src/gallium/drivers/r300/compiler/radeon_optimize.c

   1 /*
   2  * Copyright (C) 2009 Nicolai Haehnle.
   3  * Copyright 2010 Tom Stellard <tstellar@gmail.com>
   4  *
   5  * All Rights Reserved.
   6  *
   7  * Permission is hereby granted, free of charge, to any person obtaining
   8  * a copy of this software and associated documentation files (the
   9  * "Software"), to deal in the Software without restriction, including
  10  * without limitation the rights to use, copy, modify, merge, publish,
  11  * distribute, sublicense, and/or sell copies of the Software, and to
  12  * permit persons to whom the Software is furnished to do so, subject to
  13  * the following conditions:
  14  *
  15  * The above copyright notice and this permission notice (including the
  16  * next paragraph) shall be included in all copies or substantial
  17  * portions of the Software.
  18  *
  19  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  20  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  21  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
  22  * IN NO EVENT SHALL THE COPYRIGHT OWNER(S) AND/OR ITS SUPPLIERS BE
  23  * LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
  24  * OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
  25  * WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  26  *
  27  */
  28
  29 #include "radeon_dataflow.h"
  30
  31 #include "radeon_compiler.h"
  32 #include "radeon_compiler_util.h"
  33 #include "radeon_swizzle.h"
  34
  35 struct src_clobbered_reads_cb_data {
  36         rc_register_file File;
  37         unsigned int Index;
  38         unsigned int Mask;
  39         struct rc_reader_data * ReaderData;
  40 };
  41
  42 typedef void (*rc_presub_replace_fn)(struct rc_instruction *,
  43                                                 struct rc_instruction *,
  44                                                 unsigned int);
  45
  46 static struct rc_src_register chain_srcregs(struct rc_src_register outer, struct rc_src_register inner)
  47 {
  48         struct rc_src_register combine;
  49         combine.File = inner.File;
  50         combine.Index = inner.Index;
  51         combine.RelAddr = inner.RelAddr;
  52         if (outer.Abs) {
  53                 combine.Abs = 1;
  54                 combine.Negate = outer.Negate;
  55         } else {
  56                 combine.Abs = inner.Abs;
  57                 combine.Negate = swizzle_mask(outer.Swizzle, inner.Negate);
  58                 combine.Negate ^= outer.Negate;
  59         }
  60         combine.Swizzle = combine_swizzles(inner.Swizzle, outer.Swizzle);
  61         return combine;
  62 }
  63
  64 static void copy_propagate_scan_read(void * data, struct rc_instruction * inst,
  65                                                 struct rc_src_register * src)
  66 {
  67         rc_register_file file = src->File;
  68         struct rc_reader_data * reader_data = data;
  69
  70         if(!rc_inst_can_use_presub(inst,
  71                                 reader_data->Writer->U.I.PreSub.Opcode,
  72                                 rc_swizzle_to_writemask(src->Swizzle),
  73                                 src,
  74                                 &reader_data->Writer->U.I.PreSub.SrcReg[0],
  75                                 &reader_data->Writer->U.I.PreSub.SrcReg[1])) {
  76                 reader_data->Abort = 1;
  77                 return;
  78         }
  79
  80         /* XXX This could probably be handled better. */
  81         if (file == RC_FILE_ADDRESS) {
  82                 reader_data->Abort = 1;
  83                 return;
  84         }
  85
  86         /* These instructions cannot read from the constants file.
  87          * see radeonTransformTEX()
  88          */
  89         if(reader_data->Writer->U.I.SrcReg[0].File != RC_FILE_TEMPORARY &&
  90                         reader_data->Writer->U.I.SrcReg[0].File != RC_FILE_INPUT &&
  91                                 (inst->U.I.Opcode == RC_OPCODE_TEX ||
  92                                 inst->U.I.Opcode == RC_OPCODE_TXB ||
  93                                 inst->U.I.Opcode == RC_OPCODE_TXP ||
  94                                 inst->U.I.Opcode == RC_OPCODE_TXD ||
  95                                 inst->U.I.Opcode == RC_OPCODE_TXL ||
  96                                 inst->U.I.Opcode == RC_OPCODE_KIL)){
  97                 reader_data->Abort = 1;
  98                 return;
  99         }
 100 }
 101
 102 static void src_clobbered_reads_cb(
 103         void * data,
 104         struct rc_instruction * inst,
 105         struct rc_src_register * src)
 106 {
 107         struct src_clobbered_reads_cb_data * sc_data = data;
 108
 109         if (src->File == sc_data->File
 110             && src->Index == sc_data->Index
 111             && (rc_swizzle_to_writemask(src->Swizzle) & sc_data->Mask)) {
 112
 113                 sc_data->ReaderData->AbortOnRead = RC_MASK_XYZW;
 114         }
 115
 116         if (src->RelAddr && sc_data->File == RC_FILE_ADDRESS) {
 117                 sc_data->ReaderData->AbortOnRead = RC_MASK_XYZW;
 118         }
 119 }
 120
 121 static void is_src_clobbered_scan_write(
 122         void * data,
 123         struct rc_instruction * inst,
 124         rc_register_file file,
 125         unsigned int index,
 126         unsigned int mask)
 127 {
 128         struct src_clobbered_reads_cb_data sc_data;
 129         struct rc_reader_data * reader_data = data;
 130         sc_data.File = file;
 131         sc_data.Index = index;
 132         sc_data.Mask = mask;
 133         sc_data.ReaderData = reader_data;
 134         rc_for_all_reads_src(reader_data->Writer,
 135                                         src_clobbered_reads_cb, &sc_data);
 136 }
 137
 138 static void copy_propagate(struct radeon_compiler * c, struct rc_instruction * inst_mov)
 139 {
 140         struct rc_reader_data reader_data;
 141         unsigned int i;
 142
 143         if (inst_mov->U.I.DstReg.File != RC_FILE_TEMPORARY ||
 144             inst_mov->U.I.WriteALUResult ||
 145             inst_mov->U.I.SaturateMode)
 146                 return;
 147
 148         /* Get a list of all the readers of this MOV instruction. */
 149         reader_data.ExitOnAbort = 1;
 150         rc_get_readers(c, inst_mov, &reader_data,
 151                        copy_propagate_scan_read, NULL,
 152                        is_src_clobbered_scan_write);
 153
 154         if (reader_data.Abort || reader_data.ReaderCount == 0)
 155                 return;
 156
 157         /* Propagate the MOV instruction. */
 158         for (i = 0; i < reader_data.ReaderCount; i++) {
 159                 struct rc_instruction * inst = reader_data.Readers[i].Inst;
 160                 *reader_data.Readers[i].U.I.Src = chain_srcregs(*reader_data.Readers[i].U.I.Src, inst_mov->U.I.SrcReg[0]);
 161
 162                 if (inst_mov->U.I.SrcReg[0].File == RC_FILE_PRESUB)
 163                         inst->U.I.PreSub = inst_mov->U.I.PreSub;
 164         }
 165
 166         /* Finally, remove the original MOV instruction */
 167         rc_remove_instruction(inst_mov);
 168 }
 169
 170 /**
 171  * Check if a source register is actually always the same
 172  * swizzle constant.
 173  */
 174 static int is_src_uniform_constant(struct rc_src_register src,
 175                 rc_swizzle * pswz, unsigned int * pnegate)
 176 {
 177         int have_used = 0;
 178
 179         if (src.File != RC_FILE_NONE) {
 180                 *pswz = 0;
 181                 return 0;
 182         }
 183
 184         for(unsigned int chan = 0; chan < 4; ++chan) {
 185                 unsigned int swz = GET_SWZ(src.Swizzle, chan);
 186                 if (swz < 4) {
 187                         *pswz = 0;
 188                         return 0;
 189                 }
 190                 if (swz == RC_SWIZZLE_UNUSED)
 191                         continue;
 192
 193                 if (!have_used) {
 194                         *pswz = swz;
 195                         *pnegate = GET_BIT(src.Negate, chan);
 196                         have_used = 1;
 197                 } else {
 198                         if (swz != *pswz || *pnegate != GET_BIT(src.Negate, chan)) {
 199                                 *pswz = 0;
 200                                 return 0;
 201                         }
 202                 }
 203         }
 204
 205         return 1;
 206 }
 207
 208 static void constant_folding_mad(struct rc_instruction * inst)
 209 {
 210         rc_swizzle swz = 0;
 211         unsigned int negate= 0;
 212
 213         if (is_src_uniform_constant(inst->U.I.SrcReg[2], &swz, &negate)) {
 214                 if (swz == RC_SWIZZLE_ZERO) {
 215                         inst->U.I.Opcode = RC_OPCODE_MUL;
 216                         return;
 217                 }
 218         }
 219
 220         if (is_src_uniform_constant(inst->U.I.SrcReg[1], &swz, &negate)) {
 221                 if (swz == RC_SWIZZLE_ONE) {
 222                         inst->U.I.Opcode = RC_OPCODE_ADD;
 223                         if (negate)
 224                                 inst->U.I.SrcReg[0].Negate ^= RC_MASK_XYZW;
 225                         inst->U.I.SrcReg[1] = inst->U.I.SrcReg[2];
 226                         return;
 227                 } else if (swz == RC_SWIZZLE_ZERO) {
 228                         inst->U.I.Opcode = RC_OPCODE_MOV;
 229                         inst->U.I.SrcReg[0] = inst->U.I.SrcReg[2];
 230                         return;
 231                 }
 232         }
 233
 234         if (is_src_uniform_constant(inst->U.I.SrcReg[0], &swz, &negate)) {
 235                 if (swz == RC_SWIZZLE_ONE) {
 236                         inst->U.I.Opcode = RC_OPCODE_ADD;
 237                         if (negate)
 238                                 inst->U.I.SrcReg[1].Negate ^= RC_MASK_XYZW;
 239                         inst->U.I.SrcReg[0] = inst->U.I.SrcReg[2];
 240                         return;
 241                 } else if (swz == RC_SWIZZLE_ZERO) {
 242                         inst->U.I.Opcode = RC_OPCODE_MOV;
 243                         inst->U.I.SrcReg[0] = inst->U.I.SrcReg[2];
 244                         return;
 245                 }
 246         }
 247 }
 248
 249 static void constant_folding_mul(struct rc_instruction * inst)
 250 {
 251         rc_swizzle swz = 0;
 252         unsigned int negate = 0;
 253
 254         if (is_src_uniform_constant(inst->U.I.SrcReg[0], &swz, &negate)) {
 255                 if (swz == RC_SWIZZLE_ONE) {
 256                         inst->U.I.Opcode = RC_OPCODE_MOV;
 257                         inst->U.I.SrcReg[0] = inst->U.I.SrcReg[1];
 258                         if (negate)
 259                                 inst->U.I.SrcReg[0].Negate ^= RC_MASK_XYZW;
 260                         return;
 261                 } else if (swz == RC_SWIZZLE_ZERO) {
 262                         inst->U.I.Opcode = RC_OPCODE_MOV;
 263                         inst->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_0000;
 264                         return;
 265                 }
 266         }
 267
 268         if (is_src_uniform_constant(inst->U.I.SrcReg[1], &swz, &negate)) {
 269                 if (swz == RC_SWIZZLE_ONE) {
 270                         inst->U.I.Opcode = RC_OPCODE_MOV;
 271                         if (negate)
 272                                 inst->U.I.SrcReg[0].Negate ^= RC_MASK_XYZW;
 273                         return;
 274                 } else if (swz == RC_SWIZZLE_ZERO) {
 275                         inst->U.I.Opcode = RC_OPCODE_MOV;
 276                         inst->U.I.SrcReg[0].Swizzle = RC_SWIZZLE_0000;
 277                         return;
 278                 }
 279         }
 280 }
 281
 282 static void constant_folding_add(struct rc_instruction * inst)
 283 {
 284         rc_swizzle swz = 0;
 285         unsigned int negate = 0;
 286
 287         if (is_src_uniform_constant(inst->U.I.SrcReg[0], &swz, &negate)) {
 288                 if (swz == RC_SWIZZLE_ZERO) {
 289                         inst->U.I.Opcode = RC_OPCODE_MOV;
 290                         inst->U.I.SrcReg[0] = inst->U.I.SrcReg[1];
 291                         return;
 292                 }
 293         }
 294
 295         if (is_src_uniform_constant(inst->U.I.SrcReg[1], &swz, &negate)) {
 296                 if (swz == RC_SWIZZLE_ZERO) {
 297                         inst->U.I.Opcode = RC_OPCODE_MOV;
 298                         return;
 299                 }
 300         }
 301 }
 302
 303 /**
 304  * Replace 0.0, 1.0 and 0.5 immediate constants by their
 305  * respective swizzles. Simplify instructions like ADD dst, src, 0;
 306  */
 307 static void constant_folding(struct radeon_compiler * c, struct rc_instruction * inst)
 308 {
 309         const struct rc_opcode_info * opcode = rc_get_opcode_info(inst->U.I.Opcode);
 310         unsigned int i;
 311
 312         /* Replace 0.0, 1.0 and 0.5 immediates by their explicit swizzles */
 313         for(unsigned int src = 0; src < opcode->NumSrcRegs; ++src) {
 314                 struct rc_constant * constant;
 315                 struct rc_src_register newsrc;
 316                 int have_real_reference;
 317                 unsigned int chan;
 318
 319                 /* If there are only 0, 0.5, 1, or _ swizzles, mark the source as a constant. */
 320                 for (chan = 0; chan < 4; ++chan)
 321                         if (GET_SWZ(inst->U.I.SrcReg[src].Swizzle, chan) <= 3)
 322                                 break;
 323                 if (chan == 4) {
 324                         inst->U.I.SrcReg[src].File = RC_FILE_NONE;
 325                         continue;
 326                 }
 327
 328                 /* Convert immediates to swizzles. */
 329                 if (inst->U.I.SrcReg[src].File != RC_FILE_CONSTANT ||
 330                     inst->U.I.SrcReg[src].RelAddr ||
 331                     inst->U.I.SrcReg[src].Index >= c->Program.Constants.Count)
 332                         continue;
 333
 334                 constant =
 335                         &c->Program.Constants.Constants[inst->U.I.SrcReg[src].Index];
 336
 337                 if (constant->Type != RC_CONSTANT_IMMEDIATE)
 338                         continue;
 339
 340                 newsrc = inst->U.I.SrcReg[src];
 341                 have_real_reference = 0;
 342                 for (chan = 0; chan < 4; ++chan) {
 343                         unsigned int swz = GET_SWZ(newsrc.Swizzle, chan);
 344                         unsigned int newswz;
 345                         float imm;
 346                         float baseimm;
 347
 348                         if (swz >= 4)
 349                                 continue;
 350
 351                         imm = constant->u.Immediate[swz];
 352                         baseimm = imm;
 353                         if (imm < 0.0)
 354                                 baseimm = -baseimm;
 355
 356                         if (baseimm == 0.0) {
 357                                 newswz = RC_SWIZZLE_ZERO;
 358                         } else if (baseimm == 1.0) {
 359                                 newswz = RC_SWIZZLE_ONE;
 360                         } else if (baseimm == 0.5 && c->has_half_swizzles) {
 361                                 newswz = RC_SWIZZLE_HALF;
 362                         } else {
 363                                 have_real_reference = 1;
 364                                 continue;
 365                         }
 366
 367                         SET_SWZ(newsrc.Swizzle, chan, newswz);
 368                         if (imm < 0.0 && !newsrc.Abs)
 369                                 newsrc.Negate ^= 1 << chan;
 370                 }
 371
 372                 if (!have_real_reference) {
 373                         newsrc.File = RC_FILE_NONE;
 374                         newsrc.Index = 0;
 375                 }
 376
 377                 /* don't make the swizzle worse */
 378                 if (!c->SwizzleCaps->IsNative(inst->U.I.Opcode, newsrc) &&
 379                     c->SwizzleCaps->IsNative(inst->U.I.Opcode, inst->U.I.SrcReg[src]))
 380                         continue;
 381
 382                 inst->U.I.SrcReg[src] = newsrc;
 383         }
 384
 385         /* Simplify instructions based on constants */
 386         if (inst->U.I.Opcode == RC_OPCODE_MAD)
 387                 constant_folding_mad(inst);
 388
 389         /* note: MAD can simplify to MUL or ADD */
 390         if (inst->U.I.Opcode == RC_OPCODE_MUL)
 391                 constant_folding_mul(inst);
 392         else if (inst->U.I.Opcode == RC_OPCODE_ADD)
 393                 constant_folding_add(inst);
 394
 395         /* In case this instruction has been converted, make sure all of the
 396          * registers that are no longer used are empty. */
 397         opcode = rc_get_opcode_info(inst->U.I.Opcode);
 398         for(i = opcode->NumSrcRegs; i < 3; i++) {
 399                 memset(&inst->U.I.SrcReg[i], 0, sizeof(struct rc_src_register));
 400         }
 401 }
 402
 403 /**
 404  * If src and dst use the same register, this function returns a writemask that
 405  * indicates wich components are read by src.  Otherwise zero is returned.
 406  */
 407 static unsigned int src_reads_dst_mask(struct rc_src_register src,
 408                                                 struct rc_dst_register dst)
 409 {
 410         if (dst.File != src.File || dst.Index != src.Index) {
 411                 return 0;
 412         }
 413         return rc_swizzle_to_writemask(src.Swizzle);
 414 }
 415
 416 /* Return 1 if the source registers has a constant swizzle (e.g. 0, 0.5, 1.0)
 417  * in any of its channels.  Return 0 otherwise. */
 418 static int src_has_const_swz(struct rc_src_register src) {
 419         int chan;
 420         for(chan = 0; chan < 4; chan++) {
 421                 unsigned int swz = GET_SWZ(src.Swizzle, chan);
 422                 if (swz == RC_SWIZZLE_ZERO || swz == RC_SWIZZLE_HALF
 423                                                 || swz == RC_SWIZZLE_ONE) {
 424                         return 1;
 425                 }
 426         }
 427         return 0;
 428 }
 429
 430 static void presub_scan_read(
 431         void * data,
 432         struct rc_instruction * inst,
 433         struct rc_src_register * src)
 434 {
 435         struct rc_reader_data * reader_data = data;
 436         rc_presubtract_op * presub_opcode = reader_data->CbData;
 437
 438         if (!rc_inst_can_use_presub(inst, *presub_opcode,
 439                         reader_data->Writer->U.I.DstReg.WriteMask,
 440                         src,
 441                         &reader_data->Writer->U.I.SrcReg[0],
 442                         &reader_data->Writer->U.I.SrcReg[1])) {
 443                 reader_data->Abort = 1;
 444                 return;
 445         }
 446 }
 447
 448 static int presub_helper(
 449         struct radeon_compiler * c,
 450         struct rc_instruction * inst_add,
 451         rc_presubtract_op presub_opcode,
 452         rc_presub_replace_fn presub_replace)
 453 {
 454         struct rc_reader_data reader_data;
 455         unsigned int i;
 456         rc_presubtract_op cb_op = presub_opcode;
 457
 458         reader_data.CbData = &cb_op;
 459         reader_data.ExitOnAbort = 1;
 460         rc_get_readers(c, inst_add, &reader_data, presub_scan_read, NULL,
 461                                                 is_src_clobbered_scan_write);
 462
 463         if (reader_data.Abort || reader_data.ReaderCount == 0)
 464                 return 0;
 465
 466         for(i = 0; i < reader_data.ReaderCount; i++) {
 467                 unsigned int src_index;
 468                 struct rc_reader reader = reader_data.Readers[i];
 469                 const struct rc_opcode_info * info =
 470                                 rc_get_opcode_info(reader.Inst->U.I.Opcode);
 471
 472                 for (src_index = 0; src_index < info->NumSrcRegs; src_index++) {
 473                         if (&reader.Inst->U.I.SrcReg[src_index] == reader.U.I.Src)
 474                                 presub_replace(inst_add, reader.Inst, src_index);
 475                 }
 476         }
 477         return 1;
 478 }
 479
 480 /* This function assumes that inst_add->U.I.SrcReg[0] and
 481  * inst_add->U.I.SrcReg[1] aren't both negative. */
 482 static void presub_replace_add(
 483         struct rc_instruction * inst_add,
 484         struct rc_instruction * inst_reader,
 485         unsigned int src_index)
 486 {
 487         rc_presubtract_op presub_opcode;
 488         if (inst_add->U.I.SrcReg[1].Negate || inst_add->U.I.SrcReg[0].Negate)
 489                 presub_opcode = RC_PRESUB_SUB;
 490         else
 491                 presub_opcode = RC_PRESUB_ADD;
 492
 493         if (inst_add->U.I.SrcReg[1].Negate) {
 494                 inst_reader->U.I.PreSub.SrcReg[0] = inst_add->U.I.SrcReg[1];
 495                 inst_reader->U.I.PreSub.SrcReg[1] = inst_add->U.I.SrcReg[0];
 496         } else {
 497                 inst_reader->U.I.PreSub.SrcReg[0] = inst_add->U.I.SrcReg[0];
 498                 inst_reader->U.I.PreSub.SrcReg[1] = inst_add->U.I.SrcReg[1];
 499         }
 500         inst_reader->U.I.PreSub.SrcReg[0].Negate = 0;
 501         inst_reader->U.I.PreSub.SrcReg[1].Negate = 0;
 502         inst_reader->U.I.PreSub.Opcode = presub_opcode;
 503         inst_reader->U.I.SrcReg[src_index] =
 504                         chain_srcregs(inst_reader->U.I.SrcReg[src_index],
 505                                         inst_reader->U.I.PreSub.SrcReg[0]);
 506         inst_reader->U.I.SrcReg[src_index].File = RC_FILE_PRESUB;
 507         inst_reader->U.I.SrcReg[src_index].Index = presub_opcode;
 508 }
 509
 510 static int is_presub_candidate(
 511         struct radeon_compiler * c,
 512         struct rc_instruction * inst)
 513 {
 514         const struct rc_opcode_info * info = rc_get_opcode_info(inst->U.I.Opcode);
 515         unsigned int i;
 516         unsigned int is_constant[2] = {0, 0};
 517
 518         assert(inst->U.I.Opcode == RC_OPCODE_ADD);
 519
 520         if (inst->U.I.PreSub.Opcode != RC_PRESUB_NONE
 521                         || inst->U.I.SaturateMode
 522                         || inst->U.I.WriteALUResult) {
 523                 return 0;
 524         }
 525
 526         /* If both sources use a constant swizzle, then we can't convert it to
 527          * a presubtract operation.  In fact for the ADD and SUB presubtract
 528          * operations neither source can contain a constant swizzle.  This
 529          * specific case is checked in peephole_add_presub_add() when
 530          * we make sure the swizzles for both sources are equal, so we
 531          * don't need to worry about it here. */
 532         for (i = 0; i < 2; i++) {
 533                 int chan;
 534                 for (chan = 0; chan < 4; chan++) {
 535                         rc_swizzle swz =
 536                                 get_swz(inst->U.I.SrcReg[i].Swizzle, chan);
 537                         if (swz == RC_SWIZZLE_ONE
 538                                         || swz == RC_SWIZZLE_ZERO
 539                                         || swz == RC_SWIZZLE_HALF) {
 540                                 is_constant[i] = 1;
 541                         }
 542                 }
 543         }
 544         if (is_constant[0] && is_constant[1])
 545                 return 0;
 546
 547         for(i = 0; i < info->NumSrcRegs; i++) {
 548                 struct rc_src_register src = inst->U.I.SrcReg[i];
 549                 if (src_reads_dst_mask(src, inst->U.I.DstReg))
 550                         return 0;
 551
 552                 src.File = RC_FILE_PRESUB;
 553                 if (!c->SwizzleCaps->IsNative(inst->U.I.Opcode, src))
 554                         return 0;
 555         }
 556         return 1;
 557 }
 558
 559 static int peephole_add_presub_add(
 560         struct radeon_compiler * c,
 561         struct rc_instruction * inst_add)
 562 {
 563         unsigned dstmask = inst_add->U.I.DstReg.WriteMask;
 564         unsigned src0_neg = inst_add->U.I.SrcReg[0].Negate & dstmask;
 565         unsigned src1_neg = inst_add->U.I.SrcReg[1].Negate & dstmask;
 566
 567         if (inst_add->U.I.SrcReg[0].Swizzle != inst_add->U.I.SrcReg[1].Swizzle)
 568                 return 0;
 569
 570         /* src0 and src1 can't have absolute values */
 571         if (inst_add->U.I.SrcReg[0].Abs || inst_add->U.I.SrcReg[1].Abs)
 572                 return 0;
 573
 574         /* presub_replace_add() assumes only one is negative */
 575         if (inst_add->U.I.SrcReg[0].Negate && inst_add->U.I.SrcReg[1].Negate)
 576                 return 0;
 577
 578         /* if src0 is negative, at least all bits of dstmask have to be set */
 579         if (inst_add->U.I.SrcReg[0].Negate && src0_neg != dstmask)
 580                 return 0;
 581
 582         /* if src1 is negative, at least all bits of dstmask have to be set */
 583         if (inst_add->U.I.SrcReg[1].Negate && src1_neg != dstmask)
 584                 return 0;
 585
 586         if (!is_presub_candidate(c, inst_add))
 587                 return 0;
 588
 589         if (presub_helper(c, inst_add, RC_PRESUB_ADD, presub_replace_add)) {
 590                 rc_remove_instruction(inst_add);
 591                 return 1;
 592         }
 593         return 0;
 594 }
 595
 596 static void presub_replace_inv(
 597         struct rc_instruction * inst_add,
 598         struct rc_instruction * inst_reader,
 599         unsigned int src_index)
 600 {
 601         /* We must be careful not to modify inst_add, since it
 602          * is possible it will remain part of the program.*/
 603         inst_reader->U.I.PreSub.SrcReg[0] = inst_add->U.I.SrcReg[1];
 604         inst_reader->U.I.PreSub.SrcReg[0].Negate = 0;
 605         inst_reader->U.I.PreSub.Opcode = RC_PRESUB_INV;
 606         inst_reader->U.I.SrcReg[src_index] = chain_srcregs(inst_reader->U.I.SrcReg[src_index],
 607                                                 inst_reader->U.I.PreSub.SrcReg[0]);
 608
 609         inst_reader->U.I.SrcReg[src_index].File = RC_FILE_PRESUB;
 610         inst_reader->U.I.SrcReg[src_index].Index = RC_PRESUB_INV;
 611 }
 612
 613 /**
 614  * PRESUB_INV: ADD TEMP[0], none.1, -TEMP[1]
 615  * Use the presubtract 1 - src0 for all readers of TEMP[0].  The first source
 616  * of the add instruction must have the constatnt 1 swizzle.  This function
 617  * does not check const registers to see if their value is 1.0, so it should
 618  * be called after the constant_folding optimization.
 619  * @return
 620  *      0 if the ADD instruction is still part of the program.
 621  *      1 if the ADD instruction is no longer part of the program.
 622  */
 623 static int peephole_add_presub_inv(
 624         struct radeon_compiler * c,
 625         struct rc_instruction * inst_add)
 626 {
 627         unsigned int i, swz;
 628
 629         if (!is_presub_candidate(c, inst_add))
 630                 return 0;
 631
 632         /* Check if src0 is 1. */
 633         /* XXX It would be nice to use is_src_uniform_constant here, but that
 634          * function only works if the register's file is RC_FILE_NONE */
 635         for(i = 0; i < 4; i++ ) {
 636                 swz = GET_SWZ(inst_add->U.I.SrcReg[0].Swizzle, i);
 637                 if(((1 << i) & inst_add->U.I.DstReg.WriteMask)
 638                                                 && swz != RC_SWIZZLE_ONE) {
 639                         return 0;
 640                 }
 641         }
 642
 643         /* Check src1. */
 644         if ((inst_add->U.I.SrcReg[1].Negate & inst_add->U.I.DstReg.WriteMask) !=
 645                                                 inst_add->U.I.DstReg.WriteMask
 646                 || inst_add->U.I.SrcReg[1].Abs
 647                 || (inst_add->U.I.SrcReg[1].File != RC_FILE_TEMPORARY
 648                         && inst_add->U.I.SrcReg[1].File != RC_FILE_CONSTANT)
 649                 || src_has_const_swz(inst_add->U.I.SrcReg[1])) {
 650
 651                 return 0;
 652         }
 653
 654         if (presub_helper(c, inst_add, RC_PRESUB_INV, presub_replace_inv)) {
 655                 rc_remove_instruction(inst_add);
 656                 return 1;
 657         }
 658         return 0;
 659 }
 660
 661 /**
 662  * @return
 663  *      0 if inst is still part of the program.
 664  *      1 if inst is no longer part of the program.
 665  */
 666 static int peephole(struct radeon_compiler * c, struct rc_instruction * inst)
 667 {
 668         switch(inst->U.I.Opcode){
 669         case RC_OPCODE_ADD:
 670                 if (c->has_presub) {
 671                         if(peephole_add_presub_inv(c, inst))
 672                                 return 1;
 673                         if(peephole_add_presub_add(c, inst))
 674                                 return 1;
 675                 }
 676                 break;
 677         default:
 678                 break;
 679         }
 680         return 0;
 681 }
 682
 683 void rc_optimize(struct radeon_compiler * c, void *user)
 684 {
 685         struct rc_instruction * inst = c->Program.Instructions.Next;
 686         while(inst != &c->Program.Instructions) {
 687                 struct rc_instruction * cur = inst;
 688                 inst = inst->Next;
 689
 690                 constant_folding(c, cur);
 691
 692                 if(peephole(c, cur))
 693                         continue;
 694
 695                 if (cur->U.I.Opcode == RC_OPCODE_MOV) {
 696                         copy_propagate(c, cur);
 697                         /* cur may no longer be part of the program */
 698                 }
 699         }
 700 }