1; RUN: opt < %s -basicaa -slp-vectorizer -slp-threshold=-100 -instcombine -dce -S -mtriple=i386-apple-macosx10.8.0 -mcpu=corei7-avx | FileCheck %s 2 3target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:64:64-v128:128:128-a0:0:64-f80:128:128-n8:16:32-S128" 4 5 6 7; Make sure we order the operands of commutative operations so that we get 8; bigger vectorizable trees. 9 10; CHECK-LABEL: shuffle_operands1 11; CHECK: load <2 x double> 12; CHECK: fadd <2 x double> 13 14define void @shuffle_operands1(double * noalias %from, double * noalias %to, 15 double %v1, double %v2) { 16 %from_1 = getelementptr double, double *%from, i64 1 17 %v0_1 = load double , double * %from 18 %v0_2 = load double , double * %from_1 19 %v1_1 = fadd double %v0_1, %v1 20 %v1_2 = fadd double %v2, %v0_2 21 %to_2 = getelementptr double, double * %to, i64 1 22 store double %v1_1, double *%to 23 store double %v1_2, double *%to_2 24 ret void 25} 26 27; CHECK-LABEL: shuffle_preserve_broadcast 28; CHECK: %[[BCAST:[a-z0-9]+]] = insertelement <2 x double> undef, double %v0_1 29; CHECK: = insertelement <2 x double> %[[BCAST]], double %v0_1 30define void @shuffle_preserve_broadcast(double * noalias %from, 31 double * noalias %to, 32 double %v1, double %v2) { 33entry: 34br label %lp 35 36lp: 37 %p = phi double [ 1.000000e+00, %lp ], [ 0.000000e+00, %entry ] 38 %from_1 = getelementptr double, double *%from, i64 1 39 %v0_1 = load double , double * %from 40 %v0_2 = load double , double * %from_1 41 %v1_1 = fadd double %v0_1, %p 42 %v1_2 = fadd double %v0_1, %v0_2 43 %to_2 = getelementptr double, double * %to, i64 1 44 store double %v1_1, double *%to 45 store double %v1_2, double *%to_2 46br i1 undef, label %lp, label %ext 47 48ext: 49 ret void 50} 51 52; CHECK-LABEL: shuffle_preserve_broadcast2 53; CHECK: %[[BCAST:[a-z0-9]+]] = insertelement <2 x double> undef, double %v0_1 54; CHECK: = insertelement <2 x double> %[[BCAST]], double %v0_1 55define void @shuffle_preserve_broadcast2(double * noalias %from, 56 double * noalias %to, 57 double %v1, double %v2) { 58entry: 59br label %lp 60 61lp: 62 %p = phi double [ 1.000000e+00, %lp ], [ 0.000000e+00, %entry ] 63 %from_1 = getelementptr double, double *%from, i64 1 64 %v0_1 = load double , double * %from 65 %v0_2 = load double , double * %from_1 66 %v1_1 = fadd double %p, %v0_1 67 %v1_2 = fadd double %v0_2, %v0_1 68 %to_2 = getelementptr double, double * %to, i64 1 69 store double %v1_1, double *%to 70 store double %v1_2, double *%to_2 71br i1 undef, label %lp, label %ext 72 73ext: 74 ret void 75} 76 77; CHECK-LABEL: shuffle_preserve_broadcast3 78; CHECK: %[[BCAST:[a-z0-9]+]] = insertelement <2 x double> undef, double %v0_1 79; CHECK: = insertelement <2 x double> %[[BCAST]], double %v0_1 80define void @shuffle_preserve_broadcast3(double * noalias %from, 81 double * noalias %to, 82 double %v1, double %v2) { 83entry: 84br label %lp 85 86lp: 87 %p = phi double [ 1.000000e+00, %lp ], [ 0.000000e+00, %entry ] 88 %from_1 = getelementptr double, double *%from, i64 1 89 %v0_1 = load double , double * %from 90 %v0_2 = load double , double * %from_1 91 %v1_1 = fadd double %p, %v0_1 92 %v1_2 = fadd double %v0_1, %v0_2 93 %to_2 = getelementptr double, double * %to, i64 1 94 store double %v1_1, double *%to 95 store double %v1_2, double *%to_2 96br i1 undef, label %lp, label %ext 97 98ext: 99 ret void 100} 101 102 103; CHECK-LABEL: shuffle_preserve_broadcast4 104; CHECK: %[[BCAST:[a-z0-9]+]] = insertelement <2 x double> undef, double %v0_1 105; CHECK: = insertelement <2 x double> %[[BCAST]], double %v0_1 106define void @shuffle_preserve_broadcast4(double * noalias %from, 107 double * noalias %to, 108 double %v1, double %v2) { 109entry: 110br label %lp 111 112lp: 113 %p = phi double [ 1.000000e+00, %lp ], [ 0.000000e+00, %entry ] 114 %from_1 = getelementptr double, double *%from, i64 1 115 %v0_1 = load double , double * %from 116 %v0_2 = load double , double * %from_1 117 %v1_1 = fadd double %v0_2, %v0_1 118 %v1_2 = fadd double %p, %v0_1 119 %to_2 = getelementptr double, double * %to, i64 1 120 store double %v1_1, double *%to 121 store double %v1_2, double *%to_2 122br i1 undef, label %lp, label %ext 123 124ext: 125 ret void 126} 127 128; CHECK-LABEL: shuffle_preserve_broadcast5 129; CHECK: %[[BCAST:[a-z0-9]+]] = insertelement <2 x double> undef, double %v0_1 130; CHECK: = insertelement <2 x double> %[[BCAST]], double %v0_1 131define void @shuffle_preserve_broadcast5(double * noalias %from, 132 double * noalias %to, 133 double %v1, double %v2) { 134entry: 135br label %lp 136 137lp: 138 %p = phi double [ 1.000000e+00, %lp ], [ 0.000000e+00, %entry ] 139 %from_1 = getelementptr double, double *%from, i64 1 140 %v0_1 = load double , double * %from 141 %v0_2 = load double , double * %from_1 142 %v1_1 = fadd double %v0_1, %v0_2 143 %v1_2 = fadd double %p, %v0_1 144 %to_2 = getelementptr double, double * %to, i64 1 145 store double %v1_1, double *%to 146 store double %v1_2, double *%to_2 147br i1 undef, label %lp, label %ext 148 149ext: 150 ret void 151} 152 153 154; CHECK-LABEL: shuffle_preserve_broadcast6 155; CHECK: %[[BCAST:[a-z0-9]+]] = insertelement <2 x double> undef, double %v0_1 156; CHECK: = insertelement <2 x double> %[[BCAST]], double %v0_1 157define void @shuffle_preserve_broadcast6(double * noalias %from, 158 double * noalias %to, 159 double %v1, double %v2) { 160entry: 161br label %lp 162 163lp: 164 %p = phi double [ 1.000000e+00, %lp ], [ 0.000000e+00, %entry ] 165 %from_1 = getelementptr double, double *%from, i64 1 166 %v0_1 = load double , double * %from 167 %v0_2 = load double , double * %from_1 168 %v1_1 = fadd double %v0_1, %v0_2 169 %v1_2 = fadd double %v0_1, %p 170 %to_2 = getelementptr double, double * %to, i64 1 171 store double %v1_1, double *%to 172 store double %v1_2, double *%to_2 173br i1 undef, label %lp, label %ext 174 175ext: 176 ret void 177} 178 179; Make sure we don't scramble operands when we reorder them and destroy 180; 'good' source order. 181 182; CHECK-LABEL: good_load_order 183 184; CHECK: %[[V1:[0-9]+]] = load <4 x float>, <4 x float>* 185; CHECK: %[[V2:[0-9]+]] = insertelement <4 x float> undef, float %1, i32 0 186; CHECK: %[[V3:[0-9]+]] = shufflevector <4 x float> %[[V2]], <4 x float> %[[V1]], <4 x i32> <i32 0, i32 4, i32 5, i32 6> 187; CHECK: = fmul <4 x float> %[[V1]], %[[V3]] 188 189@a = common global [32000 x float] zeroinitializer, align 16 190 191define void @good_load_order() { 192entry: 193 br label %for.cond1.preheader 194 195for.cond1.preheader: 196 %0 = load float, float* getelementptr inbounds ([32000 x float], [32000 x float]* @a, i64 0, i64 0), align 16 197 br label %for.body3 198 199for.body3: 200 %1 = phi float [ %0, %for.cond1.preheader ], [ %10, %for.body3 ] 201 %indvars.iv = phi i64 [ 0, %for.cond1.preheader ], [ %indvars.iv.next, %for.body3 ] 202 %2 = add nsw i64 %indvars.iv, 1 203 %arrayidx = getelementptr inbounds [32000 x float], [32000 x float]* @a, i64 0, i64 %2 204 %3 = load float, float* %arrayidx, align 4 205 %arrayidx5 = getelementptr inbounds [32000 x float], [32000 x float]* @a, i64 0, i64 %indvars.iv 206 %mul6 = fmul float %3, %1 207 store float %mul6, float* %arrayidx5, align 4 208 %4 = add nsw i64 %indvars.iv, 2 209 %arrayidx11 = getelementptr inbounds [32000 x float], [32000 x float]* @a, i64 0, i64 %4 210 %5 = load float, float* %arrayidx11, align 4 211 %mul15 = fmul float %5, %3 212 store float %mul15, float* %arrayidx, align 4 213 %6 = add nsw i64 %indvars.iv, 3 214 %arrayidx21 = getelementptr inbounds [32000 x float], [32000 x float]* @a, i64 0, i64 %6 215 %7 = load float, float* %arrayidx21, align 4 216 %mul25 = fmul float %7, %5 217 store float %mul25, float* %arrayidx11, align 4 218 %8 = add nsw i64 %indvars.iv, 4 219 %arrayidx31 = getelementptr inbounds [32000 x float], [32000 x float]* @a, i64 0, i64 %8 220 %9 = load float, float* %arrayidx31, align 4 221 %mul35 = fmul float %9, %7 222 store float %mul35, float* %arrayidx21, align 4 223 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 5 224 %arrayidx41 = getelementptr inbounds [32000 x float], [32000 x float]* @a, i64 0, i64 %indvars.iv.next 225 %10 = load float, float* %arrayidx41, align 4 226 %mul45 = fmul float %10, %9 227 store float %mul45, float* %arrayidx31, align 4 228 %11 = trunc i64 %indvars.iv.next to i32 229 %cmp2 = icmp slt i32 %11, 31995 230 br i1 %cmp2, label %for.body3, label %for.end 231 232for.end: 233 ret void 234} 235 236; Check vectorization of following code for double data type- 237; c[0] = a[0]+b[0]; 238; c[1] = b[1]+a[1]; // swapped b[1] and a[1] 239 240; CHECK-LABEL: load_reorder_double 241; CHECK: load <2 x double>, <2 x double>* 242; CHECK: fadd <2 x double> 243define void @load_reorder_double(double* nocapture %c, double* noalias nocapture readonly %a, double* noalias nocapture readonly %b){ 244 %1 = load double, double* %a 245 %2 = load double, double* %b 246 %3 = fadd double %1, %2 247 store double %3, double* %c 248 %4 = getelementptr inbounds double, double* %b, i64 1 249 %5 = load double, double* %4 250 %6 = getelementptr inbounds double, double* %a, i64 1 251 %7 = load double, double* %6 252 %8 = fadd double %5, %7 253 %9 = getelementptr inbounds double, double* %c, i64 1 254 store double %8, double* %9 255 ret void 256} 257 258; Check vectorization of following code for float data type- 259; c[0] = a[0]+b[0]; 260; c[1] = b[1]+a[1]; // swapped b[1] and a[1] 261; c[2] = a[2]+b[2]; 262; c[3] = a[3]+b[3]; 263 264; CHECK-LABEL: load_reorder_float 265; CHECK: load <4 x float>, <4 x float>* 266; CHECK: fadd <4 x float> 267define void @load_reorder_float(float* nocapture %c, float* noalias nocapture readonly %a, float* noalias nocapture readonly %b){ 268 %1 = load float, float* %a 269 %2 = load float, float* %b 270 %3 = fadd float %1, %2 271 store float %3, float* %c 272 %4 = getelementptr inbounds float, float* %b, i64 1 273 %5 = load float, float* %4 274 %6 = getelementptr inbounds float, float* %a, i64 1 275 %7 = load float, float* %6 276 %8 = fadd float %5, %7 277 %9 = getelementptr inbounds float, float* %c, i64 1 278 store float %8, float* %9 279 %10 = getelementptr inbounds float, float* %a, i64 2 280 %11 = load float, float* %10 281 %12 = getelementptr inbounds float, float* %b, i64 2 282 %13 = load float, float* %12 283 %14 = fadd float %11, %13 284 %15 = getelementptr inbounds float, float* %c, i64 2 285 store float %14, float* %15 286 %16 = getelementptr inbounds float, float* %a, i64 3 287 %17 = load float, float* %16 288 %18 = getelementptr inbounds float, float* %b, i64 3 289 %19 = load float, float* %18 290 %20 = fadd float %17, %19 291 %21 = getelementptr inbounds float, float* %c, i64 3 292 store float %20, float* %21 293 ret void 294} 295 296; Check we properly reorder the below code so that it gets vectorized optimally- 297; a[0] = (b[0]+c[0])+d[0]; 298; a[1] = d[1]+(b[1]+c[1]); 299; a[2] = (b[2]+c[2])+d[2]; 300; a[3] = (b[3]+c[3])+d[3]; 301 302; CHECK-LABEL: opcode_reorder 303; CHECK: load <4 x float>, <4 x float>* 304; CHECK: fadd <4 x float> 305define void @opcode_reorder(float* noalias nocapture %a, float* noalias nocapture readonly %b, 306 float* noalias nocapture readonly %c,float* noalias nocapture readonly %d){ 307 %1 = load float, float* %b 308 %2 = load float, float* %c 309 %3 = fadd float %1, %2 310 %4 = load float, float* %d 311 %5 = fadd float %3, %4 312 store float %5, float* %a 313 %6 = getelementptr inbounds float, float* %d, i64 1 314 %7 = load float, float* %6 315 %8 = getelementptr inbounds float, float* %b, i64 1 316 %9 = load float, float* %8 317 %10 = getelementptr inbounds float, float* %c, i64 1 318 %11 = load float, float* %10 319 %12 = fadd float %9, %11 320 %13 = fadd float %7, %12 321 %14 = getelementptr inbounds float, float* %a, i64 1 322 store float %13, float* %14 323 %15 = getelementptr inbounds float, float* %b, i64 2 324 %16 = load float, float* %15 325 %17 = getelementptr inbounds float, float* %c, i64 2 326 %18 = load float, float* %17 327 %19 = fadd float %16, %18 328 %20 = getelementptr inbounds float, float* %d, i64 2 329 %21 = load float, float* %20 330 %22 = fadd float %19, %21 331 %23 = getelementptr inbounds float, float* %a, i64 2 332 store float %22, float* %23 333 %24 = getelementptr inbounds float, float* %b, i64 3 334 %25 = load float, float* %24 335 %26 = getelementptr inbounds float, float* %c, i64 3 336 %27 = load float, float* %26 337 %28 = fadd float %25, %27 338 %29 = getelementptr inbounds float, float* %d, i64 3 339 %30 = load float, float* %29 340 %31 = fadd float %28, %30 341 %32 = getelementptr inbounds float, float* %a, i64 3 342 store float %31, float* %32 343 ret void 344} 345