gcc/config/sh/sh-mem.cc

   1 /* Helper routines for memory move and comparison insns.
   2    Copyright (C) 2013-2016 Free Software Foundation, Inc.
   3
   4 This file is part of GCC.
   5
   6 GCC is free software; you can redistribute it and/or modify
   7 it under the terms of the GNU General Public License as published by
   8 the Free Software Foundation; either version 3, or (at your option)
   9 any later version.
  10
  11 GCC is distributed in the hope that it will be useful,
  12 but WITHOUT ANY WARRANTY; without even the implied warranty of
  13 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  14 GNU General Public License for more details.
  15
  16 You should have received a copy of the GNU General Public License
  17 along with GCC; see the file COPYING3.  If not see
  18 <http://www.gnu.org/licenses/>.  */
  19
  20 #include "config.h"
  21 #include "system.h"
  22 #include "coretypes.h"
  23 #include "tm.h"
  24 #include "function.h"
  25 #include "basic-block.h"
  26 #include "rtl.h"
  27 #include "tree.h"
  28 #include "tm_p.h"
  29 #include "emit-rtl.h"
  30 #include "explow.h"
  31 #include "expr.h"
  32
  33 /* Like force_operand, but guarantees that VALUE ends up in TARGET.  */
  34 static void
  35 force_into (rtx value, rtx target)
  36 {
  37   value = force_operand (value, target);
  38   if (! rtx_equal_p (value, target))
  39     emit_insn (gen_move_insn (target, value));
  40 }
  41
  42 /* Emit code to perform a block move.  Choose the best method.
  43
  44    OPERANDS[0] is the destination.
  45    OPERANDS[1] is the source.
  46    OPERANDS[2] is the size.
  47    OPERANDS[3] is the alignment safe to use.  */
  48 bool
  49 expand_block_move (rtx *operands)
  50 {
  51   int align = INTVAL (operands[3]);
  52   int constp = (CONST_INT_P (operands[2]));
  53   int bytes = (constp ? INTVAL (operands[2]) : 0);
  54
  55   if (! constp)
  56     return false;
  57
  58   /* If we could use mov.l to move words and dest is word-aligned, we
  59      can use movua.l for loads and still generate a relatively short
  60      and efficient sequence.  */
  61   if (TARGET_SH4A && align < 4
  62       && MEM_ALIGN (operands[0]) >= 32
  63       && can_move_by_pieces (bytes, 32))
  64     {
  65       rtx dest = copy_rtx (operands[0]);
  66       rtx src = copy_rtx (operands[1]);
  67       /* We could use different pseudos for each copied word, but
  68          since movua can only load into r0, it's kind of
  69          pointless.  */
  70       rtx temp = gen_reg_rtx (SImode);
  71       rtx src_addr = copy_addr_to_reg (XEXP (src, 0));
  72       int copied = 0;
  73
  74       while (copied + 4 <= bytes)
  75         {
  76           rtx to = adjust_address (dest, SImode, copied);
  77           rtx from = adjust_automodify_address (src, BLKmode,
  78                                                 src_addr, copied);
  79
  80           set_mem_size (from, 4);
  81           emit_insn (gen_movua (temp, from));
  82           emit_move_insn (src_addr, plus_constant (Pmode, src_addr, 4));
  83           emit_move_insn (to, temp);
  84           copied += 4;
  85         }
  86
  87       if (copied < bytes)
  88         move_by_pieces (adjust_address (dest, BLKmode, copied),
  89                         adjust_automodify_address (src, BLKmode,
  90                                                    src_addr, copied),
  91                         bytes - copied, align, 0);
  92
  93       return true;
  94     }
  95
  96   /* If it isn't a constant number of bytes, or if it doesn't have 4 byte
  97      alignment, or if it isn't a multiple of 4 bytes, then fail.  */
  98   if (align < 4 || (bytes % 4 != 0))
  99     return false;
 100
 101   if (TARGET_HARD_SH4)
 102     {
 103       if (bytes < 12)
 104         return false;
 105       else if (bytes == 12)
 106         {
 107           rtx func_addr_rtx = gen_reg_rtx (Pmode);
 108           rtx r4 = gen_rtx_REG (SImode, 4);
 109           rtx r5 = gen_rtx_REG (SImode, 5);
 110
 111           rtx lab = function_symbol (func_addr_rtx, "__movmemSI12_i4",
 112                                      SFUNC_STATIC).lab;
 113           force_into (XEXP (operands[0], 0), r4);
 114           force_into (XEXP (operands[1], 0), r5);
 115           emit_insn (gen_block_move_real_i4 (func_addr_rtx, lab));
 116           return true;
 117         }
 118       else if (! optimize_size)
 119         {
 120           rtx func_addr_rtx = gen_reg_rtx (Pmode);
 121           rtx r4 = gen_rtx_REG (SImode, 4);
 122           rtx r5 = gen_rtx_REG (SImode, 5);
 123           rtx r6 = gen_rtx_REG (SImode, 6);
 124
 125           rtx lab = function_symbol (func_addr_rtx, bytes & 4
 126                                                     ? "__movmem_i4_odd"
 127                                                     : "__movmem_i4_even",
 128                                      SFUNC_STATIC).lab;
 129           force_into (XEXP (operands[0], 0), r4);
 130           force_into (XEXP (operands[1], 0), r5);
 131
 132           int dwords = bytes >> 3;
 133           emit_insn (gen_move_insn (r6, GEN_INT (dwords - 1)));
 134           emit_insn (gen_block_lump_real_i4 (func_addr_rtx, lab));
 135           return true;
 136         }
 137       else
 138         return false;
 139     }
 140   if (bytes < 64)
 141     {
 142       char entry[30];
 143       rtx func_addr_rtx = gen_reg_rtx (Pmode);
 144       rtx r4 = gen_rtx_REG (SImode, 4);
 145       rtx r5 = gen_rtx_REG (SImode, 5);
 146
 147       sprintf (entry, "__movmemSI%d", bytes);
 148       rtx lab = function_symbol (func_addr_rtx, entry, SFUNC_STATIC).lab;
 149       force_into (XEXP (operands[0], 0), r4);
 150       force_into (XEXP (operands[1], 0), r5);
 151       emit_insn (gen_block_move_real (func_addr_rtx, lab));
 152       return true;
 153     }
 154
 155   /* This is the same number of bytes as a memcpy call, but to a different
 156      less common function name, so this will occasionally use more space.  */
 157   if (! optimize_size)
 158     {
 159       rtx func_addr_rtx = gen_reg_rtx (Pmode);
 160       int final_switch, while_loop;
 161       rtx r4 = gen_rtx_REG (SImode, 4);
 162       rtx r5 = gen_rtx_REG (SImode, 5);
 163       rtx r6 = gen_rtx_REG (SImode, 6);
 164
 165       rtx lab = function_symbol (func_addr_rtx, "__movmem", SFUNC_STATIC).lab;
 166       force_into (XEXP (operands[0], 0), r4);
 167       force_into (XEXP (operands[1], 0), r5);
 168
 169       /* r6 controls the size of the move.  16 is decremented from it
 170          for each 64 bytes moved.  Then the negative bit left over is used
 171          as an index into a list of move instructions.  e.g., a 72 byte move
 172          would be set up with size(r6) = 14, for one iteration through the
 173          big while loop, and a switch of -2 for the last part.  */
 174
 175       final_switch = 16 - ((bytes / 4) % 16);
 176       while_loop = ((bytes / 4) / 16 - 1) * 16;
 177       emit_insn (gen_move_insn (r6, GEN_INT (while_loop + final_switch)));
 178       emit_insn (gen_block_lump_real (func_addr_rtx, lab));
 179       return true;
 180     }
 181
 182   return false;
 183 }
 184
 185 static const int prob_unlikely = REG_BR_PROB_BASE / 10;
 186 static const int prob_likely = REG_BR_PROB_BASE / 4;
 187
 188 /* Emit code to perform a strcmp.
 189
 190    OPERANDS[0] is the destination.
 191    OPERANDS[1] is the first string.
 192    OPERANDS[2] is the second string.
 193    OPERANDS[3] is the known alignment.  */
 194 bool
 195 sh_expand_cmpstr (rtx *operands)
 196 {
 197   rtx addr1 = operands[1];
 198   rtx addr2 = operands[2];
 199   rtx s1_addr = copy_addr_to_reg (XEXP (addr1, 0));
 200   rtx s2_addr = copy_addr_to_reg (XEXP (addr2, 0));
 201   rtx tmp0 = gen_reg_rtx (SImode);
 202   rtx tmp1 = gen_reg_rtx (SImode);
 203   rtx tmp2 = gen_reg_rtx (SImode);
 204   rtx tmp3 = gen_reg_rtx (SImode);
 205
 206   rtx jump;
 207   rtx_code_label *L_return = gen_label_rtx ();
 208   rtx_code_label *L_loop_byte = gen_label_rtx ();
 209   rtx_code_label *L_end_loop_byte = gen_label_rtx ();
 210   rtx_code_label *L_loop_long = gen_label_rtx ();
 211   rtx_code_label *L_end_loop_long = gen_label_rtx ();
 212
 213   const unsigned int addr1_alignment = MEM_ALIGN (operands[1]) / BITS_PER_UNIT;
 214   const unsigned int addr2_alignment = MEM_ALIGN (operands[2]) / BITS_PER_UNIT;
 215
 216   if (addr1_alignment < 4 && addr2_alignment < 4)
 217     {
 218       emit_insn (gen_iorsi3 (tmp1, s1_addr, s2_addr));
 219       emit_insn (gen_tstsi_t (tmp1, GEN_INT (3)));
 220       jump = emit_jump_insn (gen_branch_false (L_loop_byte));
 221       add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 222     }
 223   else if (addr1_alignment < 4 && addr2_alignment >= 4)
 224     {
 225       emit_insn (gen_tstsi_t (s1_addr, GEN_INT (3)));
 226       jump = emit_jump_insn (gen_branch_false (L_loop_byte));
 227       add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 228     }
 229   else if (addr1_alignment >= 4 && addr2_alignment < 4)
 230     {
 231       emit_insn (gen_tstsi_t (s2_addr, GEN_INT (3)));
 232       jump = emit_jump_insn (gen_branch_false (L_loop_byte));
 233       add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 234     }
 235
 236   addr1 = adjust_automodify_address (addr1, SImode, s1_addr, 0);
 237   addr2 = adjust_automodify_address (addr2, SImode, s2_addr, 0);
 238
 239   /* tmp2 is aligned, OK to load.  */
 240   emit_move_insn (tmp3, addr2);
 241   emit_move_insn (s2_addr, plus_constant (Pmode, s2_addr, 4));
 242
 243   /* start long loop.  */
 244   emit_label (L_loop_long);
 245
 246   emit_move_insn (tmp2, tmp3);
 247
 248   /* tmp1 is aligned, OK to load.  */
 249   emit_move_insn (tmp1, addr1);
 250   emit_move_insn (s1_addr, plus_constant (Pmode, s1_addr, 4));
 251
 252   /* Is there a 0 byte ?  */
 253   emit_insn (gen_andsi3 (tmp3, tmp3, tmp1));
 254
 255   emit_insn (gen_cmpstr_t (tmp0, tmp3));
 256   jump = emit_jump_insn (gen_branch_true (L_end_loop_long));
 257   add_int_reg_note (jump, REG_BR_PROB, prob_unlikely);
 258
 259   emit_insn (gen_cmpeqsi_t (tmp1, tmp2));
 260
 261   /* tmp2 is aligned, OK to load.  */
 262   emit_move_insn (tmp3, addr2);
 263   emit_move_insn (s2_addr, plus_constant (Pmode, s2_addr, 4));
 264
 265   jump = emit_jump_insn (gen_branch_true (L_loop_long));
 266   add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 267   /* end loop.  */
 268
 269   /* Fallthu, substract words.  */
 270   if (TARGET_LITTLE_ENDIAN)
 271     {
 272       rtx low_1 = gen_lowpart (HImode, tmp1);
 273       rtx low_2 = gen_lowpart (HImode, tmp2);
 274
 275       emit_insn (gen_rotlhi3_8 (low_1, low_1));
 276       emit_insn (gen_rotlhi3_8 (low_2, low_2));
 277       emit_insn (gen_rotlsi3_16 (tmp1, tmp1));
 278       emit_insn (gen_rotlsi3_16 (tmp2, tmp2));
 279       emit_insn (gen_rotlhi3_8 (low_1, low_1));
 280       emit_insn (gen_rotlhi3_8 (low_2, low_2));
 281     }
 282
 283   jump = emit_jump_insn (gen_jump_compact (L_return));
 284   emit_barrier_after (jump);
 285
 286   emit_label (L_end_loop_long);
 287
 288   emit_move_insn (s1_addr, plus_constant (Pmode, s1_addr, -4));
 289   emit_move_insn (s2_addr, plus_constant (Pmode, s2_addr, -4));
 290
 291   /* start byte loop.  */
 292   addr1 = adjust_address (addr1, QImode, 0);
 293   addr2 = adjust_address (addr2, QImode, 0);
 294
 295   emit_label (L_loop_byte);
 296
 297   emit_insn (gen_extendqisi2 (tmp2, addr2));
 298   emit_move_insn (s2_addr, plus_constant (Pmode, s2_addr, 1));
 299
 300   emit_insn (gen_extendqisi2 (tmp1, addr1));
 301   emit_move_insn (s1_addr, plus_constant (Pmode, s1_addr, 1));
 302
 303   emit_insn (gen_cmpeqsi_t (tmp2, const0_rtx));
 304   jump = emit_jump_insn (gen_branch_true (L_end_loop_byte));
 305   add_int_reg_note (jump, REG_BR_PROB, prob_unlikely);
 306
 307   emit_insn (gen_cmpeqsi_t (tmp1, tmp2));
 308   if (flag_delayed_branch)
 309     emit_insn (gen_zero_extendqisi2 (tmp2, gen_lowpart (QImode, tmp2)));
 310   jump = emit_jump_insn (gen_branch_true (L_loop_byte));
 311   add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 312   /* end loop.  */
 313
 314   emit_label (L_end_loop_byte);
 315
 316   if (! flag_delayed_branch)
 317     emit_insn (gen_zero_extendqisi2 (tmp2, gen_lowpart (QImode, tmp2)));
 318   emit_insn (gen_zero_extendqisi2 (tmp1, gen_lowpart (QImode, tmp1)));
 319
 320   emit_label (L_return);
 321
 322   emit_insn (gen_subsi3 (operands[0], tmp1, tmp2));
 323
 324   return true;
 325 }
 326
 327 /* Emit code to perform a strncmp.
 328
 329    OPERANDS[0] is the destination.
 330    OPERANDS[1] is the first string.
 331    OPERANDS[2] is the second string.
 332    OPERANDS[3] is the length.
 333    OPERANDS[4] is the known alignment.  */
 334 bool
 335 sh_expand_cmpnstr (rtx *operands)
 336 {
 337   rtx addr1 = operands[1];
 338   rtx addr2 = operands[2];
 339   rtx s1_addr = copy_addr_to_reg (XEXP (addr1, 0));
 340   rtx s2_addr = copy_addr_to_reg (XEXP (addr2, 0));
 341   rtx tmp1 = gen_reg_rtx (SImode);
 342   rtx tmp2 = gen_reg_rtx (SImode);
 343
 344   rtx jump;
 345   rtx_code_label *L_return = gen_label_rtx ();
 346   rtx_code_label *L_loop_byte = gen_label_rtx ();
 347   rtx_code_label *L_end_loop_byte = gen_label_rtx ();
 348
 349   rtx len = force_reg (SImode, operands[3]);
 350   int constp = CONST_INT_P (operands[3]);
 351
 352   const unsigned int addr1_alignment = MEM_ALIGN (operands[1]) / BITS_PER_UNIT;
 353   const unsigned int addr2_alignment = MEM_ALIGN (operands[2]) / BITS_PER_UNIT;
 354
 355   /* Loop on a register count.  */
 356   if (constp)
 357     {
 358       rtx tmp0 = gen_reg_rtx (SImode);
 359       rtx tmp3 = gen_reg_rtx (SImode);
 360       rtx lenw = gen_reg_rtx (SImode);
 361
 362       rtx_code_label *L_loop_long = gen_label_rtx ();
 363       rtx_code_label *L_end_loop_long = gen_label_rtx ();
 364
 365       int bytes = INTVAL (operands[3]);
 366       int witers = bytes / 4;
 367
 368       if (witers > 1)
 369         {
 370           addr1 = adjust_automodify_address (addr1, SImode, s1_addr, 0);
 371           addr2 = adjust_automodify_address (addr2, SImode, s2_addr, 0);
 372
 373           emit_move_insn (tmp0, const0_rtx);
 374
 375           if (addr1_alignment < 4 && addr2_alignment < 4)
 376             {
 377               emit_insn (gen_iorsi3 (tmp1, s1_addr, s2_addr));
 378               emit_insn (gen_tstsi_t (tmp1, GEN_INT (3)));
 379               jump = emit_jump_insn (gen_branch_false (L_loop_byte));
 380               add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 381             }
 382           else if (addr1_alignment < 4 && addr2_alignment >= 4)
 383             {
 384               emit_insn (gen_tstsi_t (s1_addr, GEN_INT (3)));
 385               jump = emit_jump_insn (gen_branch_false (L_loop_byte));
 386               add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 387             }
 388           else if (addr1_alignment >= 4 && addr2_alignment < 4)
 389             {
 390               emit_insn (gen_tstsi_t (s2_addr, GEN_INT (3)));
 391               jump = emit_jump_insn (gen_branch_false (L_loop_byte));
 392               add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 393             }
 394
 395           /* word count. Do we have iterations ?  */
 396           emit_insn (gen_lshrsi3 (lenw, len, GEN_INT (2)));
 397
 398           /* start long loop.  */
 399           emit_label (L_loop_long);
 400
 401           /* tmp2 is aligned, OK to load.  */
 402           emit_move_insn (tmp2, addr2);
 403           emit_move_insn (s2_addr, plus_constant (Pmode, s2_addr,
 404                                                   GET_MODE_SIZE (SImode)));
 405
 406           /* tmp1 is aligned, OK to load.  */
 407           emit_move_insn (tmp1, addr1);
 408           emit_move_insn (s1_addr, plus_constant (Pmode, s1_addr,
 409                                                   GET_MODE_SIZE (SImode)));
 410
 411           /* Is there a 0 byte ?  */
 412           emit_insn (gen_andsi3 (tmp3, tmp2, tmp1));
 413
 414           emit_insn (gen_cmpstr_t (tmp0, tmp3));
 415           jump = emit_jump_insn (gen_branch_true (L_end_loop_long));
 416           add_int_reg_note (jump, REG_BR_PROB, prob_unlikely);
 417
 418           emit_insn (gen_cmpeqsi_t (tmp1, tmp2));
 419           jump = emit_jump_insn (gen_branch_false (L_end_loop_long));
 420           add_int_reg_note (jump, REG_BR_PROB, prob_unlikely);
 421
 422           if (TARGET_SH2)
 423             emit_insn (gen_dect (lenw, lenw));
 424           else
 425             {
 426               emit_insn (gen_addsi3 (lenw, lenw, GEN_INT (-1)));
 427               emit_insn (gen_tstsi_t (lenw, lenw));
 428             }
 429
 430           jump = emit_jump_insn (gen_branch_false (L_loop_long));
 431           add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 432
 433           int sbytes = bytes % 4;
 434
 435           /* end loop.  Reached max iterations.  */
 436           if (sbytes == 0)
 437             {
 438               emit_insn (gen_subsi3 (operands[0], tmp1, tmp2));
 439               jump = emit_jump_insn (gen_jump_compact (L_return));
 440               emit_barrier_after (jump);
 441             }
 442           else
 443             {
 444               /* Remaining bytes to check.  */
 445
 446               addr1 = adjust_automodify_address (addr1, QImode, s1_addr, 0);
 447               addr2 = adjust_automodify_address (addr2, QImode, s2_addr, 0);
 448
 449               while (sbytes--)
 450                 {
 451                   emit_insn (gen_extendqisi2 (tmp1, addr1));
 452                   emit_insn (gen_extendqisi2 (tmp2, addr2));
 453
 454                   emit_insn (gen_cmpeqsi_t (tmp2, const0_rtx));
 455                   jump = emit_jump_insn (gen_branch_true (L_end_loop_byte));
 456                   add_int_reg_note (jump, REG_BR_PROB, prob_unlikely);
 457
 458                   emit_insn (gen_cmpeqsi_t (tmp1, tmp2));
 459                   if (flag_delayed_branch)
 460                     emit_insn (gen_zero_extendqisi2 (tmp2,
 461                                                      gen_lowpart (QImode,
 462                                                                   tmp2)));
 463                   jump = emit_jump_insn (gen_branch_false (L_end_loop_byte));
 464                   add_int_reg_note (jump, REG_BR_PROB, prob_unlikely);
 465
 466                   addr1 = adjust_address (addr1, QImode,
 467                                           GET_MODE_SIZE (QImode));
 468                   addr2 = adjust_address (addr2, QImode,
 469                                           GET_MODE_SIZE (QImode));
 470                 }
 471
 472               jump = emit_jump_insn (gen_jump_compact( L_end_loop_byte));
 473               emit_barrier_after (jump);
 474             }
 475
 476           emit_label (L_end_loop_long);
 477
 478           /* Found last word.  Restart it byte per byte.  */
 479
 480           emit_move_insn (s1_addr, plus_constant (Pmode, s1_addr,
 481                                                   -GET_MODE_SIZE (SImode)));
 482           emit_move_insn (s2_addr, plus_constant (Pmode, s2_addr,
 483                                                   -GET_MODE_SIZE (SImode)));
 484
 485           /* fall thru.  */
 486         }
 487
 488       addr1 = adjust_automodify_address (addr1, QImode, s1_addr, 0);
 489       addr2 = adjust_automodify_address (addr2, QImode, s2_addr, 0);
 490
 491       while (bytes--)
 492         {
 493           emit_insn (gen_extendqisi2 (tmp1, addr1));
 494           emit_insn (gen_extendqisi2 (tmp2, addr2));
 495
 496           emit_insn (gen_cmpeqsi_t (tmp2, const0_rtx));
 497           jump = emit_jump_insn (gen_branch_true (L_end_loop_byte));
 498           add_int_reg_note (jump, REG_BR_PROB, prob_unlikely);
 499
 500           emit_insn (gen_cmpeqsi_t (tmp1, tmp2));
 501           if (flag_delayed_branch)
 502             emit_insn (gen_zero_extendqisi2 (tmp2,
 503                                              gen_lowpart (QImode, tmp2)));
 504           jump = emit_jump_insn (gen_branch_false (L_end_loop_byte));
 505           add_int_reg_note (jump, REG_BR_PROB, prob_unlikely);
 506
 507           addr1 = adjust_address (addr1, QImode, GET_MODE_SIZE (QImode));
 508           addr2 = adjust_address (addr2, QImode, GET_MODE_SIZE (QImode));
 509         }
 510
 511       jump = emit_jump_insn (gen_jump_compact( L_end_loop_byte));
 512       emit_barrier_after (jump);
 513     }
 514   else
 515     {
 516       emit_insn (gen_cmpeqsi_t (len, const0_rtx));
 517       emit_move_insn (operands[0], const0_rtx);
 518       jump = emit_jump_insn (gen_branch_true (L_return));
 519       add_int_reg_note (jump, REG_BR_PROB, prob_unlikely);
 520     }
 521
 522   addr1 = adjust_automodify_address (addr1, QImode, s1_addr, 0);
 523   addr2 = adjust_automodify_address (addr2, QImode, s2_addr, 0);
 524
 525   emit_label (L_loop_byte);
 526
 527   emit_insn (gen_extendqisi2 (tmp2, addr2));
 528   emit_move_insn (s2_addr, plus_constant (Pmode, s2_addr, 1));
 529
 530   emit_insn (gen_extendqisi2 (tmp1, addr1));
 531   emit_move_insn (s1_addr, plus_constant (Pmode, s1_addr, 1));
 532
 533   emit_insn (gen_cmpeqsi_t (tmp2, const0_rtx));
 534   jump = emit_jump_insn (gen_branch_true (L_end_loop_byte));
 535   add_int_reg_note (jump, REG_BR_PROB, prob_unlikely);
 536
 537   emit_insn (gen_cmpeqsi_t (tmp1, tmp2));
 538   if (flag_delayed_branch)
 539     emit_insn (gen_zero_extendqisi2 (tmp2, gen_lowpart (QImode, tmp2)));
 540   jump = emit_jump_insn (gen_branch_false (L_end_loop_byte));
 541   add_int_reg_note (jump, REG_BR_PROB, prob_unlikely);
 542
 543   if (TARGET_SH2)
 544     emit_insn (gen_dect (len, len));
 545   else
 546     {
 547       emit_insn (gen_addsi3 (len, len, GEN_INT (-1)));
 548       emit_insn (gen_tstsi_t (len, len));
 549     }
 550
 551   jump = emit_jump_insn (gen_branch_false (L_loop_byte));
 552   add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 553   /* end byte loop.  */
 554
 555   emit_label (L_end_loop_byte);
 556
 557   if (! flag_delayed_branch)
 558     emit_insn (gen_zero_extendqisi2 (tmp2, gen_lowpart (QImode, tmp2)));
 559   emit_insn (gen_zero_extendqisi2 (tmp1, gen_lowpart (QImode, tmp1)));
 560
 561   emit_insn (gen_subsi3 (operands[0], tmp1, tmp2));
 562
 563   emit_label (L_return);
 564
 565   return true;
 566 }
 567
 568 /* Emit code to perform a strlen.
 569
 570    OPERANDS[0] is the destination.
 571    OPERANDS[1] is the string.
 572    OPERANDS[2] is the char to search.
 573    OPERANDS[3] is the alignment.  */
 574 bool
 575 sh_expand_strlen (rtx *operands)
 576 {
 577   rtx addr1 = operands[1];
 578   rtx current_addr = copy_addr_to_reg (XEXP (addr1, 0));
 579   rtx start_addr = gen_reg_rtx (Pmode);
 580   rtx tmp0 = gen_reg_rtx (SImode);
 581   rtx tmp1 = gen_reg_rtx (SImode);
 582   rtx_code_label *L_return = gen_label_rtx ();
 583   rtx_code_label *L_loop_byte = gen_label_rtx ();
 584
 585   rtx jump;
 586   rtx_code_label *L_loop_long = gen_label_rtx ();
 587   rtx_code_label *L_end_loop_long = gen_label_rtx ();
 588
 589   int align = INTVAL (operands[3]);
 590
 591   emit_move_insn (operands[0], GEN_INT (-1));
 592
 593   /* remember start of string.  */
 594   emit_move_insn (start_addr, current_addr);
 595
 596   if (align < 4)
 597     {
 598       emit_insn (gen_tstsi_t (current_addr, GEN_INT (3)));
 599       jump = emit_jump_insn (gen_branch_false (L_loop_byte));
 600       add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 601     }
 602
 603   emit_move_insn (tmp0, operands[2]);
 604
 605   addr1 = adjust_automodify_address (addr1, SImode, current_addr, 0);
 606
 607   /* start long loop.  */
 608   emit_label (L_loop_long);
 609
 610   /* tmp1 is aligned, OK to load.  */
 611   emit_move_insn (tmp1, addr1);
 612   emit_move_insn (current_addr, plus_constant (Pmode, current_addr, 4));
 613
 614   /* Is there a 0 byte ?  */
 615   emit_insn (gen_cmpstr_t (tmp0, tmp1));
 616
 617   jump = emit_jump_insn (gen_branch_false (L_loop_long));
 618   add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 619   /* end loop.  */
 620
 621   emit_label (L_end_loop_long);
 622
 623   emit_move_insn (current_addr, plus_constant (Pmode, current_addr, -4));
 624
 625   addr1 = adjust_address (addr1, QImode, 0);
 626
 627   /* unroll remaining bytes.  */
 628   for (int i = 0; i < 4; ++i)
 629     {
 630       emit_insn (gen_extendqisi2 (tmp1, addr1));
 631       emit_move_insn (current_addr, plus_constant (Pmode, current_addr, 1));
 632       emit_insn (gen_cmpeqsi_t (tmp1, const0_rtx));
 633       jump = emit_jump_insn (gen_branch_true (L_return));
 634       add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 635     }
 636
 637   emit_barrier_after (jump);
 638
 639   /* start byte loop.  */
 640   emit_label (L_loop_byte);
 641
 642   emit_insn (gen_extendqisi2 (tmp1, addr1));
 643   emit_move_insn (current_addr, plus_constant (Pmode, current_addr, 1));
 644
 645   emit_insn (gen_cmpeqsi_t (tmp1, const0_rtx));
 646   jump = emit_jump_insn (gen_branch_false (L_loop_byte));
 647   add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 648
 649   /* end loop.  */
 650
 651   emit_label (L_return);
 652
 653   emit_insn (gen_addsi3 (start_addr, start_addr, GEN_INT (1)));
 654   emit_insn (gen_subsi3 (operands[0], current_addr, start_addr));
 655
 656   return true;
 657 }
 658
 659 /* Emit code to perform a memset.
 660
 661    OPERANDS[0] is the destination.
 662    OPERANDS[1] is the size;
 663    OPERANDS[2] is the char to search.
 664    OPERANDS[3] is the alignment.  */
 665 void
 666 sh_expand_setmem (rtx *operands)
 667 {
 668   rtx_code_label *L_loop_byte = gen_label_rtx ();
 669   rtx_code_label *L_loop_word = gen_label_rtx ();
 670   rtx_code_label *L_return = gen_label_rtx ();
 671   rtx jump;
 672   rtx dest = copy_rtx (operands[0]);
 673   rtx dest_addr = copy_addr_to_reg (XEXP (dest, 0));
 674   rtx val = force_reg (SImode, operands[2]);
 675   int align = INTVAL (operands[3]);
 676   rtx len = force_reg (SImode, operands[1]);
 677
 678   if (! CONST_INT_P (operands[1]))
 679     return;
 680
 681   int count = INTVAL (operands[1]);
 682
 683   if (CONST_INT_P (operands[2])
 684       && (INTVAL (operands[2]) == 0 || INTVAL (operands[2]) == -1) && count > 8)
 685     {
 686       rtx lenw = gen_reg_rtx (SImode);
 687
 688       if (align < 4)
 689         {
 690           emit_insn (gen_tstsi_t (dest_addr, GEN_INT (3)));
 691           jump = emit_jump_insn (gen_branch_false (L_loop_byte));
 692           add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 693         }
 694
 695       /* word count. Do we have iterations ?  */
 696       emit_insn (gen_lshrsi3 (lenw, len, GEN_INT (2)));
 697
 698       dest = adjust_automodify_address (dest, SImode, dest_addr, 0);
 699
 700       /* start loop.  */
 701       emit_label (L_loop_word);
 702
 703       if (TARGET_SH2)
 704         emit_insn (gen_dect (lenw, lenw));
 705       else
 706         {
 707           emit_insn (gen_addsi3 (lenw, lenw, GEN_INT (-1)));
 708           emit_insn (gen_tstsi_t (lenw, lenw));
 709         }
 710
 711       emit_move_insn (dest, val);
 712       emit_move_insn (dest_addr, plus_constant (Pmode, dest_addr,
 713                                                 GET_MODE_SIZE (SImode)));
 714
 715
 716       jump = emit_jump_insn (gen_branch_false (L_loop_word));
 717       add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 718       count = count % 4;
 719
 720       dest = adjust_address (dest, QImode, 0);
 721
 722       val = gen_lowpart (QImode, val);
 723
 724       while (count--)
 725         {
 726           emit_move_insn (dest, val);
 727           emit_move_insn (dest_addr, plus_constant (Pmode, dest_addr,
 728                                                     GET_MODE_SIZE (QImode)));
 729         }
 730
 731       jump = emit_jump_insn (gen_jump_compact (L_return));
 732       emit_barrier_after (jump);
 733     }
 734
 735   dest = adjust_automodify_address (dest, QImode, dest_addr, 0);
 736
 737   /* start loop.  */
 738   emit_label (L_loop_byte);
 739
 740   if (TARGET_SH2)
 741     emit_insn (gen_dect (len, len));
 742   else
 743     {
 744       emit_insn (gen_addsi3 (len, len, GEN_INT (-1)));
 745       emit_insn (gen_tstsi_t (len, len));
 746     }
 747
 748   val = gen_lowpart (QImode, val);
 749   emit_move_insn (dest, val);
 750   emit_move_insn (dest_addr, plus_constant (Pmode, dest_addr,
 751                                             GET_MODE_SIZE (QImode)));
 752
 753   jump = emit_jump_insn (gen_branch_false (L_loop_byte));
 754   add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 755
 756   emit_label (L_return);
 757 }