]], i64 [[TMP22]]
-; AVX1-NEXT:    [[TMP24:%.*]] = load i16, i16* [[ARRAYIDX4]], align 2
-; AVX1-NEXT:    [[CONV5:%.*]] = sext i16 [[TMP24]] to i32
+; AVX1-NEXT:    [[TMP85:%.*]] = shl nuw nsw i64 [[INDVARS_IV]], 1
+; AVX1-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i16, i16* [[S1]], i64 [[TMP85]]
+; AVX1-NEXT:    [[TMP86:%.*]] = load i16, i16* [[ARRAYIDX]], align 2
+; AVX1-NEXT:    [[CONV:%.*]] = sext i16 [[TMP86]] to i32
+; AVX1-NEXT:    [[ARRAYIDX4:%.*]] = getelementptr inbounds i16, i16* [[S2]], i64 [[TMP85]]
+; AVX1-NEXT:    [[TMP87:%.*]] = load i16, i16* [[ARRAYIDX4]], align 2
+; AVX1-NEXT:    [[CONV5:%.*]] = sext i16 [[TMP87]] to i32
 ; AVX1-NEXT:    [[MUL6:%.*]] = mul nsw i32 [[CONV5]], [[CONV]]
-; AVX1-NEXT:    [[TMP25:%.*]] = or i64 [[TMP22]], 1
-; AVX1-NEXT:    [[ARRAYIDX10:%.*]] = getelementptr inbounds i16, i16* [[S1]], i64 [[TMP25]]
-; AVX1-NEXT:    [[TMP26:%.*]] = load i16, i16* [[ARRAYIDX10]], align 2
-; AVX1-NEXT:    [[CONV11:%.*]] = sext i16 [[TMP26]] to i32
-; AVX1-NEXT:    [[ARRAYIDX15:%.*]] = getelementptr inbounds i16, i16* [[S2]], i64 [[TMP25]]
-; AVX1-NEXT:    [[TMP27:%.*]] = load i16, i16* [[ARRAYIDX15]], align 2
-; AVX1-NEXT:    [[CONV16:%.*]] = sext i16 [[TMP27]] to i32
+; AVX1-NEXT:    [[TMP88:%.*]] = or i64 [[TMP85]], 1
+; AVX1-NEXT:    [[ARRAYIDX10:%.*]] = getelementptr inbounds i16, i16* [[S1]], i64 [[TMP88]]
+; AVX1-NEXT:    [[TMP89:%.*]] = load i16, i16* [[ARRAYIDX10]], align 2
+; AVX1-NEXT:    [[CONV11:%.*]] = sext i16 [[TMP89]] to i32
+; AVX1-NEXT:    [[ARRAYIDX15:%.*]] = getelementptr inbounds i16, i16* [[S2]], i64 [[TMP88]]
+; AVX1-NEXT:    [[TMP90:%.*]] = load i16, i16* [[ARRAYIDX15]], align 2
+; AVX1-NEXT:    [[CONV16:%.*]] = sext i16 [[TMP90]] to i32
 ; AVX1-NEXT:    [[MUL17:%.*]] = mul nsw i32 [[CONV16]], [[CONV11]]
 ; AVX1-NEXT:    [[ADD18:%.*]] = add nsw i32 [[MUL17]], [[MUL6]]
 ; AVX1-NEXT:    [[ARRAYIDX20:%.*]] = getelementptr inbounds i32, i32* [[D1]], i64 [[INDVARS_IV]]
 ; AVX1-NEXT:    store i32 [[ADD18]], i32* [[ARRAYIDX20]], align 4
 ; AVX1-NEXT:    [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
 ; AVX1-NEXT:    [[EXITCOND_NOT:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], [[WIDE_TRIP_COUNT]]
 ; AVX1-NEXT:    br i1 [[EXITCOND_NOT]], label [[FOR_END_LOOPEXIT]], label [[FOR_BODY]], !llvm.loop [[LOOP2:![0-9]+]]
 ; AVX1:       for.end.loopexit: