1; RUN: llc < %s -mtriple=armv7-apple-darwin | FileCheck %s 2 3; PR7158 4define i32 @test_pr7158() nounwind { 5bb.nph55.bb.nph55.split_crit_edge: 6 br label %bb3 7 8bb3: ; preds = %bb3, %bb.nph55.bb.nph55.split_crit_edge 9 br i1 undef, label %bb.i19, label %bb3 10 11bb.i19: ; preds = %bb.i19, %bb3 12 %0 = insertelement <4 x float> undef, float undef, i32 3 ; <<4 x float>> [#uses=3] 13 %1 = fmul <4 x float> %0, %0 ; <<4 x float>> [#uses=1] 14 %2 = bitcast <4 x float> %1 to <2 x double> ; <<2 x double>> [#uses=0] 15 %3 = fmul <4 x float> %0, undef ; <<4 x float>> [#uses=0] 16 br label %bb.i19 17} 18 19; Check that the DAG combiner does not arbitrarily modify BUILD_VECTORs 20; after legalization. 21define void @test_illegal_build_vector() nounwind { 22entry: 23 store <2 x i64> undef, <2 x i64>* undef, align 16 24 %0 = load <16 x i8>, <16 x i8>* undef, align 16 ; <<16 x i8>> [#uses=1] 25 %1 = or <16 x i8> zeroinitializer, %0 ; <<16 x i8>> [#uses=1] 26 store <16 x i8> %1, <16 x i8>* undef, align 16 27 ret void 28} 29 30; PR22678 31; Check CONCAT_VECTORS DAG combiner pass doesn't introduce illegal types. 32define void @test_pr22678() { 33 %1 = fptoui <16 x float> undef to <16 x i8> 34 store <16 x i8> %1, <16 x i8>* undef 35 ret void 36} 37 38; Radar 8407927: Make sure that VMOVRRD gets optimized away when the result is 39; converted back to be used as a vector type. 40; CHECK-LABEL: test_vmovrrd_combine: 41define <4 x i32> @test_vmovrrd_combine() nounwind { 42entry: 43 br i1 undef, label %bb1, label %bb2 44 45bb1: 46 %0 = bitcast <2 x i64> zeroinitializer to <2 x double> 47 %1 = extractelement <2 x double> %0, i32 0 48 %2 = bitcast double %1 to i64 49 %3 = insertelement <1 x i64> undef, i64 %2, i32 0 50; CHECK-NOT: vmov s 51; CHECK: vext.8 52 %4 = shufflevector <1 x i64> %3, <1 x i64> undef, <2 x i32> <i32 0, i32 1> 53 %tmp2006.3 = bitcast <2 x i64> %4 to <16 x i8> 54 %5 = shufflevector <16 x i8> %tmp2006.3, <16 x i8> undef, <16 x i32> <i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19> 55 %tmp2004.3 = bitcast <16 x i8> %5 to <4 x i32> 56 br i1 undef, label %bb2, label %bb1 57 58bb2: 59 %result = phi <4 x i32> [ undef, %entry ], [ %tmp2004.3, %bb1 ] 60 ret <4 x i32> %result 61} 62 63; Test trying to do a ShiftCombine on illegal types. 64; The vector should be split first. 65define void @lshrIllegalType(<8 x i32>* %A) nounwind { 66 %tmp1 = load <8 x i32>, <8 x i32>* %A 67 %tmp2 = lshr <8 x i32> %tmp1, < i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3> 68 store <8 x i32> %tmp2, <8 x i32>* %A 69 ret void 70} 71 72; Test folding a binary vector operation with constant BUILD_VECTOR 73; operands with i16 elements. 74define void @test_i16_constant_fold() nounwind optsize { 75entry: 76 %0 = sext <4 x i1> zeroinitializer to <4 x i16> 77 %1 = add <4 x i16> %0, zeroinitializer 78 %2 = shufflevector <4 x i16> %1, <4 x i16> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7> 79 %3 = add <8 x i16> %2, <i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1> 80 %4 = trunc <8 x i16> %3 to <8 x i8> 81 tail call void @llvm.arm.neon.vst1.p0i8.v8i8(i8* undef, <8 x i8> %4, i32 1) 82 unreachable 83} 84 85declare void @llvm.arm.neon.vst1.p0i8.v8i8(i8*, <8 x i8>, i32) nounwind 86 87; Test that loads and stores of i64 vector elements are handled as f64 values 88; so they are not split up into i32 values. Radar 8755338. 89define void @i64_buildvector(i64* %ptr, <2 x i64>* %vp) nounwind { 90; CHECK: i64_buildvector 91; CHECK: vldr 92 %t0 = load i64, i64* %ptr, align 4 93 %t1 = insertelement <2 x i64> undef, i64 %t0, i32 0 94 store <2 x i64> %t1, <2 x i64>* %vp 95 ret void 96} 97 98define void @i64_insertelement(i64* %ptr, <2 x i64>* %vp) nounwind { 99; CHECK: i64_insertelement 100; CHECK: vldr 101 %t0 = load i64, i64* %ptr, align 4 102 %vec = load <2 x i64>, <2 x i64>* %vp 103 %t1 = insertelement <2 x i64> %vec, i64 %t0, i32 0 104 store <2 x i64> %t1, <2 x i64>* %vp 105 ret void 106} 107 108define void @i64_extractelement(i64* %ptr, <2 x i64>* %vp) nounwind { 109; CHECK: i64_extractelement 110; CHECK: vstr 111 %vec = load <2 x i64>, <2 x i64>* %vp 112 %t1 = extractelement <2 x i64> %vec, i32 0 113 store i64 %t1, i64* %ptr 114 ret void 115} 116 117; Test trying to do a AND Combine on illegal types. 118define void @andVec(<3 x i8>* %A) nounwind { 119 %tmp = load <3 x i8>, <3 x i8>* %A, align 4 120 %and = and <3 x i8> %tmp, <i8 7, i8 7, i8 7> 121 store <3 x i8> %and, <3 x i8>* %A 122 ret void 123} 124 125 126; Test trying to do an OR Combine on illegal types. 127define void @orVec(<3 x i8>* %A) nounwind { 128 %tmp = load <3 x i8>, <3 x i8>* %A, align 4 129 %or = or <3 x i8> %tmp, <i8 7, i8 7, i8 7> 130 store <3 x i8> %or, <3 x i8>* %A 131 ret void 132} 133 134; The following test was hitting an assertion in the DAG combiner when 135; constant folding the multiply because the "sext undef" was translated to 136; a BUILD_VECTOR with i32 0 operands, which did not match the i16 operands 137; of the other BUILD_VECTOR. 138define i16 @foldBuildVectors() { 139 %1 = sext <8 x i8> undef to <8 x i16> 140 %2 = mul <8 x i16> %1, <i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255, i16 255> 141 %3 = extractelement <8 x i16> %2, i32 0 142 ret i16 %3 143} 144 145; Test that we are generating vrev and vext for reverse shuffles of v8i16 146; shuffles. 147; CHECK-LABEL: reverse_v8i16: 148define void @reverse_v8i16(<8 x i16>* %loadaddr, <8 x i16>* %storeaddr) { 149 %v0 = load <8 x i16>, <8 x i16>* %loadaddr 150 ; CHECK: vrev64.16 151 ; CHECK: vext.16 152 %v1 = shufflevector <8 x i16> %v0, <8 x i16> undef, 153 <8 x i32> <i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0> 154 store <8 x i16> %v1, <8 x i16>* %storeaddr 155 ret void 156} 157 158; Test that we are generating vrev and vext for reverse shuffles of v16i8 159; shuffles. 160; CHECK-LABEL: reverse_v16i8: 161define void @reverse_v16i8(<16 x i8>* %loadaddr, <16 x i8>* %storeaddr) { 162 %v0 = load <16 x i8>, <16 x i8>* %loadaddr 163 ; CHECK: vrev64.8 164 ; CHECK: vext.8 165 %v1 = shufflevector <16 x i8> %v0, <16 x i8> undef, 166 <16 x i32> <i32 15, i32 14, i32 13, i32 12, i32 11, i32 10, i32 9, i32 8, 167 i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0> 168 store <16 x i8> %v1, <16 x i8>* %storeaddr 169 ret void 170} 171 172; <rdar://problem/14170854>. 173; vldr cannot handle unaligned loads. 174; Fall back to vld1.32, which can, instead of using the general purpose loads 175; followed by a costly sequence of instructions to build the vector register. 176; CHECK-LABEL: t3: 177; CHECK: vld1.32 {[[REG:d[0-9]+]][0]} 178; CHECK: vld1.32 {[[REG]][1]} 179; CHECK: vmull.u8 q{{[0-9]+}}, [[REG]], [[REG]] 180define <8 x i16> @t3(i8 zeroext %xf, i8* nocapture %sp0, i8* nocapture %sp1, i32* nocapture %outp) { 181entry: 182 %pix_sp0.0.cast = bitcast i8* %sp0 to i32* 183 %pix_sp0.0.copyload = load i32, i32* %pix_sp0.0.cast, align 1 184 %pix_sp1.0.cast = bitcast i8* %sp1 to i32* 185 %pix_sp1.0.copyload = load i32, i32* %pix_sp1.0.cast, align 1 186 %vecinit = insertelement <2 x i32> undef, i32 %pix_sp0.0.copyload, i32 0 187 %vecinit1 = insertelement <2 x i32> %vecinit, i32 %pix_sp1.0.copyload, i32 1 188 %0 = bitcast <2 x i32> %vecinit1 to <8 x i8> 189 %vmull.i = tail call <8 x i16> @llvm.arm.neon.vmullu.v8i16(<8 x i8> %0, <8 x i8> %0) 190 ret <8 x i16> %vmull.i 191} 192 193; Function Attrs: nounwind readnone 194declare <8 x i16> @llvm.arm.neon.vmullu.v8i16(<8 x i8>, <8 x i8>) 195 196; Check that (insert_vector_elt (load)) => (vector_load). 197; Thus, check that scalar_to_vector do not interfer with that. 198define <8 x i16> @t4(i8* nocapture %sp0) { 199; CHECK-LABEL: t4: 200; CHECK: vld1.32 {{{d[0-9]+}}[0]}, [r0] 201entry: 202 %pix_sp0.0.cast = bitcast i8* %sp0 to i32* 203 %pix_sp0.0.copyload = load i32, i32* %pix_sp0.0.cast, align 1 204 %vec = insertelement <2 x i32> undef, i32 %pix_sp0.0.copyload, i32 0 205 %0 = bitcast <2 x i32> %vec to <8 x i8> 206 %vmull.i = tail call <8 x i16> @llvm.arm.neon.vmullu.v8i16(<8 x i8> %0, <8 x i8> %0) 207 ret <8 x i16> %vmull.i 208} 209 210; Make sure vector load is used for all three loads. 211; Lowering to build vector was breaking the single use property of the load of 212; %pix_sp0.0.copyload. 213; CHECK-LABEL: t5: 214; CHECK: vld1.32 {[[REG1:d[0-9]+]][1]}, [r0] 215; CHECK: vorr [[REG2:d[0-9]+]], [[REG1]], [[REG1]] 216; CHECK: vld1.32 {[[REG1]][0]}, [r1] 217; CHECK: vld1.32 {[[REG2]][0]}, [r2] 218; CHECK: vmull.u8 q{{[0-9]+}}, [[REG1]], [[REG2]] 219define <8 x i16> @t5(i8* nocapture %sp0, i8* nocapture %sp1, i8* nocapture %sp2) { 220entry: 221 %pix_sp0.0.cast = bitcast i8* %sp0 to i32* 222 %pix_sp0.0.copyload = load i32, i32* %pix_sp0.0.cast, align 1 223 %pix_sp1.0.cast = bitcast i8* %sp1 to i32* 224 %pix_sp1.0.copyload = load i32, i32* %pix_sp1.0.cast, align 1 225 %pix_sp2.0.cast = bitcast i8* %sp2 to i32* 226 %pix_sp2.0.copyload = load i32, i32* %pix_sp2.0.cast, align 1 227 %vec = insertelement <2 x i32> undef, i32 %pix_sp0.0.copyload, i32 1 228 %vecinit1 = insertelement <2 x i32> %vec, i32 %pix_sp1.0.copyload, i32 0 229 %vecinit2 = insertelement <2 x i32> %vec, i32 %pix_sp2.0.copyload, i32 0 230 %0 = bitcast <2 x i32> %vecinit1 to <8 x i8> 231 %1 = bitcast <2 x i32> %vecinit2 to <8 x i8> 232 %vmull.i = tail call <8 x i16> @llvm.arm.neon.vmullu.v8i16(<8 x i8> %0, <8 x i8> %1) 233 ret <8 x i16> %vmull.i 234} 235 236; <rdar://problem/14989896> Make sure we manage to truncate a vector from an 237; illegal type to a legal type. 238define <2 x i8> @test_truncate(<2 x i128> %in) { 239; CHECK-LABEL: test_truncate: 240; CHECK: mov [[BASE:r[0-9]+]], sp 241; CHECK-NEXT: vld1.32 {[[REG1:d[0-9]+]][0]}, {{\[}}[[BASE]]:32] 242; CHECK-NEXT: add [[BASE2:r[0-9]+]], [[BASE]], #4 243; CHECK-NEXT: vld1.32 {[[REG1]][1]}, {{\[}}[[BASE2]]:32] 244; REG2 Should map on the same Q register as REG1, i.e., REG2 = REG1 - 1, but we 245; cannot express that. 246; CHECK-NEXT: vmov.32 [[REG2:d[0-9]+]][0], r0 247; CHECK-NEXT: vmov.32 [[REG2]][1], r1 248; The Q register used here should match floor(REG1/2), but we cannot express that. 249; CHECK-NEXT: vmovn.i64 [[RES:d[0-9]+]], q{{[0-9]+}} 250; CHECK-NEXT: vmov r0, r1, [[RES]] 251entry: 252 %res = trunc <2 x i128> %in to <2 x i8> 253 ret <2 x i8> %res 254} 255