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 "memmodel.h"
  29 #include "tm_p.h"
  30 #include "emit-rtl.h"
  31 #include "explow.h"
  32 #include "expr.h"
  33
  34 /* Like force_operand, but guarantees that VALUE ends up in TARGET.  */
  35 static void
  36 force_into (rtx value, rtx target)
  37 {
  38   value = force_operand (value, target);
  39   if (! rtx_equal_p (value, target))
  40     emit_insn (gen_move_insn (target, value));
  41 }
  42
  43 /* Emit code to perform a block move.  Choose the best method.
  44
  45    OPERANDS[0] is the destination.
  46    OPERANDS[1] is the source.
  47    OPERANDS[2] is the size.
  48    OPERANDS[3] is the alignment safe to use.  */
  49 bool
  50 expand_block_move (rtx *operands)
  51 {
  52   int align = INTVAL (operands[3]);
  53   int constp = (CONST_INT_P (operands[2]));
  54   int bytes = (constp ? INTVAL (operands[2]) : 0);
  55
  56   if (! constp)
  57     return false;
  58
  59   /* If we could use mov.l to move words and dest is word-aligned, we
  60      can use movua.l for loads and still generate a relatively short
  61      and efficient sequence.  */
  62   if (TARGET_SH4A && align < 4
  63       && MEM_ALIGN (operands[0]) >= 32
  64       && can_move_by_pieces (bytes, 32))
  65     {
  66       rtx dest = copy_rtx (operands[0]);
  67       rtx src = copy_rtx (operands[1]);
  68       /* We could use different pseudos for each copied word, but
  69          since movua can only load into r0, it's kind of
  70          pointless.  */
  71       rtx temp = gen_reg_rtx (SImode);
  72       rtx src_addr = copy_addr_to_reg (XEXP (src, 0));
  73       int copied = 0;
  74
  75       while (copied + 4 <= bytes)
  76         {
  77           rtx to = adjust_address (dest, SImode, copied);
  78           rtx from = adjust_automodify_address (src, BLKmode,
  79                                                 src_addr, copied);
  80
  81           set_mem_size (from, 4);
  82           emit_insn (gen_movua (temp, from));
  83           emit_move_insn (src_addr, plus_constant (Pmode, src_addr, 4));
  84           emit_move_insn (to, temp);
  85           copied += 4;
  86         }
  87
  88       if (copied < bytes)
  89         move_by_pieces (adjust_address (dest, BLKmode, copied),
  90                         adjust_automodify_address (src, BLKmode,
  91                                                    src_addr, copied),
  92                         bytes - copied, align, 0);
  93
  94       return true;
  95     }
  96
  97   /* If it isn't a constant number of bytes, or if it doesn't have 4 byte
  98      alignment, or if it isn't a multiple of 4 bytes, then fail.  */
  99   if (align < 4 || (bytes % 4 != 0))
 100     return false;
 101
 102   if (TARGET_HARD_SH4)
 103     {
 104       if (bytes < 12)
 105         return false;
 106       else if (bytes == 12)
 107         {
 108           rtx func_addr_rtx = gen_reg_rtx (Pmode);
 109           rtx r4 = gen_rtx_REG (SImode, 4);
 110           rtx r5 = gen_rtx_REG (SImode, 5);
 111
 112           rtx lab = function_symbol (func_addr_rtx, "__movmemSI12_i4",
 113                                      SFUNC_STATIC).lab;
 114           force_into (XEXP (operands[0], 0), r4);
 115           force_into (XEXP (operands[1], 0), r5);
 116           emit_insn (gen_block_move_real_i4 (func_addr_rtx, lab));
 117           return true;
 118         }
 119       else if (! optimize_size)
 120         {
 121           rtx func_addr_rtx = gen_reg_rtx (Pmode);
 122           rtx r4 = gen_rtx_REG (SImode, 4);
 123           rtx r5 = gen_rtx_REG (SImode, 5);
 124           rtx r6 = gen_rtx_REG (SImode, 6);
 125
 126           rtx lab = function_symbol (func_addr_rtx, bytes & 4
 127                                                     ? "__movmem_i4_odd"
 128                                                     : "__movmem_i4_even",
 129                                      SFUNC_STATIC).lab;
 130           force_into (XEXP (operands[0], 0), r4);
 131           force_into (XEXP (operands[1], 0), r5);
 132
 133           int dwords = bytes >> 3;
 134           emit_insn (gen_move_insn (r6, GEN_INT (dwords - 1)));
 135           emit_insn (gen_block_lump_real_i4 (func_addr_rtx, lab));
 136           return true;
 137         }
 138       else
 139         return false;
 140     }
 141   if (bytes < 64)
 142     {
 143       char entry[30];
 144       rtx func_addr_rtx = gen_reg_rtx (Pmode);
 145       rtx r4 = gen_rtx_REG (SImode, 4);
 146       rtx r5 = gen_rtx_REG (SImode, 5);
 147
 148       sprintf (entry, "__movmemSI%d", bytes);
 149       rtx lab = function_symbol (func_addr_rtx, entry, SFUNC_STATIC).lab;
 150       force_into (XEXP (operands[0], 0), r4);
 151       force_into (XEXP (operands[1], 0), r5);
 152       emit_insn (gen_block_move_real (func_addr_rtx, lab));
 153       return true;
 154     }
 155
 156   /* This is the same number of bytes as a memcpy call, but to a different
 157      less common function name, so this will occasionally use more space.  */
 158   if (! optimize_size)
 159     {
 160       rtx func_addr_rtx = gen_reg_rtx (Pmode);
 161       int final_switch, while_loop;
 162       rtx r4 = gen_rtx_REG (SImode, 4);
 163       rtx r5 = gen_rtx_REG (SImode, 5);
 164       rtx r6 = gen_rtx_REG (SImode, 6);
 165
 166       rtx lab = function_symbol (func_addr_rtx, "__movmem", SFUNC_STATIC).lab;
 167       force_into (XEXP (operands[0], 0), r4);
 168       force_into (XEXP (operands[1], 0), r5);
 169
 170       /* r6 controls the size of the move.  16 is decremented from it
 171          for each 64 bytes moved.  Then the negative bit left over is used
 172          as an index into a list of move instructions.  e.g., a 72 byte move
 173          would be set up with size(r6) = 14, for one iteration through the
 174          big while loop, and a switch of -2 for the last part.  */
 175
 176       final_switch = 16 - ((bytes / 4) % 16);
 177       while_loop = ((bytes / 4) / 16 - 1) * 16;
 178       emit_insn (gen_move_insn (r6, GEN_INT (while_loop + final_switch)));
 179       emit_insn (gen_block_lump_real (func_addr_rtx, lab));
 180       return true;
 181     }
 182
 183   return false;
 184 }
 185
 186 static const int prob_unlikely = REG_BR_PROB_BASE / 10;
 187 static const int prob_likely = REG_BR_PROB_BASE / 4;
 188
 189 /* Emit code to perform a strcmp.
 190
 191    OPERANDS[0] is the destination.
 192    OPERANDS[1] is the first string.
 193    OPERANDS[2] is the second string.
 194    OPERANDS[3] is the known alignment.  */
 195 bool
 196 sh_expand_cmpstr (rtx *operands)
 197 {
 198   rtx addr1 = operands[1];
 199   rtx addr2 = operands[2];
 200   rtx s1_addr = copy_addr_to_reg (XEXP (addr1, 0));
 201   rtx s2_addr = copy_addr_to_reg (XEXP (addr2, 0));
 202   rtx tmp0 = gen_reg_rtx (SImode);
 203   rtx tmp1 = gen_reg_rtx (SImode);
 204   rtx tmp2 = gen_reg_rtx (SImode);
 205   rtx tmp3 = gen_reg_rtx (SImode);
 206
 207   rtx jump;
 208   rtx_code_label *L_return = gen_label_rtx ();
 209   rtx_code_label *L_loop_byte = gen_label_rtx ();
 210   rtx_code_label *L_end_loop_byte = gen_label_rtx ();
 211   rtx_code_label *L_loop_long = gen_label_rtx ();
 212   rtx_code_label *L_end_loop_long = gen_label_rtx ();
 213
 214   const unsigned int addr1_alignment = MEM_ALIGN (operands[1]) / BITS_PER_UNIT;
 215   const unsigned int addr2_alignment = MEM_ALIGN (operands[2]) / BITS_PER_UNIT;
 216
 217   if (addr1_alignment < 4 && addr2_alignment < 4)
 218     {
 219       emit_insn (gen_iorsi3 (tmp1, s1_addr, s2_addr));
 220       emit_insn (gen_tstsi_t (tmp1, GEN_INT (3)));
 221       jump = emit_jump_insn (gen_branch_false (L_loop_byte));
 222       add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 223     }
 224   else if (addr1_alignment < 4 && addr2_alignment >= 4)
 225     {
 226       emit_insn (gen_tstsi_t (s1_addr, GEN_INT (3)));
 227       jump = emit_jump_insn (gen_branch_false (L_loop_byte));
 228       add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 229     }
 230   else if (addr1_alignment >= 4 && addr2_alignment < 4)
 231     {
 232       emit_insn (gen_tstsi_t (s2_addr, GEN_INT (3)));
 233       jump = emit_jump_insn (gen_branch_false (L_loop_byte));
 234       add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 235     }
 236
 237   addr1 = adjust_automodify_address (addr1, SImode, s1_addr, 0);
 238   addr2 = adjust_automodify_address (addr2, SImode, s2_addr, 0);
 239
 240   /* tmp2 is aligned, OK to load.  */
 241   emit_move_insn (tmp3, addr2);
 242   emit_move_insn (s2_addr, plus_constant (Pmode, s2_addr, 4));
 243
 244   /* start long loop.  */
 245   emit_label (L_loop_long);
 246
 247   emit_move_insn (tmp2, tmp3);
 248
 249   /* tmp1 is aligned, OK to load.  */
 250   emit_move_insn (tmp1, addr1);
 251   emit_move_insn (s1_addr, plus_constant (Pmode, s1_addr, 4));
 252
 253   /* Is there a 0 byte ?  */
 254   emit_insn (gen_andsi3 (tmp3, tmp3, tmp1));
 255
 256   emit_insn (gen_cmpstr_t (tmp0, tmp3));
 257   jump = emit_jump_insn (gen_branch_true (L_end_loop_long));
 258   add_int_reg_note (jump, REG_BR_PROB, prob_unlikely);
 259
 260   emit_insn (gen_cmpeqsi_t (tmp1, tmp2));
 261
 262   /* tmp2 is aligned, OK to load.  */
 263   emit_move_insn (tmp3, addr2);
 264   emit_move_insn (s2_addr, plus_constant (Pmode, s2_addr, 4));
 265
 266   jump = emit_jump_insn (gen_branch_true (L_loop_long));
 267   add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 268   /* end loop.  */
 269
 270   /* Fallthu, substract words.  */
 271   if (TARGET_LITTLE_ENDIAN)
 272     {
 273       rtx low_1 = gen_lowpart (HImode, tmp1);
 274       rtx low_2 = gen_lowpart (HImode, tmp2);
 275
 276       emit_insn (gen_rotlhi3_8 (low_1, low_1));
 277       emit_insn (gen_rotlhi3_8 (low_2, low_2));
 278       emit_insn (gen_rotlsi3_16 (tmp1, tmp1));
 279       emit_insn (gen_rotlsi3_16 (tmp2, tmp2));
 280       emit_insn (gen_rotlhi3_8 (low_1, low_1));
 281       emit_insn (gen_rotlhi3_8 (low_2, low_2));
 282     }
 283
 284   jump = emit_jump_insn (gen_jump_compact (L_return));
 285   emit_barrier_after (jump);
 286
 287   emit_label (L_end_loop_long);
 288
 289   emit_move_insn (s1_addr, plus_constant (Pmode, s1_addr, -4));
 290   emit_move_insn (s2_addr, plus_constant (Pmode, s2_addr, -4));
 291
 292   /* start byte loop.  */
 293   addr1 = adjust_address (addr1, QImode, 0);
 294   addr2 = adjust_address (addr2, QImode, 0);
 295
 296   emit_label (L_loop_byte);
 297
 298   emit_insn (gen_extendqisi2 (tmp2, addr2));
 299   emit_move_insn (s2_addr, plus_constant (Pmode, s2_addr, 1));
 300
 301   emit_insn (gen_extendqisi2 (tmp1, addr1));
 302   emit_move_insn (s1_addr, plus_constant (Pmode, s1_addr, 1));
 303
 304   emit_insn (gen_cmpeqsi_t (tmp2, const0_rtx));
 305   jump = emit_jump_insn (gen_branch_true (L_end_loop_byte));
 306   add_int_reg_note (jump, REG_BR_PROB, prob_unlikely);
 307
 308   emit_insn (gen_cmpeqsi_t (tmp1, tmp2));
 309   if (flag_delayed_branch)
 310     emit_insn (gen_zero_extendqisi2 (tmp2, gen_lowpart (QImode, tmp2)));
 311   jump = emit_jump_insn (gen_branch_true (L_loop_byte));
 312   add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 313   /* end loop.  */
 314
 315   emit_label (L_end_loop_byte);
 316
 317   if (! flag_delayed_branch)
 318     emit_insn (gen_zero_extendqisi2 (tmp2, gen_lowpart (QImode, tmp2)));
 319   emit_insn (gen_zero_extendqisi2 (tmp1, gen_lowpart (QImode, tmp1)));
 320
 321   emit_label (L_return);
 322
 323   emit_insn (gen_subsi3 (operands[0], tmp1, tmp2));
 324
 325   return true;
 326 }
 327
 328 /* Emit code to perform a strncmp.
 329
 330    OPERANDS[0] is the destination.
 331    OPERANDS[1] is the first string.
 332    OPERANDS[2] is the second string.
 333    OPERANDS[3] is the length.
 334    OPERANDS[4] is the known alignment.  */
 335 bool
 336 sh_expand_cmpnstr (rtx *operands)
 337 {
 338   rtx addr1 = operands[1];
 339   rtx addr2 = operands[2];
 340   rtx s1_addr = copy_addr_to_reg (XEXP (addr1, 0));
 341   rtx s2_addr = copy_addr_to_reg (XEXP (addr2, 0));
 342   rtx tmp1 = gen_reg_rtx (SImode);
 343   rtx tmp2 = gen_reg_rtx (SImode);
 344
 345   rtx jump;
 346   rtx_code_label *L_return = gen_label_rtx ();
 347   rtx_code_label *L_loop_byte = gen_label_rtx ();
 348   rtx_code_label *L_end_loop_byte = gen_label_rtx ();
 349
 350   rtx len = force_reg (SImode, operands[3]);
 351   int constp = CONST_INT_P (operands[3]);
 352
 353   const unsigned int addr1_alignment = MEM_ALIGN (operands[1]) / BITS_PER_UNIT;
 354   const unsigned int addr2_alignment = MEM_ALIGN (operands[2]) / BITS_PER_UNIT;
 355
 356   /* Loop on a register count.  */
 357   if (constp)
 358     {
 359       rtx tmp0 = gen_reg_rtx (SImode);
 360       rtx tmp3 = gen_reg_rtx (SImode);
 361       rtx lenw = gen_reg_rtx (SImode);
 362
 363       rtx_code_label *L_loop_long = gen_label_rtx ();
 364       rtx_code_label *L_end_loop_long = gen_label_rtx ();
 365
 366       int bytes = INTVAL (operands[3]);
 367       int witers = bytes / 4;
 368
 369       if (witers > 1)
 370         {
 371           addr1 = adjust_automodify_address (addr1, SImode, s1_addr, 0);
 372           addr2 = adjust_automodify_address (addr2, SImode, s2_addr, 0);
 373
 374           emit_move_insn (tmp0, const0_rtx);
 375
 376           if (addr1_alignment < 4 && addr2_alignment < 4)
 377             {
 378               emit_insn (gen_iorsi3 (tmp1, s1_addr, s2_addr));
 379               emit_insn (gen_tstsi_t (tmp1, GEN_INT (3)));
 380               jump = emit_jump_insn (gen_branch_false (L_loop_byte));
 381               add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 382             }
 383           else if (addr1_alignment < 4 && addr2_alignment >= 4)
 384             {
 385               emit_insn (gen_tstsi_t (s1_addr, GEN_INT (3)));
 386               jump = emit_jump_insn (gen_branch_false (L_loop_byte));
 387               add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 388             }
 389           else if (addr1_alignment >= 4 && addr2_alignment < 4)
 390             {
 391               emit_insn (gen_tstsi_t (s2_addr, GEN_INT (3)));
 392               jump = emit_jump_insn (gen_branch_false (L_loop_byte));
 393               add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 394             }
 395
 396           /* word count. Do we have iterations ?  */
 397           emit_insn (gen_lshrsi3 (lenw, len, GEN_INT (2)));
 398
 399           /* start long loop.  */
 400           emit_label (L_loop_long);
 401
 402           /* tmp2 is aligned, OK to load.  */
 403           emit_move_insn (tmp2, addr2);
 404           emit_move_insn (s2_addr, plus_constant (Pmode, s2_addr,
 405                                                   GET_MODE_SIZE (SImode)));
 406
 407           /* tmp1 is aligned, OK to load.  */
 408           emit_move_insn (tmp1, addr1);
 409           emit_move_insn (s1_addr, plus_constant (Pmode, s1_addr,
 410                                                   GET_MODE_SIZE (SImode)));
 411
 412           /* Is there a 0 byte ?  */
 413           emit_insn (gen_andsi3 (tmp3, tmp2, tmp1));
 414
 415           emit_insn (gen_cmpstr_t (tmp0, tmp3));
 416           jump = emit_jump_insn (gen_branch_true (L_end_loop_long));
 417           add_int_reg_note (jump, REG_BR_PROB, prob_unlikely);
 418
 419           emit_insn (gen_cmpeqsi_t (tmp1, tmp2));
 420           jump = emit_jump_insn (gen_branch_false (L_end_loop_long));
 421           add_int_reg_note (jump, REG_BR_PROB, prob_unlikely);
 422
 423           if (TARGET_SH2)
 424             emit_insn (gen_dect (lenw, lenw));
 425           else
 426             {
 427               emit_insn (gen_addsi3 (lenw, lenw, GEN_INT (-1)));
 428               emit_insn (gen_tstsi_t (lenw, lenw));
 429             }
 430
 431           jump = emit_jump_insn (gen_branch_false (L_loop_long));
 432           add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 433
 434           int sbytes = bytes % 4;
 435
 436           /* end loop.  Reached max iterations.  */
 437           if (sbytes == 0)
 438             {
 439               emit_insn (gen_subsi3 (operands[0], tmp1, tmp2));
 440               jump = emit_jump_insn (gen_jump_compact (L_return));
 441               emit_barrier_after (jump);
 442             }
 443           else
 444             {
 445               /* Remaining bytes to check.  */
 446
 447               addr1 = adjust_automodify_address (addr1, QImode, s1_addr, 0);
 448               addr2 = adjust_automodify_address (addr2, QImode, s2_addr, 0);
 449
 450               while (sbytes--)
 451                 {
 452                   emit_insn (gen_extendqisi2 (tmp1, addr1));
 453                   emit_insn (gen_extendqisi2 (tmp2, addr2));
 454
 455                   emit_insn (gen_cmpeqsi_t (tmp2, const0_rtx));
 456                   jump = emit_jump_insn (gen_branch_true (L_end_loop_byte));
 457                   add_int_reg_note (jump, REG_BR_PROB, prob_unlikely);
 458
 459                   emit_insn (gen_cmpeqsi_t (tmp1, tmp2));
 460                   if (flag_delayed_branch)
 461                     emit_insn (gen_zero_extendqisi2 (tmp2,
 462                                                      gen_lowpart (QImode,
 463                                                                   tmp2)));
 464                   jump = emit_jump_insn (gen_branch_false (L_end_loop_byte));
 465                   add_int_reg_note (jump, REG_BR_PROB, prob_unlikely);
 466
 467                   addr1 = adjust_address (addr1, QImode,
 468                                           GET_MODE_SIZE (QImode));
 469                   addr2 = adjust_address (addr2, QImode,
 470                                           GET_MODE_SIZE (QImode));
 471                 }
 472
 473               jump = emit_jump_insn (gen_jump_compact( L_end_loop_byte));
 474               emit_barrier_after (jump);
 475             }
 476
 477           emit_label (L_end_loop_long);
 478
 479           /* Found last word.  Restart it byte per byte.  */
 480
 481           emit_move_insn (s1_addr, plus_constant (Pmode, s1_addr,
 482                                                   -GET_MODE_SIZE (SImode)));
 483           emit_move_insn (s2_addr, plus_constant (Pmode, s2_addr,
 484                                                   -GET_MODE_SIZE (SImode)));
 485
 486           /* fall thru.  */
 487         }
 488
 489       addr1 = adjust_automodify_address (addr1, QImode, s1_addr, 0);
 490       addr2 = adjust_automodify_address (addr2, QImode, s2_addr, 0);
 491
 492       while (bytes--)
 493         {
 494           emit_insn (gen_extendqisi2 (tmp1, addr1));
 495           emit_insn (gen_extendqisi2 (tmp2, addr2));
 496
 497           emit_insn (gen_cmpeqsi_t (tmp2, const0_rtx));
 498           jump = emit_jump_insn (gen_branch_true (L_end_loop_byte));
 499           add_int_reg_note (jump, REG_BR_PROB, prob_unlikely);
 500
 501           emit_insn (gen_cmpeqsi_t (tmp1, tmp2));
 502           if (flag_delayed_branch)
 503             emit_insn (gen_zero_extendqisi2 (tmp2,
 504                                              gen_lowpart (QImode, tmp2)));
 505           jump = emit_jump_insn (gen_branch_false (L_end_loop_byte));
 506           add_int_reg_note (jump, REG_BR_PROB, prob_unlikely);
 507
 508           addr1 = adjust_address (addr1, QImode, GET_MODE_SIZE (QImode));
 509           addr2 = adjust_address (addr2, QImode, GET_MODE_SIZE (QImode));
 510         }
 511
 512       jump = emit_jump_insn (gen_jump_compact( L_end_loop_byte));
 513       emit_barrier_after (jump);
 514     }
 515   else
 516     {
 517       emit_insn (gen_cmpeqsi_t (len, const0_rtx));
 518       emit_move_insn (operands[0], const0_rtx);
 519       jump = emit_jump_insn (gen_branch_true (L_return));
 520       add_int_reg_note (jump, REG_BR_PROB, prob_unlikely);
 521     }
 522
 523   addr1 = adjust_automodify_address (addr1, QImode, s1_addr, 0);
 524   addr2 = adjust_automodify_address (addr2, QImode, s2_addr, 0);
 525
 526   emit_label (L_loop_byte);
 527
 528   emit_insn (gen_extendqisi2 (tmp2, addr2));
 529   emit_move_insn (s2_addr, plus_constant (Pmode, s2_addr, 1));
 530
 531   emit_insn (gen_extendqisi2 (tmp1, addr1));
 532   emit_move_insn (s1_addr, plus_constant (Pmode, s1_addr, 1));
 533
 534   emit_insn (gen_cmpeqsi_t (tmp2, const0_rtx));
 535   jump = emit_jump_insn (gen_branch_true (L_end_loop_byte));
 536   add_int_reg_note (jump, REG_BR_PROB, prob_unlikely);
 537
 538   emit_insn (gen_cmpeqsi_t (tmp1, tmp2));
 539   if (flag_delayed_branch)
 540     emit_insn (gen_zero_extendqisi2 (tmp2, gen_lowpart (QImode, tmp2)));
 541   jump = emit_jump_insn (gen_branch_false (L_end_loop_byte));
 542   add_int_reg_note (jump, REG_BR_PROB, prob_unlikely);
 543
 544   if (TARGET_SH2)
 545     emit_insn (gen_dect (len, len));
 546   else
 547     {
 548       emit_insn (gen_addsi3 (len, len, GEN_INT (-1)));
 549       emit_insn (gen_tstsi_t (len, len));
 550     }
 551
 552   jump = emit_jump_insn (gen_branch_false (L_loop_byte));
 553   add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 554   /* end byte loop.  */
 555
 556   emit_label (L_end_loop_byte);
 557
 558   if (! flag_delayed_branch)
 559     emit_insn (gen_zero_extendqisi2 (tmp2, gen_lowpart (QImode, tmp2)));
 560   emit_insn (gen_zero_extendqisi2 (tmp1, gen_lowpart (QImode, tmp1)));
 561
 562   emit_insn (gen_subsi3 (operands[0], tmp1, tmp2));
 563
 564   emit_label (L_return);
 565
 566   return true;
 567 }
 568
 569 /* Emit code to perform a strlen.
 570
 571    OPERANDS[0] is the destination.
 572    OPERANDS[1] is the string.
 573    OPERANDS[2] is the char to search.
 574    OPERANDS[3] is the alignment.  */
 575 bool
 576 sh_expand_strlen (rtx *operands)
 577 {
 578   rtx addr1 = operands[1];
 579   rtx current_addr = copy_addr_to_reg (XEXP (addr1, 0));
 580   rtx start_addr = gen_reg_rtx (Pmode);
 581   rtx tmp0 = gen_reg_rtx (SImode);
 582   rtx tmp1 = gen_reg_rtx (SImode);
 583   rtx_code_label *L_return = gen_label_rtx ();
 584   rtx_code_label *L_loop_byte = gen_label_rtx ();
 585
 586   rtx jump;
 587   rtx_code_label *L_loop_long = gen_label_rtx ();
 588   rtx_code_label *L_end_loop_long = gen_label_rtx ();
 589
 590   int align = INTVAL (operands[3]);
 591
 592   emit_move_insn (operands[0], GEN_INT (-1));
 593
 594   /* remember start of string.  */
 595   emit_move_insn (start_addr, current_addr);
 596
 597   if (align < 4)
 598     {
 599       emit_insn (gen_tstsi_t (current_addr, GEN_INT (3)));
 600       jump = emit_jump_insn (gen_branch_false (L_loop_byte));
 601       add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 602     }
 603
 604   emit_move_insn (tmp0, operands[2]);
 605
 606   addr1 = adjust_automodify_address (addr1, SImode, current_addr, 0);
 607
 608   /* start long loop.  */
 609   emit_label (L_loop_long);
 610
 611   /* tmp1 is aligned, OK to load.  */
 612   emit_move_insn (tmp1, addr1);
 613   emit_move_insn (current_addr, plus_constant (Pmode, current_addr, 4));
 614
 615   /* Is there a 0 byte ?  */
 616   emit_insn (gen_cmpstr_t (tmp0, tmp1));
 617
 618   jump = emit_jump_insn (gen_branch_false (L_loop_long));
 619   add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 620   /* end loop.  */
 621
 622   emit_label (L_end_loop_long);
 623
 624   emit_move_insn (current_addr, plus_constant (Pmode, current_addr, -4));
 625
 626   addr1 = adjust_address (addr1, QImode, 0);
 627
 628   /* unroll remaining bytes.  */
 629   for (int i = 0; i < 4; ++i)
 630     {
 631       emit_insn (gen_extendqisi2 (tmp1, addr1));
 632       emit_move_insn (current_addr, plus_constant (Pmode, current_addr, 1));
 633       emit_insn (gen_cmpeqsi_t (tmp1, const0_rtx));
 634       jump = emit_jump_insn (gen_branch_true (L_return));
 635       add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 636     }
 637
 638   emit_barrier_after (jump);
 639
 640   /* start byte loop.  */
 641   emit_label (L_loop_byte);
 642
 643   emit_insn (gen_extendqisi2 (tmp1, addr1));
 644   emit_move_insn (current_addr, plus_constant (Pmode, current_addr, 1));
 645
 646   emit_insn (gen_cmpeqsi_t (tmp1, const0_rtx));
 647   jump = emit_jump_insn (gen_branch_false (L_loop_byte));
 648   add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 649
 650   /* end loop.  */
 651
 652   emit_label (L_return);
 653
 654   emit_insn (gen_addsi3 (start_addr, start_addr, GEN_INT (1)));
 655   emit_insn (gen_subsi3 (operands[0], current_addr, start_addr));
 656
 657   return true;
 658 }
 659
 660 /* Emit code to perform a memset.
 661
 662    OPERANDS[0] is the destination.
 663    OPERANDS[1] is the size;
 664    OPERANDS[2] is the char to search.
 665    OPERANDS[3] is the alignment.  */
 666 void
 667 sh_expand_setmem (rtx *operands)
 668 {
 669   rtx_code_label *L_loop_byte = gen_label_rtx ();
 670   rtx_code_label *L_loop_word = gen_label_rtx ();
 671   rtx_code_label *L_return = gen_label_rtx ();
 672   rtx jump;
 673   rtx dest = copy_rtx (operands[0]);
 674   rtx dest_addr = copy_addr_to_reg (XEXP (dest, 0));
 675   rtx val = force_reg (SImode, operands[2]);
 676   int align = INTVAL (operands[3]);
 677   rtx len = force_reg (SImode, operands[1]);
 678
 679   if (! CONST_INT_P (operands[1]))
 680     return;
 681
 682   int count = INTVAL (operands[1]);
 683
 684   if (CONST_INT_P (operands[2])
 685       && (INTVAL (operands[2]) == 0 || INTVAL (operands[2]) == -1) && count > 8)
 686     {
 687       rtx lenw = gen_reg_rtx (SImode);
 688
 689       if (align < 4)
 690         {
 691           emit_insn (gen_tstsi_t (dest_addr, GEN_INT (3)));
 692           jump = emit_jump_insn (gen_branch_false (L_loop_byte));
 693           add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 694         }
 695
 696       /* word count. Do we have iterations ?  */
 697       emit_insn (gen_lshrsi3 (lenw, len, GEN_INT (2)));
 698
 699       dest = adjust_automodify_address (dest, SImode, dest_addr, 0);
 700
 701       /* start loop.  */
 702       emit_label (L_loop_word);
 703
 704       if (TARGET_SH2)
 705         emit_insn (gen_dect (lenw, lenw));
 706       else
 707         {
 708           emit_insn (gen_addsi3 (lenw, lenw, GEN_INT (-1)));
 709           emit_insn (gen_tstsi_t (lenw, lenw));
 710         }
 711
 712       emit_move_insn (dest, val);
 713       emit_move_insn (dest_addr, plus_constant (Pmode, dest_addr,
 714                                                 GET_MODE_SIZE (SImode)));
 715
 716
 717       jump = emit_jump_insn (gen_branch_false (L_loop_word));
 718       add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 719       count = count % 4;
 720
 721       dest = adjust_address (dest, QImode, 0);
 722
 723       val = gen_lowpart (QImode, val);
 724
 725       while (count--)
 726         {
 727           emit_move_insn (dest, val);
 728           emit_move_insn (dest_addr, plus_constant (Pmode, dest_addr,
 729                                                     GET_MODE_SIZE (QImode)));
 730         }
 731
 732       jump = emit_jump_insn (gen_jump_compact (L_return));
 733       emit_barrier_after (jump);
 734     }
 735
 736   dest = adjust_automodify_address (dest, QImode, dest_addr, 0);
 737
 738   /* start loop.  */
 739   emit_label (L_loop_byte);
 740
 741   if (TARGET_SH2)
 742     emit_insn (gen_dect (len, len));
 743   else
 744     {
 745       emit_insn (gen_addsi3 (len, len, GEN_INT (-1)));
 746       emit_insn (gen_tstsi_t (len, len));
 747     }
 748
 749   val = gen_lowpart (QImode, val);
 750   emit_move_insn (dest, val);
 751   emit_move_insn (dest_addr, plus_constant (Pmode, dest_addr,
 752                                             GET_MODE_SIZE (QImode)));
 753
 754   jump = emit_jump_insn (gen_branch_false (L_loop_byte));
 755   add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 756
 757   emit_label (L_return);
 758 }