t_vb_arbprogram_sse.c

winNT技术操作系统,国外开放的原代码和LIUX一样

/*
 * Mesa 3-D graphics library
 * Version:  6.3
 *
 * Copyright (C) 1999-2004  Brian Paul   All Rights Reserved.
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
 * copy of this software and associated documentation files (the "Software"),
 * to deal in the Software without restriction, including without limitation
 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
 * and/or sell copies of the Software, and to permit persons to whom the
 * Software is furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice shall be included
 * in all copies or substantial portions of the Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
 * BRIAN PAUL BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN
 * AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 */

/**
 * \file t_vb_arb_program_sse.c
 *
 * Translate simplified vertex_program representation to
 * x86/x87/SSE/SSE2 machine code using mesa's rtasm runtime assembler.
 *
 * This is very much a first attempt - build something that works.
 * There are probably better approaches for applying SSE to vertex
 * programs, and the whole thing is crying out for static analysis of
 * the programs to avoid redundant operations.
 *
 * \author Keith Whitwell
 */

#include "glheader.h"
#include "context.h"
#include "imports.h"
#include "macros.h"
#include "mtypes.h"
#include "arbprogparse.h"
#include "program.h"
#include "math/m_matrix.h"
#include "math/m_translate.h"
#include "t_context.h"
#include "t_vb_arbprogram.h"

#if defined(USE_SSE_ASM)

#include "x86/rtasm/x86sse.h"
#include "x86/common_x86_asm.h"

#define X    0
#define Y    1
#define Z    2
#define W    3

/* Reg usage:
 *
 * EAX - temp
 * EBX - point to 'm->File[0]'
 * ECX - point to 'm->File[3]'
 * EDX - holds 'm'
 * EBP,
 * ESI,
 * EDI
 */

#define DISASSEM 0

#define FAIL								\
do {									\
   _mesa_printf("x86 translation failed in %s\n", __FUNCTION__);	\
   return GL_FALSE;							\
} while (0)

struct compilation {
   struct x86_function func;
   struct tnl_compiled_program *p;   
   GLuint insn_counter;

   struct {
      GLuint file:2;
      GLuint idx:7;
      GLuint dirty:1;
      GLuint last_used:10;
   } xmm[8];

   struct {
      struct x86_reg base;
   } file[4];

   GLboolean have_sse2;
   GLshort fpucntl;
};

static INLINE GLboolean eq( struct x86_reg a,
			    struct x86_reg b )
{
   return (a.file == b.file &&
	   a.idx == b.idx &&
	   a.mod == b.mod &&
	   a.disp == b.disp);
}
      
static GLint get_offset( const void *a, const void *b )
{
   return (const char *)b - (const char *)a;
}


static struct x86_reg get_reg_ptr(GLuint file,
				  GLuint idx )
{
   struct x86_reg reg;

   switch (file) {
   case FILE_REG:
      reg = x86_make_reg(file_REG32, reg_BX);
      assert(idx != REG_UNDEF);
      break;
   case FILE_STATE_PARAM:
      reg = x86_make_reg(file_REG32, reg_CX);
      break;
   default:
      assert(0);
   }

   return x86_make_disp(reg, 16 * idx);
}
			  

static void spill( struct compilation *cp, GLuint idx )
{
   struct x86_reg oldval = get_reg_ptr(cp->xmm[idx].file,
				       cp->xmm[idx].idx);

   assert(cp->xmm[idx].dirty);
   sse_movups(&cp->func, oldval, x86_make_reg(file_XMM, idx));
   cp->xmm[idx].dirty = 0;
}

static struct x86_reg get_xmm_reg( struct compilation *cp )
{
   GLuint i;
   GLuint oldest = 0;

   for (i = 0; i < 8; i++) 
      if (cp->xmm[i].last_used < cp->xmm[oldest].last_used)
	 oldest = i;

   /* Need to write out the old value?
    */
   if (cp->xmm[oldest].dirty) 
      spill(cp, oldest);

   assert(cp->xmm[oldest].last_used != cp->insn_counter);

   cp->xmm[oldest].file = FILE_REG;
   cp->xmm[oldest].idx = REG_UNDEF;
   cp->xmm[oldest].last_used = cp->insn_counter;
   return x86_make_reg(file_XMM, oldest);
}

static void invalidate_xmm( struct compilation *cp, 
			    GLuint file, GLuint idx )
{
   GLuint i;

   /* Invalidate any old copy of this register in XMM0-7.  
    */
   for (i = 0; i < 8; i++) {
      if (cp->xmm[i].file == file && cp->xmm[i].idx == idx) {
	 cp->xmm[i].file = FILE_REG;
	 cp->xmm[i].idx = REG_UNDEF;
	 cp->xmm[i].dirty = 0;
	 break;
      }
   }
}
      

/* Return an XMM reg to receive the results of an operation.
 */
static struct x86_reg get_dst_xmm_reg( struct compilation *cp, 
				       GLuint file, GLuint idx )
{
   struct x86_reg reg;

   /* Invalidate any old copy of this register in XMM0-7.  Don't reuse
    * as this may be one of the arguments.
    */
   invalidate_xmm( cp, file, idx );

   reg = get_xmm_reg( cp );
   cp->xmm[reg.idx].file = file;
   cp->xmm[reg.idx].idx = idx;
   cp->xmm[reg.idx].dirty = 1;
   return reg;   
}

/* As above, but return a pointer.  Note - this pointer may alias
 * those returned by get_arg_ptr().
 */
static struct x86_reg get_dst_ptr( struct compilation *cp, 
				   GLuint file, GLuint idx )
{
   /* Invalidate any old copy of this register in XMM0-7.  Don't reuse
    * as this may be one of the arguments.
    */
   invalidate_xmm( cp, file, idx );

   return get_reg_ptr(file, idx);
}



/* Return an XMM reg if the argument is resident, otherwise return a
 * base+offset pointer to the saved value.
 */
static struct x86_reg get_arg( struct compilation *cp, GLuint file, GLuint idx )
{
   GLuint i;

   for (i = 0; i < 8; i++) {
      if (cp->xmm[i].file == file &&
	  cp->xmm[i].idx == idx) {
	 cp->xmm[i].last_used = cp->insn_counter;
	 return x86_make_reg(file_XMM, i);
      }
   }

   return get_reg_ptr(file, idx);
}

/* As above, but always return a pointer:
 */
static struct x86_reg get_arg_ptr( struct compilation *cp, GLuint file, GLuint idx )
{
   GLuint i;

   /* If there is a modified version of this register in one of the
    * XMM regs, write it out to memory.
    */
   for (i = 0; i < 8; i++) {
      if (cp->xmm[i].file == file && 
	  cp->xmm[i].idx == idx &&
	  cp->xmm[i].dirty) 
	 spill(cp, i);
   }

   return get_reg_ptr(file, idx);
}

/* Emulate pshufd insn in regular SSE, if necessary:
 */
static void emit_pshufd( struct compilation *cp,
			 struct x86_reg dst,
			 struct x86_reg arg0,
			 GLubyte shuf )
{
   if (cp->have_sse2) {
      sse2_pshufd(&cp->func, dst, arg0, shuf);
      cp->func.fn = 0;
   }
   else {
      if (!eq(dst, arg0)) 
	 sse_movups(&cp->func, dst, arg0);

      sse_shufps(&cp->func, dst, dst, shuf);
   }
}

static void set_fpu_round_neg_inf( struct compilation *cp )
{
   if (cp->fpucntl != RND_NEG_FPU) {
      struct x86_reg regEDX = x86_make_reg(file_REG32, reg_DX);
      struct arb_vp_machine *m = NULL;

      cp->fpucntl = RND_NEG_FPU;
      x87_fnclex(&cp->func);
      x87_fldcw(&cp->func, x86_make_disp(regEDX, get_offset(m, &m->fpucntl_rnd_neg)));
   }
}


/* Perform a reduced swizzle.  
 */
static GLboolean emit_RSW( struct compilation *cp, union instruction op ) 
{
   struct x86_reg arg0 = get_arg(cp, op.rsw.file0, op.rsw.idx0);
   struct x86_reg dst = get_dst_xmm_reg(cp, FILE_REG, op.rsw.dst);
   GLuint swz = op.rsw.swz;
   GLuint neg = op.rsw.neg;

   emit_pshufd(cp, dst, arg0, swz);
   
   if (neg) {
      struct x86_reg negs = get_arg(cp, FILE_REG, REG_SWZ);
      struct x86_reg tmp = get_xmm_reg(cp);
      /* Load 1,-1,0,0
       * Use neg as arg to pshufd
       * Multiply
       */
      emit_pshufd(cp, tmp, negs, 
		  SHUF((neg & 1) ? 1 : 0,
		       (neg & 2) ? 1 : 0,
		       (neg & 4) ? 1 : 0,
		       (neg & 8) ? 1 : 0));
      sse_mulps(&cp->func, dst, tmp);
   }

   return GL_TRUE;
}

/* Helper for writemask:
 */
static GLboolean emit_shuf_copy1( struct compilation *cp,
				  struct x86_reg dst,
				  struct x86_reg arg0,
				  struct x86_reg arg1,
				  GLubyte shuf )
{
   struct x86_reg tmp = get_xmm_reg(cp);
   sse_movups(&cp->func, dst, arg1);
   emit_pshufd(cp, dst, dst, shuf);
   emit_pshufd(cp, tmp, arg0, shuf);

   sse_movss(&cp->func, dst, tmp);

   emit_pshufd(cp, dst, dst, shuf);
   return GL_TRUE;
}


/* Helper for writemask:
 */
static GLboolean emit_shuf_copy2( struct compilation *cp,
				  struct x86_reg dst,
				  struct x86_reg arg0,
				  struct x86_reg arg1,
				  GLubyte shuf )
{
   struct x86_reg tmp = get_xmm_reg(cp);
   emit_pshufd(cp, dst, arg1, shuf);
   emit_pshufd(cp, tmp, arg0, shuf);

   sse_shufps(&cp->func, dst, tmp, SHUF(X, Y, Z, W));

   emit_pshufd(cp, dst, dst, shuf);
   return GL_TRUE;
}


static void emit_x87_ex2( struct compilation *cp )
{
   struct x86_reg st0 = x86_make_reg(file_x87, 0);
   struct x86_reg st1 = x86_make_reg(file_x87, 1);
   struct x86_reg st3 = x86_make_reg(file_x87, 3);

   set_fpu_round_neg_inf( cp );

   x87_fld(&cp->func, st0); /* a a */
   x87_fprndint( &cp->func );	/* int(a) a */
   x87_fld(&cp->func, st0); /* int(a) int(a) a */
   x87_fstp(&cp->func, st3); /* int(a) a int(a)*/
   x87_fsubp(&cp->func, st1); /* frac(a) int(a) */
   x87_f2xm1(&cp->func);    /* (2^frac(a))-1 int(a)*/
   x87_fld1(&cp->func);    /* 1 (2^frac(a))-1 int(a)*/
   x87_faddp(&cp->func, st1);	/* 2^frac(a) int(a) */
   x87_fscale(&cp->func);	/* 2^a */
}

#if 0
static GLboolean emit_MSK2( struct compilation *cp, union instruction op )
{
   struct x86_reg arg0 = get_arg(cp, op.msk.file, op.msk.arg);
   struct x86_reg arg1 = get_arg(cp, FILE_REG, op.msk.dst); /* NOTE! */
   struct x86_reg dst = get_dst_xmm_reg(cp, FILE_REG, op.msk.dst);
   
   /* make full width bitmask in tmp 
    * dst = ~tmp
    * tmp &= arg0
    * dst &= arg1
    * dst |= tmp
    */
   emit_pshufd(cp, tmp, get_arg(cp, FILE_REG, REG_NEGS), 
	       SHUF((op.msk.mask & 1) ? 2 : 0,
		    (op.msk.mask & 2) ? 2 : 0,
		    (op.msk.mask & 4) ? 2 : 0,
		    (op.msk.mask & 8) ? 2 : 0));
   sse2_pnot(&cp->func, dst, tmp);
   sse2_pand(&cp->func, arg0, tmp);
   sse2_pand(&cp->func, arg1, dst);
   sse2_por(&cp->func, tmp, dst);
   return GL_TRUE;
}
#endif


/* Used to implement write masking.  This and most of the other instructions
 * here would be easier to implement if there had been a translation
 * to a 2 argument format (dst/arg0, arg1) at the shader level before
 * attempting to translate to x86/sse code.
 */