gcc/config/sh/sh-mem.cc

   1 /* Helper routines for memory move and comparison insns.
   2    Copyright (C) 2013-2018 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 #define IN_TARGET_CODE 1
  21
  22 #include "config.h"
  23 #include "system.h"
  24 #include "coretypes.h"
  25 #include "tm.h"
  26 #include "function.h"
  27 #include "basic-block.h"
  28 #include "rtl.h"
  29 #include "tree.h"
  30 #include "memmodel.h"
  31 #include "tm_p.h"
  32 #include "emit-rtl.h"
  33 #include "explow.h"
  34 #include "expr.h"
  35
  36 /* Like force_operand, but guarantees that VALUE ends up in TARGET.  */
  37 static void
  38 force_into (rtx value, rtx target)
  39 {
  40   value = force_operand (value, target);
  41   if (! rtx_equal_p (value, target))
  42     emit_insn (gen_move_insn (target, value));
  43 }
  44
  45 /* Emit code to perform a block move.  Choose the best method.
  46
  47    OPERANDS[0] is the destination.
  48    OPERANDS[1] is the source.
  49    OPERANDS[2] is the size.
  50    OPERANDS[3] is the alignment safe to use.  */
  51 bool
  52 expand_block_move (rtx *operands)
  53 {
  54   int align = INTVAL (operands[3]);
  55   int constp = (CONST_INT_P (operands[2]));
  56   int bytes = (constp ? INTVAL (operands[2]) : 0);
  57
  58   if (! constp)
  59     return false;
  60
  61   /* If we could use mov.l to move words and dest is word-aligned, we
  62      can use movua.l for loads and still generate a relatively short
  63      and efficient sequence.  */
  64   if (TARGET_SH4A && align < 4
  65       && MEM_ALIGN (operands[0]) >= 32
  66       && can_move_by_pieces (bytes, 32))
  67     {
  68       rtx dest = copy_rtx (operands[0]);
  69       rtx src = copy_rtx (operands[1]);
  70       /* We could use different pseudos for each copied word, but
  71          since movua can only load into r0, it's kind of
  72          pointless.  */
  73       rtx temp = gen_reg_rtx (SImode);
  74       rtx src_addr = copy_addr_to_reg (XEXP (src, 0));
  75       int copied = 0;
  76
  77       while (copied + 4 <= bytes)
  78         {
  79           rtx to = adjust_address (dest, SImode, copied);
  80           rtx from = adjust_automodify_address (src, BLKmode,
  81                                                 src_addr, copied);
  82
  83           set_mem_size (from, 4);
  84           emit_insn (gen_movua (temp, from));
  85           emit_move_insn (src_addr, plus_constant (Pmode, src_addr, 4));
  86           emit_move_insn (to, temp);
  87           copied += 4;
  88         }
  89
  90       if (copied < bytes)
  91         move_by_pieces (adjust_address (dest, BLKmode, copied),
  92                         adjust_automodify_address (src, BLKmode,
  93                                                    src_addr, copied),
  94                         bytes - copied, align, 0);
  95
  96       return true;
  97     }
  98
  99   /* If it isn't a constant number of bytes, or if it doesn't have 4 byte
 100      alignment, or if it isn't a multiple of 4 bytes, then fail.  */
 101   if (align < 4 || (bytes % 4 != 0))
 102     return false;
 103
 104   if (TARGET_HARD_SH4)
 105     {
 106       if (bytes < 12)
 107         return false;
 108       else if (bytes == 12)
 109         {
 110           rtx func_addr_rtx = gen_reg_rtx (Pmode);
 111           rtx r4 = gen_rtx_REG (SImode, 4);
 112           rtx r5 = gen_rtx_REG (SImode, 5);
 113
 114           rtx lab = function_symbol (func_addr_rtx, "__movmemSI12_i4",
 115                                      SFUNC_STATIC).lab;
 116           force_into (XEXP (operands[0], 0), r4);
 117           force_into (XEXP (operands[1], 0), r5);
 118           emit_insn (gen_block_move_real_i4 (func_addr_rtx, lab));
 119           return true;
 120         }
 121       else if (! optimize_size)
 122         {
 123           rtx func_addr_rtx = gen_reg_rtx (Pmode);
 124           rtx r4 = gen_rtx_REG (SImode, 4);
 125           rtx r5 = gen_rtx_REG (SImode, 5);
 126           rtx r6 = gen_rtx_REG (SImode, 6);
 127
 128           rtx lab = function_symbol (func_addr_rtx, bytes & 4
 129                                                     ? "__movmem_i4_odd"
 130                                                     : "__movmem_i4_even",
 131                                      SFUNC_STATIC).lab;
 132           force_into (XEXP (operands[0], 0), r4);
 133           force_into (XEXP (operands[1], 0), r5);
 134
 135           int dwords = bytes >> 3;
 136           emit_insn (gen_move_insn (r6, GEN_INT (dwords - 1)));
 137           emit_insn (gen_block_lump_real_i4 (func_addr_rtx, lab));
 138           return true;
 139         }
 140       else
 141         return false;
 142     }
 143   if (bytes < 64)
 144     {
 145       char entry[30];
 146       rtx func_addr_rtx = gen_reg_rtx (Pmode);
 147       rtx r4 = gen_rtx_REG (SImode, 4);
 148       rtx r5 = gen_rtx_REG (SImode, 5);
 149
 150       sprintf (entry, "__movmemSI%d", bytes);
 151       rtx lab = function_symbol (func_addr_rtx, entry, SFUNC_STATIC).lab;
 152       force_into (XEXP (operands[0], 0), r4);
 153       force_into (XEXP (operands[1], 0), r5);
 154       emit_insn (gen_block_move_real (func_addr_rtx, lab));
 155       return true;
 156     }
 157
 158   /* This is the same number of bytes as a memcpy call, but to a different
 159      less common function name, so this will occasionally use more space.  */
 160   if (! optimize_size)
 161     {
 162       rtx func_addr_rtx = gen_reg_rtx (Pmode);
 163       int final_switch, while_loop;
 164       rtx r4 = gen_rtx_REG (SImode, 4);
 165       rtx r5 = gen_rtx_REG (SImode, 5);
 166       rtx r6 = gen_rtx_REG (SImode, 6);
 167
 168       rtx lab = function_symbol (func_addr_rtx, "__movmem", SFUNC_STATIC).lab;
 169       force_into (XEXP (operands[0], 0), r4);
 170       force_into (XEXP (operands[1], 0), r5);
 171
 172       /* r6 controls the size of the move.  16 is decremented from it
 173          for each 64 bytes moved.  Then the negative bit left over is used
 174          as an index into a list of move instructions.  e.g., a 72 byte move
 175          would be set up with size(r6) = 14, for one iteration through the
 176          big while loop, and a switch of -2 for the last part.  */
 177
 178       final_switch = 16 - ((bytes / 4) % 16);
 179       while_loop = ((bytes / 4) / 16 - 1) * 16;
 180       emit_insn (gen_move_insn (r6, GEN_INT (while_loop + final_switch)));
 181       emit_insn (gen_block_lump_real (func_addr_rtx, lab));
 182       return true;
 183     }
 184
 185   return false;
 186 }
 187
 188 static const int prob_unlikely
 189   = profile_probability::from_reg_br_prob_base (REG_BR_PROB_BASE / 10)
 190     .to_reg_br_prob_note ();
 191 static const int prob_likely
 192   = profile_probability::from_reg_br_prob_base (REG_BR_PROB_BASE / 4)
 193     .to_reg_br_prob_note ();
 194
 195 /* Emit code to perform a strcmp.
 196
 197    OPERANDS[0] is the destination.
 198    OPERANDS[1] is the first string.
 199    OPERANDS[2] is the second string.
 200    OPERANDS[3] is the known alignment.  */
 201 bool
 202 sh_expand_cmpstr (rtx *operands)
 203 {
 204   rtx addr1 = operands[1];
 205   rtx addr2 = operands[2];
 206   rtx s1_addr = copy_addr_to_reg (XEXP (addr1, 0));
 207   rtx s2_addr = copy_addr_to_reg (XEXP (addr2, 0));
 208   rtx tmp0 = gen_reg_rtx (SImode);
 209   rtx tmp1 = gen_reg_rtx (SImode);
 210   rtx tmp2 = gen_reg_rtx (SImode);
 211   rtx tmp3 = gen_reg_rtx (SImode);
 212
 213   rtx_insn *jump;
 214   rtx_code_label *L_return = gen_label_rtx ();
 215   rtx_code_label *L_loop_byte = gen_label_rtx ();
 216   rtx_code_label *L_end_loop_byte = gen_label_rtx ();
 217   rtx_code_label *L_loop_long = gen_label_rtx ();
 218   rtx_code_label *L_end_loop_long = gen_label_rtx ();
 219
 220   const unsigned int addr1_alignment = MEM_ALIGN (operands[1]) / BITS_PER_UNIT;
 221   const unsigned int addr2_alignment = MEM_ALIGN (operands[2]) / BITS_PER_UNIT;
 222
 223   if (addr1_alignment < 4 && addr2_alignment < 4)
 224     {
 225       emit_insn (gen_iorsi3 (tmp1, s1_addr, s2_addr));
 226       emit_insn (gen_tstsi_t (tmp1, 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 (s1_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   else if (addr1_alignment >= 4 && addr2_alignment < 4)
 237     {
 238       emit_insn (gen_tstsi_t (s2_addr, GEN_INT (3)));
 239       jump = emit_jump_insn (gen_branch_false (L_loop_byte));
 240       add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 241     }
 242
 243   addr1 = adjust_automodify_address (addr1, SImode, s1_addr, 0);
 244   addr2 = adjust_automodify_address (addr2, SImode, s2_addr, 0);
 245
 246   /* tmp2 is aligned, OK to load.  */
 247   emit_move_insn (tmp3, addr2);
 248   emit_move_insn (s2_addr, plus_constant (Pmode, s2_addr, 4));
 249
 250   /* start long loop.  */
 251   emit_label (L_loop_long);
 252
 253   emit_move_insn (tmp2, tmp3);
 254
 255   /* tmp1 is aligned, OK to load.  */
 256   emit_move_insn (tmp1, addr1);
 257   emit_move_insn (s1_addr, plus_constant (Pmode, s1_addr, 4));
 258
 259   /* Is there a 0 byte ?  */
 260   emit_insn (gen_andsi3 (tmp3, tmp3, tmp1));
 261
 262   emit_insn (gen_cmpstr_t (tmp0, tmp3));
 263   jump = emit_jump_insn (gen_branch_true (L_end_loop_long));
 264   add_int_reg_note (jump, REG_BR_PROB, prob_unlikely);
 265
 266   emit_insn (gen_cmpeqsi_t (tmp1, tmp2));
 267
 268   /* tmp2 is aligned, OK to load.  */
 269   emit_move_insn (tmp3, addr2);
 270   emit_move_insn (s2_addr, plus_constant (Pmode, s2_addr, 4));
 271
 272   jump = emit_jump_insn (gen_branch_true (L_loop_long));
 273   add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 274   /* end loop.  */
 275
 276   /* Fallthu, substract words.  */
 277   if (TARGET_LITTLE_ENDIAN)
 278     {
 279       rtx low_1 = gen_lowpart (HImode, tmp1);
 280       rtx low_2 = gen_lowpart (HImode, tmp2);
 281
 282       emit_insn (gen_rotlhi3_8 (low_1, low_1));
 283       emit_insn (gen_rotlhi3_8 (low_2, low_2));
 284       emit_insn (gen_rotlsi3_16 (tmp1, tmp1));
 285       emit_insn (gen_rotlsi3_16 (tmp2, tmp2));
 286       emit_insn (gen_rotlhi3_8 (low_1, low_1));
 287       emit_insn (gen_rotlhi3_8 (low_2, low_2));
 288     }
 289
 290   jump = emit_jump_insn (gen_jump_compact (L_return));
 291   emit_barrier_after (jump);
 292
 293   emit_label (L_end_loop_long);
 294
 295   emit_move_insn (s1_addr, plus_constant (Pmode, s1_addr, -4));
 296   emit_move_insn (s2_addr, plus_constant (Pmode, s2_addr, -4));
 297
 298   /* start byte loop.  */
 299   addr1 = adjust_address (addr1, QImode, 0);
 300   addr2 = adjust_address (addr2, QImode, 0);
 301
 302   emit_label (L_loop_byte);
 303
 304   emit_insn (gen_extendqisi2 (tmp2, addr2));
 305   emit_move_insn (s2_addr, plus_constant (Pmode, s2_addr, 1));
 306
 307   emit_insn (gen_extendqisi2 (tmp1, addr1));
 308   emit_move_insn (s1_addr, plus_constant (Pmode, s1_addr, 1));
 309
 310   emit_insn (gen_cmpeqsi_t (tmp2, const0_rtx));
 311   jump = emit_jump_insn (gen_branch_true (L_end_loop_byte));
 312   add_int_reg_note (jump, REG_BR_PROB, prob_unlikely);
 313
 314   emit_insn (gen_cmpeqsi_t (tmp1, tmp2));
 315   if (flag_delayed_branch)
 316     emit_insn (gen_zero_extendqisi2 (tmp2, gen_lowpart (QImode, tmp2)));
 317   jump = emit_jump_insn (gen_branch_true (L_loop_byte));
 318   add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 319   /* end loop.  */
 320
 321   emit_label (L_end_loop_byte);
 322
 323   if (! flag_delayed_branch)
 324     emit_insn (gen_zero_extendqisi2 (tmp2, gen_lowpart (QImode, tmp2)));
 325   emit_insn (gen_zero_extendqisi2 (tmp1, gen_lowpart (QImode, tmp1)));
 326
 327   emit_label (L_return);
 328
 329   emit_insn (gen_subsi3 (operands[0], tmp1, tmp2));
 330
 331   return true;
 332 }
 333
 334 /* Emit code to perform a strncmp.
 335
 336    OPERANDS[0] is the destination.
 337    OPERANDS[1] is the first string.
 338    OPERANDS[2] is the second string.
 339    OPERANDS[3] is the length.
 340    OPERANDS[4] is the known alignment.  */
 341 bool
 342 sh_expand_cmpnstr (rtx *operands)
 343 {
 344   rtx addr1 = operands[1];
 345   rtx addr2 = operands[2];
 346   rtx s1_addr = copy_addr_to_reg (XEXP (addr1, 0));
 347   rtx s2_addr = copy_addr_to_reg (XEXP (addr2, 0));
 348   rtx tmp1 = gen_reg_rtx (SImode);
 349   rtx tmp2 = gen_reg_rtx (SImode);
 350
 351   rtx_insn *jump;
 352   rtx_code_label *L_return = gen_label_rtx ();
 353   rtx_code_label *L_loop_byte = gen_label_rtx ();
 354   rtx_code_label *L_end_loop_byte = gen_label_rtx ();
 355
 356   rtx len = copy_to_mode_reg (SImode, operands[3]);
 357   int constp = CONST_INT_P (operands[3]);
 358   HOST_WIDE_INT bytes = constp ? INTVAL (operands[3]) : 0;
 359
 360   const unsigned int addr1_alignment = MEM_ALIGN (operands[1]) / BITS_PER_UNIT;
 361   const unsigned int addr2_alignment = MEM_ALIGN (operands[2]) / BITS_PER_UNIT;
 362
 363   /* Loop on a register count.  */
 364   if (constp && bytes >= 0 && bytes < 32)
 365     {
 366       rtx tmp0 = gen_reg_rtx (SImode);
 367       rtx tmp3 = gen_reg_rtx (SImode);
 368       rtx lenw = gen_reg_rtx (SImode);
 369
 370       rtx_code_label *L_loop_long = gen_label_rtx ();
 371       rtx_code_label *L_end_loop_long = gen_label_rtx ();
 372
 373       int witers = bytes / 4;
 374
 375       if (witers > 1)
 376         {
 377           addr1 = adjust_automodify_address (addr1, SImode, s1_addr, 0);
 378           addr2 = adjust_automodify_address (addr2, SImode, s2_addr, 0);
 379
 380           emit_move_insn (tmp0, const0_rtx);
 381
 382           if (addr1_alignment < 4 && addr2_alignment < 4)
 383             {
 384               emit_insn (gen_iorsi3 (tmp1, s1_addr, s2_addr));
 385               emit_insn (gen_tstsi_t (tmp1, 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 (s1_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           else if (addr1_alignment >= 4 && addr2_alignment < 4)
 396             {
 397               emit_insn (gen_tstsi_t (s2_addr, GEN_INT (3)));
 398               jump = emit_jump_insn (gen_branch_false (L_loop_byte));
 399               add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 400             }
 401
 402           /* word count. Do we have iterations ?  */
 403           emit_insn (gen_lshrsi3 (lenw, len, GEN_INT (2)));
 404
 405           /* start long loop.  */
 406           emit_label (L_loop_long);
 407
 408           /* tmp2 is aligned, OK to load.  */
 409           emit_move_insn (tmp2, addr2);
 410           emit_move_insn (s2_addr, plus_constant (Pmode, s2_addr,
 411                                                   GET_MODE_SIZE (SImode)));
 412
 413           /* tmp1 is aligned, OK to load.  */
 414           emit_move_insn (tmp1, addr1);
 415           emit_move_insn (s1_addr, plus_constant (Pmode, s1_addr,
 416                                                   GET_MODE_SIZE (SImode)));
 417
 418           /* Is there a 0 byte ?  */
 419           emit_insn (gen_andsi3 (tmp3, tmp2, tmp1));
 420
 421           emit_insn (gen_cmpstr_t (tmp0, tmp3));
 422           jump = emit_jump_insn (gen_branch_true (L_end_loop_long));
 423           add_int_reg_note (jump, REG_BR_PROB, prob_unlikely);
 424
 425           emit_insn (gen_cmpeqsi_t (tmp1, tmp2));
 426           jump = emit_jump_insn (gen_branch_false (L_end_loop_long));
 427           add_int_reg_note (jump, REG_BR_PROB, prob_unlikely);
 428
 429           if (TARGET_SH2)
 430             emit_insn (gen_dect (lenw, lenw));
 431           else
 432             {
 433               emit_insn (gen_addsi3 (lenw, lenw, GEN_INT (-1)));
 434               emit_insn (gen_tstsi_t (lenw, lenw));
 435             }
 436
 437           jump = emit_jump_insn (gen_branch_false (L_loop_long));
 438           add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 439
 440           int sbytes = bytes % 4;
 441
 442           /* end loop.  Reached max iterations.  */
 443           if (sbytes == 0)
 444             {
 445               emit_insn (gen_subsi3 (operands[0], tmp1, tmp2));
 446               jump = emit_jump_insn (gen_jump_compact (L_return));
 447               emit_barrier_after (jump);
 448             }
 449           else
 450             {
 451               /* Remaining bytes to check.  */
 452
 453               addr1 = adjust_automodify_address (addr1, QImode, s1_addr, 0);
 454               addr2 = adjust_automodify_address (addr2, QImode, s2_addr, 0);
 455
 456               while (sbytes--)
 457                 {
 458                   emit_insn (gen_extendqisi2 (tmp1, addr1));
 459                   emit_insn (gen_extendqisi2 (tmp2, addr2));
 460
 461                   emit_insn (gen_cmpeqsi_t (tmp2, const0_rtx));
 462                   jump = emit_jump_insn (gen_branch_true (L_end_loop_byte));
 463                   add_int_reg_note (jump, REG_BR_PROB, prob_unlikely);
 464
 465                   emit_insn (gen_cmpeqsi_t (tmp1, tmp2));
 466                   if (flag_delayed_branch)
 467                     emit_insn (gen_zero_extendqisi2 (tmp2,
 468                                                      gen_lowpart (QImode,
 469                                                                   tmp2)));
 470                   jump = emit_jump_insn (gen_branch_false (L_end_loop_byte));
 471                   add_int_reg_note (jump, REG_BR_PROB, prob_unlikely);
 472
 473                   addr1 = adjust_address (addr1, QImode,
 474                                           GET_MODE_SIZE (QImode));
 475                   addr2 = adjust_address (addr2, QImode,
 476                                           GET_MODE_SIZE (QImode));
 477                 }
 478
 479               jump = emit_jump_insn (gen_jump_compact( L_end_loop_byte));
 480               emit_barrier_after (jump);
 481             }
 482
 483           emit_label (L_end_loop_long);
 484
 485           /* Found last word.  Restart it byte per byte.  */
 486
 487           emit_move_insn (s1_addr, plus_constant (Pmode, s1_addr,
 488                                                   -GET_MODE_SIZE (SImode)));
 489           emit_move_insn (s2_addr, plus_constant (Pmode, s2_addr,
 490                                                   -GET_MODE_SIZE (SImode)));
 491
 492           /* fall thru.  */
 493         }
 494
 495       addr1 = adjust_automodify_address (addr1, QImode, s1_addr, 0);
 496       addr2 = adjust_automodify_address (addr2, QImode, s2_addr, 0);
 497
 498       while (bytes--)
 499         {
 500           emit_insn (gen_extendqisi2 (tmp1, addr1));
 501           emit_insn (gen_extendqisi2 (tmp2, addr2));
 502
 503           emit_insn (gen_cmpeqsi_t (tmp2, const0_rtx));
 504           jump = emit_jump_insn (gen_branch_true (L_end_loop_byte));
 505           add_int_reg_note (jump, REG_BR_PROB, prob_unlikely);
 506
 507           emit_insn (gen_cmpeqsi_t (tmp1, tmp2));
 508           if (flag_delayed_branch)
 509             emit_insn (gen_zero_extendqisi2 (tmp2,
 510                                              gen_lowpart (QImode, tmp2)));
 511           jump = emit_jump_insn (gen_branch_false (L_end_loop_byte));
 512           add_int_reg_note (jump, REG_BR_PROB, prob_unlikely);
 513
 514           addr1 = adjust_address (addr1, QImode, GET_MODE_SIZE (QImode));
 515           addr2 = adjust_address (addr2, QImode, GET_MODE_SIZE (QImode));
 516         }
 517
 518       jump = emit_jump_insn (gen_jump_compact( L_end_loop_byte));
 519       emit_barrier_after (jump);
 520     }
 521   else
 522     {
 523       emit_insn (gen_cmpeqsi_t (len, const0_rtx));
 524       emit_move_insn (operands[0], const0_rtx);
 525       jump = emit_jump_insn (gen_branch_true (L_return));
 526       add_int_reg_note (jump, REG_BR_PROB, prob_unlikely);
 527     }
 528
 529   addr1 = adjust_automodify_address (addr1, QImode, s1_addr, 0);
 530   addr2 = adjust_automodify_address (addr2, QImode, s2_addr, 0);
 531
 532   emit_label (L_loop_byte);
 533
 534   emit_insn (gen_extendqisi2 (tmp2, addr2));
 535   emit_move_insn (s2_addr, plus_constant (Pmode, s2_addr, 1));
 536
 537   emit_insn (gen_extendqisi2 (tmp1, addr1));
 538   emit_move_insn (s1_addr, plus_constant (Pmode, s1_addr, 1));
 539
 540   emit_insn (gen_cmpeqsi_t (tmp2, const0_rtx));
 541   jump = emit_jump_insn (gen_branch_true (L_end_loop_byte));
 542   add_int_reg_note (jump, REG_BR_PROB, prob_unlikely);
 543
 544   emit_insn (gen_cmpeqsi_t (tmp1, tmp2));
 545   if (flag_delayed_branch)
 546     emit_insn (gen_zero_extendqisi2 (tmp2, gen_lowpart (QImode, tmp2)));
 547   jump = emit_jump_insn (gen_branch_false (L_end_loop_byte));
 548   add_int_reg_note (jump, REG_BR_PROB, prob_unlikely);
 549
 550   if (TARGET_SH2)
 551     emit_insn (gen_dect (len, len));
 552   else
 553     {
 554       emit_insn (gen_addsi3 (len, len, GEN_INT (-1)));
 555       emit_insn (gen_tstsi_t (len, len));
 556     }
 557
 558   jump = emit_jump_insn (gen_branch_false (L_loop_byte));
 559   add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 560   /* end byte loop.  */
 561
 562   emit_label (L_end_loop_byte);
 563
 564   if (! flag_delayed_branch)
 565     emit_insn (gen_zero_extendqisi2 (tmp2, gen_lowpart (QImode, tmp2)));
 566   emit_insn (gen_zero_extendqisi2 (tmp1, gen_lowpart (QImode, tmp1)));
 567
 568   emit_insn (gen_subsi3 (operands[0], tmp1, tmp2));
 569
 570   emit_label (L_return);
 571
 572   return true;
 573 }
 574
 575 /* Emit code to perform a strlen.
 576
 577    OPERANDS[0] is the destination.
 578    OPERANDS[1] is the string.
 579    OPERANDS[2] is the char to search.
 580    OPERANDS[3] is the alignment.  */
 581 bool
 582 sh_expand_strlen (rtx *operands)
 583 {
 584   rtx addr1 = operands[1];
 585   rtx current_addr = copy_addr_to_reg (XEXP (addr1, 0));
 586   rtx start_addr = gen_reg_rtx (Pmode);
 587   rtx tmp0 = gen_reg_rtx (SImode);
 588   rtx tmp1 = gen_reg_rtx (SImode);
 589   rtx_code_label *L_return = gen_label_rtx ();
 590   rtx_code_label *L_loop_byte = gen_label_rtx ();
 591
 592   rtx_insn *jump;
 593   rtx_code_label *L_loop_long = gen_label_rtx ();
 594   rtx_code_label *L_end_loop_long = gen_label_rtx ();
 595
 596   int align = INTVAL (operands[3]);
 597
 598   emit_move_insn (operands[0], GEN_INT (-1));
 599
 600   /* remember start of string.  */
 601   emit_move_insn (start_addr, current_addr);
 602
 603   if (align < 4)
 604     {
 605       emit_insn (gen_tstsi_t (current_addr, GEN_INT (3)));
 606       jump = emit_jump_insn (gen_branch_false (L_loop_byte));
 607       add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 608     }
 609
 610   emit_move_insn (tmp0, operands[2]);
 611
 612   addr1 = adjust_automodify_address (addr1, SImode, current_addr, 0);
 613
 614   /* start long loop.  */
 615   emit_label (L_loop_long);
 616
 617   /* tmp1 is aligned, OK to load.  */
 618   emit_move_insn (tmp1, addr1);
 619   emit_move_insn (current_addr, plus_constant (Pmode, current_addr, 4));
 620
 621   /* Is there a 0 byte ?  */
 622   emit_insn (gen_cmpstr_t (tmp0, tmp1));
 623
 624   jump = emit_jump_insn (gen_branch_false (L_loop_long));
 625   add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 626   /* end loop.  */
 627
 628   emit_label (L_end_loop_long);
 629
 630   emit_move_insn (current_addr, plus_constant (Pmode, current_addr, -4));
 631
 632   addr1 = adjust_address (addr1, QImode, 0);
 633
 634   /* unroll remaining bytes.  */
 635   for (int i = 0; i < 4; ++i)
 636     {
 637       emit_insn (gen_extendqisi2 (tmp1, addr1));
 638       emit_move_insn (current_addr, plus_constant (Pmode, current_addr, 1));
 639       emit_insn (gen_cmpeqsi_t (tmp1, const0_rtx));
 640       jump = emit_jump_insn (gen_branch_true (L_return));
 641       add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 642     }
 643
 644   emit_barrier_after (jump);
 645
 646   /* start byte loop.  */
 647   emit_label (L_loop_byte);
 648
 649   emit_insn (gen_extendqisi2 (tmp1, addr1));
 650   emit_move_insn (current_addr, plus_constant (Pmode, current_addr, 1));
 651
 652   emit_insn (gen_cmpeqsi_t (tmp1, const0_rtx));
 653   jump = emit_jump_insn (gen_branch_false (L_loop_byte));
 654   add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 655
 656   /* end loop.  */
 657
 658   emit_label (L_return);
 659
 660   emit_insn (gen_addsi3 (start_addr, start_addr, GEN_INT (1)));
 661   emit_insn (gen_subsi3 (operands[0], current_addr, start_addr));
 662
 663   return true;
 664 }
 665
 666 /* Emit code to perform a memset.
 667
 668    OPERANDS[0] is the destination.
 669    OPERANDS[1] is the size;
 670    OPERANDS[2] is the char to search.
 671    OPERANDS[3] is the alignment.  */
 672 void
 673 sh_expand_setmem (rtx *operands)
 674 {
 675   rtx_code_label *L_loop_byte = gen_label_rtx ();
 676   rtx_code_label *L_loop_word = gen_label_rtx ();
 677   rtx_code_label *L_return = gen_label_rtx ();
 678   rtx_insn *jump;
 679   rtx dest = copy_rtx (operands[0]);
 680   rtx dest_addr = copy_addr_to_reg (XEXP (dest, 0));
 681   rtx val = copy_to_mode_reg (SImode, operands[2]);
 682   int align = INTVAL (operands[3]);
 683   rtx len = copy_to_mode_reg (SImode, operands[1]);
 684
 685   if (! CONST_INT_P (operands[1]))
 686     return;
 687
 688   int count = INTVAL (operands[1]);
 689
 690   if (CONST_INT_P (operands[2])
 691       && (INTVAL (operands[2]) == 0 || INTVAL (operands[2]) == -1) && count > 8)
 692     {
 693       rtx lenw = gen_reg_rtx (SImode);
 694
 695       if (align < 4)
 696         {
 697           emit_insn (gen_tstsi_t (dest_addr, GEN_INT (3)));
 698           jump = emit_jump_insn (gen_branch_false (L_loop_byte));
 699           add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 700         }
 701
 702       /* word count. Do we have iterations ?  */
 703       emit_insn (gen_lshrsi3 (lenw, len, GEN_INT (2)));
 704
 705       dest = adjust_automodify_address (dest, SImode, dest_addr, 0);
 706
 707       /* start loop.  */
 708       emit_label (L_loop_word);
 709
 710       if (TARGET_SH2)
 711         emit_insn (gen_dect (lenw, lenw));
 712       else
 713         {
 714           emit_insn (gen_addsi3 (lenw, lenw, GEN_INT (-1)));
 715           emit_insn (gen_tstsi_t (lenw, lenw));
 716         }
 717
 718       emit_move_insn (dest, val);
 719       emit_move_insn (dest_addr, plus_constant (Pmode, dest_addr,
 720                                                 GET_MODE_SIZE (SImode)));
 721
 722
 723       jump = emit_jump_insn (gen_branch_false (L_loop_word));
 724       add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 725       count = count % 4;
 726
 727       dest = adjust_address (dest, QImode, 0);
 728
 729       val = gen_lowpart (QImode, val);
 730
 731       while (count--)
 732         {
 733           emit_move_insn (dest, val);
 734           emit_move_insn (dest_addr, plus_constant (Pmode, dest_addr,
 735                                                     GET_MODE_SIZE (QImode)));
 736         }
 737
 738       jump = emit_jump_insn (gen_jump_compact (L_return));
 739       emit_barrier_after (jump);
 740     }
 741
 742   dest = adjust_automodify_address (dest, QImode, dest_addr, 0);
 743
 744   /* start loop.  */
 745   emit_label (L_loop_byte);
 746
 747   if (TARGET_SH2)
 748     emit_insn (gen_dect (len, len));
 749   else
 750     {
 751       emit_insn (gen_addsi3 (len, len, GEN_INT (-1)));
 752       emit_insn (gen_tstsi_t (len, len));
 753     }
 754
 755   val = gen_lowpart (QImode, val);
 756   emit_move_insn (dest, val);
 757   emit_move_insn (dest_addr, plus_constant (Pmode, dest_addr,
 758                                             GET_MODE_SIZE (QImode)));
 759
 760   jump = emit_jump_insn (gen_branch_false (L_loop_byte));
 761   add_int_reg_note (jump, REG_BR_PROB, prob_likely);
 762
 763   emit_label (L_return);
 764 }