[InstCombine] insert a new shuffle before its uses (PR26015)

[spatel]

Summary

Although this solves the test case in PR26015:
https://llvm.org/bugs/show_bug.cgi?id=26015

...I suspect this is not the best solution. I think we want to insert the new shuffle just ahead of the earliest ExtractElementInst that we're replacing, but I don't know how that should be implemented. Is there a way to compare/record the position of each extract that we encounter in the loop?

[• tstellarAMD]

I can't comment on the correctness of the implementation, but it fixes the regression reported in PR26015. Thanks for the quick fix.

llvm/trunk/lib/Transforms/InstCombine/InstCombineVectorOps.cpp

for (unsigned i = 0; i < NumExtElts; ++i)
     ExtendMask.push_back(ConstantInt::get(IntType, i));
   for (unsigned i = NumExtElts; i < NumInsElts; ++i)
     ExtendMask.push_back(UndefValue::get(IntType));
 
   Value *ExtVecOp = ExtElt->getVectorOperand();
   auto *WideVec = new ShuffleVectorInst(ExtVecOp, UndefValue::get(ExtVecType),
                                         ConstantVector::get(ExtendMask));
 
-  // Replace all extracts from the original narrow vector with extracts from
-  // the new wide vector.
-  WideVec->insertBefore(ExtElt);
+  // Insert the new shuffle after the vector operand of the extract is defined
+  // or at the start of the basic block, so any subsequent extracts can use it.
+  bool ReplaceAllExtUsers;
+  if (auto *ExtVecOpInst = dyn_cast<Instruction>(ExtVecOp)) {
+    WideVec->insertAfter(ExtVecOpInst);
+    ReplaceAllExtUsers = true;
+  } else {
+    // TODO: Insert at start of function, so it's always safe to replace all?
+    IC.InsertNewInstWith(WideVec, *ExtElt->getParent()->getFirstInsertionPt());
+    ReplaceAllExtUsers = false;
+  }
+
+  // Replace extracts from the original narrow vector with extracts from the new
+  // wide vector.
   for (User *U : ExtVecOp->users()) {
-    if (ExtractElementInst *OldExt = dyn_cast<ExtractElementInst>(U)) {
-      auto *NewExt = ExtractElementInst::Create(WideVec, OldExt->getOperand(1));
-      NewExt->insertAfter(WideVec);
-      IC.ReplaceInstUsesWith(*OldExt, NewExt);
-    }
-  }
-}
+    ExtractElementInst *OldExt = dyn_cast<ExtractElementInst>(U);
+    if (!OldExt ||
+        (!ReplaceAllExtUsers && OldExt->getParent() != WideVec->getParent()))
+      continue;
 
 /// We are building a shuffle to create V, which is a sequence of insertelement,
 /// extractelement pairs. If PermittedRHS is set, then we must either use it or
 /// not rely on the second vector source. Return a std::pair containing the
 /// left and right vectors of the proposed shuffle (or 0), and set the Mask
 /// parameter as required.
 ///
 /// Note: we intentionally don't try to fold earlier shuffles since they have

llvm/trunk/test/Transforms/InstCombine/insert-extract-shuffle.ll

 ; CHECK-NEXT: %[[WIDEVEC:.*]] = shufflevector <2 x float> %ext, <2 x float> undef, <8 x i32> <i32 0, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
 ; CHECK-NEXT: shufflevector <8 x float> %ins, <8 x float> %[[WIDEVEC]], <8 x i32> <i32 0, i32 1, i32 8, i32 3, i32 4, i32 5, i32 6, i32 7>
 ; CHECK-NEXT: ret <8 x float> %i1
   %e1 = extractelement <2 x float> %ext, i32 0
   %i1 = insertelement <8 x float> %ins, float %e1, i32 2
   ret <8 x float> %i1
 }
 
+; PR26015: https://llvm.org/bugs/show_bug.cgi?id=26015
+; The widening shuffle must be inserted before any uses.
+
+define <8 x i16> @pr26015(<4 x i16> %t0) {
+; CHECK-LABEL: @pr26015(
+; CHECK-NEXT:  %[[WIDEVEC:.*]] = shufflevector <4 x i16> %t0, <4 x i16> undef, <8 x i32> <i32 undef, i32 undef, i32 undef, i32 3, i32 undef, i32 undef, i32 undef, i32 undef>
+; CHECK-NEXT:  %[[EXT:.*]] = extractelement <4 x i16> %t0, i32 2
+; CHECK-NEXT:  %t2 = insertelement <8 x i16> <i16 0, i16 0, i16 0, i16 undef, i16 0, i16 0, i16 undef, i16 undef>, i16 %[[EXT]], i32 3
+; CHECK-NEXT:  %t3 = insertelement <8 x i16> %t2, i16 0, i32 6
+; CHECK-NEXT:  %t5 = shufflevector <8 x i16> %t3, <8 x i16> %[[WIDEVEC]], <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 11>
+; CHECK-NEXT:  ret <8 x i16> %t5
+  %t1 = extractelement <4 x i16> %t0, i32 2
+  %t2 = insertelement <8 x i16> zeroinitializer, i16 %t1, i32 3
+  %t3 = insertelement <8 x i16> %t2, i16 0, i32 6
+  %t4 = extractelement <4 x i16> %t0, i32 3
+  %t5 = insertelement <8 x i16> %t3, i16 %t4, i32 7
+  ret <8 x i16> %t5
+}
+
+; PR25999: https://llvm.org/bugs/show_bug.cgi?id=25999
+; TODO: The widening shuffle could be inserted at the start of the function to allow the first extract to use it.
+
+define <8 x i16> @pr25999(<4 x i16> %t0, i1 %b) {
+; CHECK-LABEL: @pr25999(
+; CHECK-NEXT:  %t1 = extractelement <4 x i16> %t0, i32 2
+; CHECK-NEXT:  br i1 %b, label %if, label %end
+; CHECK:       if:
+; CHECK-NEXT:  %[[WIDEVEC:.*]] = shufflevector <4 x i16> %t0, <4 x i16> undef, <8 x i32> <i32 undef, i32 undef, i32 undef, i32 3, i32 undef, i32 undef, i32 undef, i32 undef>
+; CHECK-NEXT:  %t2 = insertelement <8 x i16> <i16 0, i16 0, i16 0, i16 undef, i16 0, i16 0, i16 undef, i16 undef>, i16 %t1, i32 3
+; CHECK-NEXT:  %t3 = insertelement <8 x i16> %t2, i16 0, i32 6
+; CHECK-NEXT:  %t5 = shufflevector <8 x i16> %t3, <8 x i16> %[[WIDEVEC]], <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 11>
+; CHECK-NEXT:  ret <8 x i16> %t5
+; CHECK:       end:
+; CHECK-NEXT:  %a1 = add i16 %t1, 4
+; CHECK-NEXT:  %t6 = insertelement <8 x i16> <i16 undef, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0>, i16 %a1, i32 0
+; CHECK-NEXT:  ret <8 x i16> %t6
+
+  %t1 = extractelement <4 x i16> %t0, i32 2
+  br i1 %b, label %if, label %end
+
+if:
+  %t2 = insertelement <8 x i16> zeroinitializer, i16 %t1, i32 3
+  %t3 = insertelement <8 x i16> %t2, i16 0, i32 6
+  %t4 = extractelement <4 x i16> %t0, i32 3
+  %t5 = insertelement <8 x i16> %t3, i16 %t4, i32 7
+  ret <8 x i16> %t5
+
+end:
+  %a1 = add i16 %t1, 4
+  %t6 = insertelement <8 x i16> zeroinitializer, i16 %a1, i32 0
+  ret <8 x i16> %t6
+}
+