00001
00002
00003
00004
00005
00006
00007
00008
00009
00010
00011
00012
00013
00014
00015
00016
00017
00018
00019
00020
00021
00022
00023 #ifndef ALPHA_ASM_H
00024 #define ALPHA_ASM_H
00025
00026 #include <inttypes.h>
00027
00028 #if defined __GNUC__
00029 # define GNUC_PREREQ(maj, min) \
00030 ((__GNUC__ << 16) + __GNUC_MINOR__ >= ((maj) << 16) + (min))
00031 #else
00032 # define GNUC_PREREQ(maj, min) 0
00033 #endif
00034
00035 #define AMASK_BWX (1 << 0)
00036 #define AMASK_FIX (1 << 1)
00037 #define AMASK_CIX (1 << 2)
00038 #define AMASK_MVI (1 << 8)
00039
00040 #ifdef __alpha_bwx__
00041 # define HAVE_BWX() 1
00042 #else
00043 # define HAVE_BWX() (amask(AMASK_BWX) == 0)
00044 #endif
00045 #ifdef __alpha_fix__
00046 # define HAVE_FIX() 1
00047 #else
00048 # define HAVE_FIX() (amask(AMASK_FIX) == 0)
00049 #endif
00050 #ifdef __alpha_max__
00051 # define HAVE_MVI() 1
00052 #else
00053 # define HAVE_MVI() (amask(AMASK_MVI) == 0)
00054 #endif
00055 #ifdef __alpha_cix__
00056 # define HAVE_CIX() 1
00057 #else
00058 # define HAVE_CIX() (amask(AMASK_CIX) == 0)
00059 #endif
00060
00061 inline static uint64_t BYTE_VEC(uint64_t x)
00062 {
00063 x |= x << 8;
00064 x |= x << 16;
00065 x |= x << 32;
00066 return x;
00067 }
00068 inline static uint64_t WORD_VEC(uint64_t x)
00069 {
00070 x |= x << 16;
00071 x |= x << 32;
00072 return x;
00073 }
00074
00075 #define ldq(p) (*(const uint64_t *) (p))
00076 #define ldl(p) (*(const int32_t *) (p))
00077 #define stl(l, p) do { *(uint32_t *) (p) = (l); } while (0)
00078 #define stq(l, p) do { *(uint64_t *) (p) = (l); } while (0)
00079 #define sextw(x) ((int16_t) (x))
00080
00081 #ifdef __GNUC__
00082 struct unaligned_long { uint64_t l; } __attribute__((packed));
00083 #define ldq_u(p) (*(const uint64_t *) (((uint64_t) (p)) & ~7ul))
00084 #define uldq(a) (((const struct unaligned_long *) (a))->l)
00085
00086 #if GNUC_PREREQ(3,0)
00087
00088
00089
00090 # define prefetch(p) __builtin_prefetch((p), 0, 1)
00091 # define prefetch_en(p) __builtin_prefetch((p), 1, 1)
00092 # define prefetch_m(p) __builtin_prefetch((p), 0, 0)
00093 # define prefetch_men(p) __builtin_prefetch((p), 1, 0)
00094 #else
00095 # define prefetch(p) asm volatile("ldl $31,%0" : : "m"(*(const char *) (p)) : "memory")
00096 # define prefetch_en(p) asm volatile("ldq $31,%0" : : "m"(*(const char *) (p)) : "memory")
00097 # define prefetch_m(p) asm volatile("lds $f31,%0" : : "m"(*(const char *) (p)) : "memory")
00098 # define prefetch_men(p) asm volatile("ldt $f31,%0" : : "m"(*(const char *) (p)) : "memory")
00099 #endif
00100
00101 #if GNUC_PREREQ(3,3)
00102 #define cmpbge __builtin_alpha_cmpbge
00103
00104 #define extql(a, b) __builtin_alpha_extql(a, (uint64_t) (b))
00105 #define extwl(a, b) __builtin_alpha_extwl(a, (uint64_t) (b))
00106 #define extqh(a, b) __builtin_alpha_extqh(a, (uint64_t) (b))
00107 #define zap __builtin_alpha_zap
00108 #define zapnot __builtin_alpha_zapnot
00109 #define amask __builtin_alpha_amask
00110 #define implver __builtin_alpha_implver
00111 #define rpcc __builtin_alpha_rpcc
00112 #define minub8 __builtin_alpha_minub8
00113 #define minsb8 __builtin_alpha_minsb8
00114 #define minuw4 __builtin_alpha_minuw4
00115 #define minsw4 __builtin_alpha_minsw4
00116 #define maxub8 __builtin_alpha_maxub8
00117 #define maxsb8 __builtin_alpha_maxsb8
00118 #define maxuw4 __builtin_alpha_maxuw4
00119 #define maxsw4 __builtin_alpha_maxsw4
00120 #define perr __builtin_alpha_perr
00121 #define pklb __builtin_alpha_pklb
00122 #define pkwb __builtin_alpha_pkwb
00123 #define unpkbl __builtin_alpha_unpkbl
00124 #define unpkbw __builtin_alpha_unpkbw
00125 #else
00126 #define cmpbge(a, b) ({ uint64_t __r; asm ("cmpbge %r1,%2,%0" : "=r" (__r) : "rJ" (a), "rI" (b)); __r; })
00127 #define extql(a, b) ({ uint64_t __r; asm ("extql %r1,%2,%0" : "=r" (__r) : "rJ" (a), "rI" (b)); __r; })
00128 #define extwl(a, b) ({ uint64_t __r; asm ("extwl %r1,%2,%0" : "=r" (__r) : "rJ" (a), "rI" (b)); __r; })
00129 #define extqh(a, b) ({ uint64_t __r; asm ("extqh %r1,%2,%0" : "=r" (__r) : "rJ" (a), "rI" (b)); __r; })
00130 #define zap(a, b) ({ uint64_t __r; asm ("zap %r1,%2,%0" : "=r" (__r) : "rJ" (a), "rI" (b)); __r; })
00131 #define zapnot(a, b) ({ uint64_t __r; asm ("zapnot %r1,%2,%0" : "=r" (__r) : "rJ" (a), "rI" (b)); __r; })
00132 #define amask(a) ({ uint64_t __r; asm ("amask %1,%0" : "=r" (__r) : "rI" (a)); __r; })
00133 #define implver() ({ uint64_t __r; asm ("implver %0" : "=r" (__r)); __r; })
00134 #define rpcc() ({ uint64_t __r; asm volatile ("rpcc %0" : "=r" (__r)); __r; })
00135 #define minub8(a, b) ({ uint64_t __r; asm ("minub8 %r1,%2,%0" : "=r" (__r) : "%rJ" (a), "rI" (b)); __r; })
00136 #define minsb8(a, b) ({ uint64_t __r; asm ("minsb8 %r1,%2,%0" : "=r" (__r) : "%rJ" (a), "rI" (b)); __r; })
00137 #define minuw4(a, b) ({ uint64_t __r; asm ("minuw4 %r1,%2,%0" : "=r" (__r) : "%rJ" (a), "rI" (b)); __r; })
00138 #define minsw4(a, b) ({ uint64_t __r; asm ("minsw4 %r1,%2,%0" : "=r" (__r) : "%rJ" (a), "rI" (b)); __r; })
00139 #define maxub8(a, b) ({ uint64_t __r; asm ("maxub8 %r1,%2,%0" : "=r" (__r) : "%rJ" (a), "rI" (b)); __r; })
00140 #define maxsb8(a, b) ({ uint64_t __r; asm ("maxsb8 %r1,%2,%0" : "=r" (__r) : "%rJ" (a), "rI" (b)); __r; })
00141 #define maxuw4(a, b) ({ uint64_t __r; asm ("maxuw4 %r1,%2,%0" : "=r" (__r) : "%rJ" (a), "rI" (b)); __r; })
00142 #define maxsw4(a, b) ({ uint64_t __r; asm ("maxsw4 %r1,%2,%0" : "=r" (__r) : "%rJ" (a), "rI" (b)); __r; })
00143 #define perr(a, b) ({ uint64_t __r; asm ("perr %r1,%r2,%0" : "=r" (__r) : "%rJ" (a), "rJ" (b)); __r; })
00144 #define pklb(a) ({ uint64_t __r; asm ("pklb %r1,%0" : "=r" (__r) : "rJ" (a)); __r; })
00145 #define pkwb(a) ({ uint64_t __r; asm ("pkwb %r1,%0" : "=r" (__r) : "rJ" (a)); __r; })
00146 #define unpkbl(a) ({ uint64_t __r; asm ("unpkbl %r1,%0" : "=r" (__r) : "rJ" (a)); __r; })
00147 #define unpkbw(a) ({ uint64_t __r; asm ("unpkbw %r1,%0" : "=r" (__r) : "rJ" (a)); __r; })
00148 #endif
00149 #define wh64(p) asm volatile("wh64 (%0)" : : "r"(p) : "memory")
00150
00151 #elif defined(__DECC)
00152
00153 #include <c_asm.h>
00154 #define ldq_u(a) asm ("ldq_u %v0,0(%a0)", a)
00155 #define uldq(a) (*(const __unaligned uint64_t *) (a))
00156 #define cmpbge(a, b) asm ("cmpbge %a0,%a1,%v0", a, b)
00157 #define extql(a, b) asm ("extql %a0,%a1,%v0", a, b)
00158 #define extwl(a, b) asm ("extwl %a0,%a1,%v0", a, b)
00159 #define extqh(a, b) asm ("extqh %a0,%a1,%v0", a, b)
00160 #define zap(a, b) asm ("zap %a0,%a1,%v0", a, b)
00161 #define zapnot(a, b) asm ("zapnot %a0,%a1,%v0", a, b)
00162 #define amask(a) asm ("amask %a0,%v0", a)
00163 #define implver() asm ("implver %v0")
00164 #define rpcc() asm ("rpcc %v0")
00165 #define minub8(a, b) asm ("minub8 %a0,%a1,%v0", a, b)
00166 #define minsb8(a, b) asm ("minsb8 %a0,%a1,%v0", a, b)
00167 #define minuw4(a, b) asm ("minuw4 %a0,%a1,%v0", a, b)
00168 #define minsw4(a, b) asm ("minsw4 %a0,%a1,%v0", a, b)
00169 #define maxub8(a, b) asm ("maxub8 %a0,%a1,%v0", a, b)
00170 #define maxsb8(a, b) asm ("maxsb8 %a0,%a1,%v0", a, b)
00171 #define maxuw4(a, b) asm ("maxuw4 %a0,%a1,%v0", a, b)
00172 #define maxsw4(a, b) asm ("maxsw4 %a0,%a1,%v0", a, b)
00173 #define perr(a, b) asm ("perr %a0,%a1,%v0", a, b)
00174 #define pklb(a) asm ("pklb %a0,%v0", a)
00175 #define pkwb(a) asm ("pkwb %a0,%v0", a)
00176 #define unpkbl(a) asm ("unpkbl %a0,%v0", a)
00177 #define unpkbw(a) asm ("unpkbw %a0,%v0", a)
00178 #define wh64(a) asm ("wh64 %a0", a)
00179
00180 #else
00181 #error "Unknown compiler!"
00182 #endif
00183
00184 #endif