gallivm: Fix 8bit comparisons.
[mesa.git] / src / gallium / auxiliary / gallivm / 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 "util/u_cpu_detect.h"
37 #include "util/u_debug.h"
38
39 #include "lp_bld_type.h"
40 #include "lp_bld_const.h"
41 #include "lp_bld_intr.h"
42 #include "lp_bld_logic.h"
43
44
45 /*
46 * XXX
47 *
48 * Selection with vector conditional like
49 *
50 * select <4 x i1> %C, %A, %B
51 *
52 * is valid IR (e.g. llvm/test/Assembler/vector-select.ll), but it is not
53 * supported on any backend.
54 *
55 * Expanding the boolean vector to full SIMD register width, as in
56 *
57 * sext <4 x i1> %C to <4 x i32>
58 *
59 * is valid and supported (e.g., llvm/test/CodeGen/X86/vec_compare.ll), but
60 * it causes assertion failures in LLVM 2.6. It appears to work correctly on
61 * LLVM 2.7.
62 */
63
64
65 /**
66 * Build code to compare two values 'a' and 'b' of 'type' using the given func.
67 * \param func one of PIPE_FUNC_x
68 * The result values will be 0 for false or ~0 for true.
69 */
70 LLVMValueRef
71 lp_build_compare(LLVMBuilderRef builder,
72 const struct lp_type type,
73 unsigned func,
74 LLVMValueRef a,
75 LLVMValueRef b)
76 {
77 LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
78 LLVMValueRef zeros = LLVMConstNull(int_vec_type);
79 LLVMValueRef ones = LLVMConstAllOnes(int_vec_type);
80 LLVMValueRef cond;
81 LLVMValueRef res;
82
83 assert(func >= PIPE_FUNC_NEVER);
84 assert(func <= PIPE_FUNC_ALWAYS);
85
86 if(func == PIPE_FUNC_NEVER)
87 return zeros;
88 if(func == PIPE_FUNC_ALWAYS)
89 return ones;
90
91 /* TODO: optimize the constant case */
92
93 /* XXX: It is not clear if we should use the ordered or unordered operators */
94
95 #if HAVE_LLVM < 0x0207
96 #if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
97 if(type.width * type.length == 128) {
98 if(type.floating && util_cpu_caps.has_sse) {
99 /* float[4] comparison */
100 LLVMTypeRef vec_type = lp_build_vec_type(type);
101 LLVMValueRef args[3];
102 unsigned cc;
103 boolean swap;
104
105 swap = FALSE;
106 switch(func) {
107 case PIPE_FUNC_EQUAL:
108 cc = 0;
109 break;
110 case PIPE_FUNC_NOTEQUAL:
111 cc = 4;
112 break;
113 case PIPE_FUNC_LESS:
114 cc = 1;
115 break;
116 case PIPE_FUNC_LEQUAL:
117 cc = 2;
118 break;
119 case PIPE_FUNC_GREATER:
120 cc = 1;
121 swap = TRUE;
122 break;
123 case PIPE_FUNC_GEQUAL:
124 cc = 2;
125 swap = TRUE;
126 break;
127 default:
128 assert(0);
129 return lp_build_undef(type);
130 }
131
132 if(swap) {
133 args[0] = b;
134 args[1] = a;
135 }
136 else {
137 args[0] = a;
138 args[1] = b;
139 }
140
141 args[2] = LLVMConstInt(LLVMInt8Type(), cc, 0);
142 res = lp_build_intrinsic(builder,
143 "llvm.x86.sse.cmp.ps",
144 vec_type,
145 args, 3);
146 res = LLVMBuildBitCast(builder, res, int_vec_type, "");
147 return res;
148 }
149 else if(util_cpu_caps.has_sse2) {
150 /* int[4] comparison */
151 static const struct {
152 unsigned swap:1;
153 unsigned eq:1;
154 unsigned gt:1;
155 unsigned not:1;
156 } table[] = {
157 {0, 0, 0, 1}, /* PIPE_FUNC_NEVER */
158 {1, 0, 1, 0}, /* PIPE_FUNC_LESS */
159 {0, 1, 0, 0}, /* PIPE_FUNC_EQUAL */
160 {0, 0, 1, 1}, /* PIPE_FUNC_LEQUAL */
161 {0, 0, 1, 0}, /* PIPE_FUNC_GREATER */
162 {0, 1, 0, 1}, /* PIPE_FUNC_NOTEQUAL */
163 {1, 0, 1, 1}, /* PIPE_FUNC_GEQUAL */
164 {0, 0, 0, 0} /* PIPE_FUNC_ALWAYS */
165 };
166 const char *pcmpeq;
167 const char *pcmpgt;
168 LLVMValueRef args[2];
169 LLVMValueRef res;
170 LLVMTypeRef vec_type = lp_build_vec_type(type);
171
172 switch (type.width) {
173 case 8:
174 pcmpeq = "llvm.x86.sse2.pcmpeq.b";
175 pcmpgt = "llvm.x86.sse2.pcmpgt.b";
176 break;
177 case 16:
178 pcmpeq = "llvm.x86.sse2.pcmpeq.w";
179 pcmpgt = "llvm.x86.sse2.pcmpgt.w";
180 break;
181 case 32:
182 pcmpeq = "llvm.x86.sse2.pcmpeq.d";
183 pcmpgt = "llvm.x86.sse2.pcmpgt.d";
184 break;
185 default:
186 assert(0);
187 return lp_build_undef(type);
188 }
189
190 /* There are no unsigned comparison instructions. So flip the sign bit
191 * so that the results match.
192 */
193 if (table[func].gt && !type.sign) {
194 LLVMValueRef msb = lp_build_const_int_vec(type, (unsigned long long)1 << (type.width - 1));
195 a = LLVMBuildXor(builder, a, msb, "");
196 b = LLVMBuildXor(builder, b, msb, "");
197 }
198
199 if(table[func].swap) {
200 args[0] = b;
201 args[1] = a;
202 }
203 else {
204 args[0] = a;
205 args[1] = b;
206 }
207
208 if(table[func].eq)
209 res = lp_build_intrinsic(builder, pcmpeq, vec_type, args, 2);
210 else if (table[func].gt)
211 res = lp_build_intrinsic(builder, pcmpgt, vec_type, args, 2);
212 else
213 res = LLVMConstNull(vec_type);
214
215 if(table[func].not)
216 res = LLVMBuildNot(builder, res, "");
217
218 return res;
219 }
220 } /* if (type.width * type.length == 128) */
221 #endif
222 #endif /* HAVE_LLVM < 0x0207 */
223
224 if(type.floating) {
225 LLVMRealPredicate op;
226 switch(func) {
227 case PIPE_FUNC_NEVER:
228 op = LLVMRealPredicateFalse;
229 break;
230 case PIPE_FUNC_ALWAYS:
231 op = LLVMRealPredicateTrue;
232 break;
233 case PIPE_FUNC_EQUAL:
234 op = LLVMRealUEQ;
235 break;
236 case PIPE_FUNC_NOTEQUAL:
237 op = LLVMRealUNE;
238 break;
239 case PIPE_FUNC_LESS:
240 op = LLVMRealULT;
241 break;
242 case PIPE_FUNC_LEQUAL:
243 op = LLVMRealULE;
244 break;
245 case PIPE_FUNC_GREATER:
246 op = LLVMRealUGT;
247 break;
248 case PIPE_FUNC_GEQUAL:
249 op = LLVMRealUGE;
250 break;
251 default:
252 assert(0);
253 return lp_build_undef(type);
254 }
255
256 #if HAVE_LLVM >= 0x0207
257 cond = LLVMBuildFCmp(builder, op, a, b, "");
258 res = LLVMBuildSExt(builder, cond, int_vec_type, "");
259 #else
260 if (type.length == 1) {
261 cond = LLVMBuildFCmp(builder, op, a, b, "");
262 res = LLVMBuildSExt(builder, cond, int_vec_type, "");
263 }
264 else {
265 unsigned i;
266
267 res = LLVMGetUndef(int_vec_type);
268
269 debug_printf("%s: warning: using slow element-wise float"
270 " vector comparison\n", __FUNCTION__);
271 for (i = 0; i < type.length; ++i) {
272 LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
273 cond = LLVMBuildFCmp(builder, op,
274 LLVMBuildExtractElement(builder, a, index, ""),
275 LLVMBuildExtractElement(builder, b, index, ""),
276 "");
277 cond = LLVMBuildSelect(builder, cond,
278 LLVMConstExtractElement(ones, index),
279 LLVMConstExtractElement(zeros, index),
280 "");
281 res = LLVMBuildInsertElement(builder, res, cond, index, "");
282 }
283 }
284 #endif
285 }
286 else {
287 LLVMIntPredicate op;
288 switch(func) {
289 case PIPE_FUNC_EQUAL:
290 op = LLVMIntEQ;
291 break;
292 case PIPE_FUNC_NOTEQUAL:
293 op = LLVMIntNE;
294 break;
295 case PIPE_FUNC_LESS:
296 op = type.sign ? LLVMIntSLT : LLVMIntULT;
297 break;
298 case PIPE_FUNC_LEQUAL:
299 op = type.sign ? LLVMIntSLE : LLVMIntULE;
300 break;
301 case PIPE_FUNC_GREATER:
302 op = type.sign ? LLVMIntSGT : LLVMIntUGT;
303 break;
304 case PIPE_FUNC_GEQUAL:
305 op = type.sign ? LLVMIntSGE : LLVMIntUGE;
306 break;
307 default:
308 assert(0);
309 return lp_build_undef(type);
310 }
311
312 #if HAVE_LLVM >= 0x0207
313 cond = LLVMBuildICmp(builder, op, a, b, "");
314 res = LLVMBuildSExt(builder, cond, int_vec_type, "");
315 #else
316 if (type.length == 1) {
317 cond = LLVMBuildICmp(builder, op, a, b, "");
318 res = LLVMBuildSExt(builder, cond, int_vec_type, "");
319 }
320 else {
321 unsigned i;
322
323 res = LLVMGetUndef(int_vec_type);
324
325 debug_printf("%s: warning: using slow element-wise int"
326 " vector comparison\n", __FUNCTION__);
327
328 for(i = 0; i < type.length; ++i) {
329 LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
330 cond = LLVMBuildICmp(builder, op,
331 LLVMBuildExtractElement(builder, a, index, ""),
332 LLVMBuildExtractElement(builder, b, index, ""),
333 "");
334 cond = LLVMBuildSelect(builder, cond,
335 LLVMConstExtractElement(ones, index),
336 LLVMConstExtractElement(zeros, index),
337 "");
338 res = LLVMBuildInsertElement(builder, res, cond, index, "");
339 }
340 }
341 #endif
342 }
343
344 return res;
345 }
346
347
348
349 /**
350 * Build code to compare two values 'a' and 'b' using the given func.
351 * \param func one of PIPE_FUNC_x
352 * The result values will be 0 for false or ~0 for true.
353 */
354 LLVMValueRef
355 lp_build_cmp(struct lp_build_context *bld,
356 unsigned func,
357 LLVMValueRef a,
358 LLVMValueRef b)
359 {
360 return lp_build_compare(bld->builder, bld->type, func, a, b);
361 }
362
363
364 /**
365 * Return mask ? a : b;
366 *
367 * mask is a bitwise mask, composed of 0 or ~0 for each element.
368 */
369 LLVMValueRef
370 lp_build_select(struct lp_build_context *bld,
371 LLVMValueRef mask,
372 LLVMValueRef a,
373 LLVMValueRef b)
374 {
375 struct lp_type type = bld->type;
376 LLVMValueRef res;
377
378 if(a == b)
379 return a;
380
381 if (type.length == 1) {
382 mask = LLVMBuildTrunc(bld->builder, mask, LLVMInt1Type(), "");
383 res = LLVMBuildSelect(bld->builder, mask, a, b, "");
384 }
385 else {
386 if(type.floating) {
387 LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
388 a = LLVMBuildBitCast(bld->builder, a, int_vec_type, "");
389 b = LLVMBuildBitCast(bld->builder, b, int_vec_type, "");
390 }
391
392 a = LLVMBuildAnd(bld->builder, a, mask, "");
393
394 /* This often gets translated to PANDN, but sometimes the NOT is
395 * pre-computed and stored in another constant. The best strategy depends
396 * on available registers, so it is not a big deal -- hopefully LLVM does
397 * the right decision attending the rest of the program.
398 */
399 b = LLVMBuildAnd(bld->builder, b, LLVMBuildNot(bld->builder, mask, ""), "");
400
401 res = LLVMBuildOr(bld->builder, a, b, "");
402
403 if(type.floating) {
404 LLVMTypeRef vec_type = lp_build_vec_type(type);
405 res = LLVMBuildBitCast(bld->builder, res, vec_type, "");
406 }
407 }
408
409 return res;
410 }
411
412
413 LLVMValueRef
414 lp_build_select_aos(struct lp_build_context *bld,
415 LLVMValueRef a,
416 LLVMValueRef b,
417 const boolean cond[4])
418 {
419 const struct lp_type type = bld->type;
420 const unsigned n = type.length;
421 unsigned i, j;
422
423 if(a == b)
424 return a;
425 if(cond[0] && cond[1] && cond[2] && cond[3])
426 return a;
427 if(!cond[0] && !cond[1] && !cond[2] && !cond[3])
428 return b;
429 if(a == bld->undef || b == bld->undef)
430 return bld->undef;
431
432 /*
433 * There are three major ways of accomplishing this:
434 * - with a shuffle,
435 * - with a select,
436 * - or with a bit mask.
437 *
438 * Select isn't supported for vector types yet.
439 * The flip between these is empirical and might need to be.
440 */
441 if (n <= 4) {
442 /*
443 * Shuffle.
444 */
445 LLVMTypeRef elem_type = LLVMInt32Type();
446 LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
447
448 for(j = 0; j < n; j += 4)
449 for(i = 0; i < 4; ++i)
450 shuffles[j + i] = LLVMConstInt(elem_type, (cond[i] ? 0 : n) + j + i, 0);
451
452 return LLVMBuildShuffleVector(bld->builder, a, b, LLVMConstVector(shuffles, n), "");
453 }
454 else {
455 #if 0
456 /* XXX: Unfortunately select of vectors do not work */
457 /* Use a select */
458 LLVMTypeRef elem_type = LLVMInt1Type();
459 LLVMValueRef cond[LP_MAX_VECTOR_LENGTH];
460
461 for(j = 0; j < n; j += 4)
462 for(i = 0; i < 4; ++i)
463 cond[j + i] = LLVMConstInt(elem_type, cond[i] ? 1 : 0, 0);
464
465 return LLVMBuildSelect(bld->builder, LLVMConstVector(cond, n), a, b, "");
466 #else
467 LLVMValueRef mask = lp_build_const_mask_aos(type, cond);
468 return lp_build_select(bld, mask, a, b);
469 #endif
470 }
471 }
472
473
474 /** Return (a & ~b) */
475 LLVMValueRef
476 lp_build_andc(struct lp_build_context *bld, LLVMValueRef a, LLVMValueRef b)
477 {
478 b = LLVMBuildNot(bld->builder, b, "");
479 b = LLVMBuildAnd(bld->builder, a, b, "");
480 return b;
481 }