lp_bld_logic.c revision 2297bc9233be014b7b5aa037769209fbe9f6a66c 
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
38#include "lp_bld_type.h"
39#include "lp_bld_const.h"
40#include "lp_bld_intr.h"
41#include "lp_bld_logic.h"
42
43
44/**
45 * Build code to compare two values 'a' and 'b' of 'type' using the given func.
46 * \param func  one of PIPE_FUNC_x
47 */
48LLVMValueRef
49lp_build_compare(LLVMBuilderRef builder,
50                 const struct lp_type type,
51                 unsigned func,
52                 LLVMValueRef a,
53                 LLVMValueRef b)
54{
55   LLVMTypeRef vec_type = lp_build_vec_type(type);
56   LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
57   LLVMValueRef zeros = LLVMConstNull(int_vec_type);
58   LLVMValueRef ones = LLVMConstAllOnes(int_vec_type);
59   LLVMValueRef cond;
60   LLVMValueRef res;
61   unsigned i;
62
63   assert(func >= PIPE_FUNC_NEVER);
64   assert(func <= PIPE_FUNC_ALWAYS);
65
66   if(func == PIPE_FUNC_NEVER)
67      return zeros;
68   if(func == PIPE_FUNC_ALWAYS)
69      return ones;
70
71   /* TODO: optimize the constant case */
72
73   /* XXX: It is not clear if we should use the ordered or unordered operators */
74
75#if defined(PIPE_ARCH_X86) || defined(PIPE_ARCH_X86_64)
76   if(type.width * type.length == 128) {
77      if(type.floating && util_cpu_caps.has_sse) {
78         /* float[4] comparison */
79         LLVMValueRef args[3];
80         unsigned cc;
81         boolean swap;
82
83         swap = FALSE;
84         switch(func) {
85         case PIPE_FUNC_EQUAL:
86            cc = 0;
87            break;
88         case PIPE_FUNC_NOTEQUAL:
89            cc = 4;
90            break;
91         case PIPE_FUNC_LESS:
92            cc = 1;
93            break;
94         case PIPE_FUNC_LEQUAL:
95            cc = 2;
96            break;
97         case PIPE_FUNC_GREATER:
98            cc = 1;
99            swap = TRUE;
100            break;
101         case PIPE_FUNC_GEQUAL:
102            cc = 2;
103            swap = TRUE;
104            break;
105         default:
106            assert(0);
107            return lp_build_undef(type);
108         }
109
110         if(swap) {
111            args[0] = b;
112            args[1] = a;
113         }
114         else {
115            args[0] = a;
116            args[1] = b;
117         }
118
119         args[2] = LLVMConstInt(LLVMInt8Type(), cc, 0);
120         res = lp_build_intrinsic(builder,
121                                  "llvm.x86.sse.cmp.ps",
122                                  vec_type,
123                                  args, 3);
124         res = LLVMBuildBitCast(builder, res, int_vec_type, "");
125         return res;
126      }
127      else if(util_cpu_caps.has_sse2) {
128         /* int[4] comparison */
129         static const struct {
130            unsigned swap:1;
131            unsigned eq:1;
132            unsigned gt:1;
133            unsigned not:1;
134         } table[] = {
135            {0, 0, 0, 1}, /* PIPE_FUNC_NEVER */
136            {1, 0, 1, 0}, /* PIPE_FUNC_LESS */
137            {0, 1, 0, 0}, /* PIPE_FUNC_EQUAL */
138            {0, 0, 1, 1}, /* PIPE_FUNC_LEQUAL */
139            {0, 0, 1, 0}, /* PIPE_FUNC_GREATER */
140            {0, 1, 0, 1}, /* PIPE_FUNC_NOTEQUAL */
141            {1, 0, 1, 1}, /* PIPE_FUNC_GEQUAL */
142            {0, 0, 0, 0}  /* PIPE_FUNC_ALWAYS */
143         };
144         const char *pcmpeq;
145         const char *pcmpgt;
146         LLVMValueRef args[2];
147         LLVMValueRef res;
148
149         switch (type.width) {
150         case 8:
151            pcmpeq = "llvm.x86.sse2.pcmpeq.b";
152            pcmpgt = "llvm.x86.sse2.pcmpgt.b";
153            break;
154         case 16:
155            pcmpeq = "llvm.x86.sse2.pcmpeq.w";
156            pcmpgt = "llvm.x86.sse2.pcmpgt.w";
157            break;
158         case 32:
159            pcmpeq = "llvm.x86.sse2.pcmpeq.d";
160            pcmpgt = "llvm.x86.sse2.pcmpgt.d";
161            break;
162         default:
163            assert(0);
164            return lp_build_undef(type);
165         }
166
167         /* There are no signed byte and unsigned word/dword comparison
168          * instructions. So flip the sign bit so that the results match.
169          */
170         if(table[func].gt &&
171            ((type.width == 8 && type.sign) ||
172             (type.width != 8 && !type.sign))) {
173            LLVMValueRef msb = lp_build_int_const_scalar(type, (unsigned long long)1 << (type.width - 1));
174            a = LLVMBuildXor(builder, a, msb, "");
175            b = LLVMBuildXor(builder, b, msb, "");
176         }
177
178         if(table[func].swap) {
179            args[0] = b;
180            args[1] = a;
181         }
182         else {
183            args[0] = a;
184            args[1] = b;
185         }
186
187         if(table[func].eq)
188            res = lp_build_intrinsic(builder, pcmpeq, vec_type, args, 2);
189         else if (table[func].gt)
190            res = lp_build_intrinsic(builder, pcmpgt, vec_type, args, 2);
191         else
192            res = LLVMConstNull(vec_type);
193
194         if(table[func].not)
195            res = LLVMBuildNot(builder, res, "");
196
197         return res;
198      }
199   }
200#endif
201
202   if(type.floating) {
203      LLVMRealPredicate op;
204      switch(func) {
205      case PIPE_FUNC_NEVER:
206         op = LLVMRealPredicateFalse;
207         break;
208      case PIPE_FUNC_ALWAYS:
209         op = LLVMRealPredicateTrue;
210         break;
211      case PIPE_FUNC_EQUAL:
212         op = LLVMRealUEQ;
213         break;
214      case PIPE_FUNC_NOTEQUAL:
215         op = LLVMRealUNE;
216         break;
217      case PIPE_FUNC_LESS:
218         op = LLVMRealULT;
219         break;
220      case PIPE_FUNC_LEQUAL:
221         op = LLVMRealULE;
222         break;
223      case PIPE_FUNC_GREATER:
224         op = LLVMRealUGT;
225         break;
226      case PIPE_FUNC_GEQUAL:
227         op = LLVMRealUGE;
228         break;
229      default:
230         assert(0);
231         return lp_build_undef(type);
232      }
233
234#if 0
235      /* XXX: Although valid IR, no LLVM target currently support this */
236      cond = LLVMBuildFCmp(builder, op, a, b, "");
237      res = LLVMBuildSelect(builder, cond, ones, zeros, "");
238#else
239      debug_printf("%s: warning: using slow element-wise vector comparison\n",
240                   __FUNCTION__);
241      res = LLVMGetUndef(int_vec_type);
242      for(i = 0; i < type.length; ++i) {
243         LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
244         cond = LLVMBuildFCmp(builder, op,
245                              LLVMBuildExtractElement(builder, a, index, ""),
246                              LLVMBuildExtractElement(builder, b, index, ""),
247                              "");
248         cond = LLVMBuildSelect(builder, cond,
249                                LLVMConstExtractElement(ones, index),
250                                LLVMConstExtractElement(zeros, index),
251                                "");
252         res = LLVMBuildInsertElement(builder, res, cond, index, "");
253      }
254#endif
255   }
256   else {
257      LLVMIntPredicate op;
258      switch(func) {
259      case PIPE_FUNC_EQUAL:
260         op = LLVMIntEQ;
261         break;
262      case PIPE_FUNC_NOTEQUAL:
263         op = LLVMIntNE;
264         break;
265      case PIPE_FUNC_LESS:
266         op = type.sign ? LLVMIntSLT : LLVMIntULT;
267         break;
268      case PIPE_FUNC_LEQUAL:
269         op = type.sign ? LLVMIntSLE : LLVMIntULE;
270         break;
271      case PIPE_FUNC_GREATER:
272         op = type.sign ? LLVMIntSGT : LLVMIntUGT;
273         break;
274      case PIPE_FUNC_GEQUAL:
275         op = type.sign ? LLVMIntSGE : LLVMIntUGE;
276         break;
277      default:
278         assert(0);
279         return lp_build_undef(type);
280      }
281
282#if 0
283      /* XXX: Although valid IR, no LLVM target currently support this */
284      cond = LLVMBuildICmp(builder, op, a, b, "");
285      res = LLVMBuildSelect(builder, cond, ones, zeros, "");
286#else
287      debug_printf("%s: warning: using slow element-wise int vector comparison\n",
288                   __FUNCTION__);
289      res = LLVMGetUndef(int_vec_type);
290      for(i = 0; i < type.length; ++i) {
291         LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
292         cond = LLVMBuildICmp(builder, op,
293                              LLVMBuildExtractElement(builder, a, index, ""),
294                              LLVMBuildExtractElement(builder, b, index, ""),
295                              "");
296         cond = LLVMBuildSelect(builder, cond,
297                                LLVMConstExtractElement(ones, index),
298                                LLVMConstExtractElement(zeros, index),
299                                "");
300         res = LLVMBuildInsertElement(builder, res, cond, index, "");
301      }
302#endif
303   }
304
305   return res;
306}
307
308
309
310/**
311 * Build code to compare two values 'a' and 'b' using the given func.
312 * \param func  one of PIPE_FUNC_x
313 */
314LLVMValueRef
315lp_build_cmp(struct lp_build_context *bld,
316             unsigned func,
317             LLVMValueRef a,
318             LLVMValueRef b)
319{
320   return lp_build_compare(bld->builder, bld->type, func, a, b);
321}
322
323
324LLVMValueRef
325lp_build_select(struct lp_build_context *bld,
326                LLVMValueRef mask,
327                LLVMValueRef a,
328                LLVMValueRef b)
329{
330   struct lp_type type = bld->type;
331   LLVMValueRef res;
332
333   if(a == b)
334      return a;
335
336   if(type.floating) {
337      LLVMTypeRef int_vec_type = lp_build_int_vec_type(type);
338      a = LLVMBuildBitCast(bld->builder, a, int_vec_type, "");
339      b = LLVMBuildBitCast(bld->builder, b, int_vec_type, "");
340   }
341
342   a = LLVMBuildAnd(bld->builder, a, mask, "");
343
344   /* This often gets translated to PANDN, but sometimes the NOT is
345    * pre-computed and stored in another constant. The best strategy depends
346    * on available registers, so it is not a big deal -- hopefully LLVM does
347    * the right decision attending the rest of the program.
348    */
349   b = LLVMBuildAnd(bld->builder, b, LLVMBuildNot(bld->builder, mask, ""), "");
350
351   res = LLVMBuildOr(bld->builder, a, b, "");
352
353   if(type.floating) {
354      LLVMTypeRef vec_type = lp_build_vec_type(type);
355      res = LLVMBuildBitCast(bld->builder, res, vec_type, "");
356   }
357
358   return res;
359}
360
361
362LLVMValueRef
363lp_build_select_aos(struct lp_build_context *bld,
364                    LLVMValueRef a,
365                    LLVMValueRef b,
366                    const boolean cond[4])
367{
368   const struct lp_type type = bld->type;
369   const unsigned n = type.length;
370   unsigned i, j;
371
372   if(a == b)
373      return a;
374   if(cond[0] && cond[1] && cond[2] && cond[3])
375      return a;
376   if(!cond[0] && !cond[1] && !cond[2] && !cond[3])
377      return b;
378   if(a == bld->undef || b == bld->undef)
379      return bld->undef;
380
381   /*
382    * There are three major ways of accomplishing this:
383    * - with a shuffle,
384    * - with a select,
385    * - or with a bit mask.
386    *
387    * Select isn't supported for vector types yet.
388    * The flip between these is empirical and might need to be.
389    */
390   if (n <= 4) {
391      /*
392       * Shuffle.
393       */
394      LLVMTypeRef elem_type = LLVMInt32Type();
395      LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
396
397      for(j = 0; j < n; j += 4)
398         for(i = 0; i < 4; ++i)
399            shuffles[j + i] = LLVMConstInt(elem_type, (cond[i] ? 0 : n) + j + i, 0);
400
401      return LLVMBuildShuffleVector(bld->builder, a, b, LLVMConstVector(shuffles, n), "");
402   }
403   else {
404#if 0
405      /* XXX: Unfortunately select of vectors do not work */
406      /* Use a select */
407      LLVMTypeRef elem_type = LLVMInt1Type();
408      LLVMValueRef cond[LP_MAX_VECTOR_LENGTH];
409
410      for(j = 0; j < n; j += 4)
411         for(i = 0; i < 4; ++i)
412            cond[j + i] = LLVMConstInt(elem_type, cond[i] ? 1 : 0, 0);
413
414      return LLVMBuildSelect(bld->builder, LLVMConstVector(cond, n), a, b, "");
415#else
416      LLVMValueRef mask = lp_build_const_mask_aos(type, cond);
417      return lp_build_select(bld, mask, a, b);
418#endif
419   }
420}
421