Attachment #64200 for bug #85220




			"pushfl;"
			"popl	%%eax;"
			// Save current value.
			"movl	%%eax,%%ecx;"
			// Toggle bit 21.
			"xorl	$0x200000, %%eax;"
			// Load EAX into EFLAGS.

			"pushfl;"
			"popl	%%eax;"
			// Did bit 21 change?
			"xor	%%ecx, %%eax;"
			"andl	$0x200000, %%eax;"
			: "=a" (hasCPUID) // 0
			: // no input
			: "ecx" // use ecx not ebx to avoid PIC abi clobber
			);
		if (hasCPUID) {
			// Which CPUID calls are supported?
			unsigned highest;
			asm (
				"pushl	%%ebx;"
				"cpuid;"
				"popl	%%ebx;"
				: "=a" (highest) // 0
				: "0" (0) // 1: function
				: "ecx", "edx" // ebx save/restored to avoid PIC abi clobber
				);
			if (highest >= 1) {
				// Get features flags.
				unsigned features;
				asm (
					"pushl	%%ebx;"
					"cpuid;"
					"pushl	%%ebx;"
					: "=d" (features) // 0
					: "a" (1) // 1: function
					: "ecx" // ebx save/restored to avoid PIC abi clobber
					);
				PRT_DEBUG("CPU flags: " << std::hex << std::setw(8) << std::setfill('0') << std::uppercase << features);
				mmxFlag = features & 0x800000;




			"paddw	%%mm4, %%mm0;"
			"paddw	%%mm5, %%mm1;"
			
			"movd	%%eax, %%mm6;" // save eax
			"pextrw	$0, %%mm0, %%eax;"
			"movw	(%2,%%eax,2), %%ax;"
			"pinsrw	$0, %%eax, %%mm0;"

			"pextrw	$3, %%mm1, %%eax;"
			"movw	(%2,%%eax,2), %%ax;"
			"pinsrw	$3, %%eax, %%mm1;"
			"movd	%%mm6, %%eax;" // restore eax
			
			"movntq	%%mm0,   (%3,%%ecx,2);"
			"movntq	%%mm1,  8(%3,%%ecx,2);"

			, "r" (dst)   // 3
			, "m" (mask)   // 4
			: "mm0", "mm1", "mm2", "mm3", "mm4", "mm5", "mm7"
			, "ecx" // eax save/restored so available for 'r'
		);
		return;
	}

Return to bug 85220

Lines 23-29 Link Here

(-)src/HostCPU.cc.orig (-5 / +9 lines)
23	"pushfl;"	23	"pushfl;"
24	"popl %%eax;"	24	"popl %%eax;"
25	// Save current value.	25	// Save current value.
26	"movl %%eax,%%ebx;"	26	"movl %%eax,%%ecx;"
27	// Toggle bit 21.	27	// Toggle bit 21.
28	"xorl $0x200000, %%eax;"	28	"xorl $0x200000, %%eax;"
29	// Load EAX into EFLAGS.	29	// Load EAX into EFLAGS.
Lines 33-61 Link Here
33	"pushfl;"	33	"pushfl;"
34	"popl %%eax;"	34	"popl %%eax;"
35	// Did bit 21 change?	35	// Did bit 21 change?
36	"xor %%ebx, %%eax;"	36	"xor %%ecx, %%eax;"
37	"andl $0x200000, %%eax;"	37	"andl $0x200000, %%eax;"
38	: "=a" (hasCPUID) // 0	38	: "=a" (hasCPUID) // 0
39	: // no input	39	: // no input
40	: "ebx"	40	: "ecx" // use ecx not ebx to avoid PIC abi clobber
41	);	41	);
42	if (hasCPUID) {	42	if (hasCPUID) {
43	// Which CPUID calls are supported?	43	// Which CPUID calls are supported?
44	unsigned highest;	44	unsigned highest;
45	asm (	45	asm (
		46	"pushl %%ebx;"
46	"cpuid;"	47	"cpuid;"
		48	"popl %%ebx;"
47	: "=a" (highest) // 0	49	: "=a" (highest) // 0
48	: "0" (0) // 1: function	50	: "0" (0) // 1: function
49	: "ebx", "ecx", "edx"	51	: "ecx", "edx" // ebx save/restored to avoid PIC abi clobber
50	);	52	);
51	if (highest >= 1) {	53	if (highest >= 1) {
52	// Get features flags.	54	// Get features flags.
53	unsigned features;	55	unsigned features;
54	asm (	56	asm (
		57	"pushl %%ebx;"
55	"cpuid;"	58	"cpuid;"
		59	"pushl %%ebx;"
56	: "=d" (features) // 0	60	: "=d" (features) // 0
57	: "a" (1) // 1: function	61	: "a" (1) // 1: function
58	: "ebx", "ecx"	62	: "ecx" // ebx save/restored to avoid PIC abi clobber
59	);	63	);
60	PRT_DEBUG("CPU flags: " << std::hex << std::setw(8) << std::setfill('0') << std::uppercase << features);	64	PRT_DEBUG("CPU flags: " << std::hex << std::setw(8) << std::setfill('0') << std::uppercase << features);
61	mmxFlag = features & 0x800000;	65	mmxFlag = features & 0x800000;

Lines 684-689 Link Here

(-)src/video/SimpleScaler.cc.orig (-1 / +3 lines)
684	"paddw %%mm4, %%mm0;"	684	"paddw %%mm4, %%mm0;"
685	"paddw %%mm5, %%mm1;"	685	"paddw %%mm5, %%mm1;"
686		686
		687	"movd %%eax, %%mm6;" // save eax
687	"pextrw $0, %%mm0, %%eax;"	688	"pextrw $0, %%mm0, %%eax;"
688	"movw (%2,%%eax,2), %%ax;"	689	"movw (%2,%%eax,2), %%ax;"
689	"pinsrw $0, %%eax, %%mm0;"	690	"pinsrw $0, %%eax, %%mm0;"
Lines 711-716 Link Here
711	"pextrw $3, %%mm1, %%eax;"	712	"pextrw $3, %%mm1, %%eax;"
712	"movw (%2,%%eax,2), %%ax;"	713	"movw (%2,%%eax,2), %%ax;"
713	"pinsrw $3, %%eax, %%mm1;"	714	"pinsrw $3, %%eax, %%mm1;"
		715	"movd %%mm6, %%eax;" // restore eax
714		716
715	"movntq %%mm0, (%3,%%ecx,2);"	717	"movntq %%mm0, (%3,%%ecx,2);"
716	"movntq %%mm1, 8(%3,%%ecx,2);"	718	"movntq %%mm1, 8(%3,%%ecx,2);"
Lines 726-732 Link Here
726	, "r" (dst) // 3	728	, "r" (dst) // 3
727	, "m" (mask) // 4	729	, "m" (mask) // 4
728	: "mm0", "mm1", "mm2", "mm3", "mm4", "mm5", "mm7"	730	: "mm0", "mm1", "mm2", "mm3", "mm4", "mm5", "mm7"
729	, "eax", "ecx"	731	, "ecx" // eax save/restored so available for 'r'
730	);	732	);
731	return;	733	return;
732	}	734	}