[InstSimplify] Properly constrain {insert,extract}_subvector intrinsic fold

[joechrisellis]

Summary

The previous rule:

(insert_vector _, (extract_vector X, 0), 0) -> X

is not quite correct. The correct fold should be:

(insert_vector Y, (extract_vector X, 0), 0) -> X
where: Y is X, or Y is undef

This commit updates the pattern.

[joechrisellis]

Remove unused %pg variable in test.

[peterwaller-arm]

LGTM, but I missed the issue on the first iteration so please wait for a second accept.

llvm/lib/Analysis/InstructionSimplify.cpp

APInt ScaledOne =
         APInt::getOneBitSet(ReturnType->getScalarSizeInBits(),
                             cast<ConstantInt>(Op2)->getZExtValue());
     if (ScaledOne.isNonNegative() && match(Op1, m_SpecificInt(ScaledOne)))
       return Op0;
 
     return nullptr;
   }
   case Intrinsic::experimental_vector_insert: {
+    Value *Vec = Call->getArgOperand(0);
     Value *SubVec = Call->getArgOperand(1);
     Value *Idx = Call->getArgOperand(2);
     Type *ReturnType = F->getReturnType();
 
-    // (insert_vector _, (extract_vector X, 0), 0) -> X
+    // (insert_vector Y, (extract_vector X, 0), 0) -> X
+    // where: Y is X, or Y is undef
     unsigned IdxN = cast<ConstantInt>(Idx)->getZExtValue();
     Value *X = nullptr;
     if (match(SubVec, m_Intrinsic<Intrinsic::experimental_vector_extract>(
                           m_Value(X), m_Zero())) &&
-        IdxN == 0 && X->getType() == ReturnType)
+        (Q.isUndefValue(Vec) || Vec == X) && IdxN == 0 &&
+        X->getType() == ReturnType)
       return X;
 
     return nullptr;
   }
   default:
     return nullptr;
   }
 }

llvm/test/Transforms/InstSimplify/insert-vector.ll

 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
 ; RUN: opt < %s -instsimplify -S | FileCheck %s
 
-define <vscale x 16 x i8> @redundant_extract_insert_chain(<vscale x 16 x i1> %pg, <vscale x 16 x i8> %x) {
+define <vscale x 16 x i8> @redundant_extract_insert_chain(<vscale x 16 x i8> %x) {
 ; CHECK-LABEL: @redundant_extract_insert_chain(
 ; CHECK-NEXT:    ret <vscale x 16 x i8> [[X:%.*]]
 ;
   %extracted = call <32 x i8> @llvm.experimental.vector.extract.v32i8.nxv16i8(<vscale x 16 x i8> %x, i64 0)
   %inserted = call <vscale x 16 x i8> @llvm.experimental.vector.insert.nxv16i8.v32i8(<vscale x 16 x i8> undef, <32 x i8> %extracted, i64 0)
   ret <vscale x 16 x i8> %inserted
 }
 
-define <vscale x 16 x i8> @non_redundant_extract_insert_chain(<vscale x 16 x i1> %pg, <vscale x 32 x i8> %x) {
-; CHECK-LABEL: @non_redundant_extract_insert_chain(
+define <vscale x 16 x i8> @non_redundant_extract_insert_chain_0(<vscale x 32 x i8> %x) {
+; CHECK-LABEL: @non_redundant_extract_insert_chain_0(
 ; CHECK-NEXT:    [[EXTRACTED:%.*]] = call <32 x i8> @llvm.experimental.vector.extract.v32i8.nxv32i8(<vscale x 32 x i8> [[X:%.*]], i64 0)
 ; CHECK-NEXT:    [[INSERTED:%.*]] = call <vscale x 16 x i8> @llvm.experimental.vector.insert.nxv16i8.v32i8(<vscale x 16 x i8> undef, <32 x i8> [[EXTRACTED]], i64 0)
 ; CHECK-NEXT:    ret <vscale x 16 x i8> [[INSERTED]]
 ;
   %extracted = call <32 x i8> @llvm.experimental.vector.extract.v32i8.nxv32i8(<vscale x 32 x i8> %x, i64 0)
   %inserted = call <vscale x 16 x i8> @llvm.experimental.vector.insert.nxv16i8.v32i8(<vscale x 16 x i8> undef, <32 x i8> %extracted, i64 0)
   ret <vscale x 16 x i8> %inserted
 }
 
+define <vscale x 16 x i8> @non_redundant_extract_insert_chain_1(<vscale x 16 x i8> %x, <vscale x 16 x i8> %y) {
+; CHECK-LABEL: @non_redundant_extract_insert_chain_1(
+; CHECK-NEXT:    [[EXTRACTED:%.*]] = call <32 x i8> @llvm.experimental.vector.extract.v32i8.nxv16i8(<vscale x 16 x i8> [[X:%.*]], i64 0)
+; CHECK-NEXT:    [[INSERTED:%.*]] = call <vscale x 16 x i8> @llvm.experimental.vector.insert.nxv16i8.v32i8(<vscale x 16 x i8> [[Y:%.*]], <32 x i8> [[EXTRACTED]], i64 0)
+; CHECK-NEXT:    ret <vscale x 16 x i8> [[INSERTED]]
+;
+  %extracted = call <32 x i8> @llvm.experimental.vector.extract.v32i8.nxv16i8(<vscale x 16 x i8> %x, i64 0)
+  %inserted = call <vscale x 16 x i8> @llvm.experimental.vector.insert.nxv16i8.v32i8(<vscale x 16 x i8> %y, <32 x i8> %extracted, i64 0)
+  ret <vscale x 16 x i8> %inserted
+}
+
 declare <32 x i8> @llvm.experimental.vector.extract.v32i8.nxv16i8(<vscale x 16 x i8>, i64)
 declare <32 x i8> @llvm.experimental.vector.extract.v32i8.nxv32i8(<vscale x 32 x i8>, i64)
 declare <vscale x 16 x i8> @llvm.experimental.vector.insert.nxv16i8.v32i8(<vscale x 16 x i8>, <32 x i8>, i64)