[PATCH] D71220: [InstCombine] replace shuffle's insertelement operand if inserted scalar is not demanded

Sanjay Patel via Phabricator via llvm-commits llvm-commits at lists.llvm.org
Mon Dec 9 11:48:25 PST 2019


spatel created this revision.
spatel added reviewers: foad, efriedma, nlopes, aqjune.
Herald added subscribers: hiraditya, mcrosier.
Herald added a project: LLVM.

This pattern is noted as a regression from:
D70246 <https://reviews.llvm.org/D70246> 
...where we removed an over-aggressive shuffle simplification.

SimplifyDemandedVectorElts fails to catch this case when the insert has multiple uses, so I'm proposing to pattern match the minimal sequence directly. This fold does not conflict with any of our current shuffle undef/poison semantics AFAICT.


https://reviews.llvm.org/D71220

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
@@ -670,7 +670,7 @@
 ; CHECK-LABEL: @insert_undemanded_element_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:%.*]], <4 x i32> <i32 0, i32 7, i32 1, i32 4>
+; CHECK-NEXT:    [[S:%.*]] = shufflevector <4 x float> [[X]], <4 x float> [[Y:%.*]], <4 x i32> <i32 0, i32 7, i32 1, i32 4>
 ; CHECK-NEXT:    ret <4 x float> [[S]]
 ;
   %ins = insertelement <4 x float> %x, float 42.0, i32 3
@@ -683,7 +683,7 @@
 ; CHECK-LABEL: @insert_undemanded_element_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]], <4 x i32> <i32 3, i32 2, i32 1, i32 4>
+; CHECK-NEXT:    [[S:%.*]] = shufflevector <4 x float> [[Y:%.*]], <4 x float> [[X]], <4 x i32> <i32 3, i32 2, i32 1, i32 4>
 ; CHECK-NEXT:    ret <4 x float> [[S]]
 ;
   %ins = insertelement <4 x float> %x, float 42.0, i32 3
@@ -692,6 +692,8 @@
   ret <4 x float> %s
 }
 
+; Negative test - shuffle chooses the inserted constant.
+
 define <4 x float> @insert_demanded_element_op0(<4 x float> %x, <4 x float> %y) {
 ; CHECK-LABEL: @insert_demanded_element_op0(
 ; CHECK-NEXT:    [[INS:%.*]] = insertelement <4 x float> [[X:%.*]], float 4.200000e+01, i32 3
@@ -705,6 +707,8 @@
   ret <4 x float> %s
 }
 
+; Negative test - shuffle chooses the inserted constant.
+
 define <4 x float> @insert_demanded_element_op1(<4 x float> %x, <4 x float> %y) {
 ; CHECK-LABEL: @insert_demanded_element_op1(
 ; CHECK-NEXT:    [[INS:%.*]] = insertelement <4 x float> [[X:%.*]], float 4.300000e+01, i32 3
Index: llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
===================================================================
--- llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
+++ llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
@@ -1741,7 +1741,8 @@
   return new ShuffleVectorInst(X, Y, ConstantVector::get(NewMask));
 }
 
-/// Try to replace a shuffle with an insertelement.
+/// Try to replace a shuffle with an insertelement or try to replace a shuffle
+/// operand with the operand of an insertelement.
 static Instruction *foldShuffleWithInsert(ShuffleVectorInst &Shuf) {
   Value *V0 = Shuf.getOperand(0), *V1 = Shuf.getOperand(1);
   SmallVector<int, 16> Mask = Shuf.getShuffleMask();
@@ -1753,6 +1754,31 @@
   if (NumElts != (int)(V0->getType()->getVectorNumElements()))
     return nullptr;
 
+  // This is a specialization of a fold in SimplifyDemandedVectorElts. We may
+  // not be able to handle it there if the insertelement has >1 use.
+  // If the shuffle has an insertelement operand but does not choose the
+  // inserted scalar element from that value, then we can replace that shuffle
+  // operand with the source vector of the insertelement.
+  Value *X;
+  uint64_t IdxC;
+  if (match(V0, m_InsertElement(m_Value(X), m_Value(), m_ConstantInt(IdxC)))) {
+    // shuf (inselt X, ?, IdxC), ?, Mask --> shuf X, ?, Mask
+    if (none_of(Mask, [IdxC](int MaskElt) { return MaskElt == (int)IdxC; })) {
+      Shuf.setOperand(0, X);
+      return &Shuf;
+    }
+  }
+  if (match(V1, m_InsertElement(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;
+    // shuf ?, (inselt X, ?, IdxC), Mask --> shuf ?, X, Mask
+    if (none_of(Mask, [IdxC](int MaskElt) { return MaskElt == (int)IdxC; })) {
+      Shuf.setOperand(1, X);
+      return &Shuf;
+    }
+  }
+
   // shuffle (insert ?, Scalar, IndexC), V1, Mask --> insert V1, Scalar, IndexC'
   auto isShufflingScalarIntoOp1 = [&](Value *&Scalar, ConstantInt *&IndexC) {
     // We need an insertelement with a constant index.


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D71220.232905.patch
Type: text/x-patch
Size: 4236 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20191209/77c46def/attachment.bin>


More information about the llvm-commits mailing list