llvmpipe: Add a bunch of comments.
[mesa.git] / src / gallium / drivers / llvmpipe / lp_bld_logic.c
1 /**************************************************************************
2 *
3 * Copyright 2009 VMware, Inc.
4 * All Rights Reserved.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * 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, sub license, 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 portions
16 * of the Software.
17 *
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
21 * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
22 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
23 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
24 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
25 *
26 **************************************************************************/
27
28 /**
29 * @file
30 * Helper functions for logical operations.
31 *
32 * @author Jose Fonseca <jfonseca@vmware.com>
33 */
34
35
36 #include "pipe/p_defines.h"
37 #include "lp_bld_type.h"
38 #include "lp_bld_const.h"
39 #include "lp_bld_intr.h"
40 #include "lp_bld_logic.h"
41
42
43 LLVMValueRef
44 lp_build_cmp(struct lp_build_context *bld,
45 unsigned func,
46 LLVMValueRef a,
47 LLVMValueRef b)
48 {
49 const union lp_type type = bld->type;
50 LLVMTypeRef vec_type = lp_build_vec_type(type);
51 LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
52 LLVMValueRef zeros = LLVMConstNull(int_vec_type);
53 LLVMValueRef ones = LLVMConstAllOnes(int_vec_type);
54 LLVMValueRef cond;
55
56 if(func == PIPE_FUNC_NEVER)
57 return zeros;
58 if(func == PIPE_FUNC_ALWAYS)
59 return ones;
60
61 /* TODO: optimize the constant case */
62
63 /* XXX: It is not clear if we should use the ordered or unordered operators */
64
65 #if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
66 if(type.width * type.length == 128) {
67 if(type.floating) {
68 LLVMValueRef args[3];
69 unsigned cc;
70 boolean swap;
71 LLVMValueRef res;
72
73 swap = FALSE;
74 switch(func) {
75 case PIPE_FUNC_EQUAL:
76 cc = 0;
77 break;
78 case PIPE_FUNC_NOTEQUAL:
79 cc = 4;
80 break;
81 case PIPE_FUNC_LESS:
82 cc = 1;
83 break;
84 case PIPE_FUNC_LEQUAL:
85 cc = 2;
86 break;
87 case PIPE_FUNC_GREATER:
88 cc = 1;
89 swap = TRUE;
90 break;
91 case PIPE_FUNC_GEQUAL:
92 cc = 2;
93 swap = TRUE;
94 break;
95 default:
96 assert(0);
97 return bld->undef;
98 }
99
100 if(swap) {
101 args[0] = b;
102 args[1] = a;
103 }
104 else {
105 args[0] = a;
106 args[1] = b;
107 }
108
109 args[2] = LLVMConstInt(LLVMInt8Type(), cc, 0);
110 res = lp_build_intrinsic(bld->builder,
111 "llvm.x86.sse.cmp.ps",
112 vec_type,
113 args, 3);
114 res = LLVMBuildBitCast(bld->builder, res, int_vec_type, "");
115 return res;
116 }
117 else {
118 static const struct {
119 unsigned swap:1;
120 unsigned eq:1;
121 unsigned gt:1;
122 unsigned not:1;
123 } table[] = {
124 {0, 0, 0, 1}, /* PIPE_FUNC_NEVER */
125 {1, 0, 1, 0}, /* PIPE_FUNC_LESS */
126 {0, 1, 0, 0}, /* PIPE_FUNC_EQUAL */
127 {0, 0, 1, 1}, /* PIPE_FUNC_LEQUAL */
128 {0, 0, 1, 0}, /* PIPE_FUNC_GREATER */
129 {0, 1, 0, 1}, /* PIPE_FUNC_NOTEQUAL */
130 {1, 0, 1, 1}, /* PIPE_FUNC_GEQUAL */
131 {0, 0, 0, 0} /* PIPE_FUNC_ALWAYS */
132 };
133 const char *pcmpeq;
134 const char *pcmpgt;
135 LLVMValueRef args[2];
136 LLVMValueRef res;
137
138 switch (type.width) {
139 case 8:
140 pcmpeq = "llvm.x86.sse2.pcmpeq.b";
141 pcmpgt = "llvm.x86.sse2.pcmpgt.b";
142 break;
143 case 16:
144 pcmpeq = "llvm.x86.sse2.pcmpeq.w";
145 pcmpgt = "llvm.x86.sse2.pcmpgt.w";
146 break;
147 case 32:
148 pcmpeq = "llvm.x86.sse2.pcmpeq.d";
149 pcmpgt = "llvm.x86.sse2.pcmpgt.d";
150 break;
151 default:
152 assert(0);
153 return bld->undef;
154 }
155
156 /* There are no signed byte and unsigned word/dword comparison
157 * instructions. So flip the sign bit so that the results match.
158 */
159 if(table[func].gt &&
160 ((type.width == 8 && type.sign) ||
161 (type.width != 8 && !type.sign))) {
162 LLVMValueRef msb = lp_build_int_const_uni(type, (unsigned long long)1 << (type.width - 1));
163 a = LLVMBuildXor(bld->builder, a, msb, "");
164 b = LLVMBuildXor(bld->builder, b, msb, "");
165 }
166
167 if(table[func].swap) {
168 args[0] = b;
169 args[1] = a;
170 }
171 else {
172 args[0] = a;
173 args[1] = b;
174 }
175
176 if(table[func].eq)
177 res = lp_build_intrinsic(bld->builder, pcmpeq, vec_type, args, 2);
178 else if (table[func].gt)
179 res = lp_build_intrinsic(bld->builder, pcmpgt, vec_type, args, 2);
180 else
181 res = LLVMConstNull(vec_type);
182
183 if(table[func].not)
184 res = LLVMBuildNot(bld->builder, res, "");
185
186 return res;
187 }
188 }
189 #endif
190
191 if(type.floating) {
192 LLVMRealPredicate op;
193 switch(func) {
194 case PIPE_FUNC_NEVER:
195 op = LLVMRealPredicateFalse;
196 break;
197 case PIPE_FUNC_ALWAYS:
198 op = LLVMRealPredicateTrue;
199 break;
200 case PIPE_FUNC_EQUAL:
201 op = LLVMRealUEQ;
202 break;
203 case PIPE_FUNC_NOTEQUAL:
204 op = LLVMRealUNE;
205 break;
206 case PIPE_FUNC_LESS:
207 op = LLVMRealULT;
208 break;
209 case PIPE_FUNC_LEQUAL:
210 op = LLVMRealULE;
211 break;
212 case PIPE_FUNC_GREATER:
213 op = LLVMRealUGT;
214 break;
215 case PIPE_FUNC_GEQUAL:
216 op = LLVMRealUGE;
217 break;
218 default:
219 assert(0);
220 return bld->undef;
221 }
222 cond = LLVMBuildFCmp(bld->builder, op, a, b, "");
223 }
224 else {
225 LLVMIntPredicate op;
226 switch(func) {
227 case PIPE_FUNC_EQUAL:
228 op = LLVMIntEQ;
229 break;
230 case PIPE_FUNC_NOTEQUAL:
231 op = LLVMIntNE;
232 break;
233 case PIPE_FUNC_LESS:
234 op = type.sign ? LLVMIntSLT : LLVMIntULT;
235 break;
236 case PIPE_FUNC_LEQUAL:
237 op = type.sign ? LLVMIntSLE : LLVMIntULE;
238 break;
239 case PIPE_FUNC_GREATER:
240 op = type.sign ? LLVMIntSGT : LLVMIntUGT;
241 break;
242 case PIPE_FUNC_GEQUAL:
243 op = type.sign ? LLVMIntSGE : LLVMIntUGE;
244 break;
245 default:
246 assert(0);
247 return bld->undef;
248 }
249 cond = LLVMBuildICmp(bld->builder, op, a, b, "");
250 }
251
252 return LLVMBuildSelect(bld->builder, cond, ones, zeros, "");
253 }
254
255
256 LLVMValueRef
257 lp_build_select(struct lp_build_context *bld,
258 LLVMValueRef mask,
259 LLVMValueRef a,
260 LLVMValueRef b)
261 {
262 union lp_type type = bld->type;
263 LLVMValueRef res;
264
265 if(a == b)
266 return a;
267
268 if(type.floating) {
269 LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
270 a = LLVMBuildBitCast(bld->builder, a, int_vec_type, "");
271 b = LLVMBuildBitCast(bld->builder, b, int_vec_type, "");
272 }
273
274 /* TODO: On SSE4 we could do this with a single instruction -- PBLENDVB */
275
276 a = LLVMBuildAnd(bld->builder, a, mask, "");
277
278 /* This often gets translated to PANDN, but sometimes the NOT is
279 * pre-computed and stored in another constant. The best strategy depends
280 * on available registers, so it is not a big deal -- hopefully LLVM does
281 * the right decision attending the rest of the program.
282 */
283 b = LLVMBuildAnd(bld->builder, b, LLVMBuildNot(bld->builder, mask, ""), "");
284
285 res = LLVMBuildOr(bld->builder, a, b, "");
286
287 if(type.floating) {
288 LLVMTypeRef vec_type = lp_build_vec_type(type);
289 res = LLVMBuildBitCast(bld->builder, res, vec_type, "");
290 }
291
292 return res;
293 }
294
295
296 LLVMValueRef
297 lp_build_select_aos(struct lp_build_context *bld,
298 LLVMValueRef a,
299 LLVMValueRef b,
300 boolean cond[4])
301 {
302 const union lp_type type = bld->type;
303 const unsigned n = type.length;
304 unsigned i, j;
305
306 if(a == b)
307 return a;
308 if(cond[0] && cond[1] && cond[2] && cond[3])
309 return a;
310 if(!cond[0] && !cond[1] && !cond[2] && !cond[3])
311 return b;
312 if(a == bld->undef || b == bld->undef)
313 return bld->undef;
314
315 /*
316 * There are three major ways of accomplishing this:
317 * - with a shuffle,
318 * - with a select,
319 * - or with a bit mask.
320 *
321 * Select isn't supported for vector types yet.
322 * The flip between these is empirical and might need to be.
323 */
324 if (n <= 4) {
325 /*
326 * Shuffle.
327 */
328 LLVMTypeRef elem_type = LLVMInt32Type();
329 LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
330
331 for(j = 0; j < n; j += 4)
332 for(i = 0; i < 4; ++i)
333 shuffles[j + i] = LLVMConstInt(elem_type, (cond[i] ? 0 : n) + j + i, 0);
334
335 return LLVMBuildShuffleVector(bld->builder, a, b, LLVMConstVector(shuffles, n), "");
336 }
337 #if 0
338 else if(0) {
339 /* FIXME: Unfortunately select of vectors do not work */
340 /* Use a select */
341 LLVMTypeRef elem_type = LLVMInt1Type();
342 LLVMValueRef cond[LP_MAX_VECTOR_LENGTH];
343
344 for(j = 0; j < n; j += 4)
345 for(i = 0; i < 4; ++i)
346 cond[j + i] = LLVMConstInt(elem_type, cond[i] ? 1 : 0, 0);
347
348 return LLVMBuildSelect(bld->builder, LLVMConstVector(cond, n), a, b, "");
349 }
350 #endif
351 else {
352 LLVMValueRef mask = lp_build_const_mask_aos(type, cond);
353 return lp_build_select(bld, mask, a, b);
354 }
355 }