1; RUN: llc -mtriple=aarch64-gnu-linux -mcpu=cortex-a57 -enable-unsafe-fp-math -disable-post-ra < %s | FileCheck %s 2 3; Verify that the first two adds are independent regardless of how the inputs are 4; commuted. The destination registers are used as source registers for the third add. 5 6define float @reassociate_adds1(float %x0, float %x1, float %x2, float %x3) { 7; CHECK-LABEL: reassociate_adds1: 8; CHECK: fadd s0, s0, s1 9; CHECK-NEXT: fadd s1, s2, s3 10; CHECK-NEXT: fadd s0, s0, s1 11; CHECK-NEXT: ret 12 %t0 = fadd float %x0, %x1 13 %t1 = fadd float %t0, %x2 14 %t2 = fadd float %t1, %x3 15 ret float %t2 16} 17 18define float @reassociate_adds2(float %x0, float %x1, float %x2, float %x3) { 19; CHECK-LABEL: reassociate_adds2: 20; CHECK: fadd s0, s0, s1 21; CHECK-NEXT: fadd s1, s2, s3 22; CHECK-NEXT: fadd s0, s0, s1 23; CHECK-NEXT: ret 24 %t0 = fadd float %x0, %x1 25 %t1 = fadd float %x2, %t0 26 %t2 = fadd float %t1, %x3 27 ret float %t2 28} 29 30define float @reassociate_adds3(float %x0, float %x1, float %x2, float %x3) { 31; CHECK-LABEL: reassociate_adds3: 32; CHECK: s0, s0, s1 33; CHECK-NEXT: s1, s2, s3 34; CHECK-NEXT: s0, s0, s1 35; CHECK-NEXT: ret 36 %t0 = fadd float %x0, %x1 37 %t1 = fadd float %t0, %x2 38 %t2 = fadd float %x3, %t1 39 ret float %t2 40} 41 42define float @reassociate_adds4(float %x0, float %x1, float %x2, float %x3) { 43; CHECK-LABEL: reassociate_adds4: 44; CHECK: s0, s0, s1 45; CHECK-NEXT: s1, s2, s3 46; CHECK-NEXT: s0, s0, s1 47; CHECK-NEXT: ret 48 %t0 = fadd float %x0, %x1 49 %t1 = fadd float %x2, %t0 50 %t2 = fadd float %x3, %t1 51 ret float %t2 52} 53 54; Verify that we reassociate some of these ops. The optimal balanced tree of adds is not 55; produced because that would cost more compile time. 56 57define float @reassociate_adds5(float %x0, float %x1, float %x2, float %x3, float %x4, float %x5, float %x6, float %x7) { 58; CHECK-LABEL: reassociate_adds5: 59; CHECK: fadd s0, s0, s1 60; CHECK-NEXT: fadd s1, s2, s3 61; CHECK-NEXT: fadd s0, s0, s1 62; CHECK-NEXT: fadd s1, s4, s5 63; CHECK-NEXT: fadd s1, s1, s6 64; CHECK-NEXT: fadd s0, s0, s1 65; CHECK-NEXT: fadd s0, s0, s7 66; CHECK-NEXT: ret 67 %t0 = fadd float %x0, %x1 68 %t1 = fadd float %t0, %x2 69 %t2 = fadd float %t1, %x3 70 %t3 = fadd float %t2, %x4 71 %t4 = fadd float %t3, %x5 72 %t5 = fadd float %t4, %x6 73 %t6 = fadd float %t5, %x7 74 ret float %t6 75} 76 77; Verify that we only need two associative operations to reassociate the operands. 78; Also, we should reassociate such that the result of the high latency division 79; is used by the final 'add' rather than reassociating the %x3 operand with the 80; division. The latter reassociation would not improve anything. 81 82define float @reassociate_adds6(float %x0, float %x1, float %x2, float %x3) { 83; CHECK-LABEL: reassociate_adds6: 84; CHECK: fdiv s0, s0, s1 85; CHECK-NEXT: fadd s1, s2, s3 86; CHECK-NEXT: fadd s0, s0, s1 87; CHECK-NEXT: ret 88 %t0 = fdiv float %x0, %x1 89 %t1 = fadd float %x2, %t0 90 %t2 = fadd float %x3, %t1 91 ret float %t2 92} 93 94; Verify that scalar single-precision multiplies are reassociated. 95 96define float @reassociate_muls1(float %x0, float %x1, float %x2, float %x3) { 97; CHECK-LABEL: reassociate_muls1: 98; CHECK: fdiv s0, s0, s1 99; CHECK-NEXT: fmul s1, s2, s3 100; CHECK-NEXT: fmul s0, s0, s1 101; CHECK-NEXT: ret 102 %t0 = fdiv float %x0, %x1 103 %t1 = fmul float %x2, %t0 104 %t2 = fmul float %x3, %t1 105 ret float %t2 106} 107 108; Verify that scalar double-precision adds are reassociated. 109 110define double @reassociate_adds_double(double %x0, double %x1, double %x2, double %x3) { 111; CHECK-LABEL: reassociate_adds_double: 112; CHECK: fdiv d0, d0, d1 113; CHECK-NEXT: fadd d1, d2, d3 114; CHECK-NEXT: fadd d0, d0, d1 115; CHECK-NEXT: ret 116 %t0 = fdiv double %x0, %x1 117 %t1 = fadd double %x2, %t0 118 %t2 = fadd double %x3, %t1 119 ret double %t2 120} 121 122; Verify that scalar double-precision multiplies are reassociated. 123 124define double @reassociate_muls_double(double %x0, double %x1, double %x2, double %x3) { 125; CHECK-LABEL: reassociate_muls_double: 126; CHECK: fdiv d0, d0, d1 127; CHECK-NEXT: fmul d1, d2, d3 128; CHECK-NEXT: fmul d0, d0, d1 129; CHECK-NEXT: ret 130 %t0 = fdiv double %x0, %x1 131 %t1 = fmul double %x2, %t0 132 %t2 = fmul double %x3, %t1 133 ret double %t2 134} 135 136; Verify that we reassociate vector instructions too. 137 138define <4 x float> @vector_reassociate_adds1(<4 x float> %x0, <4 x float> %x1, <4 x float> %x2, <4 x float> %x3) { 139; CHECK-LABEL: vector_reassociate_adds1: 140; CHECK: fadd v0.4s, v0.4s, v1.4s 141; CHECK-NEXT: fadd v1.4s, v2.4s, v3.4s 142; CHECK-NEXT: fadd v0.4s, v0.4s, v1.4s 143; CHECK-NEXT: ret 144 %t0 = fadd <4 x float> %x0, %x1 145 %t1 = fadd <4 x float> %t0, %x2 146 %t2 = fadd <4 x float> %t1, %x3 147 ret <4 x float> %t2 148} 149 150define <4 x float> @vector_reassociate_adds2(<4 x float> %x0, <4 x float> %x1, <4 x float> %x2, <4 x float> %x3) { 151; CHECK-LABEL: vector_reassociate_adds2: 152; CHECK: fadd v0.4s, v0.4s, v1.4s 153; CHECK-NEXT: fadd v1.4s, v2.4s, v3.4s 154; CHECK-NEXT: fadd v0.4s, v0.4s, v1.4s 155 %t0 = fadd <4 x float> %x0, %x1 156 %t1 = fadd <4 x float> %x2, %t0 157 %t2 = fadd <4 x float> %t1, %x3 158 ret <4 x float> %t2 159} 160 161define <4 x float> @vector_reassociate_adds3(<4 x float> %x0, <4 x float> %x1, <4 x float> %x2, <4 x float> %x3) { 162; CHECK-LABEL: vector_reassociate_adds3: 163; CHECK: fadd v0.4s, v0.4s, v1.4s 164; CHECK-NEXT: fadd v1.4s, v2.4s, v3.4s 165; CHECK-NEXT: fadd v0.4s, v0.4s, v1.4s 166 %t0 = fadd <4 x float> %x0, %x1 167 %t1 = fadd <4 x float> %t0, %x2 168 %t2 = fadd <4 x float> %x3, %t1 169 ret <4 x float> %t2 170} 171 172define <4 x float> @vector_reassociate_adds4(<4 x float> %x0, <4 x float> %x1, <4 x float> %x2, <4 x float> %x3) { 173; CHECK-LABEL: vector_reassociate_adds4: 174; CHECK: fadd v0.4s, v0.4s, v1.4s 175; CHECK-NEXT: fadd v1.4s, v2.4s, v3.4s 176; CHECK-NEXT: fadd v0.4s, v0.4s, v1.4s 177 %t0 = fadd <4 x float> %x0, %x1 178 %t1 = fadd <4 x float> %x2, %t0 179 %t2 = fadd <4 x float> %x3, %t1 180 ret <4 x float> %t2 181} 182; Verify that 128-bit vector single-precision multiplies are reassociated. 183 184define <4 x float> @reassociate_muls_v4f32(<4 x float> %x0, <4 x float> %x1, <4 x float> %x2, <4 x float> %x3) { 185; CHECK-LABEL: reassociate_muls_v4f32: 186; CHECK: fadd v0.4s, v0.4s, v1.4s 187; CHECK-NEXT: fmul v1.4s, v2.4s, v3.4s 188; CHECK-NEXT: fmul v0.4s, v0.4s, v1.4s 189; CHECK-NEXT: ret 190 %t0 = fadd <4 x float> %x0, %x1 191 %t1 = fmul <4 x float> %x2, %t0 192 %t2 = fmul <4 x float> %x3, %t1 193 ret <4 x float> %t2 194} 195 196; Verify that 128-bit vector double-precision multiplies are reassociated. 197 198define <2 x double> @reassociate_muls_v2f64(<2 x double> %x0, <2 x double> %x1, <2 x double> %x2, <2 x double> %x3) { 199; CHECK-LABEL: reassociate_muls_v2f64: 200; CHECK: fadd v0.2d, v0.2d, v1.2d 201; CHECK-NEXT: fmul v1.2d, v2.2d, v3.2d 202; CHECK-NEXT: fmul v0.2d, v0.2d, v1.2d 203; CHECK-NEXT: ret 204 %t0 = fadd <2 x double> %x0, %x1 205 %t1 = fmul <2 x double> %x2, %t0 206 %t2 = fmul <2 x double> %x3, %t1 207 ret <2 x double> %t2 208} 209 210; PR25016: https://llvm.org/bugs/show_bug.cgi?id=25016 211; Verify that reassociation is not happening needlessly or wrongly. 212 213declare double @bar() 214 215define double @reassociate_adds_from_calls() { 216; CHECK-LABEL: reassociate_adds_from_calls: 217; CHECK: bl bar 218; CHECK-NEXT: mov v8.16b, v0.16b 219; CHECK-NEXT: bl bar 220; CHECK-NEXT: mov v9.16b, v0.16b 221; CHECK-NEXT: bl bar 222; CHECK-NEXT: mov v10.16b, v0.16b 223; CHECK-NEXT: bl bar 224; CHECK: fadd d1, d8, d9 225; CHECK-NEXT: fadd d0, d10, d0 226; CHECK-NEXT: fadd d0, d1, d0 227 %x0 = call double @bar() 228 %x1 = call double @bar() 229 %x2 = call double @bar() 230 %x3 = call double @bar() 231 %t0 = fadd double %x0, %x1 232 %t1 = fadd double %t0, %x2 233 %t2 = fadd double %t1, %x3 234 ret double %t2 235} 236 237define double @already_reassociated() { 238; CHECK-LABEL: already_reassociated: 239; CHECK: bl bar 240; CHECK-NEXT: mov v8.16b, v0.16b 241; CHECK-NEXT: bl bar 242; CHECK-NEXT: mov v9.16b, v0.16b 243; CHECK-NEXT: bl bar 244; CHECK-NEXT: mov v10.16b, v0.16b 245; CHECK-NEXT: bl bar 246; CHECK: fadd d1, d8, d9 247; CHECK-NEXT: fadd d0, d10, d0 248; CHECK-NEXT: fadd d0, d1, d0 249 %x0 = call double @bar() 250 %x1 = call double @bar() 251 %x2 = call double @bar() 252 %x3 = call double @bar() 253 %t0 = fadd double %x0, %x1 254 %t1 = fadd double %x2, %x3 255 %t2 = fadd double %t0, %t1 256 ret double %t2 257} 258 259