188 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: opt < %s -passes=loop-vectorize -force-streaming-compatible -enable-fixedwidth-autovec-in-streaming-mode -mattr=+sve -force-target-instruction-cost=1 -scalable-vectorization=off -force-vector-interleave=1 -S 2>&1 | FileCheck %s --check-prefix=SC_SVE3; RUN: opt < %s -passes=loop-vectorize -mattr=+sve -force-target-instruction-cost=1 -scalable-vectorization=off -force-vector-interleave=1 -S 2>&1 | FileCheck %s --check-prefix=NO_SC_SVE4 5target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"6target triple = "aarch64-unknown-linux-gnu"7 8@a = global [32 x i16] zeroinitializer, align 29@b = global [32 x i16] zeroinitializer, align 210@c = global [32 x i16] zeroinitializer, align 211 12define i32 @foo(i32 noundef %n, i32 noundef %lag, i32 noundef %shift) vscale_range(1,16) {13; SC_SVE-LABEL: @foo(14; SC_SVE-NEXT: entry:15; SC_SVE-NEXT: [[TMP0:%.*]] = sext i32 [[LAG:%.*]] to i6416; SC_SVE-NEXT: [[WIDE_TRIP_COUNT:%.*]] = zext i32 [[N:%.*]] to i6417; SC_SVE-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[WIDE_TRIP_COUNT]], 418; SC_SVE-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]19; SC_SVE: vector.ph:20; SC_SVE-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[WIDE_TRIP_COUNT]], 421; SC_SVE-NEXT: [[N_VEC:%.*]] = sub i64 [[WIDE_TRIP_COUNT]], [[N_MOD_VF]]22; SC_SVE-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x i32> poison, i32 [[SHIFT:%.*]], i64 023; SC_SVE-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x i32> [[BROADCAST_SPLATINSERT]], <4 x i32> poison, <4 x i32> zeroinitializer24; SC_SVE-NEXT: br label [[VECTOR_BODY:%.*]]25; SC_SVE: vector.body:26; SC_SVE-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]27; SC_SVE-NEXT: [[VEC_PHI:%.*]] = phi <4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP17:%.*]], [[VECTOR_BODY]] ]28; SC_SVE-NEXT: [[VEC_IND:%.*]] = phi <4 x i32> [ <i32 0, i32 1, i32 2, i32 3>, [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]29; SC_SVE-NEXT: [[TMP2:%.*]] = getelementptr inbounds [32 x i16], ptr @a, i64 0, i64 [[INDEX]]30; SC_SVE-NEXT: [[WIDE_LOAD:%.*]] = load <4 x i16>, ptr [[TMP2]], align 231; SC_SVE-NEXT: [[TMP4:%.*]] = sext <4 x i16> [[WIDE_LOAD]] to <4 x i32>32; SC_SVE-NEXT: [[TMP5:%.*]] = ashr <4 x i32> [[TMP4]], [[VEC_IND]]33; SC_SVE-NEXT: [[TMP6:%.*]] = add nsw i64 [[INDEX]], [[TMP0]]34; SC_SVE-NEXT: [[TMP7:%.*]] = getelementptr inbounds [32 x i16], ptr @b, i64 0, i64 [[TMP6]]35; SC_SVE-NEXT: [[WIDE_LOAD1:%.*]] = load <4 x i16>, ptr [[TMP7]], align 236; SC_SVE-NEXT: [[TMP9:%.*]] = sext <4 x i16> [[WIDE_LOAD1]] to <4 x i32>37; SC_SVE-NEXT: [[TMP10:%.*]] = shl <4 x i32> [[TMP9]], [[VEC_IND]]38; SC_SVE-NEXT: [[TMP11:%.*]] = mul nsw <4 x i32> [[TMP10]], [[TMP5]]39; SC_SVE-NEXT: [[TMP12:%.*]] = getelementptr inbounds [32 x i16], ptr @c, i64 0, i64 [[INDEX]]40; SC_SVE-NEXT: [[WIDE_LOAD2:%.*]] = load <4 x i16>, ptr [[TMP12]], align 241; SC_SVE-NEXT: [[TMP14:%.*]] = sext <4 x i16> [[WIDE_LOAD2]] to <4 x i32>42; SC_SVE-NEXT: [[TMP15:%.*]] = add nsw <4 x i32> [[TMP11]], [[TMP14]]43; SC_SVE-NEXT: [[TMP16:%.*]] = shl <4 x i32> [[TMP15]], [[BROADCAST_SPLAT]]44; SC_SVE-NEXT: [[TMP17]] = add <4 x i32> [[TMP16]], [[VEC_PHI]]45; SC_SVE-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 446; SC_SVE-NEXT: [[VEC_IND_NEXT]] = add <4 x i32> [[VEC_IND]], splat (i32 4)47; SC_SVE-NEXT: [[TMP18:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]48; SC_SVE-NEXT: br i1 [[TMP18]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]49; SC_SVE: middle.block:50; SC_SVE-NEXT: [[TMP19:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[TMP17]])51; SC_SVE-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[WIDE_TRIP_COUNT]], [[N_VEC]]52; SC_SVE-NEXT: br i1 [[CMP_N]], label [[FOR_END:%.*]], label [[SCALAR_PH]]53; SC_SVE: scalar.ph:54; SC_SVE-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]55; SC_SVE-NEXT: [[BC_MERGE_RDX:%.*]] = phi i32 [ [[TMP19]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]56; SC_SVE-NEXT: br label [[FOR_BODY:%.*]]57; SC_SVE: for.body:58; SC_SVE-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]59; SC_SVE-NEXT: [[RET_018:%.*]] = phi i32 [ [[BC_MERGE_RDX]], [[SCALAR_PH]] ], [ [[ADD9:%.*]], [[FOR_BODY]] ]60; SC_SVE-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds [32 x i16], ptr @a, i64 0, i64 [[INDVARS_IV]]61; SC_SVE-NEXT: [[TMP20:%.*]] = load i16, ptr [[ARRAYIDX]], align 262; SC_SVE-NEXT: [[CONV:%.*]] = sext i16 [[TMP20]] to i3263; SC_SVE-NEXT: [[TMP21:%.*]] = trunc i64 [[INDVARS_IV]] to i3264; SC_SVE-NEXT: [[SHR:%.*]] = ashr i32 [[CONV]], [[TMP21]]65; SC_SVE-NEXT: [[TMP22:%.*]] = add nsw i64 [[INDVARS_IV]], [[TMP0]]66; SC_SVE-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds [32 x i16], ptr @b, i64 0, i64 [[TMP22]]67; SC_SVE-NEXT: [[TMP23:%.*]] = load i16, ptr [[ARRAYIDX2]], align 268; SC_SVE-NEXT: [[CONV3:%.*]] = sext i16 [[TMP23]] to i3269; SC_SVE-NEXT: [[SHL:%.*]] = shl i32 [[CONV3]], [[TMP21]]70; SC_SVE-NEXT: [[MUL:%.*]] = mul nsw i32 [[SHL]], [[SHR]]71; SC_SVE-NEXT: [[ARRAYIDX5:%.*]] = getelementptr inbounds [32 x i16], ptr @c, i64 0, i64 [[INDVARS_IV]]72; SC_SVE-NEXT: [[TMP24:%.*]] = load i16, ptr [[ARRAYIDX5]], align 273; SC_SVE-NEXT: [[CONV6:%.*]] = sext i16 [[TMP24]] to i3274; SC_SVE-NEXT: [[ADD7:%.*]] = add nsw i32 [[MUL]], [[CONV6]]75; SC_SVE-NEXT: [[SHL8:%.*]] = shl i32 [[ADD7]], [[SHIFT]]76; SC_SVE-NEXT: [[ADD9]] = add nsw i32 [[SHL8]], [[RET_018]]77; SC_SVE-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 178; SC_SVE-NEXT: [[EXITCOND_NOT:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], [[WIDE_TRIP_COUNT]]79; SC_SVE-NEXT: br i1 [[EXITCOND_NOT]], label [[FOR_END]], label [[FOR_BODY]], !llvm.loop [[LOOP3:![0-9]+]]80; SC_SVE: for.end:81; SC_SVE-NEXT: [[RET_0_LCSSA:%.*]] = phi i32 [ [[ADD9]], [[FOR_BODY]] ], [ [[TMP19]], [[MIDDLE_BLOCK]] ]82; SC_SVE-NEXT: ret i32 [[RET_0_LCSSA]]83;84; NO_SC_SVE-LABEL: @foo(85; NO_SC_SVE-NEXT: entry:86; NO_SC_SVE-NEXT: [[TMP0:%.*]] = sext i32 [[LAG:%.*]] to i6487; NO_SC_SVE-NEXT: [[WIDE_TRIP_COUNT:%.*]] = zext i32 [[N:%.*]] to i6488; NO_SC_SVE-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[WIDE_TRIP_COUNT]], 889; NO_SC_SVE-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]90; NO_SC_SVE: vector.ph:91; NO_SC_SVE-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[WIDE_TRIP_COUNT]], 892; NO_SC_SVE-NEXT: [[N_VEC:%.*]] = sub i64 [[WIDE_TRIP_COUNT]], [[N_MOD_VF]]93; NO_SC_SVE-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <8 x i32> poison, i32 [[SHIFT:%.*]], i64 094; NO_SC_SVE-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <8 x i32> [[BROADCAST_SPLATINSERT]], <8 x i32> poison, <8 x i32> zeroinitializer95; NO_SC_SVE-NEXT: br label [[VECTOR_BODY:%.*]]96; NO_SC_SVE: vector.body:97; NO_SC_SVE-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]98; NO_SC_SVE-NEXT: [[VEC_PHI:%.*]] = phi <8 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP17:%.*]], [[VECTOR_BODY]] ]99; NO_SC_SVE-NEXT: [[VEC_IND:%.*]] = phi <8 x i32> [ <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>, [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ]100; NO_SC_SVE-NEXT: [[TMP2:%.*]] = getelementptr inbounds [32 x i16], ptr @a, i64 0, i64 [[INDEX]]101; NO_SC_SVE-NEXT: [[WIDE_LOAD:%.*]] = load <8 x i16>, ptr [[TMP2]], align 2102; NO_SC_SVE-NEXT: [[TMP4:%.*]] = sext <8 x i16> [[WIDE_LOAD]] to <8 x i32>103; NO_SC_SVE-NEXT: [[TMP5:%.*]] = ashr <8 x i32> [[TMP4]], [[VEC_IND]]104; NO_SC_SVE-NEXT: [[TMP6:%.*]] = add nsw i64 [[INDEX]], [[TMP0]]105; NO_SC_SVE-NEXT: [[TMP7:%.*]] = getelementptr inbounds [32 x i16], ptr @b, i64 0, i64 [[TMP6]]106; NO_SC_SVE-NEXT: [[WIDE_LOAD1:%.*]] = load <8 x i16>, ptr [[TMP7]], align 2107; NO_SC_SVE-NEXT: [[TMP9:%.*]] = sext <8 x i16> [[WIDE_LOAD1]] to <8 x i32>108; NO_SC_SVE-NEXT: [[TMP10:%.*]] = shl <8 x i32> [[TMP9]], [[VEC_IND]]109; NO_SC_SVE-NEXT: [[TMP11:%.*]] = mul nsw <8 x i32> [[TMP10]], [[TMP5]]110; NO_SC_SVE-NEXT: [[TMP12:%.*]] = getelementptr inbounds [32 x i16], ptr @c, i64 0, i64 [[INDEX]]111; NO_SC_SVE-NEXT: [[WIDE_LOAD2:%.*]] = load <8 x i16>, ptr [[TMP12]], align 2112; NO_SC_SVE-NEXT: [[TMP14:%.*]] = sext <8 x i16> [[WIDE_LOAD2]] to <8 x i32>113; NO_SC_SVE-NEXT: [[TMP15:%.*]] = add nsw <8 x i32> [[TMP11]], [[TMP14]]114; NO_SC_SVE-NEXT: [[TMP16:%.*]] = shl <8 x i32> [[TMP15]], [[BROADCAST_SPLAT]]115; NO_SC_SVE-NEXT: [[TMP17]] = add <8 x i32> [[TMP16]], [[VEC_PHI]]116; NO_SC_SVE-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 8117; NO_SC_SVE-NEXT: [[VEC_IND_NEXT]] = add <8 x i32> [[VEC_IND]], splat (i32 8)118; NO_SC_SVE-NEXT: [[TMP18:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]119; NO_SC_SVE-NEXT: br i1 [[TMP18]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]120; NO_SC_SVE: middle.block:121; NO_SC_SVE-NEXT: [[TMP19:%.*]] = call i32 @llvm.vector.reduce.add.v8i32(<8 x i32> [[TMP17]])122; NO_SC_SVE-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[WIDE_TRIP_COUNT]], [[N_VEC]]123; NO_SC_SVE-NEXT: br i1 [[CMP_N]], label [[FOR_END:%.*]], label [[SCALAR_PH]]124; NO_SC_SVE: scalar.ph:125; NO_SC_SVE-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]126; NO_SC_SVE-NEXT: [[BC_MERGE_RDX:%.*]] = phi i32 [ [[TMP19]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]127; NO_SC_SVE-NEXT: br label [[FOR_BODY:%.*]]128; NO_SC_SVE: for.body:129; NO_SC_SVE-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]130; NO_SC_SVE-NEXT: [[RET_018:%.*]] = phi i32 [ [[BC_MERGE_RDX]], [[SCALAR_PH]] ], [ [[ADD9:%.*]], [[FOR_BODY]] ]131; NO_SC_SVE-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds [32 x i16], ptr @a, i64 0, i64 [[INDVARS_IV]]132; NO_SC_SVE-NEXT: [[TMP20:%.*]] = load i16, ptr [[ARRAYIDX]], align 2133; NO_SC_SVE-NEXT: [[CONV:%.*]] = sext i16 [[TMP20]] to i32134; NO_SC_SVE-NEXT: [[TMP21:%.*]] = trunc i64 [[INDVARS_IV]] to i32135; NO_SC_SVE-NEXT: [[SHR:%.*]] = ashr i32 [[CONV]], [[TMP21]]136; NO_SC_SVE-NEXT: [[TMP22:%.*]] = add nsw i64 [[INDVARS_IV]], [[TMP0]]137; NO_SC_SVE-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds [32 x i16], ptr @b, i64 0, i64 [[TMP22]]138; NO_SC_SVE-NEXT: [[TMP23:%.*]] = load i16, ptr [[ARRAYIDX2]], align 2139; NO_SC_SVE-NEXT: [[CONV3:%.*]] = sext i16 [[TMP23]] to i32140; NO_SC_SVE-NEXT: [[SHL:%.*]] = shl i32 [[CONV3]], [[TMP21]]141; NO_SC_SVE-NEXT: [[MUL:%.*]] = mul nsw i32 [[SHL]], [[SHR]]142; NO_SC_SVE-NEXT: [[ARRAYIDX5:%.*]] = getelementptr inbounds [32 x i16], ptr @c, i64 0, i64 [[INDVARS_IV]]143; NO_SC_SVE-NEXT: [[TMP24:%.*]] = load i16, ptr [[ARRAYIDX5]], align 2144; NO_SC_SVE-NEXT: [[CONV6:%.*]] = sext i16 [[TMP24]] to i32145; NO_SC_SVE-NEXT: [[ADD7:%.*]] = add nsw i32 [[MUL]], [[CONV6]]146; NO_SC_SVE-NEXT: [[SHL8:%.*]] = shl i32 [[ADD7]], [[SHIFT]]147; NO_SC_SVE-NEXT: [[ADD9]] = add nsw i32 [[SHL8]], [[RET_018]]148; NO_SC_SVE-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1149; NO_SC_SVE-NEXT: [[EXITCOND_NOT:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], [[WIDE_TRIP_COUNT]]150; NO_SC_SVE-NEXT: br i1 [[EXITCOND_NOT]], label [[FOR_END]], label [[FOR_BODY]], !llvm.loop [[LOOP3:![0-9]+]]151; NO_SC_SVE: for.end:152; NO_SC_SVE-NEXT: [[RET_0_LCSSA:%.*]] = phi i32 [ [[ADD9]], [[FOR_BODY]] ], [ [[TMP19]], [[MIDDLE_BLOCK]] ]153; NO_SC_SVE-NEXT: ret i32 [[RET_0_LCSSA]]154;155entry:156 %0 = sext i32 %lag to i64157 %wide.trip.count = zext i32 %n to i64158 br label %for.body159 160for.body:161 %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]162 %ret.018 = phi i32 [ 0, %entry ], [ %add9, %for.body ]163 %arrayidx = getelementptr inbounds [32 x i16], ptr @a, i64 0, i64 %indvars.iv164 %1 = load i16, ptr %arrayidx, align 2165 %conv = sext i16 %1 to i32166 %2 = trunc i64 %indvars.iv to i32167 %shr = ashr i32 %conv, %2168 %3 = add nsw i64 %indvars.iv, %0169 %arrayidx2 = getelementptr inbounds [32 x i16], ptr @b, i64 0, i64 %3170 %4 = load i16, ptr %arrayidx2, align 2171 %conv3 = sext i16 %4 to i32172 %shl = shl i32 %conv3, %2173 %mul = mul nsw i32 %shl, %shr174 %arrayidx5 = getelementptr inbounds [32 x i16], ptr @c, i64 0, i64 %indvars.iv175 %5 = load i16, ptr %arrayidx5, align 2176 %conv6 = sext i16 %5 to i32177 %add7 = add nsw i32 %mul, %conv6178 %shl8 = shl i32 %add7, %shift179 %add9 = add nsw i32 %shl8, %ret.018180 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1181 %exitcond.not = icmp eq i64 %indvars.iv.next, %wide.trip.count182 br i1 %exitcond.not, label %for.end, label %for.body183 184for.end:185 %ret.0.lcssa = phi i32 [ %add9, %for.body ]186 ret i32 %ret.0.lcssa187}188