[llvm] [SLP]Keep originals when uniques do not reduce register uses (PR #218294)

[via llvm-commits]

llvmorg-github-actions[bot] wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-vectorizers

Author: Alexey Bataev (alexey-bataev)

<details>
<summary>Changes</summary>

When the vector of unqiues occupies the same register width as the
original (counting a sub-register vector as a full one), packing only
adds the reshuffle cost; keep the original vector. Loads are excluded
since packing may instead enable a wider contiguous load.


---
Full diff: https://github.com/llvm/llvm-project/pull/218294.diff


2 Files Affected:

- (modified) llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp (+18-1) 
- (modified) llvm/test/Transforms/SLPVectorizer/X86/shl-to-add-transformation5.ll (+5-5) 


``````````diff
diff --git a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
index 8d49f8b9fda9d..476ea4e3b2ee2 100644
--- a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
+++ b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
@@ -10981,8 +10981,25 @@ static bool tryToFindDuplicates(SmallVectorImpl<Value *> &VL,
                                          /*Extract=*/false, CostKind,
                                          AreAllValuesNonConst, UniqueValues);
     UniquesCost += ReusesCost;
-    if (UniquesCost <= InsertsCost)
+    if (UniquesCost <= InsertsCost) {
+      // Packing to a vector that occupies the same register width as the
+      // original only adds the reshuffle cost; keep the originals. Loads may
+      // instead benefit from a wider contiguous load.
+      if (S && S.getOpcode() != Instruction::Load) {
+        unsigned EltBits = S.getMainOp()
+                               ->getDataLayout()
+                               .getTypeSizeInBits(ScalarTy)
+                               .getFixedValue();
+        unsigned MinVF = R.getMinVF(EltBits);
+        auto RegWidth = [&](unsigned N) {
+          return std::max(getFullVectorNumberOfElements(TTI, ScalarTy, N),
+                          MinVF);
+        };
+        if (RegWidth(NumUniqueScalarValues) >= RegWidth(VL.size()))
+          return std::make_pair(true, true);
+      }
       return std::make_pair(true, false);
+    }
     InstructionCost CostDiff = UniquesCost - InsertsCost;
     if (CostDiff < TTI::TCC_Expensive ||
         (R.getTreeSize() == 0 && R.isReductionTree() &&
diff --git a/llvm/test/Transforms/SLPVectorizer/X86/shl-to-add-transformation5.ll b/llvm/test/Transforms/SLPVectorizer/X86/shl-to-add-transformation5.ll
index 597fed290b9b0..6d9ef6f338af5 100644
--- a/llvm/test/Transforms/SLPVectorizer/X86/shl-to-add-transformation5.ll
+++ b/llvm/test/Transforms/SLPVectorizer/X86/shl-to-add-transformation5.ll
@@ -20,11 +20,11 @@ define i32 @test(i32 %0, i32 %1) {
 ; CHECK-NEXT:    [[TMP10:%.*]] = insertelement <8 x i32> [[TMP9]], i32 [[TMP0]], i64 2
 ; CHECK-NEXT:    [[TMP11:%.*]] = insertelement <8 x i32> [[TMP10]], i32 [[SUB120_3]], i64 5
 ; CHECK-NEXT:    [[TMP12:%.*]] = shufflevector <8 x i32> [[TMP11]], <8 x i32> poison, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 5, i32 6, i32 7>
-; CHECK-NEXT:    [[TMP13:%.*]] = shufflevector <8 x i32> [[TMP11]], <8 x i32> poison, <5 x i32> <i32 2, i32 poison, i32 poison, i32 5, i32 poison>
-; CHECK-NEXT:    [[TMP14:%.*]] = insertelement <5 x i32> [[TMP13]], i32 [[DOTNEG_NEG]], i64 1
-; CHECK-NEXT:    [[TMP15:%.*]] = shufflevector <2 x i32> [[TMP7]], <2 x i32> poison, <5 x i32> <i32 0, i32 poison, i32 1, i32 poison, i32 poison>
-; CHECK-NEXT:    [[TMP16:%.*]] = shufflevector <5 x i32> [[TMP14]], <5 x i32> [[TMP15]], <5 x i32> <i32 0, i32 1, i32 5, i32 3, i32 7>
-; CHECK-NEXT:    [[TMP17:%.*]] = shufflevector <5 x i32> [[TMP16]], <5 x i32> poison, <8 x i32> <i32 0, i32 1, i32 0, i32 2, i32 0, i32 3, i32 4, i32 2>
+; CHECK-NEXT:    [[TMP13:%.*]] = shufflevector <8 x i32> [[TMP11]], <8 x i32> poison, <8 x i32> <i32 2, i32 poison, i32 2, i32 poison, i32 2, i32 5, i32 poison, i32 poison>
+; CHECK-NEXT:    [[TMP14:%.*]] = insertelement <8 x i32> [[TMP13]], i32 [[DOTNEG_NEG]], i64 1
+; CHECK-NEXT:    [[TMP15:%.*]] = shufflevector <2 x i32> [[TMP7]], <2 x i32> poison, <8 x i32> <i32 0, i32 poison, i32 poison, i32 1, i32 poison, i32 poison, i32 poison, i32 poison>
+; CHECK-NEXT:    [[TMP16:%.*]] = shufflevector <8 x i32> [[TMP14]], <8 x i32> [[TMP15]], <8 x i32> <i32 0, i32 1, i32 2, i32 8, i32 4, i32 5, i32 11, i32 7>
+; CHECK-NEXT:    [[TMP17:%.*]] = shufflevector <8 x i32> [[TMP16]], <8 x i32> poison, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 3>
 ; CHECK-NEXT:    [[TMP18:%.*]] = sub <8 x i32> [[TMP12]], [[TMP17]]
 ; CHECK-NEXT:    [[TMP19:%.*]] = add <8 x i32> [[TMP12]], [[TMP17]]
 ; CHECK-NEXT:    [[TMP20:%.*]] = shufflevector <8 x i32> [[TMP18]], <8 x i32> [[TMP19]], <8 x i32> <i32 0, i32 9, i32 10, i32 3, i32 12, i32 13, i32 6, i32 7>

``````````

</details>


https://github.com/llvm/llvm-project/pull/218294