URL
https://opencores.org/ocsvn/openrisc_me/openrisc_me/trunk
Subversion Repositories openrisc_me
[/] [openrisc/] [trunk/] [gnu-src/] [gcc-4.5.1/] [gcc/] [config/] [cris/] [mulsi3.asm] - Rev 282
Compare with Previous | Blame | View Log
;; Copyright (C) 2001, 2004 Free Software Foundation, Inc.;;;; This file is part of GCC.;;;; GCC is free software; you can redistribute it and/or modify it under;; the terms of the GNU General Public License as published by the Free;; Software Foundation; either version 3, or (at your option) any later;; version.;;;; GCC is distributed in the hope that it will be useful, but WITHOUT ANY;; WARRANTY; without even the implied warranty of MERCHANTABILITY or;; FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License;; for more details.;;;; Under Section 7 of GPL version 3, you are granted additional;; permissions described in the GCC Runtime Library Exception, version;; 3.1, as published by the Free Software Foundation.;;;; You should have received a copy of the GNU General Public License and;; a copy of the GCC Runtime Library Exception along with this program;;; see the files COPYING3 and COPYING.RUNTIME respectively. If not, see;; <http://www.gnu.org/licenses/>.;;;; This code used to be expanded through interesting expansions in;; the machine description, compiled from this code:;;;; #ifdef L_mulsi3;; long __Mul (unsigned long a, unsigned long b) __attribute__ ((__const__));;;;; /* This must be compiled with the -mexpand-mul flag, to synthesize the;; multiplication from the mstep instructions. The check for;; smaller-size multiplication pays off in the order of .5-10%;;; estimated median 1%, depending on application.;; FIXME: It can be further optimized if we go to assembler code, as;; gcc 2.7.2 adds a few unnecessary instructions and does not put the;; basic blocks in optimal order. */;; long;; __Mul (unsigned long a, unsigned long b);; {;; #if defined (__CRIS_arch_version) && __CRIS_arch_version >= 10;; /* In case other code is compiled without -march=v10, they will;; contain calls to __Mul, regardless of flags at link-time. The;; "else"-code below will work, but is unnecessarily slow. This;; sometimes cuts a few minutes off from simulation time by just;; returning a "mulu.d". */;; return a * b;;; #else;; unsigned long min;;;;; /* Get minimum via the bound insn. */;; min = a < b ? a : b;;;;; /* Can we omit computation of the high part? */;; if (min > 65535);; /* No. Perform full multiplication. */;; return a * b;;; else;; {;; /* Check if both operands are within 16 bits. */;; unsigned long max;;;;; /* Get maximum, by knowing the minimum.;; This will partition a and b into max and min.;; This is not currently something GCC understands,;; so do this trick by asm. */;; __asm__ ("xor %1,%0\n\txor %2,%0";; : "=r" (max);; : "r" (b), "r" (a), "0" (min));;;;; if (max > 65535);; /* Make GCC understand that only the low part of "min" will be;; used. */;; return max * (unsigned short) min;;; else;; /* Only the low parts of both operands are necessary. */;; return ((unsigned short) max) * (unsigned short) min;;; };; #endif /* not __CRIS_arch_version >= 10 */;; };; #endif /* L_mulsi3 */;;;; That approach was abandoned since the caveats outweighted the;; benefits. The expand-multiplication machinery is also removed, so you;; can't do this anymore.;;;; For doubters of there being any benefits, some where: insensitivity to:;; - ABI changes (mostly for experimentation);; - assembler syntax differences (mostly debug format).;; - insn scheduling issues.;; Most ABI experiments will presumably happen with arches with mul insns,;; so that argument doesn't really hold anymore, and it's unlikely there;; being new arch variants needing insn scheduling and not having mul;; insns.;; ELF and a.out have different syntax for local labels: the "wrong";; one may not be omitted from the object.#undef L#ifdef __AOUT__# define L(x) x#else# define L(x) .x#endif.global ___Mul.type ___Mul,@function___Mul:#if defined (__CRIS_arch_version) && __CRIS_arch_version >= 10;; Can't have the mulu.d last on a cache-line (in the delay-slot of the;; "ret"), due to hardware bug. See documentation for -mmul-bug-workaround.;; Not worthwhile to conditionalize here..p2alignw 2,0x050fmulu.d $r11,$r10retnop#elsemove.d $r10,$r12move.d $r11,$r9bound.d $r12,$r9cmpu.w 65535,$r9bls L(L3)move.d $r12,$r13movu.w $r11,$r9lslq 16,$r13mstep $r9,$r13mstep $r9,$r13mstep $r9,$r13mstep $r9,$r13mstep $r9,$r13mstep $r9,$r13mstep $r9,$r13mstep $r9,$r13mstep $r9,$r13mstep $r9,$r13mstep $r9,$r13mstep $r9,$r13mstep $r9,$r13mstep $r9,$r13mstep $r9,$r13mstep $r9,$r13clear.w $r10test.d $r10mstep $r9,$r10mstep $r9,$r10mstep $r9,$r10mstep $r9,$r10mstep $r9,$r10mstep $r9,$r10mstep $r9,$r10mstep $r9,$r10mstep $r9,$r10mstep $r9,$r10mstep $r9,$r10mstep $r9,$r10mstep $r9,$r10mstep $r9,$r10mstep $r9,$r10mstep $r9,$r10movu.w $r12,$r12move.d $r11,$r9clear.w $r9test.d $r9mstep $r12,$r9mstep $r12,$r9mstep $r12,$r9mstep $r12,$r9mstep $r12,$r9mstep $r12,$r9mstep $r12,$r9mstep $r12,$r9mstep $r12,$r9mstep $r12,$r9mstep $r12,$r9mstep $r12,$r9mstep $r12,$r9mstep $r12,$r9mstep $r12,$r9mstep $r12,$r9add.w $r9,$r10lslq 16,$r10retadd.d $r13,$r10L(L3):move.d $r9,$r10xor $r11,$r10xor $r12,$r10cmpu.w 65535,$r10bls L(L5)movu.w $r9,$r13movu.w $r13,$r13move.d $r10,$r9lslq 16,$r9mstep $r13,$r9mstep $r13,$r9mstep $r13,$r9mstep $r13,$r9mstep $r13,$r9mstep $r13,$r9mstep $r13,$r9mstep $r13,$r9mstep $r13,$r9mstep $r13,$r9mstep $r13,$r9mstep $r13,$r9mstep $r13,$r9mstep $r13,$r9mstep $r13,$r9mstep $r13,$r9clear.w $r10test.d $r10mstep $r13,$r10mstep $r13,$r10mstep $r13,$r10mstep $r13,$r10mstep $r13,$r10mstep $r13,$r10mstep $r13,$r10mstep $r13,$r10mstep $r13,$r10mstep $r13,$r10mstep $r13,$r10mstep $r13,$r10mstep $r13,$r10mstep $r13,$r10mstep $r13,$r10mstep $r13,$r10lslq 16,$r10retadd.d $r9,$r10L(L5):movu.w $r9,$r9lslq 16,$r10mstep $r9,$r10mstep $r9,$r10mstep $r9,$r10mstep $r9,$r10mstep $r9,$r10mstep $r9,$r10mstep $r9,$r10mstep $r9,$r10mstep $r9,$r10mstep $r9,$r10mstep $r9,$r10mstep $r9,$r10mstep $r9,$r10mstep $r9,$r10mstep $r9,$r10retmstep $r9,$r10#endifL(Lfe1):.size ___Mul,L(Lfe1)-___Mul