On a G4, the pipelining was crucial. And on a 8 channel 24 bits (well 32) audio path, it was even more so.

Compiler Abuse

/*
 *  Clip_48_8.cpp
 *  EZ8AudioDriver
 *
 *  Created by Michel on Sun Oct 06 2002.
 *  Copyright (c) 2002 __MyCompanyName__. All rights reserved.
 *
 */

#include <libkern/OSTypes.h>
#include "ice1712.h"

typedef unsigned long UInt32;

//#undef FAST

#if FAST
#define FASTPLAY	1
#define FASTCAP		1
#else
#warning Compiling in SLOW mode!
#endif

namespace EZ8 {

static inline double __clip( register double B )
{
	register double result;
	asm( "fctiw %0, %1" : "=f" (result) : "f" (B)  );
	return result;
}

void	F32L24_48_8(double *v, float *ii, UInt32 *oo, long count)
{
	register double	v0=v[0],v1=v[1],v2=v[2],v3=v[3],v4=v[4],v5=v[5],v6=v[6],v7=v[7];
	// better see that with tabs == 4!
	register double		scale		= 2147483648.0;

#if FASTPLAY
	#define _load(i)		s##i = ii[i]
	#define _volc(i)		s##i *= v##i
	#define	_clip(i)		__clip(s##i)
	#define	_d2l1(i)		s##i *= scale
	#define _d2l2(i)		o##i = (SInt32)s##i
	#define _stor(i) 		__asm__( "stwbrx %0, %1, %2" : : "r" (o##i), "b%" (i << 2), "r" (oo) : "memory" )
#else
	// this is the equivalent, without using assembly
	#define _load(i)		s##i = ii[i]
	#define _volc(i)		s##i *= v##i
	#define	_clip(i)		if (s##i > 1.0)	s##i = 1.0; else if (s##i < -1.0) s##i = -1.0
	#define	_d2l1(i)		o##i = (SInt32)(s##i * scale)
	#define _d2l2(i)		o##i = (((o##i >> 8) & 0xff) << 16) | (((o##i >> 16) & 0xff) << 8) | (((o##i >> 24) & 0xff))
	#define _stor(i) 		oo[i] = o##i
#endif

	register double	s0,s1,s2,s3,s4,s5,s6,s7;
	register UInt32	o0,o1,o2,o3,o4,o5,o6,o7;

	while (count--) {
        // staged pipeline 6x8
        _load(0);
        _load(1);
        _load(2);_volc(0);
        _load(3);_volc(1);_clip(0);
        _load(4);_volc(2);_clip(1);_d2l1(0);
        _load(5);_volc(3);_clip(2);_d2l1(1);_d2l2(0);
        _load(6);_volc(4);_clip(3);_d2l1(2);_d2l2(1);
        _load(7);_volc(5);_clip(4);_d2l1(3);_d2l2(2);_stor(0);
                 _volc(6);_clip(5);_d2l1(4);_d2l2(3);_stor(1);
                 _volc(7);_clip(6);_d2l1(5);_d2l2(4);_stor(2);
                          _clip(7);_d2l1(6);_d2l2(5);_stor(3);
                                   _d2l1(7);_d2l2(6);_stor(4);
                                            _d2l2(7);_stor(5);
                                                     _stor(6);
                                                     _stor(7);
		ii += 8;
		oo += NUM_CHANNELS_OUT;
	}
}

#undef _load
#undef _volc
#undef _volm
#undef _clip
#undef _d2l1
#undef _d2l2
#undef _stor


static inline double __fctiw( register double B )
{
	register double result;
	asm( "fctiw %0, %1" : "=f" (result) : "f" (B)  );
	return result;
}

void	L24F32_48_8(double *v, UInt32 *ii, float *oo, long count)
{
	register double	v0=v[0],v1=v[1],v2=v[2],v3=v[3],v4=v[4],v5=v[5],v6=v[6],v7=v[7];

#if FASTCAP
	union { double d; unsigned int i[2]; } transfer[2];
	transfer[0].i[0] = transfer[1].i[0] = (0x434UL - 32) << 20;
		//0x41400000UL;
	transfer[0].i[1] = 0x80000000;
	double	dBias = transfer[0].d;

	#define _load(i)	__asm__ __volatile__("lwbrx %0, %1, %2" : "=r" (s##i) : "b%" (i << 2), "r" (ii) : "memory" )
	#define _swap(i)	s##i ^= 0x80000000UL
	#define	_l2d1(_i)	transfer[_i & 1].i[1] = s##_i
	#define	_l2d2(_i)	o##_i = transfer[_i & 1].d
	#define	_l2d3(i)	o##i -= dBias
	#define _volc(i)	o##i *= v##i
	#define _stor(i)	oo[i] = o##i
#else
	register double		scale		= 4.656612873077392578125e-10;	//  1 / 2147483648.0;	// prevent unnecessary divisions
	#define _load(i)	s##i = ii[i]
	#define _swap(i)	s##i = ((s##i & 0xff) << 24) | (((s##i >> 8) & 0xff) << 16) | (((s##i >> 16) & 0xff) << 8)
	#define	_l2d1(i)	o##i = (((double)(SInt32)s##i) * scale)
	#define _l2d2(i)
	#define	_l2d3(i)	
	#define _volc(i)	o##i *= v##i
	#define _stor(i)	oo[i] = o##i
#endif

	register UInt32	s0,s1,s2,s3,s4,s5,s6,s7;
	register double	o0,o1,o2,o3,o4,o5,o6,o7;

	while (count--) {
        // staged pipeline 7x8
        _load(0);
        _load(1);
        _load(2);
        _load(3);_swap(0);
        _load(4);_swap(1);_l2d1(0);
        _load(5);_swap(2);_l2d1(1);_l2d2(0);
        _load(6);_swap(3);_l2d1(2);_l2d2(1);
        _load(7);_swap(4);_l2d1(3);_l2d2(2);_l2d3(0);
                 _swap(5);_l2d1(4);_l2d2(3);_l2d3(1);_volc(0);
                 _swap(6);_l2d1(5);_l2d2(4);_l2d3(2);_volc(1);_stor(0);
                 _swap(7);_l2d1(6);_l2d2(5);_l2d3(3);_volc(2);_stor(1);
                          _l2d1(7);_l2d2(6);_l2d3(4);_volc(3);_stor(2);
                                   _l2d2(7);_l2d3(5);_volc(4);_stor(3);
                                            _l2d3(6);_volc(5);_stor(4);
                                            _l2d3(7);_volc(6);_stor(5);
                                                     _volc(7);_stor(6);
                                                              _stor(7);
		ii += NUM_CHANNELS_IN;
		oo += 8;
	}
}


#undef _load
#undef _stor
#undef _clip
#undef _scal
#undef _swap

}	// namespace EZ8