i965: Add a note about an unsafe-looking state check.
[mesa.git] / src / mesa / drivers / dri / i965 / brw_eu.c
1 /*
2 Copyright (C) Intel Corp. 2006. All Rights Reserved.
3 Intel funded Tungsten Graphics (http://www.tungstengraphics.com) to
4 develop this 3D driver.
5
6 Permission is hereby granted, free of charge, to any person obtaining
7 a copy of this software and associated documentation files (the
8 "Software"), to deal in the Software without restriction, including
9 without limitation the rights to use, copy, modify, merge, publish,
10 distribute, sublicense, and/or sell copies of the Software, and to
11 permit persons to whom the Software is furnished to do so, subject to
12 the following conditions:
13
14 The above copyright notice and this permission notice (including the
15 next paragraph) shall be included in all copies or substantial
16 portions of the Software.
17
18 THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
19 EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
21 IN NO EVENT SHALL THE COPYRIGHT OWNER(S) AND/OR ITS SUPPLIERS BE
22 LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
23 OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
24 WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
25
26 **********************************************************************/
27 /*
28 * Authors:
29 * Keith Whitwell <keith@tungstengraphics.com>
30 */
31
32
33 #include "brw_context.h"
34 #include "brw_defines.h"
35 #include "brw_eu.h"
36
37 #include "glsl/ralloc.h"
38
39 /* Returns the corresponding conditional mod for swapping src0 and
40 * src1 in e.g. CMP.
41 */
42 uint32_t
43 brw_swap_cmod(uint32_t cmod)
44 {
45 switch (cmod) {
46 case BRW_CONDITIONAL_Z:
47 case BRW_CONDITIONAL_NZ:
48 return cmod;
49 case BRW_CONDITIONAL_G:
50 return BRW_CONDITIONAL_LE;
51 case BRW_CONDITIONAL_GE:
52 return BRW_CONDITIONAL_L;
53 case BRW_CONDITIONAL_L:
54 return BRW_CONDITIONAL_GE;
55 case BRW_CONDITIONAL_LE:
56 return BRW_CONDITIONAL_G;
57 default:
58 return ~0;
59 }
60 }
61
62
63 /* How does predicate control work when execution_size != 8? Do I
64 * need to test/set for 0xffff when execution_size is 16?
65 */
66 void brw_set_predicate_control_flag_value( struct brw_compile *p, GLuint value )
67 {
68 p->current->header.predicate_control = BRW_PREDICATE_NONE;
69
70 if (value != 0xff) {
71 if (value != p->flag_value) {
72 brw_push_insn_state(p);
73 brw_MOV(p, brw_flag_reg(), brw_imm_uw(value));
74 p->flag_value = value;
75 brw_pop_insn_state(p);
76 }
77
78 p->current->header.predicate_control = BRW_PREDICATE_NORMAL;
79 }
80 }
81
82 void brw_set_predicate_control( struct brw_compile *p, GLuint pc )
83 {
84 p->current->header.predicate_control = pc;
85 }
86
87 void brw_set_predicate_inverse(struct brw_compile *p, bool predicate_inverse)
88 {
89 p->current->header.predicate_inverse = predicate_inverse;
90 }
91
92 void brw_set_conditionalmod( struct brw_compile *p, GLuint conditional )
93 {
94 p->current->header.destreg__conditionalmod = conditional;
95 }
96
97 void brw_set_access_mode( struct brw_compile *p, GLuint access_mode )
98 {
99 p->current->header.access_mode = access_mode;
100 }
101
102 void
103 brw_set_compression_control(struct brw_compile *p,
104 enum brw_compression compression_control)
105 {
106 p->compressed = (compression_control == BRW_COMPRESSION_COMPRESSED);
107
108 if (p->brw->intel.gen >= 6) {
109 /* Since we don't use the 32-wide support in gen6, we translate
110 * the pre-gen6 compression control here.
111 */
112 switch (compression_control) {
113 case BRW_COMPRESSION_NONE:
114 /* This is the "use the first set of bits of dmask/vmask/arf
115 * according to execsize" option.
116 */
117 p->current->header.compression_control = GEN6_COMPRESSION_1Q;
118 break;
119 case BRW_COMPRESSION_2NDHALF:
120 /* For 8-wide, this is "use the second set of 8 bits." */
121 p->current->header.compression_control = GEN6_COMPRESSION_2Q;
122 break;
123 case BRW_COMPRESSION_COMPRESSED:
124 /* For 16-wide instruction compression, use the first set of 16 bits
125 * since we don't do 32-wide dispatch.
126 */
127 p->current->header.compression_control = GEN6_COMPRESSION_1H;
128 break;
129 default:
130 assert(!"not reached");
131 p->current->header.compression_control = GEN6_COMPRESSION_1H;
132 break;
133 }
134 } else {
135 p->current->header.compression_control = compression_control;
136 }
137 }
138
139 void brw_set_mask_control( struct brw_compile *p, GLuint value )
140 {
141 p->current->header.mask_control = value;
142 }
143
144 void brw_set_saturate( struct brw_compile *p, GLuint value )
145 {
146 p->current->header.saturate = value;
147 }
148
149 void brw_set_acc_write_control(struct brw_compile *p, GLuint value)
150 {
151 if (p->brw->intel.gen >= 6)
152 p->current->header.acc_wr_control = value;
153 }
154
155 void brw_push_insn_state( struct brw_compile *p )
156 {
157 assert(p->current != &p->stack[BRW_EU_MAX_INSN_STACK-1]);
158 memcpy(p->current+1, p->current, sizeof(struct brw_instruction));
159 p->compressed_stack[p->current - p->stack] = p->compressed;
160 p->current++;
161 }
162
163 void brw_pop_insn_state( struct brw_compile *p )
164 {
165 assert(p->current != p->stack);
166 p->current--;
167 p->compressed = p->compressed_stack[p->current - p->stack];
168 }
169
170
171 /***********************************************************************
172 */
173 void
174 brw_init_compile(struct brw_context *brw, struct brw_compile *p, void *mem_ctx)
175 {
176 p->brw = brw;
177 p->nr_insn = 0;
178 p->current = p->stack;
179 p->compressed = false;
180 memset(p->current, 0, sizeof(p->current[0]));
181
182 p->mem_ctx = mem_ctx;
183
184 /* Some defaults?
185 */
186 brw_set_mask_control(p, BRW_MASK_ENABLE); /* what does this do? */
187 brw_set_saturate(p, 0);
188 brw_set_compression_control(p, BRW_COMPRESSION_NONE);
189 brw_set_predicate_control_flag_value(p, 0xff);
190
191 /* Set up control flow stack */
192 p->if_stack_depth = 0;
193 p->if_stack_array_size = 16;
194 p->if_stack =
195 rzalloc_array(mem_ctx, struct brw_instruction *, p->if_stack_array_size);
196 }
197
198
199 const GLuint *brw_get_program( struct brw_compile *p,
200 GLuint *sz )
201 {
202 GLuint i;
203
204 for (i = 0; i < 8; i++)
205 brw_NOP(p);
206
207 *sz = p->nr_insn * sizeof(struct brw_instruction);
208 return (const GLuint *)p->store;
209 }
210
211
212
213 /**
214 * Subroutine calls require special attention.
215 * Mesa instructions may be expanded into multiple hardware instructions
216 * so the prog_instruction::BranchTarget field can't be used as an index
217 * into the hardware instructions.
218 *
219 * The BranchTarget field isn't needed, however. Mesa's GLSL compiler
220 * emits CAL and BGNSUB instructions with labels that can be used to map
221 * subroutine calls to actual subroutine code blocks.
222 *
223 * The structures and function here implement patching of CAL instructions
224 * so they jump to the right subroutine code...
225 */
226
227
228 /**
229 * For each OPCODE_BGNSUB we create one of these.
230 */
231 struct brw_glsl_label
232 {
233 const char *name; /**< the label string */
234 GLuint position; /**< the position of the brw instruction for this label */
235 struct brw_glsl_label *next; /**< next in linked list */
236 };
237
238
239 /**
240 * For each OPCODE_CAL we create one of these.
241 */
242 struct brw_glsl_call
243 {
244 GLuint call_inst_pos; /**< location of the CAL instruction */
245 const char *sub_name; /**< name of subroutine to call */
246 struct brw_glsl_call *next; /**< next in linked list */
247 };
248
249
250 /**
251 * Called for each OPCODE_BGNSUB.
252 */
253 void
254 brw_save_label(struct brw_compile *c, const char *name, GLuint position)
255 {
256 struct brw_glsl_label *label = CALLOC_STRUCT(brw_glsl_label);
257 label->name = name;
258 label->position = position;
259 label->next = c->first_label;
260 c->first_label = label;
261 }
262
263
264 /**
265 * Called for each OPCODE_CAL.
266 */
267 void
268 brw_save_call(struct brw_compile *c, const char *name, GLuint call_pos)
269 {
270 struct brw_glsl_call *call = CALLOC_STRUCT(brw_glsl_call);
271 call->call_inst_pos = call_pos;
272 call->sub_name = name;
273 call->next = c->first_call;
274 c->first_call = call;
275 }
276
277
278 /**
279 * Lookup a label, return label's position/offset.
280 */
281 static GLuint
282 brw_lookup_label(struct brw_compile *c, const char *name)
283 {
284 const struct brw_glsl_label *label;
285 for (label = c->first_label; label; label = label->next) {
286 if (strcmp(name, label->name) == 0) {
287 return label->position;
288 }
289 }
290 abort(); /* should never happen */
291 return ~0;
292 }
293
294
295 /**
296 * When we're done generating code, this function is called to resolve
297 * subroutine calls.
298 */
299 void
300 brw_resolve_cals(struct brw_compile *c)
301 {
302 const struct brw_glsl_call *call;
303
304 for (call = c->first_call; call; call = call->next) {
305 const GLuint sub_loc = brw_lookup_label(c, call->sub_name);
306 struct brw_instruction *brw_call_inst = &c->store[call->call_inst_pos];
307 struct brw_instruction *brw_sub_inst = &c->store[sub_loc];
308 GLint offset = brw_sub_inst - brw_call_inst;
309
310 /* patch brw_inst1 to point to brw_inst2 */
311 brw_set_src1(c, brw_call_inst, brw_imm_d(offset * 16));
312 }
313
314 /* free linked list of calls */
315 {
316 struct brw_glsl_call *call, *next;
317 for (call = c->first_call; call; call = next) {
318 next = call->next;
319 free(call);
320 }
321 c->first_call = NULL;
322 }
323
324 /* free linked list of labels */
325 {
326 struct brw_glsl_label *label, *next;
327 for (label = c->first_label; label; label = next) {
328 next = label->next;
329 free(label);
330 }
331 c->first_label = NULL;
332 }
333 }