x86sse.c

Mesa is an open-source implementation of the OpenGL specification - a system for rendering interacti

}

void sse_xorps( struct x86_function *p,
                struct x86_reg dst,
                struct x86_reg src )
{
   emit_2ub(p, X86_TWOB, 0x57);
   emit_modrm( p, dst, src );
}

void sse_cvtps2pi( struct x86_function *p,
		   struct x86_reg dst,
		   struct x86_reg src )
{
   assert(dst.file == file_MMX && 
	  (src.file == file_XMM || src.mod != mod_REG));

   p->need_emms = 1;

   emit_2ub(p, X86_TWOB, 0x2d);
   emit_modrm( p, dst, src );
}


/* Shufps can also be used to implement a reduced swizzle when dest ==
 * arg0.
 */
void sse_shufps( struct x86_function *p,
		 struct x86_reg dest,
		 struct x86_reg arg0,
		 unsigned char shuf) 
{
   emit_2ub(p, X86_TWOB, 0xC6);
   emit_modrm(p, dest, arg0);
   emit_1ub(p, shuf); 
}

void sse_cmpps( struct x86_function *p,
		struct x86_reg dest,
		struct x86_reg arg0,
		unsigned char cc) 
{
   emit_2ub(p, X86_TWOB, 0xC2);
   emit_modrm(p, dest, arg0);
   emit_1ub(p, cc); 
}

void sse_pmovmskb( struct x86_function *p,
                   struct x86_reg dest,
                   struct x86_reg src)
{
    emit_3ub(p, 0x66, X86_TWOB, 0xD7);
    emit_modrm(p, dest, src);
}

/***********************************************************************
 * SSE2 instructions
 */

/**
 * Perform a reduced swizzle:
 */
void sse2_pshufd( struct x86_function *p,
		  struct x86_reg dest,
		  struct x86_reg arg0,
		  unsigned char shuf) 
{
   emit_3ub(p, 0x66, X86_TWOB, 0x70);
   emit_modrm(p, dest, arg0);
   emit_1ub(p, shuf); 
}

void sse2_cvttps2dq( struct x86_function *p,
                     struct x86_reg dst,
                     struct x86_reg src )
{
   emit_3ub( p, 0xF3, X86_TWOB, 0x5B );
   emit_modrm( p, dst, src );
}

void sse2_cvtps2dq( struct x86_function *p,
		    struct x86_reg dst,
		    struct x86_reg src )
{
   emit_3ub(p, 0x66, X86_TWOB, 0x5B);
   emit_modrm( p, dst, src );
}

void sse2_packssdw( struct x86_function *p,
		    struct x86_reg dst,
		    struct x86_reg src )
{
   emit_3ub(p, 0x66, X86_TWOB, 0x6B);
   emit_modrm( p, dst, src );
}

void sse2_packsswb( struct x86_function *p,
		    struct x86_reg dst,
		    struct x86_reg src )
{
   emit_3ub(p, 0x66, X86_TWOB, 0x63);
   emit_modrm( p, dst, src );
}

void sse2_packuswb( struct x86_function *p,
		    struct x86_reg dst,
		    struct x86_reg src )
{
   emit_3ub(p, 0x66, X86_TWOB, 0x67);
   emit_modrm( p, dst, src );
}

void sse2_rcpps( struct x86_function *p,
                 struct x86_reg dst,
                 struct x86_reg src )
{
   emit_2ub(p, X86_TWOB, 0x53);
   emit_modrm( p, dst, src );
}

void sse2_rcpss( struct x86_function *p,
		struct x86_reg dst,
		struct x86_reg src )
{
   emit_3ub(p, 0xF3, X86_TWOB, 0x53);
   emit_modrm( p, dst, src );
}

void sse2_movd( struct x86_function *p,
		struct x86_reg dst,
		struct x86_reg src )
{
   emit_2ub(p, 0x66, X86_TWOB);
   emit_op_modrm( p, 0x6e, 0x7e, dst, src );
}




/***********************************************************************
 * x87 instructions
 */
void x87_fist( struct x86_function *p, struct x86_reg dst )
{
   emit_1ub(p, 0xdb);
   emit_modrm_noreg(p, 2, dst);
}

void x87_fistp( struct x86_function *p, struct x86_reg dst )
{
   emit_1ub(p, 0xdb);
   emit_modrm_noreg(p, 3, dst);
}

void x87_fild( struct x86_function *p, struct x86_reg arg )
{
   emit_1ub(p, 0xdf);
   emit_modrm_noreg(p, 0, arg);
}

void x87_fldz( struct x86_function *p )
{
   emit_2ub(p, 0xd9, 0xee);
}


void x87_fldcw( struct x86_function *p, struct x86_reg arg )
{
   assert(arg.file == file_REG32);
   assert(arg.mod != mod_REG);
   emit_1ub(p, 0xd9);
   emit_modrm_noreg(p, 5, arg);
}

void x87_fld1( struct x86_function *p )
{
   emit_2ub(p, 0xd9, 0xe8);
}

void x87_fldl2e( struct x86_function *p )
{
   emit_2ub(p, 0xd9, 0xea);
}

void x87_fldln2( struct x86_function *p )
{
   emit_2ub(p, 0xd9, 0xed);
}

void x87_fwait( struct x86_function *p )
{
   emit_1ub(p, 0x9b);
}

void x87_fnclex( struct x86_function *p )
{
   emit_2ub(p, 0xdb, 0xe2);
}

void x87_fclex( struct x86_function *p )
{
   x87_fwait(p);
   x87_fnclex(p);
}


static void x87_arith_op( struct x86_function *p, struct x86_reg dst, struct x86_reg arg,
			  unsigned char dst0ub0,
			  unsigned char dst0ub1,
			  unsigned char arg0ub0,
			  unsigned char arg0ub1,
			  unsigned char argmem_noreg)
{
   assert(dst.file == file_x87);

   if (arg.file == file_x87) {
      if (dst.idx == 0) 
	 emit_2ub(p, dst0ub0, dst0ub1+arg.idx);
      else if (arg.idx == 0) 
	 emit_2ub(p, arg0ub0, arg0ub1+arg.idx);
      else
	 assert(0);
   }
   else if (dst.idx == 0) {
      assert(arg.file == file_REG32);
      emit_1ub(p, 0xd8);
      emit_modrm_noreg(p, argmem_noreg, arg);
   }
   else
      assert(0);
}

void x87_fmul( struct x86_function *p, struct x86_reg dst, struct x86_reg arg )
{
   x87_arith_op(p, dst, arg, 
		0xd8, 0xc8,
		0xdc, 0xc8,
		4);
}

void x87_fsub( struct x86_function *p, struct x86_reg dst, struct x86_reg arg )
{
   x87_arith_op(p, dst, arg, 
		0xd8, 0xe0,
		0xdc, 0xe8,
		4);
}

void x87_fsubr( struct x86_function *p, struct x86_reg dst, struct x86_reg arg )
{
   x87_arith_op(p, dst, arg, 
		0xd8, 0xe8,
		0xdc, 0xe0,
		5);
}

void x87_fadd( struct x86_function *p, struct x86_reg dst, struct x86_reg arg )
{
   x87_arith_op(p, dst, arg, 
		0xd8, 0xc0,
		0xdc, 0xc0,
		0);
}

void x87_fdiv( struct x86_function *p, struct x86_reg dst, struct x86_reg arg )
{
   x87_arith_op(p, dst, arg, 
		0xd8, 0xf0,
		0xdc, 0xf8,
		6);
}

void x87_fdivr( struct x86_function *p, struct x86_reg dst, struct x86_reg arg )
{
   x87_arith_op(p, dst, arg, 
		0xd8, 0xf8,
		0xdc, 0xf0,
		7);
}

void x87_fmulp( struct x86_function *p, struct x86_reg dst )
{
   assert(dst.file == file_x87);
   assert(dst.idx >= 1);
   emit_2ub(p, 0xde, 0xc8+dst.idx);
}

void x87_fsubp( struct x86_function *p, struct x86_reg dst )
{
   assert(dst.file == file_x87);
   assert(dst.idx >= 1);
   emit_2ub(p, 0xde, 0xe8+dst.idx);
}

void x87_fsubrp( struct x86_function *p, struct x86_reg dst )
{
   assert(dst.file == file_x87);
   assert(dst.idx >= 1);
   emit_2ub(p, 0xde, 0xe0+dst.idx);
}

void x87_faddp( struct x86_function *p, struct x86_reg dst )
{
   assert(dst.file == file_x87);
   assert(dst.idx >= 1);
   emit_2ub(p, 0xde, 0xc0+dst.idx);
}

void x87_fdivp( struct x86_function *p, struct x86_reg dst )
{
   assert(dst.file == file_x87);
   assert(dst.idx >= 1);
   emit_2ub(p, 0xde, 0xf8+dst.idx);
}

void x87_fdivrp( struct x86_function *p, struct x86_reg dst )
{
   assert(dst.file == file_x87);
   assert(dst.idx >= 1);
   emit_2ub(p, 0xde, 0xf0+dst.idx);
}

void x87_fucom( struct x86_function *p, struct x86_reg arg )
{
   assert(arg.file == file_x87);
   emit_2ub(p, 0xdd, 0xe0+arg.idx);
}

void x87_fucomp( struct x86_function *p, struct x86_reg arg )
{
   assert(arg.file == file_x87);
   emit_2ub(p, 0xdd, 0xe8+arg.idx);
}

void x87_fucompp( struct x86_function *p )
{
   emit_2ub(p, 0xda, 0xe9);
}

void x87_fxch( struct x86_function *p, struct x86_reg arg )
{
   assert(arg.file == file_x87);
   emit_2ub(p, 0xd9, 0xc8+arg.idx);
}

void x87_fabs( struct x86_function *p )
{
   emit_2ub(p, 0xd9, 0xe1);
}

void x87_fchs( struct x86_function *p )
{
   emit_2ub(p, 0xd9, 0xe0);
}

void x87_fcos( struct x86_function *p )
{
   emit_2ub(p, 0xd9, 0xff);
}


void x87_fprndint( struct x86_function *p )
{
   emit_2ub(p, 0xd9, 0xfc);
}

void x87_fscale( struct x86_function *p )
{
   emit_2ub(p, 0xd9, 0xfd);
}

void x87_fsin( struct x86_function *p )
{
   emit_2ub(p, 0xd9, 0xfe);
}

void x87_fsincos( struct x86_function *p )
{
   emit_2ub(p, 0xd9, 0xfb);
}

void x87_fsqrt( struct x86_function *p )
{
   emit_2ub(p, 0xd9, 0xfa);
}

void x87_fxtract( struct x86_function *p )
{
   emit_2ub(p, 0xd9, 0xf4);
}

/* st0 = (2^st0)-1
 *
 * Restrictions: -1.0 <= st0 <= 1.0
 */
void x87_f2xm1( struct x86_function *p )
{
   emit_2ub(p, 0xd9, 0xf0);
}

/* st1 = st1 * log2(st0);
 * pop_stack;
 */
void x87_fyl2x( struct x86_function *p )
{
   emit_2ub(p, 0xd9, 0xf1);
}

/* st1 = st1 * log2(st0 + 1.0);
 * pop_stack;
 *
 * A fast operation, with restrictions: -.29 < st0 < .29 
 */
void x87_fyl2xp1( struct x86_function *p )
{
   emit_2ub(p, 0xd9, 0xf9);
}


void x87_fld( struct x86_function *p, struct x86_reg arg )
{
   if (arg.file == file_x87) 
      emit_2ub(p, 0xd9, 0xc0 + arg.idx);
   else {
      emit_1ub(p, 0xd9);
      emit_modrm_noreg(p, 0, arg);
   }
}

void x87_fst( struct x86_function *p, struct x86_reg dst )
{
   if (dst.file == file_x87) 
      emit_2ub(p, 0xdd, 0xd0 + dst.idx);
   else {
      emit_1ub(p, 0xd9);
      emit_modrm_noreg(p, 2, dst);
   }
}

void x87_fstp( struct x86_function *p, struct x86_reg dst )
{
   if (dst.file == file_x87) 
      emit_2ub(p, 0xdd, 0xd8 + dst.idx);
   else {
      emit_1ub(p, 0xd9);
      emit_modrm_noreg(p, 3, dst);
   }
}

void x87_fcom( struct x86_function *p, struct x86_reg dst )
{
   if (dst.file == file_x87) 
      emit_2ub(p, 0xd8, 0xd0 + dst.idx);
   else {
      emit_1ub(p, 0xd8);
      emit_modrm_noreg(p, 2, dst);
   }
}

void x87_fcomp( struct x86_function *p, struct x86_reg dst )
{
   if (dst.file == file_x87) 
      emit_2ub(p, 0xd8, 0xd8 + dst.idx);
   else {
      emit_1ub(p, 0xd8);
      emit_modrm_noreg(p, 3, dst);
   }
}


void x87_fnstsw( struct x86_function *p, struct x86_reg dst )
{
   assert(dst.file == file_REG32);

   if (dst.idx == reg_AX &&
       dst.mod == mod_REG) 
      emit_2ub(p, 0xdf, 0xe0);
   else {
      emit_1ub(p, 0xdd);
      emit_modrm_noreg(p, 7, dst);
   }
}




/***********************************************************************
 * MMX instructions
 */

void mmx_emms( struct x86_function *p )
{
   assert(p->need_emms);
   emit_2ub(p, 0x0f, 0x77);
   p->need_emms = 0;
}

void mmx_packssdw( struct x86_function *p,
		   struct x86_reg dst,
		   struct x86_reg src )
{
   assert(dst.file == file_MMX && 
	  (src.file == file_MMX || src.mod != mod_REG));

   p->need_emms = 1;

   emit_2ub(p, X86_TWOB, 0x6b);
   emit_modrm( p, dst, src );
}

void mmx_packuswb( struct x86_function *p,
		   struct x86_reg dst,
		   struct x86_reg src )
{
   assert(dst.file == file_MMX && 
	  (src.file == file_MMX || src.mod != mod_REG));

   p->need_emms = 1;

   emit_2ub(p, X86_TWOB, 0x67);
   emit_modrm( p, dst, src );
}

void mmx_movd( struct x86_function *p,
	       struct x86_reg dst,
	       struct x86_reg src )
{
   p->need_emms = 1;
   emit_1ub(p, X86_TWOB);
   emit_op_modrm( p, 0x6e, 0x7e, dst, src );
}

void mmx_movq( struct x86_function *p,
	       struct x86_reg dst,
	       struct x86_reg src )
{
   p->need_emms = 1;
   emit_1ub(p, X86_TWOB);
   emit_op_modrm( p, 0x6f, 0x7f, dst, src );
}


/***********************************************************************
 * Helper functions
 */


/* Retreive a reference to one of the function arguments, taking into
 * account any push/pop activity:
 */
struct x86_reg x86_fn_arg( struct x86_function *p,
			   unsigned arg )
{
   return x86_make_disp(x86_make_reg(file_REG32, reg_SP), 
			p->stack_offset + arg * 4);	/* ??? */
}


void x86_init_func( struct x86_function *p )
{
   p->size = 0;
   p->store = NULL;
   p->csr = p->store;
}

int x86_init_func_size( struct x86_function *p, unsigned code_size )
{
   p->size = code_size;
   p->store = _mesa_exec_malloc(code_size);
   p->csr = p->store;
   return p->store != NULL;
}

void x86_release_func( struct x86_function *p )
{
   _mesa_exec_free(p->store);
   p->store = NULL;
   p->csr = NULL;
   p->size = 0;
}


void (*x86_get_func( struct x86_function *p ))(void)
{
   if (DISASSEM && p->store)
      _mesa_printf("disassemble %p %p\n", p->store, p->csr);
   return (void (*)(void)) (unsigned long) p->store;
}

#else

void x86sse_dummy( void )
{
}

#endif