| 1 | /* See if various BMI2 instructions give expected results */ |
| 2 | #include <assert.h> |
| 3 | #include <stdint.h> |
| 4 | #include <stdio.h> |
| 5 | |
| 6 | #ifdef __x86_64 |
| 7 | typedef uint64_t reg_t; |
| 8 | #else |
| 9 | typedef uint32_t reg_t; |
| 10 | #endif |
| 11 | |
| 12 | #define insn1q(name, arg0) \ |
| 13 | static inline reg_t name##q(reg_t arg0) \ |
| 14 | { \ |
| 15 | reg_t result64; \ |
| 16 | asm volatile (#name "q %1, %0" : "=r"(result64) : "rm"(arg0)); \ |
| 17 | return result64; \ |
| 18 | } |
| 19 | |
| 20 | #define insn1l(name, arg0) \ |
| 21 | static inline reg_t name##l(reg_t arg0) \ |
| 22 | { \ |
| 23 | reg_t result32; \ |
| 24 | asm volatile (#name "l %k1, %k0" : "=r"(result32) : "rm"(arg0)); \ |
| 25 | return result32; \ |
| 26 | } |
| 27 | |
| 28 | #define insn2q(name, arg0, c0, arg1, c1) \ |
| 29 | static inline reg_t name##q(reg_t arg0, reg_t arg1) \ |
| 30 | { \ |
| 31 | reg_t result64; \ |
| 32 | asm volatile (#name "q %2, %1, %0" : "=r"(result64) : c0(arg0), c1(arg1)); \ |
| 33 | return result64; \ |
| 34 | } |
| 35 | |
| 36 | #define insn2l(name, arg0, c0, arg1, c1) \ |
| 37 | static inline reg_t name##l(reg_t arg0, reg_t arg1) \ |
| 38 | { \ |
| 39 | reg_t result32; \ |
| 40 | asm volatile (#name "l %k2, %k1, %k0" : "=r"(result32) : c0(arg0), c1(arg1)); \ |
| 41 | return result32; \ |
| 42 | } |
| 43 | |
| 44 | #ifdef __x86_64 |
| 45 | insn2q(pext, src, "r", mask, "rm") |
| 46 | insn2q(pdep, src, "r", mask, "rm") |
| 47 | insn2q(andn, clear, "rm", val, "r") |
| 48 | insn2q(bextr, range, "rm", val, "r") |
| 49 | insn2q(bzhi, pos, "rm", val, "r") |
| 50 | insn2q(rorx, val, "r", n, "i") |
| 51 | insn2q(sarx, val, "rm", n, "r") |
| 52 | insn2q(shlx, val, "rm", n, "r") |
| 53 | insn2q(shrx, val, "rm", n, "r") |
| 54 | insn1q(blsi, src) |
| 55 | insn1q(blsmsk, src) |
| 56 | insn1q(blsr, src) |
| 57 | #endif |
| 58 | insn2l(pext, src, "r", mask, "rm") |
| 59 | insn2l(pdep, src, "r", mask, "rm") |
| 60 | insn2l(andn, clear, "rm", val, "r") |
| 61 | insn2l(bextr, range, "rm", val, "r") |
| 62 | insn2l(bzhi, pos, "rm", val, "r") |
| 63 | insn2l(rorx, val, "r", n, "i") |
| 64 | insn2l(sarx, val, "rm", n, "r") |
| 65 | insn2l(shlx, val, "rm", n, "r") |
| 66 | insn2l(shrx, val, "rm", n, "r") |
| 67 | insn1l(blsi, src) |
| 68 | insn1l(blsmsk, src) |
| 69 | insn1l(blsr, src) |
| 70 | |
| 71 | int main(int argc, char *argv[]) { |
| 72 | uint64_t ehlo = 0x202020204f4c4845ull; |
| 73 | uint64_t mask = 0xa080800302020001ull; |
| 74 | reg_t result; |
| 75 | |
| 76 | #ifdef __x86_64 |
| 77 | /* 64 bits */ |
| 78 | result = andnq(mask, ehlo); |
| 79 | assert(result == 0x002020204d4c4844); |
| 80 | |
| 81 | result = pextq(ehlo, mask); |
| 82 | assert(result == 133); |
| 83 | |
| 84 | result = pdepq(result, mask); |
| 85 | assert(result == (ehlo & mask)); |
| 86 | |
| 87 | result = pextq(-1ull, mask); |
| 88 | assert(result == 511); /* mask has 9 bits set */ |
| 89 | |
| 90 | result = pdepq(-1ull, mask); |
| 91 | assert(result == mask); |
| 92 | |
| 93 | result = bextrq(mask, 0x3f00); |
| 94 | assert(result == (mask & ~INT64_MIN)); |
| 95 | |
| 96 | result = bextrq(mask, 0x1038); |
| 97 | assert(result == 0xa0); |
| 98 | |
| 99 | result = bextrq(mask, 0x10f8); |
| 100 | assert(result == 0); |
| 101 | |
| 102 | result = bextrq(0xfedcba9876543210ull, 0x7f00); |
| 103 | assert(result == 0xfedcba9876543210ull); |
| 104 | |
| 105 | result = blsiq(0x30); |
| 106 | assert(result == 0x10); |
| 107 | |
| 108 | result = blsiq(0x30ull << 32); |
| 109 | assert(result == 0x10ull << 32); |
| 110 | |
| 111 | result = blsmskq(0x30); |
| 112 | assert(result == 0x1f); |
| 113 | |
| 114 | result = blsrq(0x30); |
| 115 | assert(result == 0x20); |
| 116 | |
| 117 | result = blsrq(0x30ull << 32); |
| 118 | assert(result == 0x20ull << 32); |
| 119 | |
| 120 | result = bzhiq(mask, 0x3f); |
| 121 | assert(result == (mask & ~INT64_MIN)); |
| 122 | |
| 123 | result = bzhiq(mask, 0x1f); |
| 124 | assert(result == (mask & ~(-1 << 30))); |
| 125 | |
| 126 | result = bzhiq(mask, 0x40); |
| 127 | assert(result == mask); |
| 128 | |
| 129 | result = rorxq(0x2132435465768798, 8); |
| 130 | assert(result == 0x9821324354657687); |
| 131 | |
| 132 | result = sarxq(0xffeeddccbbaa9988, 8); |
| 133 | assert(result == 0xffffeeddccbbaa99); |
| 134 | |
| 135 | result = sarxq(0x77eeddccbbaa9988, 8 | 64); |
| 136 | assert(result == 0x0077eeddccbbaa99); |
| 137 | |
| 138 | result = shrxq(0xffeeddccbbaa9988, 8); |
| 139 | assert(result == 0x00ffeeddccbbaa99); |
| 140 | |
| 141 | result = shrxq(0x77eeddccbbaa9988, 8 | 192); |
| 142 | assert(result == 0x0077eeddccbbaa99); |
| 143 | |
| 144 | result = shlxq(0xffeeddccbbaa9988, 8); |
| 145 | assert(result == 0xeeddccbbaa998800); |
| 146 | #endif |
| 147 | |
| 148 | /* 32 bits */ |
| 149 | result = andnl(mask, ehlo); |
| 150 | assert(result == 0x04d4c4844); |
| 151 | |
| 152 | result = pextl((uint32_t) ehlo, mask); |
| 153 | assert(result == 5); |
| 154 | |
| 155 | result = pdepl(result, mask); |
| 156 | assert(result == (uint32_t)(ehlo & mask)); |
| 157 | |
| 158 | result = pextl(-1u, mask); |
| 159 | assert(result == 7); /* mask has 3 bits set */ |
| 160 | |
| 161 | result = pdepl(-1u, mask); |
| 162 | assert(result == (uint32_t)mask); |
| 163 | |
| 164 | result = bextrl(mask, 0x1f00); |
| 165 | assert(result == (mask & ~INT32_MIN)); |
| 166 | |
| 167 | result = bextrl(ehlo, 0x1018); |
| 168 | assert(result == 0x4f); |
| 169 | |
| 170 | result = bextrl(mask, 0x1038); |
| 171 | assert(result == 0); |
| 172 | |
| 173 | result = bextrl((reg_t)0x8f635a775ad3b9b4ull, 0x3018); |
| 174 | assert(result == 0x5a); |
| 175 | |
| 176 | result = bextrl((reg_t)0xfedcba9876543210ull, 0x7f00); |
| 177 | assert(result == 0x76543210u); |
| 178 | |
| 179 | result = bextrl(-1, 0); |
| 180 | assert(result == 0); |
| 181 | |
| 182 | result = blsil(0xffff); |
| 183 | assert(result == 1); |
| 184 | |
| 185 | result = blsmskl(0x300); |
| 186 | assert(result == 0x1ff); |
| 187 | |
| 188 | result = blsrl(0xffc); |
| 189 | assert(result == 0xff8); |
| 190 | |
| 191 | result = bzhil(mask, 0xf); |
| 192 | assert(result == 1); |
| 193 | |
| 194 | result = rorxl(0x65768798, 8); |
| 195 | assert(result == 0x98657687); |
| 196 | |
| 197 | result = sarxl(0xffeeddcc, 8); |
| 198 | assert(result == 0xffffeedd); |
| 199 | |
| 200 | result = sarxl(0x77eeddcc, 8 | 32); |
| 201 | assert(result == 0x0077eedd); |
| 202 | |
| 203 | result = shrxl(0xffeeddcc, 8); |
| 204 | assert(result == 0x00ffeedd); |
| 205 | |
| 206 | result = shrxl(0x77eeddcc, 8 | 128); |
| 207 | assert(result == 0x0077eedd); |
| 208 | |
| 209 | result = shlxl(0xffeeddcc, 8); |
| 210 | assert(result == 0xeeddcc00); |
| 211 | |
| 212 | return 0; |
| 213 | } |
| 214 |