1 #ifndef INCLUDED_ARGS_H
2 #define INCLUDED_ARGS_H
3
4 #include <immintrin.h>
5 #include <string.h>
6
7 /* Assertion macro. */
8 #define assert(test) if (!(test)) abort()
9
10 #ifdef __GNUC__
11 #define ATTRIBUTE_UNUSED __attribute__((__unused__))
12 #else
13 #define ATTRIBUTE_UNUSED
14 #endif
15
16 /* This defines the calling sequences for integers and floats. */
17 #define I0 rdi
18 #define I1 rsi
19 #define I2 rdx
20 #define I3 rcx
21 #define I4 r8
22 #define I5 r9
23 #define F0 zmm0
24 #define F1 zmm1
25 #define F2 zmm2
26 #define F3 zmm3
27 #define F4 zmm4
28 #define F5 zmm5
29 #define F6 zmm6
30 #define F7 zmm7
31
32 typedef union {
33 _Float16 __Float16[32];
34 float _float[16];
35 double _double[8];
36 long long _longlong[8];
37 int _int[16];
38 unsigned long long _ulonglong[8];
39 __m64 _m64[8];
40 __m128 _m128[4];
41 __m256 _m256[2];
42 __m512 _m512[1];
43 __m512h _m512h[1];
44 } ZMM_T;
45
46 typedef union {
47 float _float;
48 double _double;
49 long double _ldouble;
50 unsigned long long _ulonglong[2];
51 } X87_T;
52 extern void (*callthis)(void);
53 extern unsigned long long rax,rbx,rcx,rdx,rsi,rdi,rsp,rbp,r8,r9,r10,r11,r12,r13,r14,r15;
54 ZMM_T zmm_regs[32];
55 X87_T x87_regs[8];
56 extern volatile unsigned long long volatile_var;
57 extern void snapshot (void);
58 extern void snapshot_ret (void);
59 #define WRAP_CALL(N) \
60 (callthis = (void (*)()) (N), (typeof (&N)) snapshot)
61 #define WRAP_RET(N) \
62 (callthis = (void (*)()) (N), (typeof (&N)) snapshot_ret)
63
64 /* Clear all integer registers. */
65 #define clear_int_hardware_registers \
66 asm __volatile__ ("xor %%rax, %%rax\n\t" \
67 "xor %%rbx, %%rbx\n\t" \
68 "xor %%rcx, %%rcx\n\t" \
69 "xor %%rdx, %%rdx\n\t" \
70 "xor %%rsi, %%rsi\n\t" \
71 "xor %%rdi, %%rdi\n\t" \
72 "xor %%r8, %%r8\n\t" \
73 "xor %%r9, %%r9\n\t" \
74 "xor %%r10, %%r10\n\t" \
75 "xor %%r11, %%r11\n\t" \
76 "xor %%r12, %%r12\n\t" \
77 "xor %%r13, %%r13\n\t" \
78 "xor %%r14, %%r14\n\t" \
79 "xor %%r15, %%r15\n\t" \
80 ::: "rax", "rbx", "rcx", "rdx", "rsi", "rdi", "r8", \
81 "r9", "r10", "r11", "r12", "r13", "r14", "r15");
82
83 /* This is the list of registers available for passing arguments. Not all of
84 these are used or even really available. */
85 struct IntegerRegisters
86 {
87 unsigned long long rax, rbx, rcx, rdx, rsi, rdi, r8, r9, r10, r11, r12, r13, r14, r15;
88 };
89 struct FloatRegisters
90 {
91 double mm0, mm1, mm2, mm3, mm4, mm5, mm6, mm7;
92 long double st0, st1, st2, st3, st4, st5, st6, st7;
93 ZMM_T zmm0, zmm1, zmm2, zmm3, zmm4, zmm5, zmm6, zmm7, zmm8, zmm9,
94 zmm10, zmm11, zmm12, zmm13, zmm14, zmm15, zmm16, zmm17, zmm18,
95 zmm19, zmm20, zmm21, zmm22, zmm23, zmm24, zmm25, zmm26, zmm27,
96 zmm28, zmm29, zmm30, zmm31;
97 };
98
99 /* Implemented in scalarargs.c */
100 extern struct IntegerRegisters iregs;
101 extern struct FloatRegisters fregs;
102 extern unsigned int num_iregs, num_fregs;
103
104 #define check_int_arguments do { \
105 assert (num_iregs <= 0 || iregs.I0 == I0); \
106 assert (num_iregs <= 1 || iregs.I1 == I1); \
107 assert (num_iregs <= 2 || iregs.I2 == I2); \
108 assert (num_iregs <= 3 || iregs.I3 == I3); \
109 assert (num_iregs <= 4 || iregs.I4 == I4); \
110 assert (num_iregs <= 5 || iregs.I5 == I5); \
111 } while (0)
112
113 #define check_char_arguments check_int_arguments
114 #define check_short_arguments check_int_arguments
115 #define check_long_arguments check_int_arguments
116
117 /* Clear register struct. */
118 #define clear_struct_registers \
119 rax = rbx = rcx = rdx = rdi = rsi = rbp = rsp \
120 = r8 = r9 = r10 = r11 = r12 = r13 = r14 = r15 = 0; \
121 memset (&iregs, 0, sizeof (iregs)); \
122 memset (&fregs, 0, sizeof (fregs)); \
123 memset (zmm_regs, 0, sizeof (zmm_regs)); \
124 memset (x87_regs, 0, sizeof (x87_regs));
125
126 /* Clear both hardware and register structs for integers. */
127 #define clear_int_registers \
128 clear_struct_registers \
129 clear_int_hardware_registers
130
131 /* TODO: Do the checking. */
132 #define check_f_arguments(T) do { \
133 assert (num_fregs <= 0 || fregs.zmm0._ ## T [0] == zmm_regs[0]._ ## T [0]); \
134 assert (num_fregs <= 1 || fregs.zmm1._ ## T [0] == zmm_regs[1]._ ## T [0]); \
135 assert (num_fregs <= 2 || fregs.zmm2._ ## T [0] == zmm_regs[2]._ ## T [0]); \
136 assert (num_fregs <= 3 || fregs.zmm3._ ## T [0] == zmm_regs[3]._ ## T [0]); \
137 assert (num_fregs <= 4 || fregs.zmm4._ ## T [0] == zmm_regs[4]._ ## T [0]); \
138 assert (num_fregs <= 5 || fregs.zmm5._ ## T [0] == zmm_regs[5]._ ## T [0]); \
139 assert (num_fregs <= 6 || fregs.zmm6._ ## T [0] == zmm_regs[6]._ ## T [0]); \
140 assert (num_fregs <= 7 || fregs.zmm7._ ## T [0] == zmm_regs[7]._ ## T [0]); \
141 } while (0)
142
143 #define check_float_arguments check_f_arguments(float)
144 #define check_double_arguments check_f_arguments(double)
145
146 #define check_vector_arguments(T,O) do { \
147 assert (num_fregs <= 0 \
148 || memcmp (((char *) &fregs.zmm0) + (O), \
149 &zmm_regs[0], \
150 sizeof (__ ## T) - (O)) == 0); \
151 assert (num_fregs <= 1 \
152 || memcmp (((char *) &fregs.zmm1) + (O), \
153 &zmm_regs[1], \
154 sizeof (__ ## T) - (O)) == 0); \
155 assert (num_fregs <= 2 \
156 || memcmp (((char *) &fregs.zmm2) + (O), \
157 &zmm_regs[2], \
158 sizeof (__ ## T) - (O)) == 0); \
159 assert (num_fregs <= 3 \
160 || memcmp (((char *) &fregs.zmm3) + (O), \
161 &zmm_regs[3], \
162 sizeof (__ ## T) - (O)) == 0); \
163 assert (num_fregs <= 4 \
164 || memcmp (((char *) &fregs.zmm4) + (O), \
165 &zmm_regs[4], \
166 sizeof (__ ## T) - (O)) == 0); \
167 assert (num_fregs <= 5 \
168 || memcmp (((char *) &fregs.zmm5) + (O), \
169 &zmm_regs[5], \
170 sizeof (__ ## T) - (O)) == 0); \
171 assert (num_fregs <= 6 \
172 || memcmp (((char *) &fregs.zmm6) + (O), \
173 &zmm_regs[6], \
174 sizeof (__ ## T) - (O)) == 0); \
175 assert (num_fregs <= 7 \
176 || memcmp (((char *) &fregs.zmm7) + (O), \
177 &zmm_regs[7], \
178 sizeof (__ ## T) - (O)) == 0); \
179 } while (0)
180
181 #define check_m64_arguments check_vector_arguments(m64, 0)
182 #define check_m128_arguments check_vector_arguments(m128, 0)
183 #define check_m256_arguments check_vector_arguments(m256, 0)
184 #define check_m512_arguments check_vector_arguments(m512, 0)
185
186 #endif /* INCLUDED_ARGS_H */