Cell: some basic blending code
[mesa.git] / src / mesa / x86 / rtasm / x86sse.h
1
2 #ifndef _X86SSE_H_
3 #define _X86SSE_H_
4
5 #if defined(__i386__) || defined(__386__)
6
7 /* It is up to the caller to ensure that instructions issued are
8 * suitable for the host cpu. There are no checks made in this module
9 * for mmx/sse/sse2 support on the cpu.
10 */
11 struct x86_reg {
12 unsigned file:3;
13 unsigned idx:3;
14 unsigned mod:2; /* mod_REG if this is just a register */
15 int disp:24; /* only +/- 23bits of offset - should be enough... */
16 };
17
18 struct x86_function {
19 unsigned char *store;
20 unsigned char *csr;
21 unsigned stack_offset;
22 int need_emms;
23 const char *fn;
24 };
25
26 enum x86_reg_file {
27 file_REG32,
28 file_MMX,
29 file_XMM,
30 file_x87
31 };
32
33 /* Values for mod field of modr/m byte
34 */
35 enum x86_reg_mod {
36 mod_INDIRECT,
37 mod_DISP8,
38 mod_DISP32,
39 mod_REG
40 };
41
42 enum x86_reg_name {
43 reg_AX,
44 reg_CX,
45 reg_DX,
46 reg_BX,
47 reg_SP,
48 reg_BP,
49 reg_SI,
50 reg_DI
51 };
52
53
54 enum x86_cc {
55 cc_O, /* overflow */
56 cc_NO, /* not overflow */
57 cc_NAE, /* not above or equal / carry */
58 cc_AE, /* above or equal / not carry */
59 cc_E, /* equal / zero */
60 cc_NE /* not equal / not zero */
61 };
62
63 enum sse_cc {
64 cc_Equal,
65 cc_LessThan,
66 cc_LessThanEqual,
67 cc_Unordered,
68 cc_NotEqual,
69 cc_NotLessThan,
70 cc_NotLessThanEqual,
71 cc_Ordered
72 };
73
74 #define cc_Z cc_E
75 #define cc_NZ cc_NE
76
77 /* Begin/end/retreive function creation:
78 */
79
80
81 void x86_init_func( struct x86_function *p );
82 void x86_init_func_size( struct x86_function *p, unsigned code_size );
83 void x86_release_func( struct x86_function *p );
84 void (*x86_get_func( struct x86_function *p ))( void );
85
86
87
88 /* Create and manipulate registers and regmem values:
89 */
90 struct x86_reg x86_make_reg( enum x86_reg_file file,
91 enum x86_reg_name idx );
92
93 struct x86_reg x86_make_disp( struct x86_reg reg,
94 int disp );
95
96 struct x86_reg x86_deref( struct x86_reg reg );
97
98 struct x86_reg x86_get_base_reg( struct x86_reg reg );
99
100
101 /* Labels, jumps and fixup:
102 */
103 unsigned char *x86_get_label( struct x86_function *p );
104
105 void x86_jcc( struct x86_function *p,
106 enum x86_cc cc,
107 unsigned char *label );
108
109 unsigned char *x86_jcc_forward( struct x86_function *p,
110 enum x86_cc cc );
111
112 unsigned char *x86_jmp_forward( struct x86_function *p);
113
114 unsigned char *x86_call_forward( struct x86_function *p);
115
116 void x86_fixup_fwd_jump( struct x86_function *p,
117 unsigned char *fixup );
118
119 void x86_jmp( struct x86_function *p, unsigned char *label );
120
121 void x86_call( struct x86_function *p, void (*label)() );
122
123 /* michal:
124 * Temporary. As I need immediate operands, and dont want to mess with the codegen,
125 * I load the immediate into general purpose register and use it.
126 */
127 void x86_mov_reg_imm( struct x86_function *p, struct x86_reg dst, int imm );
128
129
130 /* Macro for sse_shufps() and sse2_pshufd():
131 */
132 #define SHUF(_x,_y,_z,_w) (((_x)<<0) | ((_y)<<2) | ((_z)<<4) | ((_w)<<6))
133 #define SHUF_NOOP RSW(0,1,2,3)
134 #define GET_SHUF(swz, idx) (((swz) >> ((idx)*2)) & 0x3)
135
136 void mmx_emms( struct x86_function *p );
137 void mmx_movd( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
138 void mmx_movq( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
139 void mmx_packssdw( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
140 void mmx_packuswb( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
141
142 void sse2_cvtps2dq( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
143 void sse2_cvttps2dq( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
144 void sse2_movd( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
145 void sse2_packssdw( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
146 void sse2_packsswb( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
147 void sse2_packuswb( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
148 void sse2_pshufd( struct x86_function *p, struct x86_reg dest, struct x86_reg arg0,
149 unsigned char shuf );
150 void sse2_rcpps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
151 void sse2_rcpss( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
152
153 void sse_addps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
154 void sse_addss( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
155 void sse_cvtps2pi( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
156 void sse_divss( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
157 void sse_andnps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
158 void sse_andps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
159 void sse_cmpps( struct x86_function *p, struct x86_reg dst, struct x86_reg src,
160 unsigned char cc );
161 void sse_maxps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
162 void sse_maxss( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
163 void sse_minps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
164 void sse_movaps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
165 void sse_movhlps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
166 void sse_movhps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
167 void sse_movlhps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
168 void sse_movlps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
169 void sse_movss( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
170 void sse_movups( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
171 void sse_mulps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
172 void sse_mulss( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
173 void sse_orps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
174 void sse_xorps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
175 void sse_subps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
176 void sse_rsqrtps( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
177 void sse_rsqrtss( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
178 void sse_shufps( struct x86_function *p, struct x86_reg dest, struct x86_reg arg0,
179 unsigned char shuf );
180 void sse_pmovmskb( struct x86_function *p, struct x86_reg dest, struct x86_reg src );
181
182 void x86_add( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
183 void x86_and( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
184 void x86_cmp( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
185 void x86_dec( struct x86_function *p, struct x86_reg reg );
186 void x86_inc( struct x86_function *p, struct x86_reg reg );
187 void x86_lea( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
188 void x86_mov( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
189 void x86_mul( struct x86_function *p, struct x86_reg src );
190 void x86_or( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
191 void x86_pop( struct x86_function *p, struct x86_reg reg );
192 void x86_push( struct x86_function *p, struct x86_reg reg );
193 void x86_ret( struct x86_function *p );
194 void x86_sub( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
195 void x86_test( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
196 void x86_xor( struct x86_function *p, struct x86_reg dst, struct x86_reg src );
197 void x86_sahf( struct x86_function *p );
198
199 void x87_f2xm1( struct x86_function *p );
200 void x87_fabs( struct x86_function *p );
201 void x87_fadd( struct x86_function *p, struct x86_reg dst, struct x86_reg arg );
202 void x87_faddp( struct x86_function *p, struct x86_reg dst );
203 void x87_fchs( struct x86_function *p );
204 void x87_fclex( struct x86_function *p );
205 void x87_fcom( struct x86_function *p, struct x86_reg dst );
206 void x87_fcomp( struct x86_function *p, struct x86_reg dst );
207 void x87_fcos( struct x86_function *p );
208 void x87_fdiv( struct x86_function *p, struct x86_reg dst, struct x86_reg arg );
209 void x87_fdivp( struct x86_function *p, struct x86_reg dst );
210 void x87_fdivr( struct x86_function *p, struct x86_reg dst, struct x86_reg arg );
211 void x87_fdivrp( struct x86_function *p, struct x86_reg dst );
212 void x87_fild( struct x86_function *p, struct x86_reg arg );
213 void x87_fist( struct x86_function *p, struct x86_reg dst );
214 void x87_fistp( struct x86_function *p, struct x86_reg dst );
215 void x87_fld( struct x86_function *p, struct x86_reg arg );
216 void x87_fld1( struct x86_function *p );
217 void x87_fldcw( struct x86_function *p, struct x86_reg arg );
218 void x87_fldl2e( struct x86_function *p );
219 void x87_fldln2( struct x86_function *p );
220 void x87_fldz( struct x86_function *p );
221 void x87_fmul( struct x86_function *p, struct x86_reg dst, struct x86_reg arg );
222 void x87_fmulp( struct x86_function *p, struct x86_reg dst );
223 void x87_fnclex( struct x86_function *p );
224 void x87_fprndint( struct x86_function *p );
225 void x87_fscale( struct x86_function *p );
226 void x87_fsin( struct x86_function *p );
227 void x87_fsincos( struct x86_function *p );
228 void x87_fsqrt( struct x86_function *p );
229 void x87_fst( struct x86_function *p, struct x86_reg dst );
230 void x87_fstp( struct x86_function *p, struct x86_reg dst );
231 void x87_fsub( struct x86_function *p, struct x86_reg dst, struct x86_reg arg );
232 void x87_fsubp( struct x86_function *p, struct x86_reg dst );
233 void x87_fsubr( struct x86_function *p, struct x86_reg dst, struct x86_reg arg );
234 void x87_fsubrp( struct x86_function *p, struct x86_reg dst );
235 void x87_fxch( struct x86_function *p, struct x86_reg dst );
236 void x87_fxtract( struct x86_function *p );
237 void x87_fyl2x( struct x86_function *p );
238 void x87_fyl2xp1( struct x86_function *p );
239 void x87_fwait( struct x86_function *p );
240 void x87_fnstsw( struct x86_function *p, struct x86_reg dst );
241 void x87_fucompp( struct x86_function *p );
242 void x87_fucomp( struct x86_function *p, struct x86_reg arg );
243 void x87_fucom( struct x86_function *p, struct x86_reg arg );
244
245
246
247 /* Retreive a reference to one of the function arguments, taking into
248 * account any push/pop activity. Note - doesn't track explict
249 * manipulation of ESP by other instructions.
250 */
251 struct x86_reg x86_fn_arg( struct x86_function *p, unsigned arg );
252
253 #endif
254 #endif