201 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: opt < %s -O3 -S | FileCheck --check-prefix=OLDPM %s3; RUN: opt < %s -passes="default<O3>" -S | FileCheck --check-prefix=NEWPM %s4 5target datalayout = "e-m:e-p:32:32-Fi8-i64:64-v128:64:128-a:0:32-n32-S64"6target triple = "thumbv6m-none-none-eabi"7 8; Not only should we be able to to form memsets here, the original loops9; should be deleted, too.10 11; Function Attrs: nounwind12define dso_local void @arm_fill_q7(i8 signext %value, ptr %pDst, i32 %blockSize) {13; OLDPM-LABEL: @arm_fill_q7(14; OLDPM-NEXT: entry:15; OLDPM-NEXT: [[CMP_NOT20:%.*]] = icmp ult i32 [[BLOCKSIZE:%.*]], 416; OLDPM-NEXT: br i1 [[CMP_NOT20]], label [[WHILE_END:%.*]], label [[WHILE_BODY_PREHEADER:%.*]]17; OLDPM: while.body.preheader:18; OLDPM-NEXT: [[TMP0:%.*]] = and i32 [[BLOCKSIZE]], -419; OLDPM-NEXT: call void @llvm.memset.p0.i32(ptr align 1 [[PDST:%.*]], i8 [[VALUE:%.*]], i32 [[TMP0]], i1 false)20; OLDPM-NEXT: [[SCEVGEP:%.*]] = getelementptr i8, ptr [[PDST]], i32 [[TMP0]]21; OLDPM-NEXT: br label [[WHILE_END]]22; OLDPM: while.end:23; OLDPM-NEXT: [[PDST_ADDR_0_LCSSA:%.*]] = phi ptr [ [[PDST]], [[ENTRY:%.*]] ], [ [[SCEVGEP]], [[WHILE_BODY_PREHEADER]] ]24; OLDPM-NEXT: [[REM:%.*]] = and i32 [[BLOCKSIZE]], 325; OLDPM-NEXT: [[CMP14_NOT17:%.*]] = icmp eq i32 [[REM]], 026; OLDPM-NEXT: br i1 [[CMP14_NOT17]], label [[WHILE_END18:%.*]], label [[WHILE_BODY16_PREHEADER:%.*]]27; OLDPM: while.body16.preheader:28; OLDPM-NEXT: call void @llvm.memset.p0.i32(ptr align 1 [[PDST_ADDR_0_LCSSA]], i8 [[VALUE]], i32 [[REM]], i1 false)29; OLDPM-NEXT: br label [[WHILE_END18]]30; OLDPM: while.end18:31; OLDPM-NEXT: ret void32;33; NEWPM-LABEL: @arm_fill_q7(34; NEWPM-NEXT: entry:35; NEWPM-NEXT: [[CMP_NOT17:%.*]] = icmp ult i32 [[BLOCKSIZE:%.*]], 436; NEWPM-NEXT: br i1 [[CMP_NOT17]], label [[WHILE_END:%.*]], label [[WHILE_BODY_PREHEADER:%.*]]37; NEWPM: while.body.preheader:38; NEWPM-NEXT: [[TMP0:%.*]] = and i32 [[BLOCKSIZE]], -439; NEWPM-NEXT: call void @llvm.memset.p0.i32(ptr align 1 [[PDST:%.*]], i8 [[VALUE:%.*]], i32 [[TMP0]], i1 false)40; NEWPM-NEXT: [[SCEVGEP:%.*]] = getelementptr i8, ptr [[PDST]], i32 [[TMP0]]41; NEWPM-NEXT: br label [[WHILE_END]]42; NEWPM: while.end:43; NEWPM-NEXT: [[PDST_ADDR_0_LCSSA:%.*]] = phi ptr [ [[PDST]], [[ENTRY:%.*]] ], [ [[SCEVGEP]], [[WHILE_BODY_PREHEADER]] ]44; NEWPM-NEXT: [[REM:%.*]] = and i32 [[BLOCKSIZE]], 345; NEWPM-NEXT: [[CMP14_NOT20:%.*]] = icmp eq i32 [[REM]], 046; NEWPM-NEXT: br i1 [[CMP14_NOT20]], label [[WHILE_END18:%.*]], label [[WHILE_BODY16_PREHEADER:%.*]]47; NEWPM: while.body16.preheader:48; NEWPM-NEXT: call void @llvm.memset.p0.i32(ptr align 1 [[PDST_ADDR_0_LCSSA]], i8 [[VALUE]], i32 [[REM]], i1 false)49; NEWPM-NEXT: br label [[WHILE_END18]]50; NEWPM: while.end18:51; NEWPM-NEXT: ret void52;53entry:54 %value.addr = alloca i8, align 155 %pDst.addr = alloca ptr, align 456 %blockSize.addr = alloca i32, align 457 %blkCnt = alloca i32, align 458 %packedValue = alloca i32, align 459 store i8 %value, ptr %value.addr, align 1, !tbaa !360 store ptr %pDst, ptr %pDst.addr, align 4, !tbaa !661 store i32 %blockSize, ptr %blockSize.addr, align 4, !tbaa !862 call void @llvm.lifetime.start.p0(ptr %blkCnt)63 call void @llvm.lifetime.start.p0(ptr %packedValue)64 %0 = load i8, ptr %value.addr, align 1, !tbaa !365 %conv = sext i8 %0 to i3266 %shl = shl i32 %conv, 067 %and = and i32 %shl, 25568 %1 = load i8, ptr %value.addr, align 1, !tbaa !369 %conv1 = sext i8 %1 to i3270 %shl2 = shl i32 %conv1, 871 %and3 = and i32 %shl2, 6528072 %or = or i32 %and, %and373 %2 = load i8, ptr %value.addr, align 1, !tbaa !374 %conv4 = sext i8 %2 to i3275 %shl5 = shl i32 %conv4, 1676 %and6 = and i32 %shl5, 1671168077 %or7 = or i32 %or, %and678 %3 = load i8, ptr %value.addr, align 1, !tbaa !379 %conv8 = sext i8 %3 to i3280 %shl9 = shl i32 %conv8, 2481 %and10 = and i32 %shl9, -1677721682 %or11 = or i32 %or7, %and1083 store i32 %or11, ptr %packedValue, align 4, !tbaa !884 %4 = load i32, ptr %blockSize.addr, align 4, !tbaa !885 %shr = lshr i32 %4, 286 store i32 %shr, ptr %blkCnt, align 4, !tbaa !887 br label %while.cond88 89while.cond: ; preds = %while.body, %entry90 %5 = load i32, ptr %blkCnt, align 4, !tbaa !891 %cmp = icmp ugt i32 %5, 092 br i1 %cmp, label %while.body, label %while.end93 94while.body: ; preds = %while.cond95 %6 = load i32, ptr %packedValue, align 4, !tbaa !896 call void @write_q7x4_ia(ptr %pDst.addr, i32 %6)97 %7 = load i32, ptr %blkCnt, align 4, !tbaa !898 %dec = add i32 %7, -199 store i32 %dec, ptr %blkCnt, align 4, !tbaa !8100 br label %while.cond, !llvm.loop !10101 102while.end: ; preds = %while.cond103 %8 = load i32, ptr %blockSize.addr, align 4, !tbaa !8104 %rem = urem i32 %8, 4105 store i32 %rem, ptr %blkCnt, align 4, !tbaa !8106 br label %while.cond13107 108while.cond13: ; preds = %while.body16, %while.end109 %9 = load i32, ptr %blkCnt, align 4, !tbaa !8110 %cmp14 = icmp ugt i32 %9, 0111 br i1 %cmp14, label %while.body16, label %while.end18112 113while.body16: ; preds = %while.cond13114 %10 = load i8, ptr %value.addr, align 1, !tbaa !3115 %11 = load ptr, ptr %pDst.addr, align 4, !tbaa !6116 %incdec.ptr = getelementptr inbounds i8, ptr %11, i32 1117 store ptr %incdec.ptr, ptr %pDst.addr, align 4, !tbaa !6118 store i8 %10, ptr %11, align 1, !tbaa !3119 %12 = load i32, ptr %blkCnt, align 4, !tbaa !8120 %dec17 = add i32 %12, -1121 store i32 %dec17, ptr %blkCnt, align 4, !tbaa !8122 br label %while.cond13, !llvm.loop !12123 124while.end18: ; preds = %while.cond13125 call void @llvm.lifetime.end.p0(ptr %packedValue)126 call void @llvm.lifetime.end.p0(ptr %blkCnt)127 ret void128}129 130; Function Attrs: argmemonly nofree nosync nounwind willreturn131declare void @llvm.lifetime.start.p0(ptr nocapture)132 133; Function Attrs: alwaysinline nounwind134define internal void @write_q7x4_ia(ptr %pQ7, i32 %value) {135entry:136 %pQ7.addr = alloca ptr, align 4137 %value.addr = alloca i32, align 4138 %val = alloca i32, align 4139 store ptr %pQ7, ptr %pQ7.addr, align 4, !tbaa !6140 store i32 %value, ptr %value.addr, align 4, !tbaa !8141 call void @llvm.lifetime.start.p0(ptr %val)142 %0 = load i32, ptr %value.addr, align 4, !tbaa !8143 store i32 %0, ptr %val, align 4, !tbaa !8144 %1 = load i32, ptr %val, align 4, !tbaa !8145 %and = and i32 %1, 255146 %conv = trunc i32 %and to i8147 %2 = load ptr, ptr %pQ7.addr, align 4, !tbaa !6148 %3 = load ptr, ptr %2, align 4, !tbaa !6149 store i8 %conv, ptr %3, align 1, !tbaa !3150 %4 = load i32, ptr %val, align 4, !tbaa !8151 %shr = ashr i32 %4, 8152 %and1 = and i32 %shr, 255153 %conv2 = trunc i32 %and1 to i8154 %5 = load ptr, ptr %pQ7.addr, align 4, !tbaa !6155 %6 = load ptr, ptr %5, align 4, !tbaa !6156 %arrayidx3 = getelementptr inbounds i8, ptr %6, i32 1157 store i8 %conv2, ptr %arrayidx3, align 1, !tbaa !3158 %7 = load i32, ptr %val, align 4, !tbaa !8159 %shr4 = ashr i32 %7, 16160 %and5 = and i32 %shr4, 255161 %conv6 = trunc i32 %and5 to i8162 %8 = load ptr, ptr %pQ7.addr, align 4, !tbaa !6163 %9 = load ptr, ptr %8, align 4, !tbaa !6164 %arrayidx7 = getelementptr inbounds i8, ptr %9, i32 2165 store i8 %conv6, ptr %arrayidx7, align 1, !tbaa !3166 %10 = load i32, ptr %val, align 4, !tbaa !8167 %shr8 = ashr i32 %10, 24168 %and9 = and i32 %shr8, 255169 %conv10 = trunc i32 %and9 to i8170 %11 = load ptr, ptr %pQ7.addr, align 4, !tbaa !6171 %12 = load ptr, ptr %11, align 4, !tbaa !6172 %arrayidx11 = getelementptr inbounds i8, ptr %12, i32 3173 store i8 %conv10, ptr %arrayidx11, align 1, !tbaa !3174 %13 = load ptr, ptr %pQ7.addr, align 4, !tbaa !6175 %14 = load ptr, ptr %13, align 4, !tbaa !6176 %add.ptr = getelementptr inbounds i8, ptr %14, i32 4177 store ptr %add.ptr, ptr %13, align 4, !tbaa !6178 call void @llvm.lifetime.end.p0(ptr %val)179 ret void180}181 182; Function Attrs: argmemonly nofree nosync nounwind willreturn183declare void @llvm.lifetime.end.p0(ptr nocapture)184 185!llvm.module.flags = !{!0, !1}186!llvm.ident = !{!2}187 188!0 = !{i32 1, !"wchar_size", i32 4}189!1 = !{i32 1, !"min_enum_size", i32 4}190!2 = !{!"clang version 12.0.0 (https://github.com/llvm/llvm-project.git 07f234be1ccbce131704f580aa3f117083a887f7)"}191!3 = !{!4, !4, i64 0}192!4 = !{!"omnipotent char", !5, i64 0}193!5 = !{!"Simple C/C++ TBAA"}194!6 = !{!7, !7, i64 0}195!7 = !{!"any pointer", !4, i64 0}196!8 = !{!9, !9, i64 0}197!9 = !{!"int", !4, i64 0}198!10 = distinct !{!10, !11}199!11 = !{!"llvm.loop.mustprogress"}200!12 = distinct !{!12, !11}201