1ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown#include <stdlib.h>
2ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown#include <stdio.h>
3ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
4ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Browntypedef unsigned int UInt;
5ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Browntypedef unsigned long long int ULong;
6ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
7ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brownvoid do_cmpxchg8b ( /*OUT*/
8ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown                    ULong* rdxOut,   ULong* raxOut,
9ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown		    ULong* memHiOut, ULong* memLoOut,
10ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown                    ULong* zOut,
11ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown                    /*IN*/
12ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown                    ULong rdxIn,   ULong raxIn,
13ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown                    ULong memHiIn, ULong memLoIn,
14ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown                    ULong rcxIn,   ULong rbxIn )
15ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown{
16ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   UInt mem[2];
17ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   ULong block[6];
18ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   mem[0] = (UInt)memLoIn;
19ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   mem[1] = (UInt)memHiIn;
20ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   block[0] = rdxIn;
21ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   block[1] = raxIn;
22ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   block[2] = rcxIn;
23ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   block[3] = rbxIn;
24ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   block[4] = (ULong)&mem[0];
25ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   block[5] = ~(0ULL);
26ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   __asm__ __volatile__(
27ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown          "movq %0,%%r11\n"
28ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown        "\tmovq  0(%%r11),%%rdx\n"
29ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown        "\tmovq  8(%%r11),%%rax\n"
30ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown        "\tmovq 16(%%r11),%%rcx\n"
31ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown        "\tmovq 24(%%r11),%%rbx\n"
32ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown        "\tmovq 32(%%r11),%%r10\n"
33ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown        "\tlock cmpxchg8b (%%r10)\n"
34ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown        "\tmovabsq $0,%%r10\n"
35ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown        "\tsetz %%r10b\n"
36ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown        "\tmovq %%r10,40(%%r11)\n"
37ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown        "\tmovq %%rdx,0(%%r11)\n"
38ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown        "\tmovq %%rax,8(%%r11)\n"
39ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	  : /*out*/
40ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	  : /*in*/ "r"(&block[0])
41ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	  : /*trash*/ "%r11", "%r10", "%rax", "%rbx", "%rcx", "%rdx",
42ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown                      "cc", "memory" );
43ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown    *rdxOut = block[0];
44ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown    *raxOut = block[1];
45ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown    *memLoOut = (ULong)mem[0];
46ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown    *memHiOut = (ULong)mem[1];
47ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown    *zOut = block[5];
48ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown}
49ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
50ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brownvoid try8b ( ULong d, ULong a, ULong mHi, ULong mLo, ULong c, ULong b )
51ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown{
52ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   ULong dd, aa, mmHi, mmLo, zz;
53ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   do_cmpxchg8b( &dd, &aa, &mmHi, &mmLo, &zz,
54ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown		 d,a,mHi,mLo,c,b);
55ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   printf(" Q d:a=%llx:%llx mem=%llx:%llx c:b=%llx:%llx "
56ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown          "-> z=%lld d:a=%llx:%llx mem=%llx:%llx\n",
57ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	  d,a, mHi,mLo, c,b, zz, dd,aa, mmHi,mmLo );
58ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown}
59ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
60ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brownvoid do_cmpxchg16b ( /*OUT*/
61ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown                     ULong* rdxOut,   ULong* raxOut,
62ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown		     ULong* memHiOut, ULong* memLoOut,
63ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown                     ULong* zOut,
64ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown                     /*IN*/
65ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown                     ULong rdxIn,   ULong raxIn,
66ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown                     ULong memHiIn, ULong memLoIn,
67ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown                     ULong rcxIn,   ULong rbxIn )
68ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown{
69b32f58018498ea2225959b0ba11c18f0c433deefEvgeniy Stepanov   ULong mem[2] __attribute__((aligned(16)));
70ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   ULong block[6];
71ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   mem[0] = memLoIn;
72ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   mem[1] = memHiIn;
73ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   block[0] = rdxIn;
74ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   block[1] = raxIn;
75ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   block[2] = rcxIn;
76ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   block[3] = rbxIn;
77ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   block[4] = (ULong)&mem[0];
78ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   block[5] = ~(0ULL);
79ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   __asm__ __volatile__(
80ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown          "movq %0,%%r11\n"
81ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown        "\tmovq  0(%%r11),%%rdx\n"
82ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown        "\tmovq  8(%%r11),%%rax\n"
83ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown        "\tmovq 16(%%r11),%%rcx\n"
84ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown        "\tmovq 24(%%r11),%%rbx\n"
85ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown        "\tmovq 32(%%r11),%%r10\n"
86ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown        "\t.byte 0xf0, 0x49, 0x0f, 0xc7, 0x0a\n" /* lock cmpxchg16b (%%r10) */
87ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown        "\tmovabsq $0,%%r10\n"
88ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown        "\tsetz %%r10b\n"
89ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown        "\tmovq %%r10,40(%%r11)\n"
90ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown        "\tmovq %%rdx,0(%%r11)\n"
91ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown        "\tmovq %%rax,8(%%r11)\n"
92ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	  : /*out*/
93ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	  : /*in*/ "r"(&block[0])
94ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	  : /*trash*/ "%r11", "%r10", "%rax", "%rbx", "%rcx", "%rdx",
95ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown                      "cc", "memory" );
96ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown    *rdxOut = block[0];
97ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown    *raxOut = block[1];
98ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown    *memLoOut = mem[0];
99ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown    *memHiOut = mem[1];
100ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown    *zOut = block[5];
101ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown}
102ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
103ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brownvoid try16b ( ULong d, ULong a, ULong mHi, ULong mLo, ULong c, ULong b )
104ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown{
105ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   ULong dd, aa, mmHi, mmLo, zz;
106ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   do_cmpxchg16b( &dd, &aa, &mmHi, &mmLo, &zz,
107ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown		  d,a,mHi,mLo,c,b);
108ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   printf("QQ d:a=%llx:%llx mem=%llx:%llx c:b=%llx:%llx "
109ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown          "-> z=%lld d:a=%llx:%llx mem=%llx:%llx\n",
110ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown	  d,a, mHi,mLo, c,b, zz, dd,aa, mmHi,mmLo );
111ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown}
112ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
113ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brownint main(void)
114ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown{
115ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   ULong z = 0xDEADBEEF00000000ULL;
116ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
117ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   try8b( 0,1, 5,4, 3,2 );
118ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   try8b( 0,1, 0,1, 3,2 );
119ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
120ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   try8b( 0,1, 0,4, 3,2 );
121ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   try8b( 0,1, 0,0, 3,2 );
122ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
123ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   try8b( 0,1, 5,0, 3,2 );
124ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   try8b( 0,1, 1,1, 3,2 );
125ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
126ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   try8b( 0+z,1+z, 5+z,4+z, 3+z,2+z );
127ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   try8b( 0+z,1+z, 0+z,1+z, 3+z,2+z );
128ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
129ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   try8b( 0+z,1+z, 0+z,4+z, 3+z,2+z );
130ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   try8b( 0+z,1+z, 0+z,0+z, 3+z,2+z );
131ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
132ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   try8b( 0+z,1+z, 5+z,0+z, 3+z,2+z );
133ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   try8b( 0+z,1+z, 1+z,1+z, 3+z,2+z );
134ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
135ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   try16b( 0,1, 5,4, 3,2 );
136ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   try16b( 0,1, 0,1, 3,2 );
137ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
138ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   try16b( 0,1, 0,4, 3,2 );
139ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   try16b( 0,1, 0,0, 3,2 );
140ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
141ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   try16b( 0,1, 5,0, 3,2 );
142ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   try16b( 0,1, 1,1, 3,2 );
143ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
144ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   try16b( 0+z,1+z, 5+z,4+z, 3+z,2+z );
145ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   try16b( 0+z,1+z, 0+z,1+z, 3+z,2+z );
146ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
147ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   try16b( 0+z,1+z, 0+z,4+z, 3+z,2+z );
148ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   try16b( 0+z,1+z, 0+z,0+z, 3+z,2+z );
149ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
150ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   try16b( 0+z,1+z, 5+z,0+z, 3+z,2+z );
151ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   try16b( 0+z,1+z, 1+z,1+z, 3+z,2+z );
152ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
153ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown   return 0;
154ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown}
155ed07e00d438c74b7a23c01bfffde77e3968305e4Jeff Brown
156