257 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: %if x86-registered-target %{ opt < %s -passes=slp-vectorizer -S -mtriple=x86_64-apple-macosx10.10.0 -mattr=+sse4.2 | FileCheck %s %}3; RUN: %if aarch64-registered-target %{ opt < %s -passes=slp-vectorizer -S -mtriple=aarch64-unknown-linux-gnu -mattr=+neon | FileCheck %s %}4 5; PR284746 7;void foo();8;9;int test1(unsigned int *p) {10; int sum = 0;11; #pragma nounroll12; for (int y = 0; y < 2; y++) {13; // Inner loop gets unrolled14; for (int x = 0; x < 8; x++) {15; sum += pptr 42;16; }17; // Dummy call to keep outer loop alive18; foo();19; }20; return sum;21;}22 23define i32 @test(ptr nocapture readonly %p) {24; CHECK-LABEL: @test(25; CHECK-NEXT: entry:26; CHECK-NEXT: br label [[FOR_BODY:%.*]]27; CHECK: for.body:28; CHECK-NEXT: [[SUM:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[OP_RDX:%.*]], [[FOR_BODY]] ]29; CHECK-NEXT: [[TMP1:%.*]] = load <8 x i32>, ptr [[P:%.*]], align 430; CHECK-NEXT: [[TMP2:%.*]] = mul <8 x i32> [[TMP1]], splat (i32 42)31; CHECK-NEXT: [[TMP3:%.*]] = call i32 @llvm.vector.reduce.add.v8i32(<8 x i32> [[TMP2]])32; CHECK-NEXT: [[OP_RDX]] = add i32 [[TMP3]], [[SUM]]33; CHECK-NEXT: br i1 true, label [[FOR_END:%.*]], label [[FOR_BODY]]34; CHECK: for.end:35; CHECK-NEXT: ret i32 [[OP_RDX]]36;37entry:38 %arrayidx.1 = getelementptr inbounds i32, ptr %p, i64 139 %arrayidx.2 = getelementptr inbounds i32, ptr %p, i64 240 %arrayidx.3 = getelementptr inbounds i32, ptr %p, i64 341 %arrayidx.4 = getelementptr inbounds i32, ptr %p, i64 442 %arrayidx.5 = getelementptr inbounds i32, ptr %p, i64 543 %arrayidx.6 = getelementptr inbounds i32, ptr %p, i64 644 %arrayidx.7 = getelementptr inbounds i32, ptr %p, i64 745 br label %for.body46 47for.body:48 %sum = phi i32 [ 0, %entry ], [ %add.7, %for.body ]49 %tmp = load i32, ptr %p, align 450 %mul = mul i32 %tmp, 4251 %add = add i32 %mul, %sum52 %tmp5 = load i32, ptr %arrayidx.1, align 453 %mul.1 = mul i32 %tmp5, 4254 %add.1 = add i32 %mul.1, %add55 %tmp6 = load i32, ptr %arrayidx.2, align 456 %mul.2 = mul i32 %tmp6, 4257 %add.2 = add i32 %mul.2, %add.158 %tmp7 = load i32, ptr %arrayidx.3, align 459 %mul.3 = mul i32 %tmp7, 4260 %add.3 = add i32 %mul.3, %add.261 %tmp8 = load i32, ptr %arrayidx.4, align 462 %mul.4 = mul i32 %tmp8, 4263 %add.4 = add i32 %mul.4, %add.364 %tmp9 = load i32, ptr %arrayidx.5, align 465 %mul.5 = mul i32 %tmp9, 4266 %add.5 = add i32 %mul.5, %add.467 %tmp10 = load i32, ptr %arrayidx.6, align 468 %mul.6 = mul i32 %tmp10, 4269 %add.6 = add i32 %mul.6, %add.570 %tmp11 = load i32, ptr %arrayidx.7, align 471 %mul.7 = mul i32 %tmp11, 4272 %add.7 = add i32 %mul.7, %add.673 br i1 true, label %for.end, label %for.body74 75for.end:76 ret i32 %add.777}78 79;void foo();80;81;int test2(unsigned int *p, unsigned int *q) {82; int sum = 0;83; #pragma nounroll84; for (int y = 0; y < 2; y++) {85; // Inner loop gets unrolled86; for (int x = 0; x < 8; x++) {87; sum += pptr q[x];88; }89; // Dummy call to keep outer loop alive90; foo();91; }92; return sum;93;}94 95define i32 @test2(ptr nocapture readonly %p, ptr nocapture readonly %q) {96; CHECK-LABEL: @test2(97; CHECK-NEXT: entry:98; CHECK-NEXT: br label [[FOR_BODY:%.*]]99; CHECK: for.body:100; CHECK-NEXT: [[SUM:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[OP_RDX:%.*]], [[FOR_BODY]] ]101; CHECK-NEXT: [[TMP1:%.*]] = load <8 x i32>, ptr [[P:%.*]], align 4102; CHECK-NEXT: [[TMP3:%.*]] = load <8 x i32>, ptr [[Q:%.*]], align 4103; CHECK-NEXT: [[TMP4:%.*]] = mul <8 x i32> [[TMP1]], [[TMP3]]104; CHECK-NEXT: [[TMP5:%.*]] = call i32 @llvm.vector.reduce.add.v8i32(<8 x i32> [[TMP4]])105; CHECK-NEXT: [[OP_RDX]] = add i32 [[TMP5]], [[SUM]]106; CHECK-NEXT: br i1 true, label [[FOR_END:%.*]], label [[FOR_BODY]]107; CHECK: for.end:108; CHECK-NEXT: ret i32 [[OP_RDX]]109;110entry:111 %arrayidx.p.1 = getelementptr inbounds i32, ptr %p, i64 1112 %arrayidx.p.2 = getelementptr inbounds i32, ptr %p, i64 2113 %arrayidx.p.3 = getelementptr inbounds i32, ptr %p, i64 3114 %arrayidx.p.4 = getelementptr inbounds i32, ptr %p, i64 4115 %arrayidx.p.5 = getelementptr inbounds i32, ptr %p, i64 5116 %arrayidx.p.6 = getelementptr inbounds i32, ptr %p, i64 6117 %arrayidx.p.7 = getelementptr inbounds i32, ptr %p, i64 7118 119 %arrayidx.q.1 = getelementptr inbounds i32, ptr %q, i64 1120 %arrayidx.q.2 = getelementptr inbounds i32, ptr %q, i64 2121 %arrayidx.q.3 = getelementptr inbounds i32, ptr %q, i64 3122 %arrayidx.q.4 = getelementptr inbounds i32, ptr %q, i64 4123 %arrayidx.q.5 = getelementptr inbounds i32, ptr %q, i64 5124 %arrayidx.q.6 = getelementptr inbounds i32, ptr %q, i64 6125 %arrayidx.q.7 = getelementptr inbounds i32, ptr %q, i64 7126 br label %for.body127 128for.body:129 %sum = phi i32 [ 0, %entry ], [ %add.7, %for.body ]130 %tmpp = load i32, ptr %p, align 4131 %tmpq = load i32, ptr %q, align 4132 %mul = mul i32 %tmpp, %tmpq133 %add = add i32 %mul, %sum134 %tmp5p = load i32, ptr %arrayidx.p.1, align 4135 %tmp5q = load i32, ptr %arrayidx.q.1, align 4136 %mul.1 = mul i32 %tmp5p, %tmp5q137 %add.1 = add i32 %mul.1, %add138 %tmp6p = load i32, ptr %arrayidx.p.2, align 4139 %tmp6q = load i32, ptr %arrayidx.q.2, align 4140 %mul.2 = mul i32 %tmp6p, %tmp6q141 %add.2 = add i32 %mul.2, %add.1142 %tmp7p = load i32, ptr %arrayidx.p.3, align 4143 %tmp7q = load i32, ptr %arrayidx.q.3, align 4144 %mul.3 = mul i32 %tmp7p, %tmp7q145 %add.3 = add i32 %mul.3, %add.2146 %tmp8p = load i32, ptr %arrayidx.p.4, align 4147 %tmp8q = load i32, ptr %arrayidx.q.4, align 4148 %mul.4 = mul i32 %tmp8p, %tmp8q149 %add.4 = add i32 %mul.4, %add.3150 %tmp9p = load i32, ptr %arrayidx.p.5, align 4151 %tmp9q = load i32, ptr %arrayidx.q.5, align 4152 %mul.5 = mul i32 %tmp9p, %tmp9q153 %add.5 = add i32 %mul.5, %add.4154 %tmp10p = load i32, ptr %arrayidx.p.6, align 4155 %tmp10q = load i32, ptr %arrayidx.q.6, align 4156 %mul.6 = mul i32 %tmp10p, %tmp10q157 %add.6 = add i32 %mul.6, %add.5158 %tmp11p = load i32, ptr %arrayidx.p.7, align 4159 %tmp11q = load i32, ptr %arrayidx.q.7, align 4160 %mul.7 = mul i32 %tmp11p, %tmp11q161 %add.7 = add i32 %mul.7, %add.6162 br i1 true, label %for.end, label %for.body163 164for.end:165 ret i32 %add.7166}167 168;void foo();169;170;int test3(unsigned int *p, unsigned int *q) {171; int sum = 0;172; #pragma nounroll173; for (int y = 0; y < 2; y++) {174; // Inner loop gets unrolled175; for (int x = 0; x < 8; x++) {176; sum += pptr q[7-x];177; }178; // Dummy call to keep outer loop alive179; foo();180; }181; return sum;182;}183 184define i32 @test3(ptr nocapture readonly %p, ptr nocapture readonly %q) {185; CHECK-LABEL: @test3(186; CHECK-NEXT: entry:187; CHECK-NEXT: br label [[FOR_BODY:%.*]]188; CHECK: for.body:189; CHECK-NEXT: [[SUM:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[OP_RDX:%.*]], [[FOR_BODY]] ]190; CHECK-NEXT: [[TMP1:%.*]] = load <8 x i32>, ptr [[P:%.*]], align 4191; CHECK-NEXT: [[TMP3:%.*]] = load <8 x i32>, ptr [[Q:%.*]], align 4192; CHECK-NEXT: [[SHUFFLE:%.*]] = shufflevector <8 x i32> [[TMP3]], <8 x i32> poison, <8 x i32> <i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0>193; CHECK-NEXT: [[TMP4:%.*]] = mul <8 x i32> [[TMP1]], [[SHUFFLE]]194; CHECK-NEXT: [[TMP5:%.*]] = call i32 @llvm.vector.reduce.add.v8i32(<8 x i32> [[TMP4]])195; CHECK-NEXT: [[OP_RDX]] = add i32 [[TMP5]], [[SUM]]196; CHECK-NEXT: br i1 true, label [[FOR_END:%.*]], label [[FOR_BODY]]197; CHECK: for.end:198; CHECK-NEXT: ret i32 [[OP_RDX]]199;200entry:201 %arrayidx.p.1 = getelementptr inbounds i32, ptr %p, i64 1202 %arrayidx.p.2 = getelementptr inbounds i32, ptr %p, i64 2203 %arrayidx.p.3 = getelementptr inbounds i32, ptr %p, i64 3204 %arrayidx.p.4 = getelementptr inbounds i32, ptr %p, i64 4205 %arrayidx.p.5 = getelementptr inbounds i32, ptr %p, i64 5206 %arrayidx.p.6 = getelementptr inbounds i32, ptr %p, i64 6207 %arrayidx.p.7 = getelementptr inbounds i32, ptr %p, i64 7208 209 %arrayidx.q.1 = getelementptr inbounds i32, ptr %q, i64 1210 %arrayidx.q.2 = getelementptr inbounds i32, ptr %q, i64 2211 %arrayidx.q.3 = getelementptr inbounds i32, ptr %q, i64 3212 %arrayidx.q.4 = getelementptr inbounds i32, ptr %q, i64 4213 %arrayidx.q.5 = getelementptr inbounds i32, ptr %q, i64 5214 %arrayidx.q.6 = getelementptr inbounds i32, ptr %q, i64 6215 %arrayidx.q.7 = getelementptr inbounds i32, ptr %q, i64 7216 br label %for.body217 218for.body:219 %sum = phi i32 [ 0, %entry ], [ %add.7, %for.body ]220 %tmpp = load i32, ptr %p, align 4221 %tmpq = load i32, ptr %arrayidx.q.7, align 4222 %mul = mul i32 %tmpp, %tmpq223 %add = add i32 %mul, %sum224 %tmp5p = load i32, ptr %arrayidx.p.1, align 4225 %tmp5q = load i32, ptr %arrayidx.q.6, align 4226 %mul.1 = mul i32 %tmp5p, %tmp5q227 %add.1 = add i32 %mul.1, %add228 %tmp6p = load i32, ptr %arrayidx.p.2, align 4229 %tmp6q = load i32, ptr %arrayidx.q.5, align 4230 %mul.2 = mul i32 %tmp6p, %tmp6q231 %add.2 = add i32 %mul.2, %add.1232 %tmp7p = load i32, ptr %arrayidx.p.3, align 4233 %tmp7q = load i32, ptr %arrayidx.q.4, align 4234 %mul.3 = mul i32 %tmp7p, %tmp7q235 %add.3 = add i32 %mul.3, %add.2236 %tmp8p = load i32, ptr %arrayidx.p.4, align 4237 %tmp8q = load i32, ptr %arrayidx.q.3, align 4238 %mul.4 = mul i32 %tmp8p, %tmp8q239 %add.4 = add i32 %mul.4, %add.3240 %tmp9p = load i32, ptr %arrayidx.p.5, align 4241 %tmp9q = load i32, ptr %arrayidx.q.2, align 4242 %mul.5 = mul i32 %tmp9p, %tmp9q243 %add.5 = add i32 %mul.5, %add.4244 %tmp10p = load i32, ptr %arrayidx.p.6, align 4245 %tmp10q = load i32, ptr %arrayidx.q.1, align 4246 %mul.6 = mul i32 %tmp10p, %tmp10q247 %add.6 = add i32 %mul.6, %add.5248 %tmp11p = load i32, ptr %arrayidx.p.7, align 4249 %tmp11q = load i32, ptr %q, align 4250 %mul.7 = mul i32 %tmp11p, %tmp11q251 %add.7 = add i32 %mul.7, %add.6252 br i1 true, label %for.end, label %for.body253 254for.end:255 ret i32 %add.7256}257