libgcc/config/cris/mulsi3.S

   1 ;; Copyright (C) 2001-2023 Free Software Foundation, Inc.
   2 ;;
   3 ;; This file is part of GCC.
   4 ;;
   5 ;; GCC is free software; you can redistribute it and/or modify it under
   6 ;; the terms of the GNU General Public License as published by the Free
   7 ;; Software Foundation; either version 3, or (at your option) any later
   8 ;; version.
   9 ;;
  10 ;; GCC is distributed in the hope that it will be useful, but WITHOUT ANY
  11 ;; WARRANTY; without even the implied warranty of MERCHANTABILITY or
  12 ;; FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  13 ;; for more details.
  14 ;;
  15 ;; Under Section 7 of GPL version 3, you are granted additional
  16 ;; permissions described in the GCC Runtime Library Exception, version
  17 ;; 3.1, as published by the Free Software Foundation.
  18 ;;
  19 ;; You should have received a copy of the GNU General Public License and
  20 ;; a copy of the GCC Runtime Library Exception along with this program;
  21 ;; see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
  22 ;; <http://www.gnu.org/licenses/>.
  23 ;;
  24 ;; This code used to be expanded through interesting expansions in
  25 ;; the machine description, compiled from this code:
  26 ;;
  27 ;; #ifdef L_mulsi3
  28 ;; long __Mul (unsigned long a, unsigned long b) __attribute__ ((__const__));
  29 ;;
  30 ;; /* This must be compiled with the -mexpand-mul flag, to synthesize the
  31 ;;    multiplication from the mstep instructions.  The check for
  32 ;;    smaller-size multiplication pays off in the order of .5-10%;
  33 ;;    estimated median 1%, depending on application.
  34 ;;     FIXME: It can be further optimized if we go to assembler code, as
  35 ;;    gcc 2.7.2 adds a few unnecessary instructions and does not put the
  36 ;;    basic blocks in optimal order.  */
  37 ;; long
  38 ;; __Mul (unsigned long a, unsigned long b)
  39 ;; {
  40 ;; #if defined (__CRIS_arch_version) && __CRIS_arch_version >= 10
  41 ;;   /* In case other code is compiled without -march=v10, they will
  42 ;;      contain calls to __Mul, regardless of flags at link-time.  The
  43 ;;      "else"-code below will work, but is unnecessarily slow.  This
  44 ;;      sometimes cuts a few minutes off from simulation time by just
  45 ;;      returning a "mulu.d".  */
  46 ;;   return a * b;
  47 ;; #else
  48 ;;   unsigned long min;
  49 ;;
  50 ;;   /* Get minimum via the bound insn.  */
  51 ;;   min = a < b ? a : b;
  52 ;;
  53 ;;   /* Can we omit computation of the high part?       */
  54 ;;   if (min > 65535)
  55 ;;     /* No.  Perform full multiplication.  */
  56 ;;     return a * b;
  57 ;;   else
  58 ;;     {
  59 ;;       /* Check if both operands are within 16 bits.  */
  60 ;;       unsigned long max;
  61 ;;
  62 ;;       /* Get maximum, by knowing the minimum.
  63 ;;          This will partition a and b into max and min.
  64 ;;          This is not currently something GCC understands,
  65 ;;          so do this trick by asm.  */
  66 ;;       __asm__ ("xor %1,%0\n\txor %2,%0"
  67 ;;                : "=r" (max)
  68 ;;                :  "r" (b), "r" (a), "0" (min));
  69 ;;
  70 ;;     if (max > 65535)
  71 ;;       /* Make GCC understand that only the low part of "min" will be
  72 ;;          used.  */
  73 ;;       return max * (unsigned short) min;
  74 ;;     else
  75 ;;       /* Only the low parts of both operands are necessary.  */
  76 ;;       return ((unsigned short) max) * (unsigned short) min;
  77 ;;     }
  78 ;; #endif /* not __CRIS_arch_version >= 10 */
  79 ;; }
  80 ;; #endif /* L_mulsi3 */
  81 ;;
  82 ;; That approach was abandoned since the caveats outweighted the
  83 ;; benefits.  The expand-multiplication machinery is also removed, so you
  84 ;; can't do this anymore.
  85 ;;
  86 ;; For doubters of there being any benefits, some where: insensitivity to:
  87 ;; - ABI changes (mostly for experimentation)
  88 ;; - assembler syntax differences (mostly debug format).
  89 ;; - insn scheduling issues.
  90 ;; Most ABI experiments will presumably happen with arches with mul insns,
  91 ;; so that argument doesn't really hold anymore, and it's unlikely there
  92 ;; being new arch variants needing insn scheduling and not having mul
  93 ;; insns.
  94
  95 ;; ELF and a.out have different syntax for local labels: the "wrong"
  96 ;; one may not be omitted from the object.
  97 #undef L
  98 #ifdef __AOUT__
  99 # define L(x) x
 100 #else
 101 # define L(x) .x
 102 #endif
 103
 104         .global ___Mul
 105         .type   ___Mul,@function
 106 ___Mul:
 107 #if defined (__CRIS_arch_version) && __CRIS_arch_version >= 10
 108 ;; Can't have the mulu.d last on a cache-line (in the delay-slot of the
 109 ;; "ret"), due to hardware bug.  See documentation for -mmul-bug-workaround.
 110 ;; Not worthwhile to conditionalize here.
 111         .p2alignw 2,0x050f
 112         mulu.d $r11,$r10
 113         ret
 114         nop
 115 #else
 116 ;; See if we can avoid multiplying some of the parts, knowing
 117 ;; they're zero.
 118
 119         move.d $r11,$r9
 120         bound.d $r10,$r9
 121         cmpu.w 65535,$r9
 122         bls L(L3)
 123         move.d $r10,$r12
 124
 125 ;; Nope, have to do all the parts of a 32-bit multiplication.
 126 ;; See head comment in optabs.c:expand_doubleword_mult.
 127
 128         move.d $r10,$r13
 129         movu.w $r11,$r9 ; ab*cd = (a*d + b*c)<<16 + b*d
 130         lslq 16,$r13
 131         mstep $r9,$r13  ; d*b
 132         mstep $r9,$r13
 133         mstep $r9,$r13
 134         mstep $r9,$r13
 135         mstep $r9,$r13
 136         mstep $r9,$r13
 137         mstep $r9,$r13
 138         mstep $r9,$r13
 139         mstep $r9,$r13
 140         mstep $r9,$r13
 141         mstep $r9,$r13
 142         mstep $r9,$r13
 143         mstep $r9,$r13
 144         mstep $r9,$r13
 145         mstep $r9,$r13
 146         mstep $r9,$r13
 147         clear.w $r10
 148         test.d $r10
 149         mstep $r9,$r10  ; d*a
 150         mstep $r9,$r10
 151         mstep $r9,$r10
 152         mstep $r9,$r10
 153         mstep $r9,$r10
 154         mstep $r9,$r10
 155         mstep $r9,$r10
 156         mstep $r9,$r10
 157         mstep $r9,$r10
 158         mstep $r9,$r10
 159         mstep $r9,$r10
 160         mstep $r9,$r10
 161         mstep $r9,$r10
 162         mstep $r9,$r10
 163         mstep $r9,$r10
 164         mstep $r9,$r10
 165         movu.w $r12,$r12
 166         clear.w $r11
 167         move.d $r11,$r9 ; Doubles as a "test.d" preparing for the mstep.
 168         mstep $r12,$r9  ; b*c
 169         mstep $r12,$r9
 170         mstep $r12,$r9
 171         mstep $r12,$r9
 172         mstep $r12,$r9
 173         mstep $r12,$r9
 174         mstep $r12,$r9
 175         mstep $r12,$r9
 176         mstep $r12,$r9
 177         mstep $r12,$r9
 178         mstep $r12,$r9
 179         mstep $r12,$r9
 180         mstep $r12,$r9
 181         mstep $r12,$r9
 182         mstep $r12,$r9
 183         mstep $r12,$r9
 184         add.w $r9,$r10
 185         lslq 16,$r10
 186         ret
 187         add.d $r13,$r10
 188
 189 L(L3):
 190 ;; Form the maximum in $r10, by knowing the minimum, $r9.
 191 ;; (We don't know which one of $r10 or $r11 it is.)
 192 ;; Check if the largest operand is still just 16 bits.
 193
 194         xor $r9,$r10
 195         xor $r11,$r10
 196         cmpu.w 65535,$r10
 197         bls L(L5)
 198         movu.w $r9,$r13
 199
 200 ;; We have ab*cd = (a*c)<<32 + (a*d + b*c)<<16 + b*d, but c==0
 201 ;; so we only need (a*d)<<16 + b*d with d = $r13, ab = $r10.
 202 ;; We drop the upper part of (a*d)<<16 as we're only doing a
 203 ;; 32-bit-result multiplication.
 204
 205         move.d $r10,$r9
 206         lslq 16,$r9
 207         mstep $r13,$r9  ; b*d
 208         mstep $r13,$r9
 209         mstep $r13,$r9
 210         mstep $r13,$r9
 211         mstep $r13,$r9
 212         mstep $r13,$r9
 213         mstep $r13,$r9
 214         mstep $r13,$r9
 215         mstep $r13,$r9
 216         mstep $r13,$r9
 217         mstep $r13,$r9
 218         mstep $r13,$r9
 219         mstep $r13,$r9
 220         mstep $r13,$r9
 221         mstep $r13,$r9
 222         mstep $r13,$r9
 223         clear.w $r10
 224         test.d $r10
 225         mstep $r13,$r10 ; a*d
 226         mstep $r13,$r10
 227         mstep $r13,$r10
 228         mstep $r13,$r10
 229         mstep $r13,$r10
 230         mstep $r13,$r10
 231         mstep $r13,$r10
 232         mstep $r13,$r10
 233         mstep $r13,$r10
 234         mstep $r13,$r10
 235         mstep $r13,$r10
 236         mstep $r13,$r10
 237         mstep $r13,$r10
 238         mstep $r13,$r10
 239         mstep $r13,$r10
 240         mstep $r13,$r10
 241         lslq 16,$r10
 242         ret
 243         add.d $r9,$r10
 244
 245 L(L5):
 246 ;; We have ab*cd = (a*c)<<32 + (a*d + b*c)<<16 + b*d, but a and c==0
 247 ;; so b*d (with b=$r13, a=$r10) it is.
 248
 249         lslq 16,$r10
 250         mstep $r13,$r10
 251         mstep $r13,$r10
 252         mstep $r13,$r10
 253         mstep $r13,$r10
 254         mstep $r13,$r10
 255         mstep $r13,$r10
 256         mstep $r13,$r10
 257         mstep $r13,$r10
 258         mstep $r13,$r10
 259         mstep $r13,$r10
 260         mstep $r13,$r10
 261         mstep $r13,$r10
 262         mstep $r13,$r10
 263         mstep $r13,$r10
 264         mstep $r13,$r10
 265         ret
 266         mstep $r13,$r10
 267 #endif
 268 L(Lfe1):
 269         .size   ___Mul,L(Lfe1)-___Mul