1 ; RUN: opt < %s -loop-vectorize -force-vector-unroll=1 -force-vector-width=4 -dce -instcombine -S | FileCheck %s 2 3 target 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" 4 target triple = "x86_64-apple-macosx10.8.0" 5 6 ;CHECK: @reduction_sum 7 ;CHECK: phi <4 x i32> 8 ;CHECK: load <4 x i32> 9 ;CHECK: add <4 x i32> 10 ;CHECK: shufflevector <4 x i32> %{{.*}}, <4 x i32> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef> 11 ;CHECK: add <4 x i32> 12 ;CHECK: shufflevector <4 x i32> %{{.*}}, <4 x i32> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef> 13 ;CHECK: add <4 x i32> 14 ;CHECK: extractelement <4 x i32> %{{.*}}, i32 0 15 ;CHECK: ret i32 16 define i32 @reduction_sum(i32 %n, i32* noalias nocapture %A, i32* noalias nocapture %B) nounwind uwtable readonly noinline ssp { 17 %1 = icmp sgt i32 %n, 0 18 br i1 %1, label %.lr.ph, label %._crit_edge 19 20 .lr.ph: ; preds = %0, %.lr.ph 21 %indvars.iv = phi i64 [ %indvars.iv.next, %.lr.ph ], [ 0, %0 ] 22 %sum.02 = phi i32 [ %9, %.lr.ph ], [ 0, %0 ] 23 %2 = getelementptr inbounds i32* %A, i64 %indvars.iv 24 %3 = load i32* %2, align 4 25 %4 = getelementptr inbounds i32* %B, i64 %indvars.iv 26 %5 = load i32* %4, align 4 27 %6 = trunc i64 %indvars.iv to i32 28 %7 = add i32 %sum.02, %6 29 %8 = add i32 %7, %3 30 %9 = add i32 %8, %5 31 %indvars.iv.next = add i64 %indvars.iv, 1 32 %lftr.wideiv = trunc i64 %indvars.iv.next to i32 33 %exitcond = icmp eq i32 %lftr.wideiv, %n 34 br i1 %exitcond, label %._crit_edge, label %.lr.ph 35 36 ._crit_edge: ; preds = %.lr.ph, %0 37 %sum.0.lcssa = phi i32 [ 0, %0 ], [ %9, %.lr.ph ] 38 ret i32 %sum.0.lcssa 39 } 40 41 ;CHECK: @reduction_prod 42 ;CHECK: phi <4 x i32> 43 ;CHECK: load <4 x i32> 44 ;CHECK: mul <4 x i32> 45 ;CHECK: shufflevector <4 x i32> %{{.*}}, <4 x i32> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef> 46 ;CHECK: mul <4 x i32> 47 ;CHECK: shufflevector <4 x i32> %{{.*}}, <4 x i32> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef> 48 ;CHECK: mul <4 x i32> 49 ;CHECK: extractelement <4 x i32> %{{.*}}, i32 0 50 ;CHECK: ret i32 51 define i32 @reduction_prod(i32 %n, i32* noalias nocapture %A, i32* noalias nocapture %B) nounwind uwtable readonly noinline ssp { 52 %1 = icmp sgt i32 %n, 0 53 br i1 %1, label %.lr.ph, label %._crit_edge 54 55 .lr.ph: ; preds = %0, %.lr.ph 56 %indvars.iv = phi i64 [ %indvars.iv.next, %.lr.ph ], [ 0, %0 ] 57 %prod.02 = phi i32 [ %9, %.lr.ph ], [ 1, %0 ] 58 %2 = getelementptr inbounds i32* %A, i64 %indvars.iv 59 %3 = load i32* %2, align 4 60 %4 = getelementptr inbounds i32* %B, i64 %indvars.iv 61 %5 = load i32* %4, align 4 62 %6 = trunc i64 %indvars.iv to i32 63 %7 = mul i32 %prod.02, %6 64 %8 = mul i32 %7, %3 65 %9 = mul i32 %8, %5 66 %indvars.iv.next = add i64 %indvars.iv, 1 67 %lftr.wideiv = trunc i64 %indvars.iv.next to i32 68 %exitcond = icmp eq i32 %lftr.wideiv, %n 69 br i1 %exitcond, label %._crit_edge, label %.lr.ph 70 71 ._crit_edge: ; preds = %.lr.ph, %0 72 %prod.0.lcssa = phi i32 [ 1, %0 ], [ %9, %.lr.ph ] 73 ret i32 %prod.0.lcssa 74 } 75 76 ;CHECK: @reduction_mix 77 ;CHECK: phi <4 x i32> 78 ;CHECK: load <4 x i32> 79 ;CHECK: mul nsw <4 x i32> 80 ;CHECK: shufflevector <4 x i32> %{{.*}}, <4 x i32> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef> 81 ;CHECK: add <4 x i32> 82 ;CHECK: shufflevector <4 x i32> %{{.*}}, <4 x i32> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef> 83 ;CHECK: add <4 x i32> 84 ;CHECK: extractelement <4 x i32> %{{.*}}, i32 0 85 ;CHECK: ret i32 86 define i32 @reduction_mix(i32 %n, i32* noalias nocapture %A, i32* noalias nocapture %B) nounwind uwtable readonly noinline ssp { 87 %1 = icmp sgt i32 %n, 0 88 br i1 %1, label %.lr.ph, label %._crit_edge 89 90 .lr.ph: ; preds = %0, %.lr.ph 91 %indvars.iv = phi i64 [ %indvars.iv.next, %.lr.ph ], [ 0, %0 ] 92 %sum.02 = phi i32 [ %9, %.lr.ph ], [ 0, %0 ] 93 %2 = getelementptr inbounds i32* %A, i64 %indvars.iv 94 %3 = load i32* %2, align 4 95 %4 = getelementptr inbounds i32* %B, i64 %indvars.iv 96 %5 = load i32* %4, align 4 97 %6 = mul nsw i32 %5, %3 98 %7 = trunc i64 %indvars.iv to i32 99 %8 = add i32 %sum.02, %7 100 %9 = add i32 %8, %6 101 %indvars.iv.next = add i64 %indvars.iv, 1 102 %lftr.wideiv = trunc i64 %indvars.iv.next to i32 103 %exitcond = icmp eq i32 %lftr.wideiv, %n 104 br i1 %exitcond, label %._crit_edge, label %.lr.ph 105 106 ._crit_edge: ; preds = %.lr.ph, %0 107 %sum.0.lcssa = phi i32 [ 0, %0 ], [ %9, %.lr.ph ] 108 ret i32 %sum.0.lcssa 109 } 110 111 ;CHECK: @reduction_mul 112 ;CHECK: mul <4 x i32> 113 ;CHECK: shufflevector <4 x i32> %{{.*}}, <4 x i32> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef> 114 ;CHECK: mul <4 x i32> 115 ;CHECK: shufflevector <4 x i32> %{{.*}}, <4 x i32> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef> 116 ;CHECK: mul <4 x i32> 117 ;CHECK: extractelement <4 x i32> %{{.*}}, i32 0 118 ;CHECK: ret i32 119 define i32 @reduction_mul(i32 %n, i32* noalias nocapture %A, i32* noalias nocapture %B) nounwind uwtable readonly noinline ssp { 120 %1 = icmp sgt i32 %n, 0 121 br i1 %1, label %.lr.ph, label %._crit_edge 122 123 .lr.ph: ; preds = %0, %.lr.ph 124 %indvars.iv = phi i64 [ %indvars.iv.next, %.lr.ph ], [ 0, %0 ] 125 %sum.02 = phi i32 [ %9, %.lr.ph ], [ 19, %0 ] 126 %2 = getelementptr inbounds i32* %A, i64 %indvars.iv 127 %3 = load i32* %2, align 4 128 %4 = getelementptr inbounds i32* %B, i64 %indvars.iv 129 %5 = load i32* %4, align 4 130 %6 = trunc i64 %indvars.iv to i32 131 %7 = add i32 %3, %6 132 %8 = add i32 %7, %5 133 %9 = mul i32 %8, %sum.02 134 %indvars.iv.next = add i64 %indvars.iv, 1 135 %lftr.wideiv = trunc i64 %indvars.iv.next to i32 136 %exitcond = icmp eq i32 %lftr.wideiv, %n 137 br i1 %exitcond, label %._crit_edge, label %.lr.ph 138 139 ._crit_edge: ; preds = %.lr.ph, %0 140 %sum.0.lcssa = phi i32 [ 0, %0 ], [ %9, %.lr.ph ] 141 ret i32 %sum.0.lcssa 142 } 143 144 ;CHECK: @start_at_non_zero 145 ;CHECK: phi <4 x i32> 146 ;CHECK: <i32 120, i32 0, i32 0, i32 0> 147 ;CHECK: shufflevector <4 x i32> %{{.*}}, <4 x i32> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef> 148 ;CHECK: add <4 x i32> 149 ;CHECK: shufflevector <4 x i32> %{{.*}}, <4 x i32> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef> 150 ;CHECK: add <4 x i32> 151 ;CHECK: extractelement <4 x i32> %{{.*}}, i32 0 152 ;CHECK: ret i32 153 define i32 @start_at_non_zero(i32* nocapture %in, i32* nocapture %coeff, i32* nocapture %out, i32 %n) nounwind uwtable readonly ssp { 154 entry: 155 %cmp7 = icmp sgt i32 %n, 0 156 br i1 %cmp7, label %for.body, label %for.end 157 158 for.body: ; preds = %entry, %for.body 159 %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ] 160 %sum.09 = phi i32 [ %add, %for.body ], [ 120, %entry ] 161 %arrayidx = getelementptr inbounds i32* %in, i64 %indvars.iv 162 %0 = load i32* %arrayidx, align 4 163 %arrayidx2 = getelementptr inbounds i32* %coeff, i64 %indvars.iv 164 %1 = load i32* %arrayidx2, align 4 165 %mul = mul nsw i32 %1, %0 166 %add = add nsw i32 %mul, %sum.09 167 %indvars.iv.next = add i64 %indvars.iv, 1 168 %lftr.wideiv = trunc i64 %indvars.iv.next to i32 169 %exitcond = icmp eq i32 %lftr.wideiv, %n 170 br i1 %exitcond, label %for.end, label %for.body 171 172 for.end: ; preds = %for.body, %entry 173 %sum.0.lcssa = phi i32 [ 120, %entry ], [ %add, %for.body ] 174 ret i32 %sum.0.lcssa 175 } 176 177 ;CHECK: @reduction_and 178 ;CHECK: and <4 x i32> 179 ;CHECK: <i32 -1, i32 -1, i32 -1, i32 -1> 180 ;CHECK: shufflevector <4 x i32> %{{.*}}, <4 x i32> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef> 181 ;CHECK: and <4 x i32> 182 ;CHECK: shufflevector <4 x i32> %{{.*}}, <4 x i32> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef> 183 ;CHECK: and <4 x i32> 184 ;CHECK: extractelement <4 x i32> %{{.*}}, i32 0 185 ;CHECK: ret i32 186 define i32 @reduction_and(i32 %n, i32* nocapture %A, i32* nocapture %B) nounwind uwtable readonly { 187 entry: 188 %cmp7 = icmp sgt i32 %n, 0 189 br i1 %cmp7, label %for.body, label %for.end 190 191 for.body: ; preds = %entry, %for.body 192 %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ] 193 %result.08 = phi i32 [ %and, %for.body ], [ -1, %entry ] 194 %arrayidx = getelementptr inbounds i32* %A, i64 %indvars.iv 195 %0 = load i32* %arrayidx, align 4 196 %arrayidx2 = getelementptr inbounds i32* %B, i64 %indvars.iv 197 %1 = load i32* %arrayidx2, align 4 198 %add = add nsw i32 %1, %0 199 %and = and i32 %add, %result.08 200 %indvars.iv.next = add i64 %indvars.iv, 1 201 %lftr.wideiv = trunc i64 %indvars.iv.next to i32 202 %exitcond = icmp eq i32 %lftr.wideiv, %n 203 br i1 %exitcond, label %for.end, label %for.body 204 205 for.end: ; preds = %for.body, %entry 206 %result.0.lcssa = phi i32 [ -1, %entry ], [ %and, %for.body ] 207 ret i32 %result.0.lcssa 208 } 209 210 ;CHECK: @reduction_or 211 ;CHECK: or <4 x i32> 212 ;CHECK: shufflevector <4 x i32> %{{.*}}, <4 x i32> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef> 213 ;CHECK: or <4 x i32> 214 ;CHECK: shufflevector <4 x i32> %{{.*}}, <4 x i32> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef> 215 ;CHECK: or <4 x i32> 216 ;CHECK: extractelement <4 x i32> %{{.*}}, i32 0 217 ;CHECK: ret i32 218 define i32 @reduction_or(i32 %n, i32* nocapture %A, i32* nocapture %B) nounwind uwtable readonly { 219 entry: 220 %cmp7 = icmp sgt i32 %n, 0 221 br i1 %cmp7, label %for.body, label %for.end 222 223 for.body: ; preds = %entry, %for.body 224 %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ] 225 %result.08 = phi i32 [ %or, %for.body ], [ 0, %entry ] 226 %arrayidx = getelementptr inbounds i32* %A, i64 %indvars.iv 227 %0 = load i32* %arrayidx, align 4 228 %arrayidx2 = getelementptr inbounds i32* %B, i64 %indvars.iv 229 %1 = load i32* %arrayidx2, align 4 230 %add = add nsw i32 %1, %0 231 %or = or i32 %add, %result.08 232 %indvars.iv.next = add i64 %indvars.iv, 1 233 %lftr.wideiv = trunc i64 %indvars.iv.next to i32 234 %exitcond = icmp eq i32 %lftr.wideiv, %n 235 br i1 %exitcond, label %for.end, label %for.body 236 237 for.end: ; preds = %for.body, %entry 238 %result.0.lcssa = phi i32 [ 0, %entry ], [ %or, %for.body ] 239 ret i32 %result.0.lcssa 240 } 241 242 ;CHECK: @reduction_xor 243 ;CHECK: xor <4 x i32> 244 ;CHECK: shufflevector <4 x i32> %{{.*}}, <4 x i32> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef> 245 ;CHECK: xor <4 x i32> 246 ;CHECK: shufflevector <4 x i32> %{{.*}}, <4 x i32> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef> 247 ;CHECK: xor <4 x i32> 248 ;CHECK: extractelement <4 x i32> %{{.*}}, i32 0 249 ;CHECK: ret i32 250 define i32 @reduction_xor(i32 %n, i32* nocapture %A, i32* nocapture %B) nounwind uwtable readonly { 251 entry: 252 %cmp7 = icmp sgt i32 %n, 0 253 br i1 %cmp7, label %for.body, label %for.end 254 255 for.body: ; preds = %entry, %for.body 256 %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ] 257 %result.08 = phi i32 [ %xor, %for.body ], [ 0, %entry ] 258 %arrayidx = getelementptr inbounds i32* %A, i64 %indvars.iv 259 %0 = load i32* %arrayidx, align 4 260 %arrayidx2 = getelementptr inbounds i32* %B, i64 %indvars.iv 261 %1 = load i32* %arrayidx2, align 4 262 %add = add nsw i32 %1, %0 263 %xor = xor i32 %add, %result.08 264 %indvars.iv.next = add i64 %indvars.iv, 1 265 %lftr.wideiv = trunc i64 %indvars.iv.next to i32 266 %exitcond = icmp eq i32 %lftr.wideiv, %n 267 br i1 %exitcond, label %for.end, label %for.body 268 269 for.end: ; preds = %for.body, %entry 270 %result.0.lcssa = phi i32 [ 0, %entry ], [ %xor, %for.body ] 271 ret i32 %result.0.lcssa 272 } 273 274 ; In this code the subtracted variable is on the RHS and this is not an induction variable. 275 ;CHECK: @reduction_sub_rhs 276 ;CHECK-NOT: phi <4 x i32> 277 ;CHECK-NOT: sub nsw <4 x i32> 278 ;CHECK: ret i32 279 define i32 @reduction_sub_rhs(i32 %n, i32* noalias nocapture %A) nounwind uwtable readonly { 280 entry: 281 %cmp4 = icmp sgt i32 %n, 0 282 br i1 %cmp4, label %for.body, label %for.end 283 284 for.body: ; preds = %entry, %for.body 285 %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ] 286 %x.05 = phi i32 [ %sub, %for.body ], [ 0, %entry ] 287 %arrayidx = getelementptr inbounds i32* %A, i64 %indvars.iv 288 %0 = load i32* %arrayidx, align 4 289 %sub = sub nsw i32 %0, %x.05 290 %indvars.iv.next = add i64 %indvars.iv, 1 291 %lftr.wideiv = trunc i64 %indvars.iv.next to i32 292 %exitcond = icmp eq i32 %lftr.wideiv, %n 293 br i1 %exitcond, label %for.end, label %for.body 294 295 for.end: ; preds = %for.body, %entry 296 %x.0.lcssa = phi i32 [ 0, %entry ], [ %sub, %for.body ] 297 ret i32 %x.0.lcssa 298 } 299 300 301 ; In this test the reduction variable is on the LHS and we can vectorize it. 302 ;CHECK: @reduction_sub_lhs 303 ;CHECK: phi <4 x i32> 304 ;CHECK: sub nsw <4 x i32> 305 ;CHECK: ret i32 306 define i32 @reduction_sub_lhs(i32 %n, i32* noalias nocapture %A) nounwind uwtable readonly { 307 entry: 308 %cmp4 = icmp sgt i32 %n, 0 309 br i1 %cmp4, label %for.body, label %for.end 310 311 for.body: ; preds = %entry, %for.body 312 %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ] 313 %x.05 = phi i32 [ %sub, %for.body ], [ 0, %entry ] 314 %arrayidx = getelementptr inbounds i32* %A, i64 %indvars.iv 315 %0 = load i32* %arrayidx, align 4 316 %sub = sub nsw i32 %x.05, %0 317 %indvars.iv.next = add i64 %indvars.iv, 1 318 %lftr.wideiv = trunc i64 %indvars.iv.next to i32 319 %exitcond = icmp eq i32 %lftr.wideiv, %n 320 br i1 %exitcond, label %for.end, label %for.body 321 322 for.end: ; preds = %for.body, %entry 323 %x.0.lcssa = phi i32 [ 0, %entry ], [ %sub, %for.body ] 324 ret i32 %x.0.lcssa 325 } 326
