Index: lib/Target/ARM64/ARM64ISelLowering.cpp =================================================================== --- lib/Target/ARM64/ARM64ISelLowering.cpp +++ lib/Target/ARM64/ARM64ISelLowering.cpp @@ -3981,39 +3981,33 @@ // vector sources of the shuffle are different. static bool isEXTMask(ArrayRef M, EVT VT, bool &ReverseEXT, unsigned &Imm) { - unsigned NumElts = VT.getVectorNumElements(); - ReverseEXT = false; - - // Look for the first non-undef choice and count backwards from - // that. E.g. <-1, -1, 3, ...> means that an EXT must start at 3 - 2 = 1. This - // guarantees that at least one index is correct. - const int *FirstRealElt = - std::find_if(M.begin(), M.end(), [](int Elt) { return Elt >= 0; }); - assert(FirstRealElt != M.end() && "Completely UNDEF shuffle? Why bother?"); - Imm = *FirstRealElt - (FirstRealElt - M.begin()); + // Look for the first non-undef element. + const int *FirstRealElt = std::find_if(M.begin(), M.end(), + [](int Elt) {return Elt >= 0;}); - // If this is a VEXT shuffle, the immediate value is the index of the first - // element. The other shuffle indices must be the successive elements after - // the first one. - unsigned ExpectedElt = Imm; - for (unsigned i = 1; i < NumElts; ++i) { - // Increment the expected index. If it wraps around, it may still be - // a VEXT but the source vectors must be swapped. - ExpectedElt += 1; - if (ExpectedElt == NumElts * 2) { - ExpectedElt = 0; - ReverseEXT = true; - } + // Benefit form APInt to avoid overflow when calculating the expected element. + unsigned NumElts = VT.getVectorNumElements(); + unsigned MaskBits = APInt(32, NumElts * 2).logBase2(); + APInt ExpectedElt = APInt(MaskBits, *FirstRealElt + 1); + // The following shuffle indices must be the successive elements after the + // first real element. + const int *FirstWrongElt = std::find_if(FirstRealElt + 1, M.end(), + [&](int Elt) {return Elt != ExpectedElt++ && Elt != -1;}); + if (FirstWrongElt != M.end()) + return false; - if (M[i] < 0) - continue; // ignore UNDEF indices - if (ExpectedElt != static_cast(M[i])) - return false; - } + // The index of an EXT is the first element if it is not UNDEF. + // Watch out for the beginning UNDEFs. The EXT index should be the expected + // value of the first element. + // E.g. <-1, -1, 3, ...> is treated as <1, 2, 3, ...>. + // <-1, -1, 0, 1, ...> is treated as . IDX is + // equal to the ExpectedElt. For this case, ExpectedElt is (NumElts*2 - 2). + Imm = (M[0] >= 0) ? static_cast(M[0]) : ExpectedElt.getZExtValue(); - // Adjust the index value if the source operands will be swapped. - if (ReverseEXT) + if (Imm >= NumElts) { + ReverseEXT = true; Imm -= NumElts; + } return true; } Index: test/CodeGen/ARM64/ext.ll =================================================================== --- test/CodeGen/ARM64/ext.ll +++ test/CodeGen/ARM64/ext.ll @@ -65,6 +65,15 @@ ret <8 x i8> %tmp3 } +define <8 x i8> @test_vextd_undef2(<8 x i8>* %A, <8 x i8>* %B) nounwind { +;CHECK-LABEL: test_vextd_undef2: +;CHECK: {{ext.8b.*#6}} + %tmp1 = load <8 x i8>* %A + %tmp2 = load <8 x i8>* %B + %tmp3 = shufflevector <8 x i8> %tmp1, <8 x i8> %tmp2, <8 x i32> + ret <8 x i8> %tmp3 +} + define <16 x i8> @test_vextRq_undef(<16 x i8>* %A, <16 x i8>* %B) nounwind { ;CHECK-LABEL: test_vextRq_undef: ;CHECK: {{ext.16b.*#7}} @@ -74,6 +83,14 @@ ret <16 x i8> %tmp3 } +define <8 x i16> @test_vextRq_undef2(<8 x i16>* %A) nounwind { +;CHECK-LABEL: test_vextRq_undef2: +;CHECK: {{ext.16b.*#10}} + %tmp1 = load <8 x i16>* %A + %vext = shufflevector <8 x i16> %tmp1, <8 x i16> undef, <8 x i32> + ret <8 x i16> %vext; +} + ; Tests for ReconstructShuffle function. Indices have to be carefully ; chosen to reach lowering phase as a BUILD_VECTOR.