1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py 2; RUN: opt < %s -instcombine -S | FileCheck %s 3 4; This turns into a&1 != 0 5define <2 x i1> @test1(<2 x i64> %a) { 6; CHECK-LABEL: @test1( 7; CHECK-NEXT: [[TMP1:%.*]] = and <2 x i64> %a, <i64 1, i64 1> 8; CHECK-NEXT: [[T:%.*]] = icmp ne <2 x i64> [[TMP1]], zeroinitializer 9; CHECK-NEXT: ret <2 x i1> [[T]] 10; 11 %t = trunc <2 x i64> %a to <2 x i1> 12 ret <2 x i1> %t 13} 14 15; The ashr turns into an lshr. 16define <2 x i64> @test2(<2 x i64> %a) { 17; CHECK-LABEL: @test2( 18; CHECK-NEXT: [[B:%.*]] = lshr <2 x i64> [[A:%.*]], <i64 1, i64 1> 19; CHECK-NEXT: [[TMP1:%.*]] = and <2 x i64> [[B]], <i64 32767, i64 32767> 20; CHECK-NEXT: ret <2 x i64> [[TMP1]] 21; 22 %b = and <2 x i64> %a, <i64 65535, i64 65535> 23 %t = ashr <2 x i64> %b, <i64 1, i64 1> 24 ret <2 x i64> %t 25} 26 27define <2 x i64> @test3(<4 x float> %a, <4 x float> %b) { 28; CHECK-LABEL: @test3( 29; CHECK-NEXT: [[TMP1:%.*]] = fcmp ord <4 x float> %a, %b 30; CHECK-NEXT: [[AND:%.*]] = sext <4 x i1> [[TMP1]] to <4 x i32> 31; CHECK-NEXT: [[CONV:%.*]] = bitcast <4 x i32> [[AND]] to <2 x i64> 32; CHECK-NEXT: ret <2 x i64> [[CONV]] 33; 34 %cmp = fcmp ord <4 x float> %a, zeroinitializer 35 %sext = sext <4 x i1> %cmp to <4 x i32> 36 %cmp4 = fcmp ord <4 x float> %b, zeroinitializer 37 %sext5 = sext <4 x i1> %cmp4 to <4 x i32> 38 %and = and <4 x i32> %sext, %sext5 39 %conv = bitcast <4 x i32> %and to <2 x i64> 40 ret <2 x i64> %conv 41} 42 43define <2 x i64> @test4(<4 x float> %a, <4 x float> %b) { 44; CHECK-LABEL: @test4( 45; CHECK-NEXT: [[TMP1:%.*]] = fcmp uno <4 x float> %a, %b 46; CHECK-NEXT: [[OR:%.*]] = sext <4 x i1> [[TMP1]] to <4 x i32> 47; CHECK-NEXT: [[CONV:%.*]] = bitcast <4 x i32> [[OR]] to <2 x i64> 48; CHECK-NEXT: ret <2 x i64> [[CONV]] 49; 50 %cmp = fcmp uno <4 x float> %a, zeroinitializer 51 %sext = sext <4 x i1> %cmp to <4 x i32> 52 %cmp4 = fcmp uno <4 x float> %b, zeroinitializer 53 %sext5 = sext <4 x i1> %cmp4 to <4 x i32> 54 %or = or <4 x i32> %sext, %sext5 55 %conv = bitcast <4 x i32> %or to <2 x i64> 56 ret <2 x i64> %conv 57} 58 59; rdar://7434900 60define <2 x i64> @test5(<4 x float> %a, <4 x float> %b) { 61; CHECK-LABEL: @test5( 62; CHECK-NEXT: [[CMP:%.*]] = fcmp ult <4 x float> %a, zeroinitializer 63; CHECK-NEXT: [[CMP4:%.*]] = fcmp ult <4 x float> %b, zeroinitializer 64; CHECK-NEXT: [[NARROW:%.*]] = and <4 x i1> [[CMP]], [[CMP4]] 65; CHECK-NEXT: [[AND:%.*]] = sext <4 x i1> [[NARROW]] to <4 x i32> 66; CHECK-NEXT: [[CONV:%.*]] = bitcast <4 x i32> [[AND]] to <2 x i64> 67; CHECK-NEXT: ret <2 x i64> [[CONV]] 68; 69 %cmp = fcmp ult <4 x float> %a, zeroinitializer 70 %sext = sext <4 x i1> %cmp to <4 x i32> 71 %cmp4 = fcmp ult <4 x float> %b, zeroinitializer 72 %sext5 = sext <4 x i1> %cmp4 to <4 x i32> 73 %and = and <4 x i32> %sext, %sext5 74 %conv = bitcast <4 x i32> %and to <2 x i64> 75 ret <2 x i64> %conv 76} 77 78define <2 x i64> @test6(<4 x float> %a, <4 x float> %b) { 79; CHECK-LABEL: @test6( 80; CHECK-NEXT: [[CMP:%.*]] = fcmp ult <4 x float> [[A:%.*]], zeroinitializer 81; CHECK-NEXT: [[CMP4:%.*]] = fcmp ult <4 x float> [[B:%.*]], zeroinitializer 82; CHECK-NEXT: [[NARROW:%.*]] = or <4 x i1> [[CMP]], [[CMP4]] 83; CHECK-NEXT: [[AND:%.*]] = sext <4 x i1> [[NARROW]] to <4 x i32> 84; CHECK-NEXT: [[CONV:%.*]] = bitcast <4 x i32> [[AND]] to <2 x i64> 85; CHECK-NEXT: ret <2 x i64> [[CONV]] 86; 87 %cmp = fcmp ult <4 x float> %a, zeroinitializer 88 %sext = sext <4 x i1> %cmp to <4 x i32> 89 %cmp4 = fcmp ult <4 x float> %b, zeroinitializer 90 %sext5 = sext <4 x i1> %cmp4 to <4 x i32> 91 %and = or <4 x i32> %sext, %sext5 92 %conv = bitcast <4 x i32> %and to <2 x i64> 93 ret <2 x i64> %conv 94} 95 96define <2 x i64> @test7(<4 x float> %a, <4 x float> %b) { 97; CHECK-LABEL: @test7( 98; CHECK-NEXT: [[CMP:%.*]] = fcmp ult <4 x float> [[A:%.*]], zeroinitializer 99; CHECK-NEXT: [[CMP4:%.*]] = fcmp ult <4 x float> [[B:%.*]], zeroinitializer 100; CHECK-NEXT: [[AND1:%.*]] = xor <4 x i1> [[CMP]], [[CMP4]] 101; CHECK-NEXT: [[AND:%.*]] = sext <4 x i1> [[AND1]] to <4 x i32> 102; CHECK-NEXT: [[CONV:%.*]] = bitcast <4 x i32> [[AND]] to <2 x i64> 103; CHECK-NEXT: ret <2 x i64> [[CONV]] 104; 105 %cmp = fcmp ult <4 x float> %a, zeroinitializer 106 %sext = sext <4 x i1> %cmp to <4 x i32> 107 %cmp4 = fcmp ult <4 x float> %b, zeroinitializer 108 %sext5 = sext <4 x i1> %cmp4 to <4 x i32> 109 %and = xor <4 x i32> %sext, %sext5 110 %conv = bitcast <4 x i32> %and to <2 x i64> 111 ret <2 x i64> %conv 112} 113 114define void @convert(<2 x i32>* %dst.addr, <2 x i64> %src) { 115; CHECK-LABEL: @convert( 116; CHECK-NEXT: [[VAL:%.*]] = trunc <2 x i64> %src to <2 x i32> 117; CHECK-NEXT: [[ADD:%.*]] = add <2 x i32> [[VAL]], <i32 1, i32 1> 118; CHECK-NEXT: store <2 x i32> [[ADD]], <2 x i32>* %dst.addr, align 8 119; CHECK-NEXT: ret void 120; 121 %val = trunc <2 x i64> %src to <2 x i32> 122 %add = add <2 x i32> %val, <i32 1, i32 1> 123 store <2 x i32> %add, <2 x i32>* %dst.addr 124 ret void 125} 126 127define <2 x i65> @foo(<2 x i64> %t) { 128; CHECK-LABEL: @foo( 129; CHECK-NEXT: [[TMP1:%.*]] = and <2 x i64> %t, <i64 4294967295, i64 4294967295> 130; CHECK-NEXT: [[B:%.*]] = zext <2 x i64> [[TMP1]] to <2 x i65> 131; CHECK-NEXT: ret <2 x i65> [[B]] 132; 133 %a = trunc <2 x i64> %t to <2 x i32> 134 %b = zext <2 x i32> %a to <2 x i65> 135 ret <2 x i65> %b 136} 137 138define <2 x i64> @bar(<2 x i65> %t) { 139; CHECK-LABEL: @bar( 140; CHECK-NEXT: [[A:%.*]] = trunc <2 x i65> %t to <2 x i64> 141; CHECK-NEXT: [[B:%.*]] = and <2 x i64> [[A]], <i64 4294967295, i64 4294967295> 142; CHECK-NEXT: ret <2 x i64> [[B]] 143; 144 %a = trunc <2 x i65> %t to <2 x i32> 145 %b = zext <2 x i32> %a to <2 x i64> 146 ret <2 x i64> %b 147} 148 149define <2 x i64> @bars(<2 x i65> %t) { 150; CHECK-LABEL: @bars( 151; CHECK-NEXT: [[A:%.*]] = trunc <2 x i65> %t to <2 x i64> 152; CHECK-NEXT: [[SEXT:%.*]] = shl <2 x i64> [[A]], <i64 32, i64 32> 153; CHECK-NEXT: [[B:%.*]] = ashr exact <2 x i64> [[SEXT]], <i64 32, i64 32> 154; CHECK-NEXT: ret <2 x i64> [[B]] 155; 156 %a = trunc <2 x i65> %t to <2 x i32> 157 %b = sext <2 x i32> %a to <2 x i64> 158 ret <2 x i64> %b 159} 160 161define <2 x i64> @quxs(<2 x i64> %t) { 162; CHECK-LABEL: @quxs( 163; CHECK-NEXT: [[SEXT:%.*]] = shl <2 x i64> %t, <i64 32, i64 32> 164; CHECK-NEXT: [[B:%.*]] = ashr exact <2 x i64> [[SEXT]], <i64 32, i64 32> 165; CHECK-NEXT: ret <2 x i64> [[B]] 166; 167 %a = trunc <2 x i64> %t to <2 x i32> 168 %b = sext <2 x i32> %a to <2 x i64> 169 ret <2 x i64> %b 170} 171 172define <2 x i64> @quxt(<2 x i64> %t) { 173; CHECK-LABEL: @quxt( 174; CHECK-NEXT: [[A:%.*]] = shl <2 x i64> %t, <i64 32, i64 32> 175; CHECK-NEXT: [[B:%.*]] = ashr exact <2 x i64> [[A]], <i64 32, i64 32> 176; CHECK-NEXT: ret <2 x i64> [[B]] 177; 178 %a = shl <2 x i64> %t, <i64 32, i64 32> 179 %b = ashr <2 x i64> %a, <i64 32, i64 32> 180 ret <2 x i64> %b 181} 182 183define <2 x double> @fa(<2 x double> %t) { 184; CHECK-LABEL: @fa( 185; CHECK-NEXT: [[A:%.*]] = fptrunc <2 x double> %t to <2 x float> 186; CHECK-NEXT: [[B:%.*]] = fpext <2 x float> [[A]] to <2 x double> 187; CHECK-NEXT: ret <2 x double> [[B]] 188; 189 %a = fptrunc <2 x double> %t to <2 x float> 190 %b = fpext <2 x float> %a to <2 x double> 191 ret <2 x double> %b 192} 193 194define <2 x double> @fb(<2 x double> %t) { 195; CHECK-LABEL: @fb( 196; CHECK-NEXT: [[A:%.*]] = fptoui <2 x double> %t to <2 x i64> 197; CHECK-NEXT: [[B:%.*]] = uitofp <2 x i64> [[A]] to <2 x double> 198; CHECK-NEXT: ret <2 x double> [[B]] 199; 200 %a = fptoui <2 x double> %t to <2 x i64> 201 %b = uitofp <2 x i64> %a to <2 x double> 202 ret <2 x double> %b 203} 204 205define <2 x double> @fc(<2 x double> %t) { 206; CHECK-LABEL: @fc( 207; CHECK-NEXT: [[A:%.*]] = fptosi <2 x double> %t to <2 x i64> 208; CHECK-NEXT: [[B:%.*]] = sitofp <2 x i64> [[A]] to <2 x double> 209; CHECK-NEXT: ret <2 x double> [[B]] 210; 211 %a = fptosi <2 x double> %t to <2 x i64> 212 %b = sitofp <2 x i64> %a to <2 x double> 213 ret <2 x double> %b 214} 215 216; PR9228 217define <4 x float> @f(i32 %a) { 218; CHECK-LABEL: @f( 219; CHECK-NEXT: ret <4 x float> undef 220; 221 %dim = insertelement <4 x i32> undef, i32 %a, i32 0 222 %dim30 = insertelement <4 x i32> %dim, i32 %a, i32 1 223 %dim31 = insertelement <4 x i32> %dim30, i32 %a, i32 2 224 %dim32 = insertelement <4 x i32> %dim31, i32 %a, i32 3 225 226 %offset_ptr = getelementptr <4 x float>, <4 x float>* null, i32 1 227 %offset_int = ptrtoint <4 x float>* %offset_ptr to i64 228 %sizeof32 = trunc i64 %offset_int to i32 229 230 %smearinsert33 = insertelement <4 x i32> undef, i32 %sizeof32, i32 0 231 %smearinsert34 = insertelement <4 x i32> %smearinsert33, i32 %sizeof32, i32 1 232 %smearinsert35 = insertelement <4 x i32> %smearinsert34, i32 %sizeof32, i32 2 233 %smearinsert36 = insertelement <4 x i32> %smearinsert35, i32 %sizeof32, i32 3 234 235 %delta_scale = mul <4 x i32> %dim32, %smearinsert36 236 %offset_delta = add <4 x i32> zeroinitializer, %delta_scale 237 238 %offset_varying_delta = add <4 x i32> %offset_delta, undef 239 240 ret <4 x float> undef 241} 242 243define <8 x i32> @pr24458(<8 x float> %n) { 244; CHECK-LABEL: @pr24458( 245; CHECK-NEXT: ret <8 x i32> <i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1, i32 -1> 246; 247 %notequal_b_load_.i = fcmp une <8 x float> %n, zeroinitializer 248 %equal_a_load72_.i = fcmp ueq <8 x float> %n, zeroinitializer 249 %notequal_b_load__to_boolvec.i = sext <8 x i1> %notequal_b_load_.i to <8 x i32> 250 %equal_a_load72__to_boolvec.i = sext <8 x i1> %equal_a_load72_.i to <8 x i32> 251 %wrong = or <8 x i32> %notequal_b_load__to_boolvec.i, %equal_a_load72__to_boolvec.i 252 ret <8 x i32> %wrong 253} 254 255; Hoist a trunc to a scalar if we're inserting into an undef vector. 256; trunc (inselt undef, X, Index) --> inselt undef, (trunc X), Index 257 258define <3 x i16> @trunc_inselt_undef(i32 %x) { 259; CHECK-LABEL: @trunc_inselt_undef( 260; CHECK-NEXT: [[TMP1:%.*]] = trunc i32 %x to i16 261; CHECK-NEXT: [[TRUNC:%.*]] = insertelement <3 x i16> undef, i16 [[TMP1]], i32 1 262; CHECK-NEXT: ret <3 x i16> [[TRUNC]] 263; 264 %vec = insertelement <3 x i32> undef, i32 %x, i32 1 265 %trunc = trunc <3 x i32> %vec to <3 x i16> 266 ret <3 x i16> %trunc 267} 268 269; Hoist a trunc to a scalar if we're inserting into an undef vector. 270; trunc (inselt undef, X, Index) --> inselt undef, (trunc X), Index 271 272define <2 x float> @fptrunc_inselt_undef(double %x, i32 %index) { 273; CHECK-LABEL: @fptrunc_inselt_undef( 274; CHECK-NEXT: [[TMP1:%.*]] = fptrunc double %x to float 275; CHECK-NEXT: [[TRUNC:%.*]] = insertelement <2 x float> undef, float [[TMP1]], i32 %index 276; CHECK-NEXT: ret <2 x float> [[TRUNC]] 277; 278 %vec = insertelement <2 x double> <double undef, double undef>, double %x, i32 %index 279 %trunc = fptrunc <2 x double> %vec to <2 x float> 280 ret <2 x float> %trunc 281} 282 283; TODO: Strengthen the backend, so we can have this canonicalization. 284; Insert a scalar int into a constant vector and truncate: 285; trunc (inselt C, X, Index) --> inselt C, (trunc X), Index 286 287define <3 x i16> @trunc_inselt1(i32 %x) { 288; CHECK-LABEL: @trunc_inselt1( 289; CHECK-NEXT: [[VEC:%.*]] = insertelement <3 x i32> <i32 3, i32 undef, i32 65536>, i32 %x, i32 1 290; CHECK-NEXT: [[TRUNC:%.*]] = trunc <3 x i32> [[VEC]] to <3 x i16> 291; CHECK-NEXT: ret <3 x i16> [[TRUNC]] 292; 293 %vec = insertelement <3 x i32> <i32 3, i32 -2, i32 65536>, i32 %x, i32 1 294 %trunc = trunc <3 x i32> %vec to <3 x i16> 295 ret <3 x i16> %trunc 296} 297 298; TODO: Strengthen the backend, so we can have this canonicalization. 299; Insert a scalar FP into a constant vector and FP truncate: 300; fptrunc (inselt C, X, Index) --> inselt C, (fptrunc X), Index 301 302define <2 x float> @fptrunc_inselt1(double %x, i32 %index) { 303; CHECK-LABEL: @fptrunc_inselt1( 304; CHECK-NEXT: [[VEC:%.*]] = insertelement <2 x double> <double undef, double 3.000000e+00>, double %x, i32 %index 305; CHECK-NEXT: [[TRUNC:%.*]] = fptrunc <2 x double> [[VEC]] to <2 x float> 306; CHECK-NEXT: ret <2 x float> [[TRUNC]] 307; 308 %vec = insertelement <2 x double> <double undef, double 3.0>, double %x, i32 %index 309 %trunc = fptrunc <2 x double> %vec to <2 x float> 310 ret <2 x float> %trunc 311} 312 313; TODO: Strengthen the backend, so we can have this canonicalization. 314; Insert a scalar int constant into a vector and truncate: 315; trunc (inselt X, C, Index) --> inselt (trunc X), C', Index 316 317define <8 x i16> @trunc_inselt2(<8 x i32> %x, i32 %index) { 318; CHECK-LABEL: @trunc_inselt2( 319; CHECK-NEXT: [[VEC:%.*]] = insertelement <8 x i32> %x, i32 1048576, i32 %index 320; CHECK-NEXT: [[TRUNC:%.*]] = trunc <8 x i32> [[VEC]] to <8 x i16> 321; CHECK-NEXT: ret <8 x i16> [[TRUNC]] 322; 323 %vec = insertelement <8 x i32> %x, i32 1048576, i32 %index 324 %trunc = trunc <8 x i32> %vec to <8 x i16> 325 ret <8 x i16> %trunc 326} 327 328; TODO: Strengthen the backend, so we can have this canonicalization. 329; Insert a scalar FP constant into a vector and FP truncate: 330; fptrunc (inselt X, C, Index) --> inselt (fptrunc X), C', Index 331 332define <3 x float> @fptrunc_inselt2(<3 x double> %x) { 333; CHECK-LABEL: @fptrunc_inselt2( 334; CHECK-NEXT: [[VEC:%.*]] = insertelement <3 x double> %x, double 4.000000e+00, i32 2 335; CHECK-NEXT: [[TRUNC:%.*]] = fptrunc <3 x double> [[VEC]] to <3 x float> 336; CHECK-NEXT: ret <3 x float> [[TRUNC]] 337; 338 %vec = insertelement <3 x double> %x, double 4.0, i32 2 339 %trunc = fptrunc <3 x double> %vec to <3 x float> 340 ret <3 x float> %trunc 341} 342 343