83 lines · plain
1; RUN: llc < %s -mtriple=thumbv7-none-linux-gnueabi -mcpu=cortex-a9 -mattr=+neon,+neonfp | FileCheck %s2; PR14824. The test is presented by Jiangning Liu. If the ld/st optimization algorithm is changed, this test case may fail.3; Also if the machine code for ld/st optimizor is changed, this test case may fail. If so, remove this test.4 5define void @sample_test(ptr %secondSource, ptr %source, ptr %dest) nounwind {6; CHECK: sample_test7; CHECK-NOT: vldmia8; CHECK: add9entry:10 11; Load %source12 %s0 = load <8 x i64> , ptr %source, align 6413 %arrayidx64 = getelementptr inbounds <8 x i64>, ptr %source, i32 614 %s120 = load <8 x i64> , ptr %arrayidx64, align 6415 %s122 = bitcast <8 x i64> %s120 to i51216 %data.i.i677.48.extract.shift = lshr i512 %s122, 38417 %data.i.i677.48.extract.trunc = trunc i512 %data.i.i677.48.extract.shift to i6418 %s123 = insertelement <8 x i64> undef, i64 %data.i.i677.48.extract.trunc, i32 019 %data.i.i677.32.extract.shift = lshr i512 %s122, 25620 %data.i.i677.32.extract.trunc = trunc i512 %data.i.i677.32.extract.shift to i6421 %s124 = insertelement <8 x i64> %s123, i64 %data.i.i677.32.extract.trunc, i32 122 %data.i.i677.16.extract.shift = lshr i512 %s122, 12823 %data.i.i677.16.extract.trunc = trunc i512 %data.i.i677.16.extract.shift to i6424 %s125 = insertelement <8 x i64> %s124, i64 %data.i.i677.16.extract.trunc, i32 225 %data.i.i677.56.extract.shift = lshr i512 %s122, 44826 %data.i.i677.56.extract.trunc = trunc i512 %data.i.i677.56.extract.shift to i6427 %s126 = insertelement <8 x i64> %s125, i64 %data.i.i677.56.extract.trunc, i32 328 %data.i.i677.24.extract.shift = lshr i512 %s122, 19229 %data.i.i677.24.extract.trunc = trunc i512 %data.i.i677.24.extract.shift to i6430 %s127 = insertelement <8 x i64> %s126, i64 %data.i.i677.24.extract.trunc, i32 431 %s128 = insertelement <8 x i64> %s127, i64 %data.i.i677.32.extract.trunc, i32 532 %s129 = insertelement <8 x i64> %s128, i64 %data.i.i677.16.extract.trunc, i32 633 %s130 = insertelement <8 x i64> %s129, i64 %data.i.i677.56.extract.trunc, i32 734 35; Load %secondSource36 %s1 = load <8 x i64> , ptr %secondSource, align 6437 %arrayidx67 = getelementptr inbounds <8 x i64>, ptr %secondSource, i32 638 %s121 = load <8 x i64> , ptr %arrayidx67, align 6439 %s131 = bitcast <8 x i64> %s121 to i51240 %data.i1.i676.48.extract.shift = lshr i512 %s131, 38441 %data.i1.i676.48.extract.trunc = trunc i512 %data.i1.i676.48.extract.shift to i6442 %s132 = insertelement <8 x i64> undef, i64 %data.i1.i676.48.extract.trunc, i32 043 %data.i1.i676.32.extract.shift = lshr i512 %s131, 25644 %data.i1.i676.32.extract.trunc = trunc i512 %data.i1.i676.32.extract.shift to i6445 %s133 = insertelement <8 x i64> %s132, i64 %data.i1.i676.32.extract.trunc, i32 146 %data.i1.i676.16.extract.shift = lshr i512 %s131, 12847 %data.i1.i676.16.extract.trunc = trunc i512 %data.i1.i676.16.extract.shift to i6448 %s134 = insertelement <8 x i64> %s133, i64 %data.i1.i676.16.extract.trunc, i32 249 %data.i1.i676.56.extract.shift = lshr i512 %s131, 44850 %data.i1.i676.56.extract.trunc = trunc i512 %data.i1.i676.56.extract.shift to i6451 %s135 = insertelement <8 x i64> %s134, i64 %data.i1.i676.56.extract.trunc, i32 352 %data.i1.i676.24.extract.shift = lshr i512 %s131, 19253 %data.i1.i676.24.extract.trunc = trunc i512 %data.i1.i676.24.extract.shift to i6454 %s136 = insertelement <8 x i64> %s135, i64 %data.i1.i676.24.extract.trunc, i32 455 %s137 = insertelement <8 x i64> %s136, i64 %data.i1.i676.32.extract.trunc, i32 556 %s138 = insertelement <8 x i64> %s137, i64 %data.i1.i676.16.extract.trunc, i32 657 %s139 = insertelement <8 x i64> %s138, i64 %data.i1.i676.56.extract.trunc, i32 758 59; Operations about %Source and %secondSource60 %vecinit28.i.i699 = shufflevector <8 x i64> %s139, <8 x i64> %s130, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 12, i32 undef, i32 undef, i32 undef>61 %vecinit35.i.i700 = shufflevector <8 x i64> %vecinit28.i.i699, <8 x i64> %s139, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 13, i32 undef, i32 undef>62 %vecinit42.i.i701 = shufflevector <8 x i64> %vecinit35.i.i700, <8 x i64> %s139, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 14, i32 undef>63 %vecinit49.i.i702 = shufflevector <8 x i64> %vecinit42.i.i701, <8 x i64> %s130, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 15>64 %arrayidx72 = getelementptr inbounds <8 x i64>, ptr %dest, i32 665 store <8 x i64> %vecinit49.i.i702, ptr %arrayidx72, align 6466 %arrayidx78 = getelementptr inbounds <8 x i64>, ptr %secondSource, i32 767 %s141 = load <8 x i64> , ptr %arrayidx78, align 6468 %s151 = bitcast <8 x i64> %s141 to i51269 %data.i1.i649.32.extract.shift = lshr i512 %s151, 25670 %data.i1.i649.32.extract.trunc = trunc i512 %data.i1.i649.32.extract.shift to i6471 %s152 = insertelement <8 x i64> undef, i64 %data.i1.i649.32.extract.trunc, i32 072 %s153 = insertelement <8 x i64> %s152, i64 %data.i1.i649.32.extract.trunc, i32 173 %data.i1.i649.16.extract.shift = lshr i512 %s151, 12874 %data.i1.i649.16.extract.trunc = trunc i512 %data.i1.i649.16.extract.shift to i6475 %s154 = insertelement <8 x i64> %s153, i64 %data.i1.i649.16.extract.trunc, i32 276 %data.i1.i649.8.extract.shift = lshr i512 %s151, 6477 %data.i1.i649.8.extract.trunc = trunc i512 %data.i1.i649.8.extract.shift to i6478 %s155 = insertelement <8 x i64> %s154, i64 %data.i1.i649.8.extract.trunc, i32 379 %arrayidx83 = getelementptr inbounds <8 x i64>, ptr %dest, i32 780 store <8 x i64> %s155, ptr %arrayidx83, align 6481 ret void82}83