add 3DNOW stuff from Miroslav

src/libFLAC/cpu.c

@@ -27,6 +27,10 @@ const unsigned FLAC__CPUINFO_IA32_CPUID_FXSR = 0x01000000;
 const unsigned FLAC__CPUINFO_IA32_CPUID_SSE = 0x02000000;
 const unsigned FLAC__CPUINFO_IA32_CPUID_SSE2 = 0x04000000;
 
+const unsigned FLAC__CPUINFO_IA32_CPUID_EXTENDED_AMD_3DNOW = 0x80000000;
+const unsigned FLAC__CPUINFO_IA32_CPUID_EXTENDED_AMD_EXT3DNOW = 0x40000000;
+const unsigned FLAC__CPUINFO_IA32_CPUID_EXTENDED_AMD_EXTMMX = 0x00400000;
+
 
 void FLAC__cpu_info(FLAC__CPUInfo *info)
 {
@@ -41,6 +45,11 @@ void FLAC__cpu_info(FLAC__CPUInfo *info)
 		info->data.ia32.fxsr = (cpuid & FLAC__CPUINFO_IA32_CPUID_FXSR)? true : false;
 		info->data.ia32.sse = (cpuid & FLAC__CPUINFO_IA32_CPUID_SSE)? true : false; /* @@@ also need to check for operating system support */
 		info->data.ia32.sse2 = (cpuid & FLAC__CPUINFO_IA32_CPUID_SSE2)? true : false; /* @@@ also need to check for operating system support */
+
+		cpuid = FLAC__cpu_info_extended_amd_asm_ia32();
+		info->data.ia32._3dnow = (cpuid & FLAC__CPUINFO_IA32_CPUID_EXTENDED_AMD_3DNOW)? true : false;
+		info->data.ia32.ext3dnow = (cpuid & FLAC__CPUINFO_IA32_CPUID_EXTENDED_AMD_EXT3DNOW)? true : false;
+		info->data.ia32.extmmx = (cpuid & FLAC__CPUINFO_IA32_CPUID_EXTENDED_AMD_EXTMMX)? true : false;
 	}
 #else
 	info->use_asm = false;

src/libFLAC/ia32/cpu_asm.nasm

@@ -21,20 +21,18 @@
 	data_section
 
 cglobal FLAC__cpu_info_asm_ia32
+cglobal FLAC__cpu_info_extended_amd_asm_ia32
 
 	code_section
 
 ; **********************************************************************
 ;
-	ALIGN 16
-cident FLAC__cpu_info_asm_ia32
-
-	push	ebx
 
+have_cpuid:
 	pushfd
 	pop	eax
 	mov	edx, eax
-	xor	eax, 00200000h
+	xor	eax, 0x00200000
 	push	eax
 	popfd
 	pushfd
@@ -42,12 +40,43 @@ cident FLAC__cpu_info_asm_ia32
 	cmp	eax, edx
 	jz	.no_cpuid
 	mov	eax, 1
+	jmp	.end
+.no_cpuid:
+	xor	eax, eax
+.end:
+	ret
+
+cident FLAC__cpu_info_asm_ia32
+	push	ebx
+	call	have_cpuid
+	test	eax, eax
+	jz	.no_cpuid
+	mov	eax, 1
 	cpuid
 	mov	eax, edx
-	jmp	short .end
+	jmp	.end
 .no_cpuid:
-	xor	eax, eax			; return 0
-.end:
+	xor	eax, eax
+.end
+	pop	ebx
+	ret
+
+cident FLAC__cpu_info_extended_amd_asm_ia32
+	push	ebx
+	call	have_cpuid
+	test	eax, eax
+	jz	.no_cpuid
+	mov	eax, 0x80000000
+	cpuid
+	cmp	eax, 0x80000001
+	jb	.no_cpuid
+	mov	eax, 0x80000001
+	cpuid
+	mov	eax, edx
+	jmp	.end
+.no_cpuid
+	xor	eax, eax
+.end
 	pop	ebx
 	ret
 

src/libFLAC/ia32/lpc_asm.nasm

@@ -24,6 +24,7 @@ cglobal FLAC__lpc_compute_autocorrelation_asm_ia32
 cglobal FLAC__lpc_compute_autocorrelation_asm_ia32_sse_lag_4
 cglobal FLAC__lpc_compute_autocorrelation_asm_ia32_sse_lag_8
 cglobal FLAC__lpc_compute_autocorrelation_asm_ia32_sse_lag_12
+cglobal FLAC__lpc_compute_autocorrelation_asm_ia32_3dnow
 cglobal FLAC__lpc_compute_residual_from_qlp_coefficients_asm_ia32
 cglobal FLAC__lpc_compute_residual_from_qlp_coefficients_asm_ia32_mmx
 cglobal FLAC__lpc_restore_signal_asm_ia32
@@ -592,6 +593,124 @@ cident FLAC__lpc_compute_autocorrelation_asm_ia32_sse_lag_12
 .end:
 	ret
 
+	align 16
+cident FLAC__lpc_compute_autocorrelation_asm_ia32_3dnow
+	;[ebp + 32] autoc
+	;[ebp + 28] lag
+	;[ebp + 24] data_len
+	;[ebp + 20] data
+
+	push	ebp
+	push	ebx
+	push	esi
+	push	edi
+	mov	ebp, esp
+
+	mov	esi, [ebp + 20]
+	mov	edi, [ebp + 24]
+	mov	edx, [ebp + 28]
+	mov	eax, edx
+	neg	eax
+	and	esp, byte -8
+	lea	esp, [esp + 4 * eax]
+	mov	ecx, edx
+	xor	eax, eax
+.loop0:
+	dec	ecx
+	mov	[esp + 4 * ecx], eax
+	jnz	short .loop0
+
+	mov	eax, edi
+	sub	eax, edx
+	mov	ebx, edx
+	and	ebx, byte 1
+	sub	eax, ebx
+	lea	ecx, [esi + 4 * eax - 12]
+	cmp	esi, ecx
+	mov	eax, esi
+	ja	short .loop2_pre
+	align	16		;8 nops
+.loop1_i:
+	movd	mm0, [eax]
+	movd	mm2, [eax + 4]
+	movd	mm4, [eax + 8]
+	movd	mm6, [eax + 12]
+	mov	ebx, edx
+	punpckldq	mm0, mm0
+	punpckldq	mm2, mm2
+	punpckldq	mm4, mm4
+	punpckldq	mm6, mm6
+	align	16		;3 nops
+.loop1_j:
+	sub	ebx, byte 2
+	movd	mm1, [eax + 4 * ebx]
+	movd	mm3, [eax + 4 * ebx + 4]
+	movd	mm5, [eax + 4 * ebx + 8]
+	movd	mm7, [eax + 4 * ebx + 12]
+	punpckldq	mm1, mm3
+	punpckldq	mm3, mm5
+	pfmul	mm1, mm0
+	punpckldq	mm5, mm7
+	pfmul	mm3, mm2
+	punpckldq	mm7, [eax + 4 * ebx + 16]
+	pfmul	mm5, mm4
+	pfmul	mm7, mm6
+	pfadd	mm1, mm3
+	movq	mm3, [esp + 4 * ebx]
+	pfadd	mm5, mm7
+	pfadd	mm1, mm5
+	pfadd	mm3, mm1
+	movq	[esp + 4 * ebx], mm3
+	jg	short .loop1_j
+
+	add	eax, byte 16
+	cmp	eax, ecx
+	jb	short .loop1_i
+
+.loop2_pre:
+	mov	ebx, eax
+	sub	eax, esi
+	shr	eax, 2
+	lea	ecx, [esi + 4 * edi]
+	mov	esi, ebx
+.loop2_i:
+	movd	mm0, [esi]
+	mov	ebx, edi
+	sub	ebx, eax
+	cmp	ebx, edx
+	jbe	short .loop2_j
+	mov	ebx, edx
+.loop2_j:
+	dec	ebx
+	movd	mm1, [esi + 4 * ebx]
+	pfmul	mm1, mm0
+	movd	mm2, [esp + 4 * ebx]
+	pfadd	mm1, mm2
+	movd	[esp + 4 * ebx], mm1
+
+	jnz	short .loop2_j
+
+	add	esi, byte 4
+	inc	eax
+	cmp	esi, ecx
+	jnz	short .loop2_i
+
+	mov	edi, [ebp + 32]
+.loop3:
+	dec	edx
+	mov	eax, [esp + 4 * edx]
+	mov	[edi + 4 * edx], eax
+	jnz	short .loop3
+
+	femms
+
+	mov	esp, ebp
+	pop	edi
+	pop	esi
+	pop	ebx
+	pop	ebp
+	ret
+
 ;void FLAC__lpc_compute_residual_from_qlp_coefficients(const FLAC__int32 data[], unsigned data_len, const FLAC__int32 qlp_coeff[], unsigned order, int lp_quantization, FLAC__int32 residual[])
 ;
 ;	for(i = 0; i < data_len; i++) {

src/libFLAC/include/private/cpu.h

@@ -33,13 +33,21 @@ typedef struct {
 	FLAC__bool fxsr;
 	FLAC__bool sse;
 	FLAC__bool sse2;
+	FLAC__bool _3dnow;
+	FLAC__bool ext3dnow;
+	FLAC__bool extmmx;
 } FLAC__CPUInfo_IA32;
+
 extern const unsigned FLAC__CPUINFO_IA32_CPUID_CMOV;
 extern const unsigned FLAC__CPUINFO_IA32_CPUID_MMX;
 extern const unsigned FLAC__CPUINFO_IA32_CPUID_FXSR;
 extern const unsigned FLAC__CPUINFO_IA32_CPUID_SSE;
 extern const unsigned FLAC__CPUINFO_IA32_CPUID_SSE2;
 
+extern const unsigned FLAC__CPUINFO_IA32_CPUID_EXTENDED_AMD_3DNOW;
+extern const unsigned FLAC__CPUINFO_IA32_CPUID_EXTENDED_AMD_EXT3DNOW;
+extern const unsigned FLAC__CPUINFO_IA32_CPUID_EXTENDED_AMD_EXTMMX;
+
 typedef struct {
 	FLAC__bool use_asm;
 	FLAC__CPUInfo_Type type;
@@ -54,6 +62,7 @@ void FLAC__cpu_info(FLAC__CPUInfo *info);
 #ifdef FLAC__CPU_IA32
 #ifdef FLAC__HAS_NASM
 unsigned FLAC__cpu_info_asm_ia32();
+unsigned FLAC__cpu_info_extended_amd_asm_ia32();
 #endif
 #endif
 #endif

src/libFLAC/include/private/lpc.h

@@ -44,6 +44,7 @@ void FLAC__lpc_compute_autocorrelation_asm_ia32(const FLAC__real data[], unsigne
 void FLAC__lpc_compute_autocorrelation_asm_ia32_sse_lag_4(const FLAC__real data[], unsigned data_len, unsigned lag, FLAC__real autoc[]);
 void FLAC__lpc_compute_autocorrelation_asm_ia32_sse_lag_8(const FLAC__real data[], unsigned data_len, unsigned lag, FLAC__real autoc[]);
 void FLAC__lpc_compute_autocorrelation_asm_ia32_sse_lag_12(const FLAC__real data[], unsigned data_len, unsigned lag, FLAC__real autoc[]);
+void FLAC__lpc_compute_autocorrelation_asm_ia32_3dnow(const FLAC__real data[], unsigned data_len, unsigned lag, FLAC__real autoc[]);
 #endif
 #endif
 #endif

src/libFLAC/stream_encoder.c

@@ -375,7 +375,7 @@ FLAC__StreamEncoderState FLAC__stream_encoder_init(FLAC__StreamEncoder *encoder)
 #ifdef FLAC__CPU_IA32
 	FLAC__ASSERT(encoder->private->cpuinfo.type == FLAC__CPUINFO_TYPE_IA32);
 #ifdef FLAC__HAS_NASM
-	if(0 && encoder->private->cpuinfo.data.ia32.sse) { /* SSE version lacks necessary resolution, plus SSE flag doesn't check for OS support */
+	if(0 && encoder->private->cpuinfo.data.ia32.sse) { /*@@@ SSE version lacks necessary resolution, plus SSE flag doesn't check for OS support */
 		if(encoder->protected->max_lpc_order < 4)
 			encoder->private->local_lpc_compute_autocorrelation = FLAC__lpc_compute_autocorrelation_asm_ia32_sse_lag_4;
 		else if(encoder->protected->max_lpc_order < 8)
@@ -385,6 +385,8 @@ FLAC__StreamEncoderState FLAC__stream_encoder_init(FLAC__StreamEncoder *encoder)
 		else
 			encoder->private->local_lpc_compute_autocorrelation = FLAC__lpc_compute_autocorrelation_asm_ia32;
 	}
+	else if(0 && encoder->private->cpuinfo.data.ia32._3dnow) /*@@@ turn back on in first beta after 1.0 */
+		encoder->private->local_lpc_compute_autocorrelation = FLAC__lpc_compute_autocorrelation_asm_ia32_3dnow;
 	else
 		encoder->private->local_lpc_compute_autocorrelation = FLAC__lpc_compute_autocorrelation_asm_ia32;
 	if(encoder->private->cpuinfo.data.ia32.mmx && encoder->private->cpuinfo.data.ia32.cmov)