[PATCH] D112318: [InstCombine] Extend pattern to replace shuffle's insertelement operand

[Piotr Sobczak via Phabricator via llvm-commits]

piotr created this revision.
Herald added a subscriber: hiraditya.
piotr requested review of this revision.
Herald added a project: LLVM.
Herald added a subscriber: llvm-commits.

In D71220 <https://reviews.llvm.org/D71220> a pattern was added to replace shuffle's insertelement operand
if inserted scalar is not demanded. The pattern was added only for
the case where the shuffle's mask size is equal to element's vector size.
However, that condition is not required because the pattern does not
change the shuffle vector size.

This patch extends the pattern to also include cases where shuffle's mask
size is not equal to element's vector size.


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D112318

Files:
  llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
  llvm/test/Transforms/InstCombine/insert-extract-shuffle.ll


Index: llvm/test/Transforms/InstCombine/insert-extract-shuffle.ll
===================================================================
--- llvm/test/Transforms/InstCombine/insert-extract-shuffle.ll
+++ llvm/test/Transforms/InstCombine/insert-extract-shuffle.ll
@@ -696,7 +696,7 @@
 ; CHECK-LABEL: @insert_undemanded_element_unequal_length_op0(
 ; CHECK-NEXT:    [[INS:%.*]] = insertelement <4 x float> [[X:%.*]], float 4.200000e+01, i32 3
 ; CHECK-NEXT:    call void @use(<4 x float> [[INS]])
-; CHECK-NEXT:    [[S:%.*]] = shufflevector <4 x float> [[INS]], <4 x float> [[Y:%.*]], <5 x i32> <i32 undef, i32 0, i32 7, i32 1, i32 4>
+; CHECK-NEXT:    [[S:%.*]] = shufflevector <4 x float> [[X]], <4 x float> [[Y:%.*]], <5 x i32> <i32 undef, i32 0, i32 7, i32 1, i32 4>
 ; CHECK-NEXT:    ret <5 x float> [[S]]
 ;
   %ins = insertelement <4 x float> %x, float 42.0, i32 3
@@ -709,7 +709,7 @@
 ; CHECK-LABEL: @insert_undemanded_element_unequal_length_op1(
 ; CHECK-NEXT:    [[INS:%.*]] = insertelement <4 x float> [[X:%.*]], float 4.200000e+01, i32 3
 ; CHECK-NEXT:    call void @use(<4 x float> [[INS]])
-; CHECK-NEXT:    [[S:%.*]] = shufflevector <4 x float> [[Y:%.*]], <4 x float> [[INS]], <5 x i32> <i32 undef, i32 3, i32 2, i32 1, i32 4>
+; CHECK-NEXT:    [[S:%.*]] = shufflevector <4 x float> [[Y:%.*]], <4 x float> [[X]], <5 x i32> <i32 undef, i32 3, i32 2, i32 1, i32 4>
 ; CHECK-NEXT:    ret <5 x float> [[S]]
 ;
   %ins = insertelement <4 x float> %x, float 42.0, i32 3
Index: llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
===================================================================
--- llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
+++ llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
@@ -2268,12 +2268,8 @@
   SmallVector<int, 16> Mask;
   Shuf.getShuffleMask(Mask);
 
-  // The shuffle must not change vector sizes.
-  // TODO: This restriction could be removed if the insert has only one use
-  //       (because the transform would require a new length-changing shuffle).
   int NumElts = Mask.size();
-  if (NumElts != (int)(cast<FixedVectorType>(V0->getType())->getNumElements()))
-    return nullptr;
+  int InpNumElts = cast<FixedVectorType>(V0->getType())->getNumElements();
 
   // This is a specialization of a fold in SimplifyDemandedVectorElts. We may
   // not be able to handle it there if the insertelement has >1 use.
@@ -2290,11 +2286,16 @@
   if (match(V1, m_InsertElt(m_Value(X), m_Value(), m_ConstantInt(IdxC)))) {
     // Offset the index constant by the vector width because we are checking for
     // accesses to the 2nd vector input of the shuffle.
-    IdxC += NumElts;
+    IdxC += InpNumElts;
     // shuf ?, (inselt X, ?, IdxC), Mask --> shuf ?, X, Mask
     if (!is_contained(Mask, (int)IdxC))
       return IC.replaceOperand(Shuf, 1, X);
   }
+  // The shuffle must not change vector sizes.
+  // TODO: This restriction could be removed if the insert has only one use
+  //       (because the transform would require a new length-changing shuffle).
+  if (NumElts != InpNumElts)
+    return nullptr;
 
   // shuffle (insert ?, Scalar, IndexC), V1, Mask --> insert V1, Scalar, IndexC'
   auto isShufflingScalarIntoOp1 = [&](Value *&Scalar, ConstantInt *&IndexC) {


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D112318.381555.patch
Type: text/x-patch
Size: 3238 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20211022/8ba5938d/attachment.bin>