jedi-academy/codemp/unix/linux_common.c

/** 
 * GAS syntax equivalents of the MSVC asm memory calls in common.c
 *
 * The following changes have been made to the asm:
 * 1. Registers are loaded by the inline asm arguments when possible
 * 2. Labels have been changed to local label format (0,1,etc.) to allow inlining
 *
 * HISTORY:
 *	AH - Created on 08 Dec 2000
 */

#include <unistd.h>   // AH - for size_t
#include <string.h>

// bk001207 - we need something under Linux, too. Mac?
#if 1 // defined(C_ONLY) // bk010102 - dedicated?
void Com_Memcpy (void* dest, const void* src, const size_t count) {
  memcpy(dest, src, count);
}

void Com_Memset (void* dest, const int val, const size_t count) {
  memset(dest, val, count);
}

#else

typedef enum {
  PRE_READ,         // prefetch assuming that buffer is used for reading only
  PRE_WRITE,        // prefetch assuming that buffer is used for writing only
  PRE_READ_WRITE    // prefetch assuming that buffer is used for both reading and writing
} e_prefetch;

void Com_Prefetch (const void *s, const unsigned int bytes, e_prefetch type);

void _copyDWord (unsigned int* dest, const unsigned int constant, const unsigned int count) {
	// MMX version not used on standard Pentium MMX
	// because the dword version is faster (with
	// proper destination prefetching)
		__asm__ __volatile__ ("
			//mov			eax,constant		// eax = val
			//mov			edx,dest			// dest
			//mov			ecx,count
			movd		%%eax, %%mm0
			punpckldq	%%mm0, %%mm0

			// ensure that destination is qword aligned

			testl		$7, %%edx				// qword padding?
			jz		0f	
			movl		%%eax, (%%edx)
			decl		%%ecx
			addl		$4, %%edx

0:			movl		%%ecx, %%ebx				
			andl		$0xfffffff0, %%ecx				
			jz		2f
			jmp		1f
			.align 		16

			// funny ordering here to avoid commands
			// that cross 32-byte boundaries (the
			// [edx+0] version has a special 3-byte opcode...
1:			movq		%%mm0, 8(%%edx)
			movq		%%mm0, 16(%%edx)
			movq		%%mm0, 24(%%edx)
			movq		%%mm0, 32(%%edx)
			movq		%%mm0, 40(%%edx)
			movq		%%mm0, 48(%%edx)
			movq		%%mm0, 56(%%edx)
			movq		%%mm0, (%%edx)
			addl		$64, %%edx
			subl		$16, %%ecx
			jnz		1b
2:	
			movl		%%ebx, %%ecx				// ebx = cnt
			andl		$0xfffffff0, %%ecx				// ecx = cnt&~15
			subl		%%ecx, %%ebx
			jz		6f
			cmpl		$8, %%ebx
			jl		3f

			movq		%%mm0, (%%edx)
			movq		%%mm0, 8(%%edx)
			movq		%%mm0, 16(%%edx)
			movq		%%mm0, 24(%%edx)
			addl		$32, %%edx
			subl		$8, %%ebx
			jz		6f

3:			cmpl		$4, %%ebx
			jl		4f
			
			movq		%%mm0, (%%edx)
			movq		%%mm0, 8(%%edx)
			addl		$16, %%edx
			subl		$4, %%ebx

4:			cmpl		$2, %%ebx
			jl		5f
			movq		%%mm0, (%%edx)
			addl		$8, %%edx
			subl		$2, %%ebx

5:			cmpl		$1, %%ebx
			jl		6f
			movl		%%eax, (%%edx)
6:
			emms
	"
	: : "a" (constant), "c" (count), "d" (dest)
	: "%ebx", "%edi", "%esi", "cc", "memory");
}

// optimized memory copy routine that handles all alignment
// cases and block sizes efficiently
void Com_Memcpy (void* dest, const void* src, const size_t count) {
	Com_Prefetch (src, count, PRE_READ);
	__asm__ __volatile__ ("
		pushl		%%edi
		pushl		%%esi
		//mov		ecx,count
		cmpl		$0, %%ecx						// count = 0 check (just to be on the safe side)
		je		6f
		//mov		edx,dest
		movl		%0, %%ebx
		cmpl		$32, %%ecx						// padding only?
		jl		1f

		movl		%%ecx, %%edi					
		andl		$0xfffffe00, %%edi					// edi = count&~31
		subl		$32, %%edi

		.align 16
0:
		movl		(%%ebx, %%edi, 1), %%eax
		movl		4(%%ebx, %%edi, 1), %%esi
		movl		%%eax, (%%edx, %%edi, 1)
		movl		%%esi, 4(%%edx, %%edi, 1)
		movl		8(%%ebx, %%edi, 1), %%eax
		movl		12(%%ebx, %%edi, 1), %%esi
		movl		%%eax, 8(%%edx, %%edi, 1)
		movl		%%esi, 12(%%edx, %%edi, 1)
		movl		16(%%ebx, %%edi, 1), %%eax
		movl		20(%%ebx, %%edi, 1), %%esi
		movl		%%eax, 16(%%edx, %%edi, 1)
		movl		%%esi, 20(%%edx, %%edi, 1)
		movl		24(%%ebx, %%edi, 1), %%eax
		movl		28(%%ebx, %%edi, 1), %%esi
		movl		%%eax, 24(%%edx, %%edi, 1)
		movl		%%esi, 28(%%edx, %%edi, 1)
		subl		$32, %%edi
		jge		0b
		
		movl		%%ecx, %%edi
		andl		$0xfffffe00, %%edi
		addl		%%edi, %%ebx					// increase src pointer
		addl		%%edi, %%edx					// increase dst pointer
		andl		$31, %%ecx					// new count
		jz		6f					// if count = 0, get outta here

1:
		cmpl		$16, %%ecx
		jl		2f
		movl		(%%ebx), %%eax
		movl		%%eax, (%%edx)
		movl		4(%%ebx), %%eax
		movl		%%eax, 4(%%edx)
		movl		8(%%ebx), %%eax
		movl		%%eax, 8(%%edx)
		movl		12(%%ebx), %%eax
		movl		%%eax, 12(%%edx)
		subl		$16, %%ecx
		addl		$16, %%ebx
		addl		$16, %%edx
2:
		cmpl		$8, %%ecx
		jl		3f
		movl		(%%ebx), %%eax
		movl		%%eax, (%%edx)
		movl		4(%%ebx), %%eax
		subl		$8, %%ecx
		movl		%%eax, 4(%%edx)
		addl		$8, %%ebx
		addl		$8, %%edx
3:
		cmpl		$4, %%ecx
		jl		4f
		movl		(%%ebx), %%eax	// here 4-7 bytes
		addl		$4, %%ebx
		subl		$4, %%ecx
		movl		%%eax, (%%edx)
		addl		$4, %%edx
4:							// 0-3 remaining bytes
		cmpl		$2, %%ecx
		jl		5f
		movw		(%%ebx), %%ax	// two bytes
		cmpl		$3, %%ecx				// less than 3?
		movw		%%ax, (%%edx)
		jl		6f
		movb		2(%%ebx), %%al	// last byte
		movb		%%al, 2(%%edx)
		jmp		6f
5:
		cmpl		$1, %%ecx
		jl		6f
		movb		(%%ebx), %%al
		movb		%%al, (%%edx)
6:
		popl		%%esi
		popl		%%edi
	"
	: : "m" (src), "d" (dest), "c" (count)
	: "%eax", "%ebx", "%edi", "%esi", "cc", "memory");
}

void Com_Memset (void* dest, const int val, const size_t count)
{
	unsigned int fillval;

	if (count < 8)
	{
		__asm__ __volatile__ ("
			//mov		edx,dest
			//mov		eax, val
			movb		%%al, %%ah
			movl		%%eax, %%ebx
			andl		$0xffff, %%ebx
			shll		$16, %%eax
			addl		%%ebx, %%eax	// eax now contains pattern
			//mov		ecx,count
			cmpl		$4, %%ecx
			jl		0f
			movl		%%eax, (%%edx)	// copy first dword
			addl		$4, %%edx
			subl		$4, %%ecx
	0:		cmpl		$2, %%ecx
			jl		1f
			movw		%%ax, (%%edx)	// copy 2 bytes
			addl		$2, %%edx
			subl		$2, %%ecx
	1:		cmpl		$0, %%ecx
			je		2f
			movb		%%al, (%%edx)	// copy single byte
	2:		
		"
		: : "d" (dest), "a" (val), "c" (count)
		: "%ebx", "%edi", "%esi", "cc", "memory");
		
		return;
	}

	fillval = val;
	
	fillval = fillval|(fillval<<8);
	fillval = fillval|(fillval<<16);		// fill dword with 8-bit pattern

	_copyDWord ((unsigned int*)(dest),fillval, count/4);
	
	__asm__ __volatile__ ("     		// padding of 0-3 bytes
		//mov		ecx,count
		movl		%%ecx, %%eax
		andl		$3, %%ecx
		jz		1f
		andl		$0xffffff00, %%eax
		//mov		ebx,dest
		addl		%%eax, %%edx
		movl		%0, %%eax
		cmpl		$2, %%ecx
		jl		0f
		movw		%%ax, (%%edx)
		cmpl		$2, %%ecx
		je		1f					
		movb		%%al, 2(%%edx)		
		jmp		1f
0:		
		cmpl		$0, %%ecx
		je		1f
		movb		%%al, (%%edx)
1:	
	"
	: : "m" (fillval), "c" (count), "d" (dest)
	: "%eax", "%ebx", "%edi", "%esi", "cc", "memory");	
}

void Com_Prefetch (const void *s, const unsigned int bytes, e_prefetch type)
{
	// write buffer prefetching is performed only if
	// the processor benefits from it. Read and read/write
	// prefetching is always performed.

	switch (type)
	{
		case PRE_WRITE : break;
		case PRE_READ:
		case PRE_READ_WRITE:

		__asm__ __volatile__ ("
			//mov		ebx,s
			//mov		ecx,bytes
			cmpl		$4096, %%ecx				// clamp to 4kB
			jle		0f
			movl		$4096, %%ecx
	0:
			addl		$0x1f, %%ecx
			shrl		$5, %%ecx					// number of cache lines
			jz		2f
			jmp		1f

			.align 16
	1:		testb		%%al, (%%edx)
			addl		$32, %%edx
			decl		%%ecx
			jnz		1b
	2:
		"
		: : "d" (s), "c" (bytes)
		: "%eax", "%ebx", "%edi", "%esi", "memory", "cc");
		
		break;
	}
}

#endif