test.c

version control, version control

					: "a" (p1), "0" (p), "d" (pe), "b" (p2)
					: "ecx"
				);
				do_tick();
				BAILR
			} while (!done);
		}
	}
}

void movinv32(int iter, ulong p1, ulong lb, ulong hb, int sval, int off)
{
	int i, j, k=0, done;
	volatile ulong *pe;
	volatile ulong *start, *end;
	ulong pat = 0;

/* CDH start
 *	ulong p3 = sval << 31;
 * CDH end
 */

	/* Display the current pattern */
	hprint(LINE_PAT, COL_PAT, p1);

	/* Initialize memory with the initial pattern.  */
	for (j=0; j<segs; j++) {
		start = v->map[j].start;
		end = v->map[j].end;
		pe = start;
		p = start;
		done = 0;
		k = off;
		pat = p1;
		do {
			/* Check for overflow */
			if (pe + SPINSZ > pe) {
				pe += SPINSZ;
			} else {
				pe = end;
			}
			if (pe >= end) {
				pe = end;
				done++;
			}
			if (p == pe ) {
				break;
			}
			/* Do a SPINSZ section of memory */
/* Original C code replaced with hand tuned assembly code
 *			while (p < pe) {
 *				*p = pat;
 *				if (++k >= 32) {
 *					pat = lb;
 *					k = 0;
 *				} else {
 *					pat = pat << 1;
 *					pat |= sval;
 *				}
 *				p++;
 *			}
 */
			asm __volatile__ (
				"jmp L20\n\t"
				".p2align 4,,7\n\t"

/* CDH start */
				"L20:\n\t"
				"movl %%ecx,(%%edi)\n\t"
				"incb %%bl\n\t"
				"addl $4,%%edi\n\t"
				"roll $1,%%ecx\n\t"
				"cmpl %%edx,%%edi\n\t"
				"jb L20\n\t"
				"andb $31,%%bl\n\t"
				: "=b" (k), "=D" (p), "=c" (pat)
				: "D" (p),"d" (pe),"b" (k),"c" (pat)
/* CDH end */
			);
			do_tick();
			BAILR
		} while (!done);
	}

	/* Do moving inversions test. Check for initial pattern and then
	 * write the complement for each memory location. Test from bottom
	 * up and then from the top down.  */
	for (i=0; i<iter; i++) {
		for (j=0; j<segs; j++) {
			start = v->map[j].start;
			end = v->map[j].end;
			pe = start;
			p = start;
			done = 0;
			k = off;
			pat = p1;
			do {
				/* Check for overflow */
				if (pe + SPINSZ > pe) {
					pe += SPINSZ;
				} else {
					pe = end;
				}
				if (pe >= end) {
					pe = end;
					done++;
				}
				if (p == pe ) {
					break;
				}
/* Original C code replaced with hand tuned assembly code
 *				while (p < pe) {
 *					if ((bad=*p) != pat) {
 *						error((ulong*)p, pat, bad);
 *					}
 *					*p = ~pat;
 *					if (++k >= 32) {
 *						pat = lb;
 *						k = 0;
 *					} else {
 *						pat = pat << 1;
 *						pat |= sval;
 *					}
 *					p++;
 *				}
 */
				asm __volatile__ (
					"pushl %%ebp\n\t"
					"jmp L30\n\t"

					".p2align 4,,7\n\t"
					"L30:\n\t"
					"movl (%%edi),%%ebp\n\t"
					"cmpl %%ecx,%%ebp\n\t"
					"jne L34\n\t"

/* CDH start */
					"L35:\n\t"
					"notl %%ecx\n\t"
					"movl %%ecx,(%%edi)\n\t"
					"notl %%ecx\n\t"
					"addl $4,%%edi\n\t"
					"incb %%bl\n\t"
					"roll $1,%%ecx\n\t"
					"cmpl %%edx,%%edi\n\t"
					"jb L30\n\t"
					"jmp L33\n\t"
/* CDH end */

					"L34:\n\t" \
					"pushl %%esi\n\t"
					"pushl %%eax\n\t"
					"pushl %%ebx\n\t"
					"pushl %%edx\n\t"
					"pushl %%ebp\n\t"
					"pushl %%ecx\n\t"
					"pushl %%edi\n\t"
					"call error\n\t"
					"popl %%edi\n\t"
					"popl %%ecx\n\t"
					"popl %%ebp\n\t"
					"popl %%edx\n\t"
					"popl %%ebx\n\t"
					"popl %%eax\n\t"
					"popl %%esi\n\t"
					"jmp L35\n"

/* CDH start */
					"L33:\n\t"
					"andb $31,%%bl\n\t"
					"popl %%ebp\n\t"
					: "=b" (k), "=D" (p), "=c" (pat)
					: "D" (p),"d" (pe),"b" (k),"c" (pat)
/* CDH end */
				);
				do_tick();
				BAILR
			} while (!done);
		}

		/* Since we already adjusted k and the pattern this
		 * code backs both up one step
		 */
/* CDH start */
/* Original C code replaced with hand tuned assembly code
 *		pat = lb;
 *		if ( 0 != (k = (k-1) & 31) ) {
 *			pat = (pat << k);
 *			if ( sval )
 *			pat |= ((sval << k) - 1);
 *		}
 *		k++;
 */
			asm __volatile__ (
			"decl %%ecx\n\t"
			"andl $31,%%ecx\n\t"
			"roll %%cl,%%ebx\n\t"
			"incb %%cl\n\t"
			: "=c" (k), "=b" (pat)
			: "c" (k), "b" (lb)
			);
/* CDH end */

		for (j=segs-1; j>=0; j--) {
			start = v->map[j].start;
			end = v->map[j].end;
			p = end -1;
			pe = end -1;
			done = 0;
			do {
				/* Check for underflow */
				if (pe - SPINSZ < pe) {
					pe -= SPINSZ;
				} else {
					pe = start;
				}
				if (pe <= start) {
					pe = start;
					done++;
				}
				if (p == pe ) {
					break;
				}
/* Original C code replaced with hand tuned assembly code
 *				do {
 *					if ((bad=*p) != ~pat) {
 *						error((ulong*)p, ~pat, bad);
 *					}
 *					*p = pat;
 *					if (--k <= 0) {
 *						pat = hb;
 *						k = 32;
 *					} else {
 *						pat = pat >> 1;
 *						pat |= p3;
 *					}
 *				} while (p-- > pe);
 */
				asm __volatile__ (
					"pushl %%ebp\n\t"
					"addl $4,%%edi\n\t"
					"jmp L40\n\t"

					".p2align 4,,7\n\t"
					"L40:\n\t"
					"subl $4,%%edi\n\t"
					"movl (%%edi),%%ebp\n\t"
					"notl %%ecx\n\t"
					"cmpl %%ecx,%%ebp\n\t"
					"jne L44\n\t"

/* CDH start */
					"L45:\n\t"
					"notl %%ecx\n\t"
					"movl %%ecx,(%%edi)\n\t"
					"decb %%bl\n\t"
					"rorl $1,%%ecx\n\t"
					"cmpl %%edx,%%edi\n\t"
					"ja L40\n\t"
					"jmp L43\n\t"
/* CDH end */

					"L44:\n\t" \
					"pushl %%esi\n\t"
					"pushl %%eax\n\t"
					"pushl %%ebx\n\t"
					"pushl %%edx\n\t"
					"pushl %%ebp\n\t"
					"pushl %%ecx\n\t"
					"pushl %%edi\n\t"
					"call error\n\t"
					"popl %%edi\n\t"
					"popl %%ecx\n\t"
					"popl %%ebp\n\t"
					"popl %%edx\n\t"
					"popl %%ebx\n\t"
					"popl %%eax\n\t"
					"popl %%esi\n\t"
					"jmp L45\n"

/* CDH start */
					"L43:\n\t"
					"andb $31,%%bl\n\t"
					"subl $4,%%edi\n\t"
					"popl %%ebp\n\t"
					: "=b" (k), "=D" (p), "=c" (pat)
					: "D" (p),"d" (pe),"b" (k),"c" (pat)
/* CDH end */
				);
				do_tick();
				BAILR
			} while (!done);
		}
	}
}

/*
 * Test all of memory using modulo X access pattern.
 */
void modtst(int offset, int iter, ulong p1, ulong p2)
{
	int j, k, l, done;
	volatile ulong *pe;
	volatile ulong *start, *end;

	/* Display the current pattern */
	hprint(LINE_PAT, COL_PAT-2, p1);
	cprint(LINE_PAT, COL_PAT+6, "-");
	dprint(LINE_PAT, COL_PAT+7, offset, 2, 1);

	/* Write every nth location with pattern */
	for (j=0; j<segs; j++) {
		start = v->map[j].start;
		end = v->map[j].end;
		pe = (ulong *)start;
		p = start+offset;
		done = 0;
		do {
			/* Check for overflow */
			if (pe + SPINSZ > pe) {
				pe += SPINSZ;
			} else {
				pe = end;
			}
			if (pe >= end) {
				pe = end;
				done++;
			}
			if (p == pe ) {
				break;
			}
/* Original C code replaced with hand tuned assembly code
 *			for (; p < pe; p += MOD_SZ) {
 *				*p = p1;
 *			}
 */
			asm __volatile__ (
				"jmp L60\n\t" \
				".p2align 4,,7\n\t" \

				"L60:\n\t" \
				"movl %%eax,(%%edi)\n\t" \
				"addl $80,%%edi\n\t" \
				"cmpl %%edx,%%edi\n\t" \
				"jb L60\n\t" \
				: "=D" (p)
				: "D" (p), "d" (pe), "a" (p1)
			);
			do_tick();
			BAILR
		} while (!done);
	}

	/* Write the rest of memory "iter" times with the pattern complement */
	for (l=0; l<iter; l++) {
		for (j=0; j<segs; j++) {
			start = v->map[j].start;
			end = v->map[j].end;
			pe = (ulong *)start;
			p = start;
			done = 0;
			k = 0;
			do {
				/* Check for overflow */
				if (pe + SPINSZ > pe) {
					pe += SPINSZ;
				} else {
					pe = end;
				}
				if (pe >= end) {
					pe = end;
					done++;
				}
				if (p == pe ) {
					break;
				}
/* Original C code replaced with hand tuned assembly code
 *				for (; p < pe; p++) {
 *					if (k != offset) {
 *						*p = p2;
 *					}
 *					if (++k > MOD_SZ-1) {
 *						k = 0;
 *					}
 *				}
 */
				asm __volatile__ (
					"jmp L50\n\t" \
					".p2align 4,,7\n\t" \

					"L50:\n\t" \
					"cmpl %%ebx,%%ecx\n\t" \
					"je L52\n\t" \
					  "movl %%eax,(%%edi)\n\t" \
					"L52:\n\t" \
					"incl %%ebx\n\t" \
					"cmpl $19,%%ebx\n\t" \
					"jle L53\n\t" \
					  "xorl %%ebx,%%ebx\n\t" \
					"L53:\n\t" \
					"addl $4,%%edi\n\t" \
					"cmpl %%edx,%%edi\n\t" \
					"jb L50\n\t" \
					: "=D" (p), "=b" (k)
					: "D" (p), "d" (pe), "a" (p2),
						"b" (k), "c" (offset)
				);
				do_tick();
				BAILR
			} while (!done);
		}
	}

	/* Now check every nth location */
	for (j=0; j<segs; j++) {
		start = v->map[j].start;
		end = v->map[j].end;
		pe = (ulong *)start;
		p = start+offset;
		done = 0;
		do {
			/* Check for overflow */
			if (pe + SPINSZ > pe) {
				pe += SPINSZ;
			} else {
				pe = end;
			}
			if (pe >= end) {
				pe = end;
				done++;
			}
			if (p == pe ) {
				break;
			}
/* Original C code replaced with hand tuned assembly code
 *			for (; p < pe; p += MOD_SZ) {
 *				if ((bad=*p) != p1) {
 *					error((ulong*)p, p1, bad);
 *				}
 *			}
 */
			asm __volatile__ (
				"jmp L70\n\t" \
				".p2align 4,,7\n\t" \

				"L70:\n\t" \
				"movl (%%edi),%%ecx\n\t" \
				"cmpl %%eax,%%ecx\n\t" \
				"jne L71\n\t" \
				"L72:\n\t" \
				"addl $80,%%edi\n\t" \
				"cmpl %%edx,%%edi\n\t" \
				"jb L70\n\t" \
				"jmp L73\n\t" \

				"L71:\n\t" \
				"pushl %%edx\n\t"
				"pushl %%ecx\n\t"
				"pushl %%eax\n\t"
				"pushl %%edi\n\t"
				"call error\n\t"
				"popl %%edi\n\t"
				"popl %%eax\n\t"
				"popl %%ecx\n\t"
				"popl %%edx\n\t"
				"jmp L72\n"

				"L73:\n\t" \
				: "=D" (p)
				: "D" (p), "d" (pe), "a" (p1)
				: "ecx"
			);
			do_tick();
			BAILR
		} while (!done);
	}
	cprint(LINE_PAT, COL_PAT, "          ");
}



/*
 * Test memory using block moves
 * Adapted from Robert Redelmeier's burnBX test
 */
void block_move(int iter)
{
	int i, j, done;
	ulong len;
	volatile ulong p, pe, pp;
	volatile ulong start, end;

	cprint(LINE_PAT, COL_PAT-2, "          ");

	/* Initialize memory with the initial pattern.  */
	for (j=0; j<segs; j++) {
		start = (ulong)v->map[j].start;
#ifdef USB_WAR
		/* We can't do the block move test on low memory beacuase
		 * BIOS USB support clobbers location 0x410 and 0x4e0
		 */
		if (start < 0x4f0) {
			start = 0x4f0;
		}
#endif
		end = (ulong)v->map[j].end;
		pe = start;
		p = start;
		done = 0;
		do {
			/* Check for overflow */
			if (pe + SPINSZ*4 > pe) {
				pe += SPINSZ*4;
			} else {
				pe = end;
			}
			if (pe >= end) {
				pe = end;
				done++;
			}
			if (p == pe ) {
				break;
			}
			len  = ((ulong)pe - (ulong)p) / 64;
			asm __volatile__ (
				"jmp L100\n\t"

				".p2align 4,,7\n\t"
				"L100:\n\t"
				"movl %%eax, %%edx\n\t"
				"notl %%edx\n\t"
				"movl %%eax,0(%%edi)\n\t"
				"movl %%eax,4(%%edi)\n\t"
				"movl %%eax,8(%%edi)\n\t"
				"movl %%eax,12(%%edi)\n\t"
				"movl %%edx,16(%%edi)\n\t"
				"movl %%edx,20(%%edi)\n\t"
				"movl %%eax,24(%%edi)\n\t"
				"movl %%eax,28(%%edi)\n\t"
				"movl %%eax,32(%%edi)\n\t"
				"movl %%eax,36(%%edi)\n\t"
				"movl %%edx,40(%%edi)\n\t"
				"movl %%edx,44(%%edi)\n\t"
				"movl %%eax,48(%%edi)\n\t"
				"movl %%eax,52(%%edi)\n\t"
				"movl %%edx,56(%%edi)\n\t"
				"movl %%edx,60(%%edi)\n\t"
				"rcll $1, %%eax\n\t"
				"leal 64(%%edi), %%edi\n\t"
				"decl %%ecx\n\t"
				"jnz  L100\n\t"
				: "=D" (p)
				: "D" (p), "c" (len), "a" (1)
				: "edx"
			);
			do_tick();
			BAILR
		} while (!done);
	}

	/* Now move the data around
	 * First move the data up half of the segment size we are testing
	 * Then move the data to the original location + 32 bytes
	 */
	for (j=0; j<segs; j++) {
		start = (ulong)v->map[j].start;
#ifdef USB_WAR
		/* We can't do the block move test on low memory beacuase
		 * BIOS USB support clobbers location 0x410 and 0x4e0
		 */
		if (start < 0x4f0) {
			start = 0x4f0;
		}
#endif
		end = (ulong)v->map[j].end;
		pe = start;
		p = start;
		done = 0;
		do {
			/* Check for overflow */
			if (pe + SPINSZ*4 > pe) {
				pe += SPINSZ*4;
			} else {
				pe = end;
			}
			if (pe >= end) {
				pe = end;
				done++;
			}
			if (p == pe ) {
				break;