gallivm: fix rsqrt failures
[mesa.git] / src / gallium / auxiliary / gallivm / lp_bld_const.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 /**
30 * @file
31 * Helper functions for constant building.
32 *
33 * @author Jose Fonseca <jfonseca@vmware.com>
34 */
35
36 #include <float.h>
37
38 #include "util/u_debug.h"
39 #include "util/u_math.h"
40 #include "util/u_half.h"
41
42 #include "lp_bld_type.h"
43 #include "lp_bld_const.h"
44 #include "lp_bld_init.h"
45
46
47 unsigned
48 lp_mantissa(struct lp_type type)
49 {
50 assert(type.floating);
51
52 if(type.floating) {
53 switch(type.width) {
54 case 16:
55 return 10;
56 case 32:
57 return 23;
58 case 64:
59 return 52;
60 default:
61 assert(0);
62 return 0;
63 }
64 }
65 else {
66 if(type.sign)
67 return type.width - 1;
68 else
69 return type.width;
70 }
71 }
72
73
74 /**
75 * Shift of the unity.
76 *
77 * Same as lp_const_scale(), but in terms of shifts.
78 */
79 unsigned
80 lp_const_shift(struct lp_type type)
81 {
82 if(type.floating)
83 return 0;
84 else if(type.fixed)
85 return type.width/2;
86 else if(type.norm)
87 return type.sign ? type.width - 1 : type.width;
88 else
89 return 0;
90 }
91
92
93 unsigned
94 lp_const_offset(struct lp_type type)
95 {
96 if(type.floating || type.fixed)
97 return 0;
98 else if(type.norm)
99 return 1;
100 else
101 return 0;
102 }
103
104
105 /**
106 * Scaling factor between the LLVM native value and its interpretation.
107 *
108 * This is 1.0 for all floating types and unnormalized integers, and something
109 * else for the fixed points types and normalized integers.
110 */
111 double
112 lp_const_scale(struct lp_type type)
113 {
114 unsigned long long llscale;
115 double dscale;
116
117 llscale = (unsigned long long)1 << lp_const_shift(type);
118 llscale -= lp_const_offset(type);
119 dscale = (double)llscale;
120 assert((unsigned long long)dscale == llscale);
121
122 return dscale;
123 }
124
125
126 /**
127 * Minimum value representable by the type.
128 */
129 double
130 lp_const_min(struct lp_type type)
131 {
132 unsigned bits;
133
134 if(!type.sign)
135 return 0.0;
136
137 if(type.norm)
138 return -1.0;
139
140 if (type.floating) {
141 switch(type.width) {
142 case 16:
143 return -65504;
144 case 32:
145 return -FLT_MAX;
146 case 64:
147 return -DBL_MAX;
148 default:
149 assert(0);
150 return 0.0;
151 }
152 }
153
154 if(type.fixed)
155 /* FIXME: consider the fractional bits? */
156 bits = type.width / 2 - 1;
157 else
158 bits = type.width - 1;
159
160 return (double)-((long long)1 << bits);
161 }
162
163
164 /**
165 * Maximum value representable by the type.
166 */
167 double
168 lp_const_max(struct lp_type type)
169 {
170 unsigned bits;
171
172 if(type.norm)
173 return 1.0;
174
175 if (type.floating) {
176 switch(type.width) {
177 case 16:
178 return 65504;
179 case 32:
180 return FLT_MAX;
181 case 64:
182 return DBL_MAX;
183 default:
184 assert(0);
185 return 0.0;
186 }
187 }
188
189 if(type.fixed)
190 bits = type.width / 2;
191 else
192 bits = type.width;
193
194 if(type.sign)
195 bits -= 1;
196
197 return (double)(((unsigned long long)1 << bits) - 1);
198 }
199
200
201 double
202 lp_const_eps(struct lp_type type)
203 {
204 if (type.floating) {
205 switch(type.width) {
206 case 16:
207 return 2E-10;
208 case 32:
209 return FLT_EPSILON;
210 case 64:
211 return DBL_EPSILON;
212 default:
213 assert(0);
214 return 0.0;
215 }
216 }
217 else {
218 double scale = lp_const_scale(type);
219 return 1.0/scale;
220 }
221 }
222
223
224 LLVMValueRef
225 lp_build_undef(struct gallivm_state *gallivm, struct lp_type type)
226 {
227 LLVMTypeRef vec_type = lp_build_vec_type(gallivm, type);
228 return LLVMGetUndef(vec_type);
229 }
230
231
232 LLVMValueRef
233 lp_build_zero(struct gallivm_state *gallivm, struct lp_type type)
234 {
235 if (type.length == 1) {
236 if (type.floating)
237 return lp_build_const_float(gallivm, 0.0);
238 else
239 return LLVMConstInt(LLVMIntTypeInContext(gallivm->context, type.width), 0, 0);
240 }
241 else {
242 LLVMTypeRef vec_type = lp_build_vec_type(gallivm, type);
243 return LLVMConstNull(vec_type);
244 }
245 }
246
247
248 LLVMValueRef
249 lp_build_one(struct gallivm_state *gallivm, struct lp_type type)
250 {
251 LLVMTypeRef elem_type;
252 LLVMValueRef elems[LP_MAX_VECTOR_LENGTH];
253 unsigned i;
254
255 assert(type.length <= LP_MAX_VECTOR_LENGTH);
256
257 elem_type = lp_build_elem_type(gallivm, type);
258
259 if(type.floating && type.width == 16)
260 elems[0] = LLVMConstInt(elem_type, util_float_to_half(1.0f), 0);
261 else if(type.floating)
262 elems[0] = LLVMConstReal(elem_type, 1.0);
263 else if(type.fixed)
264 elems[0] = LLVMConstInt(elem_type, 1LL << (type.width/2), 0);
265 else if(!type.norm)
266 elems[0] = LLVMConstInt(elem_type, 1, 0);
267 else if(type.sign)
268 elems[0] = LLVMConstInt(elem_type, (1LL << (type.width - 1)) - 1, 0);
269 else {
270 /* special case' -- 1.0 for normalized types is more easily attained if
271 * we start with a vector consisting of all bits set */
272 LLVMTypeRef vec_type = LLVMVectorType(elem_type, type.length);
273 LLVMValueRef vec = LLVMConstAllOnes(vec_type);
274
275 #if 0
276 if(type.sign)
277 /* TODO: Unfortunately this caused "Tried to create a shift operation
278 * on a non-integer type!" */
279 vec = LLVMConstLShr(vec, lp_build_const_int_vec(type, 1));
280 #endif
281
282 return vec;
283 }
284
285 for(i = 1; i < type.length; ++i)
286 elems[i] = elems[0];
287
288 if (type.length == 1)
289 return elems[0];
290 else
291 return LLVMConstVector(elems, type.length);
292 }
293
294
295 /**
296 * Build constant-valued element from a scalar value.
297 */
298 LLVMValueRef
299 lp_build_const_elem(struct gallivm_state *gallivm,
300 struct lp_type type,
301 double val)
302 {
303 LLVMTypeRef elem_type = lp_build_elem_type(gallivm, type);
304 LLVMValueRef elem;
305
306 if(type.floating && type.width == 16) {
307 elem = LLVMConstInt(elem_type, util_float_to_half((float)val), 0);
308 } else if(type.floating) {
309 elem = LLVMConstReal(elem_type, val);
310 }
311 else {
312 double dscale = lp_const_scale(type);
313
314 elem = LLVMConstInt(elem_type, round(val*dscale), 0);
315 }
316
317 return elem;
318 }
319
320
321 /**
322 * Build constant-valued vector from a scalar value.
323 */
324 LLVMValueRef
325 lp_build_const_vec(struct gallivm_state *gallivm, struct lp_type type,
326 double val)
327 {
328 if (type.length == 1) {
329 return lp_build_const_elem(gallivm, type, val);
330 } else {
331 LLVMValueRef elems[LP_MAX_VECTOR_LENGTH];
332 unsigned i;
333 elems[0] = lp_build_const_elem(gallivm, type, val);
334 for(i = 1; i < type.length; ++i)
335 elems[i] = elems[0];
336 return LLVMConstVector(elems, type.length);
337 }
338 }
339
340
341 LLVMValueRef
342 lp_build_const_int_vec(struct gallivm_state *gallivm, struct lp_type type,
343 long long val)
344 {
345 LLVMTypeRef elem_type = lp_build_int_elem_type(gallivm, type);
346 LLVMValueRef elems[LP_MAX_VECTOR_LENGTH];
347 unsigned i;
348
349 assert(type.length <= LP_MAX_VECTOR_LENGTH);
350
351 for(i = 0; i < type.length; ++i)
352 elems[i] = LLVMConstInt(elem_type, val, type.sign ? 1 : 0);
353
354 if (type.length == 1)
355 return elems[0];
356
357 return LLVMConstVector(elems, type.length);
358 }
359
360
361 LLVMValueRef
362 lp_build_const_aos(struct gallivm_state *gallivm,
363 struct lp_type type,
364 double r, double g, double b, double a,
365 const unsigned char *swizzle)
366 {
367 const unsigned char default_swizzle[4] = {0, 1, 2, 3};
368 LLVMValueRef elems[LP_MAX_VECTOR_LENGTH];
369 unsigned i;
370
371 assert(type.length % 4 == 0);
372 assert(type.length <= LP_MAX_VECTOR_LENGTH);
373
374 lp_build_elem_type(gallivm, type);
375
376 if(swizzle == NULL)
377 swizzle = default_swizzle;
378
379 elems[swizzle[0]] = lp_build_const_elem(gallivm, type, r);
380 elems[swizzle[1]] = lp_build_const_elem(gallivm, type, g);
381 elems[swizzle[2]] = lp_build_const_elem(gallivm, type, b);
382 elems[swizzle[3]] = lp_build_const_elem(gallivm, type, a);
383
384 for(i = 4; i < type.length; ++i)
385 elems[i] = elems[i % 4];
386
387 return LLVMConstVector(elems, type.length);
388 }
389
390
391 /**
392 * @param mask TGSI_WRITEMASK_xxx
393 */
394 LLVMValueRef
395 lp_build_const_mask_aos(struct gallivm_state *gallivm,
396 struct lp_type type,
397 unsigned mask)
398 {
399 LLVMTypeRef elem_type = LLVMIntTypeInContext(gallivm->context, type.width);
400 LLVMValueRef masks[LP_MAX_VECTOR_LENGTH];
401 unsigned i, j;
402
403 assert(type.length <= LP_MAX_VECTOR_LENGTH);
404
405 for (j = 0; j < type.length; j += 4) {
406 for( i = 0; i < 4; ++i) {
407 masks[j + i] = LLVMConstInt(elem_type,
408 mask & (1 << i) ? ~0ULL : 0,
409 1);
410 }
411 }
412
413 return LLVMConstVector(masks, type.length);
414 }
415
416
417 /**
418 * Performs lp_build_const_mask_aos, but first swizzles the mask
419 */
420 LLVMValueRef
421 lp_build_const_mask_aos_swizzled(struct gallivm_state *gallivm,
422 struct lp_type type,
423 unsigned mask,
424 const unsigned char *swizzle)
425 {
426 mask =
427 ((mask & (1 << swizzle[0])) >> swizzle[0])
428 | (((mask & (1 << swizzle[1])) >> swizzle[1]) << 1)
429 | (((mask & (1 << swizzle[2])) >> swizzle[2]) << 2)
430 | (((mask & (1 << swizzle[3])) >> swizzle[3]) << 3);
431
432 return lp_build_const_mask_aos(gallivm, type, mask);
433 }
434
435
436 /**
437 * Build a zero-terminated constant string.
438 */
439 LLVMValueRef
440 lp_build_const_string(struct gallivm_state *gallivm,
441 const char *str)
442 {
443 unsigned len = strlen(str) + 1;
444 LLVMTypeRef i8 = LLVMInt8TypeInContext(gallivm->context);
445 LLVMValueRef string = LLVMAddGlobal(gallivm->module, LLVMArrayType(i8, len), "");
446 LLVMSetGlobalConstant(string, TRUE);
447 LLVMSetLinkage(string, LLVMInternalLinkage);
448 LLVMSetInitializer(string, LLVMConstStringInContext(gallivm->context, str, len, TRUE));
449 string = LLVMConstBitCast(string, LLVMPointerType(i8, 0));
450 return string;
451 }
452
453
454 /**
455 * Build a callable function pointer.
456 *
457 * We use function pointer constants instead of LLVMAddGlobalMapping()
458 * to work around a bug in LLVM 2.6, and for efficiency/simplicity.
459 */
460 LLVMValueRef
461 lp_build_const_func_pointer(struct gallivm_state *gallivm,
462 const void *ptr,
463 LLVMTypeRef ret_type,
464 LLVMTypeRef *arg_types,
465 unsigned num_args,
466 const char *name)
467 {
468 LLVMTypeRef function_type;
469 LLVMValueRef function;
470
471 function_type = LLVMFunctionType(ret_type, arg_types, num_args, 0);
472
473 function = lp_build_const_int_pointer(gallivm, ptr);
474
475 function = LLVMBuildBitCast(gallivm->builder, function,
476 LLVMPointerType(function_type, 0),
477 name);
478
479 return function;
480 }