| 1 |
// Test sequences that can use RISBG with a zeroed first operand. |
| 2 |
// The tests here assume that RISBLG isn't available. |
| 3 |
|
| 4 |
/* Tests ported from the Llvm testsuite. */ |
| 5 |
|
| 6 |
/* { dg-do compile { target s390x-*-* } } */ |
| 7 |
/* { dg-options "-O3 -march=z10 -mzarch -fno-asynchronous-unwind-tables" } */ |
| 8 |
|
| 9 |
#define i64 signed long long |
| 10 |
#define ui64 unsigned long long |
| 11 |
#define i32 signed int |
| 12 |
#define ui32 unsigned int |
| 13 |
#define i8 signed char |
| 14 |
#define ui8 unsigned char |
| 15 |
|
| 16 |
// Test an extraction of bit 0 from a right-shifted value. |
| 17 |
i32 f1 (i32 v_foo) |
| 18 |
{ |
| 19 |
/* { dg-final { scan-assembler "f1:\n\trisbg\t%r2,%r2,64-1,128\\\+63,53\\\+1" } } */ |
| 20 |
i32 v_shr = ((ui32)v_foo) >> 10; |
| 21 |
i32 v_and = v_shr & 1; |
| 22 |
return v_and; |
| 23 |
} |
| 24 |
|
| 25 |
// ...and again with i64. |
| 26 |
i64 f2 (i64 v_foo) |
| 27 |
{ |
| 28 |
/* { dg-final { scan-assembler "f2:\n\trisbg\t%r2,%r2,64-1,128\\\+63,53\\\+1" { target { lp64 } } } } */ |
| 29 |
/* { dg-final { scan-assembler "f2:\n\trisbg\t%r3,%r3,64-1,128\\\+63,53\\\+1\n\tlhi\t%r2,0" { target { ! lp64 } } } } */ |
| 30 |
i64 v_shr = ((ui64)v_foo) >> 10; |
| 31 |
i64 v_and = v_shr & 1; |
| 32 |
return v_and; |
| 33 |
} |
| 34 |
|
| 35 |
// Test an extraction of other bits from a right-shifted value. |
| 36 |
i32 f3 (i32 v_foo) |
| 37 |
{ |
| 38 |
/* { dg-final { scan-assembler "f3:\n\trisbg\t%r2,%r2,60,128\\\+61,64-22" } } */ |
| 39 |
i32 v_shr = ((ui32)v_foo) >> 22; |
| 40 |
i32 v_and = v_shr & 12; |
| 41 |
return v_and; |
| 42 |
} |
| 43 |
|
| 44 |
// ...and again with i64. |
| 45 |
i64 f4 (i64 v_foo) |
| 46 |
{ |
| 47 |
/* { dg-final { scan-assembler "f4:\n\trisbg\t%r2,%r2,60,128\\\+61,64-22" { target { lp64 } } } } */ |
| 48 |
/* { dg-final { scan-assembler "f4:\n\trisbg\t%r3,%r3,60,128\\\+61,64-22\n\tlhi\t%r2,0" { target { ! lp64 } } } } */ |
| 49 |
i64 v_shr = ((ui64)v_foo) >> 22; |
| 50 |
i64 v_and = v_shr & 12; |
| 51 |
return v_and; |
| 52 |
} |
| 53 |
|
| 54 |
// Test an extraction of most bits from a right-shifted value. |
| 55 |
// The range should be reduced to exclude the zeroed high bits. |
| 56 |
i32 f5 (i32 v_foo) |
| 57 |
{ |
| 58 |
/* { dg-final { scan-assembler "f5:\n\trisbg\t%r2,%r2,34,128\\\+60,64-2" } } */ |
| 59 |
i32 v_shr = ((ui32)v_foo) >> 2; |
| 60 |
i32 v_and = v_shr & -8; |
| 61 |
return v_and; |
| 62 |
} |
| 63 |
|
| 64 |
// ...and again with i64. |
| 65 |
i64 f6 (i64 v_foo) |
| 66 |
{ |
| 67 |
/* { dg-final { scan-assembler "f6:\n\trisbg\t%r2,%r2,2,128\\\+60,64-2" { target { lp64 } } } } */ |
| 68 |
/* { dg-final { scan-assembler "f6:\n\trisbg\t%r3,%r2,0,0\\\+32-1,64-0-32\n\trisbg\t%r2,%r3,2,128\\\+60,64-2" { target { ! lp64 } } } } */ |
| 69 |
i64 v_shr = ((ui64)v_foo) >> 2; |
| 70 |
i64 v_and = v_shr & -8; |
| 71 |
return v_and; |
| 72 |
} |
| 73 |
|
| 74 |
// Try the next value up (mask ....1111001). This needs a separate shift |
| 75 |
// and mask. |
| 76 |
i32 f7 (i32 v_foo) |
| 77 |
{ |
| 78 |
/* Should be |
| 79 |
{ dg-final { scan-assembler "f7:\n\tsrl\t%r2,2\n\tnill\t%r2,65529" { xfail { lp64 } } } } |
| 80 |
but because a zeroextend is merged into the pattern it is actually |
| 81 |
{ dg-final { scan-assembler "f7:\n\tsrl\t%r2,2\n\tlgfi\t%r1,1073741817\n\tngr\t%r2,%r1" { target { lp64 } } } } |
| 82 |
{ dg-final { scan-assembler "f7:\n\tsrl\t%r2,2\n\tnill\t%r2,65529" { target { ! lp64 } } } } */ |
| 83 |
i32 v_shr = ((ui32)v_foo) >> 2; |
| 84 |
i32 v_and = v_shr & -7; |
| 85 |
return v_and; |
| 86 |
} |
| 87 |
|
| 88 |
// ...and again with i64. |
| 89 |
i64 f8 (i64 v_foo) |
| 90 |
{ |
| 91 |
/* { dg-final { scan-assembler "f8:\n\tsrlg\t%r2,%r2,2\n\tnill\t%r2,65529" { target { lp64 } } } } */ |
| 92 |
/* { dg-final { scan-assembler "f8:\n\trisbg\t%r3,%r2,0,0\\\+32-1,64-0-32\n\tsrlg\t%r2,%r3,2\n\tnill\t%r2,65529" { target { ! lp64 } } } } */ |
| 93 |
i64 v_shr = ((ui64)v_foo) >> 2; |
| 94 |
i64 v_and = v_shr & -7; |
| 95 |
return v_and; |
| 96 |
} |
| 97 |
|
| 98 |
// Test an extraction of bits from a left-shifted value. The range should |
| 99 |
// be reduced to exclude the zeroed low bits. |
| 100 |
i32 f9 (i32 v_foo) |
| 101 |
{ |
| 102 |
/* { dg-final { scan-assembler "f9:\n\trisbg\t%r2,%r2,56,128\\\+61,2" } } */ |
| 103 |
i32 v_shr = v_foo << 2; |
| 104 |
i32 v_and = v_shr & 255; |
| 105 |
return v_and; |
| 106 |
} |
| 107 |
|
| 108 |
// ...and again with i64. |
| 109 |
i64 f10 (i64 v_foo) |
| 110 |
{ |
| 111 |
/* { dg-final { scan-assembler "f10:\n\trisbg\t%r2,%r2,56,128\\\+61,2" { target { lp64 } } } } */ |
| 112 |
/* { dg-final { scan-assembler "f10:\n\trisbg\t%r3,%r3,56,128\\\+61,2\n\tlhi\t%r2,0" { target { ! lp64 } } } } */ |
| 113 |
i64 v_shr = v_foo << 2; |
| 114 |
i64 v_and = v_shr & 255; |
| 115 |
return v_and; |
| 116 |
} |
| 117 |
|
| 118 |
// Try a wrap-around mask (mask ....111100001111). This needs a separate shift |
| 119 |
// and mask. |
| 120 |
i32 f11 (i32 v_foo) |
| 121 |
{ |
| 122 |
/* { dg-final { scan-assembler "f11:\n\tsll\t%r2,2\n\tnill\t%r2,65295" } } */ |
| 123 |
i32 v_shr = v_foo << 2; |
| 124 |
i32 v_and = v_shr & -241; |
| 125 |
return v_and; |
| 126 |
} |
| 127 |
|
| 128 |
// ...and again with i64. |
| 129 |
i64 f12 (i64 v_foo) |
| 130 |
{ |
| 131 |
/* { dg-final { scan-assembler "f12:\n\tsllg\t%r2,%r2,2\n\tnill\t%r2,65295" { target { lp64 } } } } */ |
| 132 |
/* { dg-final { scan-assembler "f12:\n\trisbg\t%r3,%r2,0,0\\\+32-1,64-0-32\n\tsllg\t%r2,%r3,2\n\tnill\t%r2,65295" { target { ! lp64 } } } } */ |
| 133 |
i64 v_shr = v_foo << 2; |
| 134 |
i64 v_and = v_shr & -241; |
| 135 |
return v_and; |
| 136 |
} |
| 137 |
|
| 138 |
// Test an extraction from a rotated value, no mask wraparound. |
| 139 |
// This is equivalent to the lshr case, because the bits from the |
| 140 |
// shl are not used. |
| 141 |
i32 f13 (i32 v_foo) |
| 142 |
{ |
| 143 |
/* { dg-final { scan-assembler "f13:\n\trisbg\t%r2,%r2,56,128\\\+60,32\\\+14" { target { lp64 } } } } */ |
| 144 |
/* { dg-final { scan-assembler "f13:\n\trll\t%r2,%r2,14\n\tnilf\t%r2,248" { target { ! lp64 } } } } */ |
| 145 |
i32 v_parta = v_foo << 14; |
| 146 |
i32 v_partb = ((ui32)v_foo) >> 18; |
| 147 |
i32 v_rotl = v_parta | v_partb; |
| 148 |
i32 v_and = v_rotl & 248; |
| 149 |
return v_and; |
| 150 |
} |
| 151 |
|
| 152 |
// ...and again with i64. |
| 153 |
i64 f14 (i64 v_foo) |
| 154 |
{ |
| 155 |
/* { dg-final { scan-assembler "f14:\n\trisbg\t%r2,%r2,56,128\\\+60,14" { target { lp64 } } } } */ |
| 156 |
/* { dg-final { scan-assembler "f14:\n\trisbg\t%r3,%r2,56,128\\\+60,46\n\tlhi\t%r2,0" { target { ! lp64 } } } } */ |
| 157 |
i64 v_parta = v_foo << 14; |
| 158 |
i64 v_partb = ((ui64)v_foo) >> 50; |
| 159 |
i64 v_rotl = v_parta | v_partb; |
| 160 |
i64 v_and = v_rotl & 248; |
| 161 |
return v_and; |
| 162 |
} |
| 163 |
|
| 164 |
// Try a case in which only the bits from the shl are used. |
| 165 |
i32 f15 (i32 v_foo) |
| 166 |
{ |
| 167 |
/* { dg-final { scan-assembler "f15:\n\trisbg\t%r2,%r2,47,128\\\+49,14" { target { lp64 } } } } */ |
| 168 |
/* { dg-final { scan-assembler "f15:\n\trll\t%r2,%r2,14\n\tnilf\t%r2,114688" { target { ! lp64 } } } } */ |
| 169 |
i32 v_parta = v_foo << 14; |
| 170 |
i32 v_partb = ((ui32)v_foo) >> 18; |
| 171 |
i32 v_rotl = v_parta | v_partb; |
| 172 |
i32 v_and = v_rotl & 114688; |
| 173 |
return v_and; |
| 174 |
} |
| 175 |
|
| 176 |
// ...and again with i64. |
| 177 |
i64 f16 (i64 v_foo) |
| 178 |
{ |
| 179 |
/* { dg-final { scan-assembler "f16:\n\trisbg\t%r2,%r2,47,128\\\+49,14" { target { lp64 } } } } */ |
| 180 |
/* { dg-final { scan-assembler "f16:\n\trisbg\t%r3,%r3,47,128\\\+49,14\n\tlhi\t%r2,0" { target { ! lp64 } } } } */ |
| 181 |
i64 v_parta = v_foo << 14; |
| 182 |
i64 v_partb = ((ui64)v_foo) >> 50; |
| 183 |
i64 v_rotl = v_parta | v_partb; |
| 184 |
i64 v_and = v_rotl & 114688; |
| 185 |
return v_and; |
| 186 |
} |
| 187 |
|
| 188 |
// Test a 32-bit rotate in which both parts of the OR are needed. |
| 189 |
// This needs a separate shift and mask. |
| 190 |
i32 f17 (i32 v_foo) |
| 191 |
{ |
| 192 |
/* Should be |
| 193 |
{ dg-final { scan-assembler "f17:\n\trll\t%r2,%r2,4\n\tnilf\t%r2,126" { xfail { lp64 } } } } |
| 194 |
but because a zeroextend is merged into the pattern it is actually |
| 195 |
{ dg-final { scan-assembler "f17:\n\trll\t%r2,%r2,4\n\trisbg\t%r2,%r2,57,128\\\+62,0" { target { lp64 } } } } |
| 196 |
{ dg-final { scan-assembler "f17:\n\trll\t%r2,%r2,4\n\tnilf\t%r2,126" { target { ! lp64 } } } } */ |
| 197 |
i32 v_parta = v_foo << 4; |
| 198 |
i32 v_partb = ((ui32)v_foo) >> 28; |
| 199 |
i32 v_rotl = v_parta | v_partb; |
| 200 |
i32 v_and = v_rotl & 126; |
| 201 |
return v_and; |
| 202 |
} |
| 203 |
|
| 204 |
// ...and for i64, where RISBG should do the rotate too. |
| 205 |
i64 f18 (i64 v_foo) |
| 206 |
{ |
| 207 |
/* { dg-final { scan-assembler "f18:\n\trisbg\t%r2,%r2,57,128\\\+62,4" { target { lp64 } } } } */ |
| 208 |
/* { dg-final { scan-assembler "f18:\n\trisbg\t%r3,%r2,0,0\\\+32-1,64-0-32\n\tlhi\t%r2,0\n\trisbg\t%r3,%r3,57,128\\\+62,4" { target { ! lp64 } } } } */ |
| 209 |
i64 v_parta = v_foo << 4; |
| 210 |
i64 v_partb = ((ui64)v_foo) >> 60; |
| 211 |
i64 v_rotl = v_parta | v_partb; |
| 212 |
i64 v_and = v_rotl & 126; |
| 213 |
return v_and; |
| 214 |
} |
| 215 |
|
| 216 |
// Test an arithmetic shift right in which some of the sign bits are kept. |
| 217 |
// This needs a separate shift and mask on 31 bit. |
| 218 |
i32 f19 (i32 v_foo) |
| 219 |
{ |
| 220 |
/* { dg-final { scan-assembler "f19:\n\trisbg\t%r2,%r2,59,128\\+62,64-28" { target { lp64 } } } } */ |
| 221 |
/* { dg-final { scan-assembler "f19:\n\tsra\t%r2,28\n\tnilf\t%r2,30" { target { ! lp64 } } } } */ |
| 222 |
i32 v_shr = v_foo >> 28; |
| 223 |
i32 v_and = v_shr & 30; |
| 224 |
return v_and; |
| 225 |
} |
| 226 |
|
| 227 |
// ...and again with i64. In this case RISBG is the best way of doing the AND. |
| 228 |
i64 f20 (i64 v_foo) |
| 229 |
{ |
| 230 |
/* { dg-final { scan-assembler "f20:\n\tsrag\t%r2,%r2,60\n\trisbg\t%r2,%r2,59,128\\\+62,0" { target { lp64 } } } } */ |
| 231 |
/* { dg-final { scan-assembler "f20:\n\trisbg\t%r3,%r2,0,0\\\+32-1,64-0-32\n\tlhi\t%r2,0\n\tsrag\t%r3,%r3,60\n\tnilf\t%r3,30" { target { ! lp64 } } } } */ |
| 232 |
i64 v_shr = v_foo >> 60; |
| 233 |
i64 v_and = v_shr & 30; |
| 234 |
return v_and; |
| 235 |
} |
| 236 |
|
| 237 |
// Now try an arithmetic right shift in which the sign bits aren't needed. |
| 238 |
// Note: Unlike Llvm, Gcc replaces the ashrt with a lshrt in any case, using |
| 239 |
// a risbg pattern without ashrt. |
| 240 |
i32 f21 (i32 v_foo) |
| 241 |
{ |
| 242 |
/* { dg-final { scan-assembler "f21:\n\trisbg\t%r2,%r2,60,128\\\+62,64-28" } } */ |
| 243 |
i32 v_shr = v_foo >> 28; |
| 244 |
i32 v_and = v_shr & 14; |
| 245 |
return v_and; |
| 246 |
} |
| 247 |
|
| 248 |
// ...and again with i64. |
| 249 |
i64 f22 (i64 v_foo) |
| 250 |
{ |
| 251 |
/* { dg-final { scan-assembler "f22:\n\trisbg\t%r2,%r2,60,128\\\+62,64-60" { target { lp64 } } } } */ |
| 252 |
/* { dg-final { scan-assembler "f22:\n\trisbg\t%r3,%r2,60,128\\\+62,64-28\n\tlhi\t%r2,0" { target { ! lp64 } } } } */ |
| 253 |
i64 v_shr = v_foo >> 60; |
| 254 |
i64 v_and = v_shr & 14; |
| 255 |
return v_and; |
| 256 |
} |
| 257 |
|
| 258 |
// Check that we use RISBG for shifted values even if the AND is a |
| 259 |
// natural zero extension. |
| 260 |
i64 f23 (i64 v_foo) |
| 261 |
{ |
| 262 |
/* { dg-final { scan-assembler "f23:\n\trisbg\t%r2,%r2,64-8,128\\\+63,54\\\+8" { target { lp64 } } } } */ |
| 263 |
/* { dg-final { scan-assembler "f23:\n\trisbg\t%r3,%r3,64-8,128\\\+63,54\\\+8\n\tlhi\t%r2,0" { target { ! lp64 } } } } */ |
| 264 |
i64 v_shr = ((ui64)v_foo) >> 2; |
| 265 |
i64 v_and = v_shr & 255; |
| 266 |
return v_and; |
| 267 |
} |
| 268 |
|
| 269 |
// Test a case where the AND comes before a rotate. This needs a separate |
| 270 |
// mask and rotate. |
| 271 |
i32 f24 (i32 v_foo) |
| 272 |
{ |
| 273 |
/* { dg-final { scan-assembler "f24:\n\tnilf\t%r2,254\n\trll\t%r2,%r2,29\n" } } */ |
| 274 |
i32 v_and = v_foo & 254; |
| 275 |
i32 v_parta = ((ui32)v_and) >> 3; |
| 276 |
i32 v_partb = v_and << 29; |
| 277 |
i32 v_rotl = v_parta | v_partb; |
| 278 |
return v_rotl; |
| 279 |
} |
| 280 |
|
| 281 |
// ...and again with i64, where a single RISBG is enough. |
| 282 |
i64 f25 (i64 v_foo) |
| 283 |
{ |
| 284 |
/* { dg-final { scan-assembler "f25:\n\trisbg\t%r2,%r2,57,128\\\+59,3" { target { lp64 } } } } */ |
| 285 |
/* { dg-final { scan-assembler "f25:\n\trisbg\t%r3,%r3,57,128\\\+59,3\n\tlhi\t%r2,0" { target { ! lp64 } } } } */ |
| 286 |
i64 v_and = v_foo & 14; |
| 287 |
i64 v_parta = v_and << 3; |
| 288 |
i64 v_partb = ((ui64)v_and) >> 61; |
| 289 |
i64 v_rotl = v_parta | v_partb; |
| 290 |
return v_rotl; |
| 291 |
} |
| 292 |
|
| 293 |
// Test a wrap-around case in which the AND comes before a rotate. |
| 294 |
// This again needs a separate mask and rotate. |
| 295 |
i32 f26 (i32 v_foo) |
| 296 |
{ |
| 297 |
/* { dg-final { scan-assembler "f26:\n\tnill\t%r2,65487\n\trll\t%r2,%r2,5" } } */ |
| 298 |
i32 v_and = v_foo & -49; |
| 299 |
i32 v_parta = v_and << 5; |
| 300 |
i32 v_partb = ((ui32)v_and) >> 27; |
| 301 |
i32 v_rotl = v_parta | v_partb; |
| 302 |
return v_rotl; |
| 303 |
} |
| 304 |
|
| 305 |
// ...and again with i64, where a single RISBG is OK. |
| 306 |
i64 f27 (i64 v_foo) |
| 307 |
{ |
| 308 |
/* { dg-final { scan-assembler "f27:\n\trisbg\t%r2,%r2,55,128\\\+52,5" { target { lp64 } } } } */ |
| 309 |
/* { dg-final { scan-assembler "f27:\n\trisbg\t%r3,%r2,0,0\\\+32-1,64-0-32\n\trisbg\t%r2,%r3,55,128\\\+52,5" { target { ! lp64 } } } } */ |
| 310 |
i64 v_and = v_foo & -49; |
| 311 |
i64 v_parta = v_and << 5; |
| 312 |
i64 v_partb = ((ui64)v_and) >> 59; |
| 313 |
i64 v_rotl = v_parta | v_partb; |
| 314 |
return v_rotl; |
| 315 |
} |
| 316 |
|
| 317 |
// Test a case where the AND comes before a shift left. |
| 318 |
i32 f28 (i32 v_foo) |
| 319 |
{ |
| 320 |
/* { dg-final { scan-assembler "f28:\n\trisbg\t%r2,%r2,32,128\\\+45,17" } } */ |
| 321 |
i32 v_and = v_foo & 32766; |
| 322 |
i32 v_shl = v_and << 17; |
| 323 |
return v_shl; |
| 324 |
} |
| 325 |
|
| 326 |
// ...and again with i64. |
| 327 |
i64 f29 (i64 v_foo) |
| 328 |
{ |
| 329 |
/* { dg-final { scan-assembler "f29:\n\trisbg\t%r2,%r2,0,128\\\+13,49" { target { lp64 } } } } */ |
| 330 |
/* { dg-final { scan-assembler "f29:\n\trisbg\t%r\[23\],%r3,0,128\\\+13,49\n\tlr\t%r\[23\],%r\[32\]\n\tsrlg\t%r2,%r2" { target { ! lp64 } } } } */ |
| 331 |
i64 v_and = v_foo & 32766; |
| 332 |
i64 v_shl = v_and << 49; |
| 333 |
return v_shl; |
| 334 |
} |
| 335 |
|
| 336 |
// Test the next shift up from f28, in which the mask should get shortened. |
| 337 |
i32 f30 (i32 v_foo) |
| 338 |
{ |
| 339 |
/* { dg-final { scan-assembler "f30:\n\trisbg\t%r2,%r2,32,128\\\+44,18" } } */ |
| 340 |
i32 v_and = v_foo & 32766; |
| 341 |
i32 v_shl = v_and << 18; |
| 342 |
return v_shl; |
| 343 |
} |
| 344 |
|
| 345 |
// ...and again with i64. |
| 346 |
i64 f31 (i64 v_foo) |
| 347 |
{ |
| 348 |
/* { dg-final { scan-assembler "f31:\n\trisbg\t%r2,%r2,0,128\\\+12,50" { target { lp64 } } } } */ |
| 349 |
/* { dg-final { scan-assembler "f31:\n\trisbg\t%r\[23\],%r3,0,128\\\+12,50\n\tlr\t%r\[23\],%r\[32\]\n\tsrlg\t%r2,%r2" { target { ! lp64 } } } } */ |
| 350 |
i64 v_and = v_foo & 32766; |
| 351 |
i64 v_shl = v_and << 50; |
| 352 |
return v_shl; |
| 353 |
} |
| 354 |
|
| 355 |
// Test a wrap-around case in which the shift left comes after the AND. |
| 356 |
// We can't use RISBG for the shift in that case. |
| 357 |
i32 f32 (i32 v_foo) |
| 358 |
{ |
| 359 |
/* { dg-final { scan-assembler "f32:\n\tsll\t%r2,10\n\tnill\t%r2,58368" } } */ |
| 360 |
i32 v_and = v_foo & -7; |
| 361 |
i32 v_shl = v_and << 10; |
| 362 |
return v_shl; |
| 363 |
} |
| 364 |
|
| 365 |
// ...and again with i64. |
| 366 |
i64 f33 (i64 v_foo) |
| 367 |
{ |
| 368 |
/* { dg-final { scan-assembler "f33:\n\tsllg\t%r2,%r2,10\n\tnill\t%r2,58368" { target { lp64 } } } } */ |
| 369 |
/* { dg-final { scan-assembler "f33:\n\trisbg\t%r3,%r2,0,0\\\+32-1,64-0-32\n\tsllg\t%r2,%r3,10\n\tnill\t%r2,58368" { target { ! lp64 } } } } */ |
| 370 |
i64 v_and = v_foo & -7; |
| 371 |
i64 v_shl = v_and << 10; |
| 372 |
return v_shl; |
| 373 |
} |
| 374 |
|
| 375 |
// Test a case where the AND comes before a shift right. |
| 376 |
i32 f34 (i32 v_foo) |
| 377 |
{ |
| 378 |
/* { dg-final { scan-assembler "f34:\n\trisbg\t%r2,%r2,64-7,128\\\+63,48\\\+7" } } */ |
| 379 |
i32 v_and = v_foo & 65535; |
| 380 |
i32 v_shl = ((ui32)v_and) >> 9; |
| 381 |
return v_shl; |
| 382 |
} |
| 383 |
|
| 384 |
// ...and again with i64. |
| 385 |
i64 f35 (i64 v_foo) |
| 386 |
{ |
| 387 |
/* { dg-final { scan-assembler "f35:\n\trisbg\t%r2,%r2,64-7,128\\\+63,48\\\+7" { target { lp64 } } } } */ |
| 388 |
/* { dg-final { scan-assembler "f35:\n\trisbg\t%r3,%r3,64-7,128\\\+63,48\\\+7\n\tlhi\t%r2,0" { target { ! lp64 } } } } */ |
| 389 |
i64 v_and = v_foo & 65535; |
| 390 |
i64 v_shl = ((ui64)v_and) >> 9; |
| 391 |
return v_shl; |
| 392 |
} |
| 393 |
|
| 394 |
// Test a wrap-around case where the AND comes before a shift right. |
| 395 |
// We can't use RISBG for the shift in that case. |
| 396 |
i32 f36 (i32 v_foo) |
| 397 |
{ |
| 398 |
/* { dg-final { scan-assembler "f36:\n\tsrl\t%r2,1\n\tlgfi\t%r1,2147483635\n\tngr\t%r2,%r1" { target { lp64 } } } } */ |
| 399 |
/* { dg-final { scan-assembler "f36:\n\tsrl\t%r2,1\n\tnilf\t%r2,2147483635" { target { ! lp64 } } } } */ |
| 400 |
i32 v_and = v_foo & -25; |
| 401 |
i32 v_shl = ((ui32)v_and) >> 1; |
| 402 |
return v_shl; |
| 403 |
} |
| 404 |
|
| 405 |
// ...and again with i64. |
| 406 |
i64 f37 (i64 v_foo) |
| 407 |
{ |
| 408 |
/* { dg-final { scan-assembler "f37:\n\(\t.*\n\)*\tsrlg\t%r2,%r2,1\n\tng\t%r2," { target { lp64 } } } } */ |
| 409 |
/* { dg-final { scan-assembler "f37:\n\(\t.*\n\)*\trisbg\t%r3,%r2,0,0\\\+32-1,64-0-32\n\tsrlg\t%r2,%r3,1\n\tng\t%r2," { target { ! lp64 } } } } */ |
| 410 |
i64 v_and = v_foo & -25; |
| 411 |
i64 v_shl = ((ui64)v_and) >> 1; |
| 412 |
return v_shl; |
| 413 |
} |
| 414 |
|
| 415 |
// Test a combination involving a large ASHR and a shift left. We can't |
| 416 |
// use RISBG there. |
| 417 |
i64 f38 (i64 v_foo) |
| 418 |
{ |
| 419 |
/* { dg-final { scan-assembler "f38:\n\tsrag\t%r2,%r2,32\n\tsllg\t%r2,%r2,5" { target { lp64 } } } } */ |
| 420 |
/* { dg-final { scan-assembler "f38:\n\trisbg\t%r3,%r2,0,0\\\+32-1,64-0-32\n\tsrag\t%r2,%r3,32\n\tsllg\t%r2,%r2,5" { target { ! lp64 } } } } */ |
| 421 |
i64 v_ashr = v_foo >> 32; |
| 422 |
i64 v_shl = v_ashr << 5; |
| 423 |
return v_shl; |
| 424 |
} |
| 425 |
|
| 426 |
// Try a similar thing in which no shifted sign bits are kept. |
| 427 |
i64 f39 (i64 v_foo, i64 *v_dest) |
| 428 |
{ |
| 429 |
/* { dg-final { scan-assembler "f39:\n\tsrag\t%r2,%r2,35\n\(\t.*\n\)*\trisbg\t%r2,%r2,33,128\\\+61,2" { target { lp64 } } } } */ |
| 430 |
/* { dg-final { scan-assembler "f39:\n\trisbg\t%r3,%r2,0,0\\\+32-1,64-0-32\n\tlhi\t%r2,0\n\tsrag\t%r3,%r3,35\n\(\t.*\n\)*\trisbg\t%r3,%r3,33,128\\\+61,2" { target { ! lp64 } } } } */ |
| 431 |
i64 v_ashr = v_foo >> 35; |
| 432 |
*v_dest = v_ashr; |
| 433 |
i64 v_shl = v_ashr << 2; |
| 434 |
i64 v_and = v_shl & 2147483647; |
| 435 |
return v_and; |
| 436 |
} |
| 437 |
|
| 438 |
// ...and again with the next highest shift value, where one sign bit is kept. |
| 439 |
i64 f40 (i64 v_foo, i64 *v_dest) |
| 440 |
{ |
| 441 |
/* { dg-final { scan-assembler "f40:\n\tsrag\t%r2,%r2,36\n\(\t.*\n\)*\trisbg\t%r2,%r2,33,128\\\+61,2" { target { lp64 } } } } */ |
| 442 |
/* { dg-final { scan-assembler "f40:\n\trisbg\t%r3,%r2,0,0\\\+32-1,64-0-32\n\tlhi\t%r2,0\n\tsrag\t%r3,%r3,36\n\(\t.*\n\)*\trisbg\t%r3,%r3,33,128\\\+61,2" { target { ! lp64 } } } } */ |
| 443 |
i64 v_ashr = v_foo >> 36; |
| 444 |
*v_dest = v_ashr; |
| 445 |
i64 v_shl = v_ashr << 2; |
| 446 |
i64 v_and = v_shl & 2147483647; |
| 447 |
return v_and; |
| 448 |
} |
| 449 |
|
| 450 |
// Check a case where the result is zero-extended. |
| 451 |
i64 f41 (i32 v_a) |
| 452 |
{ |
| 453 |
/* { dg-final { scan-assembler "f41:\n\trisbg\t%r2,%r2,64-28,128\\\+63,34\\\+28" { target { lp64 } } } } */ |
| 454 |
/* { dg-final { scan-assembler "f41:\n\trisbg\t%r3,%r2,64-28,128\\\+63,34\\\+28\n\tlhi\t%r2,0" { target { ! lp64 } } } } */ |
| 455 |
i32 v_shl = v_a << 2; |
| 456 |
i32 v_shr = ((ui32)v_shl) >> 4; |
| 457 |
i64 v_ext = (ui64)v_shr; |
| 458 |
return v_ext; |
| 459 |
} |
| 460 |
|
| 461 |
// In this case the sign extension is converted to a pair of 32-bit shifts, |
| 462 |
// which is then extended to 64 bits. We previously used the wrong bit size |
| 463 |
// when testing whether the shifted-in bits of the shift right were significant. |
| 464 |
typedef struct { ui64 pad : 63; ui8 a : 1; } t42; |
| 465 |
i64 f42 (t42 v_x) |
| 466 |
{ |
| 467 |
/* { dg-final { scan-assembler "f42:\n\tsllg\t%r2,%r2,63\n\tsrag\t%r2,%r2,63\n\tllgcr\t%r2,%r2" { target { lp64 } } } } */ |
| 468 |
/* { dg-final { scan-assembler "f42:\n\tsllg\t%r3,%r3,63\n\tlhi\t%r2,0\n\tsrag\t%r3,%r3,63\n\tllcr\t%r3,%r3" { target { ! lp64 } } } } */ |
| 469 |
ui8 a = v_x.a << 7; |
| 470 |
i8 ext = ((i8)a) >> 7; |
| 471 |
i64 ext2 = (ui64)(ui8)ext; |
| 472 |
return ext2; |
| 473 |
} |
| 474 |
|
| 475 |
// Check that we get the case where a 64-bit shift is used by a 32-bit and. |
| 476 |
i32 f43 (i64 v_x) |
| 477 |
{ |
| 478 |
/* { dg-final { scan-assembler "f43:\n\trisbg\t%r2,%r2,32,128\\+61,32\\+20\n\tlgfr\t%r2,%r2" { target { lp64 } } } } */ |
| 479 |
/* { dg-final { scan-assembler "f43:\n\trisbg\t%r3,%r2,0,0\\\+32-1,64-0-32\n\trisbg\t%r2,%r3,32,128\\\+61,64-12" { target { ! lp64 } } } } */ |
| 480 |
i64 v_shr3 = ((ui64)v_x) >> 12; |
| 481 |
i32 v_shr3_tr = (ui32)v_shr3; |
| 482 |
i32 v_conv = v_shr3_tr & -4; |
| 483 |
return v_conv; |
| 484 |
} |
| 485 |
|
| 486 |
// Check that we don't get the case where the 32-bit and mask is not contiguous |
| 487 |
i32 f44 (i64 v_x) |
| 488 |
{ |
| 489 |
/* { dg-final { scan-assembler "f44:\n\tsrlg\t%r2,%r2,12" { target { lp64 } } } } */ |
| 490 |
/* { dg-final { scan-assembler "f44:\n\tsrlg\t%r2,%r3,12\n\tnilf\t%r2,10" { target { ! lp64 } } } } */ |
| 491 |
i64 v_shr4 = ((ui64)v_x) >> 12; |
| 492 |
i32 v_conv = (ui32)v_shr4; |
| 493 |
i32 v_and = v_conv & 10; |
| 494 |
return v_and; |
| 495 |
} |