203 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: opt < %s -passes=loop-vectorize,dce,instcombine -force-vector-interleave=1 -force-vector-width=4 -S | FileCheck %s3 4target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"5target triple = "x86_64-unknown-linux-gnu"6 7; A tricky loop:8;9; void loop(int *a, int *b) {10; for (int i = 0; i < 512; ++i) {11; a[a[i]] = b[i];12; a[i] = b[i+1];13; }14;}15 16define void @loop(ptr nocapture %a, ptr nocapture %b) nounwind uwtable {17; CHECK-LABEL: @loop(18; CHECK-NEXT: entry:19; CHECK-NEXT: br label [[FOR_BODY:%.*]]20; CHECK: for.body:21; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ 0, [[ENTRY:%.*]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]22; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i32, ptr [[B:%.*]], i64 [[INDVARS_IV]]23; CHECK-NEXT: [[TMP0:%.*]] = load i32, ptr [[ARRAYIDX]], align 424; CHECK-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i32, ptr [[A:%.*]], i64 [[INDVARS_IV]]25; CHECK-NEXT: [[TMP1:%.*]] = load i32, ptr [[ARRAYIDX2]], align 426; CHECK-NEXT: [[IDXPROM3:%.*]] = sext i32 [[TMP1]] to i6427; CHECK-NEXT: [[ARRAYIDX4:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[IDXPROM3]]28; CHECK-NEXT: store i32 [[TMP0]], ptr [[ARRAYIDX4]], align 429; CHECK-NEXT: [[INDVARS_IV_NEXT]] = add i64 [[INDVARS_IV]], 130; CHECK-NEXT: [[ARRAYIDX6:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 [[INDVARS_IV_NEXT]]31; CHECK-NEXT: [[TMP2:%.*]] = load i32, ptr [[ARRAYIDX6]], align 432; CHECK-NEXT: store i32 [[TMP2]], ptr [[ARRAYIDX2]], align 433; CHECK-NEXT: [[TMP3:%.*]] = and i64 [[INDVARS_IV_NEXT]], 429496729534; CHECK-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[TMP3]], 51235; CHECK-NEXT: br i1 [[EXITCOND]], label [[FOR_END:%.*]], label [[FOR_BODY]]36; CHECK: for.end:37; CHECK-NEXT: ret void38;39entry:40 br label %for.body41 42for.body: ; preds = %for.body, %entry43 %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]44 %arrayidx = getelementptr inbounds i32, ptr %b, i64 %indvars.iv45 %0 = load i32, ptr %arrayidx, align 446 %arrayidx2 = getelementptr inbounds i32, ptr %a, i64 %indvars.iv47 %1 = load i32, ptr %arrayidx2, align 448 %idxprom3 = sext i32 %1 to i6449 %arrayidx4 = getelementptr inbounds i32, ptr %a, i64 %idxprom350 store i32 %0, ptr %arrayidx4, align 451 %indvars.iv.next = add i64 %indvars.iv, 152 %arrayidx6 = getelementptr inbounds i32, ptr %b, i64 %indvars.iv.next53 %2 = load i32, ptr %arrayidx6, align 454 store i32 %2, ptr %arrayidx2, align 455 %lftr.wideiv = trunc i64 %indvars.iv.next to i3256 %exitcond = icmp eq i32 %lftr.wideiv, 51257 br i1 %exitcond, label %for.end, label %for.body58 59for.end: ; preds = %for.body60 ret void61}62 63; The same loop with parallel loop metadata added to the loop branch64; and the memory instructions.65 66define void @parallel_loop(ptr nocapture %a, ptr nocapture %b) nounwind uwtable {67; CHECK-LABEL: @parallel_loop(68; CHECK-NEXT: entry:69; CHECK-NEXT: br label [[VECTOR_PH:%.*]]70; CHECK: vector.ph:71; CHECK-NEXT: br label [[VECTOR_BODY:%.*]]72; CHECK: vector.body:73; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]74; CHECK-NEXT: [[TMP3:%.*]] = getelementptr inbounds i32, ptr [[B:%.*]], i64 [[INDEX]]75; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <4 x i32>, ptr [[TMP3]], align 4, !llvm.access.group [[ACC_GRP0:![0-9]+]]76; CHECK-NEXT: [[TMP1:%.*]] = extractelement <4 x i32> [[WIDE_LOAD]], i64 077; CHECK-NEXT: [[TMP21:%.*]] = extractelement <4 x i32> [[WIDE_LOAD]], i64 178; CHECK-NEXT: [[TMP22:%.*]] = extractelement <4 x i32> [[WIDE_LOAD]], i64 279; CHECK-NEXT: [[TMP23:%.*]] = extractelement <4 x i32> [[WIDE_LOAD]], i64 380; CHECK-NEXT: [[TMP5:%.*]] = getelementptr inbounds i32, ptr [[A:%.*]], i64 [[INDEX]]81; CHECK-NEXT: [[TMP2:%.*]] = getelementptr i32, ptr [[A]], i64 [[INDEX]]82; CHECK-NEXT: [[TMP8:%.*]] = getelementptr i8, ptr [[TMP2]], i64 483; CHECK-NEXT: [[TMP4:%.*]] = getelementptr i32, ptr [[A]], i64 [[INDEX]]84; CHECK-NEXT: [[TMP26:%.*]] = getelementptr i8, ptr [[TMP4]], i64 885; CHECK-NEXT: [[TMP6:%.*]] = getelementptr i32, ptr [[A]], i64 [[INDEX]]86; CHECK-NEXT: [[TMP7:%.*]] = getelementptr i8, ptr [[TMP6]], i64 1287; CHECK-NEXT: [[TMP9:%.*]] = load i32, ptr [[TMP5]], align 4, !llvm.access.group [[ACC_GRP0]]88; CHECK-NEXT: [[TMP10:%.*]] = load i32, ptr [[TMP8]], align 4, !llvm.access.group [[ACC_GRP0]]89; CHECK-NEXT: [[TMP11:%.*]] = load i32, ptr [[TMP26]], align 4, !llvm.access.group [[ACC_GRP0]]90; CHECK-NEXT: [[TMP12:%.*]] = load i32, ptr [[TMP7]], align 4, !llvm.access.group [[ACC_GRP0]]91; CHECK-NEXT: [[TMP13:%.*]] = sext i32 [[TMP9]] to i6492; CHECK-NEXT: [[TMP14:%.*]] = sext i32 [[TMP10]] to i6493; CHECK-NEXT: [[TMP15:%.*]] = sext i32 [[TMP11]] to i6494; CHECK-NEXT: [[TMP16:%.*]] = sext i32 [[TMP12]] to i6495; CHECK-NEXT: [[TMP17:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[TMP13]]96; CHECK-NEXT: [[TMP18:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[TMP14]]97; CHECK-NEXT: [[TMP19:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[TMP15]]98; CHECK-NEXT: [[TMP20:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[TMP16]]99; CHECK-NEXT: store i32 [[TMP1]], ptr [[TMP17]], align 4, !llvm.access.group [[ACC_GRP1:![0-9]+]]100; CHECK-NEXT: store i32 [[TMP21]], ptr [[TMP18]], align 4, !llvm.access.group [[ACC_GRP1]]101; CHECK-NEXT: store i32 [[TMP22]], ptr [[TMP19]], align 4, !llvm.access.group [[ACC_GRP1]]102; CHECK-NEXT: store i32 [[TMP23]], ptr [[TMP20]], align 4, !llvm.access.group [[ACC_GRP1]]103; CHECK-NEXT: [[TMP27:%.*]] = getelementptr i32, ptr [[B]], i64 [[INDEX]]104; CHECK-NEXT: [[TMP25:%.*]] = getelementptr i8, ptr [[TMP27]], i64 4105; CHECK-NEXT: [[WIDE_LOAD1:%.*]] = load <4 x i32>, ptr [[TMP25]], align 4, !llvm.access.group [[ACC_GRP0]]106; CHECK-NEXT: store <4 x i32> [[WIDE_LOAD1]], ptr [[TMP5]], align 4, !llvm.access.group [[ACC_GRP0]]107; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4108; CHECK-NEXT: [[TMP29:%.*]] = icmp eq i64 [[INDEX_NEXT]], 512109; CHECK-NEXT: br i1 [[TMP29]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP2:![0-9]+]]110; CHECK: middle.block:111; CHECK-NEXT: br label [[FOR_BODY:%.*]]112; CHECK: for.end:113; CHECK-NEXT: ret void114;115entry:116 br label %for.body117 118for.body: ; preds = %for.body, %entry119 %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]120 %arrayidx = getelementptr inbounds i32, ptr %b, i64 %indvars.iv121 %0 = load i32, ptr %arrayidx, align 4, !llvm.access.group !13122 %arrayidx2 = getelementptr inbounds i32, ptr %a, i64 %indvars.iv123 %1 = load i32, ptr %arrayidx2, align 4, !llvm.access.group !13124 %idxprom3 = sext i32 %1 to i64125 %arrayidx4 = getelementptr inbounds i32, ptr %a, i64 %idxprom3126 ; This store might have originated from inlining a function with a parallel127 ; loop. Refers to a list with the "original loop reference" (!4) also included.128 store i32 %0, ptr %arrayidx4, align 4, !llvm.access.group !15129 %indvars.iv.next = add i64 %indvars.iv, 1130 %arrayidx6 = getelementptr inbounds i32, ptr %b, i64 %indvars.iv.next131 %2 = load i32, ptr %arrayidx6, align 4, !llvm.access.group !13132 store i32 %2, ptr %arrayidx2, align 4, !llvm.access.group !13133 %lftr.wideiv = trunc i64 %indvars.iv.next to i32134 %exitcond = icmp eq i32 %lftr.wideiv, 512135 br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !3136 137for.end: ; preds = %for.body138 ret void139}140 141; The same loop with an illegal parallel loop metadata: the memory142; accesses refer to a different loop's identifier.143 144 145define void @mixed_metadata(ptr nocapture %a, ptr nocapture %b) nounwind uwtable {146; CHECK-LABEL: @mixed_metadata(147; CHECK-NEXT: entry:148; CHECK-NEXT: br label [[FOR_BODY:%.*]]149; CHECK: for.body:150; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ 0, [[ENTRY:%.*]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]151; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i32, ptr [[B:%.*]], i64 [[INDVARS_IV]]152; CHECK-NEXT: [[TMP0:%.*]] = load i32, ptr [[ARRAYIDX]], align 4, !llvm.access.group [[ACC_GRP7:![0-9]+]]153; CHECK-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds i32, ptr [[A:%.*]], i64 [[INDVARS_IV]]154; CHECK-NEXT: [[TMP1:%.*]] = load i32, ptr [[ARRAYIDX2]], align 4, !llvm.access.group [[ACC_GRP7]]155; CHECK-NEXT: [[IDXPROM3:%.*]] = sext i32 [[TMP1]] to i64156; CHECK-NEXT: [[ARRAYIDX4:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[IDXPROM3]]157; CHECK-NEXT: store i32 [[TMP0]], ptr [[ARRAYIDX4]], align 4, !llvm.access.group [[ACC_GRP8:![0-9]+]]158; CHECK-NEXT: [[INDVARS_IV_NEXT]] = add i64 [[INDVARS_IV]], 1159; CHECK-NEXT: [[ARRAYIDX6:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 [[INDVARS_IV_NEXT]]160; CHECK-NEXT: [[TMP2:%.*]] = load i32, ptr [[ARRAYIDX6]], align 4, !llvm.access.group [[ACC_GRP7]]161; CHECK-NEXT: store i32 [[TMP2]], ptr [[ARRAYIDX2]], align 4, !llvm.access.group [[ACC_GRP7]]162; CHECK-NEXT: [[TMP3:%.*]] = and i64 [[INDVARS_IV_NEXT]], 4294967295163; CHECK-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[TMP3]], 512164; CHECK-NEXT: br i1 [[EXITCOND]], label [[FOR_END:%.*]], label [[FOR_BODY]], !llvm.loop [[LOOP9:![0-9]+]]165; CHECK: for.end:166; CHECK-NEXT: ret void167;168entry:169 br label %for.body170 171for.body: ; preds = %for.body, %entry172 %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]173 %arrayidx = getelementptr inbounds i32, ptr %b, i64 %indvars.iv174 %0 = load i32, ptr %arrayidx, align 4, !llvm.access.group !16175 %arrayidx2 = getelementptr inbounds i32, ptr %a, i64 %indvars.iv176 %1 = load i32, ptr %arrayidx2, align 4, !llvm.access.group !16177 %idxprom3 = sext i32 %1 to i64178 %arrayidx4 = getelementptr inbounds i32, ptr %a, i64 %idxprom3179 ; This refers to the loop marked with !7 which we are not in at the moment.180 ; It should prevent detecting as a parallel loop.181 store i32 %0, ptr %arrayidx4, align 4, !llvm.access.group !17182 %indvars.iv.next = add i64 %indvars.iv, 1183 %arrayidx6 = getelementptr inbounds i32, ptr %b, i64 %indvars.iv.next184 %2 = load i32, ptr %arrayidx6, align 4, !llvm.access.group !16185 store i32 %2, ptr %arrayidx2, align 4, !llvm.access.group !16186 %lftr.wideiv = trunc i64 %indvars.iv.next to i32187 %exitcond = icmp eq i32 %lftr.wideiv, 512188 br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !6189 190for.end: ; preds = %for.body191 ret void192}193 194!3 = !{!3, !{!"llvm.loop.parallel_accesses", !13, !15}}195!4 = !{!4, !{!"llvm.loop.parallel_accesses", !14, !15}}196!6 = !{!6, !{!"llvm.loop.parallel_accesses", !16}}197!7 = !{!7, !{!"llvm.loop.parallel_accesses", !17}}198!13 = distinct !{}199!14 = distinct !{}200!15 = distinct !{}201!16 = distinct !{}202!17 = distinct !{}203