1; RUN: llc < %s -mtriple=x86_64-apple-darwin -mcpu=core-avx2 -mattr=+avx2 | FileCheck %s 2 3; CHECK: vpbroadcastb (% 4define <16 x i8> @BB16(i8* %ptr) nounwind uwtable readnone ssp { 5entry: 6 %q = load i8* %ptr, align 4 7 %q0 = insertelement <16 x i8> undef, i8 %q, i32 0 8 %q1 = insertelement <16 x i8> %q0, i8 %q, i32 1 9 %q2 = insertelement <16 x i8> %q1, i8 %q, i32 2 10 %q3 = insertelement <16 x i8> %q2, i8 %q, i32 3 11 %q4 = insertelement <16 x i8> %q3, i8 %q, i32 4 12 %q5 = insertelement <16 x i8> %q4, i8 %q, i32 5 13 %q6 = insertelement <16 x i8> %q5, i8 %q, i32 6 14 %q7 = insertelement <16 x i8> %q6, i8 %q, i32 7 15 %q8 = insertelement <16 x i8> %q7, i8 %q, i32 8 16 %q9 = insertelement <16 x i8> %q8, i8 %q, i32 9 17 %qa = insertelement <16 x i8> %q9, i8 %q, i32 10 18 %qb = insertelement <16 x i8> %qa, i8 %q, i32 11 19 %qc = insertelement <16 x i8> %qb, i8 %q, i32 12 20 %qd = insertelement <16 x i8> %qc, i8 %q, i32 13 21 %qe = insertelement <16 x i8> %qd, i8 %q, i32 14 22 %qf = insertelement <16 x i8> %qe, i8 %q, i32 15 23 ret <16 x i8> %qf 24} 25; CHECK: vpbroadcastb (% 26define <32 x i8> @BB32(i8* %ptr) nounwind uwtable readnone ssp { 27entry: 28 %q = load i8* %ptr, align 4 29 %q0 = insertelement <32 x i8> undef, i8 %q, i32 0 30 %q1 = insertelement <32 x i8> %q0, i8 %q, i32 1 31 %q2 = insertelement <32 x i8> %q1, i8 %q, i32 2 32 %q3 = insertelement <32 x i8> %q2, i8 %q, i32 3 33 %q4 = insertelement <32 x i8> %q3, i8 %q, i32 4 34 %q5 = insertelement <32 x i8> %q4, i8 %q, i32 5 35 %q6 = insertelement <32 x i8> %q5, i8 %q, i32 6 36 %q7 = insertelement <32 x i8> %q6, i8 %q, i32 7 37 %q8 = insertelement <32 x i8> %q7, i8 %q, i32 8 38 %q9 = insertelement <32 x i8> %q8, i8 %q, i32 9 39 %qa = insertelement <32 x i8> %q9, i8 %q, i32 10 40 %qb = insertelement <32 x i8> %qa, i8 %q, i32 11 41 %qc = insertelement <32 x i8> %qb, i8 %q, i32 12 42 %qd = insertelement <32 x i8> %qc, i8 %q, i32 13 43 %qe = insertelement <32 x i8> %qd, i8 %q, i32 14 44 %qf = insertelement <32 x i8> %qe, i8 %q, i32 15 45 46 %q20 = insertelement <32 x i8> %qf, i8 %q, i32 16 47 %q21 = insertelement <32 x i8> %q20, i8 %q, i32 17 48 %q22 = insertelement <32 x i8> %q21, i8 %q, i32 18 49 %q23 = insertelement <32 x i8> %q22, i8 %q, i32 19 50 %q24 = insertelement <32 x i8> %q23, i8 %q, i32 20 51 %q25 = insertelement <32 x i8> %q24, i8 %q, i32 21 52 %q26 = insertelement <32 x i8> %q25, i8 %q, i32 22 53 %q27 = insertelement <32 x i8> %q26, i8 %q, i32 23 54 %q28 = insertelement <32 x i8> %q27, i8 %q, i32 24 55 %q29 = insertelement <32 x i8> %q28, i8 %q, i32 25 56 %q2a = insertelement <32 x i8> %q29, i8 %q, i32 26 57 %q2b = insertelement <32 x i8> %q2a, i8 %q, i32 27 58 %q2c = insertelement <32 x i8> %q2b, i8 %q, i32 28 59 %q2d = insertelement <32 x i8> %q2c, i8 %q, i32 29 60 %q2e = insertelement <32 x i8> %q2d, i8 %q, i32 30 61 %q2f = insertelement <32 x i8> %q2e, i8 %q, i32 31 62 ret <32 x i8> %q2f 63} 64; CHECK: vpbroadcastw (% 65 66define <8 x i16> @W16(i16* %ptr) nounwind uwtable readnone ssp { 67entry: 68 %q = load i16* %ptr, align 4 69 %q0 = insertelement <8 x i16> undef, i16 %q, i32 0 70 %q1 = insertelement <8 x i16> %q0, i16 %q, i32 1 71 %q2 = insertelement <8 x i16> %q1, i16 %q, i32 2 72 %q3 = insertelement <8 x i16> %q2, i16 %q, i32 3 73 %q4 = insertelement <8 x i16> %q3, i16 %q, i32 4 74 %q5 = insertelement <8 x i16> %q4, i16 %q, i32 5 75 %q6 = insertelement <8 x i16> %q5, i16 %q, i32 6 76 %q7 = insertelement <8 x i16> %q6, i16 %q, i32 7 77 ret <8 x i16> %q7 78} 79; CHECK: vpbroadcastw (% 80define <16 x i16> @WW16(i16* %ptr) nounwind uwtable readnone ssp { 81entry: 82 %q = load i16* %ptr, align 4 83 %q0 = insertelement <16 x i16> undef, i16 %q, i32 0 84 %q1 = insertelement <16 x i16> %q0, i16 %q, i32 1 85 %q2 = insertelement <16 x i16> %q1, i16 %q, i32 2 86 %q3 = insertelement <16 x i16> %q2, i16 %q, i32 3 87 %q4 = insertelement <16 x i16> %q3, i16 %q, i32 4 88 %q5 = insertelement <16 x i16> %q4, i16 %q, i32 5 89 %q6 = insertelement <16 x i16> %q5, i16 %q, i32 6 90 %q7 = insertelement <16 x i16> %q6, i16 %q, i32 7 91 %q8 = insertelement <16 x i16> %q7, i16 %q, i32 8 92 %q9 = insertelement <16 x i16> %q8, i16 %q, i32 9 93 %qa = insertelement <16 x i16> %q9, i16 %q, i32 10 94 %qb = insertelement <16 x i16> %qa, i16 %q, i32 11 95 %qc = insertelement <16 x i16> %qb, i16 %q, i32 12 96 %qd = insertelement <16 x i16> %qc, i16 %q, i32 13 97 %qe = insertelement <16 x i16> %qd, i16 %q, i32 14 98 %qf = insertelement <16 x i16> %qe, i16 %q, i32 15 99 ret <16 x i16> %qf 100} 101; CHECK: vpbroadcastd (% 102define <4 x i32> @D32(i32* %ptr) nounwind uwtable readnone ssp { 103entry: 104 %q = load i32* %ptr, align 4 105 %q0 = insertelement <4 x i32> undef, i32 %q, i32 0 106 %q1 = insertelement <4 x i32> %q0, i32 %q, i32 1 107 %q2 = insertelement <4 x i32> %q1, i32 %q, i32 2 108 %q3 = insertelement <4 x i32> %q2, i32 %q, i32 3 109 ret <4 x i32> %q3 110} 111; CHECK: vpbroadcastd (% 112define <8 x i32> @DD32(i32* %ptr) nounwind uwtable readnone ssp { 113entry: 114 %q = load i32* %ptr, align 4 115 %q0 = insertelement <8 x i32> undef, i32 %q, i32 0 116 %q1 = insertelement <8 x i32> %q0, i32 %q, i32 1 117 %q2 = insertelement <8 x i32> %q1, i32 %q, i32 2 118 %q3 = insertelement <8 x i32> %q2, i32 %q, i32 3 119 %q4 = insertelement <8 x i32> %q3, i32 %q, i32 4 120 %q5 = insertelement <8 x i32> %q4, i32 %q, i32 5 121 %q6 = insertelement <8 x i32> %q5, i32 %q, i32 6 122 %q7 = insertelement <8 x i32> %q6, i32 %q, i32 7 123 ret <8 x i32> %q7 124} 125; CHECK: vpbroadcastq (% 126define <2 x i64> @Q64(i64* %ptr) nounwind uwtable readnone ssp { 127entry: 128 %q = load i64* %ptr, align 4 129 %q0 = insertelement <2 x i64> undef, i64 %q, i32 0 130 %q1 = insertelement <2 x i64> %q0, i64 %q, i32 1 131 ret <2 x i64> %q1 132} 133; CHECK: vpbroadcastq (% 134define <4 x i64> @QQ64(i64* %ptr) nounwind uwtable readnone ssp { 135entry: 136 %q = load i64* %ptr, align 4 137 %q0 = insertelement <4 x i64> undef, i64 %q, i32 0 138 %q1 = insertelement <4 x i64> %q0, i64 %q, i32 1 139 %q2 = insertelement <4 x i64> %q1, i64 %q, i32 2 140 %q3 = insertelement <4 x i64> %q2, i64 %q, i32 3 141 ret <4 x i64> %q3 142} 143 144; make sure that we still don't support broadcast double into 128-bit vector 145; this used to crash 146define <2 x double> @I(double* %ptr) nounwind uwtable readnone ssp { 147entry: 148 %q = load double* %ptr, align 4 149 %vecinit.i = insertelement <2 x double> undef, double %q, i32 0 150 %vecinit2.i = insertelement <2 x double> %vecinit.i, double %q, i32 1 151 ret <2 x double> %vecinit2.i 152} 153 154; CHECK: V111 155; CHECK: vpbroadcastd 156; CHECK: ret 157define <8 x i32> @V111(<8 x i32> %in) nounwind uwtable readnone ssp { 158entry: 159 %g = add <8 x i32> %in, <i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1> 160 ret <8 x i32> %g 161} 162 163; CHECK: _e2 164; CHECK: vbroadcastss 165; CHECK: ret 166define <4 x float> @_e2(float* %ptr) nounwind uwtable readnone ssp { 167 %vecinit.i = insertelement <4 x float> undef, float 0xbf80000000000000, i32 0 168 %vecinit2.i = insertelement <4 x float> %vecinit.i, float 0xbf80000000000000, i32 1 169 %vecinit4.i = insertelement <4 x float> %vecinit2.i, float 0xbf80000000000000, i32 2 170 %vecinit6.i = insertelement <4 x float> %vecinit4.i, float 0xbf80000000000000, i32 3 171 ret <4 x float> %vecinit6.i 172} 173 174; CHECK: _e4 175; CHECK-NOT: broadcast 176; CHECK: ret 177define <8 x i8> @_e4(i8* %ptr) nounwind uwtable readnone ssp { 178 %vecinit0.i = insertelement <8 x i8> undef, i8 52, i32 0 179 %vecinit1.i = insertelement <8 x i8> %vecinit0.i, i8 52, i32 1 180 %vecinit2.i = insertelement <8 x i8> %vecinit1.i, i8 52, i32 2 181 %vecinit3.i = insertelement <8 x i8> %vecinit2.i, i8 52, i32 3 182 %vecinit4.i = insertelement <8 x i8> %vecinit3.i, i8 52, i32 3 183 %vecinit5.i = insertelement <8 x i8> %vecinit4.i, i8 52, i32 3 184 %vecinit6.i = insertelement <8 x i8> %vecinit5.i, i8 52, i32 3 185 %vecinit7.i = insertelement <8 x i8> %vecinit6.i, i8 52, i32 3 186 ret <8 x i8> %vecinit7.i 187} 188