| 1 |
; RUN: opt < %s -passes=loop-interchange -cache-line-size=64 -pass-remarks-output=%t -disable-output \ |
| 2 |
; RUN: -verify-dom-info -verify-loop-info -verify-loop-lcssa |
| 3 |
; RUN: FileCheck -input-file=%t %s |
| 4 |
|
| 5 |
; Check that interchanging the loops is legal for the bitwise-or reduction. |
| 6 |
; |
| 7 |
; int b_or = 0; |
| 8 |
; for (int i = 0; i < 2; i++) |
| 9 |
; for (int j = 0; j < 2; j++) |
| 10 |
; b_or |= A[j][i]; |
| 11 |
|
| 12 |
; CHECK: --- !Pass |
| 13 |
; CHECK-NEXT: Pass: loop-interchange |
| 14 |
; CHECK-NEXT: Name: Interchanged |
| 15 |
; CHECK-NEXT: Function: reduction_or |
| 16 |
define void @reduction_or(ptr %A) { |
| 17 |
entry: |
| 18 |
br label %for.i.header |
| 19 |
|
| 20 |
for.i.header: |
| 21 |
%i = phi i32 [ 0, %entry ], [ %i.inc, %for.i.latch ] |
| 22 |
%or.i = phi i32 [ 0, %entry ], [ %or.i.lcssa, %for.i.latch ] |
| 23 |
br label %for.j |
| 24 |
|
| 25 |
for.j: |
| 26 |
%j = phi i32 [ 0, %for.i.header ], [ %j.inc, %for.j ] |
| 27 |
%or.j = phi i32 [ %or.i, %for.i.header ], [ %or.j.next, %for.j ] |
| 28 |
%idx = getelementptr inbounds [2 x [2 x i32]], ptr %A, i32 0, i32 %j, i32 %i |
| 29 |
%a = load i32, ptr %idx, align 4 |
| 30 |
%or.j.next = or i32 %or.j, %a |
| 31 |
%j.inc = add i32 %j, 1 |
| 32 |
%cmp.j = icmp slt i32 %j.inc, 2 |
| 33 |
br i1 %cmp.j, label %for.j, label %for.i.latch |
| 34 |
|
| 35 |
for.i.latch: |
| 36 |
%or.i.lcssa = phi i32 [ %or.j.next, %for.j ] |
| 37 |
%i.inc = add i32 %i, 1 |
| 38 |
%cmp.i = icmp slt i32 %i.inc, 2 |
| 39 |
br i1 %cmp.i, label %for.i.header, label %exit |
| 40 |
|
| 41 |
exit: |
| 42 |
ret void |
| 43 |
} |
| 44 |
|
| 45 |
|
| 46 |
; Check that interchanging the loops is legal for the bitwise-and reduction. |
| 47 |
; |
| 48 |
; int b_and = -1; |
| 49 |
; for (int i = 0; i < 2; i++) |
| 50 |
; for (int j = 0; j < 2; j++) |
| 51 |
; b_and &= A[j][i]; |
| 52 |
|
| 53 |
; CHECK: --- !Pass |
| 54 |
; CHECK-NEXT: Pass: loop-interchange |
| 55 |
; CHECK-NEXT: Name: Interchanged |
| 56 |
; CHECK-NEXT: Function: reduction_and |
| 57 |
define void @reduction_and(ptr %A) { |
| 58 |
entry: |
| 59 |
br label %for.i.header |
| 60 |
|
| 61 |
for.i.header: |
| 62 |
%i = phi i32 [ 0, %entry ], [ %i.inc, %for.i.latch ] |
| 63 |
%and.i = phi i32 [ -1, %entry ], [ %and.i.lcssa, %for.i.latch ] |
| 64 |
br label %for.j |
| 65 |
|
| 66 |
for.j: |
| 67 |
%j = phi i32 [ 0, %for.i.header ], [ %j.inc, %for.j ] |
| 68 |
%and.j = phi i32 [ %and.i, %for.i.header ], [ %and.j.next, %for.j ] |
| 69 |
%idx = getelementptr inbounds [2 x [2 x i32]], ptr %A, i32 0, i32 %j, i32 %i |
| 70 |
%a = load i32, ptr %idx, align 4 |
| 71 |
%and.j.next = and i32 %and.j, %a |
| 72 |
%j.inc = add i32 %j, 1 |
| 73 |
%cmp.j = icmp slt i32 %j.inc, 2 |
| 74 |
br i1 %cmp.j, label %for.j, label %for.i.latch |
| 75 |
|
| 76 |
for.i.latch: |
| 77 |
%and.i.lcssa = phi i32 [ %and.j.next, %for.j ] |
| 78 |
%i.inc = add i32 %i, 1 |
| 79 |
%cmp.i = icmp slt i32 %i.inc, 2 |
| 80 |
br i1 %cmp.i, label %for.i.header, label %exit |
| 81 |
|
| 82 |
exit: |
| 83 |
ret void |
| 84 |
} |
| 85 |
|
| 86 |
|
| 87 |
; Check that interchanging the loops is legal for the bitwise-xor reduction. |
| 88 |
; |
| 89 |
; int b_xor = 0; |
| 90 |
; for (int i = 0; i < 2; i++) |
| 91 |
; for (int j = 0; j < 2; j++) |
| 92 |
; b_xor ^= A[j][i]; |
| 93 |
|
| 94 |
; CHECK: --- !Pass |
| 95 |
; CHECK-NEXT: Pass: loop-interchange |
| 96 |
; CHECK-NEXT: Name: Interchanged |
| 97 |
; CHECK-NEXT: Function: reduction_xor |
| 98 |
define void @reduction_xor(ptr %A) { |
| 99 |
entry: |
| 100 |
br label %for.i.header |
| 101 |
|
| 102 |
for.i.header: |
| 103 |
%i = phi i32 [ 0, %entry ], [ %i.inc, %for.i.latch ] |
| 104 |
%xor.i = phi i32 [ 0, %entry ], [ %xor.i.lcssa, %for.i.latch ] |
| 105 |
br label %for.j |
| 106 |
|
| 107 |
for.j: |
| 108 |
%j = phi i32 [ 0, %for.i.header ], [ %j.inc, %for.j ] |
| 109 |
%xor.j = phi i32 [ %xor.i, %for.i.header ], [ %xor.j.next, %for.j ] |
| 110 |
%idx = getelementptr inbounds [2 x [2 x i32]], ptr %A, i32 0, i32 %j, i32 %i |
| 111 |
%a = load i32, ptr %idx, align 4 |
| 112 |
%xor.j.next = xor i32 %xor.j, %a |
| 113 |
%j.inc = add i32 %j, 1 |
| 114 |
%cmp.j = icmp slt i32 %j.inc, 2 |
| 115 |
br i1 %cmp.j, label %for.j, label %for.i.latch |
| 116 |
|
| 117 |
for.i.latch: |
| 118 |
%xor.i.lcssa = phi i32 [ %xor.j.next, %for.j ] |
| 119 |
%i.inc = add i32 %i, 1 |
| 120 |
%cmp.i = icmp slt i32 %i.inc, 2 |
| 121 |
br i1 %cmp.i, label %for.i.header, label %exit |
| 122 |
|
| 123 |
exit: |
| 124 |
ret void |
| 125 |
} |
| 126 |
|
| 127 |
|
| 128 |
; Check that interchanging the loops is legal for the signed-minimum reduction. |
| 129 |
; |
| 130 |
; int smin = init; |
| 131 |
; for (int i = 0; i < 2; i++) |
| 132 |
; for (int j = 0; j < 2; j++) |
| 133 |
; smin = (A[j][i] < smin) ? A[j][i] : smin; |
| 134 |
|
| 135 |
; CHECK: --- !Pass |
| 136 |
; CHECK-NEXT: Pass: loop-interchange |
| 137 |
; CHECK-NEXT: Name: Interchanged |
| 138 |
; CHECK-NEXT: Function: reduction_smin |
| 139 |
define void @reduction_smin(ptr %A, i32 %init) { |
| 140 |
entry: |
| 141 |
br label %for.i.header |
| 142 |
|
| 143 |
for.i.header: |
| 144 |
%i = phi i32 [ 0, %entry ], [ %i.inc, %for.i.latch ] |
| 145 |
%smin.i = phi i32 [ %init, %entry ], [ %smin.i.lcssa, %for.i.latch ] |
| 146 |
br label %for.j |
| 147 |
|
| 148 |
for.j: |
| 149 |
%j = phi i32 [ 0, %for.i.header ], [ %j.inc, %for.j ] |
| 150 |
%smin.j = phi i32 [ %smin.i, %for.i.header ], [ %smin.j.next, %for.j ] |
| 151 |
%idx = getelementptr inbounds [2 x [2 x i32]], ptr %A, i32 0, i32 %j, i32 %i |
| 152 |
%a = load i32, ptr %idx, align 4 |
| 153 |
%cmp = icmp slt i32 %a, %smin.j |
| 154 |
%smin.j.next = select i1 %cmp, i32 %a, i32 %smin.j |
| 155 |
%j.inc = add i32 %j, 1 |
| 156 |
%cmp.j = icmp slt i32 %j.inc, 2 |
| 157 |
br i1 %cmp.j, label %for.j, label %for.i.latch |
| 158 |
|
| 159 |
for.i.latch: |
| 160 |
%smin.i.lcssa = phi i32 [ %smin.j.next, %for.j ] |
| 161 |
%i.inc = add i32 %i, 1 |
| 162 |
%cmp.i = icmp slt i32 %i.inc, 2 |
| 163 |
br i1 %cmp.i, label %for.i.header, label %exit |
| 164 |
|
| 165 |
exit: |
| 166 |
ret void |
| 167 |
} |
| 168 |
|
| 169 |
|
| 170 |
; Check that interchanging the loops is legal for the signed-maximum reduction. |
| 171 |
; |
| 172 |
; int smax = init; |
| 173 |
; for (int i = 0; i < 2; i++) |
| 174 |
; for (int j = 0; j < 2; j++) |
| 175 |
; smax = (A[j][i] > smax) ? A[j][i] : smax; |
| 176 |
|
| 177 |
; CHECK: --- !Pass |
| 178 |
; CHECK-NEXT: Pass: loop-interchange |
| 179 |
; CHECK-NEXT: Name: Interchanged |
| 180 |
; CHECK-NEXT: Function: reduction_smax |
| 181 |
define void @reduction_smax(ptr %A, i32 %init) { |
| 182 |
entry: |
| 183 |
br label %for.i.header |
| 184 |
|
| 185 |
for.i.header: |
| 186 |
%i = phi i32 [ 0, %entry ], [ %i.inc, %for.i.latch ] |
| 187 |
%smax.i = phi i32 [ %init, %entry ], [ %smax.i.lcssa, %for.i.latch ] |
| 188 |
br label %for.j |
| 189 |
|
| 190 |
for.j: |
| 191 |
%j = phi i32 [ 0, %for.i.header ], [ %j.inc, %for.j ] |
| 192 |
%smax.j = phi i32 [ %smax.i, %for.i.header ], [ %smax.j.next, %for.j ] |
| 193 |
%idx = getelementptr inbounds [2 x [2 x i32]], ptr %A, i32 0, i32 %j, i32 %i |
| 194 |
%a = load i32, ptr %idx, align 4 |
| 195 |
%cmp = icmp sgt i32 %a, %smax.j |
| 196 |
%smax.j.next = select i1 %cmp, i32 %a, i32 %smax.j |
| 197 |
%j.inc = add i32 %j, 1 |
| 198 |
%cmp.j = icmp slt i32 %j.inc, 2 |
| 199 |
br i1 %cmp.j, label %for.j, label %for.i.latch |
| 200 |
|
| 201 |
for.i.latch: |
| 202 |
%smax.i.lcssa = phi i32 [ %smax.j.next, %for.j ] |
| 203 |
%i.inc = add i32 %i, 1 |
| 204 |
%cmp.i = icmp slt i32 %i.inc, 2 |
| 205 |
br i1 %cmp.i, label %for.i.header, label %exit |
| 206 |
|
| 207 |
exit: |
| 208 |
ret void |
| 209 |
} |
| 210 |
|
| 211 |
|
| 212 |
; Check that interchanging the loops is legal for the unsigned-minimum reduction. |
| 213 |
; |
| 214 |
; unsigned umin = init; |
| 215 |
; for (int i = 0; i < 2; i++) |
| 216 |
; for (int j = 0; j < 2; j++) |
| 217 |
; umin = (A[j][i] < umin) ? A[j][i] : umin; |
| 218 |
|
| 219 |
; CHECK: --- !Pass |
| 220 |
; CHECK-NEXT: Pass: loop-interchange |
| 221 |
; CHECK-NEXT: Name: Interchanged |
| 222 |
; CHECK-NEXT: Function: reduction_umin |
| 223 |
define void @reduction_umin(ptr %A, i32 %init) { |
| 224 |
entry: |
| 225 |
br label %for.i.header |
| 226 |
|
| 227 |
for.i.header: |
| 228 |
%i = phi i32 [ 0, %entry ], [ %i.inc, %for.i.latch ] |
| 229 |
%umin.i = phi i32 [ %init, %entry ], [ %umin.i.lcssa, %for.i.latch ] |
| 230 |
br label %for.j |
| 231 |
|
| 232 |
for.j: |
| 233 |
%j = phi i32 [ 0, %for.i.header ], [ %j.inc, %for.j ] |
| 234 |
%umin.j = phi i32 [ %umin.i, %for.i.header ], [ %umin.j.next, %for.j ] |
| 235 |
%idx = getelementptr inbounds [2 x [2 x i32]], ptr %A, i32 0, i32 %j, i32 %i |
| 236 |
%a = load i32, ptr %idx, align 4 |
| 237 |
%cmp = icmp ult i32 %a, %umin.j |
| 238 |
%umin.j.next = select i1 %cmp, i32 %a, i32 %umin.j |
| 239 |
%j.inc = add i32 %j, 1 |
| 240 |
%cmp.j = icmp slt i32 %j.inc, 2 |
| 241 |
br i1 %cmp.j, label %for.j, label %for.i.latch |
| 242 |
|
| 243 |
for.i.latch: |
| 244 |
%umin.i.lcssa = phi i32 [ %umin.j.next, %for.j ] |
| 245 |
%i.inc = add i32 %i, 1 |
| 246 |
%cmp.i = icmp slt i32 %i.inc, 2 |
| 247 |
br i1 %cmp.i, label %for.i.header, label %exit |
| 248 |
|
| 249 |
exit: |
| 250 |
ret void |
| 251 |
} |
| 252 |
|
| 253 |
|
| 254 |
; Check that interchanging the loops is legal for the unsigned-maximum reduction. |
| 255 |
; |
| 256 |
; unsigned umax = 0; |
| 257 |
; for (int i = 0; i < 2; i++) |
| 258 |
; for (int j = 0; j < 2; j++) |
| 259 |
; smax = (A[j][i] > smax) ? A[j][i] : smax; |
| 260 |
|
| 261 |
; CHECK: --- !Pass |
| 262 |
; CHECK-NEXT: Pass: loop-interchange |
| 263 |
; CHECK-NEXT: Name: Interchanged |
| 264 |
; CHECK-NEXT: Function: reduction_umax |
| 265 |
define void @reduction_umax(ptr %A) { |
| 266 |
entry: |
| 267 |
br label %for.i.header |
| 268 |
|
| 269 |
for.i.header: |
| 270 |
%i = phi i32 [ 0, %entry ], [ %i.inc, %for.i.latch ] |
| 271 |
%umax.i = phi i32 [ 0, %entry ], [ %umax.i.lcssa, %for.i.latch ] |
| 272 |
br label %for.j |
| 273 |
|
| 274 |
for.j: |
| 275 |
%j = phi i32 [ 0, %for.i.header ], [ %j.inc, %for.j ] |
| 276 |
%umax.j = phi i32 [ %umax.i, %for.i.header ], [ %umax.j.next, %for.j ] |
| 277 |
%idx = getelementptr inbounds [2 x [2 x i32]], ptr %A, i32 0, i32 %j, i32 %i |
| 278 |
%a = load i32, ptr %idx, align 4 |
| 279 |
%cmp = icmp ugt i32 %a, %umax.j |
| 280 |
%umax.j.next = select i1 %cmp, i32 %a, i32 %umax.j |
| 281 |
%j.inc = add i32 %j, 1 |
| 282 |
%cmp.j = icmp slt i32 %j.inc, 2 |
| 283 |
br i1 %cmp.j, label %for.j, label %for.i.latch |
| 284 |
|
| 285 |
for.i.latch: |
| 286 |
%umax.i.lcssa = phi i32 [ %umax.j.next, %for.j ] |
| 287 |
%i.inc = add i32 %i, 1 |
| 288 |
%cmp.i = icmp slt i32 %i.inc, 2 |
| 289 |
br i1 %cmp.i, label %for.i.header, label %exit |
| 290 |
|
| 291 |
exit: |
| 292 |
ret void |
| 293 |
} |
| 294 |
|
| 295 |
|
| 296 |
; Check that interchanging the loops is legal for the any-of reduction. |
| 297 |
; |
| 298 |
; int any_of = 0; |
| 299 |
; for (int i = 0; i < 2; i++) |
| 300 |
; for (int j = 0; j < 2; j++) |
| 301 |
; any_of = (A[j][i] == 42) ? 1 : any_of; |
| 302 |
|
| 303 |
; CHECK: --- !Pass |
| 304 |
; CHECK-NEXT: Pass: loop-interchange |
| 305 |
; CHECK-NEXT: Name: Interchanged |
| 306 |
; CHECK-NEXT: Function: reduction_anyof |
| 307 |
define void @reduction_anyof(ptr %A) { |
| 308 |
entry: |
| 309 |
br label %for.i.header |
| 310 |
|
| 311 |
for.i.header: |
| 312 |
%i = phi i32 [ 0, %entry ], [ %i.inc, %for.i.latch ] |
| 313 |
%anyof.i = phi i32 [ 0, %entry ], [ %anyof.i.lcssa, %for.i.latch ] |
| 314 |
br label %for.j |
| 315 |
|
| 316 |
for.j: |
| 317 |
%j = phi i32 [ 0, %for.i.header ], [ %j.inc, %for.j ] |
| 318 |
%anyof.j = phi i32 [ %anyof.i, %for.i.header ], [ %anyof.j.next, %for.j ] |
| 319 |
%idx = getelementptr inbounds [2 x [2 x i32]], ptr %A, i32 0, i32 %j, i32 %i |
| 320 |
%a = load i32, ptr %idx, align 4 |
| 321 |
%cmp = icmp eq i32 %a, 42 |
| 322 |
%anyof.j.next = select i1 %cmp, i32 1, i32 %anyof.j |
| 323 |
%j.inc = add i32 %j, 1 |
| 324 |
%cmp.j = icmp slt i32 %j.inc, 2 |
| 325 |
br i1 %cmp.j, label %for.j, label %for.i.latch |
| 326 |
|
| 327 |
for.i.latch: |
| 328 |
%anyof.i.lcssa = phi i32 [ %anyof.j.next, %for.j ] |
| 329 |
%i.inc = add i32 %i, 1 |
| 330 |
%cmp.i = icmp slt i32 %i.inc, 2 |
| 331 |
br i1 %cmp.i, label %for.i.header, label %exit |
| 332 |
|
| 333 |
exit: |
| 334 |
ret void |
| 335 |
} |