|
|
@@ -3,34 +3,40 @@
|
|
|
#include <stdint.h>
|
|
|
#include <stdio.h>
|
|
|
|
|
|
+#ifdef __x86_64
|
|
|
+typedef uint64_t reg_t;
|
|
|
+#else
|
|
|
+typedef uint32_t reg_t;
|
|
|
+#endif
|
|
|
+
|
|
|
#define insn1q(name, arg0) \
|
|
|
-static inline uint64_t name##q(uint64_t arg0) \
|
|
|
+static inline reg_t name##q(reg_t arg0) \
|
|
|
{ \
|
|
|
- uint64_t result64; \
|
|
|
+ reg_t result64; \
|
|
|
asm volatile (#name "q %1, %0" : "=r"(result64) : "rm"(arg0)); \
|
|
|
return result64; \
|
|
|
}
|
|
|
|
|
|
#define insn1l(name, arg0) \
|
|
|
-static inline uint32_t name##l(uint32_t arg0) \
|
|
|
+static inline reg_t name##l(reg_t arg0) \
|
|
|
{ \
|
|
|
- uint32_t result32; \
|
|
|
+ reg_t result32; \
|
|
|
asm volatile (#name "l %k1, %k0" : "=r"(result32) : "rm"(arg0)); \
|
|
|
return result32; \
|
|
|
}
|
|
|
|
|
|
#define insn2q(name, arg0, c0, arg1, c1) \
|
|
|
-static inline uint64_t name##q(uint64_t arg0, uint64_t arg1) \
|
|
|
+static inline reg_t name##q(reg_t arg0, reg_t arg1) \
|
|
|
{ \
|
|
|
- uint64_t result64; \
|
|
|
+ reg_t result64; \
|
|
|
asm volatile (#name "q %2, %1, %0" : "=r"(result64) : c0(arg0), c1(arg1)); \
|
|
|
return result64; \
|
|
|
}
|
|
|
|
|
|
#define insn2l(name, arg0, c0, arg1, c1) \
|
|
|
-static inline uint32_t name##l(uint32_t arg0, uint32_t arg1) \
|
|
|
+static inline reg_t name##l(reg_t arg0, reg_t arg1) \
|
|
|
{ \
|
|
|
- uint32_t result32; \
|
|
|
+ reg_t result32; \
|
|
|
asm volatile (#name "l %k2, %k1, %k0" : "=r"(result32) : c0(arg0), c1(arg1)); \
|
|
|
return result32; \
|
|
|
}
|
|
|
@@ -65,130 +71,140 @@ insn1l(blsr, src)
|
|
|
int main(int argc, char *argv[]) {
|
|
|
uint64_t ehlo = 0x202020204f4c4845ull;
|
|
|
uint64_t mask = 0xa080800302020001ull;
|
|
|
- uint32_t result32;
|
|
|
+ reg_t result;
|
|
|
|
|
|
#ifdef __x86_64
|
|
|
- uint64_t result64;
|
|
|
-
|
|
|
/* 64 bits */
|
|
|
- result64 = andnq(mask, ehlo);
|
|
|
- assert(result64 == 0x002020204d4c4844);
|
|
|
+ result = andnq(mask, ehlo);
|
|
|
+ assert(result == 0x002020204d4c4844);
|
|
|
|
|
|
- result64 = pextq(ehlo, mask);
|
|
|
- assert(result64 == 133);
|
|
|
+ result = pextq(ehlo, mask);
|
|
|
+ assert(result == 133);
|
|
|
|
|
|
- result64 = pdepq(result64, mask);
|
|
|
- assert(result64 == (ehlo & mask));
|
|
|
+ result = pdepq(result, mask);
|
|
|
+ assert(result == (ehlo & mask));
|
|
|
|
|
|
- result64 = pextq(-1ull, mask);
|
|
|
- assert(result64 == 511); /* mask has 9 bits set */
|
|
|
+ result = pextq(-1ull, mask);
|
|
|
+ assert(result == 511); /* mask has 9 bits set */
|
|
|
|
|
|
- result64 = pdepq(-1ull, mask);
|
|
|
- assert(result64 == mask);
|
|
|
+ result = pdepq(-1ull, mask);
|
|
|
+ assert(result == mask);
|
|
|
|
|
|
- result64 = bextrq(mask, 0x3f00);
|
|
|
- assert(result64 == (mask & ~INT64_MIN));
|
|
|
+ result = bextrq(mask, 0x3f00);
|
|
|
+ assert(result == (mask & ~INT64_MIN));
|
|
|
|
|
|
- result64 = bextrq(mask, 0x1038);
|
|
|
- assert(result64 == 0xa0);
|
|
|
+ result = bextrq(mask, 0x1038);
|
|
|
+ assert(result == 0xa0);
|
|
|
|
|
|
- result64 = bextrq(mask, 0x10f8);
|
|
|
- assert(result64 == 0);
|
|
|
+ result = bextrq(mask, 0x10f8);
|
|
|
+ assert(result == 0);
|
|
|
|
|
|
- result64 = blsiq(0x30);
|
|
|
- assert(result64 == 0x10);
|
|
|
+ result = bextrq(0xfedcba9876543210ull, 0x7f00);
|
|
|
+ assert(result == 0xfedcba9876543210ull);
|
|
|
|
|
|
- result64 = blsiq(0x30ull << 32);
|
|
|
- assert(result64 == 0x10ull << 32);
|
|
|
+ result = blsiq(0x30);
|
|
|
+ assert(result == 0x10);
|
|
|
|
|
|
- result64 = blsmskq(0x30);
|
|
|
- assert(result64 == 0x1f);
|
|
|
+ result = blsiq(0x30ull << 32);
|
|
|
+ assert(result == 0x10ull << 32);
|
|
|
|
|
|
- result64 = blsrq(0x30);
|
|
|
- assert(result64 == 0x20);
|
|
|
+ result = blsmskq(0x30);
|
|
|
+ assert(result == 0x1f);
|
|
|
|
|
|
- result64 = blsrq(0x30ull << 32);
|
|
|
- assert(result64 == 0x20ull << 32);
|
|
|
+ result = blsrq(0x30);
|
|
|
+ assert(result == 0x20);
|
|
|
|
|
|
- result64 = bzhiq(mask, 0x3f);
|
|
|
- assert(result64 == (mask & ~INT64_MIN));
|
|
|
+ result = blsrq(0x30ull << 32);
|
|
|
+ assert(result == 0x20ull << 32);
|
|
|
|
|
|
- result64 = bzhiq(mask, 0x1f);
|
|
|
- assert(result64 == (mask & ~(-1 << 30)));
|
|
|
+ result = bzhiq(mask, 0x3f);
|
|
|
+ assert(result == (mask & ~INT64_MIN));
|
|
|
|
|
|
- result64 = rorxq(0x2132435465768798, 8);
|
|
|
- assert(result64 == 0x9821324354657687);
|
|
|
+ result = bzhiq(mask, 0x1f);
|
|
|
+ assert(result == (mask & ~(-1 << 30)));
|
|
|
|
|
|
- result64 = sarxq(0xffeeddccbbaa9988, 8);
|
|
|
- assert(result64 == 0xffffeeddccbbaa99);
|
|
|
+ result = rorxq(0x2132435465768798, 8);
|
|
|
+ assert(result == 0x9821324354657687);
|
|
|
|
|
|
- result64 = sarxq(0x77eeddccbbaa9988, 8 | 64);
|
|
|
- assert(result64 == 0x0077eeddccbbaa99);
|
|
|
+ result = sarxq(0xffeeddccbbaa9988, 8);
|
|
|
+ assert(result == 0xffffeeddccbbaa99);
|
|
|
|
|
|
- result64 = shrxq(0xffeeddccbbaa9988, 8);
|
|
|
- assert(result64 == 0x00ffeeddccbbaa99);
|
|
|
+ result = sarxq(0x77eeddccbbaa9988, 8 | 64);
|
|
|
+ assert(result == 0x0077eeddccbbaa99);
|
|
|
|
|
|
- result64 = shrxq(0x77eeddccbbaa9988, 8 | 192);
|
|
|
- assert(result64 == 0x0077eeddccbbaa99);
|
|
|
+ result = shrxq(0xffeeddccbbaa9988, 8);
|
|
|
+ assert(result == 0x00ffeeddccbbaa99);
|
|
|
|
|
|
- result64 = shlxq(0xffeeddccbbaa9988, 8);
|
|
|
- assert(result64 == 0xeeddccbbaa998800);
|
|
|
+ result = shrxq(0x77eeddccbbaa9988, 8 | 192);
|
|
|
+ assert(result == 0x0077eeddccbbaa99);
|
|
|
+
|
|
|
+ result = shlxq(0xffeeddccbbaa9988, 8);
|
|
|
+ assert(result == 0xeeddccbbaa998800);
|
|
|
#endif
|
|
|
|
|
|
/* 32 bits */
|
|
|
- result32 = andnl(mask, ehlo);
|
|
|
- assert(result32 == 0x04d4c4844);
|
|
|
+ result = andnl(mask, ehlo);
|
|
|
+ assert(result == 0x04d4c4844);
|
|
|
+
|
|
|
+ result = pextl((uint32_t) ehlo, mask);
|
|
|
+ assert(result == 5);
|
|
|
+
|
|
|
+ result = pdepl(result, mask);
|
|
|
+ assert(result == (uint32_t)(ehlo & mask));
|
|
|
+
|
|
|
+ result = pextl(-1u, mask);
|
|
|
+ assert(result == 7); /* mask has 3 bits set */
|
|
|
|
|
|
- result32 = pextl((uint32_t) ehlo, mask);
|
|
|
- assert(result32 == 5);
|
|
|
+ result = pdepl(-1u, mask);
|
|
|
+ assert(result == (uint32_t)mask);
|
|
|
|
|
|
- result32 = pdepl(result32, mask);
|
|
|
- assert(result32 == (uint32_t)(ehlo & mask));
|
|
|
+ result = bextrl(mask, 0x1f00);
|
|
|
+ assert(result == (mask & ~INT32_MIN));
|
|
|
|
|
|
- result32 = pextl(-1u, mask);
|
|
|
- assert(result32 == 7); /* mask has 3 bits set */
|
|
|
+ result = bextrl(ehlo, 0x1018);
|
|
|
+ assert(result == 0x4f);
|
|
|
|
|
|
- result32 = pdepl(-1u, mask);
|
|
|
- assert(result32 == (uint32_t)mask);
|
|
|
+ result = bextrl(mask, 0x1038);
|
|
|
+ assert(result == 0);
|
|
|
|
|
|
- result32 = bextrl(mask, 0x1f00);
|
|
|
- assert(result32 == (mask & ~INT32_MIN));
|
|
|
+ result = bextrl((reg_t)0x8f635a775ad3b9b4ull, 0x3018);
|
|
|
+ assert(result == 0x5a);
|
|
|
|
|
|
- result32 = bextrl(ehlo, 0x1018);
|
|
|
- assert(result32 == 0x4f);
|
|
|
+ result = bextrl((reg_t)0xfedcba9876543210ull, 0x7f00);
|
|
|
+ assert(result == 0x76543210u);
|
|
|
|
|
|
- result32 = bextrl(mask, 0x1038);
|
|
|
- assert(result32 == 0);
|
|
|
+ result = bextrl(-1, 0);
|
|
|
+ assert(result == 0);
|
|
|
|
|
|
- result32 = blsil(0xffff);
|
|
|
- assert(result32 == 1);
|
|
|
+ result = blsil(0xffff);
|
|
|
+ assert(result == 1);
|
|
|
|
|
|
- result32 = blsmskl(0x300);
|
|
|
- assert(result32 == 0x1ff);
|
|
|
+ result = blsmskl(0x300);
|
|
|
+ assert(result == 0x1ff);
|
|
|
|
|
|
- result32 = blsrl(0xffc);
|
|
|
- assert(result32 == 0xff8);
|
|
|
+ result = blsrl(0xffc);
|
|
|
+ assert(result == 0xff8);
|
|
|
|
|
|
- result32 = bzhil(mask, 0xf);
|
|
|
- assert(result32 == 1);
|
|
|
+ result = bzhil(mask, 0xf);
|
|
|
+ assert(result == 1);
|
|
|
|
|
|
- result32 = rorxl(0x65768798, 8);
|
|
|
- assert(result32 == 0x98657687);
|
|
|
+ result = rorxl(0x65768798, 8);
|
|
|
+ assert(result == 0x98657687);
|
|
|
|
|
|
- result32 = sarxl(0xffeeddcc, 8);
|
|
|
- assert(result32 == 0xffffeedd);
|
|
|
+ result = sarxl(0xffeeddcc, 8);
|
|
|
+ assert(result == 0xffffeedd);
|
|
|
|
|
|
- result32 = sarxl(0x77eeddcc, 8 | 32);
|
|
|
- assert(result32 == 0x0077eedd);
|
|
|
+ result = sarxl(0x77eeddcc, 8 | 32);
|
|
|
+ assert(result == 0x0077eedd);
|
|
|
|
|
|
- result32 = shrxl(0xffeeddcc, 8);
|
|
|
- assert(result32 == 0x00ffeedd);
|
|
|
+ result = shrxl(0xffeeddcc, 8);
|
|
|
+ assert(result == 0x00ffeedd);
|
|
|
|
|
|
- result32 = shrxl(0x77eeddcc, 8 | 128);
|
|
|
- assert(result32 == 0x0077eedd);
|
|
|
+ result = shrxl(0x77eeddcc, 8 | 128);
|
|
|
+ assert(result == 0x0077eedd);
|
|
|
|
|
|
- result32 = shlxl(0xffeeddcc, 8);
|
|
|
- assert(result32 == 0xeeddcc00);
|
|
|
+ result = shlxl(0xffeeddcc, 8);
|
|
|
+ assert(result == 0xeeddcc00);
|
|
|
|
|
|
return 0;
|
|
|
}
|
Peter Maydell