121 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: opt -passes=slp-vectorizer -S < %s | FileCheck %s3 4target datalayout = "e-m:e-i64:64-i128:128-n32:64-S128"5target triple = "aarch64"6 7; This should not be matched as a load combining candidate.8; There are no 'or' operations, so it can't be a bswap or9; other pattern that we are expecting the backend to handle.10 11define void @PR50256(ptr %a, ptr %b, i32 %n) {12; CHECK-LABEL: @PR50256(13; CHECK-NEXT: [[ARRAYIDX_8:%.*]] = getelementptr inbounds i8, ptr [[A:%.*]], i64 814; CHECK-NEXT: [[ARRAYIDX3_8:%.*]] = getelementptr inbounds i16, ptr [[B:%.*]], i64 815; CHECK-NEXT: [[TMP2:%.*]] = load <8 x i8>, ptr [[A]], align 116; CHECK-NEXT: [[TMP3:%.*]] = zext <8 x i8> [[TMP2]] to <8 x i16>17; CHECK-NEXT: [[TMP4:%.*]] = shl nuw <8 x i16> [[TMP3]], splat (i16 8)18; CHECK-NEXT: [[TMP7:%.*]] = load <8 x i8>, ptr [[ARRAYIDX_8]], align 119; CHECK-NEXT: [[TMP8:%.*]] = zext <8 x i8> [[TMP7]] to <8 x i16>20; CHECK-NEXT: [[TMP9:%.*]] = shl nuw <8 x i16> [[TMP8]], splat (i16 8)21; CHECK-NEXT: store <8 x i16> [[TMP4]], ptr [[B]], align 222; CHECK-NEXT: store <8 x i16> [[TMP9]], ptr [[ARRAYIDX3_8]], align 223; CHECK-NEXT: ret void24;25 %arrayidx.1 = getelementptr inbounds i8, ptr %a, i64 126 %arrayidx.2 = getelementptr inbounds i8, ptr %a, i64 227 %arrayidx.3 = getelementptr inbounds i8, ptr %a, i64 328 %arrayidx.4 = getelementptr inbounds i8, ptr %a, i64 429 %arrayidx.5 = getelementptr inbounds i8, ptr %a, i64 530 %arrayidx.7 = getelementptr inbounds i8, ptr %a, i64 731 %arrayidx.6 = getelementptr inbounds i8, ptr %a, i64 632 %arrayidx.8 = getelementptr inbounds i8, ptr %a, i64 833 %arrayidx.9 = getelementptr inbounds i8, ptr %a, i64 934 %arrayidx.10 = getelementptr inbounds i8, ptr %a, i64 1035 %arrayidx.11 = getelementptr inbounds i8, ptr %a, i64 1136 %arrayidx.12 = getelementptr inbounds i8, ptr %a, i64 1237 %arrayidx.13 = getelementptr inbounds i8, ptr %a, i64 1338 %arrayidx.14 = getelementptr inbounds i8, ptr %a, i64 1439 %arrayidx.15 = getelementptr inbounds i8, ptr %a, i64 1540 %i = load i8, ptr %a, align 141 %i1 = load i8, ptr %arrayidx.1, align 142 %i2 = load i8, ptr %arrayidx.2, align 143 %i3 = load i8, ptr %arrayidx.3, align 144 %i4 = load i8, ptr %arrayidx.4, align 145 %i5 = load i8, ptr %arrayidx.5, align 146 %i6 = load i8, ptr %arrayidx.6, align 147 %i7 = load i8, ptr %arrayidx.7, align 148 %i8 = load i8, ptr %arrayidx.8, align 149 %i9 = load i8, ptr %arrayidx.9, align 150 %i10 = load i8, ptr %arrayidx.10, align 151 %i11 = load i8, ptr %arrayidx.11, align 152 %i12 = load i8, ptr %arrayidx.12, align 153 %i13 = load i8, ptr %arrayidx.13, align 154 %i14 = load i8, ptr %arrayidx.14, align 155 %i15 = load i8, ptr %arrayidx.15, align 156 %conv5 = zext i8 %i to i1657 %conv5.1 = zext i8 %i1 to i1658 %conv5.2 = zext i8 %i2 to i1659 %conv5.3 = zext i8 %i3 to i1660 %conv5.4 = zext i8 %i4 to i1661 %conv5.5 = zext i8 %i5 to i1662 %conv5.6 = zext i8 %i6 to i1663 %conv5.7 = zext i8 %i7 to i1664 %conv5.8 = zext i8 %i8 to i1665 %conv5.9 = zext i8 %i9 to i1666 %conv5.10 = zext i8 %i10 to i1667 %conv5.11 = zext i8 %i11 to i1668 %conv5.12 = zext i8 %i12 to i1669 %conv5.13 = zext i8 %i13 to i1670 %conv5.14 = zext i8 %i14 to i1671 %conv5.15 = zext i8 %i15 to i1672 %shl = shl nuw i16 %conv5, 873 %shl.1 = shl nuw i16 %conv5.1, 874 %shl.2 = shl nuw i16 %conv5.2, 875 %shl.3 = shl nuw i16 %conv5.3, 876 %shl.4 = shl nuw i16 %conv5.4, 877 %shl.5 = shl nuw i16 %conv5.5, 878 %shl.6 = shl nuw i16 %conv5.6, 879 %shl.7 = shl nuw i16 %conv5.7, 880 %shl.8 = shl nuw i16 %conv5.8, 881 %shl.9 = shl nuw i16 %conv5.9, 882 %shl.10 = shl nuw i16 %conv5.10, 883 %shl.11 = shl nuw i16 %conv5.11, 884 %shl.12 = shl nuw i16 %conv5.12, 885 %shl.13 = shl nuw i16 %conv5.13, 886 %shl.14 = shl nuw i16 %conv5.14, 887 %shl.15 = shl nuw i16 %conv5.15, 888 %arrayidx3.1 = getelementptr inbounds i16, ptr %b, i64 189 %arrayidx3.2 = getelementptr inbounds i16, ptr %b, i64 290 %arrayidx3.3 = getelementptr inbounds i16, ptr %b, i64 391 %arrayidx3.4 = getelementptr inbounds i16, ptr %b, i64 492 %arrayidx3.5 = getelementptr inbounds i16, ptr %b, i64 593 %arrayidx3.6 = getelementptr inbounds i16, ptr %b, i64 694 %arrayidx3.7 = getelementptr inbounds i16, ptr %b, i64 795 %arrayidx3.8 = getelementptr inbounds i16, ptr %b, i64 896 %arrayidx3.9 = getelementptr inbounds i16, ptr %b, i64 997 %arrayidx3.10 = getelementptr inbounds i16, ptr %b, i64 1098 %arrayidx3.11 = getelementptr inbounds i16, ptr %b, i64 1199 %arrayidx3.12 = getelementptr inbounds i16, ptr %b, i64 12100 %arrayidx3.13 = getelementptr inbounds i16, ptr %b, i64 13101 %arrayidx3.14 = getelementptr inbounds i16, ptr %b, i64 14102 %arrayidx3.15 = getelementptr inbounds i16, ptr %b, i64 15103 store i16 %shl, ptr %b, align 2104 store i16 %shl.1, ptr %arrayidx3.1, align 2105 store i16 %shl.2, ptr %arrayidx3.2, align 2106 store i16 %shl.3, ptr %arrayidx3.3, align 2107 store i16 %shl.4, ptr %arrayidx3.4, align 2108 store i16 %shl.5, ptr %arrayidx3.5, align 2109 store i16 %shl.6, ptr %arrayidx3.6, align 2110 store i16 %shl.7, ptr %arrayidx3.7, align 2111 store i16 %shl.8, ptr %arrayidx3.8, align 2112 store i16 %shl.9, ptr %arrayidx3.9, align 2113 store i16 %shl.10, ptr %arrayidx3.10, align 2114 store i16 %shl.11, ptr %arrayidx3.11, align 2115 store i16 %shl.12, ptr %arrayidx3.12, align 2116 store i16 %shl.13, ptr %arrayidx3.13, align 2117 store i16 %shl.14, ptr %arrayidx3.14, align 2118 store i16 %shl.15, ptr %arrayidx3.15, align 2119 ret void120}121