1; RUN: opt < %s -basicaa -slp-vectorizer -slp-threshold=-100 -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" 4target triple = "i386-apple-macosx10.9.0" 5 6;int foo(double *A, int k) { 7; double A0; 8; double A1; 9; if (k) { 10; A0 = 3; 11; A1 = 5; 12; } else { 13; A0 = A[10]; 14; A1 = A[11]; 15; } 16; A[0] = A0; 17; A[1] = A1; 18;} 19 20 21;CHECK: i32 @foo 22;CHECK: load <2 x double> 23;CHECK: phi <2 x double> 24;CHECK: store <2 x double> 25;CHECK: ret i32 undef 26define i32 @foo(double* nocapture %A, i32 %k) { 27entry: 28 %tobool = icmp eq i32 %k, 0 29 br i1 %tobool, label %if.else, label %if.end 30 31if.else: ; preds = %entry 32 %arrayidx = getelementptr inbounds double, double* %A, i64 10 33 %0 = load double, double* %arrayidx, align 8 34 %arrayidx1 = getelementptr inbounds double, double* %A, i64 11 35 %1 = load double, double* %arrayidx1, align 8 36 br label %if.end 37 38if.end: ; preds = %entry, %if.else 39 %A0.0 = phi double [ %0, %if.else ], [ 3.000000e+00, %entry ] 40 %A1.0 = phi double [ %1, %if.else ], [ 5.000000e+00, %entry ] 41 store double %A0.0, double* %A, align 8 42 %arrayidx3 = getelementptr inbounds double, double* %A, i64 1 43 store double %A1.0, double* %arrayidx3, align 8 44 ret i32 undef 45} 46 47 48;int foo(double * restrict B, double * restrict A, int n, int m) { 49; double R=A[1]; 50; double G=A[0]; 51; for (int i=0; i < 100; i++) { 52; R += 10; 53; G += 10; 54; R *= 4; 55; G *= 4; 56; R += 4; 57; G += 4; 58; } 59; B[0] = G; 60; B[1] = R; 61; return 0; 62;} 63 64;CHECK: foo2 65;CHECK: load <2 x double> 66;CHECK: phi <2 x double> 67;CHECK: fmul <2 x double> 68;CHECK: store <2 x double> 69;CHECK: ret 70define i32 @foo2(double* noalias nocapture %B, double* noalias nocapture %A, i32 %n, i32 %m) #0 { 71entry: 72 %arrayidx = getelementptr inbounds double, double* %A, i64 1 73 %0 = load double, double* %arrayidx, align 8 74 %1 = load double, double* %A, align 8 75 br label %for.body 76 77for.body: ; preds = %for.body, %entry 78 %i.019 = phi i32 [ 0, %entry ], [ %inc, %for.body ] 79 %G.018 = phi double [ %1, %entry ], [ %add5, %for.body ] 80 %R.017 = phi double [ %0, %entry ], [ %add4, %for.body ] 81 %add = fadd double %R.017, 1.000000e+01 82 %add2 = fadd double %G.018, 1.000000e+01 83 %mul = fmul double %add, 4.000000e+00 84 %mul3 = fmul double %add2, 4.000000e+00 85 %add4 = fadd double %mul, 4.000000e+00 86 %add5 = fadd double %mul3, 4.000000e+00 87 %inc = add nsw i32 %i.019, 1 88 %exitcond = icmp eq i32 %inc, 100 89 br i1 %exitcond, label %for.end, label %for.body 90 91for.end: ; preds = %for.body 92 store double %add5, double* %B, align 8 93 %arrayidx7 = getelementptr inbounds double, double* %B, i64 1 94 store double %add4, double* %arrayidx7, align 8 95 ret i32 0 96} 97 98; float foo3(float *A) { 99; 100; float R = A[0]; 101; float G = A[1]; 102; float B = A[2]; 103; float Y = A[3]; 104; float P = A[4]; 105; for (int i=0; i < 121; i+=3) { 106; R+=A[i+0]*7; 107; G+=A[i+1]*8; 108; B+=A[i+2]*9; 109; Y+=A[i+3]*10; 110; P+=A[i+4]*11; 111; } 112; 113; return R+G+B+Y+P; 114; } 115 116;CHECK: foo3 117;CHECK: phi <4 x float> 118;CHECK: fmul <4 x float> 119;CHECK: fadd <4 x float> 120;CHECK-NOT: phi <5 x float> 121;CHECK-NOT: fmul <5 x float> 122;CHECK-NOT: fadd <5 x float> 123 124define float @foo3(float* nocapture readonly %A) #0 { 125entry: 126 %0 = load float, float* %A, align 4 127 %arrayidx1 = getelementptr inbounds float, float* %A, i64 1 128 %1 = load float, float* %arrayidx1, align 4 129 %arrayidx2 = getelementptr inbounds float, float* %A, i64 2 130 %2 = load float, float* %arrayidx2, align 4 131 %arrayidx3 = getelementptr inbounds float, float* %A, i64 3 132 %3 = load float, float* %arrayidx3, align 4 133 %arrayidx4 = getelementptr inbounds float, float* %A, i64 4 134 %4 = load float, float* %arrayidx4, align 4 135 br label %for.body 136 137for.body: ; preds = %for.body, %entry 138 %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ] 139 %P.056 = phi float [ %4, %entry ], [ %add26, %for.body ] 140 %Y.055 = phi float [ %3, %entry ], [ %add21, %for.body ] 141 %B.054 = phi float [ %2, %entry ], [ %add16, %for.body ] 142 %G.053 = phi float [ %1, %entry ], [ %add11, %for.body ] 143 %R.052 = phi float [ %0, %entry ], [ %add6, %for.body ] 144 %5 = phi float [ %1, %entry ], [ %11, %for.body ] 145 %6 = phi float [ %0, %entry ], [ %9, %for.body ] 146 %mul = fmul float %6, 7.000000e+00 147 %add6 = fadd float %R.052, %mul 148 %mul10 = fmul float %5, 8.000000e+00 149 %add11 = fadd float %G.053, %mul10 150 %7 = add nsw i64 %indvars.iv, 2 151 %arrayidx14 = getelementptr inbounds float, float* %A, i64 %7 152 %8 = load float, float* %arrayidx14, align 4 153 %mul15 = fmul float %8, 9.000000e+00 154 %add16 = fadd float %B.054, %mul15 155 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 3 156 %arrayidx19 = getelementptr inbounds float, float* %A, i64 %indvars.iv.next 157 %9 = load float, float* %arrayidx19, align 4 158 %mul20 = fmul float %9, 1.000000e+01 159 %add21 = fadd float %Y.055, %mul20 160 %10 = add nsw i64 %indvars.iv, 4 161 %arrayidx24 = getelementptr inbounds float, float* %A, i64 %10 162 %11 = load float, float* %arrayidx24, align 4 163 %mul25 = fmul float %11, 1.100000e+01 164 %add26 = fadd float %P.056, %mul25 165 %12 = trunc i64 %indvars.iv.next to i32 166 %cmp = icmp slt i32 %12, 121 167 br i1 %cmp, label %for.body, label %for.end 168 169for.end: ; preds = %for.body 170 %add28 = fadd float %add6, %add11 171 %add29 = fadd float %add28, %add16 172 %add30 = fadd float %add29, %add21 173 %add31 = fadd float %add30, %add26 174 ret float %add31 175} 176 177; Make sure the order of phi nodes of different types does not prevent 178; vectorization of same typed phi nodes. 179; CHECK-LABEL: sort_phi_type 180; CHECK: phi <4 x float> 181; CHECK: fmul <4 x float> 182 183define float @sort_phi_type(float* nocapture readonly %A) { 184entry: 185 br label %for.body 186 187for.body: ; preds = %for.body, %entry 188 %Y = phi float [ 1.000000e+01, %entry ], [ %mul10, %for.body ] 189 %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ] 190 %B = phi float [ 1.000000e+01, %entry ], [ %mul15, %for.body ] 191 %G = phi float [ 1.000000e+01, %entry ], [ %mul20, %for.body ] 192 %R = phi float [ 1.000000e+01, %entry ], [ %mul25, %for.body ] 193 %mul10 = fmul float %Y, 8.000000e+00 194 %mul15 = fmul float %B, 9.000000e+00 195 %mul20 = fmul float %R, 10.000000e+01 196 %mul25 = fmul float %G, 11.100000e+01 197 %indvars.iv.next = add nsw i64 %indvars.iv, 4 198 %cmp = icmp slt i64 %indvars.iv.next, 128 199 br i1 %cmp, label %for.body, label %for.end 200 201for.end: ; preds = %for.body 202 %add28 = fadd float 1.000000e+01, %mul10 203 %add29 = fadd float %mul10, %mul15 204 %add30 = fadd float %add29, %mul20 205 %add31 = fadd float %add30, %mul25 206 ret float %add31 207} 208 209define void @test(x86_fp80* %i1, x86_fp80* %i2, x86_fp80* %o) { 210; CHECK-LABEL: @test( 211; 212; Test that we correctly recognize the discontiguous memory in arrays where the 213; size is less than the alignment, and through various different GEP formations. 214; 215; We disable the vectorization of x86_fp80 for now. 216 217entry: 218 %i1.0 = load x86_fp80, x86_fp80* %i1, align 16 219 %i1.gep1 = getelementptr x86_fp80, x86_fp80* %i1, i64 1 220 %i1.1 = load x86_fp80, x86_fp80* %i1.gep1, align 16 221; CHECK: load x86_fp80, x86_fp80* 222; CHECK: load x86_fp80, x86_fp80* 223; CHECK-NOT: insertelement <2 x x86_fp80> 224; CHECK-NOT: insertelement <2 x x86_fp80> 225 br i1 undef, label %then, label %end 226 227then: 228 %i2.gep0 = getelementptr inbounds x86_fp80, x86_fp80* %i2, i64 0 229 %i2.0 = load x86_fp80, x86_fp80* %i2.gep0, align 16 230 %i2.gep1 = getelementptr inbounds x86_fp80, x86_fp80* %i2, i64 1 231 %i2.1 = load x86_fp80, x86_fp80* %i2.gep1, align 16 232; CHECK: load x86_fp80, x86_fp80* 233; CHECK: load x86_fp80, x86_fp80* 234; CHECK-NOT: insertelement <2 x x86_fp80> 235; CHECK-NOT: insertelement <2 x x86_fp80> 236 br label %end 237 238end: 239 %phi0 = phi x86_fp80 [ %i1.0, %entry ], [ %i2.0, %then ] 240 %phi1 = phi x86_fp80 [ %i1.1, %entry ], [ %i2.1, %then ] 241; CHECK-NOT: phi <2 x x86_fp80> 242; CHECK-NOT: extractelement <2 x x86_fp80> 243; CHECK-NOT: extractelement <2 x x86_fp80> 244 store x86_fp80 %phi0, x86_fp80* %o, align 16 245 %o.gep1 = getelementptr inbounds x86_fp80, x86_fp80* %o, i64 1 246 store x86_fp80 %phi1, x86_fp80* %o.gep1, align 16 247 ret void 248} 249