1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py 2; RUN: llc -mcpu=pwr9 -verify-machineinstrs -mtriple=powerpc64le-unknown-linux-gnu < %s | FileCheck %s 3 4; void foo(ptr data, float d) { 5; long i; 6; for (i = 0; i < 8000; i++) 7; data[i] = d; 8; } 9; 10; This loop will be unrolled by 96 and vectorized on power9. 11; icmp for loop iteration index and loop trip count(384) has LSRUse for 'reg({0,+,384})'. 12; Make sure above icmp does not impact LSR choose best formulae sets based on 'reg({(192 + %0),+,384})' 13 14define void @foo(ptr nocapture %data, float %d) { 15; CHECK-LABEL: foo: 16; CHECK: # %bb.0: # %entry 17; CHECK-NEXT: xscvdpspn 0, 1 18; CHECK-NEXT: li 5, 83 19; CHECK-NEXT: addi 4, 3, 192 20; CHECK-NEXT: mtctr 5 21; CHECK-NEXT: xxspltw 0, 0, 0 22; CHECK-NEXT: .p2align 4 23; CHECK-NEXT: .LBB0_1: # %vector.body 24; CHECK-NEXT: # 25; CHECK-NEXT: stxv 0, -192(4) 26; CHECK-NEXT: stxv 0, -176(4) 27; CHECK-NEXT: stxv 0, -160(4) 28; CHECK-NEXT: stxv 0, -144(4) 29; CHECK-NEXT: stxv 0, -128(4) 30; CHECK-NEXT: stxv 0, -112(4) 31; CHECK-NEXT: stxv 0, -96(4) 32; CHECK-NEXT: stxv 0, -80(4) 33; CHECK-NEXT: stxv 0, -64(4) 34; CHECK-NEXT: stxv 0, -48(4) 35; CHECK-NEXT: stxv 0, -32(4) 36; CHECK-NEXT: stxv 0, -16(4) 37; CHECK-NEXT: stxv 0, 0(4) 38; CHECK-NEXT: stxv 0, 16(4) 39; CHECK-NEXT: stxv 0, 32(4) 40; CHECK-NEXT: stxv 0, 48(4) 41; CHECK-NEXT: stxv 0, 64(4) 42; CHECK-NEXT: stxv 0, 80(4) 43; CHECK-NEXT: stxv 0, 96(4) 44; CHECK-NEXT: stxv 0, 112(4) 45; CHECK-NEXT: stxv 0, 128(4) 46; CHECK-NEXT: stxv 0, 144(4) 47; CHECK-NEXT: stxv 0, 160(4) 48; CHECK-NEXT: stxv 0, 176(4) 49; CHECK-NEXT: addi 4, 4, 384 50; CHECK-NEXT: bdnz .LBB0_1 51; CHECK-NEXT: # %bb.2: # %for.body 52; CHECK-NEXT: stfs 1, 31872(3) 53; CHECK-NEXT: stfs 1, 31876(3) 54; CHECK-NEXT: stfs 1, 31880(3) 55; CHECK-NEXT: stfs 1, 31884(3) 56; CHECK-NEXT: stfs 1, 31888(3) 57; CHECK-NEXT: stfs 1, 31892(3) 58; CHECK-NEXT: stfs 1, 31896(3) 59; CHECK-NEXT: stfs 1, 31900(3) 60; CHECK-NEXT: stfs 1, 31904(3) 61; CHECK-NEXT: stfs 1, 31908(3) 62; CHECK-NEXT: stfs 1, 31912(3) 63; CHECK-NEXT: stfs 1, 31916(3) 64; CHECK-NEXT: stfs 1, 31920(3) 65; CHECK-NEXT: stfs 1, 31924(3) 66; CHECK-NEXT: stfs 1, 31928(3) 67; CHECK-NEXT: stfs 1, 31932(3) 68; CHECK-NEXT: stfs 1, 31936(3) 69; CHECK-NEXT: stfs 1, 31940(3) 70; CHECK-NEXT: stfs 1, 31944(3) 71; CHECK-NEXT: stfs 1, 31948(3) 72; CHECK-NEXT: stfs 1, 31952(3) 73; CHECK-NEXT: stfs 1, 31956(3) 74; CHECK-NEXT: stfs 1, 31960(3) 75; CHECK-NEXT: stfs 1, 31964(3) 76; CHECK-NEXT: stfs 1, 31968(3) 77; CHECK-NEXT: stfs 1, 31972(3) 78; CHECK-NEXT: stfs 1, 31976(3) 79; CHECK-NEXT: stfs 1, 31980(3) 80; CHECK-NEXT: stfs 1, 31984(3) 81; CHECK-NEXT: stfs 1, 31988(3) 82; CHECK-NEXT: stfs 1, 31992(3) 83; CHECK-NEXT: stfs 1, 31996(3) 84; CHECK-NEXT: blr 85 86entry: 87 %broadcast.splatinsert16 = insertelement <4 x float> undef, float %d, i32 0 88 %broadcast.splat17 = shufflevector <4 x float> %broadcast.splatinsert16, <4 x float> undef, <4 x i32> zeroinitializer 89 %broadcast.splatinsert18 = insertelement <4 x float> undef, float %d, i32 0 90 %broadcast.splat19 = shufflevector <4 x float> %broadcast.splatinsert18, <4 x float> undef, <4 x i32> zeroinitializer 91 %broadcast.splatinsert20 = insertelement <4 x float> undef, float %d, i32 0 92 %broadcast.splat21 = shufflevector <4 x float> %broadcast.splatinsert20, <4 x float> undef, <4 x i32> zeroinitializer 93 %broadcast.splatinsert22 = insertelement <4 x float> undef, float %d, i32 0 94 %broadcast.splat23 = shufflevector <4 x float> %broadcast.splatinsert22, <4 x float> undef, <4 x i32> zeroinitializer 95 %broadcast.splatinsert24 = insertelement <4 x float> undef, float %d, i32 0 96 %broadcast.splat25 = shufflevector <4 x float> %broadcast.splatinsert24, <4 x float> undef, <4 x i32> zeroinitializer 97 %broadcast.splatinsert26 = insertelement <4 x float> undef, float %d, i32 0 98 %broadcast.splat27 = shufflevector <4 x float> %broadcast.splatinsert26, <4 x float> undef, <4 x i32> zeroinitializer 99 %broadcast.splatinsert28 = insertelement <4 x float> undef, float %d, i32 0 100 %broadcast.splat29 = shufflevector <4 x float> %broadcast.splatinsert28, <4 x float> undef, <4 x i32> zeroinitializer 101 %broadcast.splatinsert30 = insertelement <4 x float> undef, float %d, i32 0 102 %broadcast.splat31 = shufflevector <4 x float> %broadcast.splatinsert30, <4 x float> undef, <4 x i32> zeroinitializer 103 %broadcast.splatinsert32 = insertelement <4 x float> undef, float %d, i32 0 104 %broadcast.splat33 = shufflevector <4 x float> %broadcast.splatinsert32, <4 x float> undef, <4 x i32> zeroinitializer 105 %broadcast.splatinsert34 = insertelement <4 x float> undef, float %d, i32 0 106 %broadcast.splat35 = shufflevector <4 x float> %broadcast.splatinsert34, <4 x float> undef, <4 x i32> zeroinitializer 107 %broadcast.splatinsert36 = insertelement <4 x float> undef, float %d, i32 0 108 %broadcast.splat37 = shufflevector <4 x float> %broadcast.splatinsert36, <4 x float> undef, <4 x i32> zeroinitializer 109 %broadcast.splatinsert38 = insertelement <4 x float> undef, float %d, i32 0 110 %broadcast.splat39 = shufflevector <4 x float> %broadcast.splatinsert38, <4 x float> undef, <4 x i32> zeroinitializer 111 br label %vector.body 112 113vector.body: ; preds = %vector.body, %entry 114 %index = phi i64 [ 0, %entry ], [ %index.next.1, %vector.body ] 115 %0 = getelementptr inbounds float, ptr %data, i64 %index 116 store <4 x float> %broadcast.splat17, ptr %0, align 4 117 %1 = getelementptr inbounds float, ptr %0, i64 4 118 store <4 x float> %broadcast.splat19, ptr %1, align 4 119 %2 = getelementptr inbounds float, ptr %0, i64 8 120 store <4 x float> %broadcast.splat21, ptr %2, align 4 121 %3 = getelementptr inbounds float, ptr %0, i64 12 122 store <4 x float> %broadcast.splat23, ptr %3, align 4 123 %4 = getelementptr inbounds float, ptr %0, i64 16 124 store <4 x float> %broadcast.splat25, ptr %4, align 4 125 %5 = getelementptr inbounds float, ptr %0, i64 20 126 store <4 x float> %broadcast.splat27, ptr %5, align 4 127 %6 = getelementptr inbounds float, ptr %0, i64 24 128 store <4 x float> %broadcast.splat29, ptr %6, align 4 129 %7 = getelementptr inbounds float, ptr %0, i64 28 130 store <4 x float> %broadcast.splat31, ptr %7, align 4 131 %8 = getelementptr inbounds float, ptr %0, i64 32 132 store <4 x float> %broadcast.splat33, ptr %8, align 4 133 %9 = getelementptr inbounds float, ptr %0, i64 36 134 store <4 x float> %broadcast.splat35, ptr %9, align 4 135 %10 = getelementptr inbounds float, ptr %0, i64 40 136 store <4 x float> %broadcast.splat37, ptr %10, align 4 137 %11 = getelementptr inbounds float, ptr %0, i64 44 138 store <4 x float> %broadcast.splat39, ptr %11, align 4 139 %index.next = add nuw nsw i64 %index, 48 140 %12 = getelementptr inbounds float, ptr %data, i64 %index.next 141 store <4 x float> %broadcast.splat17, ptr %12, align 4 142 %13 = getelementptr inbounds float, ptr %12, i64 4 143 store <4 x float> %broadcast.splat19, ptr %13, align 4 144 %14 = getelementptr inbounds float, ptr %12, i64 8 145 store <4 x float> %broadcast.splat21, ptr %14, align 4 146 %15 = getelementptr inbounds float, ptr %12, i64 12 147 store <4 x float> %broadcast.splat23, ptr %15, align 4 148 %16 = getelementptr inbounds float, ptr %12, i64 16 149 store <4 x float> %broadcast.splat25, ptr %16, align 4 150 %17 = getelementptr inbounds float, ptr %12, i64 20 151 store <4 x float> %broadcast.splat27, ptr %17, align 4 152 %18 = getelementptr inbounds float, ptr %12, i64 24 153 store <4 x float> %broadcast.splat29, ptr %18, align 4 154 %19 = getelementptr inbounds float, ptr %12, i64 28 155 store <4 x float> %broadcast.splat31, ptr %19, align 4 156 %20 = getelementptr inbounds float, ptr %12, i64 32 157 store <4 x float> %broadcast.splat33, ptr %20, align 4 158 %21 = getelementptr inbounds float, ptr %12, i64 36 159 store <4 x float> %broadcast.splat35, ptr %21, align 4 160 %22 = getelementptr inbounds float, ptr %12, i64 40 161 store <4 x float> %broadcast.splat37, ptr %22, align 4 162 %23 = getelementptr inbounds float, ptr %12, i64 44 163 store <4 x float> %broadcast.splat39, ptr %23, align 4 164 %index.next.1 = add nuw nsw i64 %index, 96 165 %24 = icmp eq i64 %index.next.1, 7968 166 br i1 %24, label %for.body, label %vector.body 167 168for.body: ; preds = %vector.body 169 %arrayidx = getelementptr inbounds float, ptr %data, i64 7968 170 store float %d, ptr %arrayidx, align 4 171 %arrayidx.1 = getelementptr inbounds float, ptr %data, i64 7969 172 store float %d, ptr %arrayidx.1, align 4 173 %arrayidx.2 = getelementptr inbounds float, ptr %data, i64 7970 174 store float %d, ptr %arrayidx.2, align 4 175 %arrayidx.3 = getelementptr inbounds float, ptr %data, i64 7971 176 store float %d, ptr %arrayidx.3, align 4 177 %arrayidx.4 = getelementptr inbounds float, ptr %data, i64 7972 178 store float %d, ptr %arrayidx.4, align 4 179 %arrayidx.5 = getelementptr inbounds float, ptr %data, i64 7973 180 store float %d, ptr %arrayidx.5, align 4 181 %arrayidx.6 = getelementptr inbounds float, ptr %data, i64 7974 182 store float %d, ptr %arrayidx.6, align 4 183 %arrayidx.7 = getelementptr inbounds float, ptr %data, i64 7975 184 store float %d, ptr %arrayidx.7, align 4 185 %arrayidx.8 = getelementptr inbounds float, ptr %data, i64 7976 186 store float %d, ptr %arrayidx.8, align 4 187 %arrayidx.9 = getelementptr inbounds float, ptr %data, i64 7977 188 store float %d, ptr %arrayidx.9, align 4 189 %arrayidx.10 = getelementptr inbounds float, ptr %data, i64 7978 190 store float %d, ptr %arrayidx.10, align 4 191 %arrayidx.11 = getelementptr inbounds float, ptr %data, i64 7979 192 store float %d, ptr %arrayidx.11, align 4 193 %arrayidx.12 = getelementptr inbounds float, ptr %data, i64 7980 194 store float %d, ptr %arrayidx.12, align 4 195 %arrayidx.13 = getelementptr inbounds float, ptr %data, i64 7981 196 store float %d, ptr %arrayidx.13, align 4 197 %arrayidx.14 = getelementptr inbounds float, ptr %data, i64 7982 198 store float %d, ptr %arrayidx.14, align 4 199 %arrayidx.15 = getelementptr inbounds float, ptr %data, i64 7983 200 store float %d, ptr %arrayidx.15, align 4 201 %arrayidx.16 = getelementptr inbounds float, ptr %data, i64 7984 202 store float %d, ptr %arrayidx.16, align 4 203 %arrayidx.17 = getelementptr inbounds float, ptr %data, i64 7985 204 store float %d, ptr %arrayidx.17, align 4 205 %arrayidx.18 = getelementptr inbounds float, ptr %data, i64 7986 206 store float %d, ptr %arrayidx.18, align 4 207 %arrayidx.19 = getelementptr inbounds float, ptr %data, i64 7987 208 store float %d, ptr %arrayidx.19, align 4 209 %arrayidx.20 = getelementptr inbounds float, ptr %data, i64 7988 210 store float %d, ptr %arrayidx.20, align 4 211 %arrayidx.21 = getelementptr inbounds float, ptr %data, i64 7989 212 store float %d, ptr %arrayidx.21, align 4 213 %arrayidx.22 = getelementptr inbounds float, ptr %data, i64 7990 214 store float %d, ptr %arrayidx.22, align 4 215 %arrayidx.23 = getelementptr inbounds float, ptr %data, i64 7991 216 store float %d, ptr %arrayidx.23, align 4 217 %arrayidx.24 = getelementptr inbounds float, ptr %data, i64 7992 218 store float %d, ptr %arrayidx.24, align 4 219 %arrayidx.25 = getelementptr inbounds float, ptr %data, i64 7993 220 store float %d, ptr %arrayidx.25, align 4 221 %arrayidx.26 = getelementptr inbounds float, ptr %data, i64 7994 222 store float %d, ptr %arrayidx.26, align 4 223 %arrayidx.27 = getelementptr inbounds float, ptr %data, i64 7995 224 store float %d, ptr %arrayidx.27, align 4 225 %arrayidx.28 = getelementptr inbounds float, ptr %data, i64 7996 226 store float %d, ptr %arrayidx.28, align 4 227 %arrayidx.29 = getelementptr inbounds float, ptr %data, i64 7997 228 store float %d, ptr %arrayidx.29, align 4 229 %arrayidx.30 = getelementptr inbounds float, ptr %data, i64 7998 230 store float %d, ptr %arrayidx.30, align 4 231 %arrayidx.31 = getelementptr inbounds float, ptr %data, i64 7999 232 store float %d, ptr %arrayidx.31, align 4 233 ret void 234} 235