mmx.h

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

/*	mmx.h

	MultiMedia eXtensions GCC interface library for IA32.

	To use this library, simply include this header file
	and compile with GCC.  You MUST have inlining enabled
	in order for mmx_ok() to work; this can be done by
	simply using -O on the GCC command line.

	Compiling with -DMMX_TRACE will cause detailed trace
	output to be sent to stderr for each mmx operation.
	This adds lots of code, and obviously slows execution to
	a crawl, but can be very useful for debugging.

	THIS SOFTWARE IS PROVIDED ``AS IS'' AND WITHOUT ANY
	EXPRESS OR IMPLIED WARRANTIES, INCLUDING, WITHOUT
	LIMITATION, THE IMPLIED WARRANTIES OF MERCHANTABILITY
	AND FITNESS FOR ANY PARTICULAR PURPOSE.

	1997-98 by H. Dietz and R. Fisher

 History:
	97-98*	R.Fisher	Early versions
	980501	R.Fisher	Original Release
	980611*	H.Dietz		Rewrite, correctly implementing inlines, and
		R.Fisher	 including direct register accesses.
	980616	R.Fisher	Release of 980611 as 980616.
	980714	R.Fisher	Minor corrections to Makefile, etc.
	980715	R.Fisher	mmx_ok() now prevents optimizer from using
				 clobbered values.
				mmx_ok() now checks if cpuid instruction is
				 available before trying to use it.
	980726*	R.Fisher	mm_support() searches for AMD 3DNow, Cyrix
				 Extended MMX, and standard MMX.  It returns a
				 value which is positive if any of these are
				 supported, and can be masked with constants to
				 see which.  mmx_ok() is now a call to this
	980726*	R.Fisher	Added i2r support for shift functions
	980919	R.Fisher	Fixed AMD extended feature recognition bug.
	980921	R.Fisher	Added definition/check for _MMX_H.
				Added "float s[2]" to mmx_t for use with
				  3DNow and EMMX.  So same mmx_t can be used.
	981013	R.Fisher	Fixed cpuid function 1 bug (looked at wrong reg)
				Fixed psllq_i2r error in mmxtest.c

	* Unreleased (internal or interim) versions

 Notes:
	It appears that the latest gas has the pand problem fixed, therefore
	  I'll undefine BROKEN_PAND by default.
	String compares may be quicker than the multiple test/jumps in vendor
	  test sequence in mmx_ok(), but I'm not concerned with that right now.

 Acknowledgments:
	Jussi Laako for pointing out the errors ultimately found to be
	  connected to the failure to notify the optimizer of clobbered values.
	Roger Hardiman for reminding us that CPUID isn't everywhere, and that
	  someone may actually try to use this on a machine without CPUID.
	  Also for suggesting code for checking this.
	Robert Dale for pointing out the AMD recognition bug.
	Jimmy Mayfield and Carl Witty for pointing out the Intel recognition
	  bug.
	Carl Witty for pointing out the psllq_i2r test bug.
*/

#ifndef _MMX_H
#define _MMX_H

//#define MMX_TRACE

/*	Warning:  at this writing, the version of GAS packaged
	with most Linux distributions does not handle the
	parallel AND operation mnemonic correctly.  If the
	symbol BROKEN_PAND is defined, a slower alternative
	coding will be used.  If execution of mmxtest results
	in an illegal instruction fault, define this symbol.
*/
#undef	BROKEN_PAND


/*	The type of an value that fits in an MMX register
	(note that long long constant values MUST be suffixed
	 by LL and unsigned long long values by ULL, lest
	 they be truncated by the compiler)
*/
typedef	union {
	long long		q;	/* Quadword (64-bit) value */
	unsigned long long	uq;	/* Unsigned Quadword */
	int			d[2];	/* 2 Doubleword (32-bit) values */
	unsigned int		ud[2];	/* 2 Unsigned Doubleword */
	short			w[4];	/* 4 Word (16-bit) values */
	unsigned short		uw[4];	/* 4 Unsigned Word */
	char			b[8];	/* 8 Byte (8-bit) values */
	unsigned char		ub[8];	/* 8 Unsigned Byte */
	float			s[2];	/* Single-precision (32-bit) value */
} mmx_t;

/*	Helper functions for the instruction macros that follow...
	(note that memory-to-register, m2r, instructions are nearly
	 as efficient as register-to-register, r2r, instructions;
	 however, memory-to-memory instructions are really simulated
	 as a convenience, and are only 1/3 as efficient)
*/
#ifdef	MMX_TRACE

/*	Include the stuff for printing a trace to stderr...
*/

#include <stdio.h>

#define	mmx_i2r(op, imm, reg) \
	{ \
		mmx_t mmx_trace; \
		mmx_trace = (imm); \
		fprintf(stderr, #op "_i2r(" #imm "=0x%016llx, ", mmx_trace.q); \
		__asm__ __volatile__ ("movq %%" #reg ", %0" \
				      : "=X" (mmx_trace) \
				      : /* nothing */ ); \
		fprintf(stderr, #reg "=0x%016llx) => ", mmx_trace.q); \
		__asm__ __volatile__ (#op " %0, %%" #reg \
				      : /* nothing */ \
				      : "X" (imm)); \
		__asm__ __volatile__ ("movq %%" #reg ", %0" \
				      : "=X" (mmx_trace) \
				      : /* nothing */ ); \
		fprintf(stderr, #reg "=0x%016llx\n", mmx_trace.q); \
	}

#define	mmx_m2r(op, mem, reg) \
	{ \
		mmx_t mmx_trace; \
		mmx_trace = (mem); \
		fprintf(stderr, #op "_m2r(" #mem "=0x%016llx, ", mmx_trace.q); \
		__asm__ __volatile__ ("movq %%" #reg ", %0" \
				      : "=X" (mmx_trace) \
				      : /* nothing */ ); \
		fprintf(stderr, #reg "=0x%016llx) => ", mmx_trace.q); \
		__asm__ __volatile__ (#op " %0, %%" #reg \
				      : /* nothing */ \
				      : "X" (mem)); \
		__asm__ __volatile__ ("movq %%" #reg ", %0" \
				      : "=X" (mmx_trace) \
				      : /* nothing */ ); \
		fprintf(stderr, #reg "=0x%016llx\n", mmx_trace.q); \
	}

#define	mmx_r2m(op, reg, mem) \
	{ \
		mmx_t mmx_trace; \
		__asm__ __volatile__ ("movq %%" #reg ", %0" \
				      : "=X" (mmx_trace) \
				      : /* nothing */ ); \
		fprintf(stderr, #op "_r2m(" #reg "=0x%016llx, ", mmx_trace.q); \
		mmx_trace = (mem); \
		fprintf(stderr, #mem "=0x%016llx) => ", mmx_trace.q); \
		__asm__ __volatile__ (#op " %%" #reg ", %0" \
				      : "=X" (mem) \
				      : /* nothing */ ); \
		mmx_trace = (mem); \
		fprintf(stderr, #mem "=0x%016llx\n", mmx_trace.q); \
	}

#define	mmx_r2r(op, regs, regd) \
	{ \
		mmx_t mmx_trace; \
		__asm__ __volatile__ ("movq %%" #regs ", %0" \
				      : "=X" (mmx_trace) \
				      : /* nothing */ ); \
		fprintf(stderr, #op "_r2r(" #regs "=0x%016llx, ", mmx_trace.q); \
		__asm__ __volatile__ ("movq %%" #regd ", %0" \
				      : "=X" (mmx_trace) \
				      : /* nothing */ ); \
		fprintf(stderr, #regd "=0x%016llx) => ", mmx_trace.q); \
		__asm__ __volatile__ (#op " %" #regs ", %" #regd); \
		__asm__ __volatile__ ("movq %%" #regd ", %0" \
				      : "=X" (mmx_trace) \
				      : /* nothing */ ); \
		fprintf(stderr, #regd "=0x%016llx\n", mmx_trace.q); \
	}

#define	mmx_m2m(op, mems, memd) \
	{ \
		mmx_t mmx_trace; \
		mmx_trace = (mems); \
		fprintf(stderr, #op "_m2m(" #mems "=0x%016llx, ", mmx_trace.q); \
		mmx_trace = (memd); \
		fprintf(stderr, #memd "=0x%016llx) => ", mmx_trace.q); \
		__asm__ __volatile__ ("movq %0, %%mm0\n\t" \
				      #op " %1, %%mm0\n\t" \
				      "movq %%mm0, %0" \
				      : "=X" (memd) \
				      : "X" (mems)); \
		mmx_trace = (memd); \
		fprintf(stderr, #memd "=0x%016llx\n", mmx_trace.q); \
	}

#else

/*	These macros are a lot simpler without the tracing...
*/

#define	mmx_i2r(op, imm, reg) \
	__asm__ __volatile__ (#op " $" #imm ", %%" #reg \
			      : /* nothing */ \
			      : /* nothing */);

#define	mmx_m2r(op, mem, reg) \
	__asm__ __volatile__ (#op " %0, %%" #reg \
			      : /* nothing */ \
			      : "X" (mem))

#define	mmx_r2m(op, reg, mem) \
	__asm__ __volatile__ (#op " %%" #reg ", %0" \
			      : "=X" (mem) \
			      : /* nothing */ )

#define	mmx_r2r(op, regs, regd) \
	__asm__ __volatile__ (#op " %" #regs ", %" #regd)

#define	mmx_m2m(op, mems, memd) \
	__asm__ __volatile__ ("movq %0, %%mm0\n\t" \
			      #op " %1, %%mm0\n\t" \
			      "movq %%mm0, %0" \
			      : "=X" (memd) \
			      : "X" (mems))

#endif


/*	1x64 MOVe Quadword
	(this is both a load and a store...
	 in fact, it is the only way to store)
*/
#define	movq_m2r(var, reg)	mmx_m2r(movq, var, reg)
#define	movq_r2m(reg, var)	mmx_r2m(movq, reg, var)
#define	movq_r2r(regs, regd)	mmx_r2r(movq, regs, regd)
#define	movq(vars, vard) \
	__asm__ __volatile__ ("movq %1, %%mm0\n\t" \
			      "movq %%mm0, %0" \
			      : "=X" (vard) \
			      : "X" (vars))


/*	1x32 MOVe Doubleword
	(like movq, this is both load and store...
	 but is most useful for moving things between
	 mmx registers and ordinary registers)
*/
#define	movd_m2r(var, reg)	mmx_m2r(movd, var, reg)
#define	movd_r2m(reg, var)	mmx_r2m(movd, reg, var)
#define	movd_r2r(regs, regd)	mmx_r2r(movd, regs, regd)
#define	movd(vars, vard) \
	__asm__ __volatile__ ("movd %1, %%mm0\n\t" \
			      "movd %%mm0, %0" \
			      : "=X" (vard) \
			      : "X" (vars))


/*	2x32, 4x16, and 8x8 Parallel ADDs
*/
#define	paddd_m2r(var, reg)	mmx_m2r(paddd, var, reg)
#define	paddd_r2r(regs, regd)	mmx_r2r(paddd, regs, regd)
#define	paddd(vars, vard)	mmx_m2m(paddd, vars, vard)

#define	paddw_m2r(var, reg)	mmx_m2r(paddw, var, reg)
#define	paddw_r2r(regs, regd)	mmx_r2r(paddw, regs, regd)
#define	paddw(vars, vard)	mmx_m2m(paddw, vars, vard)

#define	paddb_m2r(var, reg)	mmx_m2r(paddb, var, reg)
#define	paddb_r2r(regs, regd)	mmx_r2r(paddb, regs, regd)
#define	paddb(vars, vard)	mmx_m2m(paddb, vars, vard)


/*	4x16 and 8x8 Parallel ADDs using Saturation arithmetic
*/
#define	paddsw_m2r(var, reg)	mmx_m2r(paddsw, var, reg)
#define	paddsw_r2r(regs, regd)	mmx_r2r(paddsw, regs, regd)
#define	paddsw(vars, vard)	mmx_m2m(paddsw, vars, vard)

#define	paddsb_m2r(var, reg)	mmx_m2r(paddsb, var, reg)
#define	paddsb_r2r(regs, regd)	mmx_r2r(paddsb, regs, regd)