gl/mesa/src/gallium/drivers/r300/compiler/radeon_emulate_branches.c

   1 /*
   2  * Copyright 2009 Nicolai Hähnle <nhaehnle@gmail.com>
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * on the rights to use, copy, modify, merge, publish, distribute, sub
   8  * license, and/or sell copies of the Software, and to permit persons to whom
   9  * the Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  18  * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
  19  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  20  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  21  * USE OR OTHER DEALINGS IN THE SOFTWARE. */
  22
  23 #include "radeon_emulate_branches.h"
  24
  25 #include <stdio.h>
  26
  27 #include "radeon_compiler.h"
  28 #include "radeon_dataflow.h"
  29
  30 #define VERBOSE 0
  31
  32 #define DBG(...) do { if (VERBOSE) fprintf(stderr, __VA_ARGS__); } while(0)
  33
  34
  35 struct proxy_info {
  36         unsigned int Proxied:1;
  37         unsigned int Index:RC_REGISTER_INDEX_BITS;
  38 };
  39
  40 struct register_proxies {
  41         struct proxy_info Temporary[RC_REGISTER_MAX_INDEX];
  42 };
  43
  44 struct branch_info {
  45         struct rc_instruction * If;
  46         struct rc_instruction * Else;
  47 };
  48
  49 struct emulate_branch_state {
  50         struct radeon_compiler * C;
  51
  52         struct branch_info * Branches;
  53         unsigned int BranchCount;
  54         unsigned int BranchReserved;
  55 };
  56
  57
  58 static void handle_if(struct emulate_branch_state * s, struct rc_instruction * inst)
  59 {
  60         struct branch_info * branch;
  61         struct rc_instruction * inst_mov;
  62
  63         memory_pool_array_reserve(&s->C->Pool, struct branch_info,
  64                         s->Branches, s->BranchCount, s->BranchReserved, 1);
  65
  66         DBG("%s\n", __FUNCTION__);
  67
  68         branch = &s->Branches[s->BranchCount++];
  69         memset(branch, 0, sizeof(struct branch_info));
  70         branch->If = inst;
  71
  72         /* Make a safety copy of the decision register, because we will need
  73          * it at ENDIF time and it might be overwritten in both branches. */
  74         inst_mov = rc_insert_new_instruction(s->C, inst->Prev);
  75         inst_mov->U.I.Opcode = RC_OPCODE_MOV;
  76         inst_mov->U.I.DstReg.File = RC_FILE_TEMPORARY;
  77         inst_mov->U.I.DstReg.Index = rc_find_free_temporary(s->C);
  78         inst_mov->U.I.DstReg.WriteMask = RC_MASK_X;
  79         inst_mov->U.I.SrcReg[0] = inst->U.I.SrcReg[0];
  80
  81         inst->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
  82         inst->U.I.SrcReg[0].Index = inst_mov->U.I.DstReg.Index;
  83         inst->U.I.SrcReg[0].Swizzle = 0;
  84         inst->U.I.SrcReg[0].Abs = 0;
  85         inst->U.I.SrcReg[0].Negate = 0;
  86 }
  87
  88 static void handle_else(struct emulate_branch_state * s, struct rc_instruction * inst)
  89 {
  90         struct branch_info * branch;
  91
  92         if (!s->BranchCount) {
  93                 rc_error(s->C, "Encountered ELSE outside of branches");
  94                 return;
  95         }
  96
  97         DBG("%s\n", __FUNCTION__);
  98
  99         branch = &s->Branches[s->BranchCount - 1];
 100         branch->Else = inst;
 101 }
 102
 103
 104 struct state_and_proxies {
 105         struct emulate_branch_state * S;
 106         struct register_proxies * Proxies;
 107 };
 108
 109 static struct proxy_info * get_proxy_info(struct state_and_proxies * sap,
 110                         rc_register_file file, unsigned int index)
 111 {
 112         if (file == RC_FILE_TEMPORARY) {
 113                 return &sap->Proxies->Temporary[index];
 114         } else {
 115                 return 0;
 116         }
 117 }
 118
 119 static void scan_write(void * userdata, struct rc_instruction * inst,
 120                 rc_register_file file, unsigned int index, unsigned int comp)
 121 {
 122         struct state_and_proxies * sap = userdata;
 123         struct proxy_info * proxy = get_proxy_info(sap, file, index);
 124
 125         if (proxy && !proxy->Proxied) {
 126                 proxy->Proxied = 1;
 127                 proxy->Index = rc_find_free_temporary(sap->S->C);
 128         }
 129 }
 130
 131 static void remap_proxy_function(void * userdata, struct rc_instruction * inst,
 132                 rc_register_file * pfile, unsigned int * pindex)
 133 {
 134         struct state_and_proxies * sap = userdata;
 135         struct proxy_info * proxy = get_proxy_info(sap, *pfile, *pindex);
 136
 137         if (proxy && proxy->Proxied) {
 138                 *pfile = RC_FILE_TEMPORARY;
 139                 *pindex = proxy->Index;
 140         }
 141 }
 142
 143 /**
 144  * Redirect all writes in the instruction range [begin, end) to proxy
 145  * temporary registers.
 146  */
 147 static void allocate_and_insert_proxies(struct emulate_branch_state * s,
 148                 struct register_proxies * proxies,
 149                 struct rc_instruction * begin,
 150                 struct rc_instruction * end)
 151 {
 152         struct state_and_proxies sap;
 153
 154         sap.S = s;
 155         sap.Proxies = proxies;
 156
 157         for(struct rc_instruction * inst = begin; inst != end; inst = inst->Next) {
 158                 rc_for_all_writes_mask(inst, scan_write, &sap);
 159                 rc_remap_registers(inst, remap_proxy_function, &sap);
 160         }
 161
 162         for(unsigned int index = 0; index < RC_REGISTER_MAX_INDEX; ++index) {
 163                 if (proxies->Temporary[index].Proxied) {
 164                         struct rc_instruction * inst_mov = rc_insert_new_instruction(s->C, begin->Prev);
 165                         inst_mov->U.I.Opcode = RC_OPCODE_MOV;
 166                         inst_mov->U.I.DstReg.File = RC_FILE_TEMPORARY;
 167                         inst_mov->U.I.DstReg.Index = proxies->Temporary[index].Index;
 168                         inst_mov->U.I.DstReg.WriteMask = RC_MASK_XYZW;
 169                         inst_mov->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
 170                         inst_mov->U.I.SrcReg[0].Index = index;
 171                 }
 172         }
 173 }
 174
 175
 176 static void inject_cmp(struct emulate_branch_state * s,
 177                 struct rc_instruction * inst_if,
 178                 struct rc_instruction * inst_endif,
 179                 rc_register_file file, unsigned int index,
 180                 struct proxy_info ifproxy,
 181                 struct proxy_info elseproxy)
 182 {
 183         struct rc_instruction * inst_cmp = rc_insert_new_instruction(s->C, inst_endif);
 184         inst_cmp->U.I.Opcode = RC_OPCODE_CMP;
 185         inst_cmp->U.I.DstReg.File = file;
 186         inst_cmp->U.I.DstReg.Index = index;
 187         inst_cmp->U.I.DstReg.WriteMask = RC_MASK_XYZW;
 188         inst_cmp->U.I.SrcReg[0] = inst_if->U.I.SrcReg[0];
 189         inst_cmp->U.I.SrcReg[0].Abs = 1;
 190         inst_cmp->U.I.SrcReg[0].Negate = RC_MASK_XYZW;
 191         inst_cmp->U.I.SrcReg[1].File = RC_FILE_TEMPORARY;
 192         inst_cmp->U.I.SrcReg[1].Index = ifproxy.Proxied ? ifproxy.Index : index;
 193         inst_cmp->U.I.SrcReg[2].File = RC_FILE_TEMPORARY;
 194         inst_cmp->U.I.SrcReg[2].Index = elseproxy.Proxied ? elseproxy.Index : index;
 195 }
 196
 197 static void handle_endif(struct emulate_branch_state * s, struct rc_instruction * inst)
 198 {
 199         struct branch_info * branch;
 200         struct register_proxies IfProxies;
 201         struct register_proxies ElseProxies;
 202
 203         if (!s->BranchCount) {
 204                 rc_error(s->C, "Encountered ENDIF outside of branches");
 205                 return;
 206         }
 207
 208         DBG("%s\n", __FUNCTION__);
 209
 210         branch = &s->Branches[s->BranchCount - 1];
 211
 212         memset(&IfProxies, 0, sizeof(IfProxies));
 213         memset(&ElseProxies, 0, sizeof(ElseProxies));
 214
 215         allocate_and_insert_proxies(s, &IfProxies, branch->If->Next, branch->Else ? branch->Else : inst);
 216
 217         if (branch->Else)
 218                 allocate_and_insert_proxies(s, &ElseProxies, branch->Else->Next, inst);
 219
 220         /* Insert the CMP instructions at the end. */
 221         for(unsigned int index = 0; index < RC_REGISTER_MAX_INDEX; ++index) {
 222                 if (IfProxies.Temporary[index].Proxied || ElseProxies.Temporary[index].Proxied) {
 223                         inject_cmp(s, branch->If, inst, RC_FILE_TEMPORARY, index,
 224                                         IfProxies.Temporary[index], ElseProxies.Temporary[index]);
 225                 }
 226         }
 227
 228         /* Remove all traces of the branch instructions */
 229         rc_remove_instruction(branch->If);
 230         if (branch->Else)
 231                 rc_remove_instruction(branch->Else);
 232         rc_remove_instruction(inst);
 233
 234         s->BranchCount--;
 235
 236         if (VERBOSE) {
 237                 DBG("Program after ENDIF handling:\n");
 238                 rc_print_program(&s->C->Program);
 239         }
 240 }
 241
 242
 243 struct remap_output_data {
 244         unsigned int Output:RC_REGISTER_INDEX_BITS;
 245         unsigned int Temporary:RC_REGISTER_INDEX_BITS;
 246 };
 247
 248 static void remap_output_function(void * userdata, struct rc_instruction * inst,
 249                 rc_register_file * pfile, unsigned int * pindex)
 250 {
 251         struct remap_output_data * data = userdata;
 252
 253         if (*pfile == RC_FILE_OUTPUT && *pindex == data->Output) {
 254                 *pfile = RC_FILE_TEMPORARY;
 255                 *pindex = data->Temporary;
 256         }
 257 }
 258
 259
 260 /**
 261  * Output registers cannot be read from and so cannot be dealt with like
 262  * temporary registers.
 263  *
 264  * We do the simplest thing: If an output registers is written within
 265  * a branch, then *all* writes to this register are proxied to a
 266  * temporary register, and a final MOV is appended to the end of
 267  * the program.
 268  */
 269 static void fix_output_writes(struct emulate_branch_state * s, struct rc_instruction * inst)
 270 {
 271         const struct rc_opcode_info * opcode;
 272
 273         if (!s->BranchCount)
 274                 return;
 275
 276         opcode = rc_get_opcode_info(inst->U.I.Opcode);
 277
 278         if (!opcode->HasDstReg)
 279                 return;
 280
 281         if (inst->U.I.DstReg.File == RC_FILE_OUTPUT) {
 282                 struct remap_output_data remap;
 283                 struct rc_instruction * inst_mov;
 284
 285                 remap.Output = inst->U.I.DstReg.Index;
 286                 remap.Temporary = rc_find_free_temporary(s->C);
 287
 288                 for(struct rc_instruction * inst = s->C->Program.Instructions.Next;
 289                     inst != &s->C->Program.Instructions;
 290                     inst = inst->Next) {
 291                         rc_remap_registers(inst, &remap_output_function, &remap);
 292                 }
 293
 294                 inst_mov = rc_insert_new_instruction(s->C, s->C->Program.Instructions.Prev);
 295                 inst_mov->U.I.Opcode = RC_OPCODE_MOV;
 296                 inst_mov->U.I.DstReg.File = RC_FILE_OUTPUT;
 297                 inst_mov->U.I.DstReg.Index = remap.Output;
 298                 inst_mov->U.I.DstReg.WriteMask = RC_MASK_XYZW;
 299                 inst_mov->U.I.SrcReg[0].File = RC_FILE_TEMPORARY;
 300                 inst_mov->U.I.SrcReg[0].Index = remap.Temporary;
 301         }
 302 }
 303
 304 /**
 305  * Remove branch instructions; instead, execute both branches
 306  * on different register sets and choose between their results
 307  * using CMP instructions in place of the original ENDIF.
 308  */
 309 void rc_emulate_branches(struct radeon_compiler *c, void *user)
 310 {
 311         struct emulate_branch_state s;
 312         struct rc_instruction * ptr;
 313
 314         memset(&s, 0, sizeof(s));
 315         s.C = c;
 316
 317         /* Untypical loop because we may remove the current instruction */
 318         ptr = c->Program.Instructions.Next;
 319         while(ptr != &c->Program.Instructions) {
 320                 struct rc_instruction * inst = ptr;
 321                 ptr = ptr->Next;
 322
 323                 if (inst->Type == RC_INSTRUCTION_NORMAL) {
 324                         switch(inst->U.I.Opcode) {
 325                         case RC_OPCODE_IF:
 326                                 handle_if(&s, inst);
 327                                 break;
 328                         case RC_OPCODE_ELSE:
 329                                 handle_else(&s, inst);
 330                                 break;
 331                         case RC_OPCODE_ENDIF:
 332                                 handle_endif(&s, inst);
 333                                 break;
 334                         default:
 335                                 fix_output_writes(&s, inst);
 336                                 break;
 337                         }
 338                 } else {
 339                         rc_error(c, "%s: unhandled instruction type\n", __FUNCTION__);
 340                 }
 341         }
 342 }