Geoff Berry | 62c1a1e | 2016-03-02 17:58:31 +0000 | [diff] [blame] | 1 | ; RUN: llc -mtriple=aarch64-gnu-linux -mcpu=cortex-a57 -enable-unsafe-fp-math -disable-post-ra < %s | FileCheck %s |
Haicheng Wu | 08b9462 | 2016-01-07 04:01:02 +0000 | [diff] [blame] | 2 | |
Florian Hahn | ceb4494 | 2017-09-20 11:54:37 +0000 | [diff] [blame] | 3 | ; Incremental updates of the instruction depths should be enough for this test |
| 4 | ; case. |
| 5 | ; RUN: llc -mtriple=aarch64-gnu-linux -mcpu=cortex-a57 -enable-unsafe-fp-math \ |
| 6 | ; RUN: -disable-post-ra -machine-combiner-inc-threshold=0 < %s | FileCheck %s |
| 7 | |
Haicheng Wu | 08b9462 | 2016-01-07 04:01:02 +0000 | [diff] [blame] | 8 | ; Verify that the first two adds are independent regardless of how the inputs are |
| 9 | ; commuted. The destination registers are used as source registers for the third add. |
| 10 | |
| 11 | define float @reassociate_adds1(float %x0, float %x1, float %x2, float %x3) { |
| 12 | ; CHECK-LABEL: reassociate_adds1: |
| 13 | ; CHECK: fadd s0, s0, s1 |
| 14 | ; CHECK-NEXT: fadd s1, s2, s3 |
| 15 | ; CHECK-NEXT: fadd s0, s0, s1 |
| 16 | ; CHECK-NEXT: ret |
| 17 | %t0 = fadd float %x0, %x1 |
| 18 | %t1 = fadd float %t0, %x2 |
| 19 | %t2 = fadd float %t1, %x3 |
| 20 | ret float %t2 |
| 21 | } |
| 22 | |
| 23 | define float @reassociate_adds2(float %x0, float %x1, float %x2, float %x3) { |
| 24 | ; CHECK-LABEL: reassociate_adds2: |
| 25 | ; CHECK: fadd s0, s0, s1 |
| 26 | ; CHECK-NEXT: fadd s1, s2, s3 |
| 27 | ; CHECK-NEXT: fadd s0, s0, s1 |
| 28 | ; CHECK-NEXT: ret |
| 29 | %t0 = fadd float %x0, %x1 |
| 30 | %t1 = fadd float %x2, %t0 |
| 31 | %t2 = fadd float %t1, %x3 |
| 32 | ret float %t2 |
| 33 | } |
| 34 | |
| 35 | define float @reassociate_adds3(float %x0, float %x1, float %x2, float %x3) { |
| 36 | ; CHECK-LABEL: reassociate_adds3: |
| 37 | ; CHECK: s0, s0, s1 |
| 38 | ; CHECK-NEXT: s1, s2, s3 |
| 39 | ; CHECK-NEXT: s0, s0, s1 |
| 40 | ; CHECK-NEXT: ret |
| 41 | %t0 = fadd float %x0, %x1 |
| 42 | %t1 = fadd float %t0, %x2 |
| 43 | %t2 = fadd float %x3, %t1 |
| 44 | ret float %t2 |
| 45 | } |
| 46 | |
| 47 | define float @reassociate_adds4(float %x0, float %x1, float %x2, float %x3) { |
| 48 | ; CHECK-LABEL: reassociate_adds4: |
| 49 | ; CHECK: s0, s0, s1 |
| 50 | ; CHECK-NEXT: s1, s2, s3 |
| 51 | ; CHECK-NEXT: s0, s0, s1 |
| 52 | ; CHECK-NEXT: ret |
| 53 | %t0 = fadd float %x0, %x1 |
| 54 | %t1 = fadd float %x2, %t0 |
| 55 | %t2 = fadd float %x3, %t1 |
| 56 | ret float %t2 |
| 57 | } |
| 58 | |
| 59 | ; Verify that we reassociate some of these ops. The optimal balanced tree of adds is not |
| 60 | ; produced because that would cost more compile time. |
| 61 | |
| 62 | define float @reassociate_adds5(float %x0, float %x1, float %x2, float %x3, float %x4, float %x5, float %x6, float %x7) { |
| 63 | ; CHECK-LABEL: reassociate_adds5: |
| 64 | ; CHECK: fadd s0, s0, s1 |
| 65 | ; CHECK-NEXT: fadd s1, s2, s3 |
| 66 | ; CHECK-NEXT: fadd s0, s0, s1 |
| 67 | ; CHECK-NEXT: fadd s1, s4, s5 |
| 68 | ; CHECK-NEXT: fadd s1, s1, s6 |
| 69 | ; CHECK-NEXT: fadd s0, s0, s1 |
| 70 | ; CHECK-NEXT: fadd s0, s0, s7 |
| 71 | ; CHECK-NEXT: ret |
| 72 | %t0 = fadd float %x0, %x1 |
| 73 | %t1 = fadd float %t0, %x2 |
| 74 | %t2 = fadd float %t1, %x3 |
| 75 | %t3 = fadd float %t2, %x4 |
| 76 | %t4 = fadd float %t3, %x5 |
| 77 | %t5 = fadd float %t4, %x6 |
| 78 | %t6 = fadd float %t5, %x7 |
| 79 | ret float %t6 |
| 80 | } |
| 81 | |
| 82 | ; Verify that we only need two associative operations to reassociate the operands. |
| 83 | ; Also, we should reassociate such that the result of the high latency division |
| 84 | ; is used by the final 'add' rather than reassociating the %x3 operand with the |
| 85 | ; division. The latter reassociation would not improve anything. |
| 86 | |
| 87 | define float @reassociate_adds6(float %x0, float %x1, float %x2, float %x3) { |
| 88 | ; CHECK-LABEL: reassociate_adds6: |
| 89 | ; CHECK: fdiv s0, s0, s1 |
| 90 | ; CHECK-NEXT: fadd s1, s2, s3 |
| 91 | ; CHECK-NEXT: fadd s0, s0, s1 |
| 92 | ; CHECK-NEXT: ret |
| 93 | %t0 = fdiv float %x0, %x1 |
| 94 | %t1 = fadd float %x2, %t0 |
| 95 | %t2 = fadd float %x3, %t1 |
| 96 | ret float %t2 |
| 97 | } |
| 98 | |
| 99 | ; Verify that scalar single-precision multiplies are reassociated. |
| 100 | |
| 101 | define float @reassociate_muls1(float %x0, float %x1, float %x2, float %x3) { |
| 102 | ; CHECK-LABEL: reassociate_muls1: |
| 103 | ; CHECK: fdiv s0, s0, s1 |
| 104 | ; CHECK-NEXT: fmul s1, s2, s3 |
| 105 | ; CHECK-NEXT: fmul s0, s0, s1 |
| 106 | ; CHECK-NEXT: ret |
| 107 | %t0 = fdiv float %x0, %x1 |
| 108 | %t1 = fmul float %x2, %t0 |
| 109 | %t2 = fmul float %x3, %t1 |
| 110 | ret float %t2 |
| 111 | } |
| 112 | |
| 113 | ; Verify that scalar double-precision adds are reassociated. |
| 114 | |
| 115 | define double @reassociate_adds_double(double %x0, double %x1, double %x2, double %x3) { |
| 116 | ; CHECK-LABEL: reassociate_adds_double: |
| 117 | ; CHECK: fdiv d0, d0, d1 |
| 118 | ; CHECK-NEXT: fadd d1, d2, d3 |
| 119 | ; CHECK-NEXT: fadd d0, d0, d1 |
| 120 | ; CHECK-NEXT: ret |
| 121 | %t0 = fdiv double %x0, %x1 |
| 122 | %t1 = fadd double %x2, %t0 |
| 123 | %t2 = fadd double %x3, %t1 |
| 124 | ret double %t2 |
| 125 | } |
| 126 | |
| 127 | ; Verify that scalar double-precision multiplies are reassociated. |
| 128 | |
| 129 | define double @reassociate_muls_double(double %x0, double %x1, double %x2, double %x3) { |
| 130 | ; CHECK-LABEL: reassociate_muls_double: |
| 131 | ; CHECK: fdiv d0, d0, d1 |
| 132 | ; CHECK-NEXT: fmul d1, d2, d3 |
| 133 | ; CHECK-NEXT: fmul d0, d0, d1 |
| 134 | ; CHECK-NEXT: ret |
| 135 | %t0 = fdiv double %x0, %x1 |
| 136 | %t1 = fmul double %x2, %t0 |
| 137 | %t2 = fmul double %x3, %t1 |
| 138 | ret double %t2 |
| 139 | } |
| 140 | |
| 141 | ; Verify that we reassociate vector instructions too. |
| 142 | |
| 143 | define <4 x float> @vector_reassociate_adds1(<4 x float> %x0, <4 x float> %x1, <4 x float> %x2, <4 x float> %x3) { |
| 144 | ; CHECK-LABEL: vector_reassociate_adds1: |
| 145 | ; CHECK: fadd v0.4s, v0.4s, v1.4s |
| 146 | ; CHECK-NEXT: fadd v1.4s, v2.4s, v3.4s |
| 147 | ; CHECK-NEXT: fadd v0.4s, v0.4s, v1.4s |
| 148 | ; CHECK-NEXT: ret |
| 149 | %t0 = fadd <4 x float> %x0, %x1 |
| 150 | %t1 = fadd <4 x float> %t0, %x2 |
| 151 | %t2 = fadd <4 x float> %t1, %x3 |
| 152 | ret <4 x float> %t2 |
| 153 | } |
| 154 | |
| 155 | define <4 x float> @vector_reassociate_adds2(<4 x float> %x0, <4 x float> %x1, <4 x float> %x2, <4 x float> %x3) { |
| 156 | ; CHECK-LABEL: vector_reassociate_adds2: |
| 157 | ; CHECK: fadd v0.4s, v0.4s, v1.4s |
| 158 | ; CHECK-NEXT: fadd v1.4s, v2.4s, v3.4s |
| 159 | ; CHECK-NEXT: fadd v0.4s, v0.4s, v1.4s |
| 160 | %t0 = fadd <4 x float> %x0, %x1 |
| 161 | %t1 = fadd <4 x float> %x2, %t0 |
| 162 | %t2 = fadd <4 x float> %t1, %x3 |
| 163 | ret <4 x float> %t2 |
| 164 | } |
| 165 | |
| 166 | define <4 x float> @vector_reassociate_adds3(<4 x float> %x0, <4 x float> %x1, <4 x float> %x2, <4 x float> %x3) { |
| 167 | ; CHECK-LABEL: vector_reassociate_adds3: |
| 168 | ; CHECK: fadd v0.4s, v0.4s, v1.4s |
| 169 | ; CHECK-NEXT: fadd v1.4s, v2.4s, v3.4s |
| 170 | ; CHECK-NEXT: fadd v0.4s, v0.4s, v1.4s |
| 171 | %t0 = fadd <4 x float> %x0, %x1 |
| 172 | %t1 = fadd <4 x float> %t0, %x2 |
| 173 | %t2 = fadd <4 x float> %x3, %t1 |
| 174 | ret <4 x float> %t2 |
| 175 | } |
| 176 | |
| 177 | define <4 x float> @vector_reassociate_adds4(<4 x float> %x0, <4 x float> %x1, <4 x float> %x2, <4 x float> %x3) { |
| 178 | ; CHECK-LABEL: vector_reassociate_adds4: |
| 179 | ; CHECK: fadd v0.4s, v0.4s, v1.4s |
| 180 | ; CHECK-NEXT: fadd v1.4s, v2.4s, v3.4s |
| 181 | ; CHECK-NEXT: fadd v0.4s, v0.4s, v1.4s |
| 182 | %t0 = fadd <4 x float> %x0, %x1 |
| 183 | %t1 = fadd <4 x float> %x2, %t0 |
| 184 | %t2 = fadd <4 x float> %x3, %t1 |
| 185 | ret <4 x float> %t2 |
| 186 | } |
| 187 | ; Verify that 128-bit vector single-precision multiplies are reassociated. |
| 188 | |
| 189 | define <4 x float> @reassociate_muls_v4f32(<4 x float> %x0, <4 x float> %x1, <4 x float> %x2, <4 x float> %x3) { |
| 190 | ; CHECK-LABEL: reassociate_muls_v4f32: |
| 191 | ; CHECK: fadd v0.4s, v0.4s, v1.4s |
| 192 | ; CHECK-NEXT: fmul v1.4s, v2.4s, v3.4s |
| 193 | ; CHECK-NEXT: fmul v0.4s, v0.4s, v1.4s |
| 194 | ; CHECK-NEXT: ret |
| 195 | %t0 = fadd <4 x float> %x0, %x1 |
| 196 | %t1 = fmul <4 x float> %x2, %t0 |
| 197 | %t2 = fmul <4 x float> %x3, %t1 |
| 198 | ret <4 x float> %t2 |
| 199 | } |
| 200 | |
| 201 | ; Verify that 128-bit vector double-precision multiplies are reassociated. |
| 202 | |
| 203 | define <2 x double> @reassociate_muls_v2f64(<2 x double> %x0, <2 x double> %x1, <2 x double> %x2, <2 x double> %x3) { |
| 204 | ; CHECK-LABEL: reassociate_muls_v2f64: |
| 205 | ; CHECK: fadd v0.2d, v0.2d, v1.2d |
| 206 | ; CHECK-NEXT: fmul v1.2d, v2.2d, v3.2d |
| 207 | ; CHECK-NEXT: fmul v0.2d, v0.2d, v1.2d |
| 208 | ; CHECK-NEXT: ret |
| 209 | %t0 = fadd <2 x double> %x0, %x1 |
| 210 | %t1 = fmul <2 x double> %x2, %t0 |
| 211 | %t2 = fmul <2 x double> %x3, %t1 |
| 212 | ret <2 x double> %t2 |
| 213 | } |
| 214 | |
| 215 | ; PR25016: https://llvm.org/bugs/show_bug.cgi?id=25016 |
| 216 | ; Verify that reassociation is not happening needlessly or wrongly. |
| 217 | |
| 218 | declare double @bar() |
| 219 | |
| 220 | define double @reassociate_adds_from_calls() { |
| 221 | ; CHECK-LABEL: reassociate_adds_from_calls: |
| 222 | ; CHECK: bl bar |
| 223 | ; CHECK-NEXT: mov v8.16b, v0.16b |
| 224 | ; CHECK-NEXT: bl bar |
| 225 | ; CHECK-NEXT: mov v9.16b, v0.16b |
| 226 | ; CHECK-NEXT: bl bar |
| 227 | ; CHECK-NEXT: mov v10.16b, v0.16b |
| 228 | ; CHECK-NEXT: bl bar |
| 229 | ; CHECK: fadd d1, d8, d9 |
| 230 | ; CHECK-NEXT: fadd d0, d10, d0 |
| 231 | ; CHECK-NEXT: fadd d0, d1, d0 |
| 232 | %x0 = call double @bar() |
| 233 | %x1 = call double @bar() |
| 234 | %x2 = call double @bar() |
| 235 | %x3 = call double @bar() |
| 236 | %t0 = fadd double %x0, %x1 |
| 237 | %t1 = fadd double %t0, %x2 |
| 238 | %t2 = fadd double %t1, %x3 |
| 239 | ret double %t2 |
| 240 | } |
| 241 | |
| 242 | define double @already_reassociated() { |
| 243 | ; CHECK-LABEL: already_reassociated: |
| 244 | ; CHECK: bl bar |
| 245 | ; CHECK-NEXT: mov v8.16b, v0.16b |
| 246 | ; CHECK-NEXT: bl bar |
| 247 | ; CHECK-NEXT: mov v9.16b, v0.16b |
| 248 | ; CHECK-NEXT: bl bar |
| 249 | ; CHECK-NEXT: mov v10.16b, v0.16b |
| 250 | ; CHECK-NEXT: bl bar |
| 251 | ; CHECK: fadd d1, d8, d9 |
| 252 | ; CHECK-NEXT: fadd d0, d10, d0 |
| 253 | ; CHECK-NEXT: fadd d0, d1, d0 |
| 254 | %x0 = call double @bar() |
| 255 | %x1 = call double @bar() |
| 256 | %x2 = call double @bar() |
| 257 | %x3 = call double @bar() |
| 258 | %t0 = fadd double %x0, %x1 |
| 259 | %t1 = fadd double %x2, %x3 |
| 260 | %t2 = fadd double %t0, %t1 |
| 261 | ret double %t2 |
| 262 | } |
| 263 | |