154 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: opt -passes=slp-vectorizer -S -o - -mtriple=i386 -mcpu=haswell < %s | FileCheck %s3target datalayout = "e-m:e-p:32:32-f64:32:64-f80:32-n8:16:32-S128"4 5@shift = common local_unnamed_addr global [10 x i32] zeroinitializer, align 46@data = common local_unnamed_addr global [10 x ptr] zeroinitializer, align 47 8define void @flat(i32 %intensity) {9; CHECK-LABEL: @flat(10; CHECK-NEXT: entry:11; CHECK-NEXT: [[TMP0:%.*]] = load i32, ptr @shift, align 412; CHECK-NEXT: [[TMP1:%.*]] = load i32, ptr getelementptr inbounds ([10 x i32], ptr @shift, i32 0, i32 1), align 413; CHECK-NEXT: [[TMP2:%.*]] = load ptr, ptr @data, align 414; CHECK-NEXT: [[TMP3:%.*]] = load ptr, ptr getelementptr inbounds ([10 x ptr], ptr @data, i32 0, i32 1), align 415; CHECK-NEXT: [[SHR:%.*]] = lshr i32 1, [[TMP0]]16; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i8, ptr [[TMP2]], i32 [[SHR]]17; CHECK-NEXT: [[SHR1:%.*]] = lshr i32 1, [[TMP1]]18; CHECK-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i8, ptr [[TMP3]], i32 [[SHR1]]19; CHECK-NEXT: br label [[FOR_BODY:%.*]]20; CHECK: for.cond.cleanup:21; CHECK-NEXT: ret void22; CHECK: for.body:23; CHECK-NEXT: [[D1_DATA_046:%.*]] = phi ptr [ [[TMP3]], [[ENTRY:%.*]] ], [ [[ADD_PTR23_1:%.*]], [[FOR_BODY]] ]24; CHECK-NEXT: [[Y_045:%.*]] = phi i32 [ 0, [[ENTRY]] ], [ [[INC_1:%.*]], [[FOR_BODY]] ]25; CHECK-NEXT: [[TMP4:%.*]] = load i8, ptr [[ARRAYIDX]], align 126; CHECK-NEXT: [[CONV:%.*]] = zext i8 [[TMP4]] to i3227; CHECK-NEXT: [[SUB:%.*]] = add nsw i32 [[CONV]], -12828; CHECK-NEXT: [[TMP5:%.*]] = load i8, ptr [[ARRAYIDX2]], align 129; CHECK-NEXT: [[CONV3:%.*]] = zext i8 [[TMP5]] to i3230; CHECK-NEXT: [[SUB4:%.*]] = add nsw i32 [[CONV3]], -12831; CHECK-NEXT: [[CMP5:%.*]] = icmp sgt i32 [[SUB]], -132; CHECK-NEXT: [[SUB7:%.*]] = sub nsw i32 128, [[CONV]]33; CHECK-NEXT: [[COND:%.*]] = select i1 [[CMP5]], i32 [[SUB]], i32 [[SUB7]]34; CHECK-NEXT: [[CMP8:%.*]] = icmp sgt i32 [[SUB4]], -135; CHECK-NEXT: [[SUB12:%.*]] = sub nsw i32 128, [[CONV3]]36; CHECK-NEXT: [[COND14:%.*]] = select i1 [[CMP8]], i32 [[SUB4]], i32 [[SUB12]]37; CHECK-NEXT: [[ADD:%.*]] = add nsw i32 [[COND14]], [[COND]]38; CHECK-NEXT: [[IDX_NEG:%.*]] = sub nsw i32 0, [[ADD]]39; CHECK-NEXT: [[ADD_PTR:%.*]] = getelementptr inbounds i8, ptr [[D1_DATA_046]], i32 [[IDX_NEG]]40; CHECK-NEXT: [[TMP6:%.*]] = load i8, ptr [[ADD_PTR]], align 141; CHECK-NEXT: [[CONV15:%.*]] = zext i8 [[TMP6]] to i3242; CHECK-NEXT: [[ADD16:%.*]] = add nsw i32 [[CONV15]], [[INTENSITY:%.*]]43; CHECK-NEXT: [[CONV17:%.*]] = trunc i32 [[ADD16]] to i844; CHECK-NEXT: store i8 [[CONV17]], ptr [[ADD_PTR]], align 145; CHECK-NEXT: [[ADD_PTR18:%.*]] = getelementptr inbounds i8, ptr [[D1_DATA_046]], i32 [[ADD]]46; CHECK-NEXT: [[TMP7:%.*]] = load i8, ptr [[ADD_PTR18]], align 147; CHECK-NEXT: [[NOT_TOBOOL:%.*]] = icmp eq i8 [[TMP7]], 048; CHECK-NEXT: [[CONV21:%.*]] = zext i1 [[NOT_TOBOOL]] to i849; CHECK-NEXT: store i8 [[CONV21]], ptr [[ADD_PTR18]], align 150; CHECK-NEXT: [[ADD_PTR23:%.*]] = getelementptr inbounds i8, ptr [[D1_DATA_046]], i32 [[TMP1]]51; CHECK-NEXT: [[TMP8:%.*]] = load i8, ptr [[ARRAYIDX]], align 152; CHECK-NEXT: [[CONV_1:%.*]] = zext i8 [[TMP8]] to i3253; CHECK-NEXT: [[SUB_1:%.*]] = add nsw i32 [[CONV_1]], -12854; CHECK-NEXT: [[TMP9:%.*]] = load i8, ptr [[ARRAYIDX2]], align 155; CHECK-NEXT: [[CONV3_1:%.*]] = zext i8 [[TMP9]] to i3256; CHECK-NEXT: [[SUB4_1:%.*]] = add nsw i32 [[CONV3_1]], -12857; CHECK-NEXT: [[CMP5_1:%.*]] = icmp sgt i32 [[SUB_1]], -158; CHECK-NEXT: [[SUB7_1:%.*]] = sub nsw i32 128, [[CONV_1]]59; CHECK-NEXT: [[COND_1:%.*]] = select i1 [[CMP5_1]], i32 [[SUB_1]], i32 [[SUB7_1]]60; CHECK-NEXT: [[CMP8_1:%.*]] = icmp sgt i32 [[SUB4_1]], -161; CHECK-NEXT: [[SUB12_1:%.*]] = sub nsw i32 128, [[CONV3_1]]62; CHECK-NEXT: [[COND14_1:%.*]] = select i1 [[CMP8_1]], i32 [[SUB4_1]], i32 [[SUB12_1]]63; CHECK-NEXT: [[ADD_1:%.*]] = add nsw i32 [[COND14_1]], [[COND_1]]64; CHECK-NEXT: [[IDX_NEG_1:%.*]] = sub nsw i32 0, [[ADD_1]]65; CHECK-NEXT: [[ADD_PTR_1:%.*]] = getelementptr inbounds i8, ptr [[ADD_PTR23]], i32 [[IDX_NEG_1]]66; CHECK-NEXT: [[TMP10:%.*]] = load i8, ptr [[ADD_PTR_1]], align 167; CHECK-NEXT: [[CONV15_1:%.*]] = zext i8 [[TMP10]] to i3268; CHECK-NEXT: [[ADD16_1:%.*]] = add nsw i32 [[CONV15_1]], [[INTENSITY]]69; CHECK-NEXT: [[CONV17_1:%.*]] = trunc i32 [[ADD16_1]] to i870; CHECK-NEXT: store i8 [[CONV17_1]], ptr [[ADD_PTR_1]], align 171; CHECK-NEXT: [[ADD_PTR18_1:%.*]] = getelementptr inbounds i8, ptr [[ADD_PTR23]], i32 [[ADD_1]]72; CHECK-NEXT: [[TMP11:%.*]] = load i8, ptr [[ADD_PTR18_1]], align 173; CHECK-NEXT: [[NOT_TOBOOL_1:%.*]] = icmp eq i8 [[TMP11]], 074; CHECK-NEXT: [[CONV21_1:%.*]] = zext i1 [[NOT_TOBOOL_1]] to i875; CHECK-NEXT: store i8 [[CONV21_1]], ptr [[ADD_PTR18_1]], align 176; CHECK-NEXT: [[ADD_PTR23_1]] = getelementptr inbounds i8, ptr [[ADD_PTR23]], i32 [[TMP1]]77; CHECK-NEXT: [[INC_1]] = add nsw i32 [[Y_045]], 278; CHECK-NEXT: [[EXITCOND_1:%.*]] = icmp eq i32 [[INC_1]], 12879; CHECK-NEXT: br i1 [[EXITCOND_1]], label [[FOR_COND_CLEANUP:%.*]], label [[FOR_BODY]]80;81entry:82 %0 = load i32, ptr @shift, align 483 %1 = load i32, ptr getelementptr inbounds ([10 x i32], ptr @shift, i32 0, i32 1), align 484 %2 = load ptr, ptr @data, align 485 %3 = load ptr, ptr getelementptr inbounds ([10 x ptr], ptr @data, i32 0, i32 1), align 486 %shr = lshr i32 1, %087 %arrayidx = getelementptr inbounds i8, ptr %2, i32 %shr88 %shr1 = lshr i32 1, %189 %arrayidx2 = getelementptr inbounds i8, ptr %3, i32 %shr190 br label %for.body91 92for.cond.cleanup: ; preds = %for.body93 ret void94 95for.body: ; preds = %for.body, %entry96 %d1_data.046 = phi ptr [ %3, %entry ], [ %add.ptr23.1, %for.body ]97 %y.045 = phi i32 [ 0, %entry ], [ %inc.1, %for.body ]98 %4 = load i8, ptr %arrayidx, align 199 %conv = zext i8 %4 to i32100 %sub = add nsw i32 %conv, -128101 %5 = load i8, ptr %arrayidx2, align 1102 %conv3 = zext i8 %5 to i32103 %sub4 = add nsw i32 %conv3, -128104 %cmp5 = icmp sgt i32 %sub, -1105 %sub7 = sub nsw i32 128, %conv106 %cond = select i1 %cmp5, i32 %sub, i32 %sub7107 %cmp8 = icmp sgt i32 %sub4, -1108 %sub12 = sub nsw i32 128, %conv3109 %cond14 = select i1 %cmp8, i32 %sub4, i32 %sub12110 %add = add nsw i32 %cond14, %cond111 %idx.neg = sub nsw i32 0, %add112 %add.ptr = getelementptr inbounds i8, ptr %d1_data.046, i32 %idx.neg113 %6 = load i8, ptr %add.ptr, align 1114 %conv15 = zext i8 %6 to i32115 %add16 = add nsw i32 %conv15, %intensity116 %conv17 = trunc i32 %add16 to i8117 store i8 %conv17, ptr %add.ptr, align 1118 %add.ptr18 = getelementptr inbounds i8, ptr %d1_data.046, i32 %add119 %7 = load i8, ptr %add.ptr18, align 1120 %not.tobool = icmp eq i8 %7, 0121 %conv21 = zext i1 %not.tobool to i8122 store i8 %conv21, ptr %add.ptr18, align 1123 %add.ptr23 = getelementptr inbounds i8, ptr %d1_data.046, i32 %1124 %8 = load i8, ptr %arrayidx, align 1125 %conv.1 = zext i8 %8 to i32126 %sub.1 = add nsw i32 %conv.1, -128127 %9 = load i8, ptr %arrayidx2, align 1128 %conv3.1 = zext i8 %9 to i32129 %sub4.1 = add nsw i32 %conv3.1, -128130 %cmp5.1 = icmp sgt i32 %sub.1, -1131 %sub7.1 = sub nsw i32 128, %conv.1132 %cond.1 = select i1 %cmp5.1, i32 %sub.1, i32 %sub7.1133 %cmp8.1 = icmp sgt i32 %sub4.1, -1134 %sub12.1 = sub nsw i32 128, %conv3.1135 %cond14.1 = select i1 %cmp8.1, i32 %sub4.1, i32 %sub12.1136 %add.1 = add nsw i32 %cond14.1, %cond.1137 %idx.neg.1 = sub nsw i32 0, %add.1138 %add.ptr.1 = getelementptr inbounds i8, ptr %add.ptr23, i32 %idx.neg.1139 %10 = load i8, ptr %add.ptr.1, align 1140 %conv15.1 = zext i8 %10 to i32141 %add16.1 = add nsw i32 %conv15.1, %intensity142 %conv17.1 = trunc i32 %add16.1 to i8143 store i8 %conv17.1, ptr %add.ptr.1, align 1144 %add.ptr18.1 = getelementptr inbounds i8, ptr %add.ptr23, i32 %add.1145 %11 = load i8, ptr %add.ptr18.1, align 1146 %not.tobool.1 = icmp eq i8 %11, 0147 %conv21.1 = zext i1 %not.tobool.1 to i8148 store i8 %conv21.1, ptr %add.ptr18.1, align 1149 %add.ptr23.1 = getelementptr inbounds i8, ptr %add.ptr23, i32 %1150 %inc.1 = add nsw i32 %y.045, 2151 %exitcond.1 = icmp eq i32 %inc.1, 128152 br i1 %exitcond.1, label %for.cond.cleanup, label %for.body153}154